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);
204 if (Instruction *I = dyn_cast<Instruction>(New))
205 WorkList.push_back(I);
209 // EraseInstFromFunction - When dealing with an instruction that has side
210 // effects or produces a void value, we can't rely on DCE to delete the
211 // instruction. Instead, visit methods should return the value returned by
213 Instruction *EraseInstFromFunction(Instruction &I) {
214 assert(I.use_empty() && "Cannot erase instruction that is used!");
215 AddUsesToWorkList(I);
216 removeFromWorkList(&I);
218 return 0; // Don't do anything with FI
222 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
223 /// InsertBefore instruction. This is specialized a bit to avoid inserting
224 /// casts that are known to not do anything...
226 Value *InsertOperandCastBefore(Value *V, const Type *DestTy,
227 Instruction *InsertBefore);
229 // SimplifyCommutative - This performs a few simplifications for commutative
231 bool SimplifyCommutative(BinaryOperator &I);
233 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
234 uint64_t &KnownZero, uint64_t &KnownOne,
237 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
238 // PHI node as operand #0, see if we can fold the instruction into the PHI
239 // (which is only possible if all operands to the PHI are constants).
240 Instruction *FoldOpIntoPhi(Instruction &I);
242 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
243 // operator and they all are only used by the PHI, PHI together their
244 // inputs, and do the operation once, to the result of the PHI.
245 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
247 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
248 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
250 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
251 bool isSub, Instruction &I);
252 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
253 bool Inside, Instruction &IB);
254 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
257 RegisterOpt<InstCombiner> X("instcombine", "Combine redundant instructions");
260 // getComplexity: Assign a complexity or rank value to LLVM Values...
261 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
262 static unsigned getComplexity(Value *V) {
263 if (isa<Instruction>(V)) {
264 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
268 if (isa<Argument>(V)) return 3;
269 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
272 // isOnlyUse - Return true if this instruction will be deleted if we stop using
274 static bool isOnlyUse(Value *V) {
275 return V->hasOneUse() || isa<Constant>(V);
278 // getPromotedType - Return the specified type promoted as it would be to pass
279 // though a va_arg area...
280 static const Type *getPromotedType(const Type *Ty) {
281 switch (Ty->getTypeID()) {
282 case Type::SByteTyID:
283 case Type::ShortTyID: return Type::IntTy;
284 case Type::UByteTyID:
285 case Type::UShortTyID: return Type::UIntTy;
286 case Type::FloatTyID: return Type::DoubleTy;
291 /// isCast - If the specified operand is a CastInst or a constant expr cast,
292 /// return the operand value, otherwise return null.
293 static Value *isCast(Value *V) {
294 if (CastInst *I = dyn_cast<CastInst>(V))
295 return I->getOperand(0);
296 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
297 if (CE->getOpcode() == Instruction::Cast)
298 return CE->getOperand(0);
302 // SimplifyCommutative - This performs a few simplifications for commutative
305 // 1. Order operands such that they are listed from right (least complex) to
306 // left (most complex). This puts constants before unary operators before
309 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
310 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
312 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
313 bool Changed = false;
314 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
315 Changed = !I.swapOperands();
317 if (!I.isAssociative()) return Changed;
318 Instruction::BinaryOps Opcode = I.getOpcode();
319 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
320 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
321 if (isa<Constant>(I.getOperand(1))) {
322 Constant *Folded = ConstantExpr::get(I.getOpcode(),
323 cast<Constant>(I.getOperand(1)),
324 cast<Constant>(Op->getOperand(1)));
325 I.setOperand(0, Op->getOperand(0));
326 I.setOperand(1, Folded);
328 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
329 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
330 isOnlyUse(Op) && isOnlyUse(Op1)) {
331 Constant *C1 = cast<Constant>(Op->getOperand(1));
332 Constant *C2 = cast<Constant>(Op1->getOperand(1));
334 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
335 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
336 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
339 WorkList.push_back(New);
340 I.setOperand(0, New);
341 I.setOperand(1, Folded);
348 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
349 // if the LHS is a constant zero (which is the 'negate' form).
351 static inline Value *dyn_castNegVal(Value *V) {
352 if (BinaryOperator::isNeg(V))
353 return BinaryOperator::getNegArgument(V);
355 // Constants can be considered to be negated values if they can be folded.
356 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
357 return ConstantExpr::getNeg(C);
361 static inline Value *dyn_castNotVal(Value *V) {
362 if (BinaryOperator::isNot(V))
363 return BinaryOperator::getNotArgument(V);
365 // Constants can be considered to be not'ed values...
366 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
367 return ConstantExpr::getNot(C);
371 // dyn_castFoldableMul - If this value is a multiply that can be folded into
372 // other computations (because it has a constant operand), return the
373 // non-constant operand of the multiply, and set CST to point to the multiplier.
374 // Otherwise, return null.
376 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
377 if (V->hasOneUse() && V->getType()->isInteger())
378 if (Instruction *I = dyn_cast<Instruction>(V)) {
379 if (I->getOpcode() == Instruction::Mul)
380 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
381 return I->getOperand(0);
382 if (I->getOpcode() == Instruction::Shl)
383 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
384 // The multiplier is really 1 << CST.
385 Constant *One = ConstantInt::get(V->getType(), 1);
386 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
387 return I->getOperand(0);
393 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
394 /// expression, return it.
395 static User *dyn_castGetElementPtr(Value *V) {
396 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
397 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
398 if (CE->getOpcode() == Instruction::GetElementPtr)
399 return cast<User>(V);
403 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
404 static ConstantInt *AddOne(ConstantInt *C) {
405 return cast<ConstantInt>(ConstantExpr::getAdd(C,
406 ConstantInt::get(C->getType(), 1)));
408 static ConstantInt *SubOne(ConstantInt *C) {
409 return cast<ConstantInt>(ConstantExpr::getSub(C,
410 ConstantInt::get(C->getType(), 1)));
413 /// GetConstantInType - Return a ConstantInt with the specified type and value.
415 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
416 if (Ty->isUnsigned())
417 return ConstantUInt::get(Ty, Val);
418 else if (Ty->getTypeID() == Type::BoolTyID)
419 return ConstantBool::get(Val);
421 SVal <<= 64-Ty->getPrimitiveSizeInBits();
422 SVal >>= 64-Ty->getPrimitiveSizeInBits();
423 return ConstantSInt::get(Ty, SVal);
427 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
428 /// known to be either zero or one and return them in the KnownZero/KnownOne
429 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
431 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
432 uint64_t &KnownOne, unsigned Depth = 0) {
433 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
434 // we cannot optimize based on the assumption that it is zero without changing
435 // it to be an explicit zero. If we don't change it to zero, other code could
436 // optimized based on the contradictory assumption that it is non-zero.
437 // Because instcombine aggressively folds operations with undef args anyway,
438 // this won't lose us code quality.
439 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
440 // We know all of the bits for a constant!
441 KnownOne = CI->getZExtValue() & Mask;
442 KnownZero = ~KnownOne & Mask;
446 KnownZero = KnownOne = 0; // Don't know anything.
447 if (Depth == 6 || Mask == 0)
448 return; // Limit search depth.
450 uint64_t KnownZero2, KnownOne2;
451 Instruction *I = dyn_cast<Instruction>(V);
454 switch (I->getOpcode()) {
455 case Instruction::And:
456 // If either the LHS or the RHS are Zero, the result is zero.
457 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
459 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
460 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
461 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
463 // Output known-1 bits are only known if set in both the LHS & RHS.
464 KnownOne &= KnownOne2;
465 // Output known-0 are known to be clear if zero in either the LHS | RHS.
466 KnownZero |= KnownZero2;
468 case Instruction::Or:
469 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
471 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
472 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
473 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
475 // Output known-0 bits are only known if clear in both the LHS & RHS.
476 KnownZero &= KnownZero2;
477 // Output known-1 are known to be set if set in either the LHS | RHS.
478 KnownOne |= KnownOne2;
480 case Instruction::Xor: {
481 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
482 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
483 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
484 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
486 // Output known-0 bits are known if clear or set in both the LHS & RHS.
487 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
488 // Output known-1 are known to be set if set in only one of the LHS, RHS.
489 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
490 KnownZero = KnownZeroOut;
493 case Instruction::Select:
494 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
495 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
496 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
497 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
499 // Only known if known in both the LHS and RHS.
500 KnownOne &= KnownOne2;
501 KnownZero &= KnownZero2;
503 case Instruction::Cast: {
504 const Type *SrcTy = I->getOperand(0)->getType();
505 if (!SrcTy->isIntegral()) return;
507 // If this is an integer truncate or noop, just look in the input.
508 if (SrcTy->getPrimitiveSizeInBits() >=
509 I->getType()->getPrimitiveSizeInBits()) {
510 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
514 // Sign or Zero extension. Compute the bits in the result that are not
515 // present in the input.
516 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
517 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
519 // Handle zero extension.
520 if (!SrcTy->isSigned()) {
521 Mask &= SrcTy->getIntegralTypeMask();
522 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
523 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
524 // The top bits are known to be zero.
525 KnownZero |= NewBits;
528 Mask &= SrcTy->getIntegralTypeMask();
529 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
530 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
532 // If the sign bit of the input is known set or clear, then we know the
533 // top bits of the result.
534 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
535 if (KnownZero & InSignBit) { // Input sign bit known zero
536 KnownZero |= NewBits;
537 KnownOne &= ~NewBits;
538 } else if (KnownOne & InSignBit) { // Input sign bit known set
540 KnownZero &= ~NewBits;
541 } else { // Input sign bit unknown
542 KnownZero &= ~NewBits;
543 KnownOne &= ~NewBits;
548 case Instruction::Shl:
549 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
550 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
551 Mask >>= SA->getValue();
552 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
553 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
554 KnownZero <<= SA->getValue();
555 KnownOne <<= SA->getValue();
556 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
560 case Instruction::Shr:
561 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
562 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
563 // Compute the new bits that are at the top now.
564 uint64_t HighBits = (1ULL << SA->getValue())-1;
565 HighBits <<= I->getType()->getPrimitiveSizeInBits()-SA->getValue();
567 if (I->getType()->isUnsigned()) { // Unsigned shift right.
568 Mask <<= SA->getValue();
569 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
570 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
571 KnownZero >>= SA->getValue();
572 KnownOne >>= SA->getValue();
573 KnownZero |= HighBits; // high bits known zero.
575 Mask <<= SA->getValue();
576 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
577 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
578 KnownZero >>= SA->getValue();
579 KnownOne >>= SA->getValue();
581 // Handle the sign bits.
582 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
583 SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
585 if (KnownZero & SignBit) { // New bits are known zero.
586 KnownZero |= HighBits;
587 } else if (KnownOne & SignBit) { // New bits are known one.
588 KnownOne |= HighBits;
597 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
598 /// this predicate to simplify operations downstream. Mask is known to be zero
599 /// for bits that V cannot have.
600 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
601 uint64_t KnownZero, KnownOne;
602 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
603 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
604 return (KnownZero & Mask) == Mask;
607 /// ShrinkDemandedConstant - Check to see if the specified operand of the
608 /// specified instruction is a constant integer. If so, check to see if there
609 /// are any bits set in the constant that are not demanded. If so, shrink the
610 /// constant and return true.
611 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
613 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
614 if (!OpC) return false;
616 // If there are no bits set that aren't demanded, nothing to do.
617 if ((~Demanded & OpC->getZExtValue()) == 0)
620 // This is producing any bits that are not needed, shrink the RHS.
621 uint64_t Val = Demanded & OpC->getZExtValue();
622 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
626 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
627 // set of known zero and one bits, compute the maximum and minimum values that
628 // could have the specified known zero and known one bits, returning them in
630 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
633 int64_t &Min, int64_t &Max) {
634 uint64_t TypeBits = Ty->getIntegralTypeMask();
635 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
637 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
639 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
640 // bit if it is unknown.
642 Max = KnownOne|UnknownBits;
644 if (SignBit & UnknownBits) { // Sign bit is unknown
649 // Sign extend the min/max values.
650 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
651 Min = (Min << ShAmt) >> ShAmt;
652 Max = (Max << ShAmt) >> ShAmt;
655 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
656 // a set of known zero and one bits, compute the maximum and minimum values that
657 // could have the specified known zero and known one bits, returning them in
659 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
664 uint64_t TypeBits = Ty->getIntegralTypeMask();
665 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
667 // The minimum value is when the unknown bits are all zeros.
669 // The maximum value is when the unknown bits are all ones.
670 Max = KnownOne|UnknownBits;
674 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
675 /// DemandedMask bits of the result of V are ever used downstream. If we can
676 /// use this information to simplify V, do so and return true. Otherwise,
677 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
678 /// the expression (used to simplify the caller). The KnownZero/One bits may
679 /// only be accurate for those bits in the DemandedMask.
680 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
681 uint64_t &KnownZero, uint64_t &KnownOne,
683 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
684 // We know all of the bits for a constant!
685 KnownOne = CI->getZExtValue() & DemandedMask;
686 KnownZero = ~KnownOne & DemandedMask;
690 KnownZero = KnownOne = 0;
691 if (!V->hasOneUse()) { // Other users may use these bits.
692 if (Depth != 0) { // Not at the root.
693 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
694 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
697 // If this is the root being simplified, allow it to have multiple uses,
698 // just set the DemandedMask to all bits.
699 DemandedMask = V->getType()->getIntegralTypeMask();
700 } else if (DemandedMask == 0) { // Not demanding any bits from V.
701 if (V != UndefValue::get(V->getType()))
702 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
704 } else if (Depth == 6) { // Limit search depth.
708 Instruction *I = dyn_cast<Instruction>(V);
709 if (!I) return false; // Only analyze instructions.
711 uint64_t KnownZero2, KnownOne2;
712 switch (I->getOpcode()) {
714 case Instruction::And:
715 // If either the LHS or the RHS are Zero, the result is zero.
716 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
717 KnownZero, KnownOne, Depth+1))
719 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
721 // If something is known zero on the RHS, the bits aren't demanded on the
723 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
724 KnownZero2, KnownOne2, Depth+1))
726 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
728 // If all of the demanded bits are known one on one side, return the other.
729 // These bits cannot contribute to the result of the 'and'.
730 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
731 return UpdateValueUsesWith(I, I->getOperand(0));
732 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
733 return UpdateValueUsesWith(I, I->getOperand(1));
735 // If all of the demanded bits in the inputs are known zeros, return zero.
736 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
737 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
739 // If the RHS is a constant, see if we can simplify it.
740 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
741 return UpdateValueUsesWith(I, I);
743 // Output known-1 bits are only known if set in both the LHS & RHS.
744 KnownOne &= KnownOne2;
745 // Output known-0 are known to be clear if zero in either the LHS | RHS.
746 KnownZero |= KnownZero2;
748 case Instruction::Or:
749 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
750 KnownZero, KnownOne, Depth+1))
752 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
753 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
754 KnownZero2, KnownOne2, Depth+1))
756 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
758 // If all of the demanded bits are known zero on one side, return the other.
759 // These bits cannot contribute to the result of the 'or'.
760 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
761 return UpdateValueUsesWith(I, I->getOperand(0));
762 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
763 return UpdateValueUsesWith(I, I->getOperand(1));
765 // If all of the potentially set bits on one side are known to be set on
766 // the other side, just use the 'other' side.
767 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
768 (DemandedMask & (~KnownZero)))
769 return UpdateValueUsesWith(I, I->getOperand(0));
770 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
771 (DemandedMask & (~KnownZero2)))
772 return UpdateValueUsesWith(I, I->getOperand(1));
774 // If the RHS is a constant, see if we can simplify it.
775 if (ShrinkDemandedConstant(I, 1, DemandedMask))
776 return UpdateValueUsesWith(I, I);
778 // Output known-0 bits are only known if clear in both the LHS & RHS.
779 KnownZero &= KnownZero2;
780 // Output known-1 are known to be set if set in either the LHS | RHS.
781 KnownOne |= KnownOne2;
783 case Instruction::Xor: {
784 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
785 KnownZero, KnownOne, Depth+1))
787 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
788 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
789 KnownZero2, KnownOne2, Depth+1))
791 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
793 // If all of the demanded bits are known zero on one side, return the other.
794 // These bits cannot contribute to the result of the 'xor'.
795 if ((DemandedMask & KnownZero) == DemandedMask)
796 return UpdateValueUsesWith(I, I->getOperand(0));
797 if ((DemandedMask & KnownZero2) == DemandedMask)
798 return UpdateValueUsesWith(I, I->getOperand(1));
800 // Output known-0 bits are known if clear or set in both the LHS & RHS.
801 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
802 // Output known-1 are known to be set if set in only one of the LHS, RHS.
803 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
805 // If all of the unknown bits are known to be zero on one side or the other
806 // (but not both) turn this into an *inclusive* or.
807 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
808 if (uint64_t UnknownBits = DemandedMask & ~(KnownZeroOut|KnownOneOut)) {
809 if ((UnknownBits & (KnownZero|KnownZero2)) == UnknownBits) {
811 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
813 InsertNewInstBefore(Or, *I);
814 return UpdateValueUsesWith(I, Or);
818 // If all of the demanded bits on one side are known, and all of the set
819 // bits on that side are also known to be set on the other side, turn this
820 // into an AND, as we know the bits will be cleared.
821 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
822 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
823 if ((KnownOne & KnownOne2) == KnownOne) {
824 Constant *AndC = GetConstantInType(I->getType(),
825 ~KnownOne & DemandedMask);
827 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
828 InsertNewInstBefore(And, *I);
829 return UpdateValueUsesWith(I, And);
833 // If the RHS is a constant, see if we can simplify it.
834 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
835 if (ShrinkDemandedConstant(I, 1, DemandedMask))
836 return UpdateValueUsesWith(I, I);
838 KnownZero = KnownZeroOut;
839 KnownOne = KnownOneOut;
842 case Instruction::Select:
843 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
844 KnownZero, KnownOne, Depth+1))
846 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
847 KnownZero2, KnownOne2, Depth+1))
849 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
850 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
852 // If the operands are constants, see if we can simplify them.
853 if (ShrinkDemandedConstant(I, 1, DemandedMask))
854 return UpdateValueUsesWith(I, I);
855 if (ShrinkDemandedConstant(I, 2, DemandedMask))
856 return UpdateValueUsesWith(I, I);
858 // Only known if known in both the LHS and RHS.
859 KnownOne &= KnownOne2;
860 KnownZero &= KnownZero2;
862 case Instruction::Cast: {
863 const Type *SrcTy = I->getOperand(0)->getType();
864 if (!SrcTy->isIntegral()) return false;
866 // If this is an integer truncate or noop, just look in the input.
867 if (SrcTy->getPrimitiveSizeInBits() >=
868 I->getType()->getPrimitiveSizeInBits()) {
869 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
870 KnownZero, KnownOne, Depth+1))
872 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
876 // Sign or Zero extension. Compute the bits in the result that are not
877 // present in the input.
878 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
879 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
881 // Handle zero extension.
882 if (!SrcTy->isSigned()) {
883 DemandedMask &= SrcTy->getIntegralTypeMask();
884 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
885 KnownZero, KnownOne, Depth+1))
887 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
888 // The top bits are known to be zero.
889 KnownZero |= NewBits;
892 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
893 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
895 // If any of the sign extended bits are demanded, we know that the sign
897 if (NewBits & DemandedMask)
898 InputDemandedBits |= InSignBit;
900 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
901 KnownZero, KnownOne, Depth+1))
903 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
905 // If the sign bit of the input is known set or clear, then we know the
906 // top bits of the result.
908 // If the input sign bit is known zero, or if the NewBits are not demanded
909 // convert this into a zero extension.
910 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
911 // Convert to unsigned first.
913 NewVal = new CastInst(I->getOperand(0), SrcTy->getUnsignedVersion(),
914 I->getOperand(0)->getName());
915 InsertNewInstBefore(NewVal, *I);
916 // Then cast that to the destination type.
917 NewVal = new CastInst(NewVal, I->getType(), I->getName());
918 InsertNewInstBefore(NewVal, *I);
919 return UpdateValueUsesWith(I, NewVal);
920 } else if (KnownOne & InSignBit) { // Input sign bit known set
922 KnownZero &= ~NewBits;
923 } else { // Input sign bit unknown
924 KnownZero &= ~NewBits;
925 KnownOne &= ~NewBits;
930 case Instruction::Shl:
931 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
932 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> SA->getValue(),
933 KnownZero, KnownOne, Depth+1))
935 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
936 KnownZero <<= SA->getValue();
937 KnownOne <<= SA->getValue();
938 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
941 case Instruction::Shr:
942 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1))) {
943 unsigned ShAmt = SA->getValue();
945 // Compute the new bits that are at the top now.
946 uint64_t HighBits = (1ULL << ShAmt)-1;
947 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShAmt;
948 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
949 if (I->getType()->isUnsigned()) { // Unsigned shift right.
950 if (SimplifyDemandedBits(I->getOperand(0),
951 (DemandedMask << ShAmt) & TypeMask,
952 KnownZero, KnownOne, Depth+1))
954 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
955 KnownZero &= TypeMask;
956 KnownOne &= TypeMask;
959 KnownZero |= HighBits; // high bits known zero.
960 } else { // Signed shift right.
961 if (SimplifyDemandedBits(I->getOperand(0),
962 (DemandedMask << ShAmt) & TypeMask,
963 KnownZero, KnownOne, Depth+1))
965 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
966 KnownZero &= TypeMask;
967 KnownOne &= TypeMask;
968 KnownZero >>= SA->getValue();
969 KnownOne >>= SA->getValue();
971 // Handle the sign bits.
972 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
973 SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
975 // If the input sign bit is known to be zero, or if none of the top bits
976 // are demanded, turn this into an unsigned shift right.
977 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
978 // Convert the input to unsigned.
980 NewVal = new CastInst(I->getOperand(0),
981 I->getType()->getUnsignedVersion(),
982 I->getOperand(0)->getName());
983 InsertNewInstBefore(NewVal, *I);
984 // Perform the unsigned shift right.
985 NewVal = new ShiftInst(Instruction::Shr, NewVal, SA, I->getName());
986 InsertNewInstBefore(NewVal, *I);
987 // Then cast that to the destination type.
988 NewVal = new CastInst(NewVal, I->getType(), I->getName());
989 InsertNewInstBefore(NewVal, *I);
990 return UpdateValueUsesWith(I, NewVal);
991 } else if (KnownOne & SignBit) { // New bits are known one.
992 KnownOne |= HighBits;
999 // If the client is only demanding bits that we know, return the known
1001 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1002 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1006 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
1007 // true when both operands are equal...
1009 static bool isTrueWhenEqual(Instruction &I) {
1010 return I.getOpcode() == Instruction::SetEQ ||
1011 I.getOpcode() == Instruction::SetGE ||
1012 I.getOpcode() == Instruction::SetLE;
1015 /// AssociativeOpt - Perform an optimization on an associative operator. This
1016 /// function is designed to check a chain of associative operators for a
1017 /// potential to apply a certain optimization. Since the optimization may be
1018 /// applicable if the expression was reassociated, this checks the chain, then
1019 /// reassociates the expression as necessary to expose the optimization
1020 /// opportunity. This makes use of a special Functor, which must define
1021 /// 'shouldApply' and 'apply' methods.
1023 template<typename Functor>
1024 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1025 unsigned Opcode = Root.getOpcode();
1026 Value *LHS = Root.getOperand(0);
1028 // Quick check, see if the immediate LHS matches...
1029 if (F.shouldApply(LHS))
1030 return F.apply(Root);
1032 // Otherwise, if the LHS is not of the same opcode as the root, return.
1033 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1034 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1035 // Should we apply this transform to the RHS?
1036 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1038 // If not to the RHS, check to see if we should apply to the LHS...
1039 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1040 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1044 // If the functor wants to apply the optimization to the RHS of LHSI,
1045 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1047 BasicBlock *BB = Root.getParent();
1049 // Now all of the instructions are in the current basic block, go ahead
1050 // and perform the reassociation.
1051 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1053 // First move the selected RHS to the LHS of the root...
1054 Root.setOperand(0, LHSI->getOperand(1));
1056 // Make what used to be the LHS of the root be the user of the root...
1057 Value *ExtraOperand = TmpLHSI->getOperand(1);
1058 if (&Root == TmpLHSI) {
1059 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1062 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1063 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1064 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1065 BasicBlock::iterator ARI = &Root; ++ARI;
1066 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1069 // Now propagate the ExtraOperand down the chain of instructions until we
1071 while (TmpLHSI != LHSI) {
1072 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1073 // Move the instruction to immediately before the chain we are
1074 // constructing to avoid breaking dominance properties.
1075 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1076 BB->getInstList().insert(ARI, NextLHSI);
1079 Value *NextOp = NextLHSI->getOperand(1);
1080 NextLHSI->setOperand(1, ExtraOperand);
1082 ExtraOperand = NextOp;
1085 // Now that the instructions are reassociated, have the functor perform
1086 // the transformation...
1087 return F.apply(Root);
1090 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1096 // AddRHS - Implements: X + X --> X << 1
1099 AddRHS(Value *rhs) : RHS(rhs) {}
1100 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1101 Instruction *apply(BinaryOperator &Add) const {
1102 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1103 ConstantInt::get(Type::UByteTy, 1));
1107 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1109 struct AddMaskingAnd {
1111 AddMaskingAnd(Constant *c) : C2(c) {}
1112 bool shouldApply(Value *LHS) const {
1114 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1115 ConstantExpr::getAnd(C1, C2)->isNullValue();
1117 Instruction *apply(BinaryOperator &Add) const {
1118 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1122 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1124 if (isa<CastInst>(I)) {
1125 if (Constant *SOC = dyn_cast<Constant>(SO))
1126 return ConstantExpr::getCast(SOC, I.getType());
1128 return IC->InsertNewInstBefore(new CastInst(SO, I.getType(),
1129 SO->getName() + ".cast"), I);
1132 // Figure out if the constant is the left or the right argument.
1133 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1134 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1136 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1138 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1139 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1142 Value *Op0 = SO, *Op1 = ConstOperand;
1144 std::swap(Op0, Op1);
1146 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1147 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1148 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1149 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1151 assert(0 && "Unknown binary instruction type!");
1154 return IC->InsertNewInstBefore(New, I);
1157 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1158 // constant as the other operand, try to fold the binary operator into the
1159 // select arguments. This also works for Cast instructions, which obviously do
1160 // not have a second operand.
1161 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1163 // Don't modify shared select instructions
1164 if (!SI->hasOneUse()) return 0;
1165 Value *TV = SI->getOperand(1);
1166 Value *FV = SI->getOperand(2);
1168 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1169 // Bool selects with constant operands can be folded to logical ops.
1170 if (SI->getType() == Type::BoolTy) return 0;
1172 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1173 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1175 return new SelectInst(SI->getCondition(), SelectTrueVal,
1182 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1183 /// node as operand #0, see if we can fold the instruction into the PHI (which
1184 /// is only possible if all operands to the PHI are constants).
1185 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1186 PHINode *PN = cast<PHINode>(I.getOperand(0));
1187 unsigned NumPHIValues = PN->getNumIncomingValues();
1188 if (!PN->hasOneUse() || NumPHIValues == 0 ||
1189 !isa<Constant>(PN->getIncomingValue(0))) return 0;
1191 // Check to see if all of the operands of the PHI are constants. If not, we
1192 // cannot do the transformation.
1193 for (unsigned i = 1; i != NumPHIValues; ++i)
1194 if (!isa<Constant>(PN->getIncomingValue(i)))
1197 // Okay, we can do the transformation: create the new PHI node.
1198 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1200 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1201 InsertNewInstBefore(NewPN, *PN);
1203 // Next, add all of the operands to the PHI.
1204 if (I.getNumOperands() == 2) {
1205 Constant *C = cast<Constant>(I.getOperand(1));
1206 for (unsigned i = 0; i != NumPHIValues; ++i) {
1207 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
1208 NewPN->addIncoming(ConstantExpr::get(I.getOpcode(), InV, C),
1209 PN->getIncomingBlock(i));
1212 assert(isa<CastInst>(I) && "Unary op should be a cast!");
1213 const Type *RetTy = I.getType();
1214 for (unsigned i = 0; i != NumPHIValues; ++i) {
1215 Constant *InV = cast<Constant>(PN->getIncomingValue(i));
1216 NewPN->addIncoming(ConstantExpr::getCast(InV, RetTy),
1217 PN->getIncomingBlock(i));
1220 return ReplaceInstUsesWith(I, NewPN);
1223 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1224 bool Changed = SimplifyCommutative(I);
1225 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1227 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1228 // X + undef -> undef
1229 if (isa<UndefValue>(RHS))
1230 return ReplaceInstUsesWith(I, RHS);
1233 if (!I.getType()->isFloatingPoint()) { // NOTE: -0 + +0 = +0.
1234 if (RHSC->isNullValue())
1235 return ReplaceInstUsesWith(I, LHS);
1236 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1237 if (CFP->isExactlyValue(-0.0))
1238 return ReplaceInstUsesWith(I, LHS);
1241 // X + (signbit) --> X ^ signbit
1242 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1243 uint64_t Val = CI->getZExtValue();
1244 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1245 return BinaryOperator::createXor(LHS, RHS);
1248 if (isa<PHINode>(LHS))
1249 if (Instruction *NV = FoldOpIntoPhi(I))
1252 ConstantInt *XorRHS = 0;
1254 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1255 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1256 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1257 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1259 uint64_t C0080Val = 1ULL << 31;
1260 int64_t CFF80Val = -C0080Val;
1263 if (TySizeBits > Size) {
1265 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1266 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1267 if (RHSSExt == CFF80Val) {
1268 if (XorRHS->getZExtValue() == C0080Val)
1270 } else if (RHSZExt == C0080Val) {
1271 if (XorRHS->getSExtValue() == CFF80Val)
1275 // This is a sign extend if the top bits are known zero.
1276 uint64_t Mask = ~0ULL;
1277 Mask <<= 64-(TySizeBits-Size);
1278 Mask &= XorLHS->getType()->getIntegralTypeMask();
1279 if (!MaskedValueIsZero(XorLHS, Mask))
1280 Size = 0; // Not a sign ext, but can't be any others either.
1287 } while (Size >= 8);
1290 const Type *MiddleType = 0;
1293 case 32: MiddleType = Type::IntTy; break;
1294 case 16: MiddleType = Type::ShortTy; break;
1295 case 8: MiddleType = Type::SByteTy; break;
1298 Instruction *NewTrunc = new CastInst(XorLHS, MiddleType, "sext");
1299 InsertNewInstBefore(NewTrunc, I);
1300 return new CastInst(NewTrunc, I.getType());
1306 if (I.getType()->isInteger()) {
1307 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1309 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1310 if (RHSI->getOpcode() == Instruction::Sub)
1311 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1312 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1314 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1315 if (LHSI->getOpcode() == Instruction::Sub)
1316 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1317 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1322 if (Value *V = dyn_castNegVal(LHS))
1323 return BinaryOperator::createSub(RHS, V);
1326 if (!isa<Constant>(RHS))
1327 if (Value *V = dyn_castNegVal(RHS))
1328 return BinaryOperator::createSub(LHS, V);
1332 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1333 if (X == RHS) // X*C + X --> X * (C+1)
1334 return BinaryOperator::createMul(RHS, AddOne(C2));
1336 // X*C1 + X*C2 --> X * (C1+C2)
1338 if (X == dyn_castFoldableMul(RHS, C1))
1339 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1342 // X + X*C --> X * (C+1)
1343 if (dyn_castFoldableMul(RHS, C2) == LHS)
1344 return BinaryOperator::createMul(LHS, AddOne(C2));
1347 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1348 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1349 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1351 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1353 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1354 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1355 return BinaryOperator::createSub(C, X);
1358 // (X & FF00) + xx00 -> (X+xx00) & FF00
1359 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1360 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1361 if (Anded == CRHS) {
1362 // See if all bits from the first bit set in the Add RHS up are included
1363 // in the mask. First, get the rightmost bit.
1364 uint64_t AddRHSV = CRHS->getRawValue();
1366 // Form a mask of all bits from the lowest bit added through the top.
1367 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1368 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1370 // See if the and mask includes all of these bits.
1371 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getRawValue();
1373 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1374 // Okay, the xform is safe. Insert the new add pronto.
1375 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1376 LHS->getName()), I);
1377 return BinaryOperator::createAnd(NewAdd, C2);
1382 // Try to fold constant add into select arguments.
1383 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1384 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1388 return Changed ? &I : 0;
1391 // isSignBit - Return true if the value represented by the constant only has the
1392 // highest order bit set.
1393 static bool isSignBit(ConstantInt *CI) {
1394 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1395 return (CI->getRawValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1398 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1400 static Value *RemoveNoopCast(Value *V) {
1401 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1402 const Type *CTy = CI->getType();
1403 const Type *OpTy = CI->getOperand(0)->getType();
1404 if (CTy->isInteger() && OpTy->isInteger()) {
1405 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1406 return RemoveNoopCast(CI->getOperand(0));
1407 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1408 return RemoveNoopCast(CI->getOperand(0));
1413 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1414 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1416 if (Op0 == Op1) // sub X, X -> 0
1417 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1419 // If this is a 'B = x-(-A)', change to B = x+A...
1420 if (Value *V = dyn_castNegVal(Op1))
1421 return BinaryOperator::createAdd(Op0, V);
1423 if (isa<UndefValue>(Op0))
1424 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1425 if (isa<UndefValue>(Op1))
1426 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1428 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1429 // Replace (-1 - A) with (~A)...
1430 if (C->isAllOnesValue())
1431 return BinaryOperator::createNot(Op1);
1433 // C - ~X == X + (1+C)
1435 if (match(Op1, m_Not(m_Value(X))))
1436 return BinaryOperator::createAdd(X,
1437 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1438 // -((uint)X >> 31) -> ((int)X >> 31)
1439 // -((int)X >> 31) -> ((uint)X >> 31)
1440 if (C->isNullValue()) {
1441 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1442 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1443 if (SI->getOpcode() == Instruction::Shr)
1444 if (ConstantUInt *CU = dyn_cast<ConstantUInt>(SI->getOperand(1))) {
1446 if (SI->getType()->isSigned())
1447 NewTy = SI->getType()->getUnsignedVersion();
1449 NewTy = SI->getType()->getSignedVersion();
1450 // Check to see if we are shifting out everything but the sign bit.
1451 if (CU->getValue() == SI->getType()->getPrimitiveSizeInBits()-1) {
1452 // Ok, the transformation is safe. Insert a cast of the incoming
1453 // value, then the new shift, then the new cast.
1454 Instruction *FirstCast = new CastInst(SI->getOperand(0), NewTy,
1455 SI->getOperand(0)->getName());
1456 Value *InV = InsertNewInstBefore(FirstCast, I);
1457 Instruction *NewShift = new ShiftInst(Instruction::Shr, FirstCast,
1459 if (NewShift->getType() == I.getType())
1462 InV = InsertNewInstBefore(NewShift, I);
1463 return new CastInst(NewShift, I.getType());
1469 // Try to fold constant sub into select arguments.
1470 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1471 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1474 if (isa<PHINode>(Op0))
1475 if (Instruction *NV = FoldOpIntoPhi(I))
1479 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1480 if (Op1I->getOpcode() == Instruction::Add &&
1481 !Op0->getType()->isFloatingPoint()) {
1482 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1483 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1484 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1485 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
1486 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1487 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1488 // C1-(X+C2) --> (C1-C2)-X
1489 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
1490 Op1I->getOperand(0));
1494 if (Op1I->hasOneUse()) {
1495 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1496 // is not used by anyone else...
1498 if (Op1I->getOpcode() == Instruction::Sub &&
1499 !Op1I->getType()->isFloatingPoint()) {
1500 // Swap the two operands of the subexpr...
1501 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1502 Op1I->setOperand(0, IIOp1);
1503 Op1I->setOperand(1, IIOp0);
1505 // Create the new top level add instruction...
1506 return BinaryOperator::createAdd(Op0, Op1);
1509 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1511 if (Op1I->getOpcode() == Instruction::And &&
1512 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1513 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1516 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
1517 return BinaryOperator::createAnd(Op0, NewNot);
1520 // -(X sdiv C) -> (X sdiv -C)
1521 if (Op1I->getOpcode() == Instruction::Div)
1522 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
1523 if (CSI->isNullValue())
1524 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1525 return BinaryOperator::createDiv(Op1I->getOperand(0),
1526 ConstantExpr::getNeg(DivRHS));
1528 // X - X*C --> X * (1-C)
1529 ConstantInt *C2 = 0;
1530 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1532 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
1533 return BinaryOperator::createMul(Op0, CP1);
1538 if (!Op0->getType()->isFloatingPoint())
1539 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1540 if (Op0I->getOpcode() == Instruction::Add) {
1541 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1542 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1543 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1544 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1545 } else if (Op0I->getOpcode() == Instruction::Sub) {
1546 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1547 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
1551 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1552 if (X == Op1) { // X*C - X --> X * (C-1)
1553 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
1554 return BinaryOperator::createMul(Op1, CP1);
1557 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1558 if (X == dyn_castFoldableMul(Op1, C2))
1559 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
1564 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
1565 /// really just returns true if the most significant (sign) bit is set.
1566 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
1567 if (RHS->getType()->isSigned()) {
1568 // True if source is LHS < 0 or LHS <= -1
1569 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
1570 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
1572 ConstantUInt *RHSC = cast<ConstantUInt>(RHS);
1573 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
1574 // the size of the integer type.
1575 if (Opcode == Instruction::SetGE)
1576 return RHSC->getValue() ==
1577 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
1578 if (Opcode == Instruction::SetGT)
1579 return RHSC->getValue() ==
1580 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
1585 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1586 bool Changed = SimplifyCommutative(I);
1587 Value *Op0 = I.getOperand(0);
1589 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
1590 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1592 // Simplify mul instructions with a constant RHS...
1593 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
1594 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1596 // ((X << C1)*C2) == (X * (C2 << C1))
1597 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
1598 if (SI->getOpcode() == Instruction::Shl)
1599 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1600 return BinaryOperator::createMul(SI->getOperand(0),
1601 ConstantExpr::getShl(CI, ShOp));
1603 if (CI->isNullValue())
1604 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
1605 if (CI->equalsInt(1)) // X * 1 == X
1606 return ReplaceInstUsesWith(I, Op0);
1607 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1608 return BinaryOperator::createNeg(Op0, I.getName());
1610 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getRawValue();
1611 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
1612 uint64_t C = Log2_64(Val);
1613 return new ShiftInst(Instruction::Shl, Op0,
1614 ConstantUInt::get(Type::UByteTy, C));
1616 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
1617 if (Op1F->isNullValue())
1618 return ReplaceInstUsesWith(I, Op1);
1620 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1621 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1622 if (Op1F->getValue() == 1.0)
1623 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1626 // Try to fold constant mul into select arguments.
1627 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1628 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1631 if (isa<PHINode>(Op0))
1632 if (Instruction *NV = FoldOpIntoPhi(I))
1636 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1637 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
1638 return BinaryOperator::createMul(Op0v, Op1v);
1640 // If one of the operands of the multiply is a cast from a boolean value, then
1641 // we know the bool is either zero or one, so this is a 'masking' multiply.
1642 // See if we can simplify things based on how the boolean was originally
1644 CastInst *BoolCast = 0;
1645 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(0)))
1646 if (CI->getOperand(0)->getType() == Type::BoolTy)
1649 if (CastInst *CI = dyn_cast<CastInst>(I.getOperand(1)))
1650 if (CI->getOperand(0)->getType() == Type::BoolTy)
1653 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
1654 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
1655 const Type *SCOpTy = SCIOp0->getType();
1657 // If the setcc is true iff the sign bit of X is set, then convert this
1658 // multiply into a shift/and combination.
1659 if (isa<ConstantInt>(SCIOp1) &&
1660 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
1661 // Shift the X value right to turn it into "all signbits".
1662 Constant *Amt = ConstantUInt::get(Type::UByteTy,
1663 SCOpTy->getPrimitiveSizeInBits()-1);
1664 if (SCIOp0->getType()->isUnsigned()) {
1665 const Type *NewTy = SCIOp0->getType()->getSignedVersion();
1666 SCIOp0 = InsertNewInstBefore(new CastInst(SCIOp0, NewTy,
1667 SCIOp0->getName()), I);
1671 InsertNewInstBefore(new ShiftInst(Instruction::Shr, SCIOp0, Amt,
1672 BoolCast->getOperand(0)->getName()+
1675 // If the multiply type is not the same as the source type, sign extend
1676 // or truncate to the multiply type.
1677 if (I.getType() != V->getType())
1678 V = InsertNewInstBefore(new CastInst(V, I.getType(), V->getName()),I);
1680 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
1681 return BinaryOperator::createAnd(V, OtherOp);
1686 return Changed ? &I : 0;
1689 Instruction *InstCombiner::visitDiv(BinaryOperator &I) {
1690 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1692 if (isa<UndefValue>(Op0)) // undef / X -> 0
1693 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1694 if (isa<UndefValue>(Op1))
1695 return ReplaceInstUsesWith(I, Op1); // X / undef -> undef
1697 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1699 if (RHS->equalsInt(1))
1700 return ReplaceInstUsesWith(I, Op0);
1703 if (RHS->isAllOnesValue())
1704 return BinaryOperator::createNeg(Op0);
1706 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1707 if (LHS->getOpcode() == Instruction::Div)
1708 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1709 // (X / C1) / C2 -> X / (C1*C2)
1710 return BinaryOperator::createDiv(LHS->getOperand(0),
1711 ConstantExpr::getMul(RHS, LHSRHS));
1714 // Check to see if this is an unsigned division with an exact power of 2,
1715 // if so, convert to a right shift.
1716 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1717 if (uint64_t Val = C->getValue()) // Don't break X / 0
1718 if (isPowerOf2_64(Val)) {
1719 uint64_t C = Log2_64(Val);
1720 return new ShiftInst(Instruction::Shr, Op0,
1721 ConstantUInt::get(Type::UByteTy, C));
1725 if (RHS->getType()->isSigned())
1726 if (Value *LHSNeg = dyn_castNegVal(Op0))
1727 return BinaryOperator::createDiv(LHSNeg, ConstantExpr::getNeg(RHS));
1729 if (!RHS->isNullValue()) {
1730 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1731 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1733 if (isa<PHINode>(Op0))
1734 if (Instruction *NV = FoldOpIntoPhi(I))
1739 // If this is 'udiv X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1740 // transform this into: '(Cond ? (udiv X, C1) : (udiv X, C2))'.
1741 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1742 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1743 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1744 if (STO->getValue() == 0) { // Couldn't be this argument.
1745 I.setOperand(1, SFO);
1747 } else if (SFO->getValue() == 0) {
1748 I.setOperand(1, STO);
1752 uint64_t TVA = STO->getValue(), FVA = SFO->getValue();
1753 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
1754 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
1755 Constant *TC = ConstantUInt::get(Type::UByteTy, TSA);
1756 Instruction *TSI = new ShiftInst(Instruction::Shr, Op0,
1757 TC, SI->getName()+".t");
1758 TSI = InsertNewInstBefore(TSI, I);
1760 Constant *FC = ConstantUInt::get(Type::UByteTy, FSA);
1761 Instruction *FSI = new ShiftInst(Instruction::Shr, Op0,
1762 FC, SI->getName()+".f");
1763 FSI = InsertNewInstBefore(FSI, I);
1764 return new SelectInst(SI->getOperand(0), TSI, FSI);
1768 // 0 / X == 0, we don't need to preserve faults!
1769 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1770 if (LHS->equalsInt(0))
1771 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1773 if (I.getType()->isSigned()) {
1774 // If the sign bits of both operands are zero (i.e. we can prove they are
1775 // unsigned inputs), turn this into a udiv.
1776 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
1777 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1778 const Type *NTy = Op0->getType()->getUnsignedVersion();
1779 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1780 InsertNewInstBefore(LHS, I);
1782 if (Constant *R = dyn_cast<Constant>(Op1))
1783 RHS = ConstantExpr::getCast(R, NTy);
1785 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1786 Instruction *Div = BinaryOperator::createDiv(LHS, RHS, I.getName());
1787 InsertNewInstBefore(Div, I);
1788 return new CastInst(Div, I.getType());
1791 // Known to be an unsigned division.
1792 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1793 // Turn A / (C1 << N), where C1 is "1<<C2" into A >> (N+C2) [udiv only].
1794 if (RHSI->getOpcode() == Instruction::Shl &&
1795 isa<ConstantUInt>(RHSI->getOperand(0))) {
1796 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1797 if (isPowerOf2_64(C1)) {
1798 unsigned C2 = Log2_64(C1);
1799 Value *Add = RHSI->getOperand(1);
1801 Constant *C2V = ConstantUInt::get(Add->getType(), C2);
1802 Add = InsertNewInstBefore(BinaryOperator::createAdd(Add, C2V,
1805 return new ShiftInst(Instruction::Shr, Op0, Add);
1815 /// GetFactor - If we can prove that the specified value is at least a multiple
1816 /// of some factor, return that factor.
1817 static Constant *GetFactor(Value *V) {
1818 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1821 // Unless we can be tricky, we know this is a multiple of 1.
1822 Constant *Result = ConstantInt::get(V->getType(), 1);
1824 Instruction *I = dyn_cast<Instruction>(V);
1825 if (!I) return Result;
1827 if (I->getOpcode() == Instruction::Mul) {
1828 // Handle multiplies by a constant, etc.
1829 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
1830 GetFactor(I->getOperand(1)));
1831 } else if (I->getOpcode() == Instruction::Shl) {
1832 // (X<<C) -> X * (1 << C)
1833 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
1834 ShRHS = ConstantExpr::getShl(Result, ShRHS);
1835 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
1837 } else if (I->getOpcode() == Instruction::And) {
1838 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1839 // X & 0xFFF0 is known to be a multiple of 16.
1840 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
1841 if (Zeros != V->getType()->getPrimitiveSizeInBits())
1842 return ConstantExpr::getShl(Result,
1843 ConstantUInt::get(Type::UByteTy, Zeros));
1845 } else if (I->getOpcode() == Instruction::Cast) {
1846 Value *Op = I->getOperand(0);
1847 // Only handle int->int casts.
1848 if (!Op->getType()->isInteger()) return Result;
1849 return ConstantExpr::getCast(GetFactor(Op), V->getType());
1854 Instruction *InstCombiner::visitRem(BinaryOperator &I) {
1855 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1857 // 0 % X == 0, we don't need to preserve faults!
1858 if (Constant *LHS = dyn_cast<Constant>(Op0))
1859 if (LHS->isNullValue())
1860 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1862 if (isa<UndefValue>(Op0)) // undef % X -> 0
1863 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1864 if (isa<UndefValue>(Op1))
1865 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
1867 if (I.getType()->isSigned()) {
1868 if (Value *RHSNeg = dyn_castNegVal(Op1))
1869 if (!isa<ConstantSInt>(RHSNeg) ||
1870 cast<ConstantSInt>(RHSNeg)->getValue() > 0) {
1872 AddUsesToWorkList(I);
1873 I.setOperand(1, RHSNeg);
1877 // If the top bits of both operands are zero (i.e. we can prove they are
1878 // unsigned inputs), turn this into a urem.
1879 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
1880 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1881 const Type *NTy = Op0->getType()->getUnsignedVersion();
1882 Instruction *LHS = new CastInst(Op0, NTy, Op0->getName());
1883 InsertNewInstBefore(LHS, I);
1885 if (Constant *R = dyn_cast<Constant>(Op1))
1886 RHS = ConstantExpr::getCast(R, NTy);
1888 RHS = InsertNewInstBefore(new CastInst(Op1, NTy, Op1->getName()), I);
1889 Instruction *Rem = BinaryOperator::createRem(LHS, RHS, I.getName());
1890 InsertNewInstBefore(Rem, I);
1891 return new CastInst(Rem, I.getType());
1895 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1896 // X % 0 == undef, we don't need to preserve faults!
1897 if (RHS->equalsInt(0))
1898 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
1900 if (RHS->equalsInt(1)) // X % 1 == 0
1901 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1903 // Check to see if this is an unsigned remainder with an exact power of 2,
1904 // if so, convert to a bitwise and.
1905 if (ConstantUInt *C = dyn_cast<ConstantUInt>(RHS))
1906 if (isPowerOf2_64(C->getValue()))
1907 return BinaryOperator::createAnd(Op0, SubOne(C));
1909 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1910 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1911 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1913 } else if (isa<PHINode>(Op0I)) {
1914 if (Instruction *NV = FoldOpIntoPhi(I))
1918 // X*C1%C2 --> 0 iff C1%C2 == 0
1919 if (ConstantExpr::getRem(GetFactor(Op0I), RHS)->isNullValue())
1920 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1924 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1925 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1) [urem only].
1926 if (I.getType()->isUnsigned() &&
1927 RHSI->getOpcode() == Instruction::Shl &&
1928 isa<ConstantUInt>(RHSI->getOperand(0))) {
1929 unsigned C1 = cast<ConstantUInt>(RHSI->getOperand(0))->getRawValue();
1930 if (isPowerOf2_64(C1)) {
1931 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
1932 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
1934 return BinaryOperator::createAnd(Op0, Add);
1938 // If this is 'urem X, (Cond ? C1, C2)' where C1&C2 are powers of two,
1939 // transform this into: '(Cond ? (urem X, C1) : (urem X, C2))'.
1940 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1941 if (ConstantUInt *STO = dyn_cast<ConstantUInt>(SI->getOperand(1)))
1942 if (ConstantUInt *SFO = dyn_cast<ConstantUInt>(SI->getOperand(2))) {
1943 if (STO->getValue() == 0) { // Couldn't be this argument.
1944 I.setOperand(1, SFO);
1946 } else if (SFO->getValue() == 0) {
1947 I.setOperand(1, STO);
1951 if (isPowerOf2_64(STO->getValue()) && isPowerOf2_64(SFO->getValue())){
1952 Value *TrueAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1953 SubOne(STO), SI->getName()+".t"), I);
1954 Value *FalseAnd = InsertNewInstBefore(BinaryOperator::createAnd(Op0,
1955 SubOne(SFO), SI->getName()+".f"), I);
1956 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
1964 // isMaxValueMinusOne - return true if this is Max-1
1965 static bool isMaxValueMinusOne(const ConstantInt *C) {
1966 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1967 return CU->getValue() == C->getType()->getIntegralTypeMask()-1;
1969 const ConstantSInt *CS = cast<ConstantSInt>(C);
1971 // Calculate 0111111111..11111
1972 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1973 int64_t Val = INT64_MAX; // All ones
1974 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
1975 return CS->getValue() == Val-1;
1978 // isMinValuePlusOne - return true if this is Min+1
1979 static bool isMinValuePlusOne(const ConstantInt *C) {
1980 if (const ConstantUInt *CU = dyn_cast<ConstantUInt>(C))
1981 return CU->getValue() == 1;
1983 const ConstantSInt *CS = cast<ConstantSInt>(C);
1985 // Calculate 1111111111000000000000
1986 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
1987 int64_t Val = -1; // All ones
1988 Val <<= TypeBits-1; // Shift over to the right spot
1989 return CS->getValue() == Val+1;
1992 // isOneBitSet - Return true if there is exactly one bit set in the specified
1994 static bool isOneBitSet(const ConstantInt *CI) {
1995 uint64_t V = CI->getRawValue();
1996 return V && (V & (V-1)) == 0;
1999 #if 0 // Currently unused
2000 // isLowOnes - Return true if the constant is of the form 0+1+.
2001 static bool isLowOnes(const ConstantInt *CI) {
2002 uint64_t V = CI->getRawValue();
2004 // There won't be bits set in parts that the type doesn't contain.
2005 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2007 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2008 return U && V && (U & V) == 0;
2012 // isHighOnes - Return true if the constant is of the form 1+0+.
2013 // This is the same as lowones(~X).
2014 static bool isHighOnes(const ConstantInt *CI) {
2015 uint64_t V = ~CI->getRawValue();
2016 if (~V == 0) return false; // 0's does not match "1+"
2018 // There won't be bits set in parts that the type doesn't contain.
2019 V &= ConstantInt::getAllOnesValue(CI->getType())->getRawValue();
2021 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2022 return U && V && (U & V) == 0;
2026 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2027 /// are carefully arranged to allow folding of expressions such as:
2029 /// (A < B) | (A > B) --> (A != B)
2031 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2032 /// represents that the comparison is true if A == B, and bit value '1' is true
2035 static unsigned getSetCondCode(const SetCondInst *SCI) {
2036 switch (SCI->getOpcode()) {
2038 case Instruction::SetGT: return 1;
2039 case Instruction::SetEQ: return 2;
2040 case Instruction::SetGE: return 3;
2041 case Instruction::SetLT: return 4;
2042 case Instruction::SetNE: return 5;
2043 case Instruction::SetLE: return 6;
2046 assert(0 && "Invalid SetCC opcode!");
2051 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2052 /// opcode and two operands into either a constant true or false, or a brand new
2053 /// SetCC instruction.
2054 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2056 case 0: return ConstantBool::False;
2057 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2058 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2059 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2060 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2061 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2062 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2063 case 7: return ConstantBool::True;
2064 default: assert(0 && "Illegal SetCCCode!"); return 0;
2068 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2069 struct FoldSetCCLogical {
2072 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2073 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2074 bool shouldApply(Value *V) const {
2075 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2076 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2077 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2080 Instruction *apply(BinaryOperator &Log) const {
2081 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2082 if (SCI->getOperand(0) != LHS) {
2083 assert(SCI->getOperand(1) == LHS);
2084 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2087 unsigned LHSCode = getSetCondCode(SCI);
2088 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2090 switch (Log.getOpcode()) {
2091 case Instruction::And: Code = LHSCode & RHSCode; break;
2092 case Instruction::Or: Code = LHSCode | RHSCode; break;
2093 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2094 default: assert(0 && "Illegal logical opcode!"); return 0;
2097 Value *RV = getSetCCValue(Code, LHS, RHS);
2098 if (Instruction *I = dyn_cast<Instruction>(RV))
2100 // Otherwise, it's a constant boolean value...
2101 return IC.ReplaceInstUsesWith(Log, RV);
2105 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2106 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2107 // guaranteed to be either a shift instruction or a binary operator.
2108 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2109 ConstantIntegral *OpRHS,
2110 ConstantIntegral *AndRHS,
2111 BinaryOperator &TheAnd) {
2112 Value *X = Op->getOperand(0);
2113 Constant *Together = 0;
2114 if (!isa<ShiftInst>(Op))
2115 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2117 switch (Op->getOpcode()) {
2118 case Instruction::Xor:
2119 if (Op->hasOneUse()) {
2120 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2121 std::string OpName = Op->getName(); Op->setName("");
2122 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2123 InsertNewInstBefore(And, TheAnd);
2124 return BinaryOperator::createXor(And, Together);
2127 case Instruction::Or:
2128 if (Together == AndRHS) // (X | C) & C --> C
2129 return ReplaceInstUsesWith(TheAnd, AndRHS);
2131 if (Op->hasOneUse() && Together != OpRHS) {
2132 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2133 std::string Op0Name = Op->getName(); Op->setName("");
2134 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2135 InsertNewInstBefore(Or, TheAnd);
2136 return BinaryOperator::createAnd(Or, AndRHS);
2139 case Instruction::Add:
2140 if (Op->hasOneUse()) {
2141 // Adding a one to a single bit bit-field should be turned into an XOR
2142 // of the bit. First thing to check is to see if this AND is with a
2143 // single bit constant.
2144 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getRawValue();
2146 // Clear bits that are not part of the constant.
2147 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2149 // If there is only one bit set...
2150 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2151 // Ok, at this point, we know that we are masking the result of the
2152 // ADD down to exactly one bit. If the constant we are adding has
2153 // no bits set below this bit, then we can eliminate the ADD.
2154 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getRawValue();
2156 // Check to see if any bits below the one bit set in AndRHSV are set.
2157 if ((AddRHS & (AndRHSV-1)) == 0) {
2158 // If not, the only thing that can effect the output of the AND is
2159 // the bit specified by AndRHSV. If that bit is set, the effect of
2160 // the XOR is to toggle the bit. If it is clear, then the ADD has
2162 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2163 TheAnd.setOperand(0, X);
2166 std::string Name = Op->getName(); Op->setName("");
2167 // Pull the XOR out of the AND.
2168 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2169 InsertNewInstBefore(NewAnd, TheAnd);
2170 return BinaryOperator::createXor(NewAnd, AndRHS);
2177 case Instruction::Shl: {
2178 // We know that the AND will not produce any of the bits shifted in, so if
2179 // the anded constant includes them, clear them now!
2181 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2182 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2183 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2185 if (CI == ShlMask) { // Masking out bits that the shift already masks
2186 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2187 } else if (CI != AndRHS) { // Reducing bits set in and.
2188 TheAnd.setOperand(1, CI);
2193 case Instruction::Shr:
2194 // We know that the AND will not produce any of the bits shifted in, so if
2195 // the anded constant includes them, clear them now! This only applies to
2196 // unsigned shifts, because a signed shr may bring in set bits!
2198 if (AndRHS->getType()->isUnsigned()) {
2199 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2200 Constant *ShrMask = ConstantExpr::getShr(AllOne, OpRHS);
2201 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2203 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2204 return ReplaceInstUsesWith(TheAnd, Op);
2205 } else if (CI != AndRHS) {
2206 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2209 } else { // Signed shr.
2210 // See if this is shifting in some sign extension, then masking it out
2212 if (Op->hasOneUse()) {
2213 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2214 Constant *ShrMask = ConstantExpr::getUShr(AllOne, OpRHS);
2215 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2216 if (CI == AndRHS) { // Masking out bits shifted in.
2217 // Make the argument unsigned.
2218 Value *ShVal = Op->getOperand(0);
2219 ShVal = InsertCastBefore(ShVal,
2220 ShVal->getType()->getUnsignedVersion(),
2222 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::Shr, ShVal,
2223 OpRHS, Op->getName()),
2225 Value *AndRHS2 = ConstantExpr::getCast(AndRHS, ShVal->getType());
2226 ShVal = InsertNewInstBefore(BinaryOperator::createAnd(ShVal, AndRHS2,
2229 return new CastInst(ShVal, Op->getType());
2239 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2240 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2241 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2242 /// insert new instructions.
2243 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2244 bool Inside, Instruction &IB) {
2245 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2246 "Lo is not <= Hi in range emission code!");
2248 if (Lo == Hi) // Trivially false.
2249 return new SetCondInst(Instruction::SetNE, V, V);
2250 if (cast<ConstantIntegral>(Lo)->isMinValue())
2251 return new SetCondInst(Instruction::SetLT, V, Hi);
2253 Constant *AddCST = ConstantExpr::getNeg(Lo);
2254 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2255 InsertNewInstBefore(Add, IB);
2256 // Convert to unsigned for the comparison.
2257 const Type *UnsType = Add->getType()->getUnsignedVersion();
2258 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2259 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2260 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2261 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2264 if (Lo == Hi) // Trivially true.
2265 return new SetCondInst(Instruction::SetEQ, V, V);
2267 Hi = SubOne(cast<ConstantInt>(Hi));
2268 if (cast<ConstantIntegral>(Lo)->isMinValue()) // V < 0 || V >= Hi ->'V > Hi-1'
2269 return new SetCondInst(Instruction::SetGT, V, Hi);
2271 // Emit X-Lo > Hi-Lo-1
2272 Constant *AddCST = ConstantExpr::getNeg(Lo);
2273 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2274 InsertNewInstBefore(Add, IB);
2275 // Convert to unsigned for the comparison.
2276 const Type *UnsType = Add->getType()->getUnsignedVersion();
2277 Value *OffsetVal = InsertCastBefore(Add, UnsType, IB);
2278 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2279 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2280 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2283 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2284 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2285 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2286 // not, since all 1s are not contiguous.
2287 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2288 uint64_t V = Val->getRawValue();
2289 if (!isShiftedMask_64(V)) return false;
2291 // look for the first zero bit after the run of ones
2292 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2293 // look for the first non-zero bit
2294 ME = 64-CountLeadingZeros_64(V);
2300 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2301 /// where isSub determines whether the operator is a sub. If we can fold one of
2302 /// the following xforms:
2304 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2305 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2306 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2308 /// return (A +/- B).
2310 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2311 ConstantIntegral *Mask, bool isSub,
2313 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2314 if (!LHSI || LHSI->getNumOperands() != 2 ||
2315 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2317 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2319 switch (LHSI->getOpcode()) {
2321 case Instruction::And:
2322 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2323 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2324 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0)
2327 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2328 // part, we don't need any explicit masks to take them out of A. If that
2329 // is all N is, ignore it.
2331 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2332 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2334 if (MaskedValueIsZero(RHS, Mask))
2339 case Instruction::Or:
2340 case Instruction::Xor:
2341 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2342 if ((Mask->getRawValue() & Mask->getRawValue()+1) == 0 &&
2343 ConstantExpr::getAnd(N, Mask)->isNullValue())
2350 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2352 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2353 return InsertNewInstBefore(New, I);
2356 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
2357 bool Changed = SimplifyCommutative(I);
2358 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2360 if (isa<UndefValue>(Op1)) // X & undef -> 0
2361 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2365 return ReplaceInstUsesWith(I, Op1);
2367 // See if we can simplify any instructions used by the instruction whose sole
2368 // purpose is to compute bits we don't care about.
2369 uint64_t KnownZero, KnownOne;
2370 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2371 KnownZero, KnownOne))
2374 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
2375 uint64_t AndRHSMask = AndRHS->getZExtValue();
2376 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
2377 uint64_t NotAndRHS = AndRHSMask^TypeMask;
2379 // Optimize a variety of ((val OP C1) & C2) combinations...
2380 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
2381 Instruction *Op0I = cast<Instruction>(Op0);
2382 Value *Op0LHS = Op0I->getOperand(0);
2383 Value *Op0RHS = Op0I->getOperand(1);
2384 switch (Op0I->getOpcode()) {
2385 case Instruction::Xor:
2386 case Instruction::Or:
2387 // If the mask is only needed on one incoming arm, push it up.
2388 if (Op0I->hasOneUse()) {
2389 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
2390 // Not masking anything out for the LHS, move to RHS.
2391 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
2392 Op0RHS->getName()+".masked");
2393 InsertNewInstBefore(NewRHS, I);
2394 return BinaryOperator::create(
2395 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
2397 if (!isa<Constant>(Op0RHS) &&
2398 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
2399 // Not masking anything out for the RHS, move to LHS.
2400 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
2401 Op0LHS->getName()+".masked");
2402 InsertNewInstBefore(NewLHS, I);
2403 return BinaryOperator::create(
2404 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
2409 case Instruction::Add:
2410 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2411 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2412 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2413 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2414 return BinaryOperator::createAnd(V, AndRHS);
2415 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2416 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
2419 case Instruction::Sub:
2420 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2421 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2422 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2423 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2424 return BinaryOperator::createAnd(V, AndRHS);
2428 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2429 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2431 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2432 const Type *SrcTy = CI->getOperand(0)->getType();
2434 // If this is an integer truncation or change from signed-to-unsigned, and
2435 // if the source is an and/or with immediate, transform it. This
2436 // frequently occurs for bitfield accesses.
2437 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2438 if (SrcTy->getPrimitiveSizeInBits() >=
2439 I.getType()->getPrimitiveSizeInBits() &&
2440 CastOp->getNumOperands() == 2)
2441 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
2442 if (CastOp->getOpcode() == Instruction::And) {
2443 // Change: and (cast (and X, C1) to T), C2
2444 // into : and (cast X to T), trunc(C1)&C2
2445 // This will folds the two ands together, which may allow other
2447 Instruction *NewCast =
2448 new CastInst(CastOp->getOperand(0), I.getType(),
2449 CastOp->getName()+".shrunk");
2450 NewCast = InsertNewInstBefore(NewCast, I);
2452 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2453 C3 = ConstantExpr::getAnd(C3, AndRHS); // trunc(C1)&C2
2454 return BinaryOperator::createAnd(NewCast, C3);
2455 } else if (CastOp->getOpcode() == Instruction::Or) {
2456 // Change: and (cast (or X, C1) to T), C2
2457 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2458 Constant *C3=ConstantExpr::getCast(AndCI, I.getType());//trunc(C1)
2459 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
2460 return ReplaceInstUsesWith(I, AndRHS);
2465 // Try to fold constant and into select arguments.
2466 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2467 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2469 if (isa<PHINode>(Op0))
2470 if (Instruction *NV = FoldOpIntoPhi(I))
2474 Value *Op0NotVal = dyn_castNotVal(Op0);
2475 Value *Op1NotVal = dyn_castNotVal(Op1);
2477 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
2478 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2480 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2481 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2482 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
2483 I.getName()+".demorgan");
2484 InsertNewInstBefore(Or, I);
2485 return BinaryOperator::createNot(Or);
2489 Value *A = 0, *B = 0;
2490 ConstantInt *C1 = 0, *C2 = 0;
2491 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
2492 if (A == Op1 || B == Op1) // (A | ?) & A --> A
2493 return ReplaceInstUsesWith(I, Op1);
2494 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
2495 if (A == Op0 || B == Op0) // A & (A | ?) --> A
2496 return ReplaceInstUsesWith(I, Op0);
2500 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
2501 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2502 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2505 Value *LHSVal, *RHSVal;
2506 ConstantInt *LHSCst, *RHSCst;
2507 Instruction::BinaryOps LHSCC, RHSCC;
2508 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2509 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2510 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
2511 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2512 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2513 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2514 // Ensure that the larger constant is on the RHS.
2515 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2516 SetCondInst *LHS = cast<SetCondInst>(Op0);
2517 if (cast<ConstantBool>(Cmp)->getValue()) {
2518 std::swap(LHS, RHS);
2519 std::swap(LHSCst, RHSCst);
2520 std::swap(LHSCC, RHSCC);
2523 // At this point, we know we have have two setcc instructions
2524 // comparing a value against two constants and and'ing the result
2525 // together. Because of the above check, we know that we only have
2526 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2527 // FoldSetCCLogical check above), that the two constants are not
2529 assert(LHSCst != RHSCst && "Compares not folded above?");
2532 default: assert(0 && "Unknown integer condition code!");
2533 case Instruction::SetEQ:
2535 default: assert(0 && "Unknown integer condition code!");
2536 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
2537 case Instruction::SetGT: // (X == 13 & X > 15) -> false
2538 return ReplaceInstUsesWith(I, ConstantBool::False);
2539 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
2540 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
2541 return ReplaceInstUsesWith(I, LHS);
2543 case Instruction::SetNE:
2545 default: assert(0 && "Unknown integer condition code!");
2546 case Instruction::SetLT:
2547 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
2548 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
2549 break; // (X != 13 & X < 15) -> no change
2550 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
2551 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
2552 return ReplaceInstUsesWith(I, RHS);
2553 case Instruction::SetNE:
2554 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
2555 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2556 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2557 LHSVal->getName()+".off");
2558 InsertNewInstBefore(Add, I);
2559 const Type *UnsType = Add->getType()->getUnsignedVersion();
2560 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2561 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
2562 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2563 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2565 break; // (X != 13 & X != 15) -> no change
2568 case Instruction::SetLT:
2570 default: assert(0 && "Unknown integer condition code!");
2571 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
2572 case Instruction::SetGT: // (X < 13 & X > 15) -> false
2573 return ReplaceInstUsesWith(I, ConstantBool::False);
2574 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
2575 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
2576 return ReplaceInstUsesWith(I, LHS);
2578 case Instruction::SetGT:
2580 default: assert(0 && "Unknown integer condition code!");
2581 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
2582 return ReplaceInstUsesWith(I, LHS);
2583 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
2584 return ReplaceInstUsesWith(I, RHS);
2585 case Instruction::SetNE:
2586 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
2587 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
2588 break; // (X > 13 & X != 15) -> no change
2589 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
2590 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
2596 return Changed ? &I : 0;
2599 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2600 bool Changed = SimplifyCommutative(I);
2601 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2603 if (isa<UndefValue>(Op1))
2604 return ReplaceInstUsesWith(I, // X | undef -> -1
2605 ConstantIntegral::getAllOnesValue(I.getType()));
2609 return ReplaceInstUsesWith(I, Op0);
2611 // See if we can simplify any instructions used by the instruction whose sole
2612 // purpose is to compute bits we don't care about.
2613 uint64_t KnownZero, KnownOne;
2614 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2615 KnownZero, KnownOne))
2619 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2620 ConstantInt *C1 = 0; Value *X = 0;
2621 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2622 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2623 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
2625 InsertNewInstBefore(Or, I);
2626 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
2629 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2630 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
2631 std::string Op0Name = Op0->getName(); Op0->setName("");
2632 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
2633 InsertNewInstBefore(Or, I);
2634 return BinaryOperator::createXor(Or,
2635 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
2638 // Try to fold constant and into select arguments.
2639 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2640 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2642 if (isa<PHINode>(Op0))
2643 if (Instruction *NV = FoldOpIntoPhi(I))
2647 Value *A = 0, *B = 0;
2648 ConstantInt *C1 = 0, *C2 = 0;
2650 if (match(Op0, m_And(m_Value(A), m_Value(B))))
2651 if (A == Op1 || B == Op1) // (A & ?) | A --> A
2652 return ReplaceInstUsesWith(I, Op1);
2653 if (match(Op1, m_And(m_Value(A), m_Value(B))))
2654 if (A == Op0 || B == Op0) // A | (A & ?) --> A
2655 return ReplaceInstUsesWith(I, Op0);
2657 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2658 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2659 MaskedValueIsZero(Op1, C1->getZExtValue())) {
2660 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
2662 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2665 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2666 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2667 MaskedValueIsZero(Op0, C1->getZExtValue())) {
2668 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
2670 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
2673 // (A & C1)|(B & C2)
2674 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
2675 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
2677 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
2678 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
2681 // If we have: ((V + N) & C1) | (V & C2)
2682 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2683 // replace with V+N.
2684 if (C1 == ConstantExpr::getNot(C2)) {
2685 Value *V1 = 0, *V2 = 0;
2686 if ((C2->getRawValue() & (C2->getRawValue()+1)) == 0 && // C2 == 0+1+
2687 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
2688 // Add commutes, try both ways.
2689 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
2690 return ReplaceInstUsesWith(I, A);
2691 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
2692 return ReplaceInstUsesWith(I, A);
2694 // Or commutes, try both ways.
2695 if ((C1->getRawValue() & (C1->getRawValue()+1)) == 0 &&
2696 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
2697 // Add commutes, try both ways.
2698 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
2699 return ReplaceInstUsesWith(I, B);
2700 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
2701 return ReplaceInstUsesWith(I, B);
2706 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
2707 if (A == Op1) // ~A | A == -1
2708 return ReplaceInstUsesWith(I,
2709 ConstantIntegral::getAllOnesValue(I.getType()));
2713 // Note, A is still live here!
2714 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
2716 return ReplaceInstUsesWith(I,
2717 ConstantIntegral::getAllOnesValue(I.getType()));
2719 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2720 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
2721 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
2722 I.getName()+".demorgan"), I);
2723 return BinaryOperator::createNot(And);
2727 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
2728 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
2729 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2732 Value *LHSVal, *RHSVal;
2733 ConstantInt *LHSCst, *RHSCst;
2734 Instruction::BinaryOps LHSCC, RHSCC;
2735 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
2736 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
2737 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
2738 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
2739 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
2740 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
2741 // Ensure that the larger constant is on the RHS.
2742 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
2743 SetCondInst *LHS = cast<SetCondInst>(Op0);
2744 if (cast<ConstantBool>(Cmp)->getValue()) {
2745 std::swap(LHS, RHS);
2746 std::swap(LHSCst, RHSCst);
2747 std::swap(LHSCC, RHSCC);
2750 // At this point, we know we have have two setcc instructions
2751 // comparing a value against two constants and or'ing the result
2752 // together. Because of the above check, we know that we only have
2753 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
2754 // FoldSetCCLogical check above), that the two constants are not
2756 assert(LHSCst != RHSCst && "Compares not folded above?");
2759 default: assert(0 && "Unknown integer condition code!");
2760 case Instruction::SetEQ:
2762 default: assert(0 && "Unknown integer condition code!");
2763 case Instruction::SetEQ:
2764 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
2765 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2766 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
2767 LHSVal->getName()+".off");
2768 InsertNewInstBefore(Add, I);
2769 const Type *UnsType = Add->getType()->getUnsignedVersion();
2770 Value *OffsetVal = InsertCastBefore(Add, UnsType, I);
2771 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
2772 AddCST = ConstantExpr::getCast(AddCST, UnsType);
2773 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2775 break; // (X == 13 | X == 15) -> no change
2777 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
2779 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
2780 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
2781 return ReplaceInstUsesWith(I, RHS);
2784 case Instruction::SetNE:
2786 default: assert(0 && "Unknown integer condition code!");
2787 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
2788 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
2789 return ReplaceInstUsesWith(I, LHS);
2790 case Instruction::SetNE: // (X != 13 | X != 15) -> true
2791 case Instruction::SetLT: // (X != 13 | X < 15) -> true
2792 return ReplaceInstUsesWith(I, ConstantBool::True);
2795 case Instruction::SetLT:
2797 default: assert(0 && "Unknown integer condition code!");
2798 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
2800 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
2801 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
2802 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
2803 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
2804 return ReplaceInstUsesWith(I, RHS);
2807 case Instruction::SetGT:
2809 default: assert(0 && "Unknown integer condition code!");
2810 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
2811 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
2812 return ReplaceInstUsesWith(I, LHS);
2813 case Instruction::SetNE: // (X > 13 | X != 15) -> true
2814 case Instruction::SetLT: // (X > 13 | X < 15) -> true
2815 return ReplaceInstUsesWith(I, ConstantBool::True);
2821 return Changed ? &I : 0;
2824 // XorSelf - Implements: X ^ X --> 0
2827 XorSelf(Value *rhs) : RHS(rhs) {}
2828 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2829 Instruction *apply(BinaryOperator &Xor) const {
2835 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2836 bool Changed = SimplifyCommutative(I);
2837 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2839 if (isa<UndefValue>(Op1))
2840 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
2842 // xor X, X = 0, even if X is nested in a sequence of Xor's.
2843 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
2844 assert(Result == &I && "AssociativeOpt didn't work?");
2845 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2848 // See if we can simplify any instructions used by the instruction whose sole
2849 // purpose is to compute bits we don't care about.
2850 uint64_t KnownZero, KnownOne;
2851 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
2852 KnownZero, KnownOne))
2855 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
2856 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2857 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
2858 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
2859 if (RHS == ConstantBool::True && SCI->hasOneUse())
2860 return new SetCondInst(SCI->getInverseCondition(),
2861 SCI->getOperand(0), SCI->getOperand(1));
2863 // ~(c-X) == X-c-1 == X+(-c-1)
2864 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2865 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2866 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2867 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2868 ConstantInt::get(I.getType(), 1));
2869 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
2872 // ~(~X & Y) --> (X | ~Y)
2873 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
2874 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
2875 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2877 BinaryOperator::createNot(Op0I->getOperand(1),
2878 Op0I->getOperand(1)->getName()+".not");
2879 InsertNewInstBefore(NotY, I);
2880 return BinaryOperator::createOr(Op0NotVal, NotY);
2884 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2885 if (Op0I->getOpcode() == Instruction::Add) {
2886 // ~(X-c) --> (-c-1)-X
2887 if (RHS->isAllOnesValue()) {
2888 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2889 return BinaryOperator::createSub(
2890 ConstantExpr::getSub(NegOp0CI,
2891 ConstantInt::get(I.getType(), 1)),
2892 Op0I->getOperand(0));
2894 } else if (Op0I->getOpcode() == Instruction::Or) {
2895 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2896 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
2897 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2898 // Anything in both C1 and C2 is known to be zero, remove it from
2900 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2901 NewRHS = ConstantExpr::getAnd(NewRHS,
2902 ConstantExpr::getNot(CommonBits));
2903 WorkList.push_back(Op0I);
2904 I.setOperand(0, Op0I->getOperand(0));
2905 I.setOperand(1, NewRHS);
2911 // Try to fold constant and into select arguments.
2912 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2913 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2915 if (isa<PHINode>(Op0))
2916 if (Instruction *NV = FoldOpIntoPhi(I))
2920 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
2922 return ReplaceInstUsesWith(I,
2923 ConstantIntegral::getAllOnesValue(I.getType()));
2925 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
2927 return ReplaceInstUsesWith(I,
2928 ConstantIntegral::getAllOnesValue(I.getType()));
2930 if (Instruction *Op1I = dyn_cast<Instruction>(Op1))
2931 if (Op1I->getOpcode() == Instruction::Or) {
2932 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
2933 cast<BinaryOperator>(Op1I)->swapOperands();
2935 std::swap(Op0, Op1);
2936 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
2938 std::swap(Op0, Op1);
2940 } else if (Op1I->getOpcode() == Instruction::Xor) {
2941 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
2942 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
2943 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
2944 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
2947 if (Instruction *Op0I = dyn_cast<Instruction>(Op0))
2948 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
2949 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
2950 cast<BinaryOperator>(Op0I)->swapOperands();
2951 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
2952 Value *NotB = InsertNewInstBefore(BinaryOperator::createNot(Op1,
2953 Op1->getName()+".not"), I);
2954 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
2956 } else if (Op0I->getOpcode() == Instruction::Xor) {
2957 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
2958 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2959 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
2960 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2963 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
2964 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
2965 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
2968 return Changed ? &I : 0;
2971 /// MulWithOverflow - Compute Result = In1*In2, returning true if the result
2972 /// overflowed for this type.
2973 static bool MulWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2975 Result = cast<ConstantInt>(ConstantExpr::getMul(In1, In2));
2976 return !In2->isNullValue() && ConstantExpr::getDiv(Result, In2) != In1;
2979 static bool isPositive(ConstantInt *C) {
2980 return cast<ConstantSInt>(C)->getValue() >= 0;
2983 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
2984 /// overflowed for this type.
2985 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
2987 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
2989 if (In1->getType()->isUnsigned())
2990 return cast<ConstantUInt>(Result)->getValue() <
2991 cast<ConstantUInt>(In1)->getValue();
2992 if (isPositive(In1) != isPositive(In2))
2994 if (isPositive(In1))
2995 return cast<ConstantSInt>(Result)->getValue() <
2996 cast<ConstantSInt>(In1)->getValue();
2997 return cast<ConstantSInt>(Result)->getValue() >
2998 cast<ConstantSInt>(In1)->getValue();
3001 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3002 /// code necessary to compute the offset from the base pointer (without adding
3003 /// in the base pointer). Return the result as a signed integer of intptr size.
3004 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3005 TargetData &TD = IC.getTargetData();
3006 gep_type_iterator GTI = gep_type_begin(GEP);
3007 const Type *UIntPtrTy = TD.getIntPtrType();
3008 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3009 Value *Result = Constant::getNullValue(SIntPtrTy);
3011 // Build a mask for high order bits.
3012 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3014 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3015 Value *Op = GEP->getOperand(i);
3016 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3017 Constant *Scale = ConstantExpr::getCast(ConstantUInt::get(UIntPtrTy, Size),
3019 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3020 if (!OpC->isNullValue()) {
3021 OpC = ConstantExpr::getCast(OpC, SIntPtrTy);
3022 Scale = ConstantExpr::getMul(OpC, Scale);
3023 if (Constant *RC = dyn_cast<Constant>(Result))
3024 Result = ConstantExpr::getAdd(RC, Scale);
3026 // Emit an add instruction.
3027 Result = IC.InsertNewInstBefore(
3028 BinaryOperator::createAdd(Result, Scale,
3029 GEP->getName()+".offs"), I);
3033 // Convert to correct type.
3034 Op = IC.InsertNewInstBefore(new CastInst(Op, SIntPtrTy,
3035 Op->getName()+".c"), I);
3037 // We'll let instcombine(mul) convert this to a shl if possible.
3038 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3039 GEP->getName()+".idx"), I);
3041 // Emit an add instruction.
3042 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3043 GEP->getName()+".offs"), I);
3049 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3050 /// else. At this point we know that the GEP is on the LHS of the comparison.
3051 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3052 Instruction::BinaryOps Cond,
3054 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3056 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3057 if (isa<PointerType>(CI->getOperand(0)->getType()))
3058 RHS = CI->getOperand(0);
3060 Value *PtrBase = GEPLHS->getOperand(0);
3061 if (PtrBase == RHS) {
3062 // As an optimization, we don't actually have to compute the actual value of
3063 // OFFSET if this is a seteq or setne comparison, just return whether each
3064 // index is zero or not.
3065 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3066 Instruction *InVal = 0;
3067 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3068 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3070 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3071 if (isa<UndefValue>(C)) // undef index -> undef.
3072 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3073 if (C->isNullValue())
3075 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3076 EmitIt = false; // This is indexing into a zero sized array?
3077 } else if (isa<ConstantInt>(C))
3078 return ReplaceInstUsesWith(I, // No comparison is needed here.
3079 ConstantBool::get(Cond == Instruction::SetNE));
3084 new SetCondInst(Cond, GEPLHS->getOperand(i),
3085 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3089 InVal = InsertNewInstBefore(InVal, I);
3090 InsertNewInstBefore(Comp, I);
3091 if (Cond == Instruction::SetNE) // True if any are unequal
3092 InVal = BinaryOperator::createOr(InVal, Comp);
3093 else // True if all are equal
3094 InVal = BinaryOperator::createAnd(InVal, Comp);
3102 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
3103 ConstantBool::get(Cond == Instruction::SetEQ));
3106 // Only lower this if the setcc is the only user of the GEP or if we expect
3107 // the result to fold to a constant!
3108 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
3109 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
3110 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
3111 return new SetCondInst(Cond, Offset,
3112 Constant::getNullValue(Offset->getType()));
3114 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
3115 // If the base pointers are different, but the indices are the same, just
3116 // compare the base pointer.
3117 if (PtrBase != GEPRHS->getOperand(0)) {
3118 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
3119 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
3120 GEPRHS->getOperand(0)->getType();
3122 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3123 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3124 IndicesTheSame = false;
3128 // If all indices are the same, just compare the base pointers.
3130 return new SetCondInst(Cond, GEPLHS->getOperand(0),
3131 GEPRHS->getOperand(0));
3133 // Otherwise, the base pointers are different and the indices are
3134 // different, bail out.
3138 // If one of the GEPs has all zero indices, recurse.
3139 bool AllZeros = true;
3140 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
3141 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
3142 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
3147 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
3148 SetCondInst::getSwappedCondition(Cond), I);
3150 // If the other GEP has all zero indices, recurse.
3152 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3153 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
3154 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
3159 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
3161 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
3162 // If the GEPs only differ by one index, compare it.
3163 unsigned NumDifferences = 0; // Keep track of # differences.
3164 unsigned DiffOperand = 0; // The operand that differs.
3165 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
3166 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
3167 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
3168 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
3169 // Irreconcilable differences.
3173 if (NumDifferences++) break;
3178 if (NumDifferences == 0) // SAME GEP?
3179 return ReplaceInstUsesWith(I, // No comparison is needed here.
3180 ConstantBool::get(Cond == Instruction::SetEQ));
3181 else if (NumDifferences == 1) {
3182 Value *LHSV = GEPLHS->getOperand(DiffOperand);
3183 Value *RHSV = GEPRHS->getOperand(DiffOperand);
3185 // Convert the operands to signed values to make sure to perform a
3186 // signed comparison.
3187 const Type *NewTy = LHSV->getType()->getSignedVersion();
3188 if (LHSV->getType() != NewTy)
3189 LHSV = InsertNewInstBefore(new CastInst(LHSV, NewTy,
3190 LHSV->getName()), I);
3191 if (RHSV->getType() != NewTy)
3192 RHSV = InsertNewInstBefore(new CastInst(RHSV, NewTy,
3193 RHSV->getName()), I);
3194 return new SetCondInst(Cond, LHSV, RHSV);
3198 // Only lower this if the setcc is the only user of the GEP or if we expect
3199 // the result to fold to a constant!
3200 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
3201 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
3202 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
3203 Value *L = EmitGEPOffset(GEPLHS, I, *this);
3204 Value *R = EmitGEPOffset(GEPRHS, I, *this);
3205 return new SetCondInst(Cond, L, R);
3212 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
3213 bool Changed = SimplifyCommutative(I);
3214 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3215 const Type *Ty = Op0->getType();
3219 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
3221 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
3222 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
3224 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
3225 // addresses never equal each other! We already know that Op0 != Op1.
3226 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
3227 isa<ConstantPointerNull>(Op0)) &&
3228 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
3229 isa<ConstantPointerNull>(Op1)))
3230 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
3232 // setcc's with boolean values can always be turned into bitwise operations
3233 if (Ty == Type::BoolTy) {
3234 switch (I.getOpcode()) {
3235 default: assert(0 && "Invalid setcc instruction!");
3236 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
3237 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
3238 InsertNewInstBefore(Xor, I);
3239 return BinaryOperator::createNot(Xor);
3241 case Instruction::SetNE:
3242 return BinaryOperator::createXor(Op0, Op1);
3244 case Instruction::SetGT:
3245 std::swap(Op0, Op1); // Change setgt -> setlt
3247 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
3248 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3249 InsertNewInstBefore(Not, I);
3250 return BinaryOperator::createAnd(Not, Op1);
3252 case Instruction::SetGE:
3253 std::swap(Op0, Op1); // Change setge -> setle
3255 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
3256 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
3257 InsertNewInstBefore(Not, I);
3258 return BinaryOperator::createOr(Not, Op1);
3263 // See if we are doing a comparison between a constant and an instruction that
3264 // can be folded into the comparison.
3265 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3266 // Check to see if we are comparing against the minimum or maximum value...
3267 if (CI->isMinValue()) {
3268 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
3269 return ReplaceInstUsesWith(I, ConstantBool::False);
3270 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
3271 return ReplaceInstUsesWith(I, ConstantBool::True);
3272 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
3273 return BinaryOperator::createSetEQ(Op0, Op1);
3274 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
3275 return BinaryOperator::createSetNE(Op0, Op1);
3277 } else if (CI->isMaxValue()) {
3278 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
3279 return ReplaceInstUsesWith(I, ConstantBool::False);
3280 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
3281 return ReplaceInstUsesWith(I, ConstantBool::True);
3282 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
3283 return BinaryOperator::createSetEQ(Op0, Op1);
3284 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
3285 return BinaryOperator::createSetNE(Op0, Op1);
3287 // Comparing against a value really close to min or max?
3288 } else if (isMinValuePlusOne(CI)) {
3289 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
3290 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
3291 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
3292 return BinaryOperator::createSetNE(Op0, SubOne(CI));
3294 } else if (isMaxValueMinusOne(CI)) {
3295 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
3296 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
3297 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
3298 return BinaryOperator::createSetNE(Op0, AddOne(CI));
3301 // If we still have a setle or setge instruction, turn it into the
3302 // appropriate setlt or setgt instruction. Since the border cases have
3303 // already been handled above, this requires little checking.
3305 if (I.getOpcode() == Instruction::SetLE)
3306 return BinaryOperator::createSetLT(Op0, AddOne(CI));
3307 if (I.getOpcode() == Instruction::SetGE)
3308 return BinaryOperator::createSetGT(Op0, SubOne(CI));
3311 // See if we can fold the comparison based on bits known to be zero or one
3313 uint64_t KnownZero, KnownOne;
3314 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
3315 KnownZero, KnownOne, 0))
3318 // Given the known and unknown bits, compute a range that the LHS could be
3320 if (KnownOne | KnownZero) {
3321 if (Ty->isUnsigned()) { // Unsigned comparison.
3323 uint64_t RHSVal = CI->getZExtValue();
3324 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3326 switch (I.getOpcode()) { // LE/GE have been folded already.
3327 default: assert(0 && "Unknown setcc opcode!");
3328 case Instruction::SetEQ:
3329 if (Max < RHSVal || Min > RHSVal)
3330 return ReplaceInstUsesWith(I, ConstantBool::False);
3332 case Instruction::SetNE:
3333 if (Max < RHSVal || Min > RHSVal)
3334 return ReplaceInstUsesWith(I, ConstantBool::True);
3336 case Instruction::SetLT:
3337 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3338 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3340 case Instruction::SetGT:
3341 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3342 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3345 } else { // Signed comparison.
3347 int64_t RHSVal = CI->getSExtValue();
3348 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
3350 switch (I.getOpcode()) { // LE/GE have been folded already.
3351 default: assert(0 && "Unknown setcc opcode!");
3352 case Instruction::SetEQ:
3353 if (Max < RHSVal || Min > RHSVal)
3354 return ReplaceInstUsesWith(I, ConstantBool::False);
3356 case Instruction::SetNE:
3357 if (Max < RHSVal || Min > RHSVal)
3358 return ReplaceInstUsesWith(I, ConstantBool::True);
3360 case Instruction::SetLT:
3361 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3362 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3364 case Instruction::SetGT:
3365 if (Min > RHSVal) return ReplaceInstUsesWith(I, ConstantBool::True);
3366 if (Max < RHSVal) return ReplaceInstUsesWith(I, ConstantBool::False);
3373 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3374 switch (LHSI->getOpcode()) {
3375 case Instruction::And:
3376 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
3377 LHSI->getOperand(0)->hasOneUse()) {
3378 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
3379 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
3380 // happens a LOT in code produced by the C front-end, for bitfield
3382 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
3383 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
3385 // Check to see if there is a noop-cast between the shift and the and.
3387 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
3388 if (CI->getOperand(0)->getType()->isIntegral() &&
3389 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
3390 CI->getType()->getPrimitiveSizeInBits())
3391 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
3394 ConstantUInt *ShAmt;
3395 ShAmt = Shift ? dyn_cast<ConstantUInt>(Shift->getOperand(1)) : 0;
3396 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
3397 const Type *AndTy = AndCST->getType(); // Type of the and.
3399 // We can fold this as long as we can't shift unknown bits
3400 // into the mask. This can only happen with signed shift
3401 // rights, as they sign-extend.
3403 bool CanFold = Shift->getOpcode() != Instruction::Shr ||
3406 // To test for the bad case of the signed shr, see if any
3407 // of the bits shifted in could be tested after the mask.
3408 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getValue();
3409 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
3411 Constant *OShAmt = ConstantUInt::get(Type::UByteTy, ShAmtVal);
3413 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
3415 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
3421 if (Shift->getOpcode() == Instruction::Shl)
3422 NewCst = ConstantExpr::getUShr(CI, ShAmt);
3424 NewCst = ConstantExpr::getShl(CI, ShAmt);
3426 // Check to see if we are shifting out any of the bits being
3428 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
3429 // If we shifted bits out, the fold is not going to work out.
3430 // As a special case, check to see if this means that the
3431 // result is always true or false now.
3432 if (I.getOpcode() == Instruction::SetEQ)
3433 return ReplaceInstUsesWith(I, ConstantBool::False);
3434 if (I.getOpcode() == Instruction::SetNE)
3435 return ReplaceInstUsesWith(I, ConstantBool::True);
3437 I.setOperand(1, NewCst);
3438 Constant *NewAndCST;
3439 if (Shift->getOpcode() == Instruction::Shl)
3440 NewAndCST = ConstantExpr::getUShr(AndCST, ShAmt);
3442 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
3443 LHSI->setOperand(1, NewAndCST);
3445 LHSI->setOperand(0, Shift->getOperand(0));
3447 Value *NewCast = InsertCastBefore(Shift->getOperand(0), AndTy,
3449 LHSI->setOperand(0, NewCast);
3451 WorkList.push_back(Shift); // Shift is dead.
3452 AddUsesToWorkList(I);
3460 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
3461 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3462 switch (I.getOpcode()) {
3464 case Instruction::SetEQ:
3465 case Instruction::SetNE: {
3466 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3468 // Check that the shift amount is in range. If not, don't perform
3469 // undefined shifts. When the shift is visited it will be
3471 if (ShAmt->getValue() >= TypeBits)
3474 // If we are comparing against bits always shifted out, the
3475 // comparison cannot succeed.
3477 ConstantExpr::getShl(ConstantExpr::getShr(CI, ShAmt), ShAmt);
3478 if (Comp != CI) {// Comparing against a bit that we know is zero.
3479 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3480 Constant *Cst = ConstantBool::get(IsSetNE);
3481 return ReplaceInstUsesWith(I, Cst);
3484 if (LHSI->hasOneUse()) {
3485 // Otherwise strength reduce the shift into an and.
3486 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3487 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
3490 if (CI->getType()->isUnsigned()) {
3491 Mask = ConstantUInt::get(CI->getType(), Val);
3492 } else if (ShAmtVal != 0) {
3493 Mask = ConstantSInt::get(CI->getType(), Val);
3495 Mask = ConstantInt::getAllOnesValue(CI->getType());
3499 BinaryOperator::createAnd(LHSI->getOperand(0),
3500 Mask, LHSI->getName()+".mask");
3501 Value *And = InsertNewInstBefore(AndI, I);
3502 return new SetCondInst(I.getOpcode(), And,
3503 ConstantExpr::getUShr(CI, ShAmt));
3510 case Instruction::Shr: // (setcc (shr X, ShAmt), CI)
3511 if (ConstantUInt *ShAmt = dyn_cast<ConstantUInt>(LHSI->getOperand(1))) {
3512 switch (I.getOpcode()) {
3514 case Instruction::SetEQ:
3515 case Instruction::SetNE: {
3517 // Check that the shift amount is in range. If not, don't perform
3518 // undefined shifts. When the shift is visited it will be
3520 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
3521 if (ShAmt->getValue() >= TypeBits)
3524 // If we are comparing against bits always shifted out, the
3525 // comparison cannot succeed.
3527 ConstantExpr::getShr(ConstantExpr::getShl(CI, ShAmt), ShAmt);
3529 if (Comp != CI) {// Comparing against a bit that we know is zero.
3530 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
3531 Constant *Cst = ConstantBool::get(IsSetNE);
3532 return ReplaceInstUsesWith(I, Cst);
3535 if (LHSI->hasOneUse() || CI->isNullValue()) {
3536 unsigned ShAmtVal = (unsigned)ShAmt->getValue();
3538 // Otherwise strength reduce the shift into an and.
3539 uint64_t Val = ~0ULL; // All ones.
3540 Val <<= ShAmtVal; // Shift over to the right spot.
3543 if (CI->getType()->isUnsigned()) {
3544 Val &= ~0ULL >> (64-TypeBits);
3545 Mask = ConstantUInt::get(CI->getType(), Val);
3547 Mask = ConstantSInt::get(CI->getType(), Val);
3551 BinaryOperator::createAnd(LHSI->getOperand(0),
3552 Mask, LHSI->getName()+".mask");
3553 Value *And = InsertNewInstBefore(AndI, I);
3554 return new SetCondInst(I.getOpcode(), And,
3555 ConstantExpr::getShl(CI, ShAmt));
3563 case Instruction::Div:
3564 // Fold: (div X, C1) op C2 -> range check
3565 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
3566 // Fold this div into the comparison, producing a range check.
3567 // Determine, based on the divide type, what the range is being
3568 // checked. If there is an overflow on the low or high side, remember
3569 // it, otherwise compute the range [low, hi) bounding the new value.
3570 bool LoOverflow = false, HiOverflow = 0;
3571 ConstantInt *LoBound = 0, *HiBound = 0;
3574 bool ProdOV = MulWithOverflow(Prod, CI, DivRHS);
3576 Instruction::BinaryOps Opcode = I.getOpcode();
3578 if (DivRHS->isNullValue()) { // Don't hack on divide by zeros.
3579 } else if (LHSI->getType()->isUnsigned()) { // udiv
3581 LoOverflow = ProdOV;
3582 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
3583 } else if (isPositive(DivRHS)) { // Divisor is > 0.
3584 if (CI->isNullValue()) { // (X / pos) op 0
3586 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
3588 } else if (isPositive(CI)) { // (X / pos) op pos
3590 LoOverflow = ProdOV;
3591 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
3592 } else { // (X / pos) op neg
3593 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
3594 LoOverflow = AddWithOverflow(LoBound, Prod,
3595 cast<ConstantInt>(DivRHSH));
3597 HiOverflow = ProdOV;
3599 } else { // Divisor is < 0.
3600 if (CI->isNullValue()) { // (X / neg) op 0
3601 LoBound = AddOne(DivRHS);
3602 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
3603 if (HiBound == DivRHS)
3604 LoBound = 0; // - INTMIN = INTMIN
3605 } else if (isPositive(CI)) { // (X / neg) op pos
3606 HiOverflow = LoOverflow = ProdOV;
3608 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
3609 HiBound = AddOne(Prod);
3610 } else { // (X / neg) op neg
3612 LoOverflow = HiOverflow = ProdOV;
3613 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
3616 // Dividing by a negate swaps the condition.
3617 Opcode = SetCondInst::getSwappedCondition(Opcode);
3621 Value *X = LHSI->getOperand(0);
3623 default: assert(0 && "Unhandled setcc opcode!");
3624 case Instruction::SetEQ:
3625 if (LoOverflow && HiOverflow)
3626 return ReplaceInstUsesWith(I, ConstantBool::False);
3627 else if (HiOverflow)
3628 return new SetCondInst(Instruction::SetGE, X, LoBound);
3629 else if (LoOverflow)
3630 return new SetCondInst(Instruction::SetLT, X, HiBound);
3632 return InsertRangeTest(X, LoBound, HiBound, true, I);
3633 case Instruction::SetNE:
3634 if (LoOverflow && HiOverflow)
3635 return ReplaceInstUsesWith(I, ConstantBool::True);
3636 else if (HiOverflow)
3637 return new SetCondInst(Instruction::SetLT, X, LoBound);
3638 else if (LoOverflow)
3639 return new SetCondInst(Instruction::SetGE, X, HiBound);
3641 return InsertRangeTest(X, LoBound, HiBound, false, I);
3642 case Instruction::SetLT:
3644 return ReplaceInstUsesWith(I, ConstantBool::False);
3645 return new SetCondInst(Instruction::SetLT, X, LoBound);
3646 case Instruction::SetGT:
3648 return ReplaceInstUsesWith(I, ConstantBool::False);
3649 return new SetCondInst(Instruction::SetGE, X, HiBound);
3656 // Simplify seteq and setne instructions...
3657 if (I.getOpcode() == Instruction::SetEQ ||
3658 I.getOpcode() == Instruction::SetNE) {
3659 bool isSetNE = I.getOpcode() == Instruction::SetNE;
3661 // If the first operand is (and|or|xor) with a constant, and the second
3662 // operand is a constant, simplify a bit.
3663 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
3664 switch (BO->getOpcode()) {
3665 case Instruction::Rem:
3666 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3667 if (CI->isNullValue() && isa<ConstantSInt>(BO->getOperand(1)) &&
3669 cast<ConstantSInt>(BO->getOperand(1))->getValue() > 1) {
3670 int64_t V = cast<ConstantSInt>(BO->getOperand(1))->getValue();
3671 if (isPowerOf2_64(V)) {
3672 unsigned L2 = Log2_64(V);
3673 const Type *UTy = BO->getType()->getUnsignedVersion();
3674 Value *NewX = InsertNewInstBefore(new CastInst(BO->getOperand(0),
3676 Constant *RHSCst = ConstantUInt::get(UTy, 1ULL << L2);
3677 Value *NewRem =InsertNewInstBefore(BinaryOperator::createRem(NewX,
3678 RHSCst, BO->getName()), I);
3679 return BinaryOperator::create(I.getOpcode(), NewRem,
3680 Constant::getNullValue(UTy));
3685 case Instruction::Add:
3686 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3687 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3688 if (BO->hasOneUse())
3689 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3690 ConstantExpr::getSub(CI, BOp1C));
3691 } else if (CI->isNullValue()) {
3692 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3693 // efficiently invertible, or if the add has just this one use.
3694 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3696 if (Value *NegVal = dyn_castNegVal(BOp1))
3697 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
3698 else if (Value *NegVal = dyn_castNegVal(BOp0))
3699 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
3700 else if (BO->hasOneUse()) {
3701 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
3703 InsertNewInstBefore(Neg, I);
3704 return new SetCondInst(I.getOpcode(), BOp0, Neg);
3708 case Instruction::Xor:
3709 // For the xor case, we can xor two constants together, eliminating
3710 // the explicit xor.
3711 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
3712 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
3713 ConstantExpr::getXor(CI, BOC));
3716 case Instruction::Sub:
3717 // Replace (([sub|xor] A, B) != 0) with (A != B)
3718 if (CI->isNullValue())
3719 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
3723 case Instruction::Or:
3724 // If bits are being or'd in that are not present in the constant we
3725 // are comparing against, then the comparison could never succeed!
3726 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
3727 Constant *NotCI = ConstantExpr::getNot(CI);
3728 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
3729 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3733 case Instruction::And:
3734 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3735 // If bits are being compared against that are and'd out, then the
3736 // comparison can never succeed!
3737 if (!ConstantExpr::getAnd(CI,
3738 ConstantExpr::getNot(BOC))->isNullValue())
3739 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
3741 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3742 if (CI == BOC && isOneBitSet(CI))
3743 return new SetCondInst(isSetNE ? Instruction::SetEQ :
3744 Instruction::SetNE, Op0,
3745 Constant::getNullValue(CI->getType()));
3747 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
3748 // to be a signed value as appropriate.
3749 if (isSignBit(BOC)) {
3750 Value *X = BO->getOperand(0);
3751 // If 'X' is not signed, insert a cast now...
3752 if (!BOC->getType()->isSigned()) {
3753 const Type *DestTy = BOC->getType()->getSignedVersion();
3754 X = InsertCastBefore(X, DestTy, I);
3756 return new SetCondInst(isSetNE ? Instruction::SetLT :
3757 Instruction::SetGE, X,
3758 Constant::getNullValue(X->getType()));
3761 // ((X & ~7) == 0) --> X < 8
3762 if (CI->isNullValue() && isHighOnes(BOC)) {
3763 Value *X = BO->getOperand(0);
3764 Constant *NegX = ConstantExpr::getNeg(BOC);
3766 // If 'X' is signed, insert a cast now.
3767 if (NegX->getType()->isSigned()) {
3768 const Type *DestTy = NegX->getType()->getUnsignedVersion();
3769 X = InsertCastBefore(X, DestTy, I);
3770 NegX = ConstantExpr::getCast(NegX, DestTy);
3773 return new SetCondInst(isSetNE ? Instruction::SetGE :
3774 Instruction::SetLT, X, NegX);
3781 } else { // Not a SetEQ/SetNE
3782 // If the LHS is a cast from an integral value of the same size,
3783 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
3784 Value *CastOp = Cast->getOperand(0);
3785 const Type *SrcTy = CastOp->getType();
3786 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
3787 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
3788 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
3789 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
3790 "Source and destination signednesses should differ!");
3791 if (Cast->getType()->isSigned()) {
3792 // If this is a signed comparison, check for comparisons in the
3793 // vicinity of zero.
3794 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
3796 return BinaryOperator::createSetGT(CastOp,
3797 ConstantUInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
3798 else if (I.getOpcode() == Instruction::SetGT &&
3799 cast<ConstantSInt>(CI)->getValue() == -1)
3800 // X > -1 => x < 128
3801 return BinaryOperator::createSetLT(CastOp,
3802 ConstantUInt::get(SrcTy, 1ULL << (SrcTySize-1)));
3804 ConstantUInt *CUI = cast<ConstantUInt>(CI);
3805 if (I.getOpcode() == Instruction::SetLT &&
3806 CUI->getValue() == 1ULL << (SrcTySize-1))
3807 // X < 128 => X > -1
3808 return BinaryOperator::createSetGT(CastOp,
3809 ConstantSInt::get(SrcTy, -1));
3810 else if (I.getOpcode() == Instruction::SetGT &&
3811 CUI->getValue() == (1ULL << (SrcTySize-1))-1)
3813 return BinaryOperator::createSetLT(CastOp,
3814 Constant::getNullValue(SrcTy));
3821 // Handle setcc with constant RHS's that can be integer, FP or pointer.
3822 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3823 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3824 switch (LHSI->getOpcode()) {
3825 case Instruction::GetElementPtr:
3826 if (RHSC->isNullValue()) {
3827 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
3828 bool isAllZeros = true;
3829 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
3830 if (!isa<Constant>(LHSI->getOperand(i)) ||
3831 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
3836 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
3837 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3841 case Instruction::PHI:
3842 if (Instruction *NV = FoldOpIntoPhi(I))
3845 case Instruction::Select:
3846 // If either operand of the select is a constant, we can fold the
3847 // comparison into the select arms, which will cause one to be
3848 // constant folded and the select turned into a bitwise or.
3849 Value *Op1 = 0, *Op2 = 0;
3850 if (LHSI->hasOneUse()) {
3851 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3852 // Fold the known value into the constant operand.
3853 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3854 // Insert a new SetCC of the other select operand.
3855 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3856 LHSI->getOperand(2), RHSC,
3858 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3859 // Fold the known value into the constant operand.
3860 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
3861 // Insert a new SetCC of the other select operand.
3862 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
3863 LHSI->getOperand(1), RHSC,
3869 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
3874 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
3875 if (User *GEP = dyn_castGetElementPtr(Op0))
3876 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
3878 if (User *GEP = dyn_castGetElementPtr(Op1))
3879 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
3880 SetCondInst::getSwappedCondition(I.getOpcode()), I))
3883 // Test to see if the operands of the setcc are casted versions of other
3884 // values. If the cast can be stripped off both arguments, we do so now.
3885 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3886 Value *CastOp0 = CI->getOperand(0);
3887 if (CastOp0->getType()->isLosslesslyConvertibleTo(CI->getType()) &&
3888 (isa<Constant>(Op1) || isa<CastInst>(Op1)) &&
3889 (I.getOpcode() == Instruction::SetEQ ||
3890 I.getOpcode() == Instruction::SetNE)) {
3891 // We keep moving the cast from the left operand over to the right
3892 // operand, where it can often be eliminated completely.
3895 // If operand #1 is a cast instruction, see if we can eliminate it as
3897 if (CastInst *CI2 = dyn_cast<CastInst>(Op1))
3898 if (CI2->getOperand(0)->getType()->isLosslesslyConvertibleTo(
3900 Op1 = CI2->getOperand(0);
3902 // If Op1 is a constant, we can fold the cast into the constant.
3903 if (Op1->getType() != Op0->getType())
3904 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3905 Op1 = ConstantExpr::getCast(Op1C, Op0->getType());
3907 // Otherwise, cast the RHS right before the setcc
3908 Op1 = new CastInst(Op1, Op0->getType(), Op1->getName());
3909 InsertNewInstBefore(cast<Instruction>(Op1), I);
3911 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
3914 // Handle the special case of: setcc (cast bool to X), <cst>
3915 // This comes up when you have code like
3918 // For generality, we handle any zero-extension of any operand comparison
3919 // with a constant or another cast from the same type.
3920 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
3921 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
3925 if (I.getOpcode() == Instruction::SetNE ||
3926 I.getOpcode() == Instruction::SetEQ) {
3928 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
3929 (A == Op1 || B == Op1)) {
3930 // (A^B) == A -> B == 0
3931 Value *OtherVal = A == Op1 ? B : A;
3932 return BinaryOperator::create(I.getOpcode(), OtherVal,
3933 Constant::getNullValue(A->getType()));
3934 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3935 (A == Op0 || B == Op0)) {
3936 // A == (A^B) -> B == 0
3937 Value *OtherVal = A == Op0 ? B : A;
3938 return BinaryOperator::create(I.getOpcode(), OtherVal,
3939 Constant::getNullValue(A->getType()));
3940 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
3941 // (A-B) == A -> B == 0
3942 return BinaryOperator::create(I.getOpcode(), B,
3943 Constant::getNullValue(B->getType()));
3944 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
3945 // A == (A-B) -> B == 0
3946 return BinaryOperator::create(I.getOpcode(), B,
3947 Constant::getNullValue(B->getType()));
3950 return Changed ? &I : 0;
3953 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
3954 // We only handle extending casts so far.
3956 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
3957 Value *LHSCIOp = cast<CastInst>(SCI.getOperand(0))->getOperand(0);
3958 const Type *SrcTy = LHSCIOp->getType();
3959 const Type *DestTy = SCI.getOperand(0)->getType();
3962 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
3965 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
3966 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
3967 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
3969 // Is this a sign or zero extension?
3970 bool isSignSrc = SrcTy->isSigned();
3971 bool isSignDest = DestTy->isSigned();
3973 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
3974 // Not an extension from the same type?
3975 RHSCIOp = CI->getOperand(0);
3976 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
3977 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
3978 // Compute the constant that would happen if we truncated to SrcTy then
3979 // reextended to DestTy.
3980 Constant *Res = ConstantExpr::getCast(CI, SrcTy);
3982 if (ConstantExpr::getCast(Res, DestTy) == CI) {
3985 // If the value cannot be represented in the shorter type, we cannot emit
3986 // a simple comparison.
3987 if (SCI.getOpcode() == Instruction::SetEQ)
3988 return ReplaceInstUsesWith(SCI, ConstantBool::False);
3989 if (SCI.getOpcode() == Instruction::SetNE)
3990 return ReplaceInstUsesWith(SCI, ConstantBool::True);
3992 // Evaluate the comparison for LT.
3994 if (DestTy->isSigned()) {
3995 // We're performing a signed comparison.
3997 // Signed extend and signed comparison.
3998 if (cast<ConstantSInt>(CI)->getValue() < 0) // X < (small) --> false
3999 Result = ConstantBool::False;
4001 Result = ConstantBool::True; // X < (large) --> true
4003 // Unsigned extend and signed comparison.
4004 if (cast<ConstantSInt>(CI)->getValue() < 0)
4005 Result = ConstantBool::False;
4007 Result = ConstantBool::True;
4010 // We're performing an unsigned comparison.
4012 // Unsigned extend & compare -> always true.
4013 Result = ConstantBool::True;
4015 // We're performing an unsigned comp with a sign extended value.
4016 // This is true if the input is >= 0. [aka >s -1]
4017 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
4018 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
4019 NegOne, SCI.getName()), SCI);
4023 // Finally, return the value computed.
4024 if (SCI.getOpcode() == Instruction::SetLT) {
4025 return ReplaceInstUsesWith(SCI, Result);
4027 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
4028 if (Constant *CI = dyn_cast<Constant>(Result))
4029 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
4031 return BinaryOperator::createNot(Result);
4038 // Okay, just insert a compare of the reduced operands now!
4039 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
4042 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
4043 assert(I.getOperand(1)->getType() == Type::UByteTy);
4044 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4045 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4047 // shl X, 0 == X and shr X, 0 == X
4048 // shl 0, X == 0 and shr 0, X == 0
4049 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
4050 Op0 == Constant::getNullValue(Op0->getType()))
4051 return ReplaceInstUsesWith(I, Op0);
4053 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
4054 if (!isLeftShift && I.getType()->isSigned())
4055 return ReplaceInstUsesWith(I, Op0);
4056 else // undef << X -> 0 AND undef >>u X -> 0
4057 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4059 if (isa<UndefValue>(Op1)) {
4060 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
4061 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4063 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
4066 // shr int -1, X = -1 (for any arithmetic shift rights of ~0)
4068 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op0))
4069 if (CSI->isAllOnesValue())
4070 return ReplaceInstUsesWith(I, CSI);
4072 // Try to fold constant and into select arguments.
4073 if (isa<Constant>(Op0))
4074 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
4075 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4078 // See if we can turn a signed shr into an unsigned shr.
4079 if (!isLeftShift && I.getType()->isSigned()) {
4080 if (MaskedValueIsZero(Op0,
4081 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
4082 Value *V = InsertCastBefore(Op0, I.getType()->getUnsignedVersion(), I);
4083 V = InsertNewInstBefore(new ShiftInst(Instruction::Shr, V, Op1,
4085 return new CastInst(V, I.getType());
4089 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op1))
4090 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
4095 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantUInt *Op1,
4097 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4098 bool isSignedShift = Op0->getType()->isSigned();
4099 bool isUnsignedShift = !isSignedShift;
4101 // See if we can simplify any instructions used by the instruction whose sole
4102 // purpose is to compute bits we don't care about.
4103 uint64_t KnownZero, KnownOne;
4104 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
4105 KnownZero, KnownOne))
4108 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
4109 // of a signed value.
4111 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
4112 if (Op1->getValue() >= TypeBits) {
4113 if (isUnsignedShift || isLeftShift)
4114 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
4116 I.setOperand(1, ConstantUInt::get(Type::UByteTy, TypeBits-1));
4121 // ((X*C1) << C2) == (X * (C1 << C2))
4122 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
4123 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
4124 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
4125 return BinaryOperator::createMul(BO->getOperand(0),
4126 ConstantExpr::getShl(BOOp, Op1));
4128 // Try to fold constant and into select arguments.
4129 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4130 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4132 if (isa<PHINode>(Op0))
4133 if (Instruction *NV = FoldOpIntoPhi(I))
4136 if (Op0->hasOneUse()) {
4137 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
4138 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4141 switch (Op0BO->getOpcode()) {
4143 case Instruction::Add:
4144 case Instruction::And:
4145 case Instruction::Or:
4146 case Instruction::Xor:
4147 // These operators commute.
4148 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
4149 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4150 match(Op0BO->getOperand(1),
4151 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4152 Instruction *YS = new ShiftInst(Instruction::Shl,
4153 Op0BO->getOperand(0), Op1,
4155 InsertNewInstBefore(YS, I); // (Y << C)
4157 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
4158 Op0BO->getOperand(1)->getName());
4159 InsertNewInstBefore(X, I); // (X + (Y << C))
4160 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4161 C2 = ConstantExpr::getShl(C2, Op1);
4162 return BinaryOperator::createAnd(X, C2);
4165 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
4166 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4167 match(Op0BO->getOperand(1),
4168 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4169 m_ConstantInt(CC))) && V2 == Op1 &&
4170 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
4171 Instruction *YS = new ShiftInst(Instruction::Shl,
4172 Op0BO->getOperand(0), Op1,
4174 InsertNewInstBefore(YS, I); // (Y << C)
4176 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4177 V1->getName()+".mask");
4178 InsertNewInstBefore(XM, I); // X & (CC << C)
4180 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
4184 case Instruction::Sub:
4185 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4186 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4187 match(Op0BO->getOperand(0),
4188 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
4189 Instruction *YS = new ShiftInst(Instruction::Shl,
4190 Op0BO->getOperand(1), Op1,
4192 InsertNewInstBefore(YS, I); // (Y << C)
4194 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
4195 Op0BO->getOperand(0)->getName());
4196 InsertNewInstBefore(X, I); // (X + (Y << C))
4197 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
4198 C2 = ConstantExpr::getShl(C2, Op1);
4199 return BinaryOperator::createAnd(X, C2);
4202 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4203 match(Op0BO->getOperand(0),
4204 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4205 m_ConstantInt(CC))) && V2 == Op1 &&
4206 cast<BinaryOperator>(Op0BO->getOperand(0))
4207 ->getOperand(0)->hasOneUse()) {
4208 Instruction *YS = new ShiftInst(Instruction::Shl,
4209 Op0BO->getOperand(1), Op1,
4211 InsertNewInstBefore(YS, I); // (Y << C)
4213 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
4214 V1->getName()+".mask");
4215 InsertNewInstBefore(XM, I); // X & (CC << C)
4217 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
4224 // If the operand is an bitwise operator with a constant RHS, and the
4225 // shift is the only use, we can pull it out of the shift.
4226 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
4227 bool isValid = true; // Valid only for And, Or, Xor
4228 bool highBitSet = false; // Transform if high bit of constant set?
4230 switch (Op0BO->getOpcode()) {
4231 default: isValid = false; break; // Do not perform transform!
4232 case Instruction::Add:
4233 isValid = isLeftShift;
4235 case Instruction::Or:
4236 case Instruction::Xor:
4239 case Instruction::And:
4244 // If this is a signed shift right, and the high bit is modified
4245 // by the logical operation, do not perform the transformation.
4246 // The highBitSet boolean indicates the value of the high bit of
4247 // the constant which would cause it to be modified for this
4250 if (isValid && !isLeftShift && isSignedShift) {
4251 uint64_t Val = Op0C->getRawValue();
4252 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
4256 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
4258 Instruction *NewShift =
4259 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
4262 InsertNewInstBefore(NewShift, I);
4264 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
4271 // Find out if this is a shift of a shift by a constant.
4272 ShiftInst *ShiftOp = 0;
4273 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
4275 else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4276 // If this is a noop-integer case of a shift instruction, use the shift.
4277 if (CI->getOperand(0)->getType()->isInteger() &&
4278 CI->getOperand(0)->getType()->getPrimitiveSizeInBits() ==
4279 CI->getType()->getPrimitiveSizeInBits() &&
4280 isa<ShiftInst>(CI->getOperand(0))) {
4281 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
4285 if (ShiftOp && isa<ConstantUInt>(ShiftOp->getOperand(1))) {
4286 // Find the operands and properties of the input shift. Note that the
4287 // signedness of the input shift may differ from the current shift if there
4288 // is a noop cast between the two.
4289 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
4290 bool isShiftOfSignedShift = ShiftOp->getType()->isSigned();
4291 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
4293 ConstantUInt *ShiftAmt1C = cast<ConstantUInt>(ShiftOp->getOperand(1));
4295 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getValue();
4296 unsigned ShiftAmt2 = (unsigned)Op1->getValue();
4298 // Check for (A << c1) << c2 and (A >> c1) >> c2.
4299 if (isLeftShift == isShiftOfLeftShift) {
4300 // Do not fold these shifts if the first one is signed and the second one
4301 // is unsigned and this is a right shift. Further, don't do any folding
4303 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
4306 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
4307 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
4308 Amt = Op0->getType()->getPrimitiveSizeInBits();
4310 Value *Op = ShiftOp->getOperand(0);
4311 if (isShiftOfSignedShift != isSignedShift)
4312 Op = InsertNewInstBefore(new CastInst(Op, I.getType(), "tmp"), I);
4313 return new ShiftInst(I.getOpcode(), Op,
4314 ConstantUInt::get(Type::UByteTy, Amt));
4317 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
4318 // signed types, we can only support the (A >> c1) << c2 configuration,
4319 // because it can not turn an arbitrary bit of A into a sign bit.
4320 if (isUnsignedShift || isLeftShift) {
4321 // Calculate bitmask for what gets shifted off the edge.
4322 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
4324 C = ConstantExpr::getShl(C, ShiftAmt1C);
4326 C = ConstantExpr::getUShr(C, ShiftAmt1C);
4328 Value *Op = ShiftOp->getOperand(0);
4329 if (isShiftOfSignedShift != isSignedShift)
4330 Op = InsertNewInstBefore(new CastInst(Op, I.getType(),Op->getName()),I);
4333 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
4334 InsertNewInstBefore(Mask, I);
4336 // Figure out what flavor of shift we should use...
4337 if (ShiftAmt1 == ShiftAmt2) {
4338 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
4339 } else if (ShiftAmt1 < ShiftAmt2) {
4340 return new ShiftInst(I.getOpcode(), Mask,
4341 ConstantUInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
4342 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
4343 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
4344 // Make sure to emit an unsigned shift right, not a signed one.
4345 Mask = InsertNewInstBefore(new CastInst(Mask,
4346 Mask->getType()->getUnsignedVersion(),
4348 Mask = new ShiftInst(Instruction::Shr, Mask,
4349 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4350 InsertNewInstBefore(Mask, I);
4351 return new CastInst(Mask, I.getType());
4353 return new ShiftInst(ShiftOp->getOpcode(), Mask,
4354 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4357 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
4358 Op = InsertNewInstBefore(new CastInst(Mask,
4359 I.getType()->getSignedVersion(),
4360 Mask->getName()), I);
4361 Instruction *Shift =
4362 new ShiftInst(ShiftOp->getOpcode(), Op,
4363 ConstantUInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
4364 InsertNewInstBefore(Shift, I);
4366 C = ConstantIntegral::getAllOnesValue(Shift->getType());
4367 C = ConstantExpr::getShl(C, Op1);
4368 Mask = BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
4369 InsertNewInstBefore(Mask, I);
4370 return new CastInst(Mask, I.getType());
4373 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
4374 // this case, C1 == C2 and C1 is 8, 16, or 32.
4375 if (ShiftAmt1 == ShiftAmt2) {
4376 const Type *SExtType = 0;
4377 switch (ShiftAmt1) {
4378 case 8 : SExtType = Type::SByteTy; break;
4379 case 16: SExtType = Type::ShortTy; break;
4380 case 32: SExtType = Type::IntTy; break;
4384 Instruction *NewTrunc = new CastInst(ShiftOp->getOperand(0),
4386 InsertNewInstBefore(NewTrunc, I);
4387 return new CastInst(NewTrunc, I.getType());
4402 /// getCastType - In the future, we will split the cast instruction into these
4403 /// various types. Until then, we have to do the analysis here.
4404 static CastType getCastType(const Type *Src, const Type *Dest) {
4405 assert(Src->isIntegral() && Dest->isIntegral() &&
4406 "Only works on integral types!");
4407 unsigned SrcSize = Src->getPrimitiveSizeInBits();
4408 unsigned DestSize = Dest->getPrimitiveSizeInBits();
4410 if (SrcSize == DestSize) return Noop;
4411 if (SrcSize > DestSize) return Truncate;
4412 if (Src->isSigned()) return Signext;
4417 // isEliminableCastOfCast - Return true if it is valid to eliminate the CI
4420 static bool isEliminableCastOfCast(const Type *SrcTy, const Type *MidTy,
4421 const Type *DstTy, TargetData *TD) {
4423 // It is legal to eliminate the instruction if casting A->B->A if the sizes
4424 // are identical and the bits don't get reinterpreted (for example
4425 // int->float->int would not be allowed).
4426 if (SrcTy == DstTy && SrcTy->isLosslesslyConvertibleTo(MidTy))
4429 // If we are casting between pointer and integer types, treat pointers as
4430 // integers of the appropriate size for the code below.
4431 if (isa<PointerType>(SrcTy)) SrcTy = TD->getIntPtrType();
4432 if (isa<PointerType>(MidTy)) MidTy = TD->getIntPtrType();
4433 if (isa<PointerType>(DstTy)) DstTy = TD->getIntPtrType();
4435 // Allow free casting and conversion of sizes as long as the sign doesn't
4437 if (SrcTy->isIntegral() && MidTy->isIntegral() && DstTy->isIntegral()) {
4438 CastType FirstCast = getCastType(SrcTy, MidTy);
4439 CastType SecondCast = getCastType(MidTy, DstTy);
4441 // Capture the effect of these two casts. If the result is a legal cast,
4442 // the CastType is stored here, otherwise a special code is used.
4443 static const unsigned CastResult[] = {
4444 // First cast is noop
4446 // First cast is a truncate
4447 1, 1, 4, 4, // trunc->extend is not safe to eliminate
4448 // First cast is a sign ext
4449 2, 5, 2, 4, // signext->zeroext never ok
4450 // First cast is a zero ext
4454 unsigned Result = CastResult[FirstCast*4+SecondCast];
4456 default: assert(0 && "Illegal table value!");
4461 // FIXME: in the future, when LLVM has explicit sign/zeroextends and
4462 // truncates, we could eliminate more casts.
4463 return (unsigned)getCastType(SrcTy, DstTy) == Result;
4465 return false; // Not possible to eliminate this here.
4467 // Sign or zero extend followed by truncate is always ok if the result
4468 // is a truncate or noop.
4469 CastType ResultCast = getCastType(SrcTy, DstTy);
4470 if (ResultCast == Noop || ResultCast == Truncate)
4472 // Otherwise we are still growing the value, we are only safe if the
4473 // result will match the sign/zeroextendness of the result.
4474 return ResultCast == FirstCast;
4478 // If this is a cast from 'float -> double -> integer', cast from
4479 // 'float -> integer' directly, as the value isn't changed by the
4480 // float->double conversion.
4481 if (SrcTy->isFloatingPoint() && MidTy->isFloatingPoint() &&
4482 DstTy->isIntegral() &&
4483 SrcTy->getPrimitiveSize() < MidTy->getPrimitiveSize())
4489 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
4490 if (V->getType() == Ty || isa<Constant>(V)) return false;
4491 if (const CastInst *CI = dyn_cast<CastInst>(V))
4492 if (isEliminableCastOfCast(CI->getOperand(0)->getType(), CI->getType(), Ty,
4498 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
4499 /// InsertBefore instruction. This is specialized a bit to avoid inserting
4500 /// casts that are known to not do anything...
4502 Value *InstCombiner::InsertOperandCastBefore(Value *V, const Type *DestTy,
4503 Instruction *InsertBefore) {
4504 if (V->getType() == DestTy) return V;
4505 if (Constant *C = dyn_cast<Constant>(V))
4506 return ConstantExpr::getCast(C, DestTy);
4508 CastInst *CI = new CastInst(V, DestTy, V->getName());
4509 InsertNewInstBefore(CI, *InsertBefore);
4513 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
4514 /// expression. If so, decompose it, returning some value X, such that Val is
4517 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
4519 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
4520 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(Val)) {
4521 Offset = CI->getValue();
4523 return ConstantUInt::get(Type::UIntTy, 0);
4524 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
4525 if (I->getNumOperands() == 2) {
4526 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I->getOperand(1))) {
4527 if (I->getOpcode() == Instruction::Shl) {
4528 // This is a value scaled by '1 << the shift amt'.
4529 Scale = 1U << CUI->getValue();
4531 return I->getOperand(0);
4532 } else if (I->getOpcode() == Instruction::Mul) {
4533 // This value is scaled by 'CUI'.
4534 Scale = CUI->getValue();
4536 return I->getOperand(0);
4537 } else if (I->getOpcode() == Instruction::Add) {
4538 // We have X+C. Check to see if we really have (X*C2)+C1, where C1 is
4541 Value *SubVal = DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
4543 Offset += CUI->getValue();
4544 if (SubScale > 1 && (Offset % SubScale == 0)) {
4553 // Otherwise, we can't look past this.
4560 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
4561 /// try to eliminate the cast by moving the type information into the alloc.
4562 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
4563 AllocationInst &AI) {
4564 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
4565 if (!PTy) return 0; // Not casting the allocation to a pointer type.
4567 // Remove any uses of AI that are dead.
4568 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
4569 std::vector<Instruction*> DeadUsers;
4570 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
4571 Instruction *User = cast<Instruction>(*UI++);
4572 if (isInstructionTriviallyDead(User)) {
4573 while (UI != E && *UI == User)
4574 ++UI; // If this instruction uses AI more than once, don't break UI.
4576 // Add operands to the worklist.
4577 AddUsesToWorkList(*User);
4579 DEBUG(std::cerr << "IC: DCE: " << *User);
4581 User->eraseFromParent();
4582 removeFromWorkList(User);
4586 // Get the type really allocated and the type casted to.
4587 const Type *AllocElTy = AI.getAllocatedType();
4588 const Type *CastElTy = PTy->getElementType();
4589 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
4591 unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
4592 unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
4593 if (CastElTyAlign < AllocElTyAlign) return 0;
4595 // If the allocation has multiple uses, only promote it if we are strictly
4596 // increasing the alignment of the resultant allocation. If we keep it the
4597 // same, we open the door to infinite loops of various kinds.
4598 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
4600 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
4601 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
4602 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
4604 // See if we can satisfy the modulus by pulling a scale out of the array
4606 unsigned ArraySizeScale, ArrayOffset;
4607 Value *NumElements = // See if the array size is a decomposable linear expr.
4608 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
4610 // If we can now satisfy the modulus, by using a non-1 scale, we really can
4612 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
4613 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
4615 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
4620 Amt = ConstantUInt::get(Type::UIntTy, Scale);
4621 if (ConstantUInt *CI = dyn_cast<ConstantUInt>(NumElements))
4622 Amt = ConstantExpr::getMul(CI, cast<ConstantUInt>(Amt));
4623 else if (Scale != 1) {
4624 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
4625 Amt = InsertNewInstBefore(Tmp, AI);
4629 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
4630 Value *Off = ConstantUInt::get(Type::UIntTy, Offset);
4631 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
4632 Amt = InsertNewInstBefore(Tmp, AI);
4635 std::string Name = AI.getName(); AI.setName("");
4636 AllocationInst *New;
4637 if (isa<MallocInst>(AI))
4638 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
4640 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
4641 InsertNewInstBefore(New, AI);
4643 // If the allocation has multiple uses, insert a cast and change all things
4644 // that used it to use the new cast. This will also hack on CI, but it will
4646 if (!AI.hasOneUse()) {
4647 AddUsesToWorkList(AI);
4648 CastInst *NewCast = new CastInst(New, AI.getType(), "tmpcast");
4649 InsertNewInstBefore(NewCast, AI);
4650 AI.replaceAllUsesWith(NewCast);
4652 return ReplaceInstUsesWith(CI, New);
4656 // CastInst simplification
4658 Instruction *InstCombiner::visitCastInst(CastInst &CI) {
4659 Value *Src = CI.getOperand(0);
4661 // If the user is casting a value to the same type, eliminate this cast
4663 if (CI.getType() == Src->getType())
4664 return ReplaceInstUsesWith(CI, Src);
4666 if (isa<UndefValue>(Src)) // cast undef -> undef
4667 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
4669 // If casting the result of another cast instruction, try to eliminate this
4672 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
4673 Value *A = CSrc->getOperand(0);
4674 if (isEliminableCastOfCast(A->getType(), CSrc->getType(),
4675 CI.getType(), TD)) {
4676 // This instruction now refers directly to the cast's src operand. This
4677 // has a good chance of making CSrc dead.
4678 CI.setOperand(0, CSrc->getOperand(0));
4682 // If this is an A->B->A cast, and we are dealing with integral types, try
4683 // to convert this into a logical 'and' instruction.
4685 if (A->getType()->isInteger() &&
4686 CI.getType()->isInteger() && CSrc->getType()->isInteger() &&
4687 CSrc->getType()->isUnsigned() && // B->A cast must zero extend
4688 CSrc->getType()->getPrimitiveSizeInBits() <
4689 CI.getType()->getPrimitiveSizeInBits()&&
4690 A->getType()->getPrimitiveSizeInBits() ==
4691 CI.getType()->getPrimitiveSizeInBits()) {
4692 assert(CSrc->getType() != Type::ULongTy &&
4693 "Cannot have type bigger than ulong!");
4694 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
4695 Constant *AndOp = ConstantUInt::get(A->getType()->getUnsignedVersion(),
4697 AndOp = ConstantExpr::getCast(AndOp, A->getType());
4698 Instruction *And = BinaryOperator::createAnd(CSrc->getOperand(0), AndOp);
4699 if (And->getType() != CI.getType()) {
4700 And->setName(CSrc->getName()+".mask");
4701 InsertNewInstBefore(And, CI);
4702 And = new CastInst(And, CI.getType());
4708 // If this is a cast to bool, turn it into the appropriate setne instruction.
4709 if (CI.getType() == Type::BoolTy)
4710 return BinaryOperator::createSetNE(CI.getOperand(0),
4711 Constant::getNullValue(CI.getOperand(0)->getType()));
4713 // See if we can simplify any instructions used by the LHS whose sole
4714 // purpose is to compute bits we don't care about.
4715 if (CI.getType()->isInteger() && CI.getOperand(0)->getType()->isIntegral()) {
4716 uint64_t KnownZero, KnownOne;
4717 if (SimplifyDemandedBits(&CI, CI.getType()->getIntegralTypeMask(),
4718 KnownZero, KnownOne))
4722 // If casting the result of a getelementptr instruction with no offset, turn
4723 // this into a cast of the original pointer!
4725 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
4726 bool AllZeroOperands = true;
4727 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
4728 if (!isa<Constant>(GEP->getOperand(i)) ||
4729 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
4730 AllZeroOperands = false;
4733 if (AllZeroOperands) {
4734 CI.setOperand(0, GEP->getOperand(0));
4739 // If we are casting a malloc or alloca to a pointer to a type of the same
4740 // size, rewrite the allocation instruction to allocate the "right" type.
4742 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
4743 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
4746 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
4747 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
4749 if (isa<PHINode>(Src))
4750 if (Instruction *NV = FoldOpIntoPhi(CI))
4753 // If the source value is an instruction with only this use, we can attempt to
4754 // propagate the cast into the instruction. Also, only handle integral types
4756 if (Instruction *SrcI = dyn_cast<Instruction>(Src))
4757 if (SrcI->hasOneUse() && Src->getType()->isIntegral() &&
4758 CI.getType()->isInteger()) { // Don't mess with casts to bool here
4759 const Type *DestTy = CI.getType();
4760 unsigned SrcBitSize = Src->getType()->getPrimitiveSizeInBits();
4761 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
4763 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
4764 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
4766 switch (SrcI->getOpcode()) {
4767 case Instruction::Add:
4768 case Instruction::Mul:
4769 case Instruction::And:
4770 case Instruction::Or:
4771 case Instruction::Xor:
4772 // If we are discarding information, or just changing the sign, rewrite.
4773 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
4774 // Don't insert two casts if they cannot be eliminated. We allow two
4775 // casts to be inserted if the sizes are the same. This could only be
4776 // converting signedness, which is a noop.
4777 if (DestBitSize == SrcBitSize || !ValueRequiresCast(Op1, DestTy,TD) ||
4778 !ValueRequiresCast(Op0, DestTy, TD)) {
4779 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4780 Value *Op1c = InsertOperandCastBefore(Op1, DestTy, SrcI);
4781 return BinaryOperator::create(cast<BinaryOperator>(SrcI)
4782 ->getOpcode(), Op0c, Op1c);
4786 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
4787 if (SrcBitSize == 1 && SrcI->getOpcode() == Instruction::Xor &&
4788 Op1 == ConstantBool::True &&
4789 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
4790 Value *New = InsertOperandCastBefore(Op0, DestTy, &CI);
4791 return BinaryOperator::createXor(New,
4792 ConstantInt::get(CI.getType(), 1));
4795 case Instruction::Shl:
4796 // Allow changing the sign of the source operand. Do not allow changing
4797 // the size of the shift, UNLESS the shift amount is a constant. We
4798 // mush not change variable sized shifts to a smaller size, because it
4799 // is undefined to shift more bits out than exist in the value.
4800 if (DestBitSize == SrcBitSize ||
4801 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
4802 Value *Op0c = InsertOperandCastBefore(Op0, DestTy, SrcI);
4803 return new ShiftInst(Instruction::Shl, Op0c, Op1);
4806 case Instruction::Shr:
4807 // If this is a signed shr, and if all bits shifted in are about to be
4808 // truncated off, turn it into an unsigned shr to allow greater
4810 if (DestBitSize < SrcBitSize && Src->getType()->isSigned() &&
4811 isa<ConstantInt>(Op1)) {
4812 unsigned ShiftAmt = cast<ConstantUInt>(Op1)->getValue();
4813 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
4814 // Convert to unsigned.
4815 Value *N1 = InsertOperandCastBefore(Op0,
4816 Op0->getType()->getUnsignedVersion(), &CI);
4817 // Insert the new shift, which is now unsigned.
4818 N1 = InsertNewInstBefore(new ShiftInst(Instruction::Shr, N1,
4819 Op1, Src->getName()), CI);
4820 return new CastInst(N1, CI.getType());
4825 case Instruction::SetEQ:
4826 case Instruction::SetNE:
4827 // We if we are just checking for a seteq of a single bit and casting it
4828 // to an integer. If so, shift the bit to the appropriate place then
4829 // cast to integer to avoid the comparison.
4830 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
4831 uint64_t Op1CV = Op1C->getZExtValue();
4832 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
4833 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
4834 // cast (X == 1) to int --> X iff X has only the low bit set.
4835 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
4836 // cast (X != 0) to int --> X iff X has only the low bit set.
4837 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
4838 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
4839 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
4840 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
4841 // If Op1C some other power of two, convert:
4842 uint64_t KnownZero, KnownOne;
4843 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
4844 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
4846 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly one possible 1?
4847 bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
4848 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
4849 // (X&4) == 2 --> false
4850 // (X&4) != 2 --> true
4851 Constant *Res = ConstantBool::get(isSetNE);
4852 Res = ConstantExpr::getCast(Res, CI.getType());
4853 return ReplaceInstUsesWith(CI, Res);
4856 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
4859 // Perform an unsigned shr by shiftamt. Convert input to
4860 // unsigned if it is signed.
4861 if (In->getType()->isSigned())
4862 In = InsertNewInstBefore(new CastInst(In,
4863 In->getType()->getUnsignedVersion(), In->getName()),CI);
4864 // Insert the shift to put the result in the low bit.
4865 In = InsertNewInstBefore(new ShiftInst(Instruction::Shr, In,
4866 ConstantInt::get(Type::UByteTy, ShiftAmt),
4867 In->getName()+".lobit"), CI);
4870 if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
4871 Constant *One = ConstantInt::get(In->getType(), 1);
4872 In = BinaryOperator::createXor(In, One, "tmp");
4873 InsertNewInstBefore(cast<Instruction>(In), CI);
4876 if (CI.getType() == In->getType())
4877 return ReplaceInstUsesWith(CI, In);
4879 return new CastInst(In, CI.getType());
4890 /// GetSelectFoldableOperands - We want to turn code that looks like this:
4892 /// %D = select %cond, %C, %A
4894 /// %C = select %cond, %B, 0
4897 /// Assuming that the specified instruction is an operand to the select, return
4898 /// a bitmask indicating which operands of this instruction are foldable if they
4899 /// equal the other incoming value of the select.
4901 static unsigned GetSelectFoldableOperands(Instruction *I) {
4902 switch (I->getOpcode()) {
4903 case Instruction::Add:
4904 case Instruction::Mul:
4905 case Instruction::And:
4906 case Instruction::Or:
4907 case Instruction::Xor:
4908 return 3; // Can fold through either operand.
4909 case Instruction::Sub: // Can only fold on the amount subtracted.
4910 case Instruction::Shl: // Can only fold on the shift amount.
4911 case Instruction::Shr:
4914 return 0; // Cannot fold
4918 /// GetSelectFoldableConstant - For the same transformation as the previous
4919 /// function, return the identity constant that goes into the select.
4920 static Constant *GetSelectFoldableConstant(Instruction *I) {
4921 switch (I->getOpcode()) {
4922 default: assert(0 && "This cannot happen!"); abort();
4923 case Instruction::Add:
4924 case Instruction::Sub:
4925 case Instruction::Or:
4926 case Instruction::Xor:
4927 return Constant::getNullValue(I->getType());
4928 case Instruction::Shl:
4929 case Instruction::Shr:
4930 return Constant::getNullValue(Type::UByteTy);
4931 case Instruction::And:
4932 return ConstantInt::getAllOnesValue(I->getType());
4933 case Instruction::Mul:
4934 return ConstantInt::get(I->getType(), 1);
4938 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
4939 /// have the same opcode and only one use each. Try to simplify this.
4940 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
4942 if (TI->getNumOperands() == 1) {
4943 // If this is a non-volatile load or a cast from the same type,
4945 if (TI->getOpcode() == Instruction::Cast) {
4946 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
4949 return 0; // unknown unary op.
4952 // Fold this by inserting a select from the input values.
4953 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
4954 FI->getOperand(0), SI.getName()+".v");
4955 InsertNewInstBefore(NewSI, SI);
4956 return new CastInst(NewSI, TI->getType());
4959 // Only handle binary operators here.
4960 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
4963 // Figure out if the operations have any operands in common.
4964 Value *MatchOp, *OtherOpT, *OtherOpF;
4966 if (TI->getOperand(0) == FI->getOperand(0)) {
4967 MatchOp = TI->getOperand(0);
4968 OtherOpT = TI->getOperand(1);
4969 OtherOpF = FI->getOperand(1);
4970 MatchIsOpZero = true;
4971 } else if (TI->getOperand(1) == FI->getOperand(1)) {
4972 MatchOp = TI->getOperand(1);
4973 OtherOpT = TI->getOperand(0);
4974 OtherOpF = FI->getOperand(0);
4975 MatchIsOpZero = false;
4976 } else if (!TI->isCommutative()) {
4978 } else if (TI->getOperand(0) == FI->getOperand(1)) {
4979 MatchOp = TI->getOperand(0);
4980 OtherOpT = TI->getOperand(1);
4981 OtherOpF = FI->getOperand(0);
4982 MatchIsOpZero = true;
4983 } else if (TI->getOperand(1) == FI->getOperand(0)) {
4984 MatchOp = TI->getOperand(1);
4985 OtherOpT = TI->getOperand(0);
4986 OtherOpF = FI->getOperand(1);
4987 MatchIsOpZero = true;
4992 // If we reach here, they do have operations in common.
4993 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
4994 OtherOpF, SI.getName()+".v");
4995 InsertNewInstBefore(NewSI, SI);
4997 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
4999 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
5001 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
5004 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
5006 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
5010 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
5011 Value *CondVal = SI.getCondition();
5012 Value *TrueVal = SI.getTrueValue();
5013 Value *FalseVal = SI.getFalseValue();
5015 // select true, X, Y -> X
5016 // select false, X, Y -> Y
5017 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
5018 if (C == ConstantBool::True)
5019 return ReplaceInstUsesWith(SI, TrueVal);
5021 assert(C == ConstantBool::False);
5022 return ReplaceInstUsesWith(SI, FalseVal);
5025 // select C, X, X -> X
5026 if (TrueVal == FalseVal)
5027 return ReplaceInstUsesWith(SI, TrueVal);
5029 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
5030 return ReplaceInstUsesWith(SI, FalseVal);
5031 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
5032 return ReplaceInstUsesWith(SI, TrueVal);
5033 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
5034 if (isa<Constant>(TrueVal))
5035 return ReplaceInstUsesWith(SI, TrueVal);
5037 return ReplaceInstUsesWith(SI, FalseVal);
5040 if (SI.getType() == Type::BoolTy)
5041 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
5042 if (C == ConstantBool::True) {
5043 // Change: A = select B, true, C --> A = or B, C
5044 return BinaryOperator::createOr(CondVal, FalseVal);
5046 // Change: A = select B, false, C --> A = and !B, C
5048 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5049 "not."+CondVal->getName()), SI);
5050 return BinaryOperator::createAnd(NotCond, FalseVal);
5052 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
5053 if (C == ConstantBool::False) {
5054 // Change: A = select B, C, false --> A = and B, C
5055 return BinaryOperator::createAnd(CondVal, TrueVal);
5057 // Change: A = select B, C, true --> A = or !B, C
5059 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5060 "not."+CondVal->getName()), SI);
5061 return BinaryOperator::createOr(NotCond, TrueVal);
5065 // Selecting between two integer constants?
5066 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
5067 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
5068 // select C, 1, 0 -> cast C to int
5069 if (FalseValC->isNullValue() && TrueValC->getRawValue() == 1) {
5070 return new CastInst(CondVal, SI.getType());
5071 } else if (TrueValC->isNullValue() && FalseValC->getRawValue() == 1) {
5072 // select C, 0, 1 -> cast !C to int
5074 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
5075 "not."+CondVal->getName()), SI);
5076 return new CastInst(NotCond, SI.getType());
5079 // If one of the constants is zero (we know they can't both be) and we
5080 // have a setcc instruction with zero, and we have an 'and' with the
5081 // non-constant value, eliminate this whole mess. This corresponds to
5082 // cases like this: ((X & 27) ? 27 : 0)
5083 if (TrueValC->isNullValue() || FalseValC->isNullValue())
5084 if (Instruction *IC = dyn_cast<Instruction>(SI.getCondition()))
5085 if ((IC->getOpcode() == Instruction::SetEQ ||
5086 IC->getOpcode() == Instruction::SetNE) &&
5087 isa<ConstantInt>(IC->getOperand(1)) &&
5088 cast<Constant>(IC->getOperand(1))->isNullValue())
5089 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
5090 if (ICA->getOpcode() == Instruction::And &&
5091 isa<ConstantInt>(ICA->getOperand(1)) &&
5092 (ICA->getOperand(1) == TrueValC ||
5093 ICA->getOperand(1) == FalseValC) &&
5094 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
5095 // Okay, now we know that everything is set up, we just don't
5096 // know whether we have a setne or seteq and whether the true or
5097 // false val is the zero.
5098 bool ShouldNotVal = !TrueValC->isNullValue();
5099 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
5102 V = InsertNewInstBefore(BinaryOperator::create(
5103 Instruction::Xor, V, ICA->getOperand(1)), SI);
5104 return ReplaceInstUsesWith(SI, V);
5108 // See if we are selecting two values based on a comparison of the two values.
5109 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
5110 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
5111 // Transform (X == Y) ? X : Y -> Y
5112 if (SCI->getOpcode() == Instruction::SetEQ)
5113 return ReplaceInstUsesWith(SI, FalseVal);
5114 // Transform (X != Y) ? X : Y -> X
5115 if (SCI->getOpcode() == Instruction::SetNE)
5116 return ReplaceInstUsesWith(SI, TrueVal);
5117 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5119 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
5120 // Transform (X == Y) ? Y : X -> X
5121 if (SCI->getOpcode() == Instruction::SetEQ)
5122 return ReplaceInstUsesWith(SI, FalseVal);
5123 // Transform (X != Y) ? Y : X -> Y
5124 if (SCI->getOpcode() == Instruction::SetNE)
5125 return ReplaceInstUsesWith(SI, TrueVal);
5126 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
5130 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
5131 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
5132 if (TI->hasOneUse() && FI->hasOneUse()) {
5133 bool isInverse = false;
5134 Instruction *AddOp = 0, *SubOp = 0;
5136 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
5137 if (TI->getOpcode() == FI->getOpcode())
5138 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
5141 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
5142 // even legal for FP.
5143 if (TI->getOpcode() == Instruction::Sub &&
5144 FI->getOpcode() == Instruction::Add) {
5145 AddOp = FI; SubOp = TI;
5146 } else if (FI->getOpcode() == Instruction::Sub &&
5147 TI->getOpcode() == Instruction::Add) {
5148 AddOp = TI; SubOp = FI;
5152 Value *OtherAddOp = 0;
5153 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
5154 OtherAddOp = AddOp->getOperand(1);
5155 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
5156 OtherAddOp = AddOp->getOperand(0);
5160 // So at this point we know we have (Y -> OtherAddOp):
5161 // select C, (add X, Y), (sub X, Z)
5162 Value *NegVal; // Compute -Z
5163 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
5164 NegVal = ConstantExpr::getNeg(C);
5166 NegVal = InsertNewInstBefore(
5167 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
5170 Value *NewTrueOp = OtherAddOp;
5171 Value *NewFalseOp = NegVal;
5173 std::swap(NewTrueOp, NewFalseOp);
5174 Instruction *NewSel =
5175 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
5177 NewSel = InsertNewInstBefore(NewSel, SI);
5178 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
5183 // See if we can fold the select into one of our operands.
5184 if (SI.getType()->isInteger()) {
5185 // See the comment above GetSelectFoldableOperands for a description of the
5186 // transformation we are doing here.
5187 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
5188 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
5189 !isa<Constant>(FalseVal))
5190 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
5191 unsigned OpToFold = 0;
5192 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
5194 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
5199 Constant *C = GetSelectFoldableConstant(TVI);
5200 std::string Name = TVI->getName(); TVI->setName("");
5201 Instruction *NewSel =
5202 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
5204 InsertNewInstBefore(NewSel, SI);
5205 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
5206 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
5207 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
5208 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
5210 assert(0 && "Unknown instruction!!");
5215 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
5216 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
5217 !isa<Constant>(TrueVal))
5218 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
5219 unsigned OpToFold = 0;
5220 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
5222 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
5227 Constant *C = GetSelectFoldableConstant(FVI);
5228 std::string Name = FVI->getName(); FVI->setName("");
5229 Instruction *NewSel =
5230 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
5232 InsertNewInstBefore(NewSel, SI);
5233 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
5234 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
5235 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
5236 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
5238 assert(0 && "Unknown instruction!!");
5244 if (BinaryOperator::isNot(CondVal)) {
5245 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
5246 SI.setOperand(1, FalseVal);
5247 SI.setOperand(2, TrueVal);
5255 /// visitCallInst - CallInst simplification. This mostly only handles folding
5256 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
5257 /// the heavy lifting.
5259 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
5260 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
5261 if (!II) return visitCallSite(&CI);
5263 // Intrinsics cannot occur in an invoke, so handle them here instead of in
5265 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
5266 bool Changed = false;
5268 // memmove/cpy/set of zero bytes is a noop.
5269 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
5270 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
5272 // FIXME: Increase alignment here.
5274 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
5275 if (CI->getRawValue() == 1) {
5276 // Replace the instruction with just byte operations. We would
5277 // transform other cases to loads/stores, but we don't know if
5278 // alignment is sufficient.
5282 // If we have a memmove and the source operation is a constant global,
5283 // then the source and dest pointers can't alias, so we can change this
5284 // into a call to memcpy.
5285 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II))
5286 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
5287 if (GVSrc->isConstant()) {
5288 Module *M = CI.getParent()->getParent()->getParent();
5290 if (CI.getCalledFunction()->getFunctionType()->getParamType(3) ==
5292 Name = "llvm.memcpy.i32";
5294 Name = "llvm.memcpy.i64";
5295 Function *MemCpy = M->getOrInsertFunction(Name,
5296 CI.getCalledFunction()->getFunctionType());
5297 CI.setOperand(0, MemCpy);
5301 if (Changed) return II;
5302 } else if (DbgStopPointInst *SPI = dyn_cast<DbgStopPointInst>(II)) {
5303 // If this stoppoint is at the same source location as the previous
5304 // stoppoint in the chain, it is not needed.
5305 if (DbgStopPointInst *PrevSPI =
5306 dyn_cast<DbgStopPointInst>(SPI->getChain()))
5307 if (SPI->getLineNo() == PrevSPI->getLineNo() &&
5308 SPI->getColNo() == PrevSPI->getColNo()) {
5309 SPI->replaceAllUsesWith(PrevSPI);
5310 return EraseInstFromFunction(CI);
5313 switch (II->getIntrinsicID()) {
5315 case Intrinsic::stackrestore: {
5316 // If the save is right next to the restore, remove the restore. This can
5317 // happen when variable allocas are DCE'd.
5318 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
5319 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
5320 BasicBlock::iterator BI = SS;
5322 return EraseInstFromFunction(CI);
5326 // If the stack restore is in a return/unwind block and if there are no
5327 // allocas or calls between the restore and the return, nuke the restore.
5328 TerminatorInst *TI = II->getParent()->getTerminator();
5329 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
5330 BasicBlock::iterator BI = II;
5331 bool CannotRemove = false;
5332 for (++BI; &*BI != TI; ++BI) {
5333 if (isa<AllocaInst>(BI) ||
5334 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
5335 CannotRemove = true;
5340 return EraseInstFromFunction(CI);
5347 return visitCallSite(II);
5350 // InvokeInst simplification
5352 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
5353 return visitCallSite(&II);
5356 // visitCallSite - Improvements for call and invoke instructions.
5358 Instruction *InstCombiner::visitCallSite(CallSite CS) {
5359 bool Changed = false;
5361 // If the callee is a constexpr cast of a function, attempt to move the cast
5362 // to the arguments of the call/invoke.
5363 if (transformConstExprCastCall(CS)) return 0;
5365 Value *Callee = CS.getCalledValue();
5367 if (Function *CalleeF = dyn_cast<Function>(Callee))
5368 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
5369 Instruction *OldCall = CS.getInstruction();
5370 // If the call and callee calling conventions don't match, this call must
5371 // be unreachable, as the call is undefined.
5372 new StoreInst(ConstantBool::True,
5373 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
5374 if (!OldCall->use_empty())
5375 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
5376 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
5377 return EraseInstFromFunction(*OldCall);
5381 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
5382 // This instruction is not reachable, just remove it. We insert a store to
5383 // undef so that we know that this code is not reachable, despite the fact
5384 // that we can't modify the CFG here.
5385 new StoreInst(ConstantBool::True,
5386 UndefValue::get(PointerType::get(Type::BoolTy)),
5387 CS.getInstruction());
5389 if (!CS.getInstruction()->use_empty())
5390 CS.getInstruction()->
5391 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
5393 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
5394 // Don't break the CFG, insert a dummy cond branch.
5395 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
5396 ConstantBool::True, II);
5398 return EraseInstFromFunction(*CS.getInstruction());
5401 const PointerType *PTy = cast<PointerType>(Callee->getType());
5402 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
5403 if (FTy->isVarArg()) {
5404 // See if we can optimize any arguments passed through the varargs area of
5406 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
5407 E = CS.arg_end(); I != E; ++I)
5408 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
5409 // If this cast does not effect the value passed through the varargs
5410 // area, we can eliminate the use of the cast.
5411 Value *Op = CI->getOperand(0);
5412 if (CI->getType()->isLosslesslyConvertibleTo(Op->getType())) {
5419 return Changed ? CS.getInstruction() : 0;
5422 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
5423 // attempt to move the cast to the arguments of the call/invoke.
5425 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
5426 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
5427 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
5428 if (CE->getOpcode() != Instruction::Cast || !isa<Function>(CE->getOperand(0)))
5430 Function *Callee = cast<Function>(CE->getOperand(0));
5431 Instruction *Caller = CS.getInstruction();
5433 // Okay, this is a cast from a function to a different type. Unless doing so
5434 // would cause a type conversion of one of our arguments, change this call to
5435 // be a direct call with arguments casted to the appropriate types.
5437 const FunctionType *FT = Callee->getFunctionType();
5438 const Type *OldRetTy = Caller->getType();
5440 // Check to see if we are changing the return type...
5441 if (OldRetTy != FT->getReturnType()) {
5442 if (Callee->isExternal() &&
5443 !OldRetTy->isLosslesslyConvertibleTo(FT->getReturnType()) &&
5444 !Caller->use_empty())
5445 return false; // Cannot transform this return value...
5447 // If the callsite is an invoke instruction, and the return value is used by
5448 // a PHI node in a successor, we cannot change the return type of the call
5449 // because there is no place to put the cast instruction (without breaking
5450 // the critical edge). Bail out in this case.
5451 if (!Caller->use_empty())
5452 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
5453 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
5455 if (PHINode *PN = dyn_cast<PHINode>(*UI))
5456 if (PN->getParent() == II->getNormalDest() ||
5457 PN->getParent() == II->getUnwindDest())
5461 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
5462 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
5464 CallSite::arg_iterator AI = CS.arg_begin();
5465 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
5466 const Type *ParamTy = FT->getParamType(i);
5467 bool isConvertible = (*AI)->getType()->isLosslesslyConvertibleTo(ParamTy);
5468 if (Callee->isExternal() && !isConvertible) return false;
5471 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
5472 Callee->isExternal())
5473 return false; // Do not delete arguments unless we have a function body...
5475 // Okay, we decided that this is a safe thing to do: go ahead and start
5476 // inserting cast instructions as necessary...
5477 std::vector<Value*> Args;
5478 Args.reserve(NumActualArgs);
5480 AI = CS.arg_begin();
5481 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
5482 const Type *ParamTy = FT->getParamType(i);
5483 if ((*AI)->getType() == ParamTy) {
5484 Args.push_back(*AI);
5486 Args.push_back(InsertNewInstBefore(new CastInst(*AI, ParamTy, "tmp"),
5491 // If the function takes more arguments than the call was taking, add them
5493 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
5494 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
5496 // If we are removing arguments to the function, emit an obnoxious warning...
5497 if (FT->getNumParams() < NumActualArgs)
5498 if (!FT->isVarArg()) {
5499 std::cerr << "WARNING: While resolving call to function '"
5500 << Callee->getName() << "' arguments were dropped!\n";
5502 // Add all of the arguments in their promoted form to the arg list...
5503 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
5504 const Type *PTy = getPromotedType((*AI)->getType());
5505 if (PTy != (*AI)->getType()) {
5506 // Must promote to pass through va_arg area!
5507 Instruction *Cast = new CastInst(*AI, PTy, "tmp");
5508 InsertNewInstBefore(Cast, *Caller);
5509 Args.push_back(Cast);
5511 Args.push_back(*AI);
5516 if (FT->getReturnType() == Type::VoidTy)
5517 Caller->setName(""); // Void type should not have a name...
5520 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5521 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
5522 Args, Caller->getName(), Caller);
5523 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
5525 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
5526 if (cast<CallInst>(Caller)->isTailCall())
5527 cast<CallInst>(NC)->setTailCall();
5528 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
5531 // Insert a cast of the return type as necessary...
5533 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
5534 if (NV->getType() != Type::VoidTy) {
5535 NV = NC = new CastInst(NC, Caller->getType(), "tmp");
5537 // If this is an invoke instruction, we should insert it after the first
5538 // non-phi, instruction in the normal successor block.
5539 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5540 BasicBlock::iterator I = II->getNormalDest()->begin();
5541 while (isa<PHINode>(I)) ++I;
5542 InsertNewInstBefore(NC, *I);
5544 // Otherwise, it's a call, just insert cast right after the call instr
5545 InsertNewInstBefore(NC, *Caller);
5547 AddUsersToWorkList(*Caller);
5549 NV = UndefValue::get(Caller->getType());
5553 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
5554 Caller->replaceAllUsesWith(NV);
5555 Caller->getParent()->getInstList().erase(Caller);
5556 removeFromWorkList(Caller);
5561 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
5562 // operator and they all are only used by the PHI, PHI together their
5563 // inputs, and do the operation once, to the result of the PHI.
5564 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
5565 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
5567 // Scan the instruction, looking for input operations that can be folded away.
5568 // If all input operands to the phi are the same instruction (e.g. a cast from
5569 // the same type or "+42") we can pull the operation through the PHI, reducing
5570 // code size and simplifying code.
5571 Constant *ConstantOp = 0;
5572 const Type *CastSrcTy = 0;
5573 if (isa<CastInst>(FirstInst)) {
5574 CastSrcTy = FirstInst->getOperand(0)->getType();
5575 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
5576 // Can fold binop or shift if the RHS is a constant.
5577 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
5578 if (ConstantOp == 0) return 0;
5580 return 0; // Cannot fold this operation.
5583 // Check to see if all arguments are the same operation.
5584 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5585 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
5586 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
5587 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
5590 if (I->getOperand(0)->getType() != CastSrcTy)
5591 return 0; // Cast operation must match.
5592 } else if (I->getOperand(1) != ConstantOp) {
5597 // Okay, they are all the same operation. Create a new PHI node of the
5598 // correct type, and PHI together all of the LHS's of the instructions.
5599 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
5600 PN.getName()+".in");
5601 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
5603 Value *InVal = FirstInst->getOperand(0);
5604 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
5606 // Add all operands to the new PHI.
5607 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
5608 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
5609 if (NewInVal != InVal)
5611 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
5616 // The new PHI unions all of the same values together. This is really
5617 // common, so we handle it intelligently here for compile-time speed.
5621 InsertNewInstBefore(NewPN, PN);
5625 // Insert and return the new operation.
5626 if (isa<CastInst>(FirstInst))
5627 return new CastInst(PhiVal, PN.getType());
5628 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
5629 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
5631 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
5632 PhiVal, ConstantOp);
5635 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
5637 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
5638 if (PN->use_empty()) return true;
5639 if (!PN->hasOneUse()) return false;
5641 // Remember this node, and if we find the cycle, return.
5642 if (!PotentiallyDeadPHIs.insert(PN).second)
5645 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
5646 return DeadPHICycle(PU, PotentiallyDeadPHIs);
5651 // PHINode simplification
5653 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
5654 if (Value *V = PN.hasConstantValue())
5655 return ReplaceInstUsesWith(PN, V);
5657 // If the only user of this instruction is a cast instruction, and all of the
5658 // incoming values are constants, change this PHI to merge together the casted
5661 if (CastInst *CI = dyn_cast<CastInst>(PN.use_back()))
5662 if (CI->getType() != PN.getType()) { // noop casts will be folded
5663 bool AllConstant = true;
5664 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
5665 if (!isa<Constant>(PN.getIncomingValue(i))) {
5666 AllConstant = false;
5670 // Make a new PHI with all casted values.
5671 PHINode *New = new PHINode(CI->getType(), PN.getName(), &PN);
5672 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
5673 Constant *OldArg = cast<Constant>(PN.getIncomingValue(i));
5674 New->addIncoming(ConstantExpr::getCast(OldArg, New->getType()),
5675 PN.getIncomingBlock(i));
5678 // Update the cast instruction.
5679 CI->setOperand(0, New);
5680 WorkList.push_back(CI); // revisit the cast instruction to fold.
5681 WorkList.push_back(New); // Make sure to revisit the new Phi
5682 return &PN; // PN is now dead!
5686 // If all PHI operands are the same operation, pull them through the PHI,
5687 // reducing code size.
5688 if (isa<Instruction>(PN.getIncomingValue(0)) &&
5689 PN.getIncomingValue(0)->hasOneUse())
5690 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
5693 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
5694 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
5695 // PHI)... break the cycle.
5697 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
5698 std::set<PHINode*> PotentiallyDeadPHIs;
5699 PotentiallyDeadPHIs.insert(&PN);
5700 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
5701 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
5707 static Value *InsertSignExtendToPtrTy(Value *V, const Type *DTy,
5708 Instruction *InsertPoint,
5710 unsigned PS = IC->getTargetData().getPointerSize();
5711 const Type *VTy = V->getType();
5712 if (!VTy->isSigned() && VTy->getPrimitiveSize() < PS)
5713 // We must insert a cast to ensure we sign-extend.
5714 V = IC->InsertNewInstBefore(new CastInst(V, VTy->getSignedVersion(),
5715 V->getName()), *InsertPoint);
5716 return IC->InsertNewInstBefore(new CastInst(V, DTy, V->getName()),
5721 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
5722 Value *PtrOp = GEP.getOperand(0);
5723 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
5724 // If so, eliminate the noop.
5725 if (GEP.getNumOperands() == 1)
5726 return ReplaceInstUsesWith(GEP, PtrOp);
5728 if (isa<UndefValue>(GEP.getOperand(0)))
5729 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
5731 bool HasZeroPointerIndex = false;
5732 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
5733 HasZeroPointerIndex = C->isNullValue();
5735 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
5736 return ReplaceInstUsesWith(GEP, PtrOp);
5738 // Eliminate unneeded casts for indices.
5739 bool MadeChange = false;
5740 gep_type_iterator GTI = gep_type_begin(GEP);
5741 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
5742 if (isa<SequentialType>(*GTI)) {
5743 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
5744 Value *Src = CI->getOperand(0);
5745 const Type *SrcTy = Src->getType();
5746 const Type *DestTy = CI->getType();
5747 if (Src->getType()->isInteger()) {
5748 if (SrcTy->getPrimitiveSizeInBits() ==
5749 DestTy->getPrimitiveSizeInBits()) {
5750 // We can always eliminate a cast from ulong or long to the other.
5751 // We can always eliminate a cast from uint to int or the other on
5752 // 32-bit pointer platforms.
5753 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
5755 GEP.setOperand(i, Src);
5757 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
5758 SrcTy->getPrimitiveSize() == 4) {
5759 // We can always eliminate a cast from int to [u]long. We can
5760 // eliminate a cast from uint to [u]long iff the target is a 32-bit
5762 if (SrcTy->isSigned() ||
5763 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
5765 GEP.setOperand(i, Src);
5770 // If we are using a wider index than needed for this platform, shrink it
5771 // to what we need. If the incoming value needs a cast instruction,
5772 // insert it. This explicit cast can make subsequent optimizations more
5774 Value *Op = GEP.getOperand(i);
5775 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
5776 if (Constant *C = dyn_cast<Constant>(Op)) {
5777 GEP.setOperand(i, ConstantExpr::getCast(C,
5778 TD->getIntPtrType()->getSignedVersion()));
5781 Op = InsertNewInstBefore(new CastInst(Op, TD->getIntPtrType(),
5782 Op->getName()), GEP);
5783 GEP.setOperand(i, Op);
5787 // If this is a constant idx, make sure to canonicalize it to be a signed
5788 // operand, otherwise CSE and other optimizations are pessimized.
5789 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(Op)) {
5790 GEP.setOperand(i, ConstantExpr::getCast(CUI,
5791 CUI->getType()->getSignedVersion()));
5795 if (MadeChange) return &GEP;
5797 // Combine Indices - If the source pointer to this getelementptr instruction
5798 // is a getelementptr instruction, combine the indices of the two
5799 // getelementptr instructions into a single instruction.
5801 std::vector<Value*> SrcGEPOperands;
5802 if (User *Src = dyn_castGetElementPtr(PtrOp))
5803 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
5805 if (!SrcGEPOperands.empty()) {
5806 // Note that if our source is a gep chain itself that we wait for that
5807 // chain to be resolved before we perform this transformation. This
5808 // avoids us creating a TON of code in some cases.
5810 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
5811 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
5812 return 0; // Wait until our source is folded to completion.
5814 std::vector<Value *> Indices;
5816 // Find out whether the last index in the source GEP is a sequential idx.
5817 bool EndsWithSequential = false;
5818 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
5819 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
5820 EndsWithSequential = !isa<StructType>(*I);
5822 // Can we combine the two pointer arithmetics offsets?
5823 if (EndsWithSequential) {
5824 // Replace: gep (gep %P, long B), long A, ...
5825 // With: T = long A+B; gep %P, T, ...
5827 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
5828 if (SO1 == Constant::getNullValue(SO1->getType())) {
5830 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
5833 // If they aren't the same type, convert both to an integer of the
5834 // target's pointer size.
5835 if (SO1->getType() != GO1->getType()) {
5836 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
5837 SO1 = ConstantExpr::getCast(SO1C, GO1->getType());
5838 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
5839 GO1 = ConstantExpr::getCast(GO1C, SO1->getType());
5841 unsigned PS = TD->getPointerSize();
5842 if (SO1->getType()->getPrimitiveSize() == PS) {
5843 // Convert GO1 to SO1's type.
5844 GO1 = InsertSignExtendToPtrTy(GO1, SO1->getType(), &GEP, this);
5846 } else if (GO1->getType()->getPrimitiveSize() == PS) {
5847 // Convert SO1 to GO1's type.
5848 SO1 = InsertSignExtendToPtrTy(SO1, GO1->getType(), &GEP, this);
5850 const Type *PT = TD->getIntPtrType();
5851 SO1 = InsertSignExtendToPtrTy(SO1, PT, &GEP, this);
5852 GO1 = InsertSignExtendToPtrTy(GO1, PT, &GEP, this);
5856 if (isa<Constant>(SO1) && isa<Constant>(GO1))
5857 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
5859 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
5860 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
5864 // Recycle the GEP we already have if possible.
5865 if (SrcGEPOperands.size() == 2) {
5866 GEP.setOperand(0, SrcGEPOperands[0]);
5867 GEP.setOperand(1, Sum);
5870 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5871 SrcGEPOperands.end()-1);
5872 Indices.push_back(Sum);
5873 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
5875 } else if (isa<Constant>(*GEP.idx_begin()) &&
5876 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
5877 SrcGEPOperands.size() != 1) {
5878 // Otherwise we can do the fold if the first index of the GEP is a zero
5879 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
5880 SrcGEPOperands.end());
5881 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
5884 if (!Indices.empty())
5885 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
5887 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
5888 // GEP of global variable. If all of the indices for this GEP are
5889 // constants, we can promote this to a constexpr instead of an instruction.
5891 // Scan for nonconstants...
5892 std::vector<Constant*> Indices;
5893 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
5894 for (; I != E && isa<Constant>(*I); ++I)
5895 Indices.push_back(cast<Constant>(*I));
5897 if (I == E) { // If they are all constants...
5898 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
5900 // Replace all uses of the GEP with the new constexpr...
5901 return ReplaceInstUsesWith(GEP, CE);
5903 } else if (Value *X = isCast(PtrOp)) { // Is the operand a cast?
5904 if (!isa<PointerType>(X->getType())) {
5905 // Not interesting. Source pointer must be a cast from pointer.
5906 } else if (HasZeroPointerIndex) {
5907 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
5908 // into : GEP [10 x ubyte]* X, long 0, ...
5910 // This occurs when the program declares an array extern like "int X[];"
5912 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
5913 const PointerType *XTy = cast<PointerType>(X->getType());
5914 if (const ArrayType *XATy =
5915 dyn_cast<ArrayType>(XTy->getElementType()))
5916 if (const ArrayType *CATy =
5917 dyn_cast<ArrayType>(CPTy->getElementType()))
5918 if (CATy->getElementType() == XATy->getElementType()) {
5919 // At this point, we know that the cast source type is a pointer
5920 // to an array of the same type as the destination pointer
5921 // array. Because the array type is never stepped over (there
5922 // is a leading zero) we can fold the cast into this GEP.
5923 GEP.setOperand(0, X);
5926 } else if (GEP.getNumOperands() == 2) {
5927 // Transform things like:
5928 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
5929 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
5930 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
5931 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
5932 if (isa<ArrayType>(SrcElTy) &&
5933 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
5934 TD->getTypeSize(ResElTy)) {
5935 Value *V = InsertNewInstBefore(
5936 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5937 GEP.getOperand(1), GEP.getName()), GEP);
5938 return new CastInst(V, GEP.getType());
5941 // Transform things like:
5942 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
5943 // (where tmp = 8*tmp2) into:
5944 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
5946 if (isa<ArrayType>(SrcElTy) &&
5947 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
5948 uint64_t ArrayEltSize =
5949 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
5951 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
5952 // allow either a mul, shift, or constant here.
5954 ConstantInt *Scale = 0;
5955 if (ArrayEltSize == 1) {
5956 NewIdx = GEP.getOperand(1);
5957 Scale = ConstantInt::get(NewIdx->getType(), 1);
5958 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
5959 NewIdx = ConstantInt::get(CI->getType(), 1);
5961 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
5962 if (Inst->getOpcode() == Instruction::Shl &&
5963 isa<ConstantInt>(Inst->getOperand(1))) {
5964 unsigned ShAmt =cast<ConstantUInt>(Inst->getOperand(1))->getValue();
5965 if (Inst->getType()->isSigned())
5966 Scale = ConstantSInt::get(Inst->getType(), 1ULL << ShAmt);
5968 Scale = ConstantUInt::get(Inst->getType(), 1ULL << ShAmt);
5969 NewIdx = Inst->getOperand(0);
5970 } else if (Inst->getOpcode() == Instruction::Mul &&
5971 isa<ConstantInt>(Inst->getOperand(1))) {
5972 Scale = cast<ConstantInt>(Inst->getOperand(1));
5973 NewIdx = Inst->getOperand(0);
5977 // If the index will be to exactly the right offset with the scale taken
5978 // out, perform the transformation.
5979 if (Scale && Scale->getRawValue() % ArrayEltSize == 0) {
5980 if (ConstantSInt *C = dyn_cast<ConstantSInt>(Scale))
5981 Scale = ConstantSInt::get(C->getType(),
5982 (int64_t)C->getRawValue() /
5983 (int64_t)ArrayEltSize);
5985 Scale = ConstantUInt::get(Scale->getType(),
5986 Scale->getRawValue() / ArrayEltSize);
5987 if (Scale->getRawValue() != 1) {
5988 Constant *C = ConstantExpr::getCast(Scale, NewIdx->getType());
5989 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
5990 NewIdx = InsertNewInstBefore(Sc, GEP);
5993 // Insert the new GEP instruction.
5995 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
5996 NewIdx, GEP.getName());
5997 Idx = InsertNewInstBefore(Idx, GEP);
5998 return new CastInst(Idx, GEP.getType());
6007 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
6008 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
6009 if (AI.isArrayAllocation()) // Check C != 1
6010 if (const ConstantUInt *C = dyn_cast<ConstantUInt>(AI.getArraySize())) {
6011 const Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getValue());
6012 AllocationInst *New = 0;
6014 // Create and insert the replacement instruction...
6015 if (isa<MallocInst>(AI))
6016 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
6018 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
6019 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
6022 InsertNewInstBefore(New, AI);
6024 // Scan to the end of the allocation instructions, to skip over a block of
6025 // allocas if possible...
6027 BasicBlock::iterator It = New;
6028 while (isa<AllocationInst>(*It)) ++It;
6030 // Now that I is pointing to the first non-allocation-inst in the block,
6031 // insert our getelementptr instruction...
6033 Value *NullIdx = Constant::getNullValue(Type::IntTy);
6034 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
6035 New->getName()+".sub", It);
6037 // Now make everything use the getelementptr instead of the original
6039 return ReplaceInstUsesWith(AI, V);
6040 } else if (isa<UndefValue>(AI.getArraySize())) {
6041 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6044 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
6045 // Note that we only do this for alloca's, because malloc should allocate and
6046 // return a unique pointer, even for a zero byte allocation.
6047 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
6048 TD->getTypeSize(AI.getAllocatedType()) == 0)
6049 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
6054 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
6055 Value *Op = FI.getOperand(0);
6057 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
6058 if (CastInst *CI = dyn_cast<CastInst>(Op))
6059 if (isa<PointerType>(CI->getOperand(0)->getType())) {
6060 FI.setOperand(0, CI->getOperand(0));
6064 // free undef -> unreachable.
6065 if (isa<UndefValue>(Op)) {
6066 // Insert a new store to null because we cannot modify the CFG here.
6067 new StoreInst(ConstantBool::True,
6068 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
6069 return EraseInstFromFunction(FI);
6072 // If we have 'free null' delete the instruction. This can happen in stl code
6073 // when lots of inlining happens.
6074 if (isa<ConstantPointerNull>(Op))
6075 return EraseInstFromFunction(FI);
6081 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
6082 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
6083 User *CI = cast<User>(LI.getOperand(0));
6084 Value *CastOp = CI->getOperand(0);
6086 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
6087 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
6088 const Type *SrcPTy = SrcTy->getElementType();
6090 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
6091 // If the source is an array, the code below will not succeed. Check to
6092 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
6094 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
6095 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
6096 if (ASrcTy->getNumElements() != 0) {
6097 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
6098 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
6099 SrcTy = cast<PointerType>(CastOp->getType());
6100 SrcPTy = SrcTy->getElementType();
6103 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
6104 // Do not allow turning this into a load of an integer, which is then
6105 // casted to a pointer, this pessimizes pointer analysis a lot.
6106 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
6107 IC.getTargetData().getTypeSize(SrcPTy) ==
6108 IC.getTargetData().getTypeSize(DestPTy)) {
6110 // Okay, we are casting from one integer or pointer type to another of
6111 // the same size. Instead of casting the pointer before the load, cast
6112 // the result of the loaded value.
6113 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
6115 LI.isVolatile()),LI);
6116 // Now cast the result of the load.
6117 return new CastInst(NewLoad, LI.getType());
6124 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
6125 /// from this value cannot trap. If it is not obviously safe to load from the
6126 /// specified pointer, we do a quick local scan of the basic block containing
6127 /// ScanFrom, to determine if the address is already accessed.
6128 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
6129 // If it is an alloca or global variable, it is always safe to load from.
6130 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
6132 // Otherwise, be a little bit agressive by scanning the local block where we
6133 // want to check to see if the pointer is already being loaded or stored
6134 // from/to. If so, the previous load or store would have already trapped,
6135 // so there is no harm doing an extra load (also, CSE will later eliminate
6136 // the load entirely).
6137 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
6142 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
6143 if (LI->getOperand(0) == V) return true;
6144 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
6145 if (SI->getOperand(1) == V) return true;
6151 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
6152 Value *Op = LI.getOperand(0);
6154 // load (cast X) --> cast (load X) iff safe
6155 if (CastInst *CI = dyn_cast<CastInst>(Op))
6156 if (Instruction *Res = InstCombineLoadCast(*this, LI))
6159 // None of the following transforms are legal for volatile loads.
6160 if (LI.isVolatile()) return 0;
6162 if (&LI.getParent()->front() != &LI) {
6163 BasicBlock::iterator BBI = &LI; --BBI;
6164 // If the instruction immediately before this is a store to the same
6165 // address, do a simple form of store->load forwarding.
6166 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
6167 if (SI->getOperand(1) == LI.getOperand(0))
6168 return ReplaceInstUsesWith(LI, SI->getOperand(0));
6169 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
6170 if (LIB->getOperand(0) == LI.getOperand(0))
6171 return ReplaceInstUsesWith(LI, LIB);
6174 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
6175 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
6176 isa<UndefValue>(GEPI->getOperand(0))) {
6177 // Insert a new store to null instruction before the load to indicate
6178 // that this code is not reachable. We do this instead of inserting
6179 // an unreachable instruction directly because we cannot modify the
6181 new StoreInst(UndefValue::get(LI.getType()),
6182 Constant::getNullValue(Op->getType()), &LI);
6183 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
6186 if (Constant *C = dyn_cast<Constant>(Op)) {
6187 // load null/undef -> undef
6188 if ((C->isNullValue() || isa<UndefValue>(C))) {
6189 // Insert a new store to null instruction before the load to indicate that
6190 // this code is not reachable. We do this instead of inserting an
6191 // unreachable instruction directly because we cannot modify the CFG.
6192 new StoreInst(UndefValue::get(LI.getType()),
6193 Constant::getNullValue(Op->getType()), &LI);
6194 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
6197 // Instcombine load (constant global) into the value loaded.
6198 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
6199 if (GV->isConstant() && !GV->isExternal())
6200 return ReplaceInstUsesWith(LI, GV->getInitializer());
6202 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
6203 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
6204 if (CE->getOpcode() == Instruction::GetElementPtr) {
6205 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
6206 if (GV->isConstant() && !GV->isExternal())
6208 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
6209 return ReplaceInstUsesWith(LI, V);
6210 if (CE->getOperand(0)->isNullValue()) {
6211 // Insert a new store to null instruction before the load to indicate
6212 // that this code is not reachable. We do this instead of inserting
6213 // an unreachable instruction directly because we cannot modify the
6215 new StoreInst(UndefValue::get(LI.getType()),
6216 Constant::getNullValue(Op->getType()), &LI);
6217 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
6220 } else if (CE->getOpcode() == Instruction::Cast) {
6221 if (Instruction *Res = InstCombineLoadCast(*this, LI))
6226 if (Op->hasOneUse()) {
6227 // Change select and PHI nodes to select values instead of addresses: this
6228 // helps alias analysis out a lot, allows many others simplifications, and
6229 // exposes redundancy in the code.
6231 // Note that we cannot do the transformation unless we know that the
6232 // introduced loads cannot trap! Something like this is valid as long as
6233 // the condition is always false: load (select bool %C, int* null, int* %G),
6234 // but it would not be valid if we transformed it to load from null
6237 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
6238 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
6239 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
6240 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
6241 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
6242 SI->getOperand(1)->getName()+".val"), LI);
6243 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
6244 SI->getOperand(2)->getName()+".val"), LI);
6245 return new SelectInst(SI->getCondition(), V1, V2);
6248 // load (select (cond, null, P)) -> load P
6249 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
6250 if (C->isNullValue()) {
6251 LI.setOperand(0, SI->getOperand(2));
6255 // load (select (cond, P, null)) -> load P
6256 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
6257 if (C->isNullValue()) {
6258 LI.setOperand(0, SI->getOperand(1));
6262 } else if (PHINode *PN = dyn_cast<PHINode>(Op)) {
6263 // load (phi (&V1, &V2, &V3)) --> phi(load &V1, load &V2, load &V3)
6264 bool Safe = PN->getParent() == LI.getParent();
6266 // Scan all of the instructions between the PHI and the load to make
6267 // sure there are no instructions that might possibly alter the value
6268 // loaded from the PHI.
6270 BasicBlock::iterator I = &LI;
6271 for (--I; !isa<PHINode>(I); --I)
6272 if (isa<StoreInst>(I) || isa<CallInst>(I)) {
6278 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e && Safe; ++i)
6279 if (!isSafeToLoadUnconditionally(PN->getIncomingValue(i),
6280 PN->getIncomingBlock(i)->getTerminator()))
6285 PHINode *NewPN = new PHINode(LI.getType(), PN->getName());
6286 InsertNewInstBefore(NewPN, *PN);
6287 std::map<BasicBlock*,Value*> LoadMap; // Don't insert duplicate loads
6289 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
6290 BasicBlock *BB = PN->getIncomingBlock(i);
6291 Value *&TheLoad = LoadMap[BB];
6293 Value *InVal = PN->getIncomingValue(i);
6294 TheLoad = InsertNewInstBefore(new LoadInst(InVal,
6295 InVal->getName()+".val"),
6296 *BB->getTerminator());
6298 NewPN->addIncoming(TheLoad, BB);
6300 return ReplaceInstUsesWith(LI, NewPN);
6307 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
6309 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
6310 User *CI = cast<User>(SI.getOperand(1));
6311 Value *CastOp = CI->getOperand(0);
6313 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
6314 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
6315 const Type *SrcPTy = SrcTy->getElementType();
6317 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
6318 // If the source is an array, the code below will not succeed. Check to
6319 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
6321 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
6322 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
6323 if (ASrcTy->getNumElements() != 0) {
6324 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
6325 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
6326 SrcTy = cast<PointerType>(CastOp->getType());
6327 SrcPTy = SrcTy->getElementType();
6330 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
6331 IC.getTargetData().getTypeSize(SrcPTy) ==
6332 IC.getTargetData().getTypeSize(DestPTy)) {
6334 // Okay, we are casting from one integer or pointer type to another of
6335 // the same size. Instead of casting the pointer before the store, cast
6336 // the value to be stored.
6338 if (Constant *C = dyn_cast<Constant>(SI.getOperand(0)))
6339 NewCast = ConstantExpr::getCast(C, SrcPTy);
6341 NewCast = IC.InsertNewInstBefore(new CastInst(SI.getOperand(0),
6343 SI.getOperand(0)->getName()+".c"), SI);
6345 return new StoreInst(NewCast, CastOp);
6352 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
6353 Value *Val = SI.getOperand(0);
6354 Value *Ptr = SI.getOperand(1);
6356 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
6357 EraseInstFromFunction(SI);
6362 // Do really simple DSE, to catch cases where there are several consequtive
6363 // stores to the same location, separated by a few arithmetic operations. This
6364 // situation often occurs with bitfield accesses.
6365 BasicBlock::iterator BBI = &SI;
6366 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
6370 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
6371 // Prev store isn't volatile, and stores to the same location?
6372 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
6375 EraseInstFromFunction(*PrevSI);
6381 // Don't skip over loads or things that can modify memory.
6382 if (BBI->mayWriteToMemory() || isa<LoadInst>(BBI))
6387 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
6389 // store X, null -> turns into 'unreachable' in SimplifyCFG
6390 if (isa<ConstantPointerNull>(Ptr)) {
6391 if (!isa<UndefValue>(Val)) {
6392 SI.setOperand(0, UndefValue::get(Val->getType()));
6393 if (Instruction *U = dyn_cast<Instruction>(Val))
6394 WorkList.push_back(U); // Dropped a use.
6397 return 0; // Do not modify these!
6400 // store undef, Ptr -> noop
6401 if (isa<UndefValue>(Val)) {
6402 EraseInstFromFunction(SI);
6407 // If the pointer destination is a cast, see if we can fold the cast into the
6409 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
6410 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
6412 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
6413 if (CE->getOpcode() == Instruction::Cast)
6414 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
6418 // If this store is the last instruction in the basic block, and if the block
6419 // ends with an unconditional branch, try to move it to the successor block.
6421 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
6422 if (BI->isUnconditional()) {
6423 // Check to see if the successor block has exactly two incoming edges. If
6424 // so, see if the other predecessor contains a store to the same location.
6425 // if so, insert a PHI node (if needed) and move the stores down.
6426 BasicBlock *Dest = BI->getSuccessor(0);
6428 pred_iterator PI = pred_begin(Dest);
6429 BasicBlock *Other = 0;
6430 if (*PI != BI->getParent())
6433 if (PI != pred_end(Dest)) {
6434 if (*PI != BI->getParent())
6439 if (++PI != pred_end(Dest))
6442 if (Other) { // If only one other pred...
6443 BBI = Other->getTerminator();
6444 // Make sure this other block ends in an unconditional branch and that
6445 // there is an instruction before the branch.
6446 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
6447 BBI != Other->begin()) {
6449 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
6451 // If this instruction is a store to the same location.
6452 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
6453 // Okay, we know we can perform this transformation. Insert a PHI
6454 // node now if we need it.
6455 Value *MergedVal = OtherStore->getOperand(0);
6456 if (MergedVal != SI.getOperand(0)) {
6457 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
6458 PN->reserveOperandSpace(2);
6459 PN->addIncoming(SI.getOperand(0), SI.getParent());
6460 PN->addIncoming(OtherStore->getOperand(0), Other);
6461 MergedVal = InsertNewInstBefore(PN, Dest->front());
6464 // Advance to a place where it is safe to insert the new store and
6466 BBI = Dest->begin();
6467 while (isa<PHINode>(BBI)) ++BBI;
6468 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
6469 OtherStore->isVolatile()), *BBI);
6471 // Nuke the old stores.
6472 EraseInstFromFunction(SI);
6473 EraseInstFromFunction(*OtherStore);
6485 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
6486 // Change br (not X), label True, label False to: br X, label False, True
6488 BasicBlock *TrueDest;
6489 BasicBlock *FalseDest;
6490 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
6491 !isa<Constant>(X)) {
6492 // Swap Destinations and condition...
6494 BI.setSuccessor(0, FalseDest);
6495 BI.setSuccessor(1, TrueDest);
6499 // Cannonicalize setne -> seteq
6500 Instruction::BinaryOps Op; Value *Y;
6501 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
6502 TrueDest, FalseDest)))
6503 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
6504 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
6505 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
6506 std::string Name = I->getName(); I->setName("");
6507 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
6508 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
6509 // Swap Destinations and condition...
6510 BI.setCondition(NewSCC);
6511 BI.setSuccessor(0, FalseDest);
6512 BI.setSuccessor(1, TrueDest);
6513 removeFromWorkList(I);
6514 I->getParent()->getInstList().erase(I);
6515 WorkList.push_back(cast<Instruction>(NewSCC));
6522 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
6523 Value *Cond = SI.getCondition();
6524 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
6525 if (I->getOpcode() == Instruction::Add)
6526 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6527 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
6528 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
6529 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
6531 SI.setOperand(0, I->getOperand(0));
6532 WorkList.push_back(I);
6539 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
6540 if (ConstantAggregateZero *C =
6541 dyn_cast<ConstantAggregateZero>(EI.getOperand(0))) {
6542 // If packed val is constant 0, replace extract with scalar 0
6543 const Type *Ty = cast<PackedType>(C->getType())->getElementType();
6544 EI.replaceAllUsesWith(Constant::getNullValue(Ty));
6545 return ReplaceInstUsesWith(EI, Constant::getNullValue(Ty));
6547 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
6548 // If packed val is constant with uniform operands, replace EI
6549 // with that operand
6550 Constant *op0 = cast<Constant>(C->getOperand(0));
6551 for (unsigned i = 1; i < C->getNumOperands(); ++i)
6552 if (C->getOperand(i) != op0) return 0;
6553 return ReplaceInstUsesWith(EI, op0);
6555 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0)))
6556 if (I->hasOneUse()) {
6557 // Push extractelement into predecessor operation if legal and
6558 // profitable to do so
6559 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
6560 if (!isa<Constant>(BO->getOperand(0)) &&
6561 !isa<Constant>(BO->getOperand(1)))
6563 ExtractElementInst *newEI0 =
6564 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
6566 ExtractElementInst *newEI1 =
6567 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
6569 InsertNewInstBefore(newEI0, EI);
6570 InsertNewInstBefore(newEI1, EI);
6571 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
6573 switch(I->getOpcode()) {
6574 case Instruction::Load: {
6575 Value *Ptr = InsertCastBefore(I->getOperand(0),
6576 PointerType::get(EI.getType()), EI);
6577 GetElementPtrInst *GEP =
6578 new GetElementPtrInst(Ptr, EI.getOperand(1),
6579 I->getName() + ".gep");
6580 InsertNewInstBefore(GEP, EI);
6581 return new LoadInst(GEP);
6591 void InstCombiner::removeFromWorkList(Instruction *I) {
6592 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
6597 /// TryToSinkInstruction - Try to move the specified instruction from its
6598 /// current block into the beginning of DestBlock, which can only happen if it's
6599 /// safe to move the instruction past all of the instructions between it and the
6600 /// end of its block.
6601 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
6602 assert(I->hasOneUse() && "Invariants didn't hold!");
6604 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
6605 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
6607 // Do not sink alloca instructions out of the entry block.
6608 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
6611 // We can only sink load instructions if there is nothing between the load and
6612 // the end of block that could change the value.
6613 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6614 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
6616 if (Scan->mayWriteToMemory())
6620 BasicBlock::iterator InsertPos = DestBlock->begin();
6621 while (isa<PHINode>(InsertPos)) ++InsertPos;
6623 I->moveBefore(InsertPos);
6628 bool InstCombiner::runOnFunction(Function &F) {
6629 bool Changed = false;
6630 TD = &getAnalysis<TargetData>();
6633 // Populate the worklist with the reachable instructions.
6634 std::set<BasicBlock*> Visited;
6635 for (df_ext_iterator<BasicBlock*> BB = df_ext_begin(&F.front(), Visited),
6636 E = df_ext_end(&F.front(), Visited); BB != E; ++BB)
6637 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
6638 WorkList.push_back(I);
6640 // Do a quick scan over the function. If we find any blocks that are
6641 // unreachable, remove any instructions inside of them. This prevents
6642 // the instcombine code from having to deal with some bad special cases.
6643 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
6644 if (!Visited.count(BB)) {
6645 Instruction *Term = BB->getTerminator();
6646 while (Term != BB->begin()) { // Remove instrs bottom-up
6647 BasicBlock::iterator I = Term; --I;
6649 DEBUG(std::cerr << "IC: DCE: " << *I);
6652 if (!I->use_empty())
6653 I->replaceAllUsesWith(UndefValue::get(I->getType()));
6654 I->eraseFromParent();
6659 while (!WorkList.empty()) {
6660 Instruction *I = WorkList.back(); // Get an instruction from the worklist
6661 WorkList.pop_back();
6663 // Check to see if we can DCE or ConstantPropagate the instruction...
6664 // Check to see if we can DIE the instruction...
6665 if (isInstructionTriviallyDead(I)) {
6666 // Add operands to the worklist...
6667 if (I->getNumOperands() < 4)
6668 AddUsesToWorkList(*I);
6671 DEBUG(std::cerr << "IC: DCE: " << *I);
6673 I->eraseFromParent();
6674 removeFromWorkList(I);
6678 // Instruction isn't dead, see if we can constant propagate it...
6679 if (Constant *C = ConstantFoldInstruction(I)) {
6680 Value* Ptr = I->getOperand(0);
6681 if (isa<GetElementPtrInst>(I) &&
6682 cast<Constant>(Ptr)->isNullValue() &&
6683 !isa<ConstantPointerNull>(C) &&
6684 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
6685 // If this is a constant expr gep that is effectively computing an
6686 // "offsetof", fold it into 'cast int X to T*' instead of 'gep 0, 0, 12'
6687 bool isFoldableGEP = true;
6688 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
6689 if (!isa<ConstantInt>(I->getOperand(i)))
6690 isFoldableGEP = false;
6691 if (isFoldableGEP) {
6692 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
6693 std::vector<Value*>(I->op_begin()+1, I->op_end()));
6694 C = ConstantUInt::get(Type::ULongTy, Offset);
6695 C = ConstantExpr::getCast(C, TD->getIntPtrType());
6696 C = ConstantExpr::getCast(C, I->getType());
6700 DEBUG(std::cerr << "IC: ConstFold to: " << *C << " from: " << *I);
6702 // Add operands to the worklist...
6703 AddUsesToWorkList(*I);
6704 ReplaceInstUsesWith(*I, C);
6707 I->getParent()->getInstList().erase(I);
6708 removeFromWorkList(I);
6712 // See if we can trivially sink this instruction to a successor basic block.
6713 if (I->hasOneUse()) {
6714 BasicBlock *BB = I->getParent();
6715 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
6716 if (UserParent != BB) {
6717 bool UserIsSuccessor = false;
6718 // See if the user is one of our successors.
6719 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
6720 if (*SI == UserParent) {
6721 UserIsSuccessor = true;
6725 // If the user is one of our immediate successors, and if that successor
6726 // only has us as a predecessors (we'd have to split the critical edge
6727 // otherwise), we can keep going.
6728 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
6729 next(pred_begin(UserParent)) == pred_end(UserParent))
6730 // Okay, the CFG is simple enough, try to sink this instruction.
6731 Changed |= TryToSinkInstruction(I, UserParent);
6735 // Now that we have an instruction, try combining it to simplify it...
6736 if (Instruction *Result = visit(*I)) {
6738 // Should we replace the old instruction with a new one?
6740 DEBUG(std::cerr << "IC: Old = " << *I
6741 << " New = " << *Result);
6743 // Everything uses the new instruction now.
6744 I->replaceAllUsesWith(Result);
6746 // Push the new instruction and any users onto the worklist.
6747 WorkList.push_back(Result);
6748 AddUsersToWorkList(*Result);
6750 // Move the name to the new instruction first...
6751 std::string OldName = I->getName(); I->setName("");
6752 Result->setName(OldName);
6754 // Insert the new instruction into the basic block...
6755 BasicBlock *InstParent = I->getParent();
6756 BasicBlock::iterator InsertPos = I;
6758 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
6759 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
6762 InstParent->getInstList().insert(InsertPos, Result);
6764 // Make sure that we reprocess all operands now that we reduced their
6766 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6767 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6768 WorkList.push_back(OpI);
6770 // Instructions can end up on the worklist more than once. Make sure
6771 // we do not process an instruction that has been deleted.
6772 removeFromWorkList(I);
6774 // Erase the old instruction.
6775 InstParent->getInstList().erase(I);
6777 DEBUG(std::cerr << "IC: MOD = " << *I);
6779 // If the instruction was modified, it's possible that it is now dead.
6780 // if so, remove it.
6781 if (isInstructionTriviallyDead(I)) {
6782 // Make sure we process all operands now that we are reducing their
6784 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
6785 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
6786 WorkList.push_back(OpI);
6788 // Instructions may end up in the worklist more than once. Erase all
6789 // occurrences of this instruction.
6790 removeFromWorkList(I);
6791 I->eraseFromParent();
6793 WorkList.push_back(Result);
6794 AddUsersToWorkList(*Result);
6804 FunctionPass *llvm::createInstructionCombiningPass() {
6805 return new InstCombiner();