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"
56 using namespace llvm::PatternMatch;
59 Statistic<> NumCombined ("instcombine", "Number of insts combined");
60 Statistic<> NumConstProp("instcombine", "Number of constant folds");
61 Statistic<> NumDeadInst ("instcombine", "Number of dead inst eliminated");
62 Statistic<> NumSunkInst ("instcombine", "Number of instructions sunk");
64 class InstCombiner : public FunctionPass,
65 public InstVisitor<InstCombiner, Instruction*> {
66 // Worklist of all of the instructions that need to be simplified.
67 std::vector<Instruction*> WorkList;
70 /// AddUsersToWorkList - When an instruction is simplified, add all users of
71 /// the instruction to the work lists because they might get more simplified
74 void AddUsersToWorkList(Instruction &I) {
75 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
77 WorkList.push_back(cast<Instruction>(*UI));
80 /// AddUsesToWorkList - When an instruction is simplified, add operands to
81 /// the work lists because they might get more simplified now.
83 void AddUsesToWorkList(Instruction &I) {
84 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
85 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
86 WorkList.push_back(Op);
89 // removeFromWorkList - remove all instances of I from the worklist.
90 void removeFromWorkList(Instruction *I);
92 virtual bool runOnFunction(Function &F);
94 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
95 AU.addRequired<TargetData>();
99 TargetData &getTargetData() const { return *TD; }
101 // Visitation implementation - Implement instruction combining for different
102 // instruction types. The semantics are as follows:
104 // null - No change was made
105 // I - Change was made, I is still valid, I may be dead though
106 // otherwise - Change was made, replace I with returned instruction
108 Instruction *visitAdd(BinaryOperator &I);
109 Instruction *visitSub(BinaryOperator &I);
110 Instruction *visitMul(BinaryOperator &I);
111 Instruction *visitDiv(BinaryOperator &I);
112 Instruction *visitRem(BinaryOperator &I);
113 Instruction *visitAnd(BinaryOperator &I);
114 Instruction *visitOr (BinaryOperator &I);
115 Instruction *visitXor(BinaryOperator &I);
116 Instruction *visitSetCondInst(SetCondInst &I);
117 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
119 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
120 Instruction::BinaryOps Cond, Instruction &I);
121 Instruction *visitShiftInst(ShiftInst &I);
122 Instruction *visitCastInst(CastInst &CI);
123 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
125 Instruction *visitSelectInst(SelectInst &CI);
126 Instruction *visitCallInst(CallInst &CI);
127 Instruction *visitInvokeInst(InvokeInst &II);
128 Instruction *visitPHINode(PHINode &PN);
129 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
130 Instruction *visitAllocationInst(AllocationInst &AI);
131 Instruction *visitFreeInst(FreeInst &FI);
132 Instruction *visitLoadInst(LoadInst &LI);
133 Instruction *visitStoreInst(StoreInst &SI);
134 Instruction *visitBranchInst(BranchInst &BI);
135 Instruction *visitSwitchInst(SwitchInst &SI);
137 // visitInstruction - Specify what to return for unhandled instructions...
138 Instruction *visitInstruction(Instruction &I) { return 0; }
141 Instruction *visitCallSite(CallSite CS);
142 bool transformConstExprCastCall(CallSite CS);
145 // InsertNewInstBefore - insert an instruction New before instruction Old
146 // in the program. Add the new instruction to the worklist.
148 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
149 assert(New && New->getParent() == 0 &&
150 "New instruction already inserted into a basic block!");
151 BasicBlock *BB = Old.getParent();
152 BB->getInstList().insert(&Old, New); // Insert inst
153 WorkList.push_back(New); // Add to worklist
157 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
158 /// This also adds the cast to the worklist. Finally, this returns the
160 Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
161 if (V->getType() == Ty) return V;
163 Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
164 WorkList.push_back(C);
168 // ReplaceInstUsesWith - This method is to be used when an instruction is
169 // found to be dead, replacable with another preexisting expression. Here
170 // we add all uses of I to the worklist, replace all uses of I with the new
171 // value, then return I, so that the inst combiner will know that I was
174 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
175 AddUsersToWorkList(I); // Add all modified instrs to worklist
177 I.replaceAllUsesWith(V);
180 // If we are replacing the instruction with itself, this must be in a
181 // segment of unreachable code, so just clobber the instruction.
182 I.replaceAllUsesWith(UndefValue::get(I.getType()));
187 // EraseInstFromFunction - When dealing with an instruction that has side
188 // effects or produces a void value, we can't rely on DCE to delete the
189 // instruction. Instead, visit methods should return the value returned by
191 Instruction *EraseInstFromFunction(Instruction &I) {
192 assert(I.use_empty() && "Cannot erase instruction that is used!");
193 AddUsesToWorkList(I);
194 removeFromWorkList(&I);
196 return 0; // Don't do anything with FI
201 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
202 /// InsertBefore instruction. This is specialized a bit to avoid inserting
203 /// casts that are known to not do anything...
205 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
206 Instruction *InsertBefore);
208 // SimplifyCommutative - This performs a few simplifications for commutative
210 bool SimplifyCommutative(BinaryOperator &I);
213 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
214 // PHI node as operand #0, see if we can fold the instruction into the PHI
215 // (which is only possible if all operands to the PHI are constants).
216 Instruction *FoldOpIntoPhi(Instruction &I);
218 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
219 // operator and they all are only used by the PHI, PHI together their
220 // inputs, and do the operation once, to the result of the PHI.
221 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
223 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
224 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
226 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
227 bool isSub, Instruction &I);
228 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
229 bool Inside, Instruction &IB);
232 RegisterOpt<InstCombiner> X("instcombine", "Combine redundant instructions");
235 // getComplexity: Assign a complexity or rank value to LLVM Values...
236 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
237 static unsigned getComplexity(Value *V) {
238 if (isa<Instruction>(V)) {
239 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
243 if (isa<Argument>(V)) return 3;
244 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
247 // isOnlyUse - Return true if this instruction will be deleted if we stop using
249 static bool isOnlyUse(Value *V) {
250 return V->hasOneUse() || isa<Constant>(V);
253 // getPromotedType - Return the specified type promoted as it would be to pass
254 // though a va_arg area...
255 static const Type *getPromotedType(const Type *Ty) {
256 switch (Ty->getTypeID()) {
257 case Type::SByteTyID:
258 case Type::ShortTyID: return Type::IntTy;
259 case Type::UByteTyID:
260 case Type::UShortTyID: return Type::UIntTy;
261 case Type::FloatTyID: return Type::DoubleTy;
266 /// isCast - If the specified operand is a CastInst or a constant expr cast,
267 /// return the operand value, otherwise return null.
268 static Value *isCast(Value *V) {
269 if (CastInst *I = dyn_cast<CastInst>(V))
270 return I->getOperand(0);
271 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
272 if (CE->getOpcode() == Instruction::Cast)
273 return CE->getOperand(0);
277 // SimplifyCommutative - This performs a few simplifications for commutative
280 // 1. Order operands such that they are listed from right (least complex) to
281 // left (most complex). This puts constants before unary operators before
284 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
285 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
287 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
288 bool Changed = false;
289 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
290 Changed = !I.swapOperands();
292 if (!I.isAssociative()) return Changed;
293 Instruction::BinaryOps Opcode = I.getOpcode();
294 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
295 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
296 if (isa<Constant>(I.getOperand(1))) {
297 Constant *Folded = ConstantExpr::get(I.getOpcode(),
298 cast<Constant>(I.getOperand(1)),
299 cast<Constant>(Op->getOperand(1)));
300 I.setOperand(0, Op->getOperand(0));
301 I.setOperand(1, Folded);
303 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
304 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
305 isOnlyUse(Op) && isOnlyUse(Op1)) {
306 Constant *C1 = cast<Constant>(Op->getOperand(1));
307 Constant *C2 = cast<Constant>(Op1->getOperand(1));
309 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
310 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
311 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
314 WorkList.push_back(New);
315 I.setOperand(0, New);
316 I.setOperand(1, Folded);
323 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
324 // if the LHS is a constant zero (which is the 'negate' form).
326 static inline Value *dyn_castNegVal(Value *V) {
327 if (BinaryOperator::isNeg(V))
328 return BinaryOperator::getNegArgument(V);
330 // Constants can be considered to be negated values if they can be folded.
331 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
332 return ConstantExpr::getNeg(C);
336 static inline Value *dyn_castNotVal(Value *V) {
337 if (BinaryOperator::isNot(V))
338 return BinaryOperator::getNotArgument(V);
340 // Constants can be considered to be not'ed values...
341 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
342 return ConstantExpr::getNot(C);
346 // dyn_castFoldableMul - If this value is a multiply that can be folded into
347 // other computations (because it has a constant operand), return the
348 // non-constant operand of the multiply, and set CST to point to the multiplier.
349 // Otherwise, return null.
351 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
352 if (V->hasOneUse() && V->getType()->isInteger())
353 if (Instruction *I = dyn_cast<Instruction>(V)) {
354 if (I->getOpcode() == Instruction::Mul)
355 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
356 return I->getOperand(0);
357 if (I->getOpcode() == Instruction::Shl)
358 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
359 // The multiplier is really 1 << CST.
360 Constant *One = ConstantInt::get(V->getType(), 1);
361 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
362 return I->getOperand(0);
368 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
369 /// expression, return it.
370 static User *dyn_castGetElementPtr(Value *V) {
371 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
372 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
373 if (CE->getOpcode() == Instruction::GetElementPtr)
374 return cast<User>(V);
378 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
379 static ConstantInt *AddOne(ConstantInt *C) {
380 return cast<ConstantInt>(ConstantExpr::getAdd(C,
381 ConstantInt::get(C->getType(), 1)));
383 static ConstantInt *SubOne(ConstantInt *C) {
384 return cast<ConstantInt>(ConstantExpr::getSub(C,
385 ConstantInt::get(C->getType(), 1)));
388 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
389 /// this predicate to simplify operations downstream. V and Mask are known to
390 /// be the same type.
391 static bool MaskedValueIsZero(Value *V, ConstantIntegral *Mask) {
392 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
393 // we cannot optimize based on the assumption that it is zero without changing
394 // to to an explicit zero. If we don't change it to zero, other code could
395 // optimized based on the contradictory assumption that it is non-zero.
396 // Because instcombine aggressively folds operations with undef args anyway,
397 // this won't lose us code quality.
398 if (Mask->isNullValue())
400 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V))
401 return ConstantExpr::getAnd(CI, Mask)->isNullValue();
403 if (Instruction *I = dyn_cast<Instruction>(V)) {
404 switch (I->getOpcode()) {
405 case Instruction::And:
406 // (X & C1) & C2 == 0 iff C1 & C2 == 0.
407 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(I->getOperand(1))) {
408 ConstantIntegral *C1C2 =
409 cast<ConstantIntegral>(ConstantExpr::getAnd(CI, Mask));
410 if (MaskedValueIsZero(I->getOperand(0), C1C2))
413 // If either the LHS or the RHS are MaskedValueIsZero, the result is zero.
414 return MaskedValueIsZero(I->getOperand(1), Mask) ||
415 MaskedValueIsZero(I->getOperand(0), Mask);
416 case Instruction::Or:
417 case Instruction::Xor:
418 // If the LHS and the RHS are MaskedValueIsZero, the result is also zero.
419 return MaskedValueIsZero(I->getOperand(1), Mask) &&
420 MaskedValueIsZero(I->getOperand(0), Mask);
421 case Instruction::Select:
422 // If the T and F values are MaskedValueIsZero, the result is also zero.
423 return MaskedValueIsZero(I->getOperand(2), Mask) &&
424 MaskedValueIsZero(I->getOperand(1), Mask);
425 case Instruction::Cast: {
426 const Type *SrcTy = I->getOperand(0)->getType();
427 if (SrcTy == Type::BoolTy)
428 return (Mask->getRawValue() & 1) == 0;
430 if (SrcTy->isInteger()) {
431 // (cast <ty> X to int) & C2 == 0 iff <ty> could not have contained C2.
432 if (SrcTy->isUnsigned() && // Only handle zero ext.
433 ConstantExpr::getCast(Mask, SrcTy)->isNullValue())
436 // If this is a noop cast, recurse.
437 if ((SrcTy->isSigned() && SrcTy->getUnsignedVersion() == I->getType())||
438 SrcTy->getSignedVersion() == I->getType()) {
440 ConstantExpr::getCast(Mask, I->getOperand(0)->getType());
441 return MaskedValueIsZero(I->getOperand(0),
442 cast<ConstantIntegral>(NewMask));
447 case Instruction::Shl:
448 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
449 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
450 return MaskedValueIsZero(I->getOperand(0),
451 cast<ConstantIntegral>(ConstantExpr::getUShr(Mask, SA)));
453 case Instruction::Shr:
454 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
455 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
456 if (I->getType()->isUnsigned()) {
457 Constant *C1 = ConstantIntegral::getAllOnesValue(I->getType());
458 C1 = ConstantExpr::getShr(C1, SA);
459 C1 = ConstantExpr::getAnd(C1, Mask);
460 if (C1->isNullValue())
470 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
471 // true when both operands are equal...
473 static bool isTrueWhenEqual(Instruction &I) {
474 return I.getOpcode() == Instruction::SetEQ ||
475 I.getOpcode() == Instruction::SetGE ||
476 I.getOpcode() == Instruction::SetLE;
479 /// AssociativeOpt - Perform an optimization on an associative operator. This
480 /// function is designed to check a chain of associative operators for a
481 /// potential to apply a certain optimization. Since the optimization may be
482 /// applicable if the expression was reassociated, this checks the chain, then
483 /// reassociates the expression as necessary to expose the optimization
484 /// opportunity. This makes use of a special Functor, which must define
485 /// 'shouldApply' and 'apply' methods.
487 template<typename Functor>
488 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
489 unsigned Opcode = Root.getOpcode();
490 Value *LHS = Root.getOperand(0);
492 // Quick check, see if the immediate LHS matches...
493 if (F.shouldApply(LHS))
494 return F.apply(Root);
496 // Otherwise, if the LHS is not of the same opcode as the root, return.
497 Instruction *LHSI = dyn_cast<Instruction>(LHS);
498 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
499 // Should we apply this transform to the RHS?
500 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
502 // If not to the RHS, check to see if we should apply to the LHS...
503 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
504 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
508 // If the functor wants to apply the optimization to the RHS of LHSI,
509 // reassociate the expression from ((? op A) op B) to (? op (A op B))
511 BasicBlock *BB = Root.getParent();
513 // Now all of the instructions are in the current basic block, go ahead
514 // and perform the reassociation.
515 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
517 // First move the selected RHS to the LHS of the root...
518 Root.setOperand(0, LHSI->getOperand(1));
520 // Make what used to be the LHS of the root be the user of the root...
521 Value *ExtraOperand = TmpLHSI->getOperand(1);
522 if (&Root == TmpLHSI) {
523 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
526 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
527 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
528 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
529 BasicBlock::iterator ARI = &Root; ++ARI;
530 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
533 // Now propagate the ExtraOperand down the chain of instructions until we
535 while (TmpLHSI != LHSI) {
536 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
537 // Move the instruction to immediately before the chain we are
538 // constructing to avoid breaking dominance properties.
539 NextLHSI->getParent()->getInstList().remove(NextLHSI);
540 BB->getInstList().insert(ARI, NextLHSI);
543 Value *NextOp = NextLHSI->getOperand(1);
544 NextLHSI->setOperand(1, ExtraOperand);
546 ExtraOperand = NextOp;
549 // Now that the instructions are reassociated, have the functor perform
550 // the transformation...
551 return F.apply(Root);
554 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
560 // AddRHS - Implements: X + X --> X << 1
563 AddRHS(Value *rhs) : RHS(rhs) {}
564 bool shouldApply(Value *LHS) const { return LHS == RHS; }
565 Instruction *apply(BinaryOperator &Add) const {
566 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
567 ConstantInt::get(Type::UByteTy, 1));
571 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
573 struct AddMaskingAnd {
575 AddMaskingAnd(Constant *c) : C2(c) {}
576 bool shouldApply(Value *LHS) const {
578 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
579 ConstantExpr::getAnd(C1, C2)->isNullValue();
581 Instruction *apply(BinaryOperator &Add) const {
582 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
586 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
588 if (isa<CastInst>(I)) {
589 if (Constant *SOC = dyn_cast<Constant>(SO))
590 return ConstantExpr::getCast(SOC, I.getType());
592 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
593 SO->getName() + ".cast"), I);
596 // Figure out if the constant is the left or the right argument.
597 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
598 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
600 if (Constant *SOC = dyn_cast<Constant>(SO)) {
602 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
603 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
606 Value *Op0 = SO, *Op1 = ConstOperand;
610 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
611 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
612 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
613 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
615 assert(0 && "Unknown binary instruction type!");
618 return IC->InsertNewInstBefore(New, I);
621 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
622 // constant as the other operand, try to fold the binary operator into the
623 // select arguments. This also works for Cast instructions, which obviously do
624 // not have a second operand.
625 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
627 // Don't modify shared select instructions
628 if (!SI->hasOneUse()) return 0;
629 Value *TV = SI->getOperand(1);
630 Value *FV = SI->getOperand(2);
632 if (isa<Constant>(TV) || isa<Constant>(FV)) {
633 // Bool selects with constant operands can be folded to logical ops.
634 if (SI->getType() == Type::BoolTy) return 0;
636 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
637 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
639 return new SelectInst(SI->getCondition(), SelectTrueVal,
646 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
647 /// node as operand #0, see if we can fold the instruction into the PHI (which
648 /// is only possible if all operands to the PHI are constants).
649 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
650 PHINode *PN = cast<PHINode>(I.getOperand(0));
651 unsigned NumPHIValues = PN->getNumIncomingValues();
652 if (!PN->hasOneUse() || NumPHIValues == 0 ||
653 !isa<Constant>(PN->getIncomingValue(0))) return 0;
655 // Check to see if all of the operands of the PHI are constants. If not, we
656 // cannot do the transformation.
657 for (unsigned i = 1; i != NumPHIValues; ++i)
658 if (!isa<Constant>(PN->getIncomingValue(i)))
661 // Okay, we can do the transformation: create the new PHI node.
662 PHINode *NewPN = new PHINode(I.getType(), I.getName());
664 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
665 InsertNewInstBefore(NewPN, *PN);
667 // Next, add all of the operands to the PHI.
668 if (I.getNumOperands() == 2) {
669 Constant *C = cast<Constant>(I.getOperand(1));
670 for (unsigned i = 0; i != NumPHIValues; ++i) {
671 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
672 NewPN->addIncoming(ConstantExpr::get(I.getOpcode(), InV, C),
673 PN->getIncomingBlock(i));
676 assert(isa<CastInst>(I) && "Unary op should be a cast!");
677 const Type *RetTy = I.getType();
678 for (unsigned i = 0; i != NumPHIValues; ++i) {
679 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
680 NewPN->addIncoming(ConstantExpr::getCast(InV, RetTy),
681 PN->getIncomingBlock(i));
684 return ReplaceInstUsesWith(I, NewPN);
687 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
688 bool Changed = SimplifyCommutative(I);
689 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
691 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
692 // X + undef -> undef
693 if (isa<UndefValue>(RHS))
694 return ReplaceInstUsesWith(I, RHS);
697 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
698 if (RHSC->isNullValue())
699 return ReplaceInstUsesWith(I, LHS);
700 } else if (cast<ConstantFP>(RHSC)->isExactlyValue(-0.0)) {
701 return ReplaceInstUsesWith(I, LHS);
704 // X + (signbit) --> X ^ signbit
705 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
706 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
707 uint64_t Val = CI->getRawValue() & (1ULL << NumBits)-1;
708 if (Val == (1ULL << (NumBits-1)))
709 return BinaryOperator::createXor(LHS, RHS);
712 if (isa<PHINode>(LHS))
713 if (Instruction *NV = FoldOpIntoPhi(I))
718 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
719 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
720 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
721 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
723 uint64_t C0080Val = 1ULL << 31;
724 int64_t CFF80Val = -C0080Val;
727 if (TySizeBits > Size) {
729 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
730 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
731 if (RHSSExt == CFF80Val) {
732 if (XorRHS->getZExtValue() == C0080Val)
734 } else if (RHSZExt == C0080Val) {
735 if (XorRHS->getSExtValue() == CFF80Val)
739 // This is a sign extend if the top bits are known zero.
740 Constant *Mask = ConstantInt::getAllOnesValue(XorLHS->getType());
741 Mask = ConstantExpr::getShl(Mask,
742 ConstantInt::get(Type::UByteTy, 64-TySizeBits-Size));
743 if (!MaskedValueIsZero(XorLHS, cast<ConstantInt>(Mask)))
744 Size = 0; // Not a sign ext, but can't be any others either.
754 const Type *MiddleType = 0;
757 case 32: MiddleType = Type::IntTy; break;
758 case 16: MiddleType = Type::ShortTy; break;
759 case 8: MiddleType = Type::SByteTy; break;
762 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
763 InsertNewInstBefore(NewTrunc, I);
764 return new CastInst(NewTrunc, I.getType());
770 if (I.getType()->isInteger()) {
771 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
773 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
774 if (RHSI->getOpcode() == Instruction::Sub)
775 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
776 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
778 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
779 if (LHSI->getOpcode() == Instruction::Sub)
780 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
781 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
786 if (Value *V = dyn_castNegVal(LHS))
787 return BinaryOperator::createSub(RHS, V);
790 if (!isa<Constant>(RHS))
791 if (Value *V = dyn_castNegVal(RHS))
792 return BinaryOperator::createSub(LHS, V);
796 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
797 if (X == RHS) // X*C + X --> X * (C+1)
798 return BinaryOperator::createMul(RHS, AddOne(C2));
800 // X*C1 + X*C2 --> X * (C1+C2)
802 if (X == dyn_castFoldableMul(RHS, C1))
803 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
806 // X + X*C --> X * (C+1)
807 if (dyn_castFoldableMul(RHS, C2) == LHS)
808 return BinaryOperator::createMul(LHS, AddOne(C2));
811 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
812 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
813 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
815 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
817 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
818 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
819 return BinaryOperator::createSub(C, X);
822 // (X & FF00) + xx00 -> (X+xx00) & FF00
823 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
824 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
826 // See if all bits from the first bit set in the Add RHS up are included
827 // in the mask. First, get the rightmost bit.
828 uint64_t AddRHSV = CRHS->getRawValue();
830 // Form a mask of all bits from the lowest bit added through the top.
831 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
832 AddRHSHighBits &= ~0ULL >> (64-C2->getType()->getPrimitiveSizeInBits());
834 // See if the and mask includes all of these bits.
835 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
837 if (AddRHSHighBits == AddRHSHighBitsAnd) {
838 // Okay, the xform is safe. Insert the new add pronto.
839 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
841 return BinaryOperator::createAnd(NewAdd, C2);
846 // Try to fold constant add into select arguments.
847 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
848 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
852 return Changed ? &I : 0;
855 // isSignBit - Return true if the value represented by the constant only has the
856 // highest order bit set.
857 static bool isSignBit(ConstantInt *CI) {
858 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
859 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
862 /// RemoveNoopCast - Strip off nonconverting casts from the value.
864 static Value *RemoveNoopCast(Value *V) {
865 if (CastInst *CI = dyn_cast<CastInst>(V)) {
866 const Type *CTy = CI->getType();
867 const Type *OpTy = CI->getOperand(0)->getType();
868 if (CTy->isInteger() && OpTy->isInteger()) {
869 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
870 return RemoveNoopCast(CI->getOperand(0));
871 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
872 return RemoveNoopCast(CI->getOperand(0));
877 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
878 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
880 if (Op0 == Op1) // sub X, X -> 0
881 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
883 // If this is a 'B = x-(-A)', change to B = x+A...
884 if (Value *V = dyn_castNegVal(Op1))
885 return BinaryOperator::createAdd(Op0, V);
887 if (isa<UndefValue>(Op0))
888 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
889 if (isa<UndefValue>(Op1))
890 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
892 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
893 // Replace (-1 - A) with (~A)...
894 if (C->isAllOnesValue())
895 return BinaryOperator::createNot(Op1);
897 // C - ~X == X + (1+C)
899 if (match(Op1, m_Not(m_Value(X))))
900 return BinaryOperator::createAdd(X,
901 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
902 // -((uint)X >> 31) -> ((int)X >> 31)
903 // -((int)X >> 31) -> ((uint)X >> 31)
904 if (C->isNullValue()) {
905 Value *NoopCastedRHS = RemoveNoopCast(Op1);
906 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
907 if (SI->getOpcode() == Instruction::Shr)
908 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
910 if (SI->getType()->isSigned())
911 NewTy = SI->getType()->getUnsignedVersion();
913 NewTy = SI->getType()->getSignedVersion();
914 // Check to see if we are shifting out everything but the sign bit.
915 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
916 // Ok, the transformation is safe. Insert a cast of the incoming
917 // value, then the new shift, then the new cast.
918 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
919 SI->getOperand(0)->getName());
920 Value *InV = InsertNewInstBefore(FirstCast, I);
921 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
923 if (NewShift->getType() == I.getType())
926 InV = InsertNewInstBefore(NewShift, I);
927 return new CastInst(NewShift, I.getType());
933 // Try to fold constant sub into select arguments.
934 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
935 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
938 if (isa<PHINode>(Op0))
939 if (Instruction *NV = FoldOpIntoPhi(I))
943 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
944 if (Op1I->getOpcode() == Instruction::Add &&
945 !Op0->getType()->isFloatingPoint()) {
946 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
947 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
948 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
949 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
950 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
951 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
952 // C1-(X+C2) --> (C1-C2)-X
953 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
954 Op1I->getOperand(0));
958 if (Op1I->hasOneUse()) {
959 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
960 // is not used by anyone else...
962 if (Op1I->getOpcode() == Instruction::Sub &&
963 !Op1I->getType()->isFloatingPoint()) {
964 // Swap the two operands of the subexpr...
965 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
966 Op1I->setOperand(0, IIOp1);
967 Op1I->setOperand(1, IIOp0);
969 // Create the new top level add instruction...
970 return BinaryOperator::createAdd(Op0, Op1);
973 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
975 if (Op1I->getOpcode() == Instruction::And &&
976 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
977 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
980 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
981 return BinaryOperator::createAnd(Op0, NewNot);
984 // -(X sdiv C) -> (X sdiv -C)
985 if (Op1I->getOpcode() == Instruction::Div)
986 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
987 if (CSI->isNullValue())
988 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
989 return BinaryOperator::createDiv(Op1I->getOperand(0),
990 ConstantExpr::getNeg(DivRHS));
992 // X - X*C --> X * (1-C)
994 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
996 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
997 return BinaryOperator::createMul(Op0, CP1);
1002 if (!Op0->getType()->isFloatingPoint())
1003 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1004 if (Op0I->getOpcode() == Instruction::Add) {
1005 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1006 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1007 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1008 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1009 } else if (Op0I->getOpcode() == Instruction::Sub) {
1010 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1011 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1015 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1016 if (X == Op1) { // X*C - X --> X * (C-1)
1017 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1018 return BinaryOperator::createMul(Op1, CP1);
1021 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1022 if (X == dyn_castFoldableMul(Op1, C2))
1023 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1028 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1029 /// really just returns true if the most significant (sign) bit is set.
1030 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1031 if (RHS->getType()->isSigned()) {
1032 // True if source is LHS < 0 or LHS <= -1
1033 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1034 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1036 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1037 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1038 // the size of the integer type.
1039 if (Opcode == Instruction::SetGE)
1040 return RHSC->getValue() ==
1041 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1042 if (Opcode == Instruction::SetGT)
1043 return RHSC->getValue() ==
1044 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1049 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1050 bool Changed = SimplifyCommutative(I);
1051 Value *Op0 = I.getOperand(0);
1053 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1054 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1056 // Simplify mul instructions with a constant RHS...
1057 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1058 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1060 // ((X << C1)*C2) == (X * (C2 << C1))
1061 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1062 if (SI->getOpcode() == Instruction::Shl)
1063 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1064 return BinaryOperator::createMul(SI->getOperand(0),
1065 ConstantExpr::getShl(CI, ShOp));
1067 if (CI->isNullValue())
1068 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1069 if (CI->equalsInt(1)) // X * 1 == X
1070 return ReplaceInstUsesWith(I, Op0);
1071 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1072 return BinaryOperator::createNeg(Op0, I.getName());
1074 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1075 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1076 uint64_t C = Log2_64(Val);
1077 return new ShiftInst(Instruction::Shl, Op0,
1078 ConstantUInt::get(Type::UByteTy, C));
1080 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1081 if (Op1F->isNullValue())
1082 return ReplaceInstUsesWith(I, Op1);
1084 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1085 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1086 if (Op1F->getValue() == 1.0)
1087 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1090 // Try to fold constant mul into select arguments.
1091 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1092 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1095 if (isa<PHINode>(Op0))
1096 if (Instruction *NV = FoldOpIntoPhi(I))
1100 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1101 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1102 return BinaryOperator::createMul(Op0v, Op1v);
1104 // If one of the operands of the multiply is a cast from a boolean value, then
1105 // we know the bool is either zero or one, so this is a 'masking' multiply.
1106 // See if we can simplify things based on how the boolean was originally
1108 CastInst *BoolCast = 0;
1109 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1110 if (CI->getOperand(0)->getType() == Type::BoolTy)
1113 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1114 if (CI->getOperand(0)->getType() == Type::BoolTy)
1117 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1118 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1119 const Type *SCOpTy = SCIOp0->getType();
1121 // If the setcc is true iff the sign bit of X is set, then convert this
1122 // multiply into a shift/and combination.
1123 if (isa<ConstantInt>(SCIOp1) &&
1124 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1125 // Shift the X value right to turn it into "all signbits".
1126 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1127 SCOpTy->getPrimitiveSizeInBits()-1);
1128 if (SCIOp0->getType()->isUnsigned()) {
1129 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1130 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1131 SCIOp0->getName()), I);
1135 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1136 BoolCast->getOperand(0)->getName()+
1139 // If the multiply type is not the same as the source type, sign extend
1140 // or truncate to the multiply type.
1141 if (I.getType() != V->getType())
1142 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1144 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1145 return BinaryOperator::createAnd(V, OtherOp);
1150 return Changed ? &I : 0;
1153 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1154 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1156 if (isa<UndefValue>(Op0)) // undef / X -> 0
1157 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1158 if (isa<UndefValue>(Op1))
1159 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1161 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1163 if (RHS->equalsInt(1))
1164 return ReplaceInstUsesWith(I, Op0);
1167 if (RHS->isAllOnesValue())
1168 return BinaryOperator::createNeg(Op0);
1170 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1171 if (LHS->getOpcode() == Instruction::Div)
1172 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1173 // (X / C1) / C2 -> X / (C1*C2)
1174 return BinaryOperator::createDiv(LHS->getOperand(0),
1175 ConstantExpr::getMul(RHS, LHSRHS));
1178 // Check to see if this is an unsigned division with an exact power of 2,
1179 // if so, convert to a right shift.
1180 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1181 if (uint64_t Val = C->getValue()) // Don't break X / 0
1182 if (isPowerOf2_64(Val)) {
1183 uint64_t C = Log2_64(Val);
1184 return new ShiftInst(Instruction::Shr, Op0,
1185 ConstantUInt::get(Type::UByteTy, C));
1189 if (RHS->getType()->isSigned())
1190 if (Value *LHSNeg = dyn_castNegVal(Op0))
1191 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1193 if (!RHS->isNullValue()) {
1194 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1195 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1197 if (isa<PHINode>(Op0))
1198 if (Instruction *NV = FoldOpIntoPhi(I))
1203 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1204 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
1205 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1206 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1207 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1208 if (STO->getValue() == 0) { // Couldn't be this argument.
1209 I.setOperand(1, SFO);
1211 } else if (SFO->getValue() == 0) {
1212 I.setOperand(1, STO);
1216 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
1217 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
1218 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
1219 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
1220 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
1221 TC, SI->getName()+".t");
1222 TSI = InsertNewInstBefore(TSI, I);
1224 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
1225 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
1226 FC, SI->getName()+".f");
1227 FSI = InsertNewInstBefore(FSI, I);
1228 return new SelectInst(SI->getOperand(0), TSI, FSI);
1232 // 0 / X == 0, we don't need to preserve faults!
1233 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1234 if (LHS->equalsInt(0))
1235 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1241 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
1242 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1243 if (I.getType()->isSigned())
1244 if (Value *RHSNeg = dyn_castNegVal(Op1))
1245 if (!isa<ConstantSInt>(RHSNeg) ||
1246 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
1248 AddUsesToWorkList(I);
1249 I.setOperand(1, RHSNeg);
1253 if (isa<UndefValue>(Op0)) // undef % X -> 0
1254 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1255 if (isa<UndefValue>(Op1))
1256 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
1258 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1259 if (RHS->equalsInt(1)) // X % 1 == 0
1260 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1262 // Check to see if this is an unsigned remainder with an exact power of 2,
1263 // if so, convert to a bitwise and.
1264 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1265 if (uint64_t Val = C->getValue()) // Don't break X % 0 (divide by zero)
1266 if (!(Val & (Val-1))) // Power of 2
1267 return BinaryOperator::createAnd(Op0,
1268 ConstantUInt::get(I.getType(), Val-1));
1270 if (!RHS->isNullValue()) {
1271 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1272 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1274 if (isa<PHINode>(Op0))
1275 if (Instruction *NV = FoldOpIntoPhi(I))
1280 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1281 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
1282 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1283 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1284 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1285 if (STO->getValue() == 0) { // Couldn't be this argument.
1286 I.setOperand(1, SFO);
1288 } else if (SFO->getValue() == 0) {
1289 I.setOperand(1, STO);
1293 if (!(STO->getValue() & (STO->getValue()-1)) &&
1294 !(SFO->getValue() & (SFO->getValue()-1))) {
1295 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1296 SubOne(STO), SI->getName()+".t"), I);
1297 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1298 SubOne(SFO), SI->getName()+".f"), I);
1299 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
1303 // 0 % X == 0, we don't need to preserve faults!
1304 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1305 if (LHS->equalsInt(0))
1306 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1311 // isMaxValueMinusOne - return true if this is Max-1
1312 static bool isMaxValueMinusOne(const ConstantInt *C) {
1313 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C)) {
1314 // Calculate -1 casted to the right type...
1315 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1316 uint64_t Val = ~0ULL; // All ones
1317 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
1318 return CU->getValue() == Val-1;
1321 const ConstantSInt *CS = cast<ConstantSInt>(C);
1323 // Calculate 0111111111..11111
1324 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1325 int64_t Val = INT64_MAX; // All ones
1326 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
1327 return CS->getValue() == Val-1;
1330 // isMinValuePlusOne - return true if this is Min+1
1331 static bool isMinValuePlusOne(const ConstantInt *C) {
1332 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1333 return CU->getValue() == 1;
1335 const ConstantSInt *CS = cast<ConstantSInt>(C);
1337 // Calculate 1111111111000000000000
1338 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1339 int64_t Val = -1; // All ones
1340 Val <<= TypeBits-1; // Shift over to the right spot
1341 return CS->getValue() == Val+1;
1344 // isOneBitSet - Return true if there is exactly one bit set in the specified
1346 static bool isOneBitSet(const ConstantInt *CI) {
1347 uint64_t V = CI->getRawValue();
1348 return V && (V & (V-1)) == 0;
1351 #if 0 // Currently unused
1352 // isLowOnes - Return true if the constant is of the form 0+1+.
1353 static bool isLowOnes(const ConstantInt *CI) {
1354 uint64_t V = CI->getRawValue();
1356 // There won't be bits set in parts that the type doesn't contain.
1357 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1359 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1360 return U && V && (U & V) == 0;
1364 // isHighOnes - Return true if the constant is of the form 1+0+.
1365 // This is the same as lowones(~X).
1366 static bool isHighOnes(const ConstantInt *CI) {
1367 uint64_t V = ~CI->getRawValue();
1368 if (~V == 0) return false; // 0's does not match "1+"
1370 // There won't be bits set in parts that the type doesn't contain.
1371 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1373 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1374 return U && V && (U & V) == 0;
1378 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
1379 /// are carefully arranged to allow folding of expressions such as:
1381 /// (A < B) | (A > B) --> (A != B)
1383 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
1384 /// represents that the comparison is true if A == B, and bit value '1' is true
1387 static unsigned getSetCondCode(const SetCondInst *SCI) {
1388 switch (SCI->getOpcode()) {
1390 case Instruction::SetGT: return 1;
1391 case Instruction::SetEQ: return 2;
1392 case Instruction::SetGE: return 3;
1393 case Instruction::SetLT: return 4;
1394 case Instruction::SetNE: return 5;
1395 case Instruction::SetLE: return 6;
1398 assert(0 && "Invalid SetCC opcode!");
1403 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
1404 /// opcode and two operands into either a constant true or false, or a brand new
1405 /// SetCC instruction.
1406 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
1408 case 0: return ConstantBool::False;
1409 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
1410 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
1411 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
1412 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
1413 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
1414 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
1415 case 7: return ConstantBool::True;
1416 default: assert(0 && "Illegal SetCCCode!"); return 0;
1420 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
1421 struct FoldSetCCLogical {
1424 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
1425 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
1426 bool shouldApply(Value *V) const {
1427 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
1428 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
1429 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
1432 Instruction *apply(BinaryOperator &Log) const {
1433 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
1434 if (SCI->getOperand(0) != LHS) {
1435 assert(SCI->getOperand(1) == LHS);
1436 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
1439 unsigned LHSCode = getSetCondCode(SCI);
1440 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
1442 switch (Log.getOpcode()) {
1443 case Instruction::And: Code = LHSCode & RHSCode; break;
1444 case Instruction::Or: Code = LHSCode | RHSCode; break;
1445 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
1446 default: assert(0 && "Illegal logical opcode!"); return 0;
1449 Value *RV = getSetCCValue(Code, LHS, RHS);
1450 if (Instruction *I = dyn_cast<Instruction>(RV))
1452 // Otherwise, it's a constant boolean value...
1453 return IC.ReplaceInstUsesWith(Log, RV);
1457 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
1458 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
1459 // guaranteed to be either a shift instruction or a binary operator.
1460 Instruction *InstCombiner::OptAndOp(Instruction *Op,
1461 ConstantIntegral *OpRHS,
1462 ConstantIntegral *AndRHS,
1463 BinaryOperator &TheAnd) {
1464 Value *X = Op->getOperand(0);
1465 Constant *Together = 0;
1466 if (!isa<ShiftInst>(Op))
1467 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
1469 switch (Op->getOpcode()) {
1470 case Instruction::Xor:
1471 if (Op->hasOneUse()) {
1472 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1473 std::string OpName = Op->getName(); Op->setName("");
1474 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
1475 InsertNewInstBefore(And, TheAnd);
1476 return BinaryOperator::createXor(And, Together);
1479 case Instruction::Or:
1480 if (Together == AndRHS) // (X | C) & C --> C
1481 return ReplaceInstUsesWith(TheAnd, AndRHS);
1483 if (Op->hasOneUse() && Together != OpRHS) {
1484 // (X | C1) & C2 --> (X | (C1&C2)) & C2
1485 std::string Op0Name = Op->getName(); Op->setName("");
1486 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
1487 InsertNewInstBefore(Or, TheAnd);
1488 return BinaryOperator::createAnd(Or, AndRHS);
1491 case Instruction::Add:
1492 if (Op->hasOneUse()) {
1493 // Adding a one to a single bit bit-field should be turned into an XOR
1494 // of the bit. First thing to check is to see if this AND is with a
1495 // single bit constant.
1496 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
1498 // Clear bits that are not part of the constant.
1499 AndRHSV &= ~0ULL >> (64-AndRHS->getType()->getPrimitiveSizeInBits());
1501 // If there is only one bit set...
1502 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
1503 // Ok, at this point, we know that we are masking the result of the
1504 // ADD down to exactly one bit. If the constant we are adding has
1505 // no bits set below this bit, then we can eliminate the ADD.
1506 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
1508 // Check to see if any bits below the one bit set in AndRHSV are set.
1509 if ((AddRHS & (AndRHSV-1)) == 0) {
1510 // If not, the only thing that can effect the output of the AND is
1511 // the bit specified by AndRHSV. If that bit is set, the effect of
1512 // the XOR is to toggle the bit. If it is clear, then the ADD has
1514 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
1515 TheAnd.setOperand(0, X);
1518 std::string Name = Op->getName(); Op->setName("");
1519 // Pull the XOR out of the AND.
1520 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
1521 InsertNewInstBefore(NewAnd, TheAnd);
1522 return BinaryOperator::createXor(NewAnd, AndRHS);
1529 case Instruction::Shl: {
1530 // We know that the AND will not produce any of the bits shifted in, so if
1531 // the anded constant includes them, clear them now!
1533 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1534 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
1535 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
1537 if (CI == ShlMask) { // Masking out bits that the shift already masks
1538 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
1539 } else if (CI != AndRHS) { // Reducing bits set in and.
1540 TheAnd.setOperand(1, CI);
1545 case Instruction::Shr:
1546 // We know that the AND will not produce any of the bits shifted in, so if
1547 // the anded constant includes them, clear them now! This only applies to
1548 // unsigned shifts, because a signed shr may bring in set bits!
1550 if (AndRHS->getType()->isUnsigned()) {
1551 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1552 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
1553 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1555 if (CI == ShrMask) { // Masking out bits that the shift already masks.
1556 return ReplaceInstUsesWith(TheAnd, Op);
1557 } else if (CI != AndRHS) {
1558 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
1561 } else { // Signed shr.
1562 // See if this is shifting in some sign extension, then masking it out
1564 if (Op->hasOneUse()) {
1565 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1566 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
1567 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1568 if (CI == AndRHS) { // Masking out bits shifted in.
1569 // Make the argument unsigned.
1570 Value *ShVal = Op->getOperand(0);
1571 ShVal = InsertCastBefore(ShVal,
1572 ShVal->getType()->getUnsignedVersion(),
1574 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
1575 OpRHS, Op->getName()),
1577 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
1578 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
1581 return new CastInst(ShVal, Op->getType());
1591 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
1592 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
1593 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
1594 /// insert new instructions.
1595 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
1596 bool Inside, Instruction &IB) {
1597 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
1598 "Lo is not <= Hi in range emission code!");
1600 if (Lo == Hi) // Trivially false.
1601 return new SetCondInst(Instruction::SetNE, V, V);
1602 if (cast<ConstantIntegral>(Lo)->isMinValue())
1603 return new SetCondInst(Instruction::SetLT, V, Hi);
1605 Constant *AddCST = ConstantExpr::getNeg(Lo);
1606 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
1607 InsertNewInstBefore(Add, IB);
1608 // Convert to unsigned for the comparison.
1609 const Type *UnsType = Add->getType()->getUnsignedVersion();
1610 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1611 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1612 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1613 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
1616 if (Lo == Hi) // Trivially true.
1617 return new SetCondInst(Instruction::SetEQ, V, V);
1619 Hi = SubOne(cast<ConstantInt>(Hi));
1620 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
1621 return new SetCondInst(Instruction::SetGT, V, Hi);
1623 // Emit X-Lo > Hi-Lo-1
1624 Constant *AddCST = ConstantExpr::getNeg(Lo);
1625 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
1626 InsertNewInstBefore(Add, IB);
1627 // Convert to unsigned for the comparison.
1628 const Type *UnsType = Add->getType()->getUnsignedVersion();
1629 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1630 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1631 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1632 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
1635 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
1636 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
1637 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
1638 // not, since all 1s are not contiguous.
1639 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
1640 uint64_t V = Val->getRawValue();
1641 if (!isShiftedMask_64(V)) return false;
1643 // look for the first zero bit after the run of ones
1644 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
1645 // look for the first non-zero bit
1646 ME = 64-CountLeadingZeros_64(V);
1652 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
1653 /// where isSub determines whether the operator is a sub. If we can fold one of
1654 /// the following xforms:
1656 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
1657 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1658 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1660 /// return (A +/- B).
1662 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
1663 ConstantIntegral *Mask, bool isSub,
1665 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1666 if (!LHSI || LHSI->getNumOperands() != 2 ||
1667 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
1669 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
1671 switch (LHSI->getOpcode()) {
1673 case Instruction::And:
1674 if (ConstantExpr::getAnd(N, Mask) == Mask) {
1675 // If the AndRHS is a power of two minus one (0+1+), this is simple.
1676 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
1679 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
1680 // part, we don't need any explicit masks to take them out of A. If that
1681 // is all N is, ignore it.
1683 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
1684 Constant *Mask = ConstantInt::getAllOnesValue(RHS->getType());
1685 Mask = ConstantExpr::getUShr(Mask,
1686 ConstantInt::get(Type::UByteTy,
1688 if (MaskedValueIsZero(RHS, cast<ConstantIntegral>(Mask)))
1693 case Instruction::Or:
1694 case Instruction::Xor:
1695 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
1696 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
1697 ConstantExpr::getAnd(N, Mask)->isNullValue())
1704 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
1706 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
1707 return InsertNewInstBefore(New, I);
1711 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1712 bool Changed = SimplifyCommutative(I);
1713 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1715 if (isa<UndefValue>(Op1)) // X & undef -> 0
1716 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1720 return ReplaceInstUsesWith(I, Op1);
1722 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
1724 if (AndRHS->isAllOnesValue())
1725 return ReplaceInstUsesWith(I, Op0);
1727 if (MaskedValueIsZero(Op0, AndRHS)) // LHS & RHS == 0
1728 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1730 // If the mask is not masking out any bits, there is no reason to do the
1731 // and in the first place.
1732 ConstantIntegral *NotAndRHS =
1733 cast<ConstantIntegral>(ConstantExpr::getNot(AndRHS));
1734 if (MaskedValueIsZero(Op0, NotAndRHS))
1735 return ReplaceInstUsesWith(I, Op0);
1737 // Optimize a variety of ((val OP C1) & C2) combinations...
1738 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
1739 Instruction *Op0I = cast<Instruction>(Op0);
1740 Value *Op0LHS = Op0I->getOperand(0);
1741 Value *Op0RHS = Op0I->getOperand(1);
1742 switch (Op0I->getOpcode()) {
1743 case Instruction::Xor:
1744 case Instruction::Or:
1745 // (X ^ V) & C2 --> (X & C2) iff (V & C2) == 0
1746 // (X | V) & C2 --> (X & C2) iff (V & C2) == 0
1747 if (MaskedValueIsZero(Op0LHS, AndRHS))
1748 return BinaryOperator::createAnd(Op0RHS, AndRHS);
1749 if (MaskedValueIsZero(Op0RHS, AndRHS))
1750 return BinaryOperator::createAnd(Op0LHS, AndRHS);
1752 // If the mask is only needed on one incoming arm, push it up.
1753 if (Op0I->hasOneUse()) {
1754 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1755 // Not masking anything out for the LHS, move to RHS.
1756 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
1757 Op0RHS->getName()+".masked");
1758 InsertNewInstBefore(NewRHS, I);
1759 return BinaryOperator::create(
1760 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
1762 if (!isa<Constant>(NotAndRHS) &&
1763 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1764 // Not masking anything out for the RHS, move to LHS.
1765 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
1766 Op0LHS->getName()+".masked");
1767 InsertNewInstBefore(NewLHS, I);
1768 return BinaryOperator::create(
1769 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
1774 case Instruction::And:
1775 // (X & V) & C2 --> 0 iff (V & C2) == 0
1776 if (MaskedValueIsZero(Op0LHS, AndRHS) ||
1777 MaskedValueIsZero(Op0RHS, AndRHS))
1778 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1780 case Instruction::Add:
1781 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1782 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1783 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1784 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1785 return BinaryOperator::createAnd(V, AndRHS);
1786 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1787 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
1790 case Instruction::Sub:
1791 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1792 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1793 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1794 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1795 return BinaryOperator::createAnd(V, AndRHS);
1799 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1800 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1802 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
1803 const Type *SrcTy = CI->getOperand(0)->getType();
1805 // If this is an integer truncation or change from signed-to-unsigned, and
1806 // if the source is an and/or with immediate, transform it. This
1807 // frequently occurs for bitfield accesses.
1808 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
1809 if (SrcTy->getPrimitiveSizeInBits() >=
1810 I.getType()->getPrimitiveSizeInBits() &&
1811 CastOp->getNumOperands() == 2)
1812 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1)))
1813 if (CastOp->getOpcode() == Instruction::And) {
1814 // Change: and (cast (and X, C1) to T), C2
1815 // into : and (cast X to T), trunc(C1)&C2
1816 // This will folds the two ands together, which may allow other
1818 Instruction *NewCast =
1819 new CastInst(CastOp->getOperand(0), I.getType(),
1820 CastOp->getName()+".shrunk");
1821 NewCast = InsertNewInstBefore(NewCast, I);
1823 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
1824 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
1825 return BinaryOperator::createAnd(NewCast, C3);
1826 } else if (CastOp->getOpcode() == Instruction::Or) {
1827 // Change: and (cast (or X, C1) to T), C2
1828 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
1829 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
1830 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
1831 return ReplaceInstUsesWith(I, AndRHS);
1836 // If this is an integer sign or zero extension instruction.
1837 if (SrcTy->isIntegral() &&
1838 SrcTy->getPrimitiveSizeInBits() <
1839 CI->getType()->getPrimitiveSizeInBits()) {
1841 if (SrcTy->isUnsigned()) {
1842 // See if this and is clearing out bits that are known to be zero
1843 // anyway (due to the zero extension).
1844 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
1845 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
1846 Constant *Result = ConstantExpr::getAnd(Mask, AndRHS);
1847 if (Result == Mask) // The "and" isn't doing anything, remove it.
1848 return ReplaceInstUsesWith(I, CI);
1849 if (Result != AndRHS) { // Reduce the and RHS constant.
1850 I.setOperand(1, Result);
1855 if (CI->hasOneUse() && SrcTy->isInteger()) {
1856 // We can only do this if all of the sign bits brought in are masked
1857 // out. Compute this by first getting 0000011111, then inverting
1859 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
1860 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
1861 Mask = ConstantExpr::getNot(Mask); // 1's in the new bits.
1862 if (ConstantExpr::getAnd(Mask, AndRHS)->isNullValue()) {
1863 // If the and is clearing all of the sign bits, change this to a
1864 // zero extension cast. To do this, cast the cast input to
1865 // unsigned, then to the requested size.
1866 Value *CastOp = CI->getOperand(0);
1868 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
1869 CI->getName()+".uns");
1870 NC = InsertNewInstBefore(NC, I);
1871 // Finally, insert a replacement for CI.
1872 NC = new CastInst(NC, CI->getType(), CI->getName());
1874 NC = InsertNewInstBefore(NC, I);
1875 WorkList.push_back(CI); // Delete CI later.
1876 I.setOperand(0, NC);
1877 return &I; // The AND operand was modified.
1884 // Try to fold constant and into select arguments.
1885 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1886 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1888 if (isa<PHINode>(Op0))
1889 if (Instruction *NV = FoldOpIntoPhi(I))
1893 Value *Op0NotVal = dyn_castNotVal(Op0);
1894 Value *Op1NotVal = dyn_castNotVal(Op1);
1896 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
1897 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1899 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1900 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
1901 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
1902 I.getName()+".demorgan");
1903 InsertNewInstBefore(Or, I);
1904 return BinaryOperator::createNot(Or);
1907 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
1908 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
1909 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
1912 Value *LHSVal, *RHSVal;
1913 ConstantInt *LHSCst, *RHSCst;
1914 Instruction::BinaryOps LHSCC, RHSCC;
1915 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
1916 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
1917 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
1918 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
1919 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
1920 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
1921 // Ensure that the larger constant is on the RHS.
1922 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
1923 SetCondInst *LHS = cast<SetCondInst>(Op0);
1924 if (cast<ConstantBool>(Cmp)->getValue()) {
1925 std::swap(LHS, RHS);
1926 std::swap(LHSCst, RHSCst);
1927 std::swap(LHSCC, RHSCC);
1930 // At this point, we know we have have two setcc instructions
1931 // comparing a value against two constants and and'ing the result
1932 // together. Because of the above check, we know that we only have
1933 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
1934 // FoldSetCCLogical check above), that the two constants are not
1936 assert(LHSCst != RHSCst && "Compares not folded above?");
1939 default: assert(0 && "Unknown integer condition code!");
1940 case Instruction::SetEQ:
1942 default: assert(0 && "Unknown integer condition code!");
1943 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
1944 case Instruction::SetGT: // (X == 13 & X > 15) -> false
1945 return ReplaceInstUsesWith(I, ConstantBool::False);
1946 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
1947 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
1948 return ReplaceInstUsesWith(I, LHS);
1950 case Instruction::SetNE:
1952 default: assert(0 && "Unknown integer condition code!");
1953 case Instruction::SetLT:
1954 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
1955 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
1956 break; // (X != 13 & X < 15) -> no change
1957 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
1958 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
1959 return ReplaceInstUsesWith(I, RHS);
1960 case Instruction::SetNE:
1961 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
1962 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1963 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
1964 LHSVal->getName()+".off");
1965 InsertNewInstBefore(Add, I);
1966 const Type *UnsType = Add->getType()->getUnsignedVersion();
1967 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
1968 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
1969 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1970 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
1972 break; // (X != 13 & X != 15) -> no change
1975 case Instruction::SetLT:
1977 default: assert(0 && "Unknown integer condition code!");
1978 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
1979 case Instruction::SetGT: // (X < 13 & X > 15) -> false
1980 return ReplaceInstUsesWith(I, ConstantBool::False);
1981 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
1982 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
1983 return ReplaceInstUsesWith(I, LHS);
1985 case Instruction::SetGT:
1987 default: assert(0 && "Unknown integer condition code!");
1988 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
1989 return ReplaceInstUsesWith(I, LHS);
1990 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
1991 return ReplaceInstUsesWith(I, RHS);
1992 case Instruction::SetNE:
1993 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
1994 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
1995 break; // (X > 13 & X != 15) -> no change
1996 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
1997 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2003 return Changed ? &I : 0;
2006 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2007 bool Changed = SimplifyCommutative(I);
2008 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2010 if (isa<UndefValue>(Op1))
2011 return ReplaceInstUsesWith(I, // X | undef -> -1
2012 ConstantIntegral::getAllOnesValue(I.getType()));
2014 // or X, X = X or X, 0 == X
2015 if (Op0 == Op1 || Op1 == Constant::getNullValue(I.getType()))
2016 return ReplaceInstUsesWith(I, Op0);
2019 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2020 // If X is known to only contain bits that already exist in RHS, just
2021 // replace this instruction with RHS directly.
2022 if (MaskedValueIsZero(Op0,
2023 cast<ConstantIntegral>(ConstantExpr::getNot(RHS))))
2024 return ReplaceInstUsesWith(I, RHS);
2026 ConstantInt *C1; Value *X;
2027 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2028 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2029 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
2031 InsertNewInstBefore(Or, I);
2032 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
2035 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2036 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2037 std::string Op0Name = Op0->getName(); Op0->setName("");
2038 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
2039 InsertNewInstBefore(Or, I);
2040 return BinaryOperator::createXor(Or,
2041 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
2044 // Try to fold constant and into select arguments.
2045 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2046 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2048 if (isa<PHINode>(Op0))
2049 if (Instruction *NV = FoldOpIntoPhi(I))
2053 Value *A, *B; ConstantInt *C1, *C2;
2055 if (match(Op0, m_And(m_Value(A), m_Value(B))))
2056 if (A == Op1 || B == Op1) // (A & ?) | A --> A
2057 return ReplaceInstUsesWith(I, Op1);
2058 if (match(Op1, m_And(m_Value(A), m_Value(B))))
2059 if (A == Op0 || B == Op0) // A | (A & ?) --> A
2060 return ReplaceInstUsesWith(I, Op0);
2062 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2063 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2064 MaskedValueIsZero(Op1, C1)) {
2065 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
2067 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2070 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2071 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2072 MaskedValueIsZero(Op0, C1)) {
2073 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
2075 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2078 // (A & C1)|(B & C2)
2079 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
2080 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
2082 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
2083 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
2086 // If we have: ((V + N) & C1) | (V & C2)
2087 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2088 // replace with V+N.
2089 if (C1 == ConstantExpr::getNot(C2)) {
2091 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
2092 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
2093 // Add commutes, try both ways.
2094 if (V1 == B && MaskedValueIsZero(V2, C2))
2095 return ReplaceInstUsesWith(I, A);
2096 if (V2 == B && MaskedValueIsZero(V1, C2))
2097 return ReplaceInstUsesWith(I, A);
2099 // Or commutes, try both ways.
2100 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
2101 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
2102 // Add commutes, try both ways.
2103 if (V1 == A && MaskedValueIsZero(V2, C1))
2104 return ReplaceInstUsesWith(I, B);
2105 if (V2 == A && MaskedValueIsZero(V1, C1))
2106 return ReplaceInstUsesWith(I, B);
2111 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
2112 if (A == Op1) // ~A | A == -1
2113 return ReplaceInstUsesWith(I,
2114 ConstantIntegral::getAllOnesValue(I.getType()));
2118 // Note, A is still live here!
2119 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
2121 return ReplaceInstUsesWith(I,
2122 ConstantIntegral::getAllOnesValue(I.getType()));
2124 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2125 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2126 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
2127 I.getName()+".demorgan"), I);
2128 return BinaryOperator::createNot(And);
2132 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
2133 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
2134 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2137 Value *LHSVal, *RHSVal;
2138 ConstantInt *LHSCst, *RHSCst;
2139 Instruction::BinaryOps LHSCC, RHSCC;
2140 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2141 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2142 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
2143 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2144 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2145 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2146 // Ensure that the larger constant is on the RHS.
2147 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2148 SetCondInst *LHS = cast<SetCondInst>(Op0);
2149 if (cast<ConstantBool>(Cmp)->getValue()) {
2150 std::swap(LHS, RHS);
2151 std::swap(LHSCst, RHSCst);
2152 std::swap(LHSCC, RHSCC);
2155 // At this point, we know we have have two setcc instructions
2156 // comparing a value against two constants and or'ing the result
2157 // together. Because of the above check, we know that we only have
2158 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2159 // FoldSetCCLogical check above), that the two constants are not
2161 assert(LHSCst != RHSCst && "Compares not folded above?");
2164 default: assert(0 && "Unknown integer condition code!");
2165 case Instruction::SetEQ:
2167 default: assert(0 && "Unknown integer condition code!");
2168 case Instruction::SetEQ:
2169 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
2170 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2171 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2172 LHSVal->getName()+".off");
2173 InsertNewInstBefore(Add, I);
2174 const Type *UnsType = Add->getType()->getUnsignedVersion();
2175 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2176 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
2177 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2178 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2180 break; // (X == 13 | X == 15) -> no change
2182 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
2184 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
2185 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
2186 return ReplaceInstUsesWith(I, RHS);
2189 case Instruction::SetNE:
2191 default: assert(0 && "Unknown integer condition code!");
2192 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
2193 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
2194 return ReplaceInstUsesWith(I, LHS);
2195 case Instruction::SetNE: // (X != 13 | X != 15) -> true
2196 case Instruction::SetLT: // (X != 13 | X < 15) -> true
2197 return ReplaceInstUsesWith(I, ConstantBool::True);
2200 case Instruction::SetLT:
2202 default: assert(0 && "Unknown integer condition code!");
2203 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
2205 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
2206 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
2207 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
2208 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
2209 return ReplaceInstUsesWith(I, RHS);
2212 case Instruction::SetGT:
2214 default: assert(0 && "Unknown integer condition code!");
2215 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
2216 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
2217 return ReplaceInstUsesWith(I, LHS);
2218 case Instruction::SetNE: // (X > 13 | X != 15) -> true
2219 case Instruction::SetLT: // (X > 13 | X < 15) -> true
2220 return ReplaceInstUsesWith(I, ConstantBool::True);
2226 return Changed ? &I : 0;
2229 // XorSelf - Implements: X ^ X --> 0
2232 XorSelf(Value *rhs) : RHS(rhs) {}
2233 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2234 Instruction *apply(BinaryOperator &Xor) const {
2240 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2241 bool Changed = SimplifyCommutative(I);
2242 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2244 if (isa<UndefValue>(Op1))
2245 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
2247 // xor X, X = 0, even if X is nested in a sequence of Xor's.
2248 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
2249 assert(Result == &I && "AssociativeOpt didn't work?");
2250 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2253 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2255 if (RHS->isNullValue())
2256 return ReplaceInstUsesWith(I, Op0);
2258 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2259 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
2260 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
2261 if (RHS == ConstantBool::True && SCI->hasOneUse())
2262 return new SetCondInst(SCI->getInverseCondition(),
2263 SCI->getOperand(0), SCI->getOperand(1));
2265 // ~(c-X) == X-c-1 == X+(-c-1)
2266 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2267 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2268 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2269 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2270 ConstantInt::get(I.getType(), 1));
2271 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
2274 // ~(~X & Y) --> (X | ~Y)
2275 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
2276 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
2277 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2279 BinaryOperator::createNot(Op0I->getOperand(1),
2280 Op0I->getOperand(1)->getName()+".not");
2281 InsertNewInstBefore(NotY, I);
2282 return BinaryOperator::createOr(Op0NotVal, NotY);
2286 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2287 switch (Op0I->getOpcode()) {
2288 case Instruction::Add:
2289 // ~(X-c) --> (-c-1)-X
2290 if (RHS->isAllOnesValue()) {
2291 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2292 return BinaryOperator::createSub(
2293 ConstantExpr::getSub(NegOp0CI,
2294 ConstantInt::get(I.getType(), 1)),
2295 Op0I->getOperand(0));
2298 case Instruction::And:
2299 // (X & C1) ^ C2 --> (X & C1) | C2 iff (C1&C2) == 0
2300 if (ConstantExpr::getAnd(RHS, Op0CI)->isNullValue())
2301 return BinaryOperator::createOr(Op0, RHS);
2303 case Instruction::Or:
2304 // (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
2305 if (ConstantExpr::getAnd(RHS, Op0CI) == RHS)
2306 return BinaryOperator::createAnd(Op0, ConstantExpr::getNot(RHS));
2312 // Try to fold constant and into select arguments.
2313 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2314 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2316 if (isa<PHINode>(Op0))
2317 if (Instruction *NV = FoldOpIntoPhi(I))
2321 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
2323 return ReplaceInstUsesWith(I,
2324 ConstantIntegral::getAllOnesValue(I.getType()));
2326 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
2328 return ReplaceInstUsesWith(I,
2329 ConstantIntegral::getAllOnesValue(I.getType()));
2331 if (Instruction *Op1I = dyn_cast<Instruction>(Op1))
2332 if (Op1I->getOpcode() == Instruction::Or) {
2333 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
2334 cast<BinaryOperator>(Op1I)->swapOperands();
2336 std::swap(Op0, Op1);
2337 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
2339 std::swap(Op0, Op1);
2341 } else if (Op1I->getOpcode() == Instruction::Xor) {
2342 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
2343 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
2344 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
2345 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
2348 if (Instruction *Op0I = dyn_cast<Instruction>(Op0))
2349 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
2350 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
2351 cast<BinaryOperator>(Op0I)->swapOperands();
2352 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
2353 Value *NotB = InsertNewInstBefore(BinaryOperator::createNot(Op1,
2354 Op1->getName()+".not"), I);
2355 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
2357 } else if (Op0I->getOpcode() == Instruction::Xor) {
2358 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
2359 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2360 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
2361 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2364 // (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
2365 Value *A, *B; ConstantInt *C1, *C2;
2366 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
2367 match(Op1, m_And(m_Value(B), m_ConstantInt(C2))) &&
2368 ConstantExpr::getAnd(C1, C2)->isNullValue())
2369 return BinaryOperator::createOr(Op0, Op1);
2371 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
2372 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
2373 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2376 return Changed ? &I : 0;
2379 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
2380 /// overflowed for this type.
2381 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2383 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
2384 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
2387 static bool isPositive(ConstantInt *C) {
2388 return cast<ConstantSInt>(C)->getValue() >= 0;
2391 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
2392 /// overflowed for this type.
2393 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2395 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
2397 if (In1->getType()->isUnsigned())
2398 return cast<ConstantUInt>(Result)->getValue() <
2399 cast<ConstantUInt>(In1)->getValue();
2400 if (isPositive(In1) != isPositive(In2))
2402 if (isPositive(In1))
2403 return cast<ConstantSInt>(Result)->getValue() <
2404 cast<ConstantSInt>(In1)->getValue();
2405 return cast<ConstantSInt>(Result)->getValue() >
2406 cast<ConstantSInt>(In1)->getValue();
2409 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
2410 /// code necessary to compute the offset from the base pointer (without adding
2411 /// in the base pointer). Return the result as a signed integer of intptr size.
2412 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
2413 TargetData &TD = IC.getTargetData();
2414 gep_type_iterator GTI = gep_type_begin(GEP);
2415 const Type *UIntPtrTy = TD.getIntPtrType();
2416 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
2417 Value *Result = Constant::getNullValue(SIntPtrTy);
2419 // Build a mask for high order bits.
2420 uint64_t PtrSizeMask = ~0ULL;
2421 PtrSizeMask >>= 64-(TD.getPointerSize()*8);
2423 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
2424 Value *Op = GEP->getOperand(i);
2425 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
2426 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
2428 if (Constant *OpC = dyn_cast<Constant>(Op)) {
2429 if (!OpC->isNullValue()) {
2430 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
2431 Scale = ConstantExpr::getMul(OpC, Scale);
2432 if (Constant *RC = dyn_cast<Constant>(Result))
2433 Result = ConstantExpr::getAdd(RC, Scale);
2435 // Emit an add instruction.
2436 Result = IC.InsertNewInstBefore(
2437 BinaryOperator::createAdd(Result, Scale,
2438 GEP->getName()+".offs"), I);
2442 // Convert to correct type.
2443 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
2444 Op->getName()+".c"), I);
2446 // We'll let instcombine(mul) convert this to a shl if possible.
2447 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
2448 GEP->getName()+".idx"), I);
2450 // Emit an add instruction.
2451 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
2452 GEP->getName()+".offs"), I);
2458 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
2459 /// else. At this point we know that the GEP is on the LHS of the comparison.
2460 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
2461 Instruction::BinaryOps Cond,
2463 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
2465 if (CastInst *CI = dyn_cast<CastInst>(RHS))
2466 if (isa<PointerType>(CI->getOperand(0)->getType()))
2467 RHS = CI->getOperand(0);
2469 Value *PtrBase = GEPLHS->getOperand(0);
2470 if (PtrBase == RHS) {
2471 // As an optimization, we don't actually have to compute the actual value of
2472 // OFFSET if this is a seteq or setne comparison, just return whether each
2473 // index is zero or not.
2474 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
2475 Instruction *InVal = 0;
2476 gep_type_iterator GTI = gep_type_begin(GEPLHS);
2477 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
2479 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
2480 if (isa<UndefValue>(C)) // undef index -> undef.
2481 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2482 if (C->isNullValue())
2484 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
2485 EmitIt = false; // This is indexing into a zero sized array?
2486 } else if (isa<ConstantInt>(C))
2487 return ReplaceInstUsesWith(I, // No comparison is needed here.
2488 ConstantBool::get(Cond == Instruction::SetNE));
2493 new SetCondInst(Cond, GEPLHS->getOperand(i),
2494 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
2498 InVal = InsertNewInstBefore(InVal, I);
2499 InsertNewInstBefore(Comp, I);
2500 if (Cond == Instruction::SetNE) // True if any are unequal
2501 InVal = BinaryOperator::createOr(InVal, Comp);
2502 else // True if all are equal
2503 InVal = BinaryOperator::createAnd(InVal, Comp);
2511 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
2512 ConstantBool::get(Cond == Instruction::SetEQ));
2515 // Only lower this if the setcc is the only user of the GEP or if we expect
2516 // the result to fold to a constant!
2517 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
2518 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
2519 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
2520 return new SetCondInst(Cond, Offset,
2521 Constant::getNullValue(Offset->getType()));
2523 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
2524 // If the base pointers are different, but the indices are the same, just
2525 // compare the base pointer.
2526 if (PtrBase != GEPRHS->getOperand(0)) {
2527 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
2528 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
2529 GEPRHS->getOperand(0)->getType();
2531 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2532 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2533 IndicesTheSame = false;
2537 // If all indices are the same, just compare the base pointers.
2539 return new SetCondInst(Cond, GEPLHS->getOperand(0),
2540 GEPRHS->getOperand(0));
2542 // Otherwise, the base pointers are different and the indices are
2543 // different, bail out.
2547 // If one of the GEPs has all zero indices, recurse.
2548 bool AllZeros = true;
2549 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2550 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
2551 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
2556 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
2557 SetCondInst::getSwappedCondition(Cond), I);
2559 // If the other GEP has all zero indices, recurse.
2561 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2562 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
2563 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
2568 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
2570 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
2571 // If the GEPs only differ by one index, compare it.
2572 unsigned NumDifferences = 0; // Keep track of # differences.
2573 unsigned DiffOperand = 0; // The operand that differs.
2574 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2575 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2576 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
2577 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
2578 // Irreconcilable differences.
2582 if (NumDifferences++) break;
2587 if (NumDifferences == 0) // SAME GEP?
2588 return ReplaceInstUsesWith(I, // No comparison is needed here.
2589 ConstantBool::get(Cond == Instruction::SetEQ));
2590 else if (NumDifferences == 1) {
2591 Value *LHSV = GEPLHS->getOperand(DiffOperand);
2592 Value *RHSV = GEPRHS->getOperand(DiffOperand);
2594 // Convert the operands to signed values to make sure to perform a
2595 // signed comparison.
2596 const Type *NewTy = LHSV->getType()->getSignedVersion();
2597 if (LHSV->getType() != NewTy)
2598 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
2599 LHSV->getName()), I);
2600 if (RHSV->getType() != NewTy)
2601 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
2602 RHSV->getName()), I);
2603 return new SetCondInst(Cond, LHSV, RHSV);
2607 // Only lower this if the setcc is the only user of the GEP or if we expect
2608 // the result to fold to a constant!
2609 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
2610 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
2611 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
2612 Value *L = EmitGEPOffset(GEPLHS, I, *this);
2613 Value *R = EmitGEPOffset(GEPRHS, I, *this);
2614 return new SetCondInst(Cond, L, R);
2621 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
2622 bool Changed = SimplifyCommutative(I);
2623 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2624 const Type *Ty = Op0->getType();
2628 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
2630 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
2631 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
2633 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
2634 // addresses never equal each other! We already know that Op0 != Op1.
2635 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
2636 isa<ConstantPointerNull>(Op0)) &&
2637 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
2638 isa<ConstantPointerNull>(Op1)))
2639 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
2641 // setcc's with boolean values can always be turned into bitwise operations
2642 if (Ty == Type::BoolTy) {
2643 switch (I.getOpcode()) {
2644 default: assert(0 && "Invalid setcc instruction!");
2645 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
2646 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
2647 InsertNewInstBefore(Xor, I);
2648 return BinaryOperator::createNot(Xor);
2650 case Instruction::SetNE:
2651 return BinaryOperator::createXor(Op0, Op1);
2653 case Instruction::SetGT:
2654 std::swap(Op0, Op1); // Change setgt -> setlt
2656 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
2657 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2658 InsertNewInstBefore(Not, I);
2659 return BinaryOperator::createAnd(Not, Op1);
2661 case Instruction::SetGE:
2662 std::swap(Op0, Op1); // Change setge -> setle
2664 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
2665 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2666 InsertNewInstBefore(Not, I);
2667 return BinaryOperator::createOr(Not, Op1);
2672 // See if we are doing a comparison between a constant and an instruction that
2673 // can be folded into the comparison.
2674 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2675 // Check to see if we are comparing against the minimum or maximum value...
2676 if (CI->isMinValue()) {
2677 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
2678 return ReplaceInstUsesWith(I, ConstantBool::False);
2679 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
2680 return ReplaceInstUsesWith(I, ConstantBool::True);
2681 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
2682 return BinaryOperator::createSetEQ(Op0, Op1);
2683 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
2684 return BinaryOperator::createSetNE(Op0, Op1);
2686 } else if (CI->isMaxValue()) {
2687 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
2688 return ReplaceInstUsesWith(I, ConstantBool::False);
2689 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
2690 return ReplaceInstUsesWith(I, ConstantBool::True);
2691 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
2692 return BinaryOperator::createSetEQ(Op0, Op1);
2693 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
2694 return BinaryOperator::createSetNE(Op0, Op1);
2696 // Comparing against a value really close to min or max?
2697 } else if (isMinValuePlusOne(CI)) {
2698 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
2699 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
2700 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
2701 return BinaryOperator::createSetNE(Op0, SubOne(CI));
2703 } else if (isMaxValueMinusOne(CI)) {
2704 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
2705 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
2706 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
2707 return BinaryOperator::createSetNE(Op0, AddOne(CI));
2710 // If we still have a setle or setge instruction, turn it into the
2711 // appropriate setlt or setgt instruction. Since the border cases have
2712 // already been handled above, this requires little checking.
2714 if (I.getOpcode() == Instruction::SetLE)
2715 return BinaryOperator::createSetLT(Op0, AddOne(CI));
2716 if (I.getOpcode() == Instruction::SetGE)
2717 return BinaryOperator::createSetGT(Op0, SubOne(CI));
2719 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2720 switch (LHSI->getOpcode()) {
2721 case Instruction::And:
2722 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
2723 LHSI->getOperand(0)->hasOneUse()) {
2724 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
2725 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
2726 // happens a LOT in code produced by the C front-end, for bitfield
2728 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
2729 ConstantUInt *ShAmt;
2730 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
2731 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
2732 const Type *Ty = LHSI->getType();
2734 // We can fold this as long as we can't shift unknown bits
2735 // into the mask. This can only happen with signed shift
2736 // rights, as they sign-extend.
2738 bool CanFold = Shift->getOpcode() != Instruction::Shr ||
2739 Shift->getType()->isUnsigned();
2741 // To test for the bad case of the signed shr, see if any
2742 // of the bits shifted in could be tested after the mask.
2743 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
2744 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
2746 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
2748 ConstantExpr::getShl(ConstantInt::getAllOnesValue(Ty), OShAmt);
2749 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
2755 if (Shift->getOpcode() == Instruction::Shl)
2756 NewCst = ConstantExpr::getUShr(CI, ShAmt);
2758 NewCst = ConstantExpr::getShl(CI, ShAmt);
2760 // Check to see if we are shifting out any of the bits being
2762 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
2763 // If we shifted bits out, the fold is not going to work out.
2764 // As a special case, check to see if this means that the
2765 // result is always true or false now.
2766 if (I.getOpcode() == Instruction::SetEQ)
2767 return ReplaceInstUsesWith(I, ConstantBool::False);
2768 if (I.getOpcode() == Instruction::SetNE)
2769 return ReplaceInstUsesWith(I, ConstantBool::True);
2771 I.setOperand(1, NewCst);
2772 Constant *NewAndCST;
2773 if (Shift->getOpcode() == Instruction::Shl)
2774 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
2776 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
2777 LHSI->setOperand(1, NewAndCST);
2778 LHSI->setOperand(0, Shift->getOperand(0));
2779 WorkList.push_back(Shift); // Shift is dead.
2780 AddUsesToWorkList(I);
2788 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
2789 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
2790 switch (I.getOpcode()) {
2792 case Instruction::SetEQ:
2793 case Instruction::SetNE: {
2794 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
2796 // Check that the shift amount is in range. If not, don't perform
2797 // undefined shifts. When the shift is visited it will be
2799 if (ShAmt->getValue() >= TypeBits)
2802 // If we are comparing against bits always shifted out, the
2803 // comparison cannot succeed.
2805 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
2806 if (Comp != CI) {// Comparing against a bit that we know is zero.
2807 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
2808 Constant *Cst = ConstantBool::get(IsSetNE);
2809 return ReplaceInstUsesWith(I, Cst);
2812 if (LHSI->hasOneUse()) {
2813 // Otherwise strength reduce the shift into an and.
2814 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
2815 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
2818 if (CI->getType()->isUnsigned()) {
2819 Mask = ConstantUInt::get(CI->getType(), Val);
2820 } else if (ShAmtVal != 0) {
2821 Mask = ConstantSInt::get(CI->getType(), Val);
2823 Mask = ConstantInt::getAllOnesValue(CI->getType());
2827 BinaryOperator::createAnd(LHSI->getOperand(0),
2828 Mask, LHSI->getName()+".mask");
2829 Value *And = InsertNewInstBefore(AndI, I);
2830 return new SetCondInst(I.getOpcode(), And,
2831 ConstantExpr::getUShr(CI, ShAmt));
2838 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
2839 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
2840 switch (I.getOpcode()) {
2842 case Instruction::SetEQ:
2843 case Instruction::SetNE: {
2845 // Check that the shift amount is in range. If not, don't perform
2846 // undefined shifts. When the shift is visited it will be
2848 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
2849 if (ShAmt->getValue() >= TypeBits)
2852 // If we are comparing against bits always shifted out, the
2853 // comparison cannot succeed.
2855 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
2857 if (Comp != CI) {// Comparing against a bit that we know is zero.
2858 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
2859 Constant *Cst = ConstantBool::get(IsSetNE);
2860 return ReplaceInstUsesWith(I, Cst);
2863 if (LHSI->hasOneUse() || CI->isNullValue()) {
2864 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
2866 // Otherwise strength reduce the shift into an and.
2867 uint64_t Val = ~0ULL; // All ones.
2868 Val <<= ShAmtVal; // Shift over to the right spot.
2871 if (CI->getType()->isUnsigned()) {
2872 Val &= ~0ULL >> (64-TypeBits);
2873 Mask = ConstantUInt::get(CI->getType(), Val);
2875 Mask = ConstantSInt::get(CI->getType(), Val);
2879 BinaryOperator::createAnd(LHSI->getOperand(0),
2880 Mask, LHSI->getName()+".mask");
2881 Value *And = InsertNewInstBefore(AndI, I);
2882 return new SetCondInst(I.getOpcode(), And,
2883 ConstantExpr::getShl(CI, ShAmt));
2891 case Instruction::Div:
2892 // Fold: (div X, C1) op C2 -> range check
2893 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
2894 // Fold this div into the comparison, producing a range check.
2895 // Determine, based on the divide type, what the range is being
2896 // checked. If there is an overflow on the low or high side, remember
2897 // it, otherwise compute the range [low, hi) bounding the new value.
2898 bool LoOverflow = false, HiOverflow = 0;
2899 ConstantInt *LoBound = 0, *HiBound = 0;
2902 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
2904 Instruction::BinaryOps Opcode = I.getOpcode();
2906 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
2907 } else if (LHSI->getType()->isUnsigned()) { // udiv
2909 LoOverflow = ProdOV;
2910 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
2911 } else if (isPositive(DivRHS)) { // Divisor is > 0.
2912 if (CI->isNullValue()) { // (X / pos) op 0
2914 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
2916 } else if (isPositive(CI)) { // (X / pos) op pos
2918 LoOverflow = ProdOV;
2919 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
2920 } else { // (X / pos) op neg
2921 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
2922 LoOverflow = AddWithOverflow(LoBound, Prod,
2923 cast<ConstantInt>(DivRHSH));
2925 HiOverflow = ProdOV;
2927 } else { // Divisor is < 0.
2928 if (CI->isNullValue()) { // (X / neg) op 0
2929 LoBound = AddOne(DivRHS);
2930 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
2931 if (HiBound == DivRHS)
2932 LoBound = 0; // - INTMIN = INTMIN
2933 } else if (isPositive(CI)) { // (X / neg) op pos
2934 HiOverflow = LoOverflow = ProdOV;
2936 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
2937 HiBound = AddOne(Prod);
2938 } else { // (X / neg) op neg
2940 LoOverflow = HiOverflow = ProdOV;
2941 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
2944 // Dividing by a negate swaps the condition.
2945 Opcode = SetCondInst::getSwappedCondition(Opcode);
2949 Value *X = LHSI->getOperand(0);
2951 default: assert(0 && "Unhandled setcc opcode!");
2952 case Instruction::SetEQ:
2953 if (LoOverflow && HiOverflow)
2954 return ReplaceInstUsesWith(I, ConstantBool::False);
2955 else if (HiOverflow)
2956 return new SetCondInst(Instruction::SetGE, X, LoBound);
2957 else if (LoOverflow)
2958 return new SetCondInst(Instruction::SetLT, X, HiBound);
2960 return InsertRangeTest(X, LoBound, HiBound, true, I);
2961 case Instruction::SetNE:
2962 if (LoOverflow && HiOverflow)
2963 return ReplaceInstUsesWith(I, ConstantBool::True);
2964 else if (HiOverflow)
2965 return new SetCondInst(Instruction::SetLT, X, LoBound);
2966 else if (LoOverflow)
2967 return new SetCondInst(Instruction::SetGE, X, HiBound);
2969 return InsertRangeTest(X, LoBound, HiBound, false, I);
2970 case Instruction::SetLT:
2972 return ReplaceInstUsesWith(I, ConstantBool::False);
2973 return new SetCondInst(Instruction::SetLT, X, LoBound);
2974 case Instruction::SetGT:
2976 return ReplaceInstUsesWith(I, ConstantBool::False);
2977 return new SetCondInst(Instruction::SetGE, X, HiBound);
2984 // Simplify seteq and setne instructions...
2985 if (I.getOpcode() == Instruction::SetEQ ||
2986 I.getOpcode() == Instruction::SetNE) {
2987 bool isSetNE = I.getOpcode() == Instruction::SetNE;
2989 // If the first operand is (and|or|xor) with a constant, and the second
2990 // operand is a constant, simplify a bit.
2991 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
2992 switch (BO->getOpcode()) {
2993 case Instruction::Rem:
2994 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2995 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
2997 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
2998 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
2999 if (isPowerOf2_64(V)) {
3000 unsigned L2 = Log2_64(V);
3001 const Type *UTy = BO->getType()->getUnsignedVersion();
3002 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
3004 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
3005 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
3006 RHSCst, BO->getName()), I);
3007 return BinaryOperator::create(I.getOpcode(), NewRem,
3008 Constant::getNullValue(UTy));
3013 case Instruction::Add:
3014 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3015 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3016 if (BO->hasOneUse())
3017 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3018 ConstantExpr::getSub(CI, BOp1C));
3019 } else if (CI->isNullValue()) {
3020 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3021 // efficiently invertible, or if the add has just this one use.
3022 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3024 if (Value *NegVal = dyn_castNegVal(BOp1))
3025 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
3026 else if (Value *NegVal = dyn_castNegVal(BOp0))
3027 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
3028 else if (BO->hasOneUse()) {
3029 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
3031 InsertNewInstBefore(Neg, I);
3032 return new SetCondInst(I.getOpcode(), BOp0, Neg);
3036 case Instruction::Xor:
3037 // For the xor case, we can xor two constants together, eliminating
3038 // the explicit xor.
3039 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
3040 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
3041 ConstantExpr::getXor(CI, BOC));
3044 case Instruction::Sub:
3045 // Replace (([sub|xor] A, B) != 0) with (A != B)
3046 if (CI->isNullValue())
3047 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3051 case Instruction::Or:
3052 // If bits are being or'd in that are not present in the constant we
3053 // are comparing against, then the comparison could never succeed!
3054 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
3055 Constant *NotCI = ConstantExpr::getNot(CI);
3056 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
3057 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3061 case Instruction::And:
3062 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3063 // If bits are being compared against that are and'd out, then the
3064 // comparison can never succeed!
3065 if (!ConstantExpr::getAnd(CI,
3066 ConstantExpr::getNot(BOC))->isNullValue())
3067 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3069 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3070 if (CI == BOC && isOneBitSet(CI))
3071 return new SetCondInst(isSetNE ? Instruction::SetEQ :
3072 Instruction::SetNE, Op0,
3073 Constant::getNullValue(CI->getType()));
3075 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
3076 // to be a signed value as appropriate.
3077 if (isSignBit(BOC)) {
3078 Value *X = BO->getOperand(0);
3079 // If 'X' is not signed, insert a cast now...
3080 if (!BOC->getType()->isSigned()) {
3081 const Type *DestTy = BOC->getType()->getSignedVersion();
3082 X = InsertCastBefore(X, DestTy, I);
3084 return new SetCondInst(isSetNE ? Instruction::SetLT :
3085 Instruction::SetGE, X,
3086 Constant::getNullValue(X->getType()));
3089 // ((X & ~7) == 0) --> X < 8
3090 if (CI->isNullValue() && isHighOnes(BOC)) {
3091 Value *X = BO->getOperand(0);
3092 Constant *NegX = ConstantExpr::getNeg(BOC);
3094 // If 'X' is signed, insert a cast now.
3095 if (NegX->getType()->isSigned()) {
3096 const Type *DestTy = NegX->getType()->getUnsignedVersion();
3097 X = InsertCastBefore(X, DestTy, I);
3098 NegX = ConstantExpr::getCast(NegX, DestTy);
3101 return new SetCondInst(isSetNE ? Instruction::SetGE :
3102 Instruction::SetLT, X, NegX);
3109 } else { // Not a SetEQ/SetNE
3110 // If the LHS is a cast from an integral value of the same size,
3111 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
3112 Value *CastOp = Cast->getOperand(0);
3113 const Type *SrcTy = CastOp->getType();
3114 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
3115 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
3116 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
3117 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
3118 "Source and destination signednesses should differ!");
3119 if (Cast->getType()->isSigned()) {
3120 // If this is a signed comparison, check for comparisons in the
3121 // vicinity of zero.
3122 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
3124 return BinaryOperator::createSetGT(CastOp,
3125 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
3126 else if (I.getOpcode() == Instruction::SetGT &&
3127 cast<ConstantSInt>(CI)->getValue() == -1)
3128 // X > -1 => x < 128
3129 return BinaryOperator::createSetLT(CastOp,
3130 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
3132 ConstantUInt *CUI = cast<ConstantUInt>(CI);
3133 if (I.getOpcode() == Instruction::SetLT &&
3134 CUI->getValue() == 1ULL << (SrcTySize-1))
3135 // X < 128 => X > -1
3136 return BinaryOperator::createSetGT(CastOp,
3137 ConstantSInt::get(SrcTy, -1));
3138 else if (I.getOpcode() == Instruction::SetGT &&
3139 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
3141 return BinaryOperator::createSetLT(CastOp,
3142 Constant::getNullValue(SrcTy));
3149 // Handle setcc with constant RHS's that can be integer, FP or pointer.
3150 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3151 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3152 switch (LHSI->getOpcode()) {
3153 case Instruction::GetElementPtr:
3154 if (RHSC->isNullValue()) {
3155 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
3156 bool isAllZeros = true;
3157 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
3158 if (!isa<Constant>(LHSI->getOperand(i)) ||
3159 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
3164 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
3165 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3169 case Instruction::PHI:
3170 if (Instruction *NV = FoldOpIntoPhi(I))
3173 case Instruction::Select:
3174 // If either operand of the select is a constant, we can fold the
3175 // comparison into the select arms, which will cause one to be
3176 // constant folded and the select turned into a bitwise or.
3177 Value *Op1 = 0, *Op2 = 0;
3178 if (LHSI->hasOneUse()) {
3179 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3180 // Fold the known value into the constant operand.
3181 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3182 // Insert a new SetCC of the other select operand.
3183 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3184 LHSI->getOperand(2), RHSC,
3186 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3187 // Fold the known value into the constant operand.
3188 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3189 // Insert a new SetCC of the other select operand.
3190 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3191 LHSI->getOperand(1), RHSC,
3197 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
3202 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
3203 if (User *GEP = dyn_castGetElementPtr(Op0))
3204 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
3206 if (User *GEP = dyn_castGetElementPtr(Op1))
3207 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
3208 SetCondInst::getSwappedCondition(I.getOpcode()), I))
3211 // Test to see if the operands of the setcc are casted versions of other
3212 // values. If the cast can be stripped off both arguments, we do so now.
3213 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3214 Value *CastOp0 = CI->getOperand(0);
3215 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
3216 (isa<Constant>(Op1) || isa<CastInst>(Op1)) &&
3217 (I.getOpcode() == Instruction::SetEQ ||
3218 I.getOpcode() == Instruction::SetNE)) {
3219 // We keep moving the cast from the left operand over to the right
3220 // operand, where it can often be eliminated completely.
3223 // If operand #1 is a cast instruction, see if we can eliminate it as
3225 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
3226 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
3228 Op1 = CI2->getOperand(0);
3230 // If Op1 is a constant, we can fold the cast into the constant.
3231 if (Op1->getType() != Op0->getType())
3232 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3233 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
3235 // Otherwise, cast the RHS right before the setcc
3236 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
3237 InsertNewInstBefore(cast<Instruction>(Op1), I);
3239 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
3242 // Handle the special case of: setcc (cast bool to X), <cst>
3243 // This comes up when you have code like
3246 // For generality, we handle any zero-extension of any operand comparison
3247 // with a constant or another cast from the same type.
3248 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
3249 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
3252 return Changed ? &I : 0;
3255 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
3256 // We only handle extending casts so far.
3258 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
3259 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
3260 const Type *SrcTy = LHSCIOp->getType();
3261 const Type *DestTy = SCI.getOperand(0)->getType();
3264 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
3267 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
3268 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
3269 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
3271 // Is this a sign or zero extension?
3272 bool isSignSrc = SrcTy->isSigned();
3273 bool isSignDest = DestTy->isSigned();
3275 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
3276 // Not an extension from the same type?
3277 RHSCIOp = CI->getOperand(0);
3278 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
3279 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
3280 // Compute the constant that would happen if we truncated to SrcTy then
3281 // reextended to DestTy.
3282 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
3284 if (ConstantExpr::getCast(Res, DestTy) == CI) {
3287 // If the value cannot be represented in the shorter type, we cannot emit
3288 // a simple comparison.
3289 if (SCI.getOpcode() == Instruction::SetEQ)
3290 return ReplaceInstUsesWith(SCI, ConstantBool::False);
3291 if (SCI.getOpcode() == Instruction::SetNE)
3292 return ReplaceInstUsesWith(SCI, ConstantBool::True);
3294 // Evaluate the comparison for LT.
3296 if (DestTy->isSigned()) {
3297 // We're performing a signed comparison.
3299 // Signed extend and signed comparison.
3300 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
3301 Result = ConstantBool::False;
3303 Result = ConstantBool::True; // X < (large) --> true
3305 // Unsigned extend and signed comparison.
3306 if (cast<ConstantSInt>(CI)->getValue() < 0)
3307 Result = ConstantBool::False;
3309 Result = ConstantBool::True;
3312 // We're performing an unsigned comparison.
3314 // Unsigned extend & compare -> always true.
3315 Result = ConstantBool::True;
3317 // We're performing an unsigned comp with a sign extended value.
3318 // This is true if the input is >= 0. [aka >s -1]
3319 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
3320 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
3321 NegOne, SCI.getName()), SCI);
3325 // Finally, return the value computed.
3326 if (SCI.getOpcode() == Instruction::SetLT) {
3327 return ReplaceInstUsesWith(SCI, Result);
3329 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
3330 if (Constant *CI = dyn_cast<Constant>(Result))
3331 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
3333 return BinaryOperator::createNot(Result);
3340 // Okay, just insert a compare of the reduced operands now!
3341 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
3344 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
3345 assert(I.getOperand(1)->getType() == Type::UByteTy);
3346 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3347 bool isLeftShift = I.getOpcode() == Instruction::Shl;
3349 // shl X, 0 == X and shr X, 0 == X
3350 // shl 0, X == 0 and shr 0, X == 0
3351 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
3352 Op0 == Constant::getNullValue(Op0->getType()))
3353 return ReplaceInstUsesWith(I, Op0);
3355 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
3356 if (!isLeftShift && I.getType()->isSigned())
3357 return ReplaceInstUsesWith(I, Op0);
3358 else // undef << X -> 0 AND undef >>u X -> 0
3359 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3361 if (isa<UndefValue>(Op1)) {
3362 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
3363 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3365 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
3368 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
3370 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
3371 if (CSI->isAllOnesValue())
3372 return ReplaceInstUsesWith(I, CSI);
3374 // Try to fold constant and into select arguments.
3375 if (isa<Constant>(Op0))
3376 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3377 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3380 // See if we can turn a signed shr into an unsigned shr.
3381 if (!isLeftShift && I.getType()->isSigned()) {
3382 if (MaskedValueIsZero(Op0, ConstantInt::getMinValue(I.getType()))) {
3383 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
3384 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
3386 return new CastInst(V, I.getType());
3390 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1)) {
3391 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
3392 // of a signed value.
3394 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
3395 if (CUI->getValue() >= TypeBits) {
3396 if (!Op0->getType()->isSigned() || isLeftShift)
3397 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
3399 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
3404 // ((X*C1) << C2) == (X * (C1 << C2))
3405 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
3406 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
3407 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
3408 return BinaryOperator::createMul(BO->getOperand(0),
3409 ConstantExpr::getShl(BOOp, CUI));
3411 // Try to fold constant and into select arguments.
3412 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3413 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3415 if (isa<PHINode>(Op0))
3416 if (Instruction *NV = FoldOpIntoPhi(I))
3419 if (Op0->hasOneUse()) {
3420 // If this is a SHL of a sign-extending cast, see if we can turn the input
3421 // into a zero extending cast (a simple strength reduction).
3422 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3423 const Type *SrcTy = CI->getOperand(0)->getType();
3424 if (isLeftShift && SrcTy->isInteger() && SrcTy->isSigned() &&
3425 SrcTy->getPrimitiveSizeInBits() <
3426 CI->getType()->getPrimitiveSizeInBits()) {
3427 // We can change it to a zero extension if we are shifting out all of
3428 // the sign extended bits. To check this, form a mask of all of the
3429 // sign extend bits, then shift them left and see if we have anything
3431 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy); // 1111
3432 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType()); // 00001111
3433 Mask = ConstantExpr::getNot(Mask); // 1's in the sign bits: 11110000
3434 if (ConstantExpr::getShl(Mask, CUI)->isNullValue()) {
3435 // If the shift is nuking all of the sign bits, change this to a
3436 // zero extension cast. To do this, cast the cast input to
3437 // unsigned, then to the requested size.
3438 Value *CastOp = CI->getOperand(0);
3440 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
3441 CI->getName()+".uns");
3442 NC = InsertNewInstBefore(NC, I);
3443 // Finally, insert a replacement for CI.
3444 NC = new CastInst(NC, CI->getType(), CI->getName());
3446 NC = InsertNewInstBefore(NC, I);
3447 WorkList.push_back(CI); // Delete CI later.
3448 I.setOperand(0, NC);
3449 return &I; // The SHL operand was modified.
3454 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
3455 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3458 switch (Op0BO->getOpcode()) {
3460 case Instruction::Add:
3461 case Instruction::And:
3462 case Instruction::Or:
3463 case Instruction::Xor:
3464 // These operators commute.
3465 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
3466 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3467 match(Op0BO->getOperand(1),
3468 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == CUI) {
3469 Instruction *YS = new ShiftInst(Instruction::Shl,
3470 Op0BO->getOperand(0), CUI,
3472 InsertNewInstBefore(YS, I); // (Y << C)
3473 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3475 Op0BO->getOperand(1)->getName());
3476 InsertNewInstBefore(X, I); // (X + (Y << C))
3477 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3478 C2 = ConstantExpr::getShl(C2, CUI);
3479 return BinaryOperator::createAnd(X, C2);
3482 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
3483 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3484 match(Op0BO->getOperand(1),
3485 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3486 m_ConstantInt(CC))) && V2 == CUI &&
3487 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
3488 Instruction *YS = new ShiftInst(Instruction::Shl,
3489 Op0BO->getOperand(0), CUI,
3491 InsertNewInstBefore(YS, I); // (Y << C)
3493 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, CUI),
3494 V1->getName()+".mask");
3495 InsertNewInstBefore(XM, I); // X & (CC << C)
3497 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3501 case Instruction::Sub:
3502 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3503 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3504 match(Op0BO->getOperand(0),
3505 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == CUI) {
3506 Instruction *YS = new ShiftInst(Instruction::Shl,
3507 Op0BO->getOperand(1), CUI,
3509 InsertNewInstBefore(YS, I); // (Y << C)
3510 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3512 Op0BO->getOperand(0)->getName());
3513 InsertNewInstBefore(X, I); // (X + (Y << C))
3514 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3515 C2 = ConstantExpr::getShl(C2, CUI);
3516 return BinaryOperator::createAnd(X, C2);
3519 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3520 match(Op0BO->getOperand(0),
3521 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3522 m_ConstantInt(CC))) && V2 == CUI &&
3523 cast<BinaryOperator>(Op0BO->getOperand(0))->getOperand(0)->hasOneUse()) {
3524 Instruction *YS = new ShiftInst(Instruction::Shl,
3525 Op0BO->getOperand(1), CUI,
3527 InsertNewInstBefore(YS, I); // (Y << C)
3529 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, CUI),
3530 V1->getName()+".mask");
3531 InsertNewInstBefore(XM, I); // X & (CC << C)
3533 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3540 // If the operand is an bitwise operator with a constant RHS, and the
3541 // shift is the only use, we can pull it out of the shift.
3542 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
3543 bool isValid = true; // Valid only for And, Or, Xor
3544 bool highBitSet = false; // Transform if high bit of constant set?
3546 switch (Op0BO->getOpcode()) {
3547 default: isValid = false; break; // Do not perform transform!
3548 case Instruction::Add:
3549 isValid = isLeftShift;
3551 case Instruction::Or:
3552 case Instruction::Xor:
3555 case Instruction::And:
3560 // If this is a signed shift right, and the high bit is modified
3561 // by the logical operation, do not perform the transformation.
3562 // The highBitSet boolean indicates the value of the high bit of
3563 // the constant which would cause it to be modified for this
3566 if (isValid && !isLeftShift && !I.getType()->isUnsigned()) {
3567 uint64_t Val = Op0C->getRawValue();
3568 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
3572 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, CUI);
3574 Instruction *NewShift =
3575 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), CUI,
3578 InsertNewInstBefore(NewShift, I);
3580 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
3587 // If this is a shift of a shift, see if we can fold the two together...
3588 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
3589 if (ConstantUInt *ShiftAmt1C =
3590 dyn_cast<ConstantUInt>(Op0SI->getOperand(1))) {
3591 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
3592 unsigned ShiftAmt2 = (unsigned)CUI->getValue();
3594 // Check for (A << c1) << c2 and (A >> c1) >> c2
3595 if (I.getOpcode() == Op0SI->getOpcode()) {
3596 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift...
3597 if (Op0->getType()->getPrimitiveSizeInBits() < Amt)
3598 Amt = Op0->getType()->getPrimitiveSizeInBits();
3599 return new ShiftInst(I.getOpcode(), Op0SI->getOperand(0),
3600 ConstantUInt::get(Type::UByteTy, Amt));
3603 // Check for (A << c1) >> c2 or visaversa. If we are dealing with
3604 // signed types, we can only support the (A >> c1) << c2 configuration,
3605 // because it can not turn an arbitrary bit of A into a sign bit.
3606 if (I.getType()->isUnsigned() || isLeftShift) {
3607 // Calculate bitmask for what gets shifted off the edge...
3608 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
3610 C = ConstantExpr::getShl(C, ShiftAmt1C);
3612 C = ConstantExpr::getShr(C, ShiftAmt1C);
3615 BinaryOperator::createAnd(Op0SI->getOperand(0), C,
3616 Op0SI->getOperand(0)->getName()+".mask");
3617 InsertNewInstBefore(Mask, I);
3619 // Figure out what flavor of shift we should use...
3620 if (ShiftAmt1 == ShiftAmt2)
3621 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
3622 else if (ShiftAmt1 < ShiftAmt2) {
3623 return new ShiftInst(I.getOpcode(), Mask,
3624 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
3626 return new ShiftInst(Op0SI->getOpcode(), Mask,
3627 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3630 // We can handle signed (X << C1) >> C2 if it's a sign extend. In
3631 // this case, C1 == C2 and C1 is 8, 16, or 32.
3632 if (ShiftAmt1 == ShiftAmt2) {
3633 const Type *SExtType = 0;
3634 switch (ShiftAmt1) {
3635 case 8 : SExtType = Type::SByteTy; break;
3636 case 16: SExtType = Type::ShortTy; break;
3637 case 32: SExtType = Type::IntTy; break;
3641 Instruction *NewTrunc = new CastInst(Op0SI->getOperand(0),
3643 InsertNewInstBefore(NewTrunc, I);
3644 return new CastInst(NewTrunc, I.getType());
3661 /// getCastType - In the future, we will split the cast instruction into these
3662 /// various types. Until then, we have to do the analysis here.
3663 static CastType getCastType(const Type *Src, const Type *Dest) {
3664 assert(Src->isIntegral() && Dest->isIntegral() &&
3665 "Only works on integral types!");
3666 unsigned SrcSize = Src->getPrimitiveSizeInBits();
3667 unsigned DestSize = Dest->getPrimitiveSizeInBits();
3669 if (SrcSize == DestSize) return Noop;
3670 if (SrcSize > DestSize) return Truncate;
3671 if (Src->isSigned()) return Signext;
3676 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
3679 static inline bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
3680 const Type *DstTy, TargetData *TD) {
3682 // It is legal to eliminate the instruction if casting A->B->A if the sizes
3683 // are identical and the bits don't get reinterpreted (for example
3684 // int->float->int would not be allowed).
3685 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
3688 // If we are casting between pointer and integer types, treat pointers as
3689 // integers of the appropriate size for the code below.
3690 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
3691 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
3692 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
3694 // Allow free casting and conversion of sizes as long as the sign doesn't
3696 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
3697 CastType FirstCast = getCastType(SrcTy, MidTy);
3698 CastType SecondCast = getCastType(MidTy, DstTy);
3700 // Capture the effect of these two casts. If the result is a legal cast,
3701 // the CastType is stored here, otherwise a special code is used.
3702 static const unsigned CastResult[] = {
3703 // First cast is noop
3705 // First cast is a truncate
3706 1, 1, 4, 4, // trunc->extend is not safe to eliminate
3707 // First cast is a sign ext
3708 2, 5, 2, 4, // signext->zeroext never ok
3709 // First cast is a zero ext
3713 unsigned Result = CastResult[FirstCast*4+SecondCast];
3715 default: assert(0 && "Illegal table value!");
3720 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
3721 // truncates, we could eliminate more casts.
3722 return (unsigned)getCastType(SrcTy, DstTy) == Result;
3724 return false; // Not possible to eliminate this here.
3726 // Sign or zero extend followed by truncate is always ok if the result
3727 // is a truncate or noop.
3728 CastType ResultCast = getCastType(SrcTy, DstTy);
3729 if (ResultCast == Noop || ResultCast == Truncate)
3731 // Otherwise we are still growing the value, we are only safe if the
3732 // result will match the sign/zeroextendness of the result.
3733 return ResultCast == FirstCast;
3739 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
3740 if (V->getType() == Ty || isa<Constant>(V)) return false;
3741 if (const CastInst *CI = dyn_cast<CastInst>(V))
3742 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
3748 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
3749 /// InsertBefore instruction. This is specialized a bit to avoid inserting
3750 /// casts that are known to not do anything...
3752 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
3753 Instruction *InsertBefore) {
3754 if (V->getType() == DestTy) return V;
3755 if (Constant *C = dyn_cast<Constant>(V))
3756 return ConstantExpr::getCast(C, DestTy);
3758 CastInst *CI = new CastInst(V, DestTy, V->getName());
3759 InsertNewInstBefore(CI, *InsertBefore);
3763 // CastInst simplification
3765 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
3766 Value *Src = CI.getOperand(0);
3768 // If the user is casting a value to the same type, eliminate this cast
3770 if (CI.getType() == Src->getType())
3771 return ReplaceInstUsesWith(CI, Src);
3773 if (isa<UndefValue>(Src)) // cast undef -> undef
3774 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
3776 // If casting the result of another cast instruction, try to eliminate this
3779 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
3780 Value *A = CSrc->getOperand(0);
3781 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
3782 CI.getType(), TD)) {
3783 // This instruction now refers directly to the cast's src operand. This
3784 // has a good chance of making CSrc dead.
3785 CI.setOperand(0, CSrc->getOperand(0));
3789 // If this is an A->B->A cast, and we are dealing with integral types, try
3790 // to convert this into a logical 'and' instruction.
3792 if (A->getType()->isInteger() &&
3793 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
3794 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
3795 CSrc->getType()->getPrimitiveSizeInBits() <
3796 CI.getType()->getPrimitiveSizeInBits()&&
3797 A->getType()->getPrimitiveSizeInBits() ==
3798 CI.getType()->getPrimitiveSizeInBits()) {
3799 assert(CSrc->getType() != Type::ULongTy &&
3800 "Cannot have type bigger than ulong!");
3801 uint64_t AndValue = ~0ULL>>(64-CSrc->getType()->getPrimitiveSizeInBits());
3802 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
3804 AndOp = ConstantExpr::getCast(AndOp, A->getType());
3805 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
3806 if (And->getType() != CI.getType()) {
3807 And->setName(CSrc->getName()+".mask");
3808 InsertNewInstBefore(And, CI);
3809 And = new CastInst(And, CI.getType());
3815 // If this is a cast to bool, turn it into the appropriate setne instruction.
3816 if (CI.getType() == Type::BoolTy)
3817 return BinaryOperator::createSetNE(CI.getOperand(0),
3818 Constant::getNullValue(CI.getOperand(0)->getType()));
3820 // If casting the result of a getelementptr instruction with no offset, turn
3821 // this into a cast of the original pointer!
3823 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
3824 bool AllZeroOperands = true;
3825 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
3826 if (!isa<Constant>(GEP->getOperand(i)) ||
3827 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
3828 AllZeroOperands = false;
3831 if (AllZeroOperands) {
3832 CI.setOperand(0, GEP->getOperand(0));
3837 // If we are casting a malloc or alloca to a pointer to a type of the same
3838 // size, rewrite the allocation instruction to allocate the "right" type.
3840 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
3841 if (AI->hasOneUse() && !AI->isArrayAllocation())
3842 if (const PointerType *PTy = dyn_cast<PointerType>(CI.getType())) {
3843 // Get the type really allocated and the type casted to...
3844 const Type *AllocElTy = AI->getAllocatedType();
3845 const Type *CastElTy = PTy->getElementType();
3846 if (AllocElTy->isSized() && CastElTy->isSized()) {
3847 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
3848 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
3850 // If the allocation is for an even multiple of the cast type size
3851 if (CastElTySize && (AllocElTySize % CastElTySize == 0)) {
3852 Value *Amt = ConstantUInt::get(Type::UIntTy,
3853 AllocElTySize/CastElTySize);
3854 std::string Name = AI->getName(); AI->setName("");
3855 AllocationInst *New;
3856 if (isa<MallocInst>(AI))
3857 New = new MallocInst(CastElTy, Amt, Name);
3859 New = new AllocaInst(CastElTy, Amt, Name);
3860 InsertNewInstBefore(New, *AI);
3861 return ReplaceInstUsesWith(CI, New);
3866 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
3867 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
3869 if (isa<PHINode>(Src))
3870 if (Instruction *NV = FoldOpIntoPhi(CI))
3873 // If the source value is an instruction with only this use, we can attempt to
3874 // propagate the cast into the instruction. Also, only handle integral types
3876 if (Instruction *SrcI = dyn_cast<Instruction>(Src))
3877 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
3878 CI.getType()->isInteger()) { // Don't mess with casts to bool here
3879 const Type *DestTy = CI.getType();
3880 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
3881 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
3883 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
3884 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
3886 switch (SrcI->getOpcode()) {
3887 case Instruction::Add:
3888 case Instruction::Mul:
3889 case Instruction::And:
3890 case Instruction::Or:
3891 case Instruction::Xor:
3892 // If we are discarding information, or just changing the sign, rewrite.
3893 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
3894 // Don't insert two casts if they cannot be eliminated. We allow two
3895 // casts to be inserted if the sizes are the same. This could only be
3896 // converting signedness, which is a noop.
3897 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
3898 !ValueRequiresCast(Op0, DestTy, TD)) {
3899 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
3900 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
3901 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
3902 ->getOpcode(), Op0c, Op1c);
3906 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
3907 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
3908 Op1 == ConstantBool::True &&
3909 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
3910 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
3911 return BinaryOperator::createXor(New,
3912 ConstantInt::get(CI.getType(), 1));
3915 case Instruction::Shl:
3916 // Allow changing the sign of the source operand. Do not allow changing
3917 // the size of the shift, UNLESS the shift amount is a constant. We
3918 // mush not change variable sized shifts to a smaller size, because it
3919 // is undefined to shift more bits out than exist in the value.
3920 if (DestBitSize == SrcBitSize ||
3921 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
3922 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
3923 return new ShiftInst(Instruction::Shl, Op0c, Op1);
3926 case Instruction::Shr:
3927 // If this is a signed shr, and if all bits shifted in are about to be
3928 // truncated off, turn it into an unsigned shr to allow greater
3930 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
3931 isa<ConstantInt>(Op1)) {
3932 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
3933 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
3934 // Convert to unsigned.
3935 Value *N1 = InsertOperandCastBefore(Op0,
3936 Op0->getType()->getUnsignedVersion(), &CI);
3937 // Insert the new shift, which is now unsigned.
3938 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
3939 Op1, Src->getName()), CI);
3940 return new CastInst(N1, CI.getType());
3945 case Instruction::SetNE:
3946 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
3947 if (Op1C->getRawValue() == 0) {
3948 // If the input only has the low bit set, simplify directly.
3950 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
3951 // cast (X != 0) to int --> X if X&~1 == 0
3952 if (MaskedValueIsZero(Op0, cast<ConstantIntegral>(Not1))) {
3953 if (CI.getType() == Op0->getType())
3954 return ReplaceInstUsesWith(CI, Op0);
3956 return new CastInst(Op0, CI.getType());
3959 // If the input is an and with a single bit, shift then simplify.
3960 ConstantInt *AndRHS;
3961 if (match(Op0, m_And(m_Value(), m_ConstantInt(AndRHS))))
3962 if (AndRHS->getRawValue() &&
3963 (AndRHS->getRawValue() & (AndRHS->getRawValue()-1)) == 0) {
3964 unsigned ShiftAmt = Log2_64(AndRHS->getRawValue());
3965 // Perform an unsigned shr by shiftamt. Convert input to
3966 // unsigned if it is signed.
3968 if (In->getType()->isSigned())
3969 In = InsertNewInstBefore(new CastInst(In,
3970 In->getType()->getUnsignedVersion(), In->getName()),CI);
3971 // Insert the shift to put the result in the low bit.
3972 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
3973 ConstantInt::get(Type::UByteTy, ShiftAmt),
3974 In->getName()+".lobit"), CI);
3975 if (CI.getType() == In->getType())
3976 return ReplaceInstUsesWith(CI, In);
3978 return new CastInst(In, CI.getType());
3983 case Instruction::SetEQ:
3984 // We if we are just checking for a seteq of a single bit and casting it
3985 // to an integer. If so, shift the bit to the appropriate place then
3986 // cast to integer to avoid the comparison.
3987 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
3988 // Is Op1C a power of two or zero?
3989 if ((Op1C->getRawValue() & Op1C->getRawValue()-1) == 0) {
3990 // cast (X == 1) to int -> X iff X has only the low bit set.
3991 if (Op1C->getRawValue() == 1) {
3993 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
3994 if (MaskedValueIsZero(Op0, cast<ConstantIntegral>(Not1))) {
3995 if (CI.getType() == Op0->getType())
3996 return ReplaceInstUsesWith(CI, Op0);
3998 return new CastInst(Op0, CI.getType());
4009 /// GetSelectFoldableOperands - We want to turn code that looks like this:
4011 /// %D = select %cond, %C, %A
4013 /// %C = select %cond, %B, 0
4016 /// Assuming that the specified instruction is an operand to the select, return
4017 /// a bitmask indicating which operands of this instruction are foldable if they
4018 /// equal the other incoming value of the select.
4020 static unsigned GetSelectFoldableOperands(Instruction *I) {
4021 switch (I->getOpcode()) {
4022 case Instruction::Add:
4023 case Instruction::Mul:
4024 case Instruction::And:
4025 case Instruction::Or:
4026 case Instruction::Xor:
4027 return 3; // Can fold through either operand.
4028 case Instruction::Sub: // Can only fold on the amount subtracted.
4029 case Instruction::Shl: // Can only fold on the shift amount.
4030 case Instruction::Shr:
4033 return 0; // Cannot fold
4037 /// GetSelectFoldableConstant - For the same transformation as the previous
4038 /// function, return the identity constant that goes into the select.
4039 static Constant *GetSelectFoldableConstant(Instruction *I) {
4040 switch (I->getOpcode()) {
4041 default: assert(0 && "This cannot happen!"); abort();
4042 case Instruction::Add:
4043 case Instruction::Sub:
4044 case Instruction::Or:
4045 case Instruction::Xor:
4046 return Constant::getNullValue(I->getType());
4047 case Instruction::Shl:
4048 case Instruction::Shr:
4049 return Constant::getNullValue(Type::UByteTy);
4050 case Instruction::And:
4051 return ConstantInt::getAllOnesValue(I->getType());
4052 case Instruction::Mul:
4053 return ConstantInt::get(I->getType(), 1);
4057 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
4058 /// have the same opcode and only one use each. Try to simplify this.
4059 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
4061 if (TI->getNumOperands() == 1) {
4062 // If this is a non-volatile load or a cast from the same type,
4064 if (TI->getOpcode() == Instruction::Cast) {
4065 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
4068 return 0; // unknown unary op.
4071 // Fold this by inserting a select from the input values.
4072 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
4073 FI->getOperand(0), SI.getName()+".v");
4074 InsertNewInstBefore(NewSI, SI);
4075 return new CastInst(NewSI, TI->getType());
4078 // Only handle binary operators here.
4079 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
4082 // Figure out if the operations have any operands in common.
4083 Value *MatchOp, *OtherOpT, *OtherOpF;
4085 if (TI->getOperand(0) == FI->getOperand(0)) {
4086 MatchOp = TI->getOperand(0);
4087 OtherOpT = TI->getOperand(1);
4088 OtherOpF = FI->getOperand(1);
4089 MatchIsOpZero = true;
4090 } else if (TI->getOperand(1) == FI->getOperand(1)) {
4091 MatchOp = TI->getOperand(1);
4092 OtherOpT = TI->getOperand(0);
4093 OtherOpF = FI->getOperand(0);
4094 MatchIsOpZero = false;
4095 } else if (!TI->isCommutative()) {
4097 } else if (TI->getOperand(0) == FI->getOperand(1)) {
4098 MatchOp = TI->getOperand(0);
4099 OtherOpT = TI->getOperand(1);
4100 OtherOpF = FI->getOperand(0);
4101 MatchIsOpZero = true;
4102 } else if (TI->getOperand(1) == FI->getOperand(0)) {
4103 MatchOp = TI->getOperand(1);
4104 OtherOpT = TI->getOperand(0);
4105 OtherOpF = FI->getOperand(1);
4106 MatchIsOpZero = true;
4111 // If we reach here, they do have operations in common.
4112 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
4113 OtherOpF, SI.getName()+".v");
4114 InsertNewInstBefore(NewSI, SI);
4116 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
4118 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
4120 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
4123 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
4125 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
4129 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
4130 Value *CondVal = SI.getCondition();
4131 Value *TrueVal = SI.getTrueValue();
4132 Value *FalseVal = SI.getFalseValue();
4134 // select true, X, Y -> X
4135 // select false, X, Y -> Y
4136 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
4137 if (C == ConstantBool::True)
4138 return ReplaceInstUsesWith(SI, TrueVal);
4140 assert(C == ConstantBool::False);
4141 return ReplaceInstUsesWith(SI, FalseVal);
4144 // select C, X, X -> X
4145 if (TrueVal == FalseVal)
4146 return ReplaceInstUsesWith(SI, TrueVal);
4148 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
4149 return ReplaceInstUsesWith(SI, FalseVal);
4150 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
4151 return ReplaceInstUsesWith(SI, TrueVal);
4152 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
4153 if (isa<Constant>(TrueVal))
4154 return ReplaceInstUsesWith(SI, TrueVal);
4156 return ReplaceInstUsesWith(SI, FalseVal);
4159 if (SI.getType() == Type::BoolTy)
4160 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
4161 if (C == ConstantBool::True) {
4162 // Change: A = select B, true, C --> A = or B, C
4163 return BinaryOperator::createOr(CondVal, FalseVal);
4165 // Change: A = select B, false, C --> A = and !B, C
4167 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4168 "not."+CondVal->getName()), SI);
4169 return BinaryOperator::createAnd(NotCond, FalseVal);
4171 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
4172 if (C == ConstantBool::False) {
4173 // Change: A = select B, C, false --> A = and B, C
4174 return BinaryOperator::createAnd(CondVal, TrueVal);
4176 // Change: A = select B, C, true --> A = or !B, C
4178 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4179 "not."+CondVal->getName()), SI);
4180 return BinaryOperator::createOr(NotCond, TrueVal);
4184 // Selecting between two integer constants?
4185 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
4186 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
4187 // select C, 1, 0 -> cast C to int
4188 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
4189 return new CastInst(CondVal, SI.getType());
4190 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
4191 // select C, 0, 1 -> cast !C to int
4193 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4194 "not."+CondVal->getName()), SI);
4195 return new CastInst(NotCond, SI.getType());
4198 // If one of the constants is zero (we know they can't both be) and we
4199 // have a setcc instruction with zero, and we have an 'and' with the
4200 // non-constant value, eliminate this whole mess. This corresponds to
4201 // cases like this: ((X & 27) ? 27 : 0)
4202 if (TrueValC->isNullValue() || FalseValC->isNullValue())
4203 if (Instruction *IC = dyn_cast<Instruction>(SI.getCondition()))
4204 if ((IC->getOpcode() == Instruction::SetEQ ||
4205 IC->getOpcode() == Instruction::SetNE) &&
4206 isa<ConstantInt>(IC->getOperand(1)) &&
4207 cast<Constant>(IC->getOperand(1))->isNullValue())
4208 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
4209 if (ICA->getOpcode() == Instruction::And &&
4210 isa<ConstantInt>(ICA->getOperand(1)) &&
4211 (ICA->getOperand(1) == TrueValC ||
4212 ICA->getOperand(1) == FalseValC) &&
4213 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
4214 // Okay, now we know that everything is set up, we just don't
4215 // know whether we have a setne or seteq and whether the true or
4216 // false val is the zero.
4217 bool ShouldNotVal = !TrueValC->isNullValue();
4218 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
4221 V = InsertNewInstBefore(BinaryOperator::create(
4222 Instruction::Xor, V, ICA->getOperand(1)), SI);
4223 return ReplaceInstUsesWith(SI, V);
4227 // See if we are selecting two values based on a comparison of the two values.
4228 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
4229 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
4230 // Transform (X == Y) ? X : Y -> Y
4231 if (SCI->getOpcode() == Instruction::SetEQ)
4232 return ReplaceInstUsesWith(SI, FalseVal);
4233 // Transform (X != Y) ? X : Y -> X
4234 if (SCI->getOpcode() == Instruction::SetNE)
4235 return ReplaceInstUsesWith(SI, TrueVal);
4236 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4238 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
4239 // Transform (X == Y) ? Y : X -> X
4240 if (SCI->getOpcode() == Instruction::SetEQ)
4241 return ReplaceInstUsesWith(SI, FalseVal);
4242 // Transform (X != Y) ? Y : X -> Y
4243 if (SCI->getOpcode() == Instruction::SetNE)
4244 return ReplaceInstUsesWith(SI, TrueVal);
4245 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4249 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
4250 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
4251 if (TI->hasOneUse() && FI->hasOneUse()) {
4252 bool isInverse = false;
4253 Instruction *AddOp = 0, *SubOp = 0;
4255 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
4256 if (TI->getOpcode() == FI->getOpcode())
4257 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
4260 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
4261 // even legal for FP.
4262 if (TI->getOpcode() == Instruction::Sub &&
4263 FI->getOpcode() == Instruction::Add) {
4264 AddOp = FI; SubOp = TI;
4265 } else if (FI->getOpcode() == Instruction::Sub &&
4266 TI->getOpcode() == Instruction::Add) {
4267 AddOp = TI; SubOp = FI;
4271 Value *OtherAddOp = 0;
4272 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
4273 OtherAddOp = AddOp->getOperand(1);
4274 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
4275 OtherAddOp = AddOp->getOperand(0);
4279 // So at this point we know we have:
4280 // select C, (add X, Y), (sub X, ?)
4281 // We can do the transform profitably if either 'Y' = '?' or '?' is
4283 if (SubOp->getOperand(1) == AddOp ||
4284 isa<Constant>(SubOp->getOperand(1))) {
4286 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
4287 NegVal = ConstantExpr::getNeg(C);
4289 NegVal = InsertNewInstBefore(
4290 BinaryOperator::createNeg(SubOp->getOperand(1)), SI);
4293 Value *NewTrueOp = OtherAddOp;
4294 Value *NewFalseOp = NegVal;
4296 std::swap(NewTrueOp, NewFalseOp);
4297 Instruction *NewSel =
4298 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
4300 NewSel = InsertNewInstBefore(NewSel, SI);
4301 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
4307 // See if we can fold the select into one of our operands.
4308 if (SI.getType()->isInteger()) {
4309 // See the comment above GetSelectFoldableOperands for a description of the
4310 // transformation we are doing here.
4311 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
4312 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
4313 !isa<Constant>(FalseVal))
4314 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
4315 unsigned OpToFold = 0;
4316 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
4318 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
4323 Constant *C = GetSelectFoldableConstant(TVI);
4324 std::string Name = TVI->getName(); TVI->setName("");
4325 Instruction *NewSel =
4326 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
4328 InsertNewInstBefore(NewSel, SI);
4329 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
4330 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
4331 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
4332 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
4334 assert(0 && "Unknown instruction!!");
4339 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
4340 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
4341 !isa<Constant>(TrueVal))
4342 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
4343 unsigned OpToFold = 0;
4344 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
4346 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
4351 Constant *C = GetSelectFoldableConstant(FVI);
4352 std::string Name = FVI->getName(); FVI->setName("");
4353 Instruction *NewSel =
4354 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
4356 InsertNewInstBefore(NewSel, SI);
4357 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
4358 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
4359 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
4360 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
4362 assert(0 && "Unknown instruction!!");
4368 if (BinaryOperator::isNot(CondVal)) {
4369 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
4370 SI.setOperand(1, FalseVal);
4371 SI.setOperand(2, TrueVal);
4379 // CallInst simplification
4381 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
4382 // Intrinsics cannot occur in an invoke, so handle them here instead of in
4384 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&CI)) {
4385 bool Changed = false;
4387 // memmove/cpy/set of zero bytes is a noop.
4388 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
4389 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
4391 // FIXME: Increase alignment here.
4393 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
4394 if (CI->getRawValue() == 1) {
4395 // Replace the instruction with just byte operations. We would
4396 // transform other cases to loads/stores, but we don't know if
4397 // alignment is sufficient.
4401 // If we have a memmove and the source operation is a constant global,
4402 // then the source and dest pointers can't alias, so we can change this
4403 // into a call to memcpy.
4404 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI))
4405 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
4406 if (GVSrc->isConstant()) {
4407 Module *M = CI.getParent()->getParent()->getParent();
4408 Function *MemCpy = M->getOrInsertFunction("llvm.memcpy",
4409 CI.getCalledFunction()->getFunctionType());
4410 CI.setOperand(0, MemCpy);
4414 if (Changed) return &CI;
4415 } else if (DbgStopPointInst *SPI = dyn_cast<DbgStopPointInst>(&CI)) {
4416 // If this stoppoint is at the same source location as the previous
4417 // stoppoint in the chain, it is not needed.
4418 if (DbgStopPointInst *PrevSPI =
4419 dyn_cast<DbgStopPointInst>(SPI->getChain()))
4420 if (SPI->getLineNo() == PrevSPI->getLineNo() &&
4421 SPI->getColNo() == PrevSPI->getColNo()) {
4422 SPI->replaceAllUsesWith(PrevSPI);
4423 return EraseInstFromFunction(CI);
4427 return visitCallSite(&CI);
4430 // InvokeInst simplification
4432 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
4433 return visitCallSite(&II);
4436 // visitCallSite - Improvements for call and invoke instructions.
4438 Instruction *InstCombiner::visitCallSite(CallSite CS) {
4439 bool Changed = false;
4441 // If the callee is a constexpr cast of a function, attempt to move the cast
4442 // to the arguments of the call/invoke.
4443 if (transformConstExprCastCall(CS)) return 0;
4445 Value *Callee = CS.getCalledValue();
4447 if (Function *CalleeF = dyn_cast<Function>(Callee))
4448 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
4449 Instruction *OldCall = CS.getInstruction();
4450 // If the call and callee calling conventions don't match, this call must
4451 // be unreachable, as the call is undefined.
4452 new StoreInst(ConstantBool::True,
4453 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
4454 if (!OldCall->use_empty())
4455 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
4456 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
4457 return EraseInstFromFunction(*OldCall);
4461 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
4462 // This instruction is not reachable, just remove it. We insert a store to
4463 // undef so that we know that this code is not reachable, despite the fact
4464 // that we can't modify the CFG here.
4465 new StoreInst(ConstantBool::True,
4466 UndefValue::get(PointerType::get(Type::BoolTy)),
4467 CS.getInstruction());
4469 if (!CS.getInstruction()->use_empty())
4470 CS.getInstruction()->
4471 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
4473 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
4474 // Don't break the CFG, insert a dummy cond branch.
4475 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
4476 ConstantBool::True, II);
4478 return EraseInstFromFunction(*CS.getInstruction());
4481 const PointerType *PTy = cast<PointerType>(Callee->getType());
4482 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4483 if (FTy->isVarArg()) {
4484 // See if we can optimize any arguments passed through the varargs area of
4486 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
4487 E = CS.arg_end(); I != E; ++I)
4488 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
4489 // If this cast does not effect the value passed through the varargs
4490 // area, we can eliminate the use of the cast.
4491 Value *Op = CI->getOperand(0);
4492 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
4499 return Changed ? CS.getInstruction() : 0;
4502 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
4503 // attempt to move the cast to the arguments of the call/invoke.
4505 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
4506 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
4507 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
4508 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
4510 Function *Callee = cast<Function>(CE->getOperand(0));
4511 Instruction *Caller = CS.getInstruction();
4513 // Okay, this is a cast from a function to a different type. Unless doing so
4514 // would cause a type conversion of one of our arguments, change this call to
4515 // be a direct call with arguments casted to the appropriate types.
4517 const FunctionType *FT = Callee->getFunctionType();
4518 const Type *OldRetTy = Caller->getType();
4520 // Check to see if we are changing the return type...
4521 if (OldRetTy != FT->getReturnType()) {
4522 if (Callee->isExternal() &&
4523 !OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) &&
4524 !Caller->use_empty())
4525 return false; // Cannot transform this return value...
4527 // If the callsite is an invoke instruction, and the return value is used by
4528 // a PHI node in a successor, we cannot change the return type of the call
4529 // because there is no place to put the cast instruction (without breaking
4530 // the critical edge). Bail out in this case.
4531 if (!Caller->use_empty())
4532 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
4533 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
4535 if (PHINode *PN = dyn_cast<PHINode>(*UI))
4536 if (PN->getParent() == II->getNormalDest() ||
4537 PN->getParent() == II->getUnwindDest())
4541 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
4542 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
4544 CallSite::arg_iterator AI = CS.arg_begin();
4545 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
4546 const Type *ParamTy = FT->getParamType(i);
4547 bool isConvertible = (*AI)->getType()->isLosslesslyConvertibleTo(ParamTy);
4548 if (Callee->isExternal() && !isConvertible) return false;
4551 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
4552 Callee->isExternal())
4553 return false; // Do not delete arguments unless we have a function body...
4555 // Okay, we decided that this is a safe thing to do: go ahead and start
4556 // inserting cast instructions as necessary...
4557 std::vector<Value*> Args;
4558 Args.reserve(NumActualArgs);
4560 AI = CS.arg_begin();
4561 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
4562 const Type *ParamTy = FT->getParamType(i);
4563 if ((*AI)->getType() == ParamTy) {
4564 Args.push_back(*AI);
4566 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
4571 // If the function takes more arguments than the call was taking, add them
4573 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
4574 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
4576 // If we are removing arguments to the function, emit an obnoxious warning...
4577 if (FT->getNumParams() < NumActualArgs)
4578 if (!FT->isVarArg()) {
4579 std::cerr << "WARNING: While resolving call to function '"
4580 << Callee->getName() << "' arguments were dropped!\n";
4582 // Add all of the arguments in their promoted form to the arg list...
4583 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
4584 const Type *PTy = getPromotedType((*AI)->getType());
4585 if (PTy != (*AI)->getType()) {
4586 // Must promote to pass through va_arg area!
4587 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
4588 InsertNewInstBefore(Cast, *Caller);
4589 Args.push_back(Cast);
4591 Args.push_back(*AI);
4596 if (FT->getReturnType() == Type::VoidTy)
4597 Caller->setName(""); // Void type should not have a name...
4600 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4601 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
4602 Args, Caller->getName(), Caller);
4603 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
4605 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
4606 if (cast<CallInst>(Caller)->isTailCall())
4607 cast<CallInst>(NC)->setTailCall();
4608 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
4611 // Insert a cast of the return type as necessary...
4613 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
4614 if (NV->getType() != Type::VoidTy) {
4615 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
4617 // If this is an invoke instruction, we should insert it after the first
4618 // non-phi, instruction in the normal successor block.
4619 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4620 BasicBlock::iterator I = II->getNormalDest()->begin();
4621 while (isa<PHINode>(I)) ++I;
4622 InsertNewInstBefore(NC, *I);
4624 // Otherwise, it's a call, just insert cast right after the call instr
4625 InsertNewInstBefore(NC, *Caller);
4627 AddUsersToWorkList(*Caller);
4629 NV = UndefValue::get(Caller->getType());
4633 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
4634 Caller->replaceAllUsesWith(NV);
4635 Caller->getParent()->getInstList().erase(Caller);
4636 removeFromWorkList(Caller);
4641 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
4642 // operator and they all are only used by the PHI, PHI together their
4643 // inputs, and do the operation once, to the result of the PHI.
4644 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
4645 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
4647 // Scan the instruction, looking for input operations that can be folded away.
4648 // If all input operands to the phi are the same instruction (e.g. a cast from
4649 // the same type or "+42") we can pull the operation through the PHI, reducing
4650 // code size and simplifying code.
4651 Constant *ConstantOp = 0;
4652 const Type *CastSrcTy = 0;
4653 if (isa<CastInst>(FirstInst)) {
4654 CastSrcTy = FirstInst->getOperand(0)->getType();
4655 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
4656 // Can fold binop or shift if the RHS is a constant.
4657 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
4658 if (ConstantOp == 0) return 0;
4660 return 0; // Cannot fold this operation.
4663 // Check to see if all arguments are the same operation.
4664 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
4665 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
4666 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
4667 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
4670 if (I->getOperand(0)->getType() != CastSrcTy)
4671 return 0; // Cast operation must match.
4672 } else if (I->getOperand(1) != ConstantOp) {
4677 // Okay, they are all the same operation. Create a new PHI node of the
4678 // correct type, and PHI together all of the LHS's of the instructions.
4679 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
4680 PN.getName()+".in");
4681 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
4683 Value *InVal = FirstInst->getOperand(0);
4684 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
4686 // Add all operands to the new PHI.
4687 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
4688 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
4689 if (NewInVal != InVal)
4691 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
4696 // The new PHI unions all of the same values together. This is really
4697 // common, so we handle it intelligently here for compile-time speed.
4701 InsertNewInstBefore(NewPN, PN);
4705 // Insert and return the new operation.
4706 if (isa<CastInst>(FirstInst))
4707 return new CastInst(PhiVal, PN.getType());
4708 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
4709 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
4711 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
4712 PhiVal, ConstantOp);
4715 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
4717 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
4718 if (PN->use_empty()) return true;
4719 if (!PN->hasOneUse()) return false;
4721 // Remember this node, and if we find the cycle, return.
4722 if (!PotentiallyDeadPHIs.insert(PN).second)
4725 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
4726 return DeadPHICycle(PU, PotentiallyDeadPHIs);
4731 // PHINode simplification
4733 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
4734 if (Value *V = PN.hasConstantValue())
4735 return ReplaceInstUsesWith(PN, V);
4737 // If the only user of this instruction is a cast instruction, and all of the
4738 // incoming values are constants, change this PHI to merge together the casted
4741 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
4742 if (CI->getType() != PN.getType()) { // noop casts will be folded
4743 bool AllConstant = true;
4744 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
4745 if (!isa<Constant>(PN.getIncomingValue(i))) {
4746 AllConstant = false;
4750 // Make a new PHI with all casted values.
4751 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
4752 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
4753 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
4754 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
4755 PN.getIncomingBlock(i));
4758 // Update the cast instruction.
4759 CI->setOperand(0, New);
4760 WorkList.push_back(CI); // revisit the cast instruction to fold.
4761 WorkList.push_back(New); // Make sure to revisit the new Phi
4762 return &PN; // PN is now dead!
4766 // If all PHI operands are the same operation, pull them through the PHI,
4767 // reducing code size.
4768 if (isa<Instruction>(PN.getIncomingValue(0)) &&
4769 PN.getIncomingValue(0)->hasOneUse())
4770 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
4773 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
4774 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
4775 // PHI)... break the cycle.
4777 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
4778 std::set<PHINode*> PotentiallyDeadPHIs;
4779 PotentiallyDeadPHIs.insert(&PN);
4780 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
4781 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
4787 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
4788 Instruction *InsertPoint,
4790 unsigned PS = IC->getTargetData().getPointerSize();
4791 const Type *VTy = V->getType();
4792 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
4793 // We must insert a cast to ensure we sign-extend.
4794 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
4795 V->getName()), *InsertPoint);
4796 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
4801 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
4802 Value *PtrOp = GEP.getOperand(0);
4803 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
4804 // If so, eliminate the noop.
4805 if (GEP.getNumOperands() == 1)
4806 return ReplaceInstUsesWith(GEP, PtrOp);
4808 if (isa<UndefValue>(GEP.getOperand(0)))
4809 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
4811 bool HasZeroPointerIndex = false;
4812 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
4813 HasZeroPointerIndex = C->isNullValue();
4815 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
4816 return ReplaceInstUsesWith(GEP, PtrOp);
4818 // Eliminate unneeded casts for indices.
4819 bool MadeChange = false;
4820 gep_type_iterator GTI = gep_type_begin(GEP);
4821 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
4822 if (isa<SequentialType>(*GTI)) {
4823 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
4824 Value *Src = CI->getOperand(0);
4825 const Type *SrcTy = Src->getType();
4826 const Type *DestTy = CI->getType();
4827 if (Src->getType()->isInteger()) {
4828 if (SrcTy->getPrimitiveSizeInBits() ==
4829 DestTy->getPrimitiveSizeInBits()) {
4830 // We can always eliminate a cast from ulong or long to the other.
4831 // We can always eliminate a cast from uint to int or the other on
4832 // 32-bit pointer platforms.
4833 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
4835 GEP.setOperand(i, Src);
4837 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
4838 SrcTy->getPrimitiveSize() == 4) {
4839 // We can always eliminate a cast from int to [u]long. We can
4840 // eliminate a cast from uint to [u]long iff the target is a 32-bit
4842 if (SrcTy->isSigned() ||
4843 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
4845 GEP.setOperand(i, Src);
4850 // If we are using a wider index than needed for this platform, shrink it
4851 // to what we need. If the incoming value needs a cast instruction,
4852 // insert it. This explicit cast can make subsequent optimizations more
4854 Value *Op = GEP.getOperand(i);
4855 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
4856 if (Constant *C = dyn_cast<Constant>(Op)) {
4857 GEP.setOperand(i, ConstantExpr::getCast(C,
4858 TD->getIntPtrType()->getSignedVersion()));
4861 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
4862 Op->getName()), GEP);
4863 GEP.setOperand(i, Op);
4867 // If this is a constant idx, make sure to canonicalize it to be a signed
4868 // operand, otherwise CSE and other optimizations are pessimized.
4869 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
4870 GEP.setOperand(i, ConstantExpr::getCast(CUI,
4871 CUI->getType()->getSignedVersion()));
4875 if (MadeChange) return &GEP;
4877 // Combine Indices - If the source pointer to this getelementptr instruction
4878 // is a getelementptr instruction, combine the indices of the two
4879 // getelementptr instructions into a single instruction.
4881 std::vector<Value*> SrcGEPOperands;
4882 if (User *Src = dyn_castGetElementPtr(PtrOp))
4883 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
4885 if (!SrcGEPOperands.empty()) {
4886 // Note that if our source is a gep chain itself that we wait for that
4887 // chain to be resolved before we perform this transformation. This
4888 // avoids us creating a TON of code in some cases.
4890 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
4891 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
4892 return 0; // Wait until our source is folded to completion.
4894 std::vector<Value *> Indices;
4896 // Find out whether the last index in the source GEP is a sequential idx.
4897 bool EndsWithSequential = false;
4898 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
4899 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
4900 EndsWithSequential = !isa<StructType>(*I);
4902 // Can we combine the two pointer arithmetics offsets?
4903 if (EndsWithSequential) {
4904 // Replace: gep (gep %P, long B), long A, ...
4905 // With: T = long A+B; gep %P, T, ...
4907 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
4908 if (SO1 == Constant::getNullValue(SO1->getType())) {
4910 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
4913 // If they aren't the same type, convert both to an integer of the
4914 // target's pointer size.
4915 if (SO1->getType() != GO1->getType()) {
4916 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
4917 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
4918 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
4919 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
4921 unsigned PS = TD->getPointerSize();
4922 if (SO1->getType()->getPrimitiveSize() == PS) {
4923 // Convert GO1 to SO1's type.
4924 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
4926 } else if (GO1->getType()->getPrimitiveSize() == PS) {
4927 // Convert SO1 to GO1's type.
4928 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
4930 const Type *PT = TD->getIntPtrType();
4931 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
4932 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
4936 if (isa<Constant>(SO1) && isa<Constant>(GO1))
4937 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
4939 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
4940 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
4944 // Recycle the GEP we already have if possible.
4945 if (SrcGEPOperands.size() == 2) {
4946 GEP.setOperand(0, SrcGEPOperands[0]);
4947 GEP.setOperand(1, Sum);
4950 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
4951 SrcGEPOperands.end()-1);
4952 Indices.push_back(Sum);
4953 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
4955 } else if (isa<Constant>(*GEP.idx_begin()) &&
4956 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
4957 SrcGEPOperands.size() != 1) {
4958 // Otherwise we can do the fold if the first index of the GEP is a zero
4959 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
4960 SrcGEPOperands.end());
4961 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
4964 if (!Indices.empty())
4965 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
4967 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
4968 // GEP of global variable. If all of the indices for this GEP are
4969 // constants, we can promote this to a constexpr instead of an instruction.
4971 // Scan for nonconstants...
4972 std::vector<Constant*> Indices;
4973 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
4974 for (; I != E && isa<Constant>(*I); ++I)
4975 Indices.push_back(cast<Constant>(*I));
4977 if (I == E) { // If they are all constants...
4978 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
4980 // Replace all uses of the GEP with the new constexpr...
4981 return ReplaceInstUsesWith(GEP, CE);
4983 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
4984 if (!isa<PointerType>(X->getType())) {
4985 // Not interesting. Source pointer must be a cast from pointer.
4986 } else if (HasZeroPointerIndex) {
4987 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
4988 // into : GEP [10 x ubyte]* X, long 0, ...
4990 // This occurs when the program declares an array extern like "int X[];"
4992 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
4993 const PointerType *XTy = cast<PointerType>(X->getType());
4994 if (const ArrayType *XATy =
4995 dyn_cast<ArrayType>(XTy->getElementType()))
4996 if (const ArrayType *CATy =
4997 dyn_cast<ArrayType>(CPTy->getElementType()))
4998 if (CATy->getElementType() == XATy->getElementType()) {
4999 // At this point, we know that the cast source type is a pointer
5000 // to an array of the same type as the destination pointer
5001 // array. Because the array type is never stepped over (there
5002 // is a leading zero) we can fold the cast into this GEP.
5003 GEP.setOperand(0, X);
5006 } else if (GEP.getNumOperands() == 2) {
5007 // Transform things like:
5008 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
5009 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
5010 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
5011 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
5012 if (isa<ArrayType>(SrcElTy) &&
5013 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
5014 TD->getTypeSize(ResElTy)) {
5015 Value *V = InsertNewInstBefore(
5016 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5017 GEP.getOperand(1), GEP.getName()), GEP);
5018 return new CastInst(V, GEP.getType());
5021 // Transform things like:
5022 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
5023 // (where tmp = 8*tmp2) into:
5024 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
5026 if (isa<ArrayType>(SrcElTy) &&
5027 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
5028 uint64_t ArrayEltSize =
5029 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
5031 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
5032 // allow either a mul, shift, or constant here.
5034 ConstantInt *Scale = 0;
5035 if (ArrayEltSize == 1) {
5036 NewIdx = GEP.getOperand(1);
5037 Scale = ConstantInt::get(NewIdx->getType(), 1);
5038 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
5039 NewIdx = ConstantInt::get(CI->getType(), 1);
5041 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
5042 if (Inst->getOpcode() == Instruction::Shl &&
5043 isa<ConstantInt>(Inst->getOperand(1))) {
5044 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
5045 if (Inst->getType()->isSigned())
5046 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
5048 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
5049 NewIdx = Inst->getOperand(0);
5050 } else if (Inst->getOpcode() == Instruction::Mul &&
5051 isa<ConstantInt>(Inst->getOperand(1))) {
5052 Scale = cast<ConstantInt>(Inst->getOperand(1));
5053 NewIdx = Inst->getOperand(0);
5057 // If the index will be to exactly the right offset with the scale taken
5058 // out, perform the transformation.
5059 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
5060 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
5061 Scale = ConstantSInt::get(C->getType(),
5062 (int64_t)C->getRawValue() /
5063 (int64_t)ArrayEltSize);
5065 Scale = ConstantUInt::get(Scale->getType(),
5066 Scale->getRawValue() / ArrayEltSize);
5067 if (Scale->getRawValue() != 1) {
5068 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
5069 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
5070 NewIdx = InsertNewInstBefore(Sc, GEP);
5073 // Insert the new GEP instruction.
5075 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5076 NewIdx, GEP.getName());
5077 Idx = InsertNewInstBefore(Idx, GEP);
5078 return new CastInst(Idx, GEP.getType());
5087 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
5088 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
5089 if (AI.isArrayAllocation()) // Check C != 1
5090 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
5091 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
5092 AllocationInst *New = 0;
5094 // Create and insert the replacement instruction...
5095 if (isa<MallocInst>(AI))
5096 New = new MallocInst(NewTy, 0, AI.getName());
5098 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
5099 New = new AllocaInst(NewTy, 0, AI.getName());
5102 InsertNewInstBefore(New, AI);
5104 // Scan to the end of the allocation instructions, to skip over a block of
5105 // allocas if possible...
5107 BasicBlock::iterator It = New;
5108 while (isa<AllocationInst>(*It)) ++It;
5110 // Now that I is pointing to the first non-allocation-inst in the block,
5111 // insert our getelementptr instruction...
5113 Value *NullIdx = Constant::getNullValue(Type::IntTy);
5114 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
5115 New->getName()+".sub", It);
5117 // Now make everything use the getelementptr instead of the original
5119 return ReplaceInstUsesWith(AI, V);
5120 } else if (isa<UndefValue>(AI.getArraySize())) {
5121 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5124 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
5125 // Note that we only do this for alloca's, because malloc should allocate and
5126 // return a unique pointer, even for a zero byte allocation.
5127 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
5128 TD->getTypeSize(AI.getAllocatedType()) == 0)
5129 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5134 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
5135 Value *Op = FI.getOperand(0);
5137 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
5138 if (CastInst *CI = dyn_cast<CastInst>(Op))
5139 if (isa<PointerType>(CI->getOperand(0)->getType())) {
5140 FI.setOperand(0, CI->getOperand(0));
5144 // free undef -> unreachable.
5145 if (isa<UndefValue>(Op)) {
5146 // Insert a new store to null because we cannot modify the CFG here.
5147 new StoreInst(ConstantBool::True,
5148 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
5149 return EraseInstFromFunction(FI);
5152 // If we have 'free null' delete the instruction. This can happen in stl code
5153 // when lots of inlining happens.
5154 if (isa<ConstantPointerNull>(Op))
5155 return EraseInstFromFunction(FI);
5161 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
5162 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
5163 User *CI = cast<User>(LI.getOperand(0));
5164 Value *CastOp = CI->getOperand(0);
5166 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5167 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5168 const Type *SrcPTy = SrcTy->getElementType();
5170 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5171 // If the source is an array, the code below will not succeed. Check to
5172 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5174 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5175 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5176 if (ASrcTy->getNumElements() != 0) {
5177 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5178 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5179 SrcTy = cast<PointerType>(CastOp->getType());
5180 SrcPTy = SrcTy->getElementType();
5183 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5184 // Do not allow turning this into a load of an integer, which is then
5185 // casted to a pointer, this pessimizes pointer analysis a lot.
5186 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
5187 IC.getTargetData().getTypeSize(SrcPTy) ==
5188 IC.getTargetData().getTypeSize(DestPTy)) {
5190 // Okay, we are casting from one integer or pointer type to another of
5191 // the same size. Instead of casting the pointer before the load, cast
5192 // the result of the loaded value.
5193 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
5195 LI.isVolatile()),LI);
5196 // Now cast the result of the load.
5197 return new CastInst(NewLoad, LI.getType());
5204 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
5205 /// from this value cannot trap. If it is not obviously safe to load from the
5206 /// specified pointer, we do a quick local scan of the basic block containing
5207 /// ScanFrom, to determine if the address is already accessed.
5208 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
5209 // If it is an alloca or global variable, it is always safe to load from.
5210 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
5212 // Otherwise, be a little bit agressive by scanning the local block where we
5213 // want to check to see if the pointer is already being loaded or stored
5214 // from/to. If so, the previous load or store would have already trapped,
5215 // so there is no harm doing an extra load (also, CSE will later eliminate
5216 // the load entirely).
5217 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
5222 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
5223 if (LI->getOperand(0) == V) return true;
5224 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5225 if (SI->getOperand(1) == V) return true;
5231 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
5232 Value *Op = LI.getOperand(0);
5234 // load (cast X) --> cast (load X) iff safe
5235 if (CastInst *CI = dyn_cast<CastInst>(Op))
5236 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5239 // None of the following transforms are legal for volatile loads.
5240 if (LI.isVolatile()) return 0;
5242 if (&LI.getParent()->front() != &LI) {
5243 BasicBlock::iterator BBI = &LI; --BBI;
5244 // If the instruction immediately before this is a store to the same
5245 // address, do a simple form of store->load forwarding.
5246 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5247 if (SI->getOperand(1) == LI.getOperand(0))
5248 return ReplaceInstUsesWith(LI, SI->getOperand(0));
5249 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
5250 if (LIB->getOperand(0) == LI.getOperand(0))
5251 return ReplaceInstUsesWith(LI, LIB);
5254 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
5255 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
5256 isa<UndefValue>(GEPI->getOperand(0))) {
5257 // Insert a new store to null instruction before the load to indicate
5258 // that this code is not reachable. We do this instead of inserting
5259 // an unreachable instruction directly because we cannot modify the
5261 new StoreInst(UndefValue::get(LI.getType()),
5262 Constant::getNullValue(Op->getType()), &LI);
5263 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5266 if (Constant *C = dyn_cast<Constant>(Op)) {
5267 // load null/undef -> undef
5268 if ((C->isNullValue() || isa<UndefValue>(C))) {
5269 // Insert a new store to null instruction before the load to indicate that
5270 // this code is not reachable. We do this instead of inserting an
5271 // unreachable instruction directly because we cannot modify the CFG.
5272 new StoreInst(UndefValue::get(LI.getType()),
5273 Constant::getNullValue(Op->getType()), &LI);
5274 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5277 // Instcombine load (constant global) into the value loaded.
5278 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
5279 if (GV->isConstant() && !GV->isExternal())
5280 return ReplaceInstUsesWith(LI, GV->getInitializer());
5282 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
5283 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
5284 if (CE->getOpcode() == Instruction::GetElementPtr) {
5285 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
5286 if (GV->isConstant() && !GV->isExternal())
5288 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
5289 return ReplaceInstUsesWith(LI, V);
5290 if (CE->getOperand(0)->isNullValue()) {
5291 // Insert a new store to null instruction before the load to indicate
5292 // that this code is not reachable. We do this instead of inserting
5293 // an unreachable instruction directly because we cannot modify the
5295 new StoreInst(UndefValue::get(LI.getType()),
5296 Constant::getNullValue(Op->getType()), &LI);
5297 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5300 } else if (CE->getOpcode() == Instruction::Cast) {
5301 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5306 if (Op->hasOneUse()) {
5307 // Change select and PHI nodes to select values instead of addresses: this
5308 // helps alias analysis out a lot, allows many others simplifications, and
5309 // exposes redundancy in the code.
5311 // Note that we cannot do the transformation unless we know that the
5312 // introduced loads cannot trap! Something like this is valid as long as
5313 // the condition is always false: load (select bool %C, int* null, int* %G),
5314 // but it would not be valid if we transformed it to load from null
5317 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
5318 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
5319 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
5320 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
5321 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
5322 SI->getOperand(1)->getName()+".val"), LI);
5323 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
5324 SI->getOperand(2)->getName()+".val"), LI);
5325 return new SelectInst(SI->getCondition(), V1, V2);
5328 // load (select (cond, null, P)) -> load P
5329 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
5330 if (C->isNullValue()) {
5331 LI.setOperand(0, SI->getOperand(2));
5335 // load (select (cond, P, null)) -> load P
5336 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
5337 if (C->isNullValue()) {
5338 LI.setOperand(0, SI->getOperand(1));
5342 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
5343 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
5344 bool Safe = PN->getParent() == LI.getParent();
5346 // Scan all of the instructions between the PHI and the load to make
5347 // sure there are no instructions that might possibly alter the value
5348 // loaded from the PHI.
5350 BasicBlock::iterator I = &LI;
5351 for (--I; !isa<PHINode>(I); --I)
5352 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
5358 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
5359 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
5360 PN->getIncomingBlock(i)->getTerminator()))
5365 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
5366 InsertNewInstBefore(NewPN, *PN);
5367 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
5369 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
5370 BasicBlock *BB = PN->getIncomingBlock(i);
5371 Value *&TheLoad = LoadMap[BB];
5373 Value *InVal = PN->getIncomingValue(i);
5374 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
5375 InVal->getName()+".val"),
5376 *BB->getTerminator());
5378 NewPN->addIncoming(TheLoad, BB);
5380 return ReplaceInstUsesWith(LI, NewPN);
5387 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
5389 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
5390 User *CI = cast<User>(SI.getOperand(1));
5391 Value *CastOp = CI->getOperand(0);
5393 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5394 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5395 const Type *SrcPTy = SrcTy->getElementType();
5397 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5398 // If the source is an array, the code below will not succeed. Check to
5399 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5401 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5402 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5403 if (ASrcTy->getNumElements() != 0) {
5404 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5405 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5406 SrcTy = cast<PointerType>(CastOp->getType());
5407 SrcPTy = SrcTy->getElementType();
5410 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5411 IC.getTargetData().getTypeSize(SrcPTy) ==
5412 IC.getTargetData().getTypeSize(DestPTy)) {
5414 // Okay, we are casting from one integer or pointer type to another of
5415 // the same size. Instead of casting the pointer before the store, cast
5416 // the value to be stored.
5418 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
5419 NewCast = ConstantExpr::getCast(C, SrcPTy);
5421 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
5423 SI.getOperand(0)->getName()+".c"), SI);
5425 return new StoreInst(NewCast, CastOp);
5432 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
5433 Value *Val = SI.getOperand(0);
5434 Value *Ptr = SI.getOperand(1);
5436 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
5437 removeFromWorkList(&SI);
5438 SI.eraseFromParent();
5443 if (SI.isVolatile()) return 0; // Don't hack volatile loads.
5445 // store X, null -> turns into 'unreachable' in SimplifyCFG
5446 if (isa<ConstantPointerNull>(Ptr)) {
5447 if (!isa<UndefValue>(Val)) {
5448 SI.setOperand(0, UndefValue::get(Val->getType()));
5449 if (Instruction *U = dyn_cast<Instruction>(Val))
5450 WorkList.push_back(U); // Dropped a use.
5453 return 0; // Do not modify these!
5456 // store undef, Ptr -> noop
5457 if (isa<UndefValue>(Val)) {
5458 removeFromWorkList(&SI);
5459 SI.eraseFromParent();
5464 // If the pointer destination is a cast, see if we can fold the cast into the
5466 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
5467 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5469 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
5470 if (CE->getOpcode() == Instruction::Cast)
5471 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5475 // If this store is the last instruction in the basic block, and if the block
5476 // ends with an unconditional branch, try to move it to the successor block.
5477 BasicBlock::iterator BBI = &SI; ++BBI;
5478 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
5479 if (BI->isUnconditional()) {
5480 // Check to see if the successor block has exactly two incoming edges. If
5481 // so, see if the other predecessor contains a store to the same location.
5482 // if so, insert a PHI node (if needed) and move the stores down.
5483 BasicBlock *Dest = BI->getSuccessor(0);
5485 pred_iterator PI = pred_begin(Dest);
5486 BasicBlock *Other = 0;
5487 if (*PI != BI->getParent())
5490 if (PI != pred_end(Dest)) {
5491 if (*PI != BI->getParent())
5496 if (++PI != pred_end(Dest))
5499 if (Other) { // If only one other pred...
5500 BBI = Other->getTerminator();
5501 // Make sure this other block ends in an unconditional branch and that
5502 // there is an instruction before the branch.
5503 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
5504 BBI != Other->begin()) {
5506 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
5508 // If this instruction is a store to the same location.
5509 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
5510 // Okay, we know we can perform this transformation. Insert a PHI
5511 // node now if we need it.
5512 Value *MergedVal = OtherStore->getOperand(0);
5513 if (MergedVal != SI.getOperand(0)) {
5514 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
5515 PN->reserveOperandSpace(2);
5516 PN->addIncoming(SI.getOperand(0), SI.getParent());
5517 PN->addIncoming(OtherStore->getOperand(0), Other);
5518 MergedVal = InsertNewInstBefore(PN, Dest->front());
5521 // Advance to a place where it is safe to insert the new store and
5523 BBI = Dest->begin();
5524 while (isa<PHINode>(BBI)) ++BBI;
5525 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
5526 OtherStore->isVolatile()), *BBI);
5528 // Nuke the old stores.
5529 removeFromWorkList(&SI);
5530 removeFromWorkList(OtherStore);
5531 SI.eraseFromParent();
5532 OtherStore->eraseFromParent();
5544 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
5545 // Change br (not X), label True, label False to: br X, label False, True
5547 BasicBlock *TrueDest;
5548 BasicBlock *FalseDest;
5549 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
5550 !isa<Constant>(X)) {
5551 // Swap Destinations and condition...
5553 BI.setSuccessor(0, FalseDest);
5554 BI.setSuccessor(1, TrueDest);
5558 // Cannonicalize setne -> seteq
5559 Instruction::BinaryOps Op; Value *Y;
5560 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
5561 TrueDest, FalseDest)))
5562 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
5563 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
5564 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
5565 std::string Name = I->getName(); I->setName("");
5566 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
5567 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
5568 // Swap Destinations and condition...
5569 BI.setCondition(NewSCC);
5570 BI.setSuccessor(0, FalseDest);
5571 BI.setSuccessor(1, TrueDest);
5572 removeFromWorkList(I);
5573 I->getParent()->getInstList().erase(I);
5574 WorkList.push_back(cast<Instruction>(NewSCC));
5581 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
5582 Value *Cond = SI.getCondition();
5583 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
5584 if (I->getOpcode() == Instruction::Add)
5585 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
5586 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
5587 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
5588 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
5590 SI.setOperand(0, I->getOperand(0));
5591 WorkList.push_back(I);
5599 void InstCombiner::removeFromWorkList(Instruction *I) {
5600 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
5605 /// TryToSinkInstruction - Try to move the specified instruction from its
5606 /// current block into the beginning of DestBlock, which can only happen if it's
5607 /// safe to move the instruction past all of the instructions between it and the
5608 /// end of its block.
5609 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
5610 assert(I->hasOneUse() && "Invariants didn't hold!");
5612 // Cannot move control-flow-involving instructions.
5613 if (isa<PHINode>(I) || isa<InvokeInst>(I) || isa<CallInst>(I)) return false;
5615 // Do not sink alloca instructions out of the entry block.
5616 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
5619 // We can only sink load instructions if there is nothing between the load and
5620 // the end of block that could change the value.
5621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
5622 if (LI->isVolatile()) return false; // Don't sink volatile loads.
5624 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
5626 if (Scan->mayWriteToMemory())
5630 BasicBlock::iterator InsertPos = DestBlock->begin();
5631 while (isa<PHINode>(InsertPos)) ++InsertPos;
5633 I->moveBefore(InsertPos);
5638 bool InstCombiner::runOnFunction(Function &F) {
5639 bool Changed = false;
5640 TD = &getAnalysis<TargetData>();
5643 // Populate the worklist with the reachable instructions.
5644 std::set<BasicBlock*> Visited;
5645 for (df_ext_iterator<BasicBlock*> BB = df_ext_begin(&F.front(), Visited),
5646 E = df_ext_end(&F.front(), Visited); BB != E; ++BB)
5647 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
5648 WorkList.push_back(I);
5650 // Do a quick scan over the function. If we find any blocks that are
5651 // unreachable, remove any instructions inside of them. This prevents
5652 // the instcombine code from having to deal with some bad special cases.
5653 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
5654 if (!Visited.count(BB)) {
5655 Instruction *Term = BB->getTerminator();
5656 while (Term != BB->begin()) { // Remove instrs bottom-up
5657 BasicBlock::iterator I = Term; --I;
5659 DEBUG(std::cerr << "IC: DCE: " << *I);
5662 if (!I->use_empty())
5663 I->replaceAllUsesWith(UndefValue::get(I->getType()));
5664 I->eraseFromParent();
5669 while (!WorkList.empty()) {
5670 Instruction *I = WorkList.back(); // Get an instruction from the worklist
5671 WorkList.pop_back();
5673 // Check to see if we can DCE or ConstantPropagate the instruction...
5674 // Check to see if we can DIE the instruction...
5675 if (isInstructionTriviallyDead(I)) {
5676 // Add operands to the worklist...
5677 if (I->getNumOperands() < 4)
5678 AddUsesToWorkList(*I);
5681 DEBUG(std::cerr << "IC: DCE: " << *I);
5683 I->eraseFromParent();
5684 removeFromWorkList(I);
5688 // Instruction isn't dead, see if we can constant propagate it...
5689 if (Constant *C = ConstantFoldInstruction(I)) {
5690 Value* Ptr = I->getOperand(0);
5691 if (isa<GetElementPtrInst>(I) &&
5692 cast<Constant>(Ptr)->isNullValue() &&
5693 !isa<ConstantPointerNull>(C) &&
5694 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
5695 // If this is a constant expr gep that is effectively computing an
5696 // "offsetof", fold it into 'cast int X to T*' instead of 'gep 0, 0, 12'
5697 bool isFoldableGEP = true;
5698 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
5699 if (!isa<ConstantInt>(I->getOperand(i)))
5700 isFoldableGEP = false;
5701 if (isFoldableGEP) {
5702 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
5703 std::vector<Value*>(I->op_begin()+1, I->op_end()));
5704 C = ConstantUInt::get(Type::ULongTy, Offset);
5705 C = ConstantExpr::getCast(C, TD->getIntPtrType());
5706 C = ConstantExpr::getCast(C, I->getType());
5710 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
5712 // Add operands to the worklist...
5713 AddUsesToWorkList(*I);
5714 ReplaceInstUsesWith(*I, C);
5717 I->getParent()->getInstList().erase(I);
5718 removeFromWorkList(I);
5722 // See if we can trivially sink this instruction to a successor basic block.
5723 if (I->hasOneUse()) {
5724 BasicBlock *BB = I->getParent();
5725 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
5726 if (UserParent != BB) {
5727 bool UserIsSuccessor = false;
5728 // See if the user is one of our successors.
5729 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
5730 if (*SI == UserParent) {
5731 UserIsSuccessor = true;
5735 // If the user is one of our immediate successors, and if that successor
5736 // only has us as a predecessors (we'd have to split the critical edge
5737 // otherwise), we can keep going.
5738 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
5739 next(pred_begin(UserParent)) == pred_end(UserParent))
5740 // Okay, the CFG is simple enough, try to sink this instruction.
5741 Changed |= TryToSinkInstruction(I, UserParent);
5745 // Now that we have an instruction, try combining it to simplify it...
5746 if (Instruction *Result = visit(*I)) {
5748 // Should we replace the old instruction with a new one?
5750 DEBUG(std::cerr << "IC: Old = " << *I
5751 << " New = " << *Result);
5753 // Everything uses the new instruction now.
5754 I->replaceAllUsesWith(Result);
5756 // Push the new instruction and any users onto the worklist.
5757 WorkList.push_back(Result);
5758 AddUsersToWorkList(*Result);
5760 // Move the name to the new instruction first...
5761 std::string OldName = I->getName(); I->setName("");
5762 Result->setName(OldName);
5764 // Insert the new instruction into the basic block...
5765 BasicBlock *InstParent = I->getParent();
5766 BasicBlock::iterator InsertPos = I;
5768 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
5769 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
5772 InstParent->getInstList().insert(InsertPos, Result);
5774 // Make sure that we reprocess all operands now that we reduced their
5776 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
5777 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
5778 WorkList.push_back(OpI);
5780 // Instructions can end up on the worklist more than once. Make sure
5781 // we do not process an instruction that has been deleted.
5782 removeFromWorkList(I);
5784 // Erase the old instruction.
5785 InstParent->getInstList().erase(I);
5787 DEBUG(std::cerr << "IC: MOD = " << *I);
5789 // If the instruction was modified, it's possible that it is now dead.
5790 // if so, remove it.
5791 if (isInstructionTriviallyDead(I)) {
5792 // Make sure we process all operands now that we are reducing their
5794 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
5795 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
5796 WorkList.push_back(OpI);
5798 // Instructions may end up in the worklist more than once. Erase all
5799 // occurrances of this instruction.
5800 removeFromWorkList(I);
5801 I->eraseFromParent();
5803 WorkList.push_back(Result);
5804 AddUsersToWorkList(*Result);
5814 FunctionPass *llvm::createInstructionCombiningPass() {
5815 return new InstCombiner();