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"
57 using namespace llvm::PatternMatch;
60 Statistic<> NumCombined ("instcombine", "Number of insts combined");
61 Statistic<> NumConstProp("instcombine", "Number of constant folds");
62 Statistic<> NumDeadInst ("instcombine", "Number of dead inst eliminated");
63 Statistic<> NumDeadStore("instcombine", "Number of dead stores eliminated");
64 Statistic<> NumSunkInst ("instcombine", "Number of instructions sunk");
66 class VISIBILITY_HIDDEN InstCombiner
67 : public FunctionPass,
68 public InstVisitor<InstCombiner, Instruction*> {
69 // Worklist of all of the instructions that need to be simplified.
70 std::vector<Instruction*> WorkList;
73 /// AddUsersToWorkList - When an instruction is simplified, add all users of
74 /// the instruction to the work lists because they might get more simplified
77 void AddUsersToWorkList(Value &I) {
78 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
80 WorkList.push_back(cast<Instruction>(*UI));
83 /// AddUsesToWorkList - When an instruction is simplified, add operands to
84 /// the work lists because they might get more simplified now.
86 void AddUsesToWorkList(Instruction &I) {
87 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
88 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
89 WorkList.push_back(Op);
92 // removeFromWorkList - remove all instances of I from the worklist.
93 void removeFromWorkList(Instruction *I);
95 virtual bool runOnFunction(Function &F);
97 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
98 AU.addRequired<TargetData>();
99 AU.addPreservedID(LCSSAID);
100 AU.setPreservesCFG();
103 TargetData &getTargetData() const { return *TD; }
105 // Visitation implementation - Implement instruction combining for different
106 // instruction types. The semantics are as follows:
108 // null - No change was made
109 // I - Change was made, I is still valid, I may be dead though
110 // otherwise - Change was made, replace I with returned instruction
112 Instruction *visitAdd(BinaryOperator &I);
113 Instruction *visitSub(BinaryOperator &I);
114 Instruction *visitMul(BinaryOperator &I);
115 Instruction *visitDiv(BinaryOperator &I);
116 Instruction *visitRem(BinaryOperator &I);
117 Instruction *visitAnd(BinaryOperator &I);
118 Instruction *visitOr (BinaryOperator &I);
119 Instruction *visitXor(BinaryOperator &I);
120 Instruction *visitSetCondInst(SetCondInst &I);
121 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
123 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
124 Instruction::BinaryOps Cond, Instruction &I);
125 Instruction *visitShiftInst(ShiftInst &I);
126 Instruction *FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
128 Instruction *visitCastInst(CastInst &CI);
129 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
131 Instruction *visitSelectInst(SelectInst &CI);
132 Instruction *visitCallInst(CallInst &CI);
133 Instruction *visitInvokeInst(InvokeInst &II);
134 Instruction *visitPHINode(PHINode &PN);
135 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
136 Instruction *visitAllocationInst(AllocationInst &AI);
137 Instruction *visitFreeInst(FreeInst &FI);
138 Instruction *visitLoadInst(LoadInst &LI);
139 Instruction *visitStoreInst(StoreInst &SI);
140 Instruction *visitBranchInst(BranchInst &BI);
141 Instruction *visitSwitchInst(SwitchInst &SI);
142 Instruction *visitInsertElementInst(InsertElementInst &IE);
143 Instruction *visitExtractElementInst(ExtractElementInst &EI);
144 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
146 // visitInstruction - Specify what to return for unhandled instructions...
147 Instruction *visitInstruction(Instruction &I) { return 0; }
150 Instruction *visitCallSite(CallSite CS);
151 bool transformConstExprCastCall(CallSite CS);
154 // InsertNewInstBefore - insert an instruction New before instruction Old
155 // in the program. Add the new instruction to the worklist.
157 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
158 assert(New && New->getParent() == 0 &&
159 "New instruction already inserted into a basic block!");
160 BasicBlock *BB = Old.getParent();
161 BB->getInstList().insert(&Old, New); // Insert inst
162 WorkList.push_back(New); // Add to worklist
166 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
167 /// This also adds the cast to the worklist. Finally, this returns the
169 Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
170 if (V->getType() == Ty) return V;
172 if (Constant *CV = dyn_cast<Constant>(V))
173 return ConstantExpr::getCast(CV, Ty);
175 Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
176 WorkList.push_back(C);
180 // ReplaceInstUsesWith - This method is to be used when an instruction is
181 // found to be dead, replacable with another preexisting expression. Here
182 // we add all uses of I to the worklist, replace all uses of I with the new
183 // value, then return I, so that the inst combiner will know that I was
186 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
187 AddUsersToWorkList(I); // Add all modified instrs to worklist
189 I.replaceAllUsesWith(V);
192 // If we are replacing the instruction with itself, this must be in a
193 // segment of unreachable code, so just clobber the instruction.
194 I.replaceAllUsesWith(UndefValue::get(I.getType()));
199 // UpdateValueUsesWith - This method is to be used when an value is
200 // found to be replacable with another preexisting expression or was
201 // updated. Here we add all uses of I to the worklist, replace all uses of
202 // I with the new value (unless the instruction was just updated), then
203 // return true, so that the inst combiner will know that I was modified.
205 bool UpdateValueUsesWith(Value *Old, Value *New) {
206 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
208 Old->replaceAllUsesWith(New);
209 if (Instruction *I = dyn_cast<Instruction>(Old))
210 WorkList.push_back(I);
211 if (Instruction *I = dyn_cast<Instruction>(New))
212 WorkList.push_back(I);
216 // EraseInstFromFunction - When dealing with an instruction that has side
217 // effects or produces a void value, we can't rely on DCE to delete the
218 // instruction. Instead, visit methods should return the value returned by
220 Instruction *EraseInstFromFunction(Instruction &I) {
221 assert(I.use_empty() && "Cannot erase instruction that is used!");
222 AddUsesToWorkList(I);
223 removeFromWorkList(&I);
225 return 0; // Don't do anything with FI
229 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
230 /// InsertBefore instruction. This is specialized a bit to avoid inserting
231 /// casts that are known to not do anything...
233 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
234 Instruction *InsertBefore);
236 // SimplifyCommutative - This performs a few simplifications for commutative
238 bool SimplifyCommutative(BinaryOperator &I);
240 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
241 uint64_t &KnownZero, uint64_t &KnownOne,
244 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
245 // PHI node as operand #0, see if we can fold the instruction into the PHI
246 // (which is only possible if all operands to the PHI are constants).
247 Instruction *FoldOpIntoPhi(Instruction &I);
249 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
250 // operator and they all are only used by the PHI, PHI together their
251 // inputs, and do the operation once, to the result of the PHI.
252 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
254 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
255 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
257 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
258 bool isSub, Instruction &I);
259 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
260 bool Inside, Instruction &IB);
261 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
262 Instruction *MatchBSwap(BinaryOperator &I);
264 Value *EvaluateInDifferentType(Value *V, const Type *Ty);
267 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
270 // getComplexity: Assign a complexity or rank value to LLVM Values...
271 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
272 static unsigned getComplexity(Value *V) {
273 if (isa<Instruction>(V)) {
274 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
278 if (isa<Argument>(V)) return 3;
279 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
282 // isOnlyUse - Return true if this instruction will be deleted if we stop using
284 static bool isOnlyUse(Value *V) {
285 return V->hasOneUse() || isa<Constant>(V);
288 // getPromotedType - Return the specified type promoted as it would be to pass
289 // though a va_arg area...
290 static const Type *getPromotedType(const Type *Ty) {
291 switch (Ty->getTypeID()) {
292 case Type::SByteTyID:
293 case Type::ShortTyID: return Type::IntTy;
294 case Type::UByteTyID:
295 case Type::UShortTyID: return Type::UIntTy;
296 case Type::FloatTyID: return Type::DoubleTy;
301 /// isCast - If the specified operand is a CastInst or a constant expr cast,
302 /// return the operand value, otherwise return null.
303 static Value *isCast(Value *V) {
304 if (CastInst *I = dyn_cast<CastInst>(V))
305 return I->getOperand(0);
306 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
307 if (CE->getOpcode() == Instruction::Cast)
308 return CE->getOperand(0);
319 /// getCastType - In the future, we will split the cast instruction into these
320 /// various types. Until then, we have to do the analysis here.
321 static CastType getCastType(const Type *Src, const Type *Dest) {
322 assert(Src->isIntegral() && Dest->isIntegral() &&
323 "Only works on integral types!");
324 unsigned SrcSize = Src->getPrimitiveSizeInBits();
325 unsigned DestSize = Dest->getPrimitiveSizeInBits();
327 if (SrcSize == DestSize) return Noop;
328 if (SrcSize > DestSize) return Truncate;
329 if (Src->isSigned()) return Signext;
334 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
337 static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
338 const Type *DstTy, TargetData *TD) {
340 // It is legal to eliminate the instruction if casting A->B->A if the sizes
341 // are identical and the bits don't get reinterpreted (for example
342 // int->float->int would not be allowed).
343 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
346 // If we are casting between pointer and integer types, treat pointers as
347 // integers of the appropriate size for the code below.
348 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
349 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
350 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
352 // Allow free casting and conversion of sizes as long as the sign doesn't
354 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
355 CastType FirstCast = getCastType(SrcTy, MidTy);
356 CastType SecondCast = getCastType(MidTy, DstTy);
358 // Capture the effect of these two casts. If the result is a legal cast,
359 // the CastType is stored here, otherwise a special code is used.
360 static const unsigned CastResult[] = {
361 // First cast is noop
363 // First cast is a truncate
364 1, 1, 4, 4, // trunc->extend is not safe to eliminate
365 // First cast is a sign ext
366 2, 5, 2, 4, // signext->zeroext never ok
367 // First cast is a zero ext
371 unsigned Result = CastResult[FirstCast*4+SecondCast];
373 default: assert(0 && "Illegal table value!");
378 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
379 // truncates, we could eliminate more casts.
380 return (unsigned)getCastType(SrcTy, DstTy) == Result;
382 return false; // Not possible to eliminate this here.
384 // Sign or zero extend followed by truncate is always ok if the result
385 // is a truncate or noop.
386 CastType ResultCast = getCastType(SrcTy, DstTy);
387 if (ResultCast == Noop || ResultCast == Truncate)
389 // Otherwise we are still growing the value, we are only safe if the
390 // result will match the sign/zeroextendness of the result.
391 return ResultCast == FirstCast;
395 // If this is a cast from 'float -> double -> integer', cast from
396 // 'float -> integer' directly, as the value isn't changed by the
397 // float->double conversion.
398 if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
399 DstTy->isIntegral() &&
400 SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
403 // Packed type conversions don't modify bits.
404 if (isa<PackedType>(SrcTy) && isa<PackedType>(MidTy) &&isa<PackedType>(DstTy))
410 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
411 /// in any code being generated. It does not require codegen if V is simple
412 /// enough or if the cast can be folded into other casts.
413 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
414 if (V->getType() == Ty || isa<Constant>(V)) return false;
416 // If this is a noop cast, it isn't real codegen.
417 if (V->getType()->isLosslesslyConvertibleTo(Ty))
420 // If this is another cast that can be eliminated, it isn't codegen either.
421 if (const CastInst *CI = dyn_cast<CastInst>(V))
422 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
428 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
429 /// InsertBefore instruction. This is specialized a bit to avoid inserting
430 /// casts that are known to not do anything...
432 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
433 Instruction *InsertBefore) {
434 if (V->getType() == DestTy) return V;
435 if (Constant *C = dyn_cast<Constant>(V))
436 return ConstantExpr::getCast(C, DestTy);
438 CastInst *CI = new CastInst(V, DestTy, V->getName());
439 InsertNewInstBefore(CI, *InsertBefore);
443 // SimplifyCommutative - This performs a few simplifications for commutative
446 // 1. Order operands such that they are listed from right (least complex) to
447 // left (most complex). This puts constants before unary operators before
450 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
451 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
453 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
454 bool Changed = false;
455 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
456 Changed = !I.swapOperands();
458 if (!I.isAssociative()) return Changed;
459 Instruction::BinaryOps Opcode = I.getOpcode();
460 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
461 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
462 if (isa<Constant>(I.getOperand(1))) {
463 Constant *Folded = ConstantExpr::get(I.getOpcode(),
464 cast<Constant>(I.getOperand(1)),
465 cast<Constant>(Op->getOperand(1)));
466 I.setOperand(0, Op->getOperand(0));
467 I.setOperand(1, Folded);
469 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
470 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
471 isOnlyUse(Op) && isOnlyUse(Op1)) {
472 Constant *C1 = cast<Constant>(Op->getOperand(1));
473 Constant *C2 = cast<Constant>(Op1->getOperand(1));
475 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
476 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
477 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
480 WorkList.push_back(New);
481 I.setOperand(0, New);
482 I.setOperand(1, Folded);
489 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
490 // if the LHS is a constant zero (which is the 'negate' form).
492 static inline Value *dyn_castNegVal(Value *V) {
493 if (BinaryOperator::isNeg(V))
494 return BinaryOperator::getNegArgument(V);
496 // Constants can be considered to be negated values if they can be folded.
497 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
498 return ConstantExpr::getNeg(C);
502 static inline Value *dyn_castNotVal(Value *V) {
503 if (BinaryOperator::isNot(V))
504 return BinaryOperator::getNotArgument(V);
506 // Constants can be considered to be not'ed values...
507 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
508 return ConstantExpr::getNot(C);
512 // dyn_castFoldableMul - If this value is a multiply that can be folded into
513 // other computations (because it has a constant operand), return the
514 // non-constant operand of the multiply, and set CST to point to the multiplier.
515 // Otherwise, return null.
517 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
518 if (V->hasOneUse() && V->getType()->isInteger())
519 if (Instruction *I = dyn_cast<Instruction>(V)) {
520 if (I->getOpcode() == Instruction::Mul)
521 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
522 return I->getOperand(0);
523 if (I->getOpcode() == Instruction::Shl)
524 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
525 // The multiplier is really 1 << CST.
526 Constant *One = ConstantInt::get(V->getType(), 1);
527 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
528 return I->getOperand(0);
534 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
535 /// expression, return it.
536 static User *dyn_castGetElementPtr(Value *V) {
537 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
538 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
539 if (CE->getOpcode() == Instruction::GetElementPtr)
540 return cast<User>(V);
544 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
545 static ConstantInt *AddOne(ConstantInt *C) {
546 return cast<ConstantInt>(ConstantExpr::getAdd(C,
547 ConstantInt::get(C->getType(), 1)));
549 static ConstantInt *SubOne(ConstantInt *C) {
550 return cast<ConstantInt>(ConstantExpr::getSub(C,
551 ConstantInt::get(C->getType(), 1)));
554 /// GetConstantInType - Return a ConstantInt with the specified type and value.
556 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
557 if (Ty->isUnsigned())
558 return ConstantUInt::get(Ty, Val);
559 else if (Ty->getTypeID() == Type::BoolTyID)
560 return ConstantBool::get(Val);
562 SVal <<= 64-Ty->getPrimitiveSizeInBits();
563 SVal >>= 64-Ty->getPrimitiveSizeInBits();
564 return ConstantSInt::get(Ty, SVal);
568 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
569 /// known to be either zero or one and return them in the KnownZero/KnownOne
570 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
572 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
573 uint64_t &KnownOne, unsigned Depth = 0) {
574 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
575 // we cannot optimize based on the assumption that it is zero without changing
576 // it to be an explicit zero. If we don't change it to zero, other code could
577 // optimized based on the contradictory assumption that it is non-zero.
578 // Because instcombine aggressively folds operations with undef args anyway,
579 // this won't lose us code quality.
580 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
581 // We know all of the bits for a constant!
582 KnownOne = CI->getZExtValue() & Mask;
583 KnownZero = ~KnownOne & Mask;
587 KnownZero = KnownOne = 0; // Don't know anything.
588 if (Depth == 6 || Mask == 0)
589 return; // Limit search depth.
591 uint64_t KnownZero2, KnownOne2;
592 Instruction *I = dyn_cast<Instruction>(V);
595 Mask &= V->getType()->getIntegralTypeMask();
597 switch (I->getOpcode()) {
598 case Instruction::And:
599 // If either the LHS or the RHS are Zero, the result is zero.
600 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
602 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
603 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
604 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
606 // Output known-1 bits are only known if set in both the LHS & RHS.
607 KnownOne &= KnownOne2;
608 // Output known-0 are known to be clear if zero in either the LHS | RHS.
609 KnownZero |= KnownZero2;
611 case Instruction::Or:
612 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
614 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
615 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
616 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
618 // Output known-0 bits are only known if clear in both the LHS & RHS.
619 KnownZero &= KnownZero2;
620 // Output known-1 are known to be set if set in either the LHS | RHS.
621 KnownOne |= KnownOne2;
623 case Instruction::Xor: {
624 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
625 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
626 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
627 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
629 // Output known-0 bits are known if clear or set in both the LHS & RHS.
630 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
631 // Output known-1 are known to be set if set in only one of the LHS, RHS.
632 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
633 KnownZero = KnownZeroOut;
636 case Instruction::Select:
637 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
638 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
639 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
640 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
642 // Only known if known in both the LHS and RHS.
643 KnownOne &= KnownOne2;
644 KnownZero &= KnownZero2;
646 case Instruction::Cast: {
647 const Type *SrcTy = I->getOperand(0)->getType();
648 if (!SrcTy->isIntegral()) return;
650 // If this is an integer truncate or noop, just look in the input.
651 if (SrcTy->getPrimitiveSizeInBits() >=
652 I->getType()->getPrimitiveSizeInBits()) {
653 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
657 // Sign or Zero extension. Compute the bits in the result that are not
658 // present in the input.
659 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
660 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
662 // Handle zero extension.
663 if (!SrcTy->isSigned()) {
664 Mask &= SrcTy->getIntegralTypeMask();
665 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
666 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
667 // The top bits are known to be zero.
668 KnownZero |= NewBits;
671 Mask &= SrcTy->getIntegralTypeMask();
672 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
673 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
675 // If the sign bit of the input is known set or clear, then we know the
676 // top bits of the result.
677 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
678 if (KnownZero & InSignBit) { // Input sign bit known zero
679 KnownZero |= NewBits;
680 KnownOne &= ~NewBits;
681 } else if (KnownOne & InSignBit) { // Input sign bit known set
683 KnownZero &= ~NewBits;
684 } else { // Input sign bit unknown
685 KnownZero &= ~NewBits;
686 KnownOne &= ~NewBits;
691 case Instruction::Shl:
692 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
693 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
694 Mask >>= SA->getValue();
695 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
696 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
697 KnownZero <<= SA->getValue();
698 KnownOne <<= SA->getValue();
699 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
703 case Instruction::Shr:
704 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
705 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
706 // Compute the new bits that are at the top now.
707 uint64_t HighBits = (1ULL << SA->getValue())-1;
708 HighBits <<= I->getType()->getPrimitiveSizeInBits()-SA->getValue();
710 if (I->getType()->isUnsigned()) { // Unsigned shift right.
711 Mask <<= SA->getValue();
712 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
713 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
714 KnownZero >>= SA->getValue();
715 KnownOne >>= SA->getValue();
716 KnownZero |= HighBits; // high bits known zero.
718 Mask <<= SA->getValue();
719 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
720 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
721 KnownZero >>= SA->getValue();
722 KnownOne >>= SA->getValue();
724 // Handle the sign bits.
725 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
726 SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
728 if (KnownZero & SignBit) { // New bits are known zero.
729 KnownZero |= HighBits;
730 } else if (KnownOne & SignBit) { // New bits are known one.
731 KnownOne |= HighBits;
740 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
741 /// this predicate to simplify operations downstream. Mask is known to be zero
742 /// for bits that V cannot have.
743 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
744 uint64_t KnownZero, KnownOne;
745 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
746 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
747 return (KnownZero & Mask) == Mask;
750 /// ShrinkDemandedConstant - Check to see if the specified operand of the
751 /// specified instruction is a constant integer. If so, check to see if there
752 /// are any bits set in the constant that are not demanded. If so, shrink the
753 /// constant and return true.
754 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
756 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
757 if (!OpC) return false;
759 // If there are no bits set that aren't demanded, nothing to do.
760 if ((~Demanded & OpC->getZExtValue()) == 0)
763 // This is producing any bits that are not needed, shrink the RHS.
764 uint64_t Val = Demanded & OpC->getZExtValue();
765 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
769 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
770 // set of known zero and one bits, compute the maximum and minimum values that
771 // could have the specified known zero and known one bits, returning them in
773 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
776 int64_t &Min, int64_t &Max) {
777 uint64_t TypeBits = Ty->getIntegralTypeMask();
778 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
780 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
782 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
783 // bit if it is unknown.
785 Max = KnownOne|UnknownBits;
787 if (SignBit & UnknownBits) { // Sign bit is unknown
792 // Sign extend the min/max values.
793 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
794 Min = (Min << ShAmt) >> ShAmt;
795 Max = (Max << ShAmt) >> ShAmt;
798 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
799 // a set of known zero and one bits, compute the maximum and minimum values that
800 // could have the specified known zero and known one bits, returning them in
802 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
807 uint64_t TypeBits = Ty->getIntegralTypeMask();
808 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
810 // The minimum value is when the unknown bits are all zeros.
812 // The maximum value is when the unknown bits are all ones.
813 Max = KnownOne|UnknownBits;
817 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
818 /// DemandedMask bits of the result of V are ever used downstream. If we can
819 /// use this information to simplify V, do so and return true. Otherwise,
820 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
821 /// the expression (used to simplify the caller). The KnownZero/One bits may
822 /// only be accurate for those bits in the DemandedMask.
823 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
824 uint64_t &KnownZero, uint64_t &KnownOne,
826 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
827 // We know all of the bits for a constant!
828 KnownOne = CI->getZExtValue() & DemandedMask;
829 KnownZero = ~KnownOne & DemandedMask;
833 KnownZero = KnownOne = 0;
834 if (!V->hasOneUse()) { // Other users may use these bits.
835 if (Depth != 0) { // Not at the root.
836 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
837 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
840 // If this is the root being simplified, allow it to have multiple uses,
841 // just set the DemandedMask to all bits.
842 DemandedMask = V->getType()->getIntegralTypeMask();
843 } else if (DemandedMask == 0) { // Not demanding any bits from V.
844 if (V != UndefValue::get(V->getType()))
845 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
847 } else if (Depth == 6) { // Limit search depth.
851 Instruction *I = dyn_cast<Instruction>(V);
852 if (!I) return false; // Only analyze instructions.
854 DemandedMask &= V->getType()->getIntegralTypeMask();
856 uint64_t KnownZero2, KnownOne2;
857 switch (I->getOpcode()) {
859 case Instruction::And:
860 // If either the LHS or the RHS are Zero, the result is zero.
861 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
862 KnownZero, KnownOne, Depth+1))
864 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
866 // If something is known zero on the RHS, the bits aren't demanded on the
868 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
869 KnownZero2, KnownOne2, Depth+1))
871 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
873 // If all of the demanded bits are known one on one side, return the other.
874 // These bits cannot contribute to the result of the 'and'.
875 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
876 return UpdateValueUsesWith(I, I->getOperand(0));
877 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
878 return UpdateValueUsesWith(I, I->getOperand(1));
880 // If all of the demanded bits in the inputs are known zeros, return zero.
881 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
882 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
884 // If the RHS is a constant, see if we can simplify it.
885 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
886 return UpdateValueUsesWith(I, I);
888 // Output known-1 bits are only known if set in both the LHS & RHS.
889 KnownOne &= KnownOne2;
890 // Output known-0 are known to be clear if zero in either the LHS | RHS.
891 KnownZero |= KnownZero2;
893 case Instruction::Or:
894 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
895 KnownZero, KnownOne, Depth+1))
897 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
898 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
899 KnownZero2, KnownOne2, Depth+1))
901 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
903 // If all of the demanded bits are known zero on one side, return the other.
904 // These bits cannot contribute to the result of the 'or'.
905 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
906 return UpdateValueUsesWith(I, I->getOperand(0));
907 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
908 return UpdateValueUsesWith(I, I->getOperand(1));
910 // If all of the potentially set bits on one side are known to be set on
911 // the other side, just use the 'other' side.
912 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
913 (DemandedMask & (~KnownZero)))
914 return UpdateValueUsesWith(I, I->getOperand(0));
915 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
916 (DemandedMask & (~KnownZero2)))
917 return UpdateValueUsesWith(I, I->getOperand(1));
919 // If the RHS is a constant, see if we can simplify it.
920 if (ShrinkDemandedConstant(I, 1, DemandedMask))
921 return UpdateValueUsesWith(I, I);
923 // Output known-0 bits are only known if clear in both the LHS & RHS.
924 KnownZero &= KnownZero2;
925 // Output known-1 are known to be set if set in either the LHS | RHS.
926 KnownOne |= KnownOne2;
928 case Instruction::Xor: {
929 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
930 KnownZero, KnownOne, Depth+1))
932 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
933 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
934 KnownZero2, KnownOne2, Depth+1))
936 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
938 // If all of the demanded bits are known zero on one side, return the other.
939 // These bits cannot contribute to the result of the 'xor'.
940 if ((DemandedMask & KnownZero) == DemandedMask)
941 return UpdateValueUsesWith(I, I->getOperand(0));
942 if ((DemandedMask & KnownZero2) == DemandedMask)
943 return UpdateValueUsesWith(I, I->getOperand(1));
945 // Output known-0 bits are known if clear or set in both the LHS & RHS.
946 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
947 // Output known-1 are known to be set if set in only one of the LHS, RHS.
948 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
950 // If all of the unknown bits are known to be zero on one side or the other
951 // (but not both) turn this into an *inclusive* or.
952 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
953 if (uint64_t UnknownBits = DemandedMask & ~(KnownZeroOut|KnownOneOut)) {
954 if ((UnknownBits & (KnownZero|KnownZero2)) == UnknownBits) {
956 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
958 InsertNewInstBefore(Or, *I);
959 return UpdateValueUsesWith(I, Or);
963 // If all of the demanded bits on one side are known, and all of the set
964 // bits on that side are also known to be set on the other side, turn this
965 // into an AND, as we know the bits will be cleared.
966 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
967 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
968 if ((KnownOne & KnownOne2) == KnownOne) {
969 Constant *AndC = GetConstantInType(I->getType(),
970 ~KnownOne & DemandedMask);
972 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
973 InsertNewInstBefore(And, *I);
974 return UpdateValueUsesWith(I, And);
978 // If the RHS is a constant, see if we can simplify it.
979 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
980 if (ShrinkDemandedConstant(I, 1, DemandedMask))
981 return UpdateValueUsesWith(I, I);
983 KnownZero = KnownZeroOut;
984 KnownOne = KnownOneOut;
987 case Instruction::Select:
988 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
989 KnownZero, KnownOne, Depth+1))
991 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
992 KnownZero2, KnownOne2, Depth+1))
994 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
995 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
997 // If the operands are constants, see if we can simplify them.
998 if (ShrinkDemandedConstant(I, 1, DemandedMask))
999 return UpdateValueUsesWith(I, I);
1000 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1001 return UpdateValueUsesWith(I, I);
1003 // Only known if known in both the LHS and RHS.
1004 KnownOne &= KnownOne2;
1005 KnownZero &= KnownZero2;
1007 case Instruction::Cast: {
1008 const Type *SrcTy = I->getOperand(0)->getType();
1009 if (!SrcTy->isIntegral()) return false;
1011 // If this is an integer truncate or noop, just look in the input.
1012 if (SrcTy->getPrimitiveSizeInBits() >=
1013 I->getType()->getPrimitiveSizeInBits()) {
1014 // Cast to bool is a comparison against 0, which demands all bits. We
1015 // can't propagate anything useful up.
1016 if (I->getType() == Type::BoolTy)
1019 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1020 KnownZero, KnownOne, Depth+1))
1022 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1026 // Sign or Zero extension. Compute the bits in the result that are not
1027 // present in the input.
1028 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1029 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1031 // Handle zero extension.
1032 if (!SrcTy->isSigned()) {
1033 DemandedMask &= SrcTy->getIntegralTypeMask();
1034 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1035 KnownZero, KnownOne, Depth+1))
1037 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1038 // The top bits are known to be zero.
1039 KnownZero |= NewBits;
1042 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1043 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1045 // If any of the sign extended bits are demanded, we know that the sign
1047 if (NewBits & DemandedMask)
1048 InputDemandedBits |= InSignBit;
1050 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1051 KnownZero, KnownOne, Depth+1))
1053 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1055 // If the sign bit of the input is known set or clear, then we know the
1056 // top bits of the result.
1058 // If the input sign bit is known zero, or if the NewBits are not demanded
1059 // convert this into a zero extension.
1060 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1061 // Convert to unsigned first.
1062 Instruction *NewVal;
1063 NewVal = new CastInst(I->getOperand(0), SrcTy->getUnsignedVersion(),
1064 I->getOperand(0)->getName());
1065 InsertNewInstBefore(NewVal, *I);
1066 // Then cast that to the destination type.
1067 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1068 InsertNewInstBefore(NewVal, *I);
1069 return UpdateValueUsesWith(I, NewVal);
1070 } else if (KnownOne & InSignBit) { // Input sign bit known set
1071 KnownOne |= NewBits;
1072 KnownZero &= ~NewBits;
1073 } else { // Input sign bit unknown
1074 KnownZero &= ~NewBits;
1075 KnownOne &= ~NewBits;
1080 case Instruction::Shl:
1081 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
1082 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> SA->getValue(),
1083 KnownZero, KnownOne, Depth+1))
1085 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1086 KnownZero <<= SA->getValue();
1087 KnownOne <<= SA->getValue();
1088 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
1091 case Instruction::Shr:
1092 // If this is an arithmetic shift right and only the low-bit is set, we can
1093 // always convert this into a logical shr, even if the shift amount is
1094 // variable. The low bit of the shift cannot be an input sign bit unless
1095 // the shift amount is >= the size of the datatype, which is undefined.
1096 if (DemandedMask == 1 && I->getType()->isSigned()) {
1097 // Convert the input to unsigned.
1098 Instruction *NewVal = new CastInst(I->getOperand(0),
1099 I->getType()->getUnsignedVersion(),
1100 I->getOperand(0)->getName());
1101 InsertNewInstBefore(NewVal, *I);
1102 // Perform the unsigned shift right.
1103 NewVal = new ShiftInst(Instruction::Shr, NewVal, I->getOperand(1),
1105 InsertNewInstBefore(NewVal, *I);
1106 // Then cast that to the destination type.
1107 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1108 InsertNewInstBefore(NewVal, *I);
1109 return UpdateValueUsesWith(I, NewVal);
1112 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
1113 unsigned ShAmt = SA->getValue();
1115 // Compute the new bits that are at the top now.
1116 uint64_t HighBits = (1ULL << ShAmt)-1;
1117 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShAmt;
1118 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1119 if (I->getType()->isUnsigned()) { // Unsigned shift right.
1120 if (SimplifyDemandedBits(I->getOperand(0),
1121 (DemandedMask << ShAmt) & TypeMask,
1122 KnownZero, KnownOne, Depth+1))
1124 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1125 KnownZero &= TypeMask;
1126 KnownOne &= TypeMask;
1127 KnownZero >>= ShAmt;
1129 KnownZero |= HighBits; // high bits known zero.
1130 } else { // Signed shift right.
1131 if (SimplifyDemandedBits(I->getOperand(0),
1132 (DemandedMask << ShAmt) & TypeMask,
1133 KnownZero, KnownOne, Depth+1))
1135 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1136 KnownZero &= TypeMask;
1137 KnownOne &= TypeMask;
1138 KnownZero >>= SA->getValue();
1139 KnownOne >>= SA->getValue();
1141 // Handle the sign bits.
1142 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1143 SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
1145 // If the input sign bit is known to be zero, or if none of the top bits
1146 // are demanded, turn this into an unsigned shift right.
1147 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1148 // Convert the input to unsigned.
1149 Instruction *NewVal;
1150 NewVal = new CastInst(I->getOperand(0),
1151 I->getType()->getUnsignedVersion(),
1152 I->getOperand(0)->getName());
1153 InsertNewInstBefore(NewVal, *I);
1154 // Perform the unsigned shift right.
1155 NewVal = new ShiftInst(Instruction::Shr, NewVal, SA, I->getName());
1156 InsertNewInstBefore(NewVal, *I);
1157 // Then cast that to the destination type.
1158 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1159 InsertNewInstBefore(NewVal, *I);
1160 return UpdateValueUsesWith(I, NewVal);
1161 } else if (KnownOne & SignBit) { // New bits are known one.
1162 KnownOne |= HighBits;
1169 // If the client is only demanding bits that we know, return the known
1171 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1172 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1176 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
1177 // true when both operands are equal...
1179 static bool isTrueWhenEqual(Instruction &I) {
1180 return I.getOpcode() == Instruction::SetEQ ||
1181 I.getOpcode() == Instruction::SetGE ||
1182 I.getOpcode() == Instruction::SetLE;
1185 /// AssociativeOpt - Perform an optimization on an associative operator. This
1186 /// function is designed to check a chain of associative operators for a
1187 /// potential to apply a certain optimization. Since the optimization may be
1188 /// applicable if the expression was reassociated, this checks the chain, then
1189 /// reassociates the expression as necessary to expose the optimization
1190 /// opportunity. This makes use of a special Functor, which must define
1191 /// 'shouldApply' and 'apply' methods.
1193 template<typename Functor>
1194 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1195 unsigned Opcode = Root.getOpcode();
1196 Value *LHS = Root.getOperand(0);
1198 // Quick check, see if the immediate LHS matches...
1199 if (F.shouldApply(LHS))
1200 return F.apply(Root);
1202 // Otherwise, if the LHS is not of the same opcode as the root, return.
1203 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1204 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1205 // Should we apply this transform to the RHS?
1206 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1208 // If not to the RHS, check to see if we should apply to the LHS...
1209 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1210 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1214 // If the functor wants to apply the optimization to the RHS of LHSI,
1215 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1217 BasicBlock *BB = Root.getParent();
1219 // Now all of the instructions are in the current basic block, go ahead
1220 // and perform the reassociation.
1221 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1223 // First move the selected RHS to the LHS of the root...
1224 Root.setOperand(0, LHSI->getOperand(1));
1226 // Make what used to be the LHS of the root be the user of the root...
1227 Value *ExtraOperand = TmpLHSI->getOperand(1);
1228 if (&Root == TmpLHSI) {
1229 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1232 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1233 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1234 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1235 BasicBlock::iterator ARI = &Root; ++ARI;
1236 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1239 // Now propagate the ExtraOperand down the chain of instructions until we
1241 while (TmpLHSI != LHSI) {
1242 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1243 // Move the instruction to immediately before the chain we are
1244 // constructing to avoid breaking dominance properties.
1245 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1246 BB->getInstList().insert(ARI, NextLHSI);
1249 Value *NextOp = NextLHSI->getOperand(1);
1250 NextLHSI->setOperand(1, ExtraOperand);
1252 ExtraOperand = NextOp;
1255 // Now that the instructions are reassociated, have the functor perform
1256 // the transformation...
1257 return F.apply(Root);
1260 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1266 // AddRHS - Implements: X + X --> X << 1
1269 AddRHS(Value *rhs) : RHS(rhs) {}
1270 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1271 Instruction *apply(BinaryOperator &Add) const {
1272 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1273 ConstantInt::get(Type::UByteTy, 1));
1277 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1279 struct AddMaskingAnd {
1281 AddMaskingAnd(Constant *c) : C2(c) {}
1282 bool shouldApply(Value *LHS) const {
1284 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1285 ConstantExpr::getAnd(C1, C2)->isNullValue();
1287 Instruction *apply(BinaryOperator &Add) const {
1288 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1292 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1294 if (isa<CastInst>(I)) {
1295 if (Constant *SOC = dyn_cast<Constant>(SO))
1296 return ConstantExpr::getCast(SOC, I.getType());
1298 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
1299 SO->getName() + ".cast"), I);
1302 // Figure out if the constant is the left or the right argument.
1303 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1304 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1306 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1308 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1309 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1312 Value *Op0 = SO, *Op1 = ConstOperand;
1314 std::swap(Op0, Op1);
1316 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1317 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1318 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1319 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1321 assert(0 && "Unknown binary instruction type!");
1324 return IC->InsertNewInstBefore(New, I);
1327 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1328 // constant as the other operand, try to fold the binary operator into the
1329 // select arguments. This also works for Cast instructions, which obviously do
1330 // not have a second operand.
1331 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1333 // Don't modify shared select instructions
1334 if (!SI->hasOneUse()) return 0;
1335 Value *TV = SI->getOperand(1);
1336 Value *FV = SI->getOperand(2);
1338 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1339 // Bool selects with constant operands can be folded to logical ops.
1340 if (SI->getType() == Type::BoolTy) return 0;
1342 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1343 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1345 return new SelectInst(SI->getCondition(), SelectTrueVal,
1352 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1353 /// node as operand #0, see if we can fold the instruction into the PHI (which
1354 /// is only possible if all operands to the PHI are constants).
1355 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1356 PHINode *PN = cast<PHINode>(I.getOperand(0));
1357 unsigned NumPHIValues = PN->getNumIncomingValues();
1358 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1360 // Check to see if all of the operands of the PHI are constants. If there is
1361 // one non-constant value, remember the BB it is. If there is more than one
1363 BasicBlock *NonConstBB = 0;
1364 for (unsigned i = 0; i != NumPHIValues; ++i)
1365 if (!isa<Constant>(PN->getIncomingValue(i))) {
1366 if (NonConstBB) return 0; // More than one non-const value.
1367 NonConstBB = PN->getIncomingBlock(i);
1369 // If the incoming non-constant value is in I's block, we have an infinite
1371 if (NonConstBB == I.getParent())
1375 // If there is exactly one non-constant value, we can insert a copy of the
1376 // operation in that block. However, if this is a critical edge, we would be
1377 // inserting the computation one some other paths (e.g. inside a loop). Only
1378 // do this if the pred block is unconditionally branching into the phi block.
1380 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1381 if (!BI || !BI->isUnconditional()) return 0;
1384 // Okay, we can do the transformation: create the new PHI node.
1385 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1387 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1388 InsertNewInstBefore(NewPN, *PN);
1390 // Next, add all of the operands to the PHI.
1391 if (I.getNumOperands() == 2) {
1392 Constant *C = cast<Constant>(I.getOperand(1));
1393 for (unsigned i = 0; i != NumPHIValues; ++i) {
1395 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1396 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1398 assert(PN->getIncomingBlock(i) == NonConstBB);
1399 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1400 InV = BinaryOperator::create(BO->getOpcode(),
1401 PN->getIncomingValue(i), C, "phitmp",
1402 NonConstBB->getTerminator());
1403 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1404 InV = new ShiftInst(SI->getOpcode(),
1405 PN->getIncomingValue(i), C, "phitmp",
1406 NonConstBB->getTerminator());
1408 assert(0 && "Unknown binop!");
1410 WorkList.push_back(cast<Instruction>(InV));
1412 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1415 assert(isa<CastInst>(I) && "Unary op should be a cast!");
1416 const Type *RetTy = I.getType();
1417 for (unsigned i = 0; i != NumPHIValues; ++i) {
1419 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1420 InV = ConstantExpr::getCast(InC, RetTy);
1422 assert(PN->getIncomingBlock(i) == NonConstBB);
1423 InV = new CastInst(PN->getIncomingValue(i), I.getType(), "phitmp",
1424 NonConstBB->getTerminator());
1425 WorkList.push_back(cast<Instruction>(InV));
1427 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1430 return ReplaceInstUsesWith(I, NewPN);
1433 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1434 bool Changed = SimplifyCommutative(I);
1435 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1437 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1438 // X + undef -> undef
1439 if (isa<UndefValue>(RHS))
1440 return ReplaceInstUsesWith(I, RHS);
1443 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
1444 if (RHSC->isNullValue())
1445 return ReplaceInstUsesWith(I, LHS);
1446 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1447 if (CFP->isExactlyValue(-0.0))
1448 return ReplaceInstUsesWith(I, LHS);
1451 // X + (signbit) --> X ^ signbit
1452 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1453 uint64_t Val = CI->getZExtValue();
1454 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1455 return BinaryOperator::createXor(LHS, RHS);
1458 if (isa<PHINode>(LHS))
1459 if (Instruction *NV = FoldOpIntoPhi(I))
1462 ConstantInt *XorRHS = 0;
1464 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1465 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1466 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1467 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1469 uint64_t C0080Val = 1ULL << 31;
1470 int64_t CFF80Val = -C0080Val;
1473 if (TySizeBits > Size) {
1475 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1476 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1477 if (RHSSExt == CFF80Val) {
1478 if (XorRHS->getZExtValue() == C0080Val)
1480 } else if (RHSZExt == C0080Val) {
1481 if (XorRHS->getSExtValue() == CFF80Val)
1485 // This is a sign extend if the top bits are known zero.
1486 uint64_t Mask = ~0ULL;
1487 Mask <<= 64-(TySizeBits-Size);
1488 Mask &= XorLHS->getType()->getIntegralTypeMask();
1489 if (!MaskedValueIsZero(XorLHS, Mask))
1490 Size = 0; // Not a sign ext, but can't be any others either.
1497 } while (Size >= 8);
1500 const Type *MiddleType = 0;
1503 case 32: MiddleType = Type::IntTy; break;
1504 case 16: MiddleType = Type::ShortTy; break;
1505 case 8: MiddleType = Type::SByteTy; break;
1508 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
1509 InsertNewInstBefore(NewTrunc, I);
1510 return new CastInst(NewTrunc, I.getType());
1516 if (I.getType()->isInteger()) {
1517 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1519 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1520 if (RHSI->getOpcode() == Instruction::Sub)
1521 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1522 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1524 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1525 if (LHSI->getOpcode() == Instruction::Sub)
1526 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1527 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1532 if (Value *V = dyn_castNegVal(LHS))
1533 return BinaryOperator::createSub(RHS, V);
1536 if (!isa<Constant>(RHS))
1537 if (Value *V = dyn_castNegVal(RHS))
1538 return BinaryOperator::createSub(LHS, V);
1542 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1543 if (X == RHS) // X*C + X --> X * (C+1)
1544 return BinaryOperator::createMul(RHS, AddOne(C2));
1546 // X*C1 + X*C2 --> X * (C1+C2)
1548 if (X == dyn_castFoldableMul(RHS, C1))
1549 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1552 // X + X*C --> X * (C+1)
1553 if (dyn_castFoldableMul(RHS, C2) == LHS)
1554 return BinaryOperator::createMul(LHS, AddOne(C2));
1557 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1558 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1559 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1561 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1563 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1564 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1565 return BinaryOperator::createSub(C, X);
1568 // (X & FF00) + xx00 -> (X+xx00) & FF00
1569 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1570 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1571 if (Anded == CRHS) {
1572 // See if all bits from the first bit set in the Add RHS up are included
1573 // in the mask. First, get the rightmost bit.
1574 uint64_t AddRHSV = CRHS->getRawValue();
1576 // Form a mask of all bits from the lowest bit added through the top.
1577 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1578 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1580 // See if the and mask includes all of these bits.
1581 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
1583 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1584 // Okay, the xform is safe. Insert the new add pronto.
1585 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1586 LHS->getName()), I);
1587 return BinaryOperator::createAnd(NewAdd, C2);
1592 // Try to fold constant add into select arguments.
1593 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1594 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1598 // add (cast *A to intptrtype) B -> cast (GEP (cast *A to sbyte*) B) -> intptrtype
1600 CastInst* CI = dyn_cast<CastInst>(LHS);
1603 CI = dyn_cast<CastInst>(RHS);
1606 if (CI && CI->getType()->isSized() &&
1607 (CI->getType()->getPrimitiveSize() ==
1608 TD->getIntPtrType()->getPrimitiveSize())
1609 && isa<PointerType>(CI->getOperand(0)->getType())) {
1610 Value* I2 = InsertCastBefore(CI->getOperand(0),
1611 PointerType::get(Type::SByteTy), I);
1612 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1613 return new CastInst(I2, CI->getType());
1617 return Changed ? &I : 0;
1620 // isSignBit - Return true if the value represented by the constant only has the
1621 // highest order bit set.
1622 static bool isSignBit(ConstantInt *CI) {
1623 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1624 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1627 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1629 static Value *RemoveNoopCast(Value *V) {
1630 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1631 const Type *CTy = CI->getType();
1632 const Type *OpTy = CI->getOperand(0)->getType();
1633 if (CTy->isInteger() && OpTy->isInteger()) {
1634 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1635 return RemoveNoopCast(CI->getOperand(0));
1636 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1637 return RemoveNoopCast(CI->getOperand(0));
1642 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1643 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1645 if (Op0 == Op1) // sub X, X -> 0
1646 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1648 // If this is a 'B = x-(-A)', change to B = x+A...
1649 if (Value *V = dyn_castNegVal(Op1))
1650 return BinaryOperator::createAdd(Op0, V);
1652 if (isa<UndefValue>(Op0))
1653 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1654 if (isa<UndefValue>(Op1))
1655 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1657 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1658 // Replace (-1 - A) with (~A)...
1659 if (C->isAllOnesValue())
1660 return BinaryOperator::createNot(Op1);
1662 // C - ~X == X + (1+C)
1664 if (match(Op1, m_Not(m_Value(X))))
1665 return BinaryOperator::createAdd(X,
1666 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1667 // -((uint)X >> 31) -> ((int)X >> 31)
1668 // -((int)X >> 31) -> ((uint)X >> 31)
1669 if (C->isNullValue()) {
1670 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1671 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1672 if (SI->getOpcode() == Instruction::Shr)
1673 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
1675 if (SI->getType()->isSigned())
1676 NewTy = SI->getType()->getUnsignedVersion();
1678 NewTy = SI->getType()->getSignedVersion();
1679 // Check to see if we are shifting out everything but the sign bit.
1680 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
1681 // Ok, the transformation is safe. Insert a cast of the incoming
1682 // value, then the new shift, then the new cast.
1683 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
1684 SI->getOperand(0)->getName());
1685 Value *InV = InsertNewInstBefore(FirstCast, I);
1686 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
1688 if (NewShift->getType() == I.getType())
1691 InV = InsertNewInstBefore(NewShift, I);
1692 return new CastInst(NewShift, I.getType());
1698 // Try to fold constant sub into select arguments.
1699 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1700 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1703 if (isa<PHINode>(Op0))
1704 if (Instruction *NV = FoldOpIntoPhi(I))
1708 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1709 if (Op1I->getOpcode() == Instruction::Add &&
1710 !Op0->getType()->isFloatingPoint()) {
1711 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1712 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1713 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1714 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1715 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1716 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1717 // C1-(X+C2) --> (C1-C2)-X
1718 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1719 Op1I->getOperand(0));
1723 if (Op1I->hasOneUse()) {
1724 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1725 // is not used by anyone else...
1727 if (Op1I->getOpcode() == Instruction::Sub &&
1728 !Op1I->getType()->isFloatingPoint()) {
1729 // Swap the two operands of the subexpr...
1730 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1731 Op1I->setOperand(0, IIOp1);
1732 Op1I->setOperand(1, IIOp0);
1734 // Create the new top level add instruction...
1735 return BinaryOperator::createAdd(Op0, Op1);
1738 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1740 if (Op1I->getOpcode() == Instruction::And &&
1741 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1742 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1745 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1746 return BinaryOperator::createAnd(Op0, NewNot);
1749 // -(X sdiv C) -> (X sdiv -C)
1750 if (Op1I->getOpcode() == Instruction::Div)
1751 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
1752 if (CSI->isNullValue())
1753 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1754 return BinaryOperator::createDiv(Op1I->getOperand(0),
1755 ConstantExpr::getNeg(DivRHS));
1757 // X - X*C --> X * (1-C)
1758 ConstantInt *C2 = 0;
1759 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1761 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1762 return BinaryOperator::createMul(Op0, CP1);
1767 if (!Op0->getType()->isFloatingPoint())
1768 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1769 if (Op0I->getOpcode() == Instruction::Add) {
1770 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1771 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1772 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1773 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1774 } else if (Op0I->getOpcode() == Instruction::Sub) {
1775 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1776 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1780 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1781 if (X == Op1) { // X*C - X --> X * (C-1)
1782 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1783 return BinaryOperator::createMul(Op1, CP1);
1786 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1787 if (X == dyn_castFoldableMul(Op1, C2))
1788 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1793 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1794 /// really just returns true if the most significant (sign) bit is set.
1795 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1796 if (RHS->getType()->isSigned()) {
1797 // True if source is LHS < 0 or LHS <= -1
1798 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1799 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1801 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1802 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1803 // the size of the integer type.
1804 if (Opcode == Instruction::SetGE)
1805 return RHSC->getValue() ==
1806 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1807 if (Opcode == Instruction::SetGT)
1808 return RHSC->getValue() ==
1809 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1814 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1815 bool Changed = SimplifyCommutative(I);
1816 Value *Op0 = I.getOperand(0);
1818 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1819 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1821 // Simplify mul instructions with a constant RHS...
1822 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1823 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1825 // ((X << C1)*C2) == (X * (C2 << C1))
1826 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1827 if (SI->getOpcode() == Instruction::Shl)
1828 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1829 return BinaryOperator::createMul(SI->getOperand(0),
1830 ConstantExpr::getShl(CI, ShOp));
1832 if (CI->isNullValue())
1833 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1834 if (CI->equalsInt(1)) // X * 1 == X
1835 return ReplaceInstUsesWith(I, Op0);
1836 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1837 return BinaryOperator::createNeg(Op0, I.getName());
1839 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1840 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1841 uint64_t C = Log2_64(Val);
1842 return new ShiftInst(Instruction::Shl, Op0,
1843 ConstantUInt::get(Type::UByteTy, C));
1845 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1846 if (Op1F->isNullValue())
1847 return ReplaceInstUsesWith(I, Op1);
1849 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1850 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1851 if (Op1F->getValue() == 1.0)
1852 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1855 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1856 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
1857 isa<ConstantInt>(Op0I->getOperand(1))) {
1858 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
1859 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
1861 InsertNewInstBefore(Add, I);
1862 Value *C1C2 = ConstantExpr::getMul(Op1,
1863 cast<Constant>(Op0I->getOperand(1)));
1864 return BinaryOperator::createAdd(Add, C1C2);
1868 // Try to fold constant mul into select arguments.
1869 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1870 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1873 if (isa<PHINode>(Op0))
1874 if (Instruction *NV = FoldOpIntoPhi(I))
1878 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1879 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1880 return BinaryOperator::createMul(Op0v, Op1v);
1882 // If one of the operands of the multiply is a cast from a boolean value, then
1883 // we know the bool is either zero or one, so this is a 'masking' multiply.
1884 // See if we can simplify things based on how the boolean was originally
1886 CastInst *BoolCast = 0;
1887 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1888 if (CI->getOperand(0)->getType() == Type::BoolTy)
1891 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1892 if (CI->getOperand(0)->getType() == Type::BoolTy)
1895 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1896 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1897 const Type *SCOpTy = SCIOp0->getType();
1899 // If the setcc is true iff the sign bit of X is set, then convert this
1900 // multiply into a shift/and combination.
1901 if (isa<ConstantInt>(SCIOp1) &&
1902 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1903 // Shift the X value right to turn it into "all signbits".
1904 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1905 SCOpTy->getPrimitiveSizeInBits()-1);
1906 if (SCIOp0->getType()->isUnsigned()) {
1907 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1908 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1909 SCIOp0->getName()), I);
1913 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1914 BoolCast->getOperand(0)->getName()+
1917 // If the multiply type is not the same as the source type, sign extend
1918 // or truncate to the multiply type.
1919 if (I.getType() != V->getType())
1920 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1922 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1923 return BinaryOperator::createAnd(V, OtherOp);
1928 return Changed ? &I : 0;
1931 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1932 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1934 if (isa<UndefValue>(Op0)) // undef / X -> 0
1935 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1936 if (isa<UndefValue>(Op1))
1937 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1939 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1941 if (RHS->equalsInt(1))
1942 return ReplaceInstUsesWith(I, Op0);
1945 if (RHS->isAllOnesValue())
1946 return BinaryOperator::createNeg(Op0);
1948 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1949 if (LHS->getOpcode() == Instruction::Div)
1950 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1951 // (X / C1) / C2 -> X / (C1*C2)
1952 return BinaryOperator::createDiv(LHS->getOperand(0),
1953 ConstantExpr::getMul(RHS, LHSRHS));
1956 // Check to see if this is an unsigned division with an exact power of 2,
1957 // if so, convert to a right shift.
1958 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1959 if (uint64_t Val = C->getValue()) // Don't break X / 0
1960 if (isPowerOf2_64(Val)) {
1961 uint64_t C = Log2_64(Val);
1962 return new ShiftInst(Instruction::Shr, Op0,
1963 ConstantUInt::get(Type::UByteTy, C));
1967 if (RHS->getType()->isSigned())
1968 if (Value *LHSNeg = dyn_castNegVal(Op0))
1969 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1971 if (!RHS->isNullValue()) {
1972 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1973 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1975 if (isa<PHINode>(Op0))
1976 if (Instruction *NV = FoldOpIntoPhi(I))
1981 // Handle div X, Cond?Y:Z
1982 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
1983 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
1984 // same basic block, then we replace the select with Y, and the condition of
1985 // the select with false (if the cond value is in the same BB). If the
1986 // select has uses other than the div, this allows them to be simplified
1988 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
1989 if (ST->isNullValue()) {
1990 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
1991 if (CondI && CondI->getParent() == I.getParent())
1992 UpdateValueUsesWith(CondI, ConstantBool::False);
1993 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
1994 I.setOperand(1, SI->getOperand(2));
1996 UpdateValueUsesWith(SI, SI->getOperand(2));
1999 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2000 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2001 if (ST->isNullValue()) {
2002 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2003 if (CondI && CondI->getParent() == I.getParent())
2004 UpdateValueUsesWith(CondI, ConstantBool::True);
2005 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2006 I.setOperand(1, SI->getOperand(1));
2008 UpdateValueUsesWith(SI, SI->getOperand(1));
2012 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
2013 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
2014 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
2015 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
2016 // STO == 0 and SFO == 0 handled above.
2017 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
2018 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2019 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2020 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
2021 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
2022 TC, SI->getName()+".t");
2023 TSI = InsertNewInstBefore(TSI, I);
2025 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
2026 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
2027 FC, SI->getName()+".f");
2028 FSI = InsertNewInstBefore(FSI, I);
2029 return new SelectInst(SI->getOperand(0), TSI, FSI);
2034 // 0 / X == 0, we don't need to preserve faults!
2035 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2036 if (LHS->equalsInt(0))
2037 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2039 if (I.getType()->isSigned()) {
2040 // If the sign bits of both operands are zero (i.e. we can prove they are
2041 // unsigned inputs), turn this into a udiv.
2042 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2043 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2044 const Type *NTy = Op0->getType()->getUnsignedVersion();
2045 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
2046 InsertNewInstBefore(LHS, I);
2048 if (Constant *R = dyn_cast<Constant>(Op1))
2049 RHS = ConstantExpr::getCast(R, NTy);
2051 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
2052 Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
2053 InsertNewInstBefore(Div, I);
2054 return new CastInst(Div, I.getType());
2057 // Known to be an unsigned division.
2058 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2059 // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
2060 if (RHSI->getOpcode() == Instruction::Shl &&
2061 isa<ConstantUInt>(RHSI->getOperand(0))) {
2062 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
2063 if (isPowerOf2_64(C1)) {
2064 unsigned C2 = Log2_64(C1);
2065 Value *Add = RHSI->getOperand(1);
2067 Constant *C2V = ConstantUInt::get(Add->getType(), C2);
2068 Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
2071 return new ShiftInst(Instruction::Shr, Op0, Add);
2081 /// GetFactor - If we can prove that the specified value is at least a multiple
2082 /// of some factor, return that factor.
2083 static Constant *GetFactor(Value *V) {
2084 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2087 // Unless we can be tricky, we know this is a multiple of 1.
2088 Constant *Result = ConstantInt::get(V->getType(), 1);
2090 Instruction *I = dyn_cast<Instruction>(V);
2091 if (!I) return Result;
2093 if (I->getOpcode() == Instruction::Mul) {
2094 // Handle multiplies by a constant, etc.
2095 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2096 GetFactor(I->getOperand(1)));
2097 } else if (I->getOpcode() == Instruction::Shl) {
2098 // (X<<C) -> X * (1 << C)
2099 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2100 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2101 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2103 } else if (I->getOpcode() == Instruction::And) {
2104 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2105 // X & 0xFFF0 is known to be a multiple of 16.
2106 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2107 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2108 return ConstantExpr::getShl(Result,
2109 ConstantUInt::get(Type::UByteTy, Zeros));
2111 } else if (I->getOpcode() == Instruction::Cast) {
2112 Value *Op = I->getOperand(0);
2113 // Only handle int->int casts.
2114 if (!Op->getType()->isInteger()) return Result;
2115 return ConstantExpr::getCast(GetFactor(Op), V->getType());
2120 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
2121 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2123 // 0 % X == 0, we don't need to preserve faults!
2124 if (Constant *LHS = dyn_cast<Constant>(Op0))
2125 if (LHS->isNullValue())
2126 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2128 if (isa<UndefValue>(Op0)) // undef % X -> 0
2129 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2130 if (isa<UndefValue>(Op1))
2131 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2133 if (I.getType()->isSigned()) {
2134 if (Value *RHSNeg = dyn_castNegVal(Op1))
2135 if (!isa<ConstantSInt>(RHSNeg) ||
2136 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
2138 AddUsesToWorkList(I);
2139 I.setOperand(1, RHSNeg);
2143 // If the top bits of both operands are zero (i.e. we can prove they are
2144 // unsigned inputs), turn this into a urem.
2145 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2146 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2147 const Type *NTy = Op0->getType()->getUnsignedVersion();
2148 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
2149 InsertNewInstBefore(LHS, I);
2151 if (Constant *R = dyn_cast<Constant>(Op1))
2152 RHS = ConstantExpr::getCast(R, NTy);
2154 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
2155 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
2156 InsertNewInstBefore(Rem, I);
2157 return new CastInst(Rem, I.getType());
2161 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2162 // X % 0 == undef, we don't need to preserve faults!
2163 if (RHS->equalsInt(0))
2164 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2166 if (RHS->equalsInt(1)) // X % 1 == 0
2167 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2169 // Check to see if this is an unsigned remainder with an exact power of 2,
2170 // if so, convert to a bitwise and.
2171 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
2172 if (isPowerOf2_64(C->getValue()))
2173 return BinaryOperator::createAnd(Op0, SubOne(C));
2175 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2176 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2177 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2179 } else if (isa<PHINode>(Op0I)) {
2180 if (Instruction *NV = FoldOpIntoPhi(I))
2184 // X*C1%C2 --> 0 iff C1%C2 == 0
2185 if (ConstantExpr::getRem(GetFactor(Op0I), RHS)->isNullValue())
2186 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2190 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2191 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
2192 if (I.getType()->isUnsigned() &&
2193 RHSI->getOpcode() == Instruction::Shl &&
2194 isa<ConstantUInt>(RHSI->getOperand(0))) {
2195 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
2196 if (isPowerOf2_64(C1)) {
2197 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2198 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2200 return BinaryOperator::createAnd(Op0, Add);
2204 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
2205 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
2206 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2207 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2208 // the same basic block, then we replace the select with Y, and the
2209 // condition of the select with false (if the cond value is in the same
2210 // BB). If the select has uses other than the div, this allows them to be
2212 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2213 if (ST->isNullValue()) {
2214 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2215 if (CondI && CondI->getParent() == I.getParent())
2216 UpdateValueUsesWith(CondI, ConstantBool::False);
2217 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2218 I.setOperand(1, SI->getOperand(2));
2220 UpdateValueUsesWith(SI, SI->getOperand(2));
2223 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2224 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2225 if (ST->isNullValue()) {
2226 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2227 if (CondI && CondI->getParent() == I.getParent())
2228 UpdateValueUsesWith(CondI, ConstantBool::True);
2229 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2230 I.setOperand(1, SI->getOperand(1));
2232 UpdateValueUsesWith(SI, SI->getOperand(1));
2237 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
2238 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
2239 // STO == 0 and SFO == 0 handled above.
2241 if (isPowerOf2_64(STO->getValue()) && isPowerOf2_64(SFO->getValue())){
2242 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
2243 SubOne(STO), SI->getName()+".t"), I);
2244 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
2245 SubOne(SFO), SI->getName()+".f"), I);
2246 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2255 // isMaxValueMinusOne - return true if this is Max-1
2256 static bool isMaxValueMinusOne(const ConstantInt *C) {
2257 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
2258 return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
2260 const ConstantSInt *CS = cast<ConstantSInt>(C);
2262 // Calculate 0111111111..11111
2263 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2264 int64_t Val = INT64_MAX; // All ones
2265 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2266 return CS->getValue() == Val-1;
2269 // isMinValuePlusOne - return true if this is Min+1
2270 static bool isMinValuePlusOne(const ConstantInt *C) {
2271 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
2272 return CU->getValue() == 1;
2274 const ConstantSInt *CS = cast<ConstantSInt>(C);
2276 // Calculate 1111111111000000000000
2277 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2278 int64_t Val = -1; // All ones
2279 Val <<= TypeBits-1; // Shift over to the right spot
2280 return CS->getValue() == Val+1;
2283 // isOneBitSet - Return true if there is exactly one bit set in the specified
2285 static bool isOneBitSet(const ConstantInt *CI) {
2286 uint64_t V = CI->getRawValue();
2287 return V && (V & (V-1)) == 0;
2290 #if 0 // Currently unused
2291 // isLowOnes - Return true if the constant is of the form 0+1+.
2292 static bool isLowOnes(const ConstantInt *CI) {
2293 uint64_t V = CI->getRawValue();
2295 // There won't be bits set in parts that the type doesn't contain.
2296 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2298 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2299 return U && V && (U & V) == 0;
2303 // isHighOnes - Return true if the constant is of the form 1+0+.
2304 // This is the same as lowones(~X).
2305 static bool isHighOnes(const ConstantInt *CI) {
2306 uint64_t V = ~CI->getRawValue();
2307 if (~V == 0) return false; // 0's does not match "1+"
2309 // There won't be bits set in parts that the type doesn't contain.
2310 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2312 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2313 return U && V && (U & V) == 0;
2317 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2318 /// are carefully arranged to allow folding of expressions such as:
2320 /// (A < B) | (A > B) --> (A != B)
2322 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2323 /// represents that the comparison is true if A == B, and bit value '1' is true
2326 static unsigned getSetCondCode(const SetCondInst *SCI) {
2327 switch (SCI->getOpcode()) {
2329 case Instruction::SetGT: return 1;
2330 case Instruction::SetEQ: return 2;
2331 case Instruction::SetGE: return 3;
2332 case Instruction::SetLT: return 4;
2333 case Instruction::SetNE: return 5;
2334 case Instruction::SetLE: return 6;
2337 assert(0 && "Invalid SetCC opcode!");
2342 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2343 /// opcode and two operands into either a constant true or false, or a brand new
2344 /// SetCC instruction.
2345 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2347 case 0: return ConstantBool::False;
2348 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2349 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2350 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2351 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2352 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2353 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2354 case 7: return ConstantBool::True;
2355 default: assert(0 && "Illegal SetCCCode!"); return 0;
2359 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2360 struct FoldSetCCLogical {
2363 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2364 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2365 bool shouldApply(Value *V) const {
2366 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2367 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2368 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2371 Instruction *apply(BinaryOperator &Log) const {
2372 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2373 if (SCI->getOperand(0) != LHS) {
2374 assert(SCI->getOperand(1) == LHS);
2375 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2378 unsigned LHSCode = getSetCondCode(SCI);
2379 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2381 switch (Log.getOpcode()) {
2382 case Instruction::And: Code = LHSCode & RHSCode; break;
2383 case Instruction::Or: Code = LHSCode | RHSCode; break;
2384 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2385 default: assert(0 && "Illegal logical opcode!"); return 0;
2388 Value *RV = getSetCCValue(Code, LHS, RHS);
2389 if (Instruction *I = dyn_cast<Instruction>(RV))
2391 // Otherwise, it's a constant boolean value...
2392 return IC.ReplaceInstUsesWith(Log, RV);
2396 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2397 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2398 // guaranteed to be either a shift instruction or a binary operator.
2399 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2400 ConstantIntegral *OpRHS,
2401 ConstantIntegral *AndRHS,
2402 BinaryOperator &TheAnd) {
2403 Value *X = Op->getOperand(0);
2404 Constant *Together = 0;
2405 if (!isa<ShiftInst>(Op))
2406 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2408 switch (Op->getOpcode()) {
2409 case Instruction::Xor:
2410 if (Op->hasOneUse()) {
2411 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2412 std::string OpName = Op->getName(); Op->setName("");
2413 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2414 InsertNewInstBefore(And, TheAnd);
2415 return BinaryOperator::createXor(And, Together);
2418 case Instruction::Or:
2419 if (Together == AndRHS) // (X | C) & C --> C
2420 return ReplaceInstUsesWith(TheAnd, AndRHS);
2422 if (Op->hasOneUse() && Together != OpRHS) {
2423 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2424 std::string Op0Name = Op->getName(); Op->setName("");
2425 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2426 InsertNewInstBefore(Or, TheAnd);
2427 return BinaryOperator::createAnd(Or, AndRHS);
2430 case Instruction::Add:
2431 if (Op->hasOneUse()) {
2432 // Adding a one to a single bit bit-field should be turned into an XOR
2433 // of the bit. First thing to check is to see if this AND is with a
2434 // single bit constant.
2435 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
2437 // Clear bits that are not part of the constant.
2438 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2440 // If there is only one bit set...
2441 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2442 // Ok, at this point, we know that we are masking the result of the
2443 // ADD down to exactly one bit. If the constant we are adding has
2444 // no bits set below this bit, then we can eliminate the ADD.
2445 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
2447 // Check to see if any bits below the one bit set in AndRHSV are set.
2448 if ((AddRHS & (AndRHSV-1)) == 0) {
2449 // If not, the only thing that can effect the output of the AND is
2450 // the bit specified by AndRHSV. If that bit is set, the effect of
2451 // the XOR is to toggle the bit. If it is clear, then the ADD has
2453 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2454 TheAnd.setOperand(0, X);
2457 std::string Name = Op->getName(); Op->setName("");
2458 // Pull the XOR out of the AND.
2459 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2460 InsertNewInstBefore(NewAnd, TheAnd);
2461 return BinaryOperator::createXor(NewAnd, AndRHS);
2468 case Instruction::Shl: {
2469 // We know that the AND will not produce any of the bits shifted in, so if
2470 // the anded constant includes them, clear them now!
2472 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2473 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2474 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2476 if (CI == ShlMask) { // Masking out bits that the shift already masks
2477 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2478 } else if (CI != AndRHS) { // Reducing bits set in and.
2479 TheAnd.setOperand(1, CI);
2484 case Instruction::Shr:
2485 // We know that the AND will not produce any of the bits shifted in, so if
2486 // the anded constant includes them, clear them now! This only applies to
2487 // unsigned shifts, because a signed shr may bring in set bits!
2489 if (AndRHS->getType()->isUnsigned()) {
2490 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2491 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
2492 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2494 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2495 return ReplaceInstUsesWith(TheAnd, Op);
2496 } else if (CI != AndRHS) {
2497 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2500 } else { // Signed shr.
2501 // See if this is shifting in some sign extension, then masking it out
2503 if (Op->hasOneUse()) {
2504 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2505 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
2506 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2507 if (CI == AndRHS) { // Masking out bits shifted in.
2508 // Make the argument unsigned.
2509 Value *ShVal = Op->getOperand(0);
2510 ShVal = InsertCastBefore(ShVal,
2511 ShVal->getType()->getUnsignedVersion(),
2513 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
2514 OpRHS, Op->getName()),
2516 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
2517 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
2520 return new CastInst(ShVal, Op->getType());
2530 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2531 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2532 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2533 /// insert new instructions.
2534 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2535 bool Inside, Instruction &IB) {
2536 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2537 "Lo is not <= Hi in range emission code!");
2539 if (Lo == Hi) // Trivially false.
2540 return new SetCondInst(Instruction::SetNE, V, V);
2541 if (cast<ConstantIntegral>(Lo)->isMinValue())
2542 return new SetCondInst(Instruction::SetLT, V, Hi);
2544 Constant *AddCST = ConstantExpr::getNeg(Lo);
2545 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2546 InsertNewInstBefore(Add, IB);
2547 // Convert to unsigned for the comparison.
2548 const Type *UnsType = Add->getType()->getUnsignedVersion();
2549 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2550 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2551 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2552 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2555 if (Lo == Hi) // Trivially true.
2556 return new SetCondInst(Instruction::SetEQ, V, V);
2558 Hi = SubOne(cast<ConstantInt>(Hi));
2559 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
2560 return new SetCondInst(Instruction::SetGT, V, Hi);
2562 // Emit X-Lo > Hi-Lo-1
2563 Constant *AddCST = ConstantExpr::getNeg(Lo);
2564 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2565 InsertNewInstBefore(Add, IB);
2566 // Convert to unsigned for the comparison.
2567 const Type *UnsType = Add->getType()->getUnsignedVersion();
2568 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2569 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2570 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2571 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2574 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2575 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2576 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2577 // not, since all 1s are not contiguous.
2578 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2579 uint64_t V = Val->getRawValue();
2580 if (!isShiftedMask_64(V)) return false;
2582 // look for the first zero bit after the run of ones
2583 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2584 // look for the first non-zero bit
2585 ME = 64-CountLeadingZeros_64(V);
2591 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2592 /// where isSub determines whether the operator is a sub. If we can fold one of
2593 /// the following xforms:
2595 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2596 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2597 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2599 /// return (A +/- B).
2601 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2602 ConstantIntegral *Mask, bool isSub,
2604 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2605 if (!LHSI || LHSI->getNumOperands() != 2 ||
2606 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2608 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2610 switch (LHSI->getOpcode()) {
2612 case Instruction::And:
2613 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2614 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2615 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
2618 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2619 // part, we don't need any explicit masks to take them out of A. If that
2620 // is all N is, ignore it.
2622 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2623 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2625 if (MaskedValueIsZero(RHS, Mask))
2630 case Instruction::Or:
2631 case Instruction::Xor:
2632 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2633 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
2634 ConstantExpr::getAnd(N, Mask)->isNullValue())
2641 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2643 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2644 return InsertNewInstBefore(New, I);
2647 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
2648 bool Changed = SimplifyCommutative(I);
2649 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2651 if (isa<UndefValue>(Op1)) // X & undef -> 0
2652 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2656 return ReplaceInstUsesWith(I, Op1);
2658 // See if we can simplify any instructions used by the instruction whose sole
2659 // purpose is to compute bits we don't care about.
2660 uint64_t KnownZero, KnownOne;
2661 if (!isa<PackedType>(I.getType()) &&
2662 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2663 KnownZero, KnownOne))
2666 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
2667 uint64_t AndRHSMask = AndRHS->getZExtValue();
2668 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
2669 uint64_t NotAndRHS = AndRHSMask^TypeMask;
2671 // Optimize a variety of ((val OP C1) & C2) combinations...
2672 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
2673 Instruction *Op0I = cast<Instruction>(Op0);
2674 Value *Op0LHS = Op0I->getOperand(0);
2675 Value *Op0RHS = Op0I->getOperand(1);
2676 switch (Op0I->getOpcode()) {
2677 case Instruction::Xor:
2678 case Instruction::Or:
2679 // If the mask is only needed on one incoming arm, push it up.
2680 if (Op0I->hasOneUse()) {
2681 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
2682 // Not masking anything out for the LHS, move to RHS.
2683 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
2684 Op0RHS->getName()+".masked");
2685 InsertNewInstBefore(NewRHS, I);
2686 return BinaryOperator::create(
2687 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
2689 if (!isa<Constant>(Op0RHS) &&
2690 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
2691 // Not masking anything out for the RHS, move to LHS.
2692 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
2693 Op0LHS->getName()+".masked");
2694 InsertNewInstBefore(NewLHS, I);
2695 return BinaryOperator::create(
2696 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
2701 case Instruction::Add:
2702 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2703 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2704 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2705 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2706 return BinaryOperator::createAnd(V, AndRHS);
2707 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2708 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
2711 case Instruction::Sub:
2712 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2713 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2714 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2715 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2716 return BinaryOperator::createAnd(V, AndRHS);
2720 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2721 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2723 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2724 const Type *SrcTy = CI->getOperand(0)->getType();
2726 // If this is an integer truncation or change from signed-to-unsigned, and
2727 // if the source is an and/or with immediate, transform it. This
2728 // frequently occurs for bitfield accesses.
2729 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2730 if (SrcTy->getPrimitiveSizeInBits() >=
2731 I.getType()->getPrimitiveSizeInBits() &&
2732 CastOp->getNumOperands() == 2)
2733 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
2734 if (CastOp->getOpcode() == Instruction::And) {
2735 // Change: and (cast (and X, C1) to T), C2
2736 // into : and (cast X to T), trunc(C1)&C2
2737 // This will folds the two ands together, which may allow other
2739 Instruction *NewCast =
2740 new CastInst(CastOp->getOperand(0), I.getType(),
2741 CastOp->getName()+".shrunk");
2742 NewCast = InsertNewInstBefore(NewCast, I);
2744 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2745 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
2746 return BinaryOperator::createAnd(NewCast, C3);
2747 } else if (CastOp->getOpcode() == Instruction::Or) {
2748 // Change: and (cast (or X, C1) to T), C2
2749 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2750 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2751 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
2752 return ReplaceInstUsesWith(I, AndRHS);
2757 // Try to fold constant and into select arguments.
2758 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2759 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2761 if (isa<PHINode>(Op0))
2762 if (Instruction *NV = FoldOpIntoPhi(I))
2766 Value *Op0NotVal = dyn_castNotVal(Op0);
2767 Value *Op1NotVal = dyn_castNotVal(Op1);
2769 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
2770 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2772 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2773 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2774 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
2775 I.getName()+".demorgan");
2776 InsertNewInstBefore(Or, I);
2777 return BinaryOperator::createNot(Or);
2781 Value *A = 0, *B = 0;
2782 ConstantInt *C1 = 0, *C2 = 0;
2783 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
2784 if (A == Op1 || B == Op1) // (A | ?) & A --> A
2785 return ReplaceInstUsesWith(I, Op1);
2786 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
2787 if (A == Op0 || B == Op0) // A & (A | ?) --> A
2788 return ReplaceInstUsesWith(I, Op0);
2790 if (Op0->hasOneUse() &&
2791 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2792 if (A == Op1) { // (A^B)&A -> A&(A^B)
2793 I.swapOperands(); // Simplify below
2794 std::swap(Op0, Op1);
2795 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
2796 cast<BinaryOperator>(Op0)->swapOperands();
2797 I.swapOperands(); // Simplify below
2798 std::swap(Op0, Op1);
2801 if (Op1->hasOneUse() &&
2802 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2803 if (B == Op0) { // B&(A^B) -> B&(B^A)
2804 cast<BinaryOperator>(Op1)->swapOperands();
2807 if (A == Op0) { // A&(A^B) -> A & ~B
2808 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
2809 InsertNewInstBefore(NotB, I);
2810 return BinaryOperator::createAnd(A, NotB);
2816 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
2817 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2818 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2821 Value *LHSVal, *RHSVal;
2822 ConstantInt *LHSCst, *RHSCst;
2823 Instruction::BinaryOps LHSCC, RHSCC;
2824 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2825 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2826 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
2827 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2828 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2829 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2830 // Ensure that the larger constant is on the RHS.
2831 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2832 SetCondInst *LHS = cast<SetCondInst>(Op0);
2833 if (cast<ConstantBool>(Cmp)->getValue()) {
2834 std::swap(LHS, RHS);
2835 std::swap(LHSCst, RHSCst);
2836 std::swap(LHSCC, RHSCC);
2839 // At this point, we know we have have two setcc instructions
2840 // comparing a value against two constants and and'ing the result
2841 // together. Because of the above check, we know that we only have
2842 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2843 // FoldSetCCLogical check above), that the two constants are not
2845 assert(LHSCst != RHSCst && "Compares not folded above?");
2848 default: assert(0 && "Unknown integer condition code!");
2849 case Instruction::SetEQ:
2851 default: assert(0 && "Unknown integer condition code!");
2852 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
2853 case Instruction::SetGT: // (X == 13 & X > 15) -> false
2854 return ReplaceInstUsesWith(I, ConstantBool::False);
2855 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
2856 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
2857 return ReplaceInstUsesWith(I, LHS);
2859 case Instruction::SetNE:
2861 default: assert(0 && "Unknown integer condition code!");
2862 case Instruction::SetLT:
2863 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
2864 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
2865 break; // (X != 13 & X < 15) -> no change
2866 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
2867 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
2868 return ReplaceInstUsesWith(I, RHS);
2869 case Instruction::SetNE:
2870 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
2871 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2872 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2873 LHSVal->getName()+".off");
2874 InsertNewInstBefore(Add, I);
2875 const Type *UnsType = Add->getType()->getUnsignedVersion();
2876 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2877 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
2878 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2879 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2881 break; // (X != 13 & X != 15) -> no change
2884 case Instruction::SetLT:
2886 default: assert(0 && "Unknown integer condition code!");
2887 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
2888 case Instruction::SetGT: // (X < 13 & X > 15) -> false
2889 return ReplaceInstUsesWith(I, ConstantBool::False);
2890 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
2891 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
2892 return ReplaceInstUsesWith(I, LHS);
2894 case Instruction::SetGT:
2896 default: assert(0 && "Unknown integer condition code!");
2897 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
2898 return ReplaceInstUsesWith(I, LHS);
2899 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
2900 return ReplaceInstUsesWith(I, RHS);
2901 case Instruction::SetNE:
2902 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
2903 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
2904 break; // (X > 13 & X != 15) -> no change
2905 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
2906 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2912 // fold (and (cast A), (cast B)) -> (cast (and A, B))
2913 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2914 const Type *SrcTy = Op0C->getOperand(0)->getType();
2915 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2916 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
2917 // Only do this if the casts both really cause code to be generated.
2918 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
2919 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
2920 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
2921 Op1C->getOperand(0),
2923 InsertNewInstBefore(NewOp, I);
2924 return new CastInst(NewOp, I.getType());
2928 return Changed ? &I : 0;
2931 /// CollectBSwapParts - Look to see if the specified value defines a single byte
2932 /// in the result. If it does, and if the specified byte hasn't been filled in
2933 /// yet, fill it in and return false.
2934 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
2935 Instruction *I = dyn_cast<Instruction>(V);
2936 if (I == 0) return true;
2938 // If this is an or instruction, it is an inner node of the bswap.
2939 if (I->getOpcode() == Instruction::Or)
2940 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
2941 CollectBSwapParts(I->getOperand(1), ByteValues);
2943 // If this is a shift by a constant int, and it is "24", then its operand
2944 // defines a byte. We only handle unsigned types here.
2945 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
2946 // Not shifting the entire input by N-1 bytes?
2947 if (cast<ConstantInt>(I->getOperand(1))->getRawValue() !=
2948 8*(ByteValues.size()-1))
2952 if (I->getOpcode() == Instruction::Shl) {
2953 // X << 24 defines the top byte with the lowest of the input bytes.
2954 DestNo = ByteValues.size()-1;
2956 // X >>u 24 defines the low byte with the highest of the input bytes.
2960 // If the destination byte value is already defined, the values are or'd
2961 // together, which isn't a bswap (unless it's an or of the same bits).
2962 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
2964 ByteValues[DestNo] = I->getOperand(0);
2968 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
2970 Value *Shift = 0, *ShiftLHS = 0;
2971 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
2972 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
2973 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
2975 Instruction *SI = cast<Instruction>(Shift);
2977 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
2978 if (ShiftAmt->getRawValue() & 7 ||
2979 ShiftAmt->getRawValue() > 8*ByteValues.size())
2982 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
2984 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
2985 if (AndAmt->getRawValue() == uint64_t(0xFF) << 8*DestByte)
2987 // Unknown mask for bswap.
2988 if (DestByte == ByteValues.size()) return true;
2990 unsigned ShiftBytes = ShiftAmt->getRawValue()/8;
2992 if (SI->getOpcode() == Instruction::Shl)
2993 SrcByte = DestByte - ShiftBytes;
2995 SrcByte = DestByte + ShiftBytes;
2997 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
2998 if (SrcByte != ByteValues.size()-DestByte-1)
3001 // If the destination byte value is already defined, the values are or'd
3002 // together, which isn't a bswap (unless it's an or of the same bits).
3003 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3005 ByteValues[DestByte] = SI->getOperand(0);
3009 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3010 /// If so, insert the new bswap intrinsic and return it.
3011 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3012 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3013 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3016 /// ByteValues - For each byte of the result, we keep track of which value
3017 /// defines each byte.
3018 std::vector<Value*> ByteValues;
3019 ByteValues.resize(I.getType()->getPrimitiveSize());
3021 // Try to find all the pieces corresponding to the bswap.
3022 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3023 CollectBSwapParts(I.getOperand(1), ByteValues))
3026 // Check to see if all of the bytes come from the same value.
3027 Value *V = ByteValues[0];
3028 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3030 // Check to make sure that all of the bytes come from the same value.
3031 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3032 if (ByteValues[i] != V)
3035 // If they do then *success* we can turn this into a bswap. Figure out what
3036 // bswap to make it into.
3037 Module *M = I.getParent()->getParent()->getParent();
3038 const char *FnName = 0;
3039 if (I.getType() == Type::UShortTy)
3040 FnName = "llvm.bswap.i16";
3041 else if (I.getType() == Type::UIntTy)
3042 FnName = "llvm.bswap.i32";
3043 else if (I.getType() == Type::ULongTy)
3044 FnName = "llvm.bswap.i64";
3046 assert(0 && "Unknown integer type!");
3047 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3049 return new CallInst(F, V);
3053 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3054 bool Changed = SimplifyCommutative(I);
3055 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3057 if (isa<UndefValue>(Op1))
3058 return ReplaceInstUsesWith(I, // X | undef -> -1
3059 ConstantIntegral::getAllOnesValue(I.getType()));
3063 return ReplaceInstUsesWith(I, Op0);
3065 // See if we can simplify any instructions used by the instruction whose sole
3066 // purpose is to compute bits we don't care about.
3067 uint64_t KnownZero, KnownOne;
3068 if (!isa<PackedType>(I.getType()) &&
3069 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3070 KnownZero, KnownOne))
3074 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3075 ConstantInt *C1 = 0; Value *X = 0;
3076 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3077 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3078 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3080 InsertNewInstBefore(Or, I);
3081 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3084 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3085 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3086 std::string Op0Name = Op0->getName(); Op0->setName("");
3087 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3088 InsertNewInstBefore(Or, I);
3089 return BinaryOperator::createXor(Or,
3090 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3093 // Try to fold constant and into select arguments.
3094 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3095 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3097 if (isa<PHINode>(Op0))
3098 if (Instruction *NV = FoldOpIntoPhi(I))
3102 Value *A = 0, *B = 0;
3103 ConstantInt *C1 = 0, *C2 = 0;
3105 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3106 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3107 return ReplaceInstUsesWith(I, Op1);
3108 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3109 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3110 return ReplaceInstUsesWith(I, Op0);
3112 // (A | B) | C and A | (B | C) -> bswap if possible.
3113 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3114 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3115 match(Op1, m_Or(m_Value(), m_Value())) ||
3116 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3117 match(Op1, m_Shift(m_Value(), m_Value())))) {
3118 if (Instruction *BSwap = MatchBSwap(I))
3122 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3123 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3124 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3125 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3127 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3130 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3131 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3132 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3133 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3135 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3138 // (A & C1)|(B & C2)
3139 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3140 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3142 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3143 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3146 // If we have: ((V + N) & C1) | (V & C2)
3147 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3148 // replace with V+N.
3149 if (C1 == ConstantExpr::getNot(C2)) {
3150 Value *V1 = 0, *V2 = 0;
3151 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
3152 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3153 // Add commutes, try both ways.
3154 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3155 return ReplaceInstUsesWith(I, A);
3156 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3157 return ReplaceInstUsesWith(I, A);
3159 // Or commutes, try both ways.
3160 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
3161 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3162 // Add commutes, try both ways.
3163 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3164 return ReplaceInstUsesWith(I, B);
3165 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3166 return ReplaceInstUsesWith(I, B);
3171 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3172 if (A == Op1) // ~A | A == -1
3173 return ReplaceInstUsesWith(I,
3174 ConstantIntegral::getAllOnesValue(I.getType()));
3178 // Note, A is still live here!
3179 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3181 return ReplaceInstUsesWith(I,
3182 ConstantIntegral::getAllOnesValue(I.getType()));
3184 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3185 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3186 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3187 I.getName()+".demorgan"), I);
3188 return BinaryOperator::createNot(And);
3192 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
3193 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
3194 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3197 Value *LHSVal, *RHSVal;
3198 ConstantInt *LHSCst, *RHSCst;
3199 Instruction::BinaryOps LHSCC, RHSCC;
3200 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3201 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3202 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
3203 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3204 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3205 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3206 // Ensure that the larger constant is on the RHS.
3207 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3208 SetCondInst *LHS = cast<SetCondInst>(Op0);
3209 if (cast<ConstantBool>(Cmp)->getValue()) {
3210 std::swap(LHS, RHS);
3211 std::swap(LHSCst, RHSCst);
3212 std::swap(LHSCC, RHSCC);
3215 // At this point, we know we have have two setcc instructions
3216 // comparing a value against two constants and or'ing the result
3217 // together. Because of the above check, we know that we only have
3218 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3219 // FoldSetCCLogical check above), that the two constants are not
3221 assert(LHSCst != RHSCst && "Compares not folded above?");
3224 default: assert(0 && "Unknown integer condition code!");
3225 case Instruction::SetEQ:
3227 default: assert(0 && "Unknown integer condition code!");
3228 case Instruction::SetEQ:
3229 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3230 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3231 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3232 LHSVal->getName()+".off");
3233 InsertNewInstBefore(Add, I);
3234 const Type *UnsType = Add->getType()->getUnsignedVersion();
3235 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3236 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3237 AddCST = ConstantExpr::getCast(AddCST, UnsType);
3238 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
3240 break; // (X == 13 | X == 15) -> no change
3242 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
3244 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
3245 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
3246 return ReplaceInstUsesWith(I, RHS);
3249 case Instruction::SetNE:
3251 default: assert(0 && "Unknown integer condition code!");
3252 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
3253 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
3254 return ReplaceInstUsesWith(I, LHS);
3255 case Instruction::SetNE: // (X != 13 | X != 15) -> true
3256 case Instruction::SetLT: // (X != 13 | X < 15) -> true
3257 return ReplaceInstUsesWith(I, ConstantBool::True);
3260 case Instruction::SetLT:
3262 default: assert(0 && "Unknown integer condition code!");
3263 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
3265 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
3266 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
3267 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
3268 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
3269 return ReplaceInstUsesWith(I, RHS);
3272 case Instruction::SetGT:
3274 default: assert(0 && "Unknown integer condition code!");
3275 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
3276 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
3277 return ReplaceInstUsesWith(I, LHS);
3278 case Instruction::SetNE: // (X > 13 | X != 15) -> true
3279 case Instruction::SetLT: // (X > 13 | X < 15) -> true
3280 return ReplaceInstUsesWith(I, ConstantBool::True);
3286 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3287 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3288 const Type *SrcTy = Op0C->getOperand(0)->getType();
3289 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3290 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3291 // Only do this if the casts both really cause code to be generated.
3292 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3293 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3294 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3295 Op1C->getOperand(0),
3297 InsertNewInstBefore(NewOp, I);
3298 return new CastInst(NewOp, I.getType());
3303 return Changed ? &I : 0;
3306 // XorSelf - Implements: X ^ X --> 0
3309 XorSelf(Value *rhs) : RHS(rhs) {}
3310 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3311 Instruction *apply(BinaryOperator &Xor) const {
3317 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3318 bool Changed = SimplifyCommutative(I);
3319 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3321 if (isa<UndefValue>(Op1))
3322 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3324 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3325 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3326 assert(Result == &I && "AssociativeOpt didn't work?");
3327 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3330 // See if we can simplify any instructions used by the instruction whose sole
3331 // purpose is to compute bits we don't care about.
3332 uint64_t KnownZero, KnownOne;
3333 if (!isa<PackedType>(I.getType()) &&
3334 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3335 KnownZero, KnownOne))
3338 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3339 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3340 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
3341 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
3342 if (RHS == ConstantBool::True && SCI->hasOneUse())
3343 return new SetCondInst(SCI->getInverseCondition(),
3344 SCI->getOperand(0), SCI->getOperand(1));
3346 // ~(c-X) == X-c-1 == X+(-c-1)
3347 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3348 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3349 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3350 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3351 ConstantInt::get(I.getType(), 1));
3352 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3355 // ~(~X & Y) --> (X | ~Y)
3356 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3357 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3358 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3360 BinaryOperator::createNot(Op0I->getOperand(1),
3361 Op0I->getOperand(1)->getName()+".not");
3362 InsertNewInstBefore(NotY, I);
3363 return BinaryOperator::createOr(Op0NotVal, NotY);
3367 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3368 if (Op0I->getOpcode() == Instruction::Add) {
3369 // ~(X-c) --> (-c-1)-X
3370 if (RHS->isAllOnesValue()) {
3371 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3372 return BinaryOperator::createSub(
3373 ConstantExpr::getSub(NegOp0CI,
3374 ConstantInt::get(I.getType(), 1)),
3375 Op0I->getOperand(0));
3377 } else if (Op0I->getOpcode() == Instruction::Or) {
3378 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3379 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3380 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3381 // Anything in both C1 and C2 is known to be zero, remove it from
3383 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3384 NewRHS = ConstantExpr::getAnd(NewRHS,
3385 ConstantExpr::getNot(CommonBits));
3386 WorkList.push_back(Op0I);
3387 I.setOperand(0, Op0I->getOperand(0));
3388 I.setOperand(1, NewRHS);
3394 // Try to fold constant and into select arguments.
3395 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3396 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3398 if (isa<PHINode>(Op0))
3399 if (Instruction *NV = FoldOpIntoPhi(I))
3403 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3405 return ReplaceInstUsesWith(I,
3406 ConstantIntegral::getAllOnesValue(I.getType()));
3408 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3410 return ReplaceInstUsesWith(I,
3411 ConstantIntegral::getAllOnesValue(I.getType()));
3413 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3414 if (Op1I->getOpcode() == Instruction::Or) {
3415 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3416 Op1I->swapOperands();
3418 std::swap(Op0, Op1);
3419 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3420 I.swapOperands(); // Simplified below.
3421 std::swap(Op0, Op1);
3423 } else if (Op1I->getOpcode() == Instruction::Xor) {
3424 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3425 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3426 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3427 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3428 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3429 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3430 Op1I->swapOperands();
3431 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3432 I.swapOperands(); // Simplified below.
3433 std::swap(Op0, Op1);
3437 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3438 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3439 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3440 Op0I->swapOperands();
3441 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3442 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3443 InsertNewInstBefore(NotB, I);
3444 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3446 } else if (Op0I->getOpcode() == Instruction::Xor) {
3447 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3448 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3449 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3450 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3451 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3452 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3453 Op0I->swapOperands();
3454 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3455 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3456 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3457 InsertNewInstBefore(N, I);
3458 return BinaryOperator::createAnd(N, Op1);
3462 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
3463 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
3464 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3467 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3468 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3469 const Type *SrcTy = Op0C->getOperand(0)->getType();
3470 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3471 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3472 // Only do this if the casts both really cause code to be generated.
3473 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3474 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3475 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
3476 Op1C->getOperand(0),
3478 InsertNewInstBefore(NewOp, I);
3479 return new CastInst(NewOp, I.getType());
3483 return Changed ? &I : 0;
3486 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
3487 /// overflowed for this type.
3488 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3490 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
3491 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
3494 static bool isPositive(ConstantInt *C) {
3495 return cast<ConstantSInt>(C)->getValue() >= 0;
3498 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
3499 /// overflowed for this type.
3500 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3502 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
3504 if (In1->getType()->isUnsigned())
3505 return cast<ConstantUInt>(Result)->getValue() <
3506 cast<ConstantUInt>(In1)->getValue();
3507 if (isPositive(In1) != isPositive(In2))
3509 if (isPositive(In1))
3510 return cast<ConstantSInt>(Result)->getValue() <
3511 cast<ConstantSInt>(In1)->getValue();
3512 return cast<ConstantSInt>(Result)->getValue() >
3513 cast<ConstantSInt>(In1)->getValue();
3516 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3517 /// code necessary to compute the offset from the base pointer (without adding
3518 /// in the base pointer). Return the result as a signed integer of intptr size.
3519 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3520 TargetData &TD = IC.getTargetData();
3521 gep_type_iterator GTI = gep_type_begin(GEP);
3522 const Type *UIntPtrTy = TD.getIntPtrType();
3523 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3524 Value *Result = Constant::getNullValue(SIntPtrTy);
3526 // Build a mask for high order bits.
3527 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3529 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3530 Value *Op = GEP->getOperand(i);
3531 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3532 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
3534 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3535 if (!OpC->isNullValue()) {
3536 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
3537 Scale = ConstantExpr::getMul(OpC, Scale);
3538 if (Constant *RC = dyn_cast<Constant>(Result))
3539 Result = ConstantExpr::getAdd(RC, Scale);
3541 // Emit an add instruction.
3542 Result = IC.InsertNewInstBefore(
3543 BinaryOperator::createAdd(Result, Scale,
3544 GEP->getName()+".offs"), I);
3548 // Convert to correct type.
3549 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
3550 Op->getName()+".c"), I);
3552 // We'll let instcombine(mul) convert this to a shl if possible.
3553 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3554 GEP->getName()+".idx"), I);
3556 // Emit an add instruction.
3557 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3558 GEP->getName()+".offs"), I);
3564 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3565 /// else. At this point we know that the GEP is on the LHS of the comparison.
3566 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3567 Instruction::BinaryOps Cond,
3569 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3571 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3572 if (isa<PointerType>(CI->getOperand(0)->getType()))
3573 RHS = CI->getOperand(0);
3575 Value *PtrBase = GEPLHS->getOperand(0);
3576 if (PtrBase == RHS) {
3577 // As an optimization, we don't actually have to compute the actual value of
3578 // OFFSET if this is a seteq or setne comparison, just return whether each
3579 // index is zero or not.
3580 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3581 Instruction *InVal = 0;
3582 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3583 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3585 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3586 if (isa<UndefValue>(C)) // undef index -> undef.
3587 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3588 if (C->isNullValue())
3590 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3591 EmitIt = false; // This is indexing into a zero sized array?
3592 } else if (isa<ConstantInt>(C))
3593 return ReplaceInstUsesWith(I, // No comparison is needed here.
3594 ConstantBool::get(Cond == Instruction::SetNE));
3599 new SetCondInst(Cond, GEPLHS->getOperand(i),
3600 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3604 InVal = InsertNewInstBefore(InVal, I);
3605 InsertNewInstBefore(Comp, I);
3606 if (Cond == Instruction::SetNE) // True if any are unequal
3607 InVal = BinaryOperator::createOr(InVal, Comp);
3608 else // True if all are equal
3609 InVal = BinaryOperator::createAnd(InVal, Comp);
3617 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
3618 ConstantBool::get(Cond == Instruction::SetEQ));
3621 // Only lower this if the setcc is the only user of the GEP or if we expect
3622 // the result to fold to a constant!
3623 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
3624 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
3625 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
3626 return new SetCondInst(Cond, Offset,
3627 Constant::getNullValue(Offset->getType()));
3629 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
3630 // If the base pointers are different, but the indices are the same, just
3631 // compare the base pointer.
3632 if (PtrBase != GEPRHS->getOperand(0)) {
3633 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
3634 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
3635 GEPRHS->getOperand(0)->getType();
3637 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3638 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3639 IndicesTheSame = false;
3643 // If all indices are the same, just compare the base pointers.
3645 return new SetCondInst(Cond, GEPLHS->getOperand(0),
3646 GEPRHS->getOperand(0));
3648 // Otherwise, the base pointers are different and the indices are
3649 // different, bail out.
3653 // If one of the GEPs has all zero indices, recurse.
3654 bool AllZeros = true;
3655 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3656 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
3657 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
3662 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
3663 SetCondInst::getSwappedCondition(Cond), I);
3665 // If the other GEP has all zero indices, recurse.
3667 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3668 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
3669 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
3674 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
3676 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
3677 // If the GEPs only differ by one index, compare it.
3678 unsigned NumDifferences = 0; // Keep track of # differences.
3679 unsigned DiffOperand = 0; // The operand that differs.
3680 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3681 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3682 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
3683 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
3684 // Irreconcilable differences.
3688 if (NumDifferences++) break;
3693 if (NumDifferences == 0) // SAME GEP?
3694 return ReplaceInstUsesWith(I, // No comparison is needed here.
3695 ConstantBool::get(Cond == Instruction::SetEQ));
3696 else if (NumDifferences == 1) {
3697 Value *LHSV = GEPLHS->getOperand(DiffOperand);
3698 Value *RHSV = GEPRHS->getOperand(DiffOperand);
3700 // Convert the operands to signed values to make sure to perform a
3701 // signed comparison.
3702 const Type *NewTy = LHSV->getType()->getSignedVersion();
3703 if (LHSV->getType() != NewTy)
3704 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
3705 LHSV->getName()), I);
3706 if (RHSV->getType() != NewTy)
3707 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
3708 RHSV->getName()), I);
3709 return new SetCondInst(Cond, LHSV, RHSV);
3713 // Only lower this if the setcc is the only user of the GEP or if we expect
3714 // the result to fold to a constant!
3715 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
3716 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
3717 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
3718 Value *L = EmitGEPOffset(GEPLHS, I, *this);
3719 Value *R = EmitGEPOffset(GEPRHS, I, *this);
3720 return new SetCondInst(Cond, L, R);
3727 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
3728 bool Changed = SimplifyCommutative(I);
3729 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3730 const Type *Ty = Op0->getType();
3734 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
3736 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
3737 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
3739 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
3740 // addresses never equal each other! We already know that Op0 != Op1.
3741 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
3742 isa<ConstantPointerNull>(Op0)) &&
3743 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
3744 isa<ConstantPointerNull>(Op1)))
3745 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
3747 // setcc's with boolean values can always be turned into bitwise operations
3748 if (Ty == Type::BoolTy) {
3749 switch (I.getOpcode()) {
3750 default: assert(0 && "Invalid setcc instruction!");
3751 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
3752 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
3753 InsertNewInstBefore(Xor, I);
3754 return BinaryOperator::createNot(Xor);
3756 case Instruction::SetNE:
3757 return BinaryOperator::createXor(Op0, Op1);
3759 case Instruction::SetGT:
3760 std::swap(Op0, Op1); // Change setgt -> setlt
3762 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
3763 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3764 InsertNewInstBefore(Not, I);
3765 return BinaryOperator::createAnd(Not, Op1);
3767 case Instruction::SetGE:
3768 std::swap(Op0, Op1); // Change setge -> setle
3770 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
3771 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3772 InsertNewInstBefore(Not, I);
3773 return BinaryOperator::createOr(Not, Op1);
3778 // See if we are doing a comparison between a constant and an instruction that
3779 // can be folded into the comparison.
3780 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3781 // Check to see if we are comparing against the minimum or maximum value...
3782 if (CI->isMinValue()) {
3783 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
3784 return ReplaceInstUsesWith(I, ConstantBool::False);
3785 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
3786 return ReplaceInstUsesWith(I, ConstantBool::True);
3787 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
3788 return BinaryOperator::createSetEQ(Op0, Op1);
3789 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
3790 return BinaryOperator::createSetNE(Op0, Op1);
3792 } else if (CI->isMaxValue()) {
3793 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
3794 return ReplaceInstUsesWith(I, ConstantBool::False);
3795 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
3796 return ReplaceInstUsesWith(I, ConstantBool::True);
3797 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
3798 return BinaryOperator::createSetEQ(Op0, Op1);
3799 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
3800 return BinaryOperator::createSetNE(Op0, Op1);
3802 // Comparing against a value really close to min or max?
3803 } else if (isMinValuePlusOne(CI)) {
3804 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
3805 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
3806 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
3807 return BinaryOperator::createSetNE(Op0, SubOne(CI));
3809 } else if (isMaxValueMinusOne(CI)) {
3810 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
3811 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
3812 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
3813 return BinaryOperator::createSetNE(Op0, AddOne(CI));
3816 // If we still have a setle or setge instruction, turn it into the
3817 // appropriate setlt or setgt instruction. Since the border cases have
3818 // already been handled above, this requires little checking.
3820 if (I.getOpcode() == Instruction::SetLE)
3821 return BinaryOperator::createSetLT(Op0, AddOne(CI));
3822 if (I.getOpcode() == Instruction::SetGE)
3823 return BinaryOperator::createSetGT(Op0, SubOne(CI));
3826 // See if we can fold the comparison based on bits known to be zero or one
3828 uint64_t KnownZero, KnownOne;
3829 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
3830 KnownZero, KnownOne, 0))
3833 // Given the known and unknown bits, compute a range that the LHS could be
3835 if (KnownOne | KnownZero) {
3836 if (Ty->isUnsigned()) { // Unsigned comparison.
3838 uint64_t RHSVal = CI->getZExtValue();
3839 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3841 switch (I.getOpcode()) { // LE/GE have been folded already.
3842 default: assert(0 && "Unknown setcc opcode!");
3843 case Instruction::SetEQ:
3844 if (Max < RHSVal || Min > RHSVal)
3845 return ReplaceInstUsesWith(I, ConstantBool::False);
3847 case Instruction::SetNE:
3848 if (Max < RHSVal || Min > RHSVal)
3849 return ReplaceInstUsesWith(I, ConstantBool::True);
3851 case Instruction::SetLT:
3852 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3853 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3855 case Instruction::SetGT:
3856 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3857 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3860 } else { // Signed comparison.
3862 int64_t RHSVal = CI->getSExtValue();
3863 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3865 switch (I.getOpcode()) { // LE/GE have been folded already.
3866 default: assert(0 && "Unknown setcc opcode!");
3867 case Instruction::SetEQ:
3868 if (Max < RHSVal || Min > RHSVal)
3869 return ReplaceInstUsesWith(I, ConstantBool::False);
3871 case Instruction::SetNE:
3872 if (Max < RHSVal || Min > RHSVal)
3873 return ReplaceInstUsesWith(I, ConstantBool::True);
3875 case Instruction::SetLT:
3876 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3877 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3879 case Instruction::SetGT:
3880 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3881 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3888 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3889 switch (LHSI->getOpcode()) {
3890 case Instruction::And:
3891 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
3892 LHSI->getOperand(0)->hasOneUse()) {
3893 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
3895 // If an operand is an AND of a truncating cast, we can widen the
3896 // and/compare to be the input width without changing the value
3897 // produced, eliminating a cast.
3898 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
3899 // We can do this transformation if either the AND constant does not
3900 // have its sign bit set or if it is an equality comparison.
3901 // Extending a relational comparison when we're checking the sign
3902 // bit would not work.
3903 if (Cast->hasOneUse() && Cast->isTruncIntCast() &&
3905 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
3906 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
3907 ConstantInt *NewCST;
3909 if (Cast->getOperand(0)->getType()->isSigned()) {
3910 NewCST = ConstantSInt::get(Cast->getOperand(0)->getType(),
3911 AndCST->getZExtValue());
3912 NewCI = ConstantSInt::get(Cast->getOperand(0)->getType(),
3913 CI->getZExtValue());
3915 NewCST = ConstantUInt::get(Cast->getOperand(0)->getType(),
3916 AndCST->getZExtValue());
3917 NewCI = ConstantUInt::get(Cast->getOperand(0)->getType(),
3918 CI->getZExtValue());
3920 Instruction *NewAnd =
3921 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
3923 InsertNewInstBefore(NewAnd, I);
3924 return new SetCondInst(I.getOpcode(), NewAnd, NewCI);
3928 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
3929 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
3930 // happens a LOT in code produced by the C front-end, for bitfield
3932 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
3934 // Check to see if there is a noop-cast between the shift and the and.
3936 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
3937 if (CI->getOperand(0)->getType()->isIntegral() &&
3938 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
3939 CI->getType()->getPrimitiveSizeInBits())
3940 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
3943 ConstantUInt *ShAmt;
3944 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
3945 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
3946 const Type *AndTy = AndCST->getType(); // Type of the and.
3948 // We can fold this as long as we can't shift unknown bits
3949 // into the mask. This can only happen with signed shift
3950 // rights, as they sign-extend.
3952 bool CanFold = Shift->isLogicalShift();
3954 // To test for the bad case of the signed shr, see if any
3955 // of the bits shifted in could be tested after the mask.
3956 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
3957 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
3959 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
3961 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
3963 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
3969 if (Shift->getOpcode() == Instruction::Shl)
3970 NewCst = ConstantExpr::getUShr(CI, ShAmt);
3972 NewCst = ConstantExpr::getShl(CI, ShAmt);
3974 // Check to see if we are shifting out any of the bits being
3976 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
3977 // If we shifted bits out, the fold is not going to work out.
3978 // As a special case, check to see if this means that the
3979 // result is always true or false now.
3980 if (I.getOpcode() == Instruction::SetEQ)
3981 return ReplaceInstUsesWith(I, ConstantBool::False);
3982 if (I.getOpcode() == Instruction::SetNE)
3983 return ReplaceInstUsesWith(I, ConstantBool::True);
3985 I.setOperand(1, NewCst);
3986 Constant *NewAndCST;
3987 if (Shift->getOpcode() == Instruction::Shl)
3988 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
3990 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
3991 LHSI->setOperand(1, NewAndCST);
3993 LHSI->setOperand(0, Shift->getOperand(0));
3995 Value *NewCast = InsertCastBefore(Shift->getOperand(0), AndTy,
3997 LHSI->setOperand(0, NewCast);
3999 WorkList.push_back(Shift); // Shift is dead.
4000 AddUsesToWorkList(I);
4006 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4007 // preferable because it allows the C<<Y expression to be hoisted out
4008 // of a loop if Y is invariant and X is not.
4009 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4010 I.isEquality() && !Shift->isArithmeticShift() &&
4011 isa<Instruction>(Shift->getOperand(0))) {
4014 if (Shift->getOpcode() == Instruction::Shr) {
4015 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4018 // Make sure we insert a logical shift.
4019 Constant *NewAndCST = AndCST;
4020 if (AndCST->getType()->isSigned())
4021 NewAndCST = ConstantExpr::getCast(AndCST,
4022 AndCST->getType()->getUnsignedVersion());
4023 NS = new ShiftInst(Instruction::Shr, NewAndCST,
4024 Shift->getOperand(1), "tmp");
4026 InsertNewInstBefore(cast<Instruction>(NS), I);
4028 // If C's sign doesn't agree with the and, insert a cast now.
4029 if (NS->getType() != LHSI->getType())
4030 NS = InsertCastBefore(NS, LHSI->getType(), I);
4032 Value *ShiftOp = Shift->getOperand(0);
4033 if (ShiftOp->getType() != LHSI->getType())
4034 ShiftOp = InsertCastBefore(ShiftOp, LHSI->getType(), I);
4036 // Compute X & (C << Y).
4037 Instruction *NewAnd =
4038 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4039 InsertNewInstBefore(NewAnd, I);
4041 I.setOperand(0, NewAnd);
4047 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
4048 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
4049 if (I.isEquality()) {
4050 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4052 // Check that the shift amount is in range. If not, don't perform
4053 // undefined shifts. When the shift is visited it will be
4055 if (ShAmt->getValue() >= TypeBits)
4058 // If we are comparing against bits always shifted out, the
4059 // comparison cannot succeed.
4061 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
4062 if (Comp != CI) {// Comparing against a bit that we know is zero.
4063 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4064 Constant *Cst = ConstantBool::get(IsSetNE);
4065 return ReplaceInstUsesWith(I, Cst);
4068 if (LHSI->hasOneUse()) {
4069 // Otherwise strength reduce the shift into an and.
4070 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
4071 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4074 if (CI->getType()->isUnsigned()) {
4075 Mask = ConstantUInt::get(CI->getType(), Val);
4076 } else if (ShAmtVal != 0) {
4077 Mask = ConstantSInt::get(CI->getType(), Val);
4079 Mask = ConstantInt::getAllOnesValue(CI->getType());
4083 BinaryOperator::createAnd(LHSI->getOperand(0),
4084 Mask, LHSI->getName()+".mask");
4085 Value *And = InsertNewInstBefore(AndI, I);
4086 return new SetCondInst(I.getOpcode(), And,
4087 ConstantExpr::getUShr(CI, ShAmt));
4093 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
4094 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
4095 if (I.isEquality()) {
4096 // Check that the shift amount is in range. If not, don't perform
4097 // undefined shifts. When the shift is visited it will be
4099 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4100 if (ShAmt->getValue() >= TypeBits)
4103 // If we are comparing against bits always shifted out, the
4104 // comparison cannot succeed.
4106 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
4108 if (Comp != CI) {// Comparing against a bit that we know is zero.
4109 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4110 Constant *Cst = ConstantBool::get(IsSetNE);
4111 return ReplaceInstUsesWith(I, Cst);
4114 if (LHSI->hasOneUse() || CI->isNullValue()) {
4115 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
4117 // Otherwise strength reduce the shift into an and.
4118 uint64_t Val = ~0ULL; // All ones.
4119 Val <<= ShAmtVal; // Shift over to the right spot.
4122 if (CI->getType()->isUnsigned()) {
4123 Val &= ~0ULL >> (64-TypeBits);
4124 Mask = ConstantUInt::get(CI->getType(), Val);
4126 Mask = ConstantSInt::get(CI->getType(), Val);
4130 BinaryOperator::createAnd(LHSI->getOperand(0),
4131 Mask, LHSI->getName()+".mask");
4132 Value *And = InsertNewInstBefore(AndI, I);
4133 return new SetCondInst(I.getOpcode(), And,
4134 ConstantExpr::getShl(CI, ShAmt));
4140 case Instruction::Div:
4141 // Fold: (div X, C1) op C2 -> range check
4142 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4143 // Fold this div into the comparison, producing a range check.
4144 // Determine, based on the divide type, what the range is being
4145 // checked. If there is an overflow on the low or high side, remember
4146 // it, otherwise compute the range [low, hi) bounding the new value.
4147 bool LoOverflow = false, HiOverflow = 0;
4148 ConstantInt *LoBound = 0, *HiBound = 0;
4151 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
4153 Instruction::BinaryOps Opcode = I.getOpcode();
4155 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
4156 } else if (LHSI->getType()->isUnsigned()) { // udiv
4158 LoOverflow = ProdOV;
4159 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4160 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4161 if (CI->isNullValue()) { // (X / pos) op 0
4163 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4165 } else if (isPositive(CI)) { // (X / pos) op pos
4167 LoOverflow = ProdOV;
4168 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4169 } else { // (X / pos) op neg
4170 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4171 LoOverflow = AddWithOverflow(LoBound, Prod,
4172 cast<ConstantInt>(DivRHSH));
4174 HiOverflow = ProdOV;
4176 } else { // Divisor is < 0.
4177 if (CI->isNullValue()) { // (X / neg) op 0
4178 LoBound = AddOne(DivRHS);
4179 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4180 if (HiBound == DivRHS)
4181 LoBound = 0; // - INTMIN = INTMIN
4182 } else if (isPositive(CI)) { // (X / neg) op pos
4183 HiOverflow = LoOverflow = ProdOV;
4185 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4186 HiBound = AddOne(Prod);
4187 } else { // (X / neg) op neg
4189 LoOverflow = HiOverflow = ProdOV;
4190 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4193 // Dividing by a negate swaps the condition.
4194 Opcode = SetCondInst::getSwappedCondition(Opcode);
4198 Value *X = LHSI->getOperand(0);
4200 default: assert(0 && "Unhandled setcc opcode!");
4201 case Instruction::SetEQ:
4202 if (LoOverflow && HiOverflow)
4203 return ReplaceInstUsesWith(I, ConstantBool::False);
4204 else if (HiOverflow)
4205 return new SetCondInst(Instruction::SetGE, X, LoBound);
4206 else if (LoOverflow)
4207 return new SetCondInst(Instruction::SetLT, X, HiBound);
4209 return InsertRangeTest(X, LoBound, HiBound, true, I);
4210 case Instruction::SetNE:
4211 if (LoOverflow && HiOverflow)
4212 return ReplaceInstUsesWith(I, ConstantBool::True);
4213 else if (HiOverflow)
4214 return new SetCondInst(Instruction::SetLT, X, LoBound);
4215 else if (LoOverflow)
4216 return new SetCondInst(Instruction::SetGE, X, HiBound);
4218 return InsertRangeTest(X, LoBound, HiBound, false, I);
4219 case Instruction::SetLT:
4221 return ReplaceInstUsesWith(I, ConstantBool::False);
4222 return new SetCondInst(Instruction::SetLT, X, LoBound);
4223 case Instruction::SetGT:
4225 return ReplaceInstUsesWith(I, ConstantBool::False);
4226 return new SetCondInst(Instruction::SetGE, X, HiBound);
4233 // Simplify seteq and setne instructions...
4234 if (I.isEquality()) {
4235 bool isSetNE = I.getOpcode() == Instruction::SetNE;
4237 // If the first operand is (and|or|xor) with a constant, and the second
4238 // operand is a constant, simplify a bit.
4239 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4240 switch (BO->getOpcode()) {
4241 case Instruction::Rem:
4242 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4243 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
4245 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
4246 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
4247 if (isPowerOf2_64(V)) {
4248 unsigned L2 = Log2_64(V);
4249 const Type *UTy = BO->getType()->getUnsignedVersion();
4250 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
4252 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
4253 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
4254 RHSCst, BO->getName()), I);
4255 return BinaryOperator::create(I.getOpcode(), NewRem,
4256 Constant::getNullValue(UTy));
4261 case Instruction::Add:
4262 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4263 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4264 if (BO->hasOneUse())
4265 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4266 ConstantExpr::getSub(CI, BOp1C));
4267 } else if (CI->isNullValue()) {
4268 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4269 // efficiently invertible, or if the add has just this one use.
4270 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4272 if (Value *NegVal = dyn_castNegVal(BOp1))
4273 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
4274 else if (Value *NegVal = dyn_castNegVal(BOp0))
4275 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
4276 else if (BO->hasOneUse()) {
4277 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4279 InsertNewInstBefore(Neg, I);
4280 return new SetCondInst(I.getOpcode(), BOp0, Neg);
4284 case Instruction::Xor:
4285 // For the xor case, we can xor two constants together, eliminating
4286 // the explicit xor.
4287 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4288 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
4289 ConstantExpr::getXor(CI, BOC));
4292 case Instruction::Sub:
4293 // Replace (([sub|xor] A, B) != 0) with (A != B)
4294 if (CI->isNullValue())
4295 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4299 case Instruction::Or:
4300 // If bits are being or'd in that are not present in the constant we
4301 // are comparing against, then the comparison could never succeed!
4302 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4303 Constant *NotCI = ConstantExpr::getNot(CI);
4304 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4305 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4309 case Instruction::And:
4310 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4311 // If bits are being compared against that are and'd out, then the
4312 // comparison can never succeed!
4313 if (!ConstantExpr::getAnd(CI,
4314 ConstantExpr::getNot(BOC))->isNullValue())
4315 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4317 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4318 if (CI == BOC && isOneBitSet(CI))
4319 return new SetCondInst(isSetNE ? Instruction::SetEQ :
4320 Instruction::SetNE, Op0,
4321 Constant::getNullValue(CI->getType()));
4323 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
4324 // to be a signed value as appropriate.
4325 if (isSignBit(BOC)) {
4326 Value *X = BO->getOperand(0);
4327 // If 'X' is not signed, insert a cast now...
4328 if (!BOC->getType()->isSigned()) {
4329 const Type *DestTy = BOC->getType()->getSignedVersion();
4330 X = InsertCastBefore(X, DestTy, I);
4332 return new SetCondInst(isSetNE ? Instruction::SetLT :
4333 Instruction::SetGE, X,
4334 Constant::getNullValue(X->getType()));
4337 // ((X & ~7) == 0) --> X < 8
4338 if (CI->isNullValue() && isHighOnes(BOC)) {
4339 Value *X = BO->getOperand(0);
4340 Constant *NegX = ConstantExpr::getNeg(BOC);
4342 // If 'X' is signed, insert a cast now.
4343 if (NegX->getType()->isSigned()) {
4344 const Type *DestTy = NegX->getType()->getUnsignedVersion();
4345 X = InsertCastBefore(X, DestTy, I);
4346 NegX = ConstantExpr::getCast(NegX, DestTy);
4349 return new SetCondInst(isSetNE ? Instruction::SetGE :
4350 Instruction::SetLT, X, NegX);
4357 } else { // Not a SetEQ/SetNE
4358 // If the LHS is a cast from an integral value of the same size,
4359 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
4360 Value *CastOp = Cast->getOperand(0);
4361 const Type *SrcTy = CastOp->getType();
4362 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
4363 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
4364 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
4365 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
4366 "Source and destination signednesses should differ!");
4367 if (Cast->getType()->isSigned()) {
4368 // If this is a signed comparison, check for comparisons in the
4369 // vicinity of zero.
4370 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
4372 return BinaryOperator::createSetGT(CastOp,
4373 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
4374 else if (I.getOpcode() == Instruction::SetGT &&
4375 cast<ConstantSInt>(CI)->getValue() == -1)
4376 // X > -1 => x < 128
4377 return BinaryOperator::createSetLT(CastOp,
4378 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
4380 ConstantUInt *CUI = cast<ConstantUInt>(CI);
4381 if (I.getOpcode() == Instruction::SetLT &&
4382 CUI->getValue() == 1ULL << (SrcTySize-1))
4383 // X < 128 => X > -1
4384 return BinaryOperator::createSetGT(CastOp,
4385 ConstantSInt::get(SrcTy, -1));
4386 else if (I.getOpcode() == Instruction::SetGT &&
4387 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
4389 return BinaryOperator::createSetLT(CastOp,
4390 Constant::getNullValue(SrcTy));
4397 // Handle setcc with constant RHS's that can be integer, FP or pointer.
4398 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4399 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4400 switch (LHSI->getOpcode()) {
4401 case Instruction::GetElementPtr:
4402 if (RHSC->isNullValue()) {
4403 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
4404 bool isAllZeros = true;
4405 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4406 if (!isa<Constant>(LHSI->getOperand(i)) ||
4407 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4412 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
4413 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4417 case Instruction::PHI:
4418 if (Instruction *NV = FoldOpIntoPhi(I))
4421 case Instruction::Select:
4422 // If either operand of the select is a constant, we can fold the
4423 // comparison into the select arms, which will cause one to be
4424 // constant folded and the select turned into a bitwise or.
4425 Value *Op1 = 0, *Op2 = 0;
4426 if (LHSI->hasOneUse()) {
4427 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4428 // Fold the known value into the constant operand.
4429 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4430 // Insert a new SetCC of the other select operand.
4431 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4432 LHSI->getOperand(2), RHSC,
4434 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4435 // Fold the known value into the constant operand.
4436 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4437 // Insert a new SetCC of the other select operand.
4438 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4439 LHSI->getOperand(1), RHSC,
4445 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4450 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
4451 if (User *GEP = dyn_castGetElementPtr(Op0))
4452 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
4454 if (User *GEP = dyn_castGetElementPtr(Op1))
4455 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
4456 SetCondInst::getSwappedCondition(I.getOpcode()), I))
4459 // Test to see if the operands of the setcc are casted versions of other
4460 // values. If the cast can be stripped off both arguments, we do so now.
4461 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4462 Value *CastOp0 = CI->getOperand(0);
4463 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
4464 (isa<Constant>(Op1) || isa<CastInst>(Op1)) && I.isEquality()) {
4465 // We keep moving the cast from the left operand over to the right
4466 // operand, where it can often be eliminated completely.
4469 // If operand #1 is a cast instruction, see if we can eliminate it as
4471 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
4472 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
4474 Op1 = CI2->getOperand(0);
4476 // If Op1 is a constant, we can fold the cast into the constant.
4477 if (Op1->getType() != Op0->getType())
4478 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4479 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
4481 // Otherwise, cast the RHS right before the setcc
4482 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
4483 InsertNewInstBefore(cast<Instruction>(Op1), I);
4485 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
4488 // Handle the special case of: setcc (cast bool to X), <cst>
4489 // This comes up when you have code like
4492 // For generality, we handle any zero-extension of any operand comparison
4493 // with a constant or another cast from the same type.
4494 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4495 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
4499 if (I.isEquality()) {
4501 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4502 (A == Op1 || B == Op1)) {
4503 // (A^B) == A -> B == 0
4504 Value *OtherVal = A == Op1 ? B : A;
4505 return BinaryOperator::create(I.getOpcode(), OtherVal,
4506 Constant::getNullValue(A->getType()));
4507 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4508 (A == Op0 || B == Op0)) {
4509 // A == (A^B) -> B == 0
4510 Value *OtherVal = A == Op0 ? B : A;
4511 return BinaryOperator::create(I.getOpcode(), OtherVal,
4512 Constant::getNullValue(A->getType()));
4513 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4514 // (A-B) == A -> B == 0
4515 return BinaryOperator::create(I.getOpcode(), B,
4516 Constant::getNullValue(B->getType()));
4517 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4518 // A == (A-B) -> B == 0
4519 return BinaryOperator::create(I.getOpcode(), B,
4520 Constant::getNullValue(B->getType()));
4523 return Changed ? &I : 0;
4526 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
4527 // We only handle extending casts so far.
4529 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
4530 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
4531 const Type *SrcTy = LHSCIOp->getType();
4532 const Type *DestTy = SCI.getOperand(0)->getType();
4535 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
4538 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
4539 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
4540 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
4542 // Is this a sign or zero extension?
4543 bool isSignSrc = SrcTy->isSigned();
4544 bool isSignDest = DestTy->isSigned();
4546 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
4547 // Not an extension from the same type?
4548 RHSCIOp = CI->getOperand(0);
4549 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
4550 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
4551 // Compute the constant that would happen if we truncated to SrcTy then
4552 // reextended to DestTy.
4553 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
4555 if (ConstantExpr::getCast(Res, DestTy) == CI) {
4558 // If the value cannot be represented in the shorter type, we cannot emit
4559 // a simple comparison.
4560 if (SCI.getOpcode() == Instruction::SetEQ)
4561 return ReplaceInstUsesWith(SCI, ConstantBool::False);
4562 if (SCI.getOpcode() == Instruction::SetNE)
4563 return ReplaceInstUsesWith(SCI, ConstantBool::True);
4565 // Evaluate the comparison for LT.
4567 if (DestTy->isSigned()) {
4568 // We're performing a signed comparison.
4570 // Signed extend and signed comparison.
4571 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
4572 Result = ConstantBool::False;
4574 Result = ConstantBool::True; // X < (large) --> true
4576 // Unsigned extend and signed comparison.
4577 if (cast<ConstantSInt>(CI)->getValue() < 0)
4578 Result = ConstantBool::False;
4580 Result = ConstantBool::True;
4583 // We're performing an unsigned comparison.
4585 // Unsigned extend & compare -> always true.
4586 Result = ConstantBool::True;
4588 // We're performing an unsigned comp with a sign extended value.
4589 // This is true if the input is >= 0. [aka >s -1]
4590 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
4591 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
4592 NegOne, SCI.getName()), SCI);
4596 // Finally, return the value computed.
4597 if (SCI.getOpcode() == Instruction::SetLT) {
4598 return ReplaceInstUsesWith(SCI, Result);
4600 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
4601 if (Constant *CI = dyn_cast<Constant>(Result))
4602 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
4604 return BinaryOperator::createNot(Result);
4611 // Okay, just insert a compare of the reduced operands now!
4612 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
4615 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
4616 assert(I.getOperand(1)->getType() == Type::UByteTy);
4617 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4618 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4620 // shl X, 0 == X and shr X, 0 == X
4621 // shl 0, X == 0 and shr 0, X == 0
4622 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
4623 Op0 == Constant::getNullValue(Op0->getType()))
4624 return ReplaceInstUsesWith(I, Op0);
4626 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
4627 if (!isLeftShift && I.getType()->isSigned())
4628 return ReplaceInstUsesWith(I, Op0);
4629 else // undef << X -> 0 AND undef >>u X -> 0
4630 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4632 if (isa<UndefValue>(Op1)) {
4633 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
4634 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4636 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
4639 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
4641 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
4642 if (CSI->isAllOnesValue())
4643 return ReplaceInstUsesWith(I, CSI);
4645 // Try to fold constant and into select arguments.
4646 if (isa<Constant>(Op0))
4647 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
4648 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4651 // See if we can turn a signed shr into an unsigned shr.
4652 if (I.isArithmeticShift()) {
4653 if (MaskedValueIsZero(Op0,
4654 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
4655 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
4656 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
4658 return new CastInst(V, I.getType());
4662 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
4663 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
4668 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
4670 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4671 bool isSignedShift = Op0->getType()->isSigned();
4672 bool isUnsignedShift = !isSignedShift;
4674 // See if we can simplify any instructions used by the instruction whose sole
4675 // purpose is to compute bits we don't care about.
4676 uint64_t KnownZero, KnownOne;
4677 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
4678 KnownZero, KnownOne))
4681 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
4682 // of a signed value.
4684 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
4685 if (Op1->getValue() >= TypeBits) {
4686 if (isUnsignedShift || isLeftShift)
4687 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
4689 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
4694 // ((X*C1) << C2) == (X * (C1 << C2))
4695 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
4696 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
4697 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
4698 return BinaryOperator::createMul(BO->getOperand(0),
4699 ConstantExpr::getShl(BOOp, Op1));
4701 // Try to fold constant and into select arguments.
4702 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4703 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4705 if (isa<PHINode>(Op0))
4706 if (Instruction *NV = FoldOpIntoPhi(I))
4709 if (Op0->hasOneUse()) {
4710 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
4711 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4714 switch (Op0BO->getOpcode()) {
4716 case Instruction::Add:
4717 case Instruction::And:
4718 case Instruction::Or:
4719 case Instruction::Xor:
4720 // These operators commute.
4721 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
4722 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4723 match(Op0BO->getOperand(1),
4724 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4725 Instruction *YS = new ShiftInst(Instruction::Shl,
4726 Op0BO->getOperand(0), Op1,
4728 InsertNewInstBefore(YS, I); // (Y << C)
4730 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
4731 Op0BO->getOperand(1)->getName());
4732 InsertNewInstBefore(X, I); // (X + (Y << C))
4733 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4734 C2 = ConstantExpr::getShl(C2, Op1);
4735 return BinaryOperator::createAnd(X, C2);
4738 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
4739 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4740 match(Op0BO->getOperand(1),
4741 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4742 m_ConstantInt(CC))) && V2 == Op1 &&
4743 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
4744 Instruction *YS = new ShiftInst(Instruction::Shl,
4745 Op0BO->getOperand(0), Op1,
4747 InsertNewInstBefore(YS, I); // (Y << C)
4749 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4750 V1->getName()+".mask");
4751 InsertNewInstBefore(XM, I); // X & (CC << C)
4753 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
4757 case Instruction::Sub:
4758 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4759 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4760 match(Op0BO->getOperand(0),
4761 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4762 Instruction *YS = new ShiftInst(Instruction::Shl,
4763 Op0BO->getOperand(1), Op1,
4765 InsertNewInstBefore(YS, I); // (Y << C)
4767 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
4768 Op0BO->getOperand(0)->getName());
4769 InsertNewInstBefore(X, I); // (X + (Y << C))
4770 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4771 C2 = ConstantExpr::getShl(C2, Op1);
4772 return BinaryOperator::createAnd(X, C2);
4775 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
4776 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4777 match(Op0BO->getOperand(0),
4778 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4779 m_ConstantInt(CC))) && V2 == Op1 &&
4780 cast<BinaryOperator>(Op0BO->getOperand(0))
4781 ->getOperand(0)->hasOneUse()) {
4782 Instruction *YS = new ShiftInst(Instruction::Shl,
4783 Op0BO->getOperand(1), Op1,
4785 InsertNewInstBefore(YS, I); // (Y << C)
4787 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4788 V1->getName()+".mask");
4789 InsertNewInstBefore(XM, I); // X & (CC << C)
4791 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
4798 // If the operand is an bitwise operator with a constant RHS, and the
4799 // shift is the only use, we can pull it out of the shift.
4800 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
4801 bool isValid = true; // Valid only for And, Or, Xor
4802 bool highBitSet = false; // Transform if high bit of constant set?
4804 switch (Op0BO->getOpcode()) {
4805 default: isValid = false; break; // Do not perform transform!
4806 case Instruction::Add:
4807 isValid = isLeftShift;
4809 case Instruction::Or:
4810 case Instruction::Xor:
4813 case Instruction::And:
4818 // If this is a signed shift right, and the high bit is modified
4819 // by the logical operation, do not perform the transformation.
4820 // The highBitSet boolean indicates the value of the high bit of
4821 // the constant which would cause it to be modified for this
4824 if (isValid && !isLeftShift && isSignedShift) {
4825 uint64_t Val = Op0C->getRawValue();
4826 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
4830 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
4832 Instruction *NewShift =
4833 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
4836 InsertNewInstBefore(NewShift, I);
4838 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
4845 // Find out if this is a shift of a shift by a constant.
4846 ShiftInst *ShiftOp = 0;
4847 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
4849 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4850 // If this is a noop-integer case of a shift instruction, use the shift.
4851 if (CI->getOperand(0)->getType()->isInteger() &&
4852 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
4853 CI->getType()->getPrimitiveSizeInBits() &&
4854 isa<ShiftInst>(CI->getOperand(0))) {
4855 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
4859 if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
4860 // Find the operands and properties of the input shift. Note that the
4861 // signedness of the input shift may differ from the current shift if there
4862 // is a noop cast between the two.
4863 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
4864 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
4865 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
4867 ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
4869 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
4870 unsigned ShiftAmt2 = (unsigned)Op1->getValue();
4872 // Check for (A << c1) << c2 and (A >> c1) >> c2.
4873 if (isLeftShift == isShiftOfLeftShift) {
4874 // Do not fold these shifts if the first one is signed and the second one
4875 // is unsigned and this is a right shift. Further, don't do any folding
4877 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
4880 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
4881 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
4882 Amt = Op0->getType()->getPrimitiveSizeInBits();
4884 Value *Op = ShiftOp->getOperand(0);
4885 if (isShiftOfSignedShift != isSignedShift)
4886 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
4887 return new ShiftInst(I.getOpcode(), Op,
4888 ConstantUInt::get(Type::UByteTy, Amt));
4891 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
4892 // signed types, we can only support the (A >> c1) << c2 configuration,
4893 // because it can not turn an arbitrary bit of A into a sign bit.
4894 if (isUnsignedShift || isLeftShift) {
4895 // Calculate bitmask for what gets shifted off the edge.
4896 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
4898 C = ConstantExpr::getShl(C, ShiftAmt1C);
4900 C = ConstantExpr::getUShr(C, ShiftAmt1C);
4902 Value *Op = ShiftOp->getOperand(0);
4903 if (isShiftOfSignedShift != isSignedShift)
4904 Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
4907 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
4908 InsertNewInstBefore(Mask, I);
4910 // Figure out what flavor of shift we should use...
4911 if (ShiftAmt1 == ShiftAmt2) {
4912 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
4913 } else if (ShiftAmt1 < ShiftAmt2) {
4914 return new ShiftInst(I.getOpcode(), Mask,
4915 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
4916 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
4917 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
4918 // Make sure to emit an unsigned shift right, not a signed one.
4919 Mask = InsertNewInstBefore(new CastInst(Mask,
4920 Mask->getType()->getUnsignedVersion(),
4922 Mask = new ShiftInst(Instruction::Shr, Mask,
4923 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4924 InsertNewInstBefore(Mask, I);
4925 return new CastInst(Mask, I.getType());
4927 return new ShiftInst(ShiftOp->getOpcode(), Mask,
4928 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4931 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
4932 Op = InsertNewInstBefore(new CastInst(Mask,
4933 I.getType()->getSignedVersion(),
4934 Mask->getName()), I);
4935 Instruction *Shift =
4936 new ShiftInst(ShiftOp->getOpcode(), Op,
4937 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4938 InsertNewInstBefore(Shift, I);
4940 C = ConstantIntegral::getAllOnesValue(Shift->getType());
4941 C = ConstantExpr::getShl(C, Op1);
4942 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
4943 InsertNewInstBefore(Mask, I);
4944 return new CastInst(Mask, I.getType());
4947 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
4948 // this case, C1 == C2 and C1 is 8, 16, or 32.
4949 if (ShiftAmt1 == ShiftAmt2) {
4950 const Type *SExtType = 0;
4951 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
4952 case 8 : SExtType = Type::SByteTy; break;
4953 case 16: SExtType = Type::ShortTy; break;
4954 case 32: SExtType = Type::IntTy; break;
4958 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
4960 InsertNewInstBefore(NewTrunc, I);
4961 return new CastInst(NewTrunc, I.getType());
4970 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
4971 /// expression. If so, decompose it, returning some value X, such that Val is
4974 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
4976 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
4977 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
4978 Offset = CI->getValue();
4980 return ConstantUInt::get(Type::UIntTy, 0);
4981 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
4982 if (I->getNumOperands() == 2) {
4983 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
4984 if (I->getOpcode() == Instruction::Shl) {
4985 // This is a value scaled by '1 << the shift amt'.
4986 Scale = 1U << CUI->getValue();
4988 return I->getOperand(0);
4989 } else if (I->getOpcode() == Instruction::Mul) {
4990 // This value is scaled by 'CUI'.
4991 Scale = CUI->getValue();
4993 return I->getOperand(0);
4994 } else if (I->getOpcode() == Instruction::Add) {
4995 // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
4998 Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
5000 Offset += CUI->getValue();
5001 if (SubScale > 1 && (Offset % SubScale == 0)) {
5010 // Otherwise, we can't look past this.
5017 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5018 /// try to eliminate the cast by moving the type information into the alloc.
5019 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5020 AllocationInst &AI) {
5021 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5022 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5024 // Remove any uses of AI that are dead.
5025 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5026 std::vector<Instruction*> DeadUsers;
5027 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5028 Instruction *User = cast<Instruction>(*UI++);
5029 if (isInstructionTriviallyDead(User)) {
5030 while (UI != E && *UI == User)
5031 ++UI; // If this instruction uses AI more than once, don't break UI.
5033 // Add operands to the worklist.
5034 AddUsesToWorkList(*User);
5036 DEBUG(std::cerr << "IC: DCE: " << *User);
5038 User->eraseFromParent();
5039 removeFromWorkList(User);
5043 // Get the type really allocated and the type casted to.
5044 const Type *AllocElTy = AI.getAllocatedType();
5045 const Type *CastElTy = PTy->getElementType();
5046 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5048 unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
5049 unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
5050 if (CastElTyAlign < AllocElTyAlign) return 0;
5052 // If the allocation has multiple uses, only promote it if we are strictly
5053 // increasing the alignment of the resultant allocation. If we keep it the
5054 // same, we open the door to infinite loops of various kinds.
5055 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5057 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5058 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5059 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5061 // See if we can satisfy the modulus by pulling a scale out of the array
5063 unsigned ArraySizeScale, ArrayOffset;
5064 Value *NumElements = // See if the array size is a decomposable linear expr.
5065 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5067 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5069 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5070 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5072 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5077 Amt = ConstantUInt::get(Type::UIntTy, Scale);
5078 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
5079 Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
5080 else if (Scale != 1) {
5081 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5082 Amt = InsertNewInstBefore(Tmp, AI);
5086 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5087 Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
5088 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5089 Amt = InsertNewInstBefore(Tmp, AI);
5092 std::string Name = AI.getName(); AI.setName("");
5093 AllocationInst *New;
5094 if (isa<MallocInst>(AI))
5095 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5097 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5098 InsertNewInstBefore(New, AI);
5100 // If the allocation has multiple uses, insert a cast and change all things
5101 // that used it to use the new cast. This will also hack on CI, but it will
5103 if (!AI.hasOneUse()) {
5104 AddUsesToWorkList(AI);
5105 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
5106 InsertNewInstBefore(NewCast, AI);
5107 AI.replaceAllUsesWith(NewCast);
5109 return ReplaceInstUsesWith(CI, New);
5112 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5113 /// and return it without inserting any new casts. This is used by code that
5114 /// tries to decide whether promoting or shrinking integer operations to wider
5115 /// or smaller types will allow us to eliminate a truncate or extend.
5116 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5117 int &NumCastsRemoved) {
5118 if (isa<Constant>(V)) return true;
5120 Instruction *I = dyn_cast<Instruction>(V);
5121 if (!I || !I->hasOneUse()) return false;
5123 switch (I->getOpcode()) {
5124 case Instruction::And:
5125 case Instruction::Or:
5126 case Instruction::Xor:
5127 // These operators can all arbitrarily be extended or truncated.
5128 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5129 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5130 case Instruction::Cast:
5131 // If this is a cast from the destination type, we can trivially eliminate
5132 // it, and this will remove a cast overall.
5133 if (I->getOperand(0)->getType() == Ty) {
5134 // If the first operand is itself a cast, and is eliminable, do not count
5135 // this as an eliminable cast. We would prefer to eliminate those two
5137 if (CastInst *OpCast = dyn_cast<CastInst>(I->getOperand(0)))
5143 // TODO: Can handle more cases here.
5150 /// EvaluateInDifferentType - Given an expression that
5151 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5152 /// evaluate the expression.
5153 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty) {
5154 if (Constant *C = dyn_cast<Constant>(V))
5155 return ConstantExpr::getCast(C, Ty);
5157 // Otherwise, it must be an instruction.
5158 Instruction *I = cast<Instruction>(V);
5159 Instruction *Res = 0;
5160 switch (I->getOpcode()) {
5161 case Instruction::And:
5162 case Instruction::Or:
5163 case Instruction::Xor: {
5164 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
5165 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty);
5166 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5167 LHS, RHS, I->getName());
5170 case Instruction::Cast:
5171 // If this is a cast from the destination type, return the input.
5172 if (I->getOperand(0)->getType() == Ty)
5173 return I->getOperand(0);
5175 // TODO: Can handle more cases here.
5176 assert(0 && "Unreachable!");
5180 return InsertNewInstBefore(Res, *I);
5184 // CastInst simplification
5186 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
5187 Value *Src = CI.getOperand(0);
5189 // If the user is casting a value to the same type, eliminate this cast
5191 if (CI.getType() == Src->getType())
5192 return ReplaceInstUsesWith(CI, Src);
5194 if (isa<UndefValue>(Src)) // cast undef -> undef
5195 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5197 // If casting the result of another cast instruction, try to eliminate this
5200 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5201 Value *A = CSrc->getOperand(0);
5202 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
5203 CI.getType(), TD)) {
5204 // This instruction now refers directly to the cast's src operand. This
5205 // has a good chance of making CSrc dead.
5206 CI.setOperand(0, CSrc->getOperand(0));
5210 // If this is an A->B->A cast, and we are dealing with integral types, try
5211 // to convert this into a logical 'and' instruction.
5213 if (A->getType()->isInteger() &&
5214 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
5215 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
5216 CSrc->getType()->getPrimitiveSizeInBits() <
5217 CI.getType()->getPrimitiveSizeInBits()&&
5218 A->getType()->getPrimitiveSizeInBits() ==
5219 CI.getType()->getPrimitiveSizeInBits()) {
5220 assert(CSrc->getType() != Type::ULongTy &&
5221 "Cannot have type bigger than ulong!");
5222 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
5223 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
5225 AndOp = ConstantExpr::getCast(AndOp, A->getType());
5226 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
5227 if (And->getType() != CI.getType()) {
5228 And->setName(CSrc->getName()+".mask");
5229 InsertNewInstBefore(And, CI);
5230 And = new CastInst(And, CI.getType());
5236 // If this is a cast to bool, turn it into the appropriate setne instruction.
5237 if (CI.getType() == Type::BoolTy)
5238 return BinaryOperator::createSetNE(CI.getOperand(0),
5239 Constant::getNullValue(CI.getOperand(0)->getType()));
5241 // See if we can simplify any instructions used by the LHS whose sole
5242 // purpose is to compute bits we don't care about.
5243 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral()) {
5244 uint64_t KnownZero, KnownOne;
5245 if (SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask(),
5246 KnownZero, KnownOne))
5250 // If casting the result of a getelementptr instruction with no offset, turn
5251 // this into a cast of the original pointer!
5253 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5254 bool AllZeroOperands = true;
5255 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5256 if (!isa<Constant>(GEP->getOperand(i)) ||
5257 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5258 AllZeroOperands = false;
5261 if (AllZeroOperands) {
5262 CI.setOperand(0, GEP->getOperand(0));
5267 // If we are casting a malloc or alloca to a pointer to a type of the same
5268 // size, rewrite the allocation instruction to allocate the "right" type.
5270 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5271 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5274 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5275 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5277 if (isa<PHINode>(Src))
5278 if (Instruction *NV = FoldOpIntoPhi(CI))
5281 // If the source and destination are pointers, and this cast is equivalent to
5282 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
5283 // This can enhance SROA and other transforms that want type-safe pointers.
5284 if (const PointerType *DstPTy = dyn_cast<PointerType>(CI.getType()))
5285 if (const PointerType *SrcPTy = dyn_cast<PointerType>(Src->getType())) {
5286 const Type *DstTy = DstPTy->getElementType();
5287 const Type *SrcTy = SrcPTy->getElementType();
5289 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
5290 unsigned NumZeros = 0;
5291 while (SrcTy != DstTy &&
5292 isa<CompositeType>(SrcTy) && !isa<PointerType>(SrcTy) &&
5293 SrcTy->getNumContainedTypes() /* not "{}" */) {
5294 SrcTy = cast<CompositeType>(SrcTy)->getTypeAtIndex(ZeroUInt);
5298 // If we found a path from the src to dest, create the getelementptr now.
5299 if (SrcTy == DstTy) {
5300 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
5301 return new GetElementPtrInst(Src, Idxs);
5305 // If the source value is an instruction with only this use, we can attempt to
5306 // propagate the cast into the instruction. Also, only handle integral types
5308 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
5309 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
5310 CI.getType()->isInteger()) { // Don't mess with casts to bool here
5312 int NumCastsRemoved = 0;
5313 if (CanEvaluateInDifferentType(SrcI, CI.getType(), NumCastsRemoved)) {
5314 // If this cast is a truncate, evaluting in a different type always
5315 // eliminates the cast, so it is always a win. If this is a noop-cast
5316 // this just removes a noop cast which isn't pointful, but simplifies
5317 // the code. If this is a zero-extension, we need to do an AND to
5318 // maintain the clear top-part of the computation, so we require that
5319 // the input have eliminated at least one cast. If this is a sign
5320 // extension, we insert two new casts (to do the extension) so we
5321 // require that two casts have been eliminated.
5323 switch (getCastType(Src->getType(), CI.getType())) {
5324 default: assert(0 && "Unknown cast type!");
5330 DoXForm = NumCastsRemoved >= 1;
5333 DoXForm = NumCastsRemoved >= 2;
5338 Value *Res = EvaluateInDifferentType(SrcI, CI.getType());
5339 assert(Res->getType() == CI.getType());
5340 switch (getCastType(Src->getType(), CI.getType())) {
5341 default: assert(0 && "Unknown cast type!");
5344 // Just replace this cast with the result.
5345 return ReplaceInstUsesWith(CI, Res);
5347 // We need to emit an AND to clear the high bits.
5348 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5349 unsigned DestBitSize = CI.getType()->getPrimitiveSizeInBits();
5350 assert(SrcBitSize < DestBitSize && "Not a zext?");
5351 Constant *C = ConstantUInt::get(Type::ULongTy, (1 << SrcBitSize)-1);
5352 C = ConstantExpr::getCast(C, CI.getType());
5353 return BinaryOperator::createAnd(Res, C);
5356 // We need to emit a cast to truncate, then a cast to sext.
5357 return new CastInst(InsertCastBefore(Res, Src->getType(), CI),
5363 const Type *DestTy = CI.getType();
5364 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5365 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5367 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
5368 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
5370 switch (SrcI->getOpcode()) {
5371 case Instruction::Add:
5372 case Instruction::Mul:
5373 case Instruction::And:
5374 case Instruction::Or:
5375 case Instruction::Xor:
5376 // If we are discarding information, or just changing the sign, rewrite.
5377 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
5378 // Don't insert two casts if they cannot be eliminated. We allow two
5379 // casts to be inserted if the sizes are the same. This could only be
5380 // converting signedness, which is a noop.
5381 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
5382 !ValueRequiresCast(Op0, DestTy, TD)) {
5383 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5384 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5385 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
5386 ->getOpcode(), Op0c, Op1c);
5390 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
5391 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
5392 Op1 == ConstantBool::True &&
5393 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
5394 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
5395 return BinaryOperator::createXor(New,
5396 ConstantInt::get(CI.getType(), 1));
5399 case Instruction::Shl:
5400 // Allow changing the sign of the source operand. Do not allow changing
5401 // the size of the shift, UNLESS the shift amount is a constant. We
5402 // mush not change variable sized shifts to a smaller size, because it
5403 // is undefined to shift more bits out than exist in the value.
5404 if (DestBitSize == SrcBitSize ||
5405 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
5406 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5407 return new ShiftInst(Instruction::Shl, Op0c, Op1);
5410 case Instruction::Shr:
5411 // If this is a signed shr, and if all bits shifted in are about to be
5412 // truncated off, turn it into an unsigned shr to allow greater
5414 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
5415 isa<ConstantInt>(Op1)) {
5416 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
5417 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
5418 // Convert to unsigned.
5419 Value *N1 = InsertOperandCastBefore(Op0,
5420 Op0->getType()->getUnsignedVersion(), &CI);
5421 // Insert the new shift, which is now unsigned.
5422 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
5423 Op1, Src->getName()), CI);
5424 return new CastInst(N1, CI.getType());
5429 case Instruction::SetEQ:
5430 case Instruction::SetNE:
5431 // We if we are just checking for a seteq of a single bit and casting it
5432 // to an integer. If so, shift the bit to the appropriate place then
5433 // cast to integer to avoid the comparison.
5434 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
5435 uint64_t Op1CV = Op1C->getZExtValue();
5436 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
5437 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5438 // cast (X == 1) to int --> X iff X has only the low bit set.
5439 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
5440 // cast (X != 0) to int --> X iff X has only the low bit set.
5441 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
5442 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
5443 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5444 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
5445 // If Op1C some other power of two, convert:
5446 uint64_t KnownZero, KnownOne;
5447 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
5448 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
5450 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly one possible 1?
5451 bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
5452 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
5453 // (X&4) == 2 --> false
5454 // (X&4) != 2 --> true
5455 Constant *Res = ConstantBool::get(isSetNE);
5456 Res = ConstantExpr::getCast(Res, CI.getType());
5457 return ReplaceInstUsesWith(CI, Res);
5460 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
5463 // Perform an unsigned shr by shiftamt. Convert input to
5464 // unsigned if it is signed.
5465 if (In->getType()->isSigned())
5466 In = InsertNewInstBefore(new CastInst(In,
5467 In->getType()->getUnsignedVersion(), In->getName()),CI);
5468 // Insert the shift to put the result in the low bit.
5469 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
5470 ConstantInt::get(Type::UByteTy, ShiftAmt),
5471 In->getName()+".lobit"), CI);
5474 if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
5475 Constant *One = ConstantInt::get(In->getType(), 1);
5476 In = BinaryOperator::createXor(In, One, "tmp");
5477 InsertNewInstBefore(cast<Instruction>(In), CI);
5480 if (CI.getType() == In->getType())
5481 return ReplaceInstUsesWith(CI, In);
5483 return new CastInst(In, CI.getType());
5491 if (SrcI->hasOneUse()) {
5492 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(SrcI)) {
5493 // Okay, we have (cast (shuffle ..)). We know this cast is a bitconvert
5494 // because the inputs are known to be a vector. Check to see if this is
5495 // a cast to a vector with the same # elts.
5496 if (isa<PackedType>(CI.getType()) &&
5497 cast<PackedType>(CI.getType())->getNumElements() ==
5498 SVI->getType()->getNumElements()) {
5500 // If either of the operands is a cast from CI.getType(), then
5501 // evaluating the shuffle in the casted destination's type will allow
5502 // us to eliminate at least one cast.
5503 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
5504 Tmp->getOperand(0)->getType() == CI.getType()) ||
5505 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
5506 Tmp->getOperand(0)->getType() == CI.getType())) {
5507 Value *LHS = InsertOperandCastBefore(SVI->getOperand(0),
5509 Value *RHS = InsertOperandCastBefore(SVI->getOperand(1),
5511 // Return a new shuffle vector. Use the same element ID's, as we
5512 // know the vector types match #elts.
5513 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
5523 /// GetSelectFoldableOperands - We want to turn code that looks like this:
5525 /// %D = select %cond, %C, %A
5527 /// %C = select %cond, %B, 0
5530 /// Assuming that the specified instruction is an operand to the select, return
5531 /// a bitmask indicating which operands of this instruction are foldable if they
5532 /// equal the other incoming value of the select.
5534 static unsigned GetSelectFoldableOperands(Instruction *I) {
5535 switch (I->getOpcode()) {
5536 case Instruction::Add:
5537 case Instruction::Mul:
5538 case Instruction::And:
5539 case Instruction::Or:
5540 case Instruction::Xor:
5541 return 3; // Can fold through either operand.
5542 case Instruction::Sub: // Can only fold on the amount subtracted.
5543 case Instruction::Shl: // Can only fold on the shift amount.
5544 case Instruction::Shr:
5547 return 0; // Cannot fold
5551 /// GetSelectFoldableConstant - For the same transformation as the previous
5552 /// function, return the identity constant that goes into the select.
5553 static Constant *GetSelectFoldableConstant(Instruction *I) {
5554 switch (I->getOpcode()) {
5555 default: assert(0 && "This cannot happen!"); abort();
5556 case Instruction::Add:
5557 case Instruction::Sub:
5558 case Instruction::Or:
5559 case Instruction::Xor:
5560 return Constant::getNullValue(I->getType());
5561 case Instruction::Shl:
5562 case Instruction::Shr:
5563 return Constant::getNullValue(Type::UByteTy);
5564 case Instruction::And:
5565 return ConstantInt::getAllOnesValue(I->getType());
5566 case Instruction::Mul:
5567 return ConstantInt::get(I->getType(), 1);
5571 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
5572 /// have the same opcode and only one use each. Try to simplify this.
5573 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
5575 if (TI->getNumOperands() == 1) {
5576 // If this is a non-volatile load or a cast from the same type,
5578 if (TI->getOpcode() == Instruction::Cast) {
5579 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
5582 return 0; // unknown unary op.
5585 // Fold this by inserting a select from the input values.
5586 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
5587 FI->getOperand(0), SI.getName()+".v");
5588 InsertNewInstBefore(NewSI, SI);
5589 return new CastInst(NewSI, TI->getType());
5592 // Only handle binary operators here.
5593 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
5596 // Figure out if the operations have any operands in common.
5597 Value *MatchOp, *OtherOpT, *OtherOpF;
5599 if (TI->getOperand(0) == FI->getOperand(0)) {
5600 MatchOp = TI->getOperand(0);
5601 OtherOpT = TI->getOperand(1);
5602 OtherOpF = FI->getOperand(1);
5603 MatchIsOpZero = true;
5604 } else if (TI->getOperand(1) == FI->getOperand(1)) {
5605 MatchOp = TI->getOperand(1);
5606 OtherOpT = TI->getOperand(0);
5607 OtherOpF = FI->getOperand(0);
5608 MatchIsOpZero = false;
5609 } else if (!TI->isCommutative()) {
5611 } else if (TI->getOperand(0) == FI->getOperand(1)) {
5612 MatchOp = TI->getOperand(0);
5613 OtherOpT = TI->getOperand(1);
5614 OtherOpF = FI->getOperand(0);
5615 MatchIsOpZero = true;
5616 } else if (TI->getOperand(1) == FI->getOperand(0)) {
5617 MatchOp = TI->getOperand(1);
5618 OtherOpT = TI->getOperand(0);
5619 OtherOpF = FI->getOperand(1);
5620 MatchIsOpZero = true;
5625 // If we reach here, they do have operations in common.
5626 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
5627 OtherOpF, SI.getName()+".v");
5628 InsertNewInstBefore(NewSI, SI);
5630 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
5632 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
5634 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
5637 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
5639 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
5643 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
5644 Value *CondVal = SI.getCondition();
5645 Value *TrueVal = SI.getTrueValue();
5646 Value *FalseVal = SI.getFalseValue();
5648 // select true, X, Y -> X
5649 // select false, X, Y -> Y
5650 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
5651 if (C == ConstantBool::True)
5652 return ReplaceInstUsesWith(SI, TrueVal);
5654 assert(C == ConstantBool::False);
5655 return ReplaceInstUsesWith(SI, FalseVal);
5658 // select C, X, X -> X
5659 if (TrueVal == FalseVal)
5660 return ReplaceInstUsesWith(SI, TrueVal);
5662 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
5663 return ReplaceInstUsesWith(SI, FalseVal);
5664 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
5665 return ReplaceInstUsesWith(SI, TrueVal);
5666 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
5667 if (isa<Constant>(TrueVal))
5668 return ReplaceInstUsesWith(SI, TrueVal);
5670 return ReplaceInstUsesWith(SI, FalseVal);
5673 if (SI.getType() == Type::BoolTy)
5674 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
5675 if (C == ConstantBool::True) {
5676 // Change: A = select B, true, C --> A = or B, C
5677 return BinaryOperator::createOr(CondVal, FalseVal);
5679 // Change: A = select B, false, C --> A = and !B, C
5681 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5682 "not."+CondVal->getName()), SI);
5683 return BinaryOperator::createAnd(NotCond, FalseVal);
5685 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
5686 if (C == ConstantBool::False) {
5687 // Change: A = select B, C, false --> A = and B, C
5688 return BinaryOperator::createAnd(CondVal, TrueVal);
5690 // Change: A = select B, C, true --> A = or !B, C
5692 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5693 "not."+CondVal->getName()), SI);
5694 return BinaryOperator::createOr(NotCond, TrueVal);
5698 // Selecting between two integer constants?
5699 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
5700 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
5701 // select C, 1, 0 -> cast C to int
5702 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
5703 return new CastInst(CondVal, SI.getType());
5704 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
5705 // select C, 0, 1 -> cast !C to int
5707 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5708 "not."+CondVal->getName()), SI);
5709 return new CastInst(NotCond, SI.getType());
5712 if (SetCondInst *IC = dyn_cast<SetCondInst>(SI.getCondition())) {
5714 // (x <s 0) ? -1 : 0 -> sra x, 31
5715 // (x >u 2147483647) ? -1 : 0 -> sra x, 31
5716 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
5717 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
5718 bool CanXForm = false;
5719 if (CmpCst->getType()->isSigned())
5720 CanXForm = CmpCst->isNullValue() &&
5721 IC->getOpcode() == Instruction::SetLT;
5723 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
5724 CanXForm = (CmpCst->getRawValue() == ~0ULL >> (64-Bits+1)) &&
5725 IC->getOpcode() == Instruction::SetGT;
5729 // The comparison constant and the result are not neccessarily the
5730 // same width. In any case, the first step to do is make sure
5731 // that X is signed.
5732 Value *X = IC->getOperand(0);
5733 if (!X->getType()->isSigned())
5734 X = InsertCastBefore(X, X->getType()->getSignedVersion(), SI);
5736 // Now that X is signed, we have to make the all ones value. Do
5737 // this by inserting a new SRA.
5738 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
5739 Constant *ShAmt = ConstantUInt::get(Type::UByteTy, Bits-1);
5740 Instruction *SRA = new ShiftInst(Instruction::Shr, X,
5742 InsertNewInstBefore(SRA, SI);
5744 // Finally, convert to the type of the select RHS. If this is
5745 // smaller than the compare value, it will truncate the ones to
5746 // fit. If it is larger, it will sext the ones to fit.
5747 return new CastInst(SRA, SI.getType());
5752 // If one of the constants is zero (we know they can't both be) and we
5753 // have a setcc instruction with zero, and we have an 'and' with the
5754 // non-constant value, eliminate this whole mess. This corresponds to
5755 // cases like this: ((X & 27) ? 27 : 0)
5756 if (TrueValC->isNullValue() || FalseValC->isNullValue())
5757 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
5758 cast<Constant>(IC->getOperand(1))->isNullValue())
5759 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
5760 if (ICA->getOpcode() == Instruction::And &&
5761 isa<ConstantInt>(ICA->getOperand(1)) &&
5762 (ICA->getOperand(1) == TrueValC ||
5763 ICA->getOperand(1) == FalseValC) &&
5764 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
5765 // Okay, now we know that everything is set up, we just don't
5766 // know whether we have a setne or seteq and whether the true or
5767 // false val is the zero.
5768 bool ShouldNotVal = !TrueValC->isNullValue();
5769 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
5772 V = InsertNewInstBefore(BinaryOperator::create(
5773 Instruction::Xor, V, ICA->getOperand(1)), SI);
5774 return ReplaceInstUsesWith(SI, V);
5779 // See if we are selecting two values based on a comparison of the two values.
5780 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
5781 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
5782 // Transform (X == Y) ? X : Y -> Y
5783 if (SCI->getOpcode() == Instruction::SetEQ)
5784 return ReplaceInstUsesWith(SI, FalseVal);
5785 // Transform (X != Y) ? X : Y -> X
5786 if (SCI->getOpcode() == Instruction::SetNE)
5787 return ReplaceInstUsesWith(SI, TrueVal);
5788 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5790 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
5791 // Transform (X == Y) ? Y : X -> X
5792 if (SCI->getOpcode() == Instruction::SetEQ)
5793 return ReplaceInstUsesWith(SI, FalseVal);
5794 // Transform (X != Y) ? Y : X -> Y
5795 if (SCI->getOpcode() == Instruction::SetNE)
5796 return ReplaceInstUsesWith(SI, TrueVal);
5797 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5801 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
5802 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
5803 if (TI->hasOneUse() && FI->hasOneUse()) {
5804 bool isInverse = false;
5805 Instruction *AddOp = 0, *SubOp = 0;
5807 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
5808 if (TI->getOpcode() == FI->getOpcode())
5809 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
5812 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
5813 // even legal for FP.
5814 if (TI->getOpcode() == Instruction::Sub &&
5815 FI->getOpcode() == Instruction::Add) {
5816 AddOp = FI; SubOp = TI;
5817 } else if (FI->getOpcode() == Instruction::Sub &&
5818 TI->getOpcode() == Instruction::Add) {
5819 AddOp = TI; SubOp = FI;
5823 Value *OtherAddOp = 0;
5824 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
5825 OtherAddOp = AddOp->getOperand(1);
5826 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
5827 OtherAddOp = AddOp->getOperand(0);
5831 // So at this point we know we have (Y -> OtherAddOp):
5832 // select C, (add X, Y), (sub X, Z)
5833 Value *NegVal; // Compute -Z
5834 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
5835 NegVal = ConstantExpr::getNeg(C);
5837 NegVal = InsertNewInstBefore(
5838 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
5841 Value *NewTrueOp = OtherAddOp;
5842 Value *NewFalseOp = NegVal;
5844 std::swap(NewTrueOp, NewFalseOp);
5845 Instruction *NewSel =
5846 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
5848 NewSel = InsertNewInstBefore(NewSel, SI);
5849 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
5854 // See if we can fold the select into one of our operands.
5855 if (SI.getType()->isInteger()) {
5856 // See the comment above GetSelectFoldableOperands for a description of the
5857 // transformation we are doing here.
5858 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
5859 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
5860 !isa<Constant>(FalseVal))
5861 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
5862 unsigned OpToFold = 0;
5863 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
5865 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
5870 Constant *C = GetSelectFoldableConstant(TVI);
5871 std::string Name = TVI->getName(); TVI->setName("");
5872 Instruction *NewSel =
5873 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
5875 InsertNewInstBefore(NewSel, SI);
5876 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
5877 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
5878 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
5879 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
5881 assert(0 && "Unknown instruction!!");
5886 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
5887 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
5888 !isa<Constant>(TrueVal))
5889 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
5890 unsigned OpToFold = 0;
5891 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
5893 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
5898 Constant *C = GetSelectFoldableConstant(FVI);
5899 std::string Name = FVI->getName(); FVI->setName("");
5900 Instruction *NewSel =
5901 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
5903 InsertNewInstBefore(NewSel, SI);
5904 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
5905 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
5906 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
5907 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
5909 assert(0 && "Unknown instruction!!");
5915 if (BinaryOperator::isNot(CondVal)) {
5916 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
5917 SI.setOperand(1, FalseVal);
5918 SI.setOperand(2, TrueVal);
5925 /// GetKnownAlignment - If the specified pointer has an alignment that we can
5926 /// determine, return it, otherwise return 0.
5927 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
5928 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
5929 unsigned Align = GV->getAlignment();
5930 if (Align == 0 && TD)
5931 Align = TD->getTypeAlignment(GV->getType()->getElementType());
5933 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
5934 unsigned Align = AI->getAlignment();
5935 if (Align == 0 && TD) {
5936 if (isa<AllocaInst>(AI))
5937 Align = TD->getTypeAlignment(AI->getType()->getElementType());
5938 else if (isa<MallocInst>(AI)) {
5939 // Malloc returns maximally aligned memory.
5940 Align = TD->getTypeAlignment(AI->getType()->getElementType());
5941 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
5942 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::LongTy));
5946 } else if (isa<CastInst>(V) ||
5947 (isa<ConstantExpr>(V) &&
5948 cast<ConstantExpr>(V)->getOpcode() == Instruction::Cast)) {
5949 User *CI = cast<User>(V);
5950 if (isa<PointerType>(CI->getOperand(0)->getType()))
5951 return GetKnownAlignment(CI->getOperand(0), TD);
5953 } else if (isa<GetElementPtrInst>(V) ||
5954 (isa<ConstantExpr>(V) &&
5955 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
5956 User *GEPI = cast<User>(V);
5957 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
5958 if (BaseAlignment == 0) return 0;
5960 // If all indexes are zero, it is just the alignment of the base pointer.
5961 bool AllZeroOperands = true;
5962 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
5963 if (!isa<Constant>(GEPI->getOperand(i)) ||
5964 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
5965 AllZeroOperands = false;
5968 if (AllZeroOperands)
5969 return BaseAlignment;
5971 // Otherwise, if the base alignment is >= the alignment we expect for the
5972 // base pointer type, then we know that the resultant pointer is aligned at
5973 // least as much as its type requires.
5976 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
5977 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
5979 const Type *GEPTy = GEPI->getType();
5980 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
5988 /// visitCallInst - CallInst simplification. This mostly only handles folding
5989 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
5990 /// the heavy lifting.
5992 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
5993 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
5994 if (!II) return visitCallSite(&CI);
5996 // Intrinsics cannot occur in an invoke, so handle them here instead of in
5998 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
5999 bool Changed = false;
6001 // memmove/cpy/set of zero bytes is a noop.
6002 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6003 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6005 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6006 if (CI->getRawValue() == 1) {
6007 // Replace the instruction with just byte operations. We would
6008 // transform other cases to loads/stores, but we don't know if
6009 // alignment is sufficient.
6013 // If we have a memmove and the source operation is a constant global,
6014 // then the source and dest pointers can't alias, so we can change this
6015 // into a call to memcpy.
6016 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6017 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6018 if (GVSrc->isConstant()) {
6019 Module *M = CI.getParent()->getParent()->getParent();
6021 if (CI.getCalledFunction()->getFunctionType()->getParamType(3) ==
6023 Name = "llvm.memcpy.i32";
6025 Name = "llvm.memcpy.i64";
6026 Function *MemCpy = M->getOrInsertFunction(Name,
6027 CI.getCalledFunction()->getFunctionType());
6028 CI.setOperand(0, MemCpy);
6033 // If we can determine a pointer alignment that is bigger than currently
6034 // set, update the alignment.
6035 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6036 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6037 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6038 unsigned Align = std::min(Alignment1, Alignment2);
6039 if (MI->getAlignment()->getRawValue() < Align) {
6040 MI->setAlignment(ConstantUInt::get(Type::UIntTy, Align));
6043 } else if (isa<MemSetInst>(MI)) {
6044 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6045 if (MI->getAlignment()->getRawValue() < Alignment) {
6046 MI->setAlignment(ConstantUInt::get(Type::UIntTy, Alignment));
6051 if (Changed) return II;
6053 switch (II->getIntrinsicID()) {
6055 case Intrinsic::ppc_altivec_lvx:
6056 case Intrinsic::ppc_altivec_lvxl:
6057 case Intrinsic::x86_sse_loadu_ps:
6058 case Intrinsic::x86_sse2_loadu_pd:
6059 case Intrinsic::x86_sse2_loadu_dq:
6060 // Turn PPC lvx -> load if the pointer is known aligned.
6061 // Turn X86 loadups -> load if the pointer is known aligned.
6062 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6063 Value *Ptr = InsertCastBefore(II->getOperand(1),
6064 PointerType::get(II->getType()), CI);
6065 return new LoadInst(Ptr);
6068 case Intrinsic::ppc_altivec_stvx:
6069 case Intrinsic::ppc_altivec_stvxl:
6070 // Turn stvx -> store if the pointer is known aligned.
6071 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
6072 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
6073 Value *Ptr = InsertCastBefore(II->getOperand(2), OpPtrTy, CI);
6074 return new StoreInst(II->getOperand(1), Ptr);
6077 case Intrinsic::x86_sse_storeu_ps:
6078 case Intrinsic::x86_sse2_storeu_pd:
6079 case Intrinsic::x86_sse2_storeu_dq:
6080 case Intrinsic::x86_sse2_storel_dq:
6081 // Turn X86 storeu -> store if the pointer is known aligned.
6082 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6083 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
6084 Value *Ptr = InsertCastBefore(II->getOperand(1), OpPtrTy, CI);
6085 return new StoreInst(II->getOperand(2), Ptr);
6088 case Intrinsic::ppc_altivec_vperm:
6089 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
6090 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
6091 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
6093 // Check that all of the elements are integer constants or undefs.
6094 bool AllEltsOk = true;
6095 for (unsigned i = 0; i != 16; ++i) {
6096 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
6097 !isa<UndefValue>(Mask->getOperand(i))) {
6104 // Cast the input vectors to byte vectors.
6105 Value *Op0 = InsertCastBefore(II->getOperand(1), Mask->getType(), CI);
6106 Value *Op1 = InsertCastBefore(II->getOperand(2), Mask->getType(), CI);
6107 Value *Result = UndefValue::get(Op0->getType());
6109 // Only extract each element once.
6110 Value *ExtractedElts[32];
6111 memset(ExtractedElts, 0, sizeof(ExtractedElts));
6113 for (unsigned i = 0; i != 16; ++i) {
6114 if (isa<UndefValue>(Mask->getOperand(i)))
6116 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getRawValue();
6117 Idx &= 31; // Match the hardware behavior.
6119 if (ExtractedElts[Idx] == 0) {
6121 new ExtractElementInst(Idx < 16 ? Op0 : Op1,
6122 ConstantUInt::get(Type::UIntTy, Idx&15),
6124 InsertNewInstBefore(Elt, CI);
6125 ExtractedElts[Idx] = Elt;
6128 // Insert this value into the result vector.
6129 Result = new InsertElementInst(Result, ExtractedElts[Idx],
6130 ConstantUInt::get(Type::UIntTy, i),
6132 InsertNewInstBefore(cast<Instruction>(Result), CI);
6134 return new CastInst(Result, CI.getType());
6139 case Intrinsic::stackrestore: {
6140 // If the save is right next to the restore, remove the restore. This can
6141 // happen when variable allocas are DCE'd.
6142 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
6143 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
6144 BasicBlock::iterator BI = SS;
6146 return EraseInstFromFunction(CI);
6150 // If the stack restore is in a return/unwind block and if there are no
6151 // allocas or calls between the restore and the return, nuke the restore.
6152 TerminatorInst *TI = II->getParent()->getTerminator();
6153 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
6154 BasicBlock::iterator BI = II;
6155 bool CannotRemove = false;
6156 for (++BI; &*BI != TI; ++BI) {
6157 if (isa<AllocaInst>(BI) ||
6158 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
6159 CannotRemove = true;
6164 return EraseInstFromFunction(CI);
6171 return visitCallSite(II);
6174 // InvokeInst simplification
6176 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
6177 return visitCallSite(&II);
6180 // visitCallSite - Improvements for call and invoke instructions.
6182 Instruction *InstCombiner::visitCallSite(CallSite CS) {
6183 bool Changed = false;
6185 // If the callee is a constexpr cast of a function, attempt to move the cast
6186 // to the arguments of the call/invoke.
6187 if (transformConstExprCastCall(CS)) return 0;
6189 Value *Callee = CS.getCalledValue();
6191 if (Function *CalleeF = dyn_cast<Function>(Callee))
6192 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
6193 Instruction *OldCall = CS.getInstruction();
6194 // If the call and callee calling conventions don't match, this call must
6195 // be unreachable, as the call is undefined.
6196 new StoreInst(ConstantBool::True,
6197 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
6198 if (!OldCall->use_empty())
6199 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
6200 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
6201 return EraseInstFromFunction(*OldCall);
6205 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
6206 // This instruction is not reachable, just remove it. We insert a store to
6207 // undef so that we know that this code is not reachable, despite the fact
6208 // that we can't modify the CFG here.
6209 new StoreInst(ConstantBool::True,
6210 UndefValue::get(PointerType::get(Type::BoolTy)),
6211 CS.getInstruction());
6213 if (!CS.getInstruction()->use_empty())
6214 CS.getInstruction()->
6215 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
6217 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
6218 // Don't break the CFG, insert a dummy cond branch.
6219 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
6220 ConstantBool::True, II);
6222 return EraseInstFromFunction(*CS.getInstruction());
6225 const PointerType *PTy = cast<PointerType>(Callee->getType());
6226 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
6227 if (FTy->isVarArg()) {
6228 // See if we can optimize any arguments passed through the varargs area of
6230 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
6231 E = CS.arg_end(); I != E; ++I)
6232 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
6233 // If this cast does not effect the value passed through the varargs
6234 // area, we can eliminate the use of the cast.
6235 Value *Op = CI->getOperand(0);
6236 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
6243 return Changed ? CS.getInstruction() : 0;
6246 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
6247 // attempt to move the cast to the arguments of the call/invoke.
6249 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
6250 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
6251 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
6252 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
6254 Function *Callee = cast<Function>(CE->getOperand(0));
6255 Instruction *Caller = CS.getInstruction();
6257 // Okay, this is a cast from a function to a different type. Unless doing so
6258 // would cause a type conversion of one of our arguments, change this call to
6259 // be a direct call with arguments casted to the appropriate types.
6261 const FunctionType *FT = Callee->getFunctionType();
6262 const Type *OldRetTy = Caller->getType();
6264 // Check to see if we are changing the return type...
6265 if (OldRetTy != FT->getReturnType()) {
6266 if (Callee->isExternal() &&
6267 !(OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) ||
6268 (isa<PointerType>(FT->getReturnType()) &&
6269 TD->getIntPtrType()->isLosslesslyConvertibleTo(OldRetTy)))
6270 && !Caller->use_empty())
6271 return false; // Cannot transform this return value...
6273 // If the callsite is an invoke instruction, and the return value is used by
6274 // a PHI node in a successor, we cannot change the return type of the call
6275 // because there is no place to put the cast instruction (without breaking
6276 // the critical edge). Bail out in this case.
6277 if (!Caller->use_empty())
6278 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
6279 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
6281 if (PHINode *PN = dyn_cast<PHINode>(*UI))
6282 if (PN->getParent() == II->getNormalDest() ||
6283 PN->getParent() == II->getUnwindDest())
6287 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
6288 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
6290 CallSite::arg_iterator AI = CS.arg_begin();
6291 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
6292 const Type *ParamTy = FT->getParamType(i);
6293 const Type *ActTy = (*AI)->getType();
6294 ConstantSInt* c = dyn_cast<ConstantSInt>(*AI);
6295 //Either we can cast directly, or we can upconvert the argument
6296 bool isConvertible = ActTy->isLosslesslyConvertibleTo(ParamTy) ||
6297 (ParamTy->isIntegral() && ActTy->isIntegral() &&
6298 ParamTy->isSigned() == ActTy->isSigned() &&
6299 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
6300 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
6302 if (Callee->isExternal() && !isConvertible) return false;
6305 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
6306 Callee->isExternal())
6307 return false; // Do not delete arguments unless we have a function body...
6309 // Okay, we decided that this is a safe thing to do: go ahead and start
6310 // inserting cast instructions as necessary...
6311 std::vector<Value*> Args;
6312 Args.reserve(NumActualArgs);
6314 AI = CS.arg_begin();
6315 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
6316 const Type *ParamTy = FT->getParamType(i);
6317 if ((*AI)->getType() == ParamTy) {
6318 Args.push_back(*AI);
6320 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
6325 // If the function takes more arguments than the call was taking, add them
6327 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
6328 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
6330 // If we are removing arguments to the function, emit an obnoxious warning...
6331 if (FT->getNumParams() < NumActualArgs)
6332 if (!FT->isVarArg()) {
6333 std::cerr << "WARNING: While resolving call to function '"
6334 << Callee->getName() << "' arguments were dropped!\n";
6336 // Add all of the arguments in their promoted form to the arg list...
6337 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
6338 const Type *PTy = getPromotedType((*AI)->getType());
6339 if (PTy != (*AI)->getType()) {
6340 // Must promote to pass through va_arg area!
6341 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
6342 InsertNewInstBefore(Cast, *Caller);
6343 Args.push_back(Cast);
6345 Args.push_back(*AI);
6350 if (FT->getReturnType() == Type::VoidTy)
6351 Caller->setName(""); // Void type should not have a name...
6354 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6355 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
6356 Args, Caller->getName(), Caller);
6357 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
6359 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
6360 if (cast<CallInst>(Caller)->isTailCall())
6361 cast<CallInst>(NC)->setTailCall();
6362 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
6365 // Insert a cast of the return type as necessary...
6367 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
6368 if (NV->getType() != Type::VoidTy) {
6369 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
6371 // If this is an invoke instruction, we should insert it after the first
6372 // non-phi, instruction in the normal successor block.
6373 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6374 BasicBlock::iterator I = II->getNormalDest()->begin();
6375 while (isa<PHINode>(I)) ++I;
6376 InsertNewInstBefore(NC, *I);
6378 // Otherwise, it's a call, just insert cast right after the call instr
6379 InsertNewInstBefore(NC, *Caller);
6381 AddUsersToWorkList(*Caller);
6383 NV = UndefValue::get(Caller->getType());
6387 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
6388 Caller->replaceAllUsesWith(NV);
6389 Caller->getParent()->getInstList().erase(Caller);
6390 removeFromWorkList(Caller);
6395 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
6396 // operator and they all are only used by the PHI, PHI together their
6397 // inputs, and do the operation once, to the result of the PHI.
6398 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
6399 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6401 // Scan the instruction, looking for input operations that can be folded away.
6402 // If all input operands to the phi are the same instruction (e.g. a cast from
6403 // the same type or "+42") we can pull the operation through the PHI, reducing
6404 // code size and simplifying code.
6405 Constant *ConstantOp = 0;
6406 const Type *CastSrcTy = 0;
6407 if (isa<CastInst>(FirstInst)) {
6408 CastSrcTy = FirstInst->getOperand(0)->getType();
6409 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
6410 // Can fold binop or shift if the RHS is a constant.
6411 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
6412 if (ConstantOp == 0) return 0;
6414 return 0; // Cannot fold this operation.
6417 // Check to see if all arguments are the same operation.
6418 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6419 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
6420 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
6421 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
6424 if (I->getOperand(0)->getType() != CastSrcTy)
6425 return 0; // Cast operation must match.
6426 } else if (I->getOperand(1) != ConstantOp) {
6431 // Okay, they are all the same operation. Create a new PHI node of the
6432 // correct type, and PHI together all of the LHS's of the instructions.
6433 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
6434 PN.getName()+".in");
6435 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
6437 Value *InVal = FirstInst->getOperand(0);
6438 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
6440 // Add all operands to the new PHI.
6441 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6442 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
6443 if (NewInVal != InVal)
6445 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
6450 // The new PHI unions all of the same values together. This is really
6451 // common, so we handle it intelligently here for compile-time speed.
6455 InsertNewInstBefore(NewPN, PN);
6459 // Insert and return the new operation.
6460 if (isa<CastInst>(FirstInst))
6461 return new CastInst(PhiVal, PN.getType());
6462 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6463 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
6465 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
6466 PhiVal, ConstantOp);
6469 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
6471 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
6472 if (PN->use_empty()) return true;
6473 if (!PN->hasOneUse()) return false;
6475 // Remember this node, and if we find the cycle, return.
6476 if (!PotentiallyDeadPHIs.insert(PN).second)
6479 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
6480 return DeadPHICycle(PU, PotentiallyDeadPHIs);
6485 // PHINode simplification
6487 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
6488 // If LCSSA is around, don't mess with Phi nodes
6489 if (mustPreserveAnalysisID(LCSSAID)) return 0;
6491 if (Value *V = PN.hasConstantValue())
6492 return ReplaceInstUsesWith(PN, V);
6494 // If the only user of this instruction is a cast instruction, and all of the
6495 // incoming values are constants, change this PHI to merge together the casted
6498 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
6499 if (CI->getType() != PN.getType()) { // noop casts will be folded
6500 bool AllConstant = true;
6501 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
6502 if (!isa<Constant>(PN.getIncomingValue(i))) {
6503 AllConstant = false;
6507 // Make a new PHI with all casted values.
6508 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
6509 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
6510 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
6511 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
6512 PN.getIncomingBlock(i));
6515 // Update the cast instruction.
6516 CI->setOperand(0, New);
6517 WorkList.push_back(CI); // revisit the cast instruction to fold.
6518 WorkList.push_back(New); // Make sure to revisit the new Phi
6519 return &PN; // PN is now dead!
6523 // If all PHI operands are the same operation, pull them through the PHI,
6524 // reducing code size.
6525 if (isa<Instruction>(PN.getIncomingValue(0)) &&
6526 PN.getIncomingValue(0)->hasOneUse())
6527 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
6530 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
6531 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
6532 // PHI)... break the cycle.
6534 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
6535 std::set<PHINode*> PotentiallyDeadPHIs;
6536 PotentiallyDeadPHIs.insert(&PN);
6537 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
6538 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
6544 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
6545 Instruction *InsertPoint,
6547 unsigned PS = IC->getTargetData().getPointerSize();
6548 const Type *VTy = V->getType();
6549 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
6550 // We must insert a cast to ensure we sign-extend.
6551 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
6552 V->getName()), *InsertPoint);
6553 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
6558 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
6559 Value *PtrOp = GEP.getOperand(0);
6560 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
6561 // If so, eliminate the noop.
6562 if (GEP.getNumOperands() == 1)
6563 return ReplaceInstUsesWith(GEP, PtrOp);
6565 if (isa<UndefValue>(GEP.getOperand(0)))
6566 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
6568 bool HasZeroPointerIndex = false;
6569 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
6570 HasZeroPointerIndex = C->isNullValue();
6572 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
6573 return ReplaceInstUsesWith(GEP, PtrOp);
6575 // Eliminate unneeded casts for indices.
6576 bool MadeChange = false;
6577 gep_type_iterator GTI = gep_type_begin(GEP);
6578 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
6579 if (isa<SequentialType>(*GTI)) {
6580 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
6581 Value *Src = CI->getOperand(0);
6582 const Type *SrcTy = Src->getType();
6583 const Type *DestTy = CI->getType();
6584 if (Src->getType()->isInteger()) {
6585 if (SrcTy->getPrimitiveSizeInBits() ==
6586 DestTy->getPrimitiveSizeInBits()) {
6587 // We can always eliminate a cast from ulong or long to the other.
6588 // We can always eliminate a cast from uint to int or the other on
6589 // 32-bit pointer platforms.
6590 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
6592 GEP.setOperand(i, Src);
6594 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
6595 SrcTy->getPrimitiveSize() == 4) {
6596 // We can always eliminate a cast from int to [u]long. We can
6597 // eliminate a cast from uint to [u]long iff the target is a 32-bit
6599 if (SrcTy->isSigned() ||
6600 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
6602 GEP.setOperand(i, Src);
6607 // If we are using a wider index than needed for this platform, shrink it
6608 // to what we need. If the incoming value needs a cast instruction,
6609 // insert it. This explicit cast can make subsequent optimizations more
6611 Value *Op = GEP.getOperand(i);
6612 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
6613 if (Constant *C = dyn_cast<Constant>(Op)) {
6614 GEP.setOperand(i, ConstantExpr::getCast(C,
6615 TD->getIntPtrType()->getSignedVersion()));
6618 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
6619 Op->getName()), GEP);
6620 GEP.setOperand(i, Op);
6624 // If this is a constant idx, make sure to canonicalize it to be a signed
6625 // operand, otherwise CSE and other optimizations are pessimized.
6626 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
6627 GEP.setOperand(i, ConstantExpr::getCast(CUI,
6628 CUI->getType()->getSignedVersion()));
6632 if (MadeChange) return &GEP;
6634 // Combine Indices - If the source pointer to this getelementptr instruction
6635 // is a getelementptr instruction, combine the indices of the two
6636 // getelementptr instructions into a single instruction.
6638 std::vector<Value*> SrcGEPOperands;
6639 if (User *Src = dyn_castGetElementPtr(PtrOp))
6640 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
6642 if (!SrcGEPOperands.empty()) {
6643 // Note that if our source is a gep chain itself that we wait for that
6644 // chain to be resolved before we perform this transformation. This
6645 // avoids us creating a TON of code in some cases.
6647 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
6648 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
6649 return 0; // Wait until our source is folded to completion.
6651 std::vector<Value *> Indices;
6653 // Find out whether the last index in the source GEP is a sequential idx.
6654 bool EndsWithSequential = false;
6655 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
6656 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
6657 EndsWithSequential = !isa<StructType>(*I);
6659 // Can we combine the two pointer arithmetics offsets?
6660 if (EndsWithSequential) {
6661 // Replace: gep (gep %P, long B), long A, ...
6662 // With: T = long A+B; gep %P, T, ...
6664 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
6665 if (SO1 == Constant::getNullValue(SO1->getType())) {
6667 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
6670 // If they aren't the same type, convert both to an integer of the
6671 // target's pointer size.
6672 if (SO1->getType() != GO1->getType()) {
6673 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
6674 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
6675 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
6676 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
6678 unsigned PS = TD->getPointerSize();
6679 if (SO1->getType()->getPrimitiveSize() == PS) {
6680 // Convert GO1 to SO1's type.
6681 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
6683 } else if (GO1->getType()->getPrimitiveSize() == PS) {
6684 // Convert SO1 to GO1's type.
6685 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
6687 const Type *PT = TD->getIntPtrType();
6688 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
6689 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
6693 if (isa<Constant>(SO1) && isa<Constant>(GO1))
6694 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
6696 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
6697 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
6701 // Recycle the GEP we already have if possible.
6702 if (SrcGEPOperands.size() == 2) {
6703 GEP.setOperand(0, SrcGEPOperands[0]);
6704 GEP.setOperand(1, Sum);
6707 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
6708 SrcGEPOperands.end()-1);
6709 Indices.push_back(Sum);
6710 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
6712 } else if (isa<Constant>(*GEP.idx_begin()) &&
6713 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
6714 SrcGEPOperands.size() != 1) {
6715 // Otherwise we can do the fold if the first index of the GEP is a zero
6716 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
6717 SrcGEPOperands.end());
6718 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
6721 if (!Indices.empty())
6722 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
6724 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
6725 // GEP of global variable. If all of the indices for this GEP are
6726 // constants, we can promote this to a constexpr instead of an instruction.
6728 // Scan for nonconstants...
6729 std::vector<Constant*> Indices;
6730 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
6731 for (; I != E && isa<Constant>(*I); ++I)
6732 Indices.push_back(cast<Constant>(*I));
6734 if (I == E) { // If they are all constants...
6735 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
6737 // Replace all uses of the GEP with the new constexpr...
6738 return ReplaceInstUsesWith(GEP, CE);
6740 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
6741 if (!isa<PointerType>(X->getType())) {
6742 // Not interesting. Source pointer must be a cast from pointer.
6743 } else if (HasZeroPointerIndex) {
6744 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
6745 // into : GEP [10 x ubyte]* X, long 0, ...
6747 // This occurs when the program declares an array extern like "int X[];"
6749 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
6750 const PointerType *XTy = cast<PointerType>(X->getType());
6751 if (const ArrayType *XATy =
6752 dyn_cast<ArrayType>(XTy->getElementType()))
6753 if (const ArrayType *CATy =
6754 dyn_cast<ArrayType>(CPTy->getElementType()))
6755 if (CATy->getElementType() == XATy->getElementType()) {
6756 // At this point, we know that the cast source type is a pointer
6757 // to an array of the same type as the destination pointer
6758 // array. Because the array type is never stepped over (there
6759 // is a leading zero) we can fold the cast into this GEP.
6760 GEP.setOperand(0, X);
6763 } else if (GEP.getNumOperands() == 2) {
6764 // Transform things like:
6765 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
6766 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
6767 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
6768 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
6769 if (isa<ArrayType>(SrcElTy) &&
6770 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
6771 TD->getTypeSize(ResElTy)) {
6772 Value *V = InsertNewInstBefore(
6773 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
6774 GEP.getOperand(1), GEP.getName()), GEP);
6775 return new CastInst(V, GEP.getType());
6778 // Transform things like:
6779 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
6780 // (where tmp = 8*tmp2) into:
6781 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
6783 if (isa<ArrayType>(SrcElTy) &&
6784 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
6785 uint64_t ArrayEltSize =
6786 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
6788 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
6789 // allow either a mul, shift, or constant here.
6791 ConstantInt *Scale = 0;
6792 if (ArrayEltSize == 1) {
6793 NewIdx = GEP.getOperand(1);
6794 Scale = ConstantInt::get(NewIdx->getType(), 1);
6795 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
6796 NewIdx = ConstantInt::get(CI->getType(), 1);
6798 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
6799 if (Inst->getOpcode() == Instruction::Shl &&
6800 isa<ConstantInt>(Inst->getOperand(1))) {
6801 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
6802 if (Inst->getType()->isSigned())
6803 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
6805 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
6806 NewIdx = Inst->getOperand(0);
6807 } else if (Inst->getOpcode() == Instruction::Mul &&
6808 isa<ConstantInt>(Inst->getOperand(1))) {
6809 Scale = cast<ConstantInt>(Inst->getOperand(1));
6810 NewIdx = Inst->getOperand(0);
6814 // If the index will be to exactly the right offset with the scale taken
6815 // out, perform the transformation.
6816 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
6817 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
6818 Scale = ConstantSInt::get(C->getType(),
6819 (int64_t)C->getRawValue() /
6820 (int64_t)ArrayEltSize);
6822 Scale = ConstantUInt::get(Scale->getType(),
6823 Scale->getRawValue() / ArrayEltSize);
6824 if (Scale->getRawValue() != 1) {
6825 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
6826 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
6827 NewIdx = InsertNewInstBefore(Sc, GEP);
6830 // Insert the new GEP instruction.
6832 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
6833 NewIdx, GEP.getName());
6834 Idx = InsertNewInstBefore(Idx, GEP);
6835 return new CastInst(Idx, GEP.getType());
6844 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
6845 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
6846 if (AI.isArrayAllocation()) // Check C != 1
6847 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
6848 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
6849 AllocationInst *New = 0;
6851 // Create and insert the replacement instruction...
6852 if (isa<MallocInst>(AI))
6853 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
6855 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
6856 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
6859 InsertNewInstBefore(New, AI);
6861 // Scan to the end of the allocation instructions, to skip over a block of
6862 // allocas if possible...
6864 BasicBlock::iterator It = New;
6865 while (isa<AllocationInst>(*It)) ++It;
6867 // Now that I is pointing to the first non-allocation-inst in the block,
6868 // insert our getelementptr instruction...
6870 Value *NullIdx = Constant::getNullValue(Type::IntTy);
6871 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
6872 New->getName()+".sub", It);
6874 // Now make everything use the getelementptr instead of the original
6876 return ReplaceInstUsesWith(AI, V);
6877 } else if (isa<UndefValue>(AI.getArraySize())) {
6878 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6881 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
6882 // Note that we only do this for alloca's, because malloc should allocate and
6883 // return a unique pointer, even for a zero byte allocation.
6884 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
6885 TD->getTypeSize(AI.getAllocatedType()) == 0)
6886 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6891 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
6892 Value *Op = FI.getOperand(0);
6894 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
6895 if (CastInst *CI = dyn_cast<CastInst>(Op))
6896 if (isa<PointerType>(CI->getOperand(0)->getType())) {
6897 FI.setOperand(0, CI->getOperand(0));
6901 // free undef -> unreachable.
6902 if (isa<UndefValue>(Op)) {
6903 // Insert a new store to null because we cannot modify the CFG here.
6904 new StoreInst(ConstantBool::True,
6905 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
6906 return EraseInstFromFunction(FI);
6909 // If we have 'free null' delete the instruction. This can happen in stl code
6910 // when lots of inlining happens.
6911 if (isa<ConstantPointerNull>(Op))
6912 return EraseInstFromFunction(FI);
6918 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
6919 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
6920 User *CI = cast<User>(LI.getOperand(0));
6921 Value *CastOp = CI->getOperand(0);
6923 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
6924 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
6925 const Type *SrcPTy = SrcTy->getElementType();
6927 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
6928 isa<PackedType>(DestPTy)) {
6929 // If the source is an array, the code below will not succeed. Check to
6930 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
6932 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
6933 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
6934 if (ASrcTy->getNumElements() != 0) {
6935 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
6936 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
6937 SrcTy = cast<PointerType>(CastOp->getType());
6938 SrcPTy = SrcTy->getElementType();
6941 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
6942 isa<PackedType>(SrcPTy)) &&
6943 // Do not allow turning this into a load of an integer, which is then
6944 // casted to a pointer, this pessimizes pointer analysis a lot.
6945 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
6946 IC.getTargetData().getTypeSize(SrcPTy) ==
6947 IC.getTargetData().getTypeSize(DestPTy)) {
6949 // Okay, we are casting from one integer or pointer type to another of
6950 // the same size. Instead of casting the pointer before the load, cast
6951 // the result of the loaded value.
6952 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
6954 LI.isVolatile()),LI);
6955 // Now cast the result of the load.
6956 return new CastInst(NewLoad, LI.getType());
6963 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
6964 /// from this value cannot trap. If it is not obviously safe to load from the
6965 /// specified pointer, we do a quick local scan of the basic block containing
6966 /// ScanFrom, to determine if the address is already accessed.
6967 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
6968 // If it is an alloca or global variable, it is always safe to load from.
6969 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
6971 // Otherwise, be a little bit agressive by scanning the local block where we
6972 // want to check to see if the pointer is already being loaded or stored
6973 // from/to. If so, the previous load or store would have already trapped,
6974 // so there is no harm doing an extra load (also, CSE will later eliminate
6975 // the load entirely).
6976 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
6981 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
6982 if (LI->getOperand(0) == V) return true;
6983 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
6984 if (SI->getOperand(1) == V) return true;
6990 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
6991 Value *Op = LI.getOperand(0);
6993 // load (cast X) --> cast (load X) iff safe
6994 if (CastInst *CI = dyn_cast<CastInst>(Op))
6995 if (Instruction *Res = InstCombineLoadCast(*this, LI))
6998 // None of the following transforms are legal for volatile loads.
6999 if (LI.isVolatile()) return 0;
7001 if (&LI.getParent()->front() != &LI) {
7002 BasicBlock::iterator BBI = &LI; --BBI;
7003 // If the instruction immediately before this is a store to the same
7004 // address, do a simple form of store->load forwarding.
7005 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7006 if (SI->getOperand(1) == LI.getOperand(0))
7007 return ReplaceInstUsesWith(LI, SI->getOperand(0));
7008 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
7009 if (LIB->getOperand(0) == LI.getOperand(0))
7010 return ReplaceInstUsesWith(LI, LIB);
7013 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
7014 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
7015 isa<UndefValue>(GEPI->getOperand(0))) {
7016 // Insert a new store to null instruction before the load to indicate
7017 // that this code is not reachable. We do this instead of inserting
7018 // an unreachable instruction directly because we cannot modify the
7020 new StoreInst(UndefValue::get(LI.getType()),
7021 Constant::getNullValue(Op->getType()), &LI);
7022 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7025 if (Constant *C = dyn_cast<Constant>(Op)) {
7026 // load null/undef -> undef
7027 if ((C->isNullValue() || isa<UndefValue>(C))) {
7028 // Insert a new store to null instruction before the load to indicate that
7029 // this code is not reachable. We do this instead of inserting an
7030 // unreachable instruction directly because we cannot modify the CFG.
7031 new StoreInst(UndefValue::get(LI.getType()),
7032 Constant::getNullValue(Op->getType()), &LI);
7033 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7036 // Instcombine load (constant global) into the value loaded.
7037 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
7038 if (GV->isConstant() && !GV->isExternal())
7039 return ReplaceInstUsesWith(LI, GV->getInitializer());
7041 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
7042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7043 if (CE->getOpcode() == Instruction::GetElementPtr) {
7044 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
7045 if (GV->isConstant() && !GV->isExternal())
7047 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
7048 return ReplaceInstUsesWith(LI, V);
7049 if (CE->getOperand(0)->isNullValue()) {
7050 // Insert a new store to null instruction before the load to indicate
7051 // that this code is not reachable. We do this instead of inserting
7052 // an unreachable instruction directly because we cannot modify the
7054 new StoreInst(UndefValue::get(LI.getType()),
7055 Constant::getNullValue(Op->getType()), &LI);
7056 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7059 } else if (CE->getOpcode() == Instruction::Cast) {
7060 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7065 if (Op->hasOneUse()) {
7066 // Change select and PHI nodes to select values instead of addresses: this
7067 // helps alias analysis out a lot, allows many others simplifications, and
7068 // exposes redundancy in the code.
7070 // Note that we cannot do the transformation unless we know that the
7071 // introduced loads cannot trap! Something like this is valid as long as
7072 // the condition is always false: load (select bool %C, int* null, int* %G),
7073 // but it would not be valid if we transformed it to load from null
7076 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7077 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7078 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7079 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7080 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
7081 SI->getOperand(1)->getName()+".val"), LI);
7082 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
7083 SI->getOperand(2)->getName()+".val"), LI);
7084 return new SelectInst(SI->getCondition(), V1, V2);
7087 // load (select (cond, null, P)) -> load P
7088 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7089 if (C->isNullValue()) {
7090 LI.setOperand(0, SI->getOperand(2));
7094 // load (select (cond, P, null)) -> load P
7095 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7096 if (C->isNullValue()) {
7097 LI.setOperand(0, SI->getOperand(1));
7101 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
7102 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
7103 bool Safe = PN->getParent() == LI.getParent();
7105 // Scan all of the instructions between the PHI and the load to make
7106 // sure there are no instructions that might possibly alter the value
7107 // loaded from the PHI.
7109 BasicBlock::iterator I = &LI;
7110 for (--I; !isa<PHINode>(I); --I)
7111 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
7117 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
7118 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
7119 PN->getIncomingBlock(i)->getTerminator()))
7124 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
7125 InsertNewInstBefore(NewPN, *PN);
7126 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
7128 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
7129 BasicBlock *BB = PN->getIncomingBlock(i);
7130 Value *&TheLoad = LoadMap[BB];
7132 Value *InVal = PN->getIncomingValue(i);
7133 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
7134 InVal->getName()+".val"),
7135 *BB->getTerminator());
7137 NewPN->addIncoming(TheLoad, BB);
7139 return ReplaceInstUsesWith(LI, NewPN);
7146 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
7148 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7149 User *CI = cast<User>(SI.getOperand(1));
7150 Value *CastOp = CI->getOperand(0);
7152 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7153 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7154 const Type *SrcPTy = SrcTy->getElementType();
7156 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
7157 // If the source is an array, the code below will not succeed. Check to
7158 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7160 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7161 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7162 if (ASrcTy->getNumElements() != 0) {
7163 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7164 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7165 SrcTy = cast<PointerType>(CastOp->getType());
7166 SrcPTy = SrcTy->getElementType();
7169 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
7170 IC.getTargetData().getTypeSize(SrcPTy) ==
7171 IC.getTargetData().getTypeSize(DestPTy)) {
7173 // Okay, we are casting from one integer or pointer type to another of
7174 // the same size. Instead of casting the pointer before the store, cast
7175 // the value to be stored.
7177 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
7178 NewCast = ConstantExpr::getCast(C, SrcPTy);
7180 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
7182 SI.getOperand(0)->getName()+".c"), SI);
7184 return new StoreInst(NewCast, CastOp);
7191 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7192 Value *Val = SI.getOperand(0);
7193 Value *Ptr = SI.getOperand(1);
7195 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
7196 EraseInstFromFunction(SI);
7201 // Do really simple DSE, to catch cases where there are several consequtive
7202 // stores to the same location, separated by a few arithmetic operations. This
7203 // situation often occurs with bitfield accesses.
7204 BasicBlock::iterator BBI = &SI;
7205 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7209 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7210 // Prev store isn't volatile, and stores to the same location?
7211 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
7214 EraseInstFromFunction(*PrevSI);
7220 // If this is a load, we have to stop. However, if the loaded value is from
7221 // the pointer we're loading and is producing the pointer we're storing,
7222 // then *this* store is dead (X = load P; store X -> P).
7223 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7224 if (LI == Val && LI->getOperand(0) == Ptr) {
7225 EraseInstFromFunction(SI);
7229 // Otherwise, this is a load from some other location. Stores before it
7234 // Don't skip over loads or things that can modify memory.
7235 if (BBI->mayWriteToMemory())
7240 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7242 // store X, null -> turns into 'unreachable' in SimplifyCFG
7243 if (isa<ConstantPointerNull>(Ptr)) {
7244 if (!isa<UndefValue>(Val)) {
7245 SI.setOperand(0, UndefValue::get(Val->getType()));
7246 if (Instruction *U = dyn_cast<Instruction>(Val))
7247 WorkList.push_back(U); // Dropped a use.
7250 return 0; // Do not modify these!
7253 // store undef, Ptr -> noop
7254 if (isa<UndefValue>(Val)) {
7255 EraseInstFromFunction(SI);
7260 // If the pointer destination is a cast, see if we can fold the cast into the
7262 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
7263 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7265 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
7266 if (CE->getOpcode() == Instruction::Cast)
7267 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7271 // If this store is the last instruction in the basic block, and if the block
7272 // ends with an unconditional branch, try to move it to the successor block.
7274 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
7275 if (BI->isUnconditional()) {
7276 // Check to see if the successor block has exactly two incoming edges. If
7277 // so, see if the other predecessor contains a store to the same location.
7278 // if so, insert a PHI node (if needed) and move the stores down.
7279 BasicBlock *Dest = BI->getSuccessor(0);
7281 pred_iterator PI = pred_begin(Dest);
7282 BasicBlock *Other = 0;
7283 if (*PI != BI->getParent())
7286 if (PI != pred_end(Dest)) {
7287 if (*PI != BI->getParent())
7292 if (++PI != pred_end(Dest))
7295 if (Other) { // If only one other pred...
7296 BBI = Other->getTerminator();
7297 // Make sure this other block ends in an unconditional branch and that
7298 // there is an instruction before the branch.
7299 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
7300 BBI != Other->begin()) {
7302 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
7304 // If this instruction is a store to the same location.
7305 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
7306 // Okay, we know we can perform this transformation. Insert a PHI
7307 // node now if we need it.
7308 Value *MergedVal = OtherStore->getOperand(0);
7309 if (MergedVal != SI.getOperand(0)) {
7310 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
7311 PN->reserveOperandSpace(2);
7312 PN->addIncoming(SI.getOperand(0), SI.getParent());
7313 PN->addIncoming(OtherStore->getOperand(0), Other);
7314 MergedVal = InsertNewInstBefore(PN, Dest->front());
7317 // Advance to a place where it is safe to insert the new store and
7319 BBI = Dest->begin();
7320 while (isa<PHINode>(BBI)) ++BBI;
7321 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
7322 OtherStore->isVolatile()), *BBI);
7324 // Nuke the old stores.
7325 EraseInstFromFunction(SI);
7326 EraseInstFromFunction(*OtherStore);
7338 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
7339 // Change br (not X), label True, label False to: br X, label False, True
7341 BasicBlock *TrueDest;
7342 BasicBlock *FalseDest;
7343 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
7344 !isa<Constant>(X)) {
7345 // Swap Destinations and condition...
7347 BI.setSuccessor(0, FalseDest);
7348 BI.setSuccessor(1, TrueDest);
7352 // Cannonicalize setne -> seteq
7353 Instruction::BinaryOps Op; Value *Y;
7354 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
7355 TrueDest, FalseDest)))
7356 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
7357 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
7358 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
7359 std::string Name = I->getName(); I->setName("");
7360 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
7361 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
7362 // Swap Destinations and condition...
7363 BI.setCondition(NewSCC);
7364 BI.setSuccessor(0, FalseDest);
7365 BI.setSuccessor(1, TrueDest);
7366 removeFromWorkList(I);
7367 I->getParent()->getInstList().erase(I);
7368 WorkList.push_back(cast<Instruction>(NewSCC));
7375 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
7376 Value *Cond = SI.getCondition();
7377 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
7378 if (I->getOpcode() == Instruction::Add)
7379 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7380 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
7381 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
7382 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
7384 SI.setOperand(0, I->getOperand(0));
7385 WorkList.push_back(I);
7392 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
7393 /// is to leave as a vector operation.
7394 static bool CheapToScalarize(Value *V, bool isConstant) {
7395 if (isa<ConstantAggregateZero>(V))
7397 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
7398 if (isConstant) return true;
7399 // If all elts are the same, we can extract.
7400 Constant *Op0 = C->getOperand(0);
7401 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7402 if (C->getOperand(i) != Op0)
7406 Instruction *I = dyn_cast<Instruction>(V);
7407 if (!I) return false;
7409 // Insert element gets simplified to the inserted element or is deleted if
7410 // this is constant idx extract element and its a constant idx insertelt.
7411 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
7412 isa<ConstantInt>(I->getOperand(2)))
7414 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
7416 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
7417 if (BO->hasOneUse() &&
7418 (CheapToScalarize(BO->getOperand(0), isConstant) ||
7419 CheapToScalarize(BO->getOperand(1), isConstant)))
7425 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
7426 /// elements into values that are larger than the #elts in the input.
7427 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
7428 unsigned NElts = SVI->getType()->getNumElements();
7429 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
7430 return std::vector<unsigned>(NElts, 0);
7431 if (isa<UndefValue>(SVI->getOperand(2)))
7432 return std::vector<unsigned>(NElts, 2*NElts);
7434 std::vector<unsigned> Result;
7435 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
7436 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
7437 if (isa<UndefValue>(CP->getOperand(i)))
7438 Result.push_back(NElts*2); // undef -> 8
7440 Result.push_back(cast<ConstantUInt>(CP->getOperand(i))->getValue());
7444 /// FindScalarElement - Given a vector and an element number, see if the scalar
7445 /// value is already around as a register, for example if it were inserted then
7446 /// extracted from the vector.
7447 static Value *FindScalarElement(Value *V, unsigned EltNo) {
7448 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
7449 const PackedType *PTy = cast<PackedType>(V->getType());
7450 unsigned Width = PTy->getNumElements();
7451 if (EltNo >= Width) // Out of range access.
7452 return UndefValue::get(PTy->getElementType());
7454 if (isa<UndefValue>(V))
7455 return UndefValue::get(PTy->getElementType());
7456 else if (isa<ConstantAggregateZero>(V))
7457 return Constant::getNullValue(PTy->getElementType());
7458 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
7459 return CP->getOperand(EltNo);
7460 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
7461 // If this is an insert to a variable element, we don't know what it is.
7462 if (!isa<ConstantUInt>(III->getOperand(2))) return 0;
7463 unsigned IIElt = cast<ConstantUInt>(III->getOperand(2))->getValue();
7465 // If this is an insert to the element we are looking for, return the
7467 if (EltNo == IIElt) return III->getOperand(1);
7469 // Otherwise, the insertelement doesn't modify the value, recurse on its
7471 return FindScalarElement(III->getOperand(0), EltNo);
7472 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
7473 unsigned InEl = getShuffleMask(SVI)[EltNo];
7475 return FindScalarElement(SVI->getOperand(0), InEl);
7476 else if (InEl < Width*2)
7477 return FindScalarElement(SVI->getOperand(1), InEl - Width);
7479 return UndefValue::get(PTy->getElementType());
7482 // Otherwise, we don't know.
7486 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
7488 // If packed val is undef, replace extract with scalar undef.
7489 if (isa<UndefValue>(EI.getOperand(0)))
7490 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
7492 // If packed val is constant 0, replace extract with scalar 0.
7493 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
7494 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
7496 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
7497 // If packed val is constant with uniform operands, replace EI
7498 // with that operand
7499 Constant *op0 = C->getOperand(0);
7500 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7501 if (C->getOperand(i) != op0) {
7506 return ReplaceInstUsesWith(EI, op0);
7509 // If extracting a specified index from the vector, see if we can recursively
7510 // find a previously computed scalar that was inserted into the vector.
7511 if (ConstantUInt *IdxC = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
7512 if (Value *Elt = FindScalarElement(EI.getOperand(0), IdxC->getValue()))
7513 return ReplaceInstUsesWith(EI, Elt);
7516 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
7517 if (I->hasOneUse()) {
7518 // Push extractelement into predecessor operation if legal and
7519 // profitable to do so
7520 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
7521 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
7522 if (CheapToScalarize(BO, isConstantElt)) {
7523 ExtractElementInst *newEI0 =
7524 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
7525 EI.getName()+".lhs");
7526 ExtractElementInst *newEI1 =
7527 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
7528 EI.getName()+".rhs");
7529 InsertNewInstBefore(newEI0, EI);
7530 InsertNewInstBefore(newEI1, EI);
7531 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
7533 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7534 Value *Ptr = InsertCastBefore(I->getOperand(0),
7535 PointerType::get(EI.getType()), EI);
7536 GetElementPtrInst *GEP =
7537 new GetElementPtrInst(Ptr, EI.getOperand(1),
7538 I->getName() + ".gep");
7539 InsertNewInstBefore(GEP, EI);
7540 return new LoadInst(GEP);
7543 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
7544 // Extracting the inserted element?
7545 if (IE->getOperand(2) == EI.getOperand(1))
7546 return ReplaceInstUsesWith(EI, IE->getOperand(1));
7547 // If the inserted and extracted elements are constants, they must not
7548 // be the same value, extract from the pre-inserted value instead.
7549 if (isa<Constant>(IE->getOperand(2)) &&
7550 isa<Constant>(EI.getOperand(1))) {
7551 AddUsesToWorkList(EI);
7552 EI.setOperand(0, IE->getOperand(0));
7555 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
7556 // If this is extracting an element from a shufflevector, figure out where
7557 // it came from and extract from the appropriate input element instead.
7558 if (ConstantUInt *Elt = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
7559 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getValue()];
7561 if (SrcIdx < SVI->getType()->getNumElements())
7562 Src = SVI->getOperand(0);
7563 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
7564 SrcIdx -= SVI->getType()->getNumElements();
7565 Src = SVI->getOperand(1);
7567 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
7569 return new ExtractElementInst(Src,
7570 ConstantUInt::get(Type::UIntTy, SrcIdx));
7577 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
7578 /// elements from either LHS or RHS, return the shuffle mask and true.
7579 /// Otherwise, return false.
7580 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
7581 std::vector<Constant*> &Mask) {
7582 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
7583 "Invalid CollectSingleShuffleElements");
7584 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
7586 if (isa<UndefValue>(V)) {
7587 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
7589 } else if (V == LHS) {
7590 for (unsigned i = 0; i != NumElts; ++i)
7591 Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
7593 } else if (V == RHS) {
7594 for (unsigned i = 0; i != NumElts; ++i)
7595 Mask.push_back(ConstantUInt::get(Type::UIntTy, i+NumElts));
7597 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
7598 // If this is an insert of an extract from some other vector, include it.
7599 Value *VecOp = IEI->getOperand(0);
7600 Value *ScalarOp = IEI->getOperand(1);
7601 Value *IdxOp = IEI->getOperand(2);
7603 if (!isa<ConstantInt>(IdxOp))
7605 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7607 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
7608 // Okay, we can handle this if the vector we are insertinting into is
7610 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
7611 // If so, update the mask to reflect the inserted undef.
7612 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
7615 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
7616 if (isa<ConstantInt>(EI->getOperand(1)) &&
7617 EI->getOperand(0)->getType() == V->getType()) {
7618 unsigned ExtractedIdx =
7619 cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7621 // This must be extracting from either LHS or RHS.
7622 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
7623 // Okay, we can handle this if the vector we are insertinting into is
7625 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
7626 // If so, update the mask to reflect the inserted value.
7627 if (EI->getOperand(0) == LHS) {
7628 Mask[InsertedIdx & (NumElts-1)] =
7629 ConstantUInt::get(Type::UIntTy, ExtractedIdx);
7631 assert(EI->getOperand(0) == RHS);
7632 Mask[InsertedIdx & (NumElts-1)] =
7633 ConstantUInt::get(Type::UIntTy, ExtractedIdx+NumElts);
7642 // TODO: Handle shufflevector here!
7647 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
7648 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
7649 /// that computes V and the LHS value of the shuffle.
7650 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
7652 assert(isa<PackedType>(V->getType()) &&
7653 (RHS == 0 || V->getType() == RHS->getType()) &&
7654 "Invalid shuffle!");
7655 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
7657 if (isa<UndefValue>(V)) {
7658 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
7660 } else if (isa<ConstantAggregateZero>(V)) {
7661 Mask.assign(NumElts, ConstantUInt::get(Type::UIntTy, 0));
7663 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
7664 // If this is an insert of an extract from some other vector, include it.
7665 Value *VecOp = IEI->getOperand(0);
7666 Value *ScalarOp = IEI->getOperand(1);
7667 Value *IdxOp = IEI->getOperand(2);
7669 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
7670 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
7671 EI->getOperand(0)->getType() == V->getType()) {
7672 unsigned ExtractedIdx =
7673 cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7674 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7676 // Either the extracted from or inserted into vector must be RHSVec,
7677 // otherwise we'd end up with a shuffle of three inputs.
7678 if (EI->getOperand(0) == RHS || RHS == 0) {
7679 RHS = EI->getOperand(0);
7680 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
7681 Mask[InsertedIdx & (NumElts-1)] =
7682 ConstantUInt::get(Type::UIntTy, NumElts+ExtractedIdx);
7687 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
7688 // Everything but the extracted element is replaced with the RHS.
7689 for (unsigned i = 0; i != NumElts; ++i) {
7690 if (i != InsertedIdx)
7691 Mask[i] = ConstantUInt::get(Type::UIntTy, NumElts+i);
7696 // If this insertelement is a chain that comes from exactly these two
7697 // vectors, return the vector and the effective shuffle.
7698 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
7699 return EI->getOperand(0);
7704 // TODO: Handle shufflevector here!
7706 // Otherwise, can't do anything fancy. Return an identity vector.
7707 for (unsigned i = 0; i != NumElts; ++i)
7708 Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
7712 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
7713 Value *VecOp = IE.getOperand(0);
7714 Value *ScalarOp = IE.getOperand(1);
7715 Value *IdxOp = IE.getOperand(2);
7717 // If the inserted element was extracted from some other vector, and if the
7718 // indexes are constant, try to turn this into a shufflevector operation.
7719 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
7720 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
7721 EI->getOperand(0)->getType() == IE.getType()) {
7722 unsigned NumVectorElts = IE.getType()->getNumElements();
7723 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7724 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7726 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
7727 return ReplaceInstUsesWith(IE, VecOp);
7729 if (InsertedIdx >= NumVectorElts) // Out of range insert.
7730 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
7732 // If we are extracting a value from a vector, then inserting it right
7733 // back into the same place, just use the input vector.
7734 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
7735 return ReplaceInstUsesWith(IE, VecOp);
7737 // We could theoretically do this for ANY input. However, doing so could
7738 // turn chains of insertelement instructions into a chain of shufflevector
7739 // instructions, and right now we do not merge shufflevectors. As such,
7740 // only do this in a situation where it is clear that there is benefit.
7741 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
7742 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
7743 // the values of VecOp, except then one read from EIOp0.
7744 // Build a new shuffle mask.
7745 std::vector<Constant*> Mask;
7746 if (isa<UndefValue>(VecOp))
7747 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
7749 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
7750 Mask.assign(NumVectorElts, ConstantUInt::get(Type::UIntTy,
7753 Mask[InsertedIdx] = ConstantUInt::get(Type::UIntTy, ExtractedIdx);
7754 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
7755 ConstantPacked::get(Mask));
7758 // If this insertelement isn't used by some other insertelement, turn it
7759 // (and any insertelements it points to), into one big shuffle.
7760 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
7761 std::vector<Constant*> Mask;
7763 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
7764 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
7765 // We now have a shuffle of LHS, RHS, Mask.
7766 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
7775 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
7776 Value *LHS = SVI.getOperand(0);
7777 Value *RHS = SVI.getOperand(1);
7778 std::vector<unsigned> Mask = getShuffleMask(&SVI);
7780 bool MadeChange = false;
7782 if (isa<UndefValue>(SVI.getOperand(2)))
7783 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
7785 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
7786 // the undef, change them to undefs.
7788 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
7789 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
7790 if (LHS == RHS || isa<UndefValue>(LHS)) {
7791 if (isa<UndefValue>(LHS) && LHS == RHS) {
7792 // shuffle(undef,undef,mask) -> undef.
7793 return ReplaceInstUsesWith(SVI, LHS);
7796 // Remap any references to RHS to use LHS.
7797 std::vector<Constant*> Elts;
7798 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
7800 Elts.push_back(UndefValue::get(Type::UIntTy));
7802 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
7803 (Mask[i] < e && isa<UndefValue>(LHS)))
7804 Mask[i] = 2*e; // Turn into undef.
7806 Mask[i] &= (e-1); // Force to LHS.
7807 Elts.push_back(ConstantUInt::get(Type::UIntTy, Mask[i]));
7810 SVI.setOperand(0, SVI.getOperand(1));
7811 SVI.setOperand(1, UndefValue::get(RHS->getType()));
7812 SVI.setOperand(2, ConstantPacked::get(Elts));
7813 LHS = SVI.getOperand(0);
7814 RHS = SVI.getOperand(1);
7818 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
7819 bool isLHSID = true, isRHSID = true;
7821 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
7822 if (Mask[i] >= e*2) continue; // Ignore undef values.
7823 // Is this an identity shuffle of the LHS value?
7824 isLHSID &= (Mask[i] == i);
7826 // Is this an identity shuffle of the RHS value?
7827 isRHSID &= (Mask[i]-e == i);
7830 // Eliminate identity shuffles.
7831 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
7832 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
7834 // If the LHS is a shufflevector itself, see if we can combine it with this
7835 // one without producing an unusual shuffle. Here we are really conservative:
7836 // we are absolutely afraid of producing a shuffle mask not in the input
7837 // program, because the code gen may not be smart enough to turn a merged
7838 // shuffle into two specific shuffles: it may produce worse code. As such,
7839 // we only merge two shuffles if the result is one of the two input shuffle
7840 // masks. In this case, merging the shuffles just removes one instruction,
7841 // which we know is safe. This is good for things like turning:
7842 // (splat(splat)) -> splat.
7843 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
7844 if (isa<UndefValue>(RHS)) {
7845 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
7847 std::vector<unsigned> NewMask;
7848 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
7850 NewMask.push_back(2*e);
7852 NewMask.push_back(LHSMask[Mask[i]]);
7854 // If the result mask is equal to the src shuffle or this shuffle mask, do
7856 if (NewMask == LHSMask || NewMask == Mask) {
7857 std::vector<Constant*> Elts;
7858 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
7859 if (NewMask[i] >= e*2) {
7860 Elts.push_back(UndefValue::get(Type::UIntTy));
7862 Elts.push_back(ConstantUInt::get(Type::UIntTy, NewMask[i]));
7865 return new ShuffleVectorInst(LHSSVI->getOperand(0),
7866 LHSSVI->getOperand(1),
7867 ConstantPacked::get(Elts));
7872 return MadeChange ? &SVI : 0;
7877 void InstCombiner::removeFromWorkList(Instruction *I) {
7878 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
7883 /// TryToSinkInstruction - Try to move the specified instruction from its
7884 /// current block into the beginning of DestBlock, which can only happen if it's
7885 /// safe to move the instruction past all of the instructions between it and the
7886 /// end of its block.
7887 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
7888 assert(I->hasOneUse() && "Invariants didn't hold!");
7890 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
7891 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
7893 // Do not sink alloca instructions out of the entry block.
7894 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
7897 // We can only sink load instructions if there is nothing between the load and
7898 // the end of block that could change the value.
7899 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7900 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
7902 if (Scan->mayWriteToMemory())
7906 BasicBlock::iterator InsertPos = DestBlock->begin();
7907 while (isa<PHINode>(InsertPos)) ++InsertPos;
7909 I->moveBefore(InsertPos);
7914 /// OptimizeConstantExpr - Given a constant expression and target data layout
7915 /// information, symbolically evaluation the constant expr to something simpler
7917 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
7920 Constant *Ptr = CE->getOperand(0);
7921 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
7922 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
7923 // If this is a constant expr gep that is effectively computing an
7924 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
7925 bool isFoldableGEP = true;
7926 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
7927 if (!isa<ConstantInt>(CE->getOperand(i)))
7928 isFoldableGEP = false;
7929 if (isFoldableGEP) {
7930 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
7931 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
7932 Constant *C = ConstantUInt::get(Type::ULongTy, Offset);
7933 C = ConstantExpr::getCast(C, TD->getIntPtrType());
7934 return ConstantExpr::getCast(C, CE->getType());
7942 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
7943 /// all reachable code to the worklist.
7945 /// This has a couple of tricks to make the code faster and more powerful. In
7946 /// particular, we constant fold and DCE instructions as we go, to avoid adding
7947 /// them to the worklist (this significantly speeds up instcombine on code where
7948 /// many instructions are dead or constant). Additionally, if we find a branch
7949 /// whose condition is a known constant, we only visit the reachable successors.
7951 static void AddReachableCodeToWorklist(BasicBlock *BB,
7952 std::set<BasicBlock*> &Visited,
7953 std::vector<Instruction*> &WorkList,
7954 const TargetData *TD) {
7955 // We have now visited this block! If we've already been here, bail out.
7956 if (!Visited.insert(BB).second) return;
7958 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
7959 Instruction *Inst = BBI++;
7961 // DCE instruction if trivially dead.
7962 if (isInstructionTriviallyDead(Inst)) {
7964 DEBUG(std::cerr << "IC: DCE: " << *Inst);
7965 Inst->eraseFromParent();
7969 // ConstantProp instruction if trivially constant.
7970 if (Constant *C = ConstantFoldInstruction(Inst)) {
7971 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
7972 C = OptimizeConstantExpr(CE, TD);
7973 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *Inst);
7974 Inst->replaceAllUsesWith(C);
7976 Inst->eraseFromParent();
7980 WorkList.push_back(Inst);
7983 // Recursively visit successors. If this is a branch or switch on a constant,
7984 // only visit the reachable successor.
7985 TerminatorInst *TI = BB->getTerminator();
7986 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
7987 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
7988 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
7989 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
7993 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
7994 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
7995 // See if this is an explicit destination.
7996 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
7997 if (SI->getCaseValue(i) == Cond) {
7998 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
8002 // Otherwise it is the default destination.
8003 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
8008 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
8009 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
8012 bool InstCombiner::runOnFunction(Function &F) {
8013 bool Changed = false;
8014 TD = &getAnalysis<TargetData>();
8017 // Do a depth-first traversal of the function, populate the worklist with
8018 // the reachable instructions. Ignore blocks that are not reachable. Keep
8019 // track of which blocks we visit.
8020 std::set<BasicBlock*> Visited;
8021 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
8023 // Do a quick scan over the function. If we find any blocks that are
8024 // unreachable, remove any instructions inside of them. This prevents
8025 // the instcombine code from having to deal with some bad special cases.
8026 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
8027 if (!Visited.count(BB)) {
8028 Instruction *Term = BB->getTerminator();
8029 while (Term != BB->begin()) { // Remove instrs bottom-up
8030 BasicBlock::iterator I = Term; --I;
8032 DEBUG(std::cerr << "IC: DCE: " << *I);
8035 if (!I->use_empty())
8036 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8037 I->eraseFromParent();
8042 while (!WorkList.empty()) {
8043 Instruction *I = WorkList.back(); // Get an instruction from the worklist
8044 WorkList.pop_back();
8046 // Check to see if we can DCE the instruction.
8047 if (isInstructionTriviallyDead(I)) {
8048 // Add operands to the worklist.
8049 if (I->getNumOperands() < 4)
8050 AddUsesToWorkList(*I);
8053 DEBUG(std::cerr << "IC: DCE: " << *I);
8055 I->eraseFromParent();
8056 removeFromWorkList(I);
8060 // Instruction isn't dead, see if we can constant propagate it.
8061 if (Constant *C = ConstantFoldInstruction(I)) {
8062 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8063 C = OptimizeConstantExpr(CE, TD);
8064 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
8066 // Add operands to the worklist.
8067 AddUsesToWorkList(*I);
8068 ReplaceInstUsesWith(*I, C);
8071 I->eraseFromParent();
8072 removeFromWorkList(I);
8076 // See if we can trivially sink this instruction to a successor basic block.
8077 if (I->hasOneUse()) {
8078 BasicBlock *BB = I->getParent();
8079 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
8080 if (UserParent != BB) {
8081 bool UserIsSuccessor = false;
8082 // See if the user is one of our successors.
8083 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8084 if (*SI == UserParent) {
8085 UserIsSuccessor = true;
8089 // If the user is one of our immediate successors, and if that successor
8090 // only has us as a predecessors (we'd have to split the critical edge
8091 // otherwise), we can keep going.
8092 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
8093 next(pred_begin(UserParent)) == pred_end(UserParent))
8094 // Okay, the CFG is simple enough, try to sink this instruction.
8095 Changed |= TryToSinkInstruction(I, UserParent);
8099 // Now that we have an instruction, try combining it to simplify it...
8100 if (Instruction *Result = visit(*I)) {
8102 // Should we replace the old instruction with a new one?
8104 DEBUG(std::cerr << "IC: Old = " << *I
8105 << " New = " << *Result);
8107 // Everything uses the new instruction now.
8108 I->replaceAllUsesWith(Result);
8110 // Push the new instruction and any users onto the worklist.
8111 WorkList.push_back(Result);
8112 AddUsersToWorkList(*Result);
8114 // Move the name to the new instruction first...
8115 std::string OldName = I->getName(); I->setName("");
8116 Result->setName(OldName);
8118 // Insert the new instruction into the basic block...
8119 BasicBlock *InstParent = I->getParent();
8120 BasicBlock::iterator InsertPos = I;
8122 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8123 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8126 InstParent->getInstList().insert(InsertPos, Result);
8128 // Make sure that we reprocess all operands now that we reduced their
8130 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8131 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8132 WorkList.push_back(OpI);
8134 // Instructions can end up on the worklist more than once. Make sure
8135 // we do not process an instruction that has been deleted.
8136 removeFromWorkList(I);
8138 // Erase the old instruction.
8139 InstParent->getInstList().erase(I);
8141 DEBUG(std::cerr << "IC: MOD = " << *I);
8143 // If the instruction was modified, it's possible that it is now dead.
8144 // if so, remove it.
8145 if (isInstructionTriviallyDead(I)) {
8146 // Make sure we process all operands now that we are reducing their
8148 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8149 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8150 WorkList.push_back(OpI);
8152 // Instructions may end up in the worklist more than once. Erase all
8153 // occurrences of this instruction.
8154 removeFromWorkList(I);
8155 I->eraseFromParent();
8157 WorkList.push_back(Result);
8158 AddUsersToWorkList(*Result);
8168 FunctionPass *llvm::createInstructionCombiningPass() {
8169 return new InstCombiner();