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);
1607 const Type *UIntPtrTy = TD->getIntPtrType();
1608 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
1609 if((CI->getType() == UIntPtrTy || CI->getType() == SIntPtrTy)
1610 && isa<PointerType>(CI->getOperand(0)->getType())) {
1611 Instruction* I2 = new CastInst(CI->getOperand(0),
1612 PointerType::get(Type::SByteTy), "ctg", &I);
1613 WorkList.push_back(I2);
1614 I2 = new GetElementPtrInst(I2, Other, "ctg", &I);
1615 WorkList.push_back(I2);
1616 return new CastInst(I2, CI->getType());
1621 return Changed ? &I : 0;
1624 // isSignBit - Return true if the value represented by the constant only has the
1625 // highest order bit set.
1626 static bool isSignBit(ConstantInt *CI) {
1627 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1628 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1631 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1633 static Value *RemoveNoopCast(Value *V) {
1634 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1635 const Type *CTy = CI->getType();
1636 const Type *OpTy = CI->getOperand(0)->getType();
1637 if (CTy->isInteger() && OpTy->isInteger()) {
1638 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1639 return RemoveNoopCast(CI->getOperand(0));
1640 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1641 return RemoveNoopCast(CI->getOperand(0));
1646 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1647 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1649 if (Op0 == Op1) // sub X, X -> 0
1650 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1652 // If this is a 'B = x-(-A)', change to B = x+A...
1653 if (Value *V = dyn_castNegVal(Op1))
1654 return BinaryOperator::createAdd(Op0, V);
1656 if (isa<UndefValue>(Op0))
1657 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1658 if (isa<UndefValue>(Op1))
1659 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1661 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1662 // Replace (-1 - A) with (~A)...
1663 if (C->isAllOnesValue())
1664 return BinaryOperator::createNot(Op1);
1666 // C - ~X == X + (1+C)
1668 if (match(Op1, m_Not(m_Value(X))))
1669 return BinaryOperator::createAdd(X,
1670 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1671 // -((uint)X >> 31) -> ((int)X >> 31)
1672 // -((int)X >> 31) -> ((uint)X >> 31)
1673 if (C->isNullValue()) {
1674 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1675 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1676 if (SI->getOpcode() == Instruction::Shr)
1677 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
1679 if (SI->getType()->isSigned())
1680 NewTy = SI->getType()->getUnsignedVersion();
1682 NewTy = SI->getType()->getSignedVersion();
1683 // Check to see if we are shifting out everything but the sign bit.
1684 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
1685 // Ok, the transformation is safe. Insert a cast of the incoming
1686 // value, then the new shift, then the new cast.
1687 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
1688 SI->getOperand(0)->getName());
1689 Value *InV = InsertNewInstBefore(FirstCast, I);
1690 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
1692 if (NewShift->getType() == I.getType())
1695 InV = InsertNewInstBefore(NewShift, I);
1696 return new CastInst(NewShift, I.getType());
1702 // Try to fold constant sub into select arguments.
1703 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1704 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1707 if (isa<PHINode>(Op0))
1708 if (Instruction *NV = FoldOpIntoPhi(I))
1712 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1713 if (Op1I->getOpcode() == Instruction::Add &&
1714 !Op0->getType()->isFloatingPoint()) {
1715 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1716 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1717 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1718 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1719 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1720 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1721 // C1-(X+C2) --> (C1-C2)-X
1722 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1723 Op1I->getOperand(0));
1727 if (Op1I->hasOneUse()) {
1728 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1729 // is not used by anyone else...
1731 if (Op1I->getOpcode() == Instruction::Sub &&
1732 !Op1I->getType()->isFloatingPoint()) {
1733 // Swap the two operands of the subexpr...
1734 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1735 Op1I->setOperand(0, IIOp1);
1736 Op1I->setOperand(1, IIOp0);
1738 // Create the new top level add instruction...
1739 return BinaryOperator::createAdd(Op0, Op1);
1742 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1744 if (Op1I->getOpcode() == Instruction::And &&
1745 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1746 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1749 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1750 return BinaryOperator::createAnd(Op0, NewNot);
1753 // -(X sdiv C) -> (X sdiv -C)
1754 if (Op1I->getOpcode() == Instruction::Div)
1755 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
1756 if (CSI->isNullValue())
1757 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1758 return BinaryOperator::createDiv(Op1I->getOperand(0),
1759 ConstantExpr::getNeg(DivRHS));
1761 // X - X*C --> X * (1-C)
1762 ConstantInt *C2 = 0;
1763 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1765 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1766 return BinaryOperator::createMul(Op0, CP1);
1771 if (!Op0->getType()->isFloatingPoint())
1772 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1773 if (Op0I->getOpcode() == Instruction::Add) {
1774 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1775 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1776 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1777 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1778 } else if (Op0I->getOpcode() == Instruction::Sub) {
1779 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1780 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1784 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1785 if (X == Op1) { // X*C - X --> X * (C-1)
1786 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1787 return BinaryOperator::createMul(Op1, CP1);
1790 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1791 if (X == dyn_castFoldableMul(Op1, C2))
1792 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1797 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1798 /// really just returns true if the most significant (sign) bit is set.
1799 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1800 if (RHS->getType()->isSigned()) {
1801 // True if source is LHS < 0 or LHS <= -1
1802 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1803 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1805 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1806 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1807 // the size of the integer type.
1808 if (Opcode == Instruction::SetGE)
1809 return RHSC->getValue() ==
1810 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1811 if (Opcode == Instruction::SetGT)
1812 return RHSC->getValue() ==
1813 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1818 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1819 bool Changed = SimplifyCommutative(I);
1820 Value *Op0 = I.getOperand(0);
1822 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1823 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1825 // Simplify mul instructions with a constant RHS...
1826 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1827 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1829 // ((X << C1)*C2) == (X * (C2 << C1))
1830 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1831 if (SI->getOpcode() == Instruction::Shl)
1832 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1833 return BinaryOperator::createMul(SI->getOperand(0),
1834 ConstantExpr::getShl(CI, ShOp));
1836 if (CI->isNullValue())
1837 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1838 if (CI->equalsInt(1)) // X * 1 == X
1839 return ReplaceInstUsesWith(I, Op0);
1840 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1841 return BinaryOperator::createNeg(Op0, I.getName());
1843 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1844 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1845 uint64_t C = Log2_64(Val);
1846 return new ShiftInst(Instruction::Shl, Op0,
1847 ConstantUInt::get(Type::UByteTy, C));
1849 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1850 if (Op1F->isNullValue())
1851 return ReplaceInstUsesWith(I, Op1);
1853 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1854 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1855 if (Op1F->getValue() == 1.0)
1856 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1859 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1860 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
1861 isa<ConstantInt>(Op0I->getOperand(1))) {
1862 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
1863 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
1865 InsertNewInstBefore(Add, I);
1866 Value *C1C2 = ConstantExpr::getMul(Op1,
1867 cast<Constant>(Op0I->getOperand(1)));
1868 return BinaryOperator::createAdd(Add, C1C2);
1872 // Try to fold constant mul into select arguments.
1873 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1874 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1877 if (isa<PHINode>(Op0))
1878 if (Instruction *NV = FoldOpIntoPhi(I))
1882 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1883 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1884 return BinaryOperator::createMul(Op0v, Op1v);
1886 // If one of the operands of the multiply is a cast from a boolean value, then
1887 // we know the bool is either zero or one, so this is a 'masking' multiply.
1888 // See if we can simplify things based on how the boolean was originally
1890 CastInst *BoolCast = 0;
1891 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1892 if (CI->getOperand(0)->getType() == Type::BoolTy)
1895 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1896 if (CI->getOperand(0)->getType() == Type::BoolTy)
1899 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1900 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1901 const Type *SCOpTy = SCIOp0->getType();
1903 // If the setcc is true iff the sign bit of X is set, then convert this
1904 // multiply into a shift/and combination.
1905 if (isa<ConstantInt>(SCIOp1) &&
1906 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1907 // Shift the X value right to turn it into "all signbits".
1908 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1909 SCOpTy->getPrimitiveSizeInBits()-1);
1910 if (SCIOp0->getType()->isUnsigned()) {
1911 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1912 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1913 SCIOp0->getName()), I);
1917 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1918 BoolCast->getOperand(0)->getName()+
1921 // If the multiply type is not the same as the source type, sign extend
1922 // or truncate to the multiply type.
1923 if (I.getType() != V->getType())
1924 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1926 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1927 return BinaryOperator::createAnd(V, OtherOp);
1932 return Changed ? &I : 0;
1935 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1936 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1938 if (isa<UndefValue>(Op0)) // undef / X -> 0
1939 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1940 if (isa<UndefValue>(Op1))
1941 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1943 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1945 if (RHS->equalsInt(1))
1946 return ReplaceInstUsesWith(I, Op0);
1949 if (RHS->isAllOnesValue())
1950 return BinaryOperator::createNeg(Op0);
1952 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1953 if (LHS->getOpcode() == Instruction::Div)
1954 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1955 // (X / C1) / C2 -> X / (C1*C2)
1956 return BinaryOperator::createDiv(LHS->getOperand(0),
1957 ConstantExpr::getMul(RHS, LHSRHS));
1960 // Check to see if this is an unsigned division with an exact power of 2,
1961 // if so, convert to a right shift.
1962 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1963 if (uint64_t Val = C->getValue()) // Don't break X / 0
1964 if (isPowerOf2_64(Val)) {
1965 uint64_t C = Log2_64(Val);
1966 return new ShiftInst(Instruction::Shr, Op0,
1967 ConstantUInt::get(Type::UByteTy, C));
1971 if (RHS->getType()->isSigned())
1972 if (Value *LHSNeg = dyn_castNegVal(Op0))
1973 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1975 if (!RHS->isNullValue()) {
1976 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1977 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1979 if (isa<PHINode>(Op0))
1980 if (Instruction *NV = FoldOpIntoPhi(I))
1985 // Handle div X, Cond?Y:Z
1986 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
1987 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
1988 // same basic block, then we replace the select with Y, and the condition of
1989 // the select with false (if the cond value is in the same BB). If the
1990 // select has uses other than the div, this allows them to be simplified
1992 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
1993 if (ST->isNullValue()) {
1994 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
1995 if (CondI && CondI->getParent() == I.getParent())
1996 UpdateValueUsesWith(CondI, ConstantBool::False);
1997 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
1998 I.setOperand(1, SI->getOperand(2));
2000 UpdateValueUsesWith(SI, SI->getOperand(2));
2003 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2004 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2005 if (ST->isNullValue()) {
2006 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2007 if (CondI && CondI->getParent() == I.getParent())
2008 UpdateValueUsesWith(CondI, ConstantBool::True);
2009 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2010 I.setOperand(1, SI->getOperand(1));
2012 UpdateValueUsesWith(SI, SI->getOperand(1));
2016 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
2017 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
2018 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
2019 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
2020 // STO == 0 and SFO == 0 handled above.
2021 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
2022 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2023 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2024 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
2025 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
2026 TC, SI->getName()+".t");
2027 TSI = InsertNewInstBefore(TSI, I);
2029 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
2030 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
2031 FC, SI->getName()+".f");
2032 FSI = InsertNewInstBefore(FSI, I);
2033 return new SelectInst(SI->getOperand(0), TSI, FSI);
2038 // 0 / X == 0, we don't need to preserve faults!
2039 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2040 if (LHS->equalsInt(0))
2041 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2043 if (I.getType()->isSigned()) {
2044 // If the sign bits of both operands are zero (i.e. we can prove they are
2045 // unsigned inputs), turn this into a udiv.
2046 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2047 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2048 const Type *NTy = Op0->getType()->getUnsignedVersion();
2049 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
2050 InsertNewInstBefore(LHS, I);
2052 if (Constant *R = dyn_cast<Constant>(Op1))
2053 RHS = ConstantExpr::getCast(R, NTy);
2055 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
2056 Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
2057 InsertNewInstBefore(Div, I);
2058 return new CastInst(Div, I.getType());
2061 // Known to be an unsigned division.
2062 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2063 // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
2064 if (RHSI->getOpcode() == Instruction::Shl &&
2065 isa<ConstantUInt>(RHSI->getOperand(0))) {
2066 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
2067 if (isPowerOf2_64(C1)) {
2068 unsigned C2 = Log2_64(C1);
2069 Value *Add = RHSI->getOperand(1);
2071 Constant *C2V = ConstantUInt::get(Add->getType(), C2);
2072 Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
2075 return new ShiftInst(Instruction::Shr, Op0, Add);
2085 /// GetFactor - If we can prove that the specified value is at least a multiple
2086 /// of some factor, return that factor.
2087 static Constant *GetFactor(Value *V) {
2088 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2091 // Unless we can be tricky, we know this is a multiple of 1.
2092 Constant *Result = ConstantInt::get(V->getType(), 1);
2094 Instruction *I = dyn_cast<Instruction>(V);
2095 if (!I) return Result;
2097 if (I->getOpcode() == Instruction::Mul) {
2098 // Handle multiplies by a constant, etc.
2099 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2100 GetFactor(I->getOperand(1)));
2101 } else if (I->getOpcode() == Instruction::Shl) {
2102 // (X<<C) -> X * (1 << C)
2103 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2104 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2105 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2107 } else if (I->getOpcode() == Instruction::And) {
2108 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2109 // X & 0xFFF0 is known to be a multiple of 16.
2110 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2111 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2112 return ConstantExpr::getShl(Result,
2113 ConstantUInt::get(Type::UByteTy, Zeros));
2115 } else if (I->getOpcode() == Instruction::Cast) {
2116 Value *Op = I->getOperand(0);
2117 // Only handle int->int casts.
2118 if (!Op->getType()->isInteger()) return Result;
2119 return ConstantExpr::getCast(GetFactor(Op), V->getType());
2124 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
2125 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2127 // 0 % X == 0, we don't need to preserve faults!
2128 if (Constant *LHS = dyn_cast<Constant>(Op0))
2129 if (LHS->isNullValue())
2130 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2132 if (isa<UndefValue>(Op0)) // undef % X -> 0
2133 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2134 if (isa<UndefValue>(Op1))
2135 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2137 if (I.getType()->isSigned()) {
2138 if (Value *RHSNeg = dyn_castNegVal(Op1))
2139 if (!isa<ConstantSInt>(RHSNeg) ||
2140 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
2142 AddUsesToWorkList(I);
2143 I.setOperand(1, RHSNeg);
2147 // If the top bits of both operands are zero (i.e. we can prove they are
2148 // unsigned inputs), turn this into a urem.
2149 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2150 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2151 const Type *NTy = Op0->getType()->getUnsignedVersion();
2152 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
2153 InsertNewInstBefore(LHS, I);
2155 if (Constant *R = dyn_cast<Constant>(Op1))
2156 RHS = ConstantExpr::getCast(R, NTy);
2158 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
2159 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
2160 InsertNewInstBefore(Rem, I);
2161 return new CastInst(Rem, I.getType());
2165 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2166 // X % 0 == undef, we don't need to preserve faults!
2167 if (RHS->equalsInt(0))
2168 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2170 if (RHS->equalsInt(1)) // X % 1 == 0
2171 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2173 // Check to see if this is an unsigned remainder with an exact power of 2,
2174 // if so, convert to a bitwise and.
2175 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
2176 if (isPowerOf2_64(C->getValue()))
2177 return BinaryOperator::createAnd(Op0, SubOne(C));
2179 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2180 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2181 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2183 } else if (isa<PHINode>(Op0I)) {
2184 if (Instruction *NV = FoldOpIntoPhi(I))
2188 // X*C1%C2 --> 0 iff C1%C2 == 0
2189 if (ConstantExpr::getRem(GetFactor(Op0I), RHS)->isNullValue())
2190 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2194 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2195 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
2196 if (I.getType()->isUnsigned() &&
2197 RHSI->getOpcode() == Instruction::Shl &&
2198 isa<ConstantUInt>(RHSI->getOperand(0))) {
2199 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
2200 if (isPowerOf2_64(C1)) {
2201 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2202 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2204 return BinaryOperator::createAnd(Op0, Add);
2208 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
2209 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
2210 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2211 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2212 // the same basic block, then we replace the select with Y, and the
2213 // condition of the select with false (if the cond value is in the same
2214 // BB). If the select has uses other than the div, this allows them to be
2216 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2217 if (ST->isNullValue()) {
2218 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2219 if (CondI && CondI->getParent() == I.getParent())
2220 UpdateValueUsesWith(CondI, ConstantBool::False);
2221 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2222 I.setOperand(1, SI->getOperand(2));
2224 UpdateValueUsesWith(SI, SI->getOperand(2));
2227 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2228 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2229 if (ST->isNullValue()) {
2230 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2231 if (CondI && CondI->getParent() == I.getParent())
2232 UpdateValueUsesWith(CondI, ConstantBool::True);
2233 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2234 I.setOperand(1, SI->getOperand(1));
2236 UpdateValueUsesWith(SI, SI->getOperand(1));
2241 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
2242 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
2243 // STO == 0 and SFO == 0 handled above.
2245 if (isPowerOf2_64(STO->getValue()) && isPowerOf2_64(SFO->getValue())){
2246 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
2247 SubOne(STO), SI->getName()+".t"), I);
2248 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
2249 SubOne(SFO), SI->getName()+".f"), I);
2250 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2259 // isMaxValueMinusOne - return true if this is Max-1
2260 static bool isMaxValueMinusOne(const ConstantInt *C) {
2261 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
2262 return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
2264 const ConstantSInt *CS = cast<ConstantSInt>(C);
2266 // Calculate 0111111111..11111
2267 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2268 int64_t Val = INT64_MAX; // All ones
2269 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2270 return CS->getValue() == Val-1;
2273 // isMinValuePlusOne - return true if this is Min+1
2274 static bool isMinValuePlusOne(const ConstantInt *C) {
2275 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
2276 return CU->getValue() == 1;
2278 const ConstantSInt *CS = cast<ConstantSInt>(C);
2280 // Calculate 1111111111000000000000
2281 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2282 int64_t Val = -1; // All ones
2283 Val <<= TypeBits-1; // Shift over to the right spot
2284 return CS->getValue() == Val+1;
2287 // isOneBitSet - Return true if there is exactly one bit set in the specified
2289 static bool isOneBitSet(const ConstantInt *CI) {
2290 uint64_t V = CI->getRawValue();
2291 return V && (V & (V-1)) == 0;
2294 #if 0 // Currently unused
2295 // isLowOnes - Return true if the constant is of the form 0+1+.
2296 static bool isLowOnes(const ConstantInt *CI) {
2297 uint64_t V = CI->getRawValue();
2299 // There won't be bits set in parts that the type doesn't contain.
2300 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2302 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2303 return U && V && (U & V) == 0;
2307 // isHighOnes - Return true if the constant is of the form 1+0+.
2308 // This is the same as lowones(~X).
2309 static bool isHighOnes(const ConstantInt *CI) {
2310 uint64_t V = ~CI->getRawValue();
2311 if (~V == 0) return false; // 0's does not match "1+"
2313 // There won't be bits set in parts that the type doesn't contain.
2314 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2316 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2317 return U && V && (U & V) == 0;
2321 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2322 /// are carefully arranged to allow folding of expressions such as:
2324 /// (A < B) | (A > B) --> (A != B)
2326 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2327 /// represents that the comparison is true if A == B, and bit value '1' is true
2330 static unsigned getSetCondCode(const SetCondInst *SCI) {
2331 switch (SCI->getOpcode()) {
2333 case Instruction::SetGT: return 1;
2334 case Instruction::SetEQ: return 2;
2335 case Instruction::SetGE: return 3;
2336 case Instruction::SetLT: return 4;
2337 case Instruction::SetNE: return 5;
2338 case Instruction::SetLE: return 6;
2341 assert(0 && "Invalid SetCC opcode!");
2346 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2347 /// opcode and two operands into either a constant true or false, or a brand new
2348 /// SetCC instruction.
2349 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2351 case 0: return ConstantBool::False;
2352 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2353 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2354 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2355 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2356 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2357 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2358 case 7: return ConstantBool::True;
2359 default: assert(0 && "Illegal SetCCCode!"); return 0;
2363 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2364 struct FoldSetCCLogical {
2367 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2368 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2369 bool shouldApply(Value *V) const {
2370 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2371 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2372 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2375 Instruction *apply(BinaryOperator &Log) const {
2376 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2377 if (SCI->getOperand(0) != LHS) {
2378 assert(SCI->getOperand(1) == LHS);
2379 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2382 unsigned LHSCode = getSetCondCode(SCI);
2383 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2385 switch (Log.getOpcode()) {
2386 case Instruction::And: Code = LHSCode & RHSCode; break;
2387 case Instruction::Or: Code = LHSCode | RHSCode; break;
2388 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2389 default: assert(0 && "Illegal logical opcode!"); return 0;
2392 Value *RV = getSetCCValue(Code, LHS, RHS);
2393 if (Instruction *I = dyn_cast<Instruction>(RV))
2395 // Otherwise, it's a constant boolean value...
2396 return IC.ReplaceInstUsesWith(Log, RV);
2400 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2401 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2402 // guaranteed to be either a shift instruction or a binary operator.
2403 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2404 ConstantIntegral *OpRHS,
2405 ConstantIntegral *AndRHS,
2406 BinaryOperator &TheAnd) {
2407 Value *X = Op->getOperand(0);
2408 Constant *Together = 0;
2409 if (!isa<ShiftInst>(Op))
2410 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2412 switch (Op->getOpcode()) {
2413 case Instruction::Xor:
2414 if (Op->hasOneUse()) {
2415 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2416 std::string OpName = Op->getName(); Op->setName("");
2417 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2418 InsertNewInstBefore(And, TheAnd);
2419 return BinaryOperator::createXor(And, Together);
2422 case Instruction::Or:
2423 if (Together == AndRHS) // (X | C) & C --> C
2424 return ReplaceInstUsesWith(TheAnd, AndRHS);
2426 if (Op->hasOneUse() && Together != OpRHS) {
2427 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2428 std::string Op0Name = Op->getName(); Op->setName("");
2429 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2430 InsertNewInstBefore(Or, TheAnd);
2431 return BinaryOperator::createAnd(Or, AndRHS);
2434 case Instruction::Add:
2435 if (Op->hasOneUse()) {
2436 // Adding a one to a single bit bit-field should be turned into an XOR
2437 // of the bit. First thing to check is to see if this AND is with a
2438 // single bit constant.
2439 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
2441 // Clear bits that are not part of the constant.
2442 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2444 // If there is only one bit set...
2445 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2446 // Ok, at this point, we know that we are masking the result of the
2447 // ADD down to exactly one bit. If the constant we are adding has
2448 // no bits set below this bit, then we can eliminate the ADD.
2449 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
2451 // Check to see if any bits below the one bit set in AndRHSV are set.
2452 if ((AddRHS & (AndRHSV-1)) == 0) {
2453 // If not, the only thing that can effect the output of the AND is
2454 // the bit specified by AndRHSV. If that bit is set, the effect of
2455 // the XOR is to toggle the bit. If it is clear, then the ADD has
2457 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2458 TheAnd.setOperand(0, X);
2461 std::string Name = Op->getName(); Op->setName("");
2462 // Pull the XOR out of the AND.
2463 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2464 InsertNewInstBefore(NewAnd, TheAnd);
2465 return BinaryOperator::createXor(NewAnd, AndRHS);
2472 case Instruction::Shl: {
2473 // We know that the AND will not produce any of the bits shifted in, so if
2474 // the anded constant includes them, clear them now!
2476 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2477 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2478 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2480 if (CI == ShlMask) { // Masking out bits that the shift already masks
2481 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2482 } else if (CI != AndRHS) { // Reducing bits set in and.
2483 TheAnd.setOperand(1, CI);
2488 case Instruction::Shr:
2489 // We know that the AND will not produce any of the bits shifted in, so if
2490 // the anded constant includes them, clear them now! This only applies to
2491 // unsigned shifts, because a signed shr may bring in set bits!
2493 if (AndRHS->getType()->isUnsigned()) {
2494 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2495 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
2496 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2498 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2499 return ReplaceInstUsesWith(TheAnd, Op);
2500 } else if (CI != AndRHS) {
2501 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2504 } else { // Signed shr.
2505 // See if this is shifting in some sign extension, then masking it out
2507 if (Op->hasOneUse()) {
2508 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2509 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
2510 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2511 if (CI == AndRHS) { // Masking out bits shifted in.
2512 // Make the argument unsigned.
2513 Value *ShVal = Op->getOperand(0);
2514 ShVal = InsertCastBefore(ShVal,
2515 ShVal->getType()->getUnsignedVersion(),
2517 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
2518 OpRHS, Op->getName()),
2520 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
2521 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
2524 return new CastInst(ShVal, Op->getType());
2534 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2535 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2536 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2537 /// insert new instructions.
2538 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2539 bool Inside, Instruction &IB) {
2540 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2541 "Lo is not <= Hi in range emission code!");
2543 if (Lo == Hi) // Trivially false.
2544 return new SetCondInst(Instruction::SetNE, V, V);
2545 if (cast<ConstantIntegral>(Lo)->isMinValue())
2546 return new SetCondInst(Instruction::SetLT, V, Hi);
2548 Constant *AddCST = ConstantExpr::getNeg(Lo);
2549 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2550 InsertNewInstBefore(Add, IB);
2551 // Convert to unsigned for the comparison.
2552 const Type *UnsType = Add->getType()->getUnsignedVersion();
2553 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2554 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2555 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2556 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2559 if (Lo == Hi) // Trivially true.
2560 return new SetCondInst(Instruction::SetEQ, V, V);
2562 Hi = SubOne(cast<ConstantInt>(Hi));
2563 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
2564 return new SetCondInst(Instruction::SetGT, V, Hi);
2566 // Emit X-Lo > Hi-Lo-1
2567 Constant *AddCST = ConstantExpr::getNeg(Lo);
2568 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2569 InsertNewInstBefore(Add, IB);
2570 // Convert to unsigned for the comparison.
2571 const Type *UnsType = Add->getType()->getUnsignedVersion();
2572 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2573 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2574 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2575 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2578 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2579 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2580 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2581 // not, since all 1s are not contiguous.
2582 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2583 uint64_t V = Val->getRawValue();
2584 if (!isShiftedMask_64(V)) return false;
2586 // look for the first zero bit after the run of ones
2587 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2588 // look for the first non-zero bit
2589 ME = 64-CountLeadingZeros_64(V);
2595 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2596 /// where isSub determines whether the operator is a sub. If we can fold one of
2597 /// the following xforms:
2599 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2600 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2601 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2603 /// return (A +/- B).
2605 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2606 ConstantIntegral *Mask, bool isSub,
2608 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2609 if (!LHSI || LHSI->getNumOperands() != 2 ||
2610 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2612 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2614 switch (LHSI->getOpcode()) {
2616 case Instruction::And:
2617 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2618 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2619 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
2622 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2623 // part, we don't need any explicit masks to take them out of A. If that
2624 // is all N is, ignore it.
2626 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2627 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2629 if (MaskedValueIsZero(RHS, Mask))
2634 case Instruction::Or:
2635 case Instruction::Xor:
2636 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2637 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
2638 ConstantExpr::getAnd(N, Mask)->isNullValue())
2645 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2647 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2648 return InsertNewInstBefore(New, I);
2651 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
2652 bool Changed = SimplifyCommutative(I);
2653 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2655 if (isa<UndefValue>(Op1)) // X & undef -> 0
2656 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2660 return ReplaceInstUsesWith(I, Op1);
2662 // See if we can simplify any instructions used by the instruction whose sole
2663 // purpose is to compute bits we don't care about.
2664 uint64_t KnownZero, KnownOne;
2665 if (!isa<PackedType>(I.getType()) &&
2666 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2667 KnownZero, KnownOne))
2670 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
2671 uint64_t AndRHSMask = AndRHS->getZExtValue();
2672 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
2673 uint64_t NotAndRHS = AndRHSMask^TypeMask;
2675 // Optimize a variety of ((val OP C1) & C2) combinations...
2676 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
2677 Instruction *Op0I = cast<Instruction>(Op0);
2678 Value *Op0LHS = Op0I->getOperand(0);
2679 Value *Op0RHS = Op0I->getOperand(1);
2680 switch (Op0I->getOpcode()) {
2681 case Instruction::Xor:
2682 case Instruction::Or:
2683 // If the mask is only needed on one incoming arm, push it up.
2684 if (Op0I->hasOneUse()) {
2685 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
2686 // Not masking anything out for the LHS, move to RHS.
2687 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
2688 Op0RHS->getName()+".masked");
2689 InsertNewInstBefore(NewRHS, I);
2690 return BinaryOperator::create(
2691 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
2693 if (!isa<Constant>(Op0RHS) &&
2694 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
2695 // Not masking anything out for the RHS, move to LHS.
2696 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
2697 Op0LHS->getName()+".masked");
2698 InsertNewInstBefore(NewLHS, I);
2699 return BinaryOperator::create(
2700 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
2705 case Instruction::Add:
2706 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2707 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2708 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2709 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2710 return BinaryOperator::createAnd(V, AndRHS);
2711 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2712 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
2715 case Instruction::Sub:
2716 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2717 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2718 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2719 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2720 return BinaryOperator::createAnd(V, AndRHS);
2724 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2725 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2727 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2728 const Type *SrcTy = CI->getOperand(0)->getType();
2730 // If this is an integer truncation or change from signed-to-unsigned, and
2731 // if the source is an and/or with immediate, transform it. This
2732 // frequently occurs for bitfield accesses.
2733 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2734 if (SrcTy->getPrimitiveSizeInBits() >=
2735 I.getType()->getPrimitiveSizeInBits() &&
2736 CastOp->getNumOperands() == 2)
2737 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
2738 if (CastOp->getOpcode() == Instruction::And) {
2739 // Change: and (cast (and X, C1) to T), C2
2740 // into : and (cast X to T), trunc(C1)&C2
2741 // This will folds the two ands together, which may allow other
2743 Instruction *NewCast =
2744 new CastInst(CastOp->getOperand(0), I.getType(),
2745 CastOp->getName()+".shrunk");
2746 NewCast = InsertNewInstBefore(NewCast, I);
2748 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2749 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
2750 return BinaryOperator::createAnd(NewCast, C3);
2751 } else if (CastOp->getOpcode() == Instruction::Or) {
2752 // Change: and (cast (or X, C1) to T), C2
2753 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2754 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2755 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
2756 return ReplaceInstUsesWith(I, AndRHS);
2761 // Try to fold constant and into select arguments.
2762 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2763 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2765 if (isa<PHINode>(Op0))
2766 if (Instruction *NV = FoldOpIntoPhi(I))
2770 Value *Op0NotVal = dyn_castNotVal(Op0);
2771 Value *Op1NotVal = dyn_castNotVal(Op1);
2773 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
2774 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2776 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2777 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2778 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
2779 I.getName()+".demorgan");
2780 InsertNewInstBefore(Or, I);
2781 return BinaryOperator::createNot(Or);
2785 Value *A = 0, *B = 0;
2786 ConstantInt *C1 = 0, *C2 = 0;
2787 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
2788 if (A == Op1 || B == Op1) // (A | ?) & A --> A
2789 return ReplaceInstUsesWith(I, Op1);
2790 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
2791 if (A == Op0 || B == Op0) // A & (A | ?) --> A
2792 return ReplaceInstUsesWith(I, Op0);
2794 if (Op0->hasOneUse() &&
2795 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2796 if (A == Op1) { // (A^B)&A -> A&(A^B)
2797 I.swapOperands(); // Simplify below
2798 std::swap(Op0, Op1);
2799 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
2800 cast<BinaryOperator>(Op0)->swapOperands();
2801 I.swapOperands(); // Simplify below
2802 std::swap(Op0, Op1);
2805 if (Op1->hasOneUse() &&
2806 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2807 if (B == Op0) { // B&(A^B) -> B&(B^A)
2808 cast<BinaryOperator>(Op1)->swapOperands();
2811 if (A == Op0) { // A&(A^B) -> A & ~B
2812 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
2813 InsertNewInstBefore(NotB, I);
2814 return BinaryOperator::createAnd(A, NotB);
2820 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
2821 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2822 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2825 Value *LHSVal, *RHSVal;
2826 ConstantInt *LHSCst, *RHSCst;
2827 Instruction::BinaryOps LHSCC, RHSCC;
2828 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2829 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2830 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
2831 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2832 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2833 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2834 // Ensure that the larger constant is on the RHS.
2835 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2836 SetCondInst *LHS = cast<SetCondInst>(Op0);
2837 if (cast<ConstantBool>(Cmp)->getValue()) {
2838 std::swap(LHS, RHS);
2839 std::swap(LHSCst, RHSCst);
2840 std::swap(LHSCC, RHSCC);
2843 // At this point, we know we have have two setcc instructions
2844 // comparing a value against two constants and and'ing the result
2845 // together. Because of the above check, we know that we only have
2846 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2847 // FoldSetCCLogical check above), that the two constants are not
2849 assert(LHSCst != RHSCst && "Compares not folded above?");
2852 default: assert(0 && "Unknown integer condition code!");
2853 case Instruction::SetEQ:
2855 default: assert(0 && "Unknown integer condition code!");
2856 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
2857 case Instruction::SetGT: // (X == 13 & X > 15) -> false
2858 return ReplaceInstUsesWith(I, ConstantBool::False);
2859 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
2860 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
2861 return ReplaceInstUsesWith(I, LHS);
2863 case Instruction::SetNE:
2865 default: assert(0 && "Unknown integer condition code!");
2866 case Instruction::SetLT:
2867 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
2868 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
2869 break; // (X != 13 & X < 15) -> no change
2870 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
2871 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
2872 return ReplaceInstUsesWith(I, RHS);
2873 case Instruction::SetNE:
2874 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
2875 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2876 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2877 LHSVal->getName()+".off");
2878 InsertNewInstBefore(Add, I);
2879 const Type *UnsType = Add->getType()->getUnsignedVersion();
2880 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2881 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
2882 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2883 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2885 break; // (X != 13 & X != 15) -> no change
2888 case Instruction::SetLT:
2890 default: assert(0 && "Unknown integer condition code!");
2891 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
2892 case Instruction::SetGT: // (X < 13 & X > 15) -> false
2893 return ReplaceInstUsesWith(I, ConstantBool::False);
2894 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
2895 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
2896 return ReplaceInstUsesWith(I, LHS);
2898 case Instruction::SetGT:
2900 default: assert(0 && "Unknown integer condition code!");
2901 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
2902 return ReplaceInstUsesWith(I, LHS);
2903 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
2904 return ReplaceInstUsesWith(I, RHS);
2905 case Instruction::SetNE:
2906 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
2907 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
2908 break; // (X > 13 & X != 15) -> no change
2909 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
2910 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2916 // fold (and (cast A), (cast B)) -> (cast (and A, B))
2917 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2918 const Type *SrcTy = Op0C->getOperand(0)->getType();
2919 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2920 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
2921 // Only do this if the casts both really cause code to be generated.
2922 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
2923 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
2924 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
2925 Op1C->getOperand(0),
2927 InsertNewInstBefore(NewOp, I);
2928 return new CastInst(NewOp, I.getType());
2932 return Changed ? &I : 0;
2935 /// CollectBSwapParts - Look to see if the specified value defines a single byte
2936 /// in the result. If it does, and if the specified byte hasn't been filled in
2937 /// yet, fill it in and return false.
2938 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
2939 Instruction *I = dyn_cast<Instruction>(V);
2940 if (I == 0) return true;
2942 // If this is an or instruction, it is an inner node of the bswap.
2943 if (I->getOpcode() == Instruction::Or)
2944 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
2945 CollectBSwapParts(I->getOperand(1), ByteValues);
2947 // If this is a shift by a constant int, and it is "24", then its operand
2948 // defines a byte. We only handle unsigned types here.
2949 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
2950 // Not shifting the entire input by N-1 bytes?
2951 if (cast<ConstantInt>(I->getOperand(1))->getRawValue() !=
2952 8*(ByteValues.size()-1))
2956 if (I->getOpcode() == Instruction::Shl) {
2957 // X << 24 defines the top byte with the lowest of the input bytes.
2958 DestNo = ByteValues.size()-1;
2960 // X >>u 24 defines the low byte with the highest of the input bytes.
2964 // If the destination byte value is already defined, the values are or'd
2965 // together, which isn't a bswap (unless it's an or of the same bits).
2966 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
2968 ByteValues[DestNo] = I->getOperand(0);
2972 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
2974 Value *Shift = 0, *ShiftLHS = 0;
2975 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
2976 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
2977 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
2979 Instruction *SI = cast<Instruction>(Shift);
2981 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
2982 if (ShiftAmt->getRawValue() & 7 ||
2983 ShiftAmt->getRawValue() > 8*ByteValues.size())
2986 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
2988 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
2989 if (AndAmt->getRawValue() == uint64_t(0xFF) << 8*DestByte)
2991 // Unknown mask for bswap.
2992 if (DestByte == ByteValues.size()) return true;
2994 unsigned ShiftBytes = ShiftAmt->getRawValue()/8;
2996 if (SI->getOpcode() == Instruction::Shl)
2997 SrcByte = DestByte - ShiftBytes;
2999 SrcByte = DestByte + ShiftBytes;
3001 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3002 if (SrcByte != ByteValues.size()-DestByte-1)
3005 // If the destination byte value is already defined, the values are or'd
3006 // together, which isn't a bswap (unless it's an or of the same bits).
3007 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3009 ByteValues[DestByte] = SI->getOperand(0);
3013 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3014 /// If so, insert the new bswap intrinsic and return it.
3015 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3016 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3017 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3020 /// ByteValues - For each byte of the result, we keep track of which value
3021 /// defines each byte.
3022 std::vector<Value*> ByteValues;
3023 ByteValues.resize(I.getType()->getPrimitiveSize());
3025 // Try to find all the pieces corresponding to the bswap.
3026 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3027 CollectBSwapParts(I.getOperand(1), ByteValues))
3030 // Check to see if all of the bytes come from the same value.
3031 Value *V = ByteValues[0];
3032 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3034 // Check to make sure that all of the bytes come from the same value.
3035 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3036 if (ByteValues[i] != V)
3039 // If they do then *success* we can turn this into a bswap. Figure out what
3040 // bswap to make it into.
3041 Module *M = I.getParent()->getParent()->getParent();
3042 const char *FnName = 0;
3043 if (I.getType() == Type::UShortTy)
3044 FnName = "llvm.bswap.i16";
3045 else if (I.getType() == Type::UIntTy)
3046 FnName = "llvm.bswap.i32";
3047 else if (I.getType() == Type::ULongTy)
3048 FnName = "llvm.bswap.i64";
3050 assert(0 && "Unknown integer type!");
3051 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3053 return new CallInst(F, V);
3057 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3058 bool Changed = SimplifyCommutative(I);
3059 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3061 if (isa<UndefValue>(Op1))
3062 return ReplaceInstUsesWith(I, // X | undef -> -1
3063 ConstantIntegral::getAllOnesValue(I.getType()));
3067 return ReplaceInstUsesWith(I, Op0);
3069 // See if we can simplify any instructions used by the instruction whose sole
3070 // purpose is to compute bits we don't care about.
3071 uint64_t KnownZero, KnownOne;
3072 if (!isa<PackedType>(I.getType()) &&
3073 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3074 KnownZero, KnownOne))
3078 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3079 ConstantInt *C1 = 0; Value *X = 0;
3080 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3081 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3082 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3084 InsertNewInstBefore(Or, I);
3085 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3088 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3089 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3090 std::string Op0Name = Op0->getName(); Op0->setName("");
3091 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3092 InsertNewInstBefore(Or, I);
3093 return BinaryOperator::createXor(Or,
3094 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3097 // Try to fold constant and into select arguments.
3098 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3099 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3101 if (isa<PHINode>(Op0))
3102 if (Instruction *NV = FoldOpIntoPhi(I))
3106 Value *A = 0, *B = 0;
3107 ConstantInt *C1 = 0, *C2 = 0;
3109 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3110 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3111 return ReplaceInstUsesWith(I, Op1);
3112 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3113 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3114 return ReplaceInstUsesWith(I, Op0);
3116 // (A | B) | C and A | (B | C) -> bswap if possible.
3117 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3118 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3119 match(Op1, m_Or(m_Value(), m_Value())) ||
3120 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3121 match(Op1, m_Shift(m_Value(), m_Value())))) {
3122 if (Instruction *BSwap = MatchBSwap(I))
3126 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3127 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3128 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3129 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3131 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3134 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3135 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3136 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3137 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3139 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3142 // (A & C1)|(B & C2)
3143 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3144 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3146 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3147 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3150 // If we have: ((V + N) & C1) | (V & C2)
3151 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3152 // replace with V+N.
3153 if (C1 == ConstantExpr::getNot(C2)) {
3154 Value *V1 = 0, *V2 = 0;
3155 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
3156 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3157 // Add commutes, try both ways.
3158 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3159 return ReplaceInstUsesWith(I, A);
3160 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3161 return ReplaceInstUsesWith(I, A);
3163 // Or commutes, try both ways.
3164 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
3165 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3166 // Add commutes, try both ways.
3167 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3168 return ReplaceInstUsesWith(I, B);
3169 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3170 return ReplaceInstUsesWith(I, B);
3175 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3176 if (A == Op1) // ~A | A == -1
3177 return ReplaceInstUsesWith(I,
3178 ConstantIntegral::getAllOnesValue(I.getType()));
3182 // Note, A is still live here!
3183 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3185 return ReplaceInstUsesWith(I,
3186 ConstantIntegral::getAllOnesValue(I.getType()));
3188 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3189 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3190 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3191 I.getName()+".demorgan"), I);
3192 return BinaryOperator::createNot(And);
3196 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
3197 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
3198 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3201 Value *LHSVal, *RHSVal;
3202 ConstantInt *LHSCst, *RHSCst;
3203 Instruction::BinaryOps LHSCC, RHSCC;
3204 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3205 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3206 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
3207 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3208 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3209 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3210 // Ensure that the larger constant is on the RHS.
3211 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3212 SetCondInst *LHS = cast<SetCondInst>(Op0);
3213 if (cast<ConstantBool>(Cmp)->getValue()) {
3214 std::swap(LHS, RHS);
3215 std::swap(LHSCst, RHSCst);
3216 std::swap(LHSCC, RHSCC);
3219 // At this point, we know we have have two setcc instructions
3220 // comparing a value against two constants and or'ing the result
3221 // together. Because of the above check, we know that we only have
3222 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3223 // FoldSetCCLogical check above), that the two constants are not
3225 assert(LHSCst != RHSCst && "Compares not folded above?");
3228 default: assert(0 && "Unknown integer condition code!");
3229 case Instruction::SetEQ:
3231 default: assert(0 && "Unknown integer condition code!");
3232 case Instruction::SetEQ:
3233 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3234 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3235 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3236 LHSVal->getName()+".off");
3237 InsertNewInstBefore(Add, I);
3238 const Type *UnsType = Add->getType()->getUnsignedVersion();
3239 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3240 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3241 AddCST = ConstantExpr::getCast(AddCST, UnsType);
3242 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
3244 break; // (X == 13 | X == 15) -> no change
3246 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
3248 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
3249 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
3250 return ReplaceInstUsesWith(I, RHS);
3253 case Instruction::SetNE:
3255 default: assert(0 && "Unknown integer condition code!");
3256 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
3257 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
3258 return ReplaceInstUsesWith(I, LHS);
3259 case Instruction::SetNE: // (X != 13 | X != 15) -> true
3260 case Instruction::SetLT: // (X != 13 | X < 15) -> true
3261 return ReplaceInstUsesWith(I, ConstantBool::True);
3264 case Instruction::SetLT:
3266 default: assert(0 && "Unknown integer condition code!");
3267 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
3269 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
3270 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
3271 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
3272 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
3273 return ReplaceInstUsesWith(I, RHS);
3276 case Instruction::SetGT:
3278 default: assert(0 && "Unknown integer condition code!");
3279 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
3280 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
3281 return ReplaceInstUsesWith(I, LHS);
3282 case Instruction::SetNE: // (X > 13 | X != 15) -> true
3283 case Instruction::SetLT: // (X > 13 | X < 15) -> true
3284 return ReplaceInstUsesWith(I, ConstantBool::True);
3290 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3291 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3292 const Type *SrcTy = Op0C->getOperand(0)->getType();
3293 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3294 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3295 // Only do this if the casts both really cause code to be generated.
3296 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3297 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3298 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3299 Op1C->getOperand(0),
3301 InsertNewInstBefore(NewOp, I);
3302 return new CastInst(NewOp, I.getType());
3307 return Changed ? &I : 0;
3310 // XorSelf - Implements: X ^ X --> 0
3313 XorSelf(Value *rhs) : RHS(rhs) {}
3314 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3315 Instruction *apply(BinaryOperator &Xor) const {
3321 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3322 bool Changed = SimplifyCommutative(I);
3323 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3325 if (isa<UndefValue>(Op1))
3326 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3328 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3329 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3330 assert(Result == &I && "AssociativeOpt didn't work?");
3331 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3334 // See if we can simplify any instructions used by the instruction whose sole
3335 // purpose is to compute bits we don't care about.
3336 uint64_t KnownZero, KnownOne;
3337 if (!isa<PackedType>(I.getType()) &&
3338 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3339 KnownZero, KnownOne))
3342 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3343 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3344 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
3345 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
3346 if (RHS == ConstantBool::True && SCI->hasOneUse())
3347 return new SetCondInst(SCI->getInverseCondition(),
3348 SCI->getOperand(0), SCI->getOperand(1));
3350 // ~(c-X) == X-c-1 == X+(-c-1)
3351 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3352 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3353 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3354 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3355 ConstantInt::get(I.getType(), 1));
3356 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3359 // ~(~X & Y) --> (X | ~Y)
3360 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3361 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3362 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3364 BinaryOperator::createNot(Op0I->getOperand(1),
3365 Op0I->getOperand(1)->getName()+".not");
3366 InsertNewInstBefore(NotY, I);
3367 return BinaryOperator::createOr(Op0NotVal, NotY);
3371 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3372 if (Op0I->getOpcode() == Instruction::Add) {
3373 // ~(X-c) --> (-c-1)-X
3374 if (RHS->isAllOnesValue()) {
3375 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3376 return BinaryOperator::createSub(
3377 ConstantExpr::getSub(NegOp0CI,
3378 ConstantInt::get(I.getType(), 1)),
3379 Op0I->getOperand(0));
3381 } else if (Op0I->getOpcode() == Instruction::Or) {
3382 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3383 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3384 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3385 // Anything in both C1 and C2 is known to be zero, remove it from
3387 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3388 NewRHS = ConstantExpr::getAnd(NewRHS,
3389 ConstantExpr::getNot(CommonBits));
3390 WorkList.push_back(Op0I);
3391 I.setOperand(0, Op0I->getOperand(0));
3392 I.setOperand(1, NewRHS);
3398 // Try to fold constant and into select arguments.
3399 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3400 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3402 if (isa<PHINode>(Op0))
3403 if (Instruction *NV = FoldOpIntoPhi(I))
3407 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3409 return ReplaceInstUsesWith(I,
3410 ConstantIntegral::getAllOnesValue(I.getType()));
3412 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3414 return ReplaceInstUsesWith(I,
3415 ConstantIntegral::getAllOnesValue(I.getType()));
3417 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3418 if (Op1I->getOpcode() == Instruction::Or) {
3419 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3420 Op1I->swapOperands();
3422 std::swap(Op0, Op1);
3423 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3424 I.swapOperands(); // Simplified below.
3425 std::swap(Op0, Op1);
3427 } else if (Op1I->getOpcode() == Instruction::Xor) {
3428 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3429 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3430 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3431 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3432 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3433 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3434 Op1I->swapOperands();
3435 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3436 I.swapOperands(); // Simplified below.
3437 std::swap(Op0, Op1);
3441 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3442 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3443 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3444 Op0I->swapOperands();
3445 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3446 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3447 InsertNewInstBefore(NotB, I);
3448 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3450 } else if (Op0I->getOpcode() == Instruction::Xor) {
3451 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3452 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3453 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3454 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3455 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3456 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3457 Op0I->swapOperands();
3458 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3459 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3460 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3461 InsertNewInstBefore(N, I);
3462 return BinaryOperator::createAnd(N, Op1);
3466 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
3467 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
3468 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3471 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3472 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3473 const Type *SrcTy = Op0C->getOperand(0)->getType();
3474 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3475 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3476 // Only do this if the casts both really cause code to be generated.
3477 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3478 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3479 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
3480 Op1C->getOperand(0),
3482 InsertNewInstBefore(NewOp, I);
3483 return new CastInst(NewOp, I.getType());
3487 return Changed ? &I : 0;
3490 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
3491 /// overflowed for this type.
3492 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3494 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
3495 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
3498 static bool isPositive(ConstantInt *C) {
3499 return cast<ConstantSInt>(C)->getValue() >= 0;
3502 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
3503 /// overflowed for this type.
3504 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3506 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
3508 if (In1->getType()->isUnsigned())
3509 return cast<ConstantUInt>(Result)->getValue() <
3510 cast<ConstantUInt>(In1)->getValue();
3511 if (isPositive(In1) != isPositive(In2))
3513 if (isPositive(In1))
3514 return cast<ConstantSInt>(Result)->getValue() <
3515 cast<ConstantSInt>(In1)->getValue();
3516 return cast<ConstantSInt>(Result)->getValue() >
3517 cast<ConstantSInt>(In1)->getValue();
3520 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3521 /// code necessary to compute the offset from the base pointer (without adding
3522 /// in the base pointer). Return the result as a signed integer of intptr size.
3523 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3524 TargetData &TD = IC.getTargetData();
3525 gep_type_iterator GTI = gep_type_begin(GEP);
3526 const Type *UIntPtrTy = TD.getIntPtrType();
3527 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3528 Value *Result = Constant::getNullValue(SIntPtrTy);
3530 // Build a mask for high order bits.
3531 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3533 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3534 Value *Op = GEP->getOperand(i);
3535 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3536 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
3538 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3539 if (!OpC->isNullValue()) {
3540 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
3541 Scale = ConstantExpr::getMul(OpC, Scale);
3542 if (Constant *RC = dyn_cast<Constant>(Result))
3543 Result = ConstantExpr::getAdd(RC, Scale);
3545 // Emit an add instruction.
3546 Result = IC.InsertNewInstBefore(
3547 BinaryOperator::createAdd(Result, Scale,
3548 GEP->getName()+".offs"), I);
3552 // Convert to correct type.
3553 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
3554 Op->getName()+".c"), I);
3556 // We'll let instcombine(mul) convert this to a shl if possible.
3557 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3558 GEP->getName()+".idx"), I);
3560 // Emit an add instruction.
3561 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3562 GEP->getName()+".offs"), I);
3568 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3569 /// else. At this point we know that the GEP is on the LHS of the comparison.
3570 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3571 Instruction::BinaryOps Cond,
3573 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3575 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3576 if (isa<PointerType>(CI->getOperand(0)->getType()))
3577 RHS = CI->getOperand(0);
3579 Value *PtrBase = GEPLHS->getOperand(0);
3580 if (PtrBase == RHS) {
3581 // As an optimization, we don't actually have to compute the actual value of
3582 // OFFSET if this is a seteq or setne comparison, just return whether each
3583 // index is zero or not.
3584 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3585 Instruction *InVal = 0;
3586 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3587 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3589 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3590 if (isa<UndefValue>(C)) // undef index -> undef.
3591 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3592 if (C->isNullValue())
3594 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3595 EmitIt = false; // This is indexing into a zero sized array?
3596 } else if (isa<ConstantInt>(C))
3597 return ReplaceInstUsesWith(I, // No comparison is needed here.
3598 ConstantBool::get(Cond == Instruction::SetNE));
3603 new SetCondInst(Cond, GEPLHS->getOperand(i),
3604 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3608 InVal = InsertNewInstBefore(InVal, I);
3609 InsertNewInstBefore(Comp, I);
3610 if (Cond == Instruction::SetNE) // True if any are unequal
3611 InVal = BinaryOperator::createOr(InVal, Comp);
3612 else // True if all are equal
3613 InVal = BinaryOperator::createAnd(InVal, Comp);
3621 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
3622 ConstantBool::get(Cond == Instruction::SetEQ));
3625 // Only lower this if the setcc is the only user of the GEP or if we expect
3626 // the result to fold to a constant!
3627 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
3628 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
3629 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
3630 return new SetCondInst(Cond, Offset,
3631 Constant::getNullValue(Offset->getType()));
3633 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
3634 // If the base pointers are different, but the indices are the same, just
3635 // compare the base pointer.
3636 if (PtrBase != GEPRHS->getOperand(0)) {
3637 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
3638 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
3639 GEPRHS->getOperand(0)->getType();
3641 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3642 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3643 IndicesTheSame = false;
3647 // If all indices are the same, just compare the base pointers.
3649 return new SetCondInst(Cond, GEPLHS->getOperand(0),
3650 GEPRHS->getOperand(0));
3652 // Otherwise, the base pointers are different and the indices are
3653 // different, bail out.
3657 // If one of the GEPs has all zero indices, recurse.
3658 bool AllZeros = true;
3659 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3660 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
3661 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
3666 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
3667 SetCondInst::getSwappedCondition(Cond), I);
3669 // If the other GEP has all zero indices, recurse.
3671 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3672 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
3673 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
3678 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
3680 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
3681 // If the GEPs only differ by one index, compare it.
3682 unsigned NumDifferences = 0; // Keep track of # differences.
3683 unsigned DiffOperand = 0; // The operand that differs.
3684 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3685 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3686 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
3687 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
3688 // Irreconcilable differences.
3692 if (NumDifferences++) break;
3697 if (NumDifferences == 0) // SAME GEP?
3698 return ReplaceInstUsesWith(I, // No comparison is needed here.
3699 ConstantBool::get(Cond == Instruction::SetEQ));
3700 else if (NumDifferences == 1) {
3701 Value *LHSV = GEPLHS->getOperand(DiffOperand);
3702 Value *RHSV = GEPRHS->getOperand(DiffOperand);
3704 // Convert the operands to signed values to make sure to perform a
3705 // signed comparison.
3706 const Type *NewTy = LHSV->getType()->getSignedVersion();
3707 if (LHSV->getType() != NewTy)
3708 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
3709 LHSV->getName()), I);
3710 if (RHSV->getType() != NewTy)
3711 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
3712 RHSV->getName()), I);
3713 return new SetCondInst(Cond, LHSV, RHSV);
3717 // Only lower this if the setcc is the only user of the GEP or if we expect
3718 // the result to fold to a constant!
3719 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
3720 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
3721 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
3722 Value *L = EmitGEPOffset(GEPLHS, I, *this);
3723 Value *R = EmitGEPOffset(GEPRHS, I, *this);
3724 return new SetCondInst(Cond, L, R);
3731 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
3732 bool Changed = SimplifyCommutative(I);
3733 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3734 const Type *Ty = Op0->getType();
3738 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
3740 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
3741 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
3743 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
3744 // addresses never equal each other! We already know that Op0 != Op1.
3745 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
3746 isa<ConstantPointerNull>(Op0)) &&
3747 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
3748 isa<ConstantPointerNull>(Op1)))
3749 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
3751 // setcc's with boolean values can always be turned into bitwise operations
3752 if (Ty == Type::BoolTy) {
3753 switch (I.getOpcode()) {
3754 default: assert(0 && "Invalid setcc instruction!");
3755 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
3756 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
3757 InsertNewInstBefore(Xor, I);
3758 return BinaryOperator::createNot(Xor);
3760 case Instruction::SetNE:
3761 return BinaryOperator::createXor(Op0, Op1);
3763 case Instruction::SetGT:
3764 std::swap(Op0, Op1); // Change setgt -> setlt
3766 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
3767 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3768 InsertNewInstBefore(Not, I);
3769 return BinaryOperator::createAnd(Not, Op1);
3771 case Instruction::SetGE:
3772 std::swap(Op0, Op1); // Change setge -> setle
3774 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
3775 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3776 InsertNewInstBefore(Not, I);
3777 return BinaryOperator::createOr(Not, Op1);
3782 // See if we are doing a comparison between a constant and an instruction that
3783 // can be folded into the comparison.
3784 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3785 // Check to see if we are comparing against the minimum or maximum value...
3786 if (CI->isMinValue()) {
3787 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
3788 return ReplaceInstUsesWith(I, ConstantBool::False);
3789 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
3790 return ReplaceInstUsesWith(I, ConstantBool::True);
3791 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
3792 return BinaryOperator::createSetEQ(Op0, Op1);
3793 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
3794 return BinaryOperator::createSetNE(Op0, Op1);
3796 } else if (CI->isMaxValue()) {
3797 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
3798 return ReplaceInstUsesWith(I, ConstantBool::False);
3799 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
3800 return ReplaceInstUsesWith(I, ConstantBool::True);
3801 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
3802 return BinaryOperator::createSetEQ(Op0, Op1);
3803 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
3804 return BinaryOperator::createSetNE(Op0, Op1);
3806 // Comparing against a value really close to min or max?
3807 } else if (isMinValuePlusOne(CI)) {
3808 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
3809 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
3810 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
3811 return BinaryOperator::createSetNE(Op0, SubOne(CI));
3813 } else if (isMaxValueMinusOne(CI)) {
3814 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
3815 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
3816 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
3817 return BinaryOperator::createSetNE(Op0, AddOne(CI));
3820 // If we still have a setle or setge instruction, turn it into the
3821 // appropriate setlt or setgt instruction. Since the border cases have
3822 // already been handled above, this requires little checking.
3824 if (I.getOpcode() == Instruction::SetLE)
3825 return BinaryOperator::createSetLT(Op0, AddOne(CI));
3826 if (I.getOpcode() == Instruction::SetGE)
3827 return BinaryOperator::createSetGT(Op0, SubOne(CI));
3830 // See if we can fold the comparison based on bits known to be zero or one
3832 uint64_t KnownZero, KnownOne;
3833 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
3834 KnownZero, KnownOne, 0))
3837 // Given the known and unknown bits, compute a range that the LHS could be
3839 if (KnownOne | KnownZero) {
3840 if (Ty->isUnsigned()) { // Unsigned comparison.
3842 uint64_t RHSVal = CI->getZExtValue();
3843 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3845 switch (I.getOpcode()) { // LE/GE have been folded already.
3846 default: assert(0 && "Unknown setcc opcode!");
3847 case Instruction::SetEQ:
3848 if (Max < RHSVal || Min > RHSVal)
3849 return ReplaceInstUsesWith(I, ConstantBool::False);
3851 case Instruction::SetNE:
3852 if (Max < RHSVal || Min > RHSVal)
3853 return ReplaceInstUsesWith(I, ConstantBool::True);
3855 case Instruction::SetLT:
3856 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3857 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3859 case Instruction::SetGT:
3860 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3861 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3864 } else { // Signed comparison.
3866 int64_t RHSVal = CI->getSExtValue();
3867 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3869 switch (I.getOpcode()) { // LE/GE have been folded already.
3870 default: assert(0 && "Unknown setcc opcode!");
3871 case Instruction::SetEQ:
3872 if (Max < RHSVal || Min > RHSVal)
3873 return ReplaceInstUsesWith(I, ConstantBool::False);
3875 case Instruction::SetNE:
3876 if (Max < RHSVal || Min > RHSVal)
3877 return ReplaceInstUsesWith(I, ConstantBool::True);
3879 case Instruction::SetLT:
3880 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3881 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3883 case Instruction::SetGT:
3884 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3885 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3892 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3893 switch (LHSI->getOpcode()) {
3894 case Instruction::And:
3895 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
3896 LHSI->getOperand(0)->hasOneUse()) {
3897 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
3899 // If an operand is an AND of a truncating cast, we can widen the
3900 // and/compare to be the input width without changing the value
3901 // produced, eliminating a cast.
3902 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
3903 // We can do this transformation if either the AND constant does not
3904 // have its sign bit set or if it is an equality comparison.
3905 // Extending a relational comparison when we're checking the sign
3906 // bit would not work.
3907 if (Cast->hasOneUse() && Cast->isTruncIntCast() &&
3909 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
3910 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
3911 ConstantInt *NewCST;
3913 if (Cast->getOperand(0)->getType()->isSigned()) {
3914 NewCST = ConstantSInt::get(Cast->getOperand(0)->getType(),
3915 AndCST->getZExtValue());
3916 NewCI = ConstantSInt::get(Cast->getOperand(0)->getType(),
3917 CI->getZExtValue());
3919 NewCST = ConstantUInt::get(Cast->getOperand(0)->getType(),
3920 AndCST->getZExtValue());
3921 NewCI = ConstantUInt::get(Cast->getOperand(0)->getType(),
3922 CI->getZExtValue());
3924 Instruction *NewAnd =
3925 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
3927 InsertNewInstBefore(NewAnd, I);
3928 return new SetCondInst(I.getOpcode(), NewAnd, NewCI);
3932 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
3933 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
3934 // happens a LOT in code produced by the C front-end, for bitfield
3936 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
3938 // Check to see if there is a noop-cast between the shift and the and.
3940 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
3941 if (CI->getOperand(0)->getType()->isIntegral() &&
3942 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
3943 CI->getType()->getPrimitiveSizeInBits())
3944 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
3947 ConstantUInt *ShAmt;
3948 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
3949 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
3950 const Type *AndTy = AndCST->getType(); // Type of the and.
3952 // We can fold this as long as we can't shift unknown bits
3953 // into the mask. This can only happen with signed shift
3954 // rights, as they sign-extend.
3956 bool CanFold = Shift->isLogicalShift();
3958 // To test for the bad case of the signed shr, see if any
3959 // of the bits shifted in could be tested after the mask.
3960 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
3961 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
3963 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
3965 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
3967 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
3973 if (Shift->getOpcode() == Instruction::Shl)
3974 NewCst = ConstantExpr::getUShr(CI, ShAmt);
3976 NewCst = ConstantExpr::getShl(CI, ShAmt);
3978 // Check to see if we are shifting out any of the bits being
3980 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
3981 // If we shifted bits out, the fold is not going to work out.
3982 // As a special case, check to see if this means that the
3983 // result is always true or false now.
3984 if (I.getOpcode() == Instruction::SetEQ)
3985 return ReplaceInstUsesWith(I, ConstantBool::False);
3986 if (I.getOpcode() == Instruction::SetNE)
3987 return ReplaceInstUsesWith(I, ConstantBool::True);
3989 I.setOperand(1, NewCst);
3990 Constant *NewAndCST;
3991 if (Shift->getOpcode() == Instruction::Shl)
3992 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
3994 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
3995 LHSI->setOperand(1, NewAndCST);
3997 LHSI->setOperand(0, Shift->getOperand(0));
3999 Value *NewCast = InsertCastBefore(Shift->getOperand(0), AndTy,
4001 LHSI->setOperand(0, NewCast);
4003 WorkList.push_back(Shift); // Shift is dead.
4004 AddUsesToWorkList(I);
4010 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4011 // preferable because it allows the C<<Y expression to be hoisted out
4012 // of a loop if Y is invariant and X is not.
4013 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4014 I.isEquality() && !Shift->isArithmeticShift() &&
4015 isa<Instruction>(Shift->getOperand(0))) {
4018 if (Shift->getOpcode() == Instruction::Shr) {
4019 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4022 // Make sure we insert a logical shift.
4023 Constant *NewAndCST = AndCST;
4024 if (AndCST->getType()->isSigned())
4025 NewAndCST = ConstantExpr::getCast(AndCST,
4026 AndCST->getType()->getUnsignedVersion());
4027 NS = new ShiftInst(Instruction::Shr, NewAndCST,
4028 Shift->getOperand(1), "tmp");
4030 InsertNewInstBefore(cast<Instruction>(NS), I);
4032 // If C's sign doesn't agree with the and, insert a cast now.
4033 if (NS->getType() != LHSI->getType())
4034 NS = InsertCastBefore(NS, LHSI->getType(), I);
4036 Value *ShiftOp = Shift->getOperand(0);
4037 if (ShiftOp->getType() != LHSI->getType())
4038 ShiftOp = InsertCastBefore(ShiftOp, LHSI->getType(), I);
4040 // Compute X & (C << Y).
4041 Instruction *NewAnd =
4042 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4043 InsertNewInstBefore(NewAnd, I);
4045 I.setOperand(0, NewAnd);
4051 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
4052 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
4053 if (I.isEquality()) {
4054 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4056 // Check that the shift amount is in range. If not, don't perform
4057 // undefined shifts. When the shift is visited it will be
4059 if (ShAmt->getValue() >= TypeBits)
4062 // If we are comparing against bits always shifted out, the
4063 // comparison cannot succeed.
4065 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
4066 if (Comp != CI) {// Comparing against a bit that we know is zero.
4067 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4068 Constant *Cst = ConstantBool::get(IsSetNE);
4069 return ReplaceInstUsesWith(I, Cst);
4072 if (LHSI->hasOneUse()) {
4073 // Otherwise strength reduce the shift into an and.
4074 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
4075 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4078 if (CI->getType()->isUnsigned()) {
4079 Mask = ConstantUInt::get(CI->getType(), Val);
4080 } else if (ShAmtVal != 0) {
4081 Mask = ConstantSInt::get(CI->getType(), Val);
4083 Mask = ConstantInt::getAllOnesValue(CI->getType());
4087 BinaryOperator::createAnd(LHSI->getOperand(0),
4088 Mask, LHSI->getName()+".mask");
4089 Value *And = InsertNewInstBefore(AndI, I);
4090 return new SetCondInst(I.getOpcode(), And,
4091 ConstantExpr::getUShr(CI, ShAmt));
4097 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
4098 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
4099 if (I.isEquality()) {
4100 // Check that the shift amount is in range. If not, don't perform
4101 // undefined shifts. When the shift is visited it will be
4103 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4104 if (ShAmt->getValue() >= TypeBits)
4107 // If we are comparing against bits always shifted out, the
4108 // comparison cannot succeed.
4110 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
4112 if (Comp != CI) {// Comparing against a bit that we know is zero.
4113 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4114 Constant *Cst = ConstantBool::get(IsSetNE);
4115 return ReplaceInstUsesWith(I, Cst);
4118 if (LHSI->hasOneUse() || CI->isNullValue()) {
4119 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
4121 // Otherwise strength reduce the shift into an and.
4122 uint64_t Val = ~0ULL; // All ones.
4123 Val <<= ShAmtVal; // Shift over to the right spot.
4126 if (CI->getType()->isUnsigned()) {
4127 Val &= ~0ULL >> (64-TypeBits);
4128 Mask = ConstantUInt::get(CI->getType(), Val);
4130 Mask = ConstantSInt::get(CI->getType(), Val);
4134 BinaryOperator::createAnd(LHSI->getOperand(0),
4135 Mask, LHSI->getName()+".mask");
4136 Value *And = InsertNewInstBefore(AndI, I);
4137 return new SetCondInst(I.getOpcode(), And,
4138 ConstantExpr::getShl(CI, ShAmt));
4144 case Instruction::Div:
4145 // Fold: (div X, C1) op C2 -> range check
4146 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4147 // Fold this div into the comparison, producing a range check.
4148 // Determine, based on the divide type, what the range is being
4149 // checked. If there is an overflow on the low or high side, remember
4150 // it, otherwise compute the range [low, hi) bounding the new value.
4151 bool LoOverflow = false, HiOverflow = 0;
4152 ConstantInt *LoBound = 0, *HiBound = 0;
4155 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
4157 Instruction::BinaryOps Opcode = I.getOpcode();
4159 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
4160 } else if (LHSI->getType()->isUnsigned()) { // udiv
4162 LoOverflow = ProdOV;
4163 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4164 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4165 if (CI->isNullValue()) { // (X / pos) op 0
4167 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4169 } else if (isPositive(CI)) { // (X / pos) op pos
4171 LoOverflow = ProdOV;
4172 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4173 } else { // (X / pos) op neg
4174 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4175 LoOverflow = AddWithOverflow(LoBound, Prod,
4176 cast<ConstantInt>(DivRHSH));
4178 HiOverflow = ProdOV;
4180 } else { // Divisor is < 0.
4181 if (CI->isNullValue()) { // (X / neg) op 0
4182 LoBound = AddOne(DivRHS);
4183 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4184 if (HiBound == DivRHS)
4185 LoBound = 0; // - INTMIN = INTMIN
4186 } else if (isPositive(CI)) { // (X / neg) op pos
4187 HiOverflow = LoOverflow = ProdOV;
4189 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4190 HiBound = AddOne(Prod);
4191 } else { // (X / neg) op neg
4193 LoOverflow = HiOverflow = ProdOV;
4194 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4197 // Dividing by a negate swaps the condition.
4198 Opcode = SetCondInst::getSwappedCondition(Opcode);
4202 Value *X = LHSI->getOperand(0);
4204 default: assert(0 && "Unhandled setcc opcode!");
4205 case Instruction::SetEQ:
4206 if (LoOverflow && HiOverflow)
4207 return ReplaceInstUsesWith(I, ConstantBool::False);
4208 else if (HiOverflow)
4209 return new SetCondInst(Instruction::SetGE, X, LoBound);
4210 else if (LoOverflow)
4211 return new SetCondInst(Instruction::SetLT, X, HiBound);
4213 return InsertRangeTest(X, LoBound, HiBound, true, I);
4214 case Instruction::SetNE:
4215 if (LoOverflow && HiOverflow)
4216 return ReplaceInstUsesWith(I, ConstantBool::True);
4217 else if (HiOverflow)
4218 return new SetCondInst(Instruction::SetLT, X, LoBound);
4219 else if (LoOverflow)
4220 return new SetCondInst(Instruction::SetGE, X, HiBound);
4222 return InsertRangeTest(X, LoBound, HiBound, false, I);
4223 case Instruction::SetLT:
4225 return ReplaceInstUsesWith(I, ConstantBool::False);
4226 return new SetCondInst(Instruction::SetLT, X, LoBound);
4227 case Instruction::SetGT:
4229 return ReplaceInstUsesWith(I, ConstantBool::False);
4230 return new SetCondInst(Instruction::SetGE, X, HiBound);
4237 // Simplify seteq and setne instructions...
4238 if (I.isEquality()) {
4239 bool isSetNE = I.getOpcode() == Instruction::SetNE;
4241 // If the first operand is (and|or|xor) with a constant, and the second
4242 // operand is a constant, simplify a bit.
4243 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4244 switch (BO->getOpcode()) {
4245 case Instruction::Rem:
4246 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4247 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
4249 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
4250 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
4251 if (isPowerOf2_64(V)) {
4252 unsigned L2 = Log2_64(V);
4253 const Type *UTy = BO->getType()->getUnsignedVersion();
4254 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
4256 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
4257 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
4258 RHSCst, BO->getName()), I);
4259 return BinaryOperator::create(I.getOpcode(), NewRem,
4260 Constant::getNullValue(UTy));
4265 case Instruction::Add:
4266 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4267 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4268 if (BO->hasOneUse())
4269 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4270 ConstantExpr::getSub(CI, BOp1C));
4271 } else if (CI->isNullValue()) {
4272 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4273 // efficiently invertible, or if the add has just this one use.
4274 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4276 if (Value *NegVal = dyn_castNegVal(BOp1))
4277 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
4278 else if (Value *NegVal = dyn_castNegVal(BOp0))
4279 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
4280 else if (BO->hasOneUse()) {
4281 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4283 InsertNewInstBefore(Neg, I);
4284 return new SetCondInst(I.getOpcode(), BOp0, Neg);
4288 case Instruction::Xor:
4289 // For the xor case, we can xor two constants together, eliminating
4290 // the explicit xor.
4291 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4292 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
4293 ConstantExpr::getXor(CI, BOC));
4296 case Instruction::Sub:
4297 // Replace (([sub|xor] A, B) != 0) with (A != B)
4298 if (CI->isNullValue())
4299 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4303 case Instruction::Or:
4304 // If bits are being or'd in that are not present in the constant we
4305 // are comparing against, then the comparison could never succeed!
4306 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4307 Constant *NotCI = ConstantExpr::getNot(CI);
4308 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4309 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4313 case Instruction::And:
4314 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4315 // If bits are being compared against that are and'd out, then the
4316 // comparison can never succeed!
4317 if (!ConstantExpr::getAnd(CI,
4318 ConstantExpr::getNot(BOC))->isNullValue())
4319 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4321 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4322 if (CI == BOC && isOneBitSet(CI))
4323 return new SetCondInst(isSetNE ? Instruction::SetEQ :
4324 Instruction::SetNE, Op0,
4325 Constant::getNullValue(CI->getType()));
4327 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
4328 // to be a signed value as appropriate.
4329 if (isSignBit(BOC)) {
4330 Value *X = BO->getOperand(0);
4331 // If 'X' is not signed, insert a cast now...
4332 if (!BOC->getType()->isSigned()) {
4333 const Type *DestTy = BOC->getType()->getSignedVersion();
4334 X = InsertCastBefore(X, DestTy, I);
4336 return new SetCondInst(isSetNE ? Instruction::SetLT :
4337 Instruction::SetGE, X,
4338 Constant::getNullValue(X->getType()));
4341 // ((X & ~7) == 0) --> X < 8
4342 if (CI->isNullValue() && isHighOnes(BOC)) {
4343 Value *X = BO->getOperand(0);
4344 Constant *NegX = ConstantExpr::getNeg(BOC);
4346 // If 'X' is signed, insert a cast now.
4347 if (NegX->getType()->isSigned()) {
4348 const Type *DestTy = NegX->getType()->getUnsignedVersion();
4349 X = InsertCastBefore(X, DestTy, I);
4350 NegX = ConstantExpr::getCast(NegX, DestTy);
4353 return new SetCondInst(isSetNE ? Instruction::SetGE :
4354 Instruction::SetLT, X, NegX);
4361 } else { // Not a SetEQ/SetNE
4362 // If the LHS is a cast from an integral value of the same size,
4363 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
4364 Value *CastOp = Cast->getOperand(0);
4365 const Type *SrcTy = CastOp->getType();
4366 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
4367 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
4368 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
4369 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
4370 "Source and destination signednesses should differ!");
4371 if (Cast->getType()->isSigned()) {
4372 // If this is a signed comparison, check for comparisons in the
4373 // vicinity of zero.
4374 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
4376 return BinaryOperator::createSetGT(CastOp,
4377 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
4378 else if (I.getOpcode() == Instruction::SetGT &&
4379 cast<ConstantSInt>(CI)->getValue() == -1)
4380 // X > -1 => x < 128
4381 return BinaryOperator::createSetLT(CastOp,
4382 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
4384 ConstantUInt *CUI = cast<ConstantUInt>(CI);
4385 if (I.getOpcode() == Instruction::SetLT &&
4386 CUI->getValue() == 1ULL << (SrcTySize-1))
4387 // X < 128 => X > -1
4388 return BinaryOperator::createSetGT(CastOp,
4389 ConstantSInt::get(SrcTy, -1));
4390 else if (I.getOpcode() == Instruction::SetGT &&
4391 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
4393 return BinaryOperator::createSetLT(CastOp,
4394 Constant::getNullValue(SrcTy));
4401 // Handle setcc with constant RHS's that can be integer, FP or pointer.
4402 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4403 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4404 switch (LHSI->getOpcode()) {
4405 case Instruction::GetElementPtr:
4406 if (RHSC->isNullValue()) {
4407 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
4408 bool isAllZeros = true;
4409 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4410 if (!isa<Constant>(LHSI->getOperand(i)) ||
4411 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4416 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
4417 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4421 case Instruction::PHI:
4422 if (Instruction *NV = FoldOpIntoPhi(I))
4425 case Instruction::Select:
4426 // If either operand of the select is a constant, we can fold the
4427 // comparison into the select arms, which will cause one to be
4428 // constant folded and the select turned into a bitwise or.
4429 Value *Op1 = 0, *Op2 = 0;
4430 if (LHSI->hasOneUse()) {
4431 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4432 // Fold the known value into the constant operand.
4433 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4434 // Insert a new SetCC of the other select operand.
4435 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4436 LHSI->getOperand(2), RHSC,
4438 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4439 // Fold the known value into the constant operand.
4440 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4441 // Insert a new SetCC of the other select operand.
4442 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4443 LHSI->getOperand(1), RHSC,
4449 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4454 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
4455 if (User *GEP = dyn_castGetElementPtr(Op0))
4456 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
4458 if (User *GEP = dyn_castGetElementPtr(Op1))
4459 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
4460 SetCondInst::getSwappedCondition(I.getOpcode()), I))
4463 // Test to see if the operands of the setcc are casted versions of other
4464 // values. If the cast can be stripped off both arguments, we do so now.
4465 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4466 Value *CastOp0 = CI->getOperand(0);
4467 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
4468 (isa<Constant>(Op1) || isa<CastInst>(Op1)) && I.isEquality()) {
4469 // We keep moving the cast from the left operand over to the right
4470 // operand, where it can often be eliminated completely.
4473 // If operand #1 is a cast instruction, see if we can eliminate it as
4475 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
4476 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
4478 Op1 = CI2->getOperand(0);
4480 // If Op1 is a constant, we can fold the cast into the constant.
4481 if (Op1->getType() != Op0->getType())
4482 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4483 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
4485 // Otherwise, cast the RHS right before the setcc
4486 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
4487 InsertNewInstBefore(cast<Instruction>(Op1), I);
4489 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
4492 // Handle the special case of: setcc (cast bool to X), <cst>
4493 // This comes up when you have code like
4496 // For generality, we handle any zero-extension of any operand comparison
4497 // with a constant or another cast from the same type.
4498 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4499 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
4503 if (I.isEquality()) {
4505 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4506 (A == Op1 || B == Op1)) {
4507 // (A^B) == A -> B == 0
4508 Value *OtherVal = A == Op1 ? B : A;
4509 return BinaryOperator::create(I.getOpcode(), OtherVal,
4510 Constant::getNullValue(A->getType()));
4511 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4512 (A == Op0 || B == Op0)) {
4513 // A == (A^B) -> B == 0
4514 Value *OtherVal = A == Op0 ? B : A;
4515 return BinaryOperator::create(I.getOpcode(), OtherVal,
4516 Constant::getNullValue(A->getType()));
4517 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4518 // (A-B) == A -> B == 0
4519 return BinaryOperator::create(I.getOpcode(), B,
4520 Constant::getNullValue(B->getType()));
4521 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4522 // A == (A-B) -> B == 0
4523 return BinaryOperator::create(I.getOpcode(), B,
4524 Constant::getNullValue(B->getType()));
4527 return Changed ? &I : 0;
4530 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
4531 // We only handle extending casts so far.
4533 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
4534 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
4535 const Type *SrcTy = LHSCIOp->getType();
4536 const Type *DestTy = SCI.getOperand(0)->getType();
4539 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
4542 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
4543 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
4544 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
4546 // Is this a sign or zero extension?
4547 bool isSignSrc = SrcTy->isSigned();
4548 bool isSignDest = DestTy->isSigned();
4550 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
4551 // Not an extension from the same type?
4552 RHSCIOp = CI->getOperand(0);
4553 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
4554 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
4555 // Compute the constant that would happen if we truncated to SrcTy then
4556 // reextended to DestTy.
4557 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
4559 if (ConstantExpr::getCast(Res, DestTy) == CI) {
4562 // If the value cannot be represented in the shorter type, we cannot emit
4563 // a simple comparison.
4564 if (SCI.getOpcode() == Instruction::SetEQ)
4565 return ReplaceInstUsesWith(SCI, ConstantBool::False);
4566 if (SCI.getOpcode() == Instruction::SetNE)
4567 return ReplaceInstUsesWith(SCI, ConstantBool::True);
4569 // Evaluate the comparison for LT.
4571 if (DestTy->isSigned()) {
4572 // We're performing a signed comparison.
4574 // Signed extend and signed comparison.
4575 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
4576 Result = ConstantBool::False;
4578 Result = ConstantBool::True; // X < (large) --> true
4580 // Unsigned extend and signed comparison.
4581 if (cast<ConstantSInt>(CI)->getValue() < 0)
4582 Result = ConstantBool::False;
4584 Result = ConstantBool::True;
4587 // We're performing an unsigned comparison.
4589 // Unsigned extend & compare -> always true.
4590 Result = ConstantBool::True;
4592 // We're performing an unsigned comp with a sign extended value.
4593 // This is true if the input is >= 0. [aka >s -1]
4594 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
4595 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
4596 NegOne, SCI.getName()), SCI);
4600 // Finally, return the value computed.
4601 if (SCI.getOpcode() == Instruction::SetLT) {
4602 return ReplaceInstUsesWith(SCI, Result);
4604 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
4605 if (Constant *CI = dyn_cast<Constant>(Result))
4606 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
4608 return BinaryOperator::createNot(Result);
4615 // Okay, just insert a compare of the reduced operands now!
4616 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
4619 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
4620 assert(I.getOperand(1)->getType() == Type::UByteTy);
4621 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4622 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4624 // shl X, 0 == X and shr X, 0 == X
4625 // shl 0, X == 0 and shr 0, X == 0
4626 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
4627 Op0 == Constant::getNullValue(Op0->getType()))
4628 return ReplaceInstUsesWith(I, Op0);
4630 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
4631 if (!isLeftShift && I.getType()->isSigned())
4632 return ReplaceInstUsesWith(I, Op0);
4633 else // undef << X -> 0 AND undef >>u X -> 0
4634 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4636 if (isa<UndefValue>(Op1)) {
4637 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
4638 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4640 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
4643 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
4645 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
4646 if (CSI->isAllOnesValue())
4647 return ReplaceInstUsesWith(I, CSI);
4649 // Try to fold constant and into select arguments.
4650 if (isa<Constant>(Op0))
4651 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
4652 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4655 // See if we can turn a signed shr into an unsigned shr.
4656 if (I.isArithmeticShift()) {
4657 if (MaskedValueIsZero(Op0,
4658 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
4659 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
4660 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
4662 return new CastInst(V, I.getType());
4666 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
4667 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
4672 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
4674 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4675 bool isSignedShift = Op0->getType()->isSigned();
4676 bool isUnsignedShift = !isSignedShift;
4678 // See if we can simplify any instructions used by the instruction whose sole
4679 // purpose is to compute bits we don't care about.
4680 uint64_t KnownZero, KnownOne;
4681 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
4682 KnownZero, KnownOne))
4685 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
4686 // of a signed value.
4688 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
4689 if (Op1->getValue() >= TypeBits) {
4690 if (isUnsignedShift || isLeftShift)
4691 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
4693 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
4698 // ((X*C1) << C2) == (X * (C1 << C2))
4699 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
4700 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
4701 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
4702 return BinaryOperator::createMul(BO->getOperand(0),
4703 ConstantExpr::getShl(BOOp, Op1));
4705 // Try to fold constant and into select arguments.
4706 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4707 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4709 if (isa<PHINode>(Op0))
4710 if (Instruction *NV = FoldOpIntoPhi(I))
4713 if (Op0->hasOneUse()) {
4714 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
4715 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4718 switch (Op0BO->getOpcode()) {
4720 case Instruction::Add:
4721 case Instruction::And:
4722 case Instruction::Or:
4723 case Instruction::Xor:
4724 // These operators commute.
4725 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
4726 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4727 match(Op0BO->getOperand(1),
4728 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4729 Instruction *YS = new ShiftInst(Instruction::Shl,
4730 Op0BO->getOperand(0), Op1,
4732 InsertNewInstBefore(YS, I); // (Y << C)
4734 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
4735 Op0BO->getOperand(1)->getName());
4736 InsertNewInstBefore(X, I); // (X + (Y << C))
4737 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4738 C2 = ConstantExpr::getShl(C2, Op1);
4739 return BinaryOperator::createAnd(X, C2);
4742 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
4743 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4744 match(Op0BO->getOperand(1),
4745 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4746 m_ConstantInt(CC))) && V2 == Op1 &&
4747 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
4748 Instruction *YS = new ShiftInst(Instruction::Shl,
4749 Op0BO->getOperand(0), Op1,
4751 InsertNewInstBefore(YS, I); // (Y << C)
4753 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4754 V1->getName()+".mask");
4755 InsertNewInstBefore(XM, I); // X & (CC << C)
4757 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
4761 case Instruction::Sub:
4762 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4763 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4764 match(Op0BO->getOperand(0),
4765 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4766 Instruction *YS = new ShiftInst(Instruction::Shl,
4767 Op0BO->getOperand(1), Op1,
4769 InsertNewInstBefore(YS, I); // (Y << C)
4771 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
4772 Op0BO->getOperand(0)->getName());
4773 InsertNewInstBefore(X, I); // (X + (Y << C))
4774 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4775 C2 = ConstantExpr::getShl(C2, Op1);
4776 return BinaryOperator::createAnd(X, C2);
4779 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
4780 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4781 match(Op0BO->getOperand(0),
4782 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4783 m_ConstantInt(CC))) && V2 == Op1 &&
4784 cast<BinaryOperator>(Op0BO->getOperand(0))
4785 ->getOperand(0)->hasOneUse()) {
4786 Instruction *YS = new ShiftInst(Instruction::Shl,
4787 Op0BO->getOperand(1), Op1,
4789 InsertNewInstBefore(YS, I); // (Y << C)
4791 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4792 V1->getName()+".mask");
4793 InsertNewInstBefore(XM, I); // X & (CC << C)
4795 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
4802 // If the operand is an bitwise operator with a constant RHS, and the
4803 // shift is the only use, we can pull it out of the shift.
4804 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
4805 bool isValid = true; // Valid only for And, Or, Xor
4806 bool highBitSet = false; // Transform if high bit of constant set?
4808 switch (Op0BO->getOpcode()) {
4809 default: isValid = false; break; // Do not perform transform!
4810 case Instruction::Add:
4811 isValid = isLeftShift;
4813 case Instruction::Or:
4814 case Instruction::Xor:
4817 case Instruction::And:
4822 // If this is a signed shift right, and the high bit is modified
4823 // by the logical operation, do not perform the transformation.
4824 // The highBitSet boolean indicates the value of the high bit of
4825 // the constant which would cause it to be modified for this
4828 if (isValid && !isLeftShift && isSignedShift) {
4829 uint64_t Val = Op0C->getRawValue();
4830 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
4834 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
4836 Instruction *NewShift =
4837 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
4840 InsertNewInstBefore(NewShift, I);
4842 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
4849 // Find out if this is a shift of a shift by a constant.
4850 ShiftInst *ShiftOp = 0;
4851 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
4853 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4854 // If this is a noop-integer case of a shift instruction, use the shift.
4855 if (CI->getOperand(0)->getType()->isInteger() &&
4856 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
4857 CI->getType()->getPrimitiveSizeInBits() &&
4858 isa<ShiftInst>(CI->getOperand(0))) {
4859 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
4863 if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
4864 // Find the operands and properties of the input shift. Note that the
4865 // signedness of the input shift may differ from the current shift if there
4866 // is a noop cast between the two.
4867 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
4868 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
4869 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
4871 ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
4873 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
4874 unsigned ShiftAmt2 = (unsigned)Op1->getValue();
4876 // Check for (A << c1) << c2 and (A >> c1) >> c2.
4877 if (isLeftShift == isShiftOfLeftShift) {
4878 // Do not fold these shifts if the first one is signed and the second one
4879 // is unsigned and this is a right shift. Further, don't do any folding
4881 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
4884 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
4885 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
4886 Amt = Op0->getType()->getPrimitiveSizeInBits();
4888 Value *Op = ShiftOp->getOperand(0);
4889 if (isShiftOfSignedShift != isSignedShift)
4890 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
4891 return new ShiftInst(I.getOpcode(), Op,
4892 ConstantUInt::get(Type::UByteTy, Amt));
4895 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
4896 // signed types, we can only support the (A >> c1) << c2 configuration,
4897 // because it can not turn an arbitrary bit of A into a sign bit.
4898 if (isUnsignedShift || isLeftShift) {
4899 // Calculate bitmask for what gets shifted off the edge.
4900 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
4902 C = ConstantExpr::getShl(C, ShiftAmt1C);
4904 C = ConstantExpr::getUShr(C, ShiftAmt1C);
4906 Value *Op = ShiftOp->getOperand(0);
4907 if (isShiftOfSignedShift != isSignedShift)
4908 Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
4911 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
4912 InsertNewInstBefore(Mask, I);
4914 // Figure out what flavor of shift we should use...
4915 if (ShiftAmt1 == ShiftAmt2) {
4916 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
4917 } else if (ShiftAmt1 < ShiftAmt2) {
4918 return new ShiftInst(I.getOpcode(), Mask,
4919 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
4920 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
4921 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
4922 // Make sure to emit an unsigned shift right, not a signed one.
4923 Mask = InsertNewInstBefore(new CastInst(Mask,
4924 Mask->getType()->getUnsignedVersion(),
4926 Mask = new ShiftInst(Instruction::Shr, Mask,
4927 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4928 InsertNewInstBefore(Mask, I);
4929 return new CastInst(Mask, I.getType());
4931 return new ShiftInst(ShiftOp->getOpcode(), Mask,
4932 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4935 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
4936 Op = InsertNewInstBefore(new CastInst(Mask,
4937 I.getType()->getSignedVersion(),
4938 Mask->getName()), I);
4939 Instruction *Shift =
4940 new ShiftInst(ShiftOp->getOpcode(), Op,
4941 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4942 InsertNewInstBefore(Shift, I);
4944 C = ConstantIntegral::getAllOnesValue(Shift->getType());
4945 C = ConstantExpr::getShl(C, Op1);
4946 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
4947 InsertNewInstBefore(Mask, I);
4948 return new CastInst(Mask, I.getType());
4951 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
4952 // this case, C1 == C2 and C1 is 8, 16, or 32.
4953 if (ShiftAmt1 == ShiftAmt2) {
4954 const Type *SExtType = 0;
4955 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
4956 case 8 : SExtType = Type::SByteTy; break;
4957 case 16: SExtType = Type::ShortTy; break;
4958 case 32: SExtType = Type::IntTy; break;
4962 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
4964 InsertNewInstBefore(NewTrunc, I);
4965 return new CastInst(NewTrunc, I.getType());
4974 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
4975 /// expression. If so, decompose it, returning some value X, such that Val is
4978 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
4980 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
4981 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
4982 Offset = CI->getValue();
4984 return ConstantUInt::get(Type::UIntTy, 0);
4985 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
4986 if (I->getNumOperands() == 2) {
4987 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
4988 if (I->getOpcode() == Instruction::Shl) {
4989 // This is a value scaled by '1 << the shift amt'.
4990 Scale = 1U << CUI->getValue();
4992 return I->getOperand(0);
4993 } else if (I->getOpcode() == Instruction::Mul) {
4994 // This value is scaled by 'CUI'.
4995 Scale = CUI->getValue();
4997 return I->getOperand(0);
4998 } else if (I->getOpcode() == Instruction::Add) {
4999 // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
5002 Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
5004 Offset += CUI->getValue();
5005 if (SubScale > 1 && (Offset % SubScale == 0)) {
5014 // Otherwise, we can't look past this.
5021 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5022 /// try to eliminate the cast by moving the type information into the alloc.
5023 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5024 AllocationInst &AI) {
5025 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5026 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5028 // Remove any uses of AI that are dead.
5029 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5030 std::vector<Instruction*> DeadUsers;
5031 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5032 Instruction *User = cast<Instruction>(*UI++);
5033 if (isInstructionTriviallyDead(User)) {
5034 while (UI != E && *UI == User)
5035 ++UI; // If this instruction uses AI more than once, don't break UI.
5037 // Add operands to the worklist.
5038 AddUsesToWorkList(*User);
5040 DEBUG(std::cerr << "IC: DCE: " << *User);
5042 User->eraseFromParent();
5043 removeFromWorkList(User);
5047 // Get the type really allocated and the type casted to.
5048 const Type *AllocElTy = AI.getAllocatedType();
5049 const Type *CastElTy = PTy->getElementType();
5050 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5052 unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
5053 unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
5054 if (CastElTyAlign < AllocElTyAlign) return 0;
5056 // If the allocation has multiple uses, only promote it if we are strictly
5057 // increasing the alignment of the resultant allocation. If we keep it the
5058 // same, we open the door to infinite loops of various kinds.
5059 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5061 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5062 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5063 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5065 // See if we can satisfy the modulus by pulling a scale out of the array
5067 unsigned ArraySizeScale, ArrayOffset;
5068 Value *NumElements = // See if the array size is a decomposable linear expr.
5069 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5071 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5073 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5074 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5076 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5081 Amt = ConstantUInt::get(Type::UIntTy, Scale);
5082 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
5083 Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
5084 else if (Scale != 1) {
5085 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5086 Amt = InsertNewInstBefore(Tmp, AI);
5090 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5091 Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
5092 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5093 Amt = InsertNewInstBefore(Tmp, AI);
5096 std::string Name = AI.getName(); AI.setName("");
5097 AllocationInst *New;
5098 if (isa<MallocInst>(AI))
5099 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5101 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5102 InsertNewInstBefore(New, AI);
5104 // If the allocation has multiple uses, insert a cast and change all things
5105 // that used it to use the new cast. This will also hack on CI, but it will
5107 if (!AI.hasOneUse()) {
5108 AddUsesToWorkList(AI);
5109 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
5110 InsertNewInstBefore(NewCast, AI);
5111 AI.replaceAllUsesWith(NewCast);
5113 return ReplaceInstUsesWith(CI, New);
5116 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5117 /// and return it without inserting any new casts. This is used by code that
5118 /// tries to decide whether promoting or shrinking integer operations to wider
5119 /// or smaller types will allow us to eliminate a truncate or extend.
5120 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5121 int &NumCastsRemoved) {
5122 if (isa<Constant>(V)) return true;
5124 Instruction *I = dyn_cast<Instruction>(V);
5125 if (!I || !I->hasOneUse()) return false;
5127 switch (I->getOpcode()) {
5128 case Instruction::And:
5129 case Instruction::Or:
5130 case Instruction::Xor:
5131 // These operators can all arbitrarily be extended or truncated.
5132 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5133 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5134 case Instruction::Cast:
5135 // If this is a cast from the destination type, we can trivially eliminate
5136 // it, and this will remove a cast overall.
5137 if (I->getOperand(0)->getType() == Ty) {
5138 // If the first operand is itself a cast, and is eliminable, do not count
5139 // this as an eliminable cast. We would prefer to eliminate those two
5141 if (CastInst *OpCast = dyn_cast<CastInst>(I->getOperand(0)))
5147 // TODO: Can handle more cases here.
5154 /// EvaluateInDifferentType - Given an expression that
5155 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5156 /// evaluate the expression.
5157 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty) {
5158 if (Constant *C = dyn_cast<Constant>(V))
5159 return ConstantExpr::getCast(C, Ty);
5161 // Otherwise, it must be an instruction.
5162 Instruction *I = cast<Instruction>(V);
5163 Instruction *Res = 0;
5164 switch (I->getOpcode()) {
5165 case Instruction::And:
5166 case Instruction::Or:
5167 case Instruction::Xor: {
5168 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
5169 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty);
5170 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5171 LHS, RHS, I->getName());
5174 case Instruction::Cast:
5175 // If this is a cast from the destination type, return the input.
5176 if (I->getOperand(0)->getType() == Ty)
5177 return I->getOperand(0);
5179 // TODO: Can handle more cases here.
5180 assert(0 && "Unreachable!");
5184 return InsertNewInstBefore(Res, *I);
5188 // CastInst simplification
5190 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
5191 Value *Src = CI.getOperand(0);
5193 // If the user is casting a value to the same type, eliminate this cast
5195 if (CI.getType() == Src->getType())
5196 return ReplaceInstUsesWith(CI, Src);
5198 if (isa<UndefValue>(Src)) // cast undef -> undef
5199 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5201 // If casting the result of another cast instruction, try to eliminate this
5204 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5205 Value *A = CSrc->getOperand(0);
5206 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
5207 CI.getType(), TD)) {
5208 // This instruction now refers directly to the cast's src operand. This
5209 // has a good chance of making CSrc dead.
5210 CI.setOperand(0, CSrc->getOperand(0));
5214 // If this is an A->B->A cast, and we are dealing with integral types, try
5215 // to convert this into a logical 'and' instruction.
5217 if (A->getType()->isInteger() &&
5218 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
5219 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
5220 CSrc->getType()->getPrimitiveSizeInBits() <
5221 CI.getType()->getPrimitiveSizeInBits()&&
5222 A->getType()->getPrimitiveSizeInBits() ==
5223 CI.getType()->getPrimitiveSizeInBits()) {
5224 assert(CSrc->getType() != Type::ULongTy &&
5225 "Cannot have type bigger than ulong!");
5226 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
5227 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
5229 AndOp = ConstantExpr::getCast(AndOp, A->getType());
5230 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
5231 if (And->getType() != CI.getType()) {
5232 And->setName(CSrc->getName()+".mask");
5233 InsertNewInstBefore(And, CI);
5234 And = new CastInst(And, CI.getType());
5240 // If this is a cast to bool, turn it into the appropriate setne instruction.
5241 if (CI.getType() == Type::BoolTy)
5242 return BinaryOperator::createSetNE(CI.getOperand(0),
5243 Constant::getNullValue(CI.getOperand(0)->getType()));
5245 // See if we can simplify any instructions used by the LHS whose sole
5246 // purpose is to compute bits we don't care about.
5247 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral()) {
5248 uint64_t KnownZero, KnownOne;
5249 if (SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask(),
5250 KnownZero, KnownOne))
5254 // If casting the result of a getelementptr instruction with no offset, turn
5255 // this into a cast of the original pointer!
5257 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5258 bool AllZeroOperands = true;
5259 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5260 if (!isa<Constant>(GEP->getOperand(i)) ||
5261 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5262 AllZeroOperands = false;
5265 if (AllZeroOperands) {
5266 CI.setOperand(0, GEP->getOperand(0));
5271 // If we are casting a malloc or alloca to a pointer to a type of the same
5272 // size, rewrite the allocation instruction to allocate the "right" type.
5274 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5275 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5278 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5279 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5281 if (isa<PHINode>(Src))
5282 if (Instruction *NV = FoldOpIntoPhi(CI))
5285 // If the source and destination are pointers, and this cast is equivalent to
5286 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
5287 // This can enhance SROA and other transforms that want type-safe pointers.
5288 if (const PointerType *DstPTy = dyn_cast<PointerType>(CI.getType()))
5289 if (const PointerType *SrcPTy = dyn_cast<PointerType>(Src->getType())) {
5290 const Type *DstTy = DstPTy->getElementType();
5291 const Type *SrcTy = SrcPTy->getElementType();
5293 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
5294 unsigned NumZeros = 0;
5295 while (SrcTy != DstTy &&
5296 isa<CompositeType>(SrcTy) && !isa<PointerType>(SrcTy) &&
5297 SrcTy->getNumContainedTypes() /* not "{}" */) {
5298 SrcTy = cast<CompositeType>(SrcTy)->getTypeAtIndex(ZeroUInt);
5302 // If we found a path from the src to dest, create the getelementptr now.
5303 if (SrcTy == DstTy) {
5304 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
5305 return new GetElementPtrInst(Src, Idxs);
5309 // If the source value is an instruction with only this use, we can attempt to
5310 // propagate the cast into the instruction. Also, only handle integral types
5312 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
5313 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
5314 CI.getType()->isInteger()) { // Don't mess with casts to bool here
5316 int NumCastsRemoved = 0;
5317 if (CanEvaluateInDifferentType(SrcI, CI.getType(), NumCastsRemoved)) {
5318 // If this cast is a truncate, evaluting in a different type always
5319 // eliminates the cast, so it is always a win. If this is a noop-cast
5320 // this just removes a noop cast which isn't pointful, but simplifies
5321 // the code. If this is a zero-extension, we need to do an AND to
5322 // maintain the clear top-part of the computation, so we require that
5323 // the input have eliminated at least one cast. If this is a sign
5324 // extension, we insert two new casts (to do the extension) so we
5325 // require that two casts have been eliminated.
5327 switch (getCastType(Src->getType(), CI.getType())) {
5328 default: assert(0 && "Unknown cast type!");
5334 DoXForm = NumCastsRemoved >= 1;
5337 DoXForm = NumCastsRemoved >= 2;
5342 Value *Res = EvaluateInDifferentType(SrcI, CI.getType());
5343 assert(Res->getType() == CI.getType());
5344 switch (getCastType(Src->getType(), CI.getType())) {
5345 default: assert(0 && "Unknown cast type!");
5348 // Just replace this cast with the result.
5349 return ReplaceInstUsesWith(CI, Res);
5351 // We need to emit an AND to clear the high bits.
5352 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5353 unsigned DestBitSize = CI.getType()->getPrimitiveSizeInBits();
5354 assert(SrcBitSize < DestBitSize && "Not a zext?");
5355 Constant *C = ConstantUInt::get(Type::ULongTy, (1 << SrcBitSize)-1);
5356 C = ConstantExpr::getCast(C, CI.getType());
5357 return BinaryOperator::createAnd(Res, C);
5360 // We need to emit a cast to truncate, then a cast to sext.
5361 return new CastInst(InsertCastBefore(Res, Src->getType(), CI),
5367 const Type *DestTy = CI.getType();
5368 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5369 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5371 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
5372 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
5374 switch (SrcI->getOpcode()) {
5375 case Instruction::Add:
5376 case Instruction::Mul:
5377 case Instruction::And:
5378 case Instruction::Or:
5379 case Instruction::Xor:
5380 // If we are discarding information, or just changing the sign, rewrite.
5381 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
5382 // Don't insert two casts if they cannot be eliminated. We allow two
5383 // casts to be inserted if the sizes are the same. This could only be
5384 // converting signedness, which is a noop.
5385 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
5386 !ValueRequiresCast(Op0, DestTy, TD)) {
5387 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5388 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5389 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
5390 ->getOpcode(), Op0c, Op1c);
5394 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
5395 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
5396 Op1 == ConstantBool::True &&
5397 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
5398 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
5399 return BinaryOperator::createXor(New,
5400 ConstantInt::get(CI.getType(), 1));
5403 case Instruction::Shl:
5404 // Allow changing the sign of the source operand. Do not allow changing
5405 // the size of the shift, UNLESS the shift amount is a constant. We
5406 // mush not change variable sized shifts to a smaller size, because it
5407 // is undefined to shift more bits out than exist in the value.
5408 if (DestBitSize == SrcBitSize ||
5409 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
5410 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5411 return new ShiftInst(Instruction::Shl, Op0c, Op1);
5414 case Instruction::Shr:
5415 // If this is a signed shr, and if all bits shifted in are about to be
5416 // truncated off, turn it into an unsigned shr to allow greater
5418 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
5419 isa<ConstantInt>(Op1)) {
5420 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
5421 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
5422 // Convert to unsigned.
5423 Value *N1 = InsertOperandCastBefore(Op0,
5424 Op0->getType()->getUnsignedVersion(), &CI);
5425 // Insert the new shift, which is now unsigned.
5426 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
5427 Op1, Src->getName()), CI);
5428 return new CastInst(N1, CI.getType());
5433 case Instruction::SetEQ:
5434 case Instruction::SetNE:
5435 // We if we are just checking for a seteq of a single bit and casting it
5436 // to an integer. If so, shift the bit to the appropriate place then
5437 // cast to integer to avoid the comparison.
5438 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
5439 uint64_t Op1CV = Op1C->getZExtValue();
5440 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
5441 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5442 // cast (X == 1) to int --> X iff X has only the low bit set.
5443 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
5444 // cast (X != 0) to int --> X iff X has only the low bit set.
5445 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
5446 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
5447 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5448 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
5449 // If Op1C some other power of two, convert:
5450 uint64_t KnownZero, KnownOne;
5451 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
5452 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
5454 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly one possible 1?
5455 bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
5456 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
5457 // (X&4) == 2 --> false
5458 // (X&4) != 2 --> true
5459 Constant *Res = ConstantBool::get(isSetNE);
5460 Res = ConstantExpr::getCast(Res, CI.getType());
5461 return ReplaceInstUsesWith(CI, Res);
5464 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
5467 // Perform an unsigned shr by shiftamt. Convert input to
5468 // unsigned if it is signed.
5469 if (In->getType()->isSigned())
5470 In = InsertNewInstBefore(new CastInst(In,
5471 In->getType()->getUnsignedVersion(), In->getName()),CI);
5472 // Insert the shift to put the result in the low bit.
5473 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
5474 ConstantInt::get(Type::UByteTy, ShiftAmt),
5475 In->getName()+".lobit"), CI);
5478 if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
5479 Constant *One = ConstantInt::get(In->getType(), 1);
5480 In = BinaryOperator::createXor(In, One, "tmp");
5481 InsertNewInstBefore(cast<Instruction>(In), CI);
5484 if (CI.getType() == In->getType())
5485 return ReplaceInstUsesWith(CI, In);
5487 return new CastInst(In, CI.getType());
5495 if (SrcI->hasOneUse()) {
5496 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(SrcI)) {
5497 // Okay, we have (cast (shuffle ..)). We know this cast is a bitconvert
5498 // because the inputs are known to be a vector. Check to see if this is
5499 // a cast to a vector with the same # elts.
5500 if (isa<PackedType>(CI.getType()) &&
5501 cast<PackedType>(CI.getType())->getNumElements() ==
5502 SVI->getType()->getNumElements()) {
5504 // If either of the operands is a cast from CI.getType(), then
5505 // evaluating the shuffle in the casted destination's type will allow
5506 // us to eliminate at least one cast.
5507 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
5508 Tmp->getOperand(0)->getType() == CI.getType()) ||
5509 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
5510 Tmp->getOperand(0)->getType() == CI.getType())) {
5511 Value *LHS = InsertOperandCastBefore(SVI->getOperand(0),
5513 Value *RHS = InsertOperandCastBefore(SVI->getOperand(1),
5515 // Return a new shuffle vector. Use the same element ID's, as we
5516 // know the vector types match #elts.
5517 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
5527 /// GetSelectFoldableOperands - We want to turn code that looks like this:
5529 /// %D = select %cond, %C, %A
5531 /// %C = select %cond, %B, 0
5534 /// Assuming that the specified instruction is an operand to the select, return
5535 /// a bitmask indicating which operands of this instruction are foldable if they
5536 /// equal the other incoming value of the select.
5538 static unsigned GetSelectFoldableOperands(Instruction *I) {
5539 switch (I->getOpcode()) {
5540 case Instruction::Add:
5541 case Instruction::Mul:
5542 case Instruction::And:
5543 case Instruction::Or:
5544 case Instruction::Xor:
5545 return 3; // Can fold through either operand.
5546 case Instruction::Sub: // Can only fold on the amount subtracted.
5547 case Instruction::Shl: // Can only fold on the shift amount.
5548 case Instruction::Shr:
5551 return 0; // Cannot fold
5555 /// GetSelectFoldableConstant - For the same transformation as the previous
5556 /// function, return the identity constant that goes into the select.
5557 static Constant *GetSelectFoldableConstant(Instruction *I) {
5558 switch (I->getOpcode()) {
5559 default: assert(0 && "This cannot happen!"); abort();
5560 case Instruction::Add:
5561 case Instruction::Sub:
5562 case Instruction::Or:
5563 case Instruction::Xor:
5564 return Constant::getNullValue(I->getType());
5565 case Instruction::Shl:
5566 case Instruction::Shr:
5567 return Constant::getNullValue(Type::UByteTy);
5568 case Instruction::And:
5569 return ConstantInt::getAllOnesValue(I->getType());
5570 case Instruction::Mul:
5571 return ConstantInt::get(I->getType(), 1);
5575 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
5576 /// have the same opcode and only one use each. Try to simplify this.
5577 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
5579 if (TI->getNumOperands() == 1) {
5580 // If this is a non-volatile load or a cast from the same type,
5582 if (TI->getOpcode() == Instruction::Cast) {
5583 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
5586 return 0; // unknown unary op.
5589 // Fold this by inserting a select from the input values.
5590 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
5591 FI->getOperand(0), SI.getName()+".v");
5592 InsertNewInstBefore(NewSI, SI);
5593 return new CastInst(NewSI, TI->getType());
5596 // Only handle binary operators here.
5597 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
5600 // Figure out if the operations have any operands in common.
5601 Value *MatchOp, *OtherOpT, *OtherOpF;
5603 if (TI->getOperand(0) == FI->getOperand(0)) {
5604 MatchOp = TI->getOperand(0);
5605 OtherOpT = TI->getOperand(1);
5606 OtherOpF = FI->getOperand(1);
5607 MatchIsOpZero = true;
5608 } else if (TI->getOperand(1) == FI->getOperand(1)) {
5609 MatchOp = TI->getOperand(1);
5610 OtherOpT = TI->getOperand(0);
5611 OtherOpF = FI->getOperand(0);
5612 MatchIsOpZero = false;
5613 } else if (!TI->isCommutative()) {
5615 } else if (TI->getOperand(0) == FI->getOperand(1)) {
5616 MatchOp = TI->getOperand(0);
5617 OtherOpT = TI->getOperand(1);
5618 OtherOpF = FI->getOperand(0);
5619 MatchIsOpZero = true;
5620 } else if (TI->getOperand(1) == FI->getOperand(0)) {
5621 MatchOp = TI->getOperand(1);
5622 OtherOpT = TI->getOperand(0);
5623 OtherOpF = FI->getOperand(1);
5624 MatchIsOpZero = true;
5629 // If we reach here, they do have operations in common.
5630 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
5631 OtherOpF, SI.getName()+".v");
5632 InsertNewInstBefore(NewSI, SI);
5634 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
5636 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
5638 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
5641 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
5643 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
5647 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
5648 Value *CondVal = SI.getCondition();
5649 Value *TrueVal = SI.getTrueValue();
5650 Value *FalseVal = SI.getFalseValue();
5652 // select true, X, Y -> X
5653 // select false, X, Y -> Y
5654 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
5655 if (C == ConstantBool::True)
5656 return ReplaceInstUsesWith(SI, TrueVal);
5658 assert(C == ConstantBool::False);
5659 return ReplaceInstUsesWith(SI, FalseVal);
5662 // select C, X, X -> X
5663 if (TrueVal == FalseVal)
5664 return ReplaceInstUsesWith(SI, TrueVal);
5666 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
5667 return ReplaceInstUsesWith(SI, FalseVal);
5668 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
5669 return ReplaceInstUsesWith(SI, TrueVal);
5670 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
5671 if (isa<Constant>(TrueVal))
5672 return ReplaceInstUsesWith(SI, TrueVal);
5674 return ReplaceInstUsesWith(SI, FalseVal);
5677 if (SI.getType() == Type::BoolTy)
5678 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
5679 if (C == ConstantBool::True) {
5680 // Change: A = select B, true, C --> A = or B, C
5681 return BinaryOperator::createOr(CondVal, FalseVal);
5683 // Change: A = select B, false, C --> A = and !B, C
5685 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5686 "not."+CondVal->getName()), SI);
5687 return BinaryOperator::createAnd(NotCond, FalseVal);
5689 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
5690 if (C == ConstantBool::False) {
5691 // Change: A = select B, C, false --> A = and B, C
5692 return BinaryOperator::createAnd(CondVal, TrueVal);
5694 // Change: A = select B, C, true --> A = or !B, C
5696 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5697 "not."+CondVal->getName()), SI);
5698 return BinaryOperator::createOr(NotCond, TrueVal);
5702 // Selecting between two integer constants?
5703 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
5704 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
5705 // select C, 1, 0 -> cast C to int
5706 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
5707 return new CastInst(CondVal, SI.getType());
5708 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
5709 // select C, 0, 1 -> cast !C to int
5711 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5712 "not."+CondVal->getName()), SI);
5713 return new CastInst(NotCond, SI.getType());
5716 // If one of the constants is zero (we know they can't both be) and we
5717 // have a setcc instruction with zero, and we have an 'and' with the
5718 // non-constant value, eliminate this whole mess. This corresponds to
5719 // cases like this: ((X & 27) ? 27 : 0)
5720 if (TrueValC->isNullValue() || FalseValC->isNullValue())
5721 if (SetCondInst *IC = dyn_cast<SetCondInst>(SI.getCondition()))
5722 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
5723 cast<Constant>(IC->getOperand(1))->isNullValue())
5724 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
5725 if (ICA->getOpcode() == Instruction::And &&
5726 isa<ConstantInt>(ICA->getOperand(1)) &&
5727 (ICA->getOperand(1) == TrueValC ||
5728 ICA->getOperand(1) == FalseValC) &&
5729 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
5730 // Okay, now we know that everything is set up, we just don't
5731 // know whether we have a setne or seteq and whether the true or
5732 // false val is the zero.
5733 bool ShouldNotVal = !TrueValC->isNullValue();
5734 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
5737 V = InsertNewInstBefore(BinaryOperator::create(
5738 Instruction::Xor, V, ICA->getOperand(1)), SI);
5739 return ReplaceInstUsesWith(SI, V);
5743 // See if we are selecting two values based on a comparison of the two values.
5744 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
5745 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
5746 // Transform (X == Y) ? X : Y -> Y
5747 if (SCI->getOpcode() == Instruction::SetEQ)
5748 return ReplaceInstUsesWith(SI, FalseVal);
5749 // Transform (X != Y) ? X : Y -> X
5750 if (SCI->getOpcode() == Instruction::SetNE)
5751 return ReplaceInstUsesWith(SI, TrueVal);
5752 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5754 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
5755 // Transform (X == Y) ? Y : X -> X
5756 if (SCI->getOpcode() == Instruction::SetEQ)
5757 return ReplaceInstUsesWith(SI, FalseVal);
5758 // Transform (X != Y) ? Y : X -> Y
5759 if (SCI->getOpcode() == Instruction::SetNE)
5760 return ReplaceInstUsesWith(SI, TrueVal);
5761 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5765 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
5766 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
5767 if (TI->hasOneUse() && FI->hasOneUse()) {
5768 bool isInverse = false;
5769 Instruction *AddOp = 0, *SubOp = 0;
5771 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
5772 if (TI->getOpcode() == FI->getOpcode())
5773 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
5776 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
5777 // even legal for FP.
5778 if (TI->getOpcode() == Instruction::Sub &&
5779 FI->getOpcode() == Instruction::Add) {
5780 AddOp = FI; SubOp = TI;
5781 } else if (FI->getOpcode() == Instruction::Sub &&
5782 TI->getOpcode() == Instruction::Add) {
5783 AddOp = TI; SubOp = FI;
5787 Value *OtherAddOp = 0;
5788 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
5789 OtherAddOp = AddOp->getOperand(1);
5790 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
5791 OtherAddOp = AddOp->getOperand(0);
5795 // So at this point we know we have (Y -> OtherAddOp):
5796 // select C, (add X, Y), (sub X, Z)
5797 Value *NegVal; // Compute -Z
5798 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
5799 NegVal = ConstantExpr::getNeg(C);
5801 NegVal = InsertNewInstBefore(
5802 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
5805 Value *NewTrueOp = OtherAddOp;
5806 Value *NewFalseOp = NegVal;
5808 std::swap(NewTrueOp, NewFalseOp);
5809 Instruction *NewSel =
5810 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
5812 NewSel = InsertNewInstBefore(NewSel, SI);
5813 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
5818 // See if we can fold the select into one of our operands.
5819 if (SI.getType()->isInteger()) {
5820 // See the comment above GetSelectFoldableOperands for a description of the
5821 // transformation we are doing here.
5822 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
5823 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
5824 !isa<Constant>(FalseVal))
5825 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
5826 unsigned OpToFold = 0;
5827 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
5829 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
5834 Constant *C = GetSelectFoldableConstant(TVI);
5835 std::string Name = TVI->getName(); TVI->setName("");
5836 Instruction *NewSel =
5837 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
5839 InsertNewInstBefore(NewSel, SI);
5840 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
5841 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
5842 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
5843 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
5845 assert(0 && "Unknown instruction!!");
5850 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
5851 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
5852 !isa<Constant>(TrueVal))
5853 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
5854 unsigned OpToFold = 0;
5855 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
5857 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
5862 Constant *C = GetSelectFoldableConstant(FVI);
5863 std::string Name = FVI->getName(); FVI->setName("");
5864 Instruction *NewSel =
5865 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
5867 InsertNewInstBefore(NewSel, SI);
5868 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
5869 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
5870 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
5871 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
5873 assert(0 && "Unknown instruction!!");
5879 if (BinaryOperator::isNot(CondVal)) {
5880 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
5881 SI.setOperand(1, FalseVal);
5882 SI.setOperand(2, TrueVal);
5889 /// GetKnownAlignment - If the specified pointer has an alignment that we can
5890 /// determine, return it, otherwise return 0.
5891 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
5892 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
5893 unsigned Align = GV->getAlignment();
5894 if (Align == 0 && TD)
5895 Align = TD->getTypeAlignment(GV->getType()->getElementType());
5897 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
5898 unsigned Align = AI->getAlignment();
5899 if (Align == 0 && TD) {
5900 if (isa<AllocaInst>(AI))
5901 Align = TD->getTypeAlignment(AI->getType()->getElementType());
5902 else if (isa<MallocInst>(AI)) {
5903 // Malloc returns maximally aligned memory.
5904 Align = TD->getTypeAlignment(AI->getType()->getElementType());
5905 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
5906 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::LongTy));
5910 } else if (isa<CastInst>(V) ||
5911 (isa<ConstantExpr>(V) &&
5912 cast<ConstantExpr>(V)->getOpcode() == Instruction::Cast)) {
5913 User *CI = cast<User>(V);
5914 if (isa<PointerType>(CI->getOperand(0)->getType()))
5915 return GetKnownAlignment(CI->getOperand(0), TD);
5917 } else if (isa<GetElementPtrInst>(V) ||
5918 (isa<ConstantExpr>(V) &&
5919 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
5920 User *GEPI = cast<User>(V);
5921 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
5922 if (BaseAlignment == 0) return 0;
5924 // If all indexes are zero, it is just the alignment of the base pointer.
5925 bool AllZeroOperands = true;
5926 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
5927 if (!isa<Constant>(GEPI->getOperand(i)) ||
5928 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
5929 AllZeroOperands = false;
5932 if (AllZeroOperands)
5933 return BaseAlignment;
5935 // Otherwise, if the base alignment is >= the alignment we expect for the
5936 // base pointer type, then we know that the resultant pointer is aligned at
5937 // least as much as its type requires.
5940 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
5941 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
5943 const Type *GEPTy = GEPI->getType();
5944 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
5952 /// visitCallInst - CallInst simplification. This mostly only handles folding
5953 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
5954 /// the heavy lifting.
5956 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
5957 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
5958 if (!II) return visitCallSite(&CI);
5960 // Intrinsics cannot occur in an invoke, so handle them here instead of in
5962 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
5963 bool Changed = false;
5965 // memmove/cpy/set of zero bytes is a noop.
5966 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
5967 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
5969 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
5970 if (CI->getRawValue() == 1) {
5971 // Replace the instruction with just byte operations. We would
5972 // transform other cases to loads/stores, but we don't know if
5973 // alignment is sufficient.
5977 // If we have a memmove and the source operation is a constant global,
5978 // then the source and dest pointers can't alias, so we can change this
5979 // into a call to memcpy.
5980 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
5981 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
5982 if (GVSrc->isConstant()) {
5983 Module *M = CI.getParent()->getParent()->getParent();
5985 if (CI.getCalledFunction()->getFunctionType()->getParamType(3) ==
5987 Name = "llvm.memcpy.i32";
5989 Name = "llvm.memcpy.i64";
5990 Function *MemCpy = M->getOrInsertFunction(Name,
5991 CI.getCalledFunction()->getFunctionType());
5992 CI.setOperand(0, MemCpy);
5997 // If we can determine a pointer alignment that is bigger than currently
5998 // set, update the alignment.
5999 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6000 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6001 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6002 unsigned Align = std::min(Alignment1, Alignment2);
6003 if (MI->getAlignment()->getRawValue() < Align) {
6004 MI->setAlignment(ConstantUInt::get(Type::UIntTy, Align));
6007 } else if (isa<MemSetInst>(MI)) {
6008 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6009 if (MI->getAlignment()->getRawValue() < Alignment) {
6010 MI->setAlignment(ConstantUInt::get(Type::UIntTy, Alignment));
6015 if (Changed) return II;
6017 switch (II->getIntrinsicID()) {
6019 case Intrinsic::ppc_altivec_lvx:
6020 case Intrinsic::ppc_altivec_lvxl:
6021 case Intrinsic::x86_sse_loadu_ps:
6022 case Intrinsic::x86_sse2_loadu_pd:
6023 case Intrinsic::x86_sse2_loadu_dq:
6024 // Turn PPC lvx -> load if the pointer is known aligned.
6025 // Turn X86 loadups -> load if the pointer is known aligned.
6026 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6027 Value *Ptr = InsertCastBefore(II->getOperand(1),
6028 PointerType::get(II->getType()), CI);
6029 return new LoadInst(Ptr);
6032 case Intrinsic::ppc_altivec_stvx:
6033 case Intrinsic::ppc_altivec_stvxl:
6034 // Turn stvx -> store if the pointer is known aligned.
6035 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
6036 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
6037 Value *Ptr = InsertCastBefore(II->getOperand(2), OpPtrTy, CI);
6038 return new StoreInst(II->getOperand(1), Ptr);
6041 case Intrinsic::x86_sse_storeu_ps:
6042 case Intrinsic::x86_sse2_storeu_pd:
6043 case Intrinsic::x86_sse2_storeu_dq:
6044 case Intrinsic::x86_sse2_storel_dq:
6045 // Turn X86 storeu -> store if the pointer is known aligned.
6046 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6047 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
6048 Value *Ptr = InsertCastBefore(II->getOperand(1), OpPtrTy, CI);
6049 return new StoreInst(II->getOperand(2), Ptr);
6052 case Intrinsic::ppc_altivec_vperm:
6053 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
6054 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
6055 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
6057 // Check that all of the elements are integer constants or undefs.
6058 bool AllEltsOk = true;
6059 for (unsigned i = 0; i != 16; ++i) {
6060 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
6061 !isa<UndefValue>(Mask->getOperand(i))) {
6068 // Cast the input vectors to byte vectors.
6069 Value *Op0 = InsertCastBefore(II->getOperand(1), Mask->getType(), CI);
6070 Value *Op1 = InsertCastBefore(II->getOperand(2), Mask->getType(), CI);
6071 Value *Result = UndefValue::get(Op0->getType());
6073 // Only extract each element once.
6074 Value *ExtractedElts[32];
6075 memset(ExtractedElts, 0, sizeof(ExtractedElts));
6077 for (unsigned i = 0; i != 16; ++i) {
6078 if (isa<UndefValue>(Mask->getOperand(i)))
6080 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getRawValue();
6081 Idx &= 31; // Match the hardware behavior.
6083 if (ExtractedElts[Idx] == 0) {
6085 new ExtractElementInst(Idx < 16 ? Op0 : Op1,
6086 ConstantUInt::get(Type::UIntTy, Idx&15),
6088 InsertNewInstBefore(Elt, CI);
6089 ExtractedElts[Idx] = Elt;
6092 // Insert this value into the result vector.
6093 Result = new InsertElementInst(Result, ExtractedElts[Idx],
6094 ConstantUInt::get(Type::UIntTy, i),
6096 InsertNewInstBefore(cast<Instruction>(Result), CI);
6098 return new CastInst(Result, CI.getType());
6103 case Intrinsic::stackrestore: {
6104 // If the save is right next to the restore, remove the restore. This can
6105 // happen when variable allocas are DCE'd.
6106 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
6107 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
6108 BasicBlock::iterator BI = SS;
6110 return EraseInstFromFunction(CI);
6114 // If the stack restore is in a return/unwind block and if there are no
6115 // allocas or calls between the restore and the return, nuke the restore.
6116 TerminatorInst *TI = II->getParent()->getTerminator();
6117 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
6118 BasicBlock::iterator BI = II;
6119 bool CannotRemove = false;
6120 for (++BI; &*BI != TI; ++BI) {
6121 if (isa<AllocaInst>(BI) ||
6122 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
6123 CannotRemove = true;
6128 return EraseInstFromFunction(CI);
6135 return visitCallSite(II);
6138 // InvokeInst simplification
6140 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
6141 return visitCallSite(&II);
6144 // visitCallSite - Improvements for call and invoke instructions.
6146 Instruction *InstCombiner::visitCallSite(CallSite CS) {
6147 bool Changed = false;
6149 // If the callee is a constexpr cast of a function, attempt to move the cast
6150 // to the arguments of the call/invoke.
6151 if (transformConstExprCastCall(CS)) return 0;
6153 Value *Callee = CS.getCalledValue();
6155 if (Function *CalleeF = dyn_cast<Function>(Callee))
6156 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
6157 Instruction *OldCall = CS.getInstruction();
6158 // If the call and callee calling conventions don't match, this call must
6159 // be unreachable, as the call is undefined.
6160 new StoreInst(ConstantBool::True,
6161 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
6162 if (!OldCall->use_empty())
6163 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
6164 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
6165 return EraseInstFromFunction(*OldCall);
6169 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
6170 // This instruction is not reachable, just remove it. We insert a store to
6171 // undef so that we know that this code is not reachable, despite the fact
6172 // that we can't modify the CFG here.
6173 new StoreInst(ConstantBool::True,
6174 UndefValue::get(PointerType::get(Type::BoolTy)),
6175 CS.getInstruction());
6177 if (!CS.getInstruction()->use_empty())
6178 CS.getInstruction()->
6179 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
6181 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
6182 // Don't break the CFG, insert a dummy cond branch.
6183 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
6184 ConstantBool::True, II);
6186 return EraseInstFromFunction(*CS.getInstruction());
6189 const PointerType *PTy = cast<PointerType>(Callee->getType());
6190 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
6191 if (FTy->isVarArg()) {
6192 // See if we can optimize any arguments passed through the varargs area of
6194 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
6195 E = CS.arg_end(); I != E; ++I)
6196 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
6197 // If this cast does not effect the value passed through the varargs
6198 // area, we can eliminate the use of the cast.
6199 Value *Op = CI->getOperand(0);
6200 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
6207 return Changed ? CS.getInstruction() : 0;
6210 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
6211 // attempt to move the cast to the arguments of the call/invoke.
6213 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
6214 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
6215 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
6216 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
6218 Function *Callee = cast<Function>(CE->getOperand(0));
6219 Instruction *Caller = CS.getInstruction();
6221 // Okay, this is a cast from a function to a different type. Unless doing so
6222 // would cause a type conversion of one of our arguments, change this call to
6223 // be a direct call with arguments casted to the appropriate types.
6225 const FunctionType *FT = Callee->getFunctionType();
6226 const Type *OldRetTy = Caller->getType();
6228 // Check to see if we are changing the return type...
6229 if (OldRetTy != FT->getReturnType()) {
6230 if (Callee->isExternal() &&
6231 !(OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) ||
6232 (isa<PointerType>(FT->getReturnType()) &&
6233 TD->getIntPtrType()->isLosslesslyConvertibleTo(OldRetTy)))
6234 && !Caller->use_empty())
6235 return false; // Cannot transform this return value...
6237 // If the callsite is an invoke instruction, and the return value is used by
6238 // a PHI node in a successor, we cannot change the return type of the call
6239 // because there is no place to put the cast instruction (without breaking
6240 // the critical edge). Bail out in this case.
6241 if (!Caller->use_empty())
6242 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
6243 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
6245 if (PHINode *PN = dyn_cast<PHINode>(*UI))
6246 if (PN->getParent() == II->getNormalDest() ||
6247 PN->getParent() == II->getUnwindDest())
6251 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
6252 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
6254 CallSite::arg_iterator AI = CS.arg_begin();
6255 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
6256 const Type *ParamTy = FT->getParamType(i);
6257 const Type *ActTy = (*AI)->getType();
6258 ConstantSInt* c = dyn_cast<ConstantSInt>(*AI);
6259 //Either we can cast directly, or we can upconvert the argument
6260 bool isConvertible = ActTy->isLosslesslyConvertibleTo(ParamTy) ||
6261 (ParamTy->isIntegral() && ActTy->isIntegral() &&
6262 ParamTy->isSigned() == ActTy->isSigned() &&
6263 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
6264 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
6266 if (Callee->isExternal() && !isConvertible) return false;
6269 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
6270 Callee->isExternal())
6271 return false; // Do not delete arguments unless we have a function body...
6273 // Okay, we decided that this is a safe thing to do: go ahead and start
6274 // inserting cast instructions as necessary...
6275 std::vector<Value*> Args;
6276 Args.reserve(NumActualArgs);
6278 AI = CS.arg_begin();
6279 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
6280 const Type *ParamTy = FT->getParamType(i);
6281 if ((*AI)->getType() == ParamTy) {
6282 Args.push_back(*AI);
6284 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
6289 // If the function takes more arguments than the call was taking, add them
6291 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
6292 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
6294 // If we are removing arguments to the function, emit an obnoxious warning...
6295 if (FT->getNumParams() < NumActualArgs)
6296 if (!FT->isVarArg()) {
6297 std::cerr << "WARNING: While resolving call to function '"
6298 << Callee->getName() << "' arguments were dropped!\n";
6300 // Add all of the arguments in their promoted form to the arg list...
6301 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
6302 const Type *PTy = getPromotedType((*AI)->getType());
6303 if (PTy != (*AI)->getType()) {
6304 // Must promote to pass through va_arg area!
6305 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
6306 InsertNewInstBefore(Cast, *Caller);
6307 Args.push_back(Cast);
6309 Args.push_back(*AI);
6314 if (FT->getReturnType() == Type::VoidTy)
6315 Caller->setName(""); // Void type should not have a name...
6318 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6319 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
6320 Args, Caller->getName(), Caller);
6321 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
6323 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
6324 if (cast<CallInst>(Caller)->isTailCall())
6325 cast<CallInst>(NC)->setTailCall();
6326 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
6329 // Insert a cast of the return type as necessary...
6331 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
6332 if (NV->getType() != Type::VoidTy) {
6333 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
6335 // If this is an invoke instruction, we should insert it after the first
6336 // non-phi, instruction in the normal successor block.
6337 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6338 BasicBlock::iterator I = II->getNormalDest()->begin();
6339 while (isa<PHINode>(I)) ++I;
6340 InsertNewInstBefore(NC, *I);
6342 // Otherwise, it's a call, just insert cast right after the call instr
6343 InsertNewInstBefore(NC, *Caller);
6345 AddUsersToWorkList(*Caller);
6347 NV = UndefValue::get(Caller->getType());
6351 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
6352 Caller->replaceAllUsesWith(NV);
6353 Caller->getParent()->getInstList().erase(Caller);
6354 removeFromWorkList(Caller);
6359 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
6360 // operator and they all are only used by the PHI, PHI together their
6361 // inputs, and do the operation once, to the result of the PHI.
6362 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
6363 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6365 // Scan the instruction, looking for input operations that can be folded away.
6366 // If all input operands to the phi are the same instruction (e.g. a cast from
6367 // the same type or "+42") we can pull the operation through the PHI, reducing
6368 // code size and simplifying code.
6369 Constant *ConstantOp = 0;
6370 const Type *CastSrcTy = 0;
6371 if (isa<CastInst>(FirstInst)) {
6372 CastSrcTy = FirstInst->getOperand(0)->getType();
6373 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
6374 // Can fold binop or shift if the RHS is a constant.
6375 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
6376 if (ConstantOp == 0) return 0;
6378 return 0; // Cannot fold this operation.
6381 // Check to see if all arguments are the same operation.
6382 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6383 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
6384 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
6385 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
6388 if (I->getOperand(0)->getType() != CastSrcTy)
6389 return 0; // Cast operation must match.
6390 } else if (I->getOperand(1) != ConstantOp) {
6395 // Okay, they are all the same operation. Create a new PHI node of the
6396 // correct type, and PHI together all of the LHS's of the instructions.
6397 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
6398 PN.getName()+".in");
6399 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
6401 Value *InVal = FirstInst->getOperand(0);
6402 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
6404 // Add all operands to the new PHI.
6405 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6406 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
6407 if (NewInVal != InVal)
6409 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
6414 // The new PHI unions all of the same values together. This is really
6415 // common, so we handle it intelligently here for compile-time speed.
6419 InsertNewInstBefore(NewPN, PN);
6423 // Insert and return the new operation.
6424 if (isa<CastInst>(FirstInst))
6425 return new CastInst(PhiVal, PN.getType());
6426 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6427 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
6429 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
6430 PhiVal, ConstantOp);
6433 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
6435 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
6436 if (PN->use_empty()) return true;
6437 if (!PN->hasOneUse()) return false;
6439 // Remember this node, and if we find the cycle, return.
6440 if (!PotentiallyDeadPHIs.insert(PN).second)
6443 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
6444 return DeadPHICycle(PU, PotentiallyDeadPHIs);
6449 // PHINode simplification
6451 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
6452 // If LCSSA is around, don't mess with Phi nodes
6453 if (mustPreserveAnalysisID(LCSSAID)) return 0;
6455 if (Value *V = PN.hasConstantValue())
6456 return ReplaceInstUsesWith(PN, V);
6458 // If the only user of this instruction is a cast instruction, and all of the
6459 // incoming values are constants, change this PHI to merge together the casted
6462 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
6463 if (CI->getType() != PN.getType()) { // noop casts will be folded
6464 bool AllConstant = true;
6465 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
6466 if (!isa<Constant>(PN.getIncomingValue(i))) {
6467 AllConstant = false;
6471 // Make a new PHI with all casted values.
6472 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
6473 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
6474 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
6475 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
6476 PN.getIncomingBlock(i));
6479 // Update the cast instruction.
6480 CI->setOperand(0, New);
6481 WorkList.push_back(CI); // revisit the cast instruction to fold.
6482 WorkList.push_back(New); // Make sure to revisit the new Phi
6483 return &PN; // PN is now dead!
6487 // If all PHI operands are the same operation, pull them through the PHI,
6488 // reducing code size.
6489 if (isa<Instruction>(PN.getIncomingValue(0)) &&
6490 PN.getIncomingValue(0)->hasOneUse())
6491 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
6494 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
6495 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
6496 // PHI)... break the cycle.
6498 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
6499 std::set<PHINode*> PotentiallyDeadPHIs;
6500 PotentiallyDeadPHIs.insert(&PN);
6501 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
6502 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
6508 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
6509 Instruction *InsertPoint,
6511 unsigned PS = IC->getTargetData().getPointerSize();
6512 const Type *VTy = V->getType();
6513 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
6514 // We must insert a cast to ensure we sign-extend.
6515 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
6516 V->getName()), *InsertPoint);
6517 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
6522 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
6523 Value *PtrOp = GEP.getOperand(0);
6524 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
6525 // If so, eliminate the noop.
6526 if (GEP.getNumOperands() == 1)
6527 return ReplaceInstUsesWith(GEP, PtrOp);
6529 if (isa<UndefValue>(GEP.getOperand(0)))
6530 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
6532 bool HasZeroPointerIndex = false;
6533 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
6534 HasZeroPointerIndex = C->isNullValue();
6536 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
6537 return ReplaceInstUsesWith(GEP, PtrOp);
6539 // Eliminate unneeded casts for indices.
6540 bool MadeChange = false;
6541 gep_type_iterator GTI = gep_type_begin(GEP);
6542 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
6543 if (isa<SequentialType>(*GTI)) {
6544 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
6545 Value *Src = CI->getOperand(0);
6546 const Type *SrcTy = Src->getType();
6547 const Type *DestTy = CI->getType();
6548 if (Src->getType()->isInteger()) {
6549 if (SrcTy->getPrimitiveSizeInBits() ==
6550 DestTy->getPrimitiveSizeInBits()) {
6551 // We can always eliminate a cast from ulong or long to the other.
6552 // We can always eliminate a cast from uint to int or the other on
6553 // 32-bit pointer platforms.
6554 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
6556 GEP.setOperand(i, Src);
6558 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
6559 SrcTy->getPrimitiveSize() == 4) {
6560 // We can always eliminate a cast from int to [u]long. We can
6561 // eliminate a cast from uint to [u]long iff the target is a 32-bit
6563 if (SrcTy->isSigned() ||
6564 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
6566 GEP.setOperand(i, Src);
6571 // If we are using a wider index than needed for this platform, shrink it
6572 // to what we need. If the incoming value needs a cast instruction,
6573 // insert it. This explicit cast can make subsequent optimizations more
6575 Value *Op = GEP.getOperand(i);
6576 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
6577 if (Constant *C = dyn_cast<Constant>(Op)) {
6578 GEP.setOperand(i, ConstantExpr::getCast(C,
6579 TD->getIntPtrType()->getSignedVersion()));
6582 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
6583 Op->getName()), GEP);
6584 GEP.setOperand(i, Op);
6588 // If this is a constant idx, make sure to canonicalize it to be a signed
6589 // operand, otherwise CSE and other optimizations are pessimized.
6590 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
6591 GEP.setOperand(i, ConstantExpr::getCast(CUI,
6592 CUI->getType()->getSignedVersion()));
6596 if (MadeChange) return &GEP;
6598 // Combine Indices - If the source pointer to this getelementptr instruction
6599 // is a getelementptr instruction, combine the indices of the two
6600 // getelementptr instructions into a single instruction.
6602 std::vector<Value*> SrcGEPOperands;
6603 if (User *Src = dyn_castGetElementPtr(PtrOp))
6604 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
6606 if (!SrcGEPOperands.empty()) {
6607 // Note that if our source is a gep chain itself that we wait for that
6608 // chain to be resolved before we perform this transformation. This
6609 // avoids us creating a TON of code in some cases.
6611 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
6612 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
6613 return 0; // Wait until our source is folded to completion.
6615 std::vector<Value *> Indices;
6617 // Find out whether the last index in the source GEP is a sequential idx.
6618 bool EndsWithSequential = false;
6619 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
6620 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
6621 EndsWithSequential = !isa<StructType>(*I);
6623 // Can we combine the two pointer arithmetics offsets?
6624 if (EndsWithSequential) {
6625 // Replace: gep (gep %P, long B), long A, ...
6626 // With: T = long A+B; gep %P, T, ...
6628 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
6629 if (SO1 == Constant::getNullValue(SO1->getType())) {
6631 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
6634 // If they aren't the same type, convert both to an integer of the
6635 // target's pointer size.
6636 if (SO1->getType() != GO1->getType()) {
6637 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
6638 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
6639 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
6640 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
6642 unsigned PS = TD->getPointerSize();
6643 if (SO1->getType()->getPrimitiveSize() == PS) {
6644 // Convert GO1 to SO1's type.
6645 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
6647 } else if (GO1->getType()->getPrimitiveSize() == PS) {
6648 // Convert SO1 to GO1's type.
6649 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
6651 const Type *PT = TD->getIntPtrType();
6652 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
6653 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
6657 if (isa<Constant>(SO1) && isa<Constant>(GO1))
6658 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
6660 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
6661 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
6665 // Recycle the GEP we already have if possible.
6666 if (SrcGEPOperands.size() == 2) {
6667 GEP.setOperand(0, SrcGEPOperands[0]);
6668 GEP.setOperand(1, Sum);
6671 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
6672 SrcGEPOperands.end()-1);
6673 Indices.push_back(Sum);
6674 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
6676 } else if (isa<Constant>(*GEP.idx_begin()) &&
6677 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
6678 SrcGEPOperands.size() != 1) {
6679 // Otherwise we can do the fold if the first index of the GEP is a zero
6680 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
6681 SrcGEPOperands.end());
6682 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
6685 if (!Indices.empty())
6686 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
6688 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
6689 // GEP of global variable. If all of the indices for this GEP are
6690 // constants, we can promote this to a constexpr instead of an instruction.
6692 // Scan for nonconstants...
6693 std::vector<Constant*> Indices;
6694 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
6695 for (; I != E && isa<Constant>(*I); ++I)
6696 Indices.push_back(cast<Constant>(*I));
6698 if (I == E) { // If they are all constants...
6699 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
6701 // Replace all uses of the GEP with the new constexpr...
6702 return ReplaceInstUsesWith(GEP, CE);
6704 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
6705 if (!isa<PointerType>(X->getType())) {
6706 // Not interesting. Source pointer must be a cast from pointer.
6707 } else if (HasZeroPointerIndex) {
6708 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
6709 // into : GEP [10 x ubyte]* X, long 0, ...
6711 // This occurs when the program declares an array extern like "int X[];"
6713 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
6714 const PointerType *XTy = cast<PointerType>(X->getType());
6715 if (const ArrayType *XATy =
6716 dyn_cast<ArrayType>(XTy->getElementType()))
6717 if (const ArrayType *CATy =
6718 dyn_cast<ArrayType>(CPTy->getElementType()))
6719 if (CATy->getElementType() == XATy->getElementType()) {
6720 // At this point, we know that the cast source type is a pointer
6721 // to an array of the same type as the destination pointer
6722 // array. Because the array type is never stepped over (there
6723 // is a leading zero) we can fold the cast into this GEP.
6724 GEP.setOperand(0, X);
6727 } else if (GEP.getNumOperands() == 2) {
6728 // Transform things like:
6729 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
6730 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
6731 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
6732 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
6733 if (isa<ArrayType>(SrcElTy) &&
6734 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
6735 TD->getTypeSize(ResElTy)) {
6736 Value *V = InsertNewInstBefore(
6737 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
6738 GEP.getOperand(1), GEP.getName()), GEP);
6739 return new CastInst(V, GEP.getType());
6742 // Transform things like:
6743 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
6744 // (where tmp = 8*tmp2) into:
6745 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
6747 if (isa<ArrayType>(SrcElTy) &&
6748 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
6749 uint64_t ArrayEltSize =
6750 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
6752 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
6753 // allow either a mul, shift, or constant here.
6755 ConstantInt *Scale = 0;
6756 if (ArrayEltSize == 1) {
6757 NewIdx = GEP.getOperand(1);
6758 Scale = ConstantInt::get(NewIdx->getType(), 1);
6759 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
6760 NewIdx = ConstantInt::get(CI->getType(), 1);
6762 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
6763 if (Inst->getOpcode() == Instruction::Shl &&
6764 isa<ConstantInt>(Inst->getOperand(1))) {
6765 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
6766 if (Inst->getType()->isSigned())
6767 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
6769 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
6770 NewIdx = Inst->getOperand(0);
6771 } else if (Inst->getOpcode() == Instruction::Mul &&
6772 isa<ConstantInt>(Inst->getOperand(1))) {
6773 Scale = cast<ConstantInt>(Inst->getOperand(1));
6774 NewIdx = Inst->getOperand(0);
6778 // If the index will be to exactly the right offset with the scale taken
6779 // out, perform the transformation.
6780 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
6781 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
6782 Scale = ConstantSInt::get(C->getType(),
6783 (int64_t)C->getRawValue() /
6784 (int64_t)ArrayEltSize);
6786 Scale = ConstantUInt::get(Scale->getType(),
6787 Scale->getRawValue() / ArrayEltSize);
6788 if (Scale->getRawValue() != 1) {
6789 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
6790 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
6791 NewIdx = InsertNewInstBefore(Sc, GEP);
6794 // Insert the new GEP instruction.
6796 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
6797 NewIdx, GEP.getName());
6798 Idx = InsertNewInstBefore(Idx, GEP);
6799 return new CastInst(Idx, GEP.getType());
6808 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
6809 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
6810 if (AI.isArrayAllocation()) // Check C != 1
6811 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
6812 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
6813 AllocationInst *New = 0;
6815 // Create and insert the replacement instruction...
6816 if (isa<MallocInst>(AI))
6817 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
6819 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
6820 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
6823 InsertNewInstBefore(New, AI);
6825 // Scan to the end of the allocation instructions, to skip over a block of
6826 // allocas if possible...
6828 BasicBlock::iterator It = New;
6829 while (isa<AllocationInst>(*It)) ++It;
6831 // Now that I is pointing to the first non-allocation-inst in the block,
6832 // insert our getelementptr instruction...
6834 Value *NullIdx = Constant::getNullValue(Type::IntTy);
6835 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
6836 New->getName()+".sub", It);
6838 // Now make everything use the getelementptr instead of the original
6840 return ReplaceInstUsesWith(AI, V);
6841 } else if (isa<UndefValue>(AI.getArraySize())) {
6842 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6845 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
6846 // Note that we only do this for alloca's, because malloc should allocate and
6847 // return a unique pointer, even for a zero byte allocation.
6848 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
6849 TD->getTypeSize(AI.getAllocatedType()) == 0)
6850 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6855 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
6856 Value *Op = FI.getOperand(0);
6858 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
6859 if (CastInst *CI = dyn_cast<CastInst>(Op))
6860 if (isa<PointerType>(CI->getOperand(0)->getType())) {
6861 FI.setOperand(0, CI->getOperand(0));
6865 // free undef -> unreachable.
6866 if (isa<UndefValue>(Op)) {
6867 // Insert a new store to null because we cannot modify the CFG here.
6868 new StoreInst(ConstantBool::True,
6869 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
6870 return EraseInstFromFunction(FI);
6873 // If we have 'free null' delete the instruction. This can happen in stl code
6874 // when lots of inlining happens.
6875 if (isa<ConstantPointerNull>(Op))
6876 return EraseInstFromFunction(FI);
6882 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
6883 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
6884 User *CI = cast<User>(LI.getOperand(0));
6885 Value *CastOp = CI->getOperand(0);
6887 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
6888 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
6889 const Type *SrcPTy = SrcTy->getElementType();
6891 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
6892 isa<PackedType>(DestPTy)) {
6893 // If the source is an array, the code below will not succeed. Check to
6894 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
6896 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
6897 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
6898 if (ASrcTy->getNumElements() != 0) {
6899 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
6900 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
6901 SrcTy = cast<PointerType>(CastOp->getType());
6902 SrcPTy = SrcTy->getElementType();
6905 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
6906 isa<PackedType>(SrcPTy)) &&
6907 // Do not allow turning this into a load of an integer, which is then
6908 // casted to a pointer, this pessimizes pointer analysis a lot.
6909 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
6910 IC.getTargetData().getTypeSize(SrcPTy) ==
6911 IC.getTargetData().getTypeSize(DestPTy)) {
6913 // Okay, we are casting from one integer or pointer type to another of
6914 // the same size. Instead of casting the pointer before the load, cast
6915 // the result of the loaded value.
6916 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
6918 LI.isVolatile()),LI);
6919 // Now cast the result of the load.
6920 return new CastInst(NewLoad, LI.getType());
6927 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
6928 /// from this value cannot trap. If it is not obviously safe to load from the
6929 /// specified pointer, we do a quick local scan of the basic block containing
6930 /// ScanFrom, to determine if the address is already accessed.
6931 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
6932 // If it is an alloca or global variable, it is always safe to load from.
6933 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
6935 // Otherwise, be a little bit agressive by scanning the local block where we
6936 // want to check to see if the pointer is already being loaded or stored
6937 // from/to. If so, the previous load or store would have already trapped,
6938 // so there is no harm doing an extra load (also, CSE will later eliminate
6939 // the load entirely).
6940 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
6945 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
6946 if (LI->getOperand(0) == V) return true;
6947 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
6948 if (SI->getOperand(1) == V) return true;
6954 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
6955 Value *Op = LI.getOperand(0);
6957 // load (cast X) --> cast (load X) iff safe
6958 if (CastInst *CI = dyn_cast<CastInst>(Op))
6959 if (Instruction *Res = InstCombineLoadCast(*this, LI))
6962 // None of the following transforms are legal for volatile loads.
6963 if (LI.isVolatile()) return 0;
6965 if (&LI.getParent()->front() != &LI) {
6966 BasicBlock::iterator BBI = &LI; --BBI;
6967 // If the instruction immediately before this is a store to the same
6968 // address, do a simple form of store->load forwarding.
6969 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
6970 if (SI->getOperand(1) == LI.getOperand(0))
6971 return ReplaceInstUsesWith(LI, SI->getOperand(0));
6972 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
6973 if (LIB->getOperand(0) == LI.getOperand(0))
6974 return ReplaceInstUsesWith(LI, LIB);
6977 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
6978 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
6979 isa<UndefValue>(GEPI->getOperand(0))) {
6980 // Insert a new store to null instruction before the load to indicate
6981 // that this code is not reachable. We do this instead of inserting
6982 // an unreachable instruction directly because we cannot modify the
6984 new StoreInst(UndefValue::get(LI.getType()),
6985 Constant::getNullValue(Op->getType()), &LI);
6986 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
6989 if (Constant *C = dyn_cast<Constant>(Op)) {
6990 // load null/undef -> undef
6991 if ((C->isNullValue() || isa<UndefValue>(C))) {
6992 // Insert a new store to null instruction before the load to indicate that
6993 // this code is not reachable. We do this instead of inserting an
6994 // unreachable instruction directly because we cannot modify the CFG.
6995 new StoreInst(UndefValue::get(LI.getType()),
6996 Constant::getNullValue(Op->getType()), &LI);
6997 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7000 // Instcombine load (constant global) into the value loaded.
7001 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
7002 if (GV->isConstant() && !GV->isExternal())
7003 return ReplaceInstUsesWith(LI, GV->getInitializer());
7005 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
7006 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7007 if (CE->getOpcode() == Instruction::GetElementPtr) {
7008 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
7009 if (GV->isConstant() && !GV->isExternal())
7011 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
7012 return ReplaceInstUsesWith(LI, V);
7013 if (CE->getOperand(0)->isNullValue()) {
7014 // Insert a new store to null instruction before the load to indicate
7015 // that this code is not reachable. We do this instead of inserting
7016 // an unreachable instruction directly because we cannot modify the
7018 new StoreInst(UndefValue::get(LI.getType()),
7019 Constant::getNullValue(Op->getType()), &LI);
7020 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7023 } else if (CE->getOpcode() == Instruction::Cast) {
7024 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7029 if (Op->hasOneUse()) {
7030 // Change select and PHI nodes to select values instead of addresses: this
7031 // helps alias analysis out a lot, allows many others simplifications, and
7032 // exposes redundancy in the code.
7034 // Note that we cannot do the transformation unless we know that the
7035 // introduced loads cannot trap! Something like this is valid as long as
7036 // the condition is always false: load (select bool %C, int* null, int* %G),
7037 // but it would not be valid if we transformed it to load from null
7040 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7041 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7042 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7043 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7044 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
7045 SI->getOperand(1)->getName()+".val"), LI);
7046 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
7047 SI->getOperand(2)->getName()+".val"), LI);
7048 return new SelectInst(SI->getCondition(), V1, V2);
7051 // load (select (cond, null, P)) -> load P
7052 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7053 if (C->isNullValue()) {
7054 LI.setOperand(0, SI->getOperand(2));
7058 // load (select (cond, P, null)) -> load P
7059 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7060 if (C->isNullValue()) {
7061 LI.setOperand(0, SI->getOperand(1));
7065 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
7066 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
7067 bool Safe = PN->getParent() == LI.getParent();
7069 // Scan all of the instructions between the PHI and the load to make
7070 // sure there are no instructions that might possibly alter the value
7071 // loaded from the PHI.
7073 BasicBlock::iterator I = &LI;
7074 for (--I; !isa<PHINode>(I); --I)
7075 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
7081 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
7082 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
7083 PN->getIncomingBlock(i)->getTerminator()))
7088 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
7089 InsertNewInstBefore(NewPN, *PN);
7090 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
7092 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
7093 BasicBlock *BB = PN->getIncomingBlock(i);
7094 Value *&TheLoad = LoadMap[BB];
7096 Value *InVal = PN->getIncomingValue(i);
7097 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
7098 InVal->getName()+".val"),
7099 *BB->getTerminator());
7101 NewPN->addIncoming(TheLoad, BB);
7103 return ReplaceInstUsesWith(LI, NewPN);
7110 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
7112 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7113 User *CI = cast<User>(SI.getOperand(1));
7114 Value *CastOp = CI->getOperand(0);
7116 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7117 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7118 const Type *SrcPTy = SrcTy->getElementType();
7120 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
7121 // If the source is an array, the code below will not succeed. Check to
7122 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7124 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7125 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7126 if (ASrcTy->getNumElements() != 0) {
7127 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7128 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7129 SrcTy = cast<PointerType>(CastOp->getType());
7130 SrcPTy = SrcTy->getElementType();
7133 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
7134 IC.getTargetData().getTypeSize(SrcPTy) ==
7135 IC.getTargetData().getTypeSize(DestPTy)) {
7137 // Okay, we are casting from one integer or pointer type to another of
7138 // the same size. Instead of casting the pointer before the store, cast
7139 // the value to be stored.
7141 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
7142 NewCast = ConstantExpr::getCast(C, SrcPTy);
7144 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
7146 SI.getOperand(0)->getName()+".c"), SI);
7148 return new StoreInst(NewCast, CastOp);
7155 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7156 Value *Val = SI.getOperand(0);
7157 Value *Ptr = SI.getOperand(1);
7159 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
7160 EraseInstFromFunction(SI);
7165 // Do really simple DSE, to catch cases where there are several consequtive
7166 // stores to the same location, separated by a few arithmetic operations. This
7167 // situation often occurs with bitfield accesses.
7168 BasicBlock::iterator BBI = &SI;
7169 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7173 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7174 // Prev store isn't volatile, and stores to the same location?
7175 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
7178 EraseInstFromFunction(*PrevSI);
7184 // If this is a load, we have to stop. However, if the loaded value is from
7185 // the pointer we're loading and is producing the pointer we're storing,
7186 // then *this* store is dead (X = load P; store X -> P).
7187 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7188 if (LI == Val && LI->getOperand(0) == Ptr) {
7189 EraseInstFromFunction(SI);
7193 // Otherwise, this is a load from some other location. Stores before it
7198 // Don't skip over loads or things that can modify memory.
7199 if (BBI->mayWriteToMemory())
7204 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7206 // store X, null -> turns into 'unreachable' in SimplifyCFG
7207 if (isa<ConstantPointerNull>(Ptr)) {
7208 if (!isa<UndefValue>(Val)) {
7209 SI.setOperand(0, UndefValue::get(Val->getType()));
7210 if (Instruction *U = dyn_cast<Instruction>(Val))
7211 WorkList.push_back(U); // Dropped a use.
7214 return 0; // Do not modify these!
7217 // store undef, Ptr -> noop
7218 if (isa<UndefValue>(Val)) {
7219 EraseInstFromFunction(SI);
7224 // If the pointer destination is a cast, see if we can fold the cast into the
7226 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
7227 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7229 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
7230 if (CE->getOpcode() == Instruction::Cast)
7231 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7235 // If this store is the last instruction in the basic block, and if the block
7236 // ends with an unconditional branch, try to move it to the successor block.
7238 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
7239 if (BI->isUnconditional()) {
7240 // Check to see if the successor block has exactly two incoming edges. If
7241 // so, see if the other predecessor contains a store to the same location.
7242 // if so, insert a PHI node (if needed) and move the stores down.
7243 BasicBlock *Dest = BI->getSuccessor(0);
7245 pred_iterator PI = pred_begin(Dest);
7246 BasicBlock *Other = 0;
7247 if (*PI != BI->getParent())
7250 if (PI != pred_end(Dest)) {
7251 if (*PI != BI->getParent())
7256 if (++PI != pred_end(Dest))
7259 if (Other) { // If only one other pred...
7260 BBI = Other->getTerminator();
7261 // Make sure this other block ends in an unconditional branch and that
7262 // there is an instruction before the branch.
7263 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
7264 BBI != Other->begin()) {
7266 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
7268 // If this instruction is a store to the same location.
7269 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
7270 // Okay, we know we can perform this transformation. Insert a PHI
7271 // node now if we need it.
7272 Value *MergedVal = OtherStore->getOperand(0);
7273 if (MergedVal != SI.getOperand(0)) {
7274 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
7275 PN->reserveOperandSpace(2);
7276 PN->addIncoming(SI.getOperand(0), SI.getParent());
7277 PN->addIncoming(OtherStore->getOperand(0), Other);
7278 MergedVal = InsertNewInstBefore(PN, Dest->front());
7281 // Advance to a place where it is safe to insert the new store and
7283 BBI = Dest->begin();
7284 while (isa<PHINode>(BBI)) ++BBI;
7285 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
7286 OtherStore->isVolatile()), *BBI);
7288 // Nuke the old stores.
7289 EraseInstFromFunction(SI);
7290 EraseInstFromFunction(*OtherStore);
7302 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
7303 // Change br (not X), label True, label False to: br X, label False, True
7305 BasicBlock *TrueDest;
7306 BasicBlock *FalseDest;
7307 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
7308 !isa<Constant>(X)) {
7309 // Swap Destinations and condition...
7311 BI.setSuccessor(0, FalseDest);
7312 BI.setSuccessor(1, TrueDest);
7316 // Cannonicalize setne -> seteq
7317 Instruction::BinaryOps Op; Value *Y;
7318 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
7319 TrueDest, FalseDest)))
7320 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
7321 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
7322 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
7323 std::string Name = I->getName(); I->setName("");
7324 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
7325 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
7326 // Swap Destinations and condition...
7327 BI.setCondition(NewSCC);
7328 BI.setSuccessor(0, FalseDest);
7329 BI.setSuccessor(1, TrueDest);
7330 removeFromWorkList(I);
7331 I->getParent()->getInstList().erase(I);
7332 WorkList.push_back(cast<Instruction>(NewSCC));
7339 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
7340 Value *Cond = SI.getCondition();
7341 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
7342 if (I->getOpcode() == Instruction::Add)
7343 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7344 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
7345 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
7346 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
7348 SI.setOperand(0, I->getOperand(0));
7349 WorkList.push_back(I);
7356 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
7357 /// is to leave as a vector operation.
7358 static bool CheapToScalarize(Value *V, bool isConstant) {
7359 if (isa<ConstantAggregateZero>(V))
7361 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
7362 if (isConstant) return true;
7363 // If all elts are the same, we can extract.
7364 Constant *Op0 = C->getOperand(0);
7365 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7366 if (C->getOperand(i) != Op0)
7370 Instruction *I = dyn_cast<Instruction>(V);
7371 if (!I) return false;
7373 // Insert element gets simplified to the inserted element or is deleted if
7374 // this is constant idx extract element and its a constant idx insertelt.
7375 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
7376 isa<ConstantInt>(I->getOperand(2)))
7378 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
7380 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
7381 if (BO->hasOneUse() &&
7382 (CheapToScalarize(BO->getOperand(0), isConstant) ||
7383 CheapToScalarize(BO->getOperand(1), isConstant)))
7389 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
7390 /// elements into values that are larger than the #elts in the input.
7391 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
7392 unsigned NElts = SVI->getType()->getNumElements();
7393 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
7394 return std::vector<unsigned>(NElts, 0);
7395 if (isa<UndefValue>(SVI->getOperand(2)))
7396 return std::vector<unsigned>(NElts, 2*NElts);
7398 std::vector<unsigned> Result;
7399 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
7400 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
7401 if (isa<UndefValue>(CP->getOperand(i)))
7402 Result.push_back(NElts*2); // undef -> 8
7404 Result.push_back(cast<ConstantUInt>(CP->getOperand(i))->getValue());
7408 /// FindScalarElement - Given a vector and an element number, see if the scalar
7409 /// value is already around as a register, for example if it were inserted then
7410 /// extracted from the vector.
7411 static Value *FindScalarElement(Value *V, unsigned EltNo) {
7412 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
7413 const PackedType *PTy = cast<PackedType>(V->getType());
7414 unsigned Width = PTy->getNumElements();
7415 if (EltNo >= Width) // Out of range access.
7416 return UndefValue::get(PTy->getElementType());
7418 if (isa<UndefValue>(V))
7419 return UndefValue::get(PTy->getElementType());
7420 else if (isa<ConstantAggregateZero>(V))
7421 return Constant::getNullValue(PTy->getElementType());
7422 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
7423 return CP->getOperand(EltNo);
7424 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
7425 // If this is an insert to a variable element, we don't know what it is.
7426 if (!isa<ConstantUInt>(III->getOperand(2))) return 0;
7427 unsigned IIElt = cast<ConstantUInt>(III->getOperand(2))->getValue();
7429 // If this is an insert to the element we are looking for, return the
7431 if (EltNo == IIElt) return III->getOperand(1);
7433 // Otherwise, the insertelement doesn't modify the value, recurse on its
7435 return FindScalarElement(III->getOperand(0), EltNo);
7436 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
7437 unsigned InEl = getShuffleMask(SVI)[EltNo];
7439 return FindScalarElement(SVI->getOperand(0), InEl);
7440 else if (InEl < Width*2)
7441 return FindScalarElement(SVI->getOperand(1), InEl - Width);
7443 return UndefValue::get(PTy->getElementType());
7446 // Otherwise, we don't know.
7450 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
7452 // If packed val is undef, replace extract with scalar undef.
7453 if (isa<UndefValue>(EI.getOperand(0)))
7454 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
7456 // If packed val is constant 0, replace extract with scalar 0.
7457 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
7458 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
7460 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
7461 // If packed val is constant with uniform operands, replace EI
7462 // with that operand
7463 Constant *op0 = C->getOperand(0);
7464 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7465 if (C->getOperand(i) != op0) {
7470 return ReplaceInstUsesWith(EI, op0);
7473 // If extracting a specified index from the vector, see if we can recursively
7474 // find a previously computed scalar that was inserted into the vector.
7475 if (ConstantUInt *IdxC = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
7476 if (Value *Elt = FindScalarElement(EI.getOperand(0), IdxC->getValue()))
7477 return ReplaceInstUsesWith(EI, Elt);
7480 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
7481 if (I->hasOneUse()) {
7482 // Push extractelement into predecessor operation if legal and
7483 // profitable to do so
7484 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
7485 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
7486 if (CheapToScalarize(BO, isConstantElt)) {
7487 ExtractElementInst *newEI0 =
7488 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
7489 EI.getName()+".lhs");
7490 ExtractElementInst *newEI1 =
7491 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
7492 EI.getName()+".rhs");
7493 InsertNewInstBefore(newEI0, EI);
7494 InsertNewInstBefore(newEI1, EI);
7495 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
7497 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7498 Value *Ptr = InsertCastBefore(I->getOperand(0),
7499 PointerType::get(EI.getType()), EI);
7500 GetElementPtrInst *GEP =
7501 new GetElementPtrInst(Ptr, EI.getOperand(1),
7502 I->getName() + ".gep");
7503 InsertNewInstBefore(GEP, EI);
7504 return new LoadInst(GEP);
7507 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
7508 // Extracting the inserted element?
7509 if (IE->getOperand(2) == EI.getOperand(1))
7510 return ReplaceInstUsesWith(EI, IE->getOperand(1));
7511 // If the inserted and extracted elements are constants, they must not
7512 // be the same value, extract from the pre-inserted value instead.
7513 if (isa<Constant>(IE->getOperand(2)) &&
7514 isa<Constant>(EI.getOperand(1))) {
7515 AddUsesToWorkList(EI);
7516 EI.setOperand(0, IE->getOperand(0));
7519 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
7520 // If this is extracting an element from a shufflevector, figure out where
7521 // it came from and extract from the appropriate input element instead.
7522 if (ConstantUInt *Elt = dyn_cast<ConstantUInt>(EI.getOperand(1))) {
7523 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getValue()];
7525 if (SrcIdx < SVI->getType()->getNumElements())
7526 Src = SVI->getOperand(0);
7527 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
7528 SrcIdx -= SVI->getType()->getNumElements();
7529 Src = SVI->getOperand(1);
7531 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
7533 return new ExtractElementInst(Src,
7534 ConstantUInt::get(Type::UIntTy, SrcIdx));
7541 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
7542 /// elements from either LHS or RHS, return the shuffle mask and true.
7543 /// Otherwise, return false.
7544 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
7545 std::vector<Constant*> &Mask) {
7546 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
7547 "Invalid CollectSingleShuffleElements");
7548 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
7550 if (isa<UndefValue>(V)) {
7551 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
7553 } else if (V == LHS) {
7554 for (unsigned i = 0; i != NumElts; ++i)
7555 Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
7557 } else if (V == RHS) {
7558 for (unsigned i = 0; i != NumElts; ++i)
7559 Mask.push_back(ConstantUInt::get(Type::UIntTy, i+NumElts));
7561 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
7562 // If this is an insert of an extract from some other vector, include it.
7563 Value *VecOp = IEI->getOperand(0);
7564 Value *ScalarOp = IEI->getOperand(1);
7565 Value *IdxOp = IEI->getOperand(2);
7567 if (!isa<ConstantInt>(IdxOp))
7569 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7571 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
7572 // Okay, we can handle this if the vector we are insertinting into is
7574 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
7575 // If so, update the mask to reflect the inserted undef.
7576 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
7579 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
7580 if (isa<ConstantInt>(EI->getOperand(1)) &&
7581 EI->getOperand(0)->getType() == V->getType()) {
7582 unsigned ExtractedIdx =
7583 cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7585 // This must be extracting from either LHS or RHS.
7586 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
7587 // Okay, we can handle this if the vector we are insertinting into is
7589 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
7590 // If so, update the mask to reflect the inserted value.
7591 if (EI->getOperand(0) == LHS) {
7592 Mask[InsertedIdx & (NumElts-1)] =
7593 ConstantUInt::get(Type::UIntTy, ExtractedIdx);
7595 assert(EI->getOperand(0) == RHS);
7596 Mask[InsertedIdx & (NumElts-1)] =
7597 ConstantUInt::get(Type::UIntTy, ExtractedIdx+NumElts);
7606 // TODO: Handle shufflevector here!
7611 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
7612 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
7613 /// that computes V and the LHS value of the shuffle.
7614 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
7616 assert(isa<PackedType>(V->getType()) &&
7617 (RHS == 0 || V->getType() == RHS->getType()) &&
7618 "Invalid shuffle!");
7619 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
7621 if (isa<UndefValue>(V)) {
7622 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
7624 } else if (isa<ConstantAggregateZero>(V)) {
7625 Mask.assign(NumElts, ConstantUInt::get(Type::UIntTy, 0));
7627 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
7628 // If this is an insert of an extract from some other vector, include it.
7629 Value *VecOp = IEI->getOperand(0);
7630 Value *ScalarOp = IEI->getOperand(1);
7631 Value *IdxOp = IEI->getOperand(2);
7633 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
7634 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
7635 EI->getOperand(0)->getType() == V->getType()) {
7636 unsigned ExtractedIdx =
7637 cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7638 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7640 // Either the extracted from or inserted into vector must be RHSVec,
7641 // otherwise we'd end up with a shuffle of three inputs.
7642 if (EI->getOperand(0) == RHS || RHS == 0) {
7643 RHS = EI->getOperand(0);
7644 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
7645 Mask[InsertedIdx & (NumElts-1)] =
7646 ConstantUInt::get(Type::UIntTy, NumElts+ExtractedIdx);
7651 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
7652 // Everything but the extracted element is replaced with the RHS.
7653 for (unsigned i = 0; i != NumElts; ++i) {
7654 if (i != InsertedIdx)
7655 Mask[i] = ConstantUInt::get(Type::UIntTy, NumElts+i);
7660 // If this insertelement is a chain that comes from exactly these two
7661 // vectors, return the vector and the effective shuffle.
7662 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
7663 return EI->getOperand(0);
7668 // TODO: Handle shufflevector here!
7670 // Otherwise, can't do anything fancy. Return an identity vector.
7671 for (unsigned i = 0; i != NumElts; ++i)
7672 Mask.push_back(ConstantUInt::get(Type::UIntTy, i));
7676 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
7677 Value *VecOp = IE.getOperand(0);
7678 Value *ScalarOp = IE.getOperand(1);
7679 Value *IdxOp = IE.getOperand(2);
7681 // If the inserted element was extracted from some other vector, and if the
7682 // indexes are constant, try to turn this into a shufflevector operation.
7683 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
7684 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
7685 EI->getOperand(0)->getType() == IE.getType()) {
7686 unsigned NumVectorElts = IE.getType()->getNumElements();
7687 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getRawValue();
7688 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getRawValue();
7690 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
7691 return ReplaceInstUsesWith(IE, VecOp);
7693 if (InsertedIdx >= NumVectorElts) // Out of range insert.
7694 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
7696 // If we are extracting a value from a vector, then inserting it right
7697 // back into the same place, just use the input vector.
7698 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
7699 return ReplaceInstUsesWith(IE, VecOp);
7701 // We could theoretically do this for ANY input. However, doing so could
7702 // turn chains of insertelement instructions into a chain of shufflevector
7703 // instructions, and right now we do not merge shufflevectors. As such,
7704 // only do this in a situation where it is clear that there is benefit.
7705 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
7706 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
7707 // the values of VecOp, except then one read from EIOp0.
7708 // Build a new shuffle mask.
7709 std::vector<Constant*> Mask;
7710 if (isa<UndefValue>(VecOp))
7711 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
7713 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
7714 Mask.assign(NumVectorElts, ConstantUInt::get(Type::UIntTy,
7717 Mask[InsertedIdx] = ConstantUInt::get(Type::UIntTy, ExtractedIdx);
7718 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
7719 ConstantPacked::get(Mask));
7722 // If this insertelement isn't used by some other insertelement, turn it
7723 // (and any insertelements it points to), into one big shuffle.
7724 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
7725 std::vector<Constant*> Mask;
7727 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
7728 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
7729 // We now have a shuffle of LHS, RHS, Mask.
7730 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
7739 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
7740 Value *LHS = SVI.getOperand(0);
7741 Value *RHS = SVI.getOperand(1);
7742 std::vector<unsigned> Mask = getShuffleMask(&SVI);
7744 bool MadeChange = false;
7746 if (isa<UndefValue>(SVI.getOperand(2)))
7747 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
7749 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
7750 // the undef, change them to undefs.
7752 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
7753 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
7754 if (LHS == RHS || isa<UndefValue>(LHS)) {
7755 if (isa<UndefValue>(LHS) && LHS == RHS) {
7756 // shuffle(undef,undef,mask) -> undef.
7757 return ReplaceInstUsesWith(SVI, LHS);
7760 // Remap any references to RHS to use LHS.
7761 std::vector<Constant*> Elts;
7762 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
7764 Elts.push_back(UndefValue::get(Type::UIntTy));
7766 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
7767 (Mask[i] < e && isa<UndefValue>(LHS)))
7768 Mask[i] = 2*e; // Turn into undef.
7770 Mask[i] &= (e-1); // Force to LHS.
7771 Elts.push_back(ConstantUInt::get(Type::UIntTy, Mask[i]));
7774 SVI.setOperand(0, SVI.getOperand(1));
7775 SVI.setOperand(1, UndefValue::get(RHS->getType()));
7776 SVI.setOperand(2, ConstantPacked::get(Elts));
7777 LHS = SVI.getOperand(0);
7778 RHS = SVI.getOperand(1);
7782 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
7783 bool isLHSID = true, isRHSID = true;
7785 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
7786 if (Mask[i] >= e*2) continue; // Ignore undef values.
7787 // Is this an identity shuffle of the LHS value?
7788 isLHSID &= (Mask[i] == i);
7790 // Is this an identity shuffle of the RHS value?
7791 isRHSID &= (Mask[i]-e == i);
7794 // Eliminate identity shuffles.
7795 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
7796 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
7798 // If the LHS is a shufflevector itself, see if we can combine it with this
7799 // one without producing an unusual shuffle. Here we are really conservative:
7800 // we are absolutely afraid of producing a shuffle mask not in the input
7801 // program, because the code gen may not be smart enough to turn a merged
7802 // shuffle into two specific shuffles: it may produce worse code. As such,
7803 // we only merge two shuffles if the result is one of the two input shuffle
7804 // masks. In this case, merging the shuffles just removes one instruction,
7805 // which we know is safe. This is good for things like turning:
7806 // (splat(splat)) -> splat.
7807 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
7808 if (isa<UndefValue>(RHS)) {
7809 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
7811 std::vector<unsigned> NewMask;
7812 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
7814 NewMask.push_back(2*e);
7816 NewMask.push_back(LHSMask[Mask[i]]);
7818 // If the result mask is equal to the src shuffle or this shuffle mask, do
7820 if (NewMask == LHSMask || NewMask == Mask) {
7821 std::vector<Constant*> Elts;
7822 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
7823 if (NewMask[i] >= e*2) {
7824 Elts.push_back(UndefValue::get(Type::UIntTy));
7826 Elts.push_back(ConstantUInt::get(Type::UIntTy, NewMask[i]));
7829 return new ShuffleVectorInst(LHSSVI->getOperand(0),
7830 LHSSVI->getOperand(1),
7831 ConstantPacked::get(Elts));
7836 return MadeChange ? &SVI : 0;
7841 void InstCombiner::removeFromWorkList(Instruction *I) {
7842 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
7847 /// TryToSinkInstruction - Try to move the specified instruction from its
7848 /// current block into the beginning of DestBlock, which can only happen if it's
7849 /// safe to move the instruction past all of the instructions between it and the
7850 /// end of its block.
7851 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
7852 assert(I->hasOneUse() && "Invariants didn't hold!");
7854 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
7855 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
7857 // Do not sink alloca instructions out of the entry block.
7858 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
7861 // We can only sink load instructions if there is nothing between the load and
7862 // the end of block that could change the value.
7863 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7864 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
7866 if (Scan->mayWriteToMemory())
7870 BasicBlock::iterator InsertPos = DestBlock->begin();
7871 while (isa<PHINode>(InsertPos)) ++InsertPos;
7873 I->moveBefore(InsertPos);
7878 /// OptimizeConstantExpr - Given a constant expression and target data layout
7879 /// information, symbolically evaluation the constant expr to something simpler
7881 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
7884 Constant *Ptr = CE->getOperand(0);
7885 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
7886 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
7887 // If this is a constant expr gep that is effectively computing an
7888 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
7889 bool isFoldableGEP = true;
7890 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
7891 if (!isa<ConstantInt>(CE->getOperand(i)))
7892 isFoldableGEP = false;
7893 if (isFoldableGEP) {
7894 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
7895 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
7896 Constant *C = ConstantUInt::get(Type::ULongTy, Offset);
7897 C = ConstantExpr::getCast(C, TD->getIntPtrType());
7898 return ConstantExpr::getCast(C, CE->getType());
7906 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
7907 /// all reachable code to the worklist.
7909 /// This has a couple of tricks to make the code faster and more powerful. In
7910 /// particular, we constant fold and DCE instructions as we go, to avoid adding
7911 /// them to the worklist (this significantly speeds up instcombine on code where
7912 /// many instructions are dead or constant). Additionally, if we find a branch
7913 /// whose condition is a known constant, we only visit the reachable successors.
7915 static void AddReachableCodeToWorklist(BasicBlock *BB,
7916 std::set<BasicBlock*> &Visited,
7917 std::vector<Instruction*> &WorkList,
7918 const TargetData *TD) {
7919 // We have now visited this block! If we've already been here, bail out.
7920 if (!Visited.insert(BB).second) return;
7922 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
7923 Instruction *Inst = BBI++;
7925 // DCE instruction if trivially dead.
7926 if (isInstructionTriviallyDead(Inst)) {
7928 DEBUG(std::cerr << "IC: DCE: " << *Inst);
7929 Inst->eraseFromParent();
7933 // ConstantProp instruction if trivially constant.
7934 if (Constant *C = ConstantFoldInstruction(Inst)) {
7935 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
7936 C = OptimizeConstantExpr(CE, TD);
7937 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *Inst);
7938 Inst->replaceAllUsesWith(C);
7940 Inst->eraseFromParent();
7944 WorkList.push_back(Inst);
7947 // Recursively visit successors. If this is a branch or switch on a constant,
7948 // only visit the reachable successor.
7949 TerminatorInst *TI = BB->getTerminator();
7950 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
7951 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
7952 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
7953 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
7957 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
7958 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
7959 // See if this is an explicit destination.
7960 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
7961 if (SI->getCaseValue(i) == Cond) {
7962 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
7966 // Otherwise it is the default destination.
7967 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
7972 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
7973 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
7976 bool InstCombiner::runOnFunction(Function &F) {
7977 bool Changed = false;
7978 TD = &getAnalysis<TargetData>();
7981 // Do a depth-first traversal of the function, populate the worklist with
7982 // the reachable instructions. Ignore blocks that are not reachable. Keep
7983 // track of which blocks we visit.
7984 std::set<BasicBlock*> Visited;
7985 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
7987 // Do a quick scan over the function. If we find any blocks that are
7988 // unreachable, remove any instructions inside of them. This prevents
7989 // the instcombine code from having to deal with some bad special cases.
7990 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
7991 if (!Visited.count(BB)) {
7992 Instruction *Term = BB->getTerminator();
7993 while (Term != BB->begin()) { // Remove instrs bottom-up
7994 BasicBlock::iterator I = Term; --I;
7996 DEBUG(std::cerr << "IC: DCE: " << *I);
7999 if (!I->use_empty())
8000 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8001 I->eraseFromParent();
8006 while (!WorkList.empty()) {
8007 Instruction *I = WorkList.back(); // Get an instruction from the worklist
8008 WorkList.pop_back();
8010 // Check to see if we can DCE the instruction.
8011 if (isInstructionTriviallyDead(I)) {
8012 // Add operands to the worklist.
8013 if (I->getNumOperands() < 4)
8014 AddUsesToWorkList(*I);
8017 DEBUG(std::cerr << "IC: DCE: " << *I);
8019 I->eraseFromParent();
8020 removeFromWorkList(I);
8024 // Instruction isn't dead, see if we can constant propagate it.
8025 if (Constant *C = ConstantFoldInstruction(I)) {
8026 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8027 C = OptimizeConstantExpr(CE, TD);
8028 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
8030 // Add operands to the worklist.
8031 AddUsesToWorkList(*I);
8032 ReplaceInstUsesWith(*I, C);
8035 I->eraseFromParent();
8036 removeFromWorkList(I);
8040 // See if we can trivially sink this instruction to a successor basic block.
8041 if (I->hasOneUse()) {
8042 BasicBlock *BB = I->getParent();
8043 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
8044 if (UserParent != BB) {
8045 bool UserIsSuccessor = false;
8046 // See if the user is one of our successors.
8047 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8048 if (*SI == UserParent) {
8049 UserIsSuccessor = true;
8053 // If the user is one of our immediate successors, and if that successor
8054 // only has us as a predecessors (we'd have to split the critical edge
8055 // otherwise), we can keep going.
8056 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
8057 next(pred_begin(UserParent)) == pred_end(UserParent))
8058 // Okay, the CFG is simple enough, try to sink this instruction.
8059 Changed |= TryToSinkInstruction(I, UserParent);
8063 // Now that we have an instruction, try combining it to simplify it...
8064 if (Instruction *Result = visit(*I)) {
8066 // Should we replace the old instruction with a new one?
8068 DEBUG(std::cerr << "IC: Old = " << *I
8069 << " New = " << *Result);
8071 // Everything uses the new instruction now.
8072 I->replaceAllUsesWith(Result);
8074 // Push the new instruction and any users onto the worklist.
8075 WorkList.push_back(Result);
8076 AddUsersToWorkList(*Result);
8078 // Move the name to the new instruction first...
8079 std::string OldName = I->getName(); I->setName("");
8080 Result->setName(OldName);
8082 // Insert the new instruction into the basic block...
8083 BasicBlock *InstParent = I->getParent();
8084 BasicBlock::iterator InsertPos = I;
8086 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8087 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8090 InstParent->getInstList().insert(InsertPos, Result);
8092 // Make sure that we reprocess all operands now that we reduced their
8094 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8095 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8096 WorkList.push_back(OpI);
8098 // Instructions can end up on the worklist more than once. Make sure
8099 // we do not process an instruction that has been deleted.
8100 removeFromWorkList(I);
8102 // Erase the old instruction.
8103 InstParent->getInstList().erase(I);
8105 DEBUG(std::cerr << "IC: MOD = " << *I);
8107 // If the instruction was modified, it's possible that it is now dead.
8108 // if so, remove it.
8109 if (isInstructionTriviallyDead(I)) {
8110 // Make sure we process all operands now that we are reducing their
8112 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8113 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8114 WorkList.push_back(OpI);
8116 // Instructions may end up in the worklist more than once. Erase all
8117 // occurrences of this instruction.
8118 removeFromWorkList(I);
8119 I->eraseFromParent();
8121 WorkList.push_back(Result);
8122 AddUsersToWorkList(*Result);
8132 FunctionPass *llvm::createInstructionCombiningPass() {
8133 return new InstCombiner();