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/ADT/DepthFirstIterator.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<> NumSunkInst ("instcombine", "Number of instructions sunk");
65 class InstCombiner : public FunctionPass,
66 public InstVisitor<InstCombiner, Instruction*> {
67 // Worklist of all of the instructions that need to be simplified.
68 std::vector<Instruction*> WorkList;
71 /// AddUsersToWorkList - When an instruction is simplified, add all users of
72 /// the instruction to the work lists because they might get more simplified
75 void AddUsersToWorkList(Value &I) {
76 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
78 WorkList.push_back(cast<Instruction>(*UI));
81 /// AddUsesToWorkList - When an instruction is simplified, add operands to
82 /// the work lists because they might get more simplified now.
84 void AddUsesToWorkList(Instruction &I) {
85 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
86 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
87 WorkList.push_back(Op);
90 // removeFromWorkList - remove all instances of I from the worklist.
91 void removeFromWorkList(Instruction *I);
93 virtual bool runOnFunction(Function &F);
95 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
96 AU.addRequired<TargetData>();
100 TargetData &getTargetData() const { return *TD; }
102 // Visitation implementation - Implement instruction combining for different
103 // instruction types. The semantics are as follows:
105 // null - No change was made
106 // I - Change was made, I is still valid, I may be dead though
107 // otherwise - Change was made, replace I with returned instruction
109 Instruction *visitAdd(BinaryOperator &I);
110 Instruction *visitSub(BinaryOperator &I);
111 Instruction *visitMul(BinaryOperator &I);
112 Instruction *visitDiv(BinaryOperator &I);
113 Instruction *visitRem(BinaryOperator &I);
114 Instruction *visitAnd(BinaryOperator &I);
115 Instruction *visitOr (BinaryOperator &I);
116 Instruction *visitXor(BinaryOperator &I);
117 Instruction *visitSetCondInst(SetCondInst &I);
118 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
120 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
121 Instruction::BinaryOps Cond, Instruction &I);
122 Instruction *visitShiftInst(ShiftInst &I);
123 Instruction *FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
125 Instruction *visitCastInst(CastInst &CI);
126 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
128 Instruction *visitSelectInst(SelectInst &CI);
129 Instruction *visitCallInst(CallInst &CI);
130 Instruction *visitInvokeInst(InvokeInst &II);
131 Instruction *visitPHINode(PHINode &PN);
132 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
133 Instruction *visitAllocationInst(AllocationInst &AI);
134 Instruction *visitFreeInst(FreeInst &FI);
135 Instruction *visitLoadInst(LoadInst &LI);
136 Instruction *visitStoreInst(StoreInst &SI);
137 Instruction *visitBranchInst(BranchInst &BI);
138 Instruction *visitSwitchInst(SwitchInst &SI);
139 Instruction *visitExtractElementInst(ExtractElementInst &EI);
141 // visitInstruction - Specify what to return for unhandled instructions...
142 Instruction *visitInstruction(Instruction &I) { return 0; }
145 Instruction *visitCallSite(CallSite CS);
146 bool transformConstExprCastCall(CallSite CS);
149 // InsertNewInstBefore - insert an instruction New before instruction Old
150 // in the program. Add the new instruction to the worklist.
152 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
153 assert(New && New->getParent() == 0 &&
154 "New instruction already inserted into a basic block!");
155 BasicBlock *BB = Old.getParent();
156 BB->getInstList().insert(&Old, New); // Insert inst
157 WorkList.push_back(New); // Add to worklist
161 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
162 /// This also adds the cast to the worklist. Finally, this returns the
164 Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
165 if (V->getType() == Ty) return V;
167 Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
168 WorkList.push_back(C);
172 // ReplaceInstUsesWith - This method is to be used when an instruction is
173 // found to be dead, replacable with another preexisting expression. Here
174 // we add all uses of I to the worklist, replace all uses of I with the new
175 // value, then return I, so that the inst combiner will know that I was
178 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
179 AddUsersToWorkList(I); // Add all modified instrs to worklist
181 I.replaceAllUsesWith(V);
184 // If we are replacing the instruction with itself, this must be in a
185 // segment of unreachable code, so just clobber the instruction.
186 I.replaceAllUsesWith(UndefValue::get(I.getType()));
191 // UpdateValueUsesWith - This method is to be used when an value is
192 // found to be replacable with another preexisting expression or was
193 // updated. Here we add all uses of I to the worklist, replace all uses of
194 // I with the new value (unless the instruction was just updated), then
195 // return true, so that the inst combiner will know that I was modified.
197 bool UpdateValueUsesWith(Value *Old, Value *New) {
198 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
200 Old->replaceAllUsesWith(New);
201 if (Instruction *I = dyn_cast<Instruction>(Old))
202 WorkList.push_back(I);
206 // EraseInstFromFunction - When dealing with an instruction that has side
207 // effects or produces a void value, we can't rely on DCE to delete the
208 // instruction. Instead, visit methods should return the value returned by
210 Instruction *EraseInstFromFunction(Instruction &I) {
211 assert(I.use_empty() && "Cannot erase instruction that is used!");
212 AddUsesToWorkList(I);
213 removeFromWorkList(&I);
215 return 0; // Don't do anything with FI
219 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
220 /// InsertBefore instruction. This is specialized a bit to avoid inserting
221 /// casts that are known to not do anything...
223 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
224 Instruction *InsertBefore);
226 // SimplifyCommutative - This performs a few simplifications for commutative
228 bool SimplifyCommutative(BinaryOperator &I);
230 bool SimplifyDemandedBits(Value *V, uint64_t Mask, unsigned Depth = 0);
232 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
233 // PHI node as operand #0, see if we can fold the instruction into the PHI
234 // (which is only possible if all operands to the PHI are constants).
235 Instruction *FoldOpIntoPhi(Instruction &I);
237 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
238 // operator and they all are only used by the PHI, PHI together their
239 // inputs, and do the operation once, to the result of the PHI.
240 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
242 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
243 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
245 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
246 bool isSub, Instruction &I);
247 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
248 bool Inside, Instruction &IB);
249 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
252 RegisterOpt<InstCombiner> X("instcombine", "Combine redundant instructions");
255 // getComplexity: Assign a complexity or rank value to LLVM Values...
256 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
257 static unsigned getComplexity(Value *V) {
258 if (isa<Instruction>(V)) {
259 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
263 if (isa<Argument>(V)) return 3;
264 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
267 // isOnlyUse - Return true if this instruction will be deleted if we stop using
269 static bool isOnlyUse(Value *V) {
270 return V->hasOneUse() || isa<Constant>(V);
273 // getPromotedType - Return the specified type promoted as it would be to pass
274 // though a va_arg area...
275 static const Type *getPromotedType(const Type *Ty) {
276 switch (Ty->getTypeID()) {
277 case Type::SByteTyID:
278 case Type::ShortTyID: return Type::IntTy;
279 case Type::UByteTyID:
280 case Type::UShortTyID: return Type::UIntTy;
281 case Type::FloatTyID: return Type::DoubleTy;
286 /// isCast - If the specified operand is a CastInst or a constant expr cast,
287 /// return the operand value, otherwise return null.
288 static Value *isCast(Value *V) {
289 if (CastInst *I = dyn_cast<CastInst>(V))
290 return I->getOperand(0);
291 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
292 if (CE->getOpcode() == Instruction::Cast)
293 return CE->getOperand(0);
297 // SimplifyCommutative - This performs a few simplifications for commutative
300 // 1. Order operands such that they are listed from right (least complex) to
301 // left (most complex). This puts constants before unary operators before
304 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
305 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
307 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
308 bool Changed = false;
309 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
310 Changed = !I.swapOperands();
312 if (!I.isAssociative()) return Changed;
313 Instruction::BinaryOps Opcode = I.getOpcode();
314 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
315 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
316 if (isa<Constant>(I.getOperand(1))) {
317 Constant *Folded = ConstantExpr::get(I.getOpcode(),
318 cast<Constant>(I.getOperand(1)),
319 cast<Constant>(Op->getOperand(1)));
320 I.setOperand(0, Op->getOperand(0));
321 I.setOperand(1, Folded);
323 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
324 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
325 isOnlyUse(Op) && isOnlyUse(Op1)) {
326 Constant *C1 = cast<Constant>(Op->getOperand(1));
327 Constant *C2 = cast<Constant>(Op1->getOperand(1));
329 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
330 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
331 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
334 WorkList.push_back(New);
335 I.setOperand(0, New);
336 I.setOperand(1, Folded);
343 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
344 // if the LHS is a constant zero (which is the 'negate' form).
346 static inline Value *dyn_castNegVal(Value *V) {
347 if (BinaryOperator::isNeg(V))
348 return BinaryOperator::getNegArgument(V);
350 // Constants can be considered to be negated values if they can be folded.
351 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
352 return ConstantExpr::getNeg(C);
356 static inline Value *dyn_castNotVal(Value *V) {
357 if (BinaryOperator::isNot(V))
358 return BinaryOperator::getNotArgument(V);
360 // Constants can be considered to be not'ed values...
361 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
362 return ConstantExpr::getNot(C);
366 // dyn_castFoldableMul - If this value is a multiply that can be folded into
367 // other computations (because it has a constant operand), return the
368 // non-constant operand of the multiply, and set CST to point to the multiplier.
369 // Otherwise, return null.
371 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
372 if (V->hasOneUse() && V->getType()->isInteger())
373 if (Instruction *I = dyn_cast<Instruction>(V)) {
374 if (I->getOpcode() == Instruction::Mul)
375 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
376 return I->getOperand(0);
377 if (I->getOpcode() == Instruction::Shl)
378 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
379 // The multiplier is really 1 << CST.
380 Constant *One = ConstantInt::get(V->getType(), 1);
381 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
382 return I->getOperand(0);
388 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
389 /// expression, return it.
390 static User *dyn_castGetElementPtr(Value *V) {
391 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
393 if (CE->getOpcode() == Instruction::GetElementPtr)
394 return cast<User>(V);
398 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
399 static ConstantInt *AddOne(ConstantInt *C) {
400 return cast<ConstantInt>(ConstantExpr::getAdd(C,
401 ConstantInt::get(C->getType(), 1)));
403 static ConstantInt *SubOne(ConstantInt *C) {
404 return cast<ConstantInt>(ConstantExpr::getSub(C,
405 ConstantInt::get(C->getType(), 1)));
408 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
409 /// this predicate to simplify operations downstream. V and Mask are known to
410 /// be the same type.
411 static bool MaskedValueIsZero(Value *V, ConstantIntegral *Mask,
412 unsigned Depth = 0) {
413 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
414 // we cannot optimize based on the assumption that it is zero without changing
415 // to to an explicit zero. If we don't change it to zero, other code could
416 // optimized based on the contradictory assumption that it is non-zero.
417 // Because instcombine aggressively folds operations with undef args anyway,
418 // this won't lose us code quality.
419 if (Mask->isNullValue())
421 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V))
422 return ConstantExpr::getAnd(CI, Mask)->isNullValue();
424 if (Depth == 6) return false; // Limit search depth.
426 if (Instruction *I = dyn_cast<Instruction>(V)) {
427 switch (I->getOpcode()) {
428 case Instruction::And:
429 // (X & C1) & C2 == 0 iff C1 & C2 == 0.
430 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(I->getOperand(1))) {
431 ConstantIntegral *C1C2 =
432 cast<ConstantIntegral>(ConstantExpr::getAnd(CI, Mask));
433 if (MaskedValueIsZero(I->getOperand(0), C1C2, Depth+1))
436 // If either the LHS or the RHS are MaskedValueIsZero, the result is zero.
437 return MaskedValueIsZero(I->getOperand(1), Mask, Depth+1) ||
438 MaskedValueIsZero(I->getOperand(0), Mask, Depth+1);
439 case Instruction::Or:
440 case Instruction::Xor:
441 // If the LHS and the RHS are MaskedValueIsZero, the result is also zero.
442 return MaskedValueIsZero(I->getOperand(1), Mask, Depth+1) &&
443 MaskedValueIsZero(I->getOperand(0), Mask, Depth+1);
444 case Instruction::Select:
445 // If the T and F values are MaskedValueIsZero, the result is also zero.
446 return MaskedValueIsZero(I->getOperand(2), Mask, Depth+1) &&
447 MaskedValueIsZero(I->getOperand(1), Mask, Depth+1);
448 case Instruction::Cast: {
449 const Type *SrcTy = I->getOperand(0)->getType();
450 if (SrcTy == Type::BoolTy)
451 return (Mask->getRawValue() & 1) == 0;
453 if (SrcTy->isInteger()) {
454 // (cast <ty> X to int) & C2 == 0 iff <ty> could not have contained C2.
455 if (SrcTy->isUnsigned() && // Only handle zero ext.
456 ConstantExpr::getCast(Mask, SrcTy)->isNullValue())
459 // If this is a noop cast, recurse.
460 if ((SrcTy->isSigned() && SrcTy->getUnsignedVersion() == I->getType())||
461 SrcTy->getSignedVersion() == I->getType()) {
463 ConstantExpr::getCast(Mask, I->getOperand(0)->getType());
464 return MaskedValueIsZero(I->getOperand(0),
465 cast<ConstantIntegral>(NewMask), Depth+1);
470 case Instruction::Shl:
471 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
472 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
473 return MaskedValueIsZero(I->getOperand(0),
474 cast<ConstantIntegral>(ConstantExpr::getUShr(Mask, SA)),
477 case Instruction::Shr:
478 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
479 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
480 if (I->getType()->isUnsigned()) {
481 Constant *C1 = ConstantIntegral::getAllOnesValue(I->getType());
482 C1 = ConstantExpr::getShr(C1, SA);
483 C1 = ConstantExpr::getAnd(C1, Mask);
484 if (C1->isNullValue())
494 /// SimplifyDemandedBits - Look at V. At this point, we know that only the Mask
495 /// bits of the result of V are ever used downstream. If we can use this
496 /// information to simplify V, return V and set NewVal to the new value we
497 /// should use in V's place.
498 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t Mask,
500 if (!V->hasOneUse()) { // Other users may use these bits.
501 if (Depth != 0) // Not at the root.
503 // If this is the root being simplified, allow it to have multiple uses,
504 // just set the Mask to all bits.
505 Mask = V->getType()->getIntegralTypeMask();
506 } else if (Mask == 0) { // Not demanding any bits from V.
507 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
508 } else if (Depth == 6) { // Limit search depth.
512 Instruction *I = dyn_cast<Instruction>(V);
513 if (!I) return false; // Only analyze instructions.
515 switch (I->getOpcode()) {
517 case Instruction::And:
518 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
519 // Only demanding an intersection of the bits.
520 if (SimplifyDemandedBits(I->getOperand(0), RHS->getRawValue() & Mask,
523 if (~Mask & RHS->getRawValue()) {
524 // If this is producing any bits that are not needed, simplify the RHS.
525 if (I->getType()->isSigned()) {
526 int64_t Val = Mask & cast<ConstantSInt>(RHS)->getValue();
527 I->setOperand(1, ConstantSInt::get(I->getType(), Val));
529 uint64_t Val = Mask & cast<ConstantUInt>(RHS)->getValue();
530 I->setOperand(1, ConstantUInt::get(I->getType(), Val));
532 return UpdateValueUsesWith(I, I);
535 // Walk the LHS and the RHS.
536 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1) ||
537 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
538 case Instruction::Or:
539 case Instruction::Xor:
540 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
541 // If none of the [x]or'd in bits are demanded, don't both with the [x]or.
542 if ((Mask & RHS->getRawValue()) == 0)
543 return UpdateValueUsesWith(I, I->getOperand(0));
545 // Otherwise, for an OR, we only demand those bits not set by the OR.
546 if (I->getOpcode() == Instruction::Or)
547 Mask &= ~RHS->getRawValue();
548 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
550 // Walk the LHS and the RHS.
551 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1) ||
552 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
553 case Instruction::Cast: {
554 const Type *SrcTy = I->getOperand(0)->getType();
555 if (SrcTy == Type::BoolTy)
556 return SimplifyDemandedBits(I->getOperand(0), Mask&1, Depth+1);
558 if (!SrcTy->isInteger()) return false;
560 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
561 // If this is a sign-extend, treat specially.
562 if (SrcTy->isSigned() &&
563 SrcBits < I->getType()->getPrimitiveSizeInBits()) {
564 // If none of the top bits are demanded, convert this into an unsigned
565 // extend instead of a sign extend.
566 if ((Mask & ((1ULL << SrcBits)-1)) == 0) {
567 // Convert to unsigned first.
569 NewVal = new CastInst(I->getOperand(0), SrcTy->getUnsignedVersion(),
570 I->getOperand(0)->getName(), I);
571 NewVal = new CastInst(I->getOperand(0), I->getType(), I->getName());
572 return UpdateValueUsesWith(I, NewVal);
575 // Otherwise, the high-bits *are* demanded. This means that the code
576 // implicitly demands computation of the sign bit of the input, make sure
577 // we explicitly include it in Mask.
578 Mask |= 1ULL << (SrcBits-1);
581 // If this is an extension, the top bits are ignored.
582 Mask &= SrcTy->getIntegralTypeMask();
583 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
585 case Instruction::Select:
586 // Simplify the T and F values if they are not demanded.
587 return SimplifyDemandedBits(I->getOperand(2), Mask, Depth+1) ||
588 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
589 case Instruction::Shl:
590 // We only demand the low bits of the input.
591 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
592 return SimplifyDemandedBits(I->getOperand(0), Mask >> SA->getValue(),
595 case Instruction::Shr:
596 // We only demand the high bits of the input.
597 if (I->getType()->isUnsigned())
598 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
599 Mask <<= SA->getValue();
600 Mask &= I->getType()->getIntegralTypeMask();
601 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
603 // FIXME: handle signed shr, demanding the appropriate bits. If the top
604 // bits aren't demanded, strength reduce to a logical SHR instead.
610 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
611 // true when both operands are equal...
613 static bool isTrueWhenEqual(Instruction &I) {
614 return I.getOpcode() == Instruction::SetEQ ||
615 I.getOpcode() == Instruction::SetGE ||
616 I.getOpcode() == Instruction::SetLE;
619 /// AssociativeOpt - Perform an optimization on an associative operator. This
620 /// function is designed to check a chain of associative operators for a
621 /// potential to apply a certain optimization. Since the optimization may be
622 /// applicable if the expression was reassociated, this checks the chain, then
623 /// reassociates the expression as necessary to expose the optimization
624 /// opportunity. This makes use of a special Functor, which must define
625 /// 'shouldApply' and 'apply' methods.
627 template<typename Functor>
628 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
629 unsigned Opcode = Root.getOpcode();
630 Value *LHS = Root.getOperand(0);
632 // Quick check, see if the immediate LHS matches...
633 if (F.shouldApply(LHS))
634 return F.apply(Root);
636 // Otherwise, if the LHS is not of the same opcode as the root, return.
637 Instruction *LHSI = dyn_cast<Instruction>(LHS);
638 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
639 // Should we apply this transform to the RHS?
640 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
642 // If not to the RHS, check to see if we should apply to the LHS...
643 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
644 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
648 // If the functor wants to apply the optimization to the RHS of LHSI,
649 // reassociate the expression from ((? op A) op B) to (? op (A op B))
651 BasicBlock *BB = Root.getParent();
653 // Now all of the instructions are in the current basic block, go ahead
654 // and perform the reassociation.
655 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
657 // First move the selected RHS to the LHS of the root...
658 Root.setOperand(0, LHSI->getOperand(1));
660 // Make what used to be the LHS of the root be the user of the root...
661 Value *ExtraOperand = TmpLHSI->getOperand(1);
662 if (&Root == TmpLHSI) {
663 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
666 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
667 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
668 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
669 BasicBlock::iterator ARI = &Root; ++ARI;
670 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
673 // Now propagate the ExtraOperand down the chain of instructions until we
675 while (TmpLHSI != LHSI) {
676 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
677 // Move the instruction to immediately before the chain we are
678 // constructing to avoid breaking dominance properties.
679 NextLHSI->getParent()->getInstList().remove(NextLHSI);
680 BB->getInstList().insert(ARI, NextLHSI);
683 Value *NextOp = NextLHSI->getOperand(1);
684 NextLHSI->setOperand(1, ExtraOperand);
686 ExtraOperand = NextOp;
689 // Now that the instructions are reassociated, have the functor perform
690 // the transformation...
691 return F.apply(Root);
694 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
700 // AddRHS - Implements: X + X --> X << 1
703 AddRHS(Value *rhs) : RHS(rhs) {}
704 bool shouldApply(Value *LHS) const { return LHS == RHS; }
705 Instruction *apply(BinaryOperator &Add) const {
706 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
707 ConstantInt::get(Type::UByteTy, 1));
711 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
713 struct AddMaskingAnd {
715 AddMaskingAnd(Constant *c) : C2(c) {}
716 bool shouldApply(Value *LHS) const {
718 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
719 ConstantExpr::getAnd(C1, C2)->isNullValue();
721 Instruction *apply(BinaryOperator &Add) const {
722 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
726 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
728 if (isa<CastInst>(I)) {
729 if (Constant *SOC = dyn_cast<Constant>(SO))
730 return ConstantExpr::getCast(SOC, I.getType());
732 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
733 SO->getName() + ".cast"), I);
736 // Figure out if the constant is the left or the right argument.
737 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
738 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
740 if (Constant *SOC = dyn_cast<Constant>(SO)) {
742 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
743 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
746 Value *Op0 = SO, *Op1 = ConstOperand;
750 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
751 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
752 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
753 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
755 assert(0 && "Unknown binary instruction type!");
758 return IC->InsertNewInstBefore(New, I);
761 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
762 // constant as the other operand, try to fold the binary operator into the
763 // select arguments. This also works for Cast instructions, which obviously do
764 // not have a second operand.
765 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
767 // Don't modify shared select instructions
768 if (!SI->hasOneUse()) return 0;
769 Value *TV = SI->getOperand(1);
770 Value *FV = SI->getOperand(2);
772 if (isa<Constant>(TV) || isa<Constant>(FV)) {
773 // Bool selects with constant operands can be folded to logical ops.
774 if (SI->getType() == Type::BoolTy) return 0;
776 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
777 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
779 return new SelectInst(SI->getCondition(), SelectTrueVal,
786 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
787 /// node as operand #0, see if we can fold the instruction into the PHI (which
788 /// is only possible if all operands to the PHI are constants).
789 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
790 PHINode *PN = cast<PHINode>(I.getOperand(0));
791 unsigned NumPHIValues = PN->getNumIncomingValues();
792 if (!PN->hasOneUse() || NumPHIValues == 0 ||
793 !isa<Constant>(PN->getIncomingValue(0))) return 0;
795 // Check to see if all of the operands of the PHI are constants. If not, we
796 // cannot do the transformation.
797 for (unsigned i = 1; i != NumPHIValues; ++i)
798 if (!isa<Constant>(PN->getIncomingValue(i)))
801 // Okay, we can do the transformation: create the new PHI node.
802 PHINode *NewPN = new PHINode(I.getType(), I.getName());
804 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
805 InsertNewInstBefore(NewPN, *PN);
807 // Next, add all of the operands to the PHI.
808 if (I.getNumOperands() == 2) {
809 Constant *C = cast<Constant>(I.getOperand(1));
810 for (unsigned i = 0; i != NumPHIValues; ++i) {
811 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
812 NewPN->addIncoming(ConstantExpr::get(I.getOpcode(), InV, C),
813 PN->getIncomingBlock(i));
816 assert(isa<CastInst>(I) && "Unary op should be a cast!");
817 const Type *RetTy = I.getType();
818 for (unsigned i = 0; i != NumPHIValues; ++i) {
819 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
820 NewPN->addIncoming(ConstantExpr::getCast(InV, RetTy),
821 PN->getIncomingBlock(i));
824 return ReplaceInstUsesWith(I, NewPN);
827 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
828 bool Changed = SimplifyCommutative(I);
829 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
831 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
832 // X + undef -> undef
833 if (isa<UndefValue>(RHS))
834 return ReplaceInstUsesWith(I, RHS);
837 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
838 if (RHSC->isNullValue())
839 return ReplaceInstUsesWith(I, LHS);
840 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
841 if (CFP->isExactlyValue(-0.0))
842 return ReplaceInstUsesWith(I, LHS);
845 // X + (signbit) --> X ^ signbit
846 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
847 uint64_t Val = CI->getRawValue() & CI->getType()->getIntegralTypeMask();
848 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
849 return BinaryOperator::createXor(LHS, RHS);
852 if (isa<PHINode>(LHS))
853 if (Instruction *NV = FoldOpIntoPhi(I))
856 ConstantInt *XorRHS = 0;
858 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
859 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
860 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
861 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
863 uint64_t C0080Val = 1ULL << 31;
864 int64_t CFF80Val = -C0080Val;
867 if (TySizeBits > Size) {
869 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
870 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
871 if (RHSSExt == CFF80Val) {
872 if (XorRHS->getZExtValue() == C0080Val)
874 } else if (RHSZExt == C0080Val) {
875 if (XorRHS->getSExtValue() == CFF80Val)
879 // This is a sign extend if the top bits are known zero.
880 Constant *Mask = ConstantInt::getAllOnesValue(XorLHS->getType());
881 Mask = ConstantExpr::getShl(Mask,
882 ConstantInt::get(Type::UByteTy, 64-(TySizeBits-Size)));
883 if (!MaskedValueIsZero(XorLHS, cast<ConstantInt>(Mask)))
884 Size = 0; // Not a sign ext, but can't be any others either.
894 const Type *MiddleType = 0;
897 case 32: MiddleType = Type::IntTy; break;
898 case 16: MiddleType = Type::ShortTy; break;
899 case 8: MiddleType = Type::SByteTy; break;
902 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
903 InsertNewInstBefore(NewTrunc, I);
904 return new CastInst(NewTrunc, I.getType());
910 if (I.getType()->isInteger()) {
911 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
913 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
914 if (RHSI->getOpcode() == Instruction::Sub)
915 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
916 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
918 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
919 if (LHSI->getOpcode() == Instruction::Sub)
920 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
921 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
926 if (Value *V = dyn_castNegVal(LHS))
927 return BinaryOperator::createSub(RHS, V);
930 if (!isa<Constant>(RHS))
931 if (Value *V = dyn_castNegVal(RHS))
932 return BinaryOperator::createSub(LHS, V);
936 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
937 if (X == RHS) // X*C + X --> X * (C+1)
938 return BinaryOperator::createMul(RHS, AddOne(C2));
940 // X*C1 + X*C2 --> X * (C1+C2)
942 if (X == dyn_castFoldableMul(RHS, C1))
943 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
946 // X + X*C --> X * (C+1)
947 if (dyn_castFoldableMul(RHS, C2) == LHS)
948 return BinaryOperator::createMul(LHS, AddOne(C2));
951 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
952 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
953 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
955 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
957 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
958 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
959 return BinaryOperator::createSub(C, X);
962 // (X & FF00) + xx00 -> (X+xx00) & FF00
963 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
964 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
966 // See if all bits from the first bit set in the Add RHS up are included
967 // in the mask. First, get the rightmost bit.
968 uint64_t AddRHSV = CRHS->getRawValue();
970 // Form a mask of all bits from the lowest bit added through the top.
971 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
972 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
974 // See if the and mask includes all of these bits.
975 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
977 if (AddRHSHighBits == AddRHSHighBitsAnd) {
978 // Okay, the xform is safe. Insert the new add pronto.
979 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
981 return BinaryOperator::createAnd(NewAdd, C2);
986 // Try to fold constant add into select arguments.
987 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
988 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
992 return Changed ? &I : 0;
995 // isSignBit - Return true if the value represented by the constant only has the
996 // highest order bit set.
997 static bool isSignBit(ConstantInt *CI) {
998 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
999 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1002 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1004 static Value *RemoveNoopCast(Value *V) {
1005 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1006 const Type *CTy = CI->getType();
1007 const Type *OpTy = CI->getOperand(0)->getType();
1008 if (CTy->isInteger() && OpTy->isInteger()) {
1009 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1010 return RemoveNoopCast(CI->getOperand(0));
1011 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1012 return RemoveNoopCast(CI->getOperand(0));
1017 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1018 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1020 if (Op0 == Op1) // sub X, X -> 0
1021 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1023 // If this is a 'B = x-(-A)', change to B = x+A...
1024 if (Value *V = dyn_castNegVal(Op1))
1025 return BinaryOperator::createAdd(Op0, V);
1027 if (isa<UndefValue>(Op0))
1028 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1029 if (isa<UndefValue>(Op1))
1030 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1032 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1033 // Replace (-1 - A) with (~A)...
1034 if (C->isAllOnesValue())
1035 return BinaryOperator::createNot(Op1);
1037 // C - ~X == X + (1+C)
1039 if (match(Op1, m_Not(m_Value(X))))
1040 return BinaryOperator::createAdd(X,
1041 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1042 // -((uint)X >> 31) -> ((int)X >> 31)
1043 // -((int)X >> 31) -> ((uint)X >> 31)
1044 if (C->isNullValue()) {
1045 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1046 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1047 if (SI->getOpcode() == Instruction::Shr)
1048 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
1050 if (SI->getType()->isSigned())
1051 NewTy = SI->getType()->getUnsignedVersion();
1053 NewTy = SI->getType()->getSignedVersion();
1054 // Check to see if we are shifting out everything but the sign bit.
1055 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
1056 // Ok, the transformation is safe. Insert a cast of the incoming
1057 // value, then the new shift, then the new cast.
1058 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
1059 SI->getOperand(0)->getName());
1060 Value *InV = InsertNewInstBefore(FirstCast, I);
1061 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
1063 if (NewShift->getType() == I.getType())
1066 InV = InsertNewInstBefore(NewShift, I);
1067 return new CastInst(NewShift, I.getType());
1073 // Try to fold constant sub into select arguments.
1074 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1075 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1078 if (isa<PHINode>(Op0))
1079 if (Instruction *NV = FoldOpIntoPhi(I))
1083 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1084 if (Op1I->getOpcode() == Instruction::Add &&
1085 !Op0->getType()->isFloatingPoint()) {
1086 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1087 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1088 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1089 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1090 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1091 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1092 // C1-(X+C2) --> (C1-C2)-X
1093 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1094 Op1I->getOperand(0));
1098 if (Op1I->hasOneUse()) {
1099 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1100 // is not used by anyone else...
1102 if (Op1I->getOpcode() == Instruction::Sub &&
1103 !Op1I->getType()->isFloatingPoint()) {
1104 // Swap the two operands of the subexpr...
1105 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1106 Op1I->setOperand(0, IIOp1);
1107 Op1I->setOperand(1, IIOp0);
1109 // Create the new top level add instruction...
1110 return BinaryOperator::createAdd(Op0, Op1);
1113 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1115 if (Op1I->getOpcode() == Instruction::And &&
1116 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1117 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1120 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1121 return BinaryOperator::createAnd(Op0, NewNot);
1124 // -(X sdiv C) -> (X sdiv -C)
1125 if (Op1I->getOpcode() == Instruction::Div)
1126 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
1127 if (CSI->isNullValue())
1128 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1129 return BinaryOperator::createDiv(Op1I->getOperand(0),
1130 ConstantExpr::getNeg(DivRHS));
1132 // X - X*C --> X * (1-C)
1133 ConstantInt *C2 = 0;
1134 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1136 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1137 return BinaryOperator::createMul(Op0, CP1);
1142 if (!Op0->getType()->isFloatingPoint())
1143 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1144 if (Op0I->getOpcode() == Instruction::Add) {
1145 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1146 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1147 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1148 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1149 } else if (Op0I->getOpcode() == Instruction::Sub) {
1150 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1151 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1155 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1156 if (X == Op1) { // X*C - X --> X * (C-1)
1157 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1158 return BinaryOperator::createMul(Op1, CP1);
1161 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1162 if (X == dyn_castFoldableMul(Op1, C2))
1163 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1168 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1169 /// really just returns true if the most significant (sign) bit is set.
1170 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1171 if (RHS->getType()->isSigned()) {
1172 // True if source is LHS < 0 or LHS <= -1
1173 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1174 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1176 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1177 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1178 // the size of the integer type.
1179 if (Opcode == Instruction::SetGE)
1180 return RHSC->getValue() ==
1181 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1182 if (Opcode == Instruction::SetGT)
1183 return RHSC->getValue() ==
1184 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1189 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1190 bool Changed = SimplifyCommutative(I);
1191 Value *Op0 = I.getOperand(0);
1193 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1194 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1196 // Simplify mul instructions with a constant RHS...
1197 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1198 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1200 // ((X << C1)*C2) == (X * (C2 << C1))
1201 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1202 if (SI->getOpcode() == Instruction::Shl)
1203 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1204 return BinaryOperator::createMul(SI->getOperand(0),
1205 ConstantExpr::getShl(CI, ShOp));
1207 if (CI->isNullValue())
1208 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1209 if (CI->equalsInt(1)) // X * 1 == X
1210 return ReplaceInstUsesWith(I, Op0);
1211 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1212 return BinaryOperator::createNeg(Op0, I.getName());
1214 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1215 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1216 uint64_t C = Log2_64(Val);
1217 return new ShiftInst(Instruction::Shl, Op0,
1218 ConstantUInt::get(Type::UByteTy, C));
1220 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1221 if (Op1F->isNullValue())
1222 return ReplaceInstUsesWith(I, Op1);
1224 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1225 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1226 if (Op1F->getValue() == 1.0)
1227 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1230 // Try to fold constant mul into select arguments.
1231 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1232 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1235 if (isa<PHINode>(Op0))
1236 if (Instruction *NV = FoldOpIntoPhi(I))
1240 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1241 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1242 return BinaryOperator::createMul(Op0v, Op1v);
1244 // If one of the operands of the multiply is a cast from a boolean value, then
1245 // we know the bool is either zero or one, so this is a 'masking' multiply.
1246 // See if we can simplify things based on how the boolean was originally
1248 CastInst *BoolCast = 0;
1249 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1250 if (CI->getOperand(0)->getType() == Type::BoolTy)
1253 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1254 if (CI->getOperand(0)->getType() == Type::BoolTy)
1257 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1258 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1259 const Type *SCOpTy = SCIOp0->getType();
1261 // If the setcc is true iff the sign bit of X is set, then convert this
1262 // multiply into a shift/and combination.
1263 if (isa<ConstantInt>(SCIOp1) &&
1264 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1265 // Shift the X value right to turn it into "all signbits".
1266 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1267 SCOpTy->getPrimitiveSizeInBits()-1);
1268 if (SCIOp0->getType()->isUnsigned()) {
1269 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1270 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1271 SCIOp0->getName()), I);
1275 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1276 BoolCast->getOperand(0)->getName()+
1279 // If the multiply type is not the same as the source type, sign extend
1280 // or truncate to the multiply type.
1281 if (I.getType() != V->getType())
1282 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1284 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1285 return BinaryOperator::createAnd(V, OtherOp);
1290 return Changed ? &I : 0;
1293 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1294 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1296 if (isa<UndefValue>(Op0)) // undef / X -> 0
1297 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1298 if (isa<UndefValue>(Op1))
1299 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1301 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1303 if (RHS->equalsInt(1))
1304 return ReplaceInstUsesWith(I, Op0);
1307 if (RHS->isAllOnesValue())
1308 return BinaryOperator::createNeg(Op0);
1310 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1311 if (LHS->getOpcode() == Instruction::Div)
1312 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1313 // (X / C1) / C2 -> X / (C1*C2)
1314 return BinaryOperator::createDiv(LHS->getOperand(0),
1315 ConstantExpr::getMul(RHS, LHSRHS));
1318 // Check to see if this is an unsigned division with an exact power of 2,
1319 // if so, convert to a right shift.
1320 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1321 if (uint64_t Val = C->getValue()) // Don't break X / 0
1322 if (isPowerOf2_64(Val)) {
1323 uint64_t C = Log2_64(Val);
1324 return new ShiftInst(Instruction::Shr, Op0,
1325 ConstantUInt::get(Type::UByteTy, C));
1329 if (RHS->getType()->isSigned())
1330 if (Value *LHSNeg = dyn_castNegVal(Op0))
1331 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1333 if (!RHS->isNullValue()) {
1334 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1335 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1337 if (isa<PHINode>(Op0))
1338 if (Instruction *NV = FoldOpIntoPhi(I))
1343 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1344 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
1345 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1346 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1347 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1348 if (STO->getValue() == 0) { // Couldn't be this argument.
1349 I.setOperand(1, SFO);
1351 } else if (SFO->getValue() == 0) {
1352 I.setOperand(1, STO);
1356 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
1357 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
1358 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
1359 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
1360 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
1361 TC, SI->getName()+".t");
1362 TSI = InsertNewInstBefore(TSI, I);
1364 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
1365 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
1366 FC, SI->getName()+".f");
1367 FSI = InsertNewInstBefore(FSI, I);
1368 return new SelectInst(SI->getOperand(0), TSI, FSI);
1372 // 0 / X == 0, we don't need to preserve faults!
1373 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1374 if (LHS->equalsInt(0))
1375 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1377 if (I.getType()->isSigned()) {
1378 // If the top bits of both operands are zero (i.e. we can prove they are
1379 // unsigned inputs), turn this into a udiv.
1380 ConstantIntegral *MaskV = ConstantSInt::getMinValue(I.getType());
1381 if (MaskedValueIsZero(Op1, MaskV) && MaskedValueIsZero(Op0, MaskV)) {
1382 const Type *NTy = Op0->getType()->getUnsignedVersion();
1383 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1384 InsertNewInstBefore(LHS, I);
1386 if (Constant *R = dyn_cast<Constant>(Op1))
1387 RHS = ConstantExpr::getCast(R, NTy);
1389 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1390 Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
1391 InsertNewInstBefore(Div, I);
1392 return new CastInst(Div, I.getType());
1395 // Known to be an unsigned division.
1396 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1397 // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
1398 if (RHSI->getOpcode() == Instruction::Shl &&
1399 isa<ConstantUInt>(RHSI->getOperand(0))) {
1400 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1401 if (isPowerOf2_64(C1)) {
1402 unsigned C2 = Log2_64(C1);
1403 Value *Add = RHSI->getOperand(1);
1405 Constant *C2V = ConstantUInt::get(Add->getType(), C2);
1406 Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
1409 return new ShiftInst(Instruction::Shr, Op0, Add);
1419 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
1420 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1421 if (I.getType()->isSigned()) {
1422 if (Value *RHSNeg = dyn_castNegVal(Op1))
1423 if (!isa<ConstantSInt>(RHSNeg) ||
1424 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
1426 AddUsesToWorkList(I);
1427 I.setOperand(1, RHSNeg);
1431 // If the top bits of both operands are zero (i.e. we can prove they are
1432 // unsigned inputs), turn this into a urem.
1433 ConstantIntegral *MaskV = ConstantSInt::getMinValue(I.getType());
1434 if (MaskedValueIsZero(Op1, MaskV) && MaskedValueIsZero(Op0, MaskV)) {
1435 const Type *NTy = Op0->getType()->getUnsignedVersion();
1436 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1437 InsertNewInstBefore(LHS, I);
1439 if (Constant *R = dyn_cast<Constant>(Op1))
1440 RHS = ConstantExpr::getCast(R, NTy);
1442 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1443 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
1444 InsertNewInstBefore(Rem, I);
1445 return new CastInst(Rem, I.getType());
1449 if (isa<UndefValue>(Op0)) // undef % X -> 0
1450 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1451 if (isa<UndefValue>(Op1))
1452 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
1454 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1455 if (RHS->equalsInt(1)) // X % 1 == 0
1456 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1458 // Check to see if this is an unsigned remainder with an exact power of 2,
1459 // if so, convert to a bitwise and.
1460 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1461 if (uint64_t Val = C->getValue()) // Don't break X % 0 (divide by zero)
1462 if (!(Val & (Val-1))) // Power of 2
1463 return BinaryOperator::createAnd(Op0,
1464 ConstantUInt::get(I.getType(), Val-1));
1466 if (!RHS->isNullValue()) {
1467 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1468 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1470 if (isa<PHINode>(Op0))
1471 if (Instruction *NV = FoldOpIntoPhi(I))
1476 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1477 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
1478 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1479 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1480 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1481 if (STO->getValue() == 0) { // Couldn't be this argument.
1482 I.setOperand(1, SFO);
1484 } else if (SFO->getValue() == 0) {
1485 I.setOperand(1, STO);
1489 if (!(STO->getValue() & (STO->getValue()-1)) &&
1490 !(SFO->getValue() & (SFO->getValue()-1))) {
1491 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1492 SubOne(STO), SI->getName()+".t"), I);
1493 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1494 SubOne(SFO), SI->getName()+".f"), I);
1495 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
1499 // 0 % X == 0, we don't need to preserve faults!
1500 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1501 if (LHS->equalsInt(0))
1502 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1505 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1506 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
1507 if (I.getType()->isUnsigned() &&
1508 RHSI->getOpcode() == Instruction::Shl &&
1509 isa<ConstantUInt>(RHSI->getOperand(0))) {
1510 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1511 if (isPowerOf2_64(C1)) {
1512 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
1513 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
1515 return BinaryOperator::createAnd(Op0, Add);
1523 // isMaxValueMinusOne - return true if this is Max-1
1524 static bool isMaxValueMinusOne(const ConstantInt *C) {
1525 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1526 return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
1528 const ConstantSInt *CS = cast<ConstantSInt>(C);
1530 // Calculate 0111111111..11111
1531 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1532 int64_t Val = INT64_MAX; // All ones
1533 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
1534 return CS->getValue() == Val-1;
1537 // isMinValuePlusOne - return true if this is Min+1
1538 static bool isMinValuePlusOne(const ConstantInt *C) {
1539 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1540 return CU->getValue() == 1;
1542 const ConstantSInt *CS = cast<ConstantSInt>(C);
1544 // Calculate 1111111111000000000000
1545 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1546 int64_t Val = -1; // All ones
1547 Val <<= TypeBits-1; // Shift over to the right spot
1548 return CS->getValue() == Val+1;
1551 // isOneBitSet - Return true if there is exactly one bit set in the specified
1553 static bool isOneBitSet(const ConstantInt *CI) {
1554 uint64_t V = CI->getRawValue();
1555 return V && (V & (V-1)) == 0;
1558 #if 0 // Currently unused
1559 // isLowOnes - Return true if the constant is of the form 0+1+.
1560 static bool isLowOnes(const ConstantInt *CI) {
1561 uint64_t V = CI->getRawValue();
1563 // There won't be bits set in parts that the type doesn't contain.
1564 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1566 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1567 return U && V && (U & V) == 0;
1571 // isHighOnes - Return true if the constant is of the form 1+0+.
1572 // This is the same as lowones(~X).
1573 static bool isHighOnes(const ConstantInt *CI) {
1574 uint64_t V = ~CI->getRawValue();
1575 if (~V == 0) return false; // 0's does not match "1+"
1577 // There won't be bits set in parts that the type doesn't contain.
1578 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1580 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1581 return U && V && (U & V) == 0;
1585 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
1586 /// are carefully arranged to allow folding of expressions such as:
1588 /// (A < B) | (A > B) --> (A != B)
1590 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
1591 /// represents that the comparison is true if A == B, and bit value '1' is true
1594 static unsigned getSetCondCode(const SetCondInst *SCI) {
1595 switch (SCI->getOpcode()) {
1597 case Instruction::SetGT: return 1;
1598 case Instruction::SetEQ: return 2;
1599 case Instruction::SetGE: return 3;
1600 case Instruction::SetLT: return 4;
1601 case Instruction::SetNE: return 5;
1602 case Instruction::SetLE: return 6;
1605 assert(0 && "Invalid SetCC opcode!");
1610 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
1611 /// opcode and two operands into either a constant true or false, or a brand new
1612 /// SetCC instruction.
1613 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
1615 case 0: return ConstantBool::False;
1616 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
1617 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
1618 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
1619 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
1620 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
1621 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
1622 case 7: return ConstantBool::True;
1623 default: assert(0 && "Illegal SetCCCode!"); return 0;
1627 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
1628 struct FoldSetCCLogical {
1631 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
1632 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
1633 bool shouldApply(Value *V) const {
1634 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
1635 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
1636 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
1639 Instruction *apply(BinaryOperator &Log) const {
1640 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
1641 if (SCI->getOperand(0) != LHS) {
1642 assert(SCI->getOperand(1) == LHS);
1643 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
1646 unsigned LHSCode = getSetCondCode(SCI);
1647 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
1649 switch (Log.getOpcode()) {
1650 case Instruction::And: Code = LHSCode & RHSCode; break;
1651 case Instruction::Or: Code = LHSCode | RHSCode; break;
1652 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
1653 default: assert(0 && "Illegal logical opcode!"); return 0;
1656 Value *RV = getSetCCValue(Code, LHS, RHS);
1657 if (Instruction *I = dyn_cast<Instruction>(RV))
1659 // Otherwise, it's a constant boolean value...
1660 return IC.ReplaceInstUsesWith(Log, RV);
1664 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
1665 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
1666 // guaranteed to be either a shift instruction or a binary operator.
1667 Instruction *InstCombiner::OptAndOp(Instruction *Op,
1668 ConstantIntegral *OpRHS,
1669 ConstantIntegral *AndRHS,
1670 BinaryOperator &TheAnd) {
1671 Value *X = Op->getOperand(0);
1672 Constant *Together = 0;
1673 if (!isa<ShiftInst>(Op))
1674 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
1676 switch (Op->getOpcode()) {
1677 case Instruction::Xor:
1678 if (Op->hasOneUse()) {
1679 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1680 std::string OpName = Op->getName(); Op->setName("");
1681 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
1682 InsertNewInstBefore(And, TheAnd);
1683 return BinaryOperator::createXor(And, Together);
1686 case Instruction::Or:
1687 if (Together == AndRHS) // (X | C) & C --> C
1688 return ReplaceInstUsesWith(TheAnd, AndRHS);
1690 if (Op->hasOneUse() && Together != OpRHS) {
1691 // (X | C1) & C2 --> (X | (C1&C2)) & C2
1692 std::string Op0Name = Op->getName(); Op->setName("");
1693 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
1694 InsertNewInstBefore(Or, TheAnd);
1695 return BinaryOperator::createAnd(Or, AndRHS);
1698 case Instruction::Add:
1699 if (Op->hasOneUse()) {
1700 // Adding a one to a single bit bit-field should be turned into an XOR
1701 // of the bit. First thing to check is to see if this AND is with a
1702 // single bit constant.
1703 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
1705 // Clear bits that are not part of the constant.
1706 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
1708 // If there is only one bit set...
1709 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
1710 // Ok, at this point, we know that we are masking the result of the
1711 // ADD down to exactly one bit. If the constant we are adding has
1712 // no bits set below this bit, then we can eliminate the ADD.
1713 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
1715 // Check to see if any bits below the one bit set in AndRHSV are set.
1716 if ((AddRHS & (AndRHSV-1)) == 0) {
1717 // If not, the only thing that can effect the output of the AND is
1718 // the bit specified by AndRHSV. If that bit is set, the effect of
1719 // the XOR is to toggle the bit. If it is clear, then the ADD has
1721 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
1722 TheAnd.setOperand(0, X);
1725 std::string Name = Op->getName(); Op->setName("");
1726 // Pull the XOR out of the AND.
1727 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
1728 InsertNewInstBefore(NewAnd, TheAnd);
1729 return BinaryOperator::createXor(NewAnd, AndRHS);
1736 case Instruction::Shl: {
1737 // We know that the AND will not produce any of the bits shifted in, so if
1738 // the anded constant includes them, clear them now!
1740 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1741 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
1742 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
1744 if (CI == ShlMask) { // Masking out bits that the shift already masks
1745 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
1746 } else if (CI != AndRHS) { // Reducing bits set in and.
1747 TheAnd.setOperand(1, CI);
1752 case Instruction::Shr:
1753 // We know that the AND will not produce any of the bits shifted in, so if
1754 // the anded constant includes them, clear them now! This only applies to
1755 // unsigned shifts, because a signed shr may bring in set bits!
1757 if (AndRHS->getType()->isUnsigned()) {
1758 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1759 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
1760 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1762 if (CI == ShrMask) { // Masking out bits that the shift already masks.
1763 return ReplaceInstUsesWith(TheAnd, Op);
1764 } else if (CI != AndRHS) {
1765 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
1768 } else { // Signed shr.
1769 // See if this is shifting in some sign extension, then masking it out
1771 if (Op->hasOneUse()) {
1772 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1773 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
1774 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1775 if (CI == AndRHS) { // Masking out bits shifted in.
1776 // Make the argument unsigned.
1777 Value *ShVal = Op->getOperand(0);
1778 ShVal = InsertCastBefore(ShVal,
1779 ShVal->getType()->getUnsignedVersion(),
1781 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
1782 OpRHS, Op->getName()),
1784 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
1785 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
1788 return new CastInst(ShVal, Op->getType());
1798 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
1799 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
1800 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
1801 /// insert new instructions.
1802 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
1803 bool Inside, Instruction &IB) {
1804 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
1805 "Lo is not <= Hi in range emission code!");
1807 if (Lo == Hi) // Trivially false.
1808 return new SetCondInst(Instruction::SetNE, V, V);
1809 if (cast<ConstantIntegral>(Lo)->isMinValue())
1810 return new SetCondInst(Instruction::SetLT, V, Hi);
1812 Constant *AddCST = ConstantExpr::getNeg(Lo);
1813 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
1814 InsertNewInstBefore(Add, IB);
1815 // Convert to unsigned for the comparison.
1816 const Type *UnsType = Add->getType()->getUnsignedVersion();
1817 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1818 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1819 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1820 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
1823 if (Lo == Hi) // Trivially true.
1824 return new SetCondInst(Instruction::SetEQ, V, V);
1826 Hi = SubOne(cast<ConstantInt>(Hi));
1827 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
1828 return new SetCondInst(Instruction::SetGT, V, Hi);
1830 // Emit X-Lo > Hi-Lo-1
1831 Constant *AddCST = ConstantExpr::getNeg(Lo);
1832 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
1833 InsertNewInstBefore(Add, IB);
1834 // Convert to unsigned for the comparison.
1835 const Type *UnsType = Add->getType()->getUnsignedVersion();
1836 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1837 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1838 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1839 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
1842 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
1843 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
1844 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
1845 // not, since all 1s are not contiguous.
1846 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
1847 uint64_t V = Val->getRawValue();
1848 if (!isShiftedMask_64(V)) return false;
1850 // look for the first zero bit after the run of ones
1851 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
1852 // look for the first non-zero bit
1853 ME = 64-CountLeadingZeros_64(V);
1859 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
1860 /// where isSub determines whether the operator is a sub. If we can fold one of
1861 /// the following xforms:
1863 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
1864 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1865 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1867 /// return (A +/- B).
1869 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
1870 ConstantIntegral *Mask, bool isSub,
1872 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1873 if (!LHSI || LHSI->getNumOperands() != 2 ||
1874 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
1876 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
1878 switch (LHSI->getOpcode()) {
1880 case Instruction::And:
1881 if (ConstantExpr::getAnd(N, Mask) == Mask) {
1882 // If the AndRHS is a power of two minus one (0+1+), this is simple.
1883 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
1886 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
1887 // part, we don't need any explicit masks to take them out of A. If that
1888 // is all N is, ignore it.
1890 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
1891 Constant *Mask = ConstantInt::getAllOnesValue(RHS->getType());
1892 Mask = ConstantExpr::getUShr(Mask,
1893 ConstantInt::get(Type::UByteTy,
1895 if (MaskedValueIsZero(RHS, cast<ConstantIntegral>(Mask)))
1900 case Instruction::Or:
1901 case Instruction::Xor:
1902 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
1903 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
1904 ConstantExpr::getAnd(N, Mask)->isNullValue())
1911 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
1913 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
1914 return InsertNewInstBefore(New, I);
1917 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1918 bool Changed = SimplifyCommutative(I);
1919 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1921 if (isa<UndefValue>(Op1)) // X & undef -> 0
1922 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1926 return ReplaceInstUsesWith(I, Op1);
1928 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
1930 if (AndRHS->isAllOnesValue())
1931 return ReplaceInstUsesWith(I, Op0);
1933 // and (and X, c1), c2 -> and (x, c1&c2). Handle this case here, before
1934 // calling MaskedValueIsZero, to avoid inefficient cases where we traipse
1935 // through many levels of ands.
1937 Value *X = 0; ConstantInt *C1 = 0;
1938 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))))
1939 return BinaryOperator::createAnd(X, ConstantExpr::getAnd(C1, AndRHS));
1942 if (MaskedValueIsZero(Op0, AndRHS)) // LHS & RHS == 0
1943 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1945 // If the mask is not masking out any bits, there is no reason to do the
1946 // and in the first place.
1947 ConstantIntegral *NotAndRHS =
1948 cast<ConstantIntegral>(ConstantExpr::getNot(AndRHS));
1949 if (MaskedValueIsZero(Op0, NotAndRHS))
1950 return ReplaceInstUsesWith(I, Op0);
1952 // See if we can simplify any instructions used by the LHS whose sole
1953 // purpose is to compute bits we don't care about.
1954 if (SimplifyDemandedBits(Op0, AndRHS->getRawValue()))
1957 // Optimize a variety of ((val OP C1) & C2) combinations...
1958 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
1959 Instruction *Op0I = cast<Instruction>(Op0);
1960 Value *Op0LHS = Op0I->getOperand(0);
1961 Value *Op0RHS = Op0I->getOperand(1);
1962 switch (Op0I->getOpcode()) {
1963 case Instruction::Xor:
1964 case Instruction::Or:
1965 // (X ^ V) & C2 --> (X & C2) iff (V & C2) == 0
1966 // (X | V) & C2 --> (X & C2) iff (V & C2) == 0
1967 if (MaskedValueIsZero(Op0LHS, AndRHS))
1968 return BinaryOperator::createAnd(Op0RHS, AndRHS);
1969 if (MaskedValueIsZero(Op0RHS, AndRHS))
1970 return BinaryOperator::createAnd(Op0LHS, AndRHS);
1972 // If the mask is only needed on one incoming arm, push it up.
1973 if (Op0I->hasOneUse()) {
1974 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1975 // Not masking anything out for the LHS, move to RHS.
1976 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
1977 Op0RHS->getName()+".masked");
1978 InsertNewInstBefore(NewRHS, I);
1979 return BinaryOperator::create(
1980 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
1982 if (!isa<Constant>(NotAndRHS) &&
1983 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1984 // Not masking anything out for the RHS, move to LHS.
1985 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
1986 Op0LHS->getName()+".masked");
1987 InsertNewInstBefore(NewLHS, I);
1988 return BinaryOperator::create(
1989 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
1994 case Instruction::And:
1995 // (X & V) & C2 --> 0 iff (V & C2) == 0
1996 if (MaskedValueIsZero(Op0LHS, AndRHS) ||
1997 MaskedValueIsZero(Op0RHS, AndRHS))
1998 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2000 case Instruction::Add:
2001 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2002 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2003 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2004 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2005 return BinaryOperator::createAnd(V, AndRHS);
2006 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2007 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
2010 case Instruction::Sub:
2011 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2012 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2013 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2014 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2015 return BinaryOperator::createAnd(V, AndRHS);
2019 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2020 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2022 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2023 const Type *SrcTy = CI->getOperand(0)->getType();
2025 // If this is an integer truncation or change from signed-to-unsigned, and
2026 // if the source is an and/or with immediate, transform it. This
2027 // frequently occurs for bitfield accesses.
2028 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2029 if (SrcTy->getPrimitiveSizeInBits() >=
2030 I.getType()->getPrimitiveSizeInBits() &&
2031 CastOp->getNumOperands() == 2)
2032 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1)))
2033 if (CastOp->getOpcode() == Instruction::And) {
2034 // Change: and (cast (and X, C1) to T), C2
2035 // into : and (cast X to T), trunc(C1)&C2
2036 // This will folds the two ands together, which may allow other
2038 Instruction *NewCast =
2039 new CastInst(CastOp->getOperand(0), I.getType(),
2040 CastOp->getName()+".shrunk");
2041 NewCast = InsertNewInstBefore(NewCast, I);
2043 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2044 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
2045 return BinaryOperator::createAnd(NewCast, C3);
2046 } else if (CastOp->getOpcode() == Instruction::Or) {
2047 // Change: and (cast (or X, C1) to T), C2
2048 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2049 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2050 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
2051 return ReplaceInstUsesWith(I, AndRHS);
2056 // If this is an integer sign or zero extension instruction.
2057 if (SrcTy->isIntegral() &&
2058 SrcTy->getPrimitiveSizeInBits() <
2059 CI->getType()->getPrimitiveSizeInBits()) {
2061 if (SrcTy->isUnsigned()) {
2062 // See if this and is clearing out bits that are known to be zero
2063 // anyway (due to the zero extension).
2064 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
2065 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
2066 Constant *Result = ConstantExpr::getAnd(Mask, AndRHS);
2067 if (Result == Mask) // The "and" isn't doing anything, remove it.
2068 return ReplaceInstUsesWith(I, CI);
2069 if (Result != AndRHS) { // Reduce the and RHS constant.
2070 I.setOperand(1, Result);
2075 if (CI->hasOneUse() && SrcTy->isInteger()) {
2076 // We can only do this if all of the sign bits brought in are masked
2077 // out. Compute this by first getting 0000011111, then inverting
2079 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
2080 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
2081 Mask = ConstantExpr::getNot(Mask); // 1's in the new bits.
2082 if (ConstantExpr::getAnd(Mask, AndRHS)->isNullValue()) {
2083 // If the and is clearing all of the sign bits, change this to a
2084 // zero extension cast. To do this, cast the cast input to
2085 // unsigned, then to the requested size.
2086 Value *CastOp = CI->getOperand(0);
2088 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
2089 CI->getName()+".uns");
2090 NC = InsertNewInstBefore(NC, I);
2091 // Finally, insert a replacement for CI.
2092 NC = new CastInst(NC, CI->getType(), CI->getName());
2094 NC = InsertNewInstBefore(NC, I);
2095 WorkList.push_back(CI); // Delete CI later.
2096 I.setOperand(0, NC);
2097 return &I; // The AND operand was modified.
2104 // Try to fold constant and into select arguments.
2105 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2106 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2108 if (isa<PHINode>(Op0))
2109 if (Instruction *NV = FoldOpIntoPhi(I))
2113 Value *Op0NotVal = dyn_castNotVal(Op0);
2114 Value *Op1NotVal = dyn_castNotVal(Op1);
2116 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
2117 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2119 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2120 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2121 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
2122 I.getName()+".demorgan");
2123 InsertNewInstBefore(Or, I);
2124 return BinaryOperator::createNot(Or);
2127 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
2128 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2129 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2132 Value *LHSVal, *RHSVal;
2133 ConstantInt *LHSCst, *RHSCst;
2134 Instruction::BinaryOps LHSCC, RHSCC;
2135 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2136 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2137 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
2138 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2139 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2140 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2141 // Ensure that the larger constant is on the RHS.
2142 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2143 SetCondInst *LHS = cast<SetCondInst>(Op0);
2144 if (cast<ConstantBool>(Cmp)->getValue()) {
2145 std::swap(LHS, RHS);
2146 std::swap(LHSCst, RHSCst);
2147 std::swap(LHSCC, RHSCC);
2150 // At this point, we know we have have two setcc instructions
2151 // comparing a value against two constants and and'ing the result
2152 // together. Because of the above check, we know that we only have
2153 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2154 // FoldSetCCLogical check above), that the two constants are not
2156 assert(LHSCst != RHSCst && "Compares not folded above?");
2159 default: assert(0 && "Unknown integer condition code!");
2160 case Instruction::SetEQ:
2162 default: assert(0 && "Unknown integer condition code!");
2163 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
2164 case Instruction::SetGT: // (X == 13 & X > 15) -> false
2165 return ReplaceInstUsesWith(I, ConstantBool::False);
2166 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
2167 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
2168 return ReplaceInstUsesWith(I, LHS);
2170 case Instruction::SetNE:
2172 default: assert(0 && "Unknown integer condition code!");
2173 case Instruction::SetLT:
2174 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
2175 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
2176 break; // (X != 13 & X < 15) -> no change
2177 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
2178 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
2179 return ReplaceInstUsesWith(I, RHS);
2180 case Instruction::SetNE:
2181 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
2182 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2183 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2184 LHSVal->getName()+".off");
2185 InsertNewInstBefore(Add, I);
2186 const Type *UnsType = Add->getType()->getUnsignedVersion();
2187 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2188 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
2189 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2190 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2192 break; // (X != 13 & X != 15) -> no change
2195 case Instruction::SetLT:
2197 default: assert(0 && "Unknown integer condition code!");
2198 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
2199 case Instruction::SetGT: // (X < 13 & X > 15) -> false
2200 return ReplaceInstUsesWith(I, ConstantBool::False);
2201 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
2202 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
2203 return ReplaceInstUsesWith(I, LHS);
2205 case Instruction::SetGT:
2207 default: assert(0 && "Unknown integer condition code!");
2208 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
2209 return ReplaceInstUsesWith(I, LHS);
2210 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
2211 return ReplaceInstUsesWith(I, RHS);
2212 case Instruction::SetNE:
2213 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
2214 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
2215 break; // (X > 13 & X != 15) -> no change
2216 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
2217 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2223 return Changed ? &I : 0;
2226 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2227 bool Changed = SimplifyCommutative(I);
2228 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2230 if (isa<UndefValue>(Op1))
2231 return ReplaceInstUsesWith(I, // X | undef -> -1
2232 ConstantIntegral::getAllOnesValue(I.getType()));
2234 // or X, X = X or X, 0 == X
2235 if (Op0 == Op1 || Op1 == Constant::getNullValue(I.getType()))
2236 return ReplaceInstUsesWith(I, Op0);
2239 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2240 // If X is known to only contain bits that already exist in RHS, just
2241 // replace this instruction with RHS directly.
2242 if (MaskedValueIsZero(Op0,
2243 cast<ConstantIntegral>(ConstantExpr::getNot(RHS))))
2244 return ReplaceInstUsesWith(I, RHS);
2246 ConstantInt *C1 = 0; Value *X = 0;
2247 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2248 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2249 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
2251 InsertNewInstBefore(Or, I);
2252 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
2255 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2256 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2257 std::string Op0Name = Op0->getName(); Op0->setName("");
2258 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
2259 InsertNewInstBefore(Or, I);
2260 return BinaryOperator::createXor(Or,
2261 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
2264 // Try to fold constant and into select arguments.
2265 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2266 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2268 if (isa<PHINode>(Op0))
2269 if (Instruction *NV = FoldOpIntoPhi(I))
2273 Value *A = 0, *B = 0;
2274 ConstantInt *C1 = 0, *C2 = 0;
2276 if (match(Op0, m_And(m_Value(A), m_Value(B))))
2277 if (A == Op1 || B == Op1) // (A & ?) | A --> A
2278 return ReplaceInstUsesWith(I, Op1);
2279 if (match(Op1, m_And(m_Value(A), m_Value(B))))
2280 if (A == Op0 || B == Op0) // A | (A & ?) --> A
2281 return ReplaceInstUsesWith(I, Op0);
2283 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2284 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2285 MaskedValueIsZero(Op1, C1)) {
2286 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
2288 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2291 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2292 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2293 MaskedValueIsZero(Op0, C1)) {
2294 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
2296 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2299 // (A & C1)|(B & C2)
2300 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
2301 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
2303 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
2304 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
2307 // If we have: ((V + N) & C1) | (V & C2)
2308 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2309 // replace with V+N.
2310 if (C1 == ConstantExpr::getNot(C2)) {
2311 Value *V1 = 0, *V2 = 0;
2312 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
2313 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
2314 // Add commutes, try both ways.
2315 if (V1 == B && MaskedValueIsZero(V2, C2))
2316 return ReplaceInstUsesWith(I, A);
2317 if (V2 == B && MaskedValueIsZero(V1, C2))
2318 return ReplaceInstUsesWith(I, A);
2320 // Or commutes, try both ways.
2321 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
2322 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
2323 // Add commutes, try both ways.
2324 if (V1 == A && MaskedValueIsZero(V2, C1))
2325 return ReplaceInstUsesWith(I, B);
2326 if (V2 == A && MaskedValueIsZero(V1, C1))
2327 return ReplaceInstUsesWith(I, B);
2332 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
2333 if (A == Op1) // ~A | A == -1
2334 return ReplaceInstUsesWith(I,
2335 ConstantIntegral::getAllOnesValue(I.getType()));
2339 // Note, A is still live here!
2340 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
2342 return ReplaceInstUsesWith(I,
2343 ConstantIntegral::getAllOnesValue(I.getType()));
2345 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2346 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2347 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
2348 I.getName()+".demorgan"), I);
2349 return BinaryOperator::createNot(And);
2353 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
2354 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
2355 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2358 Value *LHSVal, *RHSVal;
2359 ConstantInt *LHSCst, *RHSCst;
2360 Instruction::BinaryOps LHSCC, RHSCC;
2361 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2362 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2363 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
2364 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2365 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2366 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2367 // Ensure that the larger constant is on the RHS.
2368 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2369 SetCondInst *LHS = cast<SetCondInst>(Op0);
2370 if (cast<ConstantBool>(Cmp)->getValue()) {
2371 std::swap(LHS, RHS);
2372 std::swap(LHSCst, RHSCst);
2373 std::swap(LHSCC, RHSCC);
2376 // At this point, we know we have have two setcc instructions
2377 // comparing a value against two constants and or'ing the result
2378 // together. Because of the above check, we know that we only have
2379 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2380 // FoldSetCCLogical check above), that the two constants are not
2382 assert(LHSCst != RHSCst && "Compares not folded above?");
2385 default: assert(0 && "Unknown integer condition code!");
2386 case Instruction::SetEQ:
2388 default: assert(0 && "Unknown integer condition code!");
2389 case Instruction::SetEQ:
2390 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
2391 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2392 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2393 LHSVal->getName()+".off");
2394 InsertNewInstBefore(Add, I);
2395 const Type *UnsType = Add->getType()->getUnsignedVersion();
2396 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2397 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
2398 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2399 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2401 break; // (X == 13 | X == 15) -> no change
2403 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
2405 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
2406 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
2407 return ReplaceInstUsesWith(I, RHS);
2410 case Instruction::SetNE:
2412 default: assert(0 && "Unknown integer condition code!");
2413 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
2414 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
2415 return ReplaceInstUsesWith(I, LHS);
2416 case Instruction::SetNE: // (X != 13 | X != 15) -> true
2417 case Instruction::SetLT: // (X != 13 | X < 15) -> true
2418 return ReplaceInstUsesWith(I, ConstantBool::True);
2421 case Instruction::SetLT:
2423 default: assert(0 && "Unknown integer condition code!");
2424 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
2426 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
2427 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
2428 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
2429 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
2430 return ReplaceInstUsesWith(I, RHS);
2433 case Instruction::SetGT:
2435 default: assert(0 && "Unknown integer condition code!");
2436 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
2437 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
2438 return ReplaceInstUsesWith(I, LHS);
2439 case Instruction::SetNE: // (X > 13 | X != 15) -> true
2440 case Instruction::SetLT: // (X > 13 | X < 15) -> true
2441 return ReplaceInstUsesWith(I, ConstantBool::True);
2447 return Changed ? &I : 0;
2450 // XorSelf - Implements: X ^ X --> 0
2453 XorSelf(Value *rhs) : RHS(rhs) {}
2454 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2455 Instruction *apply(BinaryOperator &Xor) const {
2461 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2462 bool Changed = SimplifyCommutative(I);
2463 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2465 if (isa<UndefValue>(Op1))
2466 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
2468 // xor X, X = 0, even if X is nested in a sequence of Xor's.
2469 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
2470 assert(Result == &I && "AssociativeOpt didn't work?");
2471 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2474 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2476 if (RHS->isNullValue())
2477 return ReplaceInstUsesWith(I, Op0);
2479 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2480 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
2481 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
2482 if (RHS == ConstantBool::True && SCI->hasOneUse())
2483 return new SetCondInst(SCI->getInverseCondition(),
2484 SCI->getOperand(0), SCI->getOperand(1));
2486 // ~(c-X) == X-c-1 == X+(-c-1)
2487 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2488 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2489 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2490 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2491 ConstantInt::get(I.getType(), 1));
2492 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
2495 // ~(~X & Y) --> (X | ~Y)
2496 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
2497 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
2498 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2500 BinaryOperator::createNot(Op0I->getOperand(1),
2501 Op0I->getOperand(1)->getName()+".not");
2502 InsertNewInstBefore(NotY, I);
2503 return BinaryOperator::createOr(Op0NotVal, NotY);
2507 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2508 switch (Op0I->getOpcode()) {
2509 case Instruction::Add:
2510 // ~(X-c) --> (-c-1)-X
2511 if (RHS->isAllOnesValue()) {
2512 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2513 return BinaryOperator::createSub(
2514 ConstantExpr::getSub(NegOp0CI,
2515 ConstantInt::get(I.getType(), 1)),
2516 Op0I->getOperand(0));
2519 case Instruction::And:
2520 // (X & C1) ^ C2 --> (X & C1) | C2 iff (C1&C2) == 0
2521 if (ConstantExpr::getAnd(RHS, Op0CI)->isNullValue())
2522 return BinaryOperator::createOr(Op0, RHS);
2524 case Instruction::Or:
2525 // (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
2526 if (ConstantExpr::getAnd(RHS, Op0CI) == RHS)
2527 return BinaryOperator::createAnd(Op0, ConstantExpr::getNot(RHS));
2533 // Try to fold constant and into select arguments.
2534 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2535 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2537 if (isa<PHINode>(Op0))
2538 if (Instruction *NV = FoldOpIntoPhi(I))
2542 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
2544 return ReplaceInstUsesWith(I,
2545 ConstantIntegral::getAllOnesValue(I.getType()));
2547 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
2549 return ReplaceInstUsesWith(I,
2550 ConstantIntegral::getAllOnesValue(I.getType()));
2552 if (Instruction *Op1I = dyn_cast<Instruction>(Op1))
2553 if (Op1I->getOpcode() == Instruction::Or) {
2554 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
2555 cast<BinaryOperator>(Op1I)->swapOperands();
2557 std::swap(Op0, Op1);
2558 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
2560 std::swap(Op0, Op1);
2562 } else if (Op1I->getOpcode() == Instruction::Xor) {
2563 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
2564 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
2565 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
2566 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
2569 if (Instruction *Op0I = dyn_cast<Instruction>(Op0))
2570 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
2571 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
2572 cast<BinaryOperator>(Op0I)->swapOperands();
2573 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
2574 Value *NotB = InsertNewInstBefore(BinaryOperator::createNot(Op1,
2575 Op1->getName()+".not"), I);
2576 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
2578 } else if (Op0I->getOpcode() == Instruction::Xor) {
2579 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
2580 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2581 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
2582 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2585 // (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
2586 ConstantInt *C1 = 0, *C2 = 0;
2587 if (match(Op0, m_And(m_Value(), m_ConstantInt(C1))) &&
2588 match(Op1, m_And(m_Value(), m_ConstantInt(C2))) &&
2589 ConstantExpr::getAnd(C1, C2)->isNullValue())
2590 return BinaryOperator::createOr(Op0, Op1);
2592 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
2593 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
2594 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2597 return Changed ? &I : 0;
2600 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
2601 /// overflowed for this type.
2602 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2604 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
2605 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
2608 static bool isPositive(ConstantInt *C) {
2609 return cast<ConstantSInt>(C)->getValue() >= 0;
2612 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
2613 /// overflowed for this type.
2614 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2616 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
2618 if (In1->getType()->isUnsigned())
2619 return cast<ConstantUInt>(Result)->getValue() <
2620 cast<ConstantUInt>(In1)->getValue();
2621 if (isPositive(In1) != isPositive(In2))
2623 if (isPositive(In1))
2624 return cast<ConstantSInt>(Result)->getValue() <
2625 cast<ConstantSInt>(In1)->getValue();
2626 return cast<ConstantSInt>(Result)->getValue() >
2627 cast<ConstantSInt>(In1)->getValue();
2630 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
2631 /// code necessary to compute the offset from the base pointer (without adding
2632 /// in the base pointer). Return the result as a signed integer of intptr size.
2633 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
2634 TargetData &TD = IC.getTargetData();
2635 gep_type_iterator GTI = gep_type_begin(GEP);
2636 const Type *UIntPtrTy = TD.getIntPtrType();
2637 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
2638 Value *Result = Constant::getNullValue(SIntPtrTy);
2640 // Build a mask for high order bits.
2641 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
2643 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
2644 Value *Op = GEP->getOperand(i);
2645 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
2646 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
2648 if (Constant *OpC = dyn_cast<Constant>(Op)) {
2649 if (!OpC->isNullValue()) {
2650 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
2651 Scale = ConstantExpr::getMul(OpC, Scale);
2652 if (Constant *RC = dyn_cast<Constant>(Result))
2653 Result = ConstantExpr::getAdd(RC, Scale);
2655 // Emit an add instruction.
2656 Result = IC.InsertNewInstBefore(
2657 BinaryOperator::createAdd(Result, Scale,
2658 GEP->getName()+".offs"), I);
2662 // Convert to correct type.
2663 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
2664 Op->getName()+".c"), I);
2666 // We'll let instcombine(mul) convert this to a shl if possible.
2667 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
2668 GEP->getName()+".idx"), I);
2670 // Emit an add instruction.
2671 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
2672 GEP->getName()+".offs"), I);
2678 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
2679 /// else. At this point we know that the GEP is on the LHS of the comparison.
2680 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
2681 Instruction::BinaryOps Cond,
2683 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
2685 if (CastInst *CI = dyn_cast<CastInst>(RHS))
2686 if (isa<PointerType>(CI->getOperand(0)->getType()))
2687 RHS = CI->getOperand(0);
2689 Value *PtrBase = GEPLHS->getOperand(0);
2690 if (PtrBase == RHS) {
2691 // As an optimization, we don't actually have to compute the actual value of
2692 // OFFSET if this is a seteq or setne comparison, just return whether each
2693 // index is zero or not.
2694 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
2695 Instruction *InVal = 0;
2696 gep_type_iterator GTI = gep_type_begin(GEPLHS);
2697 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
2699 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
2700 if (isa<UndefValue>(C)) // undef index -> undef.
2701 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2702 if (C->isNullValue())
2704 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
2705 EmitIt = false; // This is indexing into a zero sized array?
2706 } else if (isa<ConstantInt>(C))
2707 return ReplaceInstUsesWith(I, // No comparison is needed here.
2708 ConstantBool::get(Cond == Instruction::SetNE));
2713 new SetCondInst(Cond, GEPLHS->getOperand(i),
2714 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
2718 InVal = InsertNewInstBefore(InVal, I);
2719 InsertNewInstBefore(Comp, I);
2720 if (Cond == Instruction::SetNE) // True if any are unequal
2721 InVal = BinaryOperator::createOr(InVal, Comp);
2722 else // True if all are equal
2723 InVal = BinaryOperator::createAnd(InVal, Comp);
2731 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
2732 ConstantBool::get(Cond == Instruction::SetEQ));
2735 // Only lower this if the setcc is the only user of the GEP or if we expect
2736 // the result to fold to a constant!
2737 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
2738 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
2739 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
2740 return new SetCondInst(Cond, Offset,
2741 Constant::getNullValue(Offset->getType()));
2743 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
2744 // If the base pointers are different, but the indices are the same, just
2745 // compare the base pointer.
2746 if (PtrBase != GEPRHS->getOperand(0)) {
2747 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
2748 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
2749 GEPRHS->getOperand(0)->getType();
2751 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2752 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2753 IndicesTheSame = false;
2757 // If all indices are the same, just compare the base pointers.
2759 return new SetCondInst(Cond, GEPLHS->getOperand(0),
2760 GEPRHS->getOperand(0));
2762 // Otherwise, the base pointers are different and the indices are
2763 // different, bail out.
2767 // If one of the GEPs has all zero indices, recurse.
2768 bool AllZeros = true;
2769 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2770 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
2771 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
2776 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
2777 SetCondInst::getSwappedCondition(Cond), I);
2779 // If the other GEP has all zero indices, recurse.
2781 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2782 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
2783 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
2788 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
2790 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
2791 // If the GEPs only differ by one index, compare it.
2792 unsigned NumDifferences = 0; // Keep track of # differences.
2793 unsigned DiffOperand = 0; // The operand that differs.
2794 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2795 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2796 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
2797 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
2798 // Irreconcilable differences.
2802 if (NumDifferences++) break;
2807 if (NumDifferences == 0) // SAME GEP?
2808 return ReplaceInstUsesWith(I, // No comparison is needed here.
2809 ConstantBool::get(Cond == Instruction::SetEQ));
2810 else if (NumDifferences == 1) {
2811 Value *LHSV = GEPLHS->getOperand(DiffOperand);
2812 Value *RHSV = GEPRHS->getOperand(DiffOperand);
2814 // Convert the operands to signed values to make sure to perform a
2815 // signed comparison.
2816 const Type *NewTy = LHSV->getType()->getSignedVersion();
2817 if (LHSV->getType() != NewTy)
2818 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
2819 LHSV->getName()), I);
2820 if (RHSV->getType() != NewTy)
2821 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
2822 RHSV->getName()), I);
2823 return new SetCondInst(Cond, LHSV, RHSV);
2827 // Only lower this if the setcc is the only user of the GEP or if we expect
2828 // the result to fold to a constant!
2829 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
2830 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
2831 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
2832 Value *L = EmitGEPOffset(GEPLHS, I, *this);
2833 Value *R = EmitGEPOffset(GEPRHS, I, *this);
2834 return new SetCondInst(Cond, L, R);
2841 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
2842 bool Changed = SimplifyCommutative(I);
2843 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2844 const Type *Ty = Op0->getType();
2848 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
2850 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
2851 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
2853 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
2854 // addresses never equal each other! We already know that Op0 != Op1.
2855 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
2856 isa<ConstantPointerNull>(Op0)) &&
2857 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
2858 isa<ConstantPointerNull>(Op1)))
2859 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
2861 // setcc's with boolean values can always be turned into bitwise operations
2862 if (Ty == Type::BoolTy) {
2863 switch (I.getOpcode()) {
2864 default: assert(0 && "Invalid setcc instruction!");
2865 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
2866 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
2867 InsertNewInstBefore(Xor, I);
2868 return BinaryOperator::createNot(Xor);
2870 case Instruction::SetNE:
2871 return BinaryOperator::createXor(Op0, Op1);
2873 case Instruction::SetGT:
2874 std::swap(Op0, Op1); // Change setgt -> setlt
2876 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
2877 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2878 InsertNewInstBefore(Not, I);
2879 return BinaryOperator::createAnd(Not, Op1);
2881 case Instruction::SetGE:
2882 std::swap(Op0, Op1); // Change setge -> setle
2884 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
2885 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2886 InsertNewInstBefore(Not, I);
2887 return BinaryOperator::createOr(Not, Op1);
2892 // See if we are doing a comparison between a constant and an instruction that
2893 // can be folded into the comparison.
2894 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2895 // Check to see if we are comparing against the minimum or maximum value...
2896 if (CI->isMinValue()) {
2897 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
2898 return ReplaceInstUsesWith(I, ConstantBool::False);
2899 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
2900 return ReplaceInstUsesWith(I, ConstantBool::True);
2901 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
2902 return BinaryOperator::createSetEQ(Op0, Op1);
2903 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
2904 return BinaryOperator::createSetNE(Op0, Op1);
2906 } else if (CI->isMaxValue()) {
2907 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
2908 return ReplaceInstUsesWith(I, ConstantBool::False);
2909 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
2910 return ReplaceInstUsesWith(I, ConstantBool::True);
2911 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
2912 return BinaryOperator::createSetEQ(Op0, Op1);
2913 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
2914 return BinaryOperator::createSetNE(Op0, Op1);
2916 // Comparing against a value really close to min or max?
2917 } else if (isMinValuePlusOne(CI)) {
2918 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
2919 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
2920 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
2921 return BinaryOperator::createSetNE(Op0, SubOne(CI));
2923 } else if (isMaxValueMinusOne(CI)) {
2924 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
2925 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
2926 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
2927 return BinaryOperator::createSetNE(Op0, AddOne(CI));
2930 // If we still have a setle or setge instruction, turn it into the
2931 // appropriate setlt or setgt instruction. Since the border cases have
2932 // already been handled above, this requires little checking.
2934 if (I.getOpcode() == Instruction::SetLE)
2935 return BinaryOperator::createSetLT(Op0, AddOne(CI));
2936 if (I.getOpcode() == Instruction::SetGE)
2937 return BinaryOperator::createSetGT(Op0, SubOne(CI));
2939 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2940 switch (LHSI->getOpcode()) {
2941 case Instruction::And:
2942 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
2943 LHSI->getOperand(0)->hasOneUse()) {
2944 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
2945 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
2946 // happens a LOT in code produced by the C front-end, for bitfield
2948 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
2949 ConstantUInt *ShAmt;
2950 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
2951 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
2952 const Type *Ty = LHSI->getType();
2954 // We can fold this as long as we can't shift unknown bits
2955 // into the mask. This can only happen with signed shift
2956 // rights, as they sign-extend.
2958 bool CanFold = Shift->getOpcode() != Instruction::Shr ||
2959 Shift->getType()->isUnsigned();
2961 // To test for the bad case of the signed shr, see if any
2962 // of the bits shifted in could be tested after the mask.
2963 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
2964 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
2966 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
2968 ConstantExpr::getShl(ConstantInt::getAllOnesValue(Ty), OShAmt);
2969 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
2975 if (Shift->getOpcode() == Instruction::Shl)
2976 NewCst = ConstantExpr::getUShr(CI, ShAmt);
2978 NewCst = ConstantExpr::getShl(CI, ShAmt);
2980 // Check to see if we are shifting out any of the bits being
2982 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
2983 // If we shifted bits out, the fold is not going to work out.
2984 // As a special case, check to see if this means that the
2985 // result is always true or false now.
2986 if (I.getOpcode() == Instruction::SetEQ)
2987 return ReplaceInstUsesWith(I, ConstantBool::False);
2988 if (I.getOpcode() == Instruction::SetNE)
2989 return ReplaceInstUsesWith(I, ConstantBool::True);
2991 I.setOperand(1, NewCst);
2992 Constant *NewAndCST;
2993 if (Shift->getOpcode() == Instruction::Shl)
2994 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
2996 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
2997 LHSI->setOperand(1, NewAndCST);
2998 LHSI->setOperand(0, Shift->getOperand(0));
2999 WorkList.push_back(Shift); // Shift is dead.
3000 AddUsesToWorkList(I);
3008 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
3009 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3010 switch (I.getOpcode()) {
3012 case Instruction::SetEQ:
3013 case Instruction::SetNE: {
3014 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3016 // Check that the shift amount is in range. If not, don't perform
3017 // undefined shifts. When the shift is visited it will be
3019 if (ShAmt->getValue() >= TypeBits)
3022 // If we are comparing against bits always shifted out, the
3023 // comparison cannot succeed.
3025 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
3026 if (Comp != CI) {// Comparing against a bit that we know is zero.
3027 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3028 Constant *Cst = ConstantBool::get(IsSetNE);
3029 return ReplaceInstUsesWith(I, Cst);
3032 if (LHSI->hasOneUse()) {
3033 // Otherwise strength reduce the shift into an and.
3034 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3035 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
3038 if (CI->getType()->isUnsigned()) {
3039 Mask = ConstantUInt::get(CI->getType(), Val);
3040 } else if (ShAmtVal != 0) {
3041 Mask = ConstantSInt::get(CI->getType(), Val);
3043 Mask = ConstantInt::getAllOnesValue(CI->getType());
3047 BinaryOperator::createAnd(LHSI->getOperand(0),
3048 Mask, LHSI->getName()+".mask");
3049 Value *And = InsertNewInstBefore(AndI, I);
3050 return new SetCondInst(I.getOpcode(), And,
3051 ConstantExpr::getUShr(CI, ShAmt));
3058 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
3059 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3060 switch (I.getOpcode()) {
3062 case Instruction::SetEQ:
3063 case Instruction::SetNE: {
3065 // Check that the shift amount is in range. If not, don't perform
3066 // undefined shifts. When the shift is visited it will be
3068 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3069 if (ShAmt->getValue() >= TypeBits)
3072 // If we are comparing against bits always shifted out, the
3073 // comparison cannot succeed.
3075 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
3077 if (Comp != CI) {// Comparing against a bit that we know is zero.
3078 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3079 Constant *Cst = ConstantBool::get(IsSetNE);
3080 return ReplaceInstUsesWith(I, Cst);
3083 if (LHSI->hasOneUse() || CI->isNullValue()) {
3084 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3086 // Otherwise strength reduce the shift into an and.
3087 uint64_t Val = ~0ULL; // All ones.
3088 Val <<= ShAmtVal; // Shift over to the right spot.
3091 if (CI->getType()->isUnsigned()) {
3092 Val &= ~0ULL >> (64-TypeBits);
3093 Mask = ConstantUInt::get(CI->getType(), Val);
3095 Mask = ConstantSInt::get(CI->getType(), Val);
3099 BinaryOperator::createAnd(LHSI->getOperand(0),
3100 Mask, LHSI->getName()+".mask");
3101 Value *And = InsertNewInstBefore(AndI, I);
3102 return new SetCondInst(I.getOpcode(), And,
3103 ConstantExpr::getShl(CI, ShAmt));
3111 case Instruction::Div:
3112 // Fold: (div X, C1) op C2 -> range check
3113 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
3114 // Fold this div into the comparison, producing a range check.
3115 // Determine, based on the divide type, what the range is being
3116 // checked. If there is an overflow on the low or high side, remember
3117 // it, otherwise compute the range [low, hi) bounding the new value.
3118 bool LoOverflow = false, HiOverflow = 0;
3119 ConstantInt *LoBound = 0, *HiBound = 0;
3122 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
3124 Instruction::BinaryOps Opcode = I.getOpcode();
3126 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
3127 } else if (LHSI->getType()->isUnsigned()) { // udiv
3129 LoOverflow = ProdOV;
3130 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
3131 } else if (isPositive(DivRHS)) { // Divisor is > 0.
3132 if (CI->isNullValue()) { // (X / pos) op 0
3134 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
3136 } else if (isPositive(CI)) { // (X / pos) op pos
3138 LoOverflow = ProdOV;
3139 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
3140 } else { // (X / pos) op neg
3141 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
3142 LoOverflow = AddWithOverflow(LoBound, Prod,
3143 cast<ConstantInt>(DivRHSH));
3145 HiOverflow = ProdOV;
3147 } else { // Divisor is < 0.
3148 if (CI->isNullValue()) { // (X / neg) op 0
3149 LoBound = AddOne(DivRHS);
3150 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
3151 if (HiBound == DivRHS)
3152 LoBound = 0; // - INTMIN = INTMIN
3153 } else if (isPositive(CI)) { // (X / neg) op pos
3154 HiOverflow = LoOverflow = ProdOV;
3156 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
3157 HiBound = AddOne(Prod);
3158 } else { // (X / neg) op neg
3160 LoOverflow = HiOverflow = ProdOV;
3161 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
3164 // Dividing by a negate swaps the condition.
3165 Opcode = SetCondInst::getSwappedCondition(Opcode);
3169 Value *X = LHSI->getOperand(0);
3171 default: assert(0 && "Unhandled setcc opcode!");
3172 case Instruction::SetEQ:
3173 if (LoOverflow && HiOverflow)
3174 return ReplaceInstUsesWith(I, ConstantBool::False);
3175 else if (HiOverflow)
3176 return new SetCondInst(Instruction::SetGE, X, LoBound);
3177 else if (LoOverflow)
3178 return new SetCondInst(Instruction::SetLT, X, HiBound);
3180 return InsertRangeTest(X, LoBound, HiBound, true, I);
3181 case Instruction::SetNE:
3182 if (LoOverflow && HiOverflow)
3183 return ReplaceInstUsesWith(I, ConstantBool::True);
3184 else if (HiOverflow)
3185 return new SetCondInst(Instruction::SetLT, X, LoBound);
3186 else if (LoOverflow)
3187 return new SetCondInst(Instruction::SetGE, X, HiBound);
3189 return InsertRangeTest(X, LoBound, HiBound, false, I);
3190 case Instruction::SetLT:
3192 return ReplaceInstUsesWith(I, ConstantBool::False);
3193 return new SetCondInst(Instruction::SetLT, X, LoBound);
3194 case Instruction::SetGT:
3196 return ReplaceInstUsesWith(I, ConstantBool::False);
3197 return new SetCondInst(Instruction::SetGE, X, HiBound);
3204 // Simplify seteq and setne instructions...
3205 if (I.getOpcode() == Instruction::SetEQ ||
3206 I.getOpcode() == Instruction::SetNE) {
3207 bool isSetNE = I.getOpcode() == Instruction::SetNE;
3209 // If the first operand is (and|or|xor) with a constant, and the second
3210 // operand is a constant, simplify a bit.
3211 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
3212 switch (BO->getOpcode()) {
3213 case Instruction::Rem:
3214 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3215 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
3217 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
3218 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
3219 if (isPowerOf2_64(V)) {
3220 unsigned L2 = Log2_64(V);
3221 const Type *UTy = BO->getType()->getUnsignedVersion();
3222 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
3224 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
3225 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
3226 RHSCst, BO->getName()), I);
3227 return BinaryOperator::create(I.getOpcode(), NewRem,
3228 Constant::getNullValue(UTy));
3233 case Instruction::Add:
3234 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3235 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3236 if (BO->hasOneUse())
3237 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3238 ConstantExpr::getSub(CI, BOp1C));
3239 } else if (CI->isNullValue()) {
3240 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3241 // efficiently invertible, or if the add has just this one use.
3242 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3244 if (Value *NegVal = dyn_castNegVal(BOp1))
3245 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
3246 else if (Value *NegVal = dyn_castNegVal(BOp0))
3247 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
3248 else if (BO->hasOneUse()) {
3249 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
3251 InsertNewInstBefore(Neg, I);
3252 return new SetCondInst(I.getOpcode(), BOp0, Neg);
3256 case Instruction::Xor:
3257 // For the xor case, we can xor two constants together, eliminating
3258 // the explicit xor.
3259 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
3260 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
3261 ConstantExpr::getXor(CI, BOC));
3264 case Instruction::Sub:
3265 // Replace (([sub|xor] A, B) != 0) with (A != B)
3266 if (CI->isNullValue())
3267 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3271 case Instruction::Or:
3272 // If bits are being or'd in that are not present in the constant we
3273 // are comparing against, then the comparison could never succeed!
3274 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
3275 Constant *NotCI = ConstantExpr::getNot(CI);
3276 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
3277 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3281 case Instruction::And:
3282 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3283 // If bits are being compared against that are and'd out, then the
3284 // comparison can never succeed!
3285 if (!ConstantExpr::getAnd(CI,
3286 ConstantExpr::getNot(BOC))->isNullValue())
3287 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3289 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3290 if (CI == BOC && isOneBitSet(CI))
3291 return new SetCondInst(isSetNE ? Instruction::SetEQ :
3292 Instruction::SetNE, Op0,
3293 Constant::getNullValue(CI->getType()));
3295 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
3296 // to be a signed value as appropriate.
3297 if (isSignBit(BOC)) {
3298 Value *X = BO->getOperand(0);
3299 // If 'X' is not signed, insert a cast now...
3300 if (!BOC->getType()->isSigned()) {
3301 const Type *DestTy = BOC->getType()->getSignedVersion();
3302 X = InsertCastBefore(X, DestTy, I);
3304 return new SetCondInst(isSetNE ? Instruction::SetLT :
3305 Instruction::SetGE, X,
3306 Constant::getNullValue(X->getType()));
3309 // ((X & ~7) == 0) --> X < 8
3310 if (CI->isNullValue() && isHighOnes(BOC)) {
3311 Value *X = BO->getOperand(0);
3312 Constant *NegX = ConstantExpr::getNeg(BOC);
3314 // If 'X' is signed, insert a cast now.
3315 if (NegX->getType()->isSigned()) {
3316 const Type *DestTy = NegX->getType()->getUnsignedVersion();
3317 X = InsertCastBefore(X, DestTy, I);
3318 NegX = ConstantExpr::getCast(NegX, DestTy);
3321 return new SetCondInst(isSetNE ? Instruction::SetGE :
3322 Instruction::SetLT, X, NegX);
3329 } else { // Not a SetEQ/SetNE
3330 // If the LHS is a cast from an integral value of the same size,
3331 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
3332 Value *CastOp = Cast->getOperand(0);
3333 const Type *SrcTy = CastOp->getType();
3334 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
3335 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
3336 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
3337 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
3338 "Source and destination signednesses should differ!");
3339 if (Cast->getType()->isSigned()) {
3340 // If this is a signed comparison, check for comparisons in the
3341 // vicinity of zero.
3342 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
3344 return BinaryOperator::createSetGT(CastOp,
3345 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
3346 else if (I.getOpcode() == Instruction::SetGT &&
3347 cast<ConstantSInt>(CI)->getValue() == -1)
3348 // X > -1 => x < 128
3349 return BinaryOperator::createSetLT(CastOp,
3350 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
3352 ConstantUInt *CUI = cast<ConstantUInt>(CI);
3353 if (I.getOpcode() == Instruction::SetLT &&
3354 CUI->getValue() == 1ULL << (SrcTySize-1))
3355 // X < 128 => X > -1
3356 return BinaryOperator::createSetGT(CastOp,
3357 ConstantSInt::get(SrcTy, -1));
3358 else if (I.getOpcode() == Instruction::SetGT &&
3359 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
3361 return BinaryOperator::createSetLT(CastOp,
3362 Constant::getNullValue(SrcTy));
3369 // Handle setcc with constant RHS's that can be integer, FP or pointer.
3370 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3371 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3372 switch (LHSI->getOpcode()) {
3373 case Instruction::GetElementPtr:
3374 if (RHSC->isNullValue()) {
3375 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
3376 bool isAllZeros = true;
3377 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
3378 if (!isa<Constant>(LHSI->getOperand(i)) ||
3379 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
3384 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
3385 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3389 case Instruction::PHI:
3390 if (Instruction *NV = FoldOpIntoPhi(I))
3393 case Instruction::Select:
3394 // If either operand of the select is a constant, we can fold the
3395 // comparison into the select arms, which will cause one to be
3396 // constant folded and the select turned into a bitwise or.
3397 Value *Op1 = 0, *Op2 = 0;
3398 if (LHSI->hasOneUse()) {
3399 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3400 // Fold the known value into the constant operand.
3401 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3402 // Insert a new SetCC of the other select operand.
3403 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3404 LHSI->getOperand(2), RHSC,
3406 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3407 // Fold the known value into the constant operand.
3408 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3409 // Insert a new SetCC of the other select operand.
3410 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3411 LHSI->getOperand(1), RHSC,
3417 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
3422 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
3423 if (User *GEP = dyn_castGetElementPtr(Op0))
3424 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
3426 if (User *GEP = dyn_castGetElementPtr(Op1))
3427 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
3428 SetCondInst::getSwappedCondition(I.getOpcode()), I))
3431 // Test to see if the operands of the setcc are casted versions of other
3432 // values. If the cast can be stripped off both arguments, we do so now.
3433 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3434 Value *CastOp0 = CI->getOperand(0);
3435 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
3436 (isa<Constant>(Op1) || isa<CastInst>(Op1)) &&
3437 (I.getOpcode() == Instruction::SetEQ ||
3438 I.getOpcode() == Instruction::SetNE)) {
3439 // We keep moving the cast from the left operand over to the right
3440 // operand, where it can often be eliminated completely.
3443 // If operand #1 is a cast instruction, see if we can eliminate it as
3445 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
3446 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
3448 Op1 = CI2->getOperand(0);
3450 // If Op1 is a constant, we can fold the cast into the constant.
3451 if (Op1->getType() != Op0->getType())
3452 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3453 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
3455 // Otherwise, cast the RHS right before the setcc
3456 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
3457 InsertNewInstBefore(cast<Instruction>(Op1), I);
3459 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
3462 // Handle the special case of: setcc (cast bool to X), <cst>
3463 // This comes up when you have code like
3466 // For generality, we handle any zero-extension of any operand comparison
3467 // with a constant or another cast from the same type.
3468 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
3469 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
3472 return Changed ? &I : 0;
3475 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
3476 // We only handle extending casts so far.
3478 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
3479 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
3480 const Type *SrcTy = LHSCIOp->getType();
3481 const Type *DestTy = SCI.getOperand(0)->getType();
3484 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
3487 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
3488 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
3489 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
3491 // Is this a sign or zero extension?
3492 bool isSignSrc = SrcTy->isSigned();
3493 bool isSignDest = DestTy->isSigned();
3495 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
3496 // Not an extension from the same type?
3497 RHSCIOp = CI->getOperand(0);
3498 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
3499 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
3500 // Compute the constant that would happen if we truncated to SrcTy then
3501 // reextended to DestTy.
3502 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
3504 if (ConstantExpr::getCast(Res, DestTy) == CI) {
3507 // If the value cannot be represented in the shorter type, we cannot emit
3508 // a simple comparison.
3509 if (SCI.getOpcode() == Instruction::SetEQ)
3510 return ReplaceInstUsesWith(SCI, ConstantBool::False);
3511 if (SCI.getOpcode() == Instruction::SetNE)
3512 return ReplaceInstUsesWith(SCI, ConstantBool::True);
3514 // Evaluate the comparison for LT.
3516 if (DestTy->isSigned()) {
3517 // We're performing a signed comparison.
3519 // Signed extend and signed comparison.
3520 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
3521 Result = ConstantBool::False;
3523 Result = ConstantBool::True; // X < (large) --> true
3525 // Unsigned extend and signed comparison.
3526 if (cast<ConstantSInt>(CI)->getValue() < 0)
3527 Result = ConstantBool::False;
3529 Result = ConstantBool::True;
3532 // We're performing an unsigned comparison.
3534 // Unsigned extend & compare -> always true.
3535 Result = ConstantBool::True;
3537 // We're performing an unsigned comp with a sign extended value.
3538 // This is true if the input is >= 0. [aka >s -1]
3539 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
3540 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
3541 NegOne, SCI.getName()), SCI);
3545 // Finally, return the value computed.
3546 if (SCI.getOpcode() == Instruction::SetLT) {
3547 return ReplaceInstUsesWith(SCI, Result);
3549 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
3550 if (Constant *CI = dyn_cast<Constant>(Result))
3551 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
3553 return BinaryOperator::createNot(Result);
3560 // Okay, just insert a compare of the reduced operands now!
3561 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
3564 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
3565 assert(I.getOperand(1)->getType() == Type::UByteTy);
3566 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3567 bool isLeftShift = I.getOpcode() == Instruction::Shl;
3569 // shl X, 0 == X and shr X, 0 == X
3570 // shl 0, X == 0 and shr 0, X == 0
3571 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
3572 Op0 == Constant::getNullValue(Op0->getType()))
3573 return ReplaceInstUsesWith(I, Op0);
3575 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
3576 if (!isLeftShift && I.getType()->isSigned())
3577 return ReplaceInstUsesWith(I, Op0);
3578 else // undef << X -> 0 AND undef >>u X -> 0
3579 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3581 if (isa<UndefValue>(Op1)) {
3582 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
3583 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3585 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
3588 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
3590 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
3591 if (CSI->isAllOnesValue())
3592 return ReplaceInstUsesWith(I, CSI);
3594 // Try to fold constant and into select arguments.
3595 if (isa<Constant>(Op0))
3596 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3597 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3600 // See if we can turn a signed shr into an unsigned shr.
3601 if (!isLeftShift && I.getType()->isSigned()) {
3602 if (MaskedValueIsZero(Op0, ConstantInt::getMinValue(I.getType()))) {
3603 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
3604 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
3606 return new CastInst(V, I.getType());
3610 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
3611 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
3616 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
3618 bool isLeftShift = I.getOpcode() == Instruction::Shl;
3619 bool isSignedShift = Op0->getType()->isSigned();
3620 bool isUnsignedShift = !isSignedShift;
3622 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
3623 // of a signed value.
3625 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
3626 if (Op1->getValue() >= TypeBits) {
3627 if (isUnsignedShift || isLeftShift)
3628 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
3630 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
3635 // ((X*C1) << C2) == (X * (C1 << C2))
3636 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
3637 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
3638 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
3639 return BinaryOperator::createMul(BO->getOperand(0),
3640 ConstantExpr::getShl(BOOp, Op1));
3642 // Try to fold constant and into select arguments.
3643 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3644 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3646 if (isa<PHINode>(Op0))
3647 if (Instruction *NV = FoldOpIntoPhi(I))
3650 if (Op0->hasOneUse()) {
3651 // If this is a SHL of a sign-extending cast, see if we can turn the input
3652 // into a zero extending cast (a simple strength reduction).
3653 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3654 const Type *SrcTy = CI->getOperand(0)->getType();
3655 if (isLeftShift && SrcTy->isInteger() && SrcTy->isSigned() &&
3656 SrcTy->getPrimitiveSizeInBits() <
3657 CI->getType()->getPrimitiveSizeInBits()) {
3658 // We can change it to a zero extension if we are shifting out all of
3659 // the sign extended bits. To check this, form a mask of all of the
3660 // sign extend bits, then shift them left and see if we have anything
3662 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy); // 1111
3663 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType()); // 00001111
3664 Mask = ConstantExpr::getNot(Mask); // 1's in the sign bits: 11110000
3665 if (ConstantExpr::getShl(Mask, Op1)->isNullValue()) {
3666 // If the shift is nuking all of the sign bits, change this to a
3667 // zero extension cast. To do this, cast the cast input to
3668 // unsigned, then to the requested size.
3669 Value *CastOp = CI->getOperand(0);
3671 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
3672 CI->getName()+".uns");
3673 NC = InsertNewInstBefore(NC, I);
3674 // Finally, insert a replacement for CI.
3675 NC = new CastInst(NC, CI->getType(), CI->getName());
3677 NC = InsertNewInstBefore(NC, I);
3678 WorkList.push_back(CI); // Delete CI later.
3679 I.setOperand(0, NC);
3680 return &I; // The SHL operand was modified.
3685 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
3686 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3689 switch (Op0BO->getOpcode()) {
3691 case Instruction::Add:
3692 case Instruction::And:
3693 case Instruction::Or:
3694 case Instruction::Xor:
3695 // These operators commute.
3696 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
3697 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3698 match(Op0BO->getOperand(1),
3699 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
3700 Instruction *YS = new ShiftInst(Instruction::Shl,
3701 Op0BO->getOperand(0), Op1,
3703 InsertNewInstBefore(YS, I); // (Y << C)
3704 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3706 Op0BO->getOperand(1)->getName());
3707 InsertNewInstBefore(X, I); // (X + (Y << C))
3708 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3709 C2 = ConstantExpr::getShl(C2, Op1);
3710 return BinaryOperator::createAnd(X, C2);
3713 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
3714 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3715 match(Op0BO->getOperand(1),
3716 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3717 m_ConstantInt(CC))) && V2 == Op1 &&
3718 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
3719 Instruction *YS = new ShiftInst(Instruction::Shl,
3720 Op0BO->getOperand(0), Op1,
3722 InsertNewInstBefore(YS, I); // (Y << C)
3724 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
3725 V1->getName()+".mask");
3726 InsertNewInstBefore(XM, I); // X & (CC << C)
3728 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3732 case Instruction::Sub:
3733 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3734 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3735 match(Op0BO->getOperand(0),
3736 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
3737 Instruction *YS = new ShiftInst(Instruction::Shl,
3738 Op0BO->getOperand(1), Op1,
3740 InsertNewInstBefore(YS, I); // (Y << C)
3741 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3743 Op0BO->getOperand(0)->getName());
3744 InsertNewInstBefore(X, I); // (X + (Y << C))
3745 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3746 C2 = ConstantExpr::getShl(C2, Op1);
3747 return BinaryOperator::createAnd(X, C2);
3750 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3751 match(Op0BO->getOperand(0),
3752 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3753 m_ConstantInt(CC))) && V2 == Op1 &&
3754 cast<BinaryOperator>(Op0BO->getOperand(0))->getOperand(0)->hasOneUse()) {
3755 Instruction *YS = new ShiftInst(Instruction::Shl,
3756 Op0BO->getOperand(1), Op1,
3758 InsertNewInstBefore(YS, I); // (Y << C)
3760 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
3761 V1->getName()+".mask");
3762 InsertNewInstBefore(XM, I); // X & (CC << C)
3764 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3771 // If the operand is an bitwise operator with a constant RHS, and the
3772 // shift is the only use, we can pull it out of the shift.
3773 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
3774 bool isValid = true; // Valid only for And, Or, Xor
3775 bool highBitSet = false; // Transform if high bit of constant set?
3777 switch (Op0BO->getOpcode()) {
3778 default: isValid = false; break; // Do not perform transform!
3779 case Instruction::Add:
3780 isValid = isLeftShift;
3782 case Instruction::Or:
3783 case Instruction::Xor:
3786 case Instruction::And:
3791 // If this is a signed shift right, and the high bit is modified
3792 // by the logical operation, do not perform the transformation.
3793 // The highBitSet boolean indicates the value of the high bit of
3794 // the constant which would cause it to be modified for this
3797 if (isValid && !isLeftShift && isSignedShift) {
3798 uint64_t Val = Op0C->getRawValue();
3799 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
3803 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
3805 Instruction *NewShift =
3806 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
3809 InsertNewInstBefore(NewShift, I);
3811 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
3818 // Find out if this is a shift of a shift by a constant.
3819 ShiftInst *ShiftOp = 0;
3820 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
3822 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3823 // If this is a noop-integer case of a shift instruction, use the shift.
3824 if (CI->getOperand(0)->getType()->isInteger() &&
3825 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
3826 CI->getType()->getPrimitiveSizeInBits() &&
3827 isa<ShiftInst>(CI->getOperand(0))) {
3828 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
3832 if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
3833 // Find the operands and properties of the input shift. Note that the
3834 // signedness of the input shift may differ from the current shift if there
3835 // is a noop cast between the two.
3836 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
3837 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
3838 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
3840 ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
3842 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
3843 unsigned ShiftAmt2 = (unsigned)Op1->getValue();
3845 // Check for (A << c1) << c2 and (A >> c1) >> c2.
3846 if (isLeftShift == isShiftOfLeftShift) {
3847 // Do not fold these shifts if the first one is signed and the second one
3848 // is unsigned and this is a right shift. Further, don't do any folding
3850 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
3853 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
3854 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
3855 Amt = Op0->getType()->getPrimitiveSizeInBits();
3857 Value *Op = ShiftOp->getOperand(0);
3858 if (isShiftOfSignedShift != isSignedShift)
3859 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
3860 return new ShiftInst(I.getOpcode(), Op,
3861 ConstantUInt::get(Type::UByteTy, Amt));
3864 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
3865 // signed types, we can only support the (A >> c1) << c2 configuration,
3866 // because it can not turn an arbitrary bit of A into a sign bit.
3867 if (isUnsignedShift || isLeftShift) {
3868 // Calculate bitmask for what gets shifted off the edge.
3869 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
3871 C = ConstantExpr::getShl(C, ShiftAmt1C);
3873 C = ConstantExpr::getUShr(C, ShiftAmt1C);
3875 Value *Op = ShiftOp->getOperand(0);
3876 if (isShiftOfSignedShift != isSignedShift)
3877 Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
3880 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
3881 InsertNewInstBefore(Mask, I);
3883 // Figure out what flavor of shift we should use...
3884 if (ShiftAmt1 == ShiftAmt2) {
3885 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
3886 } else if (ShiftAmt1 < ShiftAmt2) {
3887 return new ShiftInst(I.getOpcode(), Mask,
3888 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
3889 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
3890 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
3891 // Make sure to emit an unsigned shift right, not a signed one.
3892 Mask = InsertNewInstBefore(new CastInst(Mask,
3893 Mask->getType()->getUnsignedVersion(),
3895 Mask = new ShiftInst(Instruction::Shr, Mask,
3896 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3897 InsertNewInstBefore(Mask, I);
3898 return new CastInst(Mask, I.getType());
3900 return new ShiftInst(ShiftOp->getOpcode(), Mask,
3901 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3904 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
3905 Op = InsertNewInstBefore(new CastInst(Mask,
3906 I.getType()->getSignedVersion(),
3907 Mask->getName()), I);
3908 Instruction *Shift =
3909 new ShiftInst(ShiftOp->getOpcode(), Op,
3910 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3911 InsertNewInstBefore(Shift, I);
3913 C = ConstantIntegral::getAllOnesValue(Shift->getType());
3914 C = ConstantExpr::getShl(C, Op1);
3915 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
3916 InsertNewInstBefore(Mask, I);
3917 return new CastInst(Mask, I.getType());
3920 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
3921 // this case, C1 == C2 and C1 is 8, 16, or 32.
3922 if (ShiftAmt1 == ShiftAmt2) {
3923 const Type *SExtType = 0;
3924 switch (ShiftAmt1) {
3925 case 8 : SExtType = Type::SByteTy; break;
3926 case 16: SExtType = Type::ShortTy; break;
3927 case 32: SExtType = Type::IntTy; break;
3931 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
3933 InsertNewInstBefore(NewTrunc, I);
3934 return new CastInst(NewTrunc, I.getType());
3949 /// getCastType - In the future, we will split the cast instruction into these
3950 /// various types. Until then, we have to do the analysis here.
3951 static CastType getCastType(const Type *Src, const Type *Dest) {
3952 assert(Src->isIntegral() && Dest->isIntegral() &&
3953 "Only works on integral types!");
3954 unsigned SrcSize = Src->getPrimitiveSizeInBits();
3955 unsigned DestSize = Dest->getPrimitiveSizeInBits();
3957 if (SrcSize == DestSize) return Noop;
3958 if (SrcSize > DestSize) return Truncate;
3959 if (Src->isSigned()) return Signext;
3964 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
3967 static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
3968 const Type *DstTy, TargetData *TD) {
3970 // It is legal to eliminate the instruction if casting A->B->A if the sizes
3971 // are identical and the bits don't get reinterpreted (for example
3972 // int->float->int would not be allowed).
3973 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
3976 // If we are casting between pointer and integer types, treat pointers as
3977 // integers of the appropriate size for the code below.
3978 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
3979 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
3980 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
3982 // Allow free casting and conversion of sizes as long as the sign doesn't
3984 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
3985 CastType FirstCast = getCastType(SrcTy, MidTy);
3986 CastType SecondCast = getCastType(MidTy, DstTy);
3988 // Capture the effect of these two casts. If the result is a legal cast,
3989 // the CastType is stored here, otherwise a special code is used.
3990 static const unsigned CastResult[] = {
3991 // First cast is noop
3993 // First cast is a truncate
3994 1, 1, 4, 4, // trunc->extend is not safe to eliminate
3995 // First cast is a sign ext
3996 2, 5, 2, 4, // signext->zeroext never ok
3997 // First cast is a zero ext
4001 unsigned Result = CastResult[FirstCast*4+SecondCast];
4003 default: assert(0 && "Illegal table value!");
4008 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
4009 // truncates, we could eliminate more casts.
4010 return (unsigned)getCastType(SrcTy, DstTy) == Result;
4012 return false; // Not possible to eliminate this here.
4014 // Sign or zero extend followed by truncate is always ok if the result
4015 // is a truncate or noop.
4016 CastType ResultCast = getCastType(SrcTy, DstTy);
4017 if (ResultCast == Noop || ResultCast == Truncate)
4019 // Otherwise we are still growing the value, we are only safe if the
4020 // result will match the sign/zeroextendness of the result.
4021 return ResultCast == FirstCast;
4025 // If this is a cast from 'float -> double -> integer', cast from
4026 // 'float -> integer' directly, as the value isn't changed by the
4027 // float->double conversion.
4028 if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
4029 DstTy->isIntegral() &&
4030 SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
4036 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
4037 if (V->getType() == Ty || isa<Constant>(V)) return false;
4038 if (const CastInst *CI = dyn_cast<CastInst>(V))
4039 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
4045 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
4046 /// InsertBefore instruction. This is specialized a bit to avoid inserting
4047 /// casts that are known to not do anything...
4049 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
4050 Instruction *InsertBefore) {
4051 if (V->getType() == DestTy) return V;
4052 if (Constant *C = dyn_cast<Constant>(V))
4053 return ConstantExpr::getCast(C, DestTy);
4055 CastInst *CI = new CastInst(V, DestTy, V->getName());
4056 InsertNewInstBefore(CI, *InsertBefore);
4060 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
4061 /// expression. If so, decompose it, returning some value X, such that Val is
4064 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
4066 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
4067 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
4068 Offset = CI->getValue();
4070 return ConstantUInt::get(Type::UIntTy, 0);
4071 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
4072 if (I->getNumOperands() == 2) {
4073 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
4074 if (I->getOpcode() == Instruction::Shl) {
4075 // This is a value scaled by '1 << the shift amt'.
4076 Scale = 1U << CUI->getValue();
4078 return I->getOperand(0);
4079 } else if (I->getOpcode() == Instruction::Mul) {
4080 // This value is scaled by 'CUI'.
4081 Scale = CUI->getValue();
4083 return I->getOperand(0);
4084 } else if (I->getOpcode() == Instruction::Add) {
4085 // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
4088 Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
4090 Offset += CUI->getValue();
4091 if (SubScale > 1 && (Offset % SubScale == 0)) {
4100 // Otherwise, we can't look past this.
4107 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
4108 /// try to eliminate the cast by moving the type information into the alloc.
4109 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
4110 AllocationInst &AI) {
4111 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
4112 if (!PTy) return 0; // Not casting the allocation to a pointer type.
4114 // Remove any uses of AI that are dead.
4115 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
4116 std::vector<Instruction*> DeadUsers;
4117 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
4118 Instruction *User = cast<Instruction>(*UI++);
4119 if (isInstructionTriviallyDead(User)) {
4120 while (UI != E && *UI == User)
4121 ++UI; // If this instruction uses AI more than once, don't break UI.
4123 // Add operands to the worklist.
4124 AddUsesToWorkList(*User);
4126 DEBUG(std::cerr << "IC: DCE: " << *User);
4128 User->eraseFromParent();
4129 removeFromWorkList(User);
4133 // Get the type really allocated and the type casted to.
4134 const Type *AllocElTy = AI.getAllocatedType();
4135 const Type *CastElTy = PTy->getElementType();
4136 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
4138 unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
4139 unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
4140 if (CastElTyAlign < AllocElTyAlign) return 0;
4142 // If the allocation has multiple uses, only promote it if we are strictly
4143 // increasing the alignment of the resultant allocation. If we keep it the
4144 // same, we open the door to infinite loops of various kinds.
4145 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
4147 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
4148 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
4149 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
4151 // See if we can satisfy the modulus by pulling a scale out of the array
4153 unsigned ArraySizeScale, ArrayOffset;
4154 Value *NumElements = // See if the array size is a decomposable linear expr.
4155 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
4157 // If we can now satisfy the modulus, by using a non-1 scale, we really can
4159 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
4160 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
4162 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
4167 Amt = ConstantUInt::get(Type::UIntTy, Scale);
4168 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
4169 Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
4170 else if (Scale != 1) {
4171 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
4172 Amt = InsertNewInstBefore(Tmp, AI);
4176 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
4177 Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
4178 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
4179 Amt = InsertNewInstBefore(Tmp, AI);
4182 std::string Name = AI.getName(); AI.setName("");
4183 AllocationInst *New;
4184 if (isa<MallocInst>(AI))
4185 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
4187 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
4188 InsertNewInstBefore(New, AI);
4190 // If the allocation has multiple uses, insert a cast and change all things
4191 // that used it to use the new cast. This will also hack on CI, but it will
4193 if (!AI.hasOneUse()) {
4194 AddUsesToWorkList(AI);
4195 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
4196 InsertNewInstBefore(NewCast, AI);
4197 AI.replaceAllUsesWith(NewCast);
4199 return ReplaceInstUsesWith(CI, New);
4203 // CastInst simplification
4205 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
4206 Value *Src = CI.getOperand(0);
4208 // If the user is casting a value to the same type, eliminate this cast
4210 if (CI.getType() == Src->getType())
4211 return ReplaceInstUsesWith(CI, Src);
4213 if (isa<UndefValue>(Src)) // cast undef -> undef
4214 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
4216 // If casting the result of another cast instruction, try to eliminate this
4219 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
4220 Value *A = CSrc->getOperand(0);
4221 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
4222 CI.getType(), TD)) {
4223 // This instruction now refers directly to the cast's src operand. This
4224 // has a good chance of making CSrc dead.
4225 CI.setOperand(0, CSrc->getOperand(0));
4229 // If this is an A->B->A cast, and we are dealing with integral types, try
4230 // to convert this into a logical 'and' instruction.
4232 if (A->getType()->isInteger() &&
4233 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
4234 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
4235 CSrc->getType()->getPrimitiveSizeInBits() <
4236 CI.getType()->getPrimitiveSizeInBits()&&
4237 A->getType()->getPrimitiveSizeInBits() ==
4238 CI.getType()->getPrimitiveSizeInBits()) {
4239 assert(CSrc->getType() != Type::ULongTy &&
4240 "Cannot have type bigger than ulong!");
4241 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
4242 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
4244 AndOp = ConstantExpr::getCast(AndOp, A->getType());
4245 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
4246 if (And->getType() != CI.getType()) {
4247 And->setName(CSrc->getName()+".mask");
4248 InsertNewInstBefore(And, CI);
4249 And = new CastInst(And, CI.getType());
4255 // If this is a cast to bool, turn it into the appropriate setne instruction.
4256 if (CI.getType() == Type::BoolTy)
4257 return BinaryOperator::createSetNE(CI.getOperand(0),
4258 Constant::getNullValue(CI.getOperand(0)->getType()));
4260 // See if we can simplify any instructions used by the LHS whose sole
4261 // purpose is to compute bits we don't care about.
4262 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral() &&
4263 SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask()))
4266 // If casting the result of a getelementptr instruction with no offset, turn
4267 // this into a cast of the original pointer!
4269 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
4270 bool AllZeroOperands = true;
4271 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
4272 if (!isa<Constant>(GEP->getOperand(i)) ||
4273 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
4274 AllZeroOperands = false;
4277 if (AllZeroOperands) {
4278 CI.setOperand(0, GEP->getOperand(0));
4283 // If we are casting a malloc or alloca to a pointer to a type of the same
4284 // size, rewrite the allocation instruction to allocate the "right" type.
4286 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
4287 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
4290 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
4291 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
4293 if (isa<PHINode>(Src))
4294 if (Instruction *NV = FoldOpIntoPhi(CI))
4297 // If the source value is an instruction with only this use, we can attempt to
4298 // propagate the cast into the instruction. Also, only handle integral types
4300 if (Instruction *SrcI = dyn_cast<Instruction>(Src))
4301 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
4302 CI.getType()->isInteger()) { // Don't mess with casts to bool here
4303 const Type *DestTy = CI.getType();
4304 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
4305 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
4307 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
4308 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
4310 switch (SrcI->getOpcode()) {
4311 case Instruction::Add:
4312 case Instruction::Mul:
4313 case Instruction::And:
4314 case Instruction::Or:
4315 case Instruction::Xor:
4316 // If we are discarding information, or just changing the sign, rewrite.
4317 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
4318 // Don't insert two casts if they cannot be eliminated. We allow two
4319 // casts to be inserted if the sizes are the same. This could only be
4320 // converting signedness, which is a noop.
4321 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
4322 !ValueRequiresCast(Op0, DestTy, TD)) {
4323 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4324 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
4325 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
4326 ->getOpcode(), Op0c, Op1c);
4330 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
4331 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
4332 Op1 == ConstantBool::True &&
4333 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
4334 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
4335 return BinaryOperator::createXor(New,
4336 ConstantInt::get(CI.getType(), 1));
4339 case Instruction::Shl:
4340 // Allow changing the sign of the source operand. Do not allow changing
4341 // the size of the shift, UNLESS the shift amount is a constant. We
4342 // mush not change variable sized shifts to a smaller size, because it
4343 // is undefined to shift more bits out than exist in the value.
4344 if (DestBitSize == SrcBitSize ||
4345 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
4346 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4347 return new ShiftInst(Instruction::Shl, Op0c, Op1);
4350 case Instruction::Shr:
4351 // If this is a signed shr, and if all bits shifted in are about to be
4352 // truncated off, turn it into an unsigned shr to allow greater
4354 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
4355 isa<ConstantInt>(Op1)) {
4356 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
4357 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
4358 // Convert to unsigned.
4359 Value *N1 = InsertOperandCastBefore(Op0,
4360 Op0->getType()->getUnsignedVersion(), &CI);
4361 // Insert the new shift, which is now unsigned.
4362 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
4363 Op1, Src->getName()), CI);
4364 return new CastInst(N1, CI.getType());
4369 case Instruction::SetNE:
4370 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
4371 if (Op1C->getRawValue() == 0) {
4372 // If the input only has the low bit set, simplify directly.
4374 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
4375 // cast (X != 0) to int --> X if X&~1 == 0
4376 if (MaskedValueIsZero(Op0, cast<ConstantIntegral>(Not1))) {
4377 if (CI.getType() == Op0->getType())
4378 return ReplaceInstUsesWith(CI, Op0);
4380 return new CastInst(Op0, CI.getType());
4383 // If the input is an and with a single bit, shift then simplify.
4384 ConstantInt *AndRHS;
4385 if (match(Op0, m_And(m_Value(), m_ConstantInt(AndRHS))))
4386 if (AndRHS->getRawValue() &&
4387 (AndRHS->getRawValue() & (AndRHS->getRawValue()-1)) == 0) {
4388 unsigned ShiftAmt = Log2_64(AndRHS->getRawValue());
4389 // Perform an unsigned shr by shiftamt. Convert input to
4390 // unsigned if it is signed.
4392 if (In->getType()->isSigned())
4393 In = InsertNewInstBefore(new CastInst(In,
4394 In->getType()->getUnsignedVersion(), In->getName()),CI);
4395 // Insert the shift to put the result in the low bit.
4396 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
4397 ConstantInt::get(Type::UByteTy, ShiftAmt),
4398 In->getName()+".lobit"), CI);
4399 if (CI.getType() == In->getType())
4400 return ReplaceInstUsesWith(CI, In);
4402 return new CastInst(In, CI.getType());
4407 case Instruction::SetEQ:
4408 // We if we are just checking for a seteq of a single bit and casting it
4409 // to an integer. If so, shift the bit to the appropriate place then
4410 // cast to integer to avoid the comparison.
4411 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
4412 // Is Op1C a power of two or zero?
4413 if ((Op1C->getRawValue() & Op1C->getRawValue()-1) == 0) {
4414 // cast (X == 1) to int -> X iff X has only the low bit set.
4415 if (Op1C->getRawValue() == 1) {
4417 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
4418 if (MaskedValueIsZero(Op0, cast<ConstantIntegral>(Not1))) {
4419 if (CI.getType() == Op0->getType())
4420 return ReplaceInstUsesWith(CI, Op0);
4422 return new CastInst(Op0, CI.getType());
4434 /// GetSelectFoldableOperands - We want to turn code that looks like this:
4436 /// %D = select %cond, %C, %A
4438 /// %C = select %cond, %B, 0
4441 /// Assuming that the specified instruction is an operand to the select, return
4442 /// a bitmask indicating which operands of this instruction are foldable if they
4443 /// equal the other incoming value of the select.
4445 static unsigned GetSelectFoldableOperands(Instruction *I) {
4446 switch (I->getOpcode()) {
4447 case Instruction::Add:
4448 case Instruction::Mul:
4449 case Instruction::And:
4450 case Instruction::Or:
4451 case Instruction::Xor:
4452 return 3; // Can fold through either operand.
4453 case Instruction::Sub: // Can only fold on the amount subtracted.
4454 case Instruction::Shl: // Can only fold on the shift amount.
4455 case Instruction::Shr:
4458 return 0; // Cannot fold
4462 /// GetSelectFoldableConstant - For the same transformation as the previous
4463 /// function, return the identity constant that goes into the select.
4464 static Constant *GetSelectFoldableConstant(Instruction *I) {
4465 switch (I->getOpcode()) {
4466 default: assert(0 && "This cannot happen!"); abort();
4467 case Instruction::Add:
4468 case Instruction::Sub:
4469 case Instruction::Or:
4470 case Instruction::Xor:
4471 return Constant::getNullValue(I->getType());
4472 case Instruction::Shl:
4473 case Instruction::Shr:
4474 return Constant::getNullValue(Type::UByteTy);
4475 case Instruction::And:
4476 return ConstantInt::getAllOnesValue(I->getType());
4477 case Instruction::Mul:
4478 return ConstantInt::get(I->getType(), 1);
4482 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
4483 /// have the same opcode and only one use each. Try to simplify this.
4484 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
4486 if (TI->getNumOperands() == 1) {
4487 // If this is a non-volatile load or a cast from the same type,
4489 if (TI->getOpcode() == Instruction::Cast) {
4490 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
4493 return 0; // unknown unary op.
4496 // Fold this by inserting a select from the input values.
4497 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
4498 FI->getOperand(0), SI.getName()+".v");
4499 InsertNewInstBefore(NewSI, SI);
4500 return new CastInst(NewSI, TI->getType());
4503 // Only handle binary operators here.
4504 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
4507 // Figure out if the operations have any operands in common.
4508 Value *MatchOp, *OtherOpT, *OtherOpF;
4510 if (TI->getOperand(0) == FI->getOperand(0)) {
4511 MatchOp = TI->getOperand(0);
4512 OtherOpT = TI->getOperand(1);
4513 OtherOpF = FI->getOperand(1);
4514 MatchIsOpZero = true;
4515 } else if (TI->getOperand(1) == FI->getOperand(1)) {
4516 MatchOp = TI->getOperand(1);
4517 OtherOpT = TI->getOperand(0);
4518 OtherOpF = FI->getOperand(0);
4519 MatchIsOpZero = false;
4520 } else if (!TI->isCommutative()) {
4522 } else if (TI->getOperand(0) == FI->getOperand(1)) {
4523 MatchOp = TI->getOperand(0);
4524 OtherOpT = TI->getOperand(1);
4525 OtherOpF = FI->getOperand(0);
4526 MatchIsOpZero = true;
4527 } else if (TI->getOperand(1) == FI->getOperand(0)) {
4528 MatchOp = TI->getOperand(1);
4529 OtherOpT = TI->getOperand(0);
4530 OtherOpF = FI->getOperand(1);
4531 MatchIsOpZero = true;
4536 // If we reach here, they do have operations in common.
4537 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
4538 OtherOpF, SI.getName()+".v");
4539 InsertNewInstBefore(NewSI, SI);
4541 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
4543 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
4545 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
4548 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
4550 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
4554 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
4555 Value *CondVal = SI.getCondition();
4556 Value *TrueVal = SI.getTrueValue();
4557 Value *FalseVal = SI.getFalseValue();
4559 // select true, X, Y -> X
4560 // select false, X, Y -> Y
4561 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
4562 if (C == ConstantBool::True)
4563 return ReplaceInstUsesWith(SI, TrueVal);
4565 assert(C == ConstantBool::False);
4566 return ReplaceInstUsesWith(SI, FalseVal);
4569 // select C, X, X -> X
4570 if (TrueVal == FalseVal)
4571 return ReplaceInstUsesWith(SI, TrueVal);
4573 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
4574 return ReplaceInstUsesWith(SI, FalseVal);
4575 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
4576 return ReplaceInstUsesWith(SI, TrueVal);
4577 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
4578 if (isa<Constant>(TrueVal))
4579 return ReplaceInstUsesWith(SI, TrueVal);
4581 return ReplaceInstUsesWith(SI, FalseVal);
4584 if (SI.getType() == Type::BoolTy)
4585 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
4586 if (C == ConstantBool::True) {
4587 // Change: A = select B, true, C --> A = or B, C
4588 return BinaryOperator::createOr(CondVal, FalseVal);
4590 // Change: A = select B, false, C --> A = and !B, C
4592 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4593 "not."+CondVal->getName()), SI);
4594 return BinaryOperator::createAnd(NotCond, FalseVal);
4596 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
4597 if (C == ConstantBool::False) {
4598 // Change: A = select B, C, false --> A = and B, C
4599 return BinaryOperator::createAnd(CondVal, TrueVal);
4601 // Change: A = select B, C, true --> A = or !B, C
4603 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4604 "not."+CondVal->getName()), SI);
4605 return BinaryOperator::createOr(NotCond, TrueVal);
4609 // Selecting between two integer constants?
4610 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
4611 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
4612 // select C, 1, 0 -> cast C to int
4613 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
4614 return new CastInst(CondVal, SI.getType());
4615 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
4616 // select C, 0, 1 -> cast !C to int
4618 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4619 "not."+CondVal->getName()), SI);
4620 return new CastInst(NotCond, SI.getType());
4623 // If one of the constants is zero (we know they can't both be) and we
4624 // have a setcc instruction with zero, and we have an 'and' with the
4625 // non-constant value, eliminate this whole mess. This corresponds to
4626 // cases like this: ((X & 27) ? 27 : 0)
4627 if (TrueValC->isNullValue() || FalseValC->isNullValue())
4628 if (Instruction *IC = dyn_cast<Instruction>(SI.getCondition()))
4629 if ((IC->getOpcode() == Instruction::SetEQ ||
4630 IC->getOpcode() == Instruction::SetNE) &&
4631 isa<ConstantInt>(IC->getOperand(1)) &&
4632 cast<Constant>(IC->getOperand(1))->isNullValue())
4633 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
4634 if (ICA->getOpcode() == Instruction::And &&
4635 isa<ConstantInt>(ICA->getOperand(1)) &&
4636 (ICA->getOperand(1) == TrueValC ||
4637 ICA->getOperand(1) == FalseValC) &&
4638 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
4639 // Okay, now we know that everything is set up, we just don't
4640 // know whether we have a setne or seteq and whether the true or
4641 // false val is the zero.
4642 bool ShouldNotVal = !TrueValC->isNullValue();
4643 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
4646 V = InsertNewInstBefore(BinaryOperator::create(
4647 Instruction::Xor, V, ICA->getOperand(1)), SI);
4648 return ReplaceInstUsesWith(SI, V);
4652 // See if we are selecting two values based on a comparison of the two values.
4653 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
4654 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
4655 // Transform (X == Y) ? X : Y -> Y
4656 if (SCI->getOpcode() == Instruction::SetEQ)
4657 return ReplaceInstUsesWith(SI, FalseVal);
4658 // Transform (X != Y) ? X : Y -> X
4659 if (SCI->getOpcode() == Instruction::SetNE)
4660 return ReplaceInstUsesWith(SI, TrueVal);
4661 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4663 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
4664 // Transform (X == Y) ? Y : X -> X
4665 if (SCI->getOpcode() == Instruction::SetEQ)
4666 return ReplaceInstUsesWith(SI, FalseVal);
4667 // Transform (X != Y) ? Y : X -> Y
4668 if (SCI->getOpcode() == Instruction::SetNE)
4669 return ReplaceInstUsesWith(SI, TrueVal);
4670 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4674 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
4675 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
4676 if (TI->hasOneUse() && FI->hasOneUse()) {
4677 bool isInverse = false;
4678 Instruction *AddOp = 0, *SubOp = 0;
4680 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
4681 if (TI->getOpcode() == FI->getOpcode())
4682 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
4685 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
4686 // even legal for FP.
4687 if (TI->getOpcode() == Instruction::Sub &&
4688 FI->getOpcode() == Instruction::Add) {
4689 AddOp = FI; SubOp = TI;
4690 } else if (FI->getOpcode() == Instruction::Sub &&
4691 TI->getOpcode() == Instruction::Add) {
4692 AddOp = TI; SubOp = FI;
4696 Value *OtherAddOp = 0;
4697 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
4698 OtherAddOp = AddOp->getOperand(1);
4699 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
4700 OtherAddOp = AddOp->getOperand(0);
4704 // So at this point we know we have:
4705 // select C, (add X, Y), (sub X, ?)
4706 // We can do the transform profitably if either 'Y' = '?' or '?' is
4708 if (SubOp->getOperand(1) == AddOp ||
4709 isa<Constant>(SubOp->getOperand(1))) {
4711 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
4712 NegVal = ConstantExpr::getNeg(C);
4714 NegVal = InsertNewInstBefore(
4715 BinaryOperator::createNeg(SubOp->getOperand(1)), SI);
4718 Value *NewTrueOp = OtherAddOp;
4719 Value *NewFalseOp = NegVal;
4721 std::swap(NewTrueOp, NewFalseOp);
4722 Instruction *NewSel =
4723 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
4725 NewSel = InsertNewInstBefore(NewSel, SI);
4726 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
4732 // See if we can fold the select into one of our operands.
4733 if (SI.getType()->isInteger()) {
4734 // See the comment above GetSelectFoldableOperands for a description of the
4735 // transformation we are doing here.
4736 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
4737 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
4738 !isa<Constant>(FalseVal))
4739 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
4740 unsigned OpToFold = 0;
4741 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
4743 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
4748 Constant *C = GetSelectFoldableConstant(TVI);
4749 std::string Name = TVI->getName(); TVI->setName("");
4750 Instruction *NewSel =
4751 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
4753 InsertNewInstBefore(NewSel, SI);
4754 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
4755 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
4756 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
4757 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
4759 assert(0 && "Unknown instruction!!");
4764 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
4765 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
4766 !isa<Constant>(TrueVal))
4767 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
4768 unsigned OpToFold = 0;
4769 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
4771 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
4776 Constant *C = GetSelectFoldableConstant(FVI);
4777 std::string Name = FVI->getName(); FVI->setName("");
4778 Instruction *NewSel =
4779 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
4781 InsertNewInstBefore(NewSel, SI);
4782 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
4783 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
4784 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
4785 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
4787 assert(0 && "Unknown instruction!!");
4793 if (BinaryOperator::isNot(CondVal)) {
4794 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
4795 SI.setOperand(1, FalseVal);
4796 SI.setOperand(2, TrueVal);
4804 /// visitCallInst - CallInst simplification. This mostly only handles folding
4805 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
4806 /// the heavy lifting.
4808 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
4809 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
4810 if (!II) return visitCallSite(&CI);
4812 // Intrinsics cannot occur in an invoke, so handle them here instead of in
4814 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
4815 bool Changed = false;
4817 // memmove/cpy/set of zero bytes is a noop.
4818 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
4819 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
4821 // FIXME: Increase alignment here.
4823 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
4824 if (CI->getRawValue() == 1) {
4825 // Replace the instruction with just byte operations. We would
4826 // transform other cases to loads/stores, but we don't know if
4827 // alignment is sufficient.
4831 // If we have a memmove and the source operation is a constant global,
4832 // then the source and dest pointers can't alias, so we can change this
4833 // into a call to memcpy.
4834 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II))
4835 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
4836 if (GVSrc->isConstant()) {
4837 Module *M = CI.getParent()->getParent()->getParent();
4838 Function *MemCpy = M->getOrInsertFunction("llvm.memcpy",
4839 CI.getCalledFunction()->getFunctionType());
4840 CI.setOperand(0, MemCpy);
4844 if (Changed) return II;
4845 } else if (DbgStopPointInst *SPI = dyn_cast<DbgStopPointInst>(II)) {
4846 // If this stoppoint is at the same source location as the previous
4847 // stoppoint in the chain, it is not needed.
4848 if (DbgStopPointInst *PrevSPI =
4849 dyn_cast<DbgStopPointInst>(SPI->getChain()))
4850 if (SPI->getLineNo() == PrevSPI->getLineNo() &&
4851 SPI->getColNo() == PrevSPI->getColNo()) {
4852 SPI->replaceAllUsesWith(PrevSPI);
4853 return EraseInstFromFunction(CI);
4856 switch (II->getIntrinsicID()) {
4858 case Intrinsic::stackrestore: {
4859 // If the save is right next to the restore, remove the restore. This can
4860 // happen when variable allocas are DCE'd.
4861 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
4862 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
4863 BasicBlock::iterator BI = SS;
4865 return EraseInstFromFunction(CI);
4869 // If the stack restore is in a return/unwind block and if there are no
4870 // allocas or calls between the restore and the return, nuke the restore.
4871 TerminatorInst *TI = II->getParent()->getTerminator();
4872 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
4873 BasicBlock::iterator BI = II;
4874 bool CannotRemove = false;
4875 for (++BI; &*BI != TI; ++BI) {
4876 if (isa<AllocaInst>(BI) ||
4877 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
4878 CannotRemove = true;
4883 return EraseInstFromFunction(CI);
4890 return visitCallSite(II);
4893 // InvokeInst simplification
4895 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
4896 return visitCallSite(&II);
4899 // visitCallSite - Improvements for call and invoke instructions.
4901 Instruction *InstCombiner::visitCallSite(CallSite CS) {
4902 bool Changed = false;
4904 // If the callee is a constexpr cast of a function, attempt to move the cast
4905 // to the arguments of the call/invoke.
4906 if (transformConstExprCastCall(CS)) return 0;
4908 Value *Callee = CS.getCalledValue();
4910 if (Function *CalleeF = dyn_cast<Function>(Callee))
4911 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
4912 Instruction *OldCall = CS.getInstruction();
4913 // If the call and callee calling conventions don't match, this call must
4914 // be unreachable, as the call is undefined.
4915 new StoreInst(ConstantBool::True,
4916 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
4917 if (!OldCall->use_empty())
4918 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
4919 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
4920 return EraseInstFromFunction(*OldCall);
4924 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
4925 // This instruction is not reachable, just remove it. We insert a store to
4926 // undef so that we know that this code is not reachable, despite the fact
4927 // that we can't modify the CFG here.
4928 new StoreInst(ConstantBool::True,
4929 UndefValue::get(PointerType::get(Type::BoolTy)),
4930 CS.getInstruction());
4932 if (!CS.getInstruction()->use_empty())
4933 CS.getInstruction()->
4934 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
4936 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
4937 // Don't break the CFG, insert a dummy cond branch.
4938 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
4939 ConstantBool::True, II);
4941 return EraseInstFromFunction(*CS.getInstruction());
4944 const PointerType *PTy = cast<PointerType>(Callee->getType());
4945 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4946 if (FTy->isVarArg()) {
4947 // See if we can optimize any arguments passed through the varargs area of
4949 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
4950 E = CS.arg_end(); I != E; ++I)
4951 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
4952 // If this cast does not effect the value passed through the varargs
4953 // area, we can eliminate the use of the cast.
4954 Value *Op = CI->getOperand(0);
4955 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
4962 return Changed ? CS.getInstruction() : 0;
4965 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
4966 // attempt to move the cast to the arguments of the call/invoke.
4968 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
4969 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
4970 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
4971 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
4973 Function *Callee = cast<Function>(CE->getOperand(0));
4974 Instruction *Caller = CS.getInstruction();
4976 // Okay, this is a cast from a function to a different type. Unless doing so
4977 // would cause a type conversion of one of our arguments, change this call to
4978 // be a direct call with arguments casted to the appropriate types.
4980 const FunctionType *FT = Callee->getFunctionType();
4981 const Type *OldRetTy = Caller->getType();
4983 // Check to see if we are changing the return type...
4984 if (OldRetTy != FT->getReturnType()) {
4985 if (Callee->isExternal() &&
4986 !OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) &&
4987 !Caller->use_empty())
4988 return false; // Cannot transform this return value...
4990 // If the callsite is an invoke instruction, and the return value is used by
4991 // a PHI node in a successor, we cannot change the return type of the call
4992 // because there is no place to put the cast instruction (without breaking
4993 // the critical edge). Bail out in this case.
4994 if (!Caller->use_empty())
4995 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
4996 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
4998 if (PHINode *PN = dyn_cast<PHINode>(*UI))
4999 if (PN->getParent() == II->getNormalDest() ||
5000 PN->getParent() == II->getUnwindDest())
5004 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
5005 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
5007 CallSite::arg_iterator AI = CS.arg_begin();
5008 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
5009 const Type *ParamTy = FT->getParamType(i);
5010 bool isConvertible = (*AI)->getType()->isLosslesslyConvertibleTo(ParamTy);
5011 if (Callee->isExternal() && !isConvertible) return false;
5014 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
5015 Callee->isExternal())
5016 return false; // Do not delete arguments unless we have a function body...
5018 // Okay, we decided that this is a safe thing to do: go ahead and start
5019 // inserting cast instructions as necessary...
5020 std::vector<Value*> Args;
5021 Args.reserve(NumActualArgs);
5023 AI = CS.arg_begin();
5024 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
5025 const Type *ParamTy = FT->getParamType(i);
5026 if ((*AI)->getType() == ParamTy) {
5027 Args.push_back(*AI);
5029 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
5034 // If the function takes more arguments than the call was taking, add them
5036 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
5037 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
5039 // If we are removing arguments to the function, emit an obnoxious warning...
5040 if (FT->getNumParams() < NumActualArgs)
5041 if (!FT->isVarArg()) {
5042 std::cerr << "WARNING: While resolving call to function '"
5043 << Callee->getName() << "' arguments were dropped!\n";
5045 // Add all of the arguments in their promoted form to the arg list...
5046 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
5047 const Type *PTy = getPromotedType((*AI)->getType());
5048 if (PTy != (*AI)->getType()) {
5049 // Must promote to pass through va_arg area!
5050 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
5051 InsertNewInstBefore(Cast, *Caller);
5052 Args.push_back(Cast);
5054 Args.push_back(*AI);
5059 if (FT->getReturnType() == Type::VoidTy)
5060 Caller->setName(""); // Void type should not have a name...
5063 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5064 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
5065 Args, Caller->getName(), Caller);
5066 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
5068 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
5069 if (cast<CallInst>(Caller)->isTailCall())
5070 cast<CallInst>(NC)->setTailCall();
5071 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
5074 // Insert a cast of the return type as necessary...
5076 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
5077 if (NV->getType() != Type::VoidTy) {
5078 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
5080 // If this is an invoke instruction, we should insert it after the first
5081 // non-phi, instruction in the normal successor block.
5082 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5083 BasicBlock::iterator I = II->getNormalDest()->begin();
5084 while (isa<PHINode>(I)) ++I;
5085 InsertNewInstBefore(NC, *I);
5087 // Otherwise, it's a call, just insert cast right after the call instr
5088 InsertNewInstBefore(NC, *Caller);
5090 AddUsersToWorkList(*Caller);
5092 NV = UndefValue::get(Caller->getType());
5096 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
5097 Caller->replaceAllUsesWith(NV);
5098 Caller->getParent()->getInstList().erase(Caller);
5099 removeFromWorkList(Caller);
5104 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
5105 // operator and they all are only used by the PHI, PHI together their
5106 // inputs, and do the operation once, to the result of the PHI.
5107 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
5108 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
5110 // Scan the instruction, looking for input operations that can be folded away.
5111 // If all input operands to the phi are the same instruction (e.g. a cast from
5112 // the same type or "+42") we can pull the operation through the PHI, reducing
5113 // code size and simplifying code.
5114 Constant *ConstantOp = 0;
5115 const Type *CastSrcTy = 0;
5116 if (isa<CastInst>(FirstInst)) {
5117 CastSrcTy = FirstInst->getOperand(0)->getType();
5118 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
5119 // Can fold binop or shift if the RHS is a constant.
5120 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
5121 if (ConstantOp == 0) return 0;
5123 return 0; // Cannot fold this operation.
5126 // Check to see if all arguments are the same operation.
5127 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5128 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
5129 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
5130 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
5133 if (I->getOperand(0)->getType() != CastSrcTy)
5134 return 0; // Cast operation must match.
5135 } else if (I->getOperand(1) != ConstantOp) {
5140 // Okay, they are all the same operation. Create a new PHI node of the
5141 // correct type, and PHI together all of the LHS's of the instructions.
5142 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
5143 PN.getName()+".in");
5144 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
5146 Value *InVal = FirstInst->getOperand(0);
5147 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
5149 // Add all operands to the new PHI.
5150 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5151 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
5152 if (NewInVal != InVal)
5154 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
5159 // The new PHI unions all of the same values together. This is really
5160 // common, so we handle it intelligently here for compile-time speed.
5164 InsertNewInstBefore(NewPN, PN);
5168 // Insert and return the new operation.
5169 if (isa<CastInst>(FirstInst))
5170 return new CastInst(PhiVal, PN.getType());
5171 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
5172 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
5174 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
5175 PhiVal, ConstantOp);
5178 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
5180 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
5181 if (PN->use_empty()) return true;
5182 if (!PN->hasOneUse()) return false;
5184 // Remember this node, and if we find the cycle, return.
5185 if (!PotentiallyDeadPHIs.insert(PN).second)
5188 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
5189 return DeadPHICycle(PU, PotentiallyDeadPHIs);
5194 // PHINode simplification
5196 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
5197 if (Value *V = PN.hasConstantValue())
5198 return ReplaceInstUsesWith(PN, V);
5200 // If the only user of this instruction is a cast instruction, and all of the
5201 // incoming values are constants, change this PHI to merge together the casted
5204 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
5205 if (CI->getType() != PN.getType()) { // noop casts will be folded
5206 bool AllConstant = true;
5207 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
5208 if (!isa<Constant>(PN.getIncomingValue(i))) {
5209 AllConstant = false;
5213 // Make a new PHI with all casted values.
5214 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
5215 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
5216 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
5217 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
5218 PN.getIncomingBlock(i));
5221 // Update the cast instruction.
5222 CI->setOperand(0, New);
5223 WorkList.push_back(CI); // revisit the cast instruction to fold.
5224 WorkList.push_back(New); // Make sure to revisit the new Phi
5225 return &PN; // PN is now dead!
5229 // If all PHI operands are the same operation, pull them through the PHI,
5230 // reducing code size.
5231 if (isa<Instruction>(PN.getIncomingValue(0)) &&
5232 PN.getIncomingValue(0)->hasOneUse())
5233 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
5236 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
5237 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
5238 // PHI)... break the cycle.
5240 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
5241 std::set<PHINode*> PotentiallyDeadPHIs;
5242 PotentiallyDeadPHIs.insert(&PN);
5243 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
5244 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
5250 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
5251 Instruction *InsertPoint,
5253 unsigned PS = IC->getTargetData().getPointerSize();
5254 const Type *VTy = V->getType();
5255 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
5256 // We must insert a cast to ensure we sign-extend.
5257 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
5258 V->getName()), *InsertPoint);
5259 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
5264 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
5265 Value *PtrOp = GEP.getOperand(0);
5266 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
5267 // If so, eliminate the noop.
5268 if (GEP.getNumOperands() == 1)
5269 return ReplaceInstUsesWith(GEP, PtrOp);
5271 if (isa<UndefValue>(GEP.getOperand(0)))
5272 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
5274 bool HasZeroPointerIndex = false;
5275 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
5276 HasZeroPointerIndex = C->isNullValue();
5278 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
5279 return ReplaceInstUsesWith(GEP, PtrOp);
5281 // Eliminate unneeded casts for indices.
5282 bool MadeChange = false;
5283 gep_type_iterator GTI = gep_type_begin(GEP);
5284 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
5285 if (isa<SequentialType>(*GTI)) {
5286 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
5287 Value *Src = CI->getOperand(0);
5288 const Type *SrcTy = Src->getType();
5289 const Type *DestTy = CI->getType();
5290 if (Src->getType()->isInteger()) {
5291 if (SrcTy->getPrimitiveSizeInBits() ==
5292 DestTy->getPrimitiveSizeInBits()) {
5293 // We can always eliminate a cast from ulong or long to the other.
5294 // We can always eliminate a cast from uint to int or the other on
5295 // 32-bit pointer platforms.
5296 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
5298 GEP.setOperand(i, Src);
5300 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
5301 SrcTy->getPrimitiveSize() == 4) {
5302 // We can always eliminate a cast from int to [u]long. We can
5303 // eliminate a cast from uint to [u]long iff the target is a 32-bit
5305 if (SrcTy->isSigned() ||
5306 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
5308 GEP.setOperand(i, Src);
5313 // If we are using a wider index than needed for this platform, shrink it
5314 // to what we need. If the incoming value needs a cast instruction,
5315 // insert it. This explicit cast can make subsequent optimizations more
5317 Value *Op = GEP.getOperand(i);
5318 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
5319 if (Constant *C = dyn_cast<Constant>(Op)) {
5320 GEP.setOperand(i, ConstantExpr::getCast(C,
5321 TD->getIntPtrType()->getSignedVersion()));
5324 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
5325 Op->getName()), GEP);
5326 GEP.setOperand(i, Op);
5330 // If this is a constant idx, make sure to canonicalize it to be a signed
5331 // operand, otherwise CSE and other optimizations are pessimized.
5332 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
5333 GEP.setOperand(i, ConstantExpr::getCast(CUI,
5334 CUI->getType()->getSignedVersion()));
5338 if (MadeChange) return &GEP;
5340 // Combine Indices - If the source pointer to this getelementptr instruction
5341 // is a getelementptr instruction, combine the indices of the two
5342 // getelementptr instructions into a single instruction.
5344 std::vector<Value*> SrcGEPOperands;
5345 if (User *Src = dyn_castGetElementPtr(PtrOp))
5346 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
5348 if (!SrcGEPOperands.empty()) {
5349 // Note that if our source is a gep chain itself that we wait for that
5350 // chain to be resolved before we perform this transformation. This
5351 // avoids us creating a TON of code in some cases.
5353 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
5354 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
5355 return 0; // Wait until our source is folded to completion.
5357 std::vector<Value *> Indices;
5359 // Find out whether the last index in the source GEP is a sequential idx.
5360 bool EndsWithSequential = false;
5361 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
5362 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
5363 EndsWithSequential = !isa<StructType>(*I);
5365 // Can we combine the two pointer arithmetics offsets?
5366 if (EndsWithSequential) {
5367 // Replace: gep (gep %P, long B), long A, ...
5368 // With: T = long A+B; gep %P, T, ...
5370 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
5371 if (SO1 == Constant::getNullValue(SO1->getType())) {
5373 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
5376 // If they aren't the same type, convert both to an integer of the
5377 // target's pointer size.
5378 if (SO1->getType() != GO1->getType()) {
5379 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
5380 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
5381 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
5382 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
5384 unsigned PS = TD->getPointerSize();
5385 if (SO1->getType()->getPrimitiveSize() == PS) {
5386 // Convert GO1 to SO1's type.
5387 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
5389 } else if (GO1->getType()->getPrimitiveSize() == PS) {
5390 // Convert SO1 to GO1's type.
5391 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
5393 const Type *PT = TD->getIntPtrType();
5394 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
5395 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
5399 if (isa<Constant>(SO1) && isa<Constant>(GO1))
5400 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
5402 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
5403 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
5407 // Recycle the GEP we already have if possible.
5408 if (SrcGEPOperands.size() == 2) {
5409 GEP.setOperand(0, SrcGEPOperands[0]);
5410 GEP.setOperand(1, Sum);
5413 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5414 SrcGEPOperands.end()-1);
5415 Indices.push_back(Sum);
5416 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
5418 } else if (isa<Constant>(*GEP.idx_begin()) &&
5419 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
5420 SrcGEPOperands.size() != 1) {
5421 // Otherwise we can do the fold if the first index of the GEP is a zero
5422 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5423 SrcGEPOperands.end());
5424 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
5427 if (!Indices.empty())
5428 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
5430 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
5431 // GEP of global variable. If all of the indices for this GEP are
5432 // constants, we can promote this to a constexpr instead of an instruction.
5434 // Scan for nonconstants...
5435 std::vector<Constant*> Indices;
5436 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
5437 for (; I != E && isa<Constant>(*I); ++I)
5438 Indices.push_back(cast<Constant>(*I));
5440 if (I == E) { // If they are all constants...
5441 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
5443 // Replace all uses of the GEP with the new constexpr...
5444 return ReplaceInstUsesWith(GEP, CE);
5446 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
5447 if (!isa<PointerType>(X->getType())) {
5448 // Not interesting. Source pointer must be a cast from pointer.
5449 } else if (HasZeroPointerIndex) {
5450 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
5451 // into : GEP [10 x ubyte]* X, long 0, ...
5453 // This occurs when the program declares an array extern like "int X[];"
5455 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
5456 const PointerType *XTy = cast<PointerType>(X->getType());
5457 if (const ArrayType *XATy =
5458 dyn_cast<ArrayType>(XTy->getElementType()))
5459 if (const ArrayType *CATy =
5460 dyn_cast<ArrayType>(CPTy->getElementType()))
5461 if (CATy->getElementType() == XATy->getElementType()) {
5462 // At this point, we know that the cast source type is a pointer
5463 // to an array of the same type as the destination pointer
5464 // array. Because the array type is never stepped over (there
5465 // is a leading zero) we can fold the cast into this GEP.
5466 GEP.setOperand(0, X);
5469 } else if (GEP.getNumOperands() == 2) {
5470 // Transform things like:
5471 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
5472 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
5473 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
5474 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
5475 if (isa<ArrayType>(SrcElTy) &&
5476 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
5477 TD->getTypeSize(ResElTy)) {
5478 Value *V = InsertNewInstBefore(
5479 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5480 GEP.getOperand(1), GEP.getName()), GEP);
5481 return new CastInst(V, GEP.getType());
5484 // Transform things like:
5485 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
5486 // (where tmp = 8*tmp2) into:
5487 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
5489 if (isa<ArrayType>(SrcElTy) &&
5490 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
5491 uint64_t ArrayEltSize =
5492 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
5494 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
5495 // allow either a mul, shift, or constant here.
5497 ConstantInt *Scale = 0;
5498 if (ArrayEltSize == 1) {
5499 NewIdx = GEP.getOperand(1);
5500 Scale = ConstantInt::get(NewIdx->getType(), 1);
5501 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
5502 NewIdx = ConstantInt::get(CI->getType(), 1);
5504 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
5505 if (Inst->getOpcode() == Instruction::Shl &&
5506 isa<ConstantInt>(Inst->getOperand(1))) {
5507 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
5508 if (Inst->getType()->isSigned())
5509 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
5511 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
5512 NewIdx = Inst->getOperand(0);
5513 } else if (Inst->getOpcode() == Instruction::Mul &&
5514 isa<ConstantInt>(Inst->getOperand(1))) {
5515 Scale = cast<ConstantInt>(Inst->getOperand(1));
5516 NewIdx = Inst->getOperand(0);
5520 // If the index will be to exactly the right offset with the scale taken
5521 // out, perform the transformation.
5522 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
5523 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
5524 Scale = ConstantSInt::get(C->getType(),
5525 (int64_t)C->getRawValue() /
5526 (int64_t)ArrayEltSize);
5528 Scale = ConstantUInt::get(Scale->getType(),
5529 Scale->getRawValue() / ArrayEltSize);
5530 if (Scale->getRawValue() != 1) {
5531 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
5532 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
5533 NewIdx = InsertNewInstBefore(Sc, GEP);
5536 // Insert the new GEP instruction.
5538 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5539 NewIdx, GEP.getName());
5540 Idx = InsertNewInstBefore(Idx, GEP);
5541 return new CastInst(Idx, GEP.getType());
5550 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
5551 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
5552 if (AI.isArrayAllocation()) // Check C != 1
5553 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
5554 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
5555 AllocationInst *New = 0;
5557 // Create and insert the replacement instruction...
5558 if (isa<MallocInst>(AI))
5559 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
5561 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
5562 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
5565 InsertNewInstBefore(New, AI);
5567 // Scan to the end of the allocation instructions, to skip over a block of
5568 // allocas if possible...
5570 BasicBlock::iterator It = New;
5571 while (isa<AllocationInst>(*It)) ++It;
5573 // Now that I is pointing to the first non-allocation-inst in the block,
5574 // insert our getelementptr instruction...
5576 Value *NullIdx = Constant::getNullValue(Type::IntTy);
5577 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
5578 New->getName()+".sub", It);
5580 // Now make everything use the getelementptr instead of the original
5582 return ReplaceInstUsesWith(AI, V);
5583 } else if (isa<UndefValue>(AI.getArraySize())) {
5584 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5587 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
5588 // Note that we only do this for alloca's, because malloc should allocate and
5589 // return a unique pointer, even for a zero byte allocation.
5590 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
5591 TD->getTypeSize(AI.getAllocatedType()) == 0)
5592 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5597 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
5598 Value *Op = FI.getOperand(0);
5600 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
5601 if (CastInst *CI = dyn_cast<CastInst>(Op))
5602 if (isa<PointerType>(CI->getOperand(0)->getType())) {
5603 FI.setOperand(0, CI->getOperand(0));
5607 // free undef -> unreachable.
5608 if (isa<UndefValue>(Op)) {
5609 // Insert a new store to null because we cannot modify the CFG here.
5610 new StoreInst(ConstantBool::True,
5611 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
5612 return EraseInstFromFunction(FI);
5615 // If we have 'free null' delete the instruction. This can happen in stl code
5616 // when lots of inlining happens.
5617 if (isa<ConstantPointerNull>(Op))
5618 return EraseInstFromFunction(FI);
5624 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
5625 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
5626 User *CI = cast<User>(LI.getOperand(0));
5627 Value *CastOp = CI->getOperand(0);
5629 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5630 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5631 const Type *SrcPTy = SrcTy->getElementType();
5633 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5634 // If the source is an array, the code below will not succeed. Check to
5635 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5637 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5638 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5639 if (ASrcTy->getNumElements() != 0) {
5640 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5641 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5642 SrcTy = cast<PointerType>(CastOp->getType());
5643 SrcPTy = SrcTy->getElementType();
5646 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5647 // Do not allow turning this into a load of an integer, which is then
5648 // casted to a pointer, this pessimizes pointer analysis a lot.
5649 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
5650 IC.getTargetData().getTypeSize(SrcPTy) ==
5651 IC.getTargetData().getTypeSize(DestPTy)) {
5653 // Okay, we are casting from one integer or pointer type to another of
5654 // the same size. Instead of casting the pointer before the load, cast
5655 // the result of the loaded value.
5656 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
5658 LI.isVolatile()),LI);
5659 // Now cast the result of the load.
5660 return new CastInst(NewLoad, LI.getType());
5667 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
5668 /// from this value cannot trap. If it is not obviously safe to load from the
5669 /// specified pointer, we do a quick local scan of the basic block containing
5670 /// ScanFrom, to determine if the address is already accessed.
5671 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
5672 // If it is an alloca or global variable, it is always safe to load from.
5673 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
5675 // Otherwise, be a little bit agressive by scanning the local block where we
5676 // want to check to see if the pointer is already being loaded or stored
5677 // from/to. If so, the previous load or store would have already trapped,
5678 // so there is no harm doing an extra load (also, CSE will later eliminate
5679 // the load entirely).
5680 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
5685 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
5686 if (LI->getOperand(0) == V) return true;
5687 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5688 if (SI->getOperand(1) == V) return true;
5694 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
5695 Value *Op = LI.getOperand(0);
5697 // load (cast X) --> cast (load X) iff safe
5698 if (CastInst *CI = dyn_cast<CastInst>(Op))
5699 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5702 // None of the following transforms are legal for volatile loads.
5703 if (LI.isVolatile()) return 0;
5705 if (&LI.getParent()->front() != &LI) {
5706 BasicBlock::iterator BBI = &LI; --BBI;
5707 // If the instruction immediately before this is a store to the same
5708 // address, do a simple form of store->load forwarding.
5709 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5710 if (SI->getOperand(1) == LI.getOperand(0))
5711 return ReplaceInstUsesWith(LI, SI->getOperand(0));
5712 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
5713 if (LIB->getOperand(0) == LI.getOperand(0))
5714 return ReplaceInstUsesWith(LI, LIB);
5717 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
5718 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
5719 isa<UndefValue>(GEPI->getOperand(0))) {
5720 // Insert a new store to null instruction before the load to indicate
5721 // that this code is not reachable. We do this instead of inserting
5722 // an unreachable instruction directly because we cannot modify the
5724 new StoreInst(UndefValue::get(LI.getType()),
5725 Constant::getNullValue(Op->getType()), &LI);
5726 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5729 if (Constant *C = dyn_cast<Constant>(Op)) {
5730 // load null/undef -> undef
5731 if ((C->isNullValue() || isa<UndefValue>(C))) {
5732 // Insert a new store to null instruction before the load to indicate that
5733 // this code is not reachable. We do this instead of inserting an
5734 // unreachable instruction directly because we cannot modify the CFG.
5735 new StoreInst(UndefValue::get(LI.getType()),
5736 Constant::getNullValue(Op->getType()), &LI);
5737 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5740 // Instcombine load (constant global) into the value loaded.
5741 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
5742 if (GV->isConstant() && !GV->isExternal())
5743 return ReplaceInstUsesWith(LI, GV->getInitializer());
5745 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
5746 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
5747 if (CE->getOpcode() == Instruction::GetElementPtr) {
5748 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
5749 if (GV->isConstant() && !GV->isExternal())
5751 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
5752 return ReplaceInstUsesWith(LI, V);
5753 if (CE->getOperand(0)->isNullValue()) {
5754 // Insert a new store to null instruction before the load to indicate
5755 // that this code is not reachable. We do this instead of inserting
5756 // an unreachable instruction directly because we cannot modify the
5758 new StoreInst(UndefValue::get(LI.getType()),
5759 Constant::getNullValue(Op->getType()), &LI);
5760 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5763 } else if (CE->getOpcode() == Instruction::Cast) {
5764 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5769 if (Op->hasOneUse()) {
5770 // Change select and PHI nodes to select values instead of addresses: this
5771 // helps alias analysis out a lot, allows many others simplifications, and
5772 // exposes redundancy in the code.
5774 // Note that we cannot do the transformation unless we know that the
5775 // introduced loads cannot trap! Something like this is valid as long as
5776 // the condition is always false: load (select bool %C, int* null, int* %G),
5777 // but it would not be valid if we transformed it to load from null
5780 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
5781 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
5782 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
5783 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
5784 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
5785 SI->getOperand(1)->getName()+".val"), LI);
5786 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
5787 SI->getOperand(2)->getName()+".val"), LI);
5788 return new SelectInst(SI->getCondition(), V1, V2);
5791 // load (select (cond, null, P)) -> load P
5792 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
5793 if (C->isNullValue()) {
5794 LI.setOperand(0, SI->getOperand(2));
5798 // load (select (cond, P, null)) -> load P
5799 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
5800 if (C->isNullValue()) {
5801 LI.setOperand(0, SI->getOperand(1));
5805 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
5806 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
5807 bool Safe = PN->getParent() == LI.getParent();
5809 // Scan all of the instructions between the PHI and the load to make
5810 // sure there are no instructions that might possibly alter the value
5811 // loaded from the PHI.
5813 BasicBlock::iterator I = &LI;
5814 for (--I; !isa<PHINode>(I); --I)
5815 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
5821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
5822 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
5823 PN->getIncomingBlock(i)->getTerminator()))
5828 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
5829 InsertNewInstBefore(NewPN, *PN);
5830 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
5832 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
5833 BasicBlock *BB = PN->getIncomingBlock(i);
5834 Value *&TheLoad = LoadMap[BB];
5836 Value *InVal = PN->getIncomingValue(i);
5837 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
5838 InVal->getName()+".val"),
5839 *BB->getTerminator());
5841 NewPN->addIncoming(TheLoad, BB);
5843 return ReplaceInstUsesWith(LI, NewPN);
5850 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
5852 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
5853 User *CI = cast<User>(SI.getOperand(1));
5854 Value *CastOp = CI->getOperand(0);
5856 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5857 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5858 const Type *SrcPTy = SrcTy->getElementType();
5860 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5861 // If the source is an array, the code below will not succeed. Check to
5862 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5864 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5865 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5866 if (ASrcTy->getNumElements() != 0) {
5867 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5868 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5869 SrcTy = cast<PointerType>(CastOp->getType());
5870 SrcPTy = SrcTy->getElementType();
5873 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5874 IC.getTargetData().getTypeSize(SrcPTy) ==
5875 IC.getTargetData().getTypeSize(DestPTy)) {
5877 // Okay, we are casting from one integer or pointer type to another of
5878 // the same size. Instead of casting the pointer before the store, cast
5879 // the value to be stored.
5881 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
5882 NewCast = ConstantExpr::getCast(C, SrcPTy);
5884 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
5886 SI.getOperand(0)->getName()+".c"), SI);
5888 return new StoreInst(NewCast, CastOp);
5895 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
5896 Value *Val = SI.getOperand(0);
5897 Value *Ptr = SI.getOperand(1);
5899 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
5900 removeFromWorkList(&SI);
5901 SI.eraseFromParent();
5906 if (SI.isVolatile()) return 0; // Don't hack volatile loads.
5908 // store X, null -> turns into 'unreachable' in SimplifyCFG
5909 if (isa<ConstantPointerNull>(Ptr)) {
5910 if (!isa<UndefValue>(Val)) {
5911 SI.setOperand(0, UndefValue::get(Val->getType()));
5912 if (Instruction *U = dyn_cast<Instruction>(Val))
5913 WorkList.push_back(U); // Dropped a use.
5916 return 0; // Do not modify these!
5919 // store undef, Ptr -> noop
5920 if (isa<UndefValue>(Val)) {
5921 removeFromWorkList(&SI);
5922 SI.eraseFromParent();
5927 // If the pointer destination is a cast, see if we can fold the cast into the
5929 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
5930 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5932 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
5933 if (CE->getOpcode() == Instruction::Cast)
5934 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5938 // If this store is the last instruction in the basic block, and if the block
5939 // ends with an unconditional branch, try to move it to the successor block.
5940 BasicBlock::iterator BBI = &SI; ++BBI;
5941 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
5942 if (BI->isUnconditional()) {
5943 // Check to see if the successor block has exactly two incoming edges. If
5944 // so, see if the other predecessor contains a store to the same location.
5945 // if so, insert a PHI node (if needed) and move the stores down.
5946 BasicBlock *Dest = BI->getSuccessor(0);
5948 pred_iterator PI = pred_begin(Dest);
5949 BasicBlock *Other = 0;
5950 if (*PI != BI->getParent())
5953 if (PI != pred_end(Dest)) {
5954 if (*PI != BI->getParent())
5959 if (++PI != pred_end(Dest))
5962 if (Other) { // If only one other pred...
5963 BBI = Other->getTerminator();
5964 // Make sure this other block ends in an unconditional branch and that
5965 // there is an instruction before the branch.
5966 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
5967 BBI != Other->begin()) {
5969 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
5971 // If this instruction is a store to the same location.
5972 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
5973 // Okay, we know we can perform this transformation. Insert a PHI
5974 // node now if we need it.
5975 Value *MergedVal = OtherStore->getOperand(0);
5976 if (MergedVal != SI.getOperand(0)) {
5977 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
5978 PN->reserveOperandSpace(2);
5979 PN->addIncoming(SI.getOperand(0), SI.getParent());
5980 PN->addIncoming(OtherStore->getOperand(0), Other);
5981 MergedVal = InsertNewInstBefore(PN, Dest->front());
5984 // Advance to a place where it is safe to insert the new store and
5986 BBI = Dest->begin();
5987 while (isa<PHINode>(BBI)) ++BBI;
5988 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
5989 OtherStore->isVolatile()), *BBI);
5991 // Nuke the old stores.
5992 removeFromWorkList(&SI);
5993 removeFromWorkList(OtherStore);
5994 SI.eraseFromParent();
5995 OtherStore->eraseFromParent();
6007 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
6008 // Change br (not X), label True, label False to: br X, label False, True
6010 BasicBlock *TrueDest;
6011 BasicBlock *FalseDest;
6012 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
6013 !isa<Constant>(X)) {
6014 // Swap Destinations and condition...
6016 BI.setSuccessor(0, FalseDest);
6017 BI.setSuccessor(1, TrueDest);
6021 // Cannonicalize setne -> seteq
6022 Instruction::BinaryOps Op; Value *Y;
6023 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
6024 TrueDest, FalseDest)))
6025 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
6026 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
6027 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
6028 std::string Name = I->getName(); I->setName("");
6029 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
6030 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
6031 // Swap Destinations and condition...
6032 BI.setCondition(NewSCC);
6033 BI.setSuccessor(0, FalseDest);
6034 BI.setSuccessor(1, TrueDest);
6035 removeFromWorkList(I);
6036 I->getParent()->getInstList().erase(I);
6037 WorkList.push_back(cast<Instruction>(NewSCC));
6044 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
6045 Value *Cond = SI.getCondition();
6046 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
6047 if (I->getOpcode() == Instruction::Add)
6048 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6049 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
6050 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
6051 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
6053 SI.setOperand(0, I->getOperand(0));
6054 WorkList.push_back(I);
6061 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
6062 if (ConstantAggregateZero *C =
6063 dyn_cast<ConstantAggregateZero>(EI.getOperand(0))) {
6064 // If packed val is constant 0, replace extract with scalar 0
6065 const Type *Ty = cast<PackedType>(C->getType())->getElementType();
6066 EI.replaceAllUsesWith(Constant::getNullValue(Ty));
6067 return ReplaceInstUsesWith(EI, Constant::getNullValue(Ty));
6069 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
6070 // If packed val is constant with uniform operands, replace EI
6071 // with that operand
6072 Constant *op0 = cast<Constant>(C->getOperand(0));
6073 for (unsigned i = 1; i < C->getNumOperands(); ++i)
6074 if (C->getOperand(i) != op0) return 0;
6075 return ReplaceInstUsesWith(EI, op0);
6077 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0)))
6078 if (I->hasOneUse()) {
6079 // Push extractelement into predecessor operation if legal and
6080 // profitable to do so
6081 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
6082 if (!isa<Constant>(BO->getOperand(0)) &&
6083 !isa<Constant>(BO->getOperand(1)))
6085 ExtractElementInst *newEI0 =
6086 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
6088 ExtractElementInst *newEI1 =
6089 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
6091 InsertNewInstBefore(newEI0, EI);
6092 InsertNewInstBefore(newEI1, EI);
6093 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
6095 switch(I->getOpcode()) {
6096 case Instruction::Load: {
6097 Value *Ptr = InsertCastBefore(I->getOperand(0),
6098 PointerType::get(EI.getType()), EI);
6099 GetElementPtrInst *GEP =
6100 new GetElementPtrInst(Ptr, EI.getOperand(1),
6101 I->getName() + ".gep");
6102 InsertNewInstBefore(GEP, EI);
6103 return new LoadInst(GEP);
6113 void InstCombiner::removeFromWorkList(Instruction *I) {
6114 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
6119 /// TryToSinkInstruction - Try to move the specified instruction from its
6120 /// current block into the beginning of DestBlock, which can only happen if it's
6121 /// safe to move the instruction past all of the instructions between it and the
6122 /// end of its block.
6123 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
6124 assert(I->hasOneUse() && "Invariants didn't hold!");
6126 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
6127 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
6129 // Do not sink alloca instructions out of the entry block.
6130 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
6133 // We can only sink load instructions if there is nothing between the load and
6134 // the end of block that could change the value.
6135 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6136 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
6138 if (Scan->mayWriteToMemory())
6142 BasicBlock::iterator InsertPos = DestBlock->begin();
6143 while (isa<PHINode>(InsertPos)) ++InsertPos;
6145 I->moveBefore(InsertPos);
6150 bool InstCombiner::runOnFunction(Function &F) {
6151 bool Changed = false;
6152 TD = &getAnalysis<TargetData>();
6155 // Populate the worklist with the reachable instructions.
6156 std::set<BasicBlock*> Visited;
6157 for (df_ext_iterator<BasicBlock*> BB = df_ext_begin(&F.front(), Visited),
6158 E = df_ext_end(&F.front(), Visited); BB != E; ++BB)
6159 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
6160 WorkList.push_back(I);
6162 // Do a quick scan over the function. If we find any blocks that are
6163 // unreachable, remove any instructions inside of them. This prevents
6164 // the instcombine code from having to deal with some bad special cases.
6165 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
6166 if (!Visited.count(BB)) {
6167 Instruction *Term = BB->getTerminator();
6168 while (Term != BB->begin()) { // Remove instrs bottom-up
6169 BasicBlock::iterator I = Term; --I;
6171 DEBUG(std::cerr << "IC: DCE: " << *I);
6174 if (!I->use_empty())
6175 I->replaceAllUsesWith(UndefValue::get(I->getType()));
6176 I->eraseFromParent();
6181 while (!WorkList.empty()) {
6182 Instruction *I = WorkList.back(); // Get an instruction from the worklist
6183 WorkList.pop_back();
6185 // Check to see if we can DCE or ConstantPropagate the instruction...
6186 // Check to see if we can DIE the instruction...
6187 if (isInstructionTriviallyDead(I)) {
6188 // Add operands to the worklist...
6189 if (I->getNumOperands() < 4)
6190 AddUsesToWorkList(*I);
6193 DEBUG(std::cerr << "IC: DCE: " << *I);
6195 I->eraseFromParent();
6196 removeFromWorkList(I);
6200 // Instruction isn't dead, see if we can constant propagate it...
6201 if (Constant *C = ConstantFoldInstruction(I)) {
6202 Value* Ptr = I->getOperand(0);
6203 if (isa<GetElementPtrInst>(I) &&
6204 cast<Constant>(Ptr)->isNullValue() &&
6205 !isa<ConstantPointerNull>(C) &&
6206 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
6207 // If this is a constant expr gep that is effectively computing an
6208 // "offsetof", fold it into 'cast int X to T*' instead of 'gep 0, 0, 12'
6209 bool isFoldableGEP = true;
6210 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
6211 if (!isa<ConstantInt>(I->getOperand(i)))
6212 isFoldableGEP = false;
6213 if (isFoldableGEP) {
6214 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
6215 std::vector<Value*>(I->op_begin()+1, I->op_end()));
6216 C = ConstantUInt::get(Type::ULongTy, Offset);
6217 C = ConstantExpr::getCast(C, TD->getIntPtrType());
6218 C = ConstantExpr::getCast(C, I->getType());
6222 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
6224 // Add operands to the worklist...
6225 AddUsesToWorkList(*I);
6226 ReplaceInstUsesWith(*I, C);
6229 I->getParent()->getInstList().erase(I);
6230 removeFromWorkList(I);
6234 // See if we can trivially sink this instruction to a successor basic block.
6235 if (I->hasOneUse()) {
6236 BasicBlock *BB = I->getParent();
6237 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
6238 if (UserParent != BB) {
6239 bool UserIsSuccessor = false;
6240 // See if the user is one of our successors.
6241 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
6242 if (*SI == UserParent) {
6243 UserIsSuccessor = true;
6247 // If the user is one of our immediate successors, and if that successor
6248 // only has us as a predecessors (we'd have to split the critical edge
6249 // otherwise), we can keep going.
6250 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
6251 next(pred_begin(UserParent)) == pred_end(UserParent))
6252 // Okay, the CFG is simple enough, try to sink this instruction.
6253 Changed |= TryToSinkInstruction(I, UserParent);
6257 // Now that we have an instruction, try combining it to simplify it...
6258 if (Instruction *Result = visit(*I)) {
6260 // Should we replace the old instruction with a new one?
6262 DEBUG(std::cerr << "IC: Old = " << *I
6263 << " New = " << *Result);
6265 // Everything uses the new instruction now.
6266 I->replaceAllUsesWith(Result);
6268 // Push the new instruction and any users onto the worklist.
6269 WorkList.push_back(Result);
6270 AddUsersToWorkList(*Result);
6272 // Move the name to the new instruction first...
6273 std::string OldName = I->getName(); I->setName("");
6274 Result->setName(OldName);
6276 // Insert the new instruction into the basic block...
6277 BasicBlock *InstParent = I->getParent();
6278 BasicBlock::iterator InsertPos = I;
6280 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
6281 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
6284 InstParent->getInstList().insert(InsertPos, Result);
6286 // Make sure that we reprocess all operands now that we reduced their
6288 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6289 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6290 WorkList.push_back(OpI);
6292 // Instructions can end up on the worklist more than once. Make sure
6293 // we do not process an instruction that has been deleted.
6294 removeFromWorkList(I);
6296 // Erase the old instruction.
6297 InstParent->getInstList().erase(I);
6299 DEBUG(std::cerr << "IC: MOD = " << *I);
6301 // If the instruction was modified, it's possible that it is now dead.
6302 // if so, remove it.
6303 if (isInstructionTriviallyDead(I)) {
6304 // Make sure we process all operands now that we are reducing their
6306 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6307 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6308 WorkList.push_back(OpI);
6310 // Instructions may end up in the worklist more than once. Erase all
6311 // occurrences of this instruction.
6312 removeFromWorkList(I);
6313 I->eraseFromParent();
6315 WorkList.push_back(Result);
6316 AddUsersToWorkList(*Result);
6326 FunctionPass *llvm::createInstructionCombiningPass() {
6327 return new InstCombiner();