1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. SetCC instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All SetCC instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CallSite.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/PatternMatch.h"
51 #include "llvm/ADT/DepthFirstIterator.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/STLExtras.h"
57 using namespace llvm::PatternMatch;
60 Statistic<> NumCombined ("instcombine", "Number of insts combined");
61 Statistic<> NumConstProp("instcombine", "Number of constant folds");
62 Statistic<> NumDeadInst ("instcombine", "Number of dead inst eliminated");
63 Statistic<> NumDeadStore("instcombine", "Number of dead stores eliminated");
64 Statistic<> NumSunkInst ("instcombine", "Number of instructions sunk");
66 class InstCombiner : public FunctionPass,
67 public InstVisitor<InstCombiner, Instruction*> {
68 // Worklist of all of the instructions that need to be simplified.
69 std::vector<Instruction*> WorkList;
72 /// AddUsersToWorkList - When an instruction is simplified, add all users of
73 /// the instruction to the work lists because they might get more simplified
76 void AddUsersToWorkList(Value &I) {
77 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
79 WorkList.push_back(cast<Instruction>(*UI));
82 /// AddUsesToWorkList - When an instruction is simplified, add operands to
83 /// the work lists because they might get more simplified now.
85 void AddUsesToWorkList(Instruction &I) {
86 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
87 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
88 WorkList.push_back(Op);
91 // removeFromWorkList - remove all instances of I from the worklist.
92 void removeFromWorkList(Instruction *I);
94 virtual bool runOnFunction(Function &F);
96 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
97 AU.addRequired<TargetData>();
101 TargetData &getTargetData() const { return *TD; }
103 // Visitation implementation - Implement instruction combining for different
104 // instruction types. The semantics are as follows:
106 // null - No change was made
107 // I - Change was made, I is still valid, I may be dead though
108 // otherwise - Change was made, replace I with returned instruction
110 Instruction *visitAdd(BinaryOperator &I);
111 Instruction *visitSub(BinaryOperator &I);
112 Instruction *visitMul(BinaryOperator &I);
113 Instruction *visitDiv(BinaryOperator &I);
114 Instruction *visitRem(BinaryOperator &I);
115 Instruction *visitAnd(BinaryOperator &I);
116 Instruction *visitOr (BinaryOperator &I);
117 Instruction *visitXor(BinaryOperator &I);
118 Instruction *visitSetCondInst(SetCondInst &I);
119 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
121 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
122 Instruction::BinaryOps Cond, Instruction &I);
123 Instruction *visitShiftInst(ShiftInst &I);
124 Instruction *FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
126 Instruction *visitCastInst(CastInst &CI);
127 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
129 Instruction *visitSelectInst(SelectInst &CI);
130 Instruction *visitCallInst(CallInst &CI);
131 Instruction *visitInvokeInst(InvokeInst &II);
132 Instruction *visitPHINode(PHINode &PN);
133 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
134 Instruction *visitAllocationInst(AllocationInst &AI);
135 Instruction *visitFreeInst(FreeInst &FI);
136 Instruction *visitLoadInst(LoadInst &LI);
137 Instruction *visitStoreInst(StoreInst &SI);
138 Instruction *visitBranchInst(BranchInst &BI);
139 Instruction *visitSwitchInst(SwitchInst &SI);
140 Instruction *visitExtractElementInst(ExtractElementInst &EI);
142 // visitInstruction - Specify what to return for unhandled instructions...
143 Instruction *visitInstruction(Instruction &I) { return 0; }
146 Instruction *visitCallSite(CallSite CS);
147 bool transformConstExprCastCall(CallSite CS);
150 // InsertNewInstBefore - insert an instruction New before instruction Old
151 // in the program. Add the new instruction to the worklist.
153 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
154 assert(New && New->getParent() == 0 &&
155 "New instruction already inserted into a basic block!");
156 BasicBlock *BB = Old.getParent();
157 BB->getInstList().insert(&Old, New); // Insert inst
158 WorkList.push_back(New); // Add to worklist
162 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
163 /// This also adds the cast to the worklist. Finally, this returns the
165 Value *InsertCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
166 if (V->getType() == Ty) return V;
168 Instruction *C = new CastInst(V, Ty, V->getName(), &Pos);
169 WorkList.push_back(C);
173 // ReplaceInstUsesWith - This method is to be used when an instruction is
174 // found to be dead, replacable with another preexisting expression. Here
175 // we add all uses of I to the worklist, replace all uses of I with the new
176 // value, then return I, so that the inst combiner will know that I was
179 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
180 AddUsersToWorkList(I); // Add all modified instrs to worklist
182 I.replaceAllUsesWith(V);
185 // If we are replacing the instruction with itself, this must be in a
186 // segment of unreachable code, so just clobber the instruction.
187 I.replaceAllUsesWith(UndefValue::get(I.getType()));
192 // UpdateValueUsesWith - This method is to be used when an value is
193 // found to be replacable with another preexisting expression or was
194 // updated. Here we add all uses of I to the worklist, replace all uses of
195 // I with the new value (unless the instruction was just updated), then
196 // return true, so that the inst combiner will know that I was modified.
198 bool UpdateValueUsesWith(Value *Old, Value *New) {
199 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
201 Old->replaceAllUsesWith(New);
202 if (Instruction *I = dyn_cast<Instruction>(Old))
203 WorkList.push_back(I);
207 // EraseInstFromFunction - When dealing with an instruction that has side
208 // effects or produces a void value, we can't rely on DCE to delete the
209 // instruction. Instead, visit methods should return the value returned by
211 Instruction *EraseInstFromFunction(Instruction &I) {
212 assert(I.use_empty() && "Cannot erase instruction that is used!");
213 AddUsesToWorkList(I);
214 removeFromWorkList(&I);
216 return 0; // Don't do anything with FI
220 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
221 /// InsertBefore instruction. This is specialized a bit to avoid inserting
222 /// casts that are known to not do anything...
224 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
225 Instruction *InsertBefore);
227 // SimplifyCommutative - This performs a few simplifications for commutative
229 bool SimplifyCommutative(BinaryOperator &I);
231 bool SimplifyDemandedBits(Value *V, uint64_t Mask, unsigned Depth = 0);
233 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
234 // PHI node as operand #0, see if we can fold the instruction into the PHI
235 // (which is only possible if all operands to the PHI are constants).
236 Instruction *FoldOpIntoPhi(Instruction &I);
238 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
239 // operator and they all are only used by the PHI, PHI together their
240 // inputs, and do the operation once, to the result of the PHI.
241 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
243 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
244 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
246 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
247 bool isSub, Instruction &I);
248 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
249 bool Inside, Instruction &IB);
250 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
253 RegisterOpt<InstCombiner> X("instcombine", "Combine redundant instructions");
256 // getComplexity: Assign a complexity or rank value to LLVM Values...
257 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
258 static unsigned getComplexity(Value *V) {
259 if (isa<Instruction>(V)) {
260 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
264 if (isa<Argument>(V)) return 3;
265 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
268 // isOnlyUse - Return true if this instruction will be deleted if we stop using
270 static bool isOnlyUse(Value *V) {
271 return V->hasOneUse() || isa<Constant>(V);
274 // getPromotedType - Return the specified type promoted as it would be to pass
275 // though a va_arg area...
276 static const Type *getPromotedType(const Type *Ty) {
277 switch (Ty->getTypeID()) {
278 case Type::SByteTyID:
279 case Type::ShortTyID: return Type::IntTy;
280 case Type::UByteTyID:
281 case Type::UShortTyID: return Type::UIntTy;
282 case Type::FloatTyID: return Type::DoubleTy;
287 /// isCast - If the specified operand is a CastInst or a constant expr cast,
288 /// return the operand value, otherwise return null.
289 static Value *isCast(Value *V) {
290 if (CastInst *I = dyn_cast<CastInst>(V))
291 return I->getOperand(0);
292 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
293 if (CE->getOpcode() == Instruction::Cast)
294 return CE->getOperand(0);
298 // SimplifyCommutative - This performs a few simplifications for commutative
301 // 1. Order operands such that they are listed from right (least complex) to
302 // left (most complex). This puts constants before unary operators before
305 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
306 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
308 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
309 bool Changed = false;
310 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
311 Changed = !I.swapOperands();
313 if (!I.isAssociative()) return Changed;
314 Instruction::BinaryOps Opcode = I.getOpcode();
315 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
316 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
317 if (isa<Constant>(I.getOperand(1))) {
318 Constant *Folded = ConstantExpr::get(I.getOpcode(),
319 cast<Constant>(I.getOperand(1)),
320 cast<Constant>(Op->getOperand(1)));
321 I.setOperand(0, Op->getOperand(0));
322 I.setOperand(1, Folded);
324 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
325 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
326 isOnlyUse(Op) && isOnlyUse(Op1)) {
327 Constant *C1 = cast<Constant>(Op->getOperand(1));
328 Constant *C2 = cast<Constant>(Op1->getOperand(1));
330 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
331 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
332 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
335 WorkList.push_back(New);
336 I.setOperand(0, New);
337 I.setOperand(1, Folded);
344 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
345 // if the LHS is a constant zero (which is the 'negate' form).
347 static inline Value *dyn_castNegVal(Value *V) {
348 if (BinaryOperator::isNeg(V))
349 return BinaryOperator::getNegArgument(V);
351 // Constants can be considered to be negated values if they can be folded.
352 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
353 return ConstantExpr::getNeg(C);
357 static inline Value *dyn_castNotVal(Value *V) {
358 if (BinaryOperator::isNot(V))
359 return BinaryOperator::getNotArgument(V);
361 // Constants can be considered to be not'ed values...
362 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
363 return ConstantExpr::getNot(C);
367 // dyn_castFoldableMul - If this value is a multiply that can be folded into
368 // other computations (because it has a constant operand), return the
369 // non-constant operand of the multiply, and set CST to point to the multiplier.
370 // Otherwise, return null.
372 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
373 if (V->hasOneUse() && V->getType()->isInteger())
374 if (Instruction *I = dyn_cast<Instruction>(V)) {
375 if (I->getOpcode() == Instruction::Mul)
376 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
377 return I->getOperand(0);
378 if (I->getOpcode() == Instruction::Shl)
379 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
380 // The multiplier is really 1 << CST.
381 Constant *One = ConstantInt::get(V->getType(), 1);
382 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
383 return I->getOperand(0);
389 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
390 /// expression, return it.
391 static User *dyn_castGetElementPtr(Value *V) {
392 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
393 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
394 if (CE->getOpcode() == Instruction::GetElementPtr)
395 return cast<User>(V);
399 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
400 static ConstantInt *AddOne(ConstantInt *C) {
401 return cast<ConstantInt>(ConstantExpr::getAdd(C,
402 ConstantInt::get(C->getType(), 1)));
404 static ConstantInt *SubOne(ConstantInt *C) {
405 return cast<ConstantInt>(ConstantExpr::getSub(C,
406 ConstantInt::get(C->getType(), 1)));
409 /// ComputeMaskedNonZeroBits - Determine which of the bits specified in Mask are
410 /// not known to be zero and return them as a bitmask. The bits that we can
411 /// guarantee to be zero are returned as zero bits in the result.
412 static uint64_t ComputeMaskedNonZeroBits(Value *V, uint64_t Mask,
413 unsigned Depth = 0) {
414 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
415 // we cannot optimize based on the assumption that it is zero without changing
416 // it to be an explicit zero. If we don't change it to zero, other code could
417 // optimized based on the contradictory assumption that it is non-zero.
418 // Because instcombine aggressively folds operations with undef args anyway,
419 // this won't lose us code quality.
420 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V))
421 return CI->getRawValue() & Mask;
422 if (Depth == 6 || Mask == 0)
423 return Mask; // Limit search depth.
425 if (Instruction *I = dyn_cast<Instruction>(V)) {
426 switch (I->getOpcode()) {
427 case Instruction::And:
428 // (X & C1) & C2 == 0 iff C1 & C2 == 0.
429 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(I->getOperand(1)))
430 return ComputeMaskedNonZeroBits(I->getOperand(0),
431 CI->getRawValue() & Mask, Depth+1);
432 // If either the LHS or the RHS are MaskedValueIsZero, the result is zero.
433 Mask = ComputeMaskedNonZeroBits(I->getOperand(1), Mask, Depth+1);
434 Mask = ComputeMaskedNonZeroBits(I->getOperand(0), Mask, Depth+1);
436 case Instruction::Or:
437 case Instruction::Xor:
438 // Any non-zero bits in the LHS or RHS are potentially non-zero in the
440 return ComputeMaskedNonZeroBits(I->getOperand(1), Mask, Depth+1) |
441 ComputeMaskedNonZeroBits(I->getOperand(0), Mask, Depth+1);
442 case Instruction::Select:
443 // Any non-zero bits in the T or F values are potentially non-zero in the
445 return ComputeMaskedNonZeroBits(I->getOperand(2), Mask, Depth+1) |
446 ComputeMaskedNonZeroBits(I->getOperand(1), Mask, Depth+1);
447 case Instruction::Cast: {
448 const Type *SrcTy = I->getOperand(0)->getType();
449 if (SrcTy == Type::BoolTy)
450 return ComputeMaskedNonZeroBits(I->getOperand(0), Mask & 1, Depth+1);
451 if (!SrcTy->isInteger()) return Mask;
453 // (cast <ty> X to int) & C2 == 0 iff <ty> could not have contained C2.
454 if (SrcTy->isUnsigned() || // Only handle zero ext/trunc/noop
455 SrcTy->getPrimitiveSizeInBits() >=
456 I->getType()->getPrimitiveSizeInBits()) {
457 Mask &= SrcTy->getIntegralTypeMask();
458 return ComputeMaskedNonZeroBits(I->getOperand(0), Mask, Depth+1);
461 // FIXME: handle sext casts.
464 case Instruction::Shl:
465 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
466 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
467 return ComputeMaskedNonZeroBits(I->getOperand(0),Mask >> SA->getValue(),
468 Depth+1) << SA->getValue();
470 case Instruction::Shr:
471 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
472 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
473 if (I->getType()->isUnsigned()) {
474 Mask <<= SA->getValue();
475 Mask &= I->getType()->getIntegralTypeMask();
476 return ComputeMaskedNonZeroBits(I->getOperand(0), Mask, Depth+1)
486 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
487 /// this predicate to simplify operations downstream. Mask is known to be zero
488 /// for bits that V cannot have.
489 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
490 return ComputeMaskedNonZeroBits(V, Mask, Depth) == 0;
493 /// SimplifyDemandedBits - Look at V. At this point, we know that only the Mask
494 /// bits of the result of V are ever used downstream. If we can use this
495 /// information to simplify V, return V and set NewVal to the new value we
496 /// should use in V's place.
497 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t Mask,
499 if (!V->hasOneUse()) { // Other users may use these bits.
500 if (Depth != 0) // Not at the root.
502 // If this is the root being simplified, allow it to have multiple uses,
503 // just set the Mask to all bits.
504 Mask = V->getType()->getIntegralTypeMask();
505 } else if (Mask == 0) { // Not demanding any bits from V.
506 if (V != UndefValue::get(V->getType()))
507 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
509 } else if (Depth == 6) { // Limit search depth.
513 Instruction *I = dyn_cast<Instruction>(V);
514 if (!I) return false; // Only analyze instructions.
516 switch (I->getOpcode()) {
518 case Instruction::And:
519 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
520 // Only demanding an intersection of the bits.
521 if (SimplifyDemandedBits(I->getOperand(0), RHS->getRawValue() & Mask,
524 if (~Mask & RHS->getZExtValue()) {
525 // If this is producing any bits that are not needed, simplify the RHS.
526 uint64_t Val = Mask & RHS->getZExtValue();
528 ConstantUInt::get(I->getType()->getUnsignedVersion(), Val);
529 if (I->getType()->isSigned())
530 RHS = ConstantExpr::getCast(RHS, I->getType());
531 I->setOperand(1, RHS);
532 return UpdateValueUsesWith(I, I);
535 // Walk the LHS and the RHS.
536 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1) ||
537 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
538 case Instruction::Or:
539 case Instruction::Xor:
540 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
541 // If none of the [x]or'd in bits are demanded, don't both with the [x]or.
542 if ((Mask & RHS->getRawValue()) == 0)
543 return UpdateValueUsesWith(I, I->getOperand(0));
545 // Otherwise, for an OR, we only demand those bits not set by the OR.
546 if (I->getOpcode() == Instruction::Or)
547 Mask &= ~RHS->getRawValue();
548 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
550 // Walk the LHS and the RHS.
551 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1) ||
552 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
553 case Instruction::Cast: {
554 const Type *SrcTy = I->getOperand(0)->getType();
555 if (SrcTy == Type::BoolTy)
556 return SimplifyDemandedBits(I->getOperand(0), Mask&1, Depth+1);
558 if (!SrcTy->isInteger()) return false;
560 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
561 // If this is a sign-extend, treat specially.
562 if (SrcTy->isSigned() &&
563 SrcBits < I->getType()->getPrimitiveSizeInBits()) {
564 // If none of the top bits are demanded, convert this into an unsigned
565 // extend instead of a sign extend.
566 if ((Mask & ((1ULL << SrcBits)-1)) == 0) {
567 // Convert to unsigned first.
569 NewVal = new CastInst(I->getOperand(0), SrcTy->getUnsignedVersion(),
570 I->getOperand(0)->getName());
571 InsertNewInstBefore(NewVal, *I);
572 NewVal = new CastInst(NewVal, I->getType(), I->getName());
573 InsertNewInstBefore(NewVal, *I);
574 return UpdateValueUsesWith(I, NewVal);
577 // Otherwise, the high-bits *are* demanded. This means that the code
578 // implicitly demands computation of the sign bit of the input, make sure
579 // we explicitly include it in Mask.
580 Mask |= 1ULL << (SrcBits-1);
583 // If this is an extension, the top bits are ignored.
584 Mask &= SrcTy->getIntegralTypeMask();
585 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
587 case Instruction::Select:
588 // Simplify the T and F values if they are not demanded.
589 return SimplifyDemandedBits(I->getOperand(2), Mask, Depth+1) ||
590 SimplifyDemandedBits(I->getOperand(1), Mask, Depth+1);
591 case Instruction::Shl:
592 // We only demand the low bits of the input.
593 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
594 return SimplifyDemandedBits(I->getOperand(0), Mask >> SA->getValue(),
597 case Instruction::Shr:
598 // We only demand the high bits of the input.
599 if (I->getType()->isUnsigned())
600 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
601 Mask <<= SA->getValue();
602 Mask &= I->getType()->getIntegralTypeMask();
603 return SimplifyDemandedBits(I->getOperand(0), Mask, Depth+1);
605 // FIXME: handle signed shr, demanding the appropriate bits. If the top
606 // bits aren't demanded, strength reduce to a logical SHR instead.
612 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
613 // true when both operands are equal...
615 static bool isTrueWhenEqual(Instruction &I) {
616 return I.getOpcode() == Instruction::SetEQ ||
617 I.getOpcode() == Instruction::SetGE ||
618 I.getOpcode() == Instruction::SetLE;
621 /// AssociativeOpt - Perform an optimization on an associative operator. This
622 /// function is designed to check a chain of associative operators for a
623 /// potential to apply a certain optimization. Since the optimization may be
624 /// applicable if the expression was reassociated, this checks the chain, then
625 /// reassociates the expression as necessary to expose the optimization
626 /// opportunity. This makes use of a special Functor, which must define
627 /// 'shouldApply' and 'apply' methods.
629 template<typename Functor>
630 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
631 unsigned Opcode = Root.getOpcode();
632 Value *LHS = Root.getOperand(0);
634 // Quick check, see if the immediate LHS matches...
635 if (F.shouldApply(LHS))
636 return F.apply(Root);
638 // Otherwise, if the LHS is not of the same opcode as the root, return.
639 Instruction *LHSI = dyn_cast<Instruction>(LHS);
640 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
641 // Should we apply this transform to the RHS?
642 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
644 // If not to the RHS, check to see if we should apply to the LHS...
645 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
646 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
650 // If the functor wants to apply the optimization to the RHS of LHSI,
651 // reassociate the expression from ((? op A) op B) to (? op (A op B))
653 BasicBlock *BB = Root.getParent();
655 // Now all of the instructions are in the current basic block, go ahead
656 // and perform the reassociation.
657 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
659 // First move the selected RHS to the LHS of the root...
660 Root.setOperand(0, LHSI->getOperand(1));
662 // Make what used to be the LHS of the root be the user of the root...
663 Value *ExtraOperand = TmpLHSI->getOperand(1);
664 if (&Root == TmpLHSI) {
665 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
668 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
669 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
670 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
671 BasicBlock::iterator ARI = &Root; ++ARI;
672 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
675 // Now propagate the ExtraOperand down the chain of instructions until we
677 while (TmpLHSI != LHSI) {
678 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
679 // Move the instruction to immediately before the chain we are
680 // constructing to avoid breaking dominance properties.
681 NextLHSI->getParent()->getInstList().remove(NextLHSI);
682 BB->getInstList().insert(ARI, NextLHSI);
685 Value *NextOp = NextLHSI->getOperand(1);
686 NextLHSI->setOperand(1, ExtraOperand);
688 ExtraOperand = NextOp;
691 // Now that the instructions are reassociated, have the functor perform
692 // the transformation...
693 return F.apply(Root);
696 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
702 // AddRHS - Implements: X + X --> X << 1
705 AddRHS(Value *rhs) : RHS(rhs) {}
706 bool shouldApply(Value *LHS) const { return LHS == RHS; }
707 Instruction *apply(BinaryOperator &Add) const {
708 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
709 ConstantInt::get(Type::UByteTy, 1));
713 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
715 struct AddMaskingAnd {
717 AddMaskingAnd(Constant *c) : C2(c) {}
718 bool shouldApply(Value *LHS) const {
720 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
721 ConstantExpr::getAnd(C1, C2)->isNullValue();
723 Instruction *apply(BinaryOperator &Add) const {
724 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
728 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
730 if (isa<CastInst>(I)) {
731 if (Constant *SOC = dyn_cast<Constant>(SO))
732 return ConstantExpr::getCast(SOC, I.getType());
734 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
735 SO->getName() + ".cast"), I);
738 // Figure out if the constant is the left or the right argument.
739 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
740 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
742 if (Constant *SOC = dyn_cast<Constant>(SO)) {
744 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
745 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
748 Value *Op0 = SO, *Op1 = ConstOperand;
752 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
753 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
754 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
755 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
757 assert(0 && "Unknown binary instruction type!");
760 return IC->InsertNewInstBefore(New, I);
763 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
764 // constant as the other operand, try to fold the binary operator into the
765 // select arguments. This also works for Cast instructions, which obviously do
766 // not have a second operand.
767 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
769 // Don't modify shared select instructions
770 if (!SI->hasOneUse()) return 0;
771 Value *TV = SI->getOperand(1);
772 Value *FV = SI->getOperand(2);
774 if (isa<Constant>(TV) || isa<Constant>(FV)) {
775 // Bool selects with constant operands can be folded to logical ops.
776 if (SI->getType() == Type::BoolTy) return 0;
778 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
779 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
781 return new SelectInst(SI->getCondition(), SelectTrueVal,
788 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
789 /// node as operand #0, see if we can fold the instruction into the PHI (which
790 /// is only possible if all operands to the PHI are constants).
791 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
792 PHINode *PN = cast<PHINode>(I.getOperand(0));
793 unsigned NumPHIValues = PN->getNumIncomingValues();
794 if (!PN->hasOneUse() || NumPHIValues == 0 ||
795 !isa<Constant>(PN->getIncomingValue(0))) return 0;
797 // Check to see if all of the operands of the PHI are constants. If not, we
798 // cannot do the transformation.
799 for (unsigned i = 1; i != NumPHIValues; ++i)
800 if (!isa<Constant>(PN->getIncomingValue(i)))
803 // Okay, we can do the transformation: create the new PHI node.
804 PHINode *NewPN = new PHINode(I.getType(), I.getName());
806 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
807 InsertNewInstBefore(NewPN, *PN);
809 // Next, add all of the operands to the PHI.
810 if (I.getNumOperands() == 2) {
811 Constant *C = cast<Constant>(I.getOperand(1));
812 for (unsigned i = 0; i != NumPHIValues; ++i) {
813 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
814 NewPN->addIncoming(ConstantExpr::get(I.getOpcode(), InV, C),
815 PN->getIncomingBlock(i));
818 assert(isa<CastInst>(I) && "Unary op should be a cast!");
819 const Type *RetTy = I.getType();
820 for (unsigned i = 0; i != NumPHIValues; ++i) {
821 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
822 NewPN->addIncoming(ConstantExpr::getCast(InV, RetTy),
823 PN->getIncomingBlock(i));
826 return ReplaceInstUsesWith(I, NewPN);
829 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
830 bool Changed = SimplifyCommutative(I);
831 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
833 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
834 // X + undef -> undef
835 if (isa<UndefValue>(RHS))
836 return ReplaceInstUsesWith(I, RHS);
839 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
840 if (RHSC->isNullValue())
841 return ReplaceInstUsesWith(I, LHS);
842 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
843 if (CFP->isExactlyValue(-0.0))
844 return ReplaceInstUsesWith(I, LHS);
847 // X + (signbit) --> X ^ signbit
848 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
849 uint64_t Val = CI->getZExtValue();
850 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
851 return BinaryOperator::createXor(LHS, RHS);
854 if (isa<PHINode>(LHS))
855 if (Instruction *NV = FoldOpIntoPhi(I))
858 ConstantInt *XorRHS = 0;
860 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
861 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
862 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
863 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
865 uint64_t C0080Val = 1ULL << 31;
866 int64_t CFF80Val = -C0080Val;
869 if (TySizeBits > Size) {
871 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
872 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
873 if (RHSSExt == CFF80Val) {
874 if (XorRHS->getZExtValue() == C0080Val)
876 } else if (RHSZExt == C0080Val) {
877 if (XorRHS->getSExtValue() == CFF80Val)
881 // This is a sign extend if the top bits are known zero.
882 uint64_t Mask = XorLHS->getType()->getIntegralTypeMask();
883 Mask <<= 64-(TySizeBits-Size);
884 if (!MaskedValueIsZero(XorLHS, Mask))
885 Size = 0; // Not a sign ext, but can't be any others either.
895 const Type *MiddleType = 0;
898 case 32: MiddleType = Type::IntTy; break;
899 case 16: MiddleType = Type::ShortTy; break;
900 case 8: MiddleType = Type::SByteTy; break;
903 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
904 InsertNewInstBefore(NewTrunc, I);
905 return new CastInst(NewTrunc, I.getType());
911 if (I.getType()->isInteger()) {
912 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
914 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
915 if (RHSI->getOpcode() == Instruction::Sub)
916 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
917 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
919 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
920 if (LHSI->getOpcode() == Instruction::Sub)
921 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
922 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
927 if (Value *V = dyn_castNegVal(LHS))
928 return BinaryOperator::createSub(RHS, V);
931 if (!isa<Constant>(RHS))
932 if (Value *V = dyn_castNegVal(RHS))
933 return BinaryOperator::createSub(LHS, V);
937 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
938 if (X == RHS) // X*C + X --> X * (C+1)
939 return BinaryOperator::createMul(RHS, AddOne(C2));
941 // X*C1 + X*C2 --> X * (C1+C2)
943 if (X == dyn_castFoldableMul(RHS, C1))
944 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
947 // X + X*C --> X * (C+1)
948 if (dyn_castFoldableMul(RHS, C2) == LHS)
949 return BinaryOperator::createMul(LHS, AddOne(C2));
952 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
953 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
954 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
956 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
958 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
959 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
960 return BinaryOperator::createSub(C, X);
963 // (X & FF00) + xx00 -> (X+xx00) & FF00
964 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
965 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
967 // See if all bits from the first bit set in the Add RHS up are included
968 // in the mask. First, get the rightmost bit.
969 uint64_t AddRHSV = CRHS->getRawValue();
971 // Form a mask of all bits from the lowest bit added through the top.
972 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
973 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
975 // See if the and mask includes all of these bits.
976 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
978 if (AddRHSHighBits == AddRHSHighBitsAnd) {
979 // Okay, the xform is safe. Insert the new add pronto.
980 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
982 return BinaryOperator::createAnd(NewAdd, C2);
987 // Try to fold constant add into select arguments.
988 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
989 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
993 return Changed ? &I : 0;
996 // isSignBit - Return true if the value represented by the constant only has the
997 // highest order bit set.
998 static bool isSignBit(ConstantInt *CI) {
999 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1000 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1003 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1005 static Value *RemoveNoopCast(Value *V) {
1006 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1007 const Type *CTy = CI->getType();
1008 const Type *OpTy = CI->getOperand(0)->getType();
1009 if (CTy->isInteger() && OpTy->isInteger()) {
1010 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1011 return RemoveNoopCast(CI->getOperand(0));
1012 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1013 return RemoveNoopCast(CI->getOperand(0));
1018 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1019 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1021 if (Op0 == Op1) // sub X, X -> 0
1022 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1024 // If this is a 'B = x-(-A)', change to B = x+A...
1025 if (Value *V = dyn_castNegVal(Op1))
1026 return BinaryOperator::createAdd(Op0, V);
1028 if (isa<UndefValue>(Op0))
1029 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1030 if (isa<UndefValue>(Op1))
1031 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1033 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1034 // Replace (-1 - A) with (~A)...
1035 if (C->isAllOnesValue())
1036 return BinaryOperator::createNot(Op1);
1038 // C - ~X == X + (1+C)
1040 if (match(Op1, m_Not(m_Value(X))))
1041 return BinaryOperator::createAdd(X,
1042 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1043 // -((uint)X >> 31) -> ((int)X >> 31)
1044 // -((int)X >> 31) -> ((uint)X >> 31)
1045 if (C->isNullValue()) {
1046 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1047 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1048 if (SI->getOpcode() == Instruction::Shr)
1049 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
1051 if (SI->getType()->isSigned())
1052 NewTy = SI->getType()->getUnsignedVersion();
1054 NewTy = SI->getType()->getSignedVersion();
1055 // Check to see if we are shifting out everything but the sign bit.
1056 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
1057 // Ok, the transformation is safe. Insert a cast of the incoming
1058 // value, then the new shift, then the new cast.
1059 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
1060 SI->getOperand(0)->getName());
1061 Value *InV = InsertNewInstBefore(FirstCast, I);
1062 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
1064 if (NewShift->getType() == I.getType())
1067 InV = InsertNewInstBefore(NewShift, I);
1068 return new CastInst(NewShift, I.getType());
1074 // Try to fold constant sub into select arguments.
1075 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1076 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1079 if (isa<PHINode>(Op0))
1080 if (Instruction *NV = FoldOpIntoPhi(I))
1084 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1085 if (Op1I->getOpcode() == Instruction::Add &&
1086 !Op0->getType()->isFloatingPoint()) {
1087 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1088 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1089 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1090 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1091 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1092 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1093 // C1-(X+C2) --> (C1-C2)-X
1094 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1095 Op1I->getOperand(0));
1099 if (Op1I->hasOneUse()) {
1100 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1101 // is not used by anyone else...
1103 if (Op1I->getOpcode() == Instruction::Sub &&
1104 !Op1I->getType()->isFloatingPoint()) {
1105 // Swap the two operands of the subexpr...
1106 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1107 Op1I->setOperand(0, IIOp1);
1108 Op1I->setOperand(1, IIOp0);
1110 // Create the new top level add instruction...
1111 return BinaryOperator::createAdd(Op0, Op1);
1114 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1116 if (Op1I->getOpcode() == Instruction::And &&
1117 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1118 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1121 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1122 return BinaryOperator::createAnd(Op0, NewNot);
1125 // -(X sdiv C) -> (X sdiv -C)
1126 if (Op1I->getOpcode() == Instruction::Div)
1127 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
1128 if (CSI->isNullValue())
1129 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1130 return BinaryOperator::createDiv(Op1I->getOperand(0),
1131 ConstantExpr::getNeg(DivRHS));
1133 // X - X*C --> X * (1-C)
1134 ConstantInt *C2 = 0;
1135 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1137 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1138 return BinaryOperator::createMul(Op0, CP1);
1143 if (!Op0->getType()->isFloatingPoint())
1144 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1145 if (Op0I->getOpcode() == Instruction::Add) {
1146 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1147 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1148 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1149 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1150 } else if (Op0I->getOpcode() == Instruction::Sub) {
1151 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1152 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1156 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1157 if (X == Op1) { // X*C - X --> X * (C-1)
1158 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1159 return BinaryOperator::createMul(Op1, CP1);
1162 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1163 if (X == dyn_castFoldableMul(Op1, C2))
1164 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1169 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1170 /// really just returns true if the most significant (sign) bit is set.
1171 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1172 if (RHS->getType()->isSigned()) {
1173 // True if source is LHS < 0 or LHS <= -1
1174 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1175 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1177 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1178 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1179 // the size of the integer type.
1180 if (Opcode == Instruction::SetGE)
1181 return RHSC->getValue() ==
1182 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1183 if (Opcode == Instruction::SetGT)
1184 return RHSC->getValue() ==
1185 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1190 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1191 bool Changed = SimplifyCommutative(I);
1192 Value *Op0 = I.getOperand(0);
1194 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1195 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1197 // Simplify mul instructions with a constant RHS...
1198 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1199 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1201 // ((X << C1)*C2) == (X * (C2 << C1))
1202 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1203 if (SI->getOpcode() == Instruction::Shl)
1204 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1205 return BinaryOperator::createMul(SI->getOperand(0),
1206 ConstantExpr::getShl(CI, ShOp));
1208 if (CI->isNullValue())
1209 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1210 if (CI->equalsInt(1)) // X * 1 == X
1211 return ReplaceInstUsesWith(I, Op0);
1212 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1213 return BinaryOperator::createNeg(Op0, I.getName());
1215 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1216 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1217 uint64_t C = Log2_64(Val);
1218 return new ShiftInst(Instruction::Shl, Op0,
1219 ConstantUInt::get(Type::UByteTy, C));
1221 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1222 if (Op1F->isNullValue())
1223 return ReplaceInstUsesWith(I, Op1);
1225 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1226 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1227 if (Op1F->getValue() == 1.0)
1228 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1231 // Try to fold constant mul into select arguments.
1232 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1233 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1236 if (isa<PHINode>(Op0))
1237 if (Instruction *NV = FoldOpIntoPhi(I))
1241 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1242 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1243 return BinaryOperator::createMul(Op0v, Op1v);
1245 // If one of the operands of the multiply is a cast from a boolean value, then
1246 // we know the bool is either zero or one, so this is a 'masking' multiply.
1247 // See if we can simplify things based on how the boolean was originally
1249 CastInst *BoolCast = 0;
1250 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1251 if (CI->getOperand(0)->getType() == Type::BoolTy)
1254 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1255 if (CI->getOperand(0)->getType() == Type::BoolTy)
1258 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1259 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1260 const Type *SCOpTy = SCIOp0->getType();
1262 // If the setcc is true iff the sign bit of X is set, then convert this
1263 // multiply into a shift/and combination.
1264 if (isa<ConstantInt>(SCIOp1) &&
1265 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1266 // Shift the X value right to turn it into "all signbits".
1267 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1268 SCOpTy->getPrimitiveSizeInBits()-1);
1269 if (SCIOp0->getType()->isUnsigned()) {
1270 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1271 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1272 SCIOp0->getName()), I);
1276 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1277 BoolCast->getOperand(0)->getName()+
1280 // If the multiply type is not the same as the source type, sign extend
1281 // or truncate to the multiply type.
1282 if (I.getType() != V->getType())
1283 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1285 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1286 return BinaryOperator::createAnd(V, OtherOp);
1291 return Changed ? &I : 0;
1294 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1295 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1297 if (isa<UndefValue>(Op0)) // undef / X -> 0
1298 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1299 if (isa<UndefValue>(Op1))
1300 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1302 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1304 if (RHS->equalsInt(1))
1305 return ReplaceInstUsesWith(I, Op0);
1308 if (RHS->isAllOnesValue())
1309 return BinaryOperator::createNeg(Op0);
1311 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1312 if (LHS->getOpcode() == Instruction::Div)
1313 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1314 // (X / C1) / C2 -> X / (C1*C2)
1315 return BinaryOperator::createDiv(LHS->getOperand(0),
1316 ConstantExpr::getMul(RHS, LHSRHS));
1319 // Check to see if this is an unsigned division with an exact power of 2,
1320 // if so, convert to a right shift.
1321 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1322 if (uint64_t Val = C->getValue()) // Don't break X / 0
1323 if (isPowerOf2_64(Val)) {
1324 uint64_t C = Log2_64(Val);
1325 return new ShiftInst(Instruction::Shr, Op0,
1326 ConstantUInt::get(Type::UByteTy, C));
1330 if (RHS->getType()->isSigned())
1331 if (Value *LHSNeg = dyn_castNegVal(Op0))
1332 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1334 if (!RHS->isNullValue()) {
1335 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1336 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1338 if (isa<PHINode>(Op0))
1339 if (Instruction *NV = FoldOpIntoPhi(I))
1344 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1345 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
1346 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1347 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1348 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1349 if (STO->getValue() == 0) { // Couldn't be this argument.
1350 I.setOperand(1, SFO);
1352 } else if (SFO->getValue() == 0) {
1353 I.setOperand(1, STO);
1357 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
1358 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
1359 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
1360 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
1361 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
1362 TC, SI->getName()+".t");
1363 TSI = InsertNewInstBefore(TSI, I);
1365 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
1366 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
1367 FC, SI->getName()+".f");
1368 FSI = InsertNewInstBefore(FSI, I);
1369 return new SelectInst(SI->getOperand(0), TSI, FSI);
1373 // 0 / X == 0, we don't need to preserve faults!
1374 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1375 if (LHS->equalsInt(0))
1376 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1378 if (I.getType()->isSigned()) {
1379 // If the sign bits of both operands are zero (i.e. we can prove they are
1380 // unsigned inputs), turn this into a udiv.
1381 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
1382 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1383 const Type *NTy = Op0->getType()->getUnsignedVersion();
1384 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1385 InsertNewInstBefore(LHS, I);
1387 if (Constant *R = dyn_cast<Constant>(Op1))
1388 RHS = ConstantExpr::getCast(R, NTy);
1390 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1391 Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
1392 InsertNewInstBefore(Div, I);
1393 return new CastInst(Div, I.getType());
1396 // Known to be an unsigned division.
1397 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1398 // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
1399 if (RHSI->getOpcode() == Instruction::Shl &&
1400 isa<ConstantUInt>(RHSI->getOperand(0))) {
1401 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1402 if (isPowerOf2_64(C1)) {
1403 unsigned C2 = Log2_64(C1);
1404 Value *Add = RHSI->getOperand(1);
1406 Constant *C2V = ConstantUInt::get(Add->getType(), C2);
1407 Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
1410 return new ShiftInst(Instruction::Shr, Op0, Add);
1420 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
1421 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1422 if (I.getType()->isSigned()) {
1423 if (Value *RHSNeg = dyn_castNegVal(Op1))
1424 if (!isa<ConstantSInt>(RHSNeg) ||
1425 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
1427 AddUsesToWorkList(I);
1428 I.setOperand(1, RHSNeg);
1432 // If the top bits of both operands are zero (i.e. we can prove they are
1433 // unsigned inputs), turn this into a urem.
1434 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
1435 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1436 const Type *NTy = Op0->getType()->getUnsignedVersion();
1437 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1438 InsertNewInstBefore(LHS, I);
1440 if (Constant *R = dyn_cast<Constant>(Op1))
1441 RHS = ConstantExpr::getCast(R, NTy);
1443 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1444 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
1445 InsertNewInstBefore(Rem, I);
1446 return new CastInst(Rem, I.getType());
1450 if (isa<UndefValue>(Op0)) // undef % X -> 0
1451 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1452 if (isa<UndefValue>(Op1))
1453 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
1455 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1456 if (RHS->equalsInt(1)) // X % 1 == 0
1457 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1459 // Check to see if this is an unsigned remainder with an exact power of 2,
1460 // if so, convert to a bitwise and.
1461 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1462 if (uint64_t Val = C->getValue()) // Don't break X % 0 (divide by zero)
1463 if (!(Val & (Val-1))) // Power of 2
1464 return BinaryOperator::createAnd(Op0,
1465 ConstantUInt::get(I.getType(), Val-1));
1467 if (!RHS->isNullValue()) {
1468 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1469 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1471 if (isa<PHINode>(Op0))
1472 if (Instruction *NV = FoldOpIntoPhi(I))
1477 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1478 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
1479 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1480 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1481 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1482 if (STO->getValue() == 0) { // Couldn't be this argument.
1483 I.setOperand(1, SFO);
1485 } else if (SFO->getValue() == 0) {
1486 I.setOperand(1, STO);
1490 if (!(STO->getValue() & (STO->getValue()-1)) &&
1491 !(SFO->getValue() & (SFO->getValue()-1))) {
1492 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1493 SubOne(STO), SI->getName()+".t"), I);
1494 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1495 SubOne(SFO), SI->getName()+".f"), I);
1496 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
1500 // 0 % X == 0, we don't need to preserve faults!
1501 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1502 if (LHS->equalsInt(0))
1503 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1506 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1507 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
1508 if (I.getType()->isUnsigned() &&
1509 RHSI->getOpcode() == Instruction::Shl &&
1510 isa<ConstantUInt>(RHSI->getOperand(0))) {
1511 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1512 if (isPowerOf2_64(C1)) {
1513 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
1514 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
1516 return BinaryOperator::createAnd(Op0, Add);
1524 // isMaxValueMinusOne - return true if this is Max-1
1525 static bool isMaxValueMinusOne(const ConstantInt *C) {
1526 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1527 return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
1529 const ConstantSInt *CS = cast<ConstantSInt>(C);
1531 // Calculate 0111111111..11111
1532 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1533 int64_t Val = INT64_MAX; // All ones
1534 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
1535 return CS->getValue() == Val-1;
1538 // isMinValuePlusOne - return true if this is Min+1
1539 static bool isMinValuePlusOne(const ConstantInt *C) {
1540 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1541 return CU->getValue() == 1;
1543 const ConstantSInt *CS = cast<ConstantSInt>(C);
1545 // Calculate 1111111111000000000000
1546 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1547 int64_t Val = -1; // All ones
1548 Val <<= TypeBits-1; // Shift over to the right spot
1549 return CS->getValue() == Val+1;
1552 // isOneBitSet - Return true if there is exactly one bit set in the specified
1554 static bool isOneBitSet(const ConstantInt *CI) {
1555 uint64_t V = CI->getRawValue();
1556 return V && (V & (V-1)) == 0;
1559 #if 0 // Currently unused
1560 // isLowOnes - Return true if the constant is of the form 0+1+.
1561 static bool isLowOnes(const ConstantInt *CI) {
1562 uint64_t V = CI->getRawValue();
1564 // There won't be bits set in parts that the type doesn't contain.
1565 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1567 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1568 return U && V && (U & V) == 0;
1572 // isHighOnes - Return true if the constant is of the form 1+0+.
1573 // This is the same as lowones(~X).
1574 static bool isHighOnes(const ConstantInt *CI) {
1575 uint64_t V = ~CI->getRawValue();
1576 if (~V == 0) return false; // 0's does not match "1+"
1578 // There won't be bits set in parts that the type doesn't contain.
1579 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
1581 uint64_t U = V+1; // If it is low ones, this should be a power of two.
1582 return U && V && (U & V) == 0;
1586 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
1587 /// are carefully arranged to allow folding of expressions such as:
1589 /// (A < B) | (A > B) --> (A != B)
1591 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
1592 /// represents that the comparison is true if A == B, and bit value '1' is true
1595 static unsigned getSetCondCode(const SetCondInst *SCI) {
1596 switch (SCI->getOpcode()) {
1598 case Instruction::SetGT: return 1;
1599 case Instruction::SetEQ: return 2;
1600 case Instruction::SetGE: return 3;
1601 case Instruction::SetLT: return 4;
1602 case Instruction::SetNE: return 5;
1603 case Instruction::SetLE: return 6;
1606 assert(0 && "Invalid SetCC opcode!");
1611 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
1612 /// opcode and two operands into either a constant true or false, or a brand new
1613 /// SetCC instruction.
1614 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
1616 case 0: return ConstantBool::False;
1617 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
1618 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
1619 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
1620 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
1621 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
1622 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
1623 case 7: return ConstantBool::True;
1624 default: assert(0 && "Illegal SetCCCode!"); return 0;
1628 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
1629 struct FoldSetCCLogical {
1632 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
1633 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
1634 bool shouldApply(Value *V) const {
1635 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
1636 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
1637 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
1640 Instruction *apply(BinaryOperator &Log) const {
1641 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
1642 if (SCI->getOperand(0) != LHS) {
1643 assert(SCI->getOperand(1) == LHS);
1644 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
1647 unsigned LHSCode = getSetCondCode(SCI);
1648 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
1650 switch (Log.getOpcode()) {
1651 case Instruction::And: Code = LHSCode & RHSCode; break;
1652 case Instruction::Or: Code = LHSCode | RHSCode; break;
1653 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
1654 default: assert(0 && "Illegal logical opcode!"); return 0;
1657 Value *RV = getSetCCValue(Code, LHS, RHS);
1658 if (Instruction *I = dyn_cast<Instruction>(RV))
1660 // Otherwise, it's a constant boolean value...
1661 return IC.ReplaceInstUsesWith(Log, RV);
1665 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
1666 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
1667 // guaranteed to be either a shift instruction or a binary operator.
1668 Instruction *InstCombiner::OptAndOp(Instruction *Op,
1669 ConstantIntegral *OpRHS,
1670 ConstantIntegral *AndRHS,
1671 BinaryOperator &TheAnd) {
1672 Value *X = Op->getOperand(0);
1673 Constant *Together = 0;
1674 if (!isa<ShiftInst>(Op))
1675 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
1677 switch (Op->getOpcode()) {
1678 case Instruction::Xor:
1679 if (Op->hasOneUse()) {
1680 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1681 std::string OpName = Op->getName(); Op->setName("");
1682 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
1683 InsertNewInstBefore(And, TheAnd);
1684 return BinaryOperator::createXor(And, Together);
1687 case Instruction::Or:
1688 if (Together == AndRHS) // (X | C) & C --> C
1689 return ReplaceInstUsesWith(TheAnd, AndRHS);
1691 if (Op->hasOneUse() && Together != OpRHS) {
1692 // (X | C1) & C2 --> (X | (C1&C2)) & C2
1693 std::string Op0Name = Op->getName(); Op->setName("");
1694 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
1695 InsertNewInstBefore(Or, TheAnd);
1696 return BinaryOperator::createAnd(Or, AndRHS);
1699 case Instruction::Add:
1700 if (Op->hasOneUse()) {
1701 // Adding a one to a single bit bit-field should be turned into an XOR
1702 // of the bit. First thing to check is to see if this AND is with a
1703 // single bit constant.
1704 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
1706 // Clear bits that are not part of the constant.
1707 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
1709 // If there is only one bit set...
1710 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
1711 // Ok, at this point, we know that we are masking the result of the
1712 // ADD down to exactly one bit. If the constant we are adding has
1713 // no bits set below this bit, then we can eliminate the ADD.
1714 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
1716 // Check to see if any bits below the one bit set in AndRHSV are set.
1717 if ((AddRHS & (AndRHSV-1)) == 0) {
1718 // If not, the only thing that can effect the output of the AND is
1719 // the bit specified by AndRHSV. If that bit is set, the effect of
1720 // the XOR is to toggle the bit. If it is clear, then the ADD has
1722 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
1723 TheAnd.setOperand(0, X);
1726 std::string Name = Op->getName(); Op->setName("");
1727 // Pull the XOR out of the AND.
1728 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
1729 InsertNewInstBefore(NewAnd, TheAnd);
1730 return BinaryOperator::createXor(NewAnd, AndRHS);
1737 case Instruction::Shl: {
1738 // We know that the AND will not produce any of the bits shifted in, so if
1739 // the anded constant includes them, clear them now!
1741 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1742 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
1743 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
1745 if (CI == ShlMask) { // Masking out bits that the shift already masks
1746 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
1747 } else if (CI != AndRHS) { // Reducing bits set in and.
1748 TheAnd.setOperand(1, CI);
1753 case Instruction::Shr:
1754 // We know that the AND will not produce any of the bits shifted in, so if
1755 // the anded constant includes them, clear them now! This only applies to
1756 // unsigned shifts, because a signed shr may bring in set bits!
1758 if (AndRHS->getType()->isUnsigned()) {
1759 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1760 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
1761 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1763 if (CI == ShrMask) { // Masking out bits that the shift already masks.
1764 return ReplaceInstUsesWith(TheAnd, Op);
1765 } else if (CI != AndRHS) {
1766 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
1769 } else { // Signed shr.
1770 // See if this is shifting in some sign extension, then masking it out
1772 if (Op->hasOneUse()) {
1773 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
1774 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
1775 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
1776 if (CI == AndRHS) { // Masking out bits shifted in.
1777 // Make the argument unsigned.
1778 Value *ShVal = Op->getOperand(0);
1779 ShVal = InsertCastBefore(ShVal,
1780 ShVal->getType()->getUnsignedVersion(),
1782 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
1783 OpRHS, Op->getName()),
1785 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
1786 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
1789 return new CastInst(ShVal, Op->getType());
1799 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
1800 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
1801 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
1802 /// insert new instructions.
1803 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
1804 bool Inside, Instruction &IB) {
1805 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
1806 "Lo is not <= Hi in range emission code!");
1808 if (Lo == Hi) // Trivially false.
1809 return new SetCondInst(Instruction::SetNE, V, V);
1810 if (cast<ConstantIntegral>(Lo)->isMinValue())
1811 return new SetCondInst(Instruction::SetLT, V, Hi);
1813 Constant *AddCST = ConstantExpr::getNeg(Lo);
1814 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
1815 InsertNewInstBefore(Add, IB);
1816 // Convert to unsigned for the comparison.
1817 const Type *UnsType = Add->getType()->getUnsignedVersion();
1818 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1819 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1820 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1821 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
1824 if (Lo == Hi) // Trivially true.
1825 return new SetCondInst(Instruction::SetEQ, V, V);
1827 Hi = SubOne(cast<ConstantInt>(Hi));
1828 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
1829 return new SetCondInst(Instruction::SetGT, V, Hi);
1831 // Emit X-Lo > Hi-Lo-1
1832 Constant *AddCST = ConstantExpr::getNeg(Lo);
1833 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
1834 InsertNewInstBefore(Add, IB);
1835 // Convert to unsigned for the comparison.
1836 const Type *UnsType = Add->getType()->getUnsignedVersion();
1837 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
1838 AddCST = ConstantExpr::getAdd(AddCST, Hi);
1839 AddCST = ConstantExpr::getCast(AddCST, UnsType);
1840 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
1843 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
1844 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
1845 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
1846 // not, since all 1s are not contiguous.
1847 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
1848 uint64_t V = Val->getRawValue();
1849 if (!isShiftedMask_64(V)) return false;
1851 // look for the first zero bit after the run of ones
1852 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
1853 // look for the first non-zero bit
1854 ME = 64-CountLeadingZeros_64(V);
1860 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
1861 /// where isSub determines whether the operator is a sub. If we can fold one of
1862 /// the following xforms:
1864 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
1865 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1866 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
1868 /// return (A +/- B).
1870 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
1871 ConstantIntegral *Mask, bool isSub,
1873 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1874 if (!LHSI || LHSI->getNumOperands() != 2 ||
1875 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
1877 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
1879 switch (LHSI->getOpcode()) {
1881 case Instruction::And:
1882 if (ConstantExpr::getAnd(N, Mask) == Mask) {
1883 // If the AndRHS is a power of two minus one (0+1+), this is simple.
1884 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
1887 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
1888 // part, we don't need any explicit masks to take them out of A. If that
1889 // is all N is, ignore it.
1891 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
1892 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
1894 if (MaskedValueIsZero(RHS, Mask))
1899 case Instruction::Or:
1900 case Instruction::Xor:
1901 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
1902 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
1903 ConstantExpr::getAnd(N, Mask)->isNullValue())
1910 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
1912 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
1913 return InsertNewInstBefore(New, I);
1916 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1917 bool Changed = SimplifyCommutative(I);
1918 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1920 if (isa<UndefValue>(Op1)) // X & undef -> 0
1921 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1925 return ReplaceInstUsesWith(I, Op1);
1927 // See if we can simplify any instructions used by the LHS whose sole
1928 // purpose is to compute bits we don't care about.
1929 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask()))
1932 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
1933 uint64_t AndRHSMask = AndRHS->getZExtValue();
1934 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
1936 if (AndRHSMask == TypeMask) // and X, -1 == X
1937 return ReplaceInstUsesWith(I, Op0);
1938 else if (AndRHSMask == 0) // and X, 0 == 0
1939 return ReplaceInstUsesWith(I, AndRHS);
1941 // and (and X, c1), c2 -> and (x, c1&c2). Handle this case here, before
1942 // calling ComputeMaskedNonZeroBits, to avoid inefficient cases where we
1943 // traipse through many levels of ands.
1945 Value *X = 0; ConstantInt *C1 = 0;
1946 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))))
1947 return BinaryOperator::createAnd(X, ConstantExpr::getAnd(C1, AndRHS));
1950 // Figure out which of the input bits are not known to be zero, and which
1951 // bits are known to be zero.
1952 uint64_t NonZeroBits = ComputeMaskedNonZeroBits(Op0, TypeMask);
1953 uint64_t ZeroBits = NonZeroBits^TypeMask;
1955 // If the mask is not masking out any bits (i.e. all of the zeros in the
1956 // mask are already known to be zero), there is no reason to do the and in
1958 uint64_t NotAndRHS = AndRHSMask^TypeMask;
1959 if ((NotAndRHS & ZeroBits) == NotAndRHS)
1960 return ReplaceInstUsesWith(I, Op0);
1962 // If the AND mask contains bits that are known zero, remove them. A
1963 // special case is when there are no bits in common, in which case we
1964 // implicitly turn this into an AND X, 0, which is later simplified into 0.
1965 if ((AndRHSMask & NonZeroBits) != AndRHSMask) {
1967 ConstantUInt::get(Type::ULongTy, AndRHSMask & NonZeroBits);
1968 I.setOperand(1, ConstantExpr::getCast(NewRHS, I.getType()));
1972 // Optimize a variety of ((val OP C1) & C2) combinations...
1973 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
1974 Instruction *Op0I = cast<Instruction>(Op0);
1975 Value *Op0LHS = Op0I->getOperand(0);
1976 Value *Op0RHS = Op0I->getOperand(1);
1977 switch (Op0I->getOpcode()) {
1978 case Instruction::Xor:
1979 case Instruction::Or:
1980 // (X ^ V) & C2 --> (X & C2) iff (V & C2) == 0
1981 // (X | V) & C2 --> (X & C2) iff (V & C2) == 0
1982 if (MaskedValueIsZero(Op0LHS, AndRHSMask))
1983 return BinaryOperator::createAnd(Op0RHS, AndRHS);
1984 if (MaskedValueIsZero(Op0RHS, AndRHSMask))
1985 return BinaryOperator::createAnd(Op0LHS, AndRHS);
1987 // If the mask is only needed on one incoming arm, push it up.
1988 if (Op0I->hasOneUse()) {
1989 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1990 // Not masking anything out for the LHS, move to RHS.
1991 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
1992 Op0RHS->getName()+".masked");
1993 InsertNewInstBefore(NewRHS, I);
1994 return BinaryOperator::create(
1995 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
1997 if (!isa<Constant>(Op0RHS) &&
1998 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1999 // Not masking anything out for the RHS, move to LHS.
2000 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
2001 Op0LHS->getName()+".masked");
2002 InsertNewInstBefore(NewLHS, I);
2003 return BinaryOperator::create(
2004 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
2009 case Instruction::And:
2010 // (X & V) & C2 --> 0 iff (V & C2) == 0
2011 if (MaskedValueIsZero(Op0LHS, AndRHSMask) ||
2012 MaskedValueIsZero(Op0RHS, AndRHSMask))
2013 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2015 case Instruction::Add:
2016 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2017 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2018 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2019 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2020 return BinaryOperator::createAnd(V, AndRHS);
2021 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2022 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
2025 case Instruction::Sub:
2026 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2027 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2028 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2029 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2030 return BinaryOperator::createAnd(V, AndRHS);
2034 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2035 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2037 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2038 const Type *SrcTy = CI->getOperand(0)->getType();
2040 // If this is an integer truncation or change from signed-to-unsigned, and
2041 // if the source is an and/or with immediate, transform it. This
2042 // frequently occurs for bitfield accesses.
2043 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2044 if (SrcTy->getPrimitiveSizeInBits() >=
2045 I.getType()->getPrimitiveSizeInBits() &&
2046 CastOp->getNumOperands() == 2)
2047 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
2048 if (CastOp->getOpcode() == Instruction::And) {
2049 // Change: and (cast (and X, C1) to T), C2
2050 // into : and (cast X to T), trunc(C1)&C2
2051 // This will folds the two ands together, which may allow other
2053 Instruction *NewCast =
2054 new CastInst(CastOp->getOperand(0), I.getType(),
2055 CastOp->getName()+".shrunk");
2056 NewCast = InsertNewInstBefore(NewCast, I);
2058 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2059 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
2060 return BinaryOperator::createAnd(NewCast, C3);
2061 } else if (CastOp->getOpcode() == Instruction::Or) {
2062 // Change: and (cast (or X, C1) to T), C2
2063 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2064 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2065 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
2066 return ReplaceInstUsesWith(I, AndRHS);
2071 // If this is an integer sign or zero extension instruction.
2072 if (SrcTy->isIntegral() &&
2073 SrcTy->getPrimitiveSizeInBits() <
2074 CI->getType()->getPrimitiveSizeInBits()) {
2076 if (SrcTy->isUnsigned()) {
2077 // See if this and is clearing out bits that are known to be zero
2078 // anyway (due to the zero extension).
2079 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
2080 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
2081 Constant *Result = ConstantExpr::getAnd(Mask, AndRHS);
2082 if (Result == Mask) // The "and" isn't doing anything, remove it.
2083 return ReplaceInstUsesWith(I, CI);
2084 if (Result != AndRHS) { // Reduce the and RHS constant.
2085 I.setOperand(1, Result);
2090 if (CI->hasOneUse() && SrcTy->isInteger()) {
2091 // We can only do this if all of the sign bits brought in are masked
2092 // out. Compute this by first getting 0000011111, then inverting
2094 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy);
2095 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType());
2096 Mask = ConstantExpr::getNot(Mask); // 1's in the new bits.
2097 if (ConstantExpr::getAnd(Mask, AndRHS)->isNullValue()) {
2098 // If the and is clearing all of the sign bits, change this to a
2099 // zero extension cast. To do this, cast the cast input to
2100 // unsigned, then to the requested size.
2101 Value *CastOp = CI->getOperand(0);
2103 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
2104 CI->getName()+".uns");
2105 NC = InsertNewInstBefore(NC, I);
2106 // Finally, insert a replacement for CI.
2107 NC = new CastInst(NC, CI->getType(), CI->getName());
2109 NC = InsertNewInstBefore(NC, I);
2110 WorkList.push_back(CI); // Delete CI later.
2111 I.setOperand(0, NC);
2112 return &I; // The AND operand was modified.
2119 // Try to fold constant and into select arguments.
2120 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2121 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2123 if (isa<PHINode>(Op0))
2124 if (Instruction *NV = FoldOpIntoPhi(I))
2128 Value *Op0NotVal = dyn_castNotVal(Op0);
2129 Value *Op1NotVal = dyn_castNotVal(Op1);
2131 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
2132 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2134 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2135 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2136 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
2137 I.getName()+".demorgan");
2138 InsertNewInstBefore(Or, I);
2139 return BinaryOperator::createNot(Or);
2142 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
2143 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2144 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2147 Value *LHSVal, *RHSVal;
2148 ConstantInt *LHSCst, *RHSCst;
2149 Instruction::BinaryOps LHSCC, RHSCC;
2150 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2151 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2152 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
2153 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2154 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2155 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2156 // Ensure that the larger constant is on the RHS.
2157 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2158 SetCondInst *LHS = cast<SetCondInst>(Op0);
2159 if (cast<ConstantBool>(Cmp)->getValue()) {
2160 std::swap(LHS, RHS);
2161 std::swap(LHSCst, RHSCst);
2162 std::swap(LHSCC, RHSCC);
2165 // At this point, we know we have have two setcc instructions
2166 // comparing a value against two constants and and'ing the result
2167 // together. Because of the above check, we know that we only have
2168 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2169 // FoldSetCCLogical check above), that the two constants are not
2171 assert(LHSCst != RHSCst && "Compares not folded above?");
2174 default: assert(0 && "Unknown integer condition code!");
2175 case Instruction::SetEQ:
2177 default: assert(0 && "Unknown integer condition code!");
2178 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
2179 case Instruction::SetGT: // (X == 13 & X > 15) -> false
2180 return ReplaceInstUsesWith(I, ConstantBool::False);
2181 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
2182 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
2183 return ReplaceInstUsesWith(I, LHS);
2185 case Instruction::SetNE:
2187 default: assert(0 && "Unknown integer condition code!");
2188 case Instruction::SetLT:
2189 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
2190 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
2191 break; // (X != 13 & X < 15) -> no change
2192 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
2193 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
2194 return ReplaceInstUsesWith(I, RHS);
2195 case Instruction::SetNE:
2196 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
2197 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2198 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2199 LHSVal->getName()+".off");
2200 InsertNewInstBefore(Add, I);
2201 const Type *UnsType = Add->getType()->getUnsignedVersion();
2202 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2203 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
2204 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2205 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2207 break; // (X != 13 & X != 15) -> no change
2210 case Instruction::SetLT:
2212 default: assert(0 && "Unknown integer condition code!");
2213 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
2214 case Instruction::SetGT: // (X < 13 & X > 15) -> false
2215 return ReplaceInstUsesWith(I, ConstantBool::False);
2216 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
2217 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
2218 return ReplaceInstUsesWith(I, LHS);
2220 case Instruction::SetGT:
2222 default: assert(0 && "Unknown integer condition code!");
2223 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
2224 return ReplaceInstUsesWith(I, LHS);
2225 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
2226 return ReplaceInstUsesWith(I, RHS);
2227 case Instruction::SetNE:
2228 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
2229 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
2230 break; // (X > 13 & X != 15) -> no change
2231 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
2232 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2238 return Changed ? &I : 0;
2241 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2242 bool Changed = SimplifyCommutative(I);
2243 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2245 if (isa<UndefValue>(Op1))
2246 return ReplaceInstUsesWith(I, // X | undef -> -1
2247 ConstantIntegral::getAllOnesValue(I.getType()));
2249 // or X, X = X or X, 0 == X
2250 if (Op0 == Op1 || Op1 == Constant::getNullValue(I.getType()))
2251 return ReplaceInstUsesWith(I, Op0);
2254 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2255 // If X is known to only contain bits that already exist in RHS, just
2256 // replace this instruction with RHS directly.
2257 if (MaskedValueIsZero(Op0,
2258 RHS->getZExtValue()^RHS->getType()->getIntegralTypeMask()))
2259 return ReplaceInstUsesWith(I, RHS);
2261 ConstantInt *C1 = 0; Value *X = 0;
2262 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2263 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2264 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
2266 InsertNewInstBefore(Or, I);
2267 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
2270 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2271 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2272 std::string Op0Name = Op0->getName(); Op0->setName("");
2273 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
2274 InsertNewInstBefore(Or, I);
2275 return BinaryOperator::createXor(Or,
2276 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
2279 // Try to fold constant and into select arguments.
2280 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2281 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2283 if (isa<PHINode>(Op0))
2284 if (Instruction *NV = FoldOpIntoPhi(I))
2288 Value *A = 0, *B = 0;
2289 ConstantInt *C1 = 0, *C2 = 0;
2291 if (match(Op0, m_And(m_Value(A), m_Value(B))))
2292 if (A == Op1 || B == Op1) // (A & ?) | A --> A
2293 return ReplaceInstUsesWith(I, Op1);
2294 if (match(Op1, m_And(m_Value(A), m_Value(B))))
2295 if (A == Op0 || B == Op0) // A | (A & ?) --> A
2296 return ReplaceInstUsesWith(I, Op0);
2298 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2299 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2300 MaskedValueIsZero(Op1, C1->getZExtValue())) {
2301 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
2303 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2306 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2307 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2308 MaskedValueIsZero(Op0, C1->getZExtValue())) {
2309 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
2311 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2314 // (A & C1)|(B & C2)
2315 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
2316 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
2318 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
2319 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
2322 // If we have: ((V + N) & C1) | (V & C2)
2323 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2324 // replace with V+N.
2325 if (C1 == ConstantExpr::getNot(C2)) {
2326 Value *V1 = 0, *V2 = 0;
2327 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
2328 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
2329 // Add commutes, try both ways.
2330 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
2331 return ReplaceInstUsesWith(I, A);
2332 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
2333 return ReplaceInstUsesWith(I, A);
2335 // Or commutes, try both ways.
2336 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
2337 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
2338 // Add commutes, try both ways.
2339 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
2340 return ReplaceInstUsesWith(I, B);
2341 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
2342 return ReplaceInstUsesWith(I, B);
2347 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
2348 if (A == Op1) // ~A | A == -1
2349 return ReplaceInstUsesWith(I,
2350 ConstantIntegral::getAllOnesValue(I.getType()));
2354 // Note, A is still live here!
2355 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
2357 return ReplaceInstUsesWith(I,
2358 ConstantIntegral::getAllOnesValue(I.getType()));
2360 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2361 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2362 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
2363 I.getName()+".demorgan"), I);
2364 return BinaryOperator::createNot(And);
2368 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
2369 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
2370 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2373 Value *LHSVal, *RHSVal;
2374 ConstantInt *LHSCst, *RHSCst;
2375 Instruction::BinaryOps LHSCC, RHSCC;
2376 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2377 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2378 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
2379 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2380 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2381 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2382 // Ensure that the larger constant is on the RHS.
2383 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2384 SetCondInst *LHS = cast<SetCondInst>(Op0);
2385 if (cast<ConstantBool>(Cmp)->getValue()) {
2386 std::swap(LHS, RHS);
2387 std::swap(LHSCst, RHSCst);
2388 std::swap(LHSCC, RHSCC);
2391 // At this point, we know we have have two setcc instructions
2392 // comparing a value against two constants and or'ing the result
2393 // together. Because of the above check, we know that we only have
2394 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2395 // FoldSetCCLogical check above), that the two constants are not
2397 assert(LHSCst != RHSCst && "Compares not folded above?");
2400 default: assert(0 && "Unknown integer condition code!");
2401 case Instruction::SetEQ:
2403 default: assert(0 && "Unknown integer condition code!");
2404 case Instruction::SetEQ:
2405 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
2406 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2407 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2408 LHSVal->getName()+".off");
2409 InsertNewInstBefore(Add, I);
2410 const Type *UnsType = Add->getType()->getUnsignedVersion();
2411 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2412 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
2413 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2414 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2416 break; // (X == 13 | X == 15) -> no change
2418 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
2420 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
2421 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
2422 return ReplaceInstUsesWith(I, RHS);
2425 case Instruction::SetNE:
2427 default: assert(0 && "Unknown integer condition code!");
2428 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
2429 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
2430 return ReplaceInstUsesWith(I, LHS);
2431 case Instruction::SetNE: // (X != 13 | X != 15) -> true
2432 case Instruction::SetLT: // (X != 13 | X < 15) -> true
2433 return ReplaceInstUsesWith(I, ConstantBool::True);
2436 case Instruction::SetLT:
2438 default: assert(0 && "Unknown integer condition code!");
2439 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
2441 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
2442 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
2443 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
2444 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
2445 return ReplaceInstUsesWith(I, RHS);
2448 case Instruction::SetGT:
2450 default: assert(0 && "Unknown integer condition code!");
2451 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
2452 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
2453 return ReplaceInstUsesWith(I, LHS);
2454 case Instruction::SetNE: // (X > 13 | X != 15) -> true
2455 case Instruction::SetLT: // (X > 13 | X < 15) -> true
2456 return ReplaceInstUsesWith(I, ConstantBool::True);
2462 return Changed ? &I : 0;
2465 // XorSelf - Implements: X ^ X --> 0
2468 XorSelf(Value *rhs) : RHS(rhs) {}
2469 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2470 Instruction *apply(BinaryOperator &Xor) const {
2476 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2477 bool Changed = SimplifyCommutative(I);
2478 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2480 if (isa<UndefValue>(Op1))
2481 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
2483 // xor X, X = 0, even if X is nested in a sequence of Xor's.
2484 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
2485 assert(Result == &I && "AssociativeOpt didn't work?");
2486 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2489 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2491 if (RHS->isNullValue())
2492 return ReplaceInstUsesWith(I, Op0);
2494 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2495 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
2496 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
2497 if (RHS == ConstantBool::True && SCI->hasOneUse())
2498 return new SetCondInst(SCI->getInverseCondition(),
2499 SCI->getOperand(0), SCI->getOperand(1));
2501 // ~(c-X) == X-c-1 == X+(-c-1)
2502 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2503 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2504 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2505 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2506 ConstantInt::get(I.getType(), 1));
2507 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
2510 // ~(~X & Y) --> (X | ~Y)
2511 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
2512 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
2513 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2515 BinaryOperator::createNot(Op0I->getOperand(1),
2516 Op0I->getOperand(1)->getName()+".not");
2517 InsertNewInstBefore(NotY, I);
2518 return BinaryOperator::createOr(Op0NotVal, NotY);
2522 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2523 switch (Op0I->getOpcode()) {
2524 case Instruction::Add:
2525 // ~(X-c) --> (-c-1)-X
2526 if (RHS->isAllOnesValue()) {
2527 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2528 return BinaryOperator::createSub(
2529 ConstantExpr::getSub(NegOp0CI,
2530 ConstantInt::get(I.getType(), 1)),
2531 Op0I->getOperand(0));
2534 case Instruction::And:
2535 // (X & C1) ^ C2 --> (X & C1) | C2 iff (C1&C2) == 0
2536 if (ConstantExpr::getAnd(RHS, Op0CI)->isNullValue())
2537 return BinaryOperator::createOr(Op0, RHS);
2539 case Instruction::Or:
2540 // (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
2541 if (ConstantExpr::getAnd(RHS, Op0CI) == RHS)
2542 return BinaryOperator::createAnd(Op0, ConstantExpr::getNot(RHS));
2548 // Try to fold constant and into select arguments.
2549 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2550 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2552 if (isa<PHINode>(Op0))
2553 if (Instruction *NV = FoldOpIntoPhi(I))
2557 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
2559 return ReplaceInstUsesWith(I,
2560 ConstantIntegral::getAllOnesValue(I.getType()));
2562 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
2564 return ReplaceInstUsesWith(I,
2565 ConstantIntegral::getAllOnesValue(I.getType()));
2567 if (Instruction *Op1I = dyn_cast<Instruction>(Op1))
2568 if (Op1I->getOpcode() == Instruction::Or) {
2569 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
2570 cast<BinaryOperator>(Op1I)->swapOperands();
2572 std::swap(Op0, Op1);
2573 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
2575 std::swap(Op0, Op1);
2577 } else if (Op1I->getOpcode() == Instruction::Xor) {
2578 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
2579 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
2580 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
2581 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
2584 if (Instruction *Op0I = dyn_cast<Instruction>(Op0))
2585 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
2586 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
2587 cast<BinaryOperator>(Op0I)->swapOperands();
2588 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
2589 Value *NotB = InsertNewInstBefore(BinaryOperator::createNot(Op1,
2590 Op1->getName()+".not"), I);
2591 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
2593 } else if (Op0I->getOpcode() == Instruction::Xor) {
2594 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
2595 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2596 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
2597 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2600 // (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
2601 ConstantInt *C1 = 0, *C2 = 0;
2602 if (match(Op0, m_And(m_Value(), m_ConstantInt(C1))) &&
2603 match(Op1, m_And(m_Value(), m_ConstantInt(C2))) &&
2604 ConstantExpr::getAnd(C1, C2)->isNullValue())
2605 return BinaryOperator::createOr(Op0, Op1);
2607 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
2608 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
2609 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2612 return Changed ? &I : 0;
2615 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
2616 /// overflowed for this type.
2617 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2619 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
2620 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
2623 static bool isPositive(ConstantInt *C) {
2624 return cast<ConstantSInt>(C)->getValue() >= 0;
2627 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
2628 /// overflowed for this type.
2629 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2631 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
2633 if (In1->getType()->isUnsigned())
2634 return cast<ConstantUInt>(Result)->getValue() <
2635 cast<ConstantUInt>(In1)->getValue();
2636 if (isPositive(In1) != isPositive(In2))
2638 if (isPositive(In1))
2639 return cast<ConstantSInt>(Result)->getValue() <
2640 cast<ConstantSInt>(In1)->getValue();
2641 return cast<ConstantSInt>(Result)->getValue() >
2642 cast<ConstantSInt>(In1)->getValue();
2645 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
2646 /// code necessary to compute the offset from the base pointer (without adding
2647 /// in the base pointer). Return the result as a signed integer of intptr size.
2648 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
2649 TargetData &TD = IC.getTargetData();
2650 gep_type_iterator GTI = gep_type_begin(GEP);
2651 const Type *UIntPtrTy = TD.getIntPtrType();
2652 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
2653 Value *Result = Constant::getNullValue(SIntPtrTy);
2655 // Build a mask for high order bits.
2656 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
2658 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
2659 Value *Op = GEP->getOperand(i);
2660 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
2661 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
2663 if (Constant *OpC = dyn_cast<Constant>(Op)) {
2664 if (!OpC->isNullValue()) {
2665 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
2666 Scale = ConstantExpr::getMul(OpC, Scale);
2667 if (Constant *RC = dyn_cast<Constant>(Result))
2668 Result = ConstantExpr::getAdd(RC, Scale);
2670 // Emit an add instruction.
2671 Result = IC.InsertNewInstBefore(
2672 BinaryOperator::createAdd(Result, Scale,
2673 GEP->getName()+".offs"), I);
2677 // Convert to correct type.
2678 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
2679 Op->getName()+".c"), I);
2681 // We'll let instcombine(mul) convert this to a shl if possible.
2682 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
2683 GEP->getName()+".idx"), I);
2685 // Emit an add instruction.
2686 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
2687 GEP->getName()+".offs"), I);
2693 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
2694 /// else. At this point we know that the GEP is on the LHS of the comparison.
2695 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
2696 Instruction::BinaryOps Cond,
2698 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
2700 if (CastInst *CI = dyn_cast<CastInst>(RHS))
2701 if (isa<PointerType>(CI->getOperand(0)->getType()))
2702 RHS = CI->getOperand(0);
2704 Value *PtrBase = GEPLHS->getOperand(0);
2705 if (PtrBase == RHS) {
2706 // As an optimization, we don't actually have to compute the actual value of
2707 // OFFSET if this is a seteq or setne comparison, just return whether each
2708 // index is zero or not.
2709 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
2710 Instruction *InVal = 0;
2711 gep_type_iterator GTI = gep_type_begin(GEPLHS);
2712 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
2714 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
2715 if (isa<UndefValue>(C)) // undef index -> undef.
2716 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2717 if (C->isNullValue())
2719 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
2720 EmitIt = false; // This is indexing into a zero sized array?
2721 } else if (isa<ConstantInt>(C))
2722 return ReplaceInstUsesWith(I, // No comparison is needed here.
2723 ConstantBool::get(Cond == Instruction::SetNE));
2728 new SetCondInst(Cond, GEPLHS->getOperand(i),
2729 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
2733 InVal = InsertNewInstBefore(InVal, I);
2734 InsertNewInstBefore(Comp, I);
2735 if (Cond == Instruction::SetNE) // True if any are unequal
2736 InVal = BinaryOperator::createOr(InVal, Comp);
2737 else // True if all are equal
2738 InVal = BinaryOperator::createAnd(InVal, Comp);
2746 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
2747 ConstantBool::get(Cond == Instruction::SetEQ));
2750 // Only lower this if the setcc is the only user of the GEP or if we expect
2751 // the result to fold to a constant!
2752 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
2753 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
2754 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
2755 return new SetCondInst(Cond, Offset,
2756 Constant::getNullValue(Offset->getType()));
2758 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
2759 // If the base pointers are different, but the indices are the same, just
2760 // compare the base pointer.
2761 if (PtrBase != GEPRHS->getOperand(0)) {
2762 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
2763 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
2764 GEPRHS->getOperand(0)->getType();
2766 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2767 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2768 IndicesTheSame = false;
2772 // If all indices are the same, just compare the base pointers.
2774 return new SetCondInst(Cond, GEPLHS->getOperand(0),
2775 GEPRHS->getOperand(0));
2777 // Otherwise, the base pointers are different and the indices are
2778 // different, bail out.
2782 // If one of the GEPs has all zero indices, recurse.
2783 bool AllZeros = true;
2784 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
2785 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
2786 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
2791 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
2792 SetCondInst::getSwappedCondition(Cond), I);
2794 // If the other GEP has all zero indices, recurse.
2796 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2797 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
2798 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
2803 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
2805 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
2806 // If the GEPs only differ by one index, compare it.
2807 unsigned NumDifferences = 0; // Keep track of # differences.
2808 unsigned DiffOperand = 0; // The operand that differs.
2809 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
2810 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
2811 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
2812 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
2813 // Irreconcilable differences.
2817 if (NumDifferences++) break;
2822 if (NumDifferences == 0) // SAME GEP?
2823 return ReplaceInstUsesWith(I, // No comparison is needed here.
2824 ConstantBool::get(Cond == Instruction::SetEQ));
2825 else if (NumDifferences == 1) {
2826 Value *LHSV = GEPLHS->getOperand(DiffOperand);
2827 Value *RHSV = GEPRHS->getOperand(DiffOperand);
2829 // Convert the operands to signed values to make sure to perform a
2830 // signed comparison.
2831 const Type *NewTy = LHSV->getType()->getSignedVersion();
2832 if (LHSV->getType() != NewTy)
2833 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
2834 LHSV->getName()), I);
2835 if (RHSV->getType() != NewTy)
2836 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
2837 RHSV->getName()), I);
2838 return new SetCondInst(Cond, LHSV, RHSV);
2842 // Only lower this if the setcc is the only user of the GEP or if we expect
2843 // the result to fold to a constant!
2844 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
2845 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
2846 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
2847 Value *L = EmitGEPOffset(GEPLHS, I, *this);
2848 Value *R = EmitGEPOffset(GEPRHS, I, *this);
2849 return new SetCondInst(Cond, L, R);
2856 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
2857 bool Changed = SimplifyCommutative(I);
2858 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2859 const Type *Ty = Op0->getType();
2863 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
2865 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
2866 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
2868 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
2869 // addresses never equal each other! We already know that Op0 != Op1.
2870 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
2871 isa<ConstantPointerNull>(Op0)) &&
2872 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
2873 isa<ConstantPointerNull>(Op1)))
2874 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
2876 // setcc's with boolean values can always be turned into bitwise operations
2877 if (Ty == Type::BoolTy) {
2878 switch (I.getOpcode()) {
2879 default: assert(0 && "Invalid setcc instruction!");
2880 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
2881 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
2882 InsertNewInstBefore(Xor, I);
2883 return BinaryOperator::createNot(Xor);
2885 case Instruction::SetNE:
2886 return BinaryOperator::createXor(Op0, Op1);
2888 case Instruction::SetGT:
2889 std::swap(Op0, Op1); // Change setgt -> setlt
2891 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
2892 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2893 InsertNewInstBefore(Not, I);
2894 return BinaryOperator::createAnd(Not, Op1);
2896 case Instruction::SetGE:
2897 std::swap(Op0, Op1); // Change setge -> setle
2899 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
2900 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
2901 InsertNewInstBefore(Not, I);
2902 return BinaryOperator::createOr(Not, Op1);
2907 // See if we are doing a comparison between a constant and an instruction that
2908 // can be folded into the comparison.
2909 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2910 // Check to see if we are comparing against the minimum or maximum value...
2911 if (CI->isMinValue()) {
2912 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
2913 return ReplaceInstUsesWith(I, ConstantBool::False);
2914 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
2915 return ReplaceInstUsesWith(I, ConstantBool::True);
2916 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
2917 return BinaryOperator::createSetEQ(Op0, Op1);
2918 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
2919 return BinaryOperator::createSetNE(Op0, Op1);
2921 } else if (CI->isMaxValue()) {
2922 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
2923 return ReplaceInstUsesWith(I, ConstantBool::False);
2924 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
2925 return ReplaceInstUsesWith(I, ConstantBool::True);
2926 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
2927 return BinaryOperator::createSetEQ(Op0, Op1);
2928 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
2929 return BinaryOperator::createSetNE(Op0, Op1);
2931 // Comparing against a value really close to min or max?
2932 } else if (isMinValuePlusOne(CI)) {
2933 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
2934 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
2935 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
2936 return BinaryOperator::createSetNE(Op0, SubOne(CI));
2938 } else if (isMaxValueMinusOne(CI)) {
2939 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
2940 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
2941 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
2942 return BinaryOperator::createSetNE(Op0, AddOne(CI));
2945 // If we still have a setle or setge instruction, turn it into the
2946 // appropriate setlt or setgt instruction. Since the border cases have
2947 // already been handled above, this requires little checking.
2949 if (I.getOpcode() == Instruction::SetLE)
2950 return BinaryOperator::createSetLT(Op0, AddOne(CI));
2951 if (I.getOpcode() == Instruction::SetGE)
2952 return BinaryOperator::createSetGT(Op0, SubOne(CI));
2954 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2955 switch (LHSI->getOpcode()) {
2956 case Instruction::And:
2957 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
2958 LHSI->getOperand(0)->hasOneUse()) {
2959 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
2960 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
2961 // happens a LOT in code produced by the C front-end, for bitfield
2963 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
2964 ConstantUInt *ShAmt;
2965 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
2966 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
2967 const Type *Ty = LHSI->getType();
2969 // We can fold this as long as we can't shift unknown bits
2970 // into the mask. This can only happen with signed shift
2971 // rights, as they sign-extend.
2973 bool CanFold = Shift->getOpcode() != Instruction::Shr ||
2974 Shift->getType()->isUnsigned();
2976 // To test for the bad case of the signed shr, see if any
2977 // of the bits shifted in could be tested after the mask.
2978 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
2979 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
2981 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
2983 ConstantExpr::getShl(ConstantInt::getAllOnesValue(Ty), OShAmt);
2984 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
2990 if (Shift->getOpcode() == Instruction::Shl)
2991 NewCst = ConstantExpr::getUShr(CI, ShAmt);
2993 NewCst = ConstantExpr::getShl(CI, ShAmt);
2995 // Check to see if we are shifting out any of the bits being
2997 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
2998 // If we shifted bits out, the fold is not going to work out.
2999 // As a special case, check to see if this means that the
3000 // result is always true or false now.
3001 if (I.getOpcode() == Instruction::SetEQ)
3002 return ReplaceInstUsesWith(I, ConstantBool::False);
3003 if (I.getOpcode() == Instruction::SetNE)
3004 return ReplaceInstUsesWith(I, ConstantBool::True);
3006 I.setOperand(1, NewCst);
3007 Constant *NewAndCST;
3008 if (Shift->getOpcode() == Instruction::Shl)
3009 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
3011 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
3012 LHSI->setOperand(1, NewAndCST);
3013 LHSI->setOperand(0, Shift->getOperand(0));
3014 WorkList.push_back(Shift); // Shift is dead.
3015 AddUsesToWorkList(I);
3023 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
3024 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3025 switch (I.getOpcode()) {
3027 case Instruction::SetEQ:
3028 case Instruction::SetNE: {
3029 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3031 // Check that the shift amount is in range. If not, don't perform
3032 // undefined shifts. When the shift is visited it will be
3034 if (ShAmt->getValue() >= TypeBits)
3037 // If we are comparing against bits always shifted out, the
3038 // comparison cannot succeed.
3040 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
3041 if (Comp != CI) {// Comparing against a bit that we know is zero.
3042 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3043 Constant *Cst = ConstantBool::get(IsSetNE);
3044 return ReplaceInstUsesWith(I, Cst);
3047 if (LHSI->hasOneUse()) {
3048 // Otherwise strength reduce the shift into an and.
3049 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3050 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
3053 if (CI->getType()->isUnsigned()) {
3054 Mask = ConstantUInt::get(CI->getType(), Val);
3055 } else if (ShAmtVal != 0) {
3056 Mask = ConstantSInt::get(CI->getType(), Val);
3058 Mask = ConstantInt::getAllOnesValue(CI->getType());
3062 BinaryOperator::createAnd(LHSI->getOperand(0),
3063 Mask, LHSI->getName()+".mask");
3064 Value *And = InsertNewInstBefore(AndI, I);
3065 return new SetCondInst(I.getOpcode(), And,
3066 ConstantExpr::getUShr(CI, ShAmt));
3073 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
3074 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3075 switch (I.getOpcode()) {
3077 case Instruction::SetEQ:
3078 case Instruction::SetNE: {
3080 // Check that the shift amount is in range. If not, don't perform
3081 // undefined shifts. When the shift is visited it will be
3083 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3084 if (ShAmt->getValue() >= TypeBits)
3087 // If we are comparing against bits always shifted out, the
3088 // comparison cannot succeed.
3090 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
3092 if (Comp != CI) {// Comparing against a bit that we know is zero.
3093 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3094 Constant *Cst = ConstantBool::get(IsSetNE);
3095 return ReplaceInstUsesWith(I, Cst);
3098 if (LHSI->hasOneUse() || CI->isNullValue()) {
3099 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3101 // Otherwise strength reduce the shift into an and.
3102 uint64_t Val = ~0ULL; // All ones.
3103 Val <<= ShAmtVal; // Shift over to the right spot.
3106 if (CI->getType()->isUnsigned()) {
3107 Val &= ~0ULL >> (64-TypeBits);
3108 Mask = ConstantUInt::get(CI->getType(), Val);
3110 Mask = ConstantSInt::get(CI->getType(), Val);
3114 BinaryOperator::createAnd(LHSI->getOperand(0),
3115 Mask, LHSI->getName()+".mask");
3116 Value *And = InsertNewInstBefore(AndI, I);
3117 return new SetCondInst(I.getOpcode(), And,
3118 ConstantExpr::getShl(CI, ShAmt));
3126 case Instruction::Div:
3127 // Fold: (div X, C1) op C2 -> range check
3128 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
3129 // Fold this div into the comparison, producing a range check.
3130 // Determine, based on the divide type, what the range is being
3131 // checked. If there is an overflow on the low or high side, remember
3132 // it, otherwise compute the range [low, hi) bounding the new value.
3133 bool LoOverflow = false, HiOverflow = 0;
3134 ConstantInt *LoBound = 0, *HiBound = 0;
3137 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
3139 Instruction::BinaryOps Opcode = I.getOpcode();
3141 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
3142 } else if (LHSI->getType()->isUnsigned()) { // udiv
3144 LoOverflow = ProdOV;
3145 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
3146 } else if (isPositive(DivRHS)) { // Divisor is > 0.
3147 if (CI->isNullValue()) { // (X / pos) op 0
3149 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
3151 } else if (isPositive(CI)) { // (X / pos) op pos
3153 LoOverflow = ProdOV;
3154 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
3155 } else { // (X / pos) op neg
3156 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
3157 LoOverflow = AddWithOverflow(LoBound, Prod,
3158 cast<ConstantInt>(DivRHSH));
3160 HiOverflow = ProdOV;
3162 } else { // Divisor is < 0.
3163 if (CI->isNullValue()) { // (X / neg) op 0
3164 LoBound = AddOne(DivRHS);
3165 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
3166 if (HiBound == DivRHS)
3167 LoBound = 0; // - INTMIN = INTMIN
3168 } else if (isPositive(CI)) { // (X / neg) op pos
3169 HiOverflow = LoOverflow = ProdOV;
3171 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
3172 HiBound = AddOne(Prod);
3173 } else { // (X / neg) op neg
3175 LoOverflow = HiOverflow = ProdOV;
3176 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
3179 // Dividing by a negate swaps the condition.
3180 Opcode = SetCondInst::getSwappedCondition(Opcode);
3184 Value *X = LHSI->getOperand(0);
3186 default: assert(0 && "Unhandled setcc opcode!");
3187 case Instruction::SetEQ:
3188 if (LoOverflow && HiOverflow)
3189 return ReplaceInstUsesWith(I, ConstantBool::False);
3190 else if (HiOverflow)
3191 return new SetCondInst(Instruction::SetGE, X, LoBound);
3192 else if (LoOverflow)
3193 return new SetCondInst(Instruction::SetLT, X, HiBound);
3195 return InsertRangeTest(X, LoBound, HiBound, true, I);
3196 case Instruction::SetNE:
3197 if (LoOverflow && HiOverflow)
3198 return ReplaceInstUsesWith(I, ConstantBool::True);
3199 else if (HiOverflow)
3200 return new SetCondInst(Instruction::SetLT, X, LoBound);
3201 else if (LoOverflow)
3202 return new SetCondInst(Instruction::SetGE, X, HiBound);
3204 return InsertRangeTest(X, LoBound, HiBound, false, I);
3205 case Instruction::SetLT:
3207 return ReplaceInstUsesWith(I, ConstantBool::False);
3208 return new SetCondInst(Instruction::SetLT, X, LoBound);
3209 case Instruction::SetGT:
3211 return ReplaceInstUsesWith(I, ConstantBool::False);
3212 return new SetCondInst(Instruction::SetGE, X, HiBound);
3219 // Simplify seteq and setne instructions...
3220 if (I.getOpcode() == Instruction::SetEQ ||
3221 I.getOpcode() == Instruction::SetNE) {
3222 bool isSetNE = I.getOpcode() == Instruction::SetNE;
3224 // If the first operand is (and|or|xor) with a constant, and the second
3225 // operand is a constant, simplify a bit.
3226 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
3227 switch (BO->getOpcode()) {
3228 case Instruction::Rem:
3229 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3230 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
3232 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
3233 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
3234 if (isPowerOf2_64(V)) {
3235 unsigned L2 = Log2_64(V);
3236 const Type *UTy = BO->getType()->getUnsignedVersion();
3237 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
3239 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
3240 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
3241 RHSCst, BO->getName()), I);
3242 return BinaryOperator::create(I.getOpcode(), NewRem,
3243 Constant::getNullValue(UTy));
3248 case Instruction::Add:
3249 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3250 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3251 if (BO->hasOneUse())
3252 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3253 ConstantExpr::getSub(CI, BOp1C));
3254 } else if (CI->isNullValue()) {
3255 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3256 // efficiently invertible, or if the add has just this one use.
3257 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3259 if (Value *NegVal = dyn_castNegVal(BOp1))
3260 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
3261 else if (Value *NegVal = dyn_castNegVal(BOp0))
3262 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
3263 else if (BO->hasOneUse()) {
3264 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
3266 InsertNewInstBefore(Neg, I);
3267 return new SetCondInst(I.getOpcode(), BOp0, Neg);
3271 case Instruction::Xor:
3272 // For the xor case, we can xor two constants together, eliminating
3273 // the explicit xor.
3274 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
3275 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
3276 ConstantExpr::getXor(CI, BOC));
3279 case Instruction::Sub:
3280 // Replace (([sub|xor] A, B) != 0) with (A != B)
3281 if (CI->isNullValue())
3282 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3286 case Instruction::Or:
3287 // If bits are being or'd in that are not present in the constant we
3288 // are comparing against, then the comparison could never succeed!
3289 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
3290 Constant *NotCI = ConstantExpr::getNot(CI);
3291 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
3292 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3296 case Instruction::And:
3297 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3298 // If bits are being compared against that are and'd out, then the
3299 // comparison can never succeed!
3300 if (!ConstantExpr::getAnd(CI,
3301 ConstantExpr::getNot(BOC))->isNullValue())
3302 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3304 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3305 if (CI == BOC && isOneBitSet(CI))
3306 return new SetCondInst(isSetNE ? Instruction::SetEQ :
3307 Instruction::SetNE, Op0,
3308 Constant::getNullValue(CI->getType()));
3310 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
3311 // to be a signed value as appropriate.
3312 if (isSignBit(BOC)) {
3313 Value *X = BO->getOperand(0);
3314 // If 'X' is not signed, insert a cast now...
3315 if (!BOC->getType()->isSigned()) {
3316 const Type *DestTy = BOC->getType()->getSignedVersion();
3317 X = InsertCastBefore(X, DestTy, I);
3319 return new SetCondInst(isSetNE ? Instruction::SetLT :
3320 Instruction::SetGE, X,
3321 Constant::getNullValue(X->getType()));
3324 // ((X & ~7) == 0) --> X < 8
3325 if (CI->isNullValue() && isHighOnes(BOC)) {
3326 Value *X = BO->getOperand(0);
3327 Constant *NegX = ConstantExpr::getNeg(BOC);
3329 // If 'X' is signed, insert a cast now.
3330 if (NegX->getType()->isSigned()) {
3331 const Type *DestTy = NegX->getType()->getUnsignedVersion();
3332 X = InsertCastBefore(X, DestTy, I);
3333 NegX = ConstantExpr::getCast(NegX, DestTy);
3336 return new SetCondInst(isSetNE ? Instruction::SetGE :
3337 Instruction::SetLT, X, NegX);
3344 } else { // Not a SetEQ/SetNE
3345 // If the LHS is a cast from an integral value of the same size,
3346 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
3347 Value *CastOp = Cast->getOperand(0);
3348 const Type *SrcTy = CastOp->getType();
3349 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
3350 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
3351 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
3352 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
3353 "Source and destination signednesses should differ!");
3354 if (Cast->getType()->isSigned()) {
3355 // If this is a signed comparison, check for comparisons in the
3356 // vicinity of zero.
3357 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
3359 return BinaryOperator::createSetGT(CastOp,
3360 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
3361 else if (I.getOpcode() == Instruction::SetGT &&
3362 cast<ConstantSInt>(CI)->getValue() == -1)
3363 // X > -1 => x < 128
3364 return BinaryOperator::createSetLT(CastOp,
3365 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
3367 ConstantUInt *CUI = cast<ConstantUInt>(CI);
3368 if (I.getOpcode() == Instruction::SetLT &&
3369 CUI->getValue() == 1ULL << (SrcTySize-1))
3370 // X < 128 => X > -1
3371 return BinaryOperator::createSetGT(CastOp,
3372 ConstantSInt::get(SrcTy, -1));
3373 else if (I.getOpcode() == Instruction::SetGT &&
3374 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
3376 return BinaryOperator::createSetLT(CastOp,
3377 Constant::getNullValue(SrcTy));
3384 // Handle setcc with constant RHS's that can be integer, FP or pointer.
3385 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3386 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3387 switch (LHSI->getOpcode()) {
3388 case Instruction::GetElementPtr:
3389 if (RHSC->isNullValue()) {
3390 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
3391 bool isAllZeros = true;
3392 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
3393 if (!isa<Constant>(LHSI->getOperand(i)) ||
3394 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
3399 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
3400 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3404 case Instruction::PHI:
3405 if (Instruction *NV = FoldOpIntoPhi(I))
3408 case Instruction::Select:
3409 // If either operand of the select is a constant, we can fold the
3410 // comparison into the select arms, which will cause one to be
3411 // constant folded and the select turned into a bitwise or.
3412 Value *Op1 = 0, *Op2 = 0;
3413 if (LHSI->hasOneUse()) {
3414 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3415 // Fold the known value into the constant operand.
3416 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3417 // Insert a new SetCC of the other select operand.
3418 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3419 LHSI->getOperand(2), RHSC,
3421 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3422 // Fold the known value into the constant operand.
3423 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3424 // Insert a new SetCC of the other select operand.
3425 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3426 LHSI->getOperand(1), RHSC,
3432 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
3437 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
3438 if (User *GEP = dyn_castGetElementPtr(Op0))
3439 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
3441 if (User *GEP = dyn_castGetElementPtr(Op1))
3442 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
3443 SetCondInst::getSwappedCondition(I.getOpcode()), I))
3446 // Test to see if the operands of the setcc are casted versions of other
3447 // values. If the cast can be stripped off both arguments, we do so now.
3448 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3449 Value *CastOp0 = CI->getOperand(0);
3450 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
3451 (isa<Constant>(Op1) || isa<CastInst>(Op1)) &&
3452 (I.getOpcode() == Instruction::SetEQ ||
3453 I.getOpcode() == Instruction::SetNE)) {
3454 // We keep moving the cast from the left operand over to the right
3455 // operand, where it can often be eliminated completely.
3458 // If operand #1 is a cast instruction, see if we can eliminate it as
3460 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
3461 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
3463 Op1 = CI2->getOperand(0);
3465 // If Op1 is a constant, we can fold the cast into the constant.
3466 if (Op1->getType() != Op0->getType())
3467 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3468 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
3470 // Otherwise, cast the RHS right before the setcc
3471 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
3472 InsertNewInstBefore(cast<Instruction>(Op1), I);
3474 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
3477 // Handle the special case of: setcc (cast bool to X), <cst>
3478 // This comes up when you have code like
3481 // For generality, we handle any zero-extension of any operand comparison
3482 // with a constant or another cast from the same type.
3483 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
3484 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
3487 return Changed ? &I : 0;
3490 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
3491 // We only handle extending casts so far.
3493 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
3494 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
3495 const Type *SrcTy = LHSCIOp->getType();
3496 const Type *DestTy = SCI.getOperand(0)->getType();
3499 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
3502 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
3503 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
3504 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
3506 // Is this a sign or zero extension?
3507 bool isSignSrc = SrcTy->isSigned();
3508 bool isSignDest = DestTy->isSigned();
3510 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
3511 // Not an extension from the same type?
3512 RHSCIOp = CI->getOperand(0);
3513 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
3514 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
3515 // Compute the constant that would happen if we truncated to SrcTy then
3516 // reextended to DestTy.
3517 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
3519 if (ConstantExpr::getCast(Res, DestTy) == CI) {
3522 // If the value cannot be represented in the shorter type, we cannot emit
3523 // a simple comparison.
3524 if (SCI.getOpcode() == Instruction::SetEQ)
3525 return ReplaceInstUsesWith(SCI, ConstantBool::False);
3526 if (SCI.getOpcode() == Instruction::SetNE)
3527 return ReplaceInstUsesWith(SCI, ConstantBool::True);
3529 // Evaluate the comparison for LT.
3531 if (DestTy->isSigned()) {
3532 // We're performing a signed comparison.
3534 // Signed extend and signed comparison.
3535 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
3536 Result = ConstantBool::False;
3538 Result = ConstantBool::True; // X < (large) --> true
3540 // Unsigned extend and signed comparison.
3541 if (cast<ConstantSInt>(CI)->getValue() < 0)
3542 Result = ConstantBool::False;
3544 Result = ConstantBool::True;
3547 // We're performing an unsigned comparison.
3549 // Unsigned extend & compare -> always true.
3550 Result = ConstantBool::True;
3552 // We're performing an unsigned comp with a sign extended value.
3553 // This is true if the input is >= 0. [aka >s -1]
3554 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
3555 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
3556 NegOne, SCI.getName()), SCI);
3560 // Finally, return the value computed.
3561 if (SCI.getOpcode() == Instruction::SetLT) {
3562 return ReplaceInstUsesWith(SCI, Result);
3564 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
3565 if (Constant *CI = dyn_cast<Constant>(Result))
3566 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
3568 return BinaryOperator::createNot(Result);
3575 // Okay, just insert a compare of the reduced operands now!
3576 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
3579 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
3580 assert(I.getOperand(1)->getType() == Type::UByteTy);
3581 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3582 bool isLeftShift = I.getOpcode() == Instruction::Shl;
3584 // shl X, 0 == X and shr X, 0 == X
3585 // shl 0, X == 0 and shr 0, X == 0
3586 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
3587 Op0 == Constant::getNullValue(Op0->getType()))
3588 return ReplaceInstUsesWith(I, Op0);
3590 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
3591 if (!isLeftShift && I.getType()->isSigned())
3592 return ReplaceInstUsesWith(I, Op0);
3593 else // undef << X -> 0 AND undef >>u X -> 0
3594 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3596 if (isa<UndefValue>(Op1)) {
3597 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
3598 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3600 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
3603 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
3605 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
3606 if (CSI->isAllOnesValue())
3607 return ReplaceInstUsesWith(I, CSI);
3609 // Try to fold constant and into select arguments.
3610 if (isa<Constant>(Op0))
3611 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3612 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3615 // See if we can turn a signed shr into an unsigned shr.
3616 if (!isLeftShift && I.getType()->isSigned()) {
3617 if (MaskedValueIsZero(Op0,
3618 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
3619 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
3620 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
3622 return new CastInst(V, I.getType());
3626 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
3627 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
3632 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
3634 bool isLeftShift = I.getOpcode() == Instruction::Shl;
3635 bool isSignedShift = Op0->getType()->isSigned();
3636 bool isUnsignedShift = !isSignedShift;
3638 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
3639 // of a signed value.
3641 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
3642 if (Op1->getValue() >= TypeBits) {
3643 if (isUnsignedShift || isLeftShift)
3644 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
3646 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
3651 // ((X*C1) << C2) == (X * (C1 << C2))
3652 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
3653 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
3654 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
3655 return BinaryOperator::createMul(BO->getOperand(0),
3656 ConstantExpr::getShl(BOOp, Op1));
3658 // Try to fold constant and into select arguments.
3659 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3660 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3662 if (isa<PHINode>(Op0))
3663 if (Instruction *NV = FoldOpIntoPhi(I))
3666 if (Op0->hasOneUse()) {
3667 // If this is a SHL of a sign-extending cast, see if we can turn the input
3668 // into a zero extending cast (a simple strength reduction).
3669 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3670 const Type *SrcTy = CI->getOperand(0)->getType();
3671 if (isLeftShift && SrcTy->isInteger() && SrcTy->isSigned() &&
3672 SrcTy->getPrimitiveSizeInBits() <
3673 CI->getType()->getPrimitiveSizeInBits()) {
3674 // We can change it to a zero extension if we are shifting out all of
3675 // the sign extended bits. To check this, form a mask of all of the
3676 // sign extend bits, then shift them left and see if we have anything
3678 Constant *Mask = ConstantIntegral::getAllOnesValue(SrcTy); // 1111
3679 Mask = ConstantExpr::getZeroExtend(Mask, CI->getType()); // 00001111
3680 Mask = ConstantExpr::getNot(Mask); // 1's in the sign bits: 11110000
3681 if (ConstantExpr::getShl(Mask, Op1)->isNullValue()) {
3682 // If the shift is nuking all of the sign bits, change this to a
3683 // zero extension cast. To do this, cast the cast input to
3684 // unsigned, then to the requested size.
3685 Value *CastOp = CI->getOperand(0);
3687 new CastInst(CastOp, CastOp->getType()->getUnsignedVersion(),
3688 CI->getName()+".uns");
3689 NC = InsertNewInstBefore(NC, I);
3690 // Finally, insert a replacement for CI.
3691 NC = new CastInst(NC, CI->getType(), CI->getName());
3693 NC = InsertNewInstBefore(NC, I);
3694 WorkList.push_back(CI); // Delete CI later.
3695 I.setOperand(0, NC);
3696 return &I; // The SHL operand was modified.
3701 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
3702 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3705 switch (Op0BO->getOpcode()) {
3707 case Instruction::Add:
3708 case Instruction::And:
3709 case Instruction::Or:
3710 case Instruction::Xor:
3711 // These operators commute.
3712 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
3713 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3714 match(Op0BO->getOperand(1),
3715 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
3716 Instruction *YS = new ShiftInst(Instruction::Shl,
3717 Op0BO->getOperand(0), Op1,
3719 InsertNewInstBefore(YS, I); // (Y << C)
3720 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3722 Op0BO->getOperand(1)->getName());
3723 InsertNewInstBefore(X, I); // (X + (Y << C))
3724 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3725 C2 = ConstantExpr::getShl(C2, Op1);
3726 return BinaryOperator::createAnd(X, C2);
3729 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
3730 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
3731 match(Op0BO->getOperand(1),
3732 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3733 m_ConstantInt(CC))) && V2 == Op1 &&
3734 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
3735 Instruction *YS = new ShiftInst(Instruction::Shl,
3736 Op0BO->getOperand(0), Op1,
3738 InsertNewInstBefore(YS, I); // (Y << C)
3740 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
3741 V1->getName()+".mask");
3742 InsertNewInstBefore(XM, I); // X & (CC << C)
3744 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3748 case Instruction::Sub:
3749 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
3750 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3751 match(Op0BO->getOperand(0),
3752 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
3753 Instruction *YS = new ShiftInst(Instruction::Shl,
3754 Op0BO->getOperand(1), Op1,
3756 InsertNewInstBefore(YS, I); // (Y << C)
3757 Instruction *X = BinaryOperator::create(Op0BO->getOpcode(), YS,
3759 Op0BO->getOperand(0)->getName());
3760 InsertNewInstBefore(X, I); // (X + (Y << C))
3761 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
3762 C2 = ConstantExpr::getShl(C2, Op1);
3763 return BinaryOperator::createAnd(X, C2);
3766 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
3767 match(Op0BO->getOperand(0),
3768 m_And(m_Shr(m_Value(V1), m_Value(V2)),
3769 m_ConstantInt(CC))) && V2 == Op1 &&
3770 cast<BinaryOperator>(Op0BO->getOperand(0))->getOperand(0)->hasOneUse()) {
3771 Instruction *YS = new ShiftInst(Instruction::Shl,
3772 Op0BO->getOperand(1), Op1,
3774 InsertNewInstBefore(YS, I); // (Y << C)
3776 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
3777 V1->getName()+".mask");
3778 InsertNewInstBefore(XM, I); // X & (CC << C)
3780 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
3787 // If the operand is an bitwise operator with a constant RHS, and the
3788 // shift is the only use, we can pull it out of the shift.
3789 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
3790 bool isValid = true; // Valid only for And, Or, Xor
3791 bool highBitSet = false; // Transform if high bit of constant set?
3793 switch (Op0BO->getOpcode()) {
3794 default: isValid = false; break; // Do not perform transform!
3795 case Instruction::Add:
3796 isValid = isLeftShift;
3798 case Instruction::Or:
3799 case Instruction::Xor:
3802 case Instruction::And:
3807 // If this is a signed shift right, and the high bit is modified
3808 // by the logical operation, do not perform the transformation.
3809 // The highBitSet boolean indicates the value of the high bit of
3810 // the constant which would cause it to be modified for this
3813 if (isValid && !isLeftShift && isSignedShift) {
3814 uint64_t Val = Op0C->getRawValue();
3815 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
3819 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
3821 Instruction *NewShift =
3822 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
3825 InsertNewInstBefore(NewShift, I);
3827 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
3834 // Find out if this is a shift of a shift by a constant.
3835 ShiftInst *ShiftOp = 0;
3836 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
3838 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3839 // If this is a noop-integer case of a shift instruction, use the shift.
3840 if (CI->getOperand(0)->getType()->isInteger() &&
3841 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
3842 CI->getType()->getPrimitiveSizeInBits() &&
3843 isa<ShiftInst>(CI->getOperand(0))) {
3844 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
3848 if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
3849 // Find the operands and properties of the input shift. Note that the
3850 // signedness of the input shift may differ from the current shift if there
3851 // is a noop cast between the two.
3852 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
3853 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
3854 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
3856 ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
3858 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
3859 unsigned ShiftAmt2 = (unsigned)Op1->getValue();
3861 // Check for (A << c1) << c2 and (A >> c1) >> c2.
3862 if (isLeftShift == isShiftOfLeftShift) {
3863 // Do not fold these shifts if the first one is signed and the second one
3864 // is unsigned and this is a right shift. Further, don't do any folding
3866 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
3869 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
3870 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
3871 Amt = Op0->getType()->getPrimitiveSizeInBits();
3873 Value *Op = ShiftOp->getOperand(0);
3874 if (isShiftOfSignedShift != isSignedShift)
3875 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
3876 return new ShiftInst(I.getOpcode(), Op,
3877 ConstantUInt::get(Type::UByteTy, Amt));
3880 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
3881 // signed types, we can only support the (A >> c1) << c2 configuration,
3882 // because it can not turn an arbitrary bit of A into a sign bit.
3883 if (isUnsignedShift || isLeftShift) {
3884 // Calculate bitmask for what gets shifted off the edge.
3885 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
3887 C = ConstantExpr::getShl(C, ShiftAmt1C);
3889 C = ConstantExpr::getUShr(C, ShiftAmt1C);
3891 Value *Op = ShiftOp->getOperand(0);
3892 if (isShiftOfSignedShift != isSignedShift)
3893 Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
3896 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
3897 InsertNewInstBefore(Mask, I);
3899 // Figure out what flavor of shift we should use...
3900 if (ShiftAmt1 == ShiftAmt2) {
3901 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
3902 } else if (ShiftAmt1 < ShiftAmt2) {
3903 return new ShiftInst(I.getOpcode(), Mask,
3904 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
3905 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
3906 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
3907 // Make sure to emit an unsigned shift right, not a signed one.
3908 Mask = InsertNewInstBefore(new CastInst(Mask,
3909 Mask->getType()->getUnsignedVersion(),
3911 Mask = new ShiftInst(Instruction::Shr, Mask,
3912 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3913 InsertNewInstBefore(Mask, I);
3914 return new CastInst(Mask, I.getType());
3916 return new ShiftInst(ShiftOp->getOpcode(), Mask,
3917 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3920 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
3921 Op = InsertNewInstBefore(new CastInst(Mask,
3922 I.getType()->getSignedVersion(),
3923 Mask->getName()), I);
3924 Instruction *Shift =
3925 new ShiftInst(ShiftOp->getOpcode(), Op,
3926 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
3927 InsertNewInstBefore(Shift, I);
3929 C = ConstantIntegral::getAllOnesValue(Shift->getType());
3930 C = ConstantExpr::getShl(C, Op1);
3931 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
3932 InsertNewInstBefore(Mask, I);
3933 return new CastInst(Mask, I.getType());
3936 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
3937 // this case, C1 == C2 and C1 is 8, 16, or 32.
3938 if (ShiftAmt1 == ShiftAmt2) {
3939 const Type *SExtType = 0;
3940 switch (ShiftAmt1) {
3941 case 8 : SExtType = Type::SByteTy; break;
3942 case 16: SExtType = Type::ShortTy; break;
3943 case 32: SExtType = Type::IntTy; break;
3947 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
3949 InsertNewInstBefore(NewTrunc, I);
3950 return new CastInst(NewTrunc, I.getType());
3965 /// getCastType - In the future, we will split the cast instruction into these
3966 /// various types. Until then, we have to do the analysis here.
3967 static CastType getCastType(const Type *Src, const Type *Dest) {
3968 assert(Src->isIntegral() && Dest->isIntegral() &&
3969 "Only works on integral types!");
3970 unsigned SrcSize = Src->getPrimitiveSizeInBits();
3971 unsigned DestSize = Dest->getPrimitiveSizeInBits();
3973 if (SrcSize == DestSize) return Noop;
3974 if (SrcSize > DestSize) return Truncate;
3975 if (Src->isSigned()) return Signext;
3980 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
3983 static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
3984 const Type *DstTy, TargetData *TD) {
3986 // It is legal to eliminate the instruction if casting A->B->A if the sizes
3987 // are identical and the bits don't get reinterpreted (for example
3988 // int->float->int would not be allowed).
3989 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
3992 // If we are casting between pointer and integer types, treat pointers as
3993 // integers of the appropriate size for the code below.
3994 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
3995 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
3996 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
3998 // Allow free casting and conversion of sizes as long as the sign doesn't
4000 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
4001 CastType FirstCast = getCastType(SrcTy, MidTy);
4002 CastType SecondCast = getCastType(MidTy, DstTy);
4004 // Capture the effect of these two casts. If the result is a legal cast,
4005 // the CastType is stored here, otherwise a special code is used.
4006 static const unsigned CastResult[] = {
4007 // First cast is noop
4009 // First cast is a truncate
4010 1, 1, 4, 4, // trunc->extend is not safe to eliminate
4011 // First cast is a sign ext
4012 2, 5, 2, 4, // signext->zeroext never ok
4013 // First cast is a zero ext
4017 unsigned Result = CastResult[FirstCast*4+SecondCast];
4019 default: assert(0 && "Illegal table value!");
4024 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
4025 // truncates, we could eliminate more casts.
4026 return (unsigned)getCastType(SrcTy, DstTy) == Result;
4028 return false; // Not possible to eliminate this here.
4030 // Sign or zero extend followed by truncate is always ok if the result
4031 // is a truncate or noop.
4032 CastType ResultCast = getCastType(SrcTy, DstTy);
4033 if (ResultCast == Noop || ResultCast == Truncate)
4035 // Otherwise we are still growing the value, we are only safe if the
4036 // result will match the sign/zeroextendness of the result.
4037 return ResultCast == FirstCast;
4041 // If this is a cast from 'float -> double -> integer', cast from
4042 // 'float -> integer' directly, as the value isn't changed by the
4043 // float->double conversion.
4044 if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
4045 DstTy->isIntegral() &&
4046 SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
4052 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
4053 if (V->getType() == Ty || isa<Constant>(V)) return false;
4054 if (const CastInst *CI = dyn_cast<CastInst>(V))
4055 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
4061 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
4062 /// InsertBefore instruction. This is specialized a bit to avoid inserting
4063 /// casts that are known to not do anything...
4065 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
4066 Instruction *InsertBefore) {
4067 if (V->getType() == DestTy) return V;
4068 if (Constant *C = dyn_cast<Constant>(V))
4069 return ConstantExpr::getCast(C, DestTy);
4071 CastInst *CI = new CastInst(V, DestTy, V->getName());
4072 InsertNewInstBefore(CI, *InsertBefore);
4076 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
4077 /// expression. If so, decompose it, returning some value X, such that Val is
4080 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
4082 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
4083 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
4084 Offset = CI->getValue();
4086 return ConstantUInt::get(Type::UIntTy, 0);
4087 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
4088 if (I->getNumOperands() == 2) {
4089 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
4090 if (I->getOpcode() == Instruction::Shl) {
4091 // This is a value scaled by '1 << the shift amt'.
4092 Scale = 1U << CUI->getValue();
4094 return I->getOperand(0);
4095 } else if (I->getOpcode() == Instruction::Mul) {
4096 // This value is scaled by 'CUI'.
4097 Scale = CUI->getValue();
4099 return I->getOperand(0);
4100 } else if (I->getOpcode() == Instruction::Add) {
4101 // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
4104 Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
4106 Offset += CUI->getValue();
4107 if (SubScale > 1 && (Offset % SubScale == 0)) {
4116 // Otherwise, we can't look past this.
4123 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
4124 /// try to eliminate the cast by moving the type information into the alloc.
4125 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
4126 AllocationInst &AI) {
4127 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
4128 if (!PTy) return 0; // Not casting the allocation to a pointer type.
4130 // Remove any uses of AI that are dead.
4131 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
4132 std::vector<Instruction*> DeadUsers;
4133 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
4134 Instruction *User = cast<Instruction>(*UI++);
4135 if (isInstructionTriviallyDead(User)) {
4136 while (UI != E && *UI == User)
4137 ++UI; // If this instruction uses AI more than once, don't break UI.
4139 // Add operands to the worklist.
4140 AddUsesToWorkList(*User);
4142 DEBUG(std::cerr << "IC: DCE: " << *User);
4144 User->eraseFromParent();
4145 removeFromWorkList(User);
4149 // Get the type really allocated and the type casted to.
4150 const Type *AllocElTy = AI.getAllocatedType();
4151 const Type *CastElTy = PTy->getElementType();
4152 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
4154 unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
4155 unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
4156 if (CastElTyAlign < AllocElTyAlign) return 0;
4158 // If the allocation has multiple uses, only promote it if we are strictly
4159 // increasing the alignment of the resultant allocation. If we keep it the
4160 // same, we open the door to infinite loops of various kinds.
4161 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
4163 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
4164 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
4165 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
4167 // See if we can satisfy the modulus by pulling a scale out of the array
4169 unsigned ArraySizeScale, ArrayOffset;
4170 Value *NumElements = // See if the array size is a decomposable linear expr.
4171 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
4173 // If we can now satisfy the modulus, by using a non-1 scale, we really can
4175 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
4176 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
4178 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
4183 Amt = ConstantUInt::get(Type::UIntTy, Scale);
4184 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
4185 Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
4186 else if (Scale != 1) {
4187 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
4188 Amt = InsertNewInstBefore(Tmp, AI);
4192 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
4193 Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
4194 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
4195 Amt = InsertNewInstBefore(Tmp, AI);
4198 std::string Name = AI.getName(); AI.setName("");
4199 AllocationInst *New;
4200 if (isa<MallocInst>(AI))
4201 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
4203 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
4204 InsertNewInstBefore(New, AI);
4206 // If the allocation has multiple uses, insert a cast and change all things
4207 // that used it to use the new cast. This will also hack on CI, but it will
4209 if (!AI.hasOneUse()) {
4210 AddUsesToWorkList(AI);
4211 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
4212 InsertNewInstBefore(NewCast, AI);
4213 AI.replaceAllUsesWith(NewCast);
4215 return ReplaceInstUsesWith(CI, New);
4219 // CastInst simplification
4221 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
4222 Value *Src = CI.getOperand(0);
4224 // If the user is casting a value to the same type, eliminate this cast
4226 if (CI.getType() == Src->getType())
4227 return ReplaceInstUsesWith(CI, Src);
4229 if (isa<UndefValue>(Src)) // cast undef -> undef
4230 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
4232 // If casting the result of another cast instruction, try to eliminate this
4235 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
4236 Value *A = CSrc->getOperand(0);
4237 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
4238 CI.getType(), TD)) {
4239 // This instruction now refers directly to the cast's src operand. This
4240 // has a good chance of making CSrc dead.
4241 CI.setOperand(0, CSrc->getOperand(0));
4245 // If this is an A->B->A cast, and we are dealing with integral types, try
4246 // to convert this into a logical 'and' instruction.
4248 if (A->getType()->isInteger() &&
4249 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
4250 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
4251 CSrc->getType()->getPrimitiveSizeInBits() <
4252 CI.getType()->getPrimitiveSizeInBits()&&
4253 A->getType()->getPrimitiveSizeInBits() ==
4254 CI.getType()->getPrimitiveSizeInBits()) {
4255 assert(CSrc->getType() != Type::ULongTy &&
4256 "Cannot have type bigger than ulong!");
4257 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
4258 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
4260 AndOp = ConstantExpr::getCast(AndOp, A->getType());
4261 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
4262 if (And->getType() != CI.getType()) {
4263 And->setName(CSrc->getName()+".mask");
4264 InsertNewInstBefore(And, CI);
4265 And = new CastInst(And, CI.getType());
4271 // If this is a cast to bool, turn it into the appropriate setne instruction.
4272 if (CI.getType() == Type::BoolTy)
4273 return BinaryOperator::createSetNE(CI.getOperand(0),
4274 Constant::getNullValue(CI.getOperand(0)->getType()));
4276 // See if we can simplify any instructions used by the LHS whose sole
4277 // purpose is to compute bits we don't care about.
4278 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral() &&
4279 SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask()))
4282 // If casting the result of a getelementptr instruction with no offset, turn
4283 // this into a cast of the original pointer!
4285 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
4286 bool AllZeroOperands = true;
4287 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
4288 if (!isa<Constant>(GEP->getOperand(i)) ||
4289 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
4290 AllZeroOperands = false;
4293 if (AllZeroOperands) {
4294 CI.setOperand(0, GEP->getOperand(0));
4299 // If we are casting a malloc or alloca to a pointer to a type of the same
4300 // size, rewrite the allocation instruction to allocate the "right" type.
4302 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
4303 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
4306 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
4307 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
4309 if (isa<PHINode>(Src))
4310 if (Instruction *NV = FoldOpIntoPhi(CI))
4313 // If the source value is an instruction with only this use, we can attempt to
4314 // propagate the cast into the instruction. Also, only handle integral types
4316 if (Instruction *SrcI = dyn_cast<Instruction>(Src))
4317 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
4318 CI.getType()->isInteger()) { // Don't mess with casts to bool here
4319 const Type *DestTy = CI.getType();
4320 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
4321 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
4323 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
4324 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
4326 switch (SrcI->getOpcode()) {
4327 case Instruction::Add:
4328 case Instruction::Mul:
4329 case Instruction::And:
4330 case Instruction::Or:
4331 case Instruction::Xor:
4332 // If we are discarding information, or just changing the sign, rewrite.
4333 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
4334 // Don't insert two casts if they cannot be eliminated. We allow two
4335 // casts to be inserted if the sizes are the same. This could only be
4336 // converting signedness, which is a noop.
4337 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
4338 !ValueRequiresCast(Op0, DestTy, TD)) {
4339 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4340 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
4341 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
4342 ->getOpcode(), Op0c, Op1c);
4346 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
4347 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
4348 Op1 == ConstantBool::True &&
4349 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
4350 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
4351 return BinaryOperator::createXor(New,
4352 ConstantInt::get(CI.getType(), 1));
4355 case Instruction::Shl:
4356 // Allow changing the sign of the source operand. Do not allow changing
4357 // the size of the shift, UNLESS the shift amount is a constant. We
4358 // mush not change variable sized shifts to a smaller size, because it
4359 // is undefined to shift more bits out than exist in the value.
4360 if (DestBitSize == SrcBitSize ||
4361 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
4362 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4363 return new ShiftInst(Instruction::Shl, Op0c, Op1);
4366 case Instruction::Shr:
4367 // If this is a signed shr, and if all bits shifted in are about to be
4368 // truncated off, turn it into an unsigned shr to allow greater
4370 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
4371 isa<ConstantInt>(Op1)) {
4372 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
4373 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
4374 // Convert to unsigned.
4375 Value *N1 = InsertOperandCastBefore(Op0,
4376 Op0->getType()->getUnsignedVersion(), &CI);
4377 // Insert the new shift, which is now unsigned.
4378 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
4379 Op1, Src->getName()), CI);
4380 return new CastInst(N1, CI.getType());
4385 case Instruction::SetNE:
4386 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
4387 if (Op1C->getRawValue() == 0) {
4388 // If the input only has the low bit set, simplify directly.
4390 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
4391 // cast (X != 0) to int --> X if X&~1 == 0
4392 if (MaskedValueIsZero(Op0,
4393 cast<ConstantIntegral>(Not1)->getZExtValue())) {
4394 if (CI.getType() == Op0->getType())
4395 return ReplaceInstUsesWith(CI, Op0);
4397 return new CastInst(Op0, CI.getType());
4400 // If the input is an and with a single bit, shift then simplify.
4401 ConstantInt *AndRHS;
4402 if (match(Op0, m_And(m_Value(), m_ConstantInt(AndRHS))))
4403 if (AndRHS->getRawValue() &&
4404 (AndRHS->getRawValue() & (AndRHS->getRawValue()-1)) == 0) {
4405 unsigned ShiftAmt = Log2_64(AndRHS->getRawValue());
4406 // Perform an unsigned shr by shiftamt. Convert input to
4407 // unsigned if it is signed.
4409 if (In->getType()->isSigned())
4410 In = InsertNewInstBefore(new CastInst(In,
4411 In->getType()->getUnsignedVersion(), In->getName()),CI);
4412 // Insert the shift to put the result in the low bit.
4413 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
4414 ConstantInt::get(Type::UByteTy, ShiftAmt),
4415 In->getName()+".lobit"), CI);
4416 if (CI.getType() == In->getType())
4417 return ReplaceInstUsesWith(CI, In);
4419 return new CastInst(In, CI.getType());
4424 case Instruction::SetEQ:
4425 // We if we are just checking for a seteq of a single bit and casting it
4426 // to an integer. If so, shift the bit to the appropriate place then
4427 // cast to integer to avoid the comparison.
4428 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
4429 // Is Op1C a power of two or zero?
4430 if ((Op1C->getRawValue() & Op1C->getRawValue()-1) == 0) {
4431 // cast (X == 1) to int -> X iff X has only the low bit set.
4432 if (Op1C->getRawValue() == 1) {
4434 ConstantExpr::getNot(ConstantInt::get(Op0->getType(), 1));
4435 if (MaskedValueIsZero(Op0,
4436 cast<ConstantIntegral>(Not1)->getZExtValue())) {
4437 if (CI.getType() == Op0->getType())
4438 return ReplaceInstUsesWith(CI, Op0);
4440 return new CastInst(Op0, CI.getType());
4452 /// GetSelectFoldableOperands - We want to turn code that looks like this:
4454 /// %D = select %cond, %C, %A
4456 /// %C = select %cond, %B, 0
4459 /// Assuming that the specified instruction is an operand to the select, return
4460 /// a bitmask indicating which operands of this instruction are foldable if they
4461 /// equal the other incoming value of the select.
4463 static unsigned GetSelectFoldableOperands(Instruction *I) {
4464 switch (I->getOpcode()) {
4465 case Instruction::Add:
4466 case Instruction::Mul:
4467 case Instruction::And:
4468 case Instruction::Or:
4469 case Instruction::Xor:
4470 return 3; // Can fold through either operand.
4471 case Instruction::Sub: // Can only fold on the amount subtracted.
4472 case Instruction::Shl: // Can only fold on the shift amount.
4473 case Instruction::Shr:
4476 return 0; // Cannot fold
4480 /// GetSelectFoldableConstant - For the same transformation as the previous
4481 /// function, return the identity constant that goes into the select.
4482 static Constant *GetSelectFoldableConstant(Instruction *I) {
4483 switch (I->getOpcode()) {
4484 default: assert(0 && "This cannot happen!"); abort();
4485 case Instruction::Add:
4486 case Instruction::Sub:
4487 case Instruction::Or:
4488 case Instruction::Xor:
4489 return Constant::getNullValue(I->getType());
4490 case Instruction::Shl:
4491 case Instruction::Shr:
4492 return Constant::getNullValue(Type::UByteTy);
4493 case Instruction::And:
4494 return ConstantInt::getAllOnesValue(I->getType());
4495 case Instruction::Mul:
4496 return ConstantInt::get(I->getType(), 1);
4500 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
4501 /// have the same opcode and only one use each. Try to simplify this.
4502 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
4504 if (TI->getNumOperands() == 1) {
4505 // If this is a non-volatile load or a cast from the same type,
4507 if (TI->getOpcode() == Instruction::Cast) {
4508 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
4511 return 0; // unknown unary op.
4514 // Fold this by inserting a select from the input values.
4515 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
4516 FI->getOperand(0), SI.getName()+".v");
4517 InsertNewInstBefore(NewSI, SI);
4518 return new CastInst(NewSI, TI->getType());
4521 // Only handle binary operators here.
4522 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
4525 // Figure out if the operations have any operands in common.
4526 Value *MatchOp, *OtherOpT, *OtherOpF;
4528 if (TI->getOperand(0) == FI->getOperand(0)) {
4529 MatchOp = TI->getOperand(0);
4530 OtherOpT = TI->getOperand(1);
4531 OtherOpF = FI->getOperand(1);
4532 MatchIsOpZero = true;
4533 } else if (TI->getOperand(1) == FI->getOperand(1)) {
4534 MatchOp = TI->getOperand(1);
4535 OtherOpT = TI->getOperand(0);
4536 OtherOpF = FI->getOperand(0);
4537 MatchIsOpZero = false;
4538 } else if (!TI->isCommutative()) {
4540 } else if (TI->getOperand(0) == FI->getOperand(1)) {
4541 MatchOp = TI->getOperand(0);
4542 OtherOpT = TI->getOperand(1);
4543 OtherOpF = FI->getOperand(0);
4544 MatchIsOpZero = true;
4545 } else if (TI->getOperand(1) == FI->getOperand(0)) {
4546 MatchOp = TI->getOperand(1);
4547 OtherOpT = TI->getOperand(0);
4548 OtherOpF = FI->getOperand(1);
4549 MatchIsOpZero = true;
4554 // If we reach here, they do have operations in common.
4555 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
4556 OtherOpF, SI.getName()+".v");
4557 InsertNewInstBefore(NewSI, SI);
4559 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
4561 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
4563 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
4566 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
4568 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
4572 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
4573 Value *CondVal = SI.getCondition();
4574 Value *TrueVal = SI.getTrueValue();
4575 Value *FalseVal = SI.getFalseValue();
4577 // select true, X, Y -> X
4578 // select false, X, Y -> Y
4579 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
4580 if (C == ConstantBool::True)
4581 return ReplaceInstUsesWith(SI, TrueVal);
4583 assert(C == ConstantBool::False);
4584 return ReplaceInstUsesWith(SI, FalseVal);
4587 // select C, X, X -> X
4588 if (TrueVal == FalseVal)
4589 return ReplaceInstUsesWith(SI, TrueVal);
4591 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
4592 return ReplaceInstUsesWith(SI, FalseVal);
4593 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
4594 return ReplaceInstUsesWith(SI, TrueVal);
4595 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
4596 if (isa<Constant>(TrueVal))
4597 return ReplaceInstUsesWith(SI, TrueVal);
4599 return ReplaceInstUsesWith(SI, FalseVal);
4602 if (SI.getType() == Type::BoolTy)
4603 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
4604 if (C == ConstantBool::True) {
4605 // Change: A = select B, true, C --> A = or B, C
4606 return BinaryOperator::createOr(CondVal, FalseVal);
4608 // Change: A = select B, false, C --> A = and !B, C
4610 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4611 "not."+CondVal->getName()), SI);
4612 return BinaryOperator::createAnd(NotCond, FalseVal);
4614 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
4615 if (C == ConstantBool::False) {
4616 // Change: A = select B, C, false --> A = and B, C
4617 return BinaryOperator::createAnd(CondVal, TrueVal);
4619 // Change: A = select B, C, true --> A = or !B, C
4621 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4622 "not."+CondVal->getName()), SI);
4623 return BinaryOperator::createOr(NotCond, TrueVal);
4627 // Selecting between two integer constants?
4628 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
4629 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
4630 // select C, 1, 0 -> cast C to int
4631 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
4632 return new CastInst(CondVal, SI.getType());
4633 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
4634 // select C, 0, 1 -> cast !C to int
4636 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
4637 "not."+CondVal->getName()), SI);
4638 return new CastInst(NotCond, SI.getType());
4641 // If one of the constants is zero (we know they can't both be) and we
4642 // have a setcc instruction with zero, and we have an 'and' with the
4643 // non-constant value, eliminate this whole mess. This corresponds to
4644 // cases like this: ((X & 27) ? 27 : 0)
4645 if (TrueValC->isNullValue() || FalseValC->isNullValue())
4646 if (Instruction *IC = dyn_cast<Instruction>(SI.getCondition()))
4647 if ((IC->getOpcode() == Instruction::SetEQ ||
4648 IC->getOpcode() == Instruction::SetNE) &&
4649 isa<ConstantInt>(IC->getOperand(1)) &&
4650 cast<Constant>(IC->getOperand(1))->isNullValue())
4651 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
4652 if (ICA->getOpcode() == Instruction::And &&
4653 isa<ConstantInt>(ICA->getOperand(1)) &&
4654 (ICA->getOperand(1) == TrueValC ||
4655 ICA->getOperand(1) == FalseValC) &&
4656 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
4657 // Okay, now we know that everything is set up, we just don't
4658 // know whether we have a setne or seteq and whether the true or
4659 // false val is the zero.
4660 bool ShouldNotVal = !TrueValC->isNullValue();
4661 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
4664 V = InsertNewInstBefore(BinaryOperator::create(
4665 Instruction::Xor, V, ICA->getOperand(1)), SI);
4666 return ReplaceInstUsesWith(SI, V);
4670 // See if we are selecting two values based on a comparison of the two values.
4671 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
4672 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
4673 // Transform (X == Y) ? X : Y -> Y
4674 if (SCI->getOpcode() == Instruction::SetEQ)
4675 return ReplaceInstUsesWith(SI, FalseVal);
4676 // Transform (X != Y) ? X : Y -> X
4677 if (SCI->getOpcode() == Instruction::SetNE)
4678 return ReplaceInstUsesWith(SI, TrueVal);
4679 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4681 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
4682 // Transform (X == Y) ? Y : X -> X
4683 if (SCI->getOpcode() == Instruction::SetEQ)
4684 return ReplaceInstUsesWith(SI, FalseVal);
4685 // Transform (X != Y) ? Y : X -> Y
4686 if (SCI->getOpcode() == Instruction::SetNE)
4687 return ReplaceInstUsesWith(SI, TrueVal);
4688 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
4692 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
4693 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
4694 if (TI->hasOneUse() && FI->hasOneUse()) {
4695 bool isInverse = false;
4696 Instruction *AddOp = 0, *SubOp = 0;
4698 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
4699 if (TI->getOpcode() == FI->getOpcode())
4700 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
4703 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
4704 // even legal for FP.
4705 if (TI->getOpcode() == Instruction::Sub &&
4706 FI->getOpcode() == Instruction::Add) {
4707 AddOp = FI; SubOp = TI;
4708 } else if (FI->getOpcode() == Instruction::Sub &&
4709 TI->getOpcode() == Instruction::Add) {
4710 AddOp = TI; SubOp = FI;
4714 Value *OtherAddOp = 0;
4715 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
4716 OtherAddOp = AddOp->getOperand(1);
4717 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
4718 OtherAddOp = AddOp->getOperand(0);
4722 // So at this point we know we have:
4723 // select C, (add X, Y), (sub X, ?)
4724 // We can do the transform profitably if either 'Y' = '?' or '?' is
4726 if (SubOp->getOperand(1) == AddOp ||
4727 isa<Constant>(SubOp->getOperand(1))) {
4729 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
4730 NegVal = ConstantExpr::getNeg(C);
4732 NegVal = InsertNewInstBefore(
4733 BinaryOperator::createNeg(SubOp->getOperand(1)), SI);
4736 Value *NewTrueOp = OtherAddOp;
4737 Value *NewFalseOp = NegVal;
4739 std::swap(NewTrueOp, NewFalseOp);
4740 Instruction *NewSel =
4741 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
4743 NewSel = InsertNewInstBefore(NewSel, SI);
4744 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
4750 // See if we can fold the select into one of our operands.
4751 if (SI.getType()->isInteger()) {
4752 // See the comment above GetSelectFoldableOperands for a description of the
4753 // transformation we are doing here.
4754 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
4755 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
4756 !isa<Constant>(FalseVal))
4757 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
4758 unsigned OpToFold = 0;
4759 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
4761 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
4766 Constant *C = GetSelectFoldableConstant(TVI);
4767 std::string Name = TVI->getName(); TVI->setName("");
4768 Instruction *NewSel =
4769 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
4771 InsertNewInstBefore(NewSel, SI);
4772 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
4773 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
4774 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
4775 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
4777 assert(0 && "Unknown instruction!!");
4782 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
4783 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
4784 !isa<Constant>(TrueVal))
4785 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
4786 unsigned OpToFold = 0;
4787 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
4789 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
4794 Constant *C = GetSelectFoldableConstant(FVI);
4795 std::string Name = FVI->getName(); FVI->setName("");
4796 Instruction *NewSel =
4797 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
4799 InsertNewInstBefore(NewSel, SI);
4800 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
4801 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
4802 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
4803 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
4805 assert(0 && "Unknown instruction!!");
4811 if (BinaryOperator::isNot(CondVal)) {
4812 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
4813 SI.setOperand(1, FalseVal);
4814 SI.setOperand(2, TrueVal);
4822 /// visitCallInst - CallInst simplification. This mostly only handles folding
4823 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
4824 /// the heavy lifting.
4826 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
4827 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
4828 if (!II) return visitCallSite(&CI);
4830 // Intrinsics cannot occur in an invoke, so handle them here instead of in
4832 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
4833 bool Changed = false;
4835 // memmove/cpy/set of zero bytes is a noop.
4836 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
4837 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
4839 // FIXME: Increase alignment here.
4841 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
4842 if (CI->getRawValue() == 1) {
4843 // Replace the instruction with just byte operations. We would
4844 // transform other cases to loads/stores, but we don't know if
4845 // alignment is sufficient.
4849 // If we have a memmove and the source operation is a constant global,
4850 // then the source and dest pointers can't alias, so we can change this
4851 // into a call to memcpy.
4852 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II))
4853 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
4854 if (GVSrc->isConstant()) {
4855 Module *M = CI.getParent()->getParent()->getParent();
4856 Function *MemCpy = M->getOrInsertFunction("llvm.memcpy",
4857 CI.getCalledFunction()->getFunctionType());
4858 CI.setOperand(0, MemCpy);
4862 if (Changed) return II;
4863 } else if (DbgStopPointInst *SPI = dyn_cast<DbgStopPointInst>(II)) {
4864 // If this stoppoint is at the same source location as the previous
4865 // stoppoint in the chain, it is not needed.
4866 if (DbgStopPointInst *PrevSPI =
4867 dyn_cast<DbgStopPointInst>(SPI->getChain()))
4868 if (SPI->getLineNo() == PrevSPI->getLineNo() &&
4869 SPI->getColNo() == PrevSPI->getColNo()) {
4870 SPI->replaceAllUsesWith(PrevSPI);
4871 return EraseInstFromFunction(CI);
4874 switch (II->getIntrinsicID()) {
4876 case Intrinsic::stackrestore: {
4877 // If the save is right next to the restore, remove the restore. This can
4878 // happen when variable allocas are DCE'd.
4879 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
4880 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
4881 BasicBlock::iterator BI = SS;
4883 return EraseInstFromFunction(CI);
4887 // If the stack restore is in a return/unwind block and if there are no
4888 // allocas or calls between the restore and the return, nuke the restore.
4889 TerminatorInst *TI = II->getParent()->getTerminator();
4890 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
4891 BasicBlock::iterator BI = II;
4892 bool CannotRemove = false;
4893 for (++BI; &*BI != TI; ++BI) {
4894 if (isa<AllocaInst>(BI) ||
4895 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
4896 CannotRemove = true;
4901 return EraseInstFromFunction(CI);
4908 return visitCallSite(II);
4911 // InvokeInst simplification
4913 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
4914 return visitCallSite(&II);
4917 // visitCallSite - Improvements for call and invoke instructions.
4919 Instruction *InstCombiner::visitCallSite(CallSite CS) {
4920 bool Changed = false;
4922 // If the callee is a constexpr cast of a function, attempt to move the cast
4923 // to the arguments of the call/invoke.
4924 if (transformConstExprCastCall(CS)) return 0;
4926 Value *Callee = CS.getCalledValue();
4928 if (Function *CalleeF = dyn_cast<Function>(Callee))
4929 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
4930 Instruction *OldCall = CS.getInstruction();
4931 // If the call and callee calling conventions don't match, this call must
4932 // be unreachable, as the call is undefined.
4933 new StoreInst(ConstantBool::True,
4934 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
4935 if (!OldCall->use_empty())
4936 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
4937 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
4938 return EraseInstFromFunction(*OldCall);
4942 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
4943 // This instruction is not reachable, just remove it. We insert a store to
4944 // undef so that we know that this code is not reachable, despite the fact
4945 // that we can't modify the CFG here.
4946 new StoreInst(ConstantBool::True,
4947 UndefValue::get(PointerType::get(Type::BoolTy)),
4948 CS.getInstruction());
4950 if (!CS.getInstruction()->use_empty())
4951 CS.getInstruction()->
4952 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
4954 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
4955 // Don't break the CFG, insert a dummy cond branch.
4956 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
4957 ConstantBool::True, II);
4959 return EraseInstFromFunction(*CS.getInstruction());
4962 const PointerType *PTy = cast<PointerType>(Callee->getType());
4963 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4964 if (FTy->isVarArg()) {
4965 // See if we can optimize any arguments passed through the varargs area of
4967 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
4968 E = CS.arg_end(); I != E; ++I)
4969 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
4970 // If this cast does not effect the value passed through the varargs
4971 // area, we can eliminate the use of the cast.
4972 Value *Op = CI->getOperand(0);
4973 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
4980 return Changed ? CS.getInstruction() : 0;
4983 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
4984 // attempt to move the cast to the arguments of the call/invoke.
4986 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
4987 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
4988 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
4989 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
4991 Function *Callee = cast<Function>(CE->getOperand(0));
4992 Instruction *Caller = CS.getInstruction();
4994 // Okay, this is a cast from a function to a different type. Unless doing so
4995 // would cause a type conversion of one of our arguments, change this call to
4996 // be a direct call with arguments casted to the appropriate types.
4998 const FunctionType *FT = Callee->getFunctionType();
4999 const Type *OldRetTy = Caller->getType();
5001 // Check to see if we are changing the return type...
5002 if (OldRetTy != FT->getReturnType()) {
5003 if (Callee->isExternal() &&
5004 !OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) &&
5005 !Caller->use_empty())
5006 return false; // Cannot transform this return value...
5008 // If the callsite is an invoke instruction, and the return value is used by
5009 // a PHI node in a successor, we cannot change the return type of the call
5010 // because there is no place to put the cast instruction (without breaking
5011 // the critical edge). Bail out in this case.
5012 if (!Caller->use_empty())
5013 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
5014 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
5016 if (PHINode *PN = dyn_cast<PHINode>(*UI))
5017 if (PN->getParent() == II->getNormalDest() ||
5018 PN->getParent() == II->getUnwindDest())
5022 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
5023 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
5025 CallSite::arg_iterator AI = CS.arg_begin();
5026 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
5027 const Type *ParamTy = FT->getParamType(i);
5028 bool isConvertible = (*AI)->getType()->isLosslesslyConvertibleTo(ParamTy);
5029 if (Callee->isExternal() && !isConvertible) return false;
5032 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
5033 Callee->isExternal())
5034 return false; // Do not delete arguments unless we have a function body...
5036 // Okay, we decided that this is a safe thing to do: go ahead and start
5037 // inserting cast instructions as necessary...
5038 std::vector<Value*> Args;
5039 Args.reserve(NumActualArgs);
5041 AI = CS.arg_begin();
5042 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
5043 const Type *ParamTy = FT->getParamType(i);
5044 if ((*AI)->getType() == ParamTy) {
5045 Args.push_back(*AI);
5047 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
5052 // If the function takes more arguments than the call was taking, add them
5054 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
5055 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
5057 // If we are removing arguments to the function, emit an obnoxious warning...
5058 if (FT->getNumParams() < NumActualArgs)
5059 if (!FT->isVarArg()) {
5060 std::cerr << "WARNING: While resolving call to function '"
5061 << Callee->getName() << "' arguments were dropped!\n";
5063 // Add all of the arguments in their promoted form to the arg list...
5064 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
5065 const Type *PTy = getPromotedType((*AI)->getType());
5066 if (PTy != (*AI)->getType()) {
5067 // Must promote to pass through va_arg area!
5068 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
5069 InsertNewInstBefore(Cast, *Caller);
5070 Args.push_back(Cast);
5072 Args.push_back(*AI);
5077 if (FT->getReturnType() == Type::VoidTy)
5078 Caller->setName(""); // Void type should not have a name...
5081 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5082 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
5083 Args, Caller->getName(), Caller);
5084 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
5086 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
5087 if (cast<CallInst>(Caller)->isTailCall())
5088 cast<CallInst>(NC)->setTailCall();
5089 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
5092 // Insert a cast of the return type as necessary...
5094 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
5095 if (NV->getType() != Type::VoidTy) {
5096 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
5098 // If this is an invoke instruction, we should insert it after the first
5099 // non-phi, instruction in the normal successor block.
5100 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5101 BasicBlock::iterator I = II->getNormalDest()->begin();
5102 while (isa<PHINode>(I)) ++I;
5103 InsertNewInstBefore(NC, *I);
5105 // Otherwise, it's a call, just insert cast right after the call instr
5106 InsertNewInstBefore(NC, *Caller);
5108 AddUsersToWorkList(*Caller);
5110 NV = UndefValue::get(Caller->getType());
5114 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
5115 Caller->replaceAllUsesWith(NV);
5116 Caller->getParent()->getInstList().erase(Caller);
5117 removeFromWorkList(Caller);
5122 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
5123 // operator and they all are only used by the PHI, PHI together their
5124 // inputs, and do the operation once, to the result of the PHI.
5125 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
5126 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
5128 // Scan the instruction, looking for input operations that can be folded away.
5129 // If all input operands to the phi are the same instruction (e.g. a cast from
5130 // the same type or "+42") we can pull the operation through the PHI, reducing
5131 // code size and simplifying code.
5132 Constant *ConstantOp = 0;
5133 const Type *CastSrcTy = 0;
5134 if (isa<CastInst>(FirstInst)) {
5135 CastSrcTy = FirstInst->getOperand(0)->getType();
5136 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
5137 // Can fold binop or shift if the RHS is a constant.
5138 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
5139 if (ConstantOp == 0) return 0;
5141 return 0; // Cannot fold this operation.
5144 // Check to see if all arguments are the same operation.
5145 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5146 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
5147 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
5148 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
5151 if (I->getOperand(0)->getType() != CastSrcTy)
5152 return 0; // Cast operation must match.
5153 } else if (I->getOperand(1) != ConstantOp) {
5158 // Okay, they are all the same operation. Create a new PHI node of the
5159 // correct type, and PHI together all of the LHS's of the instructions.
5160 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
5161 PN.getName()+".in");
5162 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
5164 Value *InVal = FirstInst->getOperand(0);
5165 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
5167 // Add all operands to the new PHI.
5168 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5169 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
5170 if (NewInVal != InVal)
5172 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
5177 // The new PHI unions all of the same values together. This is really
5178 // common, so we handle it intelligently here for compile-time speed.
5182 InsertNewInstBefore(NewPN, PN);
5186 // Insert and return the new operation.
5187 if (isa<CastInst>(FirstInst))
5188 return new CastInst(PhiVal, PN.getType());
5189 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
5190 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
5192 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
5193 PhiVal, ConstantOp);
5196 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
5198 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
5199 if (PN->use_empty()) return true;
5200 if (!PN->hasOneUse()) return false;
5202 // Remember this node, and if we find the cycle, return.
5203 if (!PotentiallyDeadPHIs.insert(PN).second)
5206 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
5207 return DeadPHICycle(PU, PotentiallyDeadPHIs);
5212 // PHINode simplification
5214 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
5215 if (Value *V = PN.hasConstantValue())
5216 return ReplaceInstUsesWith(PN, V);
5218 // If the only user of this instruction is a cast instruction, and all of the
5219 // incoming values are constants, change this PHI to merge together the casted
5222 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
5223 if (CI->getType() != PN.getType()) { // noop casts will be folded
5224 bool AllConstant = true;
5225 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
5226 if (!isa<Constant>(PN.getIncomingValue(i))) {
5227 AllConstant = false;
5231 // Make a new PHI with all casted values.
5232 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
5233 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
5234 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
5235 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
5236 PN.getIncomingBlock(i));
5239 // Update the cast instruction.
5240 CI->setOperand(0, New);
5241 WorkList.push_back(CI); // revisit the cast instruction to fold.
5242 WorkList.push_back(New); // Make sure to revisit the new Phi
5243 return &PN; // PN is now dead!
5247 // If all PHI operands are the same operation, pull them through the PHI,
5248 // reducing code size.
5249 if (isa<Instruction>(PN.getIncomingValue(0)) &&
5250 PN.getIncomingValue(0)->hasOneUse())
5251 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
5254 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
5255 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
5256 // PHI)... break the cycle.
5258 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
5259 std::set<PHINode*> PotentiallyDeadPHIs;
5260 PotentiallyDeadPHIs.insert(&PN);
5261 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
5262 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
5268 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
5269 Instruction *InsertPoint,
5271 unsigned PS = IC->getTargetData().getPointerSize();
5272 const Type *VTy = V->getType();
5273 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
5274 // We must insert a cast to ensure we sign-extend.
5275 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
5276 V->getName()), *InsertPoint);
5277 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
5282 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
5283 Value *PtrOp = GEP.getOperand(0);
5284 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
5285 // If so, eliminate the noop.
5286 if (GEP.getNumOperands() == 1)
5287 return ReplaceInstUsesWith(GEP, PtrOp);
5289 if (isa<UndefValue>(GEP.getOperand(0)))
5290 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
5292 bool HasZeroPointerIndex = false;
5293 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
5294 HasZeroPointerIndex = C->isNullValue();
5296 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
5297 return ReplaceInstUsesWith(GEP, PtrOp);
5299 // Eliminate unneeded casts for indices.
5300 bool MadeChange = false;
5301 gep_type_iterator GTI = gep_type_begin(GEP);
5302 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
5303 if (isa<SequentialType>(*GTI)) {
5304 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
5305 Value *Src = CI->getOperand(0);
5306 const Type *SrcTy = Src->getType();
5307 const Type *DestTy = CI->getType();
5308 if (Src->getType()->isInteger()) {
5309 if (SrcTy->getPrimitiveSizeInBits() ==
5310 DestTy->getPrimitiveSizeInBits()) {
5311 // We can always eliminate a cast from ulong or long to the other.
5312 // We can always eliminate a cast from uint to int or the other on
5313 // 32-bit pointer platforms.
5314 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
5316 GEP.setOperand(i, Src);
5318 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
5319 SrcTy->getPrimitiveSize() == 4) {
5320 // We can always eliminate a cast from int to [u]long. We can
5321 // eliminate a cast from uint to [u]long iff the target is a 32-bit
5323 if (SrcTy->isSigned() ||
5324 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
5326 GEP.setOperand(i, Src);
5331 // If we are using a wider index than needed for this platform, shrink it
5332 // to what we need. If the incoming value needs a cast instruction,
5333 // insert it. This explicit cast can make subsequent optimizations more
5335 Value *Op = GEP.getOperand(i);
5336 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
5337 if (Constant *C = dyn_cast<Constant>(Op)) {
5338 GEP.setOperand(i, ConstantExpr::getCast(C,
5339 TD->getIntPtrType()->getSignedVersion()));
5342 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
5343 Op->getName()), GEP);
5344 GEP.setOperand(i, Op);
5348 // If this is a constant idx, make sure to canonicalize it to be a signed
5349 // operand, otherwise CSE and other optimizations are pessimized.
5350 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
5351 GEP.setOperand(i, ConstantExpr::getCast(CUI,
5352 CUI->getType()->getSignedVersion()));
5356 if (MadeChange) return &GEP;
5358 // Combine Indices - If the source pointer to this getelementptr instruction
5359 // is a getelementptr instruction, combine the indices of the two
5360 // getelementptr instructions into a single instruction.
5362 std::vector<Value*> SrcGEPOperands;
5363 if (User *Src = dyn_castGetElementPtr(PtrOp))
5364 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
5366 if (!SrcGEPOperands.empty()) {
5367 // Note that if our source is a gep chain itself that we wait for that
5368 // chain to be resolved before we perform this transformation. This
5369 // avoids us creating a TON of code in some cases.
5371 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
5372 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
5373 return 0; // Wait until our source is folded to completion.
5375 std::vector<Value *> Indices;
5377 // Find out whether the last index in the source GEP is a sequential idx.
5378 bool EndsWithSequential = false;
5379 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
5380 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
5381 EndsWithSequential = !isa<StructType>(*I);
5383 // Can we combine the two pointer arithmetics offsets?
5384 if (EndsWithSequential) {
5385 // Replace: gep (gep %P, long B), long A, ...
5386 // With: T = long A+B; gep %P, T, ...
5388 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
5389 if (SO1 == Constant::getNullValue(SO1->getType())) {
5391 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
5394 // If they aren't the same type, convert both to an integer of the
5395 // target's pointer size.
5396 if (SO1->getType() != GO1->getType()) {
5397 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
5398 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
5399 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
5400 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
5402 unsigned PS = TD->getPointerSize();
5403 if (SO1->getType()->getPrimitiveSize() == PS) {
5404 // Convert GO1 to SO1's type.
5405 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
5407 } else if (GO1->getType()->getPrimitiveSize() == PS) {
5408 // Convert SO1 to GO1's type.
5409 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
5411 const Type *PT = TD->getIntPtrType();
5412 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
5413 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
5417 if (isa<Constant>(SO1) && isa<Constant>(GO1))
5418 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
5420 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
5421 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
5425 // Recycle the GEP we already have if possible.
5426 if (SrcGEPOperands.size() == 2) {
5427 GEP.setOperand(0, SrcGEPOperands[0]);
5428 GEP.setOperand(1, Sum);
5431 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5432 SrcGEPOperands.end()-1);
5433 Indices.push_back(Sum);
5434 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
5436 } else if (isa<Constant>(*GEP.idx_begin()) &&
5437 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
5438 SrcGEPOperands.size() != 1) {
5439 // Otherwise we can do the fold if the first index of the GEP is a zero
5440 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5441 SrcGEPOperands.end());
5442 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
5445 if (!Indices.empty())
5446 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
5448 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
5449 // GEP of global variable. If all of the indices for this GEP are
5450 // constants, we can promote this to a constexpr instead of an instruction.
5452 // Scan for nonconstants...
5453 std::vector<Constant*> Indices;
5454 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
5455 for (; I != E && isa<Constant>(*I); ++I)
5456 Indices.push_back(cast<Constant>(*I));
5458 if (I == E) { // If they are all constants...
5459 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
5461 // Replace all uses of the GEP with the new constexpr...
5462 return ReplaceInstUsesWith(GEP, CE);
5464 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
5465 if (!isa<PointerType>(X->getType())) {
5466 // Not interesting. Source pointer must be a cast from pointer.
5467 } else if (HasZeroPointerIndex) {
5468 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
5469 // into : GEP [10 x ubyte]* X, long 0, ...
5471 // This occurs when the program declares an array extern like "int X[];"
5473 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
5474 const PointerType *XTy = cast<PointerType>(X->getType());
5475 if (const ArrayType *XATy =
5476 dyn_cast<ArrayType>(XTy->getElementType()))
5477 if (const ArrayType *CATy =
5478 dyn_cast<ArrayType>(CPTy->getElementType()))
5479 if (CATy->getElementType() == XATy->getElementType()) {
5480 // At this point, we know that the cast source type is a pointer
5481 // to an array of the same type as the destination pointer
5482 // array. Because the array type is never stepped over (there
5483 // is a leading zero) we can fold the cast into this GEP.
5484 GEP.setOperand(0, X);
5487 } else if (GEP.getNumOperands() == 2) {
5488 // Transform things like:
5489 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
5490 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
5491 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
5492 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
5493 if (isa<ArrayType>(SrcElTy) &&
5494 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
5495 TD->getTypeSize(ResElTy)) {
5496 Value *V = InsertNewInstBefore(
5497 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5498 GEP.getOperand(1), GEP.getName()), GEP);
5499 return new CastInst(V, GEP.getType());
5502 // Transform things like:
5503 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
5504 // (where tmp = 8*tmp2) into:
5505 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
5507 if (isa<ArrayType>(SrcElTy) &&
5508 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
5509 uint64_t ArrayEltSize =
5510 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
5512 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
5513 // allow either a mul, shift, or constant here.
5515 ConstantInt *Scale = 0;
5516 if (ArrayEltSize == 1) {
5517 NewIdx = GEP.getOperand(1);
5518 Scale = ConstantInt::get(NewIdx->getType(), 1);
5519 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
5520 NewIdx = ConstantInt::get(CI->getType(), 1);
5522 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
5523 if (Inst->getOpcode() == Instruction::Shl &&
5524 isa<ConstantInt>(Inst->getOperand(1))) {
5525 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
5526 if (Inst->getType()->isSigned())
5527 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
5529 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
5530 NewIdx = Inst->getOperand(0);
5531 } else if (Inst->getOpcode() == Instruction::Mul &&
5532 isa<ConstantInt>(Inst->getOperand(1))) {
5533 Scale = cast<ConstantInt>(Inst->getOperand(1));
5534 NewIdx = Inst->getOperand(0);
5538 // If the index will be to exactly the right offset with the scale taken
5539 // out, perform the transformation.
5540 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
5541 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
5542 Scale = ConstantSInt::get(C->getType(),
5543 (int64_t)C->getRawValue() /
5544 (int64_t)ArrayEltSize);
5546 Scale = ConstantUInt::get(Scale->getType(),
5547 Scale->getRawValue() / ArrayEltSize);
5548 if (Scale->getRawValue() != 1) {
5549 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
5550 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
5551 NewIdx = InsertNewInstBefore(Sc, GEP);
5554 // Insert the new GEP instruction.
5556 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5557 NewIdx, GEP.getName());
5558 Idx = InsertNewInstBefore(Idx, GEP);
5559 return new CastInst(Idx, GEP.getType());
5568 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
5569 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
5570 if (AI.isArrayAllocation()) // Check C != 1
5571 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
5572 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
5573 AllocationInst *New = 0;
5575 // Create and insert the replacement instruction...
5576 if (isa<MallocInst>(AI))
5577 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
5579 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
5580 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
5583 InsertNewInstBefore(New, AI);
5585 // Scan to the end of the allocation instructions, to skip over a block of
5586 // allocas if possible...
5588 BasicBlock::iterator It = New;
5589 while (isa<AllocationInst>(*It)) ++It;
5591 // Now that I is pointing to the first non-allocation-inst in the block,
5592 // insert our getelementptr instruction...
5594 Value *NullIdx = Constant::getNullValue(Type::IntTy);
5595 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
5596 New->getName()+".sub", It);
5598 // Now make everything use the getelementptr instead of the original
5600 return ReplaceInstUsesWith(AI, V);
5601 } else if (isa<UndefValue>(AI.getArraySize())) {
5602 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5605 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
5606 // Note that we only do this for alloca's, because malloc should allocate and
5607 // return a unique pointer, even for a zero byte allocation.
5608 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
5609 TD->getTypeSize(AI.getAllocatedType()) == 0)
5610 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
5615 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
5616 Value *Op = FI.getOperand(0);
5618 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
5619 if (CastInst *CI = dyn_cast<CastInst>(Op))
5620 if (isa<PointerType>(CI->getOperand(0)->getType())) {
5621 FI.setOperand(0, CI->getOperand(0));
5625 // free undef -> unreachable.
5626 if (isa<UndefValue>(Op)) {
5627 // Insert a new store to null because we cannot modify the CFG here.
5628 new StoreInst(ConstantBool::True,
5629 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
5630 return EraseInstFromFunction(FI);
5633 // If we have 'free null' delete the instruction. This can happen in stl code
5634 // when lots of inlining happens.
5635 if (isa<ConstantPointerNull>(Op))
5636 return EraseInstFromFunction(FI);
5642 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
5643 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
5644 User *CI = cast<User>(LI.getOperand(0));
5645 Value *CastOp = CI->getOperand(0);
5647 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5648 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5649 const Type *SrcPTy = SrcTy->getElementType();
5651 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5652 // If the source is an array, the code below will not succeed. Check to
5653 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5655 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5656 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5657 if (ASrcTy->getNumElements() != 0) {
5658 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5659 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5660 SrcTy = cast<PointerType>(CastOp->getType());
5661 SrcPTy = SrcTy->getElementType();
5664 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5665 // Do not allow turning this into a load of an integer, which is then
5666 // casted to a pointer, this pessimizes pointer analysis a lot.
5667 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
5668 IC.getTargetData().getTypeSize(SrcPTy) ==
5669 IC.getTargetData().getTypeSize(DestPTy)) {
5671 // Okay, we are casting from one integer or pointer type to another of
5672 // the same size. Instead of casting the pointer before the load, cast
5673 // the result of the loaded value.
5674 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
5676 LI.isVolatile()),LI);
5677 // Now cast the result of the load.
5678 return new CastInst(NewLoad, LI.getType());
5685 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
5686 /// from this value cannot trap. If it is not obviously safe to load from the
5687 /// specified pointer, we do a quick local scan of the basic block containing
5688 /// ScanFrom, to determine if the address is already accessed.
5689 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
5690 // If it is an alloca or global variable, it is always safe to load from.
5691 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
5693 // Otherwise, be a little bit agressive by scanning the local block where we
5694 // want to check to see if the pointer is already being loaded or stored
5695 // from/to. If so, the previous load or store would have already trapped,
5696 // so there is no harm doing an extra load (also, CSE will later eliminate
5697 // the load entirely).
5698 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
5703 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
5704 if (LI->getOperand(0) == V) return true;
5705 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5706 if (SI->getOperand(1) == V) return true;
5712 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
5713 Value *Op = LI.getOperand(0);
5715 // load (cast X) --> cast (load X) iff safe
5716 if (CastInst *CI = dyn_cast<CastInst>(Op))
5717 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5720 // None of the following transforms are legal for volatile loads.
5721 if (LI.isVolatile()) return 0;
5723 if (&LI.getParent()->front() != &LI) {
5724 BasicBlock::iterator BBI = &LI; --BBI;
5725 // If the instruction immediately before this is a store to the same
5726 // address, do a simple form of store->load forwarding.
5727 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
5728 if (SI->getOperand(1) == LI.getOperand(0))
5729 return ReplaceInstUsesWith(LI, SI->getOperand(0));
5730 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
5731 if (LIB->getOperand(0) == LI.getOperand(0))
5732 return ReplaceInstUsesWith(LI, LIB);
5735 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
5736 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
5737 isa<UndefValue>(GEPI->getOperand(0))) {
5738 // Insert a new store to null instruction before the load to indicate
5739 // that this code is not reachable. We do this instead of inserting
5740 // an unreachable instruction directly because we cannot modify the
5742 new StoreInst(UndefValue::get(LI.getType()),
5743 Constant::getNullValue(Op->getType()), &LI);
5744 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5747 if (Constant *C = dyn_cast<Constant>(Op)) {
5748 // load null/undef -> undef
5749 if ((C->isNullValue() || isa<UndefValue>(C))) {
5750 // Insert a new store to null instruction before the load to indicate that
5751 // this code is not reachable. We do this instead of inserting an
5752 // unreachable instruction directly because we cannot modify the CFG.
5753 new StoreInst(UndefValue::get(LI.getType()),
5754 Constant::getNullValue(Op->getType()), &LI);
5755 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5758 // Instcombine load (constant global) into the value loaded.
5759 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
5760 if (GV->isConstant() && !GV->isExternal())
5761 return ReplaceInstUsesWith(LI, GV->getInitializer());
5763 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
5764 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
5765 if (CE->getOpcode() == Instruction::GetElementPtr) {
5766 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
5767 if (GV->isConstant() && !GV->isExternal())
5769 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
5770 return ReplaceInstUsesWith(LI, V);
5771 if (CE->getOperand(0)->isNullValue()) {
5772 // Insert a new store to null instruction before the load to indicate
5773 // that this code is not reachable. We do this instead of inserting
5774 // an unreachable instruction directly because we cannot modify the
5776 new StoreInst(UndefValue::get(LI.getType()),
5777 Constant::getNullValue(Op->getType()), &LI);
5778 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
5781 } else if (CE->getOpcode() == Instruction::Cast) {
5782 if (Instruction *Res = InstCombineLoadCast(*this, LI))
5787 if (Op->hasOneUse()) {
5788 // Change select and PHI nodes to select values instead of addresses: this
5789 // helps alias analysis out a lot, allows many others simplifications, and
5790 // exposes redundancy in the code.
5792 // Note that we cannot do the transformation unless we know that the
5793 // introduced loads cannot trap! Something like this is valid as long as
5794 // the condition is always false: load (select bool %C, int* null, int* %G),
5795 // but it would not be valid if we transformed it to load from null
5798 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
5799 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
5800 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
5801 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
5802 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
5803 SI->getOperand(1)->getName()+".val"), LI);
5804 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
5805 SI->getOperand(2)->getName()+".val"), LI);
5806 return new SelectInst(SI->getCondition(), V1, V2);
5809 // load (select (cond, null, P)) -> load P
5810 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
5811 if (C->isNullValue()) {
5812 LI.setOperand(0, SI->getOperand(2));
5816 // load (select (cond, P, null)) -> load P
5817 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
5818 if (C->isNullValue()) {
5819 LI.setOperand(0, SI->getOperand(1));
5823 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
5824 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
5825 bool Safe = PN->getParent() == LI.getParent();
5827 // Scan all of the instructions between the PHI and the load to make
5828 // sure there are no instructions that might possibly alter the value
5829 // loaded from the PHI.
5831 BasicBlock::iterator I = &LI;
5832 for (--I; !isa<PHINode>(I); --I)
5833 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
5839 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
5840 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
5841 PN->getIncomingBlock(i)->getTerminator()))
5846 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
5847 InsertNewInstBefore(NewPN, *PN);
5848 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
5850 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
5851 BasicBlock *BB = PN->getIncomingBlock(i);
5852 Value *&TheLoad = LoadMap[BB];
5854 Value *InVal = PN->getIncomingValue(i);
5855 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
5856 InVal->getName()+".val"),
5857 *BB->getTerminator());
5859 NewPN->addIncoming(TheLoad, BB);
5861 return ReplaceInstUsesWith(LI, NewPN);
5868 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
5870 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
5871 User *CI = cast<User>(SI.getOperand(1));
5872 Value *CastOp = CI->getOperand(0);
5874 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
5875 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
5876 const Type *SrcPTy = SrcTy->getElementType();
5878 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
5879 // If the source is an array, the code below will not succeed. Check to
5880 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
5882 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
5883 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
5884 if (ASrcTy->getNumElements() != 0) {
5885 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
5886 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
5887 SrcTy = cast<PointerType>(CastOp->getType());
5888 SrcPTy = SrcTy->getElementType();
5891 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
5892 IC.getTargetData().getTypeSize(SrcPTy) ==
5893 IC.getTargetData().getTypeSize(DestPTy)) {
5895 // Okay, we are casting from one integer or pointer type to another of
5896 // the same size. Instead of casting the pointer before the store, cast
5897 // the value to be stored.
5899 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
5900 NewCast = ConstantExpr::getCast(C, SrcPTy);
5902 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
5904 SI.getOperand(0)->getName()+".c"), SI);
5906 return new StoreInst(NewCast, CastOp);
5913 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
5914 Value *Val = SI.getOperand(0);
5915 Value *Ptr = SI.getOperand(1);
5917 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
5918 EraseInstFromFunction(SI);
5923 // Do really simple DSE, to catch cases where there are several consequtive
5924 // stores to the same location, separated by a few arithmetic operations. This
5925 // situation often occurs with bitfield accesses.
5926 BasicBlock::iterator BBI = &SI;
5927 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
5931 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
5932 // Prev store isn't volatile, and stores to the same location?
5933 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
5936 EraseInstFromFunction(*PrevSI);
5942 // Don't skip over loads or things that can modify memory.
5943 if (BBI->mayWriteToMemory() || isa<LoadInst>(BBI))
5948 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
5950 // store X, null -> turns into 'unreachable' in SimplifyCFG
5951 if (isa<ConstantPointerNull>(Ptr)) {
5952 if (!isa<UndefValue>(Val)) {
5953 SI.setOperand(0, UndefValue::get(Val->getType()));
5954 if (Instruction *U = dyn_cast<Instruction>(Val))
5955 WorkList.push_back(U); // Dropped a use.
5958 return 0; // Do not modify these!
5961 // store undef, Ptr -> noop
5962 if (isa<UndefValue>(Val)) {
5963 EraseInstFromFunction(SI);
5968 // If the pointer destination is a cast, see if we can fold the cast into the
5970 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
5971 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5973 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
5974 if (CE->getOpcode() == Instruction::Cast)
5975 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
5979 // If this store is the last instruction in the basic block, and if the block
5980 // ends with an unconditional branch, try to move it to the successor block.
5982 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
5983 if (BI->isUnconditional()) {
5984 // Check to see if the successor block has exactly two incoming edges. If
5985 // so, see if the other predecessor contains a store to the same location.
5986 // if so, insert a PHI node (if needed) and move the stores down.
5987 BasicBlock *Dest = BI->getSuccessor(0);
5989 pred_iterator PI = pred_begin(Dest);
5990 BasicBlock *Other = 0;
5991 if (*PI != BI->getParent())
5994 if (PI != pred_end(Dest)) {
5995 if (*PI != BI->getParent())
6000 if (++PI != pred_end(Dest))
6003 if (Other) { // If only one other pred...
6004 BBI = Other->getTerminator();
6005 // Make sure this other block ends in an unconditional branch and that
6006 // there is an instruction before the branch.
6007 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
6008 BBI != Other->begin()) {
6010 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
6012 // If this instruction is a store to the same location.
6013 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
6014 // Okay, we know we can perform this transformation. Insert a PHI
6015 // node now if we need it.
6016 Value *MergedVal = OtherStore->getOperand(0);
6017 if (MergedVal != SI.getOperand(0)) {
6018 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
6019 PN->reserveOperandSpace(2);
6020 PN->addIncoming(SI.getOperand(0), SI.getParent());
6021 PN->addIncoming(OtherStore->getOperand(0), Other);
6022 MergedVal = InsertNewInstBefore(PN, Dest->front());
6025 // Advance to a place where it is safe to insert the new store and
6027 BBI = Dest->begin();
6028 while (isa<PHINode>(BBI)) ++BBI;
6029 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
6030 OtherStore->isVolatile()), *BBI);
6032 // Nuke the old stores.
6033 EraseInstFromFunction(SI);
6034 EraseInstFromFunction(*OtherStore);
6046 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
6047 // Change br (not X), label True, label False to: br X, label False, True
6049 BasicBlock *TrueDest;
6050 BasicBlock *FalseDest;
6051 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
6052 !isa<Constant>(X)) {
6053 // Swap Destinations and condition...
6055 BI.setSuccessor(0, FalseDest);
6056 BI.setSuccessor(1, TrueDest);
6060 // Cannonicalize setne -> seteq
6061 Instruction::BinaryOps Op; Value *Y;
6062 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
6063 TrueDest, FalseDest)))
6064 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
6065 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
6066 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
6067 std::string Name = I->getName(); I->setName("");
6068 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
6069 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
6070 // Swap Destinations and condition...
6071 BI.setCondition(NewSCC);
6072 BI.setSuccessor(0, FalseDest);
6073 BI.setSuccessor(1, TrueDest);
6074 removeFromWorkList(I);
6075 I->getParent()->getInstList().erase(I);
6076 WorkList.push_back(cast<Instruction>(NewSCC));
6083 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
6084 Value *Cond = SI.getCondition();
6085 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
6086 if (I->getOpcode() == Instruction::Add)
6087 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6088 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
6089 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
6090 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
6092 SI.setOperand(0, I->getOperand(0));
6093 WorkList.push_back(I);
6100 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
6101 if (ConstantAggregateZero *C =
6102 dyn_cast<ConstantAggregateZero>(EI.getOperand(0))) {
6103 // If packed val is constant 0, replace extract with scalar 0
6104 const Type *Ty = cast<PackedType>(C->getType())->getElementType();
6105 EI.replaceAllUsesWith(Constant::getNullValue(Ty));
6106 return ReplaceInstUsesWith(EI, Constant::getNullValue(Ty));
6108 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
6109 // If packed val is constant with uniform operands, replace EI
6110 // with that operand
6111 Constant *op0 = cast<Constant>(C->getOperand(0));
6112 for (unsigned i = 1; i < C->getNumOperands(); ++i)
6113 if (C->getOperand(i) != op0) return 0;
6114 return ReplaceInstUsesWith(EI, op0);
6116 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0)))
6117 if (I->hasOneUse()) {
6118 // Push extractelement into predecessor operation if legal and
6119 // profitable to do so
6120 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
6121 if (!isa<Constant>(BO->getOperand(0)) &&
6122 !isa<Constant>(BO->getOperand(1)))
6124 ExtractElementInst *newEI0 =
6125 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
6127 ExtractElementInst *newEI1 =
6128 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
6130 InsertNewInstBefore(newEI0, EI);
6131 InsertNewInstBefore(newEI1, EI);
6132 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
6134 switch(I->getOpcode()) {
6135 case Instruction::Load: {
6136 Value *Ptr = InsertCastBefore(I->getOperand(0),
6137 PointerType::get(EI.getType()), EI);
6138 GetElementPtrInst *GEP =
6139 new GetElementPtrInst(Ptr, EI.getOperand(1),
6140 I->getName() + ".gep");
6141 InsertNewInstBefore(GEP, EI);
6142 return new LoadInst(GEP);
6152 void InstCombiner::removeFromWorkList(Instruction *I) {
6153 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
6158 /// TryToSinkInstruction - Try to move the specified instruction from its
6159 /// current block into the beginning of DestBlock, which can only happen if it's
6160 /// safe to move the instruction past all of the instructions between it and the
6161 /// end of its block.
6162 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
6163 assert(I->hasOneUse() && "Invariants didn't hold!");
6165 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
6166 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
6168 // Do not sink alloca instructions out of the entry block.
6169 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
6172 // We can only sink load instructions if there is nothing between the load and
6173 // the end of block that could change the value.
6174 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6175 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
6177 if (Scan->mayWriteToMemory())
6181 BasicBlock::iterator InsertPos = DestBlock->begin();
6182 while (isa<PHINode>(InsertPos)) ++InsertPos;
6184 I->moveBefore(InsertPos);
6189 bool InstCombiner::runOnFunction(Function &F) {
6190 bool Changed = false;
6191 TD = &getAnalysis<TargetData>();
6194 // Populate the worklist with the reachable instructions.
6195 std::set<BasicBlock*> Visited;
6196 for (df_ext_iterator<BasicBlock*> BB = df_ext_begin(&F.front(), Visited),
6197 E = df_ext_end(&F.front(), Visited); BB != E; ++BB)
6198 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
6199 WorkList.push_back(I);
6201 // Do a quick scan over the function. If we find any blocks that are
6202 // unreachable, remove any instructions inside of them. This prevents
6203 // the instcombine code from having to deal with some bad special cases.
6204 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
6205 if (!Visited.count(BB)) {
6206 Instruction *Term = BB->getTerminator();
6207 while (Term != BB->begin()) { // Remove instrs bottom-up
6208 BasicBlock::iterator I = Term; --I;
6210 DEBUG(std::cerr << "IC: DCE: " << *I);
6213 if (!I->use_empty())
6214 I->replaceAllUsesWith(UndefValue::get(I->getType()));
6215 I->eraseFromParent();
6220 while (!WorkList.empty()) {
6221 Instruction *I = WorkList.back(); // Get an instruction from the worklist
6222 WorkList.pop_back();
6224 // Check to see if we can DCE or ConstantPropagate the instruction...
6225 // Check to see if we can DIE the instruction...
6226 if (isInstructionTriviallyDead(I)) {
6227 // Add operands to the worklist...
6228 if (I->getNumOperands() < 4)
6229 AddUsesToWorkList(*I);
6232 DEBUG(std::cerr << "IC: DCE: " << *I);
6234 I->eraseFromParent();
6235 removeFromWorkList(I);
6239 // Instruction isn't dead, see if we can constant propagate it...
6240 if (Constant *C = ConstantFoldInstruction(I)) {
6241 Value* Ptr = I->getOperand(0);
6242 if (isa<GetElementPtrInst>(I) &&
6243 cast<Constant>(Ptr)->isNullValue() &&
6244 !isa<ConstantPointerNull>(C) &&
6245 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
6246 // If this is a constant expr gep that is effectively computing an
6247 // "offsetof", fold it into 'cast int X to T*' instead of 'gep 0, 0, 12'
6248 bool isFoldableGEP = true;
6249 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
6250 if (!isa<ConstantInt>(I->getOperand(i)))
6251 isFoldableGEP = false;
6252 if (isFoldableGEP) {
6253 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
6254 std::vector<Value*>(I->op_begin()+1, I->op_end()));
6255 C = ConstantUInt::get(Type::ULongTy, Offset);
6256 C = ConstantExpr::getCast(C, TD->getIntPtrType());
6257 C = ConstantExpr::getCast(C, I->getType());
6261 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
6263 // Add operands to the worklist...
6264 AddUsesToWorkList(*I);
6265 ReplaceInstUsesWith(*I, C);
6268 I->getParent()->getInstList().erase(I);
6269 removeFromWorkList(I);
6273 // See if we can trivially sink this instruction to a successor basic block.
6274 if (I->hasOneUse()) {
6275 BasicBlock *BB = I->getParent();
6276 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
6277 if (UserParent != BB) {
6278 bool UserIsSuccessor = false;
6279 // See if the user is one of our successors.
6280 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
6281 if (*SI == UserParent) {
6282 UserIsSuccessor = true;
6286 // If the user is one of our immediate successors, and if that successor
6287 // only has us as a predecessors (we'd have to split the critical edge
6288 // otherwise), we can keep going.
6289 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
6290 next(pred_begin(UserParent)) == pred_end(UserParent))
6291 // Okay, the CFG is simple enough, try to sink this instruction.
6292 Changed |= TryToSinkInstruction(I, UserParent);
6296 // Now that we have an instruction, try combining it to simplify it...
6297 if (Instruction *Result = visit(*I)) {
6299 // Should we replace the old instruction with a new one?
6301 DEBUG(std::cerr << "IC: Old = " << *I
6302 << " New = " << *Result);
6304 // Everything uses the new instruction now.
6305 I->replaceAllUsesWith(Result);
6307 // Push the new instruction and any users onto the worklist.
6308 WorkList.push_back(Result);
6309 AddUsersToWorkList(*Result);
6311 // Move the name to the new instruction first...
6312 std::string OldName = I->getName(); I->setName("");
6313 Result->setName(OldName);
6315 // Insert the new instruction into the basic block...
6316 BasicBlock *InstParent = I->getParent();
6317 BasicBlock::iterator InsertPos = I;
6319 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
6320 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
6323 InstParent->getInstList().insert(InsertPos, Result);
6325 // Make sure that we reprocess all operands now that we reduced their
6327 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6328 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6329 WorkList.push_back(OpI);
6331 // Instructions can end up on the worklist more than once. Make sure
6332 // we do not process an instruction that has been deleted.
6333 removeFromWorkList(I);
6335 // Erase the old instruction.
6336 InstParent->getInstList().erase(I);
6338 DEBUG(std::cerr << "IC: MOD = " << *I);
6340 // If the instruction was modified, it's possible that it is now dead.
6341 // if so, remove it.
6342 if (isInstructionTriviallyDead(I)) {
6343 // Make sure we process all operands now that we are reducing their
6345 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6346 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6347 WorkList.push_back(OpI);
6349 // Instructions may end up in the worklist more than once. Erase all
6350 // occurrences of this instruction.
6351 removeFromWorkList(I);
6352 I->eraseFromParent();
6354 WorkList.push_back(Result);
6355 AddUsersToWorkList(*Result);
6365 FunctionPass *llvm::createInstructionCombiningPass() {
6366 return new InstCombiner();