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 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
93 /// dead. Add all of its operands to the worklist, turning them into
94 /// undef's to reduce the number of uses of those instructions.
96 /// Return the specified operand before it is turned into an undef.
98 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
99 Value *R = I.getOperand(op);
101 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
102 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
103 WorkList.push_back(Op);
104 // Set the operand to undef to drop the use.
105 I.setOperand(i, UndefValue::get(Op->getType()));
111 // removeFromWorkList - remove all instances of I from the worklist.
112 void removeFromWorkList(Instruction *I);
114 virtual bool runOnFunction(Function &F);
116 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
117 AU.addRequired<TargetData>();
118 AU.addPreservedID(LCSSAID);
119 AU.setPreservesCFG();
122 TargetData &getTargetData() const { return *TD; }
124 // Visitation implementation - Implement instruction combining for different
125 // instruction types. The semantics are as follows:
127 // null - No change was made
128 // I - Change was made, I is still valid, I may be dead though
129 // otherwise - Change was made, replace I with returned instruction
131 Instruction *visitAdd(BinaryOperator &I);
132 Instruction *visitSub(BinaryOperator &I);
133 Instruction *visitMul(BinaryOperator &I);
134 Instruction *commonDivTransforms(BinaryOperator &I);
135 Instruction *commonIDivTransforms(BinaryOperator &I);
136 Instruction *visitUDiv(BinaryOperator &I);
137 Instruction *visitSDiv(BinaryOperator &I);
138 Instruction *visitFDiv(BinaryOperator &I);
139 Instruction *visitRem(BinaryOperator &I);
140 Instruction *visitAnd(BinaryOperator &I);
141 Instruction *visitOr (BinaryOperator &I);
142 Instruction *visitXor(BinaryOperator &I);
143 Instruction *visitSetCondInst(SetCondInst &I);
144 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
146 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
147 Instruction::BinaryOps Cond, Instruction &I);
148 Instruction *visitShiftInst(ShiftInst &I);
149 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
151 Instruction *visitCastInst(CastInst &CI);
152 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
154 Instruction *visitSelectInst(SelectInst &CI);
155 Instruction *visitCallInst(CallInst &CI);
156 Instruction *visitInvokeInst(InvokeInst &II);
157 Instruction *visitPHINode(PHINode &PN);
158 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
159 Instruction *visitAllocationInst(AllocationInst &AI);
160 Instruction *visitFreeInst(FreeInst &FI);
161 Instruction *visitLoadInst(LoadInst &LI);
162 Instruction *visitStoreInst(StoreInst &SI);
163 Instruction *visitBranchInst(BranchInst &BI);
164 Instruction *visitSwitchInst(SwitchInst &SI);
165 Instruction *visitInsertElementInst(InsertElementInst &IE);
166 Instruction *visitExtractElementInst(ExtractElementInst &EI);
167 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
169 // visitInstruction - Specify what to return for unhandled instructions...
170 Instruction *visitInstruction(Instruction &I) { return 0; }
173 Instruction *visitCallSite(CallSite CS);
174 bool transformConstExprCastCall(CallSite CS);
177 // InsertNewInstBefore - insert an instruction New before instruction Old
178 // in the program. Add the new instruction to the worklist.
180 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
181 assert(New && New->getParent() == 0 &&
182 "New instruction already inserted into a basic block!");
183 BasicBlock *BB = Old.getParent();
184 BB->getInstList().insert(&Old, New); // Insert inst
185 WorkList.push_back(New); // Add to worklist
189 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
190 /// This also adds the cast to the worklist. Finally, this returns the
192 Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
193 if (V->getType() == Ty) return V;
195 if (Constant *CV = dyn_cast<Constant>(V))
196 return ConstantExpr::getCast(CV, Ty);
198 Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
199 WorkList.push_back(C);
203 // ReplaceInstUsesWith - This method is to be used when an instruction is
204 // found to be dead, replacable with another preexisting expression. Here
205 // we add all uses of I to the worklist, replace all uses of I with the new
206 // value, then return I, so that the inst combiner will know that I was
209 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
210 AddUsersToWorkList(I); // Add all modified instrs to worklist
212 I.replaceAllUsesWith(V);
215 // If we are replacing the instruction with itself, this must be in a
216 // segment of unreachable code, so just clobber the instruction.
217 I.replaceAllUsesWith(UndefValue::get(I.getType()));
222 // UpdateValueUsesWith - This method is to be used when an value is
223 // found to be replacable with another preexisting expression or was
224 // updated. Here we add all uses of I to the worklist, replace all uses of
225 // I with the new value (unless the instruction was just updated), then
226 // return true, so that the inst combiner will know that I was modified.
228 bool UpdateValueUsesWith(Value *Old, Value *New) {
229 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
231 Old->replaceAllUsesWith(New);
232 if (Instruction *I = dyn_cast<Instruction>(Old))
233 WorkList.push_back(I);
234 if (Instruction *I = dyn_cast<Instruction>(New))
235 WorkList.push_back(I);
239 // EraseInstFromFunction - When dealing with an instruction that has side
240 // effects or produces a void value, we can't rely on DCE to delete the
241 // instruction. Instead, visit methods should return the value returned by
243 Instruction *EraseInstFromFunction(Instruction &I) {
244 assert(I.use_empty() && "Cannot erase instruction that is used!");
245 AddUsesToWorkList(I);
246 removeFromWorkList(&I);
248 return 0; // Don't do anything with FI
252 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
253 /// InsertBefore instruction. This is specialized a bit to avoid inserting
254 /// casts that are known to not do anything...
256 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
257 Instruction *InsertBefore);
259 // SimplifyCommutative - This performs a few simplifications for commutative
261 bool SimplifyCommutative(BinaryOperator &I);
263 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
264 uint64_t &KnownZero, uint64_t &KnownOne,
267 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
268 uint64_t &UndefElts, unsigned Depth = 0);
270 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
271 // PHI node as operand #0, see if we can fold the instruction into the PHI
272 // (which is only possible if all operands to the PHI are constants).
273 Instruction *FoldOpIntoPhi(Instruction &I);
275 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
276 // operator and they all are only used by the PHI, PHI together their
277 // inputs, and do the operation once, to the result of the PHI.
278 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
279 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
282 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
283 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
285 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
286 bool isSub, Instruction &I);
287 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
288 bool Inside, Instruction &IB);
289 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
290 Instruction *MatchBSwap(BinaryOperator &I);
292 Value *EvaluateInDifferentType(Value *V, const Type *Ty);
295 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
298 // getComplexity: Assign a complexity or rank value to LLVM Values...
299 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
300 static unsigned getComplexity(Value *V) {
301 if (isa<Instruction>(V)) {
302 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
306 if (isa<Argument>(V)) return 3;
307 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
310 // isOnlyUse - Return true if this instruction will be deleted if we stop using
312 static bool isOnlyUse(Value *V) {
313 return V->hasOneUse() || isa<Constant>(V);
316 // getPromotedType - Return the specified type promoted as it would be to pass
317 // though a va_arg area...
318 static const Type *getPromotedType(const Type *Ty) {
319 switch (Ty->getTypeID()) {
320 case Type::SByteTyID:
321 case Type::ShortTyID: return Type::IntTy;
322 case Type::UByteTyID:
323 case Type::UShortTyID: return Type::UIntTy;
324 case Type::FloatTyID: return Type::DoubleTy;
329 /// isCast - If the specified operand is a CastInst or a constant expr cast,
330 /// return the operand value, otherwise return null.
331 static Value *isCast(Value *V) {
332 if (CastInst *I = dyn_cast<CastInst>(V))
333 return I->getOperand(0);
334 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
335 if (CE->getOpcode() == Instruction::Cast)
336 return CE->getOperand(0);
347 /// getCastType - In the future, we will split the cast instruction into these
348 /// various types. Until then, we have to do the analysis here.
349 static CastType getCastType(const Type *Src, const Type *Dest) {
350 assert(Src->isIntegral() && Dest->isIntegral() &&
351 "Only works on integral types!");
352 unsigned SrcSize = Src->getPrimitiveSizeInBits();
353 unsigned DestSize = Dest->getPrimitiveSizeInBits();
355 if (SrcSize == DestSize) return Noop;
356 if (SrcSize > DestSize) return Truncate;
357 if (Src->isSigned()) return Signext;
362 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
365 static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
366 const Type *DstTy, TargetData *TD) {
368 // It is legal to eliminate the instruction if casting A->B->A if the sizes
369 // are identical and the bits don't get reinterpreted (for example
370 // int->float->int would not be allowed).
371 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
374 // If we are casting between pointer and integer types, treat pointers as
375 // integers of the appropriate size for the code below.
376 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
377 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
378 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
380 // Allow free casting and conversion of sizes as long as the sign doesn't
382 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
383 CastType FirstCast = getCastType(SrcTy, MidTy);
384 CastType SecondCast = getCastType(MidTy, DstTy);
386 // Capture the effect of these two casts. If the result is a legal cast,
387 // the CastType is stored here, otherwise a special code is used.
388 static const unsigned CastResult[] = {
389 // First cast is noop
391 // First cast is a truncate
392 1, 1, 4, 4, // trunc->extend is not safe to eliminate
393 // First cast is a sign ext
394 2, 5, 2, 4, // signext->zeroext never ok
395 // First cast is a zero ext
399 unsigned Result = CastResult[FirstCast*4+SecondCast];
401 default: assert(0 && "Illegal table value!");
406 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
407 // truncates, we could eliminate more casts.
408 return (unsigned)getCastType(SrcTy, DstTy) == Result;
410 return false; // Not possible to eliminate this here.
412 // Sign or zero extend followed by truncate is always ok if the result
413 // is a truncate or noop.
414 CastType ResultCast = getCastType(SrcTy, DstTy);
415 if (ResultCast == Noop || ResultCast == Truncate)
417 // Otherwise we are still growing the value, we are only safe if the
418 // result will match the sign/zeroextendness of the result.
419 return ResultCast == FirstCast;
423 // If this is a cast from 'float -> double -> integer', cast from
424 // 'float -> integer' directly, as the value isn't changed by the
425 // float->double conversion.
426 if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
427 DstTy->isIntegral() &&
428 SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
431 // Packed type conversions don't modify bits.
432 if (isa<PackedType>(SrcTy) && isa<PackedType>(MidTy) &&isa<PackedType>(DstTy))
438 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
439 /// in any code being generated. It does not require codegen if V is simple
440 /// enough or if the cast can be folded into other casts.
441 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
442 if (V->getType() == Ty || isa<Constant>(V)) return false;
444 // If this is a noop cast, it isn't real codegen.
445 if (V->getType()->isLosslesslyConvertibleTo(Ty))
448 // If this is another cast that can be eliminated, it isn't codegen either.
449 if (const CastInst *CI = dyn_cast<CastInst>(V))
450 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
456 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
457 /// InsertBefore instruction. This is specialized a bit to avoid inserting
458 /// casts that are known to not do anything...
460 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
461 Instruction *InsertBefore) {
462 if (V->getType() == DestTy) return V;
463 if (Constant *C = dyn_cast<Constant>(V))
464 return ConstantExpr::getCast(C, DestTy);
466 return InsertCastBefore(V, DestTy, *InsertBefore);
469 // SimplifyCommutative - This performs a few simplifications for commutative
472 // 1. Order operands such that they are listed from right (least complex) to
473 // left (most complex). This puts constants before unary operators before
476 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
477 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
479 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
480 bool Changed = false;
481 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
482 Changed = !I.swapOperands();
484 if (!I.isAssociative()) return Changed;
485 Instruction::BinaryOps Opcode = I.getOpcode();
486 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
487 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
488 if (isa<Constant>(I.getOperand(1))) {
489 Constant *Folded = ConstantExpr::get(I.getOpcode(),
490 cast<Constant>(I.getOperand(1)),
491 cast<Constant>(Op->getOperand(1)));
492 I.setOperand(0, Op->getOperand(0));
493 I.setOperand(1, Folded);
495 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
496 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
497 isOnlyUse(Op) && isOnlyUse(Op1)) {
498 Constant *C1 = cast<Constant>(Op->getOperand(1));
499 Constant *C2 = cast<Constant>(Op1->getOperand(1));
501 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
502 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
503 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
506 WorkList.push_back(New);
507 I.setOperand(0, New);
508 I.setOperand(1, Folded);
515 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
516 // if the LHS is a constant zero (which is the 'negate' form).
518 static inline Value *dyn_castNegVal(Value *V) {
519 if (BinaryOperator::isNeg(V))
520 return BinaryOperator::getNegArgument(V);
522 // Constants can be considered to be negated values if they can be folded.
523 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
524 return ConstantExpr::getNeg(C);
528 static inline Value *dyn_castNotVal(Value *V) {
529 if (BinaryOperator::isNot(V))
530 return BinaryOperator::getNotArgument(V);
532 // Constants can be considered to be not'ed values...
533 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
534 return ConstantExpr::getNot(C);
538 // dyn_castFoldableMul - If this value is a multiply that can be folded into
539 // other computations (because it has a constant operand), return the
540 // non-constant operand of the multiply, and set CST to point to the multiplier.
541 // Otherwise, return null.
543 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
544 if (V->hasOneUse() && V->getType()->isInteger())
545 if (Instruction *I = dyn_cast<Instruction>(V)) {
546 if (I->getOpcode() == Instruction::Mul)
547 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
548 return I->getOperand(0);
549 if (I->getOpcode() == Instruction::Shl)
550 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
551 // The multiplier is really 1 << CST.
552 Constant *One = ConstantInt::get(V->getType(), 1);
553 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
554 return I->getOperand(0);
560 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
561 /// expression, return it.
562 static User *dyn_castGetElementPtr(Value *V) {
563 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
564 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
565 if (CE->getOpcode() == Instruction::GetElementPtr)
566 return cast<User>(V);
570 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
571 static ConstantInt *AddOne(ConstantInt *C) {
572 return cast<ConstantInt>(ConstantExpr::getAdd(C,
573 ConstantInt::get(C->getType(), 1)));
575 static ConstantInt *SubOne(ConstantInt *C) {
576 return cast<ConstantInt>(ConstantExpr::getSub(C,
577 ConstantInt::get(C->getType(), 1)));
580 /// GetConstantInType - Return a ConstantInt with the specified type and value.
582 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
583 if (Ty->isUnsigned())
584 return ConstantInt::get(Ty, Val);
585 else if (Ty->getTypeID() == Type::BoolTyID)
586 return ConstantBool::get(Val);
588 SVal <<= 64-Ty->getPrimitiveSizeInBits();
589 SVal >>= 64-Ty->getPrimitiveSizeInBits();
590 return ConstantInt::get(Ty, SVal);
594 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
595 /// known to be either zero or one and return them in the KnownZero/KnownOne
596 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
598 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
599 uint64_t &KnownOne, unsigned Depth = 0) {
600 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
601 // we cannot optimize based on the assumption that it is zero without changing
602 // it to be an explicit zero. If we don't change it to zero, other code could
603 // optimized based on the contradictory assumption that it is non-zero.
604 // Because instcombine aggressively folds operations with undef args anyway,
605 // this won't lose us code quality.
606 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
607 // We know all of the bits for a constant!
608 KnownOne = CI->getZExtValue() & Mask;
609 KnownZero = ~KnownOne & Mask;
613 KnownZero = KnownOne = 0; // Don't know anything.
614 if (Depth == 6 || Mask == 0)
615 return; // Limit search depth.
617 uint64_t KnownZero2, KnownOne2;
618 Instruction *I = dyn_cast<Instruction>(V);
621 Mask &= V->getType()->getIntegralTypeMask();
623 switch (I->getOpcode()) {
624 case Instruction::And:
625 // If either the LHS or the RHS are Zero, the result is zero.
626 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
628 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
629 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
630 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
632 // Output known-1 bits are only known if set in both the LHS & RHS.
633 KnownOne &= KnownOne2;
634 // Output known-0 are known to be clear if zero in either the LHS | RHS.
635 KnownZero |= KnownZero2;
637 case Instruction::Or:
638 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
640 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
641 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
642 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
644 // Output known-0 bits are only known if clear in both the LHS & RHS.
645 KnownZero &= KnownZero2;
646 // Output known-1 are known to be set if set in either the LHS | RHS.
647 KnownOne |= KnownOne2;
649 case Instruction::Xor: {
650 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
651 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
652 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
653 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
655 // Output known-0 bits are known if clear or set in both the LHS & RHS.
656 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
657 // Output known-1 are known to be set if set in only one of the LHS, RHS.
658 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
659 KnownZero = KnownZeroOut;
662 case Instruction::Select:
663 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
664 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
665 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
666 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
668 // Only known if known in both the LHS and RHS.
669 KnownOne &= KnownOne2;
670 KnownZero &= KnownZero2;
672 case Instruction::Cast: {
673 const Type *SrcTy = I->getOperand(0)->getType();
674 if (!SrcTy->isIntegral()) return;
676 // If this is an integer truncate or noop, just look in the input.
677 if (SrcTy->getPrimitiveSizeInBits() >=
678 I->getType()->getPrimitiveSizeInBits()) {
679 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
683 // Sign or Zero extension. Compute the bits in the result that are not
684 // present in the input.
685 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
686 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
688 // Handle zero extension.
689 if (!SrcTy->isSigned()) {
690 Mask &= SrcTy->getIntegralTypeMask();
691 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
692 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
693 // The top bits are known to be zero.
694 KnownZero |= NewBits;
697 Mask &= SrcTy->getIntegralTypeMask();
698 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
699 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
701 // If the sign bit of the input is known set or clear, then we know the
702 // top bits of the result.
703 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
704 if (KnownZero & InSignBit) { // Input sign bit known zero
705 KnownZero |= NewBits;
706 KnownOne &= ~NewBits;
707 } else if (KnownOne & InSignBit) { // Input sign bit known set
709 KnownZero &= ~NewBits;
710 } else { // Input sign bit unknown
711 KnownZero &= ~NewBits;
712 KnownOne &= ~NewBits;
717 case Instruction::Shl:
718 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
719 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
720 uint64_t ShiftAmt = SA->getZExtValue();
722 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
723 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
724 KnownZero <<= ShiftAmt;
725 KnownOne <<= ShiftAmt;
726 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
730 case Instruction::Shr:
731 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
732 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
733 // Compute the new bits that are at the top now.
734 uint64_t ShiftAmt = SA->getZExtValue();
735 uint64_t HighBits = (1ULL << ShiftAmt)-1;
736 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
738 if (I->getType()->isUnsigned()) { // Unsigned shift right.
740 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
741 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
742 KnownZero >>= ShiftAmt;
743 KnownOne >>= ShiftAmt;
744 KnownZero |= HighBits; // high bits known zero.
747 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
748 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
749 KnownZero >>= ShiftAmt;
750 KnownOne >>= ShiftAmt;
752 // Handle the sign bits.
753 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
754 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
756 if (KnownZero & SignBit) { // New bits are known zero.
757 KnownZero |= HighBits;
758 } else if (KnownOne & SignBit) { // New bits are known one.
759 KnownOne |= HighBits;
768 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
769 /// this predicate to simplify operations downstream. Mask is known to be zero
770 /// for bits that V cannot have.
771 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
772 uint64_t KnownZero, KnownOne;
773 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
774 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
775 return (KnownZero & Mask) == Mask;
778 /// ShrinkDemandedConstant - Check to see if the specified operand of the
779 /// specified instruction is a constant integer. If so, check to see if there
780 /// are any bits set in the constant that are not demanded. If so, shrink the
781 /// constant and return true.
782 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
784 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
785 if (!OpC) return false;
787 // If there are no bits set that aren't demanded, nothing to do.
788 if ((~Demanded & OpC->getZExtValue()) == 0)
791 // This is producing any bits that are not needed, shrink the RHS.
792 uint64_t Val = Demanded & OpC->getZExtValue();
793 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
797 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
798 // set of known zero and one bits, compute the maximum and minimum values that
799 // could have the specified known zero and known one bits, returning them in
801 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
804 int64_t &Min, int64_t &Max) {
805 uint64_t TypeBits = Ty->getIntegralTypeMask();
806 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
808 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
810 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
811 // bit if it is unknown.
813 Max = KnownOne|UnknownBits;
815 if (SignBit & UnknownBits) { // Sign bit is unknown
820 // Sign extend the min/max values.
821 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
822 Min = (Min << ShAmt) >> ShAmt;
823 Max = (Max << ShAmt) >> ShAmt;
826 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
827 // a set of known zero and one bits, compute the maximum and minimum values that
828 // could have the specified known zero and known one bits, returning them in
830 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
835 uint64_t TypeBits = Ty->getIntegralTypeMask();
836 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
838 // The minimum value is when the unknown bits are all zeros.
840 // The maximum value is when the unknown bits are all ones.
841 Max = KnownOne|UnknownBits;
845 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
846 /// DemandedMask bits of the result of V are ever used downstream. If we can
847 /// use this information to simplify V, do so and return true. Otherwise,
848 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
849 /// the expression (used to simplify the caller). The KnownZero/One bits may
850 /// only be accurate for those bits in the DemandedMask.
851 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
852 uint64_t &KnownZero, uint64_t &KnownOne,
854 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
855 // We know all of the bits for a constant!
856 KnownOne = CI->getZExtValue() & DemandedMask;
857 KnownZero = ~KnownOne & DemandedMask;
861 KnownZero = KnownOne = 0;
862 if (!V->hasOneUse()) { // Other users may use these bits.
863 if (Depth != 0) { // Not at the root.
864 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
865 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
868 // If this is the root being simplified, allow it to have multiple uses,
869 // just set the DemandedMask to all bits.
870 DemandedMask = V->getType()->getIntegralTypeMask();
871 } else if (DemandedMask == 0) { // Not demanding any bits from V.
872 if (V != UndefValue::get(V->getType()))
873 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
875 } else if (Depth == 6) { // Limit search depth.
879 Instruction *I = dyn_cast<Instruction>(V);
880 if (!I) return false; // Only analyze instructions.
882 DemandedMask &= V->getType()->getIntegralTypeMask();
884 uint64_t KnownZero2, KnownOne2;
885 switch (I->getOpcode()) {
887 case Instruction::And:
888 // If either the LHS or the RHS are Zero, the result is zero.
889 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
890 KnownZero, KnownOne, Depth+1))
892 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
894 // If something is known zero on the RHS, the bits aren't demanded on the
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
897 KnownZero2, KnownOne2, Depth+1))
899 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
901 // If all of the demanded bits are known one on one side, return the other.
902 // These bits cannot contribute to the result of the 'and'.
903 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
904 return UpdateValueUsesWith(I, I->getOperand(0));
905 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
906 return UpdateValueUsesWith(I, I->getOperand(1));
908 // If all of the demanded bits in the inputs are known zeros, return zero.
909 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
910 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
912 // If the RHS is a constant, see if we can simplify it.
913 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
914 return UpdateValueUsesWith(I, I);
916 // Output known-1 bits are only known if set in both the LHS & RHS.
917 KnownOne &= KnownOne2;
918 // Output known-0 are known to be clear if zero in either the LHS | RHS.
919 KnownZero |= KnownZero2;
921 case Instruction::Or:
922 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
923 KnownZero, KnownOne, Depth+1))
925 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
926 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
927 KnownZero2, KnownOne2, Depth+1))
929 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
931 // If all of the demanded bits are known zero on one side, return the other.
932 // These bits cannot contribute to the result of the 'or'.
933 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
934 return UpdateValueUsesWith(I, I->getOperand(0));
935 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
936 return UpdateValueUsesWith(I, I->getOperand(1));
938 // If all of the potentially set bits on one side are known to be set on
939 // the other side, just use the 'other' side.
940 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
941 (DemandedMask & (~KnownZero)))
942 return UpdateValueUsesWith(I, I->getOperand(0));
943 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
944 (DemandedMask & (~KnownZero2)))
945 return UpdateValueUsesWith(I, I->getOperand(1));
947 // If the RHS is a constant, see if we can simplify it.
948 if (ShrinkDemandedConstant(I, 1, DemandedMask))
949 return UpdateValueUsesWith(I, I);
951 // Output known-0 bits are only known if clear in both the LHS & RHS.
952 KnownZero &= KnownZero2;
953 // Output known-1 are known to be set if set in either the LHS | RHS.
954 KnownOne |= KnownOne2;
956 case Instruction::Xor: {
957 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
958 KnownZero, KnownOne, Depth+1))
960 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
961 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
962 KnownZero2, KnownOne2, Depth+1))
964 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
966 // If all of the demanded bits are known zero on one side, return the other.
967 // These bits cannot contribute to the result of the 'xor'.
968 if ((DemandedMask & KnownZero) == DemandedMask)
969 return UpdateValueUsesWith(I, I->getOperand(0));
970 if ((DemandedMask & KnownZero2) == DemandedMask)
971 return UpdateValueUsesWith(I, I->getOperand(1));
973 // Output known-0 bits are known if clear or set in both the LHS & RHS.
974 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
975 // Output known-1 are known to be set if set in only one of the LHS, RHS.
976 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
978 // If all of the unknown bits are known to be zero on one side or the other
979 // (but not both) turn this into an *inclusive* or.
980 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
981 if (uint64_t UnknownBits = DemandedMask & ~(KnownZeroOut|KnownOneOut)) {
982 if ((UnknownBits & (KnownZero|KnownZero2)) == UnknownBits) {
984 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
986 InsertNewInstBefore(Or, *I);
987 return UpdateValueUsesWith(I, Or);
991 // If all of the demanded bits on one side are known, and all of the set
992 // bits on that side are also known to be set on the other side, turn this
993 // into an AND, as we know the bits will be cleared.
994 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
995 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
996 if ((KnownOne & KnownOne2) == KnownOne) {
997 Constant *AndC = GetConstantInType(I->getType(),
998 ~KnownOne & DemandedMask);
1000 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1001 InsertNewInstBefore(And, *I);
1002 return UpdateValueUsesWith(I, And);
1006 // If the RHS is a constant, see if we can simplify it.
1007 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1008 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1009 return UpdateValueUsesWith(I, I);
1011 KnownZero = KnownZeroOut;
1012 KnownOne = KnownOneOut;
1015 case Instruction::Select:
1016 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1017 KnownZero, KnownOne, Depth+1))
1019 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1020 KnownZero2, KnownOne2, Depth+1))
1022 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1023 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1025 // If the operands are constants, see if we can simplify them.
1026 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1027 return UpdateValueUsesWith(I, I);
1028 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1029 return UpdateValueUsesWith(I, I);
1031 // Only known if known in both the LHS and RHS.
1032 KnownOne &= KnownOne2;
1033 KnownZero &= KnownZero2;
1035 case Instruction::Cast: {
1036 const Type *SrcTy = I->getOperand(0)->getType();
1037 if (!SrcTy->isIntegral()) return false;
1039 // If this is an integer truncate or noop, just look in the input.
1040 if (SrcTy->getPrimitiveSizeInBits() >=
1041 I->getType()->getPrimitiveSizeInBits()) {
1042 // Cast to bool is a comparison against 0, which demands all bits. We
1043 // can't propagate anything useful up.
1044 if (I->getType() == Type::BoolTy)
1047 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1048 KnownZero, KnownOne, Depth+1))
1050 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1054 // Sign or Zero extension. Compute the bits in the result that are not
1055 // present in the input.
1056 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1057 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1059 // Handle zero extension.
1060 if (!SrcTy->isSigned()) {
1061 DemandedMask &= SrcTy->getIntegralTypeMask();
1062 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1063 KnownZero, KnownOne, Depth+1))
1065 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1066 // The top bits are known to be zero.
1067 KnownZero |= NewBits;
1070 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1071 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1073 // If any of the sign extended bits are demanded, we know that the sign
1075 if (NewBits & DemandedMask)
1076 InputDemandedBits |= InSignBit;
1078 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1079 KnownZero, KnownOne, Depth+1))
1081 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1083 // If the sign bit of the input is known set or clear, then we know the
1084 // top bits of the result.
1086 // If the input sign bit is known zero, or if the NewBits are not demanded
1087 // convert this into a zero extension.
1088 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1089 // Convert to unsigned first.
1091 InsertCastBefore(I->getOperand(0), SrcTy->getUnsignedVersion(), *I);
1092 // Then cast that to the destination type.
1093 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1094 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1095 return UpdateValueUsesWith(I, NewVal);
1096 } else if (KnownOne & InSignBit) { // Input sign bit known set
1097 KnownOne |= NewBits;
1098 KnownZero &= ~NewBits;
1099 } else { // Input sign bit unknown
1100 KnownZero &= ~NewBits;
1101 KnownOne &= ~NewBits;
1106 case Instruction::Shl:
1107 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1108 uint64_t ShiftAmt = SA->getZExtValue();
1109 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1110 KnownZero, KnownOne, Depth+1))
1112 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1113 KnownZero <<= ShiftAmt;
1114 KnownOne <<= ShiftAmt;
1115 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1118 case Instruction::Shr:
1119 // If this is an arithmetic shift right and only the low-bit is set, we can
1120 // always convert this into a logical shr, even if the shift amount is
1121 // variable. The low bit of the shift cannot be an input sign bit unless
1122 // the shift amount is >= the size of the datatype, which is undefined.
1123 if (DemandedMask == 1 && I->getType()->isSigned()) {
1124 // Convert the input to unsigned.
1125 Value *NewVal = InsertCastBefore(I->getOperand(0),
1126 I->getType()->getUnsignedVersion(), *I);
1127 // Perform the unsigned shift right.
1128 NewVal = new ShiftInst(Instruction::Shr, NewVal, I->getOperand(1),
1130 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1131 // Then cast that to the destination type.
1132 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1133 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1134 return UpdateValueUsesWith(I, NewVal);
1137 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1138 unsigned ShiftAmt = SA->getZExtValue();
1140 // Compute the new bits that are at the top now.
1141 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1142 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1143 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1144 if (I->getType()->isUnsigned()) { // Unsigned shift right.
1145 if (SimplifyDemandedBits(I->getOperand(0),
1146 (DemandedMask << ShiftAmt) & TypeMask,
1147 KnownZero, KnownOne, Depth+1))
1149 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1150 KnownZero &= TypeMask;
1151 KnownOne &= TypeMask;
1152 KnownZero >>= ShiftAmt;
1153 KnownOne >>= ShiftAmt;
1154 KnownZero |= HighBits; // high bits known zero.
1155 } else { // Signed shift right.
1156 if (SimplifyDemandedBits(I->getOperand(0),
1157 (DemandedMask << ShiftAmt) & TypeMask,
1158 KnownZero, KnownOne, Depth+1))
1160 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1161 KnownZero &= TypeMask;
1162 KnownOne &= TypeMask;
1163 KnownZero >>= ShiftAmt;
1164 KnownOne >>= ShiftAmt;
1166 // Handle the sign bits.
1167 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1168 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1170 // If the input sign bit is known to be zero, or if none of the top bits
1171 // are demanded, turn this into an unsigned shift right.
1172 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1173 // Convert the input to unsigned.
1174 Value *NewVal = InsertCastBefore(I->getOperand(0),
1175 I->getType()->getUnsignedVersion(), *I);
1176 // Perform the unsigned shift right.
1177 NewVal = new ShiftInst(Instruction::Shr, NewVal, SA, I->getName());
1178 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1179 // Then cast that to the destination type.
1180 NewVal = new CastInst(NewVal, I->getType(), I->getName());
1181 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1182 return UpdateValueUsesWith(I, NewVal);
1183 } else if (KnownOne & SignBit) { // New bits are known one.
1184 KnownOne |= HighBits;
1191 // If the client is only demanding bits that we know, return the known
1193 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1194 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1199 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1200 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1201 /// actually used by the caller. This method analyzes which elements of the
1202 /// operand are undef and returns that information in UndefElts.
1204 /// If the information about demanded elements can be used to simplify the
1205 /// operation, the operation is simplified, then the resultant value is
1206 /// returned. This returns null if no change was made.
1207 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1208 uint64_t &UndefElts,
1210 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1211 assert(VWidth <= 64 && "Vector too wide to analyze!");
1212 uint64_t EltMask = ~0ULL >> (64-VWidth);
1213 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1214 "Invalid DemandedElts!");
1216 if (isa<UndefValue>(V)) {
1217 // If the entire vector is undefined, just return this info.
1218 UndefElts = EltMask;
1220 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1221 UndefElts = EltMask;
1222 return UndefValue::get(V->getType());
1226 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1227 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1228 Constant *Undef = UndefValue::get(EltTy);
1230 std::vector<Constant*> Elts;
1231 for (unsigned i = 0; i != VWidth; ++i)
1232 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1233 Elts.push_back(Undef);
1234 UndefElts |= (1ULL << i);
1235 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1236 Elts.push_back(Undef);
1237 UndefElts |= (1ULL << i);
1238 } else { // Otherwise, defined.
1239 Elts.push_back(CP->getOperand(i));
1242 // If we changed the constant, return it.
1243 Constant *NewCP = ConstantPacked::get(Elts);
1244 return NewCP != CP ? NewCP : 0;
1245 } else if (isa<ConstantAggregateZero>(V)) {
1246 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1248 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1249 Constant *Zero = Constant::getNullValue(EltTy);
1250 Constant *Undef = UndefValue::get(EltTy);
1251 std::vector<Constant*> Elts;
1252 for (unsigned i = 0; i != VWidth; ++i)
1253 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1254 UndefElts = DemandedElts ^ EltMask;
1255 return ConstantPacked::get(Elts);
1258 if (!V->hasOneUse()) { // Other users may use these bits.
1259 if (Depth != 0) { // Not at the root.
1260 // TODO: Just compute the UndefElts information recursively.
1264 } else if (Depth == 10) { // Limit search depth.
1268 Instruction *I = dyn_cast<Instruction>(V);
1269 if (!I) return false; // Only analyze instructions.
1271 bool MadeChange = false;
1272 uint64_t UndefElts2;
1274 switch (I->getOpcode()) {
1277 case Instruction::InsertElement: {
1278 // If this is a variable index, we don't know which element it overwrites.
1279 // demand exactly the same input as we produce.
1280 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1282 // Note that we can't propagate undef elt info, because we don't know
1283 // which elt is getting updated.
1284 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1285 UndefElts2, Depth+1);
1286 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1290 // If this is inserting an element that isn't demanded, remove this
1292 unsigned IdxNo = Idx->getZExtValue();
1293 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1294 return AddSoonDeadInstToWorklist(*I, 0);
1296 // Otherwise, the element inserted overwrites whatever was there, so the
1297 // input demanded set is simpler than the output set.
1298 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1299 DemandedElts & ~(1ULL << IdxNo),
1300 UndefElts, Depth+1);
1301 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1303 // The inserted element is defined.
1304 UndefElts |= 1ULL << IdxNo;
1308 case Instruction::And:
1309 case Instruction::Or:
1310 case Instruction::Xor:
1311 case Instruction::Add:
1312 case Instruction::Sub:
1313 case Instruction::Mul:
1314 // div/rem demand all inputs, because they don't want divide by zero.
1315 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1316 UndefElts, Depth+1);
1317 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1318 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1319 UndefElts2, Depth+1);
1320 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1322 // Output elements are undefined if both are undefined. Consider things
1323 // like undef&0. The result is known zero, not undef.
1324 UndefElts &= UndefElts2;
1327 case Instruction::Call: {
1328 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1330 switch (II->getIntrinsicID()) {
1333 // Binary vector operations that work column-wise. A dest element is a
1334 // function of the corresponding input elements from the two inputs.
1335 case Intrinsic::x86_sse_sub_ss:
1336 case Intrinsic::x86_sse_mul_ss:
1337 case Intrinsic::x86_sse_min_ss:
1338 case Intrinsic::x86_sse_max_ss:
1339 case Intrinsic::x86_sse2_sub_sd:
1340 case Intrinsic::x86_sse2_mul_sd:
1341 case Intrinsic::x86_sse2_min_sd:
1342 case Intrinsic::x86_sse2_max_sd:
1343 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1344 UndefElts, Depth+1);
1345 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1346 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1347 UndefElts2, Depth+1);
1348 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1350 // If only the low elt is demanded and this is a scalarizable intrinsic,
1351 // scalarize it now.
1352 if (DemandedElts == 1) {
1353 switch (II->getIntrinsicID()) {
1355 case Intrinsic::x86_sse_sub_ss:
1356 case Intrinsic::x86_sse_mul_ss:
1357 case Intrinsic::x86_sse2_sub_sd:
1358 case Intrinsic::x86_sse2_mul_sd:
1359 // TODO: Lower MIN/MAX/ABS/etc
1360 Value *LHS = II->getOperand(1);
1361 Value *RHS = II->getOperand(2);
1362 // Extract the element as scalars.
1363 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1364 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1366 switch (II->getIntrinsicID()) {
1367 default: assert(0 && "Case stmts out of sync!");
1368 case Intrinsic::x86_sse_sub_ss:
1369 case Intrinsic::x86_sse2_sub_sd:
1370 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1371 II->getName()), *II);
1373 case Intrinsic::x86_sse_mul_ss:
1374 case Intrinsic::x86_sse2_mul_sd:
1375 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1376 II->getName()), *II);
1381 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1383 InsertNewInstBefore(New, *II);
1384 AddSoonDeadInstToWorklist(*II, 0);
1389 // Output elements are undefined if both are undefined. Consider things
1390 // like undef&0. The result is known zero, not undef.
1391 UndefElts &= UndefElts2;
1397 return MadeChange ? I : 0;
1400 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
1401 // true when both operands are equal...
1403 static bool isTrueWhenEqual(Instruction &I) {
1404 return I.getOpcode() == Instruction::SetEQ ||
1405 I.getOpcode() == Instruction::SetGE ||
1406 I.getOpcode() == Instruction::SetLE;
1409 /// AssociativeOpt - Perform an optimization on an associative operator. This
1410 /// function is designed to check a chain of associative operators for a
1411 /// potential to apply a certain optimization. Since the optimization may be
1412 /// applicable if the expression was reassociated, this checks the chain, then
1413 /// reassociates the expression as necessary to expose the optimization
1414 /// opportunity. This makes use of a special Functor, which must define
1415 /// 'shouldApply' and 'apply' methods.
1417 template<typename Functor>
1418 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1419 unsigned Opcode = Root.getOpcode();
1420 Value *LHS = Root.getOperand(0);
1422 // Quick check, see if the immediate LHS matches...
1423 if (F.shouldApply(LHS))
1424 return F.apply(Root);
1426 // Otherwise, if the LHS is not of the same opcode as the root, return.
1427 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1428 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1429 // Should we apply this transform to the RHS?
1430 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1432 // If not to the RHS, check to see if we should apply to the LHS...
1433 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1434 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1438 // If the functor wants to apply the optimization to the RHS of LHSI,
1439 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1441 BasicBlock *BB = Root.getParent();
1443 // Now all of the instructions are in the current basic block, go ahead
1444 // and perform the reassociation.
1445 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1447 // First move the selected RHS to the LHS of the root...
1448 Root.setOperand(0, LHSI->getOperand(1));
1450 // Make what used to be the LHS of the root be the user of the root...
1451 Value *ExtraOperand = TmpLHSI->getOperand(1);
1452 if (&Root == TmpLHSI) {
1453 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1456 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1457 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1458 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1459 BasicBlock::iterator ARI = &Root; ++ARI;
1460 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1463 // Now propagate the ExtraOperand down the chain of instructions until we
1465 while (TmpLHSI != LHSI) {
1466 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1467 // Move the instruction to immediately before the chain we are
1468 // constructing to avoid breaking dominance properties.
1469 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1470 BB->getInstList().insert(ARI, NextLHSI);
1473 Value *NextOp = NextLHSI->getOperand(1);
1474 NextLHSI->setOperand(1, ExtraOperand);
1476 ExtraOperand = NextOp;
1479 // Now that the instructions are reassociated, have the functor perform
1480 // the transformation...
1481 return F.apply(Root);
1484 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1490 // AddRHS - Implements: X + X --> X << 1
1493 AddRHS(Value *rhs) : RHS(rhs) {}
1494 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1495 Instruction *apply(BinaryOperator &Add) const {
1496 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1497 ConstantInt::get(Type::UByteTy, 1));
1501 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1503 struct AddMaskingAnd {
1505 AddMaskingAnd(Constant *c) : C2(c) {}
1506 bool shouldApply(Value *LHS) const {
1508 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1509 ConstantExpr::getAnd(C1, C2)->isNullValue();
1511 Instruction *apply(BinaryOperator &Add) const {
1512 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1516 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1518 if (isa<CastInst>(I)) {
1519 if (Constant *SOC = dyn_cast<Constant>(SO))
1520 return ConstantExpr::getCast(SOC, I.getType());
1522 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
1523 SO->getName() + ".cast"), I);
1526 // Figure out if the constant is the left or the right argument.
1527 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1528 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1530 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1532 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1533 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1536 Value *Op0 = SO, *Op1 = ConstOperand;
1538 std::swap(Op0, Op1);
1540 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1541 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1542 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1543 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1545 assert(0 && "Unknown binary instruction type!");
1548 return IC->InsertNewInstBefore(New, I);
1551 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1552 // constant as the other operand, try to fold the binary operator into the
1553 // select arguments. This also works for Cast instructions, which obviously do
1554 // not have a second operand.
1555 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1557 // Don't modify shared select instructions
1558 if (!SI->hasOneUse()) return 0;
1559 Value *TV = SI->getOperand(1);
1560 Value *FV = SI->getOperand(2);
1562 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1563 // Bool selects with constant operands can be folded to logical ops.
1564 if (SI->getType() == Type::BoolTy) return 0;
1566 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1567 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1569 return new SelectInst(SI->getCondition(), SelectTrueVal,
1576 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1577 /// node as operand #0, see if we can fold the instruction into the PHI (which
1578 /// is only possible if all operands to the PHI are constants).
1579 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1580 PHINode *PN = cast<PHINode>(I.getOperand(0));
1581 unsigned NumPHIValues = PN->getNumIncomingValues();
1582 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1584 // Check to see if all of the operands of the PHI are constants. If there is
1585 // one non-constant value, remember the BB it is. If there is more than one
1587 BasicBlock *NonConstBB = 0;
1588 for (unsigned i = 0; i != NumPHIValues; ++i)
1589 if (!isa<Constant>(PN->getIncomingValue(i))) {
1590 if (NonConstBB) return 0; // More than one non-const value.
1591 NonConstBB = PN->getIncomingBlock(i);
1593 // If the incoming non-constant value is in I's block, we have an infinite
1595 if (NonConstBB == I.getParent())
1599 // If there is exactly one non-constant value, we can insert a copy of the
1600 // operation in that block. However, if this is a critical edge, we would be
1601 // inserting the computation one some other paths (e.g. inside a loop). Only
1602 // do this if the pred block is unconditionally branching into the phi block.
1604 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1605 if (!BI || !BI->isUnconditional()) return 0;
1608 // Okay, we can do the transformation: create the new PHI node.
1609 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1611 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1612 InsertNewInstBefore(NewPN, *PN);
1614 // Next, add all of the operands to the PHI.
1615 if (I.getNumOperands() == 2) {
1616 Constant *C = cast<Constant>(I.getOperand(1));
1617 for (unsigned i = 0; i != NumPHIValues; ++i) {
1619 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1620 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1622 assert(PN->getIncomingBlock(i) == NonConstBB);
1623 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1624 InV = BinaryOperator::create(BO->getOpcode(),
1625 PN->getIncomingValue(i), C, "phitmp",
1626 NonConstBB->getTerminator());
1627 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1628 InV = new ShiftInst(SI->getOpcode(),
1629 PN->getIncomingValue(i), C, "phitmp",
1630 NonConstBB->getTerminator());
1632 assert(0 && "Unknown binop!");
1634 WorkList.push_back(cast<Instruction>(InV));
1636 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1639 assert(isa<CastInst>(I) && "Unary op should be a cast!");
1640 const Type *RetTy = I.getType();
1641 for (unsigned i = 0; i != NumPHIValues; ++i) {
1643 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1644 InV = ConstantExpr::getCast(InC, RetTy);
1646 assert(PN->getIncomingBlock(i) == NonConstBB);
1647 InV = new CastInst(PN->getIncomingValue(i), I.getType(), "phitmp",
1648 NonConstBB->getTerminator());
1649 WorkList.push_back(cast<Instruction>(InV));
1651 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1654 return ReplaceInstUsesWith(I, NewPN);
1657 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1658 bool Changed = SimplifyCommutative(I);
1659 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1661 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1662 // X + undef -> undef
1663 if (isa<UndefValue>(RHS))
1664 return ReplaceInstUsesWith(I, RHS);
1667 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
1668 if (RHSC->isNullValue())
1669 return ReplaceInstUsesWith(I, LHS);
1670 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1671 if (CFP->isExactlyValue(-0.0))
1672 return ReplaceInstUsesWith(I, LHS);
1675 // X + (signbit) --> X ^ signbit
1676 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1677 uint64_t Val = CI->getZExtValue();
1678 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1679 return BinaryOperator::createXor(LHS, RHS);
1682 if (isa<PHINode>(LHS))
1683 if (Instruction *NV = FoldOpIntoPhi(I))
1686 ConstantInt *XorRHS = 0;
1688 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1689 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1690 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1691 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1693 uint64_t C0080Val = 1ULL << 31;
1694 int64_t CFF80Val = -C0080Val;
1697 if (TySizeBits > Size) {
1699 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1700 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1701 if (RHSSExt == CFF80Val) {
1702 if (XorRHS->getZExtValue() == C0080Val)
1704 } else if (RHSZExt == C0080Val) {
1705 if (XorRHS->getSExtValue() == CFF80Val)
1709 // This is a sign extend if the top bits are known zero.
1710 uint64_t Mask = ~0ULL;
1711 Mask <<= 64-(TySizeBits-Size);
1712 Mask &= XorLHS->getType()->getIntegralTypeMask();
1713 if (!MaskedValueIsZero(XorLHS, Mask))
1714 Size = 0; // Not a sign ext, but can't be any others either.
1721 } while (Size >= 8);
1724 const Type *MiddleType = 0;
1727 case 32: MiddleType = Type::IntTy; break;
1728 case 16: MiddleType = Type::ShortTy; break;
1729 case 8: MiddleType = Type::SByteTy; break;
1732 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
1733 InsertNewInstBefore(NewTrunc, I);
1734 return new CastInst(NewTrunc, I.getType());
1740 if (I.getType()->isInteger()) {
1741 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1743 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1744 if (RHSI->getOpcode() == Instruction::Sub)
1745 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1746 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1748 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1749 if (LHSI->getOpcode() == Instruction::Sub)
1750 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1751 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1756 if (Value *V = dyn_castNegVal(LHS))
1757 return BinaryOperator::createSub(RHS, V);
1760 if (!isa<Constant>(RHS))
1761 if (Value *V = dyn_castNegVal(RHS))
1762 return BinaryOperator::createSub(LHS, V);
1766 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1767 if (X == RHS) // X*C + X --> X * (C+1)
1768 return BinaryOperator::createMul(RHS, AddOne(C2));
1770 // X*C1 + X*C2 --> X * (C1+C2)
1772 if (X == dyn_castFoldableMul(RHS, C1))
1773 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1776 // X + X*C --> X * (C+1)
1777 if (dyn_castFoldableMul(RHS, C2) == LHS)
1778 return BinaryOperator::createMul(LHS, AddOne(C2));
1781 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1782 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1783 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1785 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1787 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1788 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1789 return BinaryOperator::createSub(C, X);
1792 // (X & FF00) + xx00 -> (X+xx00) & FF00
1793 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1794 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1795 if (Anded == CRHS) {
1796 // See if all bits from the first bit set in the Add RHS up are included
1797 // in the mask. First, get the rightmost bit.
1798 uint64_t AddRHSV = CRHS->getZExtValue();
1800 // Form a mask of all bits from the lowest bit added through the top.
1801 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1802 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1804 // See if the and mask includes all of these bits.
1805 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1807 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1808 // Okay, the xform is safe. Insert the new add pronto.
1809 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1810 LHS->getName()), I);
1811 return BinaryOperator::createAnd(NewAdd, C2);
1816 // Try to fold constant add into select arguments.
1817 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1818 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1822 // add (cast *A to intptrtype) B ->
1823 // cast (GEP (cast *A to sbyte*) B) ->
1826 CastInst* CI = dyn_cast<CastInst>(LHS);
1829 CI = dyn_cast<CastInst>(RHS);
1832 if (CI && CI->getType()->isSized() &&
1833 (CI->getType()->getPrimitiveSize() ==
1834 TD->getIntPtrType()->getPrimitiveSize())
1835 && isa<PointerType>(CI->getOperand(0)->getType())) {
1836 Value* I2 = InsertCastBefore(CI->getOperand(0),
1837 PointerType::get(Type::SByteTy), I);
1838 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1839 return new CastInst(I2, CI->getType());
1843 return Changed ? &I : 0;
1846 // isSignBit - Return true if the value represented by the constant only has the
1847 // highest order bit set.
1848 static bool isSignBit(ConstantInt *CI) {
1849 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1850 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1853 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1855 static Value *RemoveNoopCast(Value *V) {
1856 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1857 const Type *CTy = CI->getType();
1858 const Type *OpTy = CI->getOperand(0)->getType();
1859 if (CTy->isInteger() && OpTy->isInteger()) {
1860 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1861 return RemoveNoopCast(CI->getOperand(0));
1862 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1863 return RemoveNoopCast(CI->getOperand(0));
1868 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1869 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1871 if (Op0 == Op1) // sub X, X -> 0
1872 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1874 // If this is a 'B = x-(-A)', change to B = x+A...
1875 if (Value *V = dyn_castNegVal(Op1))
1876 return BinaryOperator::createAdd(Op0, V);
1878 if (isa<UndefValue>(Op0))
1879 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1880 if (isa<UndefValue>(Op1))
1881 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1883 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1884 // Replace (-1 - A) with (~A)...
1885 if (C->isAllOnesValue())
1886 return BinaryOperator::createNot(Op1);
1888 // C - ~X == X + (1+C)
1890 if (match(Op1, m_Not(m_Value(X))))
1891 return BinaryOperator::createAdd(X,
1892 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1893 // -((uint)X >> 31) -> ((int)X >> 31)
1894 // -((int)X >> 31) -> ((uint)X >> 31)
1895 if (C->isNullValue()) {
1896 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1897 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1898 if (SI->getOpcode() == Instruction::Shr)
1899 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1901 if (SI->getType()->isSigned())
1902 NewTy = SI->getType()->getUnsignedVersion();
1904 NewTy = SI->getType()->getSignedVersion();
1905 // Check to see if we are shifting out everything but the sign bit.
1906 if (CU->getZExtValue() ==
1907 SI->getType()->getPrimitiveSizeInBits()-1) {
1908 // Ok, the transformation is safe. Insert a cast of the incoming
1909 // value, then the new shift, then the new cast.
1910 Value *InV = InsertCastBefore(SI->getOperand(0), NewTy, I);
1911 Instruction *NewShift = new ShiftInst(Instruction::Shr, InV,
1913 if (NewShift->getType() == I.getType())
1916 InsertNewInstBefore(NewShift, I);
1917 return new CastInst(NewShift, I.getType());
1923 // Try to fold constant sub into select arguments.
1924 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1925 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1928 if (isa<PHINode>(Op0))
1929 if (Instruction *NV = FoldOpIntoPhi(I))
1933 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1934 if (Op1I->getOpcode() == Instruction::Add &&
1935 !Op0->getType()->isFloatingPoint()) {
1936 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1937 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1938 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1939 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1940 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1941 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1942 // C1-(X+C2) --> (C1-C2)-X
1943 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1944 Op1I->getOperand(0));
1948 if (Op1I->hasOneUse()) {
1949 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1950 // is not used by anyone else...
1952 if (Op1I->getOpcode() == Instruction::Sub &&
1953 !Op1I->getType()->isFloatingPoint()) {
1954 // Swap the two operands of the subexpr...
1955 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1956 Op1I->setOperand(0, IIOp1);
1957 Op1I->setOperand(1, IIOp0);
1959 // Create the new top level add instruction...
1960 return BinaryOperator::createAdd(Op0, Op1);
1963 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1965 if (Op1I->getOpcode() == Instruction::And &&
1966 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1967 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1970 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1971 return BinaryOperator::createAnd(Op0, NewNot);
1974 // 0 - (X sdiv C) -> (X sdiv -C)
1975 if (Op1I->getOpcode() == Instruction::SDiv)
1976 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
1977 if (CSI->isNullValue())
1978 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1979 return BinaryOperator::createSDiv(Op1I->getOperand(0),
1980 ConstantExpr::getNeg(DivRHS));
1982 // X - X*C --> X * (1-C)
1983 ConstantInt *C2 = 0;
1984 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1986 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1987 return BinaryOperator::createMul(Op0, CP1);
1992 if (!Op0->getType()->isFloatingPoint())
1993 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1994 if (Op0I->getOpcode() == Instruction::Add) {
1995 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1996 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1997 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1998 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1999 } else if (Op0I->getOpcode() == Instruction::Sub) {
2000 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2001 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2005 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2006 if (X == Op1) { // X*C - X --> X * (C-1)
2007 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2008 return BinaryOperator::createMul(Op1, CP1);
2011 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2012 if (X == dyn_castFoldableMul(Op1, C2))
2013 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2018 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
2019 /// really just returns true if the most significant (sign) bit is set.
2020 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
2021 if (RHS->getType()->isSigned()) {
2022 // True if source is LHS < 0 or LHS <= -1
2023 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
2024 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
2026 ConstantInt *RHSC = cast<ConstantInt>(RHS);
2027 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
2028 // the size of the integer type.
2029 if (Opcode == Instruction::SetGE)
2030 return RHSC->getZExtValue() ==
2031 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
2032 if (Opcode == Instruction::SetGT)
2033 return RHSC->getZExtValue() ==
2034 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2039 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2040 bool Changed = SimplifyCommutative(I);
2041 Value *Op0 = I.getOperand(0);
2043 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2044 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2046 // Simplify mul instructions with a constant RHS...
2047 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2048 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2050 // ((X << C1)*C2) == (X * (C2 << C1))
2051 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2052 if (SI->getOpcode() == Instruction::Shl)
2053 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2054 return BinaryOperator::createMul(SI->getOperand(0),
2055 ConstantExpr::getShl(CI, ShOp));
2057 if (CI->isNullValue())
2058 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2059 if (CI->equalsInt(1)) // X * 1 == X
2060 return ReplaceInstUsesWith(I, Op0);
2061 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2062 return BinaryOperator::createNeg(Op0, I.getName());
2064 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2065 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2066 uint64_t C = Log2_64(Val);
2067 return new ShiftInst(Instruction::Shl, Op0,
2068 ConstantInt::get(Type::UByteTy, C));
2070 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2071 if (Op1F->isNullValue())
2072 return ReplaceInstUsesWith(I, Op1);
2074 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2075 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2076 if (Op1F->getValue() == 1.0)
2077 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2080 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2081 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2082 isa<ConstantInt>(Op0I->getOperand(1))) {
2083 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2084 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2086 InsertNewInstBefore(Add, I);
2087 Value *C1C2 = ConstantExpr::getMul(Op1,
2088 cast<Constant>(Op0I->getOperand(1)));
2089 return BinaryOperator::createAdd(Add, C1C2);
2093 // Try to fold constant mul into select arguments.
2094 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2095 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2098 if (isa<PHINode>(Op0))
2099 if (Instruction *NV = FoldOpIntoPhi(I))
2103 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2104 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2105 return BinaryOperator::createMul(Op0v, Op1v);
2107 // If one of the operands of the multiply is a cast from a boolean value, then
2108 // we know the bool is either zero or one, so this is a 'masking' multiply.
2109 // See if we can simplify things based on how the boolean was originally
2111 CastInst *BoolCast = 0;
2112 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
2113 if (CI->getOperand(0)->getType() == Type::BoolTy)
2116 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
2117 if (CI->getOperand(0)->getType() == Type::BoolTy)
2120 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
2121 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2122 const Type *SCOpTy = SCIOp0->getType();
2124 // If the setcc is true iff the sign bit of X is set, then convert this
2125 // multiply into a shift/and combination.
2126 if (isa<ConstantInt>(SCIOp1) &&
2127 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
2128 // Shift the X value right to turn it into "all signbits".
2129 Constant *Amt = ConstantInt::get(Type::UByteTy,
2130 SCOpTy->getPrimitiveSizeInBits()-1);
2131 if (SCIOp0->getType()->isUnsigned()) {
2132 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
2133 SCIOp0 = InsertCastBefore(SCIOp0, NewTy, I);
2137 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
2138 BoolCast->getOperand(0)->getName()+
2141 // If the multiply type is not the same as the source type, sign extend
2142 // or truncate to the multiply type.
2143 if (I.getType() != V->getType())
2144 V = InsertCastBefore(V, I.getType(), I);
2146 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2147 return BinaryOperator::createAnd(V, OtherOp);
2152 return Changed ? &I : 0;
2155 /// This function implements the transforms on div instructions that work
2156 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2157 /// used by the visitors to those instructions.
2158 /// @brief Transforms common to all three div instructions
2159 Instruction* InstCombiner::commonDivTransforms(BinaryOperator &I) {
2160 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2163 if (isa<UndefValue>(Op0))
2164 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2166 // X / undef -> undef
2167 if (isa<UndefValue>(Op1))
2168 return ReplaceInstUsesWith(I, Op1);
2170 // Handle cases involving: div X, (select Cond, Y, Z)
2171 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2172 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2173 // same basic block, then we replace the select with Y, and the condition
2174 // of the select with false (if the cond value is in the same BB). If the
2175 // select has uses other than the div, this allows them to be simplified
2176 // also. Note that div X, Y is just as good as div X, 0 (undef)
2177 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2178 if (ST->isNullValue()) {
2179 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2180 if (CondI && CondI->getParent() == I.getParent())
2181 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2182 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2183 I.setOperand(1, SI->getOperand(2));
2185 UpdateValueUsesWith(SI, SI->getOperand(2));
2189 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2190 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2191 if (ST->isNullValue()) {
2192 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2193 if (CondI && CondI->getParent() == I.getParent())
2194 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2195 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2196 I.setOperand(1, SI->getOperand(1));
2198 UpdateValueUsesWith(SI, SI->getOperand(1));
2206 /// This function implements the transforms common to both integer division
2207 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2208 /// division instructions.
2209 /// @brief Common integer divide transforms
2210 Instruction* InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2211 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2213 if (Instruction *Common = commonDivTransforms(I))
2216 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2218 if (RHS->equalsInt(1))
2219 return ReplaceInstUsesWith(I, Op0);
2221 // (X / C1) / C2 -> X / (C1*C2)
2222 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2223 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2224 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2225 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2226 ConstantExpr::getMul(RHS, LHSRHS));
2229 if (!RHS->isNullValue()) { // avoid X udiv 0
2230 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2231 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2233 if (isa<PHINode>(Op0))
2234 if (Instruction *NV = FoldOpIntoPhi(I))
2239 // 0 / X == 0, we don't need to preserve faults!
2240 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2241 if (LHS->equalsInt(0))
2242 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2247 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2248 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2250 // Handle the integer div common cases
2251 if (Instruction *Common = commonIDivTransforms(I))
2254 // X udiv C^2 -> X >> C
2255 // Check to see if this is an unsigned division with an exact power of 2,
2256 // if so, convert to a right shift.
2257 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2258 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2259 if (isPowerOf2_64(Val)) {
2260 uint64_t ShiftAmt = Log2_64(Val);
2262 const Type* XTy = X->getType();
2263 bool isSigned = XTy->isSigned();
2265 X = InsertCastBefore(X, XTy->getUnsignedVersion(), I);
2266 Instruction* Result =
2267 new ShiftInst(Instruction::Shr, X,
2268 ConstantInt::get(Type::UByteTy, ShiftAmt));
2271 InsertNewInstBefore(Result, I);
2272 return new CastInst(Result, XTy->getSignedVersion(), I.getName());
2276 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2277 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2278 if (RHSI->getOpcode() == Instruction::Shl &&
2279 isa<ConstantInt>(RHSI->getOperand(0))) {
2280 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2281 if (isPowerOf2_64(C1)) {
2282 Value *N = RHSI->getOperand(1);
2283 const Type* NTy = N->getType();
2284 bool isSigned = NTy->isSigned();
2285 if (uint64_t C2 = Log2_64(C1)) {
2287 NTy = NTy->getUnsignedVersion();
2288 N = InsertCastBefore(N, NTy, I);
2290 Constant *C2V = ConstantInt::get(NTy, C2);
2291 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2293 Instruction* Result = new ShiftInst(Instruction::Shr, Op0, N);
2296 InsertNewInstBefore(Result, I);
2297 return new CastInst(Result, NTy->getSignedVersion(), I.getName());
2302 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2303 // where C1&C2 are powers of two.
2304 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2305 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2306 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2307 if (!STO->isNullValue() && !STO->isNullValue()) {
2308 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2309 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2310 // Compute the shift amounts
2311 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2312 // Make sure we get the unsigned version of X
2314 const Type* origXTy = X->getType();
2315 bool isSigned = origXTy->isSigned();
2317 X = InsertCastBefore(X, X->getType()->getUnsignedVersion(), I);
2318 // Construct the "on true" case of the select
2319 Constant *TC = ConstantInt::get(Type::UByteTy, TSA);
2321 new ShiftInst(Instruction::Shr, X, TC, SI->getName()+".t");
2322 TSI = InsertNewInstBefore(TSI, I);
2324 // Construct the "on false" case of the select
2325 Constant *FC = ConstantInt::get(Type::UByteTy, FSA);
2327 new ShiftInst(Instruction::Shr, X, FC, SI->getName()+".f");
2328 FSI = InsertNewInstBefore(FSI, I);
2330 // construct the select instruction and return it.
2332 new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2335 InsertNewInstBefore(NewSI, I);
2336 return new CastInst(NewSI, origXTy, NewSI->getName());
2343 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2344 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2346 // Handle the integer div common cases
2347 if (Instruction *Common = commonIDivTransforms(I))
2350 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2352 if (RHS->isAllOnesValue())
2353 return BinaryOperator::createNeg(Op0);
2356 if (Value *LHSNeg = dyn_castNegVal(Op0))
2357 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2360 // If the sign bits of both operands are zero (i.e. we can prove they are
2361 // unsigned inputs), turn this into a udiv.
2362 if (I.getType()->isInteger()) {
2363 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2364 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2365 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2372 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2373 return commonDivTransforms(I);
2376 /// GetFactor - If we can prove that the specified value is at least a multiple
2377 /// of some factor, return that factor.
2378 static Constant *GetFactor(Value *V) {
2379 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2382 // Unless we can be tricky, we know this is a multiple of 1.
2383 Constant *Result = ConstantInt::get(V->getType(), 1);
2385 Instruction *I = dyn_cast<Instruction>(V);
2386 if (!I) return Result;
2388 if (I->getOpcode() == Instruction::Mul) {
2389 // Handle multiplies by a constant, etc.
2390 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2391 GetFactor(I->getOperand(1)));
2392 } else if (I->getOpcode() == Instruction::Shl) {
2393 // (X<<C) -> X * (1 << C)
2394 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2395 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2396 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2398 } else if (I->getOpcode() == Instruction::And) {
2399 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2400 // X & 0xFFF0 is known to be a multiple of 16.
2401 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2402 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2403 return ConstantExpr::getShl(Result,
2404 ConstantInt::get(Type::UByteTy, Zeros));
2406 } else if (I->getOpcode() == Instruction::Cast) {
2407 Value *Op = I->getOperand(0);
2408 // Only handle int->int casts.
2409 if (!Op->getType()->isInteger()) return Result;
2410 return ConstantExpr::getCast(GetFactor(Op), V->getType());
2415 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
2416 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2418 // 0 % X == 0, we don't need to preserve faults!
2419 if (Constant *LHS = dyn_cast<Constant>(Op0))
2420 if (LHS->isNullValue())
2421 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2423 if (isa<UndefValue>(Op0)) // undef % X -> 0
2424 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2425 if (isa<UndefValue>(Op1))
2426 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2428 if (I.getType()->isSigned()) {
2429 if (Value *RHSNeg = dyn_castNegVal(Op1))
2430 if (!isa<ConstantInt>(RHSNeg) || !RHSNeg->getType()->isSigned() ||
2431 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2433 AddUsesToWorkList(I);
2434 I.setOperand(1, RHSNeg);
2438 // If the top bits of both operands are zero (i.e. we can prove they are
2439 // unsigned inputs), turn this into a urem.
2440 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2441 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2442 const Type *NTy = Op0->getType()->getUnsignedVersion();
2443 Value *LHS = InsertCastBefore(Op0, NTy, I);
2445 if (Constant *R = dyn_cast<Constant>(Op1))
2446 RHS = ConstantExpr::getCast(R, NTy);
2448 RHS = InsertCastBefore(Op1, NTy, I);
2449 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
2450 InsertNewInstBefore(Rem, I);
2451 return new CastInst(Rem, I.getType());
2455 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2456 // X % 0 == undef, we don't need to preserve faults!
2457 if (RHS->equalsInt(0))
2458 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2460 if (RHS->equalsInt(1)) // X % 1 == 0
2461 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2463 // Check to see if this is an unsigned remainder with an exact power of 2,
2464 // if so, convert to a bitwise and.
2465 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2466 if (RHS->getType()->isUnsigned())
2467 if (isPowerOf2_64(C->getZExtValue()))
2468 return BinaryOperator::createAnd(Op0, SubOne(C));
2470 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2471 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2472 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2474 } else if (isa<PHINode>(Op0I)) {
2475 if (Instruction *NV = FoldOpIntoPhi(I))
2479 // X*C1%C2 --> 0 iff C1%C2 == 0
2480 if (ConstantExpr::getRem(GetFactor(Op0I), RHS)->isNullValue())
2481 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2485 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2486 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
2487 if (I.getType()->isUnsigned() &&
2488 RHSI->getOpcode() == Instruction::Shl &&
2489 isa<ConstantInt>(RHSI->getOperand(0)) &&
2490 RHSI->getOperand(0)->getType()->isUnsigned()) {
2491 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2492 if (isPowerOf2_64(C1)) {
2493 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2494 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2496 return BinaryOperator::createAnd(Op0, Add);
2500 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
2501 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
2502 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2503 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2504 // the same basic block, then we replace the select with Y, and the
2505 // condition of the select with false (if the cond value is in the same
2506 // BB). If the select has uses other than the div, this allows them to be
2508 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2509 if (ST->isNullValue()) {
2510 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2511 if (CondI && CondI->getParent() == I.getParent())
2512 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2513 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2514 I.setOperand(1, SI->getOperand(2));
2516 UpdateValueUsesWith(SI, SI->getOperand(2));
2519 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2520 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2521 if (ST->isNullValue()) {
2522 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2523 if (CondI && CondI->getParent() == I.getParent())
2524 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2525 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2526 I.setOperand(1, SI->getOperand(1));
2528 UpdateValueUsesWith(SI, SI->getOperand(1));
2533 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2534 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2535 if (STO->getType()->isUnsigned() && SFO->getType()->isUnsigned()) {
2536 // STO == 0 and SFO == 0 handled above.
2537 if (isPowerOf2_64(STO->getZExtValue()) &&
2538 isPowerOf2_64(SFO->getZExtValue())) {
2539 Value *TrueAnd = InsertNewInstBefore(
2540 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"),
2542 Value *FalseAnd = InsertNewInstBefore(
2543 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"),
2545 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2554 // isMaxValueMinusOne - return true if this is Max-1
2555 static bool isMaxValueMinusOne(const ConstantInt *C) {
2556 if (C->getType()->isUnsigned())
2557 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2559 // Calculate 0111111111..11111
2560 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2561 int64_t Val = INT64_MAX; // All ones
2562 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2563 return C->getSExtValue() == Val-1;
2566 // isMinValuePlusOne - return true if this is Min+1
2567 static bool isMinValuePlusOne(const ConstantInt *C) {
2568 if (C->getType()->isUnsigned())
2569 return C->getZExtValue() == 1;
2571 // Calculate 1111111111000000000000
2572 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2573 int64_t Val = -1; // All ones
2574 Val <<= TypeBits-1; // Shift over to the right spot
2575 return C->getSExtValue() == Val+1;
2578 // isOneBitSet - Return true if there is exactly one bit set in the specified
2580 static bool isOneBitSet(const ConstantInt *CI) {
2581 uint64_t V = CI->getZExtValue();
2582 return V && (V & (V-1)) == 0;
2585 #if 0 // Currently unused
2586 // isLowOnes - Return true if the constant is of the form 0+1+.
2587 static bool isLowOnes(const ConstantInt *CI) {
2588 uint64_t V = CI->getZExtValue();
2590 // There won't be bits set in parts that the type doesn't contain.
2591 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2593 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2594 return U && V && (U & V) == 0;
2598 // isHighOnes - Return true if the constant is of the form 1+0+.
2599 // This is the same as lowones(~X).
2600 static bool isHighOnes(const ConstantInt *CI) {
2601 uint64_t V = ~CI->getZExtValue();
2602 if (~V == 0) return false; // 0's does not match "1+"
2604 // There won't be bits set in parts that the type doesn't contain.
2605 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2607 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2608 return U && V && (U & V) == 0;
2612 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2613 /// are carefully arranged to allow folding of expressions such as:
2615 /// (A < B) | (A > B) --> (A != B)
2617 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2618 /// represents that the comparison is true if A == B, and bit value '1' is true
2621 static unsigned getSetCondCode(const SetCondInst *SCI) {
2622 switch (SCI->getOpcode()) {
2624 case Instruction::SetGT: return 1;
2625 case Instruction::SetEQ: return 2;
2626 case Instruction::SetGE: return 3;
2627 case Instruction::SetLT: return 4;
2628 case Instruction::SetNE: return 5;
2629 case Instruction::SetLE: return 6;
2632 assert(0 && "Invalid SetCC opcode!");
2637 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2638 /// opcode and two operands into either a constant true or false, or a brand new
2639 /// SetCC instruction.
2640 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2642 case 0: return ConstantBool::getFalse();
2643 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2644 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2645 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2646 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2647 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2648 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2649 case 7: return ConstantBool::getTrue();
2650 default: assert(0 && "Illegal SetCCCode!"); return 0;
2654 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2655 struct FoldSetCCLogical {
2658 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2659 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2660 bool shouldApply(Value *V) const {
2661 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2662 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2663 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2666 Instruction *apply(BinaryOperator &Log) const {
2667 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2668 if (SCI->getOperand(0) != LHS) {
2669 assert(SCI->getOperand(1) == LHS);
2670 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2673 unsigned LHSCode = getSetCondCode(SCI);
2674 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2676 switch (Log.getOpcode()) {
2677 case Instruction::And: Code = LHSCode & RHSCode; break;
2678 case Instruction::Or: Code = LHSCode | RHSCode; break;
2679 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2680 default: assert(0 && "Illegal logical opcode!"); return 0;
2683 Value *RV = getSetCCValue(Code, LHS, RHS);
2684 if (Instruction *I = dyn_cast<Instruction>(RV))
2686 // Otherwise, it's a constant boolean value...
2687 return IC.ReplaceInstUsesWith(Log, RV);
2691 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2692 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2693 // guaranteed to be either a shift instruction or a binary operator.
2694 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2695 ConstantIntegral *OpRHS,
2696 ConstantIntegral *AndRHS,
2697 BinaryOperator &TheAnd) {
2698 Value *X = Op->getOperand(0);
2699 Constant *Together = 0;
2700 if (!isa<ShiftInst>(Op))
2701 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2703 switch (Op->getOpcode()) {
2704 case Instruction::Xor:
2705 if (Op->hasOneUse()) {
2706 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2707 std::string OpName = Op->getName(); Op->setName("");
2708 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2709 InsertNewInstBefore(And, TheAnd);
2710 return BinaryOperator::createXor(And, Together);
2713 case Instruction::Or:
2714 if (Together == AndRHS) // (X | C) & C --> C
2715 return ReplaceInstUsesWith(TheAnd, AndRHS);
2717 if (Op->hasOneUse() && Together != OpRHS) {
2718 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2719 std::string Op0Name = Op->getName(); Op->setName("");
2720 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2721 InsertNewInstBefore(Or, TheAnd);
2722 return BinaryOperator::createAnd(Or, AndRHS);
2725 case Instruction::Add:
2726 if (Op->hasOneUse()) {
2727 // Adding a one to a single bit bit-field should be turned into an XOR
2728 // of the bit. First thing to check is to see if this AND is with a
2729 // single bit constant.
2730 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2732 // Clear bits that are not part of the constant.
2733 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2735 // If there is only one bit set...
2736 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2737 // Ok, at this point, we know that we are masking the result of the
2738 // ADD down to exactly one bit. If the constant we are adding has
2739 // no bits set below this bit, then we can eliminate the ADD.
2740 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2742 // Check to see if any bits below the one bit set in AndRHSV are set.
2743 if ((AddRHS & (AndRHSV-1)) == 0) {
2744 // If not, the only thing that can effect the output of the AND is
2745 // the bit specified by AndRHSV. If that bit is set, the effect of
2746 // the XOR is to toggle the bit. If it is clear, then the ADD has
2748 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2749 TheAnd.setOperand(0, X);
2752 std::string Name = Op->getName(); Op->setName("");
2753 // Pull the XOR out of the AND.
2754 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2755 InsertNewInstBefore(NewAnd, TheAnd);
2756 return BinaryOperator::createXor(NewAnd, AndRHS);
2763 case Instruction::Shl: {
2764 // We know that the AND will not produce any of the bits shifted in, so if
2765 // the anded constant includes them, clear them now!
2767 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2768 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2769 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2771 if (CI == ShlMask) { // Masking out bits that the shift already masks
2772 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2773 } else if (CI != AndRHS) { // Reducing bits set in and.
2774 TheAnd.setOperand(1, CI);
2779 case Instruction::Shr:
2780 // We know that the AND will not produce any of the bits shifted in, so if
2781 // the anded constant includes them, clear them now! This only applies to
2782 // unsigned shifts, because a signed shr may bring in set bits!
2784 if (AndRHS->getType()->isUnsigned()) {
2785 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2786 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
2787 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2789 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2790 return ReplaceInstUsesWith(TheAnd, Op);
2791 } else if (CI != AndRHS) {
2792 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2795 } else { // Signed shr.
2796 // See if this is shifting in some sign extension, then masking it out
2798 if (Op->hasOneUse()) {
2799 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2800 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
2801 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2802 if (CI == AndRHS) { // Masking out bits shifted in.
2803 // Make the argument unsigned.
2804 Value *ShVal = Op->getOperand(0);
2805 ShVal = InsertCastBefore(ShVal,
2806 ShVal->getType()->getUnsignedVersion(),
2808 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
2809 OpRHS, Op->getName()),
2811 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
2812 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
2815 return new CastInst(ShVal, Op->getType());
2825 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2826 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2827 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2828 /// insert new instructions.
2829 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2830 bool Inside, Instruction &IB) {
2831 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2832 "Lo is not <= Hi in range emission code!");
2834 if (Lo == Hi) // Trivially false.
2835 return new SetCondInst(Instruction::SetNE, V, V);
2836 if (cast<ConstantIntegral>(Lo)->isMinValue())
2837 return new SetCondInst(Instruction::SetLT, V, Hi);
2839 Constant *AddCST = ConstantExpr::getNeg(Lo);
2840 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2841 InsertNewInstBefore(Add, IB);
2842 // Convert to unsigned for the comparison.
2843 const Type *UnsType = Add->getType()->getUnsignedVersion();
2844 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2845 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2846 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2847 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2850 if (Lo == Hi) // Trivially true.
2851 return new SetCondInst(Instruction::SetEQ, V, V);
2853 Hi = SubOne(cast<ConstantInt>(Hi));
2855 // V < 0 || V >= Hi ->'V > Hi-1'
2856 if (cast<ConstantIntegral>(Lo)->isMinValue())
2857 return new SetCondInst(Instruction::SetGT, V, Hi);
2859 // Emit X-Lo > Hi-Lo-1
2860 Constant *AddCST = ConstantExpr::getNeg(Lo);
2861 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2862 InsertNewInstBefore(Add, IB);
2863 // Convert to unsigned for the comparison.
2864 const Type *UnsType = Add->getType()->getUnsignedVersion();
2865 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2866 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2867 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2868 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2871 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2872 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2873 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2874 // not, since all 1s are not contiguous.
2875 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2876 uint64_t V = Val->getZExtValue();
2877 if (!isShiftedMask_64(V)) return false;
2879 // look for the first zero bit after the run of ones
2880 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2881 // look for the first non-zero bit
2882 ME = 64-CountLeadingZeros_64(V);
2888 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2889 /// where isSub determines whether the operator is a sub. If we can fold one of
2890 /// the following xforms:
2892 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2893 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2894 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2896 /// return (A +/- B).
2898 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2899 ConstantIntegral *Mask, bool isSub,
2901 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2902 if (!LHSI || LHSI->getNumOperands() != 2 ||
2903 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2905 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2907 switch (LHSI->getOpcode()) {
2909 case Instruction::And:
2910 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2911 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2912 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
2915 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2916 // part, we don't need any explicit masks to take them out of A. If that
2917 // is all N is, ignore it.
2919 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2920 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2922 if (MaskedValueIsZero(RHS, Mask))
2927 case Instruction::Or:
2928 case Instruction::Xor:
2929 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2930 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
2931 ConstantExpr::getAnd(N, Mask)->isNullValue())
2938 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2940 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2941 return InsertNewInstBefore(New, I);
2944 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
2945 bool Changed = SimplifyCommutative(I);
2946 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2948 if (isa<UndefValue>(Op1)) // X & undef -> 0
2949 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2953 return ReplaceInstUsesWith(I, Op1);
2955 // See if we can simplify any instructions used by the instruction whose sole
2956 // purpose is to compute bits we don't care about.
2957 uint64_t KnownZero, KnownOne;
2958 if (!isa<PackedType>(I.getType()) &&
2959 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2960 KnownZero, KnownOne))
2963 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
2964 uint64_t AndRHSMask = AndRHS->getZExtValue();
2965 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
2966 uint64_t NotAndRHS = AndRHSMask^TypeMask;
2968 // Optimize a variety of ((val OP C1) & C2) combinations...
2969 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
2970 Instruction *Op0I = cast<Instruction>(Op0);
2971 Value *Op0LHS = Op0I->getOperand(0);
2972 Value *Op0RHS = Op0I->getOperand(1);
2973 switch (Op0I->getOpcode()) {
2974 case Instruction::Xor:
2975 case Instruction::Or:
2976 // If the mask is only needed on one incoming arm, push it up.
2977 if (Op0I->hasOneUse()) {
2978 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
2979 // Not masking anything out for the LHS, move to RHS.
2980 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
2981 Op0RHS->getName()+".masked");
2982 InsertNewInstBefore(NewRHS, I);
2983 return BinaryOperator::create(
2984 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
2986 if (!isa<Constant>(Op0RHS) &&
2987 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
2988 // Not masking anything out for the RHS, move to LHS.
2989 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
2990 Op0LHS->getName()+".masked");
2991 InsertNewInstBefore(NewLHS, I);
2992 return BinaryOperator::create(
2993 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
2998 case Instruction::Add:
2999 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3000 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3001 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3002 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3003 return BinaryOperator::createAnd(V, AndRHS);
3004 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3005 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3008 case Instruction::Sub:
3009 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3010 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3011 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3012 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3013 return BinaryOperator::createAnd(V, AndRHS);
3017 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3018 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3020 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3021 const Type *SrcTy = CI->getOperand(0)->getType();
3023 // If this is an integer truncation or change from signed-to-unsigned, and
3024 // if the source is an and/or with immediate, transform it. This
3025 // frequently occurs for bitfield accesses.
3026 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3027 if (SrcTy->getPrimitiveSizeInBits() >=
3028 I.getType()->getPrimitiveSizeInBits() &&
3029 CastOp->getNumOperands() == 2)
3030 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3031 if (CastOp->getOpcode() == Instruction::And) {
3032 // Change: and (cast (and X, C1) to T), C2
3033 // into : and (cast X to T), trunc(C1)&C2
3034 // This will folds the two ands together, which may allow other
3036 Instruction *NewCast =
3037 new CastInst(CastOp->getOperand(0), I.getType(),
3038 CastOp->getName()+".shrunk");
3039 NewCast = InsertNewInstBefore(NewCast, I);
3041 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
3042 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
3043 return BinaryOperator::createAnd(NewCast, C3);
3044 } else if (CastOp->getOpcode() == Instruction::Or) {
3045 // Change: and (cast (or X, C1) to T), C2
3046 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3047 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
3048 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3049 return ReplaceInstUsesWith(I, AndRHS);
3054 // Try to fold constant and into select arguments.
3055 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3056 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3058 if (isa<PHINode>(Op0))
3059 if (Instruction *NV = FoldOpIntoPhi(I))
3063 Value *Op0NotVal = dyn_castNotVal(Op0);
3064 Value *Op1NotVal = dyn_castNotVal(Op1);
3066 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3067 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3069 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3070 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3071 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3072 I.getName()+".demorgan");
3073 InsertNewInstBefore(Or, I);
3074 return BinaryOperator::createNot(Or);
3078 Value *A = 0, *B = 0;
3079 ConstantInt *C1 = 0, *C2 = 0;
3080 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3081 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3082 return ReplaceInstUsesWith(I, Op1);
3083 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3084 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3085 return ReplaceInstUsesWith(I, Op0);
3087 if (Op0->hasOneUse() &&
3088 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3089 if (A == Op1) { // (A^B)&A -> A&(A^B)
3090 I.swapOperands(); // Simplify below
3091 std::swap(Op0, Op1);
3092 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3093 cast<BinaryOperator>(Op0)->swapOperands();
3094 I.swapOperands(); // Simplify below
3095 std::swap(Op0, Op1);
3098 if (Op1->hasOneUse() &&
3099 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3100 if (B == Op0) { // B&(A^B) -> B&(B^A)
3101 cast<BinaryOperator>(Op1)->swapOperands();
3104 if (A == Op0) { // A&(A^B) -> A & ~B
3105 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3106 InsertNewInstBefore(NotB, I);
3107 return BinaryOperator::createAnd(A, NotB);
3113 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
3114 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
3115 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3118 Value *LHSVal, *RHSVal;
3119 ConstantInt *LHSCst, *RHSCst;
3120 Instruction::BinaryOps LHSCC, RHSCC;
3121 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3122 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3123 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
3124 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3125 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3126 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3127 // Ensure that the larger constant is on the RHS.
3128 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3129 SetCondInst *LHS = cast<SetCondInst>(Op0);
3130 if (cast<ConstantBool>(Cmp)->getValue()) {
3131 std::swap(LHS, RHS);
3132 std::swap(LHSCst, RHSCst);
3133 std::swap(LHSCC, RHSCC);
3136 // At this point, we know we have have two setcc instructions
3137 // comparing a value against two constants and and'ing the result
3138 // together. Because of the above check, we know that we only have
3139 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3140 // FoldSetCCLogical check above), that the two constants are not
3142 assert(LHSCst != RHSCst && "Compares not folded above?");
3145 default: assert(0 && "Unknown integer condition code!");
3146 case Instruction::SetEQ:
3148 default: assert(0 && "Unknown integer condition code!");
3149 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
3150 case Instruction::SetGT: // (X == 13 & X > 15) -> false
3151 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3152 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
3153 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
3154 return ReplaceInstUsesWith(I, LHS);
3156 case Instruction::SetNE:
3158 default: assert(0 && "Unknown integer condition code!");
3159 case Instruction::SetLT:
3160 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
3161 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
3162 break; // (X != 13 & X < 15) -> no change
3163 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
3164 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
3165 return ReplaceInstUsesWith(I, RHS);
3166 case Instruction::SetNE:
3167 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
3168 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3169 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3170 LHSVal->getName()+".off");
3171 InsertNewInstBefore(Add, I);
3172 const Type *UnsType = Add->getType()->getUnsignedVersion();
3173 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3174 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
3175 AddCST = ConstantExpr::getCast(AddCST, UnsType);
3176 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
3178 break; // (X != 13 & X != 15) -> no change
3181 case Instruction::SetLT:
3183 default: assert(0 && "Unknown integer condition code!");
3184 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
3185 case Instruction::SetGT: // (X < 13 & X > 15) -> false
3186 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3187 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
3188 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
3189 return ReplaceInstUsesWith(I, LHS);
3191 case Instruction::SetGT:
3193 default: assert(0 && "Unknown integer condition code!");
3194 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
3195 return ReplaceInstUsesWith(I, LHS);
3196 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
3197 return ReplaceInstUsesWith(I, RHS);
3198 case Instruction::SetNE:
3199 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
3200 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
3201 break; // (X > 13 & X != 15) -> no change
3202 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
3203 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
3209 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3210 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3211 const Type *SrcTy = Op0C->getOperand(0)->getType();
3212 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3213 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3214 // Only do this if the casts both really cause code to be generated.
3215 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3216 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3217 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3218 Op1C->getOperand(0),
3220 InsertNewInstBefore(NewOp, I);
3221 return new CastInst(NewOp, I.getType());
3225 return Changed ? &I : 0;
3228 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3229 /// in the result. If it does, and if the specified byte hasn't been filled in
3230 /// yet, fill it in and return false.
3231 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3232 Instruction *I = dyn_cast<Instruction>(V);
3233 if (I == 0) return true;
3235 // If this is an or instruction, it is an inner node of the bswap.
3236 if (I->getOpcode() == Instruction::Or)
3237 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3238 CollectBSwapParts(I->getOperand(1), ByteValues);
3240 // If this is a shift by a constant int, and it is "24", then its operand
3241 // defines a byte. We only handle unsigned types here.
3242 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3243 // Not shifting the entire input by N-1 bytes?
3244 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3245 8*(ByteValues.size()-1))
3249 if (I->getOpcode() == Instruction::Shl) {
3250 // X << 24 defines the top byte with the lowest of the input bytes.
3251 DestNo = ByteValues.size()-1;
3253 // X >>u 24 defines the low byte with the highest of the input bytes.
3257 // If the destination byte value is already defined, the values are or'd
3258 // together, which isn't a bswap (unless it's an or of the same bits).
3259 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3261 ByteValues[DestNo] = I->getOperand(0);
3265 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3267 Value *Shift = 0, *ShiftLHS = 0;
3268 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3269 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3270 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3272 Instruction *SI = cast<Instruction>(Shift);
3274 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3275 if (ShiftAmt->getZExtValue() & 7 ||
3276 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3279 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3281 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3282 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3284 // Unknown mask for bswap.
3285 if (DestByte == ByteValues.size()) return true;
3287 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3289 if (SI->getOpcode() == Instruction::Shl)
3290 SrcByte = DestByte - ShiftBytes;
3292 SrcByte = DestByte + ShiftBytes;
3294 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3295 if (SrcByte != ByteValues.size()-DestByte-1)
3298 // If the destination byte value is already defined, the values are or'd
3299 // together, which isn't a bswap (unless it's an or of the same bits).
3300 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3302 ByteValues[DestByte] = SI->getOperand(0);
3306 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3307 /// If so, insert the new bswap intrinsic and return it.
3308 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3309 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3310 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3313 /// ByteValues - For each byte of the result, we keep track of which value
3314 /// defines each byte.
3315 std::vector<Value*> ByteValues;
3316 ByteValues.resize(I.getType()->getPrimitiveSize());
3318 // Try to find all the pieces corresponding to the bswap.
3319 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3320 CollectBSwapParts(I.getOperand(1), ByteValues))
3323 // Check to see if all of the bytes come from the same value.
3324 Value *V = ByteValues[0];
3325 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3327 // Check to make sure that all of the bytes come from the same value.
3328 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3329 if (ByteValues[i] != V)
3332 // If they do then *success* we can turn this into a bswap. Figure out what
3333 // bswap to make it into.
3334 Module *M = I.getParent()->getParent()->getParent();
3335 const char *FnName = 0;
3336 if (I.getType() == Type::UShortTy)
3337 FnName = "llvm.bswap.i16";
3338 else if (I.getType() == Type::UIntTy)
3339 FnName = "llvm.bswap.i32";
3340 else if (I.getType() == Type::ULongTy)
3341 FnName = "llvm.bswap.i64";
3343 assert(0 && "Unknown integer type!");
3344 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3346 return new CallInst(F, V);
3350 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3351 bool Changed = SimplifyCommutative(I);
3352 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3354 if (isa<UndefValue>(Op1))
3355 return ReplaceInstUsesWith(I, // X | undef -> -1
3356 ConstantIntegral::getAllOnesValue(I.getType()));
3360 return ReplaceInstUsesWith(I, Op0);
3362 // See if we can simplify any instructions used by the instruction whose sole
3363 // purpose is to compute bits we don't care about.
3364 uint64_t KnownZero, KnownOne;
3365 if (!isa<PackedType>(I.getType()) &&
3366 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3367 KnownZero, KnownOne))
3371 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3372 ConstantInt *C1 = 0; Value *X = 0;
3373 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3374 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3375 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3377 InsertNewInstBefore(Or, I);
3378 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3381 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3382 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3383 std::string Op0Name = Op0->getName(); Op0->setName("");
3384 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3385 InsertNewInstBefore(Or, I);
3386 return BinaryOperator::createXor(Or,
3387 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3390 // Try to fold constant and into select arguments.
3391 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3392 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3394 if (isa<PHINode>(Op0))
3395 if (Instruction *NV = FoldOpIntoPhi(I))
3399 Value *A = 0, *B = 0;
3400 ConstantInt *C1 = 0, *C2 = 0;
3402 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3403 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3404 return ReplaceInstUsesWith(I, Op1);
3405 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3406 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3407 return ReplaceInstUsesWith(I, Op0);
3409 // (A | B) | C and A | (B | C) -> bswap if possible.
3410 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3411 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3412 match(Op1, m_Or(m_Value(), m_Value())) ||
3413 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3414 match(Op1, m_Shift(m_Value(), m_Value())))) {
3415 if (Instruction *BSwap = MatchBSwap(I))
3419 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3420 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3421 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3422 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3424 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3427 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3428 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3429 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3430 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3432 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3435 // (A & C1)|(B & C2)
3436 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3437 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3439 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3440 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3443 // If we have: ((V + N) & C1) | (V & C2)
3444 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3445 // replace with V+N.
3446 if (C1 == ConstantExpr::getNot(C2)) {
3447 Value *V1 = 0, *V2 = 0;
3448 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3449 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3450 // Add commutes, try both ways.
3451 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3452 return ReplaceInstUsesWith(I, A);
3453 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3454 return ReplaceInstUsesWith(I, A);
3456 // Or commutes, try both ways.
3457 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3458 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3459 // Add commutes, try both ways.
3460 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3461 return ReplaceInstUsesWith(I, B);
3462 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3463 return ReplaceInstUsesWith(I, B);
3468 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3469 if (A == Op1) // ~A | A == -1
3470 return ReplaceInstUsesWith(I,
3471 ConstantIntegral::getAllOnesValue(I.getType()));
3475 // Note, A is still live here!
3476 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3478 return ReplaceInstUsesWith(I,
3479 ConstantIntegral::getAllOnesValue(I.getType()));
3481 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3482 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3483 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3484 I.getName()+".demorgan"), I);
3485 return BinaryOperator::createNot(And);
3489 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
3490 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
3491 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3494 Value *LHSVal, *RHSVal;
3495 ConstantInt *LHSCst, *RHSCst;
3496 Instruction::BinaryOps LHSCC, RHSCC;
3497 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3498 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3499 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
3500 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3501 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3502 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3503 // Ensure that the larger constant is on the RHS.
3504 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3505 SetCondInst *LHS = cast<SetCondInst>(Op0);
3506 if (cast<ConstantBool>(Cmp)->getValue()) {
3507 std::swap(LHS, RHS);
3508 std::swap(LHSCst, RHSCst);
3509 std::swap(LHSCC, RHSCC);
3512 // At this point, we know we have have two setcc instructions
3513 // comparing a value against two constants and or'ing the result
3514 // together. Because of the above check, we know that we only have
3515 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3516 // FoldSetCCLogical check above), that the two constants are not
3518 assert(LHSCst != RHSCst && "Compares not folded above?");
3521 default: assert(0 && "Unknown integer condition code!");
3522 case Instruction::SetEQ:
3524 default: assert(0 && "Unknown integer condition code!");
3525 case Instruction::SetEQ:
3526 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3527 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3528 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3529 LHSVal->getName()+".off");
3530 InsertNewInstBefore(Add, I);
3531 const Type *UnsType = Add->getType()->getUnsignedVersion();
3532 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
3533 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3534 AddCST = ConstantExpr::getCast(AddCST, UnsType);
3535 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
3537 break; // (X == 13 | X == 15) -> no change
3539 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
3541 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
3542 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
3543 return ReplaceInstUsesWith(I, RHS);
3546 case Instruction::SetNE:
3548 default: assert(0 && "Unknown integer condition code!");
3549 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
3550 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
3551 return ReplaceInstUsesWith(I, LHS);
3552 case Instruction::SetNE: // (X != 13 | X != 15) -> true
3553 case Instruction::SetLT: // (X != 13 | X < 15) -> true
3554 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3557 case Instruction::SetLT:
3559 default: assert(0 && "Unknown integer condition code!");
3560 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
3562 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
3563 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
3564 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
3565 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
3566 return ReplaceInstUsesWith(I, RHS);
3569 case Instruction::SetGT:
3571 default: assert(0 && "Unknown integer condition code!");
3572 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
3573 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
3574 return ReplaceInstUsesWith(I, LHS);
3575 case Instruction::SetNE: // (X > 13 | X != 15) -> true
3576 case Instruction::SetLT: // (X > 13 | X < 15) -> true
3577 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3583 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3584 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3585 const Type *SrcTy = Op0C->getOperand(0)->getType();
3586 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3587 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3588 // Only do this if the casts both really cause code to be generated.
3589 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3590 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3591 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3592 Op1C->getOperand(0),
3594 InsertNewInstBefore(NewOp, I);
3595 return new CastInst(NewOp, I.getType());
3600 return Changed ? &I : 0;
3603 // XorSelf - Implements: X ^ X --> 0
3606 XorSelf(Value *rhs) : RHS(rhs) {}
3607 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3608 Instruction *apply(BinaryOperator &Xor) const {
3614 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3615 bool Changed = SimplifyCommutative(I);
3616 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3618 if (isa<UndefValue>(Op1))
3619 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3621 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3622 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3623 assert(Result == &I && "AssociativeOpt didn't work?");
3624 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3627 // See if we can simplify any instructions used by the instruction whose sole
3628 // purpose is to compute bits we don't care about.
3629 uint64_t KnownZero, KnownOne;
3630 if (!isa<PackedType>(I.getType()) &&
3631 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3632 KnownZero, KnownOne))
3635 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3636 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3637 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
3638 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
3639 if (RHS == ConstantBool::getTrue() && SCI->hasOneUse())
3640 return new SetCondInst(SCI->getInverseCondition(),
3641 SCI->getOperand(0), SCI->getOperand(1));
3643 // ~(c-X) == X-c-1 == X+(-c-1)
3644 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3645 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3646 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3647 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3648 ConstantInt::get(I.getType(), 1));
3649 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3652 // ~(~X & Y) --> (X | ~Y)
3653 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3654 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3655 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3657 BinaryOperator::createNot(Op0I->getOperand(1),
3658 Op0I->getOperand(1)->getName()+".not");
3659 InsertNewInstBefore(NotY, I);
3660 return BinaryOperator::createOr(Op0NotVal, NotY);
3664 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3665 if (Op0I->getOpcode() == Instruction::Add) {
3666 // ~(X-c) --> (-c-1)-X
3667 if (RHS->isAllOnesValue()) {
3668 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3669 return BinaryOperator::createSub(
3670 ConstantExpr::getSub(NegOp0CI,
3671 ConstantInt::get(I.getType(), 1)),
3672 Op0I->getOperand(0));
3674 } else if (Op0I->getOpcode() == Instruction::Or) {
3675 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3676 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3677 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3678 // Anything in both C1 and C2 is known to be zero, remove it from
3680 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3681 NewRHS = ConstantExpr::getAnd(NewRHS,
3682 ConstantExpr::getNot(CommonBits));
3683 WorkList.push_back(Op0I);
3684 I.setOperand(0, Op0I->getOperand(0));
3685 I.setOperand(1, NewRHS);
3691 // Try to fold constant and into select arguments.
3692 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3693 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3695 if (isa<PHINode>(Op0))
3696 if (Instruction *NV = FoldOpIntoPhi(I))
3700 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3702 return ReplaceInstUsesWith(I,
3703 ConstantIntegral::getAllOnesValue(I.getType()));
3705 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3707 return ReplaceInstUsesWith(I,
3708 ConstantIntegral::getAllOnesValue(I.getType()));
3710 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3711 if (Op1I->getOpcode() == Instruction::Or) {
3712 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3713 Op1I->swapOperands();
3715 std::swap(Op0, Op1);
3716 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3717 I.swapOperands(); // Simplified below.
3718 std::swap(Op0, Op1);
3720 } else if (Op1I->getOpcode() == Instruction::Xor) {
3721 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3722 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3723 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3724 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3725 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3726 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3727 Op1I->swapOperands();
3728 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3729 I.swapOperands(); // Simplified below.
3730 std::swap(Op0, Op1);
3734 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3735 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3736 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3737 Op0I->swapOperands();
3738 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3739 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3740 InsertNewInstBefore(NotB, I);
3741 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3743 } else if (Op0I->getOpcode() == Instruction::Xor) {
3744 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3745 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3746 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3747 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3748 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3749 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3750 Op0I->swapOperands();
3751 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3752 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3753 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3754 InsertNewInstBefore(N, I);
3755 return BinaryOperator::createAnd(N, Op1);
3759 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
3760 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
3761 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3764 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3765 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3766 const Type *SrcTy = Op0C->getOperand(0)->getType();
3767 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3768 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3769 // Only do this if the casts both really cause code to be generated.
3770 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3771 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3772 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
3773 Op1C->getOperand(0),
3775 InsertNewInstBefore(NewOp, I);
3776 return new CastInst(NewOp, I.getType());
3780 return Changed ? &I : 0;
3783 static bool isPositive(ConstantInt *C) {
3784 return C->getSExtValue() >= 0;
3787 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
3788 /// overflowed for this type.
3789 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3791 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
3793 if (In1->getType()->isUnsigned())
3794 return cast<ConstantInt>(Result)->getZExtValue() <
3795 cast<ConstantInt>(In1)->getZExtValue();
3796 if (isPositive(In1) != isPositive(In2))
3798 if (isPositive(In1))
3799 return cast<ConstantInt>(Result)->getSExtValue() <
3800 cast<ConstantInt>(In1)->getSExtValue();
3801 return cast<ConstantInt>(Result)->getSExtValue() >
3802 cast<ConstantInt>(In1)->getSExtValue();
3805 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3806 /// code necessary to compute the offset from the base pointer (without adding
3807 /// in the base pointer). Return the result as a signed integer of intptr size.
3808 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3809 TargetData &TD = IC.getTargetData();
3810 gep_type_iterator GTI = gep_type_begin(GEP);
3811 const Type *UIntPtrTy = TD.getIntPtrType();
3812 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3813 Value *Result = Constant::getNullValue(SIntPtrTy);
3815 // Build a mask for high order bits.
3816 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3818 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3819 Value *Op = GEP->getOperand(i);
3820 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3821 Constant *Scale = ConstantExpr::getCast(ConstantInt::get(UIntPtrTy, Size),
3823 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3824 if (!OpC->isNullValue()) {
3825 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
3826 Scale = ConstantExpr::getMul(OpC, Scale);
3827 if (Constant *RC = dyn_cast<Constant>(Result))
3828 Result = ConstantExpr::getAdd(RC, Scale);
3830 // Emit an add instruction.
3831 Result = IC.InsertNewInstBefore(
3832 BinaryOperator::createAdd(Result, Scale,
3833 GEP->getName()+".offs"), I);
3837 // Convert to correct type.
3838 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
3839 Op->getName()+".c"), I);
3841 // We'll let instcombine(mul) convert this to a shl if possible.
3842 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3843 GEP->getName()+".idx"), I);
3845 // Emit an add instruction.
3846 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3847 GEP->getName()+".offs"), I);
3853 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3854 /// else. At this point we know that the GEP is on the LHS of the comparison.
3855 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3856 Instruction::BinaryOps Cond,
3858 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3860 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3861 if (isa<PointerType>(CI->getOperand(0)->getType()))
3862 RHS = CI->getOperand(0);
3864 Value *PtrBase = GEPLHS->getOperand(0);
3865 if (PtrBase == RHS) {
3866 // As an optimization, we don't actually have to compute the actual value of
3867 // OFFSET if this is a seteq or setne comparison, just return whether each
3868 // index is zero or not.
3869 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3870 Instruction *InVal = 0;
3871 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3872 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3874 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3875 if (isa<UndefValue>(C)) // undef index -> undef.
3876 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3877 if (C->isNullValue())
3879 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3880 EmitIt = false; // This is indexing into a zero sized array?
3881 } else if (isa<ConstantInt>(C))
3882 return ReplaceInstUsesWith(I, // No comparison is needed here.
3883 ConstantBool::get(Cond == Instruction::SetNE));
3888 new SetCondInst(Cond, GEPLHS->getOperand(i),
3889 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3893 InVal = InsertNewInstBefore(InVal, I);
3894 InsertNewInstBefore(Comp, I);
3895 if (Cond == Instruction::SetNE) // True if any are unequal
3896 InVal = BinaryOperator::createOr(InVal, Comp);
3897 else // True if all are equal
3898 InVal = BinaryOperator::createAnd(InVal, Comp);
3906 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
3907 ConstantBool::get(Cond == Instruction::SetEQ));
3910 // Only lower this if the setcc is the only user of the GEP or if we expect
3911 // the result to fold to a constant!
3912 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
3913 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
3914 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
3915 return new SetCondInst(Cond, Offset,
3916 Constant::getNullValue(Offset->getType()));
3918 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
3919 // If the base pointers are different, but the indices are the same, just
3920 // compare the base pointer.
3921 if (PtrBase != GEPRHS->getOperand(0)) {
3922 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
3923 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
3924 GEPRHS->getOperand(0)->getType();
3926 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3927 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3928 IndicesTheSame = false;
3932 // If all indices are the same, just compare the base pointers.
3934 return new SetCondInst(Cond, GEPLHS->getOperand(0),
3935 GEPRHS->getOperand(0));
3937 // Otherwise, the base pointers are different and the indices are
3938 // different, bail out.
3942 // If one of the GEPs has all zero indices, recurse.
3943 bool AllZeros = true;
3944 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3945 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
3946 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
3951 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
3952 SetCondInst::getSwappedCondition(Cond), I);
3954 // If the other GEP has all zero indices, recurse.
3956 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3957 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
3958 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
3963 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
3965 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
3966 // If the GEPs only differ by one index, compare it.
3967 unsigned NumDifferences = 0; // Keep track of # differences.
3968 unsigned DiffOperand = 0; // The operand that differs.
3969 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3970 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3971 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
3972 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
3973 // Irreconcilable differences.
3977 if (NumDifferences++) break;
3982 if (NumDifferences == 0) // SAME GEP?
3983 return ReplaceInstUsesWith(I, // No comparison is needed here.
3984 ConstantBool::get(Cond == Instruction::SetEQ));
3985 else if (NumDifferences == 1) {
3986 Value *LHSV = GEPLHS->getOperand(DiffOperand);
3987 Value *RHSV = GEPRHS->getOperand(DiffOperand);
3989 // Convert the operands to signed values to make sure to perform a
3990 // signed comparison.
3991 const Type *NewTy = LHSV->getType()->getSignedVersion();
3992 if (LHSV->getType() != NewTy)
3993 LHSV = InsertCastBefore(LHSV, NewTy, I);
3994 if (RHSV->getType() != NewTy)
3995 RHSV = InsertCastBefore(RHSV, NewTy, I);
3996 return new SetCondInst(Cond, LHSV, RHSV);
4000 // Only lower this if the setcc is the only user of the GEP or if we expect
4001 // the result to fold to a constant!
4002 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4003 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4004 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4005 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4006 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4007 return new SetCondInst(Cond, L, R);
4014 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
4015 bool Changed = SimplifyCommutative(I);
4016 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4017 const Type *Ty = Op0->getType();
4021 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
4023 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
4024 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
4026 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4027 // addresses never equal each other! We already know that Op0 != Op1.
4028 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4029 isa<ConstantPointerNull>(Op0)) &&
4030 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4031 isa<ConstantPointerNull>(Op1)))
4032 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
4034 // setcc's with boolean values can always be turned into bitwise operations
4035 if (Ty == Type::BoolTy) {
4036 switch (I.getOpcode()) {
4037 default: assert(0 && "Invalid setcc instruction!");
4038 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
4039 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4040 InsertNewInstBefore(Xor, I);
4041 return BinaryOperator::createNot(Xor);
4043 case Instruction::SetNE:
4044 return BinaryOperator::createXor(Op0, Op1);
4046 case Instruction::SetGT:
4047 std::swap(Op0, Op1); // Change setgt -> setlt
4049 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
4050 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4051 InsertNewInstBefore(Not, I);
4052 return BinaryOperator::createAnd(Not, Op1);
4054 case Instruction::SetGE:
4055 std::swap(Op0, Op1); // Change setge -> setle
4057 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
4058 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4059 InsertNewInstBefore(Not, I);
4060 return BinaryOperator::createOr(Not, Op1);
4065 // See if we are doing a comparison between a constant and an instruction that
4066 // can be folded into the comparison.
4067 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4068 // Check to see if we are comparing against the minimum or maximum value...
4069 if (CI->isMinValue()) {
4070 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
4071 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4072 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
4073 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4074 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
4075 return BinaryOperator::createSetEQ(Op0, Op1);
4076 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
4077 return BinaryOperator::createSetNE(Op0, Op1);
4079 } else if (CI->isMaxValue()) {
4080 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
4081 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4082 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
4083 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4084 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
4085 return BinaryOperator::createSetEQ(Op0, Op1);
4086 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
4087 return BinaryOperator::createSetNE(Op0, Op1);
4089 // Comparing against a value really close to min or max?
4090 } else if (isMinValuePlusOne(CI)) {
4091 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
4092 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
4093 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
4094 return BinaryOperator::createSetNE(Op0, SubOne(CI));
4096 } else if (isMaxValueMinusOne(CI)) {
4097 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
4098 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
4099 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
4100 return BinaryOperator::createSetNE(Op0, AddOne(CI));
4103 // If we still have a setle or setge instruction, turn it into the
4104 // appropriate setlt or setgt instruction. Since the border cases have
4105 // already been handled above, this requires little checking.
4107 if (I.getOpcode() == Instruction::SetLE)
4108 return BinaryOperator::createSetLT(Op0, AddOne(CI));
4109 if (I.getOpcode() == Instruction::SetGE)
4110 return BinaryOperator::createSetGT(Op0, SubOne(CI));
4113 // See if we can fold the comparison based on bits known to be zero or one
4115 uint64_t KnownZero, KnownOne;
4116 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4117 KnownZero, KnownOne, 0))
4120 // Given the known and unknown bits, compute a range that the LHS could be
4122 if (KnownOne | KnownZero) {
4123 if (Ty->isUnsigned()) { // Unsigned comparison.
4125 uint64_t RHSVal = CI->getZExtValue();
4126 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4128 switch (I.getOpcode()) { // LE/GE have been folded already.
4129 default: assert(0 && "Unknown setcc opcode!");
4130 case Instruction::SetEQ:
4131 if (Max < RHSVal || Min > RHSVal)
4132 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4134 case Instruction::SetNE:
4135 if (Max < RHSVal || Min > RHSVal)
4136 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4138 case Instruction::SetLT:
4140 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4142 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4144 case Instruction::SetGT:
4146 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4148 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4151 } else { // Signed comparison.
4153 int64_t RHSVal = CI->getSExtValue();
4154 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4156 switch (I.getOpcode()) { // LE/GE have been folded already.
4157 default: assert(0 && "Unknown setcc opcode!");
4158 case Instruction::SetEQ:
4159 if (Max < RHSVal || Min > RHSVal)
4160 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4162 case Instruction::SetNE:
4163 if (Max < RHSVal || Min > RHSVal)
4164 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4166 case Instruction::SetLT:
4168 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4170 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4172 case Instruction::SetGT:
4174 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4176 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4182 // Since the RHS is a constantInt (CI), if the left hand side is an
4183 // instruction, see if that instruction also has constants so that the
4184 // instruction can be folded into the setcc
4185 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4186 switch (LHSI->getOpcode()) {
4187 case Instruction::And:
4188 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4189 LHSI->getOperand(0)->hasOneUse()) {
4190 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4192 // If an operand is an AND of a truncating cast, we can widen the
4193 // and/compare to be the input width without changing the value
4194 // produced, eliminating a cast.
4195 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4196 // We can do this transformation if either the AND constant does not
4197 // have its sign bit set or if it is an equality comparison.
4198 // Extending a relational comparison when we're checking the sign
4199 // bit would not work.
4200 if (Cast->hasOneUse() && Cast->isTruncIntCast() &&
4202 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4203 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4204 ConstantInt *NewCST;
4206 if (Cast->getOperand(0)->getType()->isSigned()) {
4207 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4208 AndCST->getZExtValue());
4209 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4210 CI->getZExtValue());
4212 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4213 AndCST->getZExtValue());
4214 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4215 CI->getZExtValue());
4217 Instruction *NewAnd =
4218 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4220 InsertNewInstBefore(NewAnd, I);
4221 return new SetCondInst(I.getOpcode(), NewAnd, NewCI);
4225 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4226 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4227 // happens a LOT in code produced by the C front-end, for bitfield
4229 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4231 // Check to see if there is a noop-cast between the shift and the and.
4233 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4234 if (CI->getOperand(0)->getType()->isIntegral() &&
4235 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
4236 CI->getType()->getPrimitiveSizeInBits())
4237 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4241 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4242 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4243 const Type *AndTy = AndCST->getType(); // Type of the and.
4245 // We can fold this as long as we can't shift unknown bits
4246 // into the mask. This can only happen with signed shift
4247 // rights, as they sign-extend.
4249 bool CanFold = Shift->isLogicalShift();
4251 // To test for the bad case of the signed shr, see if any
4252 // of the bits shifted in could be tested after the mask.
4253 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4254 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4256 Constant *OShAmt = ConstantInt::get(Type::UByteTy, ShAmtVal);
4258 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4260 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4266 if (Shift->getOpcode() == Instruction::Shl)
4267 NewCst = ConstantExpr::getUShr(CI, ShAmt);
4269 NewCst = ConstantExpr::getShl(CI, ShAmt);
4271 // Check to see if we are shifting out any of the bits being
4273 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4274 // If we shifted bits out, the fold is not going to work out.
4275 // As a special case, check to see if this means that the
4276 // result is always true or false now.
4277 if (I.getOpcode() == Instruction::SetEQ)
4278 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4279 if (I.getOpcode() == Instruction::SetNE)
4280 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4282 I.setOperand(1, NewCst);
4283 Constant *NewAndCST;
4284 if (Shift->getOpcode() == Instruction::Shl)
4285 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
4287 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4288 LHSI->setOperand(1, NewAndCST);
4290 LHSI->setOperand(0, Shift->getOperand(0));
4292 Value *NewCast = InsertCastBefore(Shift->getOperand(0), AndTy,
4294 LHSI->setOperand(0, NewCast);
4296 WorkList.push_back(Shift); // Shift is dead.
4297 AddUsesToWorkList(I);
4303 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4304 // preferable because it allows the C<<Y expression to be hoisted out
4305 // of a loop if Y is invariant and X is not.
4306 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4307 I.isEquality() && !Shift->isArithmeticShift() &&
4308 isa<Instruction>(Shift->getOperand(0))) {
4311 if (Shift->getOpcode() == Instruction::Shr) {
4312 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4315 // Make sure we insert a logical shift.
4316 Constant *NewAndCST = AndCST;
4317 if (AndCST->getType()->isSigned())
4318 NewAndCST = ConstantExpr::getCast(AndCST,
4319 AndCST->getType()->getUnsignedVersion());
4320 NS = new ShiftInst(Instruction::Shr, NewAndCST,
4321 Shift->getOperand(1), "tmp");
4323 InsertNewInstBefore(cast<Instruction>(NS), I);
4325 // If C's sign doesn't agree with the and, insert a cast now.
4326 if (NS->getType() != LHSI->getType())
4327 NS = InsertCastBefore(NS, LHSI->getType(), I);
4329 Value *ShiftOp = Shift->getOperand(0);
4330 if (ShiftOp->getType() != LHSI->getType())
4331 ShiftOp = InsertCastBefore(ShiftOp, LHSI->getType(), I);
4333 // Compute X & (C << Y).
4334 Instruction *NewAnd =
4335 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4336 InsertNewInstBefore(NewAnd, I);
4338 I.setOperand(0, NewAnd);
4344 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
4345 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4346 if (I.isEquality()) {
4347 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4349 // Check that the shift amount is in range. If not, don't perform
4350 // undefined shifts. When the shift is visited it will be
4352 if (ShAmt->getZExtValue() >= TypeBits)
4355 // If we are comparing against bits always shifted out, the
4356 // comparison cannot succeed.
4358 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
4359 if (Comp != CI) {// Comparing against a bit that we know is zero.
4360 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4361 Constant *Cst = ConstantBool::get(IsSetNE);
4362 return ReplaceInstUsesWith(I, Cst);
4365 if (LHSI->hasOneUse()) {
4366 // Otherwise strength reduce the shift into an and.
4367 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4368 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4371 if (CI->getType()->isUnsigned()) {
4372 Mask = ConstantInt::get(CI->getType(), Val);
4373 } else if (ShAmtVal != 0) {
4374 Mask = ConstantInt::get(CI->getType(), Val);
4376 Mask = ConstantInt::getAllOnesValue(CI->getType());
4380 BinaryOperator::createAnd(LHSI->getOperand(0),
4381 Mask, LHSI->getName()+".mask");
4382 Value *And = InsertNewInstBefore(AndI, I);
4383 return new SetCondInst(I.getOpcode(), And,
4384 ConstantExpr::getUShr(CI, ShAmt));
4390 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
4391 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4392 if (I.isEquality()) {
4393 // Check that the shift amount is in range. If not, don't perform
4394 // undefined shifts. When the shift is visited it will be
4396 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4397 if (ShAmt->getZExtValue() >= TypeBits)
4400 // If we are comparing against bits always shifted out, the
4401 // comparison cannot succeed.
4403 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
4405 if (Comp != CI) {// Comparing against a bit that we know is zero.
4406 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4407 Constant *Cst = ConstantBool::get(IsSetNE);
4408 return ReplaceInstUsesWith(I, Cst);
4411 if (LHSI->hasOneUse() || CI->isNullValue()) {
4412 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4414 // Otherwise strength reduce the shift into an and.
4415 uint64_t Val = ~0ULL; // All ones.
4416 Val <<= ShAmtVal; // Shift over to the right spot.
4419 if (CI->getType()->isUnsigned()) {
4420 Val &= ~0ULL >> (64-TypeBits);
4421 Mask = ConstantInt::get(CI->getType(), Val);
4423 Mask = ConstantInt::get(CI->getType(), Val);
4427 BinaryOperator::createAnd(LHSI->getOperand(0),
4428 Mask, LHSI->getName()+".mask");
4429 Value *And = InsertNewInstBefore(AndI, I);
4430 return new SetCondInst(I.getOpcode(), And,
4431 ConstantExpr::getShl(CI, ShAmt));
4437 case Instruction::SDiv:
4438 case Instruction::UDiv:
4439 // Fold: setcc ([us]div X, C1), C2 -> range test
4440 // Fold this div into the comparison, producing a range check.
4441 // Determine, based on the divide type, what the range is being
4442 // checked. If there is an overflow on the low or high side, remember
4443 // it, otherwise compute the range [low, hi) bounding the new value.
4444 // See: InsertRangeTest above for the kinds of replacements possible.
4445 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4446 // FIXME: If the operand types don't match the type of the divide
4447 // then don't attempt this transform. The code below doesn't have the
4448 // logic to deal with a signed divide and an unsigned compare (and
4449 // vice versa). This is because (x /s C1) <s C2 produces different
4450 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4451 // (x /u C1) <u C2. Simply casting the operands and result won't
4452 // work. :( The if statement below tests that condition and bails
4454 const Type* DivRHSTy = DivRHS->getType();
4455 unsigned DivOpCode = LHSI->getOpcode();
4456 if (I.isEquality() &&
4457 ((DivOpCode == Instruction::SDiv && DivRHSTy->isUnsigned()) ||
4458 (DivOpCode == Instruction::UDiv && DivRHSTy->isSigned())))
4461 // Initialize the variables that will indicate the nature of the
4463 bool LoOverflow = false, HiOverflow = false;
4464 ConstantInt *LoBound = 0, *HiBound = 0;
4466 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4467 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4468 // C2 (CI). By solving for X we can turn this into a range check
4469 // instead of computing a divide.
4471 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4473 // Determine if the product overflows by seeing if the product is
4474 // not equal to the divide. Make sure we do the same kind of divide
4475 // as in the LHS instruction that we're folding.
4476 bool ProdOV = !DivRHS->isNullValue() &&
4477 (DivOpCode == Instruction::SDiv ?
4478 ConstantExpr::getSDiv(Prod, DivRHS) :
4479 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4481 // Get the SetCC opcode
4482 Instruction::BinaryOps Opcode = I.getOpcode();
4484 if (DivRHS->isNullValue()) {
4485 // Don't hack on divide by zeros!
4486 } else if (DivOpCode == Instruction::UDiv) { // udiv
4488 LoOverflow = ProdOV;
4489 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4490 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4491 if (CI->isNullValue()) { // (X / pos) op 0
4493 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4495 } else if (isPositive(CI)) { // (X / pos) op pos
4497 LoOverflow = ProdOV;
4498 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4499 } else { // (X / pos) op neg
4500 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4501 LoOverflow = AddWithOverflow(LoBound, Prod,
4502 cast<ConstantInt>(DivRHSH));
4504 HiOverflow = ProdOV;
4506 } else { // Divisor is < 0.
4507 if (CI->isNullValue()) { // (X / neg) op 0
4508 LoBound = AddOne(DivRHS);
4509 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4510 if (HiBound == DivRHS)
4511 LoBound = 0; // - INTMIN = INTMIN
4512 } else if (isPositive(CI)) { // (X / neg) op pos
4513 HiOverflow = LoOverflow = ProdOV;
4515 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4516 HiBound = AddOne(Prod);
4517 } else { // (X / neg) op neg
4519 LoOverflow = HiOverflow = ProdOV;
4520 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4523 // Dividing by a negate swaps the condition.
4524 Opcode = SetCondInst::getSwappedCondition(Opcode);
4528 Value *X = LHSI->getOperand(0);
4530 default: assert(0 && "Unhandled setcc opcode!");
4531 case Instruction::SetEQ:
4532 if (LoOverflow && HiOverflow)
4533 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4534 else if (HiOverflow)
4535 return new SetCondInst(Instruction::SetGE, X, LoBound);
4536 else if (LoOverflow)
4537 return new SetCondInst(Instruction::SetLT, X, HiBound);
4539 return InsertRangeTest(X, LoBound, HiBound, true, I);
4540 case Instruction::SetNE:
4541 if (LoOverflow && HiOverflow)
4542 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4543 else if (HiOverflow)
4544 return new SetCondInst(Instruction::SetLT, X, LoBound);
4545 else if (LoOverflow)
4546 return new SetCondInst(Instruction::SetGE, X, HiBound);
4548 return InsertRangeTest(X, LoBound, HiBound, false, I);
4549 case Instruction::SetLT:
4551 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4552 return new SetCondInst(Instruction::SetLT, X, LoBound);
4553 case Instruction::SetGT:
4555 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4556 return new SetCondInst(Instruction::SetGE, X, HiBound);
4563 // Simplify seteq and setne instructions...
4564 if (I.isEquality()) {
4565 bool isSetNE = I.getOpcode() == Instruction::SetNE;
4567 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4568 // the second operand is a constant, simplify a bit.
4569 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4570 switch (BO->getOpcode()) {
4572 case Instruction::SRem:
4573 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4574 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4576 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4577 if (V > 1 && isPowerOf2_64(V)) {
4578 Value *NewRem = InsertNewInstBefore(
4579 BinaryOperator::createURem(BO->getOperand(0),
4582 return BinaryOperator::create(
4583 I.getOpcode(), NewRem,
4584 Constant::getNullValue(NewRem->getType()));
4590 case Instruction::Rem:
4591 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4592 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4593 BO->hasOneUse() && BO->getOperand(1)->getType()->isSigned()) {
4594 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4595 if (V > 1 && isPowerOf2_64(V)) {
4596 unsigned L2 = Log2_64(V);
4597 const Type *UTy = BO->getType()->getUnsignedVersion();
4598 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
4600 Constant *RHSCst = ConstantInt::get(UTy, 1ULL << L2);
4601 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
4602 RHSCst, BO->getName()), I);
4603 return BinaryOperator::create(I.getOpcode(), NewRem,
4604 Constant::getNullValue(UTy));
4608 case Instruction::Add:
4609 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4610 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4611 if (BO->hasOneUse())
4612 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4613 ConstantExpr::getSub(CI, BOp1C));
4614 } else if (CI->isNullValue()) {
4615 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4616 // efficiently invertible, or if the add has just this one use.
4617 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4619 if (Value *NegVal = dyn_castNegVal(BOp1))
4620 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
4621 else if (Value *NegVal = dyn_castNegVal(BOp0))
4622 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
4623 else if (BO->hasOneUse()) {
4624 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4626 InsertNewInstBefore(Neg, I);
4627 return new SetCondInst(I.getOpcode(), BOp0, Neg);
4631 case Instruction::Xor:
4632 // For the xor case, we can xor two constants together, eliminating
4633 // the explicit xor.
4634 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4635 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
4636 ConstantExpr::getXor(CI, BOC));
4639 case Instruction::Sub:
4640 // Replace (([sub|xor] A, B) != 0) with (A != B)
4641 if (CI->isNullValue())
4642 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4646 case Instruction::Or:
4647 // If bits are being or'd in that are not present in the constant we
4648 // are comparing against, then the comparison could never succeed!
4649 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4650 Constant *NotCI = ConstantExpr::getNot(CI);
4651 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4652 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4656 case Instruction::And:
4657 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4658 // If bits are being compared against that are and'd out, then the
4659 // comparison can never succeed!
4660 if (!ConstantExpr::getAnd(CI,
4661 ConstantExpr::getNot(BOC))->isNullValue())
4662 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4664 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4665 if (CI == BOC && isOneBitSet(CI))
4666 return new SetCondInst(isSetNE ? Instruction::SetEQ :
4667 Instruction::SetNE, Op0,
4668 Constant::getNullValue(CI->getType()));
4670 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
4671 // to be a signed value as appropriate.
4672 if (isSignBit(BOC)) {
4673 Value *X = BO->getOperand(0);
4674 // If 'X' is not signed, insert a cast now...
4675 if (!BOC->getType()->isSigned()) {
4676 const Type *DestTy = BOC->getType()->getSignedVersion();
4677 X = InsertCastBefore(X, DestTy, I);
4679 return new SetCondInst(isSetNE ? Instruction::SetLT :
4680 Instruction::SetGE, X,
4681 Constant::getNullValue(X->getType()));
4684 // ((X & ~7) == 0) --> X < 8
4685 if (CI->isNullValue() && isHighOnes(BOC)) {
4686 Value *X = BO->getOperand(0);
4687 Constant *NegX = ConstantExpr::getNeg(BOC);
4689 // If 'X' is signed, insert a cast now.
4690 if (NegX->getType()->isSigned()) {
4691 const Type *DestTy = NegX->getType()->getUnsignedVersion();
4692 X = InsertCastBefore(X, DestTy, I);
4693 NegX = ConstantExpr::getCast(NegX, DestTy);
4696 return new SetCondInst(isSetNE ? Instruction::SetGE :
4697 Instruction::SetLT, X, NegX);
4704 } else { // Not a SetEQ/SetNE
4705 // If the LHS is a cast from an integral value of the same size,
4706 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
4707 Value *CastOp = Cast->getOperand(0);
4708 const Type *SrcTy = CastOp->getType();
4709 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
4710 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
4711 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
4712 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
4713 "Source and destination signednesses should differ!");
4714 if (Cast->getType()->isSigned()) {
4715 // If this is a signed comparison, check for comparisons in the
4716 // vicinity of zero.
4717 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
4719 return BinaryOperator::createSetGT(CastOp,
4720 ConstantInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
4721 else if (I.getOpcode() == Instruction::SetGT &&
4722 cast<ConstantInt>(CI)->getSExtValue() == -1)
4723 // X > -1 => x < 128
4724 return BinaryOperator::createSetLT(CastOp,
4725 ConstantInt::get(SrcTy, 1ULL << (SrcTySize-1)));
4727 ConstantInt *CUI = cast<ConstantInt>(CI);
4728 if (I.getOpcode() == Instruction::SetLT &&
4729 CUI->getZExtValue() == 1ULL << (SrcTySize-1))
4730 // X < 128 => X > -1
4731 return BinaryOperator::createSetGT(CastOp,
4732 ConstantInt::get(SrcTy, -1));
4733 else if (I.getOpcode() == Instruction::SetGT &&
4734 CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
4736 return BinaryOperator::createSetLT(CastOp,
4737 Constant::getNullValue(SrcTy));
4744 // Handle setcc with constant RHS's that can be integer, FP or pointer.
4745 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4746 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4747 switch (LHSI->getOpcode()) {
4748 case Instruction::GetElementPtr:
4749 if (RHSC->isNullValue()) {
4750 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
4751 bool isAllZeros = true;
4752 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4753 if (!isa<Constant>(LHSI->getOperand(i)) ||
4754 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4759 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
4760 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4764 case Instruction::PHI:
4765 if (Instruction *NV = FoldOpIntoPhi(I))
4768 case Instruction::Select:
4769 // If either operand of the select is a constant, we can fold the
4770 // comparison into the select arms, which will cause one to be
4771 // constant folded and the select turned into a bitwise or.
4772 Value *Op1 = 0, *Op2 = 0;
4773 if (LHSI->hasOneUse()) {
4774 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4775 // Fold the known value into the constant operand.
4776 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4777 // Insert a new SetCC of the other select operand.
4778 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4779 LHSI->getOperand(2), RHSC,
4781 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4782 // Fold the known value into the constant operand.
4783 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4784 // Insert a new SetCC of the other select operand.
4785 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4786 LHSI->getOperand(1), RHSC,
4792 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4797 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
4798 if (User *GEP = dyn_castGetElementPtr(Op0))
4799 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
4801 if (User *GEP = dyn_castGetElementPtr(Op1))
4802 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
4803 SetCondInst::getSwappedCondition(I.getOpcode()), I))
4806 // Test to see if the operands of the setcc are casted versions of other
4807 // values. If the cast can be stripped off both arguments, we do so now.
4808 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4809 Value *CastOp0 = CI->getOperand(0);
4810 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
4811 (isa<Constant>(Op1) || isa<CastInst>(Op1)) && I.isEquality()) {
4812 // We keep moving the cast from the left operand over to the right
4813 // operand, where it can often be eliminated completely.
4816 // If operand #1 is a cast instruction, see if we can eliminate it as
4818 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
4819 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
4821 Op1 = CI2->getOperand(0);
4823 // If Op1 is a constant, we can fold the cast into the constant.
4824 if (Op1->getType() != Op0->getType())
4825 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4826 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
4828 // Otherwise, cast the RHS right before the setcc
4829 Op1 = InsertCastBefore(Op1, Op0->getType(), I);
4831 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
4834 // Handle the special case of: setcc (cast bool to X), <cst>
4835 // This comes up when you have code like
4838 // For generality, we handle any zero-extension of any operand comparison
4839 // with a constant or another cast from the same type.
4840 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4841 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
4845 if (I.isEquality()) {
4847 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4848 (A == Op1 || B == Op1)) {
4849 // (A^B) == A -> B == 0
4850 Value *OtherVal = A == Op1 ? B : A;
4851 return BinaryOperator::create(I.getOpcode(), OtherVal,
4852 Constant::getNullValue(A->getType()));
4853 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4854 (A == Op0 || B == Op0)) {
4855 // A == (A^B) -> B == 0
4856 Value *OtherVal = A == Op0 ? B : A;
4857 return BinaryOperator::create(I.getOpcode(), OtherVal,
4858 Constant::getNullValue(A->getType()));
4859 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4860 // (A-B) == A -> B == 0
4861 return BinaryOperator::create(I.getOpcode(), B,
4862 Constant::getNullValue(B->getType()));
4863 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4864 // A == (A-B) -> B == 0
4865 return BinaryOperator::create(I.getOpcode(), B,
4866 Constant::getNullValue(B->getType()));
4869 return Changed ? &I : 0;
4872 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
4873 // We only handle extending casts so far.
4875 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
4876 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
4877 const Type *SrcTy = LHSCIOp->getType();
4878 const Type *DestTy = SCI.getOperand(0)->getType();
4881 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
4884 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
4885 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
4886 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
4888 // Is this a sign or zero extension?
4889 bool isSignSrc = SrcTy->isSigned();
4890 bool isSignDest = DestTy->isSigned();
4892 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
4893 // Not an extension from the same type?
4894 RHSCIOp = CI->getOperand(0);
4895 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
4896 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
4897 // Compute the constant that would happen if we truncated to SrcTy then
4898 // reextended to DestTy.
4899 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
4901 if (ConstantExpr::getCast(Res, DestTy) == CI) {
4902 // Make sure that src sign and dest sign match. For example,
4904 // %A = cast short %X to uint
4905 // %B = setgt uint %A, 1330
4907 // It is incorrect to transform this into
4909 // %B = setgt short %X, 1330
4911 // because %A may have negative value.
4912 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
4913 // OR operation is EQ/NE.
4914 if (isSignSrc == isSignDest || SrcTy == Type::BoolTy || SCI.isEquality())
4919 // If the value cannot be represented in the shorter type, we cannot emit
4920 // a simple comparison.
4921 if (SCI.getOpcode() == Instruction::SetEQ)
4922 return ReplaceInstUsesWith(SCI, ConstantBool::getFalse());
4923 if (SCI.getOpcode() == Instruction::SetNE)
4924 return ReplaceInstUsesWith(SCI, ConstantBool::getTrue());
4926 // Evaluate the comparison for LT.
4928 if (DestTy->isSigned()) {
4929 // We're performing a signed comparison.
4931 // Signed extend and signed comparison.
4932 if (cast<ConstantInt>(CI)->getSExtValue() < 0)// X < (small) --> false
4933 Result = ConstantBool::getFalse();
4935 Result = ConstantBool::getTrue(); // X < (large) --> true
4937 // Unsigned extend and signed comparison.
4938 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
4939 Result = ConstantBool::getFalse();
4941 Result = ConstantBool::getTrue();
4944 // We're performing an unsigned comparison.
4946 // Unsigned extend & compare -> always true.
4947 Result = ConstantBool::getTrue();
4949 // We're performing an unsigned comp with a sign extended value.
4950 // This is true if the input is >= 0. [aka >s -1]
4951 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
4952 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
4953 NegOne, SCI.getName()), SCI);
4957 // Finally, return the value computed.
4958 if (SCI.getOpcode() == Instruction::SetLT) {
4959 return ReplaceInstUsesWith(SCI, Result);
4961 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
4962 if (Constant *CI = dyn_cast<Constant>(Result))
4963 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
4965 return BinaryOperator::createNot(Result);
4972 // Okay, just insert a compare of the reduced operands now!
4973 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
4976 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
4977 assert(I.getOperand(1)->getType() == Type::UByteTy);
4978 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4979 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4981 // shl X, 0 == X and shr X, 0 == X
4982 // shl 0, X == 0 and shr 0, X == 0
4983 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
4984 Op0 == Constant::getNullValue(Op0->getType()))
4985 return ReplaceInstUsesWith(I, Op0);
4987 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
4988 if (!isLeftShift && I.getType()->isSigned())
4989 return ReplaceInstUsesWith(I, Op0);
4990 else // undef << X -> 0 AND undef >>u X -> 0
4991 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4993 if (isa<UndefValue>(Op1)) {
4994 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
4995 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4997 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
5000 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
5002 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5003 if (CSI->isAllOnesValue() && Op0->getType()->isSigned())
5004 return ReplaceInstUsesWith(I, CSI);
5006 // Try to fold constant and into select arguments.
5007 if (isa<Constant>(Op0))
5008 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5009 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5012 // See if we can turn a signed shr into an unsigned shr.
5013 if (I.isArithmeticShift()) {
5014 if (MaskedValueIsZero(Op0,
5015 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5016 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
5017 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
5019 return new CastInst(V, I.getType());
5023 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5024 if (CUI->getType()->isUnsigned())
5025 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5030 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5032 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5033 bool isSignedShift = Op0->getType()->isSigned();
5034 bool isUnsignedShift = !isSignedShift;
5036 // See if we can simplify any instructions used by the instruction whose sole
5037 // purpose is to compute bits we don't care about.
5038 uint64_t KnownZero, KnownOne;
5039 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5040 KnownZero, KnownOne))
5043 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5044 // of a signed value.
5046 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5047 if (Op1->getZExtValue() >= TypeBits) {
5048 if (isUnsignedShift || isLeftShift)
5049 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5051 I.setOperand(1, ConstantInt::get(Type::UByteTy, TypeBits-1));
5056 // ((X*C1) << C2) == (X * (C1 << C2))
5057 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5058 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5059 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5060 return BinaryOperator::createMul(BO->getOperand(0),
5061 ConstantExpr::getShl(BOOp, Op1));
5063 // Try to fold constant and into select arguments.
5064 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5065 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5067 if (isa<PHINode>(Op0))
5068 if (Instruction *NV = FoldOpIntoPhi(I))
5071 if (Op0->hasOneUse()) {
5072 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5073 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5076 switch (Op0BO->getOpcode()) {
5078 case Instruction::Add:
5079 case Instruction::And:
5080 case Instruction::Or:
5081 case Instruction::Xor:
5082 // These operators commute.
5083 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5084 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5085 match(Op0BO->getOperand(1),
5086 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5087 Instruction *YS = new ShiftInst(Instruction::Shl,
5088 Op0BO->getOperand(0), Op1,
5090 InsertNewInstBefore(YS, I); // (Y << C)
5092 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5093 Op0BO->getOperand(1)->getName());
5094 InsertNewInstBefore(X, I); // (X + (Y << C))
5095 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5096 C2 = ConstantExpr::getShl(C2, Op1);
5097 return BinaryOperator::createAnd(X, C2);
5100 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5101 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5102 match(Op0BO->getOperand(1),
5103 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5104 m_ConstantInt(CC))) && V2 == Op1 &&
5105 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5106 Instruction *YS = new ShiftInst(Instruction::Shl,
5107 Op0BO->getOperand(0), Op1,
5109 InsertNewInstBefore(YS, I); // (Y << C)
5111 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5112 V1->getName()+".mask");
5113 InsertNewInstBefore(XM, I); // X & (CC << C)
5115 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5119 case Instruction::Sub:
5120 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5121 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5122 match(Op0BO->getOperand(0),
5123 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5124 Instruction *YS = new ShiftInst(Instruction::Shl,
5125 Op0BO->getOperand(1), Op1,
5127 InsertNewInstBefore(YS, I); // (Y << C)
5129 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5130 Op0BO->getOperand(0)->getName());
5131 InsertNewInstBefore(X, I); // (X + (Y << C))
5132 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5133 C2 = ConstantExpr::getShl(C2, Op1);
5134 return BinaryOperator::createAnd(X, C2);
5137 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5138 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5139 match(Op0BO->getOperand(0),
5140 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5141 m_ConstantInt(CC))) && V2 == Op1 &&
5142 cast<BinaryOperator>(Op0BO->getOperand(0))
5143 ->getOperand(0)->hasOneUse()) {
5144 Instruction *YS = new ShiftInst(Instruction::Shl,
5145 Op0BO->getOperand(1), Op1,
5147 InsertNewInstBefore(YS, I); // (Y << C)
5149 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5150 V1->getName()+".mask");
5151 InsertNewInstBefore(XM, I); // X & (CC << C)
5153 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5160 // If the operand is an bitwise operator with a constant RHS, and the
5161 // shift is the only use, we can pull it out of the shift.
5162 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5163 bool isValid = true; // Valid only for And, Or, Xor
5164 bool highBitSet = false; // Transform if high bit of constant set?
5166 switch (Op0BO->getOpcode()) {
5167 default: isValid = false; break; // Do not perform transform!
5168 case Instruction::Add:
5169 isValid = isLeftShift;
5171 case Instruction::Or:
5172 case Instruction::Xor:
5175 case Instruction::And:
5180 // If this is a signed shift right, and the high bit is modified
5181 // by the logical operation, do not perform the transformation.
5182 // The highBitSet boolean indicates the value of the high bit of
5183 // the constant which would cause it to be modified for this
5186 if (isValid && !isLeftShift && isSignedShift) {
5187 uint64_t Val = Op0C->getZExtValue();
5188 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5192 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5194 Instruction *NewShift =
5195 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5198 InsertNewInstBefore(NewShift, I);
5200 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5207 // Find out if this is a shift of a shift by a constant.
5208 ShiftInst *ShiftOp = 0;
5209 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5211 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
5212 // If this is a noop-integer case of a shift instruction, use the shift.
5213 if (CI->getOperand(0)->getType()->isInteger() &&
5214 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
5215 CI->getType()->getPrimitiveSizeInBits() &&
5216 isa<ShiftInst>(CI->getOperand(0))) {
5217 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5221 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5222 // Find the operands and properties of the input shift. Note that the
5223 // signedness of the input shift may differ from the current shift if there
5224 // is a noop cast between the two.
5225 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5226 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
5227 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5229 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5231 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5232 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5234 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5235 if (isLeftShift == isShiftOfLeftShift) {
5236 // Do not fold these shifts if the first one is signed and the second one
5237 // is unsigned and this is a right shift. Further, don't do any folding
5239 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5242 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5243 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5244 Amt = Op0->getType()->getPrimitiveSizeInBits();
5246 Value *Op = ShiftOp->getOperand(0);
5247 if (isShiftOfSignedShift != isSignedShift)
5248 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
5249 return new ShiftInst(I.getOpcode(), Op,
5250 ConstantInt::get(Type::UByteTy, Amt));
5253 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5254 // signed types, we can only support the (A >> c1) << c2 configuration,
5255 // because it can not turn an arbitrary bit of A into a sign bit.
5256 if (isUnsignedShift || isLeftShift) {
5257 // Calculate bitmask for what gets shifted off the edge.
5258 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
5260 C = ConstantExpr::getShl(C, ShiftAmt1C);
5262 C = ConstantExpr::getUShr(C, ShiftAmt1C);
5264 Value *Op = ShiftOp->getOperand(0);
5265 if (isShiftOfSignedShift != isSignedShift)
5266 Op = InsertCastBefore(Op, I.getType(), I);
5269 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5270 InsertNewInstBefore(Mask, I);
5272 // Figure out what flavor of shift we should use...
5273 if (ShiftAmt1 == ShiftAmt2) {
5274 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5275 } else if (ShiftAmt1 < ShiftAmt2) {
5276 return new ShiftInst(I.getOpcode(), Mask,
5277 ConstantInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
5278 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5279 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5280 // Make sure to emit an unsigned shift right, not a signed one.
5281 Mask = InsertNewInstBefore(new CastInst(Mask,
5282 Mask->getType()->getUnsignedVersion(),
5284 Mask = new ShiftInst(Instruction::Shr, Mask,
5285 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5286 InsertNewInstBefore(Mask, I);
5287 return new CastInst(Mask, I.getType());
5289 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5290 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5293 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5294 Op = InsertCastBefore(Mask, I.getType()->getSignedVersion(), I);
5295 Instruction *Shift =
5296 new ShiftInst(ShiftOp->getOpcode(), Op,
5297 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5298 InsertNewInstBefore(Shift, I);
5300 C = ConstantIntegral::getAllOnesValue(Shift->getType());
5301 C = ConstantExpr::getShl(C, Op1);
5302 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5303 InsertNewInstBefore(Mask, I);
5304 return new CastInst(Mask, I.getType());
5307 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5308 // this case, C1 == C2 and C1 is 8, 16, or 32.
5309 if (ShiftAmt1 == ShiftAmt2) {
5310 const Type *SExtType = 0;
5311 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5312 case 8 : SExtType = Type::SByteTy; break;
5313 case 16: SExtType = Type::ShortTy; break;
5314 case 32: SExtType = Type::IntTy; break;
5318 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
5320 InsertNewInstBefore(NewTrunc, I);
5321 return new CastInst(NewTrunc, I.getType());
5330 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5331 /// expression. If so, decompose it, returning some value X, such that Val is
5334 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5336 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
5337 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5338 if (CI->getType()->isUnsigned()) {
5339 Offset = CI->getZExtValue();
5341 return ConstantInt::get(Type::UIntTy, 0);
5343 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5344 if (I->getNumOperands() == 2) {
5345 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5346 if (CUI->getType()->isUnsigned()) {
5347 if (I->getOpcode() == Instruction::Shl) {
5348 // This is a value scaled by '1 << the shift amt'.
5349 Scale = 1U << CUI->getZExtValue();
5351 return I->getOperand(0);
5352 } else if (I->getOpcode() == Instruction::Mul) {
5353 // This value is scaled by 'CUI'.
5354 Scale = CUI->getZExtValue();
5356 return I->getOperand(0);
5357 } else if (I->getOpcode() == Instruction::Add) {
5358 // We have X+C. Check to see if we really have (X*C2)+C1,
5359 // where C1 is divisible by C2.
5362 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5363 Offset += CUI->getZExtValue();
5364 if (SubScale > 1 && (Offset % SubScale == 0)) {
5374 // Otherwise, we can't look past this.
5381 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5382 /// try to eliminate the cast by moving the type information into the alloc.
5383 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5384 AllocationInst &AI) {
5385 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5386 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5388 // Remove any uses of AI that are dead.
5389 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5390 std::vector<Instruction*> DeadUsers;
5391 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5392 Instruction *User = cast<Instruction>(*UI++);
5393 if (isInstructionTriviallyDead(User)) {
5394 while (UI != E && *UI == User)
5395 ++UI; // If this instruction uses AI more than once, don't break UI.
5397 // Add operands to the worklist.
5398 AddUsesToWorkList(*User);
5400 DEBUG(std::cerr << "IC: DCE: " << *User);
5402 User->eraseFromParent();
5403 removeFromWorkList(User);
5407 // Get the type really allocated and the type casted to.
5408 const Type *AllocElTy = AI.getAllocatedType();
5409 const Type *CastElTy = PTy->getElementType();
5410 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5412 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5413 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5414 if (CastElTyAlign < AllocElTyAlign) return 0;
5416 // If the allocation has multiple uses, only promote it if we are strictly
5417 // increasing the alignment of the resultant allocation. If we keep it the
5418 // same, we open the door to infinite loops of various kinds.
5419 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5421 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5422 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5423 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5425 // See if we can satisfy the modulus by pulling a scale out of the array
5427 unsigned ArraySizeScale, ArrayOffset;
5428 Value *NumElements = // See if the array size is a decomposable linear expr.
5429 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5431 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5433 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5434 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5436 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5441 // If the allocation size is constant, form a constant mul expression
5442 Amt = ConstantInt::get(Type::UIntTy, Scale);
5443 if (isa<ConstantInt>(NumElements) && NumElements->getType()->isUnsigned())
5444 Amt = ConstantExpr::getMul(
5445 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5446 // otherwise multiply the amount and the number of elements
5447 else if (Scale != 1) {
5448 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5449 Amt = InsertNewInstBefore(Tmp, AI);
5453 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5454 Value *Off = ConstantInt::get(Type::UIntTy, Offset);
5455 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5456 Amt = InsertNewInstBefore(Tmp, AI);
5459 std::string Name = AI.getName(); AI.setName("");
5460 AllocationInst *New;
5461 if (isa<MallocInst>(AI))
5462 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5464 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5465 InsertNewInstBefore(New, AI);
5467 // If the allocation has multiple uses, insert a cast and change all things
5468 // that used it to use the new cast. This will also hack on CI, but it will
5470 if (!AI.hasOneUse()) {
5471 AddUsesToWorkList(AI);
5472 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
5473 InsertNewInstBefore(NewCast, AI);
5474 AI.replaceAllUsesWith(NewCast);
5476 return ReplaceInstUsesWith(CI, New);
5479 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5480 /// and return it without inserting any new casts. This is used by code that
5481 /// tries to decide whether promoting or shrinking integer operations to wider
5482 /// or smaller types will allow us to eliminate a truncate or extend.
5483 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5484 int &NumCastsRemoved) {
5485 if (isa<Constant>(V)) return true;
5487 Instruction *I = dyn_cast<Instruction>(V);
5488 if (!I || !I->hasOneUse()) return false;
5490 switch (I->getOpcode()) {
5491 case Instruction::And:
5492 case Instruction::Or:
5493 case Instruction::Xor:
5494 // These operators can all arbitrarily be extended or truncated.
5495 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5496 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5497 case Instruction::Cast:
5498 // If this is a cast from the destination type, we can trivially eliminate
5499 // it, and this will remove a cast overall.
5500 if (I->getOperand(0)->getType() == Ty) {
5501 // If the first operand is itself a cast, and is eliminable, do not count
5502 // this as an eliminable cast. We would prefer to eliminate those two
5504 if (CastInst *OpCast = dyn_cast<CastInst>(I->getOperand(0)))
5510 // TODO: Can handle more cases here.
5517 /// EvaluateInDifferentType - Given an expression that
5518 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5519 /// evaluate the expression.
5520 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty) {
5521 if (Constant *C = dyn_cast<Constant>(V))
5522 return ConstantExpr::getCast(C, Ty);
5524 // Otherwise, it must be an instruction.
5525 Instruction *I = cast<Instruction>(V);
5526 Instruction *Res = 0;
5527 switch (I->getOpcode()) {
5528 case Instruction::And:
5529 case Instruction::Or:
5530 case Instruction::Xor: {
5531 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty);
5532 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty);
5533 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5534 LHS, RHS, I->getName());
5537 case Instruction::Cast:
5538 // If this is a cast from the destination type, return the input.
5539 if (I->getOperand(0)->getType() == Ty)
5540 return I->getOperand(0);
5542 // TODO: Can handle more cases here.
5543 assert(0 && "Unreachable!");
5547 return InsertNewInstBefore(Res, *I);
5551 // CastInst simplification
5553 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
5554 Value *Src = CI.getOperand(0);
5556 // If the user is casting a value to the same type, eliminate this cast
5558 if (CI.getType() == Src->getType())
5559 return ReplaceInstUsesWith(CI, Src);
5561 if (isa<UndefValue>(Src)) // cast undef -> undef
5562 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5564 // If casting the result of another cast instruction, try to eliminate this
5567 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5568 Value *A = CSrc->getOperand(0);
5569 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
5570 CI.getType(), TD)) {
5571 // This instruction now refers directly to the cast's src operand. This
5572 // has a good chance of making CSrc dead.
5573 CI.setOperand(0, CSrc->getOperand(0));
5577 // If this is an A->B->A cast, and we are dealing with integral types, try
5578 // to convert this into a logical 'and' instruction.
5580 if (A->getType()->isInteger() &&
5581 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
5582 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
5583 CSrc->getType()->getPrimitiveSizeInBits() <
5584 CI.getType()->getPrimitiveSizeInBits()&&
5585 A->getType()->getPrimitiveSizeInBits() ==
5586 CI.getType()->getPrimitiveSizeInBits()) {
5587 assert(CSrc->getType() != Type::ULongTy &&
5588 "Cannot have type bigger than ulong!");
5589 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
5590 Constant *AndOp = ConstantInt::get(A->getType()->getUnsignedVersion(),
5592 AndOp = ConstantExpr::getCast(AndOp, A->getType());
5593 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
5594 if (And->getType() != CI.getType()) {
5595 And->setName(CSrc->getName()+".mask");
5596 InsertNewInstBefore(And, CI);
5597 And = new CastInst(And, CI.getType());
5603 // If this is a cast to bool, turn it into the appropriate setne instruction.
5604 if (CI.getType() == Type::BoolTy)
5605 return BinaryOperator::createSetNE(CI.getOperand(0),
5606 Constant::getNullValue(CI.getOperand(0)->getType()));
5608 // See if we can simplify any instructions used by the LHS whose sole
5609 // purpose is to compute bits we don't care about.
5610 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral()) {
5611 uint64_t KnownZero, KnownOne;
5612 if (SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask(),
5613 KnownZero, KnownOne))
5617 // If casting the result of a getelementptr instruction with no offset, turn
5618 // this into a cast of the original pointer!
5620 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5621 bool AllZeroOperands = true;
5622 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5623 if (!isa<Constant>(GEP->getOperand(i)) ||
5624 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5625 AllZeroOperands = false;
5628 if (AllZeroOperands) {
5629 CI.setOperand(0, GEP->getOperand(0));
5634 // If we are casting a malloc or alloca to a pointer to a type of the same
5635 // size, rewrite the allocation instruction to allocate the "right" type.
5637 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5638 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5641 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5642 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5644 if (isa<PHINode>(Src))
5645 if (Instruction *NV = FoldOpIntoPhi(CI))
5648 // If the source and destination are pointers, and this cast is equivalent to
5649 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
5650 // This can enhance SROA and other transforms that want type-safe pointers.
5651 if (const PointerType *DstPTy = dyn_cast<PointerType>(CI.getType()))
5652 if (const PointerType *SrcPTy = dyn_cast<PointerType>(Src->getType())) {
5653 const Type *DstTy = DstPTy->getElementType();
5654 const Type *SrcTy = SrcPTy->getElementType();
5656 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
5657 unsigned NumZeros = 0;
5658 while (SrcTy != DstTy &&
5659 isa<CompositeType>(SrcTy) && !isa<PointerType>(SrcTy) &&
5660 SrcTy->getNumContainedTypes() /* not "{}" */) {
5661 SrcTy = cast<CompositeType>(SrcTy)->getTypeAtIndex(ZeroUInt);
5665 // If we found a path from the src to dest, create the getelementptr now.
5666 if (SrcTy == DstTy) {
5667 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
5668 return new GetElementPtrInst(Src, Idxs);
5672 // If the source value is an instruction with only this use, we can attempt to
5673 // propagate the cast into the instruction. Also, only handle integral types
5675 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
5676 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
5677 CI.getType()->isInteger()) { // Don't mess with casts to bool here
5679 int NumCastsRemoved = 0;
5680 if (CanEvaluateInDifferentType(SrcI, CI.getType(), NumCastsRemoved)) {
5681 // If this cast is a truncate, evaluting in a different type always
5682 // eliminates the cast, so it is always a win. If this is a noop-cast
5683 // this just removes a noop cast which isn't pointful, but simplifies
5684 // the code. If this is a zero-extension, we need to do an AND to
5685 // maintain the clear top-part of the computation, so we require that
5686 // the input have eliminated at least one cast. If this is a sign
5687 // extension, we insert two new casts (to do the extension) so we
5688 // require that two casts have been eliminated.
5690 switch (getCastType(Src->getType(), CI.getType())) {
5691 default: assert(0 && "Unknown cast type!");
5697 DoXForm = NumCastsRemoved >= 1;
5700 DoXForm = NumCastsRemoved >= 2;
5705 Value *Res = EvaluateInDifferentType(SrcI, CI.getType());
5706 assert(Res->getType() == CI.getType());
5707 switch (getCastType(Src->getType(), CI.getType())) {
5708 default: assert(0 && "Unknown cast type!");
5711 // Just replace this cast with the result.
5712 return ReplaceInstUsesWith(CI, Res);
5714 // We need to emit an AND to clear the high bits.
5715 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5716 unsigned DestBitSize = CI.getType()->getPrimitiveSizeInBits();
5717 assert(SrcBitSize < DestBitSize && "Not a zext?");
5719 ConstantInt::get(Type::ULongTy, (1ULL << SrcBitSize)-1);
5720 C = ConstantExpr::getCast(C, CI.getType());
5721 return BinaryOperator::createAnd(Res, C);
5724 // We need to emit a cast to truncate, then a cast to sext.
5725 return new CastInst(InsertCastBefore(Res, Src->getType(), CI),
5731 const Type *DestTy = CI.getType();
5732 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
5733 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5735 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
5736 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
5738 switch (SrcI->getOpcode()) {
5739 case Instruction::Add:
5740 case Instruction::Mul:
5741 case Instruction::And:
5742 case Instruction::Or:
5743 case Instruction::Xor:
5744 // If we are discarding information, or just changing the sign, rewrite.
5745 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
5746 // Don't insert two casts if they cannot be eliminated. We allow two
5747 // casts to be inserted if the sizes are the same. This could only be
5748 // converting signedness, which is a noop.
5749 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
5750 !ValueRequiresCast(Op0, DestTy, TD)) {
5751 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5752 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5753 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
5754 ->getOpcode(), Op0c, Op1c);
5758 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
5759 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
5760 Op1 == ConstantBool::getTrue() &&
5761 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
5762 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
5763 return BinaryOperator::createXor(New,
5764 ConstantInt::get(CI.getType(), 1));
5767 case Instruction::SDiv:
5768 case Instruction::UDiv:
5769 // If we are just changing the sign, rewrite.
5770 if (DestBitSize == SrcBitSize) {
5771 // Don't insert two casts if they cannot be eliminated. We allow two
5772 // casts to be inserted if the sizes are the same. This could only be
5773 // converting signedness, which is a noop.
5774 if (!ValueRequiresCast(Op1, DestTy,TD) ||
5775 !ValueRequiresCast(Op0, DestTy, TD)) {
5776 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5777 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
5778 return BinaryOperator::create(
5779 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
5784 case Instruction::Shl:
5785 // Allow changing the sign of the source operand. Do not allow changing
5786 // the size of the shift, UNLESS the shift amount is a constant. We
5787 // mush not change variable sized shifts to a smaller size, because it
5788 // is undefined to shift more bits out than exist in the value.
5789 if (DestBitSize == SrcBitSize ||
5790 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
5791 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
5792 return new ShiftInst(Instruction::Shl, Op0c, Op1);
5795 case Instruction::Shr:
5796 // If this is a signed shr, and if all bits shifted in are about to be
5797 // truncated off, turn it into an unsigned shr to allow greater
5799 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
5800 isa<ConstantInt>(Op1)) {
5801 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
5802 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
5803 // Convert to unsigned.
5804 Value *N1 = InsertOperandCastBefore(Op0,
5805 Op0->getType()->getUnsignedVersion(), &CI);
5806 // Insert the new shift, which is now unsigned.
5807 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
5808 Op1, Src->getName()), CI);
5809 return new CastInst(N1, CI.getType());
5814 case Instruction::SetEQ:
5815 case Instruction::SetNE:
5816 // We if we are just checking for a seteq of a single bit and casting it
5817 // to an integer. If so, shift the bit to the appropriate place then
5818 // cast to integer to avoid the comparison.
5819 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
5820 uint64_t Op1CV = Op1C->getZExtValue();
5821 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
5822 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5823 // cast (X == 1) to int --> X iff X has only the low bit set.
5824 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
5825 // cast (X != 0) to int --> X iff X has only the low bit set.
5826 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
5827 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
5828 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5829 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
5830 // If Op1C some other power of two, convert:
5831 uint64_t KnownZero, KnownOne;
5832 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
5833 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
5835 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly one possible 1?
5836 bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
5837 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
5838 // (X&4) == 2 --> false
5839 // (X&4) != 2 --> true
5840 Constant *Res = ConstantBool::get(isSetNE);
5841 Res = ConstantExpr::getCast(Res, CI.getType());
5842 return ReplaceInstUsesWith(CI, Res);
5845 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
5848 // Perform an unsigned shr by shiftamt. Convert input to
5849 // unsigned if it is signed.
5850 if (In->getType()->isSigned())
5851 In = InsertCastBefore(
5852 In, In->getType()->getUnsignedVersion(), CI);
5853 // Insert the shift to put the result in the low bit.
5854 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
5855 ConstantInt::get(Type::UByteTy, ShiftAmt),
5856 In->getName()+".lobit"), CI);
5859 if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
5860 Constant *One = ConstantInt::get(In->getType(), 1);
5861 In = BinaryOperator::createXor(In, One, "tmp");
5862 InsertNewInstBefore(cast<Instruction>(In), CI);
5865 if (CI.getType() == In->getType())
5866 return ReplaceInstUsesWith(CI, In);
5868 return new CastInst(In, CI.getType());
5876 if (SrcI->hasOneUse()) {
5877 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(SrcI)) {
5878 // Okay, we have (cast (shuffle ..)). We know this cast is a bitconvert
5879 // because the inputs are known to be a vector. Check to see if this is
5880 // a cast to a vector with the same # elts.
5881 if (isa<PackedType>(CI.getType()) &&
5882 cast<PackedType>(CI.getType())->getNumElements() ==
5883 SVI->getType()->getNumElements()) {
5885 // If either of the operands is a cast from CI.getType(), then
5886 // evaluating the shuffle in the casted destination's type will allow
5887 // us to eliminate at least one cast.
5888 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
5889 Tmp->getOperand(0)->getType() == CI.getType()) ||
5890 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
5891 Tmp->getOperand(0)->getType() == CI.getType())) {
5892 Value *LHS = InsertOperandCastBefore(SVI->getOperand(0),
5894 Value *RHS = InsertOperandCastBefore(SVI->getOperand(1),
5896 // Return a new shuffle vector. Use the same element ID's, as we
5897 // know the vector types match #elts.
5898 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
5908 /// GetSelectFoldableOperands - We want to turn code that looks like this:
5910 /// %D = select %cond, %C, %A
5912 /// %C = select %cond, %B, 0
5915 /// Assuming that the specified instruction is an operand to the select, return
5916 /// a bitmask indicating which operands of this instruction are foldable if they
5917 /// equal the other incoming value of the select.
5919 static unsigned GetSelectFoldableOperands(Instruction *I) {
5920 switch (I->getOpcode()) {
5921 case Instruction::Add:
5922 case Instruction::Mul:
5923 case Instruction::And:
5924 case Instruction::Or:
5925 case Instruction::Xor:
5926 return 3; // Can fold through either operand.
5927 case Instruction::Sub: // Can only fold on the amount subtracted.
5928 case Instruction::Shl: // Can only fold on the shift amount.
5929 case Instruction::Shr:
5932 return 0; // Cannot fold
5936 /// GetSelectFoldableConstant - For the same transformation as the previous
5937 /// function, return the identity constant that goes into the select.
5938 static Constant *GetSelectFoldableConstant(Instruction *I) {
5939 switch (I->getOpcode()) {
5940 default: assert(0 && "This cannot happen!"); abort();
5941 case Instruction::Add:
5942 case Instruction::Sub:
5943 case Instruction::Or:
5944 case Instruction::Xor:
5945 return Constant::getNullValue(I->getType());
5946 case Instruction::Shl:
5947 case Instruction::Shr:
5948 return Constant::getNullValue(Type::UByteTy);
5949 case Instruction::And:
5950 return ConstantInt::getAllOnesValue(I->getType());
5951 case Instruction::Mul:
5952 return ConstantInt::get(I->getType(), 1);
5956 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
5957 /// have the same opcode and only one use each. Try to simplify this.
5958 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
5960 if (TI->getNumOperands() == 1) {
5961 // If this is a non-volatile load or a cast from the same type,
5963 if (TI->getOpcode() == Instruction::Cast) {
5964 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
5967 return 0; // unknown unary op.
5970 // Fold this by inserting a select from the input values.
5971 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
5972 FI->getOperand(0), SI.getName()+".v");
5973 InsertNewInstBefore(NewSI, SI);
5974 return new CastInst(NewSI, TI->getType());
5977 // Only handle binary operators here.
5978 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
5981 // Figure out if the operations have any operands in common.
5982 Value *MatchOp, *OtherOpT, *OtherOpF;
5984 if (TI->getOperand(0) == FI->getOperand(0)) {
5985 MatchOp = TI->getOperand(0);
5986 OtherOpT = TI->getOperand(1);
5987 OtherOpF = FI->getOperand(1);
5988 MatchIsOpZero = true;
5989 } else if (TI->getOperand(1) == FI->getOperand(1)) {
5990 MatchOp = TI->getOperand(1);
5991 OtherOpT = TI->getOperand(0);
5992 OtherOpF = FI->getOperand(0);
5993 MatchIsOpZero = false;
5994 } else if (!TI->isCommutative()) {
5996 } else if (TI->getOperand(0) == FI->getOperand(1)) {
5997 MatchOp = TI->getOperand(0);
5998 OtherOpT = TI->getOperand(1);
5999 OtherOpF = FI->getOperand(0);
6000 MatchIsOpZero = true;
6001 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6002 MatchOp = TI->getOperand(1);
6003 OtherOpT = TI->getOperand(0);
6004 OtherOpF = FI->getOperand(1);
6005 MatchIsOpZero = true;
6010 // If we reach here, they do have operations in common.
6011 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6012 OtherOpF, SI.getName()+".v");
6013 InsertNewInstBefore(NewSI, SI);
6015 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6017 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6019 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6022 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6024 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6028 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6029 Value *CondVal = SI.getCondition();
6030 Value *TrueVal = SI.getTrueValue();
6031 Value *FalseVal = SI.getFalseValue();
6033 // select true, X, Y -> X
6034 // select false, X, Y -> Y
6035 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
6036 return ReplaceInstUsesWith(SI, C->getValue() ? TrueVal : FalseVal);
6038 // select C, X, X -> X
6039 if (TrueVal == FalseVal)
6040 return ReplaceInstUsesWith(SI, TrueVal);
6042 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6043 return ReplaceInstUsesWith(SI, FalseVal);
6044 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6045 return ReplaceInstUsesWith(SI, TrueVal);
6046 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6047 if (isa<Constant>(TrueVal))
6048 return ReplaceInstUsesWith(SI, TrueVal);
6050 return ReplaceInstUsesWith(SI, FalseVal);
6053 if (SI.getType() == Type::BoolTy)
6054 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
6055 if (C->getValue()) {
6056 // Change: A = select B, true, C --> A = or B, C
6057 return BinaryOperator::createOr(CondVal, FalseVal);
6059 // Change: A = select B, false, C --> A = and !B, C
6061 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6062 "not."+CondVal->getName()), SI);
6063 return BinaryOperator::createAnd(NotCond, FalseVal);
6065 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
6066 if (C->getValue() == false) {
6067 // Change: A = select B, C, false --> A = and B, C
6068 return BinaryOperator::createAnd(CondVal, TrueVal);
6070 // Change: A = select B, C, true --> A = or !B, C
6072 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6073 "not."+CondVal->getName()), SI);
6074 return BinaryOperator::createOr(NotCond, TrueVal);
6078 // Selecting between two integer constants?
6079 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6080 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6081 // select C, 1, 0 -> cast C to int
6082 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6083 return new CastInst(CondVal, SI.getType());
6084 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6085 // select C, 0, 1 -> cast !C to int
6087 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6088 "not."+CondVal->getName()), SI);
6089 return new CastInst(NotCond, SI.getType());
6092 if (SetCondInst *IC = dyn_cast<SetCondInst>(SI.getCondition())) {
6094 // (x <s 0) ? -1 : 0 -> sra x, 31
6095 // (x >u 2147483647) ? -1 : 0 -> sra x, 31
6096 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6097 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6098 bool CanXForm = false;
6099 if (CmpCst->getType()->isSigned())
6100 CanXForm = CmpCst->isNullValue() &&
6101 IC->getOpcode() == Instruction::SetLT;
6103 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6104 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6105 IC->getOpcode() == Instruction::SetGT;
6109 // The comparison constant and the result are not neccessarily the
6110 // same width. In any case, the first step to do is make sure
6111 // that X is signed.
6112 Value *X = IC->getOperand(0);
6113 if (!X->getType()->isSigned())
6114 X = InsertCastBefore(X, X->getType()->getSignedVersion(), SI);
6116 // Now that X is signed, we have to make the all ones value. Do
6117 // this by inserting a new SRA.
6118 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6119 Constant *ShAmt = ConstantInt::get(Type::UByteTy, Bits-1);
6120 Instruction *SRA = new ShiftInst(Instruction::Shr, X,
6122 InsertNewInstBefore(SRA, SI);
6124 // Finally, convert to the type of the select RHS. If this is
6125 // smaller than the compare value, it will truncate the ones to
6126 // fit. If it is larger, it will sext the ones to fit.
6127 return new CastInst(SRA, SI.getType());
6132 // If one of the constants is zero (we know they can't both be) and we
6133 // have a setcc instruction with zero, and we have an 'and' with the
6134 // non-constant value, eliminate this whole mess. This corresponds to
6135 // cases like this: ((X & 27) ? 27 : 0)
6136 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6137 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6138 cast<Constant>(IC->getOperand(1))->isNullValue())
6139 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6140 if (ICA->getOpcode() == Instruction::And &&
6141 isa<ConstantInt>(ICA->getOperand(1)) &&
6142 (ICA->getOperand(1) == TrueValC ||
6143 ICA->getOperand(1) == FalseValC) &&
6144 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6145 // Okay, now we know that everything is set up, we just don't
6146 // know whether we have a setne or seteq and whether the true or
6147 // false val is the zero.
6148 bool ShouldNotVal = !TrueValC->isNullValue();
6149 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
6152 V = InsertNewInstBefore(BinaryOperator::create(
6153 Instruction::Xor, V, ICA->getOperand(1)), SI);
6154 return ReplaceInstUsesWith(SI, V);
6159 // See if we are selecting two values based on a comparison of the two values.
6160 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
6161 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
6162 // Transform (X == Y) ? X : Y -> Y
6163 if (SCI->getOpcode() == Instruction::SetEQ)
6164 return ReplaceInstUsesWith(SI, FalseVal);
6165 // Transform (X != Y) ? X : Y -> X
6166 if (SCI->getOpcode() == Instruction::SetNE)
6167 return ReplaceInstUsesWith(SI, TrueVal);
6168 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6170 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
6171 // Transform (X == Y) ? Y : X -> X
6172 if (SCI->getOpcode() == Instruction::SetEQ)
6173 return ReplaceInstUsesWith(SI, FalseVal);
6174 // Transform (X != Y) ? Y : X -> Y
6175 if (SCI->getOpcode() == Instruction::SetNE)
6176 return ReplaceInstUsesWith(SI, TrueVal);
6177 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6181 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6182 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6183 if (TI->hasOneUse() && FI->hasOneUse()) {
6184 bool isInverse = false;
6185 Instruction *AddOp = 0, *SubOp = 0;
6187 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6188 if (TI->getOpcode() == FI->getOpcode())
6189 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6192 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6193 // even legal for FP.
6194 if (TI->getOpcode() == Instruction::Sub &&
6195 FI->getOpcode() == Instruction::Add) {
6196 AddOp = FI; SubOp = TI;
6197 } else if (FI->getOpcode() == Instruction::Sub &&
6198 TI->getOpcode() == Instruction::Add) {
6199 AddOp = TI; SubOp = FI;
6203 Value *OtherAddOp = 0;
6204 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6205 OtherAddOp = AddOp->getOperand(1);
6206 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6207 OtherAddOp = AddOp->getOperand(0);
6211 // So at this point we know we have (Y -> OtherAddOp):
6212 // select C, (add X, Y), (sub X, Z)
6213 Value *NegVal; // Compute -Z
6214 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6215 NegVal = ConstantExpr::getNeg(C);
6217 NegVal = InsertNewInstBefore(
6218 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6221 Value *NewTrueOp = OtherAddOp;
6222 Value *NewFalseOp = NegVal;
6224 std::swap(NewTrueOp, NewFalseOp);
6225 Instruction *NewSel =
6226 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6228 NewSel = InsertNewInstBefore(NewSel, SI);
6229 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6234 // See if we can fold the select into one of our operands.
6235 if (SI.getType()->isInteger()) {
6236 // See the comment above GetSelectFoldableOperands for a description of the
6237 // transformation we are doing here.
6238 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6239 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6240 !isa<Constant>(FalseVal))
6241 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6242 unsigned OpToFold = 0;
6243 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6245 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6250 Constant *C = GetSelectFoldableConstant(TVI);
6251 std::string Name = TVI->getName(); TVI->setName("");
6252 Instruction *NewSel =
6253 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6255 InsertNewInstBefore(NewSel, SI);
6256 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6257 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6258 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6259 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6261 assert(0 && "Unknown instruction!!");
6266 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6267 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6268 !isa<Constant>(TrueVal))
6269 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6270 unsigned OpToFold = 0;
6271 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6273 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6278 Constant *C = GetSelectFoldableConstant(FVI);
6279 std::string Name = FVI->getName(); FVI->setName("");
6280 Instruction *NewSel =
6281 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6283 InsertNewInstBefore(NewSel, SI);
6284 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6285 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6286 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6287 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6289 assert(0 && "Unknown instruction!!");
6295 if (BinaryOperator::isNot(CondVal)) {
6296 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6297 SI.setOperand(1, FalseVal);
6298 SI.setOperand(2, TrueVal);
6305 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6306 /// determine, return it, otherwise return 0.
6307 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6308 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6309 unsigned Align = GV->getAlignment();
6310 if (Align == 0 && TD)
6311 Align = TD->getTypeAlignment(GV->getType()->getElementType());
6313 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6314 unsigned Align = AI->getAlignment();
6315 if (Align == 0 && TD) {
6316 if (isa<AllocaInst>(AI))
6317 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6318 else if (isa<MallocInst>(AI)) {
6319 // Malloc returns maximally aligned memory.
6320 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6321 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
6322 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::LongTy));
6326 } else if (isa<CastInst>(V) ||
6327 (isa<ConstantExpr>(V) &&
6328 cast<ConstantExpr>(V)->getOpcode() == Instruction::Cast)) {
6329 User *CI = cast<User>(V);
6330 if (isa<PointerType>(CI->getOperand(0)->getType()))
6331 return GetKnownAlignment(CI->getOperand(0), TD);
6333 } else if (isa<GetElementPtrInst>(V) ||
6334 (isa<ConstantExpr>(V) &&
6335 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6336 User *GEPI = cast<User>(V);
6337 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6338 if (BaseAlignment == 0) return 0;
6340 // If all indexes are zero, it is just the alignment of the base pointer.
6341 bool AllZeroOperands = true;
6342 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6343 if (!isa<Constant>(GEPI->getOperand(i)) ||
6344 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6345 AllZeroOperands = false;
6348 if (AllZeroOperands)
6349 return BaseAlignment;
6351 // Otherwise, if the base alignment is >= the alignment we expect for the
6352 // base pointer type, then we know that the resultant pointer is aligned at
6353 // least as much as its type requires.
6356 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6357 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
6359 const Type *GEPTy = GEPI->getType();
6360 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
6368 /// visitCallInst - CallInst simplification. This mostly only handles folding
6369 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6370 /// the heavy lifting.
6372 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6373 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6374 if (!II) return visitCallSite(&CI);
6376 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6378 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6379 bool Changed = false;
6381 // memmove/cpy/set of zero bytes is a noop.
6382 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6383 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6385 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6386 if (CI->getZExtValue() == 1) {
6387 // Replace the instruction with just byte operations. We would
6388 // transform other cases to loads/stores, but we don't know if
6389 // alignment is sufficient.
6393 // If we have a memmove and the source operation is a constant global,
6394 // then the source and dest pointers can't alias, so we can change this
6395 // into a call to memcpy.
6396 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6397 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6398 if (GVSrc->isConstant()) {
6399 Module *M = CI.getParent()->getParent()->getParent();
6401 if (CI.getCalledFunction()->getFunctionType()->getParamType(3) ==
6403 Name = "llvm.memcpy.i32";
6405 Name = "llvm.memcpy.i64";
6406 Function *MemCpy = M->getOrInsertFunction(Name,
6407 CI.getCalledFunction()->getFunctionType());
6408 CI.setOperand(0, MemCpy);
6413 // If we can determine a pointer alignment that is bigger than currently
6414 // set, update the alignment.
6415 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6416 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6417 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6418 unsigned Align = std::min(Alignment1, Alignment2);
6419 if (MI->getAlignment()->getZExtValue() < Align) {
6420 MI->setAlignment(ConstantInt::get(Type::UIntTy, Align));
6423 } else if (isa<MemSetInst>(MI)) {
6424 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6425 if (MI->getAlignment()->getZExtValue() < Alignment) {
6426 MI->setAlignment(ConstantInt::get(Type::UIntTy, Alignment));
6431 if (Changed) return II;
6433 switch (II->getIntrinsicID()) {
6435 case Intrinsic::ppc_altivec_lvx:
6436 case Intrinsic::ppc_altivec_lvxl:
6437 case Intrinsic::x86_sse_loadu_ps:
6438 case Intrinsic::x86_sse2_loadu_pd:
6439 case Intrinsic::x86_sse2_loadu_dq:
6440 // Turn PPC lvx -> load if the pointer is known aligned.
6441 // Turn X86 loadups -> load if the pointer is known aligned.
6442 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6443 Value *Ptr = InsertCastBefore(II->getOperand(1),
6444 PointerType::get(II->getType()), CI);
6445 return new LoadInst(Ptr);
6448 case Intrinsic::ppc_altivec_stvx:
6449 case Intrinsic::ppc_altivec_stvxl:
6450 // Turn stvx -> store if the pointer is known aligned.
6451 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
6452 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
6453 Value *Ptr = InsertCastBefore(II->getOperand(2), OpPtrTy, CI);
6454 return new StoreInst(II->getOperand(1), Ptr);
6457 case Intrinsic::x86_sse_storeu_ps:
6458 case Intrinsic::x86_sse2_storeu_pd:
6459 case Intrinsic::x86_sse2_storeu_dq:
6460 case Intrinsic::x86_sse2_storel_dq:
6461 // Turn X86 storeu -> store if the pointer is known aligned.
6462 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6463 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
6464 Value *Ptr = InsertCastBefore(II->getOperand(1), OpPtrTy, CI);
6465 return new StoreInst(II->getOperand(2), Ptr);
6469 case Intrinsic::x86_sse_cvttss2si: {
6470 // These intrinsics only demands the 0th element of its input vector. If
6471 // we can simplify the input based on that, do so now.
6473 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
6475 II->setOperand(1, V);
6481 case Intrinsic::ppc_altivec_vperm:
6482 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
6483 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
6484 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
6486 // Check that all of the elements are integer constants or undefs.
6487 bool AllEltsOk = true;
6488 for (unsigned i = 0; i != 16; ++i) {
6489 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
6490 !isa<UndefValue>(Mask->getOperand(i))) {
6497 // Cast the input vectors to byte vectors.
6498 Value *Op0 = InsertCastBefore(II->getOperand(1), Mask->getType(), CI);
6499 Value *Op1 = InsertCastBefore(II->getOperand(2), Mask->getType(), CI);
6500 Value *Result = UndefValue::get(Op0->getType());
6502 // Only extract each element once.
6503 Value *ExtractedElts[32];
6504 memset(ExtractedElts, 0, sizeof(ExtractedElts));
6506 for (unsigned i = 0; i != 16; ++i) {
6507 if (isa<UndefValue>(Mask->getOperand(i)))
6509 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
6510 Idx &= 31; // Match the hardware behavior.
6512 if (ExtractedElts[Idx] == 0) {
6514 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
6515 InsertNewInstBefore(Elt, CI);
6516 ExtractedElts[Idx] = Elt;
6519 // Insert this value into the result vector.
6520 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
6521 InsertNewInstBefore(cast<Instruction>(Result), CI);
6523 return new CastInst(Result, CI.getType());
6528 case Intrinsic::stackrestore: {
6529 // If the save is right next to the restore, remove the restore. This can
6530 // happen when variable allocas are DCE'd.
6531 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
6532 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
6533 BasicBlock::iterator BI = SS;
6535 return EraseInstFromFunction(CI);
6539 // If the stack restore is in a return/unwind block and if there are no
6540 // allocas or calls between the restore and the return, nuke the restore.
6541 TerminatorInst *TI = II->getParent()->getTerminator();
6542 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
6543 BasicBlock::iterator BI = II;
6544 bool CannotRemove = false;
6545 for (++BI; &*BI != TI; ++BI) {
6546 if (isa<AllocaInst>(BI) ||
6547 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
6548 CannotRemove = true;
6553 return EraseInstFromFunction(CI);
6560 return visitCallSite(II);
6563 // InvokeInst simplification
6565 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
6566 return visitCallSite(&II);
6569 // visitCallSite - Improvements for call and invoke instructions.
6571 Instruction *InstCombiner::visitCallSite(CallSite CS) {
6572 bool Changed = false;
6574 // If the callee is a constexpr cast of a function, attempt to move the cast
6575 // to the arguments of the call/invoke.
6576 if (transformConstExprCastCall(CS)) return 0;
6578 Value *Callee = CS.getCalledValue();
6580 if (Function *CalleeF = dyn_cast<Function>(Callee))
6581 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
6582 Instruction *OldCall = CS.getInstruction();
6583 // If the call and callee calling conventions don't match, this call must
6584 // be unreachable, as the call is undefined.
6585 new StoreInst(ConstantBool::getTrue(),
6586 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
6587 if (!OldCall->use_empty())
6588 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
6589 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
6590 return EraseInstFromFunction(*OldCall);
6594 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
6595 // This instruction is not reachable, just remove it. We insert a store to
6596 // undef so that we know that this code is not reachable, despite the fact
6597 // that we can't modify the CFG here.
6598 new StoreInst(ConstantBool::getTrue(),
6599 UndefValue::get(PointerType::get(Type::BoolTy)),
6600 CS.getInstruction());
6602 if (!CS.getInstruction()->use_empty())
6603 CS.getInstruction()->
6604 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
6606 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
6607 // Don't break the CFG, insert a dummy cond branch.
6608 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
6609 ConstantBool::getTrue(), II);
6611 return EraseInstFromFunction(*CS.getInstruction());
6614 const PointerType *PTy = cast<PointerType>(Callee->getType());
6615 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
6616 if (FTy->isVarArg()) {
6617 // See if we can optimize any arguments passed through the varargs area of
6619 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
6620 E = CS.arg_end(); I != E; ++I)
6621 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
6622 // If this cast does not effect the value passed through the varargs
6623 // area, we can eliminate the use of the cast.
6624 Value *Op = CI->getOperand(0);
6625 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
6632 return Changed ? CS.getInstruction() : 0;
6635 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
6636 // attempt to move the cast to the arguments of the call/invoke.
6638 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
6639 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
6640 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
6641 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
6643 Function *Callee = cast<Function>(CE->getOperand(0));
6644 Instruction *Caller = CS.getInstruction();
6646 // Okay, this is a cast from a function to a different type. Unless doing so
6647 // would cause a type conversion of one of our arguments, change this call to
6648 // be a direct call with arguments casted to the appropriate types.
6650 const FunctionType *FT = Callee->getFunctionType();
6651 const Type *OldRetTy = Caller->getType();
6653 // Check to see if we are changing the return type...
6654 if (OldRetTy != FT->getReturnType()) {
6655 if (Callee->isExternal() &&
6656 !(OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) ||
6657 (isa<PointerType>(FT->getReturnType()) &&
6658 TD->getIntPtrType()->isLosslesslyConvertibleTo(OldRetTy)))
6659 && !Caller->use_empty())
6660 return false; // Cannot transform this return value...
6662 // If the callsite is an invoke instruction, and the return value is used by
6663 // a PHI node in a successor, we cannot change the return type of the call
6664 // because there is no place to put the cast instruction (without breaking
6665 // the critical edge). Bail out in this case.
6666 if (!Caller->use_empty())
6667 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
6668 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
6670 if (PHINode *PN = dyn_cast<PHINode>(*UI))
6671 if (PN->getParent() == II->getNormalDest() ||
6672 PN->getParent() == II->getUnwindDest())
6676 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
6677 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
6679 CallSite::arg_iterator AI = CS.arg_begin();
6680 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
6681 const Type *ParamTy = FT->getParamType(i);
6682 const Type *ActTy = (*AI)->getType();
6683 ConstantInt* c = dyn_cast<ConstantInt>(*AI);
6684 //Either we can cast directly, or we can upconvert the argument
6685 bool isConvertible = ActTy->isLosslesslyConvertibleTo(ParamTy) ||
6686 (ParamTy->isIntegral() && ActTy->isIntegral() &&
6687 ParamTy->isSigned() == ActTy->isSigned() &&
6688 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
6689 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
6690 c->getSExtValue() > 0);
6691 if (Callee->isExternal() && !isConvertible) return false;
6694 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
6695 Callee->isExternal())
6696 return false; // Do not delete arguments unless we have a function body...
6698 // Okay, we decided that this is a safe thing to do: go ahead and start
6699 // inserting cast instructions as necessary...
6700 std::vector<Value*> Args;
6701 Args.reserve(NumActualArgs);
6703 AI = CS.arg_begin();
6704 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
6705 const Type *ParamTy = FT->getParamType(i);
6706 if ((*AI)->getType() == ParamTy) {
6707 Args.push_back(*AI);
6709 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
6714 // If the function takes more arguments than the call was taking, add them
6716 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
6717 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
6719 // If we are removing arguments to the function, emit an obnoxious warning...
6720 if (FT->getNumParams() < NumActualArgs)
6721 if (!FT->isVarArg()) {
6722 std::cerr << "WARNING: While resolving call to function '"
6723 << Callee->getName() << "' arguments were dropped!\n";
6725 // Add all of the arguments in their promoted form to the arg list...
6726 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
6727 const Type *PTy = getPromotedType((*AI)->getType());
6728 if (PTy != (*AI)->getType()) {
6729 // Must promote to pass through va_arg area!
6730 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
6731 InsertNewInstBefore(Cast, *Caller);
6732 Args.push_back(Cast);
6734 Args.push_back(*AI);
6739 if (FT->getReturnType() == Type::VoidTy)
6740 Caller->setName(""); // Void type should not have a name...
6743 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6744 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
6745 Args, Caller->getName(), Caller);
6746 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
6748 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
6749 if (cast<CallInst>(Caller)->isTailCall())
6750 cast<CallInst>(NC)->setTailCall();
6751 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
6754 // Insert a cast of the return type as necessary...
6756 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
6757 if (NV->getType() != Type::VoidTy) {
6758 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
6760 // If this is an invoke instruction, we should insert it after the first
6761 // non-phi, instruction in the normal successor block.
6762 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6763 BasicBlock::iterator I = II->getNormalDest()->begin();
6764 while (isa<PHINode>(I)) ++I;
6765 InsertNewInstBefore(NC, *I);
6767 // Otherwise, it's a call, just insert cast right after the call instr
6768 InsertNewInstBefore(NC, *Caller);
6770 AddUsersToWorkList(*Caller);
6772 NV = UndefValue::get(Caller->getType());
6776 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
6777 Caller->replaceAllUsesWith(NV);
6778 Caller->getParent()->getInstList().erase(Caller);
6779 removeFromWorkList(Caller);
6783 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
6784 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
6785 /// and a single binop.
6786 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
6787 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6788 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
6789 isa<GetElementPtrInst>(FirstInst));
6790 unsigned Opc = FirstInst->getOpcode();
6791 const Type *LHSType = FirstInst->getOperand(0)->getType();
6792 const Type *RHSType = FirstInst->getOperand(1)->getType();
6794 // Scan to see if all operands are the same opcode, all have one use, and all
6795 // kill their operands (i.e. the operands have one use).
6796 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
6797 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
6798 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
6799 // Verify type of the LHS matches so we don't fold setcc's of different
6800 // types or GEP's with different index types.
6801 I->getOperand(0)->getType() != LHSType ||
6802 I->getOperand(1)->getType() != RHSType)
6806 // Otherwise, this is safe and profitable to transform. Create two phi nodes.
6807 PHINode *NewLHS = new PHINode(FirstInst->getOperand(0)->getType(),
6808 FirstInst->getOperand(0)->getName()+".pn");
6809 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
6810 PHINode *NewRHS = new PHINode(FirstInst->getOperand(1)->getType(),
6811 FirstInst->getOperand(1)->getName()+".pn");
6812 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
6814 Value *InLHS = FirstInst->getOperand(0);
6815 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
6816 Value *InRHS = FirstInst->getOperand(1);
6817 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
6819 // Add all operands to the new PHsI.
6820 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6821 Value *NewInLHS = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
6822 Value *NewInRHS = cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
6823 if (NewInLHS != InLHS) InLHS = 0;
6824 if (NewInRHS != InRHS) InRHS = 0;
6825 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
6826 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
6831 // The new PHI unions all of the same values together. This is really
6832 // common, so we handle it intelligently here for compile-time speed.
6836 InsertNewInstBefore(NewLHS, PN);
6841 // The new PHI unions all of the same values together. This is really
6842 // common, so we handle it intelligently here for compile-time speed.
6846 InsertNewInstBefore(NewRHS, PN);
6850 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6851 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
6852 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
6853 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
6855 assert(isa<GetElementPtrInst>(FirstInst));
6856 return new GetElementPtrInst(LHSVal, RHSVal);
6860 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
6861 /// of the block that defines it. This means that it must be obvious the value
6862 /// of the load is not changed from the point of the load to the end of the
6864 static bool isSafeToSinkLoad(LoadInst *L) {
6865 BasicBlock::iterator BBI = L, E = L->getParent()->end();
6867 for (++BBI; BBI != E; ++BBI)
6868 if (BBI->mayWriteToMemory())
6874 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
6875 // operator and they all are only used by the PHI, PHI together their
6876 // inputs, and do the operation once, to the result of the PHI.
6877 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
6878 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6880 // Scan the instruction, looking for input operations that can be folded away.
6881 // If all input operands to the phi are the same instruction (e.g. a cast from
6882 // the same type or "+42") we can pull the operation through the PHI, reducing
6883 // code size and simplifying code.
6884 Constant *ConstantOp = 0;
6885 const Type *CastSrcTy = 0;
6886 bool isVolatile = false;
6887 if (isa<CastInst>(FirstInst)) {
6888 CastSrcTy = FirstInst->getOperand(0)->getType();
6889 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
6890 // Can fold binop or shift here if the RHS is a constant, otherwise call
6891 // FoldPHIArgBinOpIntoPHI.
6892 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
6893 if (ConstantOp == 0)
6894 return FoldPHIArgBinOpIntoPHI(PN);
6895 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
6896 isVolatile = LI->isVolatile();
6897 // We can't sink the load if the loaded value could be modified between the
6898 // load and the PHI.
6899 if (LI->getParent() != PN.getIncomingBlock(0) ||
6900 !isSafeToSinkLoad(LI))
6902 } else if (isa<GetElementPtrInst>(FirstInst)) {
6903 if (FirstInst->getNumOperands() == 2)
6904 return FoldPHIArgBinOpIntoPHI(PN);
6905 // Can't handle general GEPs yet.
6908 return 0; // Cannot fold this operation.
6911 // Check to see if all arguments are the same operation.
6912 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6913 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
6914 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
6915 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
6918 if (I->getOperand(0)->getType() != CastSrcTy)
6919 return 0; // Cast operation must match.
6920 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6921 // We can't sink the load if the loaded value could be modified between the
6922 // load and the PHI.
6923 if (LI->isVolatile() != isVolatile ||
6924 LI->getParent() != PN.getIncomingBlock(i) ||
6925 !isSafeToSinkLoad(LI))
6927 } else if (I->getOperand(1) != ConstantOp) {
6932 // Okay, they are all the same operation. Create a new PHI node of the
6933 // correct type, and PHI together all of the LHS's of the instructions.
6934 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
6935 PN.getName()+".in");
6936 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
6938 Value *InVal = FirstInst->getOperand(0);
6939 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
6941 // Add all operands to the new PHI.
6942 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6943 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
6944 if (NewInVal != InVal)
6946 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
6951 // The new PHI unions all of the same values together. This is really
6952 // common, so we handle it intelligently here for compile-time speed.
6956 InsertNewInstBefore(NewPN, PN);
6960 // Insert and return the new operation.
6961 if (isa<CastInst>(FirstInst))
6962 return new CastInst(PhiVal, PN.getType());
6963 else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst))
6964 return new LoadInst(PhiVal, "", isVolatile);
6965 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6966 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
6968 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
6969 PhiVal, ConstantOp);
6972 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
6974 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
6975 if (PN->use_empty()) return true;
6976 if (!PN->hasOneUse()) return false;
6978 // Remember this node, and if we find the cycle, return.
6979 if (!PotentiallyDeadPHIs.insert(PN).second)
6982 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
6983 return DeadPHICycle(PU, PotentiallyDeadPHIs);
6988 // PHINode simplification
6990 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
6991 // If LCSSA is around, don't mess with Phi nodes
6992 if (mustPreserveAnalysisID(LCSSAID)) return 0;
6994 if (Value *V = PN.hasConstantValue())
6995 return ReplaceInstUsesWith(PN, V);
6997 // If the only user of this instruction is a cast instruction, and all of the
6998 // incoming values are constants, change this PHI to merge together the casted
7001 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
7002 if (CI->getType() != PN.getType()) { // noop casts will be folded
7003 bool AllConstant = true;
7004 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
7005 if (!isa<Constant>(PN.getIncomingValue(i))) {
7006 AllConstant = false;
7010 // Make a new PHI with all casted values.
7011 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
7012 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
7013 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
7014 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
7015 PN.getIncomingBlock(i));
7018 // Update the cast instruction.
7019 CI->setOperand(0, New);
7020 WorkList.push_back(CI); // revisit the cast instruction to fold.
7021 WorkList.push_back(New); // Make sure to revisit the new Phi
7022 return &PN; // PN is now dead!
7026 // If all PHI operands are the same operation, pull them through the PHI,
7027 // reducing code size.
7028 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7029 PN.getIncomingValue(0)->hasOneUse())
7030 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7033 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7034 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7035 // PHI)... break the cycle.
7037 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7038 std::set<PHINode*> PotentiallyDeadPHIs;
7039 PotentiallyDeadPHIs.insert(&PN);
7040 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7041 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7047 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
7048 Instruction *InsertPoint,
7050 unsigned PS = IC->getTargetData().getPointerSize();
7051 const Type *VTy = V->getType();
7052 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
7053 // We must insert a cast to ensure we sign-extend.
7054 V = IC->InsertCastBefore(V, VTy->getSignedVersion(), *InsertPoint);
7055 return IC->InsertCastBefore(V, DTy, *InsertPoint);
7059 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7060 Value *PtrOp = GEP.getOperand(0);
7061 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7062 // If so, eliminate the noop.
7063 if (GEP.getNumOperands() == 1)
7064 return ReplaceInstUsesWith(GEP, PtrOp);
7066 if (isa<UndefValue>(GEP.getOperand(0)))
7067 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7069 bool HasZeroPointerIndex = false;
7070 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7071 HasZeroPointerIndex = C->isNullValue();
7073 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7074 return ReplaceInstUsesWith(GEP, PtrOp);
7076 // Eliminate unneeded casts for indices.
7077 bool MadeChange = false;
7078 gep_type_iterator GTI = gep_type_begin(GEP);
7079 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7080 if (isa<SequentialType>(*GTI)) {
7081 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7082 Value *Src = CI->getOperand(0);
7083 const Type *SrcTy = Src->getType();
7084 const Type *DestTy = CI->getType();
7085 if (Src->getType()->isInteger()) {
7086 if (SrcTy->getPrimitiveSizeInBits() ==
7087 DestTy->getPrimitiveSizeInBits()) {
7088 // We can always eliminate a cast from ulong or long to the other.
7089 // We can always eliminate a cast from uint to int or the other on
7090 // 32-bit pointer platforms.
7091 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7093 GEP.setOperand(i, Src);
7095 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
7096 SrcTy->getPrimitiveSize() == 4) {
7097 // We can always eliminate a cast from int to [u]long. We can
7098 // eliminate a cast from uint to [u]long iff the target is a 32-bit
7100 if (SrcTy->isSigned() ||
7101 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7103 GEP.setOperand(i, Src);
7108 // If we are using a wider index than needed for this platform, shrink it
7109 // to what we need. If the incoming value needs a cast instruction,
7110 // insert it. This explicit cast can make subsequent optimizations more
7112 Value *Op = GEP.getOperand(i);
7113 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
7114 if (Constant *C = dyn_cast<Constant>(Op)) {
7115 GEP.setOperand(i, ConstantExpr::getCast(C,
7116 TD->getIntPtrType()->getSignedVersion()));
7119 Op = InsertCastBefore(Op, TD->getIntPtrType(), GEP);
7120 GEP.setOperand(i, Op);
7124 // If this is a constant idx, make sure to canonicalize it to be a signed
7125 // operand, otherwise CSE and other optimizations are pessimized.
7126 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op))
7127 if (CUI->getType()->isUnsigned()) {
7129 ConstantExpr::getCast(CUI, CUI->getType()->getSignedVersion()));
7133 if (MadeChange) return &GEP;
7135 // Combine Indices - If the source pointer to this getelementptr instruction
7136 // is a getelementptr instruction, combine the indices of the two
7137 // getelementptr instructions into a single instruction.
7139 std::vector<Value*> SrcGEPOperands;
7140 if (User *Src = dyn_castGetElementPtr(PtrOp))
7141 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7143 if (!SrcGEPOperands.empty()) {
7144 // Note that if our source is a gep chain itself that we wait for that
7145 // chain to be resolved before we perform this transformation. This
7146 // avoids us creating a TON of code in some cases.
7148 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7149 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7150 return 0; // Wait until our source is folded to completion.
7152 std::vector<Value *> Indices;
7154 // Find out whether the last index in the source GEP is a sequential idx.
7155 bool EndsWithSequential = false;
7156 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7157 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7158 EndsWithSequential = !isa<StructType>(*I);
7160 // Can we combine the two pointer arithmetics offsets?
7161 if (EndsWithSequential) {
7162 // Replace: gep (gep %P, long B), long A, ...
7163 // With: T = long A+B; gep %P, T, ...
7165 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7166 if (SO1 == Constant::getNullValue(SO1->getType())) {
7168 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7171 // If they aren't the same type, convert both to an integer of the
7172 // target's pointer size.
7173 if (SO1->getType() != GO1->getType()) {
7174 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7175 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
7176 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7177 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
7179 unsigned PS = TD->getPointerSize();
7180 if (SO1->getType()->getPrimitiveSize() == PS) {
7181 // Convert GO1 to SO1's type.
7182 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
7184 } else if (GO1->getType()->getPrimitiveSize() == PS) {
7185 // Convert SO1 to GO1's type.
7186 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
7188 const Type *PT = TD->getIntPtrType();
7189 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
7190 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
7194 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7195 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7197 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7198 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7202 // Recycle the GEP we already have if possible.
7203 if (SrcGEPOperands.size() == 2) {
7204 GEP.setOperand(0, SrcGEPOperands[0]);
7205 GEP.setOperand(1, Sum);
7208 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7209 SrcGEPOperands.end()-1);
7210 Indices.push_back(Sum);
7211 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7213 } else if (isa<Constant>(*GEP.idx_begin()) &&
7214 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7215 SrcGEPOperands.size() != 1) {
7216 // Otherwise we can do the fold if the first index of the GEP is a zero
7217 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7218 SrcGEPOperands.end());
7219 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7222 if (!Indices.empty())
7223 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7225 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7226 // GEP of global variable. If all of the indices for this GEP are
7227 // constants, we can promote this to a constexpr instead of an instruction.
7229 // Scan for nonconstants...
7230 std::vector<Constant*> Indices;
7231 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7232 for (; I != E && isa<Constant>(*I); ++I)
7233 Indices.push_back(cast<Constant>(*I));
7235 if (I == E) { // If they are all constants...
7236 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7238 // Replace all uses of the GEP with the new constexpr...
7239 return ReplaceInstUsesWith(GEP, CE);
7241 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
7242 if (!isa<PointerType>(X->getType())) {
7243 // Not interesting. Source pointer must be a cast from pointer.
7244 } else if (HasZeroPointerIndex) {
7245 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7246 // into : GEP [10 x ubyte]* X, long 0, ...
7248 // This occurs when the program declares an array extern like "int X[];"
7250 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7251 const PointerType *XTy = cast<PointerType>(X->getType());
7252 if (const ArrayType *XATy =
7253 dyn_cast<ArrayType>(XTy->getElementType()))
7254 if (const ArrayType *CATy =
7255 dyn_cast<ArrayType>(CPTy->getElementType()))
7256 if (CATy->getElementType() == XATy->getElementType()) {
7257 // At this point, we know that the cast source type is a pointer
7258 // to an array of the same type as the destination pointer
7259 // array. Because the array type is never stepped over (there
7260 // is a leading zero) we can fold the cast into this GEP.
7261 GEP.setOperand(0, X);
7264 } else if (GEP.getNumOperands() == 2) {
7265 // Transform things like:
7266 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7267 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7268 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7269 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7270 if (isa<ArrayType>(SrcElTy) &&
7271 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7272 TD->getTypeSize(ResElTy)) {
7273 Value *V = InsertNewInstBefore(
7274 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7275 GEP.getOperand(1), GEP.getName()), GEP);
7276 return new CastInst(V, GEP.getType());
7279 // Transform things like:
7280 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7281 // (where tmp = 8*tmp2) into:
7282 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7284 if (isa<ArrayType>(SrcElTy) &&
7285 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
7286 uint64_t ArrayEltSize =
7287 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7289 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7290 // allow either a mul, shift, or constant here.
7292 ConstantInt *Scale = 0;
7293 if (ArrayEltSize == 1) {
7294 NewIdx = GEP.getOperand(1);
7295 Scale = ConstantInt::get(NewIdx->getType(), 1);
7296 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7297 NewIdx = ConstantInt::get(CI->getType(), 1);
7299 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7300 if (Inst->getOpcode() == Instruction::Shl &&
7301 isa<ConstantInt>(Inst->getOperand(1))) {
7303 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7304 if (Inst->getType()->isSigned())
7305 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7307 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7308 NewIdx = Inst->getOperand(0);
7309 } else if (Inst->getOpcode() == Instruction::Mul &&
7310 isa<ConstantInt>(Inst->getOperand(1))) {
7311 Scale = cast<ConstantInt>(Inst->getOperand(1));
7312 NewIdx = Inst->getOperand(0);
7316 // If the index will be to exactly the right offset with the scale taken
7317 // out, perform the transformation.
7318 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7319 if (ConstantInt *C = dyn_cast<ConstantInt>(Scale))
7320 Scale = ConstantInt::get(Scale->getType(),
7321 Scale->getZExtValue() / ArrayEltSize);
7322 if (Scale->getZExtValue() != 1) {
7323 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
7324 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7325 NewIdx = InsertNewInstBefore(Sc, GEP);
7328 // Insert the new GEP instruction.
7330 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7331 NewIdx, GEP.getName());
7332 Idx = InsertNewInstBefore(Idx, GEP);
7333 return new CastInst(Idx, GEP.getType());
7342 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7343 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7344 if (AI.isArrayAllocation()) // Check C != 1
7345 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7347 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7348 AllocationInst *New = 0;
7350 // Create and insert the replacement instruction...
7351 if (isa<MallocInst>(AI))
7352 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7354 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7355 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7358 InsertNewInstBefore(New, AI);
7360 // Scan to the end of the allocation instructions, to skip over a block of
7361 // allocas if possible...
7363 BasicBlock::iterator It = New;
7364 while (isa<AllocationInst>(*It)) ++It;
7366 // Now that I is pointing to the first non-allocation-inst in the block,
7367 // insert our getelementptr instruction...
7369 Value *NullIdx = Constant::getNullValue(Type::IntTy);
7370 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7371 New->getName()+".sub", It);
7373 // Now make everything use the getelementptr instead of the original
7375 return ReplaceInstUsesWith(AI, V);
7376 } else if (isa<UndefValue>(AI.getArraySize())) {
7377 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7380 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7381 // Note that we only do this for alloca's, because malloc should allocate and
7382 // return a unique pointer, even for a zero byte allocation.
7383 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7384 TD->getTypeSize(AI.getAllocatedType()) == 0)
7385 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7390 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7391 Value *Op = FI.getOperand(0);
7393 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7394 if (CastInst *CI = dyn_cast<CastInst>(Op))
7395 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7396 FI.setOperand(0, CI->getOperand(0));
7400 // free undef -> unreachable.
7401 if (isa<UndefValue>(Op)) {
7402 // Insert a new store to null because we cannot modify the CFG here.
7403 new StoreInst(ConstantBool::getTrue(),
7404 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
7405 return EraseInstFromFunction(FI);
7408 // If we have 'free null' delete the instruction. This can happen in stl code
7409 // when lots of inlining happens.
7410 if (isa<ConstantPointerNull>(Op))
7411 return EraseInstFromFunction(FI);
7417 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7418 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7419 User *CI = cast<User>(LI.getOperand(0));
7420 Value *CastOp = CI->getOperand(0);
7422 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7423 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7424 const Type *SrcPTy = SrcTy->getElementType();
7426 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7427 isa<PackedType>(DestPTy)) {
7428 // If the source is an array, the code below will not succeed. Check to
7429 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7431 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7432 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7433 if (ASrcTy->getNumElements() != 0) {
7434 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7435 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7436 SrcTy = cast<PointerType>(CastOp->getType());
7437 SrcPTy = SrcTy->getElementType();
7440 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7441 isa<PackedType>(SrcPTy)) &&
7442 // Do not allow turning this into a load of an integer, which is then
7443 // casted to a pointer, this pessimizes pointer analysis a lot.
7444 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7445 IC.getTargetData().getTypeSize(SrcPTy) ==
7446 IC.getTargetData().getTypeSize(DestPTy)) {
7448 // Okay, we are casting from one integer or pointer type to another of
7449 // the same size. Instead of casting the pointer before the load, cast
7450 // the result of the loaded value.
7451 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
7453 LI.isVolatile()),LI);
7454 // Now cast the result of the load.
7455 return new CastInst(NewLoad, LI.getType());
7462 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
7463 /// from this value cannot trap. If it is not obviously safe to load from the
7464 /// specified pointer, we do a quick local scan of the basic block containing
7465 /// ScanFrom, to determine if the address is already accessed.
7466 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
7467 // If it is an alloca or global variable, it is always safe to load from.
7468 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
7470 // Otherwise, be a little bit agressive by scanning the local block where we
7471 // want to check to see if the pointer is already being loaded or stored
7472 // from/to. If so, the previous load or store would have already trapped,
7473 // so there is no harm doing an extra load (also, CSE will later eliminate
7474 // the load entirely).
7475 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
7480 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7481 if (LI->getOperand(0) == V) return true;
7482 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7483 if (SI->getOperand(1) == V) return true;
7489 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
7490 Value *Op = LI.getOperand(0);
7492 // load (cast X) --> cast (load X) iff safe
7493 if (CastInst *CI = dyn_cast<CastInst>(Op))
7494 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7497 // None of the following transforms are legal for volatile loads.
7498 if (LI.isVolatile()) return 0;
7500 if (&LI.getParent()->front() != &LI) {
7501 BasicBlock::iterator BBI = &LI; --BBI;
7502 // If the instruction immediately before this is a store to the same
7503 // address, do a simple form of store->load forwarding.
7504 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7505 if (SI->getOperand(1) == LI.getOperand(0))
7506 return ReplaceInstUsesWith(LI, SI->getOperand(0));
7507 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
7508 if (LIB->getOperand(0) == LI.getOperand(0))
7509 return ReplaceInstUsesWith(LI, LIB);
7512 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
7513 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
7514 isa<UndefValue>(GEPI->getOperand(0))) {
7515 // Insert a new store to null instruction before the load to indicate
7516 // that this code is not reachable. We do this instead of inserting
7517 // an unreachable instruction directly because we cannot modify the
7519 new StoreInst(UndefValue::get(LI.getType()),
7520 Constant::getNullValue(Op->getType()), &LI);
7521 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7524 if (Constant *C = dyn_cast<Constant>(Op)) {
7525 // load null/undef -> undef
7526 if ((C->isNullValue() || isa<UndefValue>(C))) {
7527 // Insert a new store to null instruction before the load to indicate that
7528 // this code is not reachable. We do this instead of inserting an
7529 // unreachable instruction directly because we cannot modify the CFG.
7530 new StoreInst(UndefValue::get(LI.getType()),
7531 Constant::getNullValue(Op->getType()), &LI);
7532 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7535 // Instcombine load (constant global) into the value loaded.
7536 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
7537 if (GV->isConstant() && !GV->isExternal())
7538 return ReplaceInstUsesWith(LI, GV->getInitializer());
7540 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
7541 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7542 if (CE->getOpcode() == Instruction::GetElementPtr) {
7543 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
7544 if (GV->isConstant() && !GV->isExternal())
7546 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
7547 return ReplaceInstUsesWith(LI, V);
7548 if (CE->getOperand(0)->isNullValue()) {
7549 // Insert a new store to null instruction before the load to indicate
7550 // that this code is not reachable. We do this instead of inserting
7551 // an unreachable instruction directly because we cannot modify the
7553 new StoreInst(UndefValue::get(LI.getType()),
7554 Constant::getNullValue(Op->getType()), &LI);
7555 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7558 } else if (CE->getOpcode() == Instruction::Cast) {
7559 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7564 if (Op->hasOneUse()) {
7565 // Change select and PHI nodes to select values instead of addresses: this
7566 // helps alias analysis out a lot, allows many others simplifications, and
7567 // exposes redundancy in the code.
7569 // Note that we cannot do the transformation unless we know that the
7570 // introduced loads cannot trap! Something like this is valid as long as
7571 // the condition is always false: load (select bool %C, int* null, int* %G),
7572 // but it would not be valid if we transformed it to load from null
7575 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7576 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7577 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7578 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7579 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
7580 SI->getOperand(1)->getName()+".val"), LI);
7581 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
7582 SI->getOperand(2)->getName()+".val"), LI);
7583 return new SelectInst(SI->getCondition(), V1, V2);
7586 // load (select (cond, null, P)) -> load P
7587 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7588 if (C->isNullValue()) {
7589 LI.setOperand(0, SI->getOperand(2));
7593 // load (select (cond, P, null)) -> load P
7594 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7595 if (C->isNullValue()) {
7596 LI.setOperand(0, SI->getOperand(1));
7604 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
7606 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7607 User *CI = cast<User>(SI.getOperand(1));
7608 Value *CastOp = CI->getOperand(0);
7610 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7611 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7612 const Type *SrcPTy = SrcTy->getElementType();
7614 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
7615 // If the source is an array, the code below will not succeed. Check to
7616 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7618 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7619 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7620 if (ASrcTy->getNumElements() != 0) {
7621 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7622 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7623 SrcTy = cast<PointerType>(CastOp->getType());
7624 SrcPTy = SrcTy->getElementType();
7627 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
7628 IC.getTargetData().getTypeSize(SrcPTy) ==
7629 IC.getTargetData().getTypeSize(DestPTy)) {
7631 // Okay, we are casting from one integer or pointer type to another of
7632 // the same size. Instead of casting the pointer before the store, cast
7633 // the value to be stored.
7635 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
7636 NewCast = ConstantExpr::getCast(C, SrcPTy);
7638 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
7640 SI.getOperand(0)->getName()+".c"), SI);
7642 return new StoreInst(NewCast, CastOp);
7649 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7650 Value *Val = SI.getOperand(0);
7651 Value *Ptr = SI.getOperand(1);
7653 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
7654 EraseInstFromFunction(SI);
7659 // Do really simple DSE, to catch cases where there are several consequtive
7660 // stores to the same location, separated by a few arithmetic operations. This
7661 // situation often occurs with bitfield accesses.
7662 BasicBlock::iterator BBI = &SI;
7663 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7667 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7668 // Prev store isn't volatile, and stores to the same location?
7669 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
7672 EraseInstFromFunction(*PrevSI);
7678 // If this is a load, we have to stop. However, if the loaded value is from
7679 // the pointer we're loading and is producing the pointer we're storing,
7680 // then *this* store is dead (X = load P; store X -> P).
7681 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7682 if (LI == Val && LI->getOperand(0) == Ptr) {
7683 EraseInstFromFunction(SI);
7687 // Otherwise, this is a load from some other location. Stores before it
7692 // Don't skip over loads or things that can modify memory.
7693 if (BBI->mayWriteToMemory())
7698 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7700 // store X, null -> turns into 'unreachable' in SimplifyCFG
7701 if (isa<ConstantPointerNull>(Ptr)) {
7702 if (!isa<UndefValue>(Val)) {
7703 SI.setOperand(0, UndefValue::get(Val->getType()));
7704 if (Instruction *U = dyn_cast<Instruction>(Val))
7705 WorkList.push_back(U); // Dropped a use.
7708 return 0; // Do not modify these!
7711 // store undef, Ptr -> noop
7712 if (isa<UndefValue>(Val)) {
7713 EraseInstFromFunction(SI);
7718 // If the pointer destination is a cast, see if we can fold the cast into the
7720 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
7721 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7723 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
7724 if (CE->getOpcode() == Instruction::Cast)
7725 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7729 // If this store is the last instruction in the basic block, and if the block
7730 // ends with an unconditional branch, try to move it to the successor block.
7732 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
7733 if (BI->isUnconditional()) {
7734 // Check to see if the successor block has exactly two incoming edges. If
7735 // so, see if the other predecessor contains a store to the same location.
7736 // if so, insert a PHI node (if needed) and move the stores down.
7737 BasicBlock *Dest = BI->getSuccessor(0);
7739 pred_iterator PI = pred_begin(Dest);
7740 BasicBlock *Other = 0;
7741 if (*PI != BI->getParent())
7744 if (PI != pred_end(Dest)) {
7745 if (*PI != BI->getParent())
7750 if (++PI != pred_end(Dest))
7753 if (Other) { // If only one other pred...
7754 BBI = Other->getTerminator();
7755 // Make sure this other block ends in an unconditional branch and that
7756 // there is an instruction before the branch.
7757 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
7758 BBI != Other->begin()) {
7760 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
7762 // If this instruction is a store to the same location.
7763 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
7764 // Okay, we know we can perform this transformation. Insert a PHI
7765 // node now if we need it.
7766 Value *MergedVal = OtherStore->getOperand(0);
7767 if (MergedVal != SI.getOperand(0)) {
7768 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
7769 PN->reserveOperandSpace(2);
7770 PN->addIncoming(SI.getOperand(0), SI.getParent());
7771 PN->addIncoming(OtherStore->getOperand(0), Other);
7772 MergedVal = InsertNewInstBefore(PN, Dest->front());
7775 // Advance to a place where it is safe to insert the new store and
7777 BBI = Dest->begin();
7778 while (isa<PHINode>(BBI)) ++BBI;
7779 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
7780 OtherStore->isVolatile()), *BBI);
7782 // Nuke the old stores.
7783 EraseInstFromFunction(SI);
7784 EraseInstFromFunction(*OtherStore);
7796 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
7797 // Change br (not X), label True, label False to: br X, label False, True
7799 BasicBlock *TrueDest;
7800 BasicBlock *FalseDest;
7801 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
7802 !isa<Constant>(X)) {
7803 // Swap Destinations and condition...
7805 BI.setSuccessor(0, FalseDest);
7806 BI.setSuccessor(1, TrueDest);
7810 // Cannonicalize setne -> seteq
7811 Instruction::BinaryOps Op; Value *Y;
7812 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
7813 TrueDest, FalseDest)))
7814 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
7815 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
7816 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
7817 std::string Name = I->getName(); I->setName("");
7818 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
7819 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
7820 // Swap Destinations and condition...
7821 BI.setCondition(NewSCC);
7822 BI.setSuccessor(0, FalseDest);
7823 BI.setSuccessor(1, TrueDest);
7824 removeFromWorkList(I);
7825 I->getParent()->getInstList().erase(I);
7826 WorkList.push_back(cast<Instruction>(NewSCC));
7833 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
7834 Value *Cond = SI.getCondition();
7835 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
7836 if (I->getOpcode() == Instruction::Add)
7837 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7838 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
7839 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
7840 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
7842 SI.setOperand(0, I->getOperand(0));
7843 WorkList.push_back(I);
7850 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
7851 /// is to leave as a vector operation.
7852 static bool CheapToScalarize(Value *V, bool isConstant) {
7853 if (isa<ConstantAggregateZero>(V))
7855 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
7856 if (isConstant) return true;
7857 // If all elts are the same, we can extract.
7858 Constant *Op0 = C->getOperand(0);
7859 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7860 if (C->getOperand(i) != Op0)
7864 Instruction *I = dyn_cast<Instruction>(V);
7865 if (!I) return false;
7867 // Insert element gets simplified to the inserted element or is deleted if
7868 // this is constant idx extract element and its a constant idx insertelt.
7869 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
7870 isa<ConstantInt>(I->getOperand(2)))
7872 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
7874 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
7875 if (BO->hasOneUse() &&
7876 (CheapToScalarize(BO->getOperand(0), isConstant) ||
7877 CheapToScalarize(BO->getOperand(1), isConstant)))
7883 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
7884 /// elements into values that are larger than the #elts in the input.
7885 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
7886 unsigned NElts = SVI->getType()->getNumElements();
7887 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
7888 return std::vector<unsigned>(NElts, 0);
7889 if (isa<UndefValue>(SVI->getOperand(2)))
7890 return std::vector<unsigned>(NElts, 2*NElts);
7892 std::vector<unsigned> Result;
7893 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
7894 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
7895 if (isa<UndefValue>(CP->getOperand(i)))
7896 Result.push_back(NElts*2); // undef -> 8
7898 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
7902 /// FindScalarElement - Given a vector and an element number, see if the scalar
7903 /// value is already around as a register, for example if it were inserted then
7904 /// extracted from the vector.
7905 static Value *FindScalarElement(Value *V, unsigned EltNo) {
7906 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
7907 const PackedType *PTy = cast<PackedType>(V->getType());
7908 unsigned Width = PTy->getNumElements();
7909 if (EltNo >= Width) // Out of range access.
7910 return UndefValue::get(PTy->getElementType());
7912 if (isa<UndefValue>(V))
7913 return UndefValue::get(PTy->getElementType());
7914 else if (isa<ConstantAggregateZero>(V))
7915 return Constant::getNullValue(PTy->getElementType());
7916 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
7917 return CP->getOperand(EltNo);
7918 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
7919 // If this is an insert to a variable element, we don't know what it is.
7920 if (!isa<ConstantInt>(III->getOperand(2)))
7922 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
7924 // If this is an insert to the element we are looking for, return the
7927 return III->getOperand(1);
7929 // Otherwise, the insertelement doesn't modify the value, recurse on its
7931 return FindScalarElement(III->getOperand(0), EltNo);
7932 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
7933 unsigned InEl = getShuffleMask(SVI)[EltNo];
7935 return FindScalarElement(SVI->getOperand(0), InEl);
7936 else if (InEl < Width*2)
7937 return FindScalarElement(SVI->getOperand(1), InEl - Width);
7939 return UndefValue::get(PTy->getElementType());
7942 // Otherwise, we don't know.
7946 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
7948 // If packed val is undef, replace extract with scalar undef.
7949 if (isa<UndefValue>(EI.getOperand(0)))
7950 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
7952 // If packed val is constant 0, replace extract with scalar 0.
7953 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
7954 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
7956 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
7957 // If packed val is constant with uniform operands, replace EI
7958 // with that operand
7959 Constant *op0 = C->getOperand(0);
7960 for (unsigned i = 1; i < C->getNumOperands(); ++i)
7961 if (C->getOperand(i) != op0) {
7966 return ReplaceInstUsesWith(EI, op0);
7969 // If extracting a specified index from the vector, see if we can recursively
7970 // find a previously computed scalar that was inserted into the vector.
7971 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
7972 // This instruction only demands the single element from the input vector.
7973 // If the input vector has a single use, simplify it based on this use
7975 uint64_t IndexVal = IdxC->getZExtValue();
7976 if (EI.getOperand(0)->hasOneUse()) {
7978 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
7981 EI.setOperand(0, V);
7986 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
7987 return ReplaceInstUsesWith(EI, Elt);
7990 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
7991 if (I->hasOneUse()) {
7992 // Push extractelement into predecessor operation if legal and
7993 // profitable to do so
7994 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
7995 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
7996 if (CheapToScalarize(BO, isConstantElt)) {
7997 ExtractElementInst *newEI0 =
7998 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
7999 EI.getName()+".lhs");
8000 ExtractElementInst *newEI1 =
8001 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8002 EI.getName()+".rhs");
8003 InsertNewInstBefore(newEI0, EI);
8004 InsertNewInstBefore(newEI1, EI);
8005 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8007 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8008 Value *Ptr = InsertCastBefore(I->getOperand(0),
8009 PointerType::get(EI.getType()), EI);
8010 GetElementPtrInst *GEP =
8011 new GetElementPtrInst(Ptr, EI.getOperand(1),
8012 I->getName() + ".gep");
8013 InsertNewInstBefore(GEP, EI);
8014 return new LoadInst(GEP);
8017 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8018 // Extracting the inserted element?
8019 if (IE->getOperand(2) == EI.getOperand(1))
8020 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8021 // If the inserted and extracted elements are constants, they must not
8022 // be the same value, extract from the pre-inserted value instead.
8023 if (isa<Constant>(IE->getOperand(2)) &&
8024 isa<Constant>(EI.getOperand(1))) {
8025 AddUsesToWorkList(EI);
8026 EI.setOperand(0, IE->getOperand(0));
8029 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8030 // If this is extracting an element from a shufflevector, figure out where
8031 // it came from and extract from the appropriate input element instead.
8032 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8033 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8035 if (SrcIdx < SVI->getType()->getNumElements())
8036 Src = SVI->getOperand(0);
8037 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8038 SrcIdx -= SVI->getType()->getNumElements();
8039 Src = SVI->getOperand(1);
8041 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8043 return new ExtractElementInst(Src, SrcIdx);
8050 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8051 /// elements from either LHS or RHS, return the shuffle mask and true.
8052 /// Otherwise, return false.
8053 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8054 std::vector<Constant*> &Mask) {
8055 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8056 "Invalid CollectSingleShuffleElements");
8057 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8059 if (isa<UndefValue>(V)) {
8060 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8062 } else if (V == LHS) {
8063 for (unsigned i = 0; i != NumElts; ++i)
8064 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8066 } else if (V == RHS) {
8067 for (unsigned i = 0; i != NumElts; ++i)
8068 Mask.push_back(ConstantInt::get(Type::UIntTy, i+NumElts));
8070 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8071 // If this is an insert of an extract from some other vector, include it.
8072 Value *VecOp = IEI->getOperand(0);
8073 Value *ScalarOp = IEI->getOperand(1);
8074 Value *IdxOp = IEI->getOperand(2);
8076 if (!isa<ConstantInt>(IdxOp))
8078 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8080 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8081 // Okay, we can handle this if the vector we are insertinting into is
8083 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8084 // If so, update the mask to reflect the inserted undef.
8085 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
8088 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8089 if (isa<ConstantInt>(EI->getOperand(1)) &&
8090 EI->getOperand(0)->getType() == V->getType()) {
8091 unsigned ExtractedIdx =
8092 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8094 // This must be extracting from either LHS or RHS.
8095 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8096 // Okay, we can handle this if the vector we are insertinting into is
8098 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8099 // If so, update the mask to reflect the inserted value.
8100 if (EI->getOperand(0) == LHS) {
8101 Mask[InsertedIdx & (NumElts-1)] =
8102 ConstantInt::get(Type::UIntTy, ExtractedIdx);
8104 assert(EI->getOperand(0) == RHS);
8105 Mask[InsertedIdx & (NumElts-1)] =
8106 ConstantInt::get(Type::UIntTy, ExtractedIdx+NumElts);
8115 // TODO: Handle shufflevector here!
8120 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8121 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8122 /// that computes V and the LHS value of the shuffle.
8123 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8125 assert(isa<PackedType>(V->getType()) &&
8126 (RHS == 0 || V->getType() == RHS->getType()) &&
8127 "Invalid shuffle!");
8128 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8130 if (isa<UndefValue>(V)) {
8131 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8133 } else if (isa<ConstantAggregateZero>(V)) {
8134 Mask.assign(NumElts, ConstantInt::get(Type::UIntTy, 0));
8136 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8137 // If this is an insert of an extract from some other vector, include it.
8138 Value *VecOp = IEI->getOperand(0);
8139 Value *ScalarOp = IEI->getOperand(1);
8140 Value *IdxOp = IEI->getOperand(2);
8142 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8143 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8144 EI->getOperand(0)->getType() == V->getType()) {
8145 unsigned ExtractedIdx =
8146 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8147 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8149 // Either the extracted from or inserted into vector must be RHSVec,
8150 // otherwise we'd end up with a shuffle of three inputs.
8151 if (EI->getOperand(0) == RHS || RHS == 0) {
8152 RHS = EI->getOperand(0);
8153 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8154 Mask[InsertedIdx & (NumElts-1)] =
8155 ConstantInt::get(Type::UIntTy, NumElts+ExtractedIdx);
8160 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8161 // Everything but the extracted element is replaced with the RHS.
8162 for (unsigned i = 0; i != NumElts; ++i) {
8163 if (i != InsertedIdx)
8164 Mask[i] = ConstantInt::get(Type::UIntTy, NumElts+i);
8169 // If this insertelement is a chain that comes from exactly these two
8170 // vectors, return the vector and the effective shuffle.
8171 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8172 return EI->getOperand(0);
8177 // TODO: Handle shufflevector here!
8179 // Otherwise, can't do anything fancy. Return an identity vector.
8180 for (unsigned i = 0; i != NumElts; ++i)
8181 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8185 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8186 Value *VecOp = IE.getOperand(0);
8187 Value *ScalarOp = IE.getOperand(1);
8188 Value *IdxOp = IE.getOperand(2);
8190 // If the inserted element was extracted from some other vector, and if the
8191 // indexes are constant, try to turn this into a shufflevector operation.
8192 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8193 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8194 EI->getOperand(0)->getType() == IE.getType()) {
8195 unsigned NumVectorElts = IE.getType()->getNumElements();
8196 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8197 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8199 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8200 return ReplaceInstUsesWith(IE, VecOp);
8202 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8203 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8205 // If we are extracting a value from a vector, then inserting it right
8206 // back into the same place, just use the input vector.
8207 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8208 return ReplaceInstUsesWith(IE, VecOp);
8210 // We could theoretically do this for ANY input. However, doing so could
8211 // turn chains of insertelement instructions into a chain of shufflevector
8212 // instructions, and right now we do not merge shufflevectors. As such,
8213 // only do this in a situation where it is clear that there is benefit.
8214 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8215 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8216 // the values of VecOp, except then one read from EIOp0.
8217 // Build a new shuffle mask.
8218 std::vector<Constant*> Mask;
8219 if (isa<UndefValue>(VecOp))
8220 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
8222 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8223 Mask.assign(NumVectorElts, ConstantInt::get(Type::UIntTy,
8226 Mask[InsertedIdx] = ConstantInt::get(Type::UIntTy, ExtractedIdx);
8227 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8228 ConstantPacked::get(Mask));
8231 // If this insertelement isn't used by some other insertelement, turn it
8232 // (and any insertelements it points to), into one big shuffle.
8233 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8234 std::vector<Constant*> Mask;
8236 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8237 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8238 // We now have a shuffle of LHS, RHS, Mask.
8239 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8248 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8249 Value *LHS = SVI.getOperand(0);
8250 Value *RHS = SVI.getOperand(1);
8251 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8253 bool MadeChange = false;
8255 // Undefined shuffle mask -> undefined value.
8256 if (isa<UndefValue>(SVI.getOperand(2)))
8257 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8259 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
8260 // the undef, change them to undefs.
8262 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8263 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8264 if (LHS == RHS || isa<UndefValue>(LHS)) {
8265 if (isa<UndefValue>(LHS) && LHS == RHS) {
8266 // shuffle(undef,undef,mask) -> undef.
8267 return ReplaceInstUsesWith(SVI, LHS);
8270 // Remap any references to RHS to use LHS.
8271 std::vector<Constant*> Elts;
8272 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8274 Elts.push_back(UndefValue::get(Type::UIntTy));
8276 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8277 (Mask[i] < e && isa<UndefValue>(LHS)))
8278 Mask[i] = 2*e; // Turn into undef.
8280 Mask[i] &= (e-1); // Force to LHS.
8281 Elts.push_back(ConstantInt::get(Type::UIntTy, Mask[i]));
8284 SVI.setOperand(0, SVI.getOperand(1));
8285 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8286 SVI.setOperand(2, ConstantPacked::get(Elts));
8287 LHS = SVI.getOperand(0);
8288 RHS = SVI.getOperand(1);
8292 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8293 bool isLHSID = true, isRHSID = true;
8295 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8296 if (Mask[i] >= e*2) continue; // Ignore undef values.
8297 // Is this an identity shuffle of the LHS value?
8298 isLHSID &= (Mask[i] == i);
8300 // Is this an identity shuffle of the RHS value?
8301 isRHSID &= (Mask[i]-e == i);
8304 // Eliminate identity shuffles.
8305 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8306 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8308 // If the LHS is a shufflevector itself, see if we can combine it with this
8309 // one without producing an unusual shuffle. Here we are really conservative:
8310 // we are absolutely afraid of producing a shuffle mask not in the input
8311 // program, because the code gen may not be smart enough to turn a merged
8312 // shuffle into two specific shuffles: it may produce worse code. As such,
8313 // we only merge two shuffles if the result is one of the two input shuffle
8314 // masks. In this case, merging the shuffles just removes one instruction,
8315 // which we know is safe. This is good for things like turning:
8316 // (splat(splat)) -> splat.
8317 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8318 if (isa<UndefValue>(RHS)) {
8319 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8321 std::vector<unsigned> NewMask;
8322 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8324 NewMask.push_back(2*e);
8326 NewMask.push_back(LHSMask[Mask[i]]);
8328 // If the result mask is equal to the src shuffle or this shuffle mask, do
8330 if (NewMask == LHSMask || NewMask == Mask) {
8331 std::vector<Constant*> Elts;
8332 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8333 if (NewMask[i] >= e*2) {
8334 Elts.push_back(UndefValue::get(Type::UIntTy));
8336 Elts.push_back(ConstantInt::get(Type::UIntTy, NewMask[i]));
8339 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8340 LHSSVI->getOperand(1),
8341 ConstantPacked::get(Elts));
8346 return MadeChange ? &SVI : 0;
8351 void InstCombiner::removeFromWorkList(Instruction *I) {
8352 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8357 /// TryToSinkInstruction - Try to move the specified instruction from its
8358 /// current block into the beginning of DestBlock, which can only happen if it's
8359 /// safe to move the instruction past all of the instructions between it and the
8360 /// end of its block.
8361 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8362 assert(I->hasOneUse() && "Invariants didn't hold!");
8364 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8365 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
8367 // Do not sink alloca instructions out of the entry block.
8368 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
8371 // We can only sink load instructions if there is nothing between the load and
8372 // the end of block that could change the value.
8373 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8374 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
8376 if (Scan->mayWriteToMemory())
8380 BasicBlock::iterator InsertPos = DestBlock->begin();
8381 while (isa<PHINode>(InsertPos)) ++InsertPos;
8383 I->moveBefore(InsertPos);
8388 /// OptimizeConstantExpr - Given a constant expression and target data layout
8389 /// information, symbolically evaluation the constant expr to something simpler
8391 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
8394 Constant *Ptr = CE->getOperand(0);
8395 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
8396 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
8397 // If this is a constant expr gep that is effectively computing an
8398 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
8399 bool isFoldableGEP = true;
8400 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
8401 if (!isa<ConstantInt>(CE->getOperand(i)))
8402 isFoldableGEP = false;
8403 if (isFoldableGEP) {
8404 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
8405 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
8406 Constant *C = ConstantInt::get(Type::ULongTy, Offset);
8407 C = ConstantExpr::getCast(C, TD->getIntPtrType());
8408 return ConstantExpr::getCast(C, CE->getType());
8416 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
8417 /// all reachable code to the worklist.
8419 /// This has a couple of tricks to make the code faster and more powerful. In
8420 /// particular, we constant fold and DCE instructions as we go, to avoid adding
8421 /// them to the worklist (this significantly speeds up instcombine on code where
8422 /// many instructions are dead or constant). Additionally, if we find a branch
8423 /// whose condition is a known constant, we only visit the reachable successors.
8425 static void AddReachableCodeToWorklist(BasicBlock *BB,
8426 std::set<BasicBlock*> &Visited,
8427 std::vector<Instruction*> &WorkList,
8428 const TargetData *TD) {
8429 // We have now visited this block! If we've already been here, bail out.
8430 if (!Visited.insert(BB).second) return;
8432 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
8433 Instruction *Inst = BBI++;
8435 // DCE instruction if trivially dead.
8436 if (isInstructionTriviallyDead(Inst)) {
8438 DEBUG(std::cerr << "IC: DCE: " << *Inst);
8439 Inst->eraseFromParent();
8443 // ConstantProp instruction if trivially constant.
8444 if (Constant *C = ConstantFoldInstruction(Inst)) {
8445 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8446 C = OptimizeConstantExpr(CE, TD);
8447 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *Inst);
8448 Inst->replaceAllUsesWith(C);
8450 Inst->eraseFromParent();
8454 WorkList.push_back(Inst);
8457 // Recursively visit successors. If this is a branch or switch on a constant,
8458 // only visit the reachable successor.
8459 TerminatorInst *TI = BB->getTerminator();
8460 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
8461 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
8462 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
8463 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
8467 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
8468 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
8469 // See if this is an explicit destination.
8470 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
8471 if (SI->getCaseValue(i) == Cond) {
8472 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
8476 // Otherwise it is the default destination.
8477 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
8482 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
8483 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
8486 bool InstCombiner::runOnFunction(Function &F) {
8487 bool Changed = false;
8488 TD = &getAnalysis<TargetData>();
8491 // Do a depth-first traversal of the function, populate the worklist with
8492 // the reachable instructions. Ignore blocks that are not reachable. Keep
8493 // track of which blocks we visit.
8494 std::set<BasicBlock*> Visited;
8495 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
8497 // Do a quick scan over the function. If we find any blocks that are
8498 // unreachable, remove any instructions inside of them. This prevents
8499 // the instcombine code from having to deal with some bad special cases.
8500 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
8501 if (!Visited.count(BB)) {
8502 Instruction *Term = BB->getTerminator();
8503 while (Term != BB->begin()) { // Remove instrs bottom-up
8504 BasicBlock::iterator I = Term; --I;
8506 DEBUG(std::cerr << "IC: DCE: " << *I);
8509 if (!I->use_empty())
8510 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8511 I->eraseFromParent();
8516 while (!WorkList.empty()) {
8517 Instruction *I = WorkList.back(); // Get an instruction from the worklist
8518 WorkList.pop_back();
8520 // Check to see if we can DCE the instruction.
8521 if (isInstructionTriviallyDead(I)) {
8522 // Add operands to the worklist.
8523 if (I->getNumOperands() < 4)
8524 AddUsesToWorkList(*I);
8527 DEBUG(std::cerr << "IC: DCE: " << *I);
8529 I->eraseFromParent();
8530 removeFromWorkList(I);
8534 // Instruction isn't dead, see if we can constant propagate it.
8535 if (Constant *C = ConstantFoldInstruction(I)) {
8536 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8537 C = OptimizeConstantExpr(CE, TD);
8538 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
8540 // Add operands to the worklist.
8541 AddUsesToWorkList(*I);
8542 ReplaceInstUsesWith(*I, C);
8545 I->eraseFromParent();
8546 removeFromWorkList(I);
8550 // See if we can trivially sink this instruction to a successor basic block.
8551 if (I->hasOneUse()) {
8552 BasicBlock *BB = I->getParent();
8553 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
8554 if (UserParent != BB) {
8555 bool UserIsSuccessor = false;
8556 // See if the user is one of our successors.
8557 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8558 if (*SI == UserParent) {
8559 UserIsSuccessor = true;
8563 // If the user is one of our immediate successors, and if that successor
8564 // only has us as a predecessors (we'd have to split the critical edge
8565 // otherwise), we can keep going.
8566 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
8567 next(pred_begin(UserParent)) == pred_end(UserParent))
8568 // Okay, the CFG is simple enough, try to sink this instruction.
8569 Changed |= TryToSinkInstruction(I, UserParent);
8573 // Now that we have an instruction, try combining it to simplify it...
8574 if (Instruction *Result = visit(*I)) {
8576 // Should we replace the old instruction with a new one?
8578 DEBUG(std::cerr << "IC: Old = " << *I
8579 << " New = " << *Result);
8581 // Everything uses the new instruction now.
8582 I->replaceAllUsesWith(Result);
8584 // Push the new instruction and any users onto the worklist.
8585 WorkList.push_back(Result);
8586 AddUsersToWorkList(*Result);
8588 // Move the name to the new instruction first...
8589 std::string OldName = I->getName(); I->setName("");
8590 Result->setName(OldName);
8592 // Insert the new instruction into the basic block...
8593 BasicBlock *InstParent = I->getParent();
8594 BasicBlock::iterator InsertPos = I;
8596 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8597 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8600 InstParent->getInstList().insert(InsertPos, Result);
8602 // Make sure that we reprocess all operands now that we reduced their
8604 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8605 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8606 WorkList.push_back(OpI);
8608 // Instructions can end up on the worklist more than once. Make sure
8609 // we do not process an instruction that has been deleted.
8610 removeFromWorkList(I);
8612 // Erase the old instruction.
8613 InstParent->getInstList().erase(I);
8615 DEBUG(std::cerr << "IC: MOD = " << *I);
8617 // If the instruction was modified, it's possible that it is now dead.
8618 // if so, remove it.
8619 if (isInstructionTriviallyDead(I)) {
8620 // Make sure we process all operands now that we are reducing their
8622 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8623 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8624 WorkList.push_back(OpI);
8626 // Instructions may end up in the worklist more than once. Erase all
8627 // occurrences of this instruction.
8628 removeFromWorkList(I);
8629 I->eraseFromParent();
8631 WorkList.push_back(Result);
8632 AddUsersToWorkList(*Result);
8642 FunctionPass *llvm::createInstructionCombiningPass() {
8643 return new InstCombiner();