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;
59 Statistic NumCombined ("instcombine", "Number of insts combined");
60 Statistic NumConstProp("instcombine", "Number of constant folds");
61 Statistic NumDeadInst ("instcombine", "Number of dead inst eliminated");
62 Statistic NumDeadStore("instcombine", "Number of dead stores eliminated");
63 Statistic NumSunkInst ("instcombine", "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(Value *V, const Type *Ty, Instruction &Pos) {
209 if (V->getType() == Ty) return V;
211 if (Constant *CV = dyn_cast<Constant>(V))
212 return ConstantExpr::getCast(CV, Ty);
214 Instruction *C = CastInst::createInferredCast(V, Ty, V->getName(), &Pos);
215 WorkList.push_back(C);
219 // ReplaceInstUsesWith - This method is to be used when an instruction is
220 // found to be dead, replacable with another preexisting expression. Here
221 // we add all uses of I to the worklist, replace all uses of I with the new
222 // value, then return I, so that the inst combiner will know that I was
225 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
226 AddUsersToWorkList(I); // Add all modified instrs to worklist
228 I.replaceAllUsesWith(V);
231 // If we are replacing the instruction with itself, this must be in a
232 // segment of unreachable code, so just clobber the instruction.
233 I.replaceAllUsesWith(UndefValue::get(I.getType()));
238 // UpdateValueUsesWith - This method is to be used when an value is
239 // found to be replacable with another preexisting expression or was
240 // updated. Here we add all uses of I to the worklist, replace all uses of
241 // I with the new value (unless the instruction was just updated), then
242 // return true, so that the inst combiner will know that I was modified.
244 bool UpdateValueUsesWith(Value *Old, Value *New) {
245 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
247 Old->replaceAllUsesWith(New);
248 if (Instruction *I = dyn_cast<Instruction>(Old))
249 WorkList.push_back(I);
250 if (Instruction *I = dyn_cast<Instruction>(New))
251 WorkList.push_back(I);
255 // EraseInstFromFunction - When dealing with an instruction that has side
256 // effects or produces a void value, we can't rely on DCE to delete the
257 // instruction. Instead, visit methods should return the value returned by
259 Instruction *EraseInstFromFunction(Instruction &I) {
260 assert(I.use_empty() && "Cannot erase instruction that is used!");
261 AddUsesToWorkList(I);
262 removeFromWorkList(&I);
264 return 0; // Don't do anything with FI
268 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
269 /// InsertBefore instruction. This is specialized a bit to avoid inserting
270 /// casts that are known to not do anything...
272 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
273 Instruction *InsertBefore);
275 // SimplifyCommutative - This performs a few simplifications for commutative
277 bool SimplifyCommutative(BinaryOperator &I);
279 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
280 uint64_t &KnownZero, uint64_t &KnownOne,
283 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
284 uint64_t &UndefElts, unsigned Depth = 0);
286 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
287 // PHI node as operand #0, see if we can fold the instruction into the PHI
288 // (which is only possible if all operands to the PHI are constants).
289 Instruction *FoldOpIntoPhi(Instruction &I);
291 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
292 // operator and they all are only used by the PHI, PHI together their
293 // inputs, and do the operation once, to the result of the PHI.
294 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
295 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
298 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
299 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
301 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
302 bool isSub, Instruction &I);
303 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
304 bool Inside, Instruction &IB);
305 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
306 Instruction *MatchBSwap(BinaryOperator &I);
308 Value *EvaluateInDifferentType(Value *V, const Type *Ty);
311 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
314 // getComplexity: Assign a complexity or rank value to LLVM Values...
315 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
316 static unsigned getComplexity(Value *V) {
317 if (isa<Instruction>(V)) {
318 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
322 if (isa<Argument>(V)) return 3;
323 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
326 // isOnlyUse - Return true if this instruction will be deleted if we stop using
328 static bool isOnlyUse(Value *V) {
329 return V->hasOneUse() || isa<Constant>(V);
332 // getPromotedType - Return the specified type promoted as it would be to pass
333 // though a va_arg area...
334 static const Type *getPromotedType(const Type *Ty) {
335 switch (Ty->getTypeID()) {
336 case Type::SByteTyID:
337 case Type::ShortTyID: return Type::IntTy;
338 case Type::UByteTyID:
339 case Type::UShortTyID: return Type::UIntTy;
340 case Type::FloatTyID: return Type::DoubleTy;
345 /// getBitCastOperand - If the specified operand is a CastInst or a constant
346 /// expression bitcast, return the operand value, otherwise return null.
347 static Value *getBitCastOperand(Value *V) {
348 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
349 return I->getOperand(0);
350 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
351 if (CE->getOpcode() == Instruction::BitCast)
352 return CE->getOperand(0);
356 /// This function is a wrapper around CastInst::isEliminableCastPair. It
357 /// simply extracts arguments and returns what that function returns.
358 /// @Determine if it is valid to eliminate a Convert pair
359 static Instruction::CastOps
360 isEliminableCastPair(
361 const CastInst *CI, ///< The first cast instruction
362 unsigned opcode, ///< The opcode of the second cast instruction
363 const Type *DstTy, ///< The target type for the second cast instruction
364 TargetData *TD ///< The target data for pointer size
367 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
368 const Type *MidTy = CI->getType(); // B from above
370 // Get the opcodes of the two Cast instructions
371 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
372 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
374 return Instruction::CastOps(
375 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
376 DstTy, TD->getIntPtrType()));
379 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
380 /// in any code being generated. It does not require codegen if V is simple
381 /// enough or if the cast can be folded into other casts.
382 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
383 if (V->getType() == Ty || isa<Constant>(V)) return false;
385 // If this is a noop cast, it isn't real codegen.
386 if (V->getType()->canLosslesslyBitCastTo(Ty))
389 // If this is another cast that can be eliminated, it isn't codegen either.
390 if (const CastInst *CI = dyn_cast<CastInst>(V))
391 if (isEliminableCastPair(CI, CastInst::getCastOpcode(
392 V, V->getType()->isSigned(), Ty, Ty->isSigned()), Ty, TD))
397 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
398 /// InsertBefore instruction. This is specialized a bit to avoid inserting
399 /// casts that are known to not do anything...
401 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
402 Instruction *InsertBefore) {
403 if (V->getType() == DestTy) return V;
404 if (Constant *C = dyn_cast<Constant>(V))
405 return ConstantExpr::getCast(C, DestTy);
407 return InsertCastBefore(V, DestTy, *InsertBefore);
410 // SimplifyCommutative - This performs a few simplifications for commutative
413 // 1. Order operands such that they are listed from right (least complex) to
414 // left (most complex). This puts constants before unary operators before
417 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
418 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
420 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
421 bool Changed = false;
422 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
423 Changed = !I.swapOperands();
425 if (!I.isAssociative()) return Changed;
426 Instruction::BinaryOps Opcode = I.getOpcode();
427 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
428 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
429 if (isa<Constant>(I.getOperand(1))) {
430 Constant *Folded = ConstantExpr::get(I.getOpcode(),
431 cast<Constant>(I.getOperand(1)),
432 cast<Constant>(Op->getOperand(1)));
433 I.setOperand(0, Op->getOperand(0));
434 I.setOperand(1, Folded);
436 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
437 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
438 isOnlyUse(Op) && isOnlyUse(Op1)) {
439 Constant *C1 = cast<Constant>(Op->getOperand(1));
440 Constant *C2 = cast<Constant>(Op1->getOperand(1));
442 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
443 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
444 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
447 WorkList.push_back(New);
448 I.setOperand(0, New);
449 I.setOperand(1, Folded);
456 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
457 // if the LHS is a constant zero (which is the 'negate' form).
459 static inline Value *dyn_castNegVal(Value *V) {
460 if (BinaryOperator::isNeg(V))
461 return BinaryOperator::getNegArgument(V);
463 // Constants can be considered to be negated values if they can be folded.
464 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
465 return ConstantExpr::getNeg(C);
469 static inline Value *dyn_castNotVal(Value *V) {
470 if (BinaryOperator::isNot(V))
471 return BinaryOperator::getNotArgument(V);
473 // Constants can be considered to be not'ed values...
474 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
475 return ConstantExpr::getNot(C);
479 // dyn_castFoldableMul - If this value is a multiply that can be folded into
480 // other computations (because it has a constant operand), return the
481 // non-constant operand of the multiply, and set CST to point to the multiplier.
482 // Otherwise, return null.
484 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
485 if (V->hasOneUse() && V->getType()->isInteger())
486 if (Instruction *I = dyn_cast<Instruction>(V)) {
487 if (I->getOpcode() == Instruction::Mul)
488 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
489 return I->getOperand(0);
490 if (I->getOpcode() == Instruction::Shl)
491 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
492 // The multiplier is really 1 << CST.
493 Constant *One = ConstantInt::get(V->getType(), 1);
494 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
495 return I->getOperand(0);
501 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
502 /// expression, return it.
503 static User *dyn_castGetElementPtr(Value *V) {
504 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
505 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
506 if (CE->getOpcode() == Instruction::GetElementPtr)
507 return cast<User>(V);
511 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
512 static ConstantInt *AddOne(ConstantInt *C) {
513 return cast<ConstantInt>(ConstantExpr::getAdd(C,
514 ConstantInt::get(C->getType(), 1)));
516 static ConstantInt *SubOne(ConstantInt *C) {
517 return cast<ConstantInt>(ConstantExpr::getSub(C,
518 ConstantInt::get(C->getType(), 1)));
521 /// GetConstantInType - Return a ConstantInt with the specified type and value.
523 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
524 if (Ty->isUnsigned())
525 return ConstantInt::get(Ty, Val);
526 else if (Ty->getTypeID() == Type::BoolTyID)
527 return ConstantBool::get(Val);
529 SVal <<= 64-Ty->getPrimitiveSizeInBits();
530 SVal >>= 64-Ty->getPrimitiveSizeInBits();
531 return ConstantInt::get(Ty, SVal);
535 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
536 /// known to be either zero or one and return them in the KnownZero/KnownOne
537 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
539 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
540 uint64_t &KnownOne, unsigned Depth = 0) {
541 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
542 // we cannot optimize based on the assumption that it is zero without changing
543 // it to be an explicit zero. If we don't change it to zero, other code could
544 // optimized based on the contradictory assumption that it is non-zero.
545 // Because instcombine aggressively folds operations with undef args anyway,
546 // this won't lose us code quality.
547 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
548 // We know all of the bits for a constant!
549 KnownOne = CI->getZExtValue() & Mask;
550 KnownZero = ~KnownOne & Mask;
554 KnownZero = KnownOne = 0; // Don't know anything.
555 if (Depth == 6 || Mask == 0)
556 return; // Limit search depth.
558 uint64_t KnownZero2, KnownOne2;
559 Instruction *I = dyn_cast<Instruction>(V);
562 Mask &= V->getType()->getIntegralTypeMask();
564 switch (I->getOpcode()) {
565 case Instruction::And:
566 // If either the LHS or the RHS are Zero, the result is zero.
567 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
569 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
570 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
571 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
573 // Output known-1 bits are only known if set in both the LHS & RHS.
574 KnownOne &= KnownOne2;
575 // Output known-0 are known to be clear if zero in either the LHS | RHS.
576 KnownZero |= KnownZero2;
578 case Instruction::Or:
579 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
581 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
582 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
583 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
585 // Output known-0 bits are only known if clear in both the LHS & RHS.
586 KnownZero &= KnownZero2;
587 // Output known-1 are known to be set if set in either the LHS | RHS.
588 KnownOne |= KnownOne2;
590 case Instruction::Xor: {
591 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
592 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
593 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
594 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
596 // Output known-0 bits are known if clear or set in both the LHS & RHS.
597 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
598 // Output known-1 are known to be set if set in only one of the LHS, RHS.
599 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
600 KnownZero = KnownZeroOut;
603 case Instruction::Select:
604 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
605 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
606 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
607 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
609 // Only known if known in both the LHS and RHS.
610 KnownOne &= KnownOne2;
611 KnownZero &= KnownZero2;
613 case Instruction::FPTrunc:
614 case Instruction::FPExt:
615 case Instruction::FPToUI:
616 case Instruction::FPToSI:
617 case Instruction::SIToFP:
618 case Instruction::PtrToInt:
619 case Instruction::UIToFP:
620 case Instruction::IntToPtr:
621 return; // Can't work with floating point or pointers
622 case Instruction::Trunc:
623 // All these have integer operands
624 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
626 case Instruction::BitCast: {
627 const Type *SrcTy = I->getOperand(0)->getType();
628 if (SrcTy->isIntegral()) {
629 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
634 case Instruction::ZExt: {
635 // Compute the bits in the result that are not present in the input.
636 const Type *SrcTy = I->getOperand(0)->getType();
637 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
638 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
640 Mask &= SrcTy->getIntegralTypeMask();
641 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
642 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
643 // The top bits are known to be zero.
644 KnownZero |= NewBits;
647 case Instruction::SExt: {
648 // Compute the bits in the result that are not present in the input.
649 const Type *SrcTy = I->getOperand(0)->getType();
650 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
651 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
653 Mask &= SrcTy->getIntegralTypeMask();
654 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
655 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
657 // If the sign bit of the input is known set or clear, then we know the
658 // top bits of the result.
659 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
660 if (KnownZero & InSignBit) { // Input sign bit known zero
661 KnownZero |= NewBits;
662 KnownOne &= ~NewBits;
663 } else if (KnownOne & InSignBit) { // Input sign bit known set
665 KnownZero &= ~NewBits;
666 } else { // Input sign bit unknown
667 KnownZero &= ~NewBits;
668 KnownOne &= ~NewBits;
672 case Instruction::Shl:
673 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
674 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
675 uint64_t ShiftAmt = SA->getZExtValue();
677 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
678 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
679 KnownZero <<= ShiftAmt;
680 KnownOne <<= ShiftAmt;
681 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
685 case Instruction::LShr:
686 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
687 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
688 // Compute the new bits that are at the top now.
689 uint64_t ShiftAmt = SA->getZExtValue();
690 uint64_t HighBits = (1ULL << ShiftAmt)-1;
691 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
693 // Unsigned shift right.
695 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
696 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
697 KnownZero >>= ShiftAmt;
698 KnownOne >>= ShiftAmt;
699 KnownZero |= HighBits; // high bits known zero.
703 case Instruction::AShr:
704 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
705 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
706 // Compute the new bits that are at the top now.
707 uint64_t ShiftAmt = SA->getZExtValue();
708 uint64_t HighBits = (1ULL << ShiftAmt)-1;
709 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
711 // Signed shift right.
713 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
714 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
715 KnownZero >>= ShiftAmt;
716 KnownOne >>= ShiftAmt;
718 // Handle the sign bits.
719 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
720 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
722 if (KnownZero & SignBit) { // New bits are known zero.
723 KnownZero |= HighBits;
724 } else if (KnownOne & SignBit) { // New bits are known one.
725 KnownOne |= HighBits;
733 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
734 /// this predicate to simplify operations downstream. Mask is known to be zero
735 /// for bits that V cannot have.
736 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
737 uint64_t KnownZero, KnownOne;
738 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
739 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
740 return (KnownZero & Mask) == Mask;
743 /// ShrinkDemandedConstant - Check to see if the specified operand of the
744 /// specified instruction is a constant integer. If so, check to see if there
745 /// are any bits set in the constant that are not demanded. If so, shrink the
746 /// constant and return true.
747 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
749 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
750 if (!OpC) return false;
752 // If there are no bits set that aren't demanded, nothing to do.
753 if ((~Demanded & OpC->getZExtValue()) == 0)
756 // This is producing any bits that are not needed, shrink the RHS.
757 uint64_t Val = Demanded & OpC->getZExtValue();
758 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
762 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
763 // set of known zero and one bits, compute the maximum and minimum values that
764 // could have the specified known zero and known one bits, returning them in
766 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
769 int64_t &Min, int64_t &Max) {
770 uint64_t TypeBits = Ty->getIntegralTypeMask();
771 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
773 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
775 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
776 // bit if it is unknown.
778 Max = KnownOne|UnknownBits;
780 if (SignBit & UnknownBits) { // Sign bit is unknown
785 // Sign extend the min/max values.
786 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
787 Min = (Min << ShAmt) >> ShAmt;
788 Max = (Max << ShAmt) >> ShAmt;
791 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
792 // a set of known zero and one bits, compute the maximum and minimum values that
793 // could have the specified known zero and known one bits, returning them in
795 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
800 uint64_t TypeBits = Ty->getIntegralTypeMask();
801 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
803 // The minimum value is when the unknown bits are all zeros.
805 // The maximum value is when the unknown bits are all ones.
806 Max = KnownOne|UnknownBits;
810 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
811 /// DemandedMask bits of the result of V are ever used downstream. If we can
812 /// use this information to simplify V, do so and return true. Otherwise,
813 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
814 /// the expression (used to simplify the caller). The KnownZero/One bits may
815 /// only be accurate for those bits in the DemandedMask.
816 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
817 uint64_t &KnownZero, uint64_t &KnownOne,
819 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
820 // We know all of the bits for a constant!
821 KnownOne = CI->getZExtValue() & DemandedMask;
822 KnownZero = ~KnownOne & DemandedMask;
826 KnownZero = KnownOne = 0;
827 if (!V->hasOneUse()) { // Other users may use these bits.
828 if (Depth != 0) { // Not at the root.
829 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
830 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
833 // If this is the root being simplified, allow it to have multiple uses,
834 // just set the DemandedMask to all bits.
835 DemandedMask = V->getType()->getIntegralTypeMask();
836 } else if (DemandedMask == 0) { // Not demanding any bits from V.
837 if (V != UndefValue::get(V->getType()))
838 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
840 } else if (Depth == 6) { // Limit search depth.
844 Instruction *I = dyn_cast<Instruction>(V);
845 if (!I) return false; // Only analyze instructions.
847 DemandedMask &= V->getType()->getIntegralTypeMask();
849 uint64_t KnownZero2 = 0, KnownOne2 = 0;
850 switch (I->getOpcode()) {
852 case Instruction::And:
853 // If either the LHS or the RHS are Zero, the result is zero.
854 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
855 KnownZero, KnownOne, Depth+1))
857 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
859 // If something is known zero on the RHS, the bits aren't demanded on the
861 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
862 KnownZero2, KnownOne2, Depth+1))
864 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
866 // If all of the demanded bits are known 1 on one side, return the other.
867 // These bits cannot contribute to the result of the 'and'.
868 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
869 return UpdateValueUsesWith(I, I->getOperand(0));
870 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
871 return UpdateValueUsesWith(I, I->getOperand(1));
873 // If all of the demanded bits in the inputs are known zeros, return zero.
874 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
875 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
877 // If the RHS is a constant, see if we can simplify it.
878 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
879 return UpdateValueUsesWith(I, I);
881 // Output known-1 bits are only known if set in both the LHS & RHS.
882 KnownOne &= KnownOne2;
883 // Output known-0 are known to be clear if zero in either the LHS | RHS.
884 KnownZero |= KnownZero2;
886 case Instruction::Or:
887 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
888 KnownZero, KnownOne, Depth+1))
890 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
891 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
892 KnownZero2, KnownOne2, Depth+1))
894 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
896 // If all of the demanded bits are known zero on one side, return the other.
897 // These bits cannot contribute to the result of the 'or'.
898 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
899 return UpdateValueUsesWith(I, I->getOperand(0));
900 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
901 return UpdateValueUsesWith(I, I->getOperand(1));
903 // If all of the potentially set bits on one side are known to be set on
904 // the other side, just use the 'other' side.
905 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
906 (DemandedMask & (~KnownZero)))
907 return UpdateValueUsesWith(I, I->getOperand(0));
908 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
909 (DemandedMask & (~KnownZero2)))
910 return UpdateValueUsesWith(I, I->getOperand(1));
912 // If the RHS is a constant, see if we can simplify it.
913 if (ShrinkDemandedConstant(I, 1, DemandedMask))
914 return UpdateValueUsesWith(I, I);
916 // Output known-0 bits are only known if clear in both the LHS & RHS.
917 KnownZero &= KnownZero2;
918 // Output known-1 are known to be set if set in either the LHS | RHS.
919 KnownOne |= KnownOne2;
921 case Instruction::Xor: {
922 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
923 KnownZero, KnownOne, Depth+1))
925 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
926 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
927 KnownZero2, KnownOne2, Depth+1))
929 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
931 // If all of the demanded bits are known zero on one side, return the other.
932 // These bits cannot contribute to the result of the 'xor'.
933 if ((DemandedMask & KnownZero) == DemandedMask)
934 return UpdateValueUsesWith(I, I->getOperand(0));
935 if ((DemandedMask & KnownZero2) == DemandedMask)
936 return UpdateValueUsesWith(I, I->getOperand(1));
938 // Output known-0 bits are known if clear or set in both the LHS & RHS.
939 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
940 // Output known-1 are known to be set if set in only one of the LHS, RHS.
941 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
943 // If all of the demanded bits are known to be zero on one side or the
944 // other, turn this into an *inclusive* or.
945 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
946 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
948 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
950 InsertNewInstBefore(Or, *I);
951 return UpdateValueUsesWith(I, Or);
954 // If all of the demanded bits on one side are known, and all of the set
955 // bits on that side are also known to be set on the other side, turn this
956 // into an AND, as we know the bits will be cleared.
957 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
958 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
959 if ((KnownOne & KnownOne2) == KnownOne) {
960 Constant *AndC = GetConstantInType(I->getType(),
961 ~KnownOne & DemandedMask);
963 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
964 InsertNewInstBefore(And, *I);
965 return UpdateValueUsesWith(I, And);
969 // If the RHS is a constant, see if we can simplify it.
970 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
971 if (ShrinkDemandedConstant(I, 1, DemandedMask))
972 return UpdateValueUsesWith(I, I);
974 KnownZero = KnownZeroOut;
975 KnownOne = KnownOneOut;
978 case Instruction::Select:
979 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
980 KnownZero, KnownOne, Depth+1))
982 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
983 KnownZero2, KnownOne2, Depth+1))
985 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
986 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
988 // If the operands are constants, see if we can simplify them.
989 if (ShrinkDemandedConstant(I, 1, DemandedMask))
990 return UpdateValueUsesWith(I, I);
991 if (ShrinkDemandedConstant(I, 2, DemandedMask))
992 return UpdateValueUsesWith(I, I);
994 // Only known if known in both the LHS and RHS.
995 KnownOne &= KnownOne2;
996 KnownZero &= KnownZero2;
998 case Instruction::Trunc:
999 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1000 KnownZero, KnownOne, Depth+1))
1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1004 case Instruction::BitCast:
1005 if (!I->getOperand(0)->getType()->isIntegral())
1008 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1009 KnownZero, KnownOne, Depth+1))
1011 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1013 case Instruction::ZExt: {
1014 // Compute the bits in the result that are not present in the input.
1015 const Type *SrcTy = I->getOperand(0)->getType();
1016 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1017 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1019 DemandedMask &= SrcTy->getIntegralTypeMask();
1020 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1021 KnownZero, KnownOne, Depth+1))
1023 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1024 // The top bits are known to be zero.
1025 KnownZero |= NewBits;
1028 case Instruction::SExt: {
1029 // Compute the bits in the result that are not present in the input.
1030 const Type *SrcTy = I->getOperand(0)->getType();
1031 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1032 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1034 // Get the sign bit for the source type
1035 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1036 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1038 // If any of the sign extended bits are demanded, we know that the sign
1040 if (NewBits & DemandedMask)
1041 InputDemandedBits |= InSignBit;
1043 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1044 KnownZero, KnownOne, Depth+1))
1046 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1048 // If the sign bit of the input is known set or clear, then we know the
1049 // top bits of the result.
1051 // If the input sign bit is known zero, or if the NewBits are not demanded
1052 // convert this into a zero extension.
1053 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1054 // Convert to ZExt cast
1055 CastInst *NewCast = CastInst::create(
1056 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1057 return UpdateValueUsesWith(I, NewCast);
1058 } else if (KnownOne & InSignBit) { // Input sign bit known set
1059 KnownOne |= NewBits;
1060 KnownZero &= ~NewBits;
1061 } else { // Input sign bit unknown
1062 KnownZero &= ~NewBits;
1063 KnownOne &= ~NewBits;
1067 case Instruction::Add:
1068 // If there is a constant on the RHS, there are a variety of xformations
1070 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1071 // If null, this should be simplified elsewhere. Some of the xforms here
1072 // won't work if the RHS is zero.
1073 if (RHS->isNullValue())
1076 // Figure out what the input bits are. If the top bits of the and result
1077 // are not demanded, then the add doesn't demand them from its input
1080 // Shift the demanded mask up so that it's at the top of the uint64_t.
1081 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1082 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1084 // If the top bit of the output is demanded, demand everything from the
1085 // input. Otherwise, we demand all the input bits except NLZ top bits.
1086 uint64_t InDemandedBits = ~0ULL >> 64-BitWidth+NLZ;
1088 // Find information about known zero/one bits in the input.
1089 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1090 KnownZero2, KnownOne2, Depth+1))
1093 // If the RHS of the add has bits set that can't affect the input, reduce
1095 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1096 return UpdateValueUsesWith(I, I);
1098 // Avoid excess work.
1099 if (KnownZero2 == 0 && KnownOne2 == 0)
1102 // Turn it into OR if input bits are zero.
1103 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1105 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1107 InsertNewInstBefore(Or, *I);
1108 return UpdateValueUsesWith(I, Or);
1111 // We can say something about the output known-zero and known-one bits,
1112 // depending on potential carries from the input constant and the
1113 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1114 // bits set and the RHS constant is 0x01001, then we know we have a known
1115 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1117 // To compute this, we first compute the potential carry bits. These are
1118 // the bits which may be modified. I'm not aware of a better way to do
1120 uint64_t RHSVal = RHS->getZExtValue();
1122 bool CarryIn = false;
1123 uint64_t CarryBits = 0;
1124 uint64_t CurBit = 1;
1125 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1126 // Record the current carry in.
1127 if (CarryIn) CarryBits |= CurBit;
1131 // This bit has a carry out unless it is "zero + zero" or
1132 // "zero + anything" with no carry in.
1133 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1134 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1135 } else if (!CarryIn &&
1136 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1137 CarryOut = false; // 0 + anything has no carry out if no carry in.
1139 // Otherwise, we have to assume we have a carry out.
1143 // This stage's carry out becomes the next stage's carry-in.
1147 // Now that we know which bits have carries, compute the known-1/0 sets.
1149 // Bits are known one if they are known zero in one operand and one in the
1150 // other, and there is no input carry.
1151 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1153 // Bits are known zero if they are known zero in both operands and there
1154 // is no input carry.
1155 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1158 case Instruction::Shl:
1159 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1160 uint64_t ShiftAmt = SA->getZExtValue();
1161 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1162 KnownZero, KnownOne, Depth+1))
1164 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1165 KnownZero <<= ShiftAmt;
1166 KnownOne <<= ShiftAmt;
1167 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1170 case Instruction::LShr:
1171 // For a logical shift right
1172 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1173 unsigned ShiftAmt = SA->getZExtValue();
1175 // Compute the new bits that are at the top now.
1176 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1177 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1178 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1179 // Unsigned shift right.
1180 if (SimplifyDemandedBits(I->getOperand(0),
1181 (DemandedMask << ShiftAmt) & TypeMask,
1182 KnownZero, KnownOne, Depth+1))
1184 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1185 KnownZero &= TypeMask;
1186 KnownOne &= TypeMask;
1187 KnownZero >>= ShiftAmt;
1188 KnownOne >>= ShiftAmt;
1189 KnownZero |= HighBits; // high bits known zero.
1192 case Instruction::AShr:
1193 // If this is an arithmetic shift right and only the low-bit is set, we can
1194 // always convert this into a logical shr, even if the shift amount is
1195 // variable. The low bit of the shift cannot be an input sign bit unless
1196 // the shift amount is >= the size of the datatype, which is undefined.
1197 if (DemandedMask == 1) {
1198 // Perform the logical shift right.
1199 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1200 I->getOperand(1), I->getName());
1201 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1202 return UpdateValueUsesWith(I, NewVal);
1205 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1206 unsigned ShiftAmt = SA->getZExtValue();
1208 // Compute the new bits that are at the top now.
1209 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1210 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1211 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1212 // Signed shift right.
1213 if (SimplifyDemandedBits(I->getOperand(0),
1214 (DemandedMask << ShiftAmt) & TypeMask,
1215 KnownZero, KnownOne, Depth+1))
1217 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1218 KnownZero &= TypeMask;
1219 KnownOne &= TypeMask;
1220 KnownZero >>= ShiftAmt;
1221 KnownOne >>= ShiftAmt;
1223 // Handle the sign bits.
1224 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1225 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1227 // If the input sign bit is known to be zero, or if none of the top bits
1228 // are demanded, turn this into an unsigned shift right.
1229 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1230 // Perform the logical shift right.
1231 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1233 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1234 return UpdateValueUsesWith(I, NewVal);
1235 } else if (KnownOne & SignBit) { // New bits are known one.
1236 KnownOne |= HighBits;
1242 // If the client is only demanding bits that we know, return the known
1244 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1245 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1250 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1251 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1252 /// actually used by the caller. This method analyzes which elements of the
1253 /// operand are undef and returns that information in UndefElts.
1255 /// If the information about demanded elements can be used to simplify the
1256 /// operation, the operation is simplified, then the resultant value is
1257 /// returned. This returns null if no change was made.
1258 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1259 uint64_t &UndefElts,
1261 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1262 assert(VWidth <= 64 && "Vector too wide to analyze!");
1263 uint64_t EltMask = ~0ULL >> (64-VWidth);
1264 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1265 "Invalid DemandedElts!");
1267 if (isa<UndefValue>(V)) {
1268 // If the entire vector is undefined, just return this info.
1269 UndefElts = EltMask;
1271 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1272 UndefElts = EltMask;
1273 return UndefValue::get(V->getType());
1277 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1278 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1279 Constant *Undef = UndefValue::get(EltTy);
1281 std::vector<Constant*> Elts;
1282 for (unsigned i = 0; i != VWidth; ++i)
1283 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1284 Elts.push_back(Undef);
1285 UndefElts |= (1ULL << i);
1286 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1287 Elts.push_back(Undef);
1288 UndefElts |= (1ULL << i);
1289 } else { // Otherwise, defined.
1290 Elts.push_back(CP->getOperand(i));
1293 // If we changed the constant, return it.
1294 Constant *NewCP = ConstantPacked::get(Elts);
1295 return NewCP != CP ? NewCP : 0;
1296 } else if (isa<ConstantAggregateZero>(V)) {
1297 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1299 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1300 Constant *Zero = Constant::getNullValue(EltTy);
1301 Constant *Undef = UndefValue::get(EltTy);
1302 std::vector<Constant*> Elts;
1303 for (unsigned i = 0; i != VWidth; ++i)
1304 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1305 UndefElts = DemandedElts ^ EltMask;
1306 return ConstantPacked::get(Elts);
1309 if (!V->hasOneUse()) { // Other users may use these bits.
1310 if (Depth != 0) { // Not at the root.
1311 // TODO: Just compute the UndefElts information recursively.
1315 } else if (Depth == 10) { // Limit search depth.
1319 Instruction *I = dyn_cast<Instruction>(V);
1320 if (!I) return false; // Only analyze instructions.
1322 bool MadeChange = false;
1323 uint64_t UndefElts2;
1325 switch (I->getOpcode()) {
1328 case Instruction::InsertElement: {
1329 // If this is a variable index, we don't know which element it overwrites.
1330 // demand exactly the same input as we produce.
1331 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1333 // Note that we can't propagate undef elt info, because we don't know
1334 // which elt is getting updated.
1335 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1336 UndefElts2, Depth+1);
1337 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1341 // If this is inserting an element that isn't demanded, remove this
1343 unsigned IdxNo = Idx->getZExtValue();
1344 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1345 return AddSoonDeadInstToWorklist(*I, 0);
1347 // Otherwise, the element inserted overwrites whatever was there, so the
1348 // input demanded set is simpler than the output set.
1349 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1350 DemandedElts & ~(1ULL << IdxNo),
1351 UndefElts, Depth+1);
1352 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1354 // The inserted element is defined.
1355 UndefElts |= 1ULL << IdxNo;
1359 case Instruction::And:
1360 case Instruction::Or:
1361 case Instruction::Xor:
1362 case Instruction::Add:
1363 case Instruction::Sub:
1364 case Instruction::Mul:
1365 // div/rem demand all inputs, because they don't want divide by zero.
1366 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1367 UndefElts, Depth+1);
1368 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1369 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1370 UndefElts2, Depth+1);
1371 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1373 // Output elements are undefined if both are undefined. Consider things
1374 // like undef&0. The result is known zero, not undef.
1375 UndefElts &= UndefElts2;
1378 case Instruction::Call: {
1379 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1381 switch (II->getIntrinsicID()) {
1384 // Binary vector operations that work column-wise. A dest element is a
1385 // function of the corresponding input elements from the two inputs.
1386 case Intrinsic::x86_sse_sub_ss:
1387 case Intrinsic::x86_sse_mul_ss:
1388 case Intrinsic::x86_sse_min_ss:
1389 case Intrinsic::x86_sse_max_ss:
1390 case Intrinsic::x86_sse2_sub_sd:
1391 case Intrinsic::x86_sse2_mul_sd:
1392 case Intrinsic::x86_sse2_min_sd:
1393 case Intrinsic::x86_sse2_max_sd:
1394 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1395 UndefElts, Depth+1);
1396 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1397 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1398 UndefElts2, Depth+1);
1399 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1401 // If only the low elt is demanded and this is a scalarizable intrinsic,
1402 // scalarize it now.
1403 if (DemandedElts == 1) {
1404 switch (II->getIntrinsicID()) {
1406 case Intrinsic::x86_sse_sub_ss:
1407 case Intrinsic::x86_sse_mul_ss:
1408 case Intrinsic::x86_sse2_sub_sd:
1409 case Intrinsic::x86_sse2_mul_sd:
1410 // TODO: Lower MIN/MAX/ABS/etc
1411 Value *LHS = II->getOperand(1);
1412 Value *RHS = II->getOperand(2);
1413 // Extract the element as scalars.
1414 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1415 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1417 switch (II->getIntrinsicID()) {
1418 default: assert(0 && "Case stmts out of sync!");
1419 case Intrinsic::x86_sse_sub_ss:
1420 case Intrinsic::x86_sse2_sub_sd:
1421 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1422 II->getName()), *II);
1424 case Intrinsic::x86_sse_mul_ss:
1425 case Intrinsic::x86_sse2_mul_sd:
1426 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1427 II->getName()), *II);
1432 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1434 InsertNewInstBefore(New, *II);
1435 AddSoonDeadInstToWorklist(*II, 0);
1440 // Output elements are undefined if both are undefined. Consider things
1441 // like undef&0. The result is known zero, not undef.
1442 UndefElts &= UndefElts2;
1448 return MadeChange ? I : 0;
1451 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
1452 // true when both operands are equal...
1454 static bool isTrueWhenEqual(Instruction &I) {
1455 return I.getOpcode() == Instruction::SetEQ ||
1456 I.getOpcode() == Instruction::SetGE ||
1457 I.getOpcode() == Instruction::SetLE;
1460 /// AssociativeOpt - Perform an optimization on an associative operator. This
1461 /// function is designed to check a chain of associative operators for a
1462 /// potential to apply a certain optimization. Since the optimization may be
1463 /// applicable if the expression was reassociated, this checks the chain, then
1464 /// reassociates the expression as necessary to expose the optimization
1465 /// opportunity. This makes use of a special Functor, which must define
1466 /// 'shouldApply' and 'apply' methods.
1468 template<typename Functor>
1469 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1470 unsigned Opcode = Root.getOpcode();
1471 Value *LHS = Root.getOperand(0);
1473 // Quick check, see if the immediate LHS matches...
1474 if (F.shouldApply(LHS))
1475 return F.apply(Root);
1477 // Otherwise, if the LHS is not of the same opcode as the root, return.
1478 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1479 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1480 // Should we apply this transform to the RHS?
1481 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1483 // If not to the RHS, check to see if we should apply to the LHS...
1484 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1485 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1489 // If the functor wants to apply the optimization to the RHS of LHSI,
1490 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1492 BasicBlock *BB = Root.getParent();
1494 // Now all of the instructions are in the current basic block, go ahead
1495 // and perform the reassociation.
1496 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1498 // First move the selected RHS to the LHS of the root...
1499 Root.setOperand(0, LHSI->getOperand(1));
1501 // Make what used to be the LHS of the root be the user of the root...
1502 Value *ExtraOperand = TmpLHSI->getOperand(1);
1503 if (&Root == TmpLHSI) {
1504 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1507 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1508 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1509 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1510 BasicBlock::iterator ARI = &Root; ++ARI;
1511 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1514 // Now propagate the ExtraOperand down the chain of instructions until we
1516 while (TmpLHSI != LHSI) {
1517 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1518 // Move the instruction to immediately before the chain we are
1519 // constructing to avoid breaking dominance properties.
1520 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1521 BB->getInstList().insert(ARI, NextLHSI);
1524 Value *NextOp = NextLHSI->getOperand(1);
1525 NextLHSI->setOperand(1, ExtraOperand);
1527 ExtraOperand = NextOp;
1530 // Now that the instructions are reassociated, have the functor perform
1531 // the transformation...
1532 return F.apply(Root);
1535 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1541 // AddRHS - Implements: X + X --> X << 1
1544 AddRHS(Value *rhs) : RHS(rhs) {}
1545 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1546 Instruction *apply(BinaryOperator &Add) const {
1547 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1548 ConstantInt::get(Type::UByteTy, 1));
1552 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1554 struct AddMaskingAnd {
1556 AddMaskingAnd(Constant *c) : C2(c) {}
1557 bool shouldApply(Value *LHS) const {
1559 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1560 ConstantExpr::getAnd(C1, C2)->isNullValue();
1562 Instruction *apply(BinaryOperator &Add) const {
1563 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1567 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1569 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1570 if (Constant *SOC = dyn_cast<Constant>(SO))
1571 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1573 return IC->InsertNewInstBefore(CastInst::create(
1574 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1577 // Figure out if the constant is the left or the right argument.
1578 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1579 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1581 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1583 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1584 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1587 Value *Op0 = SO, *Op1 = ConstOperand;
1589 std::swap(Op0, Op1);
1591 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1592 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1593 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1594 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1596 assert(0 && "Unknown binary instruction type!");
1599 return IC->InsertNewInstBefore(New, I);
1602 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1603 // constant as the other operand, try to fold the binary operator into the
1604 // select arguments. This also works for Cast instructions, which obviously do
1605 // not have a second operand.
1606 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1608 // Don't modify shared select instructions
1609 if (!SI->hasOneUse()) return 0;
1610 Value *TV = SI->getOperand(1);
1611 Value *FV = SI->getOperand(2);
1613 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1614 // Bool selects with constant operands can be folded to logical ops.
1615 if (SI->getType() == Type::BoolTy) return 0;
1617 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1618 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1620 return new SelectInst(SI->getCondition(), SelectTrueVal,
1627 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1628 /// node as operand #0, see if we can fold the instruction into the PHI (which
1629 /// is only possible if all operands to the PHI are constants).
1630 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1631 PHINode *PN = cast<PHINode>(I.getOperand(0));
1632 unsigned NumPHIValues = PN->getNumIncomingValues();
1633 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1635 // Check to see if all of the operands of the PHI are constants. If there is
1636 // one non-constant value, remember the BB it is. If there is more than one
1638 BasicBlock *NonConstBB = 0;
1639 for (unsigned i = 0; i != NumPHIValues; ++i)
1640 if (!isa<Constant>(PN->getIncomingValue(i))) {
1641 if (NonConstBB) return 0; // More than one non-const value.
1642 NonConstBB = PN->getIncomingBlock(i);
1644 // If the incoming non-constant value is in I's block, we have an infinite
1646 if (NonConstBB == I.getParent())
1650 // If there is exactly one non-constant value, we can insert a copy of the
1651 // operation in that block. However, if this is a critical edge, we would be
1652 // inserting the computation one some other paths (e.g. inside a loop). Only
1653 // do this if the pred block is unconditionally branching into the phi block.
1655 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1656 if (!BI || !BI->isUnconditional()) return 0;
1659 // Okay, we can do the transformation: create the new PHI node.
1660 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1662 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1663 InsertNewInstBefore(NewPN, *PN);
1665 // Next, add all of the operands to the PHI.
1666 if (I.getNumOperands() == 2) {
1667 Constant *C = cast<Constant>(I.getOperand(1));
1668 for (unsigned i = 0; i != NumPHIValues; ++i) {
1670 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1671 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1673 assert(PN->getIncomingBlock(i) == NonConstBB);
1674 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1675 InV = BinaryOperator::create(BO->getOpcode(),
1676 PN->getIncomingValue(i), C, "phitmp",
1677 NonConstBB->getTerminator());
1678 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1679 InV = new ShiftInst(SI->getOpcode(),
1680 PN->getIncomingValue(i), C, "phitmp",
1681 NonConstBB->getTerminator());
1683 assert(0 && "Unknown binop!");
1685 WorkList.push_back(cast<Instruction>(InV));
1687 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1690 CastInst *CI = cast<CastInst>(&I);
1691 const Type *RetTy = CI->getType();
1692 for (unsigned i = 0; i != NumPHIValues; ++i) {
1694 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1695 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1697 assert(PN->getIncomingBlock(i) == NonConstBB);
1698 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1699 I.getType(), "phitmp",
1700 NonConstBB->getTerminator());
1701 WorkList.push_back(cast<Instruction>(InV));
1703 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1706 return ReplaceInstUsesWith(I, NewPN);
1709 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1710 bool Changed = SimplifyCommutative(I);
1711 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1713 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1714 // X + undef -> undef
1715 if (isa<UndefValue>(RHS))
1716 return ReplaceInstUsesWith(I, RHS);
1719 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1720 if (RHSC->isNullValue())
1721 return ReplaceInstUsesWith(I, LHS);
1722 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1723 if (CFP->isExactlyValue(-0.0))
1724 return ReplaceInstUsesWith(I, LHS);
1727 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1728 // X + (signbit) --> X ^ signbit
1729 uint64_t Val = CI->getZExtValue();
1730 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1731 return BinaryOperator::createXor(LHS, RHS);
1733 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1734 // (X & 254)+1 -> (X&254)|1
1735 uint64_t KnownZero, KnownOne;
1736 if (!isa<PackedType>(I.getType()) &&
1737 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
1738 KnownZero, KnownOne))
1742 if (isa<PHINode>(LHS))
1743 if (Instruction *NV = FoldOpIntoPhi(I))
1746 ConstantInt *XorRHS = 0;
1748 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1749 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1750 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1751 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1753 uint64_t C0080Val = 1ULL << 31;
1754 int64_t CFF80Val = -C0080Val;
1757 if (TySizeBits > Size) {
1759 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1760 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1761 if (RHSSExt == CFF80Val) {
1762 if (XorRHS->getZExtValue() == C0080Val)
1764 } else if (RHSZExt == C0080Val) {
1765 if (XorRHS->getSExtValue() == CFF80Val)
1769 // This is a sign extend if the top bits are known zero.
1770 uint64_t Mask = ~0ULL;
1771 Mask <<= 64-(TySizeBits-Size);
1772 Mask &= XorLHS->getType()->getIntegralTypeMask();
1773 if (!MaskedValueIsZero(XorLHS, Mask))
1774 Size = 0; // Not a sign ext, but can't be any others either.
1781 } while (Size >= 8);
1784 const Type *MiddleType = 0;
1787 case 32: MiddleType = Type::IntTy; break;
1788 case 16: MiddleType = Type::ShortTy; break;
1789 case 8: MiddleType = Type::SByteTy; break;
1792 Instruction *NewTrunc =
1793 CastInst::createInferredCast(XorLHS, MiddleType, "sext");
1794 InsertNewInstBefore(NewTrunc, I);
1795 return new SExtInst(NewTrunc, I.getType());
1801 if (I.getType()->isInteger()) {
1802 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1804 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1805 if (RHSI->getOpcode() == Instruction::Sub)
1806 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1807 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1809 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1810 if (LHSI->getOpcode() == Instruction::Sub)
1811 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1812 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1817 if (Value *V = dyn_castNegVal(LHS))
1818 return BinaryOperator::createSub(RHS, V);
1821 if (!isa<Constant>(RHS))
1822 if (Value *V = dyn_castNegVal(RHS))
1823 return BinaryOperator::createSub(LHS, V);
1827 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1828 if (X == RHS) // X*C + X --> X * (C+1)
1829 return BinaryOperator::createMul(RHS, AddOne(C2));
1831 // X*C1 + X*C2 --> X * (C1+C2)
1833 if (X == dyn_castFoldableMul(RHS, C1))
1834 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1837 // X + X*C --> X * (C+1)
1838 if (dyn_castFoldableMul(RHS, C2) == LHS)
1839 return BinaryOperator::createMul(LHS, AddOne(C2));
1842 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1843 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1844 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1846 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1848 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1849 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1850 return BinaryOperator::createSub(C, X);
1853 // (X & FF00) + xx00 -> (X+xx00) & FF00
1854 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1855 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1856 if (Anded == CRHS) {
1857 // See if all bits from the first bit set in the Add RHS up are included
1858 // in the mask. First, get the rightmost bit.
1859 uint64_t AddRHSV = CRHS->getZExtValue();
1861 // Form a mask of all bits from the lowest bit added through the top.
1862 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1863 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1865 // See if the and mask includes all of these bits.
1866 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1868 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1869 // Okay, the xform is safe. Insert the new add pronto.
1870 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1871 LHS->getName()), I);
1872 return BinaryOperator::createAnd(NewAdd, C2);
1877 // Try to fold constant add into select arguments.
1878 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1879 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1883 // add (cast *A to intptrtype) B ->
1884 // cast (GEP (cast *A to sbyte*) B) ->
1887 CastInst *CI = dyn_cast<CastInst>(LHS);
1890 CI = dyn_cast<CastInst>(RHS);
1893 if (CI && CI->getType()->isSized() &&
1894 (CI->getType()->getPrimitiveSize() ==
1895 TD->getIntPtrType()->getPrimitiveSize())
1896 && isa<PointerType>(CI->getOperand(0)->getType())) {
1897 Value *I2 = InsertCastBefore(CI->getOperand(0),
1898 PointerType::get(Type::SByteTy), I);
1899 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1900 return new PtrToIntInst(I2, CI->getType());
1904 return Changed ? &I : 0;
1907 // isSignBit - Return true if the value represented by the constant only has the
1908 // highest order bit set.
1909 static bool isSignBit(ConstantInt *CI) {
1910 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1911 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1914 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1916 static Value *RemoveNoopCast(Value *V) {
1917 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1918 const Type *CTy = CI->getType();
1919 const Type *OpTy = CI->getOperand(0)->getType();
1920 if (CTy->isInteger() && OpTy->isInteger()) {
1921 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1922 return RemoveNoopCast(CI->getOperand(0));
1923 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1924 return RemoveNoopCast(CI->getOperand(0));
1929 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1930 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1932 if (Op0 == Op1) // sub X, X -> 0
1933 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1935 // If this is a 'B = x-(-A)', change to B = x+A...
1936 if (Value *V = dyn_castNegVal(Op1))
1937 return BinaryOperator::createAdd(Op0, V);
1939 if (isa<UndefValue>(Op0))
1940 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1941 if (isa<UndefValue>(Op1))
1942 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1944 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1945 // Replace (-1 - A) with (~A)...
1946 if (C->isAllOnesValue())
1947 return BinaryOperator::createNot(Op1);
1949 // C - ~X == X + (1+C)
1951 if (match(Op1, m_Not(m_Value(X))))
1952 return BinaryOperator::createAdd(X,
1953 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1954 // -((uint)X >> 31) -> ((int)X >> 31)
1955 // -((int)X >> 31) -> ((uint)X >> 31)
1956 if (C->isNullValue()) {
1957 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1958 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1959 if (SI->getOpcode() == Instruction::LShr) {
1960 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1961 // Check to see if we are shifting out everything but the sign bit.
1962 if (CU->getZExtValue() ==
1963 SI->getType()->getPrimitiveSizeInBits()-1) {
1964 // Ok, the transformation is safe. Insert AShr.
1965 return new ShiftInst(Instruction::AShr, SI->getOperand(0),
1970 else if (SI->getOpcode() == Instruction::AShr) {
1971 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1972 // Check to see if we are shifting out everything but the sign bit.
1973 if (CU->getZExtValue() ==
1974 SI->getType()->getPrimitiveSizeInBits()-1) {
1975 // Ok, the transformation is safe. Insert LShr.
1976 return new ShiftInst(Instruction::LShr, SI->getOperand(0),
1983 // Try to fold constant sub into select arguments.
1984 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1985 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1988 if (isa<PHINode>(Op0))
1989 if (Instruction *NV = FoldOpIntoPhi(I))
1993 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1994 if (Op1I->getOpcode() == Instruction::Add &&
1995 !Op0->getType()->isFPOrFPVector()) {
1996 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1997 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1998 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1999 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2000 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2001 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2002 // C1-(X+C2) --> (C1-C2)-X
2003 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2004 Op1I->getOperand(0));
2008 if (Op1I->hasOneUse()) {
2009 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2010 // is not used by anyone else...
2012 if (Op1I->getOpcode() == Instruction::Sub &&
2013 !Op1I->getType()->isFPOrFPVector()) {
2014 // Swap the two operands of the subexpr...
2015 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2016 Op1I->setOperand(0, IIOp1);
2017 Op1I->setOperand(1, IIOp0);
2019 // Create the new top level add instruction...
2020 return BinaryOperator::createAdd(Op0, Op1);
2023 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2025 if (Op1I->getOpcode() == Instruction::And &&
2026 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2027 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2030 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2031 return BinaryOperator::createAnd(Op0, NewNot);
2034 // 0 - (X sdiv C) -> (X sdiv -C)
2035 if (Op1I->getOpcode() == Instruction::SDiv)
2036 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2037 if (CSI->isNullValue())
2038 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2039 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2040 ConstantExpr::getNeg(DivRHS));
2042 // X - X*C --> X * (1-C)
2043 ConstantInt *C2 = 0;
2044 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2046 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2047 return BinaryOperator::createMul(Op0, CP1);
2052 if (!Op0->getType()->isFPOrFPVector())
2053 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2054 if (Op0I->getOpcode() == Instruction::Add) {
2055 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2056 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2057 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2058 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2059 } else if (Op0I->getOpcode() == Instruction::Sub) {
2060 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2061 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2065 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2066 if (X == Op1) { // X*C - X --> X * (C-1)
2067 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2068 return BinaryOperator::createMul(Op1, CP1);
2071 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2072 if (X == dyn_castFoldableMul(Op1, C2))
2073 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2078 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
2079 /// really just returns true if the most significant (sign) bit is set.
2080 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
2081 if (RHS->getType()->isSigned()) {
2082 // True if source is LHS < 0 or LHS <= -1
2083 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
2084 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
2086 ConstantInt *RHSC = cast<ConstantInt>(RHS);
2087 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
2088 // the size of the integer type.
2089 if (Opcode == Instruction::SetGE)
2090 return RHSC->getZExtValue() ==
2091 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
2092 if (Opcode == Instruction::SetGT)
2093 return RHSC->getZExtValue() ==
2094 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2099 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2100 bool Changed = SimplifyCommutative(I);
2101 Value *Op0 = I.getOperand(0);
2103 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2104 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2106 // Simplify mul instructions with a constant RHS...
2107 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2108 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2110 // ((X << C1)*C2) == (X * (C2 << C1))
2111 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2112 if (SI->getOpcode() == Instruction::Shl)
2113 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2114 return BinaryOperator::createMul(SI->getOperand(0),
2115 ConstantExpr::getShl(CI, ShOp));
2117 if (CI->isNullValue())
2118 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2119 if (CI->equalsInt(1)) // X * 1 == X
2120 return ReplaceInstUsesWith(I, Op0);
2121 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2122 return BinaryOperator::createNeg(Op0, I.getName());
2124 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2125 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2126 uint64_t C = Log2_64(Val);
2127 return new ShiftInst(Instruction::Shl, Op0,
2128 ConstantInt::get(Type::UByteTy, C));
2130 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2131 if (Op1F->isNullValue())
2132 return ReplaceInstUsesWith(I, Op1);
2134 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2135 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2136 if (Op1F->getValue() == 1.0)
2137 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2140 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2141 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2142 isa<ConstantInt>(Op0I->getOperand(1))) {
2143 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2144 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2146 InsertNewInstBefore(Add, I);
2147 Value *C1C2 = ConstantExpr::getMul(Op1,
2148 cast<Constant>(Op0I->getOperand(1)));
2149 return BinaryOperator::createAdd(Add, C1C2);
2153 // Try to fold constant mul into select arguments.
2154 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2155 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2158 if (isa<PHINode>(Op0))
2159 if (Instruction *NV = FoldOpIntoPhi(I))
2163 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2164 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2165 return BinaryOperator::createMul(Op0v, Op1v);
2167 // If one of the operands of the multiply is a cast from a boolean value, then
2168 // we know the bool is either zero or one, so this is a 'masking' multiply.
2169 // See if we can simplify things based on how the boolean was originally
2171 CastInst *BoolCast = 0;
2172 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
2173 if (CI->getOperand(0)->getType() == Type::BoolTy)
2176 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
2177 if (CI->getOperand(0)->getType() == Type::BoolTy)
2180 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
2181 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2182 const Type *SCOpTy = SCIOp0->getType();
2184 // If the setcc is true iff the sign bit of X is set, then convert this
2185 // multiply into a shift/and combination.
2186 if (isa<ConstantInt>(SCIOp1) &&
2187 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
2188 // Shift the X value right to turn it into "all signbits".
2189 Constant *Amt = ConstantInt::get(Type::UByteTy,
2190 SCOpTy->getPrimitiveSizeInBits()-1);
2191 if (SCIOp0->getType()->isUnsigned()) {
2192 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
2193 SCIOp0 = InsertCastBefore(SCIOp0, NewTy, I);
2197 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2198 BoolCast->getOperand(0)->getName()+
2201 // If the multiply type is not the same as the source type, sign extend
2202 // or truncate to the multiply type.
2203 if (I.getType() != V->getType())
2204 V = InsertCastBefore(V, I.getType(), I);
2206 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2207 return BinaryOperator::createAnd(V, OtherOp);
2212 return Changed ? &I : 0;
2215 /// This function implements the transforms on div instructions that work
2216 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2217 /// used by the visitors to those instructions.
2218 /// @brief Transforms common to all three div instructions
2219 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2220 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2223 if (isa<UndefValue>(Op0))
2224 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2226 // X / undef -> undef
2227 if (isa<UndefValue>(Op1))
2228 return ReplaceInstUsesWith(I, Op1);
2230 // Handle cases involving: div X, (select Cond, Y, Z)
2231 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2232 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2233 // same basic block, then we replace the select with Y, and the condition
2234 // of the select with false (if the cond value is in the same BB). If the
2235 // select has uses other than the div, this allows them to be simplified
2236 // also. Note that div X, Y is just as good as div X, 0 (undef)
2237 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2238 if (ST->isNullValue()) {
2239 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2240 if (CondI && CondI->getParent() == I.getParent())
2241 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2242 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2243 I.setOperand(1, SI->getOperand(2));
2245 UpdateValueUsesWith(SI, SI->getOperand(2));
2249 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2250 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2251 if (ST->isNullValue()) {
2252 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2253 if (CondI && CondI->getParent() == I.getParent())
2254 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2255 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2256 I.setOperand(1, SI->getOperand(1));
2258 UpdateValueUsesWith(SI, SI->getOperand(1));
2266 /// This function implements the transforms common to both integer division
2267 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2268 /// division instructions.
2269 /// @brief Common integer divide transforms
2270 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2271 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2273 if (Instruction *Common = commonDivTransforms(I))
2276 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2278 if (RHS->equalsInt(1))
2279 return ReplaceInstUsesWith(I, Op0);
2281 // (X / C1) / C2 -> X / (C1*C2)
2282 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2283 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2284 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2285 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2286 ConstantExpr::getMul(RHS, LHSRHS));
2289 if (!RHS->isNullValue()) { // avoid X udiv 0
2290 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2291 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2293 if (isa<PHINode>(Op0))
2294 if (Instruction *NV = FoldOpIntoPhi(I))
2299 // 0 / X == 0, we don't need to preserve faults!
2300 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2301 if (LHS->equalsInt(0))
2302 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2307 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2308 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2310 // Handle the integer div common cases
2311 if (Instruction *Common = commonIDivTransforms(I))
2314 // X udiv C^2 -> X >> C
2315 // Check to see if this is an unsigned division with an exact power of 2,
2316 // if so, convert to a right shift.
2317 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2318 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2319 if (isPowerOf2_64(Val)) {
2320 uint64_t ShiftAmt = Log2_64(Val);
2321 return new ShiftInst(Instruction::LShr, Op0,
2322 ConstantInt::get(Type::UByteTy, ShiftAmt));
2326 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2327 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2328 if (RHSI->getOpcode() == Instruction::Shl &&
2329 isa<ConstantInt>(RHSI->getOperand(0))) {
2330 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2331 if (isPowerOf2_64(C1)) {
2332 Value *N = RHSI->getOperand(1);
2333 const Type *NTy = N->getType();
2334 if (uint64_t C2 = Log2_64(C1)) {
2335 Constant *C2V = ConstantInt::get(NTy, C2);
2336 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2338 return new ShiftInst(Instruction::LShr, Op0, N);
2343 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2344 // where C1&C2 are powers of two.
2345 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2346 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2347 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2348 if (!STO->isNullValue() && !STO->isNullValue()) {
2349 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2350 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2351 // Compute the shift amounts
2352 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2353 // Construct the "on true" case of the select
2354 Constant *TC = ConstantInt::get(Type::UByteTy, TSA);
2356 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2357 TSI = InsertNewInstBefore(TSI, I);
2359 // Construct the "on false" case of the select
2360 Constant *FC = ConstantInt::get(Type::UByteTy, FSA);
2362 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2363 FSI = InsertNewInstBefore(FSI, I);
2365 // construct the select instruction and return it.
2366 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2373 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2374 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2376 // Handle the integer div common cases
2377 if (Instruction *Common = commonIDivTransforms(I))
2380 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2382 if (RHS->isAllOnesValue())
2383 return BinaryOperator::createNeg(Op0);
2386 if (Value *LHSNeg = dyn_castNegVal(Op0))
2387 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2390 // If the sign bits of both operands are zero (i.e. we can prove they are
2391 // unsigned inputs), turn this into a udiv.
2392 if (I.getType()->isInteger()) {
2393 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2394 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2395 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2402 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2403 return commonDivTransforms(I);
2406 /// GetFactor - If we can prove that the specified value is at least a multiple
2407 /// of some factor, return that factor.
2408 static Constant *GetFactor(Value *V) {
2409 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2412 // Unless we can be tricky, we know this is a multiple of 1.
2413 Constant *Result = ConstantInt::get(V->getType(), 1);
2415 Instruction *I = dyn_cast<Instruction>(V);
2416 if (!I) return Result;
2418 if (I->getOpcode() == Instruction::Mul) {
2419 // Handle multiplies by a constant, etc.
2420 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2421 GetFactor(I->getOperand(1)));
2422 } else if (I->getOpcode() == Instruction::Shl) {
2423 // (X<<C) -> X * (1 << C)
2424 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2425 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2426 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2428 } else if (I->getOpcode() == Instruction::And) {
2429 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2430 // X & 0xFFF0 is known to be a multiple of 16.
2431 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2432 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2433 return ConstantExpr::getShl(Result,
2434 ConstantInt::get(Type::UByteTy, Zeros));
2436 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2437 // Only handle int->int casts.
2438 if (!CI->isIntegerCast())
2440 Value *Op = CI->getOperand(0);
2441 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2446 /// This function implements the transforms on rem instructions that work
2447 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2448 /// is used by the visitors to those instructions.
2449 /// @brief Transforms common to all three rem instructions
2450 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2451 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2453 // 0 % X == 0, we don't need to preserve faults!
2454 if (Constant *LHS = dyn_cast<Constant>(Op0))
2455 if (LHS->isNullValue())
2456 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2458 if (isa<UndefValue>(Op0)) // undef % X -> 0
2459 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2460 if (isa<UndefValue>(Op1))
2461 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2463 // Handle cases involving: rem X, (select Cond, Y, Z)
2464 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2465 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2466 // the same basic block, then we replace the select with Y, and the
2467 // condition of the select with false (if the cond value is in the same
2468 // BB). If the select has uses other than the div, this allows them to be
2470 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2471 if (ST->isNullValue()) {
2472 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2473 if (CondI && CondI->getParent() == I.getParent())
2474 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2475 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2476 I.setOperand(1, SI->getOperand(2));
2478 UpdateValueUsesWith(SI, SI->getOperand(2));
2481 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2482 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2483 if (ST->isNullValue()) {
2484 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2485 if (CondI && CondI->getParent() == I.getParent())
2486 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2487 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2488 I.setOperand(1, SI->getOperand(1));
2490 UpdateValueUsesWith(SI, SI->getOperand(1));
2498 /// This function implements the transforms common to both integer remainder
2499 /// instructions (urem and srem). It is called by the visitors to those integer
2500 /// remainder instructions.
2501 /// @brief Common integer remainder transforms
2502 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2503 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2505 if (Instruction *common = commonRemTransforms(I))
2508 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2509 // X % 0 == undef, we don't need to preserve faults!
2510 if (RHS->equalsInt(0))
2511 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2513 if (RHS->equalsInt(1)) // X % 1 == 0
2514 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2516 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2517 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2518 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2520 } else if (isa<PHINode>(Op0I)) {
2521 if (Instruction *NV = FoldOpIntoPhi(I))
2524 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2525 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2526 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2533 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2534 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2536 if (Instruction *common = commonIRemTransforms(I))
2539 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2540 // X urem C^2 -> X and C
2541 // Check to see if this is an unsigned remainder with an exact power of 2,
2542 // if so, convert to a bitwise and.
2543 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2544 if (isPowerOf2_64(C->getZExtValue()))
2545 return BinaryOperator::createAnd(Op0, SubOne(C));
2548 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2549 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2550 if (RHSI->getOpcode() == Instruction::Shl &&
2551 isa<ConstantInt>(RHSI->getOperand(0))) {
2552 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2553 if (isPowerOf2_64(C1)) {
2554 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2555 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2557 return BinaryOperator::createAnd(Op0, Add);
2562 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2563 // where C1&C2 are powers of two.
2564 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2565 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2566 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2567 // STO == 0 and SFO == 0 handled above.
2568 if (isPowerOf2_64(STO->getZExtValue()) &&
2569 isPowerOf2_64(SFO->getZExtValue())) {
2570 Value *TrueAnd = InsertNewInstBefore(
2571 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2572 Value *FalseAnd = InsertNewInstBefore(
2573 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2574 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2582 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2583 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2585 if (Instruction *common = commonIRemTransforms(I))
2588 if (Value *RHSNeg = dyn_castNegVal(Op1))
2589 if (!isa<ConstantInt>(RHSNeg) ||
2590 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2592 AddUsesToWorkList(I);
2593 I.setOperand(1, RHSNeg);
2597 // If the top bits of both operands are zero (i.e. we can prove they are
2598 // unsigned inputs), turn this into a urem.
2599 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2600 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2601 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2602 return BinaryOperator::createURem(Op0, Op1, I.getName());
2608 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2609 return commonRemTransforms(I);
2612 // isMaxValueMinusOne - return true if this is Max-1
2613 static bool isMaxValueMinusOne(const ConstantInt *C) {
2614 if (C->getType()->isUnsigned())
2615 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2617 // Calculate 0111111111..11111
2618 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2619 int64_t Val = INT64_MAX; // All ones
2620 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2621 return C->getSExtValue() == Val-1;
2624 // isMinValuePlusOne - return true if this is Min+1
2625 static bool isMinValuePlusOne(const ConstantInt *C) {
2626 if (C->getType()->isUnsigned())
2627 return C->getZExtValue() == 1;
2629 // Calculate 1111111111000000000000
2630 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2631 int64_t Val = -1; // All ones
2632 Val <<= TypeBits-1; // Shift over to the right spot
2633 return C->getSExtValue() == Val+1;
2636 // isOneBitSet - Return true if there is exactly one bit set in the specified
2638 static bool isOneBitSet(const ConstantInt *CI) {
2639 uint64_t V = CI->getZExtValue();
2640 return V && (V & (V-1)) == 0;
2643 #if 0 // Currently unused
2644 // isLowOnes - Return true if the constant is of the form 0+1+.
2645 static bool isLowOnes(const ConstantInt *CI) {
2646 uint64_t V = CI->getZExtValue();
2648 // There won't be bits set in parts that the type doesn't contain.
2649 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2651 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2652 return U && V && (U & V) == 0;
2656 // isHighOnes - Return true if the constant is of the form 1+0+.
2657 // This is the same as lowones(~X).
2658 static bool isHighOnes(const ConstantInt *CI) {
2659 uint64_t V = ~CI->getZExtValue();
2660 if (~V == 0) return false; // 0's does not match "1+"
2662 // There won't be bits set in parts that the type doesn't contain.
2663 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2665 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2666 return U && V && (U & V) == 0;
2670 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2671 /// are carefully arranged to allow folding of expressions such as:
2673 /// (A < B) | (A > B) --> (A != B)
2675 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2676 /// represents that the comparison is true if A == B, and bit value '1' is true
2679 static unsigned getSetCondCode(const SetCondInst *SCI) {
2680 switch (SCI->getOpcode()) {
2682 case Instruction::SetGT: return 1;
2683 case Instruction::SetEQ: return 2;
2684 case Instruction::SetGE: return 3;
2685 case Instruction::SetLT: return 4;
2686 case Instruction::SetNE: return 5;
2687 case Instruction::SetLE: return 6;
2690 assert(0 && "Invalid SetCC opcode!");
2695 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2696 /// opcode and two operands into either a constant true or false, or a brand new
2697 /// SetCC instruction.
2698 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2700 case 0: return ConstantBool::getFalse();
2701 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2702 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2703 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2704 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2705 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2706 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2707 case 7: return ConstantBool::getTrue();
2708 default: assert(0 && "Illegal SetCCCode!"); return 0;
2712 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2714 struct FoldSetCCLogical {
2717 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2718 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2719 bool shouldApply(Value *V) const {
2720 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2721 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2722 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2725 Instruction *apply(BinaryOperator &Log) const {
2726 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2727 if (SCI->getOperand(0) != LHS) {
2728 assert(SCI->getOperand(1) == LHS);
2729 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2732 unsigned LHSCode = getSetCondCode(SCI);
2733 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2735 switch (Log.getOpcode()) {
2736 case Instruction::And: Code = LHSCode & RHSCode; break;
2737 case Instruction::Or: Code = LHSCode | RHSCode; break;
2738 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2739 default: assert(0 && "Illegal logical opcode!"); return 0;
2742 Value *RV = getSetCCValue(Code, LHS, RHS);
2743 if (Instruction *I = dyn_cast<Instruction>(RV))
2745 // Otherwise, it's a constant boolean value...
2746 return IC.ReplaceInstUsesWith(Log, RV);
2749 } // end anonymous namespace
2751 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2752 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2753 // guaranteed to be either a shift instruction or a binary operator.
2754 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2755 ConstantIntegral *OpRHS,
2756 ConstantIntegral *AndRHS,
2757 BinaryOperator &TheAnd) {
2758 Value *X = Op->getOperand(0);
2759 Constant *Together = 0;
2760 if (!isa<ShiftInst>(Op))
2761 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2763 switch (Op->getOpcode()) {
2764 case Instruction::Xor:
2765 if (Op->hasOneUse()) {
2766 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2767 std::string OpName = Op->getName(); Op->setName("");
2768 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2769 InsertNewInstBefore(And, TheAnd);
2770 return BinaryOperator::createXor(And, Together);
2773 case Instruction::Or:
2774 if (Together == AndRHS) // (X | C) & C --> C
2775 return ReplaceInstUsesWith(TheAnd, AndRHS);
2777 if (Op->hasOneUse() && Together != OpRHS) {
2778 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2779 std::string Op0Name = Op->getName(); Op->setName("");
2780 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2781 InsertNewInstBefore(Or, TheAnd);
2782 return BinaryOperator::createAnd(Or, AndRHS);
2785 case Instruction::Add:
2786 if (Op->hasOneUse()) {
2787 // Adding a one to a single bit bit-field should be turned into an XOR
2788 // of the bit. First thing to check is to see if this AND is with a
2789 // single bit constant.
2790 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2792 // Clear bits that are not part of the constant.
2793 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2795 // If there is only one bit set...
2796 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2797 // Ok, at this point, we know that we are masking the result of the
2798 // ADD down to exactly one bit. If the constant we are adding has
2799 // no bits set below this bit, then we can eliminate the ADD.
2800 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2802 // Check to see if any bits below the one bit set in AndRHSV are set.
2803 if ((AddRHS & (AndRHSV-1)) == 0) {
2804 // If not, the only thing that can effect the output of the AND is
2805 // the bit specified by AndRHSV. If that bit is set, the effect of
2806 // the XOR is to toggle the bit. If it is clear, then the ADD has
2808 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2809 TheAnd.setOperand(0, X);
2812 std::string Name = Op->getName(); Op->setName("");
2813 // Pull the XOR out of the AND.
2814 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2815 InsertNewInstBefore(NewAnd, TheAnd);
2816 return BinaryOperator::createXor(NewAnd, AndRHS);
2823 case Instruction::Shl: {
2824 // We know that the AND will not produce any of the bits shifted in, so if
2825 // the anded constant includes them, clear them now!
2827 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2828 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2829 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2831 if (CI == ShlMask) { // Masking out bits that the shift already masks
2832 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2833 } else if (CI != AndRHS) { // Reducing bits set in and.
2834 TheAnd.setOperand(1, CI);
2839 case Instruction::LShr:
2841 // We know that the AND will not produce any of the bits shifted in, so if
2842 // the anded constant includes them, clear them now! This only applies to
2843 // unsigned shifts, because a signed shr may bring in set bits!
2845 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2846 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2847 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2849 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2850 return ReplaceInstUsesWith(TheAnd, Op);
2851 } else if (CI != AndRHS) {
2852 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2857 case Instruction::AShr:
2859 // See if this is shifting in some sign extension, then masking it out
2861 if (Op->hasOneUse()) {
2862 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2863 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2864 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2865 if (CI == AndRHS) { // Masking out bits shifted in.
2866 // Make the argument unsigned.
2867 Value *ShVal = Op->getOperand(0);
2868 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2869 OpRHS, Op->getName()),
2871 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
2872 return BinaryOperator::createAnd(ShVal, AndRHS2, TheAnd.getName());
2881 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2882 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2883 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2884 /// insert new instructions.
2885 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2886 bool Inside, Instruction &IB) {
2887 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2888 "Lo is not <= Hi in range emission code!");
2890 if (Lo == Hi) // Trivially false.
2891 return new SetCondInst(Instruction::SetNE, V, V);
2892 if (cast<ConstantIntegral>(Lo)->isMinValue())
2893 return new SetCondInst(Instruction::SetLT, V, Hi);
2895 Constant *AddCST = ConstantExpr::getNeg(Lo);
2896 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2897 InsertNewInstBefore(Add, IB);
2898 // Convert to unsigned for the comparison.
2899 const Type *UnsType = Add->getType()->getUnsignedVersion();
2900 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2901 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2902 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2903 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2906 if (Lo == Hi) // Trivially true.
2907 return new SetCondInst(Instruction::SetEQ, V, V);
2909 Hi = SubOne(cast<ConstantInt>(Hi));
2911 // V < 0 || V >= Hi ->'V > Hi-1'
2912 if (cast<ConstantIntegral>(Lo)->isMinValue())
2913 return new SetCondInst(Instruction::SetGT, V, Hi);
2915 // Emit X-Lo > Hi-Lo-1
2916 Constant *AddCST = ConstantExpr::getNeg(Lo);
2917 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2918 InsertNewInstBefore(Add, IB);
2919 // Convert to unsigned for the comparison.
2920 const Type *UnsType = Add->getType()->getUnsignedVersion();
2921 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2922 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2923 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2924 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2927 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2928 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2929 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2930 // not, since all 1s are not contiguous.
2931 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2932 uint64_t V = Val->getZExtValue();
2933 if (!isShiftedMask_64(V)) return false;
2935 // look for the first zero bit after the run of ones
2936 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2937 // look for the first non-zero bit
2938 ME = 64-CountLeadingZeros_64(V);
2944 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2945 /// where isSub determines whether the operator is a sub. If we can fold one of
2946 /// the following xforms:
2948 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2949 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2950 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2952 /// return (A +/- B).
2954 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2955 ConstantIntegral *Mask, bool isSub,
2957 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2958 if (!LHSI || LHSI->getNumOperands() != 2 ||
2959 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2961 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2963 switch (LHSI->getOpcode()) {
2965 case Instruction::And:
2966 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2967 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2968 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
2971 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2972 // part, we don't need any explicit masks to take them out of A. If that
2973 // is all N is, ignore it.
2975 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2976 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2978 if (MaskedValueIsZero(RHS, Mask))
2983 case Instruction::Or:
2984 case Instruction::Xor:
2985 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2986 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
2987 ConstantExpr::getAnd(N, Mask)->isNullValue())
2994 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2996 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2997 return InsertNewInstBefore(New, I);
3000 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3001 bool Changed = SimplifyCommutative(I);
3002 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3004 if (isa<UndefValue>(Op1)) // X & undef -> 0
3005 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3009 return ReplaceInstUsesWith(I, Op1);
3011 // See if we can simplify any instructions used by the instruction whose sole
3012 // purpose is to compute bits we don't care about.
3013 uint64_t KnownZero, KnownOne;
3014 if (!isa<PackedType>(I.getType()) &&
3015 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3016 KnownZero, KnownOne))
3019 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
3020 uint64_t AndRHSMask = AndRHS->getZExtValue();
3021 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
3022 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3024 // Optimize a variety of ((val OP C1) & C2) combinations...
3025 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3026 Instruction *Op0I = cast<Instruction>(Op0);
3027 Value *Op0LHS = Op0I->getOperand(0);
3028 Value *Op0RHS = Op0I->getOperand(1);
3029 switch (Op0I->getOpcode()) {
3030 case Instruction::Xor:
3031 case Instruction::Or:
3032 // If the mask is only needed on one incoming arm, push it up.
3033 if (Op0I->hasOneUse()) {
3034 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3035 // Not masking anything out for the LHS, move to RHS.
3036 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3037 Op0RHS->getName()+".masked");
3038 InsertNewInstBefore(NewRHS, I);
3039 return BinaryOperator::create(
3040 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3042 if (!isa<Constant>(Op0RHS) &&
3043 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3044 // Not masking anything out for the RHS, move to LHS.
3045 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3046 Op0LHS->getName()+".masked");
3047 InsertNewInstBefore(NewLHS, I);
3048 return BinaryOperator::create(
3049 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3054 case Instruction::Add:
3055 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3056 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3057 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3058 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3059 return BinaryOperator::createAnd(V, AndRHS);
3060 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3061 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3064 case Instruction::Sub:
3065 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3066 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3067 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3068 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3069 return BinaryOperator::createAnd(V, AndRHS);
3073 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3074 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3076 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3077 // If this is an integer truncation or change from signed-to-unsigned, and
3078 // if the source is an and/or with immediate, transform it. This
3079 // frequently occurs for bitfield accesses.
3080 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3081 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3082 CastOp->getNumOperands() == 2)
3083 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3084 if (CastOp->getOpcode() == Instruction::And) {
3085 // Change: and (cast (and X, C1) to T), C2
3086 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3087 // This will fold the two constants together, which may allow
3088 // other simplifications.
3089 Instruction *NewCast =
3090 CastInst::createInferredCast(CastOp->getOperand(0), I.getType(),
3091 CastOp->getName()+".shrunk");
3092 NewCast = InsertNewInstBefore(NewCast, I);
3093 // trunc_or_bitcast(C1)&C2
3094 Instruction::CastOps opc = (
3095 AndCI->getType()->getPrimitiveSizeInBits() ==
3096 I.getType()->getPrimitiveSizeInBits() ?
3097 Instruction::BitCast : Instruction::Trunc);
3098 Constant *C3 = ConstantExpr::getCast(opc, AndCI, I.getType());
3099 C3 = ConstantExpr::getAnd(C3, AndRHS);
3100 return BinaryOperator::createAnd(NewCast, C3);
3101 } else if (CastOp->getOpcode() == Instruction::Or) {
3102 // Change: and (cast (or X, C1) to T), C2
3103 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3104 Constant *C3 = ConstantExpr::getCast(AndCI, I.getType());
3105 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3106 return ReplaceInstUsesWith(I, AndRHS);
3111 // Try to fold constant and into select arguments.
3112 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3113 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3115 if (isa<PHINode>(Op0))
3116 if (Instruction *NV = FoldOpIntoPhi(I))
3120 Value *Op0NotVal = dyn_castNotVal(Op0);
3121 Value *Op1NotVal = dyn_castNotVal(Op1);
3123 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3124 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3126 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3127 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3128 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3129 I.getName()+".demorgan");
3130 InsertNewInstBefore(Or, I);
3131 return BinaryOperator::createNot(Or);
3135 Value *A = 0, *B = 0;
3136 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3137 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3138 return ReplaceInstUsesWith(I, Op1);
3139 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3140 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3141 return ReplaceInstUsesWith(I, Op0);
3143 if (Op0->hasOneUse() &&
3144 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3145 if (A == Op1) { // (A^B)&A -> A&(A^B)
3146 I.swapOperands(); // Simplify below
3147 std::swap(Op0, Op1);
3148 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3149 cast<BinaryOperator>(Op0)->swapOperands();
3150 I.swapOperands(); // Simplify below
3151 std::swap(Op0, Op1);
3154 if (Op1->hasOneUse() &&
3155 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3156 if (B == Op0) { // B&(A^B) -> B&(B^A)
3157 cast<BinaryOperator>(Op1)->swapOperands();
3160 if (A == Op0) { // A&(A^B) -> A & ~B
3161 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3162 InsertNewInstBefore(NotB, I);
3163 return BinaryOperator::createAnd(A, NotB);
3169 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
3170 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
3171 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3174 Value *LHSVal, *RHSVal;
3175 ConstantInt *LHSCst, *RHSCst;
3176 Instruction::BinaryOps LHSCC, RHSCC;
3177 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3178 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3179 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
3180 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3181 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3182 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3183 // Ensure that the larger constant is on the RHS.
3184 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3185 SetCondInst *LHS = cast<SetCondInst>(Op0);
3186 if (cast<ConstantBool>(Cmp)->getValue()) {
3187 std::swap(LHS, RHS);
3188 std::swap(LHSCst, RHSCst);
3189 std::swap(LHSCC, RHSCC);
3192 // At this point, we know we have have two setcc instructions
3193 // comparing a value against two constants and and'ing the result
3194 // together. Because of the above check, we know that we only have
3195 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3196 // FoldSetCCLogical check above), that the two constants are not
3198 assert(LHSCst != RHSCst && "Compares not folded above?");
3201 default: assert(0 && "Unknown integer condition code!");
3202 case Instruction::SetEQ:
3204 default: assert(0 && "Unknown integer condition code!");
3205 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
3206 case Instruction::SetGT: // (X == 13 & X > 15) -> false
3207 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3208 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
3209 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
3210 return ReplaceInstUsesWith(I, LHS);
3212 case Instruction::SetNE:
3214 default: assert(0 && "Unknown integer condition code!");
3215 case Instruction::SetLT:
3216 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
3217 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
3218 break; // (X != 13 & X < 15) -> no change
3219 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
3220 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
3221 return ReplaceInstUsesWith(I, RHS);
3222 case Instruction::SetNE:
3223 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
3224 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3225 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3226 LHSVal->getName()+".off");
3227 InsertNewInstBefore(Add, I);
3228 const Type *UnsType = Add->getType()->getUnsignedVersion();
3229 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3230 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
3231 AddCST = ConstantExpr::getCast(AddCST, UnsType);
3232 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
3234 break; // (X != 13 & X != 15) -> no change
3237 case Instruction::SetLT:
3239 default: assert(0 && "Unknown integer condition code!");
3240 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
3241 case Instruction::SetGT: // (X < 13 & X > 15) -> false
3242 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3243 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
3244 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
3245 return ReplaceInstUsesWith(I, LHS);
3247 case Instruction::SetGT:
3249 default: assert(0 && "Unknown integer condition code!");
3250 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
3251 return ReplaceInstUsesWith(I, LHS);
3252 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
3253 return ReplaceInstUsesWith(I, RHS);
3254 case Instruction::SetNE:
3255 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
3256 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
3257 break; // (X > 13 & X != 15) -> no change
3258 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
3259 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
3265 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3266 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
3267 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
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::createInferredCast(NewOp, I.getType());
3282 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3283 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3284 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3285 if (SI0->getOpcode() == SI1->getOpcode() &&
3286 SI0->getOperand(1) == SI1->getOperand(1) &&
3287 (SI0->hasOneUse() || SI1->hasOneUse())) {
3288 Instruction *NewOp =
3289 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3291 SI0->getName()), I);
3292 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3296 return Changed ? &I : 0;
3299 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3300 /// in the result. If it does, and if the specified byte hasn't been filled in
3301 /// yet, fill it in and return false.
3302 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3303 Instruction *I = dyn_cast<Instruction>(V);
3304 if (I == 0) return true;
3306 // If this is an or instruction, it is an inner node of the bswap.
3307 if (I->getOpcode() == Instruction::Or)
3308 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3309 CollectBSwapParts(I->getOperand(1), ByteValues);
3311 // If this is a shift by a constant int, and it is "24", then its operand
3312 // defines a byte. We only handle unsigned types here.
3313 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3314 // Not shifting the entire input by N-1 bytes?
3315 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3316 8*(ByteValues.size()-1))
3320 if (I->getOpcode() == Instruction::Shl) {
3321 // X << 24 defines the top byte with the lowest of the input bytes.
3322 DestNo = ByteValues.size()-1;
3324 // X >>u 24 defines the low byte with the highest of the input bytes.
3328 // If the destination byte value is already defined, the values are or'd
3329 // together, which isn't a bswap (unless it's an or of the same bits).
3330 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3332 ByteValues[DestNo] = I->getOperand(0);
3336 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3338 Value *Shift = 0, *ShiftLHS = 0;
3339 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3340 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3341 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3343 Instruction *SI = cast<Instruction>(Shift);
3345 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3346 if (ShiftAmt->getZExtValue() & 7 ||
3347 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3350 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3352 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3353 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3355 // Unknown mask for bswap.
3356 if (DestByte == ByteValues.size()) return true;
3358 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3360 if (SI->getOpcode() == Instruction::Shl)
3361 SrcByte = DestByte - ShiftBytes;
3363 SrcByte = DestByte + ShiftBytes;
3365 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3366 if (SrcByte != ByteValues.size()-DestByte-1)
3369 // If the destination byte value is already defined, the values are or'd
3370 // together, which isn't a bswap (unless it's an or of the same bits).
3371 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3373 ByteValues[DestByte] = SI->getOperand(0);
3377 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3378 /// If so, insert the new bswap intrinsic and return it.
3379 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3380 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3381 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3384 /// ByteValues - For each byte of the result, we keep track of which value
3385 /// defines each byte.
3386 std::vector<Value*> ByteValues;
3387 ByteValues.resize(I.getType()->getPrimitiveSize());
3389 // Try to find all the pieces corresponding to the bswap.
3390 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3391 CollectBSwapParts(I.getOperand(1), ByteValues))
3394 // Check to see if all of the bytes come from the same value.
3395 Value *V = ByteValues[0];
3396 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3398 // Check to make sure that all of the bytes come from the same value.
3399 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3400 if (ByteValues[i] != V)
3403 // If they do then *success* we can turn this into a bswap. Figure out what
3404 // bswap to make it into.
3405 Module *M = I.getParent()->getParent()->getParent();
3406 const char *FnName = 0;
3407 if (I.getType() == Type::UShortTy)
3408 FnName = "llvm.bswap.i16";
3409 else if (I.getType() == Type::UIntTy)
3410 FnName = "llvm.bswap.i32";
3411 else if (I.getType() == Type::ULongTy)
3412 FnName = "llvm.bswap.i64";
3414 assert(0 && "Unknown integer type!");
3415 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3417 return new CallInst(F, V);
3421 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3422 bool Changed = SimplifyCommutative(I);
3423 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3425 if (isa<UndefValue>(Op1))
3426 return ReplaceInstUsesWith(I, // X | undef -> -1
3427 ConstantIntegral::getAllOnesValue(I.getType()));
3431 return ReplaceInstUsesWith(I, Op0);
3433 // See if we can simplify any instructions used by the instruction whose sole
3434 // purpose is to compute bits we don't care about.
3435 uint64_t KnownZero, KnownOne;
3436 if (!isa<PackedType>(I.getType()) &&
3437 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3438 KnownZero, KnownOne))
3442 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3443 ConstantInt *C1 = 0; Value *X = 0;
3444 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3445 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3446 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3448 InsertNewInstBefore(Or, I);
3449 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3452 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3453 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3454 std::string Op0Name = Op0->getName(); Op0->setName("");
3455 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3456 InsertNewInstBefore(Or, I);
3457 return BinaryOperator::createXor(Or,
3458 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3461 // Try to fold constant and into select arguments.
3462 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3463 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3465 if (isa<PHINode>(Op0))
3466 if (Instruction *NV = FoldOpIntoPhi(I))
3470 Value *A = 0, *B = 0;
3471 ConstantInt *C1 = 0, *C2 = 0;
3473 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3474 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3475 return ReplaceInstUsesWith(I, Op1);
3476 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3477 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3478 return ReplaceInstUsesWith(I, Op0);
3480 // (A | B) | C and A | (B | C) -> bswap if possible.
3481 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3482 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3483 match(Op1, m_Or(m_Value(), m_Value())) ||
3484 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3485 match(Op1, m_Shift(m_Value(), m_Value())))) {
3486 if (Instruction *BSwap = MatchBSwap(I))
3490 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3491 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3492 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3493 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3495 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3498 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3499 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3500 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3501 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3503 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3506 // (A & C1)|(B & C2)
3507 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3508 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3510 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3511 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3514 // If we have: ((V + N) & C1) | (V & C2)
3515 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3516 // replace with V+N.
3517 if (C1 == ConstantExpr::getNot(C2)) {
3518 Value *V1 = 0, *V2 = 0;
3519 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3520 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3521 // Add commutes, try both ways.
3522 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3523 return ReplaceInstUsesWith(I, A);
3524 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3525 return ReplaceInstUsesWith(I, A);
3527 // Or commutes, try both ways.
3528 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3529 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3530 // Add commutes, try both ways.
3531 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3532 return ReplaceInstUsesWith(I, B);
3533 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3534 return ReplaceInstUsesWith(I, B);
3539 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3540 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3541 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3542 if (SI0->getOpcode() == SI1->getOpcode() &&
3543 SI0->getOperand(1) == SI1->getOperand(1) &&
3544 (SI0->hasOneUse() || SI1->hasOneUse())) {
3545 Instruction *NewOp =
3546 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3548 SI0->getName()), I);
3549 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3553 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3554 if (A == Op1) // ~A | A == -1
3555 return ReplaceInstUsesWith(I,
3556 ConstantIntegral::getAllOnesValue(I.getType()));
3560 // Note, A is still live here!
3561 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3563 return ReplaceInstUsesWith(I,
3564 ConstantIntegral::getAllOnesValue(I.getType()));
3566 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3567 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3568 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3569 I.getName()+".demorgan"), I);
3570 return BinaryOperator::createNot(And);
3574 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
3575 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
3576 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3579 Value *LHSVal, *RHSVal;
3580 ConstantInt *LHSCst, *RHSCst;
3581 Instruction::BinaryOps LHSCC, RHSCC;
3582 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3583 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3584 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
3585 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3586 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3587 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3588 // Ensure that the larger constant is on the RHS.
3589 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3590 SetCondInst *LHS = cast<SetCondInst>(Op0);
3591 if (cast<ConstantBool>(Cmp)->getValue()) {
3592 std::swap(LHS, RHS);
3593 std::swap(LHSCst, RHSCst);
3594 std::swap(LHSCC, RHSCC);
3597 // At this point, we know we have have two setcc instructions
3598 // comparing a value against two constants and or'ing the result
3599 // together. Because of the above check, we know that we only have
3600 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3601 // FoldSetCCLogical check above), that the two constants are not
3603 assert(LHSCst != RHSCst && "Compares not folded above?");
3606 default: assert(0 && "Unknown integer condition code!");
3607 case Instruction::SetEQ:
3609 default: assert(0 && "Unknown integer condition code!");
3610 case Instruction::SetEQ:
3611 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3612 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3613 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3614 LHSVal->getName()+".off");
3615 InsertNewInstBefore(Add, I);
3616 const Type *UnsType = Add->getType()->getUnsignedVersion();
3617 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3618 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3619 AddCST = ConstantExpr::getCast(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 const Type *SrcTy = Op0C->getOperand(0)->getType();
3671 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3672 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3673 // Only do this if the casts both really cause code to be generated.
3674 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3675 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3676 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3677 Op1C->getOperand(0),
3679 InsertNewInstBefore(NewOp, I);
3680 return CastInst::createInferredCast(NewOp, I.getType());
3685 return Changed ? &I : 0;
3688 // XorSelf - Implements: X ^ X --> 0
3691 XorSelf(Value *rhs) : RHS(rhs) {}
3692 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3693 Instruction *apply(BinaryOperator &Xor) const {
3699 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3700 bool Changed = SimplifyCommutative(I);
3701 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3703 if (isa<UndefValue>(Op1))
3704 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3706 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3707 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3708 assert(Result == &I && "AssociativeOpt didn't work?");
3709 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3712 // See if we can simplify any instructions used by the instruction whose sole
3713 // purpose is to compute bits we don't care about.
3714 uint64_t KnownZero, KnownOne;
3715 if (!isa<PackedType>(I.getType()) &&
3716 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3717 KnownZero, KnownOne))
3720 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3721 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3722 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
3723 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
3724 if (RHS == ConstantBool::getTrue() && SCI->hasOneUse())
3725 return new SetCondInst(SCI->getInverseCondition(),
3726 SCI->getOperand(0), SCI->getOperand(1));
3728 // ~(c-X) == X-c-1 == X+(-c-1)
3729 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3730 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3731 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3732 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3733 ConstantInt::get(I.getType(), 1));
3734 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3737 // ~(~X & Y) --> (X | ~Y)
3738 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3739 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3740 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3742 BinaryOperator::createNot(Op0I->getOperand(1),
3743 Op0I->getOperand(1)->getName()+".not");
3744 InsertNewInstBefore(NotY, I);
3745 return BinaryOperator::createOr(Op0NotVal, NotY);
3749 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3750 if (Op0I->getOpcode() == Instruction::Add) {
3751 // ~(X-c) --> (-c-1)-X
3752 if (RHS->isAllOnesValue()) {
3753 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3754 return BinaryOperator::createSub(
3755 ConstantExpr::getSub(NegOp0CI,
3756 ConstantInt::get(I.getType(), 1)),
3757 Op0I->getOperand(0));
3759 } else if (Op0I->getOpcode() == Instruction::Or) {
3760 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3761 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3762 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3763 // Anything in both C1 and C2 is known to be zero, remove it from
3765 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3766 NewRHS = ConstantExpr::getAnd(NewRHS,
3767 ConstantExpr::getNot(CommonBits));
3768 WorkList.push_back(Op0I);
3769 I.setOperand(0, Op0I->getOperand(0));
3770 I.setOperand(1, NewRHS);
3776 // Try to fold constant and into select arguments.
3777 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3778 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3780 if (isa<PHINode>(Op0))
3781 if (Instruction *NV = FoldOpIntoPhi(I))
3785 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3787 return ReplaceInstUsesWith(I,
3788 ConstantIntegral::getAllOnesValue(I.getType()));
3790 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3792 return ReplaceInstUsesWith(I,
3793 ConstantIntegral::getAllOnesValue(I.getType()));
3795 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3796 if (Op1I->getOpcode() == Instruction::Or) {
3797 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3798 Op1I->swapOperands();
3800 std::swap(Op0, Op1);
3801 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3802 I.swapOperands(); // Simplified below.
3803 std::swap(Op0, Op1);
3805 } else if (Op1I->getOpcode() == Instruction::Xor) {
3806 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3807 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3808 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3809 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3810 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3811 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3812 Op1I->swapOperands();
3813 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3814 I.swapOperands(); // Simplified below.
3815 std::swap(Op0, Op1);
3819 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3820 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3821 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3822 Op0I->swapOperands();
3823 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3824 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3825 InsertNewInstBefore(NotB, I);
3826 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3828 } else if (Op0I->getOpcode() == Instruction::Xor) {
3829 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3830 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3831 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3832 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3833 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3834 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3835 Op0I->swapOperands();
3836 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3837 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3838 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3839 InsertNewInstBefore(N, I);
3840 return BinaryOperator::createAnd(N, Op1);
3844 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
3845 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
3846 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3849 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3850 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3851 const Type *SrcTy = Op0C->getOperand(0)->getType();
3852 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3853 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3854 // Only do this if the casts both really cause code to be generated.
3855 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3856 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3857 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
3858 Op1C->getOperand(0),
3860 InsertNewInstBefore(NewOp, I);
3861 return CastInst::createInferredCast(NewOp, I.getType());
3865 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
3866 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3867 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3868 if (SI0->getOpcode() == SI1->getOpcode() &&
3869 SI0->getOperand(1) == SI1->getOperand(1) &&
3870 (SI0->hasOneUse() || SI1->hasOneUse())) {
3871 Instruction *NewOp =
3872 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
3874 SI0->getName()), I);
3875 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3879 return Changed ? &I : 0;
3882 static bool isPositive(ConstantInt *C) {
3883 return C->getSExtValue() >= 0;
3886 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
3887 /// overflowed for this type.
3888 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3890 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
3892 if (In1->getType()->isUnsigned())
3893 return cast<ConstantInt>(Result)->getZExtValue() <
3894 cast<ConstantInt>(In1)->getZExtValue();
3895 if (isPositive(In1) != isPositive(In2))
3897 if (isPositive(In1))
3898 return cast<ConstantInt>(Result)->getSExtValue() <
3899 cast<ConstantInt>(In1)->getSExtValue();
3900 return cast<ConstantInt>(Result)->getSExtValue() >
3901 cast<ConstantInt>(In1)->getSExtValue();
3904 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3905 /// code necessary to compute the offset from the base pointer (without adding
3906 /// in the base pointer). Return the result as a signed integer of intptr size.
3907 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3908 TargetData &TD = IC.getTargetData();
3909 gep_type_iterator GTI = gep_type_begin(GEP);
3910 const Type *UIntPtrTy = TD.getIntPtrType();
3911 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3912 Value *Result = Constant::getNullValue(SIntPtrTy);
3914 // Build a mask for high order bits.
3915 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3917 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3918 Value *Op = GEP->getOperand(i);
3919 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3920 Constant *Scale = ConstantExpr::getCast(ConstantInt::get(UIntPtrTy, Size),
3922 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3923 if (!OpC->isNullValue()) {
3924 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
3925 Scale = ConstantExpr::getMul(OpC, Scale);
3926 if (Constant *RC = dyn_cast<Constant>(Result))
3927 Result = ConstantExpr::getAdd(RC, Scale);
3929 // Emit an add instruction.
3930 Result = IC.InsertNewInstBefore(
3931 BinaryOperator::createAdd(Result, Scale,
3932 GEP->getName()+".offs"), I);
3936 // Convert to correct type.
3937 Op = IC.InsertNewInstBefore(CastInst::createInferredCast(Op, SIntPtrTy,
3938 Op->getName()+".c"), I);
3940 // We'll let instcombine(mul) convert this to a shl if possible.
3941 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3942 GEP->getName()+".idx"), I);
3944 // Emit an add instruction.
3945 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3946 GEP->getName()+".offs"), I);
3952 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3953 /// else. At this point we know that the GEP is on the LHS of the comparison.
3954 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3955 Instruction::BinaryOps Cond,
3957 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3959 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3960 if (isa<PointerType>(CI->getOperand(0)->getType()))
3961 RHS = CI->getOperand(0);
3963 Value *PtrBase = GEPLHS->getOperand(0);
3964 if (PtrBase == RHS) {
3965 // As an optimization, we don't actually have to compute the actual value of
3966 // OFFSET if this is a seteq or setne comparison, just return whether each
3967 // index is zero or not.
3968 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3969 Instruction *InVal = 0;
3970 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3971 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3973 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3974 if (isa<UndefValue>(C)) // undef index -> undef.
3975 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3976 if (C->isNullValue())
3978 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3979 EmitIt = false; // This is indexing into a zero sized array?
3980 } else if (isa<ConstantInt>(C))
3981 return ReplaceInstUsesWith(I, // No comparison is needed here.
3982 ConstantBool::get(Cond == Instruction::SetNE));
3987 new SetCondInst(Cond, GEPLHS->getOperand(i),
3988 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3992 InVal = InsertNewInstBefore(InVal, I);
3993 InsertNewInstBefore(Comp, I);
3994 if (Cond == Instruction::SetNE) // True if any are unequal
3995 InVal = BinaryOperator::createOr(InVal, Comp);
3996 else // True if all are equal
3997 InVal = BinaryOperator::createAnd(InVal, Comp);
4005 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
4006 ConstantBool::get(Cond == Instruction::SetEQ));
4009 // Only lower this if the setcc is the only user of the GEP or if we expect
4010 // the result to fold to a constant!
4011 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4012 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4013 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4014 return new SetCondInst(Cond, Offset,
4015 Constant::getNullValue(Offset->getType()));
4017 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4018 // If the base pointers are different, but the indices are the same, just
4019 // compare the base pointer.
4020 if (PtrBase != GEPRHS->getOperand(0)) {
4021 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4022 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4023 GEPRHS->getOperand(0)->getType();
4025 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4026 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4027 IndicesTheSame = false;
4031 // If all indices are the same, just compare the base pointers.
4033 return new SetCondInst(Cond, GEPLHS->getOperand(0),
4034 GEPRHS->getOperand(0));
4036 // Otherwise, the base pointers are different and the indices are
4037 // different, bail out.
4041 // If one of the GEPs has all zero indices, recurse.
4042 bool AllZeros = true;
4043 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4044 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4045 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4050 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
4051 SetCondInst::getSwappedCondition(Cond), I);
4053 // If the other GEP has all zero indices, recurse.
4055 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4056 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4057 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4062 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4064 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4065 // If the GEPs only differ by one index, compare it.
4066 unsigned NumDifferences = 0; // Keep track of # differences.
4067 unsigned DiffOperand = 0; // The operand that differs.
4068 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4069 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4070 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4071 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4072 // Irreconcilable differences.
4076 if (NumDifferences++) break;
4081 if (NumDifferences == 0) // SAME GEP?
4082 return ReplaceInstUsesWith(I, // No comparison is needed here.
4083 ConstantBool::get(Cond == Instruction::SetEQ));
4084 else if (NumDifferences == 1) {
4085 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4086 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4088 // Convert the operands to signed values to make sure to perform a
4089 // signed comparison.
4090 const Type *NewTy = LHSV->getType()->getSignedVersion();
4091 if (LHSV->getType() != NewTy)
4092 LHSV = InsertCastBefore(LHSV, NewTy, I);
4093 if (RHSV->getType() != NewTy)
4094 RHSV = InsertCastBefore(RHSV, NewTy, I);
4095 return new SetCondInst(Cond, LHSV, RHSV);
4099 // Only lower this if the setcc is the only user of the GEP or if we expect
4100 // the result to fold to a constant!
4101 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4102 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4103 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4104 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4105 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4106 return new SetCondInst(Cond, L, R);
4113 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
4114 bool Changed = SimplifyCommutative(I);
4115 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4116 const Type *Ty = Op0->getType();
4120 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
4122 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
4123 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
4125 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4126 // addresses never equal each other! We already know that Op0 != Op1.
4127 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4128 isa<ConstantPointerNull>(Op0)) &&
4129 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4130 isa<ConstantPointerNull>(Op1)))
4131 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
4133 // setcc's with boolean values can always be turned into bitwise operations
4134 if (Ty == Type::BoolTy) {
4135 switch (I.getOpcode()) {
4136 default: assert(0 && "Invalid setcc instruction!");
4137 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
4138 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4139 InsertNewInstBefore(Xor, I);
4140 return BinaryOperator::createNot(Xor);
4142 case Instruction::SetNE:
4143 return BinaryOperator::createXor(Op0, Op1);
4145 case Instruction::SetGT:
4146 std::swap(Op0, Op1); // Change setgt -> setlt
4148 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
4149 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4150 InsertNewInstBefore(Not, I);
4151 return BinaryOperator::createAnd(Not, Op1);
4153 case Instruction::SetGE:
4154 std::swap(Op0, Op1); // Change setge -> setle
4156 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
4157 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4158 InsertNewInstBefore(Not, I);
4159 return BinaryOperator::createOr(Not, Op1);
4164 // See if we are doing a comparison between a constant and an instruction that
4165 // can be folded into the comparison.
4166 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4167 // Check to see if we are comparing against the minimum or maximum value...
4168 if (CI->isMinValue()) {
4169 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
4170 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4171 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
4172 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4173 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
4174 return BinaryOperator::createSetEQ(Op0, Op1);
4175 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
4176 return BinaryOperator::createSetNE(Op0, Op1);
4178 } else if (CI->isMaxValue()) {
4179 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
4180 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4181 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
4182 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4183 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
4184 return BinaryOperator::createSetEQ(Op0, Op1);
4185 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
4186 return BinaryOperator::createSetNE(Op0, Op1);
4188 // Comparing against a value really close to min or max?
4189 } else if (isMinValuePlusOne(CI)) {
4190 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
4191 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
4192 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
4193 return BinaryOperator::createSetNE(Op0, SubOne(CI));
4195 } else if (isMaxValueMinusOne(CI)) {
4196 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
4197 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
4198 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
4199 return BinaryOperator::createSetNE(Op0, AddOne(CI));
4202 // If we still have a setle or setge instruction, turn it into the
4203 // appropriate setlt or setgt instruction. Since the border cases have
4204 // already been handled above, this requires little checking.
4206 if (I.getOpcode() == Instruction::SetLE)
4207 return BinaryOperator::createSetLT(Op0, AddOne(CI));
4208 if (I.getOpcode() == Instruction::SetGE)
4209 return BinaryOperator::createSetGT(Op0, SubOne(CI));
4212 // See if we can fold the comparison based on bits known to be zero or one
4214 uint64_t KnownZero, KnownOne;
4215 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4216 KnownZero, KnownOne, 0))
4219 // Given the known and unknown bits, compute a range that the LHS could be
4221 if (KnownOne | KnownZero) {
4222 if (Ty->isUnsigned()) { // Unsigned comparison.
4224 uint64_t RHSVal = CI->getZExtValue();
4225 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4227 switch (I.getOpcode()) { // LE/GE have been folded already.
4228 default: assert(0 && "Unknown setcc opcode!");
4229 case Instruction::SetEQ:
4230 if (Max < RHSVal || Min > RHSVal)
4231 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4233 case Instruction::SetNE:
4234 if (Max < RHSVal || Min > RHSVal)
4235 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4237 case Instruction::SetLT:
4239 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4241 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4243 case Instruction::SetGT:
4245 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4247 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4250 } else { // Signed comparison.
4252 int64_t RHSVal = CI->getSExtValue();
4253 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4255 switch (I.getOpcode()) { // LE/GE have been folded already.
4256 default: assert(0 && "Unknown setcc opcode!");
4257 case Instruction::SetEQ:
4258 if (Max < RHSVal || Min > RHSVal)
4259 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4261 case Instruction::SetNE:
4262 if (Max < RHSVal || Min > RHSVal)
4263 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4265 case Instruction::SetLT:
4267 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4269 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4271 case Instruction::SetGT:
4273 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4275 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4281 // Since the RHS is a constantInt (CI), if the left hand side is an
4282 // instruction, see if that instruction also has constants so that the
4283 // instruction can be folded into the setcc
4284 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4285 switch (LHSI->getOpcode()) {
4286 case Instruction::And:
4287 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4288 LHSI->getOperand(0)->hasOneUse()) {
4289 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4291 // If an operand is an AND of a truncating cast, we can widen the
4292 // and/compare to be the input width without changing the value
4293 // produced, eliminating a cast.
4294 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4295 // We can do this transformation if either the AND constant does not
4296 // have its sign bit set or if it is an equality comparison.
4297 // Extending a relational comparison when we're checking the sign
4298 // bit would not work.
4299 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4301 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4302 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4303 ConstantInt *NewCST;
4305 if (Cast->getOperand(0)->getType()->isSigned()) {
4306 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4307 AndCST->getZExtValue());
4308 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4309 CI->getZExtValue());
4311 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4312 AndCST->getZExtValue());
4313 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4314 CI->getZExtValue());
4316 Instruction *NewAnd =
4317 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4319 InsertNewInstBefore(NewAnd, I);
4320 return new SetCondInst(I.getOpcode(), NewAnd, NewCI);
4324 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4325 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4326 // happens a LOT in code produced by the C front-end, for bitfield
4328 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4330 // Check to see if there is a noop-cast between the shift and the and.
4332 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4333 if (CI->getOperand(0)->getType()->isIntegral() &&
4334 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
4335 CI->getType()->getPrimitiveSizeInBits())
4336 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4340 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4341 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4342 const Type *AndTy = AndCST->getType(); // Type of the and.
4344 // We can fold this as long as we can't shift unknown bits
4345 // into the mask. This can only happen with signed shift
4346 // rights, as they sign-extend.
4348 bool CanFold = Shift->isLogicalShift();
4350 // To test for the bad case of the signed shr, see if any
4351 // of the bits shifted in could be tested after the mask.
4352 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4353 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4355 Constant *OShAmt = ConstantInt::get(Type::UByteTy, ShAmtVal);
4357 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4359 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4365 if (Shift->getOpcode() == Instruction::Shl)
4366 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4368 NewCst = ConstantExpr::getShl(CI, ShAmt);
4370 // Check to see if we are shifting out any of the bits being
4372 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4373 // If we shifted bits out, the fold is not going to work out.
4374 // As a special case, check to see if this means that the
4375 // result is always true or false now.
4376 if (I.getOpcode() == Instruction::SetEQ)
4377 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4378 if (I.getOpcode() == Instruction::SetNE)
4379 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4381 I.setOperand(1, NewCst);
4382 Constant *NewAndCST;
4383 if (Shift->getOpcode() == Instruction::Shl)
4384 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4386 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4387 LHSI->setOperand(1, NewAndCST);
4389 LHSI->setOperand(0, Shift->getOperand(0));
4391 Value *NewCast = InsertCastBefore(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 // Make sure we insert a logical shift.
4415 Constant *NewAndCST = AndCST;
4416 if (AndCST->getType()->isSigned())
4417 NewAndCST = ConstantExpr::getCast(AndCST,
4418 AndCST->getType()->getUnsignedVersion());
4419 NS = new ShiftInst(Instruction::LShr, NewAndCST,
4420 Shift->getOperand(1), "tmp");
4422 InsertNewInstBefore(cast<Instruction>(NS), I);
4424 // If C's sign doesn't agree with the and, insert a cast now.
4425 if (NS->getType() != LHSI->getType())
4426 NS = InsertCastBefore(NS, LHSI->getType(), I);
4428 Value *ShiftOp = Shift->getOperand(0);
4429 if (ShiftOp->getType() != LHSI->getType())
4430 ShiftOp = InsertCastBefore(ShiftOp, LHSI->getType(), I);
4432 // Compute X & (C << Y).
4433 Instruction *NewAnd =
4434 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4435 InsertNewInstBefore(NewAnd, I);
4437 I.setOperand(0, NewAnd);
4443 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
4444 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4445 if (I.isEquality()) {
4446 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4448 // Check that the shift amount is in range. If not, don't perform
4449 // undefined shifts. When the shift is visited it will be
4451 if (ShAmt->getZExtValue() >= TypeBits)
4454 // If we are comparing against bits always shifted out, the
4455 // comparison cannot succeed.
4457 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4458 if (Comp != CI) {// Comparing against a bit that we know is zero.
4459 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4460 Constant *Cst = ConstantBool::get(IsSetNE);
4461 return ReplaceInstUsesWith(I, Cst);
4464 if (LHSI->hasOneUse()) {
4465 // Otherwise strength reduce the shift into an and.
4466 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4467 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4470 if (CI->getType()->isUnsigned()) {
4471 Mask = ConstantInt::get(CI->getType(), Val);
4472 } else if (ShAmtVal != 0) {
4473 Mask = ConstantInt::get(CI->getType(), Val);
4475 Mask = ConstantInt::getAllOnesValue(CI->getType());
4479 BinaryOperator::createAnd(LHSI->getOperand(0),
4480 Mask, LHSI->getName()+".mask");
4481 Value *And = InsertNewInstBefore(AndI, I);
4482 return new SetCondInst(I.getOpcode(), And,
4483 ConstantExpr::getLShr(CI, ShAmt));
4489 case Instruction::LShr: // (setcc (shr X, ShAmt), CI)
4490 case Instruction::AShr:
4491 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4492 if (I.isEquality()) {
4493 // Check that the shift amount is in range. If not, don't perform
4494 // undefined shifts. When the shift is visited it will be
4496 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4497 if (ShAmt->getZExtValue() >= TypeBits)
4500 // If we are comparing against bits always shifted out, the
4501 // comparison cannot succeed.
4503 if (CI->getType()->isUnsigned())
4504 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4507 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4510 if (Comp != CI) {// Comparing against a bit that we know is zero.
4511 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4512 Constant *Cst = ConstantBool::get(IsSetNE);
4513 return ReplaceInstUsesWith(I, Cst);
4516 if (LHSI->hasOneUse() || CI->isNullValue()) {
4517 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4519 // Otherwise strength reduce the shift into an and.
4520 uint64_t Val = ~0ULL; // All ones.
4521 Val <<= ShAmtVal; // Shift over to the right spot.
4524 if (CI->getType()->isUnsigned()) {
4525 Val &= ~0ULL >> (64-TypeBits);
4526 Mask = ConstantInt::get(CI->getType(), Val);
4528 Mask = ConstantInt::get(CI->getType(), Val);
4532 BinaryOperator::createAnd(LHSI->getOperand(0),
4533 Mask, LHSI->getName()+".mask");
4534 Value *And = InsertNewInstBefore(AndI, I);
4535 return new SetCondInst(I.getOpcode(), And,
4536 ConstantExpr::getShl(CI, ShAmt));
4542 case Instruction::SDiv:
4543 case Instruction::UDiv:
4544 // Fold: setcc ([us]div X, C1), C2 -> range test
4545 // Fold this div into the comparison, producing a range check.
4546 // Determine, based on the divide type, what the range is being
4547 // checked. If there is an overflow on the low or high side, remember
4548 // it, otherwise compute the range [low, hi) bounding the new value.
4549 // See: InsertRangeTest above for the kinds of replacements possible.
4550 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4551 // FIXME: If the operand types don't match the type of the divide
4552 // then don't attempt this transform. The code below doesn't have the
4553 // logic to deal with a signed divide and an unsigned compare (and
4554 // vice versa). This is because (x /s C1) <s C2 produces different
4555 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4556 // (x /u C1) <u C2. Simply casting the operands and result won't
4557 // work. :( The if statement below tests that condition and bails
4559 const Type *DivRHSTy = DivRHS->getType();
4560 unsigned DivOpCode = LHSI->getOpcode();
4561 if (I.isEquality() &&
4562 ((DivOpCode == Instruction::SDiv && DivRHSTy->isUnsigned()) ||
4563 (DivOpCode == Instruction::UDiv && DivRHSTy->isSigned())))
4566 // Initialize the variables that will indicate the nature of the
4568 bool LoOverflow = false, HiOverflow = false;
4569 ConstantInt *LoBound = 0, *HiBound = 0;
4571 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4572 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4573 // C2 (CI). By solving for X we can turn this into a range check
4574 // instead of computing a divide.
4576 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4578 // Determine if the product overflows by seeing if the product is
4579 // not equal to the divide. Make sure we do the same kind of divide
4580 // as in the LHS instruction that we're folding.
4581 bool ProdOV = !DivRHS->isNullValue() &&
4582 (DivOpCode == Instruction::SDiv ?
4583 ConstantExpr::getSDiv(Prod, DivRHS) :
4584 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4586 // Get the SetCC opcode
4587 Instruction::BinaryOps Opcode = I.getOpcode();
4589 if (DivRHS->isNullValue()) {
4590 // Don't hack on divide by zeros!
4591 } else if (DivOpCode == Instruction::UDiv) { // udiv
4593 LoOverflow = ProdOV;
4594 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4595 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4596 if (CI->isNullValue()) { // (X / pos) op 0
4598 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4600 } else if (isPositive(CI)) { // (X / pos) op pos
4602 LoOverflow = ProdOV;
4603 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4604 } else { // (X / pos) op neg
4605 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4606 LoOverflow = AddWithOverflow(LoBound, Prod,
4607 cast<ConstantInt>(DivRHSH));
4609 HiOverflow = ProdOV;
4611 } else { // Divisor is < 0.
4612 if (CI->isNullValue()) { // (X / neg) op 0
4613 LoBound = AddOne(DivRHS);
4614 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4615 if (HiBound == DivRHS)
4616 LoBound = 0; // - INTMIN = INTMIN
4617 } else if (isPositive(CI)) { // (X / neg) op pos
4618 HiOverflow = LoOverflow = ProdOV;
4620 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4621 HiBound = AddOne(Prod);
4622 } else { // (X / neg) op neg
4624 LoOverflow = HiOverflow = ProdOV;
4625 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4628 // Dividing by a negate swaps the condition.
4629 Opcode = SetCondInst::getSwappedCondition(Opcode);
4633 Value *X = LHSI->getOperand(0);
4635 default: assert(0 && "Unhandled setcc opcode!");
4636 case Instruction::SetEQ:
4637 if (LoOverflow && HiOverflow)
4638 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4639 else if (HiOverflow)
4640 return new SetCondInst(Instruction::SetGE, X, LoBound);
4641 else if (LoOverflow)
4642 return new SetCondInst(Instruction::SetLT, X, HiBound);
4644 return InsertRangeTest(X, LoBound, HiBound, true, I);
4645 case Instruction::SetNE:
4646 if (LoOverflow && HiOverflow)
4647 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4648 else if (HiOverflow)
4649 return new SetCondInst(Instruction::SetLT, X, LoBound);
4650 else if (LoOverflow)
4651 return new SetCondInst(Instruction::SetGE, X, HiBound);
4653 return InsertRangeTest(X, LoBound, HiBound, false, I);
4654 case Instruction::SetLT:
4656 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4657 return new SetCondInst(Instruction::SetLT, X, LoBound);
4658 case Instruction::SetGT:
4660 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4661 return new SetCondInst(Instruction::SetGE, X, HiBound);
4668 // Simplify seteq and setne instructions with integer constant RHS.
4669 if (I.isEquality()) {
4670 bool isSetNE = I.getOpcode() == Instruction::SetNE;
4672 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4673 // the second operand is a constant, simplify a bit.
4674 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4675 switch (BO->getOpcode()) {
4676 case Instruction::SRem:
4677 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4678 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4680 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4681 if (V > 1 && isPowerOf2_64(V)) {
4682 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4683 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4684 return BinaryOperator::create(I.getOpcode(), NewRem,
4685 Constant::getNullValue(BO->getType()));
4689 case Instruction::Add:
4690 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4691 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4692 if (BO->hasOneUse())
4693 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4694 ConstantExpr::getSub(CI, BOp1C));
4695 } else if (CI->isNullValue()) {
4696 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4697 // efficiently invertible, or if the add has just this one use.
4698 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4700 if (Value *NegVal = dyn_castNegVal(BOp1))
4701 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
4702 else if (Value *NegVal = dyn_castNegVal(BOp0))
4703 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
4704 else if (BO->hasOneUse()) {
4705 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4707 InsertNewInstBefore(Neg, I);
4708 return new SetCondInst(I.getOpcode(), BOp0, Neg);
4712 case Instruction::Xor:
4713 // For the xor case, we can xor two constants together, eliminating
4714 // the explicit xor.
4715 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4716 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
4717 ConstantExpr::getXor(CI, BOC));
4720 case Instruction::Sub:
4721 // Replace (([sub|xor] A, B) != 0) with (A != B)
4722 if (CI->isNullValue())
4723 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4727 case Instruction::Or:
4728 // If bits are being or'd in that are not present in the constant we
4729 // are comparing against, then the comparison could never succeed!
4730 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4731 Constant *NotCI = ConstantExpr::getNot(CI);
4732 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4733 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4737 case Instruction::And:
4738 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4739 // If bits are being compared against that are and'd out, then the
4740 // comparison can never succeed!
4741 if (!ConstantExpr::getAnd(CI,
4742 ConstantExpr::getNot(BOC))->isNullValue())
4743 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4745 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4746 if (CI == BOC && isOneBitSet(CI))
4747 return new SetCondInst(isSetNE ? Instruction::SetEQ :
4748 Instruction::SetNE, Op0,
4749 Constant::getNullValue(CI->getType()));
4751 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
4752 // to be a signed value as appropriate.
4753 if (isSignBit(BOC)) {
4754 Value *X = BO->getOperand(0);
4755 // If 'X' is not signed, insert a cast now...
4756 if (!BOC->getType()->isSigned()) {
4757 const Type *DestTy = BOC->getType()->getSignedVersion();
4758 X = InsertCastBefore(X, DestTy, I);
4760 return new SetCondInst(isSetNE ? Instruction::SetLT :
4761 Instruction::SetGE, X,
4762 Constant::getNullValue(X->getType()));
4765 // ((X & ~7) == 0) --> X < 8
4766 if (CI->isNullValue() && isHighOnes(BOC)) {
4767 Value *X = BO->getOperand(0);
4768 Constant *NegX = ConstantExpr::getNeg(BOC);
4770 // If 'X' is signed, insert a cast now.
4771 if (NegX->getType()->isSigned()) {
4772 const Type *DestTy = NegX->getType()->getUnsignedVersion();
4773 X = InsertCastBefore(X, DestTy, I);
4774 NegX = ConstantExpr::getCast(NegX, DestTy);
4777 return new SetCondInst(isSetNE ? Instruction::SetGE :
4778 Instruction::SetLT, X, NegX);
4784 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
4785 // Handle set{eq|ne} <intrinsic>, intcst.
4786 switch (II->getIntrinsicID()) {
4788 case Intrinsic::bswap_i16: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4789 WorkList.push_back(II); // Dead?
4790 I.setOperand(0, II->getOperand(1));
4791 I.setOperand(1, ConstantInt::get(Type::UShortTy,
4792 ByteSwap_16(CI->getZExtValue())));
4794 case Intrinsic::bswap_i32: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4795 WorkList.push_back(II); // Dead?
4796 I.setOperand(0, II->getOperand(1));
4797 I.setOperand(1, ConstantInt::get(Type::UIntTy,
4798 ByteSwap_32(CI->getZExtValue())));
4800 case Intrinsic::bswap_i64: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4801 WorkList.push_back(II); // Dead?
4802 I.setOperand(0, II->getOperand(1));
4803 I.setOperand(1, ConstantInt::get(Type::ULongTy,
4804 ByteSwap_64(CI->getZExtValue())));
4808 } else { // Not a SetEQ/SetNE
4809 // If the LHS is a cast from an integral value of the same size,
4810 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
4811 Value *CastOp = Cast->getOperand(0);
4812 const Type *SrcTy = CastOp->getType();
4813 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
4814 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
4815 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
4816 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
4817 "Source and destination signednesses should differ!");
4818 if (Cast->getType()->isSigned()) {
4819 // If this is a signed comparison, check for comparisons in the
4820 // vicinity of zero.
4821 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
4823 return BinaryOperator::createSetGT(CastOp,
4824 ConstantInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
4825 else if (I.getOpcode() == Instruction::SetGT &&
4826 cast<ConstantInt>(CI)->getSExtValue() == -1)
4827 // X > -1 => x < 128
4828 return BinaryOperator::createSetLT(CastOp,
4829 ConstantInt::get(SrcTy, 1ULL << (SrcTySize-1)));
4831 ConstantInt *CUI = cast<ConstantInt>(CI);
4832 if (I.getOpcode() == Instruction::SetLT &&
4833 CUI->getZExtValue() == 1ULL << (SrcTySize-1))
4834 // X < 128 => X > -1
4835 return BinaryOperator::createSetGT(CastOp,
4836 ConstantInt::get(SrcTy, -1));
4837 else if (I.getOpcode() == Instruction::SetGT &&
4838 CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
4840 return BinaryOperator::createSetLT(CastOp,
4841 Constant::getNullValue(SrcTy));
4848 // Handle setcc with constant RHS's that can be integer, FP or pointer.
4849 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4850 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4851 switch (LHSI->getOpcode()) {
4852 case Instruction::GetElementPtr:
4853 if (RHSC->isNullValue()) {
4854 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
4855 bool isAllZeros = true;
4856 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4857 if (!isa<Constant>(LHSI->getOperand(i)) ||
4858 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4863 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
4864 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4868 case Instruction::PHI:
4869 if (Instruction *NV = FoldOpIntoPhi(I))
4872 case Instruction::Select:
4873 // If either operand of the select is a constant, we can fold the
4874 // comparison into the select arms, which will cause one to be
4875 // constant folded and the select turned into a bitwise or.
4876 Value *Op1 = 0, *Op2 = 0;
4877 if (LHSI->hasOneUse()) {
4878 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4879 // Fold the known value into the constant operand.
4880 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4881 // Insert a new SetCC of the other select operand.
4882 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4883 LHSI->getOperand(2), RHSC,
4885 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4886 // Fold the known value into the constant operand.
4887 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4888 // Insert a new SetCC of the other select operand.
4889 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4890 LHSI->getOperand(1), RHSC,
4896 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4901 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
4902 if (User *GEP = dyn_castGetElementPtr(Op0))
4903 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
4905 if (User *GEP = dyn_castGetElementPtr(Op1))
4906 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
4907 SetCondInst::getSwappedCondition(I.getOpcode()), I))
4910 // Test to see if the operands of the setcc are casted versions of other
4911 // values. If the cast can be stripped off both arguments, we do so now.
4912 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4913 Value *CastOp0 = CI->getOperand(0);
4914 if (CI->isLosslessCast() && I.isEquality() &&
4915 (isa<Constant>(Op1) || isa<CastInst>(Op1))) {
4916 // We keep moving the cast from the left operand over to the right
4917 // operand, where it can often be eliminated completely.
4920 // If operand #1 is a cast instruction, see if we can eliminate it as
4922 if (CastInst *CI2 = dyn_cast<CastInst>(Op1)) {
4923 Value *CI2Op0 = CI2->getOperand(0);
4924 if (CI2Op0->getType()->canLosslesslyBitCastTo(Op0->getType()))
4928 // If Op1 is a constant, we can fold the cast into the constant.
4929 if (Op1->getType() != Op0->getType())
4930 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4931 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
4933 // Otherwise, cast the RHS right before the setcc
4934 Op1 = InsertCastBefore(Op1, Op0->getType(), I);
4936 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
4939 // Handle the special case of: setcc (cast bool to X), <cst>
4940 // This comes up when you have code like
4943 // For generality, we handle any zero-extension of any operand comparison
4944 // with a constant or another cast from the same type.
4945 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4946 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
4950 if (I.isEquality()) {
4952 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4953 (A == Op1 || B == Op1)) {
4954 // (A^B) == A -> B == 0
4955 Value *OtherVal = A == Op1 ? B : A;
4956 return BinaryOperator::create(I.getOpcode(), OtherVal,
4957 Constant::getNullValue(A->getType()));
4958 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4959 (A == Op0 || B == Op0)) {
4960 // A == (A^B) -> B == 0
4961 Value *OtherVal = A == Op0 ? B : A;
4962 return BinaryOperator::create(I.getOpcode(), OtherVal,
4963 Constant::getNullValue(A->getType()));
4964 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4965 // (A-B) == A -> B == 0
4966 return BinaryOperator::create(I.getOpcode(), B,
4967 Constant::getNullValue(B->getType()));
4968 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4969 // A == (A-B) -> B == 0
4970 return BinaryOperator::create(I.getOpcode(), B,
4971 Constant::getNullValue(B->getType()));
4975 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4976 if (Op0->hasOneUse() && Op1->hasOneUse() &&
4977 match(Op0, m_And(m_Value(A), m_Value(B))) &&
4978 match(Op1, m_And(m_Value(C), m_Value(D)))) {
4979 Value *X = 0, *Y = 0, *Z = 0;
4982 X = B; Y = D; Z = A;
4983 } else if (A == D) {
4984 X = B; Y = C; Z = A;
4985 } else if (B == C) {
4986 X = A; Y = D; Z = B;
4987 } else if (B == D) {
4988 X = A; Y = C; Z = B;
4991 if (X) { // Build (X^Y) & Z
4992 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
4993 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
4994 I.setOperand(0, Op1);
4995 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5000 return Changed ? &I : 0;
5003 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
5004 // We only handle extending casts so far.
5006 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
5007 const CastInst *LHSCI = cast<CastInst>(SCI.getOperand(0));
5008 Value *LHSCIOp = LHSCI->getOperand(0);
5009 const Type *SrcTy = LHSCIOp->getType();
5010 const Type *DestTy = SCI.getOperand(0)->getType();
5013 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
5016 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
5017 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
5018 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
5020 // Is this a sign or zero extension?
5021 bool isSignSrc = SrcTy->isSigned();
5022 bool isSignDest = DestTy->isSigned();
5024 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
5025 // Not an extension from the same type?
5026 RHSCIOp = CI->getOperand(0);
5027 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
5028 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
5029 // Compute the constant that would happen if we truncated to SrcTy then
5030 // reextended to DestTy.
5031 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5032 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5035 // Make sure that src sign and dest sign match. For example,
5037 // %A = cast short %X to uint
5038 // %B = setgt uint %A, 1330
5040 // It is incorrect to transform this into
5042 // %B = setgt short %X, 1330
5044 // because %A may have negative value.
5045 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5046 // OR operation is EQ/NE.
5047 if (isSignSrc == isSignDest || SrcTy == Type::BoolTy || SCI.isEquality())
5052 // If the value cannot be represented in the shorter type, we cannot emit
5053 // a simple comparison.
5054 if (SCI.getOpcode() == Instruction::SetEQ)
5055 return ReplaceInstUsesWith(SCI, ConstantBool::getFalse());
5056 if (SCI.getOpcode() == Instruction::SetNE)
5057 return ReplaceInstUsesWith(SCI, ConstantBool::getTrue());
5059 // Evaluate the comparison for LT.
5061 if (DestTy->isSigned()) {
5062 // We're performing a signed comparison.
5064 // Signed extend and signed comparison.
5065 if (cast<ConstantInt>(CI)->getSExtValue() < 0)// X < (small) --> false
5066 Result = ConstantBool::getFalse();
5068 Result = ConstantBool::getTrue(); // X < (large) --> true
5070 // Unsigned extend and signed comparison.
5071 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5072 Result = ConstantBool::getFalse();
5074 Result = ConstantBool::getTrue();
5077 // We're performing an unsigned comparison.
5079 // Unsigned extend & compare -> always true.
5080 Result = ConstantBool::getTrue();
5082 // We're performing an unsigned comp with a sign extended value.
5083 // This is true if the input is >= 0. [aka >s -1]
5084 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
5085 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
5086 NegOne, SCI.getName()), SCI);
5090 // Finally, return the value computed.
5091 if (SCI.getOpcode() == Instruction::SetLT) {
5092 return ReplaceInstUsesWith(SCI, Result);
5094 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
5095 if (Constant *CI = dyn_cast<Constant>(Result))
5096 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
5098 return BinaryOperator::createNot(Result);
5105 // Okay, just insert a compare of the reduced operands now!
5106 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
5109 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5110 assert(I.getOperand(1)->getType() == Type::UByteTy);
5111 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5112 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5114 // shl X, 0 == X and shr X, 0 == X
5115 // shl 0, X == 0 and shr 0, X == 0
5116 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
5117 Op0 == Constant::getNullValue(Op0->getType()))
5118 return ReplaceInstUsesWith(I, Op0);
5120 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
5121 if (!isLeftShift && I.getType()->isSigned())
5122 return ReplaceInstUsesWith(I, Op0);
5123 else // undef << X -> 0 AND undef >>u X -> 0
5124 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5126 if (isa<UndefValue>(Op1)) {
5127 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
5128 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5130 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
5133 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5134 if (I.getOpcode() == Instruction::AShr)
5135 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5136 if (CSI->isAllOnesValue())
5137 return ReplaceInstUsesWith(I, CSI);
5139 // Try to fold constant and into select arguments.
5140 if (isa<Constant>(Op0))
5141 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5142 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5145 // See if we can turn a signed shr into an unsigned shr.
5146 if (I.isArithmeticShift()) {
5147 if (MaskedValueIsZero(Op0,
5148 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5149 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5153 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5154 if (CUI->getType()->isUnsigned())
5155 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5160 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5162 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5163 bool isSignedShift = isLeftShift ? Op0->getType()->isSigned() :
5164 I.getOpcode() == Instruction::AShr;
5165 bool isUnsignedShift = !isSignedShift;
5167 // See if we can simplify any instructions used by the instruction whose sole
5168 // purpose is to compute bits we don't care about.
5169 uint64_t KnownZero, KnownOne;
5170 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5171 KnownZero, KnownOne))
5174 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5175 // of a signed value.
5177 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5178 if (Op1->getZExtValue() >= TypeBits) {
5179 if (isUnsignedShift || isLeftShift)
5180 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5182 I.setOperand(1, ConstantInt::get(Type::UByteTy, TypeBits-1));
5187 // ((X*C1) << C2) == (X * (C1 << C2))
5188 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5189 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5190 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5191 return BinaryOperator::createMul(BO->getOperand(0),
5192 ConstantExpr::getShl(BOOp, Op1));
5194 // Try to fold constant and into select arguments.
5195 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5196 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5198 if (isa<PHINode>(Op0))
5199 if (Instruction *NV = FoldOpIntoPhi(I))
5202 if (Op0->hasOneUse()) {
5203 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5204 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5207 switch (Op0BO->getOpcode()) {
5209 case Instruction::Add:
5210 case Instruction::And:
5211 case Instruction::Or:
5212 case Instruction::Xor:
5213 // These operators commute.
5214 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5215 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5216 match(Op0BO->getOperand(1),
5217 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5218 Instruction *YS = new ShiftInst(Instruction::Shl,
5219 Op0BO->getOperand(0), Op1,
5221 InsertNewInstBefore(YS, I); // (Y << C)
5223 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5224 Op0BO->getOperand(1)->getName());
5225 InsertNewInstBefore(X, I); // (X + (Y << C))
5226 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5227 C2 = ConstantExpr::getShl(C2, Op1);
5228 return BinaryOperator::createAnd(X, C2);
5231 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5232 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5233 match(Op0BO->getOperand(1),
5234 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5235 m_ConstantInt(CC))) && V2 == Op1 &&
5236 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5237 Instruction *YS = new ShiftInst(Instruction::Shl,
5238 Op0BO->getOperand(0), Op1,
5240 InsertNewInstBefore(YS, I); // (Y << C)
5242 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5243 V1->getName()+".mask");
5244 InsertNewInstBefore(XM, I); // X & (CC << C)
5246 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5250 case Instruction::Sub:
5251 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5252 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5253 match(Op0BO->getOperand(0),
5254 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5255 Instruction *YS = new ShiftInst(Instruction::Shl,
5256 Op0BO->getOperand(1), Op1,
5258 InsertNewInstBefore(YS, I); // (Y << C)
5260 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5261 Op0BO->getOperand(0)->getName());
5262 InsertNewInstBefore(X, I); // (X + (Y << C))
5263 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5264 C2 = ConstantExpr::getShl(C2, Op1);
5265 return BinaryOperator::createAnd(X, C2);
5268 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5269 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5270 match(Op0BO->getOperand(0),
5271 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5272 m_ConstantInt(CC))) && V2 == Op1 &&
5273 cast<BinaryOperator>(Op0BO->getOperand(0))
5274 ->getOperand(0)->hasOneUse()) {
5275 Instruction *YS = new ShiftInst(Instruction::Shl,
5276 Op0BO->getOperand(1), Op1,
5278 InsertNewInstBefore(YS, I); // (Y << C)
5280 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5281 V1->getName()+".mask");
5282 InsertNewInstBefore(XM, I); // X & (CC << C)
5284 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5291 // If the operand is an bitwise operator with a constant RHS, and the
5292 // shift is the only use, we can pull it out of the shift.
5293 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5294 bool isValid = true; // Valid only for And, Or, Xor
5295 bool highBitSet = false; // Transform if high bit of constant set?
5297 switch (Op0BO->getOpcode()) {
5298 default: isValid = false; break; // Do not perform transform!
5299 case Instruction::Add:
5300 isValid = isLeftShift;
5302 case Instruction::Or:
5303 case Instruction::Xor:
5306 case Instruction::And:
5311 // If this is a signed shift right, and the high bit is modified
5312 // by the logical operation, do not perform the transformation.
5313 // The highBitSet boolean indicates the value of the high bit of
5314 // the constant which would cause it to be modified for this
5317 if (isValid && !isLeftShift && isSignedShift) {
5318 uint64_t Val = Op0C->getZExtValue();
5319 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5323 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5325 Instruction *NewShift =
5326 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5329 InsertNewInstBefore(NewShift, I);
5331 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5338 // Find out if this is a shift of a shift by a constant.
5339 ShiftInst *ShiftOp = 0;
5340 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5342 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5343 // If this is a noop-integer cast of a shift instruction, use the shift.
5344 if (isa<ShiftInst>(CI->getOperand(0))) {
5345 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5349 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5350 // Find the operands and properties of the input shift. Note that the
5351 // signedness of the input shift may differ from the current shift if there
5352 // is a noop cast between the two.
5353 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5354 bool isShiftOfSignedShift = isShiftOfLeftShift ?
5355 ShiftOp->getType()->isSigned() :
5356 ShiftOp->getOpcode() == Instruction::AShr;
5357 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5359 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5361 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5362 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5364 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5365 if (isLeftShift == isShiftOfLeftShift) {
5366 // Do not fold these shifts if the first one is signed and the second one
5367 // is unsigned and this is a right shift. Further, don't do any folding
5369 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5372 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5373 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5374 Amt = Op0->getType()->getPrimitiveSizeInBits();
5376 Value *Op = ShiftOp->getOperand(0);
5377 if (isShiftOfSignedShift != isSignedShift)
5378 Op = InsertNewInstBefore(
5379 CastInst::createInferredCast(Op, I.getType(), "tmp"), I);
5380 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5381 ConstantInt::get(Type::UByteTy, Amt));
5382 if (I.getType() == ShiftResult->getType())
5384 InsertNewInstBefore(ShiftResult, I);
5385 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5388 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5389 // signed types, we can only support the (A >> c1) << c2 configuration,
5390 // because it can not turn an arbitrary bit of A into a sign bit.
5391 if (isUnsignedShift || isLeftShift) {
5392 // Calculate bitmask for what gets shifted off the edge.
5393 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
5395 C = ConstantExpr::getShl(C, ShiftAmt1C);
5397 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5399 Value *Op = ShiftOp->getOperand(0);
5400 if (Op->getType() != C->getType())
5401 Op = InsertCastBefore(Op, I.getType(), I);
5404 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5405 InsertNewInstBefore(Mask, I);
5407 // Figure out what flavor of shift we should use...
5408 if (ShiftAmt1 == ShiftAmt2) {
5409 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5410 } else if (ShiftAmt1 < ShiftAmt2) {
5411 return new ShiftInst(I.getOpcode(), Mask,
5412 ConstantInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
5413 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5414 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5415 return new ShiftInst(Instruction::LShr, Mask,
5416 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5418 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5419 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5422 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5423 Op = InsertCastBefore(Mask, I.getType()->getSignedVersion(), I);
5424 Instruction *Shift =
5425 new ShiftInst(ShiftOp->getOpcode(), Op,
5426 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5427 InsertNewInstBefore(Shift, I);
5429 C = ConstantIntegral::getAllOnesValue(Shift->getType());
5430 C = ConstantExpr::getShl(C, Op1);
5431 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5432 InsertNewInstBefore(Mask, I);
5433 return CastInst::create(Instruction::BitCast, Mask, I.getType());
5436 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5437 // this case, C1 == C2 and C1 is 8, 16, or 32.
5438 if (ShiftAmt1 == ShiftAmt2) {
5439 const Type *SExtType = 0;
5440 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5441 case 8 : SExtType = Type::SByteTy; break;
5442 case 16: SExtType = Type::ShortTy; break;
5443 case 32: SExtType = Type::IntTy; break;
5447 Instruction *NewTrunc =
5448 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5449 InsertNewInstBefore(NewTrunc, I);
5450 return new SExtInst(NewTrunc, I.getType());
5459 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5460 /// expression. If so, decompose it, returning some value X, such that Val is
5463 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5465 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
5466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5467 if (CI->getType()->isUnsigned()) {
5468 Offset = CI->getZExtValue();
5470 return ConstantInt::get(Type::UIntTy, 0);
5472 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5473 if (I->getNumOperands() == 2) {
5474 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5475 if (CUI->getType()->isUnsigned()) {
5476 if (I->getOpcode() == Instruction::Shl) {
5477 // This is a value scaled by '1 << the shift amt'.
5478 Scale = 1U << CUI->getZExtValue();
5480 return I->getOperand(0);
5481 } else if (I->getOpcode() == Instruction::Mul) {
5482 // This value is scaled by 'CUI'.
5483 Scale = CUI->getZExtValue();
5485 return I->getOperand(0);
5486 } else if (I->getOpcode() == Instruction::Add) {
5487 // We have X+C. Check to see if we really have (X*C2)+C1,
5488 // where C1 is divisible by C2.
5491 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5492 Offset += CUI->getZExtValue();
5493 if (SubScale > 1 && (Offset % SubScale == 0)) {
5503 // Otherwise, we can't look past this.
5510 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5511 /// try to eliminate the cast by moving the type information into the alloc.
5512 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5513 AllocationInst &AI) {
5514 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5515 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5517 // Remove any uses of AI that are dead.
5518 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5519 std::vector<Instruction*> DeadUsers;
5520 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5521 Instruction *User = cast<Instruction>(*UI++);
5522 if (isInstructionTriviallyDead(User)) {
5523 while (UI != E && *UI == User)
5524 ++UI; // If this instruction uses AI more than once, don't break UI.
5526 // Add operands to the worklist.
5527 AddUsesToWorkList(*User);
5529 DOUT << "IC: DCE: " << *User;
5531 User->eraseFromParent();
5532 removeFromWorkList(User);
5536 // Get the type really allocated and the type casted to.
5537 const Type *AllocElTy = AI.getAllocatedType();
5538 const Type *CastElTy = PTy->getElementType();
5539 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5541 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5542 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5543 if (CastElTyAlign < AllocElTyAlign) return 0;
5545 // If the allocation has multiple uses, only promote it if we are strictly
5546 // increasing the alignment of the resultant allocation. If we keep it the
5547 // same, we open the door to infinite loops of various kinds.
5548 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5550 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5551 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5552 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5554 // See if we can satisfy the modulus by pulling a scale out of the array
5556 unsigned ArraySizeScale, ArrayOffset;
5557 Value *NumElements = // See if the array size is a decomposable linear expr.
5558 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5560 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5562 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5563 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5565 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5570 // If the allocation size is constant, form a constant mul expression
5571 Amt = ConstantInt::get(Type::UIntTy, Scale);
5572 if (isa<ConstantInt>(NumElements) && NumElements->getType()->isUnsigned())
5573 Amt = ConstantExpr::getMul(
5574 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5575 // otherwise multiply the amount and the number of elements
5576 else if (Scale != 1) {
5577 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5578 Amt = InsertNewInstBefore(Tmp, AI);
5582 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5583 Value *Off = ConstantInt::get(Type::UIntTy, Offset);
5584 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5585 Amt = InsertNewInstBefore(Tmp, AI);
5588 std::string Name = AI.getName(); AI.setName("");
5589 AllocationInst *New;
5590 if (isa<MallocInst>(AI))
5591 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5593 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5594 InsertNewInstBefore(New, AI);
5596 // If the allocation has multiple uses, insert a cast and change all things
5597 // that used it to use the new cast. This will also hack on CI, but it will
5599 if (!AI.hasOneUse()) {
5600 AddUsesToWorkList(AI);
5601 // New is the allocation instruction, pointer typed. AI is the original
5602 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5603 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5604 InsertNewInstBefore(NewCast, AI);
5605 AI.replaceAllUsesWith(NewCast);
5607 return ReplaceInstUsesWith(CI, New);
5610 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5611 /// and return it without inserting any new casts. This is used by code that
5612 /// tries to decide whether promoting or shrinking integer operations to wider
5613 /// or smaller types will allow us to eliminate a truncate or extend.
5614 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5615 int &NumCastsRemoved) {
5616 if (isa<Constant>(V)) return true;
5618 Instruction *I = dyn_cast<Instruction>(V);
5619 if (!I || !I->hasOneUse()) return false;
5621 switch (I->getOpcode()) {
5622 case Instruction::And:
5623 case Instruction::Or:
5624 case Instruction::Xor:
5625 // These operators can all arbitrarily be extended or truncated.
5626 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5627 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5628 case Instruction::AShr:
5629 case Instruction::LShr:
5630 case Instruction::Shl:
5631 // If this is just a bitcast changing the sign of the operation, we can
5632 // convert if the operand can be converted.
5633 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5634 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5636 case Instruction::Trunc:
5637 case Instruction::ZExt:
5638 case Instruction::SExt:
5639 case Instruction::BitCast:
5640 // If this is a cast from the destination type, we can trivially eliminate
5641 // it, and this will remove a cast overall.
5642 if (I->getOperand(0)->getType() == Ty) {
5643 // If the first operand is itself a cast, and is eliminable, do not count
5644 // this as an eliminable cast. We would prefer to eliminate those two
5646 if (isa<CastInst>(I->getOperand(0)))
5654 // TODO: Can handle more cases here.
5661 /// EvaluateInDifferentType - Given an expression that
5662 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5663 /// evaluate the expression.
5664 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty) {
5665 if (Constant *C = dyn_cast<Constant>(V))
5666 return ConstantExpr::getCast(C, Ty);
5668 // Otherwise, it must be an instruction.
5669 Instruction *I = cast<Instruction>(V);
5670 Instruction *Res = 0;
5671 switch (I->getOpcode()) {
5672 case Instruction::And:
5673 case Instruction::Or:
5674 case Instruction::Xor: {
5675 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
5676 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty);
5677 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5678 LHS, RHS, I->getName());
5681 case Instruction::AShr:
5682 case Instruction::LShr:
5683 case Instruction::Shl: {
5684 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
5685 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5686 I->getOperand(1), I->getName());
5689 case Instruction::Trunc:
5690 case Instruction::ZExt:
5691 case Instruction::SExt:
5692 case Instruction::BitCast:
5693 // If the source type of the cast is the type we're trying for then we can
5694 // just return the source. There's no need to insert it because its not new.
5695 if (I->getOperand(0)->getType() == Ty)
5696 return I->getOperand(0);
5698 // Some other kind of cast, which shouldn't happen, so just ..
5701 // TODO: Can handle more cases here.
5702 assert(0 && "Unreachable!");
5706 return InsertNewInstBefore(Res, *I);
5709 /// @brief Implement the transforms common to all CastInst visitors.
5710 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5711 Value *Src = CI.getOperand(0);
5713 // Casting undef to anything results in undef so might as just replace it and
5714 // get rid of the cast.
5715 if (isa<UndefValue>(Src)) // cast undef -> undef
5716 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5718 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5719 // eliminate it now.
5720 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5721 if (Instruction::CastOps opc =
5722 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5723 // The first cast (CSrc) is eliminable so we need to fix up or replace
5724 // the second cast (CI). CSrc will then have a good chance of being dead.
5725 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
5729 // If casting the result of a getelementptr instruction with no offset, turn
5730 // this into a cast of the original pointer!
5732 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5733 bool AllZeroOperands = true;
5734 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5735 if (!isa<Constant>(GEP->getOperand(i)) ||
5736 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5737 AllZeroOperands = false;
5740 if (AllZeroOperands) {
5741 // Changing the cast operand is usually not a good idea but it is safe
5742 // here because the pointer operand is being replaced with another
5743 // pointer operand so the opcode doesn't need to change.
5744 CI.setOperand(0, GEP->getOperand(0));
5749 // If we are casting a malloc or alloca to a pointer to a type of the same
5750 // size, rewrite the allocation instruction to allocate the "right" type.
5751 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5752 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5755 // If we are casting a select then fold the cast into the select
5756 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5757 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5760 // If we are casting a PHI then fold the cast into the PHI
5761 if (isa<PHINode>(Src))
5762 if (Instruction *NV = FoldOpIntoPhi(CI))
5768 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
5769 /// integers. This function implements the common transforms for all those
5771 /// @brief Implement the transforms common to CastInst with integer operands
5772 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
5773 if (Instruction *Result = commonCastTransforms(CI))
5776 Value *Src = CI.getOperand(0);
5777 const Type *SrcTy = Src->getType();
5778 const Type *DestTy = CI.getType();
5779 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
5780 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5782 // See if we can simplify any instructions used by the LHS whose sole
5783 // purpose is to compute bits we don't care about.
5784 uint64_t KnownZero = 0, KnownOne = 0;
5785 if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
5786 KnownZero, KnownOne))
5789 // If the source isn't an instruction or has more than one use then we
5790 // can't do anything more.
5791 if (!isa<Instruction>(Src) || !Src->hasOneUse())
5794 // Attempt to propagate the cast into the instruction.
5795 Instruction *SrcI = cast<Instruction>(Src);
5796 int NumCastsRemoved = 0;
5797 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
5798 // If this cast is a truncate, evaluting in a different type always
5799 // eliminates the cast, so it is always a win. If this is a noop-cast
5800 // this just removes a noop cast which isn't pointful, but simplifies
5801 // the code. If this is a zero-extension, we need to do an AND to
5802 // maintain the clear top-part of the computation, so we require that
5803 // the input have eliminated at least one cast. If this is a sign
5804 // extension, we insert two new casts (to do the extension) so we
5805 // require that two casts have been eliminated.
5806 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
5808 switch (CI.getOpcode()) {
5809 case Instruction::Trunc:
5812 case Instruction::ZExt:
5813 DoXForm = NumCastsRemoved >= 1;
5815 case Instruction::SExt:
5816 DoXForm = NumCastsRemoved >= 2;
5818 case Instruction::BitCast:
5822 // All the others use floating point so we shouldn't actually
5823 // get here because of the check above.
5824 assert(!"Unknown cast type .. unreachable");
5830 Value *Res = EvaluateInDifferentType(SrcI, DestTy);
5831 assert(Res->getType() == DestTy);
5832 switch (CI.getOpcode()) {
5833 default: assert(0 && "Unknown cast type!");
5834 case Instruction::Trunc:
5835 case Instruction::BitCast:
5836 // Just replace this cast with the result.
5837 return ReplaceInstUsesWith(CI, Res);
5838 case Instruction::ZExt: {
5839 // We need to emit an AND to clear the high bits.
5840 assert(SrcBitSize < DestBitSize && "Not a zext?");
5842 ConstantInt::get(Type::ULongTy, (1ULL << SrcBitSize)-1);
5843 if (DestBitSize < 64)
5844 C = ConstantExpr::getTrunc(C, DestTy);
5846 assert(DestBitSize == 64);
5847 C = ConstantExpr::getBitCast(C, DestTy);
5849 return BinaryOperator::createAnd(Res, C);
5851 case Instruction::SExt:
5852 // We need to emit a cast to truncate, then a cast to sext.
5853 return CastInst::create(Instruction::SExt,
5854 InsertCastBefore(Res, Src->getType(), CI), DestTy);
5859 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
5860 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
5862 switch (SrcI->getOpcode()) {
5863 case Instruction::Add:
5864 case Instruction::Mul:
5865 case Instruction::And:
5866 case Instruction::Or:
5867 case Instruction::Xor:
5868 // If we are discarding information, or just changing the sign,
5870 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
5871 // Don't insert two casts if they cannot be eliminated. We allow
5872 // two casts to be inserted if the sizes are the same. This could
5873 // only be converting signedness, which is a noop.
5874 if (DestBitSize == SrcBitSize ||
5875 !ValueRequiresCast(Op1, DestTy,TD) ||
5876 !ValueRequiresCast(Op0, DestTy, TD)) {
5877 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5878 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5879 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
5880 ->getOpcode(), Op0c, Op1c);
5884 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
5885 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
5886 SrcI->getOpcode() == Instruction::Xor &&
5887 Op1 == ConstantBool::getTrue() &&
5888 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
5889 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
5890 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
5893 case Instruction::SDiv:
5894 case Instruction::UDiv:
5895 case Instruction::SRem:
5896 case Instruction::URem:
5897 // If we are just changing the sign, rewrite.
5898 if (DestBitSize == SrcBitSize) {
5899 // Don't insert two casts if they cannot be eliminated. We allow
5900 // two casts to be inserted if the sizes are the same. This could
5901 // only be converting signedness, which is a noop.
5902 if (!ValueRequiresCast(Op1, DestTy,TD) ||
5903 !ValueRequiresCast(Op0, DestTy, TD)) {
5904 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5905 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5906 return BinaryOperator::create(
5907 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
5912 case Instruction::Shl:
5913 // Allow changing the sign of the source operand. Do not allow
5914 // changing the size of the shift, UNLESS the shift amount is a
5915 // constant. We must not change variable sized shifts to a smaller
5916 // size, because it is undefined to shift more bits out than exist
5918 if (DestBitSize == SrcBitSize ||
5919 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
5920 Value *Op0c = InsertOperandCastBefore(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::getZeroExtend(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::createInferredCast(In, CI.getType());
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 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcI->getOperand(0), Mask)) {
6018 if (ShAmt >= DestBitWidth) // All zeros.
6019 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6021 // Okay, we can shrink this. Truncate the input, then return a new
6023 Value *V = InsertCastBefore(SrcI->getOperand(0), Ty, CI);
6024 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6026 } else { // This is a variable shr.
6028 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6029 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6030 // loop-invariant and CSE'd.
6031 if (CI.getType() == Type::BoolTy && SrcI->hasOneUse()) {
6032 Value *One = ConstantInt::get(SrcI->getType(), 1);
6034 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6035 SrcI->getOperand(1),
6037 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6038 SrcI->getOperand(0),
6040 Value *Zero = Constant::getNullValue(V->getType());
6041 return BinaryOperator::createSetNE(V, Zero);
6051 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6052 // If one of the common conversion will work ..
6053 if (Instruction *Result = commonIntCastTransforms(CI))
6056 Value *Src = CI.getOperand(0);
6058 // If this is a cast of a cast
6059 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6060 // If the operand of the ZEXT is a TRUNC then we are dealing with integral
6061 // types and we can convert this to a logical AND if the sizes are just
6062 // right. This will be much cheaper than the pair of casts.
6063 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6064 // types and if the sizes are just right we can convert this into a logical
6065 // 'and' which will be much cheaper than the pair of casts.
6066 if (isa<TruncInst>(CSrc)) {
6067 // Get the sizes of the types involved
6068 Value *A = CSrc->getOperand(0);
6069 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6070 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6071 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6072 // If we're actually extending zero bits and the trunc is a no-op
6073 if (MidSize < DstSize && SrcSize == DstSize) {
6074 // Replace both of the casts with an And of the type mask.
6075 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
6076 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6078 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6079 // Unfortunately, if the type changed, we need to cast it back.
6080 if (And->getType() != CI.getType()) {
6081 And->setName(CSrc->getName()+".mask");
6082 InsertNewInstBefore(And, CI);
6083 And = CastInst::createInferredCast(And, CI.getType());
6093 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6094 return commonIntCastTransforms(CI);
6097 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6098 return commonCastTransforms(CI);
6101 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6102 return commonCastTransforms(CI);
6105 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6106 return commonCastTransforms(CI);
6109 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6110 return commonCastTransforms(CI);
6113 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6114 return commonCastTransforms(CI);
6117 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6118 return commonCastTransforms(CI);
6121 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6122 return commonCastTransforms(CI);
6125 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6126 return commonCastTransforms(CI);
6129 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6131 // If the operands are integer typed then apply the integer transforms,
6132 // otherwise just apply the common ones.
6133 Value *Src = CI.getOperand(0);
6134 const Type *SrcTy = Src->getType();
6135 const Type *DestTy = CI.getType();
6137 if (SrcTy->isInteger() && DestTy->isInteger()) {
6138 if (Instruction *Result = commonIntCastTransforms(CI))
6141 if (Instruction *Result = commonCastTransforms(CI))
6146 // Get rid of casts from one type to the same type. These are useless and can
6147 // be replaced by the operand.
6148 if (DestTy == Src->getType())
6149 return ReplaceInstUsesWith(CI, Src);
6151 // If the source and destination are pointers, and this cast is equivalent to
6152 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6153 // This can enhance SROA and other transforms that want type-safe pointers.
6154 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6155 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6156 const Type *DstElTy = DstPTy->getElementType();
6157 const Type *SrcElTy = SrcPTy->getElementType();
6159 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
6160 unsigned NumZeros = 0;
6161 while (SrcElTy != DstElTy &&
6162 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6163 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6164 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6168 // If we found a path from the src to dest, create the getelementptr now.
6169 if (SrcElTy == DstElTy) {
6170 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6171 return new GetElementPtrInst(Src, Idxs);
6176 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6177 if (SVI->hasOneUse()) {
6178 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6179 // a bitconvert to a vector with the same # elts.
6180 if (isa<PackedType>(DestTy) &&
6181 cast<PackedType>(DestTy)->getNumElements() ==
6182 SVI->getType()->getNumElements()) {
6184 // If either of the operands is a cast from CI.getType(), then
6185 // evaluating the shuffle in the casted destination's type will allow
6186 // us to eliminate at least one cast.
6187 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6188 Tmp->getOperand(0)->getType() == DestTy) ||
6189 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6190 Tmp->getOperand(0)->getType() == DestTy)) {
6191 Value *LHS = InsertOperandCastBefore(SVI->getOperand(0), DestTy, &CI);
6192 Value *RHS = InsertOperandCastBefore(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(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(II->getOperand(2), OpPtrTy, CI);
6751 return new StoreInst(II->getOperand(1), Ptr);
6754 case Intrinsic::x86_sse_storeu_ps:
6755 case Intrinsic::x86_sse2_storeu_pd:
6756 case Intrinsic::x86_sse2_storeu_dq:
6757 case Intrinsic::x86_sse2_storel_dq:
6758 // Turn X86 storeu -> store if the pointer is known aligned.
6759 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6760 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
6761 Value *Ptr = InsertCastBefore(II->getOperand(1), OpPtrTy, CI);
6762 return new StoreInst(II->getOperand(2), Ptr);
6766 case Intrinsic::x86_sse_cvttss2si: {
6767 // These intrinsics only demands the 0th element of its input vector. If
6768 // we can simplify the input based on that, do so now.
6770 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
6772 II->setOperand(1, V);
6778 case Intrinsic::ppc_altivec_vperm:
6779 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
6780 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
6781 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
6783 // Check that all of the elements are integer constants or undefs.
6784 bool AllEltsOk = true;
6785 for (unsigned i = 0; i != 16; ++i) {
6786 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
6787 !isa<UndefValue>(Mask->getOperand(i))) {
6794 // Cast the input vectors to byte vectors.
6795 Value *Op0 = InsertCastBefore(II->getOperand(1), Mask->getType(), CI);
6796 Value *Op1 = InsertCastBefore(II->getOperand(2), Mask->getType(), CI);
6797 Value *Result = UndefValue::get(Op0->getType());
6799 // Only extract each element once.
6800 Value *ExtractedElts[32];
6801 memset(ExtractedElts, 0, sizeof(ExtractedElts));
6803 for (unsigned i = 0; i != 16; ++i) {
6804 if (isa<UndefValue>(Mask->getOperand(i)))
6806 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
6807 Idx &= 31; // Match the hardware behavior.
6809 if (ExtractedElts[Idx] == 0) {
6811 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
6812 InsertNewInstBefore(Elt, CI);
6813 ExtractedElts[Idx] = Elt;
6816 // Insert this value into the result vector.
6817 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
6818 InsertNewInstBefore(cast<Instruction>(Result), CI);
6820 return CastInst::create(Instruction::BitCast, Result, CI.getType());
6825 case Intrinsic::stackrestore: {
6826 // If the save is right next to the restore, remove the restore. This can
6827 // happen when variable allocas are DCE'd.
6828 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
6829 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
6830 BasicBlock::iterator BI = SS;
6832 return EraseInstFromFunction(CI);
6836 // If the stack restore is in a return/unwind block and if there are no
6837 // allocas or calls between the restore and the return, nuke the restore.
6838 TerminatorInst *TI = II->getParent()->getTerminator();
6839 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
6840 BasicBlock::iterator BI = II;
6841 bool CannotRemove = false;
6842 for (++BI; &*BI != TI; ++BI) {
6843 if (isa<AllocaInst>(BI) ||
6844 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
6845 CannotRemove = true;
6850 return EraseInstFromFunction(CI);
6857 return visitCallSite(II);
6860 // InvokeInst simplification
6862 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
6863 return visitCallSite(&II);
6866 // visitCallSite - Improvements for call and invoke instructions.
6868 Instruction *InstCombiner::visitCallSite(CallSite CS) {
6869 bool Changed = false;
6871 // If the callee is a constexpr cast of a function, attempt to move the cast
6872 // to the arguments of the call/invoke.
6873 if (transformConstExprCastCall(CS)) return 0;
6875 Value *Callee = CS.getCalledValue();
6877 if (Function *CalleeF = dyn_cast<Function>(Callee))
6878 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
6879 Instruction *OldCall = CS.getInstruction();
6880 // If the call and callee calling conventions don't match, this call must
6881 // be unreachable, as the call is undefined.
6882 new StoreInst(ConstantBool::getTrue(),
6883 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
6884 if (!OldCall->use_empty())
6885 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
6886 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
6887 return EraseInstFromFunction(*OldCall);
6891 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
6892 // This instruction is not reachable, just remove it. We insert a store to
6893 // undef so that we know that this code is not reachable, despite the fact
6894 // that we can't modify the CFG here.
6895 new StoreInst(ConstantBool::getTrue(),
6896 UndefValue::get(PointerType::get(Type::BoolTy)),
6897 CS.getInstruction());
6899 if (!CS.getInstruction()->use_empty())
6900 CS.getInstruction()->
6901 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
6903 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
6904 // Don't break the CFG, insert a dummy cond branch.
6905 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
6906 ConstantBool::getTrue(), II);
6908 return EraseInstFromFunction(*CS.getInstruction());
6911 const PointerType *PTy = cast<PointerType>(Callee->getType());
6912 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
6913 if (FTy->isVarArg()) {
6914 // See if we can optimize any arguments passed through the varargs area of
6916 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
6917 E = CS.arg_end(); I != E; ++I)
6918 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
6919 // If this cast does not effect the value passed through the varargs
6920 // area, we can eliminate the use of the cast.
6921 Value *Op = CI->getOperand(0);
6922 if (CI->isLosslessCast()) {
6929 return Changed ? CS.getInstruction() : 0;
6932 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
6933 // attempt to move the cast to the arguments of the call/invoke.
6935 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
6936 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
6937 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
6938 if (CE->getOpcode() != Instruction::BitCast ||
6939 !isa<Function>(CE->getOperand(0)))
6941 Function *Callee = cast<Function>(CE->getOperand(0));
6942 Instruction *Caller = CS.getInstruction();
6944 // Okay, this is a cast from a function to a different type. Unless doing so
6945 // would cause a type conversion of one of our arguments, change this call to
6946 // be a direct call with arguments casted to the appropriate types.
6948 const FunctionType *FT = Callee->getFunctionType();
6949 const Type *OldRetTy = Caller->getType();
6951 // Check to see if we are changing the return type...
6952 if (OldRetTy != FT->getReturnType()) {
6953 if (Callee->isExternal() &&
6954 !Caller->use_empty() &&
6955 !(OldRetTy->canLosslesslyBitCastTo(FT->getReturnType()) ||
6956 (isa<PointerType>(FT->getReturnType()) &&
6957 TD->getIntPtrType()->canLosslesslyBitCastTo(OldRetTy)))
6959 return false; // Cannot transform this return value...
6961 // If the callsite is an invoke instruction, and the return value is used by
6962 // a PHI node in a successor, we cannot change the return type of the call
6963 // because there is no place to put the cast instruction (without breaking
6964 // the critical edge). Bail out in this case.
6965 if (!Caller->use_empty())
6966 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
6967 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
6969 if (PHINode *PN = dyn_cast<PHINode>(*UI))
6970 if (PN->getParent() == II->getNormalDest() ||
6971 PN->getParent() == II->getUnwindDest())
6975 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
6976 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
6978 CallSite::arg_iterator AI = CS.arg_begin();
6979 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
6980 const Type *ParamTy = FT->getParamType(i);
6981 const Type *ActTy = (*AI)->getType();
6982 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
6983 //Either we can cast directly, or we can upconvert the argument
6984 bool isConvertible = ActTy->canLosslesslyBitCastTo(ParamTy) ||
6985 (ParamTy->isIntegral() && ActTy->isIntegral() &&
6986 ParamTy->isSigned() == ActTy->isSigned() &&
6987 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
6988 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
6989 c->getSExtValue() > 0);
6990 if (Callee->isExternal() && !isConvertible) return false;
6993 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
6994 Callee->isExternal())
6995 return false; // Do not delete arguments unless we have a function body...
6997 // Okay, we decided that this is a safe thing to do: go ahead and start
6998 // inserting cast instructions as necessary...
6999 std::vector<Value*> Args;
7000 Args.reserve(NumActualArgs);
7002 AI = CS.arg_begin();
7003 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7004 const Type *ParamTy = FT->getParamType(i);
7005 if ((*AI)->getType() == ParamTy) {
7006 Args.push_back(*AI);
7008 CastInst *NewCast = CastInst::createInferredCast(*AI, ParamTy, "tmp");
7009 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7013 // If the function takes more arguments than the call was taking, add them
7015 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7016 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7018 // If we are removing arguments to the function, emit an obnoxious warning...
7019 if (FT->getNumParams() < NumActualArgs)
7020 if (!FT->isVarArg()) {
7021 llvm_cerr << "WARNING: While resolving call to function '"
7022 << Callee->getName() << "' arguments were dropped!\n";
7024 // Add all of the arguments in their promoted form to the arg list...
7025 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7026 const Type *PTy = getPromotedType((*AI)->getType());
7027 if (PTy != (*AI)->getType()) {
7028 // Must promote to pass through va_arg area!
7029 Instruction *Cast = CastInst::createInferredCast(*AI, PTy, "tmp");
7030 InsertNewInstBefore(Cast, *Caller);
7031 Args.push_back(Cast);
7033 Args.push_back(*AI);
7038 if (FT->getReturnType() == Type::VoidTy)
7039 Caller->setName(""); // Void type should not have a name...
7042 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7043 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7044 Args, Caller->getName(), Caller);
7045 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7047 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7048 if (cast<CallInst>(Caller)->isTailCall())
7049 cast<CallInst>(NC)->setTailCall();
7050 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7053 // Insert a cast of the return type as necessary...
7055 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7056 if (NV->getType() != Type::VoidTy) {
7057 NV = NC = CastInst::createInferredCast(NC, Caller->getType(), "tmp");
7059 // If this is an invoke instruction, we should insert it after the first
7060 // non-phi, instruction in the normal successor block.
7061 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7062 BasicBlock::iterator I = II->getNormalDest()->begin();
7063 while (isa<PHINode>(I)) ++I;
7064 InsertNewInstBefore(NC, *I);
7066 // Otherwise, it's a call, just insert cast right after the call instr
7067 InsertNewInstBefore(NC, *Caller);
7069 AddUsersToWorkList(*Caller);
7071 NV = UndefValue::get(Caller->getType());
7075 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7076 Caller->replaceAllUsesWith(NV);
7077 Caller->getParent()->getInstList().erase(Caller);
7078 removeFromWorkList(Caller);
7082 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7083 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7084 /// and a single binop.
7085 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7086 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7087 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7088 isa<GetElementPtrInst>(FirstInst));
7089 unsigned Opc = FirstInst->getOpcode();
7090 Value *LHSVal = FirstInst->getOperand(0);
7091 Value *RHSVal = FirstInst->getOperand(1);
7093 const Type *LHSType = LHSVal->getType();
7094 const Type *RHSType = RHSVal->getType();
7096 // Scan to see if all operands are the same opcode, all have one use, and all
7097 // kill their operands (i.e. the operands have one use).
7098 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7099 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7100 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7101 // Verify type of the LHS matches so we don't fold setcc's of different
7102 // types or GEP's with different index types.
7103 I->getOperand(0)->getType() != LHSType ||
7104 I->getOperand(1)->getType() != RHSType)
7107 // Keep track of which operand needs a phi node.
7108 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7109 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7112 // Otherwise, this is safe to transform, determine if it is profitable.
7114 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7115 // Indexes are often folded into load/store instructions, so we don't want to
7116 // hide them behind a phi.
7117 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7120 Value *InLHS = FirstInst->getOperand(0);
7121 Value *InRHS = FirstInst->getOperand(1);
7122 PHINode *NewLHS = 0, *NewRHS = 0;
7124 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7125 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7126 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7127 InsertNewInstBefore(NewLHS, PN);
7132 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7133 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7134 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7135 InsertNewInstBefore(NewRHS, PN);
7139 // Add all operands to the new PHIs.
7140 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7142 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7143 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7146 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7147 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7151 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7152 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7153 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7154 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7156 assert(isa<GetElementPtrInst>(FirstInst));
7157 return new GetElementPtrInst(LHSVal, RHSVal);
7161 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7162 /// of the block that defines it. This means that it must be obvious the value
7163 /// of the load is not changed from the point of the load to the end of the
7165 static bool isSafeToSinkLoad(LoadInst *L) {
7166 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7168 for (++BBI; BBI != E; ++BBI)
7169 if (BBI->mayWriteToMemory())
7175 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7176 // operator and they all are only used by the PHI, PHI together their
7177 // inputs, and do the operation once, to the result of the PHI.
7178 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7179 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7181 // Scan the instruction, looking for input operations that can be folded away.
7182 // If all input operands to the phi are the same instruction (e.g. a cast from
7183 // the same type or "+42") we can pull the operation through the PHI, reducing
7184 // code size and simplifying code.
7185 Constant *ConstantOp = 0;
7186 const Type *CastSrcTy = 0;
7187 bool isVolatile = false;
7188 if (isa<CastInst>(FirstInst)) {
7189 CastSrcTy = FirstInst->getOperand(0)->getType();
7190 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
7191 // Can fold binop or shift here if the RHS is a constant, otherwise call
7192 // FoldPHIArgBinOpIntoPHI.
7193 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7194 if (ConstantOp == 0)
7195 return FoldPHIArgBinOpIntoPHI(PN);
7196 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7197 isVolatile = LI->isVolatile();
7198 // We can't sink the load if the loaded value could be modified between the
7199 // load and the PHI.
7200 if (LI->getParent() != PN.getIncomingBlock(0) ||
7201 !isSafeToSinkLoad(LI))
7203 } else if (isa<GetElementPtrInst>(FirstInst)) {
7204 if (FirstInst->getNumOperands() == 2)
7205 return FoldPHIArgBinOpIntoPHI(PN);
7206 // Can't handle general GEPs yet.
7209 return 0; // Cannot fold this operation.
7212 // Check to see if all arguments are the same operation.
7213 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7214 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7215 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7216 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
7219 if (I->getOperand(0)->getType() != CastSrcTy)
7220 return 0; // Cast operation must match.
7221 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7222 // We can't sink the load if the loaded value could be modified between the
7223 // load and the PHI.
7224 if (LI->isVolatile() != isVolatile ||
7225 LI->getParent() != PN.getIncomingBlock(i) ||
7226 !isSafeToSinkLoad(LI))
7228 } else if (I->getOperand(1) != ConstantOp) {
7233 // Okay, they are all the same operation. Create a new PHI node of the
7234 // correct type, and PHI together all of the LHS's of the instructions.
7235 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7236 PN.getName()+".in");
7237 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7239 Value *InVal = FirstInst->getOperand(0);
7240 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7242 // Add all operands to the new PHI.
7243 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7244 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7245 if (NewInVal != InVal)
7247 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7252 // The new PHI unions all of the same values together. This is really
7253 // common, so we handle it intelligently here for compile-time speed.
7257 InsertNewInstBefore(NewPN, PN);
7261 // Insert and return the new operation.
7262 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7263 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7264 else if (isa<LoadInst>(FirstInst))
7265 return new LoadInst(PhiVal, "", isVolatile);
7266 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7267 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7269 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7270 PhiVal, ConstantOp);
7273 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7275 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7276 if (PN->use_empty()) return true;
7277 if (!PN->hasOneUse()) return false;
7279 // Remember this node, and if we find the cycle, return.
7280 if (!PotentiallyDeadPHIs.insert(PN).second)
7283 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7284 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7289 // PHINode simplification
7291 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7292 // If LCSSA is around, don't mess with Phi nodes
7293 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7295 if (Value *V = PN.hasConstantValue())
7296 return ReplaceInstUsesWith(PN, V);
7298 // If all PHI operands are the same operation, pull them through the PHI,
7299 // reducing code size.
7300 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7301 PN.getIncomingValue(0)->hasOneUse())
7302 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7305 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7306 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7307 // PHI)... break the cycle.
7309 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7310 std::set<PHINode*> PotentiallyDeadPHIs;
7311 PotentiallyDeadPHIs.insert(&PN);
7312 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7313 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7319 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
7320 Instruction *InsertPoint,
7322 unsigned PS = IC->getTargetData().getPointerSize();
7323 const Type *VTy = V->getType();
7324 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
7325 // We must insert a cast to ensure we sign-extend.
7326 V = IC->InsertCastBefore(V, VTy->getSignedVersion(), *InsertPoint);
7327 return IC->InsertCastBefore(V, DTy, *InsertPoint);
7331 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7332 Value *PtrOp = GEP.getOperand(0);
7333 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7334 // If so, eliminate the noop.
7335 if (GEP.getNumOperands() == 1)
7336 return ReplaceInstUsesWith(GEP, PtrOp);
7338 if (isa<UndefValue>(GEP.getOperand(0)))
7339 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7341 bool HasZeroPointerIndex = false;
7342 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7343 HasZeroPointerIndex = C->isNullValue();
7345 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7346 return ReplaceInstUsesWith(GEP, PtrOp);
7348 // Eliminate unneeded casts for indices.
7349 bool MadeChange = false;
7350 gep_type_iterator GTI = gep_type_begin(GEP);
7351 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7352 if (isa<SequentialType>(*GTI)) {
7353 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7354 Value *Src = CI->getOperand(0);
7355 const Type *SrcTy = Src->getType();
7356 const Type *DestTy = CI->getType();
7357 if (Src->getType()->isInteger()) {
7358 if (SrcTy->getPrimitiveSizeInBits() ==
7359 DestTy->getPrimitiveSizeInBits()) {
7360 // We can always eliminate a cast from ulong or long to the other.
7361 // We can always eliminate a cast from uint to int or the other on
7362 // 32-bit pointer platforms.
7363 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7365 GEP.setOperand(i, Src);
7367 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
7368 SrcTy->getPrimitiveSize() == 4) {
7369 // We can always eliminate a cast from int to [u]long. We can
7370 // eliminate a cast from uint to [u]long iff the target is a 32-bit
7372 if (SrcTy->isSigned() ||
7373 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7375 GEP.setOperand(i, Src);
7380 // If we are using a wider index than needed for this platform, shrink it
7381 // to what we need. If the incoming value needs a cast instruction,
7382 // insert it. This explicit cast can make subsequent optimizations more
7384 Value *Op = GEP.getOperand(i);
7385 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
7386 if (Constant *C = dyn_cast<Constant>(Op)) {
7387 GEP.setOperand(i, ConstantExpr::getCast(C,
7388 TD->getIntPtrType()->getSignedVersion()));
7391 Op = InsertCastBefore(Op, TD->getIntPtrType(), GEP);
7392 GEP.setOperand(i, Op);
7396 // If this is a constant idx, make sure to canonicalize it to be a signed
7397 // operand, otherwise CSE and other optimizations are pessimized.
7398 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op))
7399 if (CUI->getType()->isUnsigned()) {
7401 ConstantExpr::getCast(CUI, CUI->getType()->getSignedVersion()));
7405 if (MadeChange) return &GEP;
7407 // Combine Indices - If the source pointer to this getelementptr instruction
7408 // is a getelementptr instruction, combine the indices of the two
7409 // getelementptr instructions into a single instruction.
7411 std::vector<Value*> SrcGEPOperands;
7412 if (User *Src = dyn_castGetElementPtr(PtrOp))
7413 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7415 if (!SrcGEPOperands.empty()) {
7416 // Note that if our source is a gep chain itself that we wait for that
7417 // chain to be resolved before we perform this transformation. This
7418 // avoids us creating a TON of code in some cases.
7420 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7421 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7422 return 0; // Wait until our source is folded to completion.
7424 std::vector<Value *> Indices;
7426 // Find out whether the last index in the source GEP is a sequential idx.
7427 bool EndsWithSequential = false;
7428 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7429 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7430 EndsWithSequential = !isa<StructType>(*I);
7432 // Can we combine the two pointer arithmetics offsets?
7433 if (EndsWithSequential) {
7434 // Replace: gep (gep %P, long B), long A, ...
7435 // With: T = long A+B; gep %P, T, ...
7437 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7438 if (SO1 == Constant::getNullValue(SO1->getType())) {
7440 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7443 // If they aren't the same type, convert both to an integer of the
7444 // target's pointer size.
7445 if (SO1->getType() != GO1->getType()) {
7446 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7447 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
7448 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7449 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
7451 unsigned PS = TD->getPointerSize();
7452 if (SO1->getType()->getPrimitiveSize() == PS) {
7453 // Convert GO1 to SO1's type.
7454 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
7456 } else if (GO1->getType()->getPrimitiveSize() == PS) {
7457 // Convert SO1 to GO1's type.
7458 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
7460 const Type *PT = TD->getIntPtrType();
7461 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
7462 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
7466 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7467 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7469 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7470 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7474 // Recycle the GEP we already have if possible.
7475 if (SrcGEPOperands.size() == 2) {
7476 GEP.setOperand(0, SrcGEPOperands[0]);
7477 GEP.setOperand(1, Sum);
7480 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7481 SrcGEPOperands.end()-1);
7482 Indices.push_back(Sum);
7483 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7485 } else if (isa<Constant>(*GEP.idx_begin()) &&
7486 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7487 SrcGEPOperands.size() != 1) {
7488 // Otherwise we can do the fold if the first index of the GEP is a zero
7489 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7490 SrcGEPOperands.end());
7491 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7494 if (!Indices.empty())
7495 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7497 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7498 // GEP of global variable. If all of the indices for this GEP are
7499 // constants, we can promote this to a constexpr instead of an instruction.
7501 // Scan for nonconstants...
7502 std::vector<Constant*> Indices;
7503 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7504 for (; I != E && isa<Constant>(*I); ++I)
7505 Indices.push_back(cast<Constant>(*I));
7507 if (I == E) { // If they are all constants...
7508 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7510 // Replace all uses of the GEP with the new constexpr...
7511 return ReplaceInstUsesWith(GEP, CE);
7513 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7514 if (!isa<PointerType>(X->getType())) {
7515 // Not interesting. Source pointer must be a cast from pointer.
7516 } else if (HasZeroPointerIndex) {
7517 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7518 // into : GEP [10 x ubyte]* X, long 0, ...
7520 // This occurs when the program declares an array extern like "int X[];"
7522 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7523 const PointerType *XTy = cast<PointerType>(X->getType());
7524 if (const ArrayType *XATy =
7525 dyn_cast<ArrayType>(XTy->getElementType()))
7526 if (const ArrayType *CATy =
7527 dyn_cast<ArrayType>(CPTy->getElementType()))
7528 if (CATy->getElementType() == XATy->getElementType()) {
7529 // At this point, we know that the cast source type is a pointer
7530 // to an array of the same type as the destination pointer
7531 // array. Because the array type is never stepped over (there
7532 // is a leading zero) we can fold the cast into this GEP.
7533 GEP.setOperand(0, X);
7536 } else if (GEP.getNumOperands() == 2) {
7537 // Transform things like:
7538 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7539 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7540 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7541 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7542 if (isa<ArrayType>(SrcElTy) &&
7543 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7544 TD->getTypeSize(ResElTy)) {
7545 Value *V = InsertNewInstBefore(
7546 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7547 GEP.getOperand(1), GEP.getName()), GEP);
7548 // V and GEP are both pointer types --> BitCast
7549 return new BitCastInst(V, GEP.getType());
7552 // Transform things like:
7553 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7554 // (where tmp = 8*tmp2) into:
7555 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7557 if (isa<ArrayType>(SrcElTy) &&
7558 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
7559 uint64_t ArrayEltSize =
7560 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7562 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7563 // allow either a mul, shift, or constant here.
7565 ConstantInt *Scale = 0;
7566 if (ArrayEltSize == 1) {
7567 NewIdx = GEP.getOperand(1);
7568 Scale = ConstantInt::get(NewIdx->getType(), 1);
7569 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7570 NewIdx = ConstantInt::get(CI->getType(), 1);
7572 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7573 if (Inst->getOpcode() == Instruction::Shl &&
7574 isa<ConstantInt>(Inst->getOperand(1))) {
7576 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7577 if (Inst->getType()->isSigned())
7578 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7580 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7581 NewIdx = Inst->getOperand(0);
7582 } else if (Inst->getOpcode() == Instruction::Mul &&
7583 isa<ConstantInt>(Inst->getOperand(1))) {
7584 Scale = cast<ConstantInt>(Inst->getOperand(1));
7585 NewIdx = Inst->getOperand(0);
7589 // If the index will be to exactly the right offset with the scale taken
7590 // out, perform the transformation.
7591 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7592 if (isa<ConstantInt>(Scale))
7593 Scale = ConstantInt::get(Scale->getType(),
7594 Scale->getZExtValue() / ArrayEltSize);
7595 if (Scale->getZExtValue() != 1) {
7596 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
7597 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7598 NewIdx = InsertNewInstBefore(Sc, GEP);
7601 // Insert the new GEP instruction.
7602 Instruction *NewGEP =
7603 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7604 NewIdx, GEP.getName());
7605 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7606 // The NewGEP must be pointer typed, so must the old one -> BitCast
7607 return new BitCastInst(NewGEP, GEP.getType());
7616 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7617 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7618 if (AI.isArrayAllocation()) // Check C != 1
7619 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7621 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7622 AllocationInst *New = 0;
7624 // Create and insert the replacement instruction...
7625 if (isa<MallocInst>(AI))
7626 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7628 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7629 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7632 InsertNewInstBefore(New, AI);
7634 // Scan to the end of the allocation instructions, to skip over a block of
7635 // allocas if possible...
7637 BasicBlock::iterator It = New;
7638 while (isa<AllocationInst>(*It)) ++It;
7640 // Now that I is pointing to the first non-allocation-inst in the block,
7641 // insert our getelementptr instruction...
7643 Value *NullIdx = Constant::getNullValue(Type::IntTy);
7644 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7645 New->getName()+".sub", It);
7647 // Now make everything use the getelementptr instead of the original
7649 return ReplaceInstUsesWith(AI, V);
7650 } else if (isa<UndefValue>(AI.getArraySize())) {
7651 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7654 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7655 // Note that we only do this for alloca's, because malloc should allocate and
7656 // return a unique pointer, even for a zero byte allocation.
7657 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7658 TD->getTypeSize(AI.getAllocatedType()) == 0)
7659 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7664 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7665 Value *Op = FI.getOperand(0);
7667 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7668 if (CastInst *CI = dyn_cast<CastInst>(Op))
7669 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7670 FI.setOperand(0, CI->getOperand(0));
7674 // free undef -> unreachable.
7675 if (isa<UndefValue>(Op)) {
7676 // Insert a new store to null because we cannot modify the CFG here.
7677 new StoreInst(ConstantBool::getTrue(),
7678 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
7679 return EraseInstFromFunction(FI);
7682 // If we have 'free null' delete the instruction. This can happen in stl code
7683 // when lots of inlining happens.
7684 if (isa<ConstantPointerNull>(Op))
7685 return EraseInstFromFunction(FI);
7691 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7692 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7693 User *CI = cast<User>(LI.getOperand(0));
7694 Value *CastOp = CI->getOperand(0);
7696 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7697 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7698 const Type *SrcPTy = SrcTy->getElementType();
7700 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7701 isa<PackedType>(DestPTy)) {
7702 // If the source is an array, the code below will not succeed. Check to
7703 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7705 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7706 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7707 if (ASrcTy->getNumElements() != 0) {
7708 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7709 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7710 SrcTy = cast<PointerType>(CastOp->getType());
7711 SrcPTy = SrcTy->getElementType();
7714 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7715 isa<PackedType>(SrcPTy)) &&
7716 // Do not allow turning this into a load of an integer, which is then
7717 // casted to a pointer, this pessimizes pointer analysis a lot.
7718 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7719 IC.getTargetData().getTypeSize(SrcPTy) ==
7720 IC.getTargetData().getTypeSize(DestPTy)) {
7722 // Okay, we are casting from one integer or pointer type to another of
7723 // the same size. Instead of casting the pointer before the load, cast
7724 // the result of the loaded value.
7725 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
7727 LI.isVolatile()),LI);
7728 // Now cast the result of the load.
7729 return CastInst::createInferredCast(NewLoad, LI.getType());
7736 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
7737 /// from this value cannot trap. If it is not obviously safe to load from the
7738 /// specified pointer, we do a quick local scan of the basic block containing
7739 /// ScanFrom, to determine if the address is already accessed.
7740 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
7741 // If it is an alloca or global variable, it is always safe to load from.
7742 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
7744 // Otherwise, be a little bit agressive by scanning the local block where we
7745 // want to check to see if the pointer is already being loaded or stored
7746 // from/to. If so, the previous load or store would have already trapped,
7747 // so there is no harm doing an extra load (also, CSE will later eliminate
7748 // the load entirely).
7749 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
7754 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7755 if (LI->getOperand(0) == V) return true;
7756 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7757 if (SI->getOperand(1) == V) return true;
7763 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
7764 Value *Op = LI.getOperand(0);
7766 // load (cast X) --> cast (load X) iff safe
7767 if (isa<CastInst>(Op))
7768 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7771 // None of the following transforms are legal for volatile loads.
7772 if (LI.isVolatile()) return 0;
7774 if (&LI.getParent()->front() != &LI) {
7775 BasicBlock::iterator BBI = &LI; --BBI;
7776 // If the instruction immediately before this is a store to the same
7777 // address, do a simple form of store->load forwarding.
7778 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7779 if (SI->getOperand(1) == LI.getOperand(0))
7780 return ReplaceInstUsesWith(LI, SI->getOperand(0));
7781 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
7782 if (LIB->getOperand(0) == LI.getOperand(0))
7783 return ReplaceInstUsesWith(LI, LIB);
7786 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
7787 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
7788 isa<UndefValue>(GEPI->getOperand(0))) {
7789 // Insert a new store to null instruction before the load to indicate
7790 // that this code is not reachable. We do this instead of inserting
7791 // an unreachable instruction directly because we cannot modify the
7793 new StoreInst(UndefValue::get(LI.getType()),
7794 Constant::getNullValue(Op->getType()), &LI);
7795 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7798 if (Constant *C = dyn_cast<Constant>(Op)) {
7799 // load null/undef -> undef
7800 if ((C->isNullValue() || isa<UndefValue>(C))) {
7801 // Insert a new store to null instruction before the load to indicate that
7802 // this code is not reachable. We do this instead of inserting an
7803 // unreachable instruction directly because we cannot modify the CFG.
7804 new StoreInst(UndefValue::get(LI.getType()),
7805 Constant::getNullValue(Op->getType()), &LI);
7806 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7809 // Instcombine load (constant global) into the value loaded.
7810 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
7811 if (GV->isConstant() && !GV->isExternal())
7812 return ReplaceInstUsesWith(LI, GV->getInitializer());
7814 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
7815 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7816 if (CE->getOpcode() == Instruction::GetElementPtr) {
7817 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
7818 if (GV->isConstant() && !GV->isExternal())
7820 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
7821 return ReplaceInstUsesWith(LI, V);
7822 if (CE->getOperand(0)->isNullValue()) {
7823 // Insert a new store to null instruction before the load to indicate
7824 // that this code is not reachable. We do this instead of inserting
7825 // an unreachable instruction directly because we cannot modify the
7827 new StoreInst(UndefValue::get(LI.getType()),
7828 Constant::getNullValue(Op->getType()), &LI);
7829 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7832 } else if (CE->isCast()) {
7833 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7838 if (Op->hasOneUse()) {
7839 // Change select and PHI nodes to select values instead of addresses: this
7840 // helps alias analysis out a lot, allows many others simplifications, and
7841 // exposes redundancy in the code.
7843 // Note that we cannot do the transformation unless we know that the
7844 // introduced loads cannot trap! Something like this is valid as long as
7845 // the condition is always false: load (select bool %C, int* null, int* %G),
7846 // but it would not be valid if we transformed it to load from null
7849 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7850 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7851 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7852 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7853 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
7854 SI->getOperand(1)->getName()+".val"), LI);
7855 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
7856 SI->getOperand(2)->getName()+".val"), LI);
7857 return new SelectInst(SI->getCondition(), V1, V2);
7860 // load (select (cond, null, P)) -> load P
7861 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7862 if (C->isNullValue()) {
7863 LI.setOperand(0, SI->getOperand(2));
7867 // load (select (cond, P, null)) -> load P
7868 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7869 if (C->isNullValue()) {
7870 LI.setOperand(0, SI->getOperand(1));
7878 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
7880 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7881 User *CI = cast<User>(SI.getOperand(1));
7882 Value *CastOp = CI->getOperand(0);
7884 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7885 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7886 const Type *SrcPTy = SrcTy->getElementType();
7888 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
7889 // If the source is an array, the code below will not succeed. Check to
7890 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7892 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7893 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7894 if (ASrcTy->getNumElements() != 0) {
7895 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7896 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7897 SrcTy = cast<PointerType>(CastOp->getType());
7898 SrcPTy = SrcTy->getElementType();
7901 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
7902 IC.getTargetData().getTypeSize(SrcPTy) ==
7903 IC.getTargetData().getTypeSize(DestPTy)) {
7905 // Okay, we are casting from one integer or pointer type to another of
7906 // the same size. Instead of casting the pointer before the store, cast
7907 // the value to be stored.
7909 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
7910 NewCast = ConstantExpr::getCast(C, SrcPTy);
7912 NewCast = IC.InsertNewInstBefore(
7913 CastInst::createInferredCast(SI.getOperand(0), SrcPTy,
7914 SI.getOperand(0)->getName()+".c"), SI);
7916 return new StoreInst(NewCast, CastOp);
7923 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7924 Value *Val = SI.getOperand(0);
7925 Value *Ptr = SI.getOperand(1);
7927 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
7928 EraseInstFromFunction(SI);
7933 // Do really simple DSE, to catch cases where there are several consequtive
7934 // stores to the same location, separated by a few arithmetic operations. This
7935 // situation often occurs with bitfield accesses.
7936 BasicBlock::iterator BBI = &SI;
7937 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7941 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7942 // Prev store isn't volatile, and stores to the same location?
7943 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
7946 EraseInstFromFunction(*PrevSI);
7952 // If this is a load, we have to stop. However, if the loaded value is from
7953 // the pointer we're loading and is producing the pointer we're storing,
7954 // then *this* store is dead (X = load P; store X -> P).
7955 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7956 if (LI == Val && LI->getOperand(0) == Ptr) {
7957 EraseInstFromFunction(SI);
7961 // Otherwise, this is a load from some other location. Stores before it
7966 // Don't skip over loads or things that can modify memory.
7967 if (BBI->mayWriteToMemory())
7972 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7974 // store X, null -> turns into 'unreachable' in SimplifyCFG
7975 if (isa<ConstantPointerNull>(Ptr)) {
7976 if (!isa<UndefValue>(Val)) {
7977 SI.setOperand(0, UndefValue::get(Val->getType()));
7978 if (Instruction *U = dyn_cast<Instruction>(Val))
7979 WorkList.push_back(U); // Dropped a use.
7982 return 0; // Do not modify these!
7985 // store undef, Ptr -> noop
7986 if (isa<UndefValue>(Val)) {
7987 EraseInstFromFunction(SI);
7992 // If the pointer destination is a cast, see if we can fold the cast into the
7994 if (isa<CastInst>(Ptr))
7995 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7997 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
7999 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8003 // If this store is the last instruction in the basic block, and if the block
8004 // ends with an unconditional branch, try to move it to the successor block.
8006 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8007 if (BI->isUnconditional()) {
8008 // Check to see if the successor block has exactly two incoming edges. If
8009 // so, see if the other predecessor contains a store to the same location.
8010 // if so, insert a PHI node (if needed) and move the stores down.
8011 BasicBlock *Dest = BI->getSuccessor(0);
8013 pred_iterator PI = pred_begin(Dest);
8014 BasicBlock *Other = 0;
8015 if (*PI != BI->getParent())
8018 if (PI != pred_end(Dest)) {
8019 if (*PI != BI->getParent())
8024 if (++PI != pred_end(Dest))
8027 if (Other) { // If only one other pred...
8028 BBI = Other->getTerminator();
8029 // Make sure this other block ends in an unconditional branch and that
8030 // there is an instruction before the branch.
8031 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8032 BBI != Other->begin()) {
8034 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8036 // If this instruction is a store to the same location.
8037 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8038 // Okay, we know we can perform this transformation. Insert a PHI
8039 // node now if we need it.
8040 Value *MergedVal = OtherStore->getOperand(0);
8041 if (MergedVal != SI.getOperand(0)) {
8042 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8043 PN->reserveOperandSpace(2);
8044 PN->addIncoming(SI.getOperand(0), SI.getParent());
8045 PN->addIncoming(OtherStore->getOperand(0), Other);
8046 MergedVal = InsertNewInstBefore(PN, Dest->front());
8049 // Advance to a place where it is safe to insert the new store and
8051 BBI = Dest->begin();
8052 while (isa<PHINode>(BBI)) ++BBI;
8053 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8054 OtherStore->isVolatile()), *BBI);
8056 // Nuke the old stores.
8057 EraseInstFromFunction(SI);
8058 EraseInstFromFunction(*OtherStore);
8070 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8071 // Change br (not X), label True, label False to: br X, label False, True
8073 BasicBlock *TrueDest;
8074 BasicBlock *FalseDest;
8075 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8076 !isa<Constant>(X)) {
8077 // Swap Destinations and condition...
8079 BI.setSuccessor(0, FalseDest);
8080 BI.setSuccessor(1, TrueDest);
8084 // Cannonicalize setne -> seteq
8085 Instruction::BinaryOps Op; Value *Y;
8086 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
8087 TrueDest, FalseDest)))
8088 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
8089 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
8090 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
8091 std::string Name = I->getName(); I->setName("");
8092 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
8093 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
8094 // Swap Destinations and condition...
8095 BI.setCondition(NewSCC);
8096 BI.setSuccessor(0, FalseDest);
8097 BI.setSuccessor(1, TrueDest);
8098 removeFromWorkList(I);
8099 I->getParent()->getInstList().erase(I);
8100 WorkList.push_back(cast<Instruction>(NewSCC));
8107 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8108 Value *Cond = SI.getCondition();
8109 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8110 if (I->getOpcode() == Instruction::Add)
8111 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8112 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8113 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8114 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8116 SI.setOperand(0, I->getOperand(0));
8117 WorkList.push_back(I);
8124 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8125 /// is to leave as a vector operation.
8126 static bool CheapToScalarize(Value *V, bool isConstant) {
8127 if (isa<ConstantAggregateZero>(V))
8129 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8130 if (isConstant) return true;
8131 // If all elts are the same, we can extract.
8132 Constant *Op0 = C->getOperand(0);
8133 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8134 if (C->getOperand(i) != Op0)
8138 Instruction *I = dyn_cast<Instruction>(V);
8139 if (!I) return false;
8141 // Insert element gets simplified to the inserted element or is deleted if
8142 // this is constant idx extract element and its a constant idx insertelt.
8143 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8144 isa<ConstantInt>(I->getOperand(2)))
8146 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8148 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8149 if (BO->hasOneUse() &&
8150 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8151 CheapToScalarize(BO->getOperand(1), isConstant)))
8157 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8158 /// elements into values that are larger than the #elts in the input.
8159 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8160 unsigned NElts = SVI->getType()->getNumElements();
8161 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8162 return std::vector<unsigned>(NElts, 0);
8163 if (isa<UndefValue>(SVI->getOperand(2)))
8164 return std::vector<unsigned>(NElts, 2*NElts);
8166 std::vector<unsigned> Result;
8167 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8168 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8169 if (isa<UndefValue>(CP->getOperand(i)))
8170 Result.push_back(NElts*2); // undef -> 8
8172 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8176 /// FindScalarElement - Given a vector and an element number, see if the scalar
8177 /// value is already around as a register, for example if it were inserted then
8178 /// extracted from the vector.
8179 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8180 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8181 const PackedType *PTy = cast<PackedType>(V->getType());
8182 unsigned Width = PTy->getNumElements();
8183 if (EltNo >= Width) // Out of range access.
8184 return UndefValue::get(PTy->getElementType());
8186 if (isa<UndefValue>(V))
8187 return UndefValue::get(PTy->getElementType());
8188 else if (isa<ConstantAggregateZero>(V))
8189 return Constant::getNullValue(PTy->getElementType());
8190 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8191 return CP->getOperand(EltNo);
8192 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8193 // If this is an insert to a variable element, we don't know what it is.
8194 if (!isa<ConstantInt>(III->getOperand(2)))
8196 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8198 // If this is an insert to the element we are looking for, return the
8201 return III->getOperand(1);
8203 // Otherwise, the insertelement doesn't modify the value, recurse on its
8205 return FindScalarElement(III->getOperand(0), EltNo);
8206 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8207 unsigned InEl = getShuffleMask(SVI)[EltNo];
8209 return FindScalarElement(SVI->getOperand(0), InEl);
8210 else if (InEl < Width*2)
8211 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8213 return UndefValue::get(PTy->getElementType());
8216 // Otherwise, we don't know.
8220 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8222 // If packed val is undef, replace extract with scalar undef.
8223 if (isa<UndefValue>(EI.getOperand(0)))
8224 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8226 // If packed val is constant 0, replace extract with scalar 0.
8227 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8228 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8230 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8231 // If packed val is constant with uniform operands, replace EI
8232 // with that operand
8233 Constant *op0 = C->getOperand(0);
8234 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8235 if (C->getOperand(i) != op0) {
8240 return ReplaceInstUsesWith(EI, op0);
8243 // If extracting a specified index from the vector, see if we can recursively
8244 // find a previously computed scalar that was inserted into the vector.
8245 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8246 // This instruction only demands the single element from the input vector.
8247 // If the input vector has a single use, simplify it based on this use
8249 uint64_t IndexVal = IdxC->getZExtValue();
8250 if (EI.getOperand(0)->hasOneUse()) {
8252 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8255 EI.setOperand(0, V);
8260 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8261 return ReplaceInstUsesWith(EI, Elt);
8264 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8265 if (I->hasOneUse()) {
8266 // Push extractelement into predecessor operation if legal and
8267 // profitable to do so
8268 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8269 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8270 if (CheapToScalarize(BO, isConstantElt)) {
8271 ExtractElementInst *newEI0 =
8272 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8273 EI.getName()+".lhs");
8274 ExtractElementInst *newEI1 =
8275 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8276 EI.getName()+".rhs");
8277 InsertNewInstBefore(newEI0, EI);
8278 InsertNewInstBefore(newEI1, EI);
8279 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8281 } else if (isa<LoadInst>(I)) {
8282 Value *Ptr = InsertCastBefore(I->getOperand(0),
8283 PointerType::get(EI.getType()), EI);
8284 GetElementPtrInst *GEP =
8285 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8286 InsertNewInstBefore(GEP, EI);
8287 return new LoadInst(GEP);
8290 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8291 // Extracting the inserted element?
8292 if (IE->getOperand(2) == EI.getOperand(1))
8293 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8294 // If the inserted and extracted elements are constants, they must not
8295 // be the same value, extract from the pre-inserted value instead.
8296 if (isa<Constant>(IE->getOperand(2)) &&
8297 isa<Constant>(EI.getOperand(1))) {
8298 AddUsesToWorkList(EI);
8299 EI.setOperand(0, IE->getOperand(0));
8302 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8303 // If this is extracting an element from a shufflevector, figure out where
8304 // it came from and extract from the appropriate input element instead.
8305 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8306 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8308 if (SrcIdx < SVI->getType()->getNumElements())
8309 Src = SVI->getOperand(0);
8310 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8311 SrcIdx -= SVI->getType()->getNumElements();
8312 Src = SVI->getOperand(1);
8314 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8316 return new ExtractElementInst(Src, SrcIdx);
8323 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8324 /// elements from either LHS or RHS, return the shuffle mask and true.
8325 /// Otherwise, return false.
8326 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8327 std::vector<Constant*> &Mask) {
8328 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8329 "Invalid CollectSingleShuffleElements");
8330 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8332 if (isa<UndefValue>(V)) {
8333 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8335 } else if (V == LHS) {
8336 for (unsigned i = 0; i != NumElts; ++i)
8337 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8339 } else if (V == RHS) {
8340 for (unsigned i = 0; i != NumElts; ++i)
8341 Mask.push_back(ConstantInt::get(Type::UIntTy, i+NumElts));
8343 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8344 // If this is an insert of an extract from some other vector, include it.
8345 Value *VecOp = IEI->getOperand(0);
8346 Value *ScalarOp = IEI->getOperand(1);
8347 Value *IdxOp = IEI->getOperand(2);
8349 if (!isa<ConstantInt>(IdxOp))
8351 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8353 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8354 // Okay, we can handle this if the vector we are insertinting into is
8356 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8357 // If so, update the mask to reflect the inserted undef.
8358 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
8361 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8362 if (isa<ConstantInt>(EI->getOperand(1)) &&
8363 EI->getOperand(0)->getType() == V->getType()) {
8364 unsigned ExtractedIdx =
8365 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8367 // This must be extracting from either LHS or RHS.
8368 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8369 // Okay, we can handle this if the vector we are insertinting into is
8371 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8372 // If so, update the mask to reflect the inserted value.
8373 if (EI->getOperand(0) == LHS) {
8374 Mask[InsertedIdx & (NumElts-1)] =
8375 ConstantInt::get(Type::UIntTy, ExtractedIdx);
8377 assert(EI->getOperand(0) == RHS);
8378 Mask[InsertedIdx & (NumElts-1)] =
8379 ConstantInt::get(Type::UIntTy, ExtractedIdx+NumElts);
8388 // TODO: Handle shufflevector here!
8393 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8394 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8395 /// that computes V and the LHS value of the shuffle.
8396 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8398 assert(isa<PackedType>(V->getType()) &&
8399 (RHS == 0 || V->getType() == RHS->getType()) &&
8400 "Invalid shuffle!");
8401 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8403 if (isa<UndefValue>(V)) {
8404 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8406 } else if (isa<ConstantAggregateZero>(V)) {
8407 Mask.assign(NumElts, ConstantInt::get(Type::UIntTy, 0));
8409 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8410 // If this is an insert of an extract from some other vector, include it.
8411 Value *VecOp = IEI->getOperand(0);
8412 Value *ScalarOp = IEI->getOperand(1);
8413 Value *IdxOp = IEI->getOperand(2);
8415 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8416 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8417 EI->getOperand(0)->getType() == V->getType()) {
8418 unsigned ExtractedIdx =
8419 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8420 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8422 // Either the extracted from or inserted into vector must be RHSVec,
8423 // otherwise we'd end up with a shuffle of three inputs.
8424 if (EI->getOperand(0) == RHS || RHS == 0) {
8425 RHS = EI->getOperand(0);
8426 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8427 Mask[InsertedIdx & (NumElts-1)] =
8428 ConstantInt::get(Type::UIntTy, NumElts+ExtractedIdx);
8433 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8434 // Everything but the extracted element is replaced with the RHS.
8435 for (unsigned i = 0; i != NumElts; ++i) {
8436 if (i != InsertedIdx)
8437 Mask[i] = ConstantInt::get(Type::UIntTy, NumElts+i);
8442 // If this insertelement is a chain that comes from exactly these two
8443 // vectors, return the vector and the effective shuffle.
8444 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8445 return EI->getOperand(0);
8450 // TODO: Handle shufflevector here!
8452 // Otherwise, can't do anything fancy. Return an identity vector.
8453 for (unsigned i = 0; i != NumElts; ++i)
8454 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8458 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8459 Value *VecOp = IE.getOperand(0);
8460 Value *ScalarOp = IE.getOperand(1);
8461 Value *IdxOp = IE.getOperand(2);
8463 // If the inserted element was extracted from some other vector, and if the
8464 // indexes are constant, try to turn this into a shufflevector operation.
8465 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8466 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8467 EI->getOperand(0)->getType() == IE.getType()) {
8468 unsigned NumVectorElts = IE.getType()->getNumElements();
8469 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8470 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8472 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8473 return ReplaceInstUsesWith(IE, VecOp);
8475 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8476 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8478 // If we are extracting a value from a vector, then inserting it right
8479 // back into the same place, just use the input vector.
8480 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8481 return ReplaceInstUsesWith(IE, VecOp);
8483 // We could theoretically do this for ANY input. However, doing so could
8484 // turn chains of insertelement instructions into a chain of shufflevector
8485 // instructions, and right now we do not merge shufflevectors. As such,
8486 // only do this in a situation where it is clear that there is benefit.
8487 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8488 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8489 // the values of VecOp, except then one read from EIOp0.
8490 // Build a new shuffle mask.
8491 std::vector<Constant*> Mask;
8492 if (isa<UndefValue>(VecOp))
8493 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
8495 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8496 Mask.assign(NumVectorElts, ConstantInt::get(Type::UIntTy,
8499 Mask[InsertedIdx] = ConstantInt::get(Type::UIntTy, ExtractedIdx);
8500 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8501 ConstantPacked::get(Mask));
8504 // If this insertelement isn't used by some other insertelement, turn it
8505 // (and any insertelements it points to), into one big shuffle.
8506 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8507 std::vector<Constant*> Mask;
8509 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8510 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8511 // We now have a shuffle of LHS, RHS, Mask.
8512 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8521 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8522 Value *LHS = SVI.getOperand(0);
8523 Value *RHS = SVI.getOperand(1);
8524 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8526 bool MadeChange = false;
8528 // Undefined shuffle mask -> undefined value.
8529 if (isa<UndefValue>(SVI.getOperand(2)))
8530 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8532 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
8533 // the undef, change them to undefs.
8535 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8536 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8537 if (LHS == RHS || isa<UndefValue>(LHS)) {
8538 if (isa<UndefValue>(LHS) && LHS == RHS) {
8539 // shuffle(undef,undef,mask) -> undef.
8540 return ReplaceInstUsesWith(SVI, LHS);
8543 // Remap any references to RHS to use LHS.
8544 std::vector<Constant*> Elts;
8545 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8547 Elts.push_back(UndefValue::get(Type::UIntTy));
8549 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8550 (Mask[i] < e && isa<UndefValue>(LHS)))
8551 Mask[i] = 2*e; // Turn into undef.
8553 Mask[i] &= (e-1); // Force to LHS.
8554 Elts.push_back(ConstantInt::get(Type::UIntTy, Mask[i]));
8557 SVI.setOperand(0, SVI.getOperand(1));
8558 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8559 SVI.setOperand(2, ConstantPacked::get(Elts));
8560 LHS = SVI.getOperand(0);
8561 RHS = SVI.getOperand(1);
8565 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8566 bool isLHSID = true, isRHSID = true;
8568 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8569 if (Mask[i] >= e*2) continue; // Ignore undef values.
8570 // Is this an identity shuffle of the LHS value?
8571 isLHSID &= (Mask[i] == i);
8573 // Is this an identity shuffle of the RHS value?
8574 isRHSID &= (Mask[i]-e == i);
8577 // Eliminate identity shuffles.
8578 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8579 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8581 // If the LHS is a shufflevector itself, see if we can combine it with this
8582 // one without producing an unusual shuffle. Here we are really conservative:
8583 // we are absolutely afraid of producing a shuffle mask not in the input
8584 // program, because the code gen may not be smart enough to turn a merged
8585 // shuffle into two specific shuffles: it may produce worse code. As such,
8586 // we only merge two shuffles if the result is one of the two input shuffle
8587 // masks. In this case, merging the shuffles just removes one instruction,
8588 // which we know is safe. This is good for things like turning:
8589 // (splat(splat)) -> splat.
8590 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8591 if (isa<UndefValue>(RHS)) {
8592 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8594 std::vector<unsigned> NewMask;
8595 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8597 NewMask.push_back(2*e);
8599 NewMask.push_back(LHSMask[Mask[i]]);
8601 // If the result mask is equal to the src shuffle or this shuffle mask, do
8603 if (NewMask == LHSMask || NewMask == Mask) {
8604 std::vector<Constant*> Elts;
8605 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8606 if (NewMask[i] >= e*2) {
8607 Elts.push_back(UndefValue::get(Type::UIntTy));
8609 Elts.push_back(ConstantInt::get(Type::UIntTy, NewMask[i]));
8612 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8613 LHSSVI->getOperand(1),
8614 ConstantPacked::get(Elts));
8619 return MadeChange ? &SVI : 0;
8624 void InstCombiner::removeFromWorkList(Instruction *I) {
8625 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8630 /// TryToSinkInstruction - Try to move the specified instruction from its
8631 /// current block into the beginning of DestBlock, which can only happen if it's
8632 /// safe to move the instruction past all of the instructions between it and the
8633 /// end of its block.
8634 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8635 assert(I->hasOneUse() && "Invariants didn't hold!");
8637 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8638 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
8640 // Do not sink alloca instructions out of the entry block.
8641 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
8644 // We can only sink load instructions if there is nothing between the load and
8645 // the end of block that could change the value.
8646 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8647 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
8649 if (Scan->mayWriteToMemory())
8653 BasicBlock::iterator InsertPos = DestBlock->begin();
8654 while (isa<PHINode>(InsertPos)) ++InsertPos;
8656 I->moveBefore(InsertPos);
8661 /// OptimizeConstantExpr - Given a constant expression and target data layout
8662 /// information, symbolically evaluation the constant expr to something simpler
8664 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
8667 Constant *Ptr = CE->getOperand(0);
8668 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
8669 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
8670 // If this is a constant expr gep that is effectively computing an
8671 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
8672 bool isFoldableGEP = true;
8673 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
8674 if (!isa<ConstantInt>(CE->getOperand(i)))
8675 isFoldableGEP = false;
8676 if (isFoldableGEP) {
8677 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
8678 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
8679 Constant *C = ConstantInt::get(Type::ULongTy, Offset);
8680 C = ConstantExpr::getCast(C, TD->getIntPtrType());
8681 return ConstantExpr::getCast(C, CE->getType());
8689 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
8690 /// all reachable code to the worklist.
8692 /// This has a couple of tricks to make the code faster and more powerful. In
8693 /// particular, we constant fold and DCE instructions as we go, to avoid adding
8694 /// them to the worklist (this significantly speeds up instcombine on code where
8695 /// many instructions are dead or constant). Additionally, if we find a branch
8696 /// whose condition is a known constant, we only visit the reachable successors.
8698 static void AddReachableCodeToWorklist(BasicBlock *BB,
8699 std::set<BasicBlock*> &Visited,
8700 std::vector<Instruction*> &WorkList,
8701 const TargetData *TD) {
8702 // We have now visited this block! If we've already been here, bail out.
8703 if (!Visited.insert(BB).second) return;
8705 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
8706 Instruction *Inst = BBI++;
8708 // DCE instruction if trivially dead.
8709 if (isInstructionTriviallyDead(Inst)) {
8711 DOUT << "IC: DCE: " << *Inst;
8712 Inst->eraseFromParent();
8716 // ConstantProp instruction if trivially constant.
8717 if (Constant *C = ConstantFoldInstruction(Inst)) {
8718 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8719 C = OptimizeConstantExpr(CE, TD);
8720 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
8721 Inst->replaceAllUsesWith(C);
8723 Inst->eraseFromParent();
8727 WorkList.push_back(Inst);
8730 // Recursively visit successors. If this is a branch or switch on a constant,
8731 // only visit the reachable successor.
8732 TerminatorInst *TI = BB->getTerminator();
8733 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
8734 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
8735 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
8736 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
8740 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
8741 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
8742 // See if this is an explicit destination.
8743 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
8744 if (SI->getCaseValue(i) == Cond) {
8745 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
8749 // Otherwise it is the default destination.
8750 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
8755 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
8756 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
8759 bool InstCombiner::runOnFunction(Function &F) {
8760 bool Changed = false;
8761 TD = &getAnalysis<TargetData>();
8764 // Do a depth-first traversal of the function, populate the worklist with
8765 // the reachable instructions. Ignore blocks that are not reachable. Keep
8766 // track of which blocks we visit.
8767 std::set<BasicBlock*> Visited;
8768 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
8770 // Do a quick scan over the function. If we find any blocks that are
8771 // unreachable, remove any instructions inside of them. This prevents
8772 // the instcombine code from having to deal with some bad special cases.
8773 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
8774 if (!Visited.count(BB)) {
8775 Instruction *Term = BB->getTerminator();
8776 while (Term != BB->begin()) { // Remove instrs bottom-up
8777 BasicBlock::iterator I = Term; --I;
8779 DOUT << "IC: DCE: " << *I;
8782 if (!I->use_empty())
8783 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8784 I->eraseFromParent();
8789 while (!WorkList.empty()) {
8790 Instruction *I = WorkList.back(); // Get an instruction from the worklist
8791 WorkList.pop_back();
8793 // Check to see if we can DCE the instruction.
8794 if (isInstructionTriviallyDead(I)) {
8795 // Add operands to the worklist.
8796 if (I->getNumOperands() < 4)
8797 AddUsesToWorkList(*I);
8800 DOUT << "IC: DCE: " << *I;
8802 I->eraseFromParent();
8803 removeFromWorkList(I);
8807 // Instruction isn't dead, see if we can constant propagate it.
8808 if (Constant *C = ConstantFoldInstruction(I)) {
8809 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8810 C = OptimizeConstantExpr(CE, TD);
8811 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
8813 // Add operands to the worklist.
8814 AddUsesToWorkList(*I);
8815 ReplaceInstUsesWith(*I, C);
8818 I->eraseFromParent();
8819 removeFromWorkList(I);
8823 // See if we can trivially sink this instruction to a successor basic block.
8824 if (I->hasOneUse()) {
8825 BasicBlock *BB = I->getParent();
8826 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
8827 if (UserParent != BB) {
8828 bool UserIsSuccessor = false;
8829 // See if the user is one of our successors.
8830 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8831 if (*SI == UserParent) {
8832 UserIsSuccessor = true;
8836 // If the user is one of our immediate successors, and if that successor
8837 // only has us as a predecessors (we'd have to split the critical edge
8838 // otherwise), we can keep going.
8839 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
8840 next(pred_begin(UserParent)) == pred_end(UserParent))
8841 // Okay, the CFG is simple enough, try to sink this instruction.
8842 Changed |= TryToSinkInstruction(I, UserParent);
8846 // Now that we have an instruction, try combining it to simplify it...
8847 if (Instruction *Result = visit(*I)) {
8849 // Should we replace the old instruction with a new one?
8851 DOUT << "IC: Old = " << *I
8852 << " New = " << *Result;
8854 // Everything uses the new instruction now.
8855 I->replaceAllUsesWith(Result);
8857 // Push the new instruction and any users onto the worklist.
8858 WorkList.push_back(Result);
8859 AddUsersToWorkList(*Result);
8861 // Move the name to the new instruction first...
8862 std::string OldName = I->getName(); I->setName("");
8863 Result->setName(OldName);
8865 // Insert the new instruction into the basic block...
8866 BasicBlock *InstParent = I->getParent();
8867 BasicBlock::iterator InsertPos = I;
8869 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8870 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8873 InstParent->getInstList().insert(InsertPos, Result);
8875 // Make sure that we reprocess all operands now that we reduced their
8877 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8878 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8879 WorkList.push_back(OpI);
8881 // Instructions can end up on the worklist more than once. Make sure
8882 // we do not process an instruction that has been deleted.
8883 removeFromWorkList(I);
8885 // Erase the old instruction.
8886 InstParent->getInstList().erase(I);
8888 DOUT << "IC: MOD = " << *I;
8890 // If the instruction was modified, it's possible that it is now dead.
8891 // if so, remove it.
8892 if (isInstructionTriviallyDead(I)) {
8893 // Make sure we process all operands now that we are reducing their
8895 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8896 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8897 WorkList.push_back(OpI);
8899 // Instructions may end up in the worklist more than once. Erase all
8900 // occurrences of this instruction.
8901 removeFromWorkList(I);
8902 I->eraseFromParent();
8904 WorkList.push_back(Result);
8905 AddUsersToWorkList(*Result);
8915 FunctionPass *llvm::createInstructionCombiningPass() {
8916 return new InstCombiner();