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. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CallSite.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/PatternMatch.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/STLExtras.h"
56 using namespace llvm::PatternMatch;
58 STATISTIC(NumCombined , "Number of insts combined");
59 STATISTIC(NumConstProp, "Number of constant folds");
60 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
61 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
62 STATISTIC(NumSunkInst , "Number of instructions sunk");
65 class VISIBILITY_HIDDEN InstCombiner
66 : 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 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
92 /// dead. Add all of its operands to the worklist, turning them into
93 /// undef's to reduce the number of uses of those instructions.
95 /// Return the specified operand before it is turned into an undef.
97 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
98 Value *R = I.getOperand(op);
100 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
101 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
102 WorkList.push_back(Op);
103 // Set the operand to undef to drop the use.
104 I.setOperand(i, UndefValue::get(Op->getType()));
110 // removeFromWorkList - remove all instances of I from the worklist.
111 void removeFromWorkList(Instruction *I);
113 virtual bool runOnFunction(Function &F);
115 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
116 AU.addRequired<TargetData>();
117 AU.addPreservedID(LCSSAID);
118 AU.setPreservesCFG();
121 TargetData &getTargetData() const { return *TD; }
123 // Visitation implementation - Implement instruction combining for different
124 // instruction types. The semantics are as follows:
126 // null - No change was made
127 // I - Change was made, I is still valid, I may be dead though
128 // otherwise - Change was made, replace I with returned instruction
130 Instruction *visitAdd(BinaryOperator &I);
131 Instruction *visitSub(BinaryOperator &I);
132 Instruction *visitMul(BinaryOperator &I);
133 Instruction *visitURem(BinaryOperator &I);
134 Instruction *visitSRem(BinaryOperator &I);
135 Instruction *visitFRem(BinaryOperator &I);
136 Instruction *commonRemTransforms(BinaryOperator &I);
137 Instruction *commonIRemTransforms(BinaryOperator &I);
138 Instruction *commonDivTransforms(BinaryOperator &I);
139 Instruction *commonIDivTransforms(BinaryOperator &I);
140 Instruction *visitUDiv(BinaryOperator &I);
141 Instruction *visitSDiv(BinaryOperator &I);
142 Instruction *visitFDiv(BinaryOperator &I);
143 Instruction *visitAnd(BinaryOperator &I);
144 Instruction *visitOr (BinaryOperator &I);
145 Instruction *visitXor(BinaryOperator &I);
146 Instruction *visitFCmpInst(FCmpInst &I);
147 Instruction *visitICmpInst(ICmpInst &I);
148 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
150 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
151 ICmpInst::Predicate Cond, Instruction &I);
152 Instruction *visitShiftInst(ShiftInst &I);
153 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
155 Instruction *commonCastTransforms(CastInst &CI);
156 Instruction *commonIntCastTransforms(CastInst &CI);
157 Instruction *visitTrunc(CastInst &CI);
158 Instruction *visitZExt(CastInst &CI);
159 Instruction *visitSExt(CastInst &CI);
160 Instruction *visitFPTrunc(CastInst &CI);
161 Instruction *visitFPExt(CastInst &CI);
162 Instruction *visitFPToUI(CastInst &CI);
163 Instruction *visitFPToSI(CastInst &CI);
164 Instruction *visitUIToFP(CastInst &CI);
165 Instruction *visitSIToFP(CastInst &CI);
166 Instruction *visitPtrToInt(CastInst &CI);
167 Instruction *visitIntToPtr(CastInst &CI);
168 Instruction *visitBitCast(CastInst &CI);
169 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
171 Instruction *visitSelectInst(SelectInst &CI);
172 Instruction *visitCallInst(CallInst &CI);
173 Instruction *visitInvokeInst(InvokeInst &II);
174 Instruction *visitPHINode(PHINode &PN);
175 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
176 Instruction *visitAllocationInst(AllocationInst &AI);
177 Instruction *visitFreeInst(FreeInst &FI);
178 Instruction *visitLoadInst(LoadInst &LI);
179 Instruction *visitStoreInst(StoreInst &SI);
180 Instruction *visitBranchInst(BranchInst &BI);
181 Instruction *visitSwitchInst(SwitchInst &SI);
182 Instruction *visitInsertElementInst(InsertElementInst &IE);
183 Instruction *visitExtractElementInst(ExtractElementInst &EI);
184 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
186 // visitInstruction - Specify what to return for unhandled instructions...
187 Instruction *visitInstruction(Instruction &I) { return 0; }
190 Instruction *visitCallSite(CallSite CS);
191 bool transformConstExprCastCall(CallSite CS);
194 // InsertNewInstBefore - insert an instruction New before instruction Old
195 // in the program. Add the new instruction to the worklist.
197 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
198 assert(New && New->getParent() == 0 &&
199 "New instruction already inserted into a basic block!");
200 BasicBlock *BB = Old.getParent();
201 BB->getInstList().insert(&Old, New); // Insert inst
202 WorkList.push_back(New); // Add to worklist
206 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
207 /// This also adds the cast to the worklist. Finally, this returns the
209 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
211 if (V->getType() == Ty) return V;
213 if (Constant *CV = dyn_cast<Constant>(V))
214 return ConstantExpr::getCast(opc, CV, Ty);
216 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
217 WorkList.push_back(C);
221 // ReplaceInstUsesWith - This method is to be used when an instruction is
222 // found to be dead, replacable with another preexisting expression. Here
223 // we add all uses of I to the worklist, replace all uses of I with the new
224 // value, then return I, so that the inst combiner will know that I was
227 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
228 AddUsersToWorkList(I); // Add all modified instrs to worklist
230 I.replaceAllUsesWith(V);
233 // If we are replacing the instruction with itself, this must be in a
234 // segment of unreachable code, so just clobber the instruction.
235 I.replaceAllUsesWith(UndefValue::get(I.getType()));
240 // UpdateValueUsesWith - This method is to be used when an value is
241 // found to be replacable with another preexisting expression or was
242 // updated. Here we add all uses of I to the worklist, replace all uses of
243 // I with the new value (unless the instruction was just updated), then
244 // return true, so that the inst combiner will know that I was modified.
246 bool UpdateValueUsesWith(Value *Old, Value *New) {
247 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
249 Old->replaceAllUsesWith(New);
250 if (Instruction *I = dyn_cast<Instruction>(Old))
251 WorkList.push_back(I);
252 if (Instruction *I = dyn_cast<Instruction>(New))
253 WorkList.push_back(I);
257 // EraseInstFromFunction - When dealing with an instruction that has side
258 // effects or produces a void value, we can't rely on DCE to delete the
259 // instruction. Instead, visit methods should return the value returned by
261 Instruction *EraseInstFromFunction(Instruction &I) {
262 assert(I.use_empty() && "Cannot erase instruction that is used!");
263 AddUsesToWorkList(I);
264 removeFromWorkList(&I);
266 return 0; // Don't do anything with FI
270 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
271 /// InsertBefore instruction. This is specialized a bit to avoid inserting
272 /// casts that are known to not do anything...
274 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
275 Value *V, const Type *DestTy,
276 Instruction *InsertBefore);
278 /// SimplifyCommutative - This performs a few simplifications for
279 /// commutative operators.
280 bool SimplifyCommutative(BinaryOperator &I);
282 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
283 /// most-complex to least-complex order.
284 bool SimplifyCompare(CmpInst &I);
286 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
287 uint64_t &KnownZero, uint64_t &KnownOne,
290 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
291 uint64_t &UndefElts, unsigned Depth = 0);
293 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
294 // PHI node as operand #0, see if we can fold the instruction into the PHI
295 // (which is only possible if all operands to the PHI are constants).
296 Instruction *FoldOpIntoPhi(Instruction &I);
298 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
299 // operator and they all are only used by the PHI, PHI together their
300 // inputs, and do the operation once, to the result of the PHI.
301 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
302 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
305 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
306 ConstantInt *AndRHS, BinaryOperator &TheAnd);
308 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
309 bool isSub, Instruction &I);
310 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
311 bool isSigned, bool Inside, Instruction &IB);
312 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
313 Instruction *MatchBSwap(BinaryOperator &I);
315 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
318 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
321 // getComplexity: Assign a complexity or rank value to LLVM Values...
322 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
323 static unsigned getComplexity(Value *V) {
324 if (isa<Instruction>(V)) {
325 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
329 if (isa<Argument>(V)) return 3;
330 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
333 // isOnlyUse - Return true if this instruction will be deleted if we stop using
335 static bool isOnlyUse(Value *V) {
336 return V->hasOneUse() || isa<Constant>(V);
339 // getPromotedType - Return the specified type promoted as it would be to pass
340 // though a va_arg area...
341 static const Type *getPromotedType(const Type *Ty) {
342 switch (Ty->getTypeID()) {
344 case Type::Int16TyID: return Type::Int32Ty;
345 case Type::FloatTyID: return Type::DoubleTy;
350 /// getBitCastOperand - If the specified operand is a CastInst or a constant
351 /// expression bitcast, return the operand value, otherwise return null.
352 static Value *getBitCastOperand(Value *V) {
353 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
354 return I->getOperand(0);
355 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
356 if (CE->getOpcode() == Instruction::BitCast)
357 return CE->getOperand(0);
361 /// This function is a wrapper around CastInst::isEliminableCastPair. It
362 /// simply extracts arguments and returns what that function returns.
363 /// @Determine if it is valid to eliminate a Convert pair
364 static Instruction::CastOps
365 isEliminableCastPair(
366 const CastInst *CI, ///< The first cast instruction
367 unsigned opcode, ///< The opcode of the second cast instruction
368 const Type *DstTy, ///< The target type for the second cast instruction
369 TargetData *TD ///< The target data for pointer size
372 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
373 const Type *MidTy = CI->getType(); // B from above
375 // Get the opcodes of the two Cast instructions
376 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
377 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
379 return Instruction::CastOps(
380 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
381 DstTy, TD->getIntPtrType()));
384 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
385 /// in any code being generated. It does not require codegen if V is simple
386 /// enough or if the cast can be folded into other casts.
387 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
388 const Type *Ty, TargetData *TD) {
389 if (V->getType() == Ty || isa<Constant>(V)) return false;
391 // If this is another cast that can be eliminated, it isn't codegen either.
392 if (const CastInst *CI = dyn_cast<CastInst>(V))
393 if (isEliminableCastPair(CI, opcode, Ty, TD))
398 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
399 /// InsertBefore instruction. This is specialized a bit to avoid inserting
400 /// casts that are known to not do anything...
402 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
403 Value *V, const Type *DestTy,
404 Instruction *InsertBefore) {
405 if (V->getType() == DestTy) return V;
406 if (Constant *C = dyn_cast<Constant>(V))
407 return ConstantExpr::getCast(opcode, C, DestTy);
409 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
412 // SimplifyCommutative - This performs a few simplifications for commutative
415 // 1. Order operands such that they are listed from right (least complex) to
416 // left (most complex). This puts constants before unary operators before
419 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
420 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
422 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
423 bool Changed = false;
424 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
425 Changed = !I.swapOperands();
427 if (!I.isAssociative()) return Changed;
428 Instruction::BinaryOps Opcode = I.getOpcode();
429 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
430 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
431 if (isa<Constant>(I.getOperand(1))) {
432 Constant *Folded = ConstantExpr::get(I.getOpcode(),
433 cast<Constant>(I.getOperand(1)),
434 cast<Constant>(Op->getOperand(1)));
435 I.setOperand(0, Op->getOperand(0));
436 I.setOperand(1, Folded);
438 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
439 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
440 isOnlyUse(Op) && isOnlyUse(Op1)) {
441 Constant *C1 = cast<Constant>(Op->getOperand(1));
442 Constant *C2 = cast<Constant>(Op1->getOperand(1));
444 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
445 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
446 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
449 WorkList.push_back(New);
450 I.setOperand(0, New);
451 I.setOperand(1, Folded);
458 /// SimplifyCompare - For a CmpInst this function just orders the operands
459 /// so that theyare listed from right (least complex) to left (most complex).
460 /// This puts constants before unary operators before binary operators.
461 bool InstCombiner::SimplifyCompare(CmpInst &I) {
462 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
465 // Compare instructions are not associative so there's nothing else we can do.
469 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
470 // if the LHS is a constant zero (which is the 'negate' form).
472 static inline Value *dyn_castNegVal(Value *V) {
473 if (BinaryOperator::isNeg(V))
474 return BinaryOperator::getNegArgument(V);
476 // Constants can be considered to be negated values if they can be folded.
477 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
478 return ConstantExpr::getNeg(C);
482 static inline Value *dyn_castNotVal(Value *V) {
483 if (BinaryOperator::isNot(V))
484 return BinaryOperator::getNotArgument(V);
486 // Constants can be considered to be not'ed values...
487 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
488 return ConstantExpr::getNot(C);
492 // dyn_castFoldableMul - If this value is a multiply that can be folded into
493 // other computations (because it has a constant operand), return the
494 // non-constant operand of the multiply, and set CST to point to the multiplier.
495 // Otherwise, return null.
497 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
498 if (V->hasOneUse() && V->getType()->isInteger())
499 if (Instruction *I = dyn_cast<Instruction>(V)) {
500 if (I->getOpcode() == Instruction::Mul)
501 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
502 return I->getOperand(0);
503 if (I->getOpcode() == Instruction::Shl)
504 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
505 // The multiplier is really 1 << CST.
506 Constant *One = ConstantInt::get(V->getType(), 1);
507 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
508 return I->getOperand(0);
514 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
515 /// expression, return it.
516 static User *dyn_castGetElementPtr(Value *V) {
517 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
518 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
519 if (CE->getOpcode() == Instruction::GetElementPtr)
520 return cast<User>(V);
524 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
525 static ConstantInt *AddOne(ConstantInt *C) {
526 return cast<ConstantInt>(ConstantExpr::getAdd(C,
527 ConstantInt::get(C->getType(), 1)));
529 static ConstantInt *SubOne(ConstantInt *C) {
530 return cast<ConstantInt>(ConstantExpr::getSub(C,
531 ConstantInt::get(C->getType(), 1)));
535 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
536 /// known to be either zero or one and return them in the KnownZero/KnownOne
537 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
539 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
540 uint64_t &KnownOne, unsigned Depth = 0) {
541 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
542 // we cannot optimize based on the assumption that it is zero without changing
543 // it to be an explicit zero. If we don't change it to zero, other code could
544 // optimized based on the contradictory assumption that it is non-zero.
545 // Because instcombine aggressively folds operations with undef args anyway,
546 // this won't lose us code quality.
547 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
548 // We know all of the bits for a constant!
549 KnownOne = CI->getZExtValue() & Mask;
550 KnownZero = ~KnownOne & Mask;
554 KnownZero = KnownOne = 0; // Don't know anything.
555 if (Depth == 6 || Mask == 0)
556 return; // Limit search depth.
558 uint64_t KnownZero2, KnownOne2;
559 Instruction *I = dyn_cast<Instruction>(V);
562 Mask &= V->getType()->getIntegralTypeMask();
564 switch (I->getOpcode()) {
565 case Instruction::And:
566 // If either the LHS or the RHS are Zero, the result is zero.
567 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
569 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
570 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
571 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
573 // Output known-1 bits are only known if set in both the LHS & RHS.
574 KnownOne &= KnownOne2;
575 // Output known-0 are known to be clear if zero in either the LHS | RHS.
576 KnownZero |= KnownZero2;
578 case Instruction::Or:
579 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
581 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
582 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
583 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
585 // Output known-0 bits are only known if clear in both the LHS & RHS.
586 KnownZero &= KnownZero2;
587 // Output known-1 are known to be set if set in either the LHS | RHS.
588 KnownOne |= KnownOne2;
590 case Instruction::Xor: {
591 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
592 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
593 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
594 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
596 // Output known-0 bits are known if clear or set in both the LHS & RHS.
597 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
598 // Output known-1 are known to be set if set in only one of the LHS, RHS.
599 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
600 KnownZero = KnownZeroOut;
603 case Instruction::Select:
604 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
605 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
606 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
607 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
609 // Only known if known in both the LHS and RHS.
610 KnownOne &= KnownOne2;
611 KnownZero &= KnownZero2;
613 case Instruction::FPTrunc:
614 case Instruction::FPExt:
615 case Instruction::FPToUI:
616 case Instruction::FPToSI:
617 case Instruction::SIToFP:
618 case Instruction::PtrToInt:
619 case Instruction::UIToFP:
620 case Instruction::IntToPtr:
621 return; // Can't work with floating point or pointers
622 case Instruction::Trunc:
623 // All these have integer operands
624 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
626 case Instruction::BitCast: {
627 const Type *SrcTy = I->getOperand(0)->getType();
628 if (SrcTy->isIntegral()) {
629 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
634 case Instruction::ZExt: {
635 // Compute the bits in the result that are not present in the input.
636 const Type *SrcTy = I->getOperand(0)->getType();
637 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
638 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
640 Mask &= SrcTy->getIntegralTypeMask();
641 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
642 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
643 // The top bits are known to be zero.
644 KnownZero |= NewBits;
647 case Instruction::SExt: {
648 // Compute the bits in the result that are not present in the input.
649 const Type *SrcTy = I->getOperand(0)->getType();
650 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
651 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
653 Mask &= SrcTy->getIntegralTypeMask();
654 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
655 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
657 // If the sign bit of the input is known set or clear, then we know the
658 // top bits of the result.
659 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
660 if (KnownZero & InSignBit) { // Input sign bit known zero
661 KnownZero |= NewBits;
662 KnownOne &= ~NewBits;
663 } else if (KnownOne & InSignBit) { // Input sign bit known set
665 KnownZero &= ~NewBits;
666 } else { // Input sign bit unknown
667 KnownZero &= ~NewBits;
668 KnownOne &= ~NewBits;
672 case Instruction::Shl:
673 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
674 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
675 uint64_t ShiftAmt = SA->getZExtValue();
677 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
678 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
679 KnownZero <<= ShiftAmt;
680 KnownOne <<= ShiftAmt;
681 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
685 case Instruction::LShr:
686 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
687 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
688 // Compute the new bits that are at the top now.
689 uint64_t ShiftAmt = SA->getZExtValue();
690 uint64_t HighBits = (1ULL << ShiftAmt)-1;
691 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
693 // Unsigned shift right.
695 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
696 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
697 KnownZero >>= ShiftAmt;
698 KnownOne >>= ShiftAmt;
699 KnownZero |= HighBits; // high bits known zero.
703 case Instruction::AShr:
704 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
705 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
706 // Compute the new bits that are at the top now.
707 uint64_t ShiftAmt = SA->getZExtValue();
708 uint64_t HighBits = (1ULL << ShiftAmt)-1;
709 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
711 // Signed shift right.
713 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
714 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
715 KnownZero >>= ShiftAmt;
716 KnownOne >>= ShiftAmt;
718 // Handle the sign bits.
719 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
720 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
722 if (KnownZero & SignBit) { // New bits are known zero.
723 KnownZero |= HighBits;
724 } else if (KnownOne & SignBit) { // New bits are known one.
725 KnownOne |= HighBits;
733 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
734 /// this predicate to simplify operations downstream. Mask is known to be zero
735 /// for bits that V cannot have.
736 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
737 uint64_t KnownZero, KnownOne;
738 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
739 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
740 return (KnownZero & Mask) == Mask;
743 /// ShrinkDemandedConstant - Check to see if the specified operand of the
744 /// specified instruction is a constant integer. If so, check to see if there
745 /// are any bits set in the constant that are not demanded. If so, shrink the
746 /// constant and return true.
747 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
749 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
750 if (!OpC) return false;
752 // If there are no bits set that aren't demanded, nothing to do.
753 if ((~Demanded & OpC->getZExtValue()) == 0)
756 // This is producing any bits that are not needed, shrink the RHS.
757 uint64_t Val = Demanded & OpC->getZExtValue();
758 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
762 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
763 // set of known zero and one bits, compute the maximum and minimum values that
764 // could have the specified known zero and known one bits, returning them in
766 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
769 int64_t &Min, int64_t &Max) {
770 uint64_t TypeBits = Ty->getIntegralTypeMask();
771 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
773 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
775 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
776 // bit if it is unknown.
778 Max = KnownOne|UnknownBits;
780 if (SignBit & UnknownBits) { // Sign bit is unknown
785 // Sign extend the min/max values.
786 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
787 Min = (Min << ShAmt) >> ShAmt;
788 Max = (Max << ShAmt) >> ShAmt;
791 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
792 // a set of known zero and one bits, compute the maximum and minimum values that
793 // could have the specified known zero and known one bits, returning them in
795 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
800 uint64_t TypeBits = Ty->getIntegralTypeMask();
801 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
803 // The minimum value is when the unknown bits are all zeros.
805 // The maximum value is when the unknown bits are all ones.
806 Max = KnownOne|UnknownBits;
810 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
811 /// DemandedMask bits of the result of V are ever used downstream. If we can
812 /// use this information to simplify V, do so and return true. Otherwise,
813 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
814 /// the expression (used to simplify the caller). The KnownZero/One bits may
815 /// only be accurate for those bits in the DemandedMask.
816 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
817 uint64_t &KnownZero, uint64_t &KnownOne,
819 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
820 // We know all of the bits for a constant!
821 KnownOne = CI->getZExtValue() & DemandedMask;
822 KnownZero = ~KnownOne & DemandedMask;
826 KnownZero = KnownOne = 0;
827 if (!V->hasOneUse()) { // Other users may use these bits.
828 if (Depth != 0) { // Not at the root.
829 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
830 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
833 // If this is the root being simplified, allow it to have multiple uses,
834 // just set the DemandedMask to all bits.
835 DemandedMask = V->getType()->getIntegralTypeMask();
836 } else if (DemandedMask == 0) { // Not demanding any bits from V.
837 if (V != UndefValue::get(V->getType()))
838 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
840 } else if (Depth == 6) { // Limit search depth.
844 Instruction *I = dyn_cast<Instruction>(V);
845 if (!I) return false; // Only analyze instructions.
847 DemandedMask &= V->getType()->getIntegralTypeMask();
849 uint64_t KnownZero2 = 0, KnownOne2 = 0;
850 switch (I->getOpcode()) {
852 case Instruction::And:
853 // If either the LHS or the RHS are Zero, the result is zero.
854 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
855 KnownZero, KnownOne, Depth+1))
857 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
859 // If something is known zero on the RHS, the bits aren't demanded on the
861 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
862 KnownZero2, KnownOne2, Depth+1))
864 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
866 // If all of the demanded bits are known 1 on one side, return the other.
867 // These bits cannot contribute to the result of the 'and'.
868 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
869 return UpdateValueUsesWith(I, I->getOperand(0));
870 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
871 return UpdateValueUsesWith(I, I->getOperand(1));
873 // If all of the demanded bits in the inputs are known zeros, return zero.
874 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
875 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
877 // If the RHS is a constant, see if we can simplify it.
878 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
879 return UpdateValueUsesWith(I, I);
881 // Output known-1 bits are only known if set in both the LHS & RHS.
882 KnownOne &= KnownOne2;
883 // Output known-0 are known to be clear if zero in either the LHS | RHS.
884 KnownZero |= KnownZero2;
886 case Instruction::Or:
887 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
888 KnownZero, KnownOne, Depth+1))
890 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
891 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
892 KnownZero2, KnownOne2, Depth+1))
894 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
896 // If all of the demanded bits are known zero on one side, return the other.
897 // These bits cannot contribute to the result of the 'or'.
898 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
899 return UpdateValueUsesWith(I, I->getOperand(0));
900 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
901 return UpdateValueUsesWith(I, I->getOperand(1));
903 // If all of the potentially set bits on one side are known to be set on
904 // the other side, just use the 'other' side.
905 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
906 (DemandedMask & (~KnownZero)))
907 return UpdateValueUsesWith(I, I->getOperand(0));
908 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
909 (DemandedMask & (~KnownZero2)))
910 return UpdateValueUsesWith(I, I->getOperand(1));
912 // If the RHS is a constant, see if we can simplify it.
913 if (ShrinkDemandedConstant(I, 1, DemandedMask))
914 return UpdateValueUsesWith(I, I);
916 // Output known-0 bits are only known if clear in both the LHS & RHS.
917 KnownZero &= KnownZero2;
918 // Output known-1 are known to be set if set in either the LHS | RHS.
919 KnownOne |= KnownOne2;
921 case Instruction::Xor: {
922 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
923 KnownZero, KnownOne, Depth+1))
925 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
926 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
927 KnownZero2, KnownOne2, Depth+1))
929 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
931 // If all of the demanded bits are known zero on one side, return the other.
932 // These bits cannot contribute to the result of the 'xor'.
933 if ((DemandedMask & KnownZero) == DemandedMask)
934 return UpdateValueUsesWith(I, I->getOperand(0));
935 if ((DemandedMask & KnownZero2) == DemandedMask)
936 return UpdateValueUsesWith(I, I->getOperand(1));
938 // Output known-0 bits are known if clear or set in both the LHS & RHS.
939 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
940 // Output known-1 are known to be set if set in only one of the LHS, RHS.
941 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
943 // If all of the demanded bits are known to be zero on one side or the
944 // other, turn this into an *inclusive* or.
945 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
946 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
948 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
950 InsertNewInstBefore(Or, *I);
951 return UpdateValueUsesWith(I, Or);
954 // If all of the demanded bits on one side are known, and all of the set
955 // bits on that side are also known to be set on the other side, turn this
956 // into an AND, as we know the bits will be cleared.
957 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
958 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
959 if ((KnownOne & KnownOne2) == KnownOne) {
960 Constant *AndC = ConstantInt::get(I->getType(),
961 ~KnownOne & DemandedMask);
963 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
964 InsertNewInstBefore(And, *I);
965 return UpdateValueUsesWith(I, And);
969 // If the RHS is a constant, see if we can simplify it.
970 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
971 if (ShrinkDemandedConstant(I, 1, DemandedMask))
972 return UpdateValueUsesWith(I, I);
974 KnownZero = KnownZeroOut;
975 KnownOne = KnownOneOut;
978 case Instruction::Select:
979 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
980 KnownZero, KnownOne, Depth+1))
982 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
983 KnownZero2, KnownOne2, Depth+1))
985 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
986 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
988 // If the operands are constants, see if we can simplify them.
989 if (ShrinkDemandedConstant(I, 1, DemandedMask))
990 return UpdateValueUsesWith(I, I);
991 if (ShrinkDemandedConstant(I, 2, DemandedMask))
992 return UpdateValueUsesWith(I, I);
994 // Only known if known in both the LHS and RHS.
995 KnownOne &= KnownOne2;
996 KnownZero &= KnownZero2;
998 case Instruction::Trunc:
999 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1000 KnownZero, KnownOne, Depth+1))
1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1004 case Instruction::BitCast:
1005 if (!I->getOperand(0)->getType()->isIntegral())
1008 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1009 KnownZero, KnownOne, Depth+1))
1011 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1013 case Instruction::ZExt: {
1014 // Compute the bits in the result that are not present in the input.
1015 const Type *SrcTy = I->getOperand(0)->getType();
1016 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1017 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1019 DemandedMask &= SrcTy->getIntegralTypeMask();
1020 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1021 KnownZero, KnownOne, Depth+1))
1023 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1024 // The top bits are known to be zero.
1025 KnownZero |= NewBits;
1028 case Instruction::SExt: {
1029 // Compute the bits in the result that are not present in the input.
1030 const Type *SrcTy = I->getOperand(0)->getType();
1031 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1032 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1034 // Get the sign bit for the source type
1035 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1036 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1038 // If any of the sign extended bits are demanded, we know that the sign
1040 if (NewBits & DemandedMask)
1041 InputDemandedBits |= InSignBit;
1043 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1044 KnownZero, KnownOne, Depth+1))
1046 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1048 // If the sign bit of the input is known set or clear, then we know the
1049 // top bits of the result.
1051 // If the input sign bit is known zero, or if the NewBits are not demanded
1052 // convert this into a zero extension.
1053 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1054 // Convert to ZExt cast
1055 CastInst *NewCast = CastInst::create(
1056 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1057 return UpdateValueUsesWith(I, NewCast);
1058 } else if (KnownOne & InSignBit) { // Input sign bit known set
1059 KnownOne |= NewBits;
1060 KnownZero &= ~NewBits;
1061 } else { // Input sign bit unknown
1062 KnownZero &= ~NewBits;
1063 KnownOne &= ~NewBits;
1067 case Instruction::Add:
1068 // If there is a constant on the RHS, there are a variety of xformations
1070 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1071 // If null, this should be simplified elsewhere. Some of the xforms here
1072 // won't work if the RHS is zero.
1073 if (RHS->isNullValue())
1076 // Figure out what the input bits are. If the top bits of the and result
1077 // are not demanded, then the add doesn't demand them from its input
1080 // Shift the demanded mask up so that it's at the top of the uint64_t.
1081 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1082 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1084 // If the top bit of the output is demanded, demand everything from the
1085 // input. Otherwise, we demand all the input bits except NLZ top bits.
1086 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1088 // Find information about known zero/one bits in the input.
1089 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1090 KnownZero2, KnownOne2, Depth+1))
1093 // If the RHS of the add has bits set that can't affect the input, reduce
1095 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1096 return UpdateValueUsesWith(I, I);
1098 // Avoid excess work.
1099 if (KnownZero2 == 0 && KnownOne2 == 0)
1102 // Turn it into OR if input bits are zero.
1103 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1105 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1107 InsertNewInstBefore(Or, *I);
1108 return UpdateValueUsesWith(I, Or);
1111 // We can say something about the output known-zero and known-one bits,
1112 // depending on potential carries from the input constant and the
1113 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1114 // bits set and the RHS constant is 0x01001, then we know we have a known
1115 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1117 // To compute this, we first compute the potential carry bits. These are
1118 // the bits which may be modified. I'm not aware of a better way to do
1120 uint64_t RHSVal = RHS->getZExtValue();
1122 bool CarryIn = false;
1123 uint64_t CarryBits = 0;
1124 uint64_t CurBit = 1;
1125 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1126 // Record the current carry in.
1127 if (CarryIn) CarryBits |= CurBit;
1131 // This bit has a carry out unless it is "zero + zero" or
1132 // "zero + anything" with no carry in.
1133 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1134 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1135 } else if (!CarryIn &&
1136 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1137 CarryOut = false; // 0 + anything has no carry out if no carry in.
1139 // Otherwise, we have to assume we have a carry out.
1143 // This stage's carry out becomes the next stage's carry-in.
1147 // Now that we know which bits have carries, compute the known-1/0 sets.
1149 // Bits are known one if they are known zero in one operand and one in the
1150 // other, and there is no input carry.
1151 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1153 // Bits are known zero if they are known zero in both operands and there
1154 // is no input carry.
1155 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1158 case Instruction::Shl:
1159 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1160 uint64_t ShiftAmt = SA->getZExtValue();
1161 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1162 KnownZero, KnownOne, Depth+1))
1164 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1165 KnownZero <<= ShiftAmt;
1166 KnownOne <<= ShiftAmt;
1167 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1170 case Instruction::LShr:
1171 // For a logical shift right
1172 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1173 unsigned ShiftAmt = SA->getZExtValue();
1175 // Compute the new bits that are at the top now.
1176 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1177 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1178 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1179 // Unsigned shift right.
1180 if (SimplifyDemandedBits(I->getOperand(0),
1181 (DemandedMask << ShiftAmt) & TypeMask,
1182 KnownZero, KnownOne, Depth+1))
1184 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1185 KnownZero &= TypeMask;
1186 KnownOne &= TypeMask;
1187 KnownZero >>= ShiftAmt;
1188 KnownOne >>= ShiftAmt;
1189 KnownZero |= HighBits; // high bits known zero.
1192 case Instruction::AShr:
1193 // If this is an arithmetic shift right and only the low-bit is set, we can
1194 // always convert this into a logical shr, even if the shift amount is
1195 // variable. The low bit of the shift cannot be an input sign bit unless
1196 // the shift amount is >= the size of the datatype, which is undefined.
1197 if (DemandedMask == 1) {
1198 // Perform the logical shift right.
1199 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1200 I->getOperand(1), I->getName());
1201 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1202 return UpdateValueUsesWith(I, NewVal);
1205 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1206 unsigned ShiftAmt = SA->getZExtValue();
1208 // Compute the new bits that are at the top now.
1209 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1210 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1211 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1212 // Signed shift right.
1213 if (SimplifyDemandedBits(I->getOperand(0),
1214 (DemandedMask << ShiftAmt) & TypeMask,
1215 KnownZero, KnownOne, Depth+1))
1217 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1218 KnownZero &= TypeMask;
1219 KnownOne &= TypeMask;
1220 KnownZero >>= ShiftAmt;
1221 KnownOne >>= ShiftAmt;
1223 // Handle the sign bits.
1224 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1225 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1227 // If the input sign bit is known to be zero, or if none of the top bits
1228 // are demanded, turn this into an unsigned shift right.
1229 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1230 // Perform the logical shift right.
1231 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1233 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1234 return UpdateValueUsesWith(I, NewVal);
1235 } else if (KnownOne & SignBit) { // New bits are known one.
1236 KnownOne |= HighBits;
1242 // If the client is only demanding bits that we know, return the known
1244 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1245 return UpdateValueUsesWith(I, ConstantInt::get(I->getType(), KnownOne));
1250 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1251 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1252 /// actually used by the caller. This method analyzes which elements of the
1253 /// operand are undef and returns that information in UndefElts.
1255 /// If the information about demanded elements can be used to simplify the
1256 /// operation, the operation is simplified, then the resultant value is
1257 /// returned. This returns null if no change was made.
1258 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1259 uint64_t &UndefElts,
1261 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1262 assert(VWidth <= 64 && "Vector too wide to analyze!");
1263 uint64_t EltMask = ~0ULL >> (64-VWidth);
1264 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1265 "Invalid DemandedElts!");
1267 if (isa<UndefValue>(V)) {
1268 // If the entire vector is undefined, just return this info.
1269 UndefElts = EltMask;
1271 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1272 UndefElts = EltMask;
1273 return UndefValue::get(V->getType());
1277 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1278 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1279 Constant *Undef = UndefValue::get(EltTy);
1281 std::vector<Constant*> Elts;
1282 for (unsigned i = 0; i != VWidth; ++i)
1283 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1284 Elts.push_back(Undef);
1285 UndefElts |= (1ULL << i);
1286 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1287 Elts.push_back(Undef);
1288 UndefElts |= (1ULL << i);
1289 } else { // Otherwise, defined.
1290 Elts.push_back(CP->getOperand(i));
1293 // If we changed the constant, return it.
1294 Constant *NewCP = ConstantPacked::get(Elts);
1295 return NewCP != CP ? NewCP : 0;
1296 } else if (isa<ConstantAggregateZero>(V)) {
1297 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1299 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1300 Constant *Zero = Constant::getNullValue(EltTy);
1301 Constant *Undef = UndefValue::get(EltTy);
1302 std::vector<Constant*> Elts;
1303 for (unsigned i = 0; i != VWidth; ++i)
1304 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1305 UndefElts = DemandedElts ^ EltMask;
1306 return ConstantPacked::get(Elts);
1309 if (!V->hasOneUse()) { // Other users may use these bits.
1310 if (Depth != 0) { // Not at the root.
1311 // TODO: Just compute the UndefElts information recursively.
1315 } else if (Depth == 10) { // Limit search depth.
1319 Instruction *I = dyn_cast<Instruction>(V);
1320 if (!I) return false; // Only analyze instructions.
1322 bool MadeChange = false;
1323 uint64_t UndefElts2;
1325 switch (I->getOpcode()) {
1328 case Instruction::InsertElement: {
1329 // If this is a variable index, we don't know which element it overwrites.
1330 // demand exactly the same input as we produce.
1331 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1333 // Note that we can't propagate undef elt info, because we don't know
1334 // which elt is getting updated.
1335 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1336 UndefElts2, Depth+1);
1337 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1341 // If this is inserting an element that isn't demanded, remove this
1343 unsigned IdxNo = Idx->getZExtValue();
1344 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1345 return AddSoonDeadInstToWorklist(*I, 0);
1347 // Otherwise, the element inserted overwrites whatever was there, so the
1348 // input demanded set is simpler than the output set.
1349 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1350 DemandedElts & ~(1ULL << IdxNo),
1351 UndefElts, Depth+1);
1352 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1354 // The inserted element is defined.
1355 UndefElts |= 1ULL << IdxNo;
1359 case Instruction::And:
1360 case Instruction::Or:
1361 case Instruction::Xor:
1362 case Instruction::Add:
1363 case Instruction::Sub:
1364 case Instruction::Mul:
1365 // div/rem demand all inputs, because they don't want divide by zero.
1366 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1367 UndefElts, Depth+1);
1368 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1369 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1370 UndefElts2, Depth+1);
1371 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1373 // Output elements are undefined if both are undefined. Consider things
1374 // like undef&0. The result is known zero, not undef.
1375 UndefElts &= UndefElts2;
1378 case Instruction::Call: {
1379 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1381 switch (II->getIntrinsicID()) {
1384 // Binary vector operations that work column-wise. A dest element is a
1385 // function of the corresponding input elements from the two inputs.
1386 case Intrinsic::x86_sse_sub_ss:
1387 case Intrinsic::x86_sse_mul_ss:
1388 case Intrinsic::x86_sse_min_ss:
1389 case Intrinsic::x86_sse_max_ss:
1390 case Intrinsic::x86_sse2_sub_sd:
1391 case Intrinsic::x86_sse2_mul_sd:
1392 case Intrinsic::x86_sse2_min_sd:
1393 case Intrinsic::x86_sse2_max_sd:
1394 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1395 UndefElts, Depth+1);
1396 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1397 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1398 UndefElts2, Depth+1);
1399 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1401 // If only the low elt is demanded and this is a scalarizable intrinsic,
1402 // scalarize it now.
1403 if (DemandedElts == 1) {
1404 switch (II->getIntrinsicID()) {
1406 case Intrinsic::x86_sse_sub_ss:
1407 case Intrinsic::x86_sse_mul_ss:
1408 case Intrinsic::x86_sse2_sub_sd:
1409 case Intrinsic::x86_sse2_mul_sd:
1410 // TODO: Lower MIN/MAX/ABS/etc
1411 Value *LHS = II->getOperand(1);
1412 Value *RHS = II->getOperand(2);
1413 // Extract the element as scalars.
1414 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1415 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1417 switch (II->getIntrinsicID()) {
1418 default: assert(0 && "Case stmts out of sync!");
1419 case Intrinsic::x86_sse_sub_ss:
1420 case Intrinsic::x86_sse2_sub_sd:
1421 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1422 II->getName()), *II);
1424 case Intrinsic::x86_sse_mul_ss:
1425 case Intrinsic::x86_sse2_mul_sd:
1426 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1427 II->getName()), *II);
1432 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1434 InsertNewInstBefore(New, *II);
1435 AddSoonDeadInstToWorklist(*II, 0);
1440 // Output elements are undefined if both are undefined. Consider things
1441 // like undef&0. The result is known zero, not undef.
1442 UndefElts &= UndefElts2;
1448 return MadeChange ? I : 0;
1451 /// @returns true if the specified compare instruction is
1452 /// true when both operands are equal...
1453 /// @brief Determine if the ICmpInst returns true if both operands are equal
1454 static bool isTrueWhenEqual(ICmpInst &ICI) {
1455 ICmpInst::Predicate pred = ICI.getPredicate();
1456 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1457 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1458 pred == ICmpInst::ICMP_SLE;
1461 /// @returns true if the specified compare instruction is
1462 /// true when both operands are equal...
1463 /// @brief Determine if the FCmpInst returns true if both operands are equal
1464 static bool isTrueWhenEqual(FCmpInst &FCI) {
1465 FCmpInst::Predicate pred = FCI.getPredicate();
1466 return pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ ||
1467 pred == FCmpInst::FCMP_OGE || pred == FCmpInst::FCMP_UGE ||
1468 pred == FCmpInst::FCMP_OLE || pred == FCmpInst::FCMP_ULE;
1471 /// AssociativeOpt - Perform an optimization on an associative operator. This
1472 /// function is designed to check a chain of associative operators for a
1473 /// potential to apply a certain optimization. Since the optimization may be
1474 /// applicable if the expression was reassociated, this checks the chain, then
1475 /// reassociates the expression as necessary to expose the optimization
1476 /// opportunity. This makes use of a special Functor, which must define
1477 /// 'shouldApply' and 'apply' methods.
1479 template<typename Functor>
1480 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1481 unsigned Opcode = Root.getOpcode();
1482 Value *LHS = Root.getOperand(0);
1484 // Quick check, see if the immediate LHS matches...
1485 if (F.shouldApply(LHS))
1486 return F.apply(Root);
1488 // Otherwise, if the LHS is not of the same opcode as the root, return.
1489 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1490 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1491 // Should we apply this transform to the RHS?
1492 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1494 // If not to the RHS, check to see if we should apply to the LHS...
1495 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1496 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1500 // If the functor wants to apply the optimization to the RHS of LHSI,
1501 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1503 BasicBlock *BB = Root.getParent();
1505 // Now all of the instructions are in the current basic block, go ahead
1506 // and perform the reassociation.
1507 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1509 // First move the selected RHS to the LHS of the root...
1510 Root.setOperand(0, LHSI->getOperand(1));
1512 // Make what used to be the LHS of the root be the user of the root...
1513 Value *ExtraOperand = TmpLHSI->getOperand(1);
1514 if (&Root == TmpLHSI) {
1515 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1518 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1519 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1520 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1521 BasicBlock::iterator ARI = &Root; ++ARI;
1522 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1525 // Now propagate the ExtraOperand down the chain of instructions until we
1527 while (TmpLHSI != LHSI) {
1528 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1529 // Move the instruction to immediately before the chain we are
1530 // constructing to avoid breaking dominance properties.
1531 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1532 BB->getInstList().insert(ARI, NextLHSI);
1535 Value *NextOp = NextLHSI->getOperand(1);
1536 NextLHSI->setOperand(1, ExtraOperand);
1538 ExtraOperand = NextOp;
1541 // Now that the instructions are reassociated, have the functor perform
1542 // the transformation...
1543 return F.apply(Root);
1546 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1552 // AddRHS - Implements: X + X --> X << 1
1555 AddRHS(Value *rhs) : RHS(rhs) {}
1556 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1557 Instruction *apply(BinaryOperator &Add) const {
1558 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1559 ConstantInt::get(Type::Int8Ty, 1));
1563 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1565 struct AddMaskingAnd {
1567 AddMaskingAnd(Constant *c) : C2(c) {}
1568 bool shouldApply(Value *LHS) const {
1570 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1571 ConstantExpr::getAnd(C1, C2)->isNullValue();
1573 Instruction *apply(BinaryOperator &Add) const {
1574 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1578 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1580 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1581 if (Constant *SOC = dyn_cast<Constant>(SO))
1582 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1584 return IC->InsertNewInstBefore(CastInst::create(
1585 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1588 // Figure out if the constant is the left or the right argument.
1589 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1590 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1592 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1594 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1595 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1598 Value *Op0 = SO, *Op1 = ConstOperand;
1600 std::swap(Op0, Op1);
1602 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1603 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1604 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1605 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1606 SO->getName()+".cmp");
1607 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1608 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1610 assert(0 && "Unknown binary instruction type!");
1613 return IC->InsertNewInstBefore(New, I);
1616 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1617 // constant as the other operand, try to fold the binary operator into the
1618 // select arguments. This also works for Cast instructions, which obviously do
1619 // not have a second operand.
1620 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1622 // Don't modify shared select instructions
1623 if (!SI->hasOneUse()) return 0;
1624 Value *TV = SI->getOperand(1);
1625 Value *FV = SI->getOperand(2);
1627 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1628 // Bool selects with constant operands can be folded to logical ops.
1629 if (SI->getType() == Type::Int1Ty) return 0;
1631 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1632 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1634 return new SelectInst(SI->getCondition(), SelectTrueVal,
1641 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1642 /// node as operand #0, see if we can fold the instruction into the PHI (which
1643 /// is only possible if all operands to the PHI are constants).
1644 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1645 PHINode *PN = cast<PHINode>(I.getOperand(0));
1646 unsigned NumPHIValues = PN->getNumIncomingValues();
1647 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1649 // Check to see if all of the operands of the PHI are constants. If there is
1650 // one non-constant value, remember the BB it is. If there is more than one
1652 BasicBlock *NonConstBB = 0;
1653 for (unsigned i = 0; i != NumPHIValues; ++i)
1654 if (!isa<Constant>(PN->getIncomingValue(i))) {
1655 if (NonConstBB) return 0; // More than one non-const value.
1656 NonConstBB = PN->getIncomingBlock(i);
1658 // If the incoming non-constant value is in I's block, we have an infinite
1660 if (NonConstBB == I.getParent())
1664 // If there is exactly one non-constant value, we can insert a copy of the
1665 // operation in that block. However, if this is a critical edge, we would be
1666 // inserting the computation one some other paths (e.g. inside a loop). Only
1667 // do this if the pred block is unconditionally branching into the phi block.
1669 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1670 if (!BI || !BI->isUnconditional()) return 0;
1673 // Okay, we can do the transformation: create the new PHI node.
1674 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1676 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1677 InsertNewInstBefore(NewPN, *PN);
1679 // Next, add all of the operands to the PHI.
1680 if (I.getNumOperands() == 2) {
1681 Constant *C = cast<Constant>(I.getOperand(1));
1682 for (unsigned i = 0; i != NumPHIValues; ++i) {
1684 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1685 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1686 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1688 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1690 assert(PN->getIncomingBlock(i) == NonConstBB);
1691 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1692 InV = BinaryOperator::create(BO->getOpcode(),
1693 PN->getIncomingValue(i), C, "phitmp",
1694 NonConstBB->getTerminator());
1695 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1696 InV = CmpInst::create(CI->getOpcode(),
1698 PN->getIncomingValue(i), C, "phitmp",
1699 NonConstBB->getTerminator());
1700 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1701 InV = new ShiftInst(SI->getOpcode(),
1702 PN->getIncomingValue(i), C, "phitmp",
1703 NonConstBB->getTerminator());
1705 assert(0 && "Unknown binop!");
1707 WorkList.push_back(cast<Instruction>(InV));
1709 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1712 CastInst *CI = cast<CastInst>(&I);
1713 const Type *RetTy = CI->getType();
1714 for (unsigned i = 0; i != NumPHIValues; ++i) {
1716 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1717 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1719 assert(PN->getIncomingBlock(i) == NonConstBB);
1720 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1721 I.getType(), "phitmp",
1722 NonConstBB->getTerminator());
1723 WorkList.push_back(cast<Instruction>(InV));
1725 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1728 return ReplaceInstUsesWith(I, NewPN);
1731 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1732 bool Changed = SimplifyCommutative(I);
1733 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1735 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1736 // X + undef -> undef
1737 if (isa<UndefValue>(RHS))
1738 return ReplaceInstUsesWith(I, RHS);
1741 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1742 if (RHSC->isNullValue())
1743 return ReplaceInstUsesWith(I, LHS);
1744 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1745 if (CFP->isExactlyValue(-0.0))
1746 return ReplaceInstUsesWith(I, LHS);
1749 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1750 // X + (signbit) --> X ^ signbit
1751 uint64_t Val = CI->getZExtValue();
1752 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1753 return BinaryOperator::createXor(LHS, RHS);
1755 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1756 // (X & 254)+1 -> (X&254)|1
1757 uint64_t KnownZero, KnownOne;
1758 if (!isa<PackedType>(I.getType()) &&
1759 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
1760 KnownZero, KnownOne))
1764 if (isa<PHINode>(LHS))
1765 if (Instruction *NV = FoldOpIntoPhi(I))
1768 ConstantInt *XorRHS = 0;
1770 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1771 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1772 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1773 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1775 uint64_t C0080Val = 1ULL << 31;
1776 int64_t CFF80Val = -C0080Val;
1779 if (TySizeBits > Size) {
1781 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1782 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1783 if (RHSSExt == CFF80Val) {
1784 if (XorRHS->getZExtValue() == C0080Val)
1786 } else if (RHSZExt == C0080Val) {
1787 if (XorRHS->getSExtValue() == CFF80Val)
1791 // This is a sign extend if the top bits are known zero.
1792 uint64_t Mask = ~0ULL;
1793 Mask <<= 64-(TySizeBits-Size);
1794 Mask &= XorLHS->getType()->getIntegralTypeMask();
1795 if (!MaskedValueIsZero(XorLHS, Mask))
1796 Size = 0; // Not a sign ext, but can't be any others either.
1803 } while (Size >= 8);
1806 const Type *MiddleType = 0;
1809 case 32: MiddleType = Type::Int32Ty; break;
1810 case 16: MiddleType = Type::Int16Ty; break;
1811 case 8: MiddleType = Type::Int8Ty; break;
1814 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1815 InsertNewInstBefore(NewTrunc, I);
1816 return new SExtInst(NewTrunc, I.getType());
1822 if (I.getType()->isInteger()) {
1823 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1825 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1826 if (RHSI->getOpcode() == Instruction::Sub)
1827 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1828 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1830 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1831 if (LHSI->getOpcode() == Instruction::Sub)
1832 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1833 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1838 if (Value *V = dyn_castNegVal(LHS))
1839 return BinaryOperator::createSub(RHS, V);
1842 if (!isa<Constant>(RHS))
1843 if (Value *V = dyn_castNegVal(RHS))
1844 return BinaryOperator::createSub(LHS, V);
1848 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1849 if (X == RHS) // X*C + X --> X * (C+1)
1850 return BinaryOperator::createMul(RHS, AddOne(C2));
1852 // X*C1 + X*C2 --> X * (C1+C2)
1854 if (X == dyn_castFoldableMul(RHS, C1))
1855 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1858 // X + X*C --> X * (C+1)
1859 if (dyn_castFoldableMul(RHS, C2) == LHS)
1860 return BinaryOperator::createMul(LHS, AddOne(C2));
1862 // X + ~X --> -1 since ~X = -X-1
1863 if (dyn_castNotVal(LHS) == RHS ||
1864 dyn_castNotVal(RHS) == LHS)
1865 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1868 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1869 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1870 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1873 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1875 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1876 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1877 return BinaryOperator::createSub(C, X);
1880 // (X & FF00) + xx00 -> (X+xx00) & FF00
1881 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1882 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1883 if (Anded == CRHS) {
1884 // See if all bits from the first bit set in the Add RHS up are included
1885 // in the mask. First, get the rightmost bit.
1886 uint64_t AddRHSV = CRHS->getZExtValue();
1888 // Form a mask of all bits from the lowest bit added through the top.
1889 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1890 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1892 // See if the and mask includes all of these bits.
1893 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1895 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1896 // Okay, the xform is safe. Insert the new add pronto.
1897 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1898 LHS->getName()), I);
1899 return BinaryOperator::createAnd(NewAdd, C2);
1904 // Try to fold constant add into select arguments.
1905 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1906 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1910 // add (cast *A to intptrtype) B ->
1911 // cast (GEP (cast *A to sbyte*) B) ->
1914 CastInst *CI = dyn_cast<CastInst>(LHS);
1917 CI = dyn_cast<CastInst>(RHS);
1920 if (CI && CI->getType()->isSized() &&
1921 (CI->getType()->getPrimitiveSizeInBits() ==
1922 TD->getIntPtrType()->getPrimitiveSizeInBits())
1923 && isa<PointerType>(CI->getOperand(0)->getType())) {
1924 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1925 PointerType::get(Type::Int8Ty), I);
1926 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1927 return new PtrToIntInst(I2, CI->getType());
1931 return Changed ? &I : 0;
1934 // isSignBit - Return true if the value represented by the constant only has the
1935 // highest order bit set.
1936 static bool isSignBit(ConstantInt *CI) {
1937 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1938 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1941 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1943 static Value *RemoveNoopCast(Value *V) {
1944 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1945 const Type *CTy = CI->getType();
1946 const Type *OpTy = CI->getOperand(0)->getType();
1947 if (CTy->isInteger() && OpTy->isInteger()) {
1948 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1949 return RemoveNoopCast(CI->getOperand(0));
1950 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1951 return RemoveNoopCast(CI->getOperand(0));
1956 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1957 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1959 if (Op0 == Op1) // sub X, X -> 0
1960 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1962 // If this is a 'B = x-(-A)', change to B = x+A...
1963 if (Value *V = dyn_castNegVal(Op1))
1964 return BinaryOperator::createAdd(Op0, V);
1966 if (isa<UndefValue>(Op0))
1967 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1968 if (isa<UndefValue>(Op1))
1969 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1971 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1972 // Replace (-1 - A) with (~A)...
1973 if (C->isAllOnesValue())
1974 return BinaryOperator::createNot(Op1);
1976 // C - ~X == X + (1+C)
1978 if (match(Op1, m_Not(m_Value(X))))
1979 return BinaryOperator::createAdd(X,
1980 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1981 // -((uint)X >> 31) -> ((int)X >> 31)
1982 // -((int)X >> 31) -> ((uint)X >> 31)
1983 if (C->isNullValue()) {
1984 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1985 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1986 if (SI->getOpcode() == Instruction::LShr) {
1987 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1988 // Check to see if we are shifting out everything but the sign bit.
1989 if (CU->getZExtValue() ==
1990 SI->getType()->getPrimitiveSizeInBits()-1) {
1991 // Ok, the transformation is safe. Insert AShr.
1992 // FIXME: Once integer types are signless, this cast should be
1994 Value *ShiftOp = SI->getOperand(0);
1995 return new ShiftInst(Instruction::AShr, ShiftOp, CU,
2000 else if (SI->getOpcode() == Instruction::AShr) {
2001 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2002 // Check to see if we are shifting out everything but the sign bit.
2003 if (CU->getZExtValue() ==
2004 SI->getType()->getPrimitiveSizeInBits()-1) {
2006 // Ok, the transformation is safe. Insert LShr.
2007 return new ShiftInst(Instruction::LShr, SI->getOperand(0), CU,
2014 // Try to fold constant sub into select arguments.
2015 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2016 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2019 if (isa<PHINode>(Op0))
2020 if (Instruction *NV = FoldOpIntoPhi(I))
2024 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2025 if (Op1I->getOpcode() == Instruction::Add &&
2026 !Op0->getType()->isFPOrFPVector()) {
2027 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2028 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2029 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2030 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2031 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2032 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2033 // C1-(X+C2) --> (C1-C2)-X
2034 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2035 Op1I->getOperand(0));
2039 if (Op1I->hasOneUse()) {
2040 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2041 // is not used by anyone else...
2043 if (Op1I->getOpcode() == Instruction::Sub &&
2044 !Op1I->getType()->isFPOrFPVector()) {
2045 // Swap the two operands of the subexpr...
2046 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2047 Op1I->setOperand(0, IIOp1);
2048 Op1I->setOperand(1, IIOp0);
2050 // Create the new top level add instruction...
2051 return BinaryOperator::createAdd(Op0, Op1);
2054 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2056 if (Op1I->getOpcode() == Instruction::And &&
2057 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2058 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2061 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2062 return BinaryOperator::createAnd(Op0, NewNot);
2065 // 0 - (X sdiv C) -> (X sdiv -C)
2066 if (Op1I->getOpcode() == Instruction::SDiv)
2067 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2068 if (CSI->isNullValue())
2069 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2070 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2071 ConstantExpr::getNeg(DivRHS));
2073 // X - X*C --> X * (1-C)
2074 ConstantInt *C2 = 0;
2075 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2077 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2078 return BinaryOperator::createMul(Op0, CP1);
2083 if (!Op0->getType()->isFPOrFPVector())
2084 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2085 if (Op0I->getOpcode() == Instruction::Add) {
2086 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2087 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2088 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2089 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2090 } else if (Op0I->getOpcode() == Instruction::Sub) {
2091 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2092 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2096 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2097 if (X == Op1) { // X*C - X --> X * (C-1)
2098 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2099 return BinaryOperator::createMul(Op1, CP1);
2102 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2103 if (X == dyn_castFoldableMul(Op1, C2))
2104 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2109 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2110 /// really just returns true if the most significant (sign) bit is set.
2111 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2113 case ICmpInst::ICMP_SLT:
2114 // True if LHS s< RHS and RHS == 0
2115 return RHS->isNullValue();
2116 case ICmpInst::ICMP_SLE:
2117 // True if LHS s<= RHS and RHS == -1
2118 return RHS->isAllOnesValue();
2119 case ICmpInst::ICMP_UGE:
2120 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2121 return RHS->getZExtValue() == (1ULL <<
2122 (RHS->getType()->getPrimitiveSizeInBits()-1));
2123 case ICmpInst::ICMP_UGT:
2124 // True if LHS u> RHS and RHS == high-bit-mask - 1
2125 return RHS->getZExtValue() ==
2126 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2132 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2133 bool Changed = SimplifyCommutative(I);
2134 Value *Op0 = I.getOperand(0);
2136 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2137 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2139 // Simplify mul instructions with a constant RHS...
2140 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2141 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2143 // ((X << C1)*C2) == (X * (C2 << C1))
2144 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2145 if (SI->getOpcode() == Instruction::Shl)
2146 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2147 return BinaryOperator::createMul(SI->getOperand(0),
2148 ConstantExpr::getShl(CI, ShOp));
2150 if (CI->isNullValue())
2151 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2152 if (CI->equalsInt(1)) // X * 1 == X
2153 return ReplaceInstUsesWith(I, Op0);
2154 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2155 return BinaryOperator::createNeg(Op0, I.getName());
2157 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2158 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2159 uint64_t C = Log2_64(Val);
2160 return new ShiftInst(Instruction::Shl, Op0,
2161 ConstantInt::get(Type::Int8Ty, C));
2163 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2164 if (Op1F->isNullValue())
2165 return ReplaceInstUsesWith(I, Op1);
2167 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2168 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2169 if (Op1F->getValue() == 1.0)
2170 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2173 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2174 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2175 isa<ConstantInt>(Op0I->getOperand(1))) {
2176 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2177 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2179 InsertNewInstBefore(Add, I);
2180 Value *C1C2 = ConstantExpr::getMul(Op1,
2181 cast<Constant>(Op0I->getOperand(1)));
2182 return BinaryOperator::createAdd(Add, C1C2);
2186 // Try to fold constant mul into select arguments.
2187 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2188 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2191 if (isa<PHINode>(Op0))
2192 if (Instruction *NV = FoldOpIntoPhi(I))
2196 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2197 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2198 return BinaryOperator::createMul(Op0v, Op1v);
2200 // If one of the operands of the multiply is a cast from a boolean value, then
2201 // we know the bool is either zero or one, so this is a 'masking' multiply.
2202 // See if we can simplify things based on how the boolean was originally
2204 CastInst *BoolCast = 0;
2205 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2206 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2209 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2210 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2213 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2214 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2215 const Type *SCOpTy = SCIOp0->getType();
2217 // If the icmp is true iff the sign bit of X is set, then convert this
2218 // multiply into a shift/and combination.
2219 if (isa<ConstantInt>(SCIOp1) &&
2220 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2221 // Shift the X value right to turn it into "all signbits".
2222 Constant *Amt = ConstantInt::get(Type::Int8Ty,
2223 SCOpTy->getPrimitiveSizeInBits()-1);
2225 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2226 BoolCast->getOperand(0)->getName()+
2229 // If the multiply type is not the same as the source type, sign extend
2230 // or truncate to the multiply type.
2231 if (I.getType() != V->getType()) {
2232 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2233 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2234 Instruction::CastOps opcode =
2235 (SrcBits == DstBits ? Instruction::BitCast :
2236 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2237 V = InsertCastBefore(opcode, V, I.getType(), I);
2240 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2241 return BinaryOperator::createAnd(V, OtherOp);
2246 return Changed ? &I : 0;
2249 /// This function implements the transforms on div instructions that work
2250 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2251 /// used by the visitors to those instructions.
2252 /// @brief Transforms common to all three div instructions
2253 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2254 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2257 if (isa<UndefValue>(Op0))
2258 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2260 // X / undef -> undef
2261 if (isa<UndefValue>(Op1))
2262 return ReplaceInstUsesWith(I, Op1);
2264 // Handle cases involving: div X, (select Cond, Y, Z)
2265 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2266 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2267 // same basic block, then we replace the select with Y, and the condition
2268 // of the select with false (if the cond value is in the same BB). If the
2269 // select has uses other than the div, this allows them to be simplified
2270 // also. Note that div X, Y is just as good as div X, 0 (undef)
2271 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2272 if (ST->isNullValue()) {
2273 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2274 if (CondI && CondI->getParent() == I.getParent())
2275 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2276 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2277 I.setOperand(1, SI->getOperand(2));
2279 UpdateValueUsesWith(SI, SI->getOperand(2));
2283 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2284 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2285 if (ST->isNullValue()) {
2286 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2287 if (CondI && CondI->getParent() == I.getParent())
2288 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2289 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2290 I.setOperand(1, SI->getOperand(1));
2292 UpdateValueUsesWith(SI, SI->getOperand(1));
2300 /// This function implements the transforms common to both integer division
2301 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2302 /// division instructions.
2303 /// @brief Common integer divide transforms
2304 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2305 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2307 if (Instruction *Common = commonDivTransforms(I))
2310 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2312 if (RHS->equalsInt(1))
2313 return ReplaceInstUsesWith(I, Op0);
2315 // (X / C1) / C2 -> X / (C1*C2)
2316 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2317 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2318 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2319 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2320 ConstantExpr::getMul(RHS, LHSRHS));
2323 if (!RHS->isNullValue()) { // avoid X udiv 0
2324 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2325 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2327 if (isa<PHINode>(Op0))
2328 if (Instruction *NV = FoldOpIntoPhi(I))
2333 // 0 / X == 0, we don't need to preserve faults!
2334 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2335 if (LHS->equalsInt(0))
2336 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2341 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2342 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2344 // Handle the integer div common cases
2345 if (Instruction *Common = commonIDivTransforms(I))
2348 // X udiv C^2 -> X >> C
2349 // Check to see if this is an unsigned division with an exact power of 2,
2350 // if so, convert to a right shift.
2351 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2352 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2353 if (isPowerOf2_64(Val)) {
2354 uint64_t ShiftAmt = Log2_64(Val);
2355 return new ShiftInst(Instruction::LShr, Op0,
2356 ConstantInt::get(Type::Int8Ty, ShiftAmt));
2360 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2361 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2362 if (RHSI->getOpcode() == Instruction::Shl &&
2363 isa<ConstantInt>(RHSI->getOperand(0))) {
2364 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2365 if (isPowerOf2_64(C1)) {
2366 Value *N = RHSI->getOperand(1);
2367 const Type *NTy = N->getType();
2368 if (uint64_t C2 = Log2_64(C1)) {
2369 Constant *C2V = ConstantInt::get(NTy, C2);
2370 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2372 return new ShiftInst(Instruction::LShr, Op0, N);
2377 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2378 // where C1&C2 are powers of two.
2379 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2380 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2381 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2382 if (!STO->isNullValue() && !STO->isNullValue()) {
2383 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2384 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2385 // Compute the shift amounts
2386 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2387 // Construct the "on true" case of the select
2388 Constant *TC = ConstantInt::get(Type::Int8Ty, TSA);
2390 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2391 TSI = InsertNewInstBefore(TSI, I);
2393 // Construct the "on false" case of the select
2394 Constant *FC = ConstantInt::get(Type::Int8Ty, FSA);
2396 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2397 FSI = InsertNewInstBefore(FSI, I);
2399 // construct the select instruction and return it.
2400 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2407 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2408 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2410 // Handle the integer div common cases
2411 if (Instruction *Common = commonIDivTransforms(I))
2414 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2416 if (RHS->isAllOnesValue())
2417 return BinaryOperator::createNeg(Op0);
2420 if (Value *LHSNeg = dyn_castNegVal(Op0))
2421 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2424 // If the sign bits of both operands are zero (i.e. we can prove they are
2425 // unsigned inputs), turn this into a udiv.
2426 if (I.getType()->isInteger()) {
2427 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2428 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2429 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2436 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2437 return commonDivTransforms(I);
2440 /// GetFactor - If we can prove that the specified value is at least a multiple
2441 /// of some factor, return that factor.
2442 static Constant *GetFactor(Value *V) {
2443 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2446 // Unless we can be tricky, we know this is a multiple of 1.
2447 Constant *Result = ConstantInt::get(V->getType(), 1);
2449 Instruction *I = dyn_cast<Instruction>(V);
2450 if (!I) return Result;
2452 if (I->getOpcode() == Instruction::Mul) {
2453 // Handle multiplies by a constant, etc.
2454 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2455 GetFactor(I->getOperand(1)));
2456 } else if (I->getOpcode() == Instruction::Shl) {
2457 // (X<<C) -> X * (1 << C)
2458 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2459 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2460 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2462 } else if (I->getOpcode() == Instruction::And) {
2463 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2464 // X & 0xFFF0 is known to be a multiple of 16.
2465 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2466 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2467 return ConstantExpr::getShl(Result,
2468 ConstantInt::get(Type::Int8Ty, Zeros));
2470 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2471 // Only handle int->int casts.
2472 if (!CI->isIntegerCast())
2474 Value *Op = CI->getOperand(0);
2475 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2480 /// This function implements the transforms on rem instructions that work
2481 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2482 /// is used by the visitors to those instructions.
2483 /// @brief Transforms common to all three rem instructions
2484 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2485 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2487 // 0 % X == 0, we don't need to preserve faults!
2488 if (Constant *LHS = dyn_cast<Constant>(Op0))
2489 if (LHS->isNullValue())
2490 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2492 if (isa<UndefValue>(Op0)) // undef % X -> 0
2493 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2494 if (isa<UndefValue>(Op1))
2495 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2497 // Handle cases involving: rem X, (select Cond, Y, Z)
2498 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2499 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2500 // the same basic block, then we replace the select with Y, and the
2501 // condition of the select with false (if the cond value is in the same
2502 // BB). If the select has uses other than the div, this allows them to be
2504 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2505 if (ST->isNullValue()) {
2506 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2507 if (CondI && CondI->getParent() == I.getParent())
2508 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2509 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2510 I.setOperand(1, SI->getOperand(2));
2512 UpdateValueUsesWith(SI, SI->getOperand(2));
2515 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2516 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2517 if (ST->isNullValue()) {
2518 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2519 if (CondI && CondI->getParent() == I.getParent())
2520 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2521 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2522 I.setOperand(1, SI->getOperand(1));
2524 UpdateValueUsesWith(SI, SI->getOperand(1));
2532 /// This function implements the transforms common to both integer remainder
2533 /// instructions (urem and srem). It is called by the visitors to those integer
2534 /// remainder instructions.
2535 /// @brief Common integer remainder transforms
2536 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2537 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2539 if (Instruction *common = commonRemTransforms(I))
2542 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2543 // X % 0 == undef, we don't need to preserve faults!
2544 if (RHS->equalsInt(0))
2545 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2547 if (RHS->equalsInt(1)) // X % 1 == 0
2548 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2550 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2551 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2552 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2554 } else if (isa<PHINode>(Op0I)) {
2555 if (Instruction *NV = FoldOpIntoPhi(I))
2558 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2559 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2560 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2567 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2568 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 if (Instruction *common = commonIRemTransforms(I))
2573 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2574 // X urem C^2 -> X and C
2575 // Check to see if this is an unsigned remainder with an exact power of 2,
2576 // if so, convert to a bitwise and.
2577 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2578 if (isPowerOf2_64(C->getZExtValue()))
2579 return BinaryOperator::createAnd(Op0, SubOne(C));
2582 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2583 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2584 if (RHSI->getOpcode() == Instruction::Shl &&
2585 isa<ConstantInt>(RHSI->getOperand(0))) {
2586 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2587 if (isPowerOf2_64(C1)) {
2588 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2589 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2591 return BinaryOperator::createAnd(Op0, Add);
2596 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2597 // where C1&C2 are powers of two.
2598 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2599 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2600 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2601 // STO == 0 and SFO == 0 handled above.
2602 if (isPowerOf2_64(STO->getZExtValue()) &&
2603 isPowerOf2_64(SFO->getZExtValue())) {
2604 Value *TrueAnd = InsertNewInstBefore(
2605 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2606 Value *FalseAnd = InsertNewInstBefore(
2607 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2608 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2616 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2617 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2619 if (Instruction *common = commonIRemTransforms(I))
2622 if (Value *RHSNeg = dyn_castNegVal(Op1))
2623 if (!isa<ConstantInt>(RHSNeg) ||
2624 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2626 AddUsesToWorkList(I);
2627 I.setOperand(1, RHSNeg);
2631 // If the top bits of both operands are zero (i.e. we can prove they are
2632 // unsigned inputs), turn this into a urem.
2633 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2634 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2635 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2636 return BinaryOperator::createURem(Op0, Op1, I.getName());
2642 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2643 return commonRemTransforms(I);
2646 // isMaxValueMinusOne - return true if this is Max-1
2647 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2649 // Calculate 0111111111..11111
2650 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2651 int64_t Val = INT64_MAX; // All ones
2652 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2653 return C->getSExtValue() == Val-1;
2655 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2658 // isMinValuePlusOne - return true if this is Min+1
2659 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2661 // Calculate 1111111111000000000000
2662 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2663 int64_t Val = -1; // All ones
2664 Val <<= TypeBits-1; // Shift over to the right spot
2665 return C->getSExtValue() == Val+1;
2667 return C->getZExtValue() == 1; // unsigned
2670 // isOneBitSet - Return true if there is exactly one bit set in the specified
2672 static bool isOneBitSet(const ConstantInt *CI) {
2673 uint64_t V = CI->getZExtValue();
2674 return V && (V & (V-1)) == 0;
2677 #if 0 // Currently unused
2678 // isLowOnes - Return true if the constant is of the form 0+1+.
2679 static bool isLowOnes(const ConstantInt *CI) {
2680 uint64_t V = CI->getZExtValue();
2682 // There won't be bits set in parts that the type doesn't contain.
2683 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2685 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2686 return U && V && (U & V) == 0;
2690 // isHighOnes - Return true if the constant is of the form 1+0+.
2691 // This is the same as lowones(~X).
2692 static bool isHighOnes(const ConstantInt *CI) {
2693 uint64_t V = ~CI->getZExtValue();
2694 if (~V == 0) return false; // 0's does not match "1+"
2696 // There won't be bits set in parts that the type doesn't contain.
2697 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2699 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2700 return U && V && (U & V) == 0;
2703 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2704 /// are carefully arranged to allow folding of expressions such as:
2706 /// (A < B) | (A > B) --> (A != B)
2708 /// Note that this is only valid if the first and second predicates have the
2709 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2711 /// Three bits are used to represent the condition, as follows:
2716 /// <=> Value Definition
2717 /// 000 0 Always false
2724 /// 111 7 Always true
2726 static unsigned getICmpCode(const ICmpInst *ICI) {
2727 switch (ICI->getPredicate()) {
2729 case ICmpInst::ICMP_UGT: return 1; // 001
2730 case ICmpInst::ICMP_SGT: return 1; // 001
2731 case ICmpInst::ICMP_EQ: return 2; // 010
2732 case ICmpInst::ICMP_UGE: return 3; // 011
2733 case ICmpInst::ICMP_SGE: return 3; // 011
2734 case ICmpInst::ICMP_ULT: return 4; // 100
2735 case ICmpInst::ICMP_SLT: return 4; // 100
2736 case ICmpInst::ICMP_NE: return 5; // 101
2737 case ICmpInst::ICMP_ULE: return 6; // 110
2738 case ICmpInst::ICMP_SLE: return 6; // 110
2741 assert(0 && "Invalid ICmp predicate!");
2746 /// getICmpValue - This is the complement of getICmpCode, which turns an
2747 /// opcode and two operands into either a constant true or false, or a brand
2748 /// new /// ICmp instruction. The sign is passed in to determine which kind
2749 /// of predicate to use in new icmp instructions.
2750 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2752 default: assert(0 && "Illegal ICmp code!");
2753 case 0: return ConstantInt::getFalse();
2756 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2758 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2759 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2762 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2764 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2767 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2769 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2770 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2773 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2775 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2776 case 7: return ConstantInt::getTrue();
2780 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2781 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2782 (ICmpInst::isSignedPredicate(p1) &&
2783 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2784 (ICmpInst::isSignedPredicate(p2) &&
2785 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2789 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2790 struct FoldICmpLogical {
2793 ICmpInst::Predicate pred;
2794 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2795 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2796 pred(ICI->getPredicate()) {}
2797 bool shouldApply(Value *V) const {
2798 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2799 if (PredicatesFoldable(pred, ICI->getPredicate()))
2800 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2801 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2804 Instruction *apply(Instruction &Log) const {
2805 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2806 if (ICI->getOperand(0) != LHS) {
2807 assert(ICI->getOperand(1) == LHS);
2808 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2811 unsigned LHSCode = getICmpCode(ICI);
2812 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2814 switch (Log.getOpcode()) {
2815 case Instruction::And: Code = LHSCode & RHSCode; break;
2816 case Instruction::Or: Code = LHSCode | RHSCode; break;
2817 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2818 default: assert(0 && "Illegal logical opcode!"); return 0;
2821 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2822 if (Instruction *I = dyn_cast<Instruction>(RV))
2824 // Otherwise, it's a constant boolean value...
2825 return IC.ReplaceInstUsesWith(Log, RV);
2828 } // end anonymous namespace
2830 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2831 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2832 // guaranteed to be either a shift instruction or a binary operator.
2833 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2835 ConstantInt *AndRHS,
2836 BinaryOperator &TheAnd) {
2837 Value *X = Op->getOperand(0);
2838 Constant *Together = 0;
2839 if (!isa<ShiftInst>(Op))
2840 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2842 switch (Op->getOpcode()) {
2843 case Instruction::Xor:
2844 if (Op->hasOneUse()) {
2845 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2846 std::string OpName = Op->getName(); Op->setName("");
2847 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2848 InsertNewInstBefore(And, TheAnd);
2849 return BinaryOperator::createXor(And, Together);
2852 case Instruction::Or:
2853 if (Together == AndRHS) // (X | C) & C --> C
2854 return ReplaceInstUsesWith(TheAnd, AndRHS);
2856 if (Op->hasOneUse() && Together != OpRHS) {
2857 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2858 std::string Op0Name = Op->getName(); Op->setName("");
2859 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2860 InsertNewInstBefore(Or, TheAnd);
2861 return BinaryOperator::createAnd(Or, AndRHS);
2864 case Instruction::Add:
2865 if (Op->hasOneUse()) {
2866 // Adding a one to a single bit bit-field should be turned into an XOR
2867 // of the bit. First thing to check is to see if this AND is with a
2868 // single bit constant.
2869 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2871 // Clear bits that are not part of the constant.
2872 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2874 // If there is only one bit set...
2875 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2876 // Ok, at this point, we know that we are masking the result of the
2877 // ADD down to exactly one bit. If the constant we are adding has
2878 // no bits set below this bit, then we can eliminate the ADD.
2879 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2881 // Check to see if any bits below the one bit set in AndRHSV are set.
2882 if ((AddRHS & (AndRHSV-1)) == 0) {
2883 // If not, the only thing that can effect the output of the AND is
2884 // the bit specified by AndRHSV. If that bit is set, the effect of
2885 // the XOR is to toggle the bit. If it is clear, then the ADD has
2887 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2888 TheAnd.setOperand(0, X);
2891 std::string Name = Op->getName(); Op->setName("");
2892 // Pull the XOR out of the AND.
2893 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2894 InsertNewInstBefore(NewAnd, TheAnd);
2895 return BinaryOperator::createXor(NewAnd, AndRHS);
2902 case Instruction::Shl: {
2903 // We know that the AND will not produce any of the bits shifted in, so if
2904 // the anded constant includes them, clear them now!
2906 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2907 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2908 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2910 if (CI == ShlMask) { // Masking out bits that the shift already masks
2911 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2912 } else if (CI != AndRHS) { // Reducing bits set in and.
2913 TheAnd.setOperand(1, CI);
2918 case Instruction::LShr:
2920 // We know that the AND will not produce any of the bits shifted in, so if
2921 // the anded constant includes them, clear them now! This only applies to
2922 // unsigned shifts, because a signed shr may bring in set bits!
2924 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2925 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2926 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2928 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2929 return ReplaceInstUsesWith(TheAnd, Op);
2930 } else if (CI != AndRHS) {
2931 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2936 case Instruction::AShr:
2938 // See if this is shifting in some sign extension, then masking it out
2940 if (Op->hasOneUse()) {
2941 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2942 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2943 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2944 if (C == AndRHS) { // Masking out bits shifted in.
2945 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2946 // Make the argument unsigned.
2947 Value *ShVal = Op->getOperand(0);
2948 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2949 OpRHS, Op->getName()), TheAnd);
2950 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2959 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2960 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2961 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2962 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2963 /// insert new instructions.
2964 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2965 bool isSigned, bool Inside,
2967 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2968 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getBoolValue() &&
2969 "Lo is not <= Hi in range emission code!");
2972 if (Lo == Hi) // Trivially false.
2973 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2975 // V >= Min && V < Hi --> V < Hi
2976 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2977 ICmpInst::Predicate pred = (isSigned ?
2978 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2979 return new ICmpInst(pred, V, Hi);
2982 // Emit V-Lo <u Hi-Lo
2983 Constant *NegLo = ConstantExpr::getNeg(Lo);
2984 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2985 InsertNewInstBefore(Add, IB);
2986 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2987 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2990 if (Lo == Hi) // Trivially true.
2991 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2993 // V < Min || V >= Hi ->'V > Hi-1'
2994 Hi = SubOne(cast<ConstantInt>(Hi));
2995 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2996 ICmpInst::Predicate pred = (isSigned ?
2997 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2998 return new ICmpInst(pred, V, Hi);
3001 // Emit V-Lo > Hi-1-Lo
3002 Constant *NegLo = ConstantExpr::getNeg(Lo);
3003 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3004 InsertNewInstBefore(Add, IB);
3005 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3006 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3009 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3010 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3011 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3012 // not, since all 1s are not contiguous.
3013 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3014 uint64_t V = Val->getZExtValue();
3015 if (!isShiftedMask_64(V)) return false;
3017 // look for the first zero bit after the run of ones
3018 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3019 // look for the first non-zero bit
3020 ME = 64-CountLeadingZeros_64(V);
3026 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3027 /// where isSub determines whether the operator is a sub. If we can fold one of
3028 /// the following xforms:
3030 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3031 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3032 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3034 /// return (A +/- B).
3036 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3037 ConstantInt *Mask, bool isSub,
3039 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3040 if (!LHSI || LHSI->getNumOperands() != 2 ||
3041 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3043 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3045 switch (LHSI->getOpcode()) {
3047 case Instruction::And:
3048 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3049 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3050 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3053 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3054 // part, we don't need any explicit masks to take them out of A. If that
3055 // is all N is, ignore it.
3057 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3058 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
3060 if (MaskedValueIsZero(RHS, Mask))
3065 case Instruction::Or:
3066 case Instruction::Xor:
3067 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3068 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3069 ConstantExpr::getAnd(N, Mask)->isNullValue())
3076 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3078 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3079 return InsertNewInstBefore(New, I);
3082 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3083 bool Changed = SimplifyCommutative(I);
3084 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3086 if (isa<UndefValue>(Op1)) // X & undef -> 0
3087 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3091 return ReplaceInstUsesWith(I, Op1);
3093 // See if we can simplify any instructions used by the instruction whose sole
3094 // purpose is to compute bits we don't care about.
3095 uint64_t KnownZero, KnownOne;
3096 if (!isa<PackedType>(I.getType()) &&
3097 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3098 KnownZero, KnownOne))
3101 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3102 uint64_t AndRHSMask = AndRHS->getZExtValue();
3103 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
3104 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3106 // Optimize a variety of ((val OP C1) & C2) combinations...
3107 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3108 Instruction *Op0I = cast<Instruction>(Op0);
3109 Value *Op0LHS = Op0I->getOperand(0);
3110 Value *Op0RHS = Op0I->getOperand(1);
3111 switch (Op0I->getOpcode()) {
3112 case Instruction::Xor:
3113 case Instruction::Or:
3114 // If the mask is only needed on one incoming arm, push it up.
3115 if (Op0I->hasOneUse()) {
3116 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3117 // Not masking anything out for the LHS, move to RHS.
3118 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3119 Op0RHS->getName()+".masked");
3120 InsertNewInstBefore(NewRHS, I);
3121 return BinaryOperator::create(
3122 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3124 if (!isa<Constant>(Op0RHS) &&
3125 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3126 // Not masking anything out for the RHS, move to LHS.
3127 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3128 Op0LHS->getName()+".masked");
3129 InsertNewInstBefore(NewLHS, I);
3130 return BinaryOperator::create(
3131 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3136 case Instruction::Add:
3137 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3138 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3139 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3140 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3141 return BinaryOperator::createAnd(V, AndRHS);
3142 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3143 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3146 case Instruction::Sub:
3147 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3148 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3149 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3150 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3151 return BinaryOperator::createAnd(V, AndRHS);
3155 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3156 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3158 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3159 // If this is an integer truncation or change from signed-to-unsigned, and
3160 // if the source is an and/or with immediate, transform it. This
3161 // frequently occurs for bitfield accesses.
3162 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3163 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3164 CastOp->getNumOperands() == 2)
3165 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3166 if (CastOp->getOpcode() == Instruction::And) {
3167 // Change: and (cast (and X, C1) to T), C2
3168 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3169 // This will fold the two constants together, which may allow
3170 // other simplifications.
3171 Instruction *NewCast = CastInst::createTruncOrBitCast(
3172 CastOp->getOperand(0), I.getType(),
3173 CastOp->getName()+".shrunk");
3174 NewCast = InsertNewInstBefore(NewCast, I);
3175 // trunc_or_bitcast(C1)&C2
3176 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3177 C3 = ConstantExpr::getAnd(C3, AndRHS);
3178 return BinaryOperator::createAnd(NewCast, C3);
3179 } else if (CastOp->getOpcode() == Instruction::Or) {
3180 // Change: and (cast (or X, C1) to T), C2
3181 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3182 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3183 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3184 return ReplaceInstUsesWith(I, AndRHS);
3189 // Try to fold constant and into select arguments.
3190 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3191 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3193 if (isa<PHINode>(Op0))
3194 if (Instruction *NV = FoldOpIntoPhi(I))
3198 Value *Op0NotVal = dyn_castNotVal(Op0);
3199 Value *Op1NotVal = dyn_castNotVal(Op1);
3201 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3202 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3204 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3205 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3206 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3207 I.getName()+".demorgan");
3208 InsertNewInstBefore(Or, I);
3209 return BinaryOperator::createNot(Or);
3213 Value *A = 0, *B = 0;
3214 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3215 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3216 return ReplaceInstUsesWith(I, Op1);
3217 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3218 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3219 return ReplaceInstUsesWith(I, Op0);
3221 if (Op0->hasOneUse() &&
3222 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3223 if (A == Op1) { // (A^B)&A -> A&(A^B)
3224 I.swapOperands(); // Simplify below
3225 std::swap(Op0, Op1);
3226 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3227 cast<BinaryOperator>(Op0)->swapOperands();
3228 I.swapOperands(); // Simplify below
3229 std::swap(Op0, Op1);
3232 if (Op1->hasOneUse() &&
3233 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3234 if (B == Op0) { // B&(A^B) -> B&(B^A)
3235 cast<BinaryOperator>(Op1)->swapOperands();
3238 if (A == Op0) { // A&(A^B) -> A & ~B
3239 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3240 InsertNewInstBefore(NotB, I);
3241 return BinaryOperator::createAnd(A, NotB);
3246 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3247 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3248 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3251 Value *LHSVal, *RHSVal;
3252 ConstantInt *LHSCst, *RHSCst;
3253 ICmpInst::Predicate LHSCC, RHSCC;
3254 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3255 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3256 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3257 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3258 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3259 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3260 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3261 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3262 // Ensure that the larger constant is on the RHS.
3263 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3264 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3265 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3266 ICmpInst *LHS = cast<ICmpInst>(Op0);
3267 if (cast<ConstantInt>(Cmp)->getBoolValue()) {
3268 std::swap(LHS, RHS);
3269 std::swap(LHSCst, RHSCst);
3270 std::swap(LHSCC, RHSCC);
3273 // At this point, we know we have have two icmp instructions
3274 // comparing a value against two constants and and'ing the result
3275 // together. Because of the above check, we know that we only have
3276 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3277 // (from the FoldICmpLogical check above), that the two constants
3278 // are not equal and that the larger constant is on the RHS
3279 assert(LHSCst != RHSCst && "Compares not folded above?");
3282 default: assert(0 && "Unknown integer condition code!");
3283 case ICmpInst::ICMP_EQ:
3285 default: assert(0 && "Unknown integer condition code!");
3286 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3287 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3288 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3289 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3290 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3291 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3292 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3293 return ReplaceInstUsesWith(I, LHS);
3295 case ICmpInst::ICMP_NE:
3297 default: assert(0 && "Unknown integer condition code!");
3298 case ICmpInst::ICMP_ULT:
3299 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3300 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3301 break; // (X != 13 & X u< 15) -> no change
3302 case ICmpInst::ICMP_SLT:
3303 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3304 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3305 break; // (X != 13 & X s< 15) -> no change
3306 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3307 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3308 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3309 return ReplaceInstUsesWith(I, RHS);
3310 case ICmpInst::ICMP_NE:
3311 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3312 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3313 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3314 LHSVal->getName()+".off");
3315 InsertNewInstBefore(Add, I);
3316 return new ICmpInst(ICmpInst::ICMP_UGT, Add, AddCST);
3318 break; // (X != 13 & X != 15) -> no change
3321 case ICmpInst::ICMP_ULT:
3323 default: assert(0 && "Unknown integer condition code!");
3324 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3325 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3326 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3327 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3329 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3330 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3331 return ReplaceInstUsesWith(I, LHS);
3332 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3336 case ICmpInst::ICMP_SLT:
3338 default: assert(0 && "Unknown integer condition code!");
3339 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3340 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3341 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3342 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3344 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3345 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3346 return ReplaceInstUsesWith(I, LHS);
3347 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3351 case ICmpInst::ICMP_UGT:
3353 default: assert(0 && "Unknown integer condition code!");
3354 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3355 return ReplaceInstUsesWith(I, LHS);
3356 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3357 return ReplaceInstUsesWith(I, RHS);
3358 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3360 case ICmpInst::ICMP_NE:
3361 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3362 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3363 break; // (X u> 13 & X != 15) -> no change
3364 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3365 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3367 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3371 case ICmpInst::ICMP_SGT:
3373 default: assert(0 && "Unknown integer condition code!");
3374 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3375 return ReplaceInstUsesWith(I, LHS);
3376 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3377 return ReplaceInstUsesWith(I, RHS);
3378 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3380 case ICmpInst::ICMP_NE:
3381 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3382 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3383 break; // (X s> 13 & X != 15) -> no change
3384 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3385 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3387 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3395 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3396 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3397 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3398 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3399 const Type *SrcTy = Op0C->getOperand(0)->getType();
3400 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3401 // Only do this if the casts both really cause code to be generated.
3402 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3404 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3406 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3407 Op1C->getOperand(0),
3409 InsertNewInstBefore(NewOp, I);
3410 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3414 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3415 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3416 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3417 if (SI0->getOpcode() == SI1->getOpcode() &&
3418 SI0->getOperand(1) == SI1->getOperand(1) &&
3419 (SI0->hasOneUse() || SI1->hasOneUse())) {
3420 Instruction *NewOp =
3421 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3423 SI0->getName()), I);
3424 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3428 return Changed ? &I : 0;
3431 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3432 /// in the result. If it does, and if the specified byte hasn't been filled in
3433 /// yet, fill it in and return false.
3434 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3435 Instruction *I = dyn_cast<Instruction>(V);
3436 if (I == 0) return true;
3438 // If this is an or instruction, it is an inner node of the bswap.
3439 if (I->getOpcode() == Instruction::Or)
3440 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3441 CollectBSwapParts(I->getOperand(1), ByteValues);
3443 // If this is a shift by a constant int, and it is "24", then its operand
3444 // defines a byte. We only handle unsigned types here.
3445 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3446 // Not shifting the entire input by N-1 bytes?
3447 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3448 8*(ByteValues.size()-1))
3452 if (I->getOpcode() == Instruction::Shl) {
3453 // X << 24 defines the top byte with the lowest of the input bytes.
3454 DestNo = ByteValues.size()-1;
3456 // X >>u 24 defines the low byte with the highest of the input bytes.
3460 // If the destination byte value is already defined, the values are or'd
3461 // together, which isn't a bswap (unless it's an or of the same bits).
3462 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3464 ByteValues[DestNo] = I->getOperand(0);
3468 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3470 Value *Shift = 0, *ShiftLHS = 0;
3471 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3472 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3473 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3475 Instruction *SI = cast<Instruction>(Shift);
3477 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3478 if (ShiftAmt->getZExtValue() & 7 ||
3479 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3482 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3484 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3485 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3487 // Unknown mask for bswap.
3488 if (DestByte == ByteValues.size()) return true;
3490 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3492 if (SI->getOpcode() == Instruction::Shl)
3493 SrcByte = DestByte - ShiftBytes;
3495 SrcByte = DestByte + ShiftBytes;
3497 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3498 if (SrcByte != ByteValues.size()-DestByte-1)
3501 // If the destination byte value is already defined, the values are or'd
3502 // together, which isn't a bswap (unless it's an or of the same bits).
3503 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3505 ByteValues[DestByte] = SI->getOperand(0);
3509 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3510 /// If so, insert the new bswap intrinsic and return it.
3511 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3512 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3513 if (I.getType() == Type::Int8Ty)
3516 /// ByteValues - For each byte of the result, we keep track of which value
3517 /// defines each byte.
3518 std::vector<Value*> ByteValues;
3519 ByteValues.resize(I.getType()->getPrimitiveSize());
3521 // Try to find all the pieces corresponding to the bswap.
3522 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3523 CollectBSwapParts(I.getOperand(1), ByteValues))
3526 // Check to see if all of the bytes come from the same value.
3527 Value *V = ByteValues[0];
3528 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3530 // Check to make sure that all of the bytes come from the same value.
3531 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3532 if (ByteValues[i] != V)
3535 // If they do then *success* we can turn this into a bswap. Figure out what
3536 // bswap to make it into.
3537 Module *M = I.getParent()->getParent()->getParent();
3538 const char *FnName = 0;
3539 if (I.getType() == Type::Int16Ty)
3540 FnName = "llvm.bswap.i16";
3541 else if (I.getType() == Type::Int32Ty)
3542 FnName = "llvm.bswap.i32";
3543 else if (I.getType() == Type::Int64Ty)
3544 FnName = "llvm.bswap.i64";
3546 assert(0 && "Unknown integer type!");
3547 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3548 return new CallInst(F, V);
3552 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3553 bool Changed = SimplifyCommutative(I);
3554 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3556 if (isa<UndefValue>(Op1))
3557 return ReplaceInstUsesWith(I, // X | undef -> -1
3558 ConstantInt::getAllOnesValue(I.getType()));
3562 return ReplaceInstUsesWith(I, Op0);
3564 // See if we can simplify any instructions used by the instruction whose sole
3565 // purpose is to compute bits we don't care about.
3566 uint64_t KnownZero, KnownOne;
3567 if (!isa<PackedType>(I.getType()) &&
3568 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3569 KnownZero, KnownOne))
3573 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3574 ConstantInt *C1 = 0; Value *X = 0;
3575 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3576 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3577 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3579 InsertNewInstBefore(Or, I);
3580 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3583 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3584 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3585 std::string Op0Name = Op0->getName(); Op0->setName("");
3586 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3587 InsertNewInstBefore(Or, I);
3588 return BinaryOperator::createXor(Or,
3589 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3592 // Try to fold constant and into select arguments.
3593 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3594 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3596 if (isa<PHINode>(Op0))
3597 if (Instruction *NV = FoldOpIntoPhi(I))
3601 Value *A = 0, *B = 0;
3602 ConstantInt *C1 = 0, *C2 = 0;
3604 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3605 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3606 return ReplaceInstUsesWith(I, Op1);
3607 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3608 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3609 return ReplaceInstUsesWith(I, Op0);
3611 // (A | B) | C and A | (B | C) -> bswap if possible.
3612 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3613 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3614 match(Op1, m_Or(m_Value(), m_Value())) ||
3615 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3616 match(Op1, m_Shift(m_Value(), m_Value())))) {
3617 if (Instruction *BSwap = MatchBSwap(I))
3621 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3622 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3623 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3624 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3626 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3629 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3630 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3631 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3632 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3634 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3637 // (A & C1)|(B & C2)
3638 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3639 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3641 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3642 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3645 // If we have: ((V + N) & C1) | (V & C2)
3646 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3647 // replace with V+N.
3648 if (C1 == ConstantExpr::getNot(C2)) {
3649 Value *V1 = 0, *V2 = 0;
3650 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3651 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3652 // Add commutes, try both ways.
3653 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3654 return ReplaceInstUsesWith(I, A);
3655 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3656 return ReplaceInstUsesWith(I, A);
3658 // Or commutes, try both ways.
3659 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3660 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3661 // Add commutes, try both ways.
3662 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3663 return ReplaceInstUsesWith(I, B);
3664 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3665 return ReplaceInstUsesWith(I, B);
3670 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3671 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3672 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3673 if (SI0->getOpcode() == SI1->getOpcode() &&
3674 SI0->getOperand(1) == SI1->getOperand(1) &&
3675 (SI0->hasOneUse() || SI1->hasOneUse())) {
3676 Instruction *NewOp =
3677 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3679 SI0->getName()), I);
3680 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3684 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3685 if (A == Op1) // ~A | A == -1
3686 return ReplaceInstUsesWith(I,
3687 ConstantInt::getAllOnesValue(I.getType()));
3691 // Note, A is still live here!
3692 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3694 return ReplaceInstUsesWith(I,
3695 ConstantInt::getAllOnesValue(I.getType()));
3697 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3698 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3699 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3700 I.getName()+".demorgan"), I);
3701 return BinaryOperator::createNot(And);
3705 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3706 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3707 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3710 Value *LHSVal, *RHSVal;
3711 ConstantInt *LHSCst, *RHSCst;
3712 ICmpInst::Predicate LHSCC, RHSCC;
3713 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3714 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3715 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3716 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3717 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3718 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3719 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3720 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3721 // Ensure that the larger constant is on the RHS.
3722 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3723 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3724 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3725 ICmpInst *LHS = cast<ICmpInst>(Op0);
3726 if (cast<ConstantInt>(Cmp)->getBoolValue()) {
3727 std::swap(LHS, RHS);
3728 std::swap(LHSCst, RHSCst);
3729 std::swap(LHSCC, RHSCC);
3732 // At this point, we know we have have two icmp instructions
3733 // comparing a value against two constants and or'ing the result
3734 // together. Because of the above check, we know that we only have
3735 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3736 // FoldICmpLogical check above), that the two constants are not
3738 assert(LHSCst != RHSCst && "Compares not folded above?");
3741 default: assert(0 && "Unknown integer condition code!");
3742 case ICmpInst::ICMP_EQ:
3744 default: assert(0 && "Unknown integer condition code!");
3745 case ICmpInst::ICMP_EQ:
3746 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3747 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3748 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3749 LHSVal->getName()+".off");
3750 InsertNewInstBefore(Add, I);
3751 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3752 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3754 break; // (X == 13 | X == 15) -> no change
3755 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3756 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3758 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3759 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3760 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3761 return ReplaceInstUsesWith(I, RHS);
3764 case ICmpInst::ICMP_NE:
3766 default: assert(0 && "Unknown integer condition code!");
3767 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3768 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3769 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3770 return ReplaceInstUsesWith(I, LHS);
3771 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3772 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3773 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3774 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3777 case ICmpInst::ICMP_ULT:
3779 default: assert(0 && "Unknown integer condition code!");
3780 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3782 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3783 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3785 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3787 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3788 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3789 return ReplaceInstUsesWith(I, RHS);
3790 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3794 case ICmpInst::ICMP_SLT:
3796 default: assert(0 && "Unknown integer condition code!");
3797 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3799 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3800 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3802 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3804 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3805 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3806 return ReplaceInstUsesWith(I, RHS);
3807 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3811 case ICmpInst::ICMP_UGT:
3813 default: assert(0 && "Unknown integer condition code!");
3814 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3815 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3816 return ReplaceInstUsesWith(I, LHS);
3817 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3819 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3820 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3821 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3822 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3826 case ICmpInst::ICMP_SGT:
3828 default: assert(0 && "Unknown integer condition code!");
3829 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3830 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3831 return ReplaceInstUsesWith(I, LHS);
3832 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3834 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3835 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3836 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3837 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3845 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3846 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3847 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3848 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3849 const Type *SrcTy = Op0C->getOperand(0)->getType();
3850 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3851 // Only do this if the casts both really cause code to be generated.
3852 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3854 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3856 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3857 Op1C->getOperand(0),
3859 InsertNewInstBefore(NewOp, I);
3860 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3865 return Changed ? &I : 0;
3868 // XorSelf - Implements: X ^ X --> 0
3871 XorSelf(Value *rhs) : RHS(rhs) {}
3872 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3873 Instruction *apply(BinaryOperator &Xor) const {
3879 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3880 bool Changed = SimplifyCommutative(I);
3881 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3883 if (isa<UndefValue>(Op1))
3884 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3886 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3887 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3888 assert(Result == &I && "AssociativeOpt didn't work?");
3889 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3892 // See if we can simplify any instructions used by the instruction whose sole
3893 // purpose is to compute bits we don't care about.
3894 uint64_t KnownZero, KnownOne;
3895 if (!isa<PackedType>(I.getType()) &&
3896 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3897 KnownZero, KnownOne))
3900 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3901 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3902 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3903 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3904 return new ICmpInst(ICI->getInversePredicate(),
3905 ICI->getOperand(0), ICI->getOperand(1));
3907 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3908 // ~(c-X) == X-c-1 == X+(-c-1)
3909 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3910 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3911 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3912 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3913 ConstantInt::get(I.getType(), 1));
3914 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3917 // ~(~X & Y) --> (X | ~Y)
3918 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3919 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3920 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3922 BinaryOperator::createNot(Op0I->getOperand(1),
3923 Op0I->getOperand(1)->getName()+".not");
3924 InsertNewInstBefore(NotY, I);
3925 return BinaryOperator::createOr(Op0NotVal, NotY);
3929 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3930 if (Op0I->getOpcode() == Instruction::Add) {
3931 // ~(X-c) --> (-c-1)-X
3932 if (RHS->isAllOnesValue()) {
3933 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3934 return BinaryOperator::createSub(
3935 ConstantExpr::getSub(NegOp0CI,
3936 ConstantInt::get(I.getType(), 1)),
3937 Op0I->getOperand(0));
3939 } else if (Op0I->getOpcode() == Instruction::Or) {
3940 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3941 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3942 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3943 // Anything in both C1 and C2 is known to be zero, remove it from
3945 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3946 NewRHS = ConstantExpr::getAnd(NewRHS,
3947 ConstantExpr::getNot(CommonBits));
3948 WorkList.push_back(Op0I);
3949 I.setOperand(0, Op0I->getOperand(0));
3950 I.setOperand(1, NewRHS);
3956 // Try to fold constant and into select arguments.
3957 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3958 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3960 if (isa<PHINode>(Op0))
3961 if (Instruction *NV = FoldOpIntoPhi(I))
3965 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3967 return ReplaceInstUsesWith(I,
3968 ConstantInt::getAllOnesValue(I.getType()));
3970 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3972 return ReplaceInstUsesWith(I,
3973 ConstantInt::getAllOnesValue(I.getType()));
3975 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3976 if (Op1I->getOpcode() == Instruction::Or) {
3977 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3978 Op1I->swapOperands();
3980 std::swap(Op0, Op1);
3981 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3982 I.swapOperands(); // Simplified below.
3983 std::swap(Op0, Op1);
3985 } else if (Op1I->getOpcode() == Instruction::Xor) {
3986 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3987 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3988 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3989 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3990 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3991 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3992 Op1I->swapOperands();
3993 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3994 I.swapOperands(); // Simplified below.
3995 std::swap(Op0, Op1);
3999 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
4000 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
4001 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
4002 Op0I->swapOperands();
4003 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
4004 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
4005 InsertNewInstBefore(NotB, I);
4006 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
4008 } else if (Op0I->getOpcode() == Instruction::Xor) {
4009 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
4010 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
4011 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
4012 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
4013 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
4014 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
4015 Op0I->swapOperands();
4016 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4017 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4018 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4019 InsertNewInstBefore(N, I);
4020 return BinaryOperator::createAnd(N, Op1);
4024 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4025 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4026 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4029 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4030 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4031 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4032 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4033 const Type *SrcTy = Op0C->getOperand(0)->getType();
4034 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
4035 // Only do this if the casts both really cause code to be generated.
4036 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4038 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4040 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4041 Op1C->getOperand(0),
4043 InsertNewInstBefore(NewOp, I);
4044 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4048 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4049 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
4050 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
4051 if (SI0->getOpcode() == SI1->getOpcode() &&
4052 SI0->getOperand(1) == SI1->getOperand(1) &&
4053 (SI0->hasOneUse() || SI1->hasOneUse())) {
4054 Instruction *NewOp =
4055 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4057 SI0->getName()), I);
4058 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
4062 return Changed ? &I : 0;
4065 static bool isPositive(ConstantInt *C) {
4066 return C->getSExtValue() >= 0;
4069 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4070 /// overflowed for this type.
4071 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4073 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4075 return cast<ConstantInt>(Result)->getZExtValue() <
4076 cast<ConstantInt>(In1)->getZExtValue();
4079 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4080 /// code necessary to compute the offset from the base pointer (without adding
4081 /// in the base pointer). Return the result as a signed integer of intptr size.
4082 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4083 TargetData &TD = IC.getTargetData();
4084 gep_type_iterator GTI = gep_type_begin(GEP);
4085 const Type *IntPtrTy = TD.getIntPtrType();
4086 Value *Result = Constant::getNullValue(IntPtrTy);
4088 // Build a mask for high order bits.
4089 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4091 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4092 Value *Op = GEP->getOperand(i);
4093 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4094 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4095 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4096 if (!OpC->isNullValue()) {
4097 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4098 Scale = ConstantExpr::getMul(OpC, Scale);
4099 if (Constant *RC = dyn_cast<Constant>(Result))
4100 Result = ConstantExpr::getAdd(RC, Scale);
4102 // Emit an add instruction.
4103 Result = IC.InsertNewInstBefore(
4104 BinaryOperator::createAdd(Result, Scale,
4105 GEP->getName()+".offs"), I);
4109 // Convert to correct type.
4110 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4111 Op->getName()+".c"), I);
4113 // We'll let instcombine(mul) convert this to a shl if possible.
4114 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4115 GEP->getName()+".idx"), I);
4117 // Emit an add instruction.
4118 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4119 GEP->getName()+".offs"), I);
4125 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4126 /// else. At this point we know that the GEP is on the LHS of the comparison.
4127 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4128 ICmpInst::Predicate Cond,
4130 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4132 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4133 if (isa<PointerType>(CI->getOperand(0)->getType()))
4134 RHS = CI->getOperand(0);
4136 Value *PtrBase = GEPLHS->getOperand(0);
4137 if (PtrBase == RHS) {
4138 // As an optimization, we don't actually have to compute the actual value of
4139 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4140 // each index is zero or not.
4141 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4142 Instruction *InVal = 0;
4143 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4144 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4146 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4147 if (isa<UndefValue>(C)) // undef index -> undef.
4148 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4149 if (C->isNullValue())
4151 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4152 EmitIt = false; // This is indexing into a zero sized array?
4153 } else if (isa<ConstantInt>(C))
4154 return ReplaceInstUsesWith(I, // No comparison is needed here.
4155 ConstantInt::get(Cond == ICmpInst::ICMP_NE));
4160 new ICmpInst(Cond, GEPLHS->getOperand(i),
4161 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4165 InVal = InsertNewInstBefore(InVal, I);
4166 InsertNewInstBefore(Comp, I);
4167 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4168 InVal = BinaryOperator::createOr(InVal, Comp);
4169 else // True if all are equal
4170 InVal = BinaryOperator::createAnd(InVal, Comp);
4178 // No comparison is needed here, all indexes = 0
4179 ReplaceInstUsesWith(I, ConstantInt::get(Cond == ICmpInst::ICMP_EQ));
4182 // Only lower this if the icmp is the only user of the GEP or if we expect
4183 // the result to fold to a constant!
4184 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4185 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4186 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4187 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4188 Constant::getNullValue(Offset->getType()));
4190 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4191 // If the base pointers are different, but the indices are the same, just
4192 // compare the base pointer.
4193 if (PtrBase != GEPRHS->getOperand(0)) {
4194 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4195 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4196 GEPRHS->getOperand(0)->getType();
4198 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4199 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4200 IndicesTheSame = false;
4204 // If all indices are the same, just compare the base pointers.
4206 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4207 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4209 // Otherwise, the base pointers are different and the indices are
4210 // different, bail out.
4214 // If one of the GEPs has all zero indices, recurse.
4215 bool AllZeros = true;
4216 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4217 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4218 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4223 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4224 ICmpInst::getSwappedPredicate(Cond), I);
4226 // If the other GEP has all zero indices, recurse.
4228 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4229 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4230 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4235 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4237 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4238 // If the GEPs only differ by one index, compare it.
4239 unsigned NumDifferences = 0; // Keep track of # differences.
4240 unsigned DiffOperand = 0; // The operand that differs.
4241 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4242 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4243 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4244 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4245 // Irreconcilable differences.
4249 if (NumDifferences++) break;
4254 if (NumDifferences == 0) // SAME GEP?
4255 return ReplaceInstUsesWith(I, // No comparison is needed here.
4256 ConstantInt::get(Cond == ICmpInst::ICMP_EQ));
4257 else if (NumDifferences == 1) {
4258 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4259 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4260 // Make sure we do a signed comparison here.
4261 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4265 // Only lower this if the icmp is the only user of the GEP or if we expect
4266 // the result to fold to a constant!
4267 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4268 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4269 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4270 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4271 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4272 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4278 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4279 bool Changed = SimplifyCompare(I);
4280 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4284 return ReplaceInstUsesWith(I, ConstantInt::get(isTrueWhenEqual(I)));
4286 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4287 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4289 // Handle fcmp with constant RHS
4290 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4291 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4292 switch (LHSI->getOpcode()) {
4293 case Instruction::PHI:
4294 if (Instruction *NV = FoldOpIntoPhi(I))
4297 case Instruction::Select:
4298 // If either operand of the select is a constant, we can fold the
4299 // comparison into the select arms, which will cause one to be
4300 // constant folded and the select turned into a bitwise or.
4301 Value *Op1 = 0, *Op2 = 0;
4302 if (LHSI->hasOneUse()) {
4303 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4304 // Fold the known value into the constant operand.
4305 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4306 // Insert a new FCmp of the other select operand.
4307 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4308 LHSI->getOperand(2), RHSC,
4310 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4311 // Fold the known value into the constant operand.
4312 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4313 // Insert a new FCmp of the other select operand.
4314 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4315 LHSI->getOperand(1), RHSC,
4321 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4326 return Changed ? &I : 0;
4329 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4330 bool Changed = SimplifyCompare(I);
4331 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4332 const Type *Ty = Op0->getType();
4336 return ReplaceInstUsesWith(I, ConstantInt::get(isTrueWhenEqual(I)));
4338 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4339 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4341 // icmp of GlobalValues can never equal each other as long as they aren't
4342 // external weak linkage type.
4343 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4344 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4345 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4346 return ReplaceInstUsesWith(I, ConstantInt::get(!isTrueWhenEqual(I)));
4348 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4349 // addresses never equal each other! We already know that Op0 != Op1.
4350 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4351 isa<ConstantPointerNull>(Op0)) &&
4352 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4353 isa<ConstantPointerNull>(Op1)))
4354 return ReplaceInstUsesWith(I, ConstantInt::get(!isTrueWhenEqual(I)));
4356 // icmp's with boolean values can always be turned into bitwise operations
4357 if (Ty == Type::Int1Ty) {
4358 switch (I.getPredicate()) {
4359 default: assert(0 && "Invalid icmp instruction!");
4360 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4361 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4362 InsertNewInstBefore(Xor, I);
4363 return BinaryOperator::createNot(Xor);
4365 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4366 return BinaryOperator::createXor(Op0, Op1);
4368 case ICmpInst::ICMP_UGT:
4369 case ICmpInst::ICMP_SGT:
4370 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4372 case ICmpInst::ICMP_ULT:
4373 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4374 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4375 InsertNewInstBefore(Not, I);
4376 return BinaryOperator::createAnd(Not, Op1);
4378 case ICmpInst::ICMP_UGE:
4379 case ICmpInst::ICMP_SGE:
4380 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4382 case ICmpInst::ICMP_ULE:
4383 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4384 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4385 InsertNewInstBefore(Not, I);
4386 return BinaryOperator::createOr(Not, Op1);
4391 // See if we are doing a comparison between a constant and an instruction that
4392 // can be folded into the comparison.
4393 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4394 switch (I.getPredicate()) {
4396 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4397 if (CI->isMinValue(false))
4398 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4399 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4400 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4401 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4402 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4405 case ICmpInst::ICMP_SLT:
4406 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4407 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4408 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4409 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4410 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4411 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4414 case ICmpInst::ICMP_UGT:
4415 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4416 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4417 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4418 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4419 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4420 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4423 case ICmpInst::ICMP_SGT:
4424 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4425 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4426 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4427 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4428 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4429 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4432 case ICmpInst::ICMP_ULE:
4433 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4434 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4435 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4436 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4437 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4438 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4441 case ICmpInst::ICMP_SLE:
4442 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4443 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4444 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4445 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4446 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4447 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4450 case ICmpInst::ICMP_UGE:
4451 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4452 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4453 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4454 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4455 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4456 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4459 case ICmpInst::ICMP_SGE:
4460 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4461 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4462 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4463 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4464 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4465 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4469 // If we still have a icmp le or icmp ge instruction, turn it into the
4470 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4471 // already been handled above, this requires little checking.
4473 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4474 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4475 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4476 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4477 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4478 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4479 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4480 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4482 // See if we can fold the comparison based on bits known to be zero or one
4484 uint64_t KnownZero, KnownOne;
4485 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4486 KnownZero, KnownOne, 0))
4489 // Given the known and unknown bits, compute a range that the LHS could be
4491 if (KnownOne | KnownZero) {
4492 // Compute the Min, Max and RHS values based on the known bits. For the
4493 // EQ and NE we use unsigned values.
4494 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4495 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4496 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4497 SRHSVal = CI->getSExtValue();
4498 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4501 URHSVal = CI->getZExtValue();
4502 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4505 switch (I.getPredicate()) { // LE/GE have been folded already.
4506 default: assert(0 && "Unknown icmp opcode!");
4507 case ICmpInst::ICMP_EQ:
4508 if (UMax < URHSVal || UMin > URHSVal)
4509 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4511 case ICmpInst::ICMP_NE:
4512 if (UMax < URHSVal || UMin > URHSVal)
4513 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4515 case ICmpInst::ICMP_ULT:
4517 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4519 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4521 case ICmpInst::ICMP_UGT:
4523 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4525 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4527 case ICmpInst::ICMP_SLT:
4529 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4531 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4533 case ICmpInst::ICMP_SGT:
4535 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4537 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4542 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4543 // instruction, see if that instruction also has constants so that the
4544 // instruction can be folded into the icmp
4545 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4546 switch (LHSI->getOpcode()) {
4547 case Instruction::And:
4548 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4549 LHSI->getOperand(0)->hasOneUse()) {
4550 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4552 // If the LHS is an AND of a truncating cast, we can widen the
4553 // and/compare to be the input width without changing the value
4554 // produced, eliminating a cast.
4555 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4556 // We can do this transformation if either the AND constant does not
4557 // have its sign bit set or if it is an equality comparison.
4558 // Extending a relational comparison when we're checking the sign
4559 // bit would not work.
4560 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4562 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4563 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4564 ConstantInt *NewCST;
4566 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4567 AndCST->getZExtValue());
4568 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4569 CI->getZExtValue());
4570 Instruction *NewAnd =
4571 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4573 InsertNewInstBefore(NewAnd, I);
4574 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4578 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4579 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4580 // happens a LOT in code produced by the C front-end, for bitfield
4582 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4584 // Check to see if there is a noop-cast between the shift and the and.
4586 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4587 if (CI->getOpcode() == Instruction::BitCast)
4588 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4592 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4593 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4594 const Type *AndTy = AndCST->getType(); // Type of the and.
4596 // We can fold this as long as we can't shift unknown bits
4597 // into the mask. This can only happen with signed shift
4598 // rights, as they sign-extend.
4600 bool CanFold = Shift->isLogicalShift();
4602 // To test for the bad case of the signed shr, see if any
4603 // of the bits shifted in could be tested after the mask.
4604 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4605 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4607 Constant *OShAmt = ConstantInt::get(Type::Int8Ty, ShAmtVal);
4609 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4611 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4617 if (Shift->getOpcode() == Instruction::Shl)
4618 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4620 NewCst = ConstantExpr::getShl(CI, ShAmt);
4622 // Check to see if we are shifting out any of the bits being
4624 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4625 // If we shifted bits out, the fold is not going to work out.
4626 // As a special case, check to see if this means that the
4627 // result is always true or false now.
4628 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4629 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4630 if (I.getPredicate() == ICmpInst::ICMP_NE)
4631 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4633 I.setOperand(1, NewCst);
4634 Constant *NewAndCST;
4635 if (Shift->getOpcode() == Instruction::Shl)
4636 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4638 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4639 LHSI->setOperand(1, NewAndCST);
4640 LHSI->setOperand(0, Shift->getOperand(0));
4641 WorkList.push_back(Shift); // Shift is dead.
4642 AddUsesToWorkList(I);
4648 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4649 // preferable because it allows the C<<Y expression to be hoisted out
4650 // of a loop if Y is invariant and X is not.
4651 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4652 I.isEquality() && !Shift->isArithmeticShift() &&
4653 isa<Instruction>(Shift->getOperand(0))) {
4656 if (Shift->getOpcode() == Instruction::LShr) {
4657 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4660 // Insert a logical shift.
4661 NS = new ShiftInst(Instruction::LShr, AndCST,
4662 Shift->getOperand(1), "tmp");
4664 InsertNewInstBefore(cast<Instruction>(NS), I);
4666 // Compute X & (C << Y).
4667 Instruction *NewAnd = BinaryOperator::createAnd(
4668 Shift->getOperand(0), NS, LHSI->getName());
4669 InsertNewInstBefore(NewAnd, I);
4671 I.setOperand(0, NewAnd);
4677 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4678 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4679 if (I.isEquality()) {
4680 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4682 // Check that the shift amount is in range. If not, don't perform
4683 // undefined shifts. When the shift is visited it will be
4685 if (ShAmt->getZExtValue() >= TypeBits)
4688 // If we are comparing against bits always shifted out, the
4689 // comparison cannot succeed.
4691 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4692 if (Comp != CI) {// Comparing against a bit that we know is zero.
4693 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4694 Constant *Cst = ConstantInt::get(IsICMP_NE);
4695 return ReplaceInstUsesWith(I, Cst);
4698 if (LHSI->hasOneUse()) {
4699 // Otherwise strength reduce the shift into an and.
4700 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4701 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4702 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4705 BinaryOperator::createAnd(LHSI->getOperand(0),
4706 Mask, LHSI->getName()+".mask");
4707 Value *And = InsertNewInstBefore(AndI, I);
4708 return new ICmpInst(I.getPredicate(), And,
4709 ConstantExpr::getLShr(CI, ShAmt));
4715 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4716 case Instruction::AShr:
4717 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4718 if (I.isEquality()) {
4719 // Check that the shift amount is in range. If not, don't perform
4720 // undefined shifts. When the shift is visited it will be
4722 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4723 if (ShAmt->getZExtValue() >= TypeBits)
4726 // If we are comparing against bits always shifted out, the
4727 // comparison cannot succeed.
4729 if (LHSI->getOpcode() == Instruction::LShr)
4730 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4733 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4736 if (Comp != CI) {// Comparing against a bit that we know is zero.
4737 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4738 Constant *Cst = ConstantInt::get(IsICMP_NE);
4739 return ReplaceInstUsesWith(I, Cst);
4742 if (LHSI->hasOneUse() || CI->isNullValue()) {
4743 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4745 // Otherwise strength reduce the shift into an and.
4746 uint64_t Val = ~0ULL; // All ones.
4747 Val <<= ShAmtVal; // Shift over to the right spot.
4748 Val &= ~0ULL >> (64-TypeBits);
4749 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4752 BinaryOperator::createAnd(LHSI->getOperand(0),
4753 Mask, LHSI->getName()+".mask");
4754 Value *And = InsertNewInstBefore(AndI, I);
4755 return new ICmpInst(I.getPredicate(), And,
4756 ConstantExpr::getShl(CI, ShAmt));
4762 case Instruction::SDiv:
4763 case Instruction::UDiv:
4764 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4765 // Fold this div into the comparison, producing a range check.
4766 // Determine, based on the divide type, what the range is being
4767 // checked. If there is an overflow on the low or high side, remember
4768 // it, otherwise compute the range [low, hi) bounding the new value.
4769 // See: InsertRangeTest above for the kinds of replacements possible.
4770 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4771 // FIXME: If the operand types don't match the type of the divide
4772 // then don't attempt this transform. The code below doesn't have the
4773 // logic to deal with a signed divide and an unsigned compare (and
4774 // vice versa). This is because (x /s C1) <s C2 produces different
4775 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4776 // (x /u C1) <u C2. Simply casting the operands and result won't
4777 // work. :( The if statement below tests that condition and bails
4779 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4780 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4783 // Initialize the variables that will indicate the nature of the
4785 bool LoOverflow = false, HiOverflow = false;
4786 ConstantInt *LoBound = 0, *HiBound = 0;
4788 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4789 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4790 // C2 (CI). By solving for X we can turn this into a range check
4791 // instead of computing a divide.
4793 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4795 // Determine if the product overflows by seeing if the product is
4796 // not equal to the divide. Make sure we do the same kind of divide
4797 // as in the LHS instruction that we're folding.
4798 bool ProdOV = !DivRHS->isNullValue() &&
4799 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4800 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4802 // Get the ICmp opcode
4803 ICmpInst::Predicate predicate = I.getPredicate();
4805 if (DivRHS->isNullValue()) {
4806 // Don't hack on divide by zeros!
4807 } else if (!DivIsSigned) { // udiv
4809 LoOverflow = ProdOV;
4810 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4811 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4812 if (CI->isNullValue()) { // (X / pos) op 0
4814 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4816 } else if (isPositive(CI)) { // (X / pos) op pos
4818 LoOverflow = ProdOV;
4819 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4820 } else { // (X / pos) op neg
4821 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4822 LoOverflow = AddWithOverflow(LoBound, Prod,
4823 cast<ConstantInt>(DivRHSH));
4825 HiOverflow = ProdOV;
4827 } else { // Divisor is < 0.
4828 if (CI->isNullValue()) { // (X / neg) op 0
4829 LoBound = AddOne(DivRHS);
4830 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4831 if (HiBound == DivRHS)
4832 LoBound = 0; // - INTMIN = INTMIN
4833 } else if (isPositive(CI)) { // (X / neg) op pos
4834 HiOverflow = LoOverflow = ProdOV;
4836 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4837 HiBound = AddOne(Prod);
4838 } else { // (X / neg) op neg
4840 LoOverflow = HiOverflow = ProdOV;
4841 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4844 // Dividing by a negate swaps the condition.
4845 predicate = ICmpInst::getSwappedPredicate(predicate);
4849 Value *X = LHSI->getOperand(0);
4850 switch (predicate) {
4851 default: assert(0 && "Unhandled icmp opcode!");
4852 case ICmpInst::ICMP_EQ:
4853 if (LoOverflow && HiOverflow)
4854 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4855 else if (HiOverflow)
4856 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4857 ICmpInst::ICMP_UGE, X, LoBound);
4858 else if (LoOverflow)
4859 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4860 ICmpInst::ICMP_ULT, X, HiBound);
4862 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4864 case ICmpInst::ICMP_NE:
4865 if (LoOverflow && HiOverflow)
4866 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4867 else if (HiOverflow)
4868 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4869 ICmpInst::ICMP_ULT, X, LoBound);
4870 else if (LoOverflow)
4871 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4872 ICmpInst::ICMP_UGE, X, HiBound);
4874 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4876 case ICmpInst::ICMP_ULT:
4877 case ICmpInst::ICMP_SLT:
4879 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4880 return new ICmpInst(predicate, X, LoBound);
4881 case ICmpInst::ICMP_UGT:
4882 case ICmpInst::ICMP_SGT:
4884 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4885 if (predicate == ICmpInst::ICMP_UGT)
4886 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4888 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4895 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4896 if (I.isEquality()) {
4897 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4899 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4900 // the second operand is a constant, simplify a bit.
4901 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4902 switch (BO->getOpcode()) {
4903 case Instruction::SRem:
4904 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4905 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4907 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4908 if (V > 1 && isPowerOf2_64(V)) {
4909 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4910 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4911 return new ICmpInst(I.getPredicate(), NewRem,
4912 Constant::getNullValue(BO->getType()));
4916 case Instruction::Add:
4917 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4918 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4919 if (BO->hasOneUse())
4920 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4921 ConstantExpr::getSub(CI, BOp1C));
4922 } else if (CI->isNullValue()) {
4923 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4924 // efficiently invertible, or if the add has just this one use.
4925 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4927 if (Value *NegVal = dyn_castNegVal(BOp1))
4928 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4929 else if (Value *NegVal = dyn_castNegVal(BOp0))
4930 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4931 else if (BO->hasOneUse()) {
4932 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4934 InsertNewInstBefore(Neg, I);
4935 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4939 case Instruction::Xor:
4940 // For the xor case, we can xor two constants together, eliminating
4941 // the explicit xor.
4942 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4943 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4944 ConstantExpr::getXor(CI, BOC));
4947 case Instruction::Sub:
4948 // Replace (([sub|xor] A, B) != 0) with (A != B)
4949 if (CI->isNullValue())
4950 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4954 case Instruction::Or:
4955 // If bits are being or'd in that are not present in the constant we
4956 // are comparing against, then the comparison could never succeed!
4957 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4958 Constant *NotCI = ConstantExpr::getNot(CI);
4959 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4960 return ReplaceInstUsesWith(I, ConstantInt::get(isICMP_NE));
4964 case Instruction::And:
4965 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4966 // If bits are being compared against that are and'd out, then the
4967 // comparison can never succeed!
4968 if (!ConstantExpr::getAnd(CI,
4969 ConstantExpr::getNot(BOC))->isNullValue())
4970 return ReplaceInstUsesWith(I, ConstantInt::get(isICMP_NE));
4972 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4973 if (CI == BOC && isOneBitSet(CI))
4974 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
4975 ICmpInst::ICMP_NE, Op0,
4976 Constant::getNullValue(CI->getType()));
4978 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
4979 if (isSignBit(BOC)) {
4980 Value *X = BO->getOperand(0);
4981 Constant *Zero = Constant::getNullValue(X->getType());
4982 ICmpInst::Predicate pred = isICMP_NE ?
4983 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
4984 return new ICmpInst(pred, X, Zero);
4987 // ((X & ~7) == 0) --> X < 8
4988 if (CI->isNullValue() && isHighOnes(BOC)) {
4989 Value *X = BO->getOperand(0);
4990 Constant *NegX = ConstantExpr::getNeg(BOC);
4991 ICmpInst::Predicate pred = isICMP_NE ?
4992 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4993 return new ICmpInst(pred, X, NegX);
4999 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5000 // Handle set{eq|ne} <intrinsic>, intcst.
5001 switch (II->getIntrinsicID()) {
5003 case Intrinsic::bswap_i16:
5004 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5005 WorkList.push_back(II); // Dead?
5006 I.setOperand(0, II->getOperand(1));
5007 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5008 ByteSwap_16(CI->getZExtValue())));
5010 case Intrinsic::bswap_i32:
5011 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5012 WorkList.push_back(II); // Dead?
5013 I.setOperand(0, II->getOperand(1));
5014 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5015 ByteSwap_32(CI->getZExtValue())));
5017 case Intrinsic::bswap_i64:
5018 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5019 WorkList.push_back(II); // Dead?
5020 I.setOperand(0, II->getOperand(1));
5021 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5022 ByteSwap_64(CI->getZExtValue())));
5026 } else { // Not a ICMP_EQ/ICMP_NE
5027 // If the LHS is a cast from an integral value of the same size, then
5028 // since we know the RHS is a constant, try to simlify.
5029 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5030 Value *CastOp = Cast->getOperand(0);
5031 const Type *SrcTy = CastOp->getType();
5032 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5033 if (SrcTy->isInteger() &&
5034 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5035 // If this is an unsigned comparison, try to make the comparison use
5036 // smaller constant values.
5037 switch (I.getPredicate()) {
5039 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5040 ConstantInt *CUI = cast<ConstantInt>(CI);
5041 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5042 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5043 ConstantInt::get(SrcTy, -1));
5046 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5047 ConstantInt *CUI = cast<ConstantInt>(CI);
5048 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5049 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5050 Constant::getNullValue(SrcTy));
5060 // Handle icmp with constant RHS
5061 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5062 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5063 switch (LHSI->getOpcode()) {
5064 case Instruction::GetElementPtr:
5065 if (RHSC->isNullValue()) {
5066 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5067 bool isAllZeros = true;
5068 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5069 if (!isa<Constant>(LHSI->getOperand(i)) ||
5070 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5075 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5076 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5080 case Instruction::PHI:
5081 if (Instruction *NV = FoldOpIntoPhi(I))
5084 case Instruction::Select:
5085 // If either operand of the select is a constant, we can fold the
5086 // comparison into the select arms, which will cause one to be
5087 // constant folded and the select turned into a bitwise or.
5088 Value *Op1 = 0, *Op2 = 0;
5089 if (LHSI->hasOneUse()) {
5090 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5091 // Fold the known value into the constant operand.
5092 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5093 // Insert a new ICmp of the other select operand.
5094 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5095 LHSI->getOperand(2), RHSC,
5097 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5098 // Fold the known value into the constant operand.
5099 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5100 // Insert a new ICmp of the other select operand.
5101 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5102 LHSI->getOperand(1), RHSC,
5108 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5113 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5114 if (User *GEP = dyn_castGetElementPtr(Op0))
5115 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5117 if (User *GEP = dyn_castGetElementPtr(Op1))
5118 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5119 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5122 // Test to see if the operands of the icmp are casted versions of other
5123 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5125 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5126 if (isa<PointerType>(Op0->getType()) &&
5127 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5128 // We keep moving the cast from the left operand over to the right
5129 // operand, where it can often be eliminated completely.
5130 Op0 = CI->getOperand(0);
5132 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5133 // so eliminate it as well.
5134 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5135 Op1 = CI2->getOperand(0);
5137 // If Op1 is a constant, we can fold the cast into the constant.
5138 if (Op0->getType() != Op1->getType())
5139 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5140 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5142 // Otherwise, cast the RHS right before the icmp
5143 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5145 return new ICmpInst(I.getPredicate(), Op0, Op1);
5149 if (isa<CastInst>(Op0)) {
5150 // Handle the special case of: icmp (cast bool to X), <cst>
5151 // This comes up when you have code like
5154 // For generality, we handle any zero-extension of any operand comparison
5155 // with a constant or another cast from the same type.
5156 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5157 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5161 if (I.isEquality()) {
5162 Value *A, *B, *C, *D;
5163 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5164 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5165 Value *OtherVal = A == Op1 ? B : A;
5166 return new ICmpInst(I.getPredicate(), OtherVal,
5167 Constant::getNullValue(A->getType()));
5170 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5171 // A^c1 == C^c2 --> A == C^(c1^c2)
5172 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5173 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5174 if (Op1->hasOneUse()) {
5175 Constant *NC = ConstantExpr::getXor(C1, C2);
5176 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5177 return new ICmpInst(I.getPredicate(), A,
5178 InsertNewInstBefore(Xor, I));
5181 // A^B == A^D -> B == D
5182 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5183 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5184 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5185 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5189 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5190 (A == Op0 || B == Op0)) {
5191 // A == (A^B) -> B == 0
5192 Value *OtherVal = A == Op0 ? B : A;
5193 return new ICmpInst(I.getPredicate(), OtherVal,
5194 Constant::getNullValue(A->getType()));
5196 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5197 // (A-B) == A -> B == 0
5198 return new ICmpInst(I.getPredicate(), B,
5199 Constant::getNullValue(B->getType()));
5201 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5202 // A == (A-B) -> B == 0
5203 return new ICmpInst(I.getPredicate(), B,
5204 Constant::getNullValue(B->getType()));
5207 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5208 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5209 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5210 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5211 Value *X = 0, *Y = 0, *Z = 0;
5214 X = B; Y = D; Z = A;
5215 } else if (A == D) {
5216 X = B; Y = C; Z = A;
5217 } else if (B == C) {
5218 X = A; Y = D; Z = B;
5219 } else if (B == D) {
5220 X = A; Y = C; Z = B;
5223 if (X) { // Build (X^Y) & Z
5224 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5225 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5226 I.setOperand(0, Op1);
5227 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5232 return Changed ? &I : 0;
5235 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5236 // We only handle extending casts so far.
5238 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5239 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5240 Value *LHSCIOp = LHSCI->getOperand(0);
5241 const Type *SrcTy = LHSCIOp->getType();
5242 const Type *DestTy = LHSCI->getType();
5245 // We only handle extension cast instructions, so far. Enforce this.
5246 if (LHSCI->getOpcode() != Instruction::ZExt &&
5247 LHSCI->getOpcode() != Instruction::SExt)
5250 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5251 bool isSignedCmp = ICI.isSignedPredicate();
5253 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5254 // Not an extension from the same type?
5255 RHSCIOp = CI->getOperand(0);
5256 if (RHSCIOp->getType() != LHSCIOp->getType())
5259 // Okay, just insert a compare of the reduced operands now!
5260 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5263 // If we aren't dealing with a constant on the RHS, exit early
5264 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5268 // Compute the constant that would happen if we truncated to SrcTy then
5269 // reextended to DestTy.
5270 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5271 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5273 // If the re-extended constant didn't change...
5275 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5276 // For example, we might have:
5277 // %A = sext short %X to uint
5278 // %B = icmp ugt uint %A, 1330
5279 // It is incorrect to transform this into
5280 // %B = icmp ugt short %X, 1330
5281 // because %A may have negative value.
5283 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5284 // OR operation is EQ/NE.
5285 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5286 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5291 // The re-extended constant changed so the constant cannot be represented
5292 // in the shorter type. Consequently, we cannot emit a simple comparison.
5294 // First, handle some easy cases. We know the result cannot be equal at this
5295 // point so handle the ICI.isEquality() cases
5296 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5297 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5298 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5299 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5301 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5302 // should have been folded away previously and not enter in here.
5305 // We're performing a signed comparison.
5306 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5307 Result = ConstantInt::getFalse(); // X < (small) --> false
5309 Result = ConstantInt::getTrue(); // X < (large) --> true
5311 // We're performing an unsigned comparison.
5313 // We're performing an unsigned comp with a sign extended value.
5314 // This is true if the input is >= 0. [aka >s -1]
5315 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5316 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5317 NegOne, ICI.getName()), ICI);
5319 // Unsigned extend & unsigned compare -> always true.
5320 Result = ConstantInt::getTrue();
5324 // Finally, return the value computed.
5325 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5326 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5327 return ReplaceInstUsesWith(ICI, Result);
5329 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5330 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5331 "ICmp should be folded!");
5332 if (Constant *CI = dyn_cast<Constant>(Result))
5333 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5335 return BinaryOperator::createNot(Result);
5339 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5340 assert(I.getOperand(1)->getType() == Type::Int8Ty);
5341 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5343 // shl X, 0 == X and shr X, 0 == X
5344 // shl 0, X == 0 and shr 0, X == 0
5345 if (Op1 == Constant::getNullValue(Type::Int8Ty) ||
5346 Op0 == Constant::getNullValue(Op0->getType()))
5347 return ReplaceInstUsesWith(I, Op0);
5349 if (isa<UndefValue>(Op0)) {
5350 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5351 return ReplaceInstUsesWith(I, Op0);
5352 else // undef << X -> 0, undef >>u X -> 0
5353 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5355 if (isa<UndefValue>(Op1)) {
5356 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5357 return ReplaceInstUsesWith(I, Op0);
5358 else // X << undef, X >>u undef -> 0
5359 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5362 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5363 if (I.getOpcode() == Instruction::AShr)
5364 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5365 if (CSI->isAllOnesValue())
5366 return ReplaceInstUsesWith(I, CSI);
5368 // Try to fold constant and into select arguments.
5369 if (isa<Constant>(Op0))
5370 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5371 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5374 // See if we can turn a signed shr into an unsigned shr.
5375 if (I.isArithmeticShift()) {
5376 if (MaskedValueIsZero(Op0,
5377 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5378 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5382 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5383 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5388 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5390 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5391 bool isSignedShift = I.getOpcode() == Instruction::AShr;
5392 bool isUnsignedShift = !isSignedShift;
5394 // See if we can simplify any instructions used by the instruction whose sole
5395 // purpose is to compute bits we don't care about.
5396 uint64_t KnownZero, KnownOne;
5397 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5398 KnownZero, KnownOne))
5401 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5402 // of a signed value.
5404 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5405 if (Op1->getZExtValue() >= TypeBits) {
5406 if (isUnsignedShift || isLeftShift)
5407 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5409 I.setOperand(1, ConstantInt::get(Type::Int8Ty, TypeBits-1));
5414 // ((X*C1) << C2) == (X * (C1 << C2))
5415 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5416 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5417 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5418 return BinaryOperator::createMul(BO->getOperand(0),
5419 ConstantExpr::getShl(BOOp, Op1));
5421 // Try to fold constant and into select arguments.
5422 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5423 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5425 if (isa<PHINode>(Op0))
5426 if (Instruction *NV = FoldOpIntoPhi(I))
5429 if (Op0->hasOneUse()) {
5430 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5431 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5434 switch (Op0BO->getOpcode()) {
5436 case Instruction::Add:
5437 case Instruction::And:
5438 case Instruction::Or:
5439 case Instruction::Xor:
5440 // These operators commute.
5441 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5442 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5443 match(Op0BO->getOperand(1),
5444 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5445 Instruction *YS = new ShiftInst(Instruction::Shl,
5446 Op0BO->getOperand(0), Op1,
5448 InsertNewInstBefore(YS, I); // (Y << C)
5450 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5451 Op0BO->getOperand(1)->getName());
5452 InsertNewInstBefore(X, I); // (X + (Y << C))
5453 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5454 C2 = ConstantExpr::getShl(C2, Op1);
5455 return BinaryOperator::createAnd(X, C2);
5458 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5459 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5460 match(Op0BO->getOperand(1),
5461 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5462 m_ConstantInt(CC))) && V2 == Op1 &&
5463 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5464 Instruction *YS = new ShiftInst(Instruction::Shl,
5465 Op0BO->getOperand(0), Op1,
5467 InsertNewInstBefore(YS, I); // (Y << C)
5469 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5470 V1->getName()+".mask");
5471 InsertNewInstBefore(XM, I); // X & (CC << C)
5473 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5477 case Instruction::Sub:
5478 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5479 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5480 match(Op0BO->getOperand(0),
5481 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5482 Instruction *YS = new ShiftInst(Instruction::Shl,
5483 Op0BO->getOperand(1), Op1,
5485 InsertNewInstBefore(YS, I); // (Y << C)
5487 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5488 Op0BO->getOperand(0)->getName());
5489 InsertNewInstBefore(X, I); // (X + (Y << C))
5490 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5491 C2 = ConstantExpr::getShl(C2, Op1);
5492 return BinaryOperator::createAnd(X, C2);
5495 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5496 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5497 match(Op0BO->getOperand(0),
5498 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5499 m_ConstantInt(CC))) && V2 == Op1 &&
5500 cast<BinaryOperator>(Op0BO->getOperand(0))
5501 ->getOperand(0)->hasOneUse()) {
5502 Instruction *YS = new ShiftInst(Instruction::Shl,
5503 Op0BO->getOperand(1), Op1,
5505 InsertNewInstBefore(YS, I); // (Y << C)
5507 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5508 V1->getName()+".mask");
5509 InsertNewInstBefore(XM, I); // X & (CC << C)
5511 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5518 // If the operand is an bitwise operator with a constant RHS, and the
5519 // shift is the only use, we can pull it out of the shift.
5520 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5521 bool isValid = true; // Valid only for And, Or, Xor
5522 bool highBitSet = false; // Transform if high bit of constant set?
5524 switch (Op0BO->getOpcode()) {
5525 default: isValid = false; break; // Do not perform transform!
5526 case Instruction::Add:
5527 isValid = isLeftShift;
5529 case Instruction::Or:
5530 case Instruction::Xor:
5533 case Instruction::And:
5538 // If this is a signed shift right, and the high bit is modified
5539 // by the logical operation, do not perform the transformation.
5540 // The highBitSet boolean indicates the value of the high bit of
5541 // the constant which would cause it to be modified for this
5544 if (isValid && !isLeftShift && isSignedShift) {
5545 uint64_t Val = Op0C->getZExtValue();
5546 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5550 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5552 Instruction *NewShift =
5553 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5556 InsertNewInstBefore(NewShift, I);
5558 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5565 // Find out if this is a shift of a shift by a constant.
5566 ShiftInst *ShiftOp = 0;
5567 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5569 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5570 // If this is a noop-integer cast of a shift instruction, use the shift.
5571 if (isa<ShiftInst>(CI->getOperand(0))) {
5572 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5576 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5577 // Find the operands and properties of the input shift. Note that the
5578 // signedness of the input shift may differ from the current shift if there
5579 // is a noop cast between the two.
5580 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5581 bool isShiftOfSignedShift = ShiftOp->getOpcode() == Instruction::AShr;
5582 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5584 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5586 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5587 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5589 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5590 if (isLeftShift == isShiftOfLeftShift) {
5591 // Do not fold these shifts if the first one is signed and the second one
5592 // is unsigned and this is a right shift. Further, don't do any folding
5594 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5597 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5598 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5599 Amt = Op0->getType()->getPrimitiveSizeInBits();
5601 Value *Op = ShiftOp->getOperand(0);
5602 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5603 ConstantInt::get(Type::Int8Ty, Amt));
5604 if (I.getType() == ShiftResult->getType())
5606 InsertNewInstBefore(ShiftResult, I);
5607 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5610 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5611 // signed types, we can only support the (A >> c1) << c2 configuration,
5612 // because it can not turn an arbitrary bit of A into a sign bit.
5613 if (isUnsignedShift || isLeftShift) {
5614 // Calculate bitmask for what gets shifted off the edge.
5615 Constant *C = ConstantInt::getAllOnesValue(I.getType());
5617 C = ConstantExpr::getShl(C, ShiftAmt1C);
5619 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5621 Value *Op = ShiftOp->getOperand(0);
5624 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5625 InsertNewInstBefore(Mask, I);
5627 // Figure out what flavor of shift we should use...
5628 if (ShiftAmt1 == ShiftAmt2) {
5629 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5630 } else if (ShiftAmt1 < ShiftAmt2) {
5631 return new ShiftInst(I.getOpcode(), Mask,
5632 ConstantInt::get(Type::Int8Ty, ShiftAmt2-ShiftAmt1));
5633 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5634 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5635 return new ShiftInst(Instruction::LShr, Mask,
5636 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5638 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5639 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5642 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5643 Instruction *Shift =
5644 new ShiftInst(ShiftOp->getOpcode(), Mask,
5645 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5646 InsertNewInstBefore(Shift, I);
5648 C = ConstantInt::getAllOnesValue(Shift->getType());
5649 C = ConstantExpr::getShl(C, Op1);
5650 return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5653 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5654 // this case, C1 == C2 and C1 is 8, 16, or 32.
5655 if (ShiftAmt1 == ShiftAmt2) {
5656 const Type *SExtType = 0;
5657 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5658 case 8 : SExtType = Type::Int8Ty; break;
5659 case 16: SExtType = Type::Int16Ty; break;
5660 case 32: SExtType = Type::Int32Ty; break;
5664 Instruction *NewTrunc =
5665 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5666 InsertNewInstBefore(NewTrunc, I);
5667 return new SExtInst(NewTrunc, I.getType());
5676 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5677 /// expression. If so, decompose it, returning some value X, such that Val is
5680 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5682 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5683 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5684 Offset = CI->getZExtValue();
5686 return ConstantInt::get(Type::Int32Ty, 0);
5687 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5688 if (I->getNumOperands() == 2) {
5689 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5690 if (I->getOpcode() == Instruction::Shl) {
5691 // This is a value scaled by '1 << the shift amt'.
5692 Scale = 1U << CUI->getZExtValue();
5694 return I->getOperand(0);
5695 } else if (I->getOpcode() == Instruction::Mul) {
5696 // This value is scaled by 'CUI'.
5697 Scale = CUI->getZExtValue();
5699 return I->getOperand(0);
5700 } else if (I->getOpcode() == Instruction::Add) {
5701 // We have X+C. Check to see if we really have (X*C2)+C1,
5702 // where C1 is divisible by C2.
5705 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5706 Offset += CUI->getZExtValue();
5707 if (SubScale > 1 && (Offset % SubScale == 0)) {
5716 // Otherwise, we can't look past this.
5723 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5724 /// try to eliminate the cast by moving the type information into the alloc.
5725 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5726 AllocationInst &AI) {
5727 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5728 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5730 // Remove any uses of AI that are dead.
5731 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5732 std::vector<Instruction*> DeadUsers;
5733 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5734 Instruction *User = cast<Instruction>(*UI++);
5735 if (isInstructionTriviallyDead(User)) {
5736 while (UI != E && *UI == User)
5737 ++UI; // If this instruction uses AI more than once, don't break UI.
5739 // Add operands to the worklist.
5740 AddUsesToWorkList(*User);
5742 DOUT << "IC: DCE: " << *User;
5744 User->eraseFromParent();
5745 removeFromWorkList(User);
5749 // Get the type really allocated and the type casted to.
5750 const Type *AllocElTy = AI.getAllocatedType();
5751 const Type *CastElTy = PTy->getElementType();
5752 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5754 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5755 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5756 if (CastElTyAlign < AllocElTyAlign) return 0;
5758 // If the allocation has multiple uses, only promote it if we are strictly
5759 // increasing the alignment of the resultant allocation. If we keep it the
5760 // same, we open the door to infinite loops of various kinds.
5761 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5763 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5764 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5765 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5767 // See if we can satisfy the modulus by pulling a scale out of the array
5769 unsigned ArraySizeScale, ArrayOffset;
5770 Value *NumElements = // See if the array size is a decomposable linear expr.
5771 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5773 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5775 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5776 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5778 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5783 // If the allocation size is constant, form a constant mul expression
5784 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5785 if (isa<ConstantInt>(NumElements))
5786 Amt = ConstantExpr::getMul(
5787 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5788 // otherwise multiply the amount and the number of elements
5789 else if (Scale != 1) {
5790 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5791 Amt = InsertNewInstBefore(Tmp, AI);
5795 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5796 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5797 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5798 Amt = InsertNewInstBefore(Tmp, AI);
5801 std::string Name = AI.getName(); AI.setName("");
5802 AllocationInst *New;
5803 if (isa<MallocInst>(AI))
5804 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5806 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5807 InsertNewInstBefore(New, AI);
5809 // If the allocation has multiple uses, insert a cast and change all things
5810 // that used it to use the new cast. This will also hack on CI, but it will
5812 if (!AI.hasOneUse()) {
5813 AddUsesToWorkList(AI);
5814 // New is the allocation instruction, pointer typed. AI is the original
5815 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5816 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5817 InsertNewInstBefore(NewCast, AI);
5818 AI.replaceAllUsesWith(NewCast);
5820 return ReplaceInstUsesWith(CI, New);
5823 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5824 /// and return it without inserting any new casts. This is used by code that
5825 /// tries to decide whether promoting or shrinking integer operations to wider
5826 /// or smaller types will allow us to eliminate a truncate or extend.
5827 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5828 int &NumCastsRemoved) {
5829 if (isa<Constant>(V)) return true;
5831 Instruction *I = dyn_cast<Instruction>(V);
5832 if (!I || !I->hasOneUse()) return false;
5834 switch (I->getOpcode()) {
5835 case Instruction::And:
5836 case Instruction::Or:
5837 case Instruction::Xor:
5838 // These operators can all arbitrarily be extended or truncated.
5839 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5840 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5841 case Instruction::AShr:
5842 case Instruction::LShr:
5843 case Instruction::Shl:
5844 // If this is just a bitcast changing the sign of the operation, we can
5845 // convert if the operand can be converted.
5846 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5847 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5849 case Instruction::Trunc:
5850 case Instruction::ZExt:
5851 case Instruction::SExt:
5852 case Instruction::BitCast:
5853 // If this is a cast from the destination type, we can trivially eliminate
5854 // it, and this will remove a cast overall.
5855 if (I->getOperand(0)->getType() == Ty) {
5856 // If the first operand is itself a cast, and is eliminable, do not count
5857 // this as an eliminable cast. We would prefer to eliminate those two
5859 if (isa<CastInst>(I->getOperand(0)))
5867 // TODO: Can handle more cases here.
5874 /// EvaluateInDifferentType - Given an expression that
5875 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5876 /// evaluate the expression.
5877 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5879 if (Constant *C = dyn_cast<Constant>(V))
5880 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5882 // Otherwise, it must be an instruction.
5883 Instruction *I = cast<Instruction>(V);
5884 Instruction *Res = 0;
5885 switch (I->getOpcode()) {
5886 case Instruction::And:
5887 case Instruction::Or:
5888 case Instruction::Xor: {
5889 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5890 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5891 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5892 LHS, RHS, I->getName());
5895 case Instruction::AShr:
5896 case Instruction::LShr:
5897 case Instruction::Shl: {
5898 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5899 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5900 I->getOperand(1), I->getName());
5903 case Instruction::Trunc:
5904 case Instruction::ZExt:
5905 case Instruction::SExt:
5906 case Instruction::BitCast:
5907 // If the source type of the cast is the type we're trying for then we can
5908 // just return the source. There's no need to insert it because its not new.
5909 if (I->getOperand(0)->getType() == Ty)
5910 return I->getOperand(0);
5912 // Some other kind of cast, which shouldn't happen, so just ..
5915 // TODO: Can handle more cases here.
5916 assert(0 && "Unreachable!");
5920 return InsertNewInstBefore(Res, *I);
5923 /// @brief Implement the transforms common to all CastInst visitors.
5924 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5925 Value *Src = CI.getOperand(0);
5927 // Casting undef to anything results in undef so might as just replace it and
5928 // get rid of the cast.
5929 if (isa<UndefValue>(Src)) // cast undef -> undef
5930 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5932 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5933 // eliminate it now.
5934 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5935 if (Instruction::CastOps opc =
5936 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5937 // The first cast (CSrc) is eliminable so we need to fix up or replace
5938 // the second cast (CI). CSrc will then have a good chance of being dead.
5939 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
5943 // If casting the result of a getelementptr instruction with no offset, turn
5944 // this into a cast of the original pointer!
5946 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5947 bool AllZeroOperands = true;
5948 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5949 if (!isa<Constant>(GEP->getOperand(i)) ||
5950 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5951 AllZeroOperands = false;
5954 if (AllZeroOperands) {
5955 // Changing the cast operand is usually not a good idea but it is safe
5956 // here because the pointer operand is being replaced with another
5957 // pointer operand so the opcode doesn't need to change.
5958 CI.setOperand(0, GEP->getOperand(0));
5963 // If we are casting a malloc or alloca to a pointer to a type of the same
5964 // size, rewrite the allocation instruction to allocate the "right" type.
5965 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5966 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5969 // If we are casting a select then fold the cast into the select
5970 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5971 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5974 // If we are casting a PHI then fold the cast into the PHI
5975 if (isa<PHINode>(Src))
5976 if (Instruction *NV = FoldOpIntoPhi(CI))
5982 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
5983 /// integers. This function implements the common transforms for all those
5985 /// @brief Implement the transforms common to CastInst with integer operands
5986 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
5987 if (Instruction *Result = commonCastTransforms(CI))
5990 Value *Src = CI.getOperand(0);
5991 const Type *SrcTy = Src->getType();
5992 const Type *DestTy = CI.getType();
5993 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
5994 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5996 // See if we can simplify any instructions used by the LHS whose sole
5997 // purpose is to compute bits we don't care about.
5998 uint64_t KnownZero = 0, KnownOne = 0;
5999 if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
6000 KnownZero, KnownOne))
6003 // If the source isn't an instruction or has more than one use then we
6004 // can't do anything more.
6005 Instruction *SrcI = dyn_cast<Instruction>(Src);
6006 if (!SrcI || !Src->hasOneUse())
6009 // Attempt to propagate the cast into the instruction.
6010 int NumCastsRemoved = 0;
6011 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6012 // If this cast is a truncate, evaluting in a different type always
6013 // eliminates the cast, so it is always a win. If this is a noop-cast
6014 // this just removes a noop cast which isn't pointful, but simplifies
6015 // the code. If this is a zero-extension, we need to do an AND to
6016 // maintain the clear top-part of the computation, so we require that
6017 // the input have eliminated at least one cast. If this is a sign
6018 // extension, we insert two new casts (to do the extension) so we
6019 // require that two casts have been eliminated.
6020 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6022 switch (CI.getOpcode()) {
6023 case Instruction::Trunc:
6026 case Instruction::ZExt:
6027 DoXForm = NumCastsRemoved >= 1;
6029 case Instruction::SExt:
6030 DoXForm = NumCastsRemoved >= 2;
6032 case Instruction::BitCast:
6036 // All the others use floating point so we shouldn't actually
6037 // get here because of the check above.
6038 assert(!"Unknown cast type .. unreachable");
6044 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6045 CI.getOpcode() == Instruction::SExt);
6046 assert(Res->getType() == DestTy);
6047 switch (CI.getOpcode()) {
6048 default: assert(0 && "Unknown cast type!");
6049 case Instruction::Trunc:
6050 case Instruction::BitCast:
6051 // Just replace this cast with the result.
6052 return ReplaceInstUsesWith(CI, Res);
6053 case Instruction::ZExt: {
6054 // We need to emit an AND to clear the high bits.
6055 assert(SrcBitSize < DestBitSize && "Not a zext?");
6057 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6058 if (DestBitSize < 64)
6059 C = ConstantExpr::getTrunc(C, DestTy);
6060 return BinaryOperator::createAnd(Res, C);
6062 case Instruction::SExt:
6063 // We need to emit a cast to truncate, then a cast to sext.
6064 return CastInst::create(Instruction::SExt,
6065 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6071 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6072 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6074 switch (SrcI->getOpcode()) {
6075 case Instruction::Add:
6076 case Instruction::Mul:
6077 case Instruction::And:
6078 case Instruction::Or:
6079 case Instruction::Xor:
6080 // If we are discarding information, or just changing the sign,
6082 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6083 // Don't insert two casts if they cannot be eliminated. We allow
6084 // two casts to be inserted if the sizes are the same. This could
6085 // only be converting signedness, which is a noop.
6086 if (DestBitSize == SrcBitSize ||
6087 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6088 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6089 Instruction::CastOps opcode = CI.getOpcode();
6090 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6091 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6092 return BinaryOperator::create(
6093 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6097 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6098 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6099 SrcI->getOpcode() == Instruction::Xor &&
6100 Op1 == ConstantInt::getTrue() &&
6101 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6102 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6103 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6106 case Instruction::SDiv:
6107 case Instruction::UDiv:
6108 case Instruction::SRem:
6109 case Instruction::URem:
6110 // If we are just changing the sign, rewrite.
6111 if (DestBitSize == SrcBitSize) {
6112 // Don't insert two casts if they cannot be eliminated. We allow
6113 // two casts to be inserted if the sizes are the same. This could
6114 // only be converting signedness, which is a noop.
6115 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6116 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6117 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6119 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6121 return BinaryOperator::create(
6122 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6127 case Instruction::Shl:
6128 // Allow changing the sign of the source operand. Do not allow
6129 // changing the size of the shift, UNLESS the shift amount is a
6130 // constant. We must not change variable sized shifts to a smaller
6131 // size, because it is undefined to shift more bits out than exist
6133 if (DestBitSize == SrcBitSize ||
6134 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6135 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6136 Instruction::BitCast : Instruction::Trunc);
6137 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6138 return new ShiftInst(Instruction::Shl, Op0c, Op1);
6141 case Instruction::AShr:
6142 // If this is a signed shr, and if all bits shifted in are about to be
6143 // truncated off, turn it into an unsigned shr to allow greater
6145 if (DestBitSize < SrcBitSize &&
6146 isa<ConstantInt>(Op1)) {
6147 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6148 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6149 // Insert the new logical shift right.
6150 return new ShiftInst(Instruction::LShr, Op0, Op1);
6155 case Instruction::ICmp:
6156 // If we are just checking for a icmp eq of a single bit and casting it
6157 // to an integer, then shift the bit to the appropriate place and then
6158 // cast to integer to avoid the comparison.
6159 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6160 uint64_t Op1CV = Op1C->getZExtValue();
6161 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6162 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6163 // cast (X == 1) to int --> X iff X has only the low bit set.
6164 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6165 // cast (X != 0) to int --> X iff X has only the low bit set.
6166 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6167 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6168 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6169 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6170 // If Op1C some other power of two, convert:
6171 uint64_t KnownZero, KnownOne;
6172 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
6173 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6175 // This only works for EQ and NE
6176 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6177 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6180 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6181 bool isNE = pred == ICmpInst::ICMP_NE;
6182 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6183 // (X&4) == 2 --> false
6184 // (X&4) != 2 --> true
6185 Constant *Res = ConstantInt::get(isNE);
6186 Res = ConstantExpr::getZExt(Res, CI.getType());
6187 return ReplaceInstUsesWith(CI, Res);
6190 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6193 // Perform a logical shr by shiftamt.
6194 // Insert the shift to put the result in the low bit.
6195 In = InsertNewInstBefore(
6196 new ShiftInst(Instruction::LShr, In,
6197 ConstantInt::get(Type::Int8Ty, ShiftAmt),
6198 In->getName()+".lobit"), CI);
6201 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6202 Constant *One = ConstantInt::get(In->getType(), 1);
6203 In = BinaryOperator::createXor(In, One, "tmp");
6204 InsertNewInstBefore(cast<Instruction>(In), CI);
6207 if (CI.getType() == In->getType())
6208 return ReplaceInstUsesWith(CI, In);
6210 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6219 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6220 if (Instruction *Result = commonIntCastTransforms(CI))
6223 Value *Src = CI.getOperand(0);
6224 const Type *Ty = CI.getType();
6225 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6227 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6228 switch (SrcI->getOpcode()) {
6230 case Instruction::LShr:
6231 // We can shrink lshr to something smaller if we know the bits shifted in
6232 // are already zeros.
6233 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6234 unsigned ShAmt = ShAmtV->getZExtValue();
6236 // Get a mask for the bits shifting in.
6237 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6238 Value* SrcIOp0 = SrcI->getOperand(0);
6239 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6240 if (ShAmt >= DestBitWidth) // All zeros.
6241 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6243 // Okay, we can shrink this. Truncate the input, then return a new
6245 Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6246 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6248 } else { // This is a variable shr.
6250 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6251 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6252 // loop-invariant and CSE'd.
6253 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6254 Value *One = ConstantInt::get(SrcI->getType(), 1);
6256 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6257 SrcI->getOperand(1),
6259 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6260 SrcI->getOperand(0),
6262 Value *Zero = Constant::getNullValue(V->getType());
6263 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6273 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6274 // If one of the common conversion will work ..
6275 if (Instruction *Result = commonIntCastTransforms(CI))
6278 Value *Src = CI.getOperand(0);
6280 // If this is a cast of a cast
6281 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6282 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6283 // types and if the sizes are just right we can convert this into a logical
6284 // 'and' which will be much cheaper than the pair of casts.
6285 if (isa<TruncInst>(CSrc)) {
6286 // Get the sizes of the types involved
6287 Value *A = CSrc->getOperand(0);
6288 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6289 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6290 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6291 // If we're actually extending zero bits and the trunc is a no-op
6292 if (MidSize < DstSize && SrcSize == DstSize) {
6293 // Replace both of the casts with an And of the type mask.
6294 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
6295 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6297 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6298 // Unfortunately, if the type changed, we need to cast it back.
6299 if (And->getType() != CI.getType()) {
6300 And->setName(CSrc->getName()+".mask");
6301 InsertNewInstBefore(And, CI);
6302 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6312 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6313 return commonIntCastTransforms(CI);
6316 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6317 return commonCastTransforms(CI);
6320 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6321 return commonCastTransforms(CI);
6324 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6325 return commonCastTransforms(CI);
6328 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6329 return commonCastTransforms(CI);
6332 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6333 return commonCastTransforms(CI);
6336 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6337 return commonCastTransforms(CI);
6340 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6341 return commonCastTransforms(CI);
6344 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6345 return commonCastTransforms(CI);
6348 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6350 // If the operands are integer typed then apply the integer transforms,
6351 // otherwise just apply the common ones.
6352 Value *Src = CI.getOperand(0);
6353 const Type *SrcTy = Src->getType();
6354 const Type *DestTy = CI.getType();
6356 if (SrcTy->isInteger() && DestTy->isInteger()) {
6357 if (Instruction *Result = commonIntCastTransforms(CI))
6360 if (Instruction *Result = commonCastTransforms(CI))
6365 // Get rid of casts from one type to the same type. These are useless and can
6366 // be replaced by the operand.
6367 if (DestTy == Src->getType())
6368 return ReplaceInstUsesWith(CI, Src);
6370 // If the source and destination are pointers, and this cast is equivalent to
6371 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6372 // This can enhance SROA and other transforms that want type-safe pointers.
6373 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6374 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6375 const Type *DstElTy = DstPTy->getElementType();
6376 const Type *SrcElTy = SrcPTy->getElementType();
6378 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6379 unsigned NumZeros = 0;
6380 while (SrcElTy != DstElTy &&
6381 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6382 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6383 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6387 // If we found a path from the src to dest, create the getelementptr now.
6388 if (SrcElTy == DstElTy) {
6389 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6390 return new GetElementPtrInst(Src, Idxs);
6395 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6396 if (SVI->hasOneUse()) {
6397 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6398 // a bitconvert to a vector with the same # elts.
6399 if (isa<PackedType>(DestTy) &&
6400 cast<PackedType>(DestTy)->getNumElements() ==
6401 SVI->getType()->getNumElements()) {
6403 // If either of the operands is a cast from CI.getType(), then
6404 // evaluating the shuffle in the casted destination's type will allow
6405 // us to eliminate at least one cast.
6406 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6407 Tmp->getOperand(0)->getType() == DestTy) ||
6408 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6409 Tmp->getOperand(0)->getType() == DestTy)) {
6410 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6411 SVI->getOperand(0), DestTy, &CI);
6412 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6413 SVI->getOperand(1), DestTy, &CI);
6414 // Return a new shuffle vector. Use the same element ID's, as we
6415 // know the vector types match #elts.
6416 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6424 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6426 /// %D = select %cond, %C, %A
6428 /// %C = select %cond, %B, 0
6431 /// Assuming that the specified instruction is an operand to the select, return
6432 /// a bitmask indicating which operands of this instruction are foldable if they
6433 /// equal the other incoming value of the select.
6435 static unsigned GetSelectFoldableOperands(Instruction *I) {
6436 switch (I->getOpcode()) {
6437 case Instruction::Add:
6438 case Instruction::Mul:
6439 case Instruction::And:
6440 case Instruction::Or:
6441 case Instruction::Xor:
6442 return 3; // Can fold through either operand.
6443 case Instruction::Sub: // Can only fold on the amount subtracted.
6444 case Instruction::Shl: // Can only fold on the shift amount.
6445 case Instruction::LShr:
6446 case Instruction::AShr:
6449 return 0; // Cannot fold
6453 /// GetSelectFoldableConstant - For the same transformation as the previous
6454 /// function, return the identity constant that goes into the select.
6455 static Constant *GetSelectFoldableConstant(Instruction *I) {
6456 switch (I->getOpcode()) {
6457 default: assert(0 && "This cannot happen!"); abort();
6458 case Instruction::Add:
6459 case Instruction::Sub:
6460 case Instruction::Or:
6461 case Instruction::Xor:
6462 return Constant::getNullValue(I->getType());
6463 case Instruction::Shl:
6464 case Instruction::LShr:
6465 case Instruction::AShr:
6466 return Constant::getNullValue(Type::Int8Ty);
6467 case Instruction::And:
6468 return ConstantInt::getAllOnesValue(I->getType());
6469 case Instruction::Mul:
6470 return ConstantInt::get(I->getType(), 1);
6474 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6475 /// have the same opcode and only one use each. Try to simplify this.
6476 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6478 if (TI->getNumOperands() == 1) {
6479 // If this is a non-volatile load or a cast from the same type,
6482 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6485 return 0; // unknown unary op.
6488 // Fold this by inserting a select from the input values.
6489 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6490 FI->getOperand(0), SI.getName()+".v");
6491 InsertNewInstBefore(NewSI, SI);
6492 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6496 // Only handle binary, compare and shift operators here.
6497 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
6500 // Figure out if the operations have any operands in common.
6501 Value *MatchOp, *OtherOpT, *OtherOpF;
6503 if (TI->getOperand(0) == FI->getOperand(0)) {
6504 MatchOp = TI->getOperand(0);
6505 OtherOpT = TI->getOperand(1);
6506 OtherOpF = FI->getOperand(1);
6507 MatchIsOpZero = true;
6508 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6509 MatchOp = TI->getOperand(1);
6510 OtherOpT = TI->getOperand(0);
6511 OtherOpF = FI->getOperand(0);
6512 MatchIsOpZero = false;
6513 } else if (!TI->isCommutative()) {
6515 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6516 MatchOp = TI->getOperand(0);
6517 OtherOpT = TI->getOperand(1);
6518 OtherOpF = FI->getOperand(0);
6519 MatchIsOpZero = true;
6520 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6521 MatchOp = TI->getOperand(1);
6522 OtherOpT = TI->getOperand(0);
6523 OtherOpF = FI->getOperand(1);
6524 MatchIsOpZero = true;
6529 // If we reach here, they do have operations in common.
6530 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6531 OtherOpF, SI.getName()+".v");
6532 InsertNewInstBefore(NewSI, SI);
6534 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6536 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6538 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6541 assert(isa<ShiftInst>(TI) && "Should only have Shift here");
6543 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6545 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6548 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6549 Value *CondVal = SI.getCondition();
6550 Value *TrueVal = SI.getTrueValue();
6551 Value *FalseVal = SI.getFalseValue();
6553 // select true, X, Y -> X
6554 // select false, X, Y -> Y
6555 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6556 return ReplaceInstUsesWith(SI, C->getBoolValue() ? TrueVal : FalseVal);
6558 // select C, X, X -> X
6559 if (TrueVal == FalseVal)
6560 return ReplaceInstUsesWith(SI, TrueVal);
6562 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6563 return ReplaceInstUsesWith(SI, FalseVal);
6564 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6565 return ReplaceInstUsesWith(SI, TrueVal);
6566 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6567 if (isa<Constant>(TrueVal))
6568 return ReplaceInstUsesWith(SI, TrueVal);
6570 return ReplaceInstUsesWith(SI, FalseVal);
6573 if (SI.getType() == Type::Int1Ty) {
6575 if ((C = dyn_cast<ConstantInt>(TrueVal)) &&
6576 C->getType() == Type::Int1Ty) {
6577 if (C->getBoolValue()) {
6578 // Change: A = select B, true, C --> A = or B, C
6579 return BinaryOperator::createOr(CondVal, FalseVal);
6581 // Change: A = select B, false, C --> A = and !B, C
6583 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6584 "not."+CondVal->getName()), SI);
6585 return BinaryOperator::createAnd(NotCond, FalseVal);
6587 } else if ((C = dyn_cast<ConstantInt>(FalseVal)) &&
6588 C->getType() == Type::Int1Ty) {
6589 if (C->getBoolValue() == false) {
6590 // Change: A = select B, C, false --> A = and B, C
6591 return BinaryOperator::createAnd(CondVal, TrueVal);
6593 // Change: A = select B, C, true --> A = or !B, C
6595 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6596 "not."+CondVal->getName()), SI);
6597 return BinaryOperator::createOr(NotCond, TrueVal);
6602 // Selecting between two integer constants?
6603 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6604 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6605 // select C, 1, 0 -> cast C to int
6606 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6607 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6608 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6609 // select C, 0, 1 -> cast !C to int
6611 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6612 "not."+CondVal->getName()), SI);
6613 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6616 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6618 // (x <s 0) ? -1 : 0 -> ashr x, 31
6619 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6620 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6621 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6622 bool CanXForm = false;
6623 if (IC->isSignedPredicate())
6624 CanXForm = CmpCst->isNullValue() &&
6625 IC->getPredicate() == ICmpInst::ICMP_SLT;
6627 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6628 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6629 IC->getPredicate() == ICmpInst::ICMP_UGT;
6633 // The comparison constant and the result are not neccessarily the
6634 // same width. Make an all-ones value by inserting a AShr.
6635 Value *X = IC->getOperand(0);
6636 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6637 Constant *ShAmt = ConstantInt::get(Type::Int8Ty, Bits-1);
6638 Instruction *SRA = new ShiftInst(Instruction::AShr, X,
6640 InsertNewInstBefore(SRA, SI);
6642 // Finally, convert to the type of the select RHS. We figure out
6643 // if this requires a SExt, Trunc or BitCast based on the sizes.
6644 Instruction::CastOps opc = Instruction::BitCast;
6645 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6646 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6647 if (SRASize < SISize)
6648 opc = Instruction::SExt;
6649 else if (SRASize > SISize)
6650 opc = Instruction::Trunc;
6651 return CastInst::create(opc, SRA, SI.getType());
6656 // If one of the constants is zero (we know they can't both be) and we
6657 // have a fcmp instruction with zero, and we have an 'and' with the
6658 // non-constant value, eliminate this whole mess. This corresponds to
6659 // cases like this: ((X & 27) ? 27 : 0)
6660 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6661 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6662 cast<Constant>(IC->getOperand(1))->isNullValue())
6663 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6664 if (ICA->getOpcode() == Instruction::And &&
6665 isa<ConstantInt>(ICA->getOperand(1)) &&
6666 (ICA->getOperand(1) == TrueValC ||
6667 ICA->getOperand(1) == FalseValC) &&
6668 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6669 // Okay, now we know that everything is set up, we just don't
6670 // know whether we have a icmp_ne or icmp_eq and whether the
6671 // true or false val is the zero.
6672 bool ShouldNotVal = !TrueValC->isNullValue();
6673 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6676 V = InsertNewInstBefore(BinaryOperator::create(
6677 Instruction::Xor, V, ICA->getOperand(1)), SI);
6678 return ReplaceInstUsesWith(SI, V);
6683 // See if we are selecting two values based on a comparison of the two values.
6684 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6685 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6686 // Transform (X == Y) ? X : Y -> Y
6687 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6688 return ReplaceInstUsesWith(SI, FalseVal);
6689 // Transform (X != Y) ? X : Y -> X
6690 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6691 return ReplaceInstUsesWith(SI, TrueVal);
6692 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6694 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6695 // Transform (X == Y) ? Y : X -> X
6696 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6697 return ReplaceInstUsesWith(SI, FalseVal);
6698 // Transform (X != Y) ? Y : X -> Y
6699 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6700 return ReplaceInstUsesWith(SI, TrueVal);
6701 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6705 // See if we are selecting two values based on a comparison of the two values.
6706 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6707 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6708 // Transform (X == Y) ? X : Y -> Y
6709 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6710 return ReplaceInstUsesWith(SI, FalseVal);
6711 // Transform (X != Y) ? X : Y -> X
6712 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6713 return ReplaceInstUsesWith(SI, TrueVal);
6714 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6716 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6717 // Transform (X == Y) ? Y : X -> X
6718 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6719 return ReplaceInstUsesWith(SI, FalseVal);
6720 // Transform (X != Y) ? Y : X -> Y
6721 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6722 return ReplaceInstUsesWith(SI, TrueVal);
6723 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6727 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6728 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6729 if (TI->hasOneUse() && FI->hasOneUse()) {
6730 Instruction *AddOp = 0, *SubOp = 0;
6732 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6733 if (TI->getOpcode() == FI->getOpcode())
6734 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6737 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6738 // even legal for FP.
6739 if (TI->getOpcode() == Instruction::Sub &&
6740 FI->getOpcode() == Instruction::Add) {
6741 AddOp = FI; SubOp = TI;
6742 } else if (FI->getOpcode() == Instruction::Sub &&
6743 TI->getOpcode() == Instruction::Add) {
6744 AddOp = TI; SubOp = FI;
6748 Value *OtherAddOp = 0;
6749 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6750 OtherAddOp = AddOp->getOperand(1);
6751 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6752 OtherAddOp = AddOp->getOperand(0);
6756 // So at this point we know we have (Y -> OtherAddOp):
6757 // select C, (add X, Y), (sub X, Z)
6758 Value *NegVal; // Compute -Z
6759 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6760 NegVal = ConstantExpr::getNeg(C);
6762 NegVal = InsertNewInstBefore(
6763 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6766 Value *NewTrueOp = OtherAddOp;
6767 Value *NewFalseOp = NegVal;
6769 std::swap(NewTrueOp, NewFalseOp);
6770 Instruction *NewSel =
6771 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6773 NewSel = InsertNewInstBefore(NewSel, SI);
6774 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6779 // See if we can fold the select into one of our operands.
6780 if (SI.getType()->isInteger()) {
6781 // See the comment above GetSelectFoldableOperands for a description of the
6782 // transformation we are doing here.
6783 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6784 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6785 !isa<Constant>(FalseVal))
6786 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6787 unsigned OpToFold = 0;
6788 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6790 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6795 Constant *C = GetSelectFoldableConstant(TVI);
6796 std::string Name = TVI->getName(); TVI->setName("");
6797 Instruction *NewSel =
6798 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6800 InsertNewInstBefore(NewSel, SI);
6801 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6802 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6803 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6804 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6806 assert(0 && "Unknown instruction!!");
6811 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6812 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6813 !isa<Constant>(TrueVal))
6814 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6815 unsigned OpToFold = 0;
6816 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6818 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6823 Constant *C = GetSelectFoldableConstant(FVI);
6824 std::string Name = FVI->getName(); FVI->setName("");
6825 Instruction *NewSel =
6826 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6828 InsertNewInstBefore(NewSel, SI);
6829 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6830 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6831 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6832 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6834 assert(0 && "Unknown instruction!!");
6840 if (BinaryOperator::isNot(CondVal)) {
6841 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6842 SI.setOperand(1, FalseVal);
6843 SI.setOperand(2, TrueVal);
6850 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6851 /// determine, return it, otherwise return 0.
6852 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6853 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6854 unsigned Align = GV->getAlignment();
6855 if (Align == 0 && TD)
6856 Align = TD->getTypeAlignment(GV->getType()->getElementType());
6858 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6859 unsigned Align = AI->getAlignment();
6860 if (Align == 0 && TD) {
6861 if (isa<AllocaInst>(AI))
6862 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6863 else if (isa<MallocInst>(AI)) {
6864 // Malloc returns maximally aligned memory.
6865 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6866 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
6867 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::Int64Ty));
6871 } else if (isa<BitCastInst>(V) ||
6872 (isa<ConstantExpr>(V) &&
6873 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6874 User *CI = cast<User>(V);
6875 if (isa<PointerType>(CI->getOperand(0)->getType()))
6876 return GetKnownAlignment(CI->getOperand(0), TD);
6878 } else if (isa<GetElementPtrInst>(V) ||
6879 (isa<ConstantExpr>(V) &&
6880 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6881 User *GEPI = cast<User>(V);
6882 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6883 if (BaseAlignment == 0) return 0;
6885 // If all indexes are zero, it is just the alignment of the base pointer.
6886 bool AllZeroOperands = true;
6887 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6888 if (!isa<Constant>(GEPI->getOperand(i)) ||
6889 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6890 AllZeroOperands = false;
6893 if (AllZeroOperands)
6894 return BaseAlignment;
6896 // Otherwise, if the base alignment is >= the alignment we expect for the
6897 // base pointer type, then we know that the resultant pointer is aligned at
6898 // least as much as its type requires.
6901 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6902 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
6904 const Type *GEPTy = GEPI->getType();
6905 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
6913 /// visitCallInst - CallInst simplification. This mostly only handles folding
6914 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6915 /// the heavy lifting.
6917 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6918 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6919 if (!II) return visitCallSite(&CI);
6921 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6923 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6924 bool Changed = false;
6926 // memmove/cpy/set of zero bytes is a noop.
6927 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6928 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6930 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6931 if (CI->getZExtValue() == 1) {
6932 // Replace the instruction with just byte operations. We would
6933 // transform other cases to loads/stores, but we don't know if
6934 // alignment is sufficient.
6938 // If we have a memmove and the source operation is a constant global,
6939 // then the source and dest pointers can't alias, so we can change this
6940 // into a call to memcpy.
6941 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6942 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6943 if (GVSrc->isConstant()) {
6944 Module *M = CI.getParent()->getParent()->getParent();
6946 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
6948 Name = "llvm.memcpy.i32";
6950 Name = "llvm.memcpy.i64";
6951 Constant *MemCpy = M->getOrInsertFunction(Name,
6952 CI.getCalledFunction()->getFunctionType());
6953 CI.setOperand(0, MemCpy);
6958 // If we can determine a pointer alignment that is bigger than currently
6959 // set, update the alignment.
6960 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6961 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6962 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6963 unsigned Align = std::min(Alignment1, Alignment2);
6964 if (MI->getAlignment()->getZExtValue() < Align) {
6965 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
6968 } else if (isa<MemSetInst>(MI)) {
6969 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6970 if (MI->getAlignment()->getZExtValue() < Alignment) {
6971 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
6976 if (Changed) return II;
6978 switch (II->getIntrinsicID()) {
6980 case Intrinsic::ppc_altivec_lvx:
6981 case Intrinsic::ppc_altivec_lvxl:
6982 case Intrinsic::x86_sse_loadu_ps:
6983 case Intrinsic::x86_sse2_loadu_pd:
6984 case Intrinsic::x86_sse2_loadu_dq:
6985 // Turn PPC lvx -> load if the pointer is known aligned.
6986 // Turn X86 loadups -> load if the pointer is known aligned.
6987 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6988 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
6989 PointerType::get(II->getType()), CI);
6990 return new LoadInst(Ptr);
6993 case Intrinsic::ppc_altivec_stvx:
6994 case Intrinsic::ppc_altivec_stvxl:
6995 // Turn stvx -> store if the pointer is known aligned.
6996 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
6997 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
6998 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7000 return new StoreInst(II->getOperand(1), Ptr);
7003 case Intrinsic::x86_sse_storeu_ps:
7004 case Intrinsic::x86_sse2_storeu_pd:
7005 case Intrinsic::x86_sse2_storeu_dq:
7006 case Intrinsic::x86_sse2_storel_dq:
7007 // Turn X86 storeu -> store if the pointer is known aligned.
7008 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7009 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7010 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7012 return new StoreInst(II->getOperand(2), Ptr);
7016 case Intrinsic::x86_sse_cvttss2si: {
7017 // These intrinsics only demands the 0th element of its input vector. If
7018 // we can simplify the input based on that, do so now.
7020 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7022 II->setOperand(1, V);
7028 case Intrinsic::ppc_altivec_vperm:
7029 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7030 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
7031 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7033 // Check that all of the elements are integer constants or undefs.
7034 bool AllEltsOk = true;
7035 for (unsigned i = 0; i != 16; ++i) {
7036 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7037 !isa<UndefValue>(Mask->getOperand(i))) {
7044 // Cast the input vectors to byte vectors.
7045 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7046 II->getOperand(1), Mask->getType(), CI);
7047 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7048 II->getOperand(2), Mask->getType(), CI);
7049 Value *Result = UndefValue::get(Op0->getType());
7051 // Only extract each element once.
7052 Value *ExtractedElts[32];
7053 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7055 for (unsigned i = 0; i != 16; ++i) {
7056 if (isa<UndefValue>(Mask->getOperand(i)))
7058 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7059 Idx &= 31; // Match the hardware behavior.
7061 if (ExtractedElts[Idx] == 0) {
7063 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7064 InsertNewInstBefore(Elt, CI);
7065 ExtractedElts[Idx] = Elt;
7068 // Insert this value into the result vector.
7069 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7070 InsertNewInstBefore(cast<Instruction>(Result), CI);
7072 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7077 case Intrinsic::stackrestore: {
7078 // If the save is right next to the restore, remove the restore. This can
7079 // happen when variable allocas are DCE'd.
7080 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7081 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7082 BasicBlock::iterator BI = SS;
7084 return EraseInstFromFunction(CI);
7088 // If the stack restore is in a return/unwind block and if there are no
7089 // allocas or calls between the restore and the return, nuke the restore.
7090 TerminatorInst *TI = II->getParent()->getTerminator();
7091 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7092 BasicBlock::iterator BI = II;
7093 bool CannotRemove = false;
7094 for (++BI; &*BI != TI; ++BI) {
7095 if (isa<AllocaInst>(BI) ||
7096 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7097 CannotRemove = true;
7102 return EraseInstFromFunction(CI);
7109 return visitCallSite(II);
7112 // InvokeInst simplification
7114 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7115 return visitCallSite(&II);
7118 // visitCallSite - Improvements for call and invoke instructions.
7120 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7121 bool Changed = false;
7123 // If the callee is a constexpr cast of a function, attempt to move the cast
7124 // to the arguments of the call/invoke.
7125 if (transformConstExprCastCall(CS)) return 0;
7127 Value *Callee = CS.getCalledValue();
7129 if (Function *CalleeF = dyn_cast<Function>(Callee))
7130 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7131 Instruction *OldCall = CS.getInstruction();
7132 // If the call and callee calling conventions don't match, this call must
7133 // be unreachable, as the call is undefined.
7134 new StoreInst(ConstantInt::getTrue(),
7135 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7136 if (!OldCall->use_empty())
7137 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7138 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7139 return EraseInstFromFunction(*OldCall);
7143 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7144 // This instruction is not reachable, just remove it. We insert a store to
7145 // undef so that we know that this code is not reachable, despite the fact
7146 // that we can't modify the CFG here.
7147 new StoreInst(ConstantInt::getTrue(),
7148 UndefValue::get(PointerType::get(Type::Int1Ty)),
7149 CS.getInstruction());
7151 if (!CS.getInstruction()->use_empty())
7152 CS.getInstruction()->
7153 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7155 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7156 // Don't break the CFG, insert a dummy cond branch.
7157 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7158 ConstantInt::getTrue(), II);
7160 return EraseInstFromFunction(*CS.getInstruction());
7163 const PointerType *PTy = cast<PointerType>(Callee->getType());
7164 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7165 if (FTy->isVarArg()) {
7166 // See if we can optimize any arguments passed through the varargs area of
7168 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7169 E = CS.arg_end(); I != E; ++I)
7170 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7171 // If this cast does not effect the value passed through the varargs
7172 // area, we can eliminate the use of the cast.
7173 Value *Op = CI->getOperand(0);
7174 if (CI->isLosslessCast()) {
7181 return Changed ? CS.getInstruction() : 0;
7184 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7185 // attempt to move the cast to the arguments of the call/invoke.
7187 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7188 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7189 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7190 if (CE->getOpcode() != Instruction::BitCast ||
7191 !isa<Function>(CE->getOperand(0)))
7193 Function *Callee = cast<Function>(CE->getOperand(0));
7194 Instruction *Caller = CS.getInstruction();
7196 // Okay, this is a cast from a function to a different type. Unless doing so
7197 // would cause a type conversion of one of our arguments, change this call to
7198 // be a direct call with arguments casted to the appropriate types.
7200 const FunctionType *FT = Callee->getFunctionType();
7201 const Type *OldRetTy = Caller->getType();
7203 // Check to see if we are changing the return type...
7204 if (OldRetTy != FT->getReturnType()) {
7205 if (Callee->isExternal() && !Caller->use_empty() &&
7206 OldRetTy != FT->getReturnType() &&
7207 // Conversion is ok if changing from pointer to int of same size.
7208 !(isa<PointerType>(FT->getReturnType()) &&
7209 TD->getIntPtrType() == OldRetTy))
7210 return false; // Cannot transform this return value.
7212 // If the callsite is an invoke instruction, and the return value is used by
7213 // a PHI node in a successor, we cannot change the return type of the call
7214 // because there is no place to put the cast instruction (without breaking
7215 // the critical edge). Bail out in this case.
7216 if (!Caller->use_empty())
7217 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7218 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7220 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7221 if (PN->getParent() == II->getNormalDest() ||
7222 PN->getParent() == II->getUnwindDest())
7226 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7227 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7229 CallSite::arg_iterator AI = CS.arg_begin();
7230 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7231 const Type *ParamTy = FT->getParamType(i);
7232 const Type *ActTy = (*AI)->getType();
7233 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7234 //Either we can cast directly, or we can upconvert the argument
7235 bool isConvertible = ActTy == ParamTy ||
7236 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7237 (ParamTy->isIntegral() && ActTy->isIntegral() &&
7238 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7239 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7240 && c->getSExtValue() > 0);
7241 if (Callee->isExternal() && !isConvertible) return false;
7244 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7245 Callee->isExternal())
7246 return false; // Do not delete arguments unless we have a function body...
7248 // Okay, we decided that this is a safe thing to do: go ahead and start
7249 // inserting cast instructions as necessary...
7250 std::vector<Value*> Args;
7251 Args.reserve(NumActualArgs);
7253 AI = CS.arg_begin();
7254 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7255 const Type *ParamTy = FT->getParamType(i);
7256 if ((*AI)->getType() == ParamTy) {
7257 Args.push_back(*AI);
7259 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7260 false, ParamTy, false);
7261 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7262 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7266 // If the function takes more arguments than the call was taking, add them
7268 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7269 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7271 // If we are removing arguments to the function, emit an obnoxious warning...
7272 if (FT->getNumParams() < NumActualArgs)
7273 if (!FT->isVarArg()) {
7274 cerr << "WARNING: While resolving call to function '"
7275 << Callee->getName() << "' arguments were dropped!\n";
7277 // Add all of the arguments in their promoted form to the arg list...
7278 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7279 const Type *PTy = getPromotedType((*AI)->getType());
7280 if (PTy != (*AI)->getType()) {
7281 // Must promote to pass through va_arg area!
7282 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7284 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7285 InsertNewInstBefore(Cast, *Caller);
7286 Args.push_back(Cast);
7288 Args.push_back(*AI);
7293 if (FT->getReturnType() == Type::VoidTy)
7294 Caller->setName(""); // Void type should not have a name...
7297 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7298 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7299 Args, Caller->getName(), Caller);
7300 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7302 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7303 if (cast<CallInst>(Caller)->isTailCall())
7304 cast<CallInst>(NC)->setTailCall();
7305 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7308 // Insert a cast of the return type as necessary...
7310 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7311 if (NV->getType() != Type::VoidTy) {
7312 const Type *CallerTy = Caller->getType();
7313 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7315 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7317 // If this is an invoke instruction, we should insert it after the first
7318 // non-phi, instruction in the normal successor block.
7319 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7320 BasicBlock::iterator I = II->getNormalDest()->begin();
7321 while (isa<PHINode>(I)) ++I;
7322 InsertNewInstBefore(NC, *I);
7324 // Otherwise, it's a call, just insert cast right after the call instr
7325 InsertNewInstBefore(NC, *Caller);
7327 AddUsersToWorkList(*Caller);
7329 NV = UndefValue::get(Caller->getType());
7333 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7334 Caller->replaceAllUsesWith(NV);
7335 Caller->getParent()->getInstList().erase(Caller);
7336 removeFromWorkList(Caller);
7340 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7341 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7342 /// and a single binop.
7343 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7344 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7345 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7346 isa<GetElementPtrInst>(FirstInst) || isa<CmpInst>(FirstInst));
7347 unsigned Opc = FirstInst->getOpcode();
7348 Value *LHSVal = FirstInst->getOperand(0);
7349 Value *RHSVal = FirstInst->getOperand(1);
7351 const Type *LHSType = LHSVal->getType();
7352 const Type *RHSType = RHSVal->getType();
7354 // Scan to see if all operands are the same opcode, all have one use, and all
7355 // kill their operands (i.e. the operands have one use).
7356 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7357 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7358 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7359 // Verify type of the LHS matches so we don't fold cmp's of different
7360 // types or GEP's with different index types.
7361 I->getOperand(0)->getType() != LHSType ||
7362 I->getOperand(1)->getType() != RHSType)
7365 // If they are CmpInst instructions, check their predicates
7366 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7367 if (cast<CmpInst>(I)->getPredicate() !=
7368 cast<CmpInst>(FirstInst)->getPredicate())
7371 // Keep track of which operand needs a phi node.
7372 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7373 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7376 // Otherwise, this is safe to transform, determine if it is profitable.
7378 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7379 // Indexes are often folded into load/store instructions, so we don't want to
7380 // hide them behind a phi.
7381 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7384 Value *InLHS = FirstInst->getOperand(0);
7385 Value *InRHS = FirstInst->getOperand(1);
7386 PHINode *NewLHS = 0, *NewRHS = 0;
7388 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7389 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7390 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7391 InsertNewInstBefore(NewLHS, PN);
7396 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7397 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7398 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7399 InsertNewInstBefore(NewRHS, PN);
7403 // Add all operands to the new PHIs.
7404 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7406 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7407 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7410 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7411 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7415 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7416 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7417 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7418 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7420 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7421 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7423 assert(isa<GetElementPtrInst>(FirstInst));
7424 return new GetElementPtrInst(LHSVal, RHSVal);
7428 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7429 /// of the block that defines it. This means that it must be obvious the value
7430 /// of the load is not changed from the point of the load to the end of the
7432 static bool isSafeToSinkLoad(LoadInst *L) {
7433 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7435 for (++BBI; BBI != E; ++BBI)
7436 if (BBI->mayWriteToMemory())
7442 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7443 // operator and they all are only used by the PHI, PHI together their
7444 // inputs, and do the operation once, to the result of the PHI.
7445 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7446 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7448 // Scan the instruction, looking for input operations that can be folded away.
7449 // If all input operands to the phi are the same instruction (e.g. a cast from
7450 // the same type or "+42") we can pull the operation through the PHI, reducing
7451 // code size and simplifying code.
7452 Constant *ConstantOp = 0;
7453 const Type *CastSrcTy = 0;
7454 bool isVolatile = false;
7455 if (isa<CastInst>(FirstInst)) {
7456 CastSrcTy = FirstInst->getOperand(0)->getType();
7457 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7458 isa<CmpInst>(FirstInst)) {
7459 // Can fold binop, compare or shift here if the RHS is a constant,
7460 // otherwise call FoldPHIArgBinOpIntoPHI.
7461 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7462 if (ConstantOp == 0)
7463 return FoldPHIArgBinOpIntoPHI(PN);
7464 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7465 isVolatile = LI->isVolatile();
7466 // We can't sink the load if the loaded value could be modified between the
7467 // load and the PHI.
7468 if (LI->getParent() != PN.getIncomingBlock(0) ||
7469 !isSafeToSinkLoad(LI))
7471 } else if (isa<GetElementPtrInst>(FirstInst)) {
7472 if (FirstInst->getNumOperands() == 2)
7473 return FoldPHIArgBinOpIntoPHI(PN);
7474 // Can't handle general GEPs yet.
7477 return 0; // Cannot fold this operation.
7480 // Check to see if all arguments are the same operation.
7481 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7482 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7483 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7484 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7487 if (I->getOperand(0)->getType() != CastSrcTy)
7488 return 0; // Cast operation must match.
7489 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7490 // We can't sink the load if the loaded value could be modified between
7491 // the load and the PHI.
7492 if (LI->isVolatile() != isVolatile ||
7493 LI->getParent() != PN.getIncomingBlock(i) ||
7494 !isSafeToSinkLoad(LI))
7496 } else if (I->getOperand(1) != ConstantOp) {
7501 // Okay, they are all the same operation. Create a new PHI node of the
7502 // correct type, and PHI together all of the LHS's of the instructions.
7503 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7504 PN.getName()+".in");
7505 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7507 Value *InVal = FirstInst->getOperand(0);
7508 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7510 // Add all operands to the new PHI.
7511 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7512 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7513 if (NewInVal != InVal)
7515 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7520 // The new PHI unions all of the same values together. This is really
7521 // common, so we handle it intelligently here for compile-time speed.
7525 InsertNewInstBefore(NewPN, PN);
7529 // Insert and return the new operation.
7530 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7531 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7532 else if (isa<LoadInst>(FirstInst))
7533 return new LoadInst(PhiVal, "", isVolatile);
7534 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7535 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7536 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7537 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7538 PhiVal, ConstantOp);
7540 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7541 PhiVal, ConstantOp);
7544 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7546 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7547 if (PN->use_empty()) return true;
7548 if (!PN->hasOneUse()) return false;
7550 // Remember this node, and if we find the cycle, return.
7551 if (!PotentiallyDeadPHIs.insert(PN).second)
7554 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7555 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7560 // PHINode simplification
7562 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7563 // If LCSSA is around, don't mess with Phi nodes
7564 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7566 if (Value *V = PN.hasConstantValue())
7567 return ReplaceInstUsesWith(PN, V);
7569 // If all PHI operands are the same operation, pull them through the PHI,
7570 // reducing code size.
7571 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7572 PN.getIncomingValue(0)->hasOneUse())
7573 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7576 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7577 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7578 // PHI)... break the cycle.
7580 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7581 std::set<PHINode*> PotentiallyDeadPHIs;
7582 PotentiallyDeadPHIs.insert(&PN);
7583 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7584 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7590 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7591 Instruction *InsertPoint,
7593 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7594 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7595 // We must cast correctly to the pointer type. Ensure that we
7596 // sign extend the integer value if it is smaller as this is
7597 // used for address computation.
7598 Instruction::CastOps opcode =
7599 (VTySize < PtrSize ? Instruction::SExt :
7600 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7601 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7605 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7606 Value *PtrOp = GEP.getOperand(0);
7607 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7608 // If so, eliminate the noop.
7609 if (GEP.getNumOperands() == 1)
7610 return ReplaceInstUsesWith(GEP, PtrOp);
7612 if (isa<UndefValue>(GEP.getOperand(0)))
7613 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7615 bool HasZeroPointerIndex = false;
7616 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7617 HasZeroPointerIndex = C->isNullValue();
7619 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7620 return ReplaceInstUsesWith(GEP, PtrOp);
7622 // Eliminate unneeded casts for indices.
7623 bool MadeChange = false;
7624 gep_type_iterator GTI = gep_type_begin(GEP);
7625 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7626 if (isa<SequentialType>(*GTI)) {
7627 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7628 Value *Src = CI->getOperand(0);
7629 const Type *SrcTy = Src->getType();
7630 const Type *DestTy = CI->getType();
7631 if (Src->getType()->isInteger()) {
7632 if (SrcTy->getPrimitiveSizeInBits() ==
7633 DestTy->getPrimitiveSizeInBits()) {
7634 // We can always eliminate a cast from ulong or long to the other.
7635 // We can always eliminate a cast from uint to int or the other on
7636 // 32-bit pointer platforms.
7637 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7639 GEP.setOperand(i, Src);
7641 } else if (SrcTy->getPrimitiveSizeInBits() <
7642 DestTy->getPrimitiveSizeInBits() &&
7643 SrcTy->getPrimitiveSize() == 4) {
7644 // We can eliminate a cast from [u]int to [u]long iff the target
7645 // is a 32-bit pointer target.
7646 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7648 GEP.setOperand(i, Src);
7653 // If we are using a wider index than needed for this platform, shrink it
7654 // to what we need. If the incoming value needs a cast instruction,
7655 // insert it. This explicit cast can make subsequent optimizations more
7657 Value *Op = GEP.getOperand(i);
7658 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
7659 if (Constant *C = dyn_cast<Constant>(Op)) {
7660 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7663 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7665 GEP.setOperand(i, Op);
7669 if (MadeChange) return &GEP;
7671 // Combine Indices - If the source pointer to this getelementptr instruction
7672 // is a getelementptr instruction, combine the indices of the two
7673 // getelementptr instructions into a single instruction.
7675 std::vector<Value*> SrcGEPOperands;
7676 if (User *Src = dyn_castGetElementPtr(PtrOp))
7677 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7679 if (!SrcGEPOperands.empty()) {
7680 // Note that if our source is a gep chain itself that we wait for that
7681 // chain to be resolved before we perform this transformation. This
7682 // avoids us creating a TON of code in some cases.
7684 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7685 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7686 return 0; // Wait until our source is folded to completion.
7688 std::vector<Value *> Indices;
7690 // Find out whether the last index in the source GEP is a sequential idx.
7691 bool EndsWithSequential = false;
7692 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7693 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7694 EndsWithSequential = !isa<StructType>(*I);
7696 // Can we combine the two pointer arithmetics offsets?
7697 if (EndsWithSequential) {
7698 // Replace: gep (gep %P, long B), long A, ...
7699 // With: T = long A+B; gep %P, T, ...
7701 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7702 if (SO1 == Constant::getNullValue(SO1->getType())) {
7704 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7707 // If they aren't the same type, convert both to an integer of the
7708 // target's pointer size.
7709 if (SO1->getType() != GO1->getType()) {
7710 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7711 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7712 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7713 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7715 unsigned PS = TD->getPointerSize();
7716 if (SO1->getType()->getPrimitiveSize() == PS) {
7717 // Convert GO1 to SO1's type.
7718 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7720 } else if (GO1->getType()->getPrimitiveSize() == PS) {
7721 // Convert SO1 to GO1's type.
7722 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7724 const Type *PT = TD->getIntPtrType();
7725 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7726 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7730 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7731 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7733 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7734 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7738 // Recycle the GEP we already have if possible.
7739 if (SrcGEPOperands.size() == 2) {
7740 GEP.setOperand(0, SrcGEPOperands[0]);
7741 GEP.setOperand(1, Sum);
7744 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7745 SrcGEPOperands.end()-1);
7746 Indices.push_back(Sum);
7747 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7749 } else if (isa<Constant>(*GEP.idx_begin()) &&
7750 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7751 SrcGEPOperands.size() != 1) {
7752 // Otherwise we can do the fold if the first index of the GEP is a zero
7753 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7754 SrcGEPOperands.end());
7755 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7758 if (!Indices.empty())
7759 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7761 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7762 // GEP of global variable. If all of the indices for this GEP are
7763 // constants, we can promote this to a constexpr instead of an instruction.
7765 // Scan for nonconstants...
7766 std::vector<Constant*> Indices;
7767 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7768 for (; I != E && isa<Constant>(*I); ++I)
7769 Indices.push_back(cast<Constant>(*I));
7771 if (I == E) { // If they are all constants...
7772 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7774 // Replace all uses of the GEP with the new constexpr...
7775 return ReplaceInstUsesWith(GEP, CE);
7777 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7778 if (!isa<PointerType>(X->getType())) {
7779 // Not interesting. Source pointer must be a cast from pointer.
7780 } else if (HasZeroPointerIndex) {
7781 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7782 // into : GEP [10 x ubyte]* X, long 0, ...
7784 // This occurs when the program declares an array extern like "int X[];"
7786 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7787 const PointerType *XTy = cast<PointerType>(X->getType());
7788 if (const ArrayType *XATy =
7789 dyn_cast<ArrayType>(XTy->getElementType()))
7790 if (const ArrayType *CATy =
7791 dyn_cast<ArrayType>(CPTy->getElementType()))
7792 if (CATy->getElementType() == XATy->getElementType()) {
7793 // At this point, we know that the cast source type is a pointer
7794 // to an array of the same type as the destination pointer
7795 // array. Because the array type is never stepped over (there
7796 // is a leading zero) we can fold the cast into this GEP.
7797 GEP.setOperand(0, X);
7800 } else if (GEP.getNumOperands() == 2) {
7801 // Transform things like:
7802 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7803 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7804 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7805 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7806 if (isa<ArrayType>(SrcElTy) &&
7807 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7808 TD->getTypeSize(ResElTy)) {
7809 Value *V = InsertNewInstBefore(
7810 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7811 GEP.getOperand(1), GEP.getName()), GEP);
7812 // V and GEP are both pointer types --> BitCast
7813 return new BitCastInst(V, GEP.getType());
7816 // Transform things like:
7817 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7818 // (where tmp = 8*tmp2) into:
7819 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7821 if (isa<ArrayType>(SrcElTy) &&
7822 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7823 uint64_t ArrayEltSize =
7824 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7826 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7827 // allow either a mul, shift, or constant here.
7829 ConstantInt *Scale = 0;
7830 if (ArrayEltSize == 1) {
7831 NewIdx = GEP.getOperand(1);
7832 Scale = ConstantInt::get(NewIdx->getType(), 1);
7833 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7834 NewIdx = ConstantInt::get(CI->getType(), 1);
7836 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7837 if (Inst->getOpcode() == Instruction::Shl &&
7838 isa<ConstantInt>(Inst->getOperand(1))) {
7840 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7841 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7842 NewIdx = Inst->getOperand(0);
7843 } else if (Inst->getOpcode() == Instruction::Mul &&
7844 isa<ConstantInt>(Inst->getOperand(1))) {
7845 Scale = cast<ConstantInt>(Inst->getOperand(1));
7846 NewIdx = Inst->getOperand(0);
7850 // If the index will be to exactly the right offset with the scale taken
7851 // out, perform the transformation.
7852 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7853 if (isa<ConstantInt>(Scale))
7854 Scale = ConstantInt::get(Scale->getType(),
7855 Scale->getZExtValue() / ArrayEltSize);
7856 if (Scale->getZExtValue() != 1) {
7857 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7859 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7860 NewIdx = InsertNewInstBefore(Sc, GEP);
7863 // Insert the new GEP instruction.
7864 Instruction *NewGEP =
7865 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7866 NewIdx, GEP.getName());
7867 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7868 // The NewGEP must be pointer typed, so must the old one -> BitCast
7869 return new BitCastInst(NewGEP, GEP.getType());
7878 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7879 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7880 if (AI.isArrayAllocation()) // Check C != 1
7881 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7883 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7884 AllocationInst *New = 0;
7886 // Create and insert the replacement instruction...
7887 if (isa<MallocInst>(AI))
7888 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7890 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7891 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7894 InsertNewInstBefore(New, AI);
7896 // Scan to the end of the allocation instructions, to skip over a block of
7897 // allocas if possible...
7899 BasicBlock::iterator It = New;
7900 while (isa<AllocationInst>(*It)) ++It;
7902 // Now that I is pointing to the first non-allocation-inst in the block,
7903 // insert our getelementptr instruction...
7905 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
7906 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7907 New->getName()+".sub", It);
7909 // Now make everything use the getelementptr instead of the original
7911 return ReplaceInstUsesWith(AI, V);
7912 } else if (isa<UndefValue>(AI.getArraySize())) {
7913 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7916 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7917 // Note that we only do this for alloca's, because malloc should allocate and
7918 // return a unique pointer, even for a zero byte allocation.
7919 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7920 TD->getTypeSize(AI.getAllocatedType()) == 0)
7921 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7926 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7927 Value *Op = FI.getOperand(0);
7929 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7930 if (CastInst *CI = dyn_cast<CastInst>(Op))
7931 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7932 FI.setOperand(0, CI->getOperand(0));
7936 // free undef -> unreachable.
7937 if (isa<UndefValue>(Op)) {
7938 // Insert a new store to null because we cannot modify the CFG here.
7939 new StoreInst(ConstantInt::getTrue(),
7940 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
7941 return EraseInstFromFunction(FI);
7944 // If we have 'free null' delete the instruction. This can happen in stl code
7945 // when lots of inlining happens.
7946 if (isa<ConstantPointerNull>(Op))
7947 return EraseInstFromFunction(FI);
7953 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7954 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7955 User *CI = cast<User>(LI.getOperand(0));
7956 Value *CastOp = CI->getOperand(0);
7958 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7959 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7960 const Type *SrcPTy = SrcTy->getElementType();
7962 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7963 isa<PackedType>(DestPTy)) {
7964 // If the source is an array, the code below will not succeed. Check to
7965 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7967 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7968 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7969 if (ASrcTy->getNumElements() != 0) {
7970 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
7971 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7972 SrcTy = cast<PointerType>(CastOp->getType());
7973 SrcPTy = SrcTy->getElementType();
7976 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7977 isa<PackedType>(SrcPTy)) &&
7978 // Do not allow turning this into a load of an integer, which is then
7979 // casted to a pointer, this pessimizes pointer analysis a lot.
7980 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7981 IC.getTargetData().getTypeSize(SrcPTy) ==
7982 IC.getTargetData().getTypeSize(DestPTy)) {
7984 // Okay, we are casting from one integer or pointer type to another of
7985 // the same size. Instead of casting the pointer before the load, cast
7986 // the result of the loaded value.
7987 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
7989 LI.isVolatile()),LI);
7990 // Now cast the result of the load.
7991 return new BitCastInst(NewLoad, LI.getType());
7998 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
7999 /// from this value cannot trap. If it is not obviously safe to load from the
8000 /// specified pointer, we do a quick local scan of the basic block containing
8001 /// ScanFrom, to determine if the address is already accessed.
8002 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8003 // If it is an alloca or global variable, it is always safe to load from.
8004 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8006 // Otherwise, be a little bit agressive by scanning the local block where we
8007 // want to check to see if the pointer is already being loaded or stored
8008 // from/to. If so, the previous load or store would have already trapped,
8009 // so there is no harm doing an extra load (also, CSE will later eliminate
8010 // the load entirely).
8011 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8016 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8017 if (LI->getOperand(0) == V) return true;
8018 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8019 if (SI->getOperand(1) == V) return true;
8025 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8026 Value *Op = LI.getOperand(0);
8028 // load (cast X) --> cast (load X) iff safe
8029 if (isa<CastInst>(Op))
8030 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8033 // None of the following transforms are legal for volatile loads.
8034 if (LI.isVolatile()) return 0;
8036 if (&LI.getParent()->front() != &LI) {
8037 BasicBlock::iterator BBI = &LI; --BBI;
8038 // If the instruction immediately before this is a store to the same
8039 // address, do a simple form of store->load forwarding.
8040 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8041 if (SI->getOperand(1) == LI.getOperand(0))
8042 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8043 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8044 if (LIB->getOperand(0) == LI.getOperand(0))
8045 return ReplaceInstUsesWith(LI, LIB);
8048 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8049 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8050 isa<UndefValue>(GEPI->getOperand(0))) {
8051 // Insert a new store to null instruction before the load to indicate
8052 // that this code is not reachable. We do this instead of inserting
8053 // an unreachable instruction directly because we cannot modify the
8055 new StoreInst(UndefValue::get(LI.getType()),
8056 Constant::getNullValue(Op->getType()), &LI);
8057 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8060 if (Constant *C = dyn_cast<Constant>(Op)) {
8061 // load null/undef -> undef
8062 if ((C->isNullValue() || isa<UndefValue>(C))) {
8063 // Insert a new store to null instruction before the load to indicate that
8064 // this code is not reachable. We do this instead of inserting an
8065 // unreachable instruction directly because we cannot modify the CFG.
8066 new StoreInst(UndefValue::get(LI.getType()),
8067 Constant::getNullValue(Op->getType()), &LI);
8068 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8071 // Instcombine load (constant global) into the value loaded.
8072 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8073 if (GV->isConstant() && !GV->isExternal())
8074 return ReplaceInstUsesWith(LI, GV->getInitializer());
8076 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8077 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8078 if (CE->getOpcode() == Instruction::GetElementPtr) {
8079 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8080 if (GV->isConstant() && !GV->isExternal())
8082 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8083 return ReplaceInstUsesWith(LI, V);
8084 if (CE->getOperand(0)->isNullValue()) {
8085 // Insert a new store to null instruction before the load to indicate
8086 // that this code is not reachable. We do this instead of inserting
8087 // an unreachable instruction directly because we cannot modify the
8089 new StoreInst(UndefValue::get(LI.getType()),
8090 Constant::getNullValue(Op->getType()), &LI);
8091 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8094 } else if (CE->isCast()) {
8095 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8100 if (Op->hasOneUse()) {
8101 // Change select and PHI nodes to select values instead of addresses: this
8102 // helps alias analysis out a lot, allows many others simplifications, and
8103 // exposes redundancy in the code.
8105 // Note that we cannot do the transformation unless we know that the
8106 // introduced loads cannot trap! Something like this is valid as long as
8107 // the condition is always false: load (select bool %C, int* null, int* %G),
8108 // but it would not be valid if we transformed it to load from null
8111 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8112 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8113 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8114 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8115 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8116 SI->getOperand(1)->getName()+".val"), LI);
8117 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8118 SI->getOperand(2)->getName()+".val"), LI);
8119 return new SelectInst(SI->getCondition(), V1, V2);
8122 // load (select (cond, null, P)) -> load P
8123 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8124 if (C->isNullValue()) {
8125 LI.setOperand(0, SI->getOperand(2));
8129 // load (select (cond, P, null)) -> load P
8130 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8131 if (C->isNullValue()) {
8132 LI.setOperand(0, SI->getOperand(1));
8140 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
8142 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8143 User *CI = cast<User>(SI.getOperand(1));
8144 Value *CastOp = CI->getOperand(0);
8146 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8147 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8148 const Type *SrcPTy = SrcTy->getElementType();
8150 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8151 // If the source is an array, the code below will not succeed. Check to
8152 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8154 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8155 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8156 if (ASrcTy->getNumElements() != 0) {
8157 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
8158 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8159 SrcTy = cast<PointerType>(CastOp->getType());
8160 SrcPTy = SrcTy->getElementType();
8163 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8164 IC.getTargetData().getTypeSize(SrcPTy) ==
8165 IC.getTargetData().getTypeSize(DestPTy)) {
8167 // Okay, we are casting from one integer or pointer type to another of
8168 // the same size. Instead of casting the pointer before the store, cast
8169 // the value to be stored.
8171 Instruction::CastOps opcode = Instruction::BitCast;
8172 Value *SIOp0 = SI.getOperand(0);
8173 if (isa<PointerType>(SrcPTy)) {
8174 if (SIOp0->getType()->isIntegral())
8175 opcode = Instruction::IntToPtr;
8176 } else if (SrcPTy->isIntegral()) {
8177 if (isa<PointerType>(SIOp0->getType()))
8178 opcode = Instruction::PtrToInt;
8180 if (Constant *C = dyn_cast<Constant>(SIOp0))
8181 NewCast = ConstantExpr::getCast(opcode, C, SrcPTy);
8183 NewCast = IC.InsertNewInstBefore(
8184 CastInst::create(opcode, SIOp0, SrcPTy, SIOp0->getName()+".c"), SI);
8185 return new StoreInst(NewCast, CastOp);
8192 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8193 Value *Val = SI.getOperand(0);
8194 Value *Ptr = SI.getOperand(1);
8196 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8197 EraseInstFromFunction(SI);
8202 // Do really simple DSE, to catch cases where there are several consequtive
8203 // stores to the same location, separated by a few arithmetic operations. This
8204 // situation often occurs with bitfield accesses.
8205 BasicBlock::iterator BBI = &SI;
8206 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8210 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8211 // Prev store isn't volatile, and stores to the same location?
8212 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8215 EraseInstFromFunction(*PrevSI);
8221 // If this is a load, we have to stop. However, if the loaded value is from
8222 // the pointer we're loading and is producing the pointer we're storing,
8223 // then *this* store is dead (X = load P; store X -> P).
8224 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8225 if (LI == Val && LI->getOperand(0) == Ptr) {
8226 EraseInstFromFunction(SI);
8230 // Otherwise, this is a load from some other location. Stores before it
8235 // Don't skip over loads or things that can modify memory.
8236 if (BBI->mayWriteToMemory())
8241 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8243 // store X, null -> turns into 'unreachable' in SimplifyCFG
8244 if (isa<ConstantPointerNull>(Ptr)) {
8245 if (!isa<UndefValue>(Val)) {
8246 SI.setOperand(0, UndefValue::get(Val->getType()));
8247 if (Instruction *U = dyn_cast<Instruction>(Val))
8248 WorkList.push_back(U); // Dropped a use.
8251 return 0; // Do not modify these!
8254 // store undef, Ptr -> noop
8255 if (isa<UndefValue>(Val)) {
8256 EraseInstFromFunction(SI);
8261 // If the pointer destination is a cast, see if we can fold the cast into the
8263 if (isa<CastInst>(Ptr))
8264 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8266 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8268 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8272 // If this store is the last instruction in the basic block, and if the block
8273 // ends with an unconditional branch, try to move it to the successor block.
8275 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8276 if (BI->isUnconditional()) {
8277 // Check to see if the successor block has exactly two incoming edges. If
8278 // so, see if the other predecessor contains a store to the same location.
8279 // if so, insert a PHI node (if needed) and move the stores down.
8280 BasicBlock *Dest = BI->getSuccessor(0);
8282 pred_iterator PI = pred_begin(Dest);
8283 BasicBlock *Other = 0;
8284 if (*PI != BI->getParent())
8287 if (PI != pred_end(Dest)) {
8288 if (*PI != BI->getParent())
8293 if (++PI != pred_end(Dest))
8296 if (Other) { // If only one other pred...
8297 BBI = Other->getTerminator();
8298 // Make sure this other block ends in an unconditional branch and that
8299 // there is an instruction before the branch.
8300 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8301 BBI != Other->begin()) {
8303 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8305 // If this instruction is a store to the same location.
8306 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8307 // Okay, we know we can perform this transformation. Insert a PHI
8308 // node now if we need it.
8309 Value *MergedVal = OtherStore->getOperand(0);
8310 if (MergedVal != SI.getOperand(0)) {
8311 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8312 PN->reserveOperandSpace(2);
8313 PN->addIncoming(SI.getOperand(0), SI.getParent());
8314 PN->addIncoming(OtherStore->getOperand(0), Other);
8315 MergedVal = InsertNewInstBefore(PN, Dest->front());
8318 // Advance to a place where it is safe to insert the new store and
8320 BBI = Dest->begin();
8321 while (isa<PHINode>(BBI)) ++BBI;
8322 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8323 OtherStore->isVolatile()), *BBI);
8325 // Nuke the old stores.
8326 EraseInstFromFunction(SI);
8327 EraseInstFromFunction(*OtherStore);
8339 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8340 // Change br (not X), label True, label False to: br X, label False, True
8342 BasicBlock *TrueDest;
8343 BasicBlock *FalseDest;
8344 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8345 !isa<Constant>(X)) {
8346 // Swap Destinations and condition...
8348 BI.setSuccessor(0, FalseDest);
8349 BI.setSuccessor(1, TrueDest);
8353 // Cannonicalize fcmp_one -> fcmp_oeq
8354 FCmpInst::Predicate FPred; Value *Y;
8355 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8356 TrueDest, FalseDest)))
8357 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8358 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8359 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8360 std::string Name = I->getName(); I->setName("");
8361 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8362 Value *NewSCC = new FCmpInst(NewPred, X, Y, Name, I);
8363 // Swap Destinations and condition...
8364 BI.setCondition(NewSCC);
8365 BI.setSuccessor(0, FalseDest);
8366 BI.setSuccessor(1, TrueDest);
8367 removeFromWorkList(I);
8368 I->getParent()->getInstList().erase(I);
8369 WorkList.push_back(cast<Instruction>(NewSCC));
8373 // Cannonicalize icmp_ne -> icmp_eq
8374 ICmpInst::Predicate IPred;
8375 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8376 TrueDest, FalseDest)))
8377 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8378 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8379 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8380 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8381 std::string Name = I->getName(); I->setName("");
8382 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8383 Value *NewSCC = new ICmpInst(NewPred, X, Y, Name, I);
8384 // Swap Destinations and condition...
8385 BI.setCondition(NewSCC);
8386 BI.setSuccessor(0, FalseDest);
8387 BI.setSuccessor(1, TrueDest);
8388 removeFromWorkList(I);
8389 I->getParent()->getInstList().erase(I);
8390 WorkList.push_back(cast<Instruction>(NewSCC));
8397 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8398 Value *Cond = SI.getCondition();
8399 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8400 if (I->getOpcode() == Instruction::Add)
8401 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8402 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8403 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8404 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8406 SI.setOperand(0, I->getOperand(0));
8407 WorkList.push_back(I);
8414 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8415 /// is to leave as a vector operation.
8416 static bool CheapToScalarize(Value *V, bool isConstant) {
8417 if (isa<ConstantAggregateZero>(V))
8419 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8420 if (isConstant) return true;
8421 // If all elts are the same, we can extract.
8422 Constant *Op0 = C->getOperand(0);
8423 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8424 if (C->getOperand(i) != Op0)
8428 Instruction *I = dyn_cast<Instruction>(V);
8429 if (!I) return false;
8431 // Insert element gets simplified to the inserted element or is deleted if
8432 // this is constant idx extract element and its a constant idx insertelt.
8433 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8434 isa<ConstantInt>(I->getOperand(2)))
8436 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8438 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8439 if (BO->hasOneUse() &&
8440 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8441 CheapToScalarize(BO->getOperand(1), isConstant)))
8443 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8444 if (CI->hasOneUse() &&
8445 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8446 CheapToScalarize(CI->getOperand(1), isConstant)))
8452 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8453 /// elements into values that are larger than the #elts in the input.
8454 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8455 unsigned NElts = SVI->getType()->getNumElements();
8456 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8457 return std::vector<unsigned>(NElts, 0);
8458 if (isa<UndefValue>(SVI->getOperand(2)))
8459 return std::vector<unsigned>(NElts, 2*NElts);
8461 std::vector<unsigned> Result;
8462 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8463 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8464 if (isa<UndefValue>(CP->getOperand(i)))
8465 Result.push_back(NElts*2); // undef -> 8
8467 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8471 /// FindScalarElement - Given a vector and an element number, see if the scalar
8472 /// value is already around as a register, for example if it were inserted then
8473 /// extracted from the vector.
8474 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8475 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8476 const PackedType *PTy = cast<PackedType>(V->getType());
8477 unsigned Width = PTy->getNumElements();
8478 if (EltNo >= Width) // Out of range access.
8479 return UndefValue::get(PTy->getElementType());
8481 if (isa<UndefValue>(V))
8482 return UndefValue::get(PTy->getElementType());
8483 else if (isa<ConstantAggregateZero>(V))
8484 return Constant::getNullValue(PTy->getElementType());
8485 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8486 return CP->getOperand(EltNo);
8487 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8488 // If this is an insert to a variable element, we don't know what it is.
8489 if (!isa<ConstantInt>(III->getOperand(2)))
8491 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8493 // If this is an insert to the element we are looking for, return the
8496 return III->getOperand(1);
8498 // Otherwise, the insertelement doesn't modify the value, recurse on its
8500 return FindScalarElement(III->getOperand(0), EltNo);
8501 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8502 unsigned InEl = getShuffleMask(SVI)[EltNo];
8504 return FindScalarElement(SVI->getOperand(0), InEl);
8505 else if (InEl < Width*2)
8506 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8508 return UndefValue::get(PTy->getElementType());
8511 // Otherwise, we don't know.
8515 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8517 // If packed val is undef, replace extract with scalar undef.
8518 if (isa<UndefValue>(EI.getOperand(0)))
8519 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8521 // If packed val is constant 0, replace extract with scalar 0.
8522 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8523 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8525 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8526 // If packed val is constant with uniform operands, replace EI
8527 // with that operand
8528 Constant *op0 = C->getOperand(0);
8529 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8530 if (C->getOperand(i) != op0) {
8535 return ReplaceInstUsesWith(EI, op0);
8538 // If extracting a specified index from the vector, see if we can recursively
8539 // find a previously computed scalar that was inserted into the vector.
8540 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8541 // This instruction only demands the single element from the input vector.
8542 // If the input vector has a single use, simplify it based on this use
8544 uint64_t IndexVal = IdxC->getZExtValue();
8545 if (EI.getOperand(0)->hasOneUse()) {
8547 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8550 EI.setOperand(0, V);
8555 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8556 return ReplaceInstUsesWith(EI, Elt);
8559 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8560 if (I->hasOneUse()) {
8561 // Push extractelement into predecessor operation if legal and
8562 // profitable to do so
8563 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8564 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8565 if (CheapToScalarize(BO, isConstantElt)) {
8566 ExtractElementInst *newEI0 =
8567 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8568 EI.getName()+".lhs");
8569 ExtractElementInst *newEI1 =
8570 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8571 EI.getName()+".rhs");
8572 InsertNewInstBefore(newEI0, EI);
8573 InsertNewInstBefore(newEI1, EI);
8574 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8576 } else if (isa<LoadInst>(I)) {
8577 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8578 PointerType::get(EI.getType()), EI);
8579 GetElementPtrInst *GEP =
8580 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8581 InsertNewInstBefore(GEP, EI);
8582 return new LoadInst(GEP);
8585 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8586 // Extracting the inserted element?
8587 if (IE->getOperand(2) == EI.getOperand(1))
8588 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8589 // If the inserted and extracted elements are constants, they must not
8590 // be the same value, extract from the pre-inserted value instead.
8591 if (isa<Constant>(IE->getOperand(2)) &&
8592 isa<Constant>(EI.getOperand(1))) {
8593 AddUsesToWorkList(EI);
8594 EI.setOperand(0, IE->getOperand(0));
8597 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8598 // If this is extracting an element from a shufflevector, figure out where
8599 // it came from and extract from the appropriate input element instead.
8600 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8601 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8603 if (SrcIdx < SVI->getType()->getNumElements())
8604 Src = SVI->getOperand(0);
8605 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8606 SrcIdx -= SVI->getType()->getNumElements();
8607 Src = SVI->getOperand(1);
8609 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8611 return new ExtractElementInst(Src, SrcIdx);
8618 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8619 /// elements from either LHS or RHS, return the shuffle mask and true.
8620 /// Otherwise, return false.
8621 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8622 std::vector<Constant*> &Mask) {
8623 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8624 "Invalid CollectSingleShuffleElements");
8625 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8627 if (isa<UndefValue>(V)) {
8628 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8630 } else if (V == LHS) {
8631 for (unsigned i = 0; i != NumElts; ++i)
8632 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8634 } else if (V == RHS) {
8635 for (unsigned i = 0; i != NumElts; ++i)
8636 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8638 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8639 // If this is an insert of an extract from some other vector, include it.
8640 Value *VecOp = IEI->getOperand(0);
8641 Value *ScalarOp = IEI->getOperand(1);
8642 Value *IdxOp = IEI->getOperand(2);
8644 if (!isa<ConstantInt>(IdxOp))
8646 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8648 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8649 // Okay, we can handle this if the vector we are insertinting into is
8651 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8652 // If so, update the mask to reflect the inserted undef.
8653 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8656 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8657 if (isa<ConstantInt>(EI->getOperand(1)) &&
8658 EI->getOperand(0)->getType() == V->getType()) {
8659 unsigned ExtractedIdx =
8660 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8662 // This must be extracting from either LHS or RHS.
8663 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8664 // Okay, we can handle this if the vector we are insertinting into is
8666 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8667 // If so, update the mask to reflect the inserted value.
8668 if (EI->getOperand(0) == LHS) {
8669 Mask[InsertedIdx & (NumElts-1)] =
8670 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8672 assert(EI->getOperand(0) == RHS);
8673 Mask[InsertedIdx & (NumElts-1)] =
8674 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8683 // TODO: Handle shufflevector here!
8688 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8689 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8690 /// that computes V and the LHS value of the shuffle.
8691 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8693 assert(isa<PackedType>(V->getType()) &&
8694 (RHS == 0 || V->getType() == RHS->getType()) &&
8695 "Invalid shuffle!");
8696 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8698 if (isa<UndefValue>(V)) {
8699 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8701 } else if (isa<ConstantAggregateZero>(V)) {
8702 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8704 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8705 // If this is an insert of an extract from some other vector, include it.
8706 Value *VecOp = IEI->getOperand(0);
8707 Value *ScalarOp = IEI->getOperand(1);
8708 Value *IdxOp = IEI->getOperand(2);
8710 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8711 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8712 EI->getOperand(0)->getType() == V->getType()) {
8713 unsigned ExtractedIdx =
8714 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8715 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8717 // Either the extracted from or inserted into vector must be RHSVec,
8718 // otherwise we'd end up with a shuffle of three inputs.
8719 if (EI->getOperand(0) == RHS || RHS == 0) {
8720 RHS = EI->getOperand(0);
8721 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8722 Mask[InsertedIdx & (NumElts-1)] =
8723 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8728 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8729 // Everything but the extracted element is replaced with the RHS.
8730 for (unsigned i = 0; i != NumElts; ++i) {
8731 if (i != InsertedIdx)
8732 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8737 // If this insertelement is a chain that comes from exactly these two
8738 // vectors, return the vector and the effective shuffle.
8739 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8740 return EI->getOperand(0);
8745 // TODO: Handle shufflevector here!
8747 // Otherwise, can't do anything fancy. Return an identity vector.
8748 for (unsigned i = 0; i != NumElts; ++i)
8749 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8753 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8754 Value *VecOp = IE.getOperand(0);
8755 Value *ScalarOp = IE.getOperand(1);
8756 Value *IdxOp = IE.getOperand(2);
8758 // If the inserted element was extracted from some other vector, and if the
8759 // indexes are constant, try to turn this into a shufflevector operation.
8760 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8761 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8762 EI->getOperand(0)->getType() == IE.getType()) {
8763 unsigned NumVectorElts = IE.getType()->getNumElements();
8764 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8765 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8767 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8768 return ReplaceInstUsesWith(IE, VecOp);
8770 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8771 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8773 // If we are extracting a value from a vector, then inserting it right
8774 // back into the same place, just use the input vector.
8775 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8776 return ReplaceInstUsesWith(IE, VecOp);
8778 // We could theoretically do this for ANY input. However, doing so could
8779 // turn chains of insertelement instructions into a chain of shufflevector
8780 // instructions, and right now we do not merge shufflevectors. As such,
8781 // only do this in a situation where it is clear that there is benefit.
8782 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8783 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8784 // the values of VecOp, except then one read from EIOp0.
8785 // Build a new shuffle mask.
8786 std::vector<Constant*> Mask;
8787 if (isa<UndefValue>(VecOp))
8788 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8790 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8791 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8794 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8795 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8796 ConstantPacked::get(Mask));
8799 // If this insertelement isn't used by some other insertelement, turn it
8800 // (and any insertelements it points to), into one big shuffle.
8801 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8802 std::vector<Constant*> Mask;
8804 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8805 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8806 // We now have a shuffle of LHS, RHS, Mask.
8807 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8816 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8817 Value *LHS = SVI.getOperand(0);
8818 Value *RHS = SVI.getOperand(1);
8819 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8821 bool MadeChange = false;
8823 // Undefined shuffle mask -> undefined value.
8824 if (isa<UndefValue>(SVI.getOperand(2)))
8825 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8827 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8828 // the undef, change them to undefs.
8829 if (isa<UndefValue>(SVI.getOperand(1))) {
8830 // Scan to see if there are any references to the RHS. If so, replace them
8831 // with undef element refs and set MadeChange to true.
8832 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8833 if (Mask[i] >= e && Mask[i] != 2*e) {
8840 // Remap any references to RHS to use LHS.
8841 std::vector<Constant*> Elts;
8842 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8844 Elts.push_back(UndefValue::get(Type::Int32Ty));
8846 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8848 SVI.setOperand(2, ConstantPacked::get(Elts));
8852 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8853 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8854 if (LHS == RHS || isa<UndefValue>(LHS)) {
8855 if (isa<UndefValue>(LHS) && LHS == RHS) {
8856 // shuffle(undef,undef,mask) -> undef.
8857 return ReplaceInstUsesWith(SVI, LHS);
8860 // Remap any references to RHS to use LHS.
8861 std::vector<Constant*> Elts;
8862 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8864 Elts.push_back(UndefValue::get(Type::Int32Ty));
8866 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8867 (Mask[i] < e && isa<UndefValue>(LHS)))
8868 Mask[i] = 2*e; // Turn into undef.
8870 Mask[i] &= (e-1); // Force to LHS.
8871 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8874 SVI.setOperand(0, SVI.getOperand(1));
8875 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8876 SVI.setOperand(2, ConstantPacked::get(Elts));
8877 LHS = SVI.getOperand(0);
8878 RHS = SVI.getOperand(1);
8882 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8883 bool isLHSID = true, isRHSID = true;
8885 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8886 if (Mask[i] >= e*2) continue; // Ignore undef values.
8887 // Is this an identity shuffle of the LHS value?
8888 isLHSID &= (Mask[i] == i);
8890 // Is this an identity shuffle of the RHS value?
8891 isRHSID &= (Mask[i]-e == i);
8894 // Eliminate identity shuffles.
8895 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8896 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8898 // If the LHS is a shufflevector itself, see if we can combine it with this
8899 // one without producing an unusual shuffle. Here we are really conservative:
8900 // we are absolutely afraid of producing a shuffle mask not in the input
8901 // program, because the code gen may not be smart enough to turn a merged
8902 // shuffle into two specific shuffles: it may produce worse code. As such,
8903 // we only merge two shuffles if the result is one of the two input shuffle
8904 // masks. In this case, merging the shuffles just removes one instruction,
8905 // which we know is safe. This is good for things like turning:
8906 // (splat(splat)) -> splat.
8907 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8908 if (isa<UndefValue>(RHS)) {
8909 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8911 std::vector<unsigned> NewMask;
8912 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8914 NewMask.push_back(2*e);
8916 NewMask.push_back(LHSMask[Mask[i]]);
8918 // If the result mask is equal to the src shuffle or this shuffle mask, do
8920 if (NewMask == LHSMask || NewMask == Mask) {
8921 std::vector<Constant*> Elts;
8922 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8923 if (NewMask[i] >= e*2) {
8924 Elts.push_back(UndefValue::get(Type::Int32Ty));
8926 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
8929 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8930 LHSSVI->getOperand(1),
8931 ConstantPacked::get(Elts));
8936 return MadeChange ? &SVI : 0;
8941 void InstCombiner::removeFromWorkList(Instruction *I) {
8942 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8947 /// TryToSinkInstruction - Try to move the specified instruction from its
8948 /// current block into the beginning of DestBlock, which can only happen if it's
8949 /// safe to move the instruction past all of the instructions between it and the
8950 /// end of its block.
8951 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8952 assert(I->hasOneUse() && "Invariants didn't hold!");
8954 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8955 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
8957 // Do not sink alloca instructions out of the entry block.
8958 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
8961 // We can only sink load instructions if there is nothing between the load and
8962 // the end of block that could change the value.
8963 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8964 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
8966 if (Scan->mayWriteToMemory())
8970 BasicBlock::iterator InsertPos = DestBlock->begin();
8971 while (isa<PHINode>(InsertPos)) ++InsertPos;
8973 I->moveBefore(InsertPos);
8978 /// OptimizeConstantExpr - Given a constant expression and target data layout
8979 /// information, symbolically evaluate the constant expr to something simpler
8981 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
8984 Constant *Ptr = CE->getOperand(0);
8985 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
8986 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
8987 // If this is a constant expr gep that is effectively computing an
8988 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
8989 bool isFoldableGEP = true;
8990 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
8991 if (!isa<ConstantInt>(CE->getOperand(i)))
8992 isFoldableGEP = false;
8993 if (isFoldableGEP) {
8994 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
8995 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
8996 Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
8997 return ConstantExpr::getIntToPtr(C, CE->getType());
9005 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9006 /// all reachable code to the worklist.
9008 /// This has a couple of tricks to make the code faster and more powerful. In
9009 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9010 /// them to the worklist (this significantly speeds up instcombine on code where
9011 /// many instructions are dead or constant). Additionally, if we find a branch
9012 /// whose condition is a known constant, we only visit the reachable successors.
9014 static void AddReachableCodeToWorklist(BasicBlock *BB,
9015 std::set<BasicBlock*> &Visited,
9016 std::vector<Instruction*> &WorkList,
9017 const TargetData *TD) {
9018 // We have now visited this block! If we've already been here, bail out.
9019 if (!Visited.insert(BB).second) return;
9021 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9022 Instruction *Inst = BBI++;
9024 // DCE instruction if trivially dead.
9025 if (isInstructionTriviallyDead(Inst)) {
9027 DOUT << "IC: DCE: " << *Inst;
9028 Inst->eraseFromParent();
9032 // ConstantProp instruction if trivially constant.
9033 if (Constant *C = ConstantFoldInstruction(Inst)) {
9034 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9035 C = OptimizeConstantExpr(CE, TD);
9036 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9037 Inst->replaceAllUsesWith(C);
9039 Inst->eraseFromParent();
9043 WorkList.push_back(Inst);
9046 // Recursively visit successors. If this is a branch or switch on a constant,
9047 // only visit the reachable successor.
9048 TerminatorInst *TI = BB->getTerminator();
9049 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9050 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition()) &&
9051 BI->getCondition()->getType() == Type::Int1Ty) {
9052 bool CondVal = cast<ConstantInt>(BI->getCondition())->getBoolValue();
9053 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9057 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9058 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9059 // See if this is an explicit destination.
9060 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9061 if (SI->getCaseValue(i) == Cond) {
9062 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9066 // Otherwise it is the default destination.
9067 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9072 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9073 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9076 bool InstCombiner::runOnFunction(Function &F) {
9077 bool Changed = false;
9078 TD = &getAnalysis<TargetData>();
9081 // Do a depth-first traversal of the function, populate the worklist with
9082 // the reachable instructions. Ignore blocks that are not reachable. Keep
9083 // track of which blocks we visit.
9084 std::set<BasicBlock*> Visited;
9085 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9087 // Do a quick scan over the function. If we find any blocks that are
9088 // unreachable, remove any instructions inside of them. This prevents
9089 // the instcombine code from having to deal with some bad special cases.
9090 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9091 if (!Visited.count(BB)) {
9092 Instruction *Term = BB->getTerminator();
9093 while (Term != BB->begin()) { // Remove instrs bottom-up
9094 BasicBlock::iterator I = Term; --I;
9096 DOUT << "IC: DCE: " << *I;
9099 if (!I->use_empty())
9100 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9101 I->eraseFromParent();
9106 while (!WorkList.empty()) {
9107 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9108 WorkList.pop_back();
9110 // Check to see if we can DCE the instruction.
9111 if (isInstructionTriviallyDead(I)) {
9112 // Add operands to the worklist.
9113 if (I->getNumOperands() < 4)
9114 AddUsesToWorkList(*I);
9117 DOUT << "IC: DCE: " << *I;
9119 I->eraseFromParent();
9120 removeFromWorkList(I);
9124 // Instruction isn't dead, see if we can constant propagate it.
9125 if (Constant *C = ConstantFoldInstruction(I)) {
9126 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9127 C = OptimizeConstantExpr(CE, TD);
9128 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9130 // Add operands to the worklist.
9131 AddUsesToWorkList(*I);
9132 ReplaceInstUsesWith(*I, C);
9135 I->eraseFromParent();
9136 removeFromWorkList(I);
9140 // See if we can trivially sink this instruction to a successor basic block.
9141 if (I->hasOneUse()) {
9142 BasicBlock *BB = I->getParent();
9143 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9144 if (UserParent != BB) {
9145 bool UserIsSuccessor = false;
9146 // See if the user is one of our successors.
9147 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9148 if (*SI == UserParent) {
9149 UserIsSuccessor = true;
9153 // If the user is one of our immediate successors, and if that successor
9154 // only has us as a predecessors (we'd have to split the critical edge
9155 // otherwise), we can keep going.
9156 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9157 next(pred_begin(UserParent)) == pred_end(UserParent))
9158 // Okay, the CFG is simple enough, try to sink this instruction.
9159 Changed |= TryToSinkInstruction(I, UserParent);
9163 // Now that we have an instruction, try combining it to simplify it...
9164 if (Instruction *Result = visit(*I)) {
9166 // Should we replace the old instruction with a new one?
9168 DOUT << "IC: Old = " << *I
9169 << " New = " << *Result;
9171 // Everything uses the new instruction now.
9172 I->replaceAllUsesWith(Result);
9174 // Push the new instruction and any users onto the worklist.
9175 WorkList.push_back(Result);
9176 AddUsersToWorkList(*Result);
9178 // Move the name to the new instruction first...
9179 std::string OldName = I->getName(); I->setName("");
9180 Result->setName(OldName);
9182 // Insert the new instruction into the basic block...
9183 BasicBlock *InstParent = I->getParent();
9184 BasicBlock::iterator InsertPos = I;
9186 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9187 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9190 InstParent->getInstList().insert(InsertPos, Result);
9192 // Make sure that we reprocess all operands now that we reduced their
9194 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9195 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9196 WorkList.push_back(OpI);
9198 // Instructions can end up on the worklist more than once. Make sure
9199 // we do not process an instruction that has been deleted.
9200 removeFromWorkList(I);
9202 // Erase the old instruction.
9203 InstParent->getInstList().erase(I);
9205 DOUT << "IC: MOD = " << *I;
9207 // If the instruction was modified, it's possible that it is now dead.
9208 // if so, remove it.
9209 if (isInstructionTriviallyDead(I)) {
9210 // Make sure we process all operands now that we are reducing their
9212 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9213 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9214 WorkList.push_back(OpI);
9216 // Instructions may end up in the worklist more than once. Erase all
9217 // occurrences of this instruction.
9218 removeFromWorkList(I);
9219 I->eraseFromParent();
9221 WorkList.push_back(Result);
9222 AddUsersToWorkList(*Result);
9232 FunctionPass *llvm::createInstructionCombiningPass() {
9233 return new InstCombiner();