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/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/Statistic.h"
54 #include "llvm/ADT/STLExtras.h"
57 using namespace llvm::PatternMatch;
59 STATISTIC(NumCombined , "Number of insts combined");
60 STATISTIC(NumConstProp, "Number of constant folds");
61 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
62 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
63 STATISTIC(NumSunkInst , "Number of instructions sunk");
66 class VISIBILITY_HIDDEN InstCombiner
67 : public FunctionPass,
68 public InstVisitor<InstCombiner, Instruction*> {
69 // Worklist of all of the instructions that need to be simplified.
70 std::vector<Instruction*> WorkList;
73 /// AddUsersToWorkList - When an instruction is simplified, add all users of
74 /// the instruction to the work lists because they might get more simplified
77 void AddUsersToWorkList(Value &I) {
78 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
80 WorkList.push_back(cast<Instruction>(*UI));
83 /// AddUsesToWorkList - When an instruction is simplified, add operands to
84 /// the work lists because they might get more simplified now.
86 void AddUsesToWorkList(Instruction &I) {
87 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
88 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
89 WorkList.push_back(Op);
92 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
93 /// dead. Add all of its operands to the worklist, turning them into
94 /// undef's to reduce the number of uses of those instructions.
96 /// Return the specified operand before it is turned into an undef.
98 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
99 Value *R = I.getOperand(op);
101 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
102 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
103 WorkList.push_back(Op);
104 // Set the operand to undef to drop the use.
105 I.setOperand(i, UndefValue::get(Op->getType()));
111 // removeFromWorkList - remove all instances of I from the worklist.
112 void removeFromWorkList(Instruction *I);
114 virtual bool runOnFunction(Function &F);
116 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
117 AU.addRequired<TargetData>();
118 AU.addPreservedID(LCSSAID);
119 AU.setPreservesCFG();
122 TargetData &getTargetData() const { return *TD; }
124 // Visitation implementation - Implement instruction combining for different
125 // instruction types. The semantics are as follows:
127 // null - No change was made
128 // I - Change was made, I is still valid, I may be dead though
129 // otherwise - Change was made, replace I with returned instruction
131 Instruction *visitAdd(BinaryOperator &I);
132 Instruction *visitSub(BinaryOperator &I);
133 Instruction *visitMul(BinaryOperator &I);
134 Instruction *visitURem(BinaryOperator &I);
135 Instruction *visitSRem(BinaryOperator &I);
136 Instruction *visitFRem(BinaryOperator &I);
137 Instruction *commonRemTransforms(BinaryOperator &I);
138 Instruction *commonIRemTransforms(BinaryOperator &I);
139 Instruction *commonDivTransforms(BinaryOperator &I);
140 Instruction *commonIDivTransforms(BinaryOperator &I);
141 Instruction *visitUDiv(BinaryOperator &I);
142 Instruction *visitSDiv(BinaryOperator &I);
143 Instruction *visitFDiv(BinaryOperator &I);
144 Instruction *visitAnd(BinaryOperator &I);
145 Instruction *visitOr (BinaryOperator &I);
146 Instruction *visitXor(BinaryOperator &I);
147 Instruction *visitFCmpInst(FCmpInst &I);
148 Instruction *visitICmpInst(ICmpInst &I);
149 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
151 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
152 ICmpInst::Predicate Cond, Instruction &I);
153 Instruction *visitShiftInst(ShiftInst &I);
154 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
156 Instruction *commonCastTransforms(CastInst &CI);
157 Instruction *commonIntCastTransforms(CastInst &CI);
158 Instruction *visitTrunc(CastInst &CI);
159 Instruction *visitZExt(CastInst &CI);
160 Instruction *visitSExt(CastInst &CI);
161 Instruction *visitFPTrunc(CastInst &CI);
162 Instruction *visitFPExt(CastInst &CI);
163 Instruction *visitFPToUI(CastInst &CI);
164 Instruction *visitFPToSI(CastInst &CI);
165 Instruction *visitUIToFP(CastInst &CI);
166 Instruction *visitSIToFP(CastInst &CI);
167 Instruction *visitPtrToInt(CastInst &CI);
168 Instruction *visitIntToPtr(CastInst &CI);
169 Instruction *visitBitCast(CastInst &CI);
170 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
172 Instruction *visitSelectInst(SelectInst &CI);
173 Instruction *visitCallInst(CallInst &CI);
174 Instruction *visitInvokeInst(InvokeInst &II);
175 Instruction *visitPHINode(PHINode &PN);
176 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
177 Instruction *visitAllocationInst(AllocationInst &AI);
178 Instruction *visitFreeInst(FreeInst &FI);
179 Instruction *visitLoadInst(LoadInst &LI);
180 Instruction *visitStoreInst(StoreInst &SI);
181 Instruction *visitBranchInst(BranchInst &BI);
182 Instruction *visitSwitchInst(SwitchInst &SI);
183 Instruction *visitInsertElementInst(InsertElementInst &IE);
184 Instruction *visitExtractElementInst(ExtractElementInst &EI);
185 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
187 // visitInstruction - Specify what to return for unhandled instructions...
188 Instruction *visitInstruction(Instruction &I) { return 0; }
191 Instruction *visitCallSite(CallSite CS);
192 bool transformConstExprCastCall(CallSite CS);
195 // InsertNewInstBefore - insert an instruction New before instruction Old
196 // in the program. Add the new instruction to the worklist.
198 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
199 assert(New && New->getParent() == 0 &&
200 "New instruction already inserted into a basic block!");
201 BasicBlock *BB = Old.getParent();
202 BB->getInstList().insert(&Old, New); // Insert inst
203 WorkList.push_back(New); // Add to worklist
207 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
208 /// This also adds the cast to the worklist. Finally, this returns the
210 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
212 if (V->getType() == Ty) return V;
214 if (Constant *CV = dyn_cast<Constant>(V))
215 return ConstantExpr::getCast(opc, CV, Ty);
217 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
218 WorkList.push_back(C);
222 // ReplaceInstUsesWith - This method is to be used when an instruction is
223 // found to be dead, replacable with another preexisting expression. Here
224 // we add all uses of I to the worklist, replace all uses of I with the new
225 // value, then return I, so that the inst combiner will know that I was
228 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
229 AddUsersToWorkList(I); // Add all modified instrs to worklist
231 I.replaceAllUsesWith(V);
234 // If we are replacing the instruction with itself, this must be in a
235 // segment of unreachable code, so just clobber the instruction.
236 I.replaceAllUsesWith(UndefValue::get(I.getType()));
241 // UpdateValueUsesWith - This method is to be used when an value is
242 // found to be replacable with another preexisting expression or was
243 // updated. Here we add all uses of I to the worklist, replace all uses of
244 // I with the new value (unless the instruction was just updated), then
245 // return true, so that the inst combiner will know that I was modified.
247 bool UpdateValueUsesWith(Value *Old, Value *New) {
248 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
250 Old->replaceAllUsesWith(New);
251 if (Instruction *I = dyn_cast<Instruction>(Old))
252 WorkList.push_back(I);
253 if (Instruction *I = dyn_cast<Instruction>(New))
254 WorkList.push_back(I);
258 // EraseInstFromFunction - When dealing with an instruction that has side
259 // effects or produces a void value, we can't rely on DCE to delete the
260 // instruction. Instead, visit methods should return the value returned by
262 Instruction *EraseInstFromFunction(Instruction &I) {
263 assert(I.use_empty() && "Cannot erase instruction that is used!");
264 AddUsesToWorkList(I);
265 removeFromWorkList(&I);
267 return 0; // Don't do anything with FI
271 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
272 /// InsertBefore instruction. This is specialized a bit to avoid inserting
273 /// casts that are known to not do anything...
275 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
276 Value *V, const Type *DestTy,
277 Instruction *InsertBefore);
279 /// SimplifyCommutative - This performs a few simplifications for
280 /// commutative operators.
281 bool SimplifyCommutative(BinaryOperator &I);
283 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
284 /// most-complex to least-complex order.
285 bool SimplifyCompare(CmpInst &I);
287 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
288 uint64_t &KnownZero, uint64_t &KnownOne,
291 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
292 uint64_t &UndefElts, unsigned Depth = 0);
294 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
295 // PHI node as operand #0, see if we can fold the instruction into the PHI
296 // (which is only possible if all operands to the PHI are constants).
297 Instruction *FoldOpIntoPhi(Instruction &I);
299 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
300 // operator and they all are only used by the PHI, PHI together their
301 // inputs, and do the operation once, to the result of the PHI.
302 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
303 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
306 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
307 ConstantInt *AndRHS, BinaryOperator &TheAnd);
309 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
310 bool isSub, Instruction &I);
311 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
312 bool isSigned, bool Inside, Instruction &IB);
313 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
314 Instruction *MatchBSwap(BinaryOperator &I);
316 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
319 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
322 // getComplexity: Assign a complexity or rank value to LLVM Values...
323 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
324 static unsigned getComplexity(Value *V) {
325 if (isa<Instruction>(V)) {
326 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
330 if (isa<Argument>(V)) return 3;
331 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
334 // isOnlyUse - Return true if this instruction will be deleted if we stop using
336 static bool isOnlyUse(Value *V) {
337 return V->hasOneUse() || isa<Constant>(V);
340 // getPromotedType - Return the specified type promoted as it would be to pass
341 // though a va_arg area...
342 static const Type *getPromotedType(const Type *Ty) {
343 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
344 if (ITy->getBitWidth() < 32)
345 return Type::Int32Ty;
346 } else if (Ty == Type::FloatTy)
347 return Type::DoubleTy;
351 /// getBitCastOperand - If the specified operand is a CastInst or a constant
352 /// expression bitcast, return the operand value, otherwise return null.
353 static Value *getBitCastOperand(Value *V) {
354 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
355 return I->getOperand(0);
356 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
357 if (CE->getOpcode() == Instruction::BitCast)
358 return CE->getOperand(0);
362 /// This function is a wrapper around CastInst::isEliminableCastPair. It
363 /// simply extracts arguments and returns what that function returns.
364 /// @Determine if it is valid to eliminate a Convert pair
365 static Instruction::CastOps
366 isEliminableCastPair(
367 const CastInst *CI, ///< The first cast instruction
368 unsigned opcode, ///< The opcode of the second cast instruction
369 const Type *DstTy, ///< The target type for the second cast instruction
370 TargetData *TD ///< The target data for pointer size
373 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
374 const Type *MidTy = CI->getType(); // B from above
376 // Get the opcodes of the two Cast instructions
377 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
378 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
380 return Instruction::CastOps(
381 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
382 DstTy, TD->getIntPtrType()));
385 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
386 /// in any code being generated. It does not require codegen if V is simple
387 /// enough or if the cast can be folded into other casts.
388 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
389 const Type *Ty, TargetData *TD) {
390 if (V->getType() == Ty || isa<Constant>(V)) return false;
392 // If this is another cast that can be eliminated, it isn't codegen either.
393 if (const CastInst *CI = dyn_cast<CastInst>(V))
394 if (isEliminableCastPair(CI, opcode, Ty, TD))
399 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
400 /// InsertBefore instruction. This is specialized a bit to avoid inserting
401 /// casts that are known to not do anything...
403 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
404 Value *V, const Type *DestTy,
405 Instruction *InsertBefore) {
406 if (V->getType() == DestTy) return V;
407 if (Constant *C = dyn_cast<Constant>(V))
408 return ConstantExpr::getCast(opcode, C, DestTy);
410 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
413 // SimplifyCommutative - This performs a few simplifications for commutative
416 // 1. Order operands such that they are listed from right (least complex) to
417 // left (most complex). This puts constants before unary operators before
420 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
421 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
423 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
424 bool Changed = false;
425 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
426 Changed = !I.swapOperands();
428 if (!I.isAssociative()) return Changed;
429 Instruction::BinaryOps Opcode = I.getOpcode();
430 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
431 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
432 if (isa<Constant>(I.getOperand(1))) {
433 Constant *Folded = ConstantExpr::get(I.getOpcode(),
434 cast<Constant>(I.getOperand(1)),
435 cast<Constant>(Op->getOperand(1)));
436 I.setOperand(0, Op->getOperand(0));
437 I.setOperand(1, Folded);
439 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
440 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
441 isOnlyUse(Op) && isOnlyUse(Op1)) {
442 Constant *C1 = cast<Constant>(Op->getOperand(1));
443 Constant *C2 = cast<Constant>(Op1->getOperand(1));
445 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
446 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
447 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
450 WorkList.push_back(New);
451 I.setOperand(0, New);
452 I.setOperand(1, Folded);
459 /// SimplifyCompare - For a CmpInst this function just orders the operands
460 /// so that theyare listed from right (least complex) to left (most complex).
461 /// This puts constants before unary operators before binary operators.
462 bool InstCombiner::SimplifyCompare(CmpInst &I) {
463 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
466 // Compare instructions are not associative so there's nothing else we can do.
470 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
471 // if the LHS is a constant zero (which is the 'negate' form).
473 static inline Value *dyn_castNegVal(Value *V) {
474 if (BinaryOperator::isNeg(V))
475 return BinaryOperator::getNegArgument(V);
477 // Constants can be considered to be negated values if they can be folded.
478 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
479 return ConstantExpr::getNeg(C);
483 static inline Value *dyn_castNotVal(Value *V) {
484 if (BinaryOperator::isNot(V))
485 return BinaryOperator::getNotArgument(V);
487 // Constants can be considered to be not'ed values...
488 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
489 return ConstantExpr::getNot(C);
493 // dyn_castFoldableMul - If this value is a multiply that can be folded into
494 // other computations (because it has a constant operand), return the
495 // non-constant operand of the multiply, and set CST to point to the multiplier.
496 // Otherwise, return null.
498 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
499 if (V->hasOneUse() && V->getType()->isInteger())
500 if (Instruction *I = dyn_cast<Instruction>(V)) {
501 if (I->getOpcode() == Instruction::Mul)
502 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
503 return I->getOperand(0);
504 if (I->getOpcode() == Instruction::Shl)
505 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
506 // The multiplier is really 1 << CST.
507 Constant *One = ConstantInt::get(V->getType(), 1);
508 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
509 return I->getOperand(0);
515 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
516 /// expression, return it.
517 static User *dyn_castGetElementPtr(Value *V) {
518 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
519 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
520 if (CE->getOpcode() == Instruction::GetElementPtr)
521 return cast<User>(V);
525 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
526 static ConstantInt *AddOne(ConstantInt *C) {
527 return cast<ConstantInt>(ConstantExpr::getAdd(C,
528 ConstantInt::get(C->getType(), 1)));
530 static ConstantInt *SubOne(ConstantInt *C) {
531 return cast<ConstantInt>(ConstantExpr::getSub(C,
532 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 &= cast<IntegerType>(V->getType())->getBitMask();
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->isInteger()) {
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 IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
637 uint64_t NotIn = ~SrcTy->getBitMask();
638 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
640 Mask &= SrcTy->getBitMask();
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 IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
650 uint64_t NotIn = ~SrcTy->getBitMask();
651 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
653 Mask &= SrcTy->getBitMask();
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 = cast<IntegerType>(Ty)->getBitMask();
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 = cast<IntegerType>(Ty)->getBitMask();
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 = cast<IntegerType>(V->getType())->getBitMask();
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 &= cast<IntegerType>(V->getType())->getBitMask();
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()->isInteger())
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 IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1016 uint64_t NotIn = ~SrcTy->getBitMask();
1017 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1019 DemandedMask &= SrcTy->getBitMask();
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 IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1031 uint64_t NotIn = ~SrcTy->getBitMask();
1032 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1034 // Get the sign bit for the source type
1035 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1036 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
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 = cast<IntegerType>(I->getType())->getBitMask();
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 = cast<IntegerType>(I->getType())->getBitMask();
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 /// AssociativeOpt - Perform an optimization on an associative operator. This
1462 /// function is designed to check a chain of associative operators for a
1463 /// potential to apply a certain optimization. Since the optimization may be
1464 /// applicable if the expression was reassociated, this checks the chain, then
1465 /// reassociates the expression as necessary to expose the optimization
1466 /// opportunity. This makes use of a special Functor, which must define
1467 /// 'shouldApply' and 'apply' methods.
1469 template<typename Functor>
1470 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1471 unsigned Opcode = Root.getOpcode();
1472 Value *LHS = Root.getOperand(0);
1474 // Quick check, see if the immediate LHS matches...
1475 if (F.shouldApply(LHS))
1476 return F.apply(Root);
1478 // Otherwise, if the LHS is not of the same opcode as the root, return.
1479 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1480 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1481 // Should we apply this transform to the RHS?
1482 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1484 // If not to the RHS, check to see if we should apply to the LHS...
1485 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1486 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1490 // If the functor wants to apply the optimization to the RHS of LHSI,
1491 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1493 BasicBlock *BB = Root.getParent();
1495 // Now all of the instructions are in the current basic block, go ahead
1496 // and perform the reassociation.
1497 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1499 // First move the selected RHS to the LHS of the root...
1500 Root.setOperand(0, LHSI->getOperand(1));
1502 // Make what used to be the LHS of the root be the user of the root...
1503 Value *ExtraOperand = TmpLHSI->getOperand(1);
1504 if (&Root == TmpLHSI) {
1505 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1508 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1509 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1510 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1511 BasicBlock::iterator ARI = &Root; ++ARI;
1512 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1515 // Now propagate the ExtraOperand down the chain of instructions until we
1517 while (TmpLHSI != LHSI) {
1518 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1519 // Move the instruction to immediately before the chain we are
1520 // constructing to avoid breaking dominance properties.
1521 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1522 BB->getInstList().insert(ARI, NextLHSI);
1525 Value *NextOp = NextLHSI->getOperand(1);
1526 NextLHSI->setOperand(1, ExtraOperand);
1528 ExtraOperand = NextOp;
1531 // Now that the instructions are reassociated, have the functor perform
1532 // the transformation...
1533 return F.apply(Root);
1536 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1542 // AddRHS - Implements: X + X --> X << 1
1545 AddRHS(Value *rhs) : RHS(rhs) {}
1546 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1547 Instruction *apply(BinaryOperator &Add) const {
1548 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1549 ConstantInt::get(Type::Int8Ty, 1));
1553 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1555 struct AddMaskingAnd {
1557 AddMaskingAnd(Constant *c) : C2(c) {}
1558 bool shouldApply(Value *LHS) const {
1560 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1561 ConstantExpr::getAnd(C1, C2)->isNullValue();
1563 Instruction *apply(BinaryOperator &Add) const {
1564 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1568 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1570 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1571 if (Constant *SOC = dyn_cast<Constant>(SO))
1572 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1574 return IC->InsertNewInstBefore(CastInst::create(
1575 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1578 // Figure out if the constant is the left or the right argument.
1579 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1580 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1582 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1584 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1585 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1588 Value *Op0 = SO, *Op1 = ConstOperand;
1590 std::swap(Op0, Op1);
1592 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1593 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1594 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1595 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1596 SO->getName()+".cmp");
1597 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1598 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1600 assert(0 && "Unknown binary instruction type!");
1603 return IC->InsertNewInstBefore(New, I);
1606 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1607 // constant as the other operand, try to fold the binary operator into the
1608 // select arguments. This also works for Cast instructions, which obviously do
1609 // not have a second operand.
1610 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1612 // Don't modify shared select instructions
1613 if (!SI->hasOneUse()) return 0;
1614 Value *TV = SI->getOperand(1);
1615 Value *FV = SI->getOperand(2);
1617 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1618 // Bool selects with constant operands can be folded to logical ops.
1619 if (SI->getType() == Type::Int1Ty) return 0;
1621 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1622 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1624 return new SelectInst(SI->getCondition(), SelectTrueVal,
1631 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1632 /// node as operand #0, see if we can fold the instruction into the PHI (which
1633 /// is only possible if all operands to the PHI are constants).
1634 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1635 PHINode *PN = cast<PHINode>(I.getOperand(0));
1636 unsigned NumPHIValues = PN->getNumIncomingValues();
1637 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1639 // Check to see if all of the operands of the PHI are constants. If there is
1640 // one non-constant value, remember the BB it is. If there is more than one
1642 BasicBlock *NonConstBB = 0;
1643 for (unsigned i = 0; i != NumPHIValues; ++i)
1644 if (!isa<Constant>(PN->getIncomingValue(i))) {
1645 if (NonConstBB) return 0; // More than one non-const value.
1646 NonConstBB = PN->getIncomingBlock(i);
1648 // If the incoming non-constant value is in I's block, we have an infinite
1650 if (NonConstBB == I.getParent())
1654 // If there is exactly one non-constant value, we can insert a copy of the
1655 // operation in that block. However, if this is a critical edge, we would be
1656 // inserting the computation one some other paths (e.g. inside a loop). Only
1657 // do this if the pred block is unconditionally branching into the phi block.
1659 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1660 if (!BI || !BI->isUnconditional()) return 0;
1663 // Okay, we can do the transformation: create the new PHI node.
1664 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1666 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1667 InsertNewInstBefore(NewPN, *PN);
1669 // Next, add all of the operands to the PHI.
1670 if (I.getNumOperands() == 2) {
1671 Constant *C = cast<Constant>(I.getOperand(1));
1672 for (unsigned i = 0; i != NumPHIValues; ++i) {
1674 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1675 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1676 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1678 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1680 assert(PN->getIncomingBlock(i) == NonConstBB);
1681 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1682 InV = BinaryOperator::create(BO->getOpcode(),
1683 PN->getIncomingValue(i), C, "phitmp",
1684 NonConstBB->getTerminator());
1685 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1686 InV = CmpInst::create(CI->getOpcode(),
1688 PN->getIncomingValue(i), C, "phitmp",
1689 NonConstBB->getTerminator());
1690 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1691 InV = new ShiftInst(SI->getOpcode(),
1692 PN->getIncomingValue(i), C, "phitmp",
1693 NonConstBB->getTerminator());
1695 assert(0 && "Unknown binop!");
1697 WorkList.push_back(cast<Instruction>(InV));
1699 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1702 CastInst *CI = cast<CastInst>(&I);
1703 const Type *RetTy = CI->getType();
1704 for (unsigned i = 0; i != NumPHIValues; ++i) {
1706 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1707 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1709 assert(PN->getIncomingBlock(i) == NonConstBB);
1710 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1711 I.getType(), "phitmp",
1712 NonConstBB->getTerminator());
1713 WorkList.push_back(cast<Instruction>(InV));
1715 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1718 return ReplaceInstUsesWith(I, NewPN);
1721 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1722 bool Changed = SimplifyCommutative(I);
1723 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1725 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1726 // X + undef -> undef
1727 if (isa<UndefValue>(RHS))
1728 return ReplaceInstUsesWith(I, RHS);
1731 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1732 if (RHSC->isNullValue())
1733 return ReplaceInstUsesWith(I, LHS);
1734 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1735 if (CFP->isExactlyValue(-0.0))
1736 return ReplaceInstUsesWith(I, LHS);
1739 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1740 // X + (signbit) --> X ^ signbit
1741 uint64_t Val = CI->getZExtValue();
1742 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1743 return BinaryOperator::createXor(LHS, RHS);
1745 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1746 // (X & 254)+1 -> (X&254)|1
1747 uint64_t KnownZero, KnownOne;
1748 if (!isa<PackedType>(I.getType()) &&
1749 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
1750 KnownZero, KnownOne))
1754 if (isa<PHINode>(LHS))
1755 if (Instruction *NV = FoldOpIntoPhi(I))
1758 ConstantInt *XorRHS = 0;
1760 if (isa<ConstantInt>(RHSC) &&
1761 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1762 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1763 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1764 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1766 uint64_t C0080Val = 1ULL << 31;
1767 int64_t CFF80Val = -C0080Val;
1770 if (TySizeBits > Size) {
1772 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1773 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1774 if (RHSSExt == CFF80Val) {
1775 if (XorRHS->getZExtValue() == C0080Val)
1777 } else if (RHSZExt == C0080Val) {
1778 if (XorRHS->getSExtValue() == CFF80Val)
1782 // This is a sign extend if the top bits are known zero.
1783 uint64_t Mask = ~0ULL;
1784 Mask <<= 64-(TySizeBits-Size);
1785 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
1786 if (!MaskedValueIsZero(XorLHS, Mask))
1787 Size = 0; // Not a sign ext, but can't be any others either.
1794 } while (Size >= 8);
1797 const Type *MiddleType = 0;
1800 case 32: MiddleType = Type::Int32Ty; break;
1801 case 16: MiddleType = Type::Int16Ty; break;
1802 case 8: MiddleType = Type::Int8Ty; break;
1805 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1806 InsertNewInstBefore(NewTrunc, I);
1807 return new SExtInst(NewTrunc, I.getType());
1813 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1814 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1816 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1817 if (RHSI->getOpcode() == Instruction::Sub)
1818 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1819 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1821 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1822 if (LHSI->getOpcode() == Instruction::Sub)
1823 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1824 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1829 if (Value *V = dyn_castNegVal(LHS))
1830 return BinaryOperator::createSub(RHS, V);
1833 if (!isa<Constant>(RHS))
1834 if (Value *V = dyn_castNegVal(RHS))
1835 return BinaryOperator::createSub(LHS, V);
1839 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1840 if (X == RHS) // X*C + X --> X * (C+1)
1841 return BinaryOperator::createMul(RHS, AddOne(C2));
1843 // X*C1 + X*C2 --> X * (C1+C2)
1845 if (X == dyn_castFoldableMul(RHS, C1))
1846 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1849 // X + X*C --> X * (C+1)
1850 if (dyn_castFoldableMul(RHS, C2) == LHS)
1851 return BinaryOperator::createMul(LHS, AddOne(C2));
1853 // X + ~X --> -1 since ~X = -X-1
1854 if (dyn_castNotVal(LHS) == RHS ||
1855 dyn_castNotVal(RHS) == LHS)
1856 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1859 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1860 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1861 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1864 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1866 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1867 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1868 return BinaryOperator::createSub(C, X);
1871 // (X & FF00) + xx00 -> (X+xx00) & FF00
1872 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1873 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1874 if (Anded == CRHS) {
1875 // See if all bits from the first bit set in the Add RHS up are included
1876 // in the mask. First, get the rightmost bit.
1877 uint64_t AddRHSV = CRHS->getZExtValue();
1879 // Form a mask of all bits from the lowest bit added through the top.
1880 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1881 AddRHSHighBits &= C2->getType()->getBitMask();
1883 // See if the and mask includes all of these bits.
1884 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1886 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1887 // Okay, the xform is safe. Insert the new add pronto.
1888 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1889 LHS->getName()), I);
1890 return BinaryOperator::createAnd(NewAdd, C2);
1895 // Try to fold constant add into select arguments.
1896 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1897 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1901 // add (cast *A to intptrtype) B ->
1902 // cast (GEP (cast *A to sbyte*) B) ->
1905 CastInst *CI = dyn_cast<CastInst>(LHS);
1908 CI = dyn_cast<CastInst>(RHS);
1911 if (CI && CI->getType()->isSized() &&
1912 (CI->getType()->getPrimitiveSizeInBits() ==
1913 TD->getIntPtrType()->getPrimitiveSizeInBits())
1914 && isa<PointerType>(CI->getOperand(0)->getType())) {
1915 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1916 PointerType::get(Type::Int8Ty), I);
1917 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1918 return new PtrToIntInst(I2, CI->getType());
1922 return Changed ? &I : 0;
1925 // isSignBit - Return true if the value represented by the constant only has the
1926 // highest order bit set.
1927 static bool isSignBit(ConstantInt *CI) {
1928 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1929 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1932 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1933 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1935 if (Op0 == Op1) // sub X, X -> 0
1936 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1938 // If this is a 'B = x-(-A)', change to B = x+A...
1939 if (Value *V = dyn_castNegVal(Op1))
1940 return BinaryOperator::createAdd(Op0, V);
1942 if (isa<UndefValue>(Op0))
1943 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1944 if (isa<UndefValue>(Op1))
1945 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1947 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1948 // Replace (-1 - A) with (~A)...
1949 if (C->isAllOnesValue())
1950 return BinaryOperator::createNot(Op1);
1952 // C - ~X == X + (1+C)
1954 if (match(Op1, m_Not(m_Value(X))))
1955 return BinaryOperator::createAdd(X,
1956 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1957 // -(X >>u 31) -> (X >>s 31)
1958 // -(X >>s 31) -> (X >>u 31)
1959 if (C->isNullValue()) {
1960 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op1))
1961 if (SI->getOpcode() == Instruction::LShr) {
1962 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1963 // Check to see if we are shifting out everything but the sign bit.
1964 if (CU->getZExtValue() ==
1965 SI->getType()->getPrimitiveSizeInBits()-1) {
1966 // Ok, the transformation is safe. Insert AShr.
1967 return new ShiftInst(Instruction::AShr, SI->getOperand(0), CU,
1972 else if (SI->getOpcode() == Instruction::AShr) {
1973 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1974 // Check to see if we are shifting out everything but the sign bit.
1975 if (CU->getZExtValue() ==
1976 SI->getType()->getPrimitiveSizeInBits()-1) {
1977 // Ok, the transformation is safe. Insert LShr.
1978 return new ShiftInst(Instruction::LShr, SI->getOperand(0), CU,
1985 // Try to fold constant sub into select arguments.
1986 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1987 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1990 if (isa<PHINode>(Op0))
1991 if (Instruction *NV = FoldOpIntoPhi(I))
1995 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1996 if (Op1I->getOpcode() == Instruction::Add &&
1997 !Op0->getType()->isFPOrFPVector()) {
1998 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1999 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2000 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2001 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2002 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2003 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2004 // C1-(X+C2) --> (C1-C2)-X
2005 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2006 Op1I->getOperand(0));
2010 if (Op1I->hasOneUse()) {
2011 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2012 // is not used by anyone else...
2014 if (Op1I->getOpcode() == Instruction::Sub &&
2015 !Op1I->getType()->isFPOrFPVector()) {
2016 // Swap the two operands of the subexpr...
2017 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2018 Op1I->setOperand(0, IIOp1);
2019 Op1I->setOperand(1, IIOp0);
2021 // Create the new top level add instruction...
2022 return BinaryOperator::createAdd(Op0, Op1);
2025 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2027 if (Op1I->getOpcode() == Instruction::And &&
2028 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2029 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2032 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2033 return BinaryOperator::createAnd(Op0, NewNot);
2036 // 0 - (X sdiv C) -> (X sdiv -C)
2037 if (Op1I->getOpcode() == Instruction::SDiv)
2038 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2039 if (CSI->isNullValue())
2040 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2041 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2042 ConstantExpr::getNeg(DivRHS));
2044 // X - X*C --> X * (1-C)
2045 ConstantInt *C2 = 0;
2046 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2048 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2049 return BinaryOperator::createMul(Op0, CP1);
2054 if (!Op0->getType()->isFPOrFPVector())
2055 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2056 if (Op0I->getOpcode() == Instruction::Add) {
2057 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2058 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2059 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2060 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2061 } else if (Op0I->getOpcode() == Instruction::Sub) {
2062 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2063 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2067 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2068 if (X == Op1) { // X*C - X --> X * (C-1)
2069 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2070 return BinaryOperator::createMul(Op1, CP1);
2073 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2074 if (X == dyn_castFoldableMul(Op1, C2))
2075 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2080 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2081 /// really just returns true if the most significant (sign) bit is set.
2082 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2084 case ICmpInst::ICMP_SLT:
2085 // True if LHS s< RHS and RHS == 0
2086 return RHS->isNullValue();
2087 case ICmpInst::ICMP_SLE:
2088 // True if LHS s<= RHS and RHS == -1
2089 return RHS->isAllOnesValue();
2090 case ICmpInst::ICMP_UGE:
2091 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2092 return RHS->getZExtValue() == (1ULL <<
2093 (RHS->getType()->getPrimitiveSizeInBits()-1));
2094 case ICmpInst::ICMP_UGT:
2095 // True if LHS u> RHS and RHS == high-bit-mask - 1
2096 return RHS->getZExtValue() ==
2097 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2103 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2104 bool Changed = SimplifyCommutative(I);
2105 Value *Op0 = I.getOperand(0);
2107 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2108 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2110 // Simplify mul instructions with a constant RHS...
2111 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2112 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2114 // ((X << C1)*C2) == (X * (C2 << C1))
2115 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2116 if (SI->getOpcode() == Instruction::Shl)
2117 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2118 return BinaryOperator::createMul(SI->getOperand(0),
2119 ConstantExpr::getShl(CI, ShOp));
2121 if (CI->isNullValue())
2122 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2123 if (CI->equalsInt(1)) // X * 1 == X
2124 return ReplaceInstUsesWith(I, Op0);
2125 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2126 return BinaryOperator::createNeg(Op0, I.getName());
2128 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2129 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2130 uint64_t C = Log2_64(Val);
2131 return new ShiftInst(Instruction::Shl, Op0,
2132 ConstantInt::get(Type::Int8Ty, C));
2134 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2135 if (Op1F->isNullValue())
2136 return ReplaceInstUsesWith(I, Op1);
2138 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2139 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2140 if (Op1F->getValue() == 1.0)
2141 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2144 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2145 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2146 isa<ConstantInt>(Op0I->getOperand(1))) {
2147 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2148 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2150 InsertNewInstBefore(Add, I);
2151 Value *C1C2 = ConstantExpr::getMul(Op1,
2152 cast<Constant>(Op0I->getOperand(1)));
2153 return BinaryOperator::createAdd(Add, C1C2);
2157 // Try to fold constant mul into select arguments.
2158 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2159 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2162 if (isa<PHINode>(Op0))
2163 if (Instruction *NV = FoldOpIntoPhi(I))
2167 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2168 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2169 return BinaryOperator::createMul(Op0v, Op1v);
2171 // If one of the operands of the multiply is a cast from a boolean value, then
2172 // we know the bool is either zero or one, so this is a 'masking' multiply.
2173 // See if we can simplify things based on how the boolean was originally
2175 CastInst *BoolCast = 0;
2176 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2177 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2180 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2181 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2184 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2185 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2186 const Type *SCOpTy = SCIOp0->getType();
2188 // If the icmp is true iff the sign bit of X is set, then convert this
2189 // multiply into a shift/and combination.
2190 if (isa<ConstantInt>(SCIOp1) &&
2191 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2192 // Shift the X value right to turn it into "all signbits".
2193 Constant *Amt = ConstantInt::get(Type::Int8Ty,
2194 SCOpTy->getPrimitiveSizeInBits()-1);
2196 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2197 BoolCast->getOperand(0)->getName()+
2200 // If the multiply type is not the same as the source type, sign extend
2201 // or truncate to the multiply type.
2202 if (I.getType() != V->getType()) {
2203 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2204 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2205 Instruction::CastOps opcode =
2206 (SrcBits == DstBits ? Instruction::BitCast :
2207 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2208 V = InsertCastBefore(opcode, V, I.getType(), I);
2211 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2212 return BinaryOperator::createAnd(V, OtherOp);
2217 return Changed ? &I : 0;
2220 /// This function implements the transforms on div instructions that work
2221 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2222 /// used by the visitors to those instructions.
2223 /// @brief Transforms common to all three div instructions
2224 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2225 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2228 if (isa<UndefValue>(Op0))
2229 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2231 // X / undef -> undef
2232 if (isa<UndefValue>(Op1))
2233 return ReplaceInstUsesWith(I, Op1);
2235 // Handle cases involving: div X, (select Cond, Y, Z)
2236 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2237 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2238 // same basic block, then we replace the select with Y, and the condition
2239 // of the select with false (if the cond value is in the same BB). If the
2240 // select has uses other than the div, this allows them to be simplified
2241 // also. Note that div X, Y is just as good as div X, 0 (undef)
2242 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2243 if (ST->isNullValue()) {
2244 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2245 if (CondI && CondI->getParent() == I.getParent())
2246 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2247 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2248 I.setOperand(1, SI->getOperand(2));
2250 UpdateValueUsesWith(SI, SI->getOperand(2));
2254 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2255 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2256 if (ST->isNullValue()) {
2257 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2258 if (CondI && CondI->getParent() == I.getParent())
2259 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2260 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2261 I.setOperand(1, SI->getOperand(1));
2263 UpdateValueUsesWith(SI, SI->getOperand(1));
2271 /// This function implements the transforms common to both integer division
2272 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2273 /// division instructions.
2274 /// @brief Common integer divide transforms
2275 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2276 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2278 if (Instruction *Common = commonDivTransforms(I))
2281 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2283 if (RHS->equalsInt(1))
2284 return ReplaceInstUsesWith(I, Op0);
2286 // (X / C1) / C2 -> X / (C1*C2)
2287 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2288 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2289 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2290 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2291 ConstantExpr::getMul(RHS, LHSRHS));
2294 if (!RHS->isNullValue()) { // avoid X udiv 0
2295 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2296 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2298 if (isa<PHINode>(Op0))
2299 if (Instruction *NV = FoldOpIntoPhi(I))
2304 // 0 / X == 0, we don't need to preserve faults!
2305 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2306 if (LHS->equalsInt(0))
2307 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2312 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2313 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2315 // Handle the integer div common cases
2316 if (Instruction *Common = commonIDivTransforms(I))
2319 // X udiv C^2 -> X >> C
2320 // Check to see if this is an unsigned division with an exact power of 2,
2321 // if so, convert to a right shift.
2322 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2323 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2324 if (isPowerOf2_64(Val)) {
2325 uint64_t ShiftAmt = Log2_64(Val);
2326 return new ShiftInst(Instruction::LShr, Op0,
2327 ConstantInt::get(Type::Int8Ty, ShiftAmt));
2331 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2332 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2333 if (RHSI->getOpcode() == Instruction::Shl &&
2334 isa<ConstantInt>(RHSI->getOperand(0))) {
2335 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2336 if (isPowerOf2_64(C1)) {
2337 Value *N = RHSI->getOperand(1);
2338 const Type *NTy = N->getType();
2339 if (uint64_t C2 = Log2_64(C1)) {
2340 Constant *C2V = ConstantInt::get(NTy, C2);
2341 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2343 return new ShiftInst(Instruction::LShr, Op0, N);
2348 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2349 // where C1&C2 are powers of two.
2350 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2351 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2352 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2353 if (!STO->isNullValue() && !STO->isNullValue()) {
2354 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2355 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2356 // Compute the shift amounts
2357 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2358 // Construct the "on true" case of the select
2359 Constant *TC = ConstantInt::get(Type::Int8Ty, TSA);
2361 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2362 TSI = InsertNewInstBefore(TSI, I);
2364 // Construct the "on false" case of the select
2365 Constant *FC = ConstantInt::get(Type::Int8Ty, FSA);
2367 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2368 FSI = InsertNewInstBefore(FSI, I);
2370 // construct the select instruction and return it.
2371 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2378 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2379 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2381 // Handle the integer div common cases
2382 if (Instruction *Common = commonIDivTransforms(I))
2385 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2387 if (RHS->isAllOnesValue())
2388 return BinaryOperator::createNeg(Op0);
2391 if (Value *LHSNeg = dyn_castNegVal(Op0))
2392 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2395 // If the sign bits of both operands are zero (i.e. we can prove they are
2396 // unsigned inputs), turn this into a udiv.
2397 if (I.getType()->isInteger()) {
2398 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2399 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2400 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2407 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2408 return commonDivTransforms(I);
2411 /// GetFactor - If we can prove that the specified value is at least a multiple
2412 /// of some factor, return that factor.
2413 static Constant *GetFactor(Value *V) {
2414 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2417 // Unless we can be tricky, we know this is a multiple of 1.
2418 Constant *Result = ConstantInt::get(V->getType(), 1);
2420 Instruction *I = dyn_cast<Instruction>(V);
2421 if (!I) return Result;
2423 if (I->getOpcode() == Instruction::Mul) {
2424 // Handle multiplies by a constant, etc.
2425 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2426 GetFactor(I->getOperand(1)));
2427 } else if (I->getOpcode() == Instruction::Shl) {
2428 // (X<<C) -> X * (1 << C)
2429 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2430 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2431 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2433 } else if (I->getOpcode() == Instruction::And) {
2434 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2435 // X & 0xFFF0 is known to be a multiple of 16.
2436 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2437 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2438 return ConstantExpr::getShl(Result,
2439 ConstantInt::get(Type::Int8Ty, Zeros));
2441 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2442 // Only handle int->int casts.
2443 if (!CI->isIntegerCast())
2445 Value *Op = CI->getOperand(0);
2446 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2451 /// This function implements the transforms on rem instructions that work
2452 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2453 /// is used by the visitors to those instructions.
2454 /// @brief Transforms common to all three rem instructions
2455 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2456 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2458 // 0 % X == 0, we don't need to preserve faults!
2459 if (Constant *LHS = dyn_cast<Constant>(Op0))
2460 if (LHS->isNullValue())
2461 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2463 if (isa<UndefValue>(Op0)) // undef % X -> 0
2464 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2465 if (isa<UndefValue>(Op1))
2466 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2468 // Handle cases involving: rem X, (select Cond, Y, Z)
2469 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2470 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2471 // the same basic block, then we replace the select with Y, and the
2472 // condition of the select with false (if the cond value is in the same
2473 // BB). If the select has uses other than the div, this allows them to be
2475 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2476 if (ST->isNullValue()) {
2477 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2478 if (CondI && CondI->getParent() == I.getParent())
2479 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2480 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2481 I.setOperand(1, SI->getOperand(2));
2483 UpdateValueUsesWith(SI, SI->getOperand(2));
2486 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2487 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2488 if (ST->isNullValue()) {
2489 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2490 if (CondI && CondI->getParent() == I.getParent())
2491 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2492 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2493 I.setOperand(1, SI->getOperand(1));
2495 UpdateValueUsesWith(SI, SI->getOperand(1));
2503 /// This function implements the transforms common to both integer remainder
2504 /// instructions (urem and srem). It is called by the visitors to those integer
2505 /// remainder instructions.
2506 /// @brief Common integer remainder transforms
2507 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2508 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2510 if (Instruction *common = commonRemTransforms(I))
2513 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2514 // X % 0 == undef, we don't need to preserve faults!
2515 if (RHS->equalsInt(0))
2516 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2518 if (RHS->equalsInt(1)) // X % 1 == 0
2519 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2521 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2522 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2523 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2525 } else if (isa<PHINode>(Op0I)) {
2526 if (Instruction *NV = FoldOpIntoPhi(I))
2529 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2530 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2531 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2538 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2539 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2541 if (Instruction *common = commonIRemTransforms(I))
2544 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2545 // X urem C^2 -> X and C
2546 // Check to see if this is an unsigned remainder with an exact power of 2,
2547 // if so, convert to a bitwise and.
2548 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2549 if (isPowerOf2_64(C->getZExtValue()))
2550 return BinaryOperator::createAnd(Op0, SubOne(C));
2553 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2554 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2555 if (RHSI->getOpcode() == Instruction::Shl &&
2556 isa<ConstantInt>(RHSI->getOperand(0))) {
2557 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2558 if (isPowerOf2_64(C1)) {
2559 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2560 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2562 return BinaryOperator::createAnd(Op0, Add);
2567 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2568 // where C1&C2 are powers of two.
2569 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2570 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2571 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2572 // STO == 0 and SFO == 0 handled above.
2573 if (isPowerOf2_64(STO->getZExtValue()) &&
2574 isPowerOf2_64(SFO->getZExtValue())) {
2575 Value *TrueAnd = InsertNewInstBefore(
2576 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2577 Value *FalseAnd = InsertNewInstBefore(
2578 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2579 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2587 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2588 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2590 if (Instruction *common = commonIRemTransforms(I))
2593 if (Value *RHSNeg = dyn_castNegVal(Op1))
2594 if (!isa<ConstantInt>(RHSNeg) ||
2595 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2597 AddUsesToWorkList(I);
2598 I.setOperand(1, RHSNeg);
2602 // If the top bits of both operands are zero (i.e. we can prove they are
2603 // unsigned inputs), turn this into a urem.
2604 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2605 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2606 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2607 return BinaryOperator::createURem(Op0, Op1, I.getName());
2613 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2614 return commonRemTransforms(I);
2617 // isMaxValueMinusOne - return true if this is Max-1
2618 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2620 // Calculate 0111111111..11111
2621 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2622 int64_t Val = INT64_MAX; // All ones
2623 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2624 return C->getSExtValue() == Val-1;
2626 return C->getZExtValue() == C->getType()->getBitMask()-1;
2629 // isMinValuePlusOne - return true if this is Min+1
2630 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2632 // Calculate 1111111111000000000000
2633 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2634 int64_t Val = -1; // All ones
2635 Val <<= TypeBits-1; // Shift over to the right spot
2636 return C->getSExtValue() == Val+1;
2638 return C->getZExtValue() == 1; // unsigned
2641 // isOneBitSet - Return true if there is exactly one bit set in the specified
2643 static bool isOneBitSet(const ConstantInt *CI) {
2644 uint64_t V = CI->getZExtValue();
2645 return V && (V & (V-1)) == 0;
2648 #if 0 // Currently unused
2649 // isLowOnes - Return true if the constant is of the form 0+1+.
2650 static bool isLowOnes(const ConstantInt *CI) {
2651 uint64_t V = CI->getZExtValue();
2653 // There won't be bits set in parts that the type doesn't contain.
2654 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2656 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2657 return U && V && (U & V) == 0;
2661 // isHighOnes - Return true if the constant is of the form 1+0+.
2662 // This is the same as lowones(~X).
2663 static bool isHighOnes(const ConstantInt *CI) {
2664 uint64_t V = ~CI->getZExtValue();
2665 if (~V == 0) return false; // 0's does not match "1+"
2667 // There won't be bits set in parts that the type doesn't contain.
2668 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2670 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2671 return U && V && (U & V) == 0;
2674 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2675 /// are carefully arranged to allow folding of expressions such as:
2677 /// (A < B) | (A > B) --> (A != B)
2679 /// Note that this is only valid if the first and second predicates have the
2680 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2682 /// Three bits are used to represent the condition, as follows:
2687 /// <=> Value Definition
2688 /// 000 0 Always false
2695 /// 111 7 Always true
2697 static unsigned getICmpCode(const ICmpInst *ICI) {
2698 switch (ICI->getPredicate()) {
2700 case ICmpInst::ICMP_UGT: return 1; // 001
2701 case ICmpInst::ICMP_SGT: return 1; // 001
2702 case ICmpInst::ICMP_EQ: return 2; // 010
2703 case ICmpInst::ICMP_UGE: return 3; // 011
2704 case ICmpInst::ICMP_SGE: return 3; // 011
2705 case ICmpInst::ICMP_ULT: return 4; // 100
2706 case ICmpInst::ICMP_SLT: return 4; // 100
2707 case ICmpInst::ICMP_NE: return 5; // 101
2708 case ICmpInst::ICMP_ULE: return 6; // 110
2709 case ICmpInst::ICMP_SLE: return 6; // 110
2712 assert(0 && "Invalid ICmp predicate!");
2717 /// getICmpValue - This is the complement of getICmpCode, which turns an
2718 /// opcode and two operands into either a constant true or false, or a brand
2719 /// new /// ICmp instruction. The sign is passed in to determine which kind
2720 /// of predicate to use in new icmp instructions.
2721 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2723 default: assert(0 && "Illegal ICmp code!");
2724 case 0: return ConstantInt::getFalse();
2727 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2729 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2730 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2733 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2735 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2738 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2740 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2741 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2744 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2746 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2747 case 7: return ConstantInt::getTrue();
2751 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2752 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2753 (ICmpInst::isSignedPredicate(p1) &&
2754 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2755 (ICmpInst::isSignedPredicate(p2) &&
2756 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2760 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2761 struct FoldICmpLogical {
2764 ICmpInst::Predicate pred;
2765 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2766 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2767 pred(ICI->getPredicate()) {}
2768 bool shouldApply(Value *V) const {
2769 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2770 if (PredicatesFoldable(pred, ICI->getPredicate()))
2771 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2772 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2775 Instruction *apply(Instruction &Log) const {
2776 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2777 if (ICI->getOperand(0) != LHS) {
2778 assert(ICI->getOperand(1) == LHS);
2779 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2782 unsigned LHSCode = getICmpCode(ICI);
2783 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2785 switch (Log.getOpcode()) {
2786 case Instruction::And: Code = LHSCode & RHSCode; break;
2787 case Instruction::Or: Code = LHSCode | RHSCode; break;
2788 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2789 default: assert(0 && "Illegal logical opcode!"); return 0;
2792 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2793 if (Instruction *I = dyn_cast<Instruction>(RV))
2795 // Otherwise, it's a constant boolean value...
2796 return IC.ReplaceInstUsesWith(Log, RV);
2799 } // end anonymous namespace
2801 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2802 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2803 // guaranteed to be either a shift instruction or a binary operator.
2804 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2806 ConstantInt *AndRHS,
2807 BinaryOperator &TheAnd) {
2808 Value *X = Op->getOperand(0);
2809 Constant *Together = 0;
2810 if (!isa<ShiftInst>(Op))
2811 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2813 switch (Op->getOpcode()) {
2814 case Instruction::Xor:
2815 if (Op->hasOneUse()) {
2816 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2817 std::string OpName = Op->getName(); Op->setName("");
2818 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2819 InsertNewInstBefore(And, TheAnd);
2820 return BinaryOperator::createXor(And, Together);
2823 case Instruction::Or:
2824 if (Together == AndRHS) // (X | C) & C --> C
2825 return ReplaceInstUsesWith(TheAnd, AndRHS);
2827 if (Op->hasOneUse() && Together != OpRHS) {
2828 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2829 std::string Op0Name = Op->getName(); Op->setName("");
2830 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2831 InsertNewInstBefore(Or, TheAnd);
2832 return BinaryOperator::createAnd(Or, AndRHS);
2835 case Instruction::Add:
2836 if (Op->hasOneUse()) {
2837 // Adding a one to a single bit bit-field should be turned into an XOR
2838 // of the bit. First thing to check is to see if this AND is with a
2839 // single bit constant.
2840 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2842 // Clear bits that are not part of the constant.
2843 AndRHSV &= AndRHS->getType()->getBitMask();
2845 // If there is only one bit set...
2846 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2847 // Ok, at this point, we know that we are masking the result of the
2848 // ADD down to exactly one bit. If the constant we are adding has
2849 // no bits set below this bit, then we can eliminate the ADD.
2850 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2852 // Check to see if any bits below the one bit set in AndRHSV are set.
2853 if ((AddRHS & (AndRHSV-1)) == 0) {
2854 // If not, the only thing that can effect the output of the AND is
2855 // the bit specified by AndRHSV. If that bit is set, the effect of
2856 // the XOR is to toggle the bit. If it is clear, then the ADD has
2858 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2859 TheAnd.setOperand(0, X);
2862 std::string Name = Op->getName(); Op->setName("");
2863 // Pull the XOR out of the AND.
2864 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2865 InsertNewInstBefore(NewAnd, TheAnd);
2866 return BinaryOperator::createXor(NewAnd, AndRHS);
2873 case Instruction::Shl: {
2874 // We know that the AND will not produce any of the bits shifted in, so if
2875 // the anded constant includes them, clear them now!
2877 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2878 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2879 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2881 if (CI == ShlMask) { // Masking out bits that the shift already masks
2882 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2883 } else if (CI != AndRHS) { // Reducing bits set in and.
2884 TheAnd.setOperand(1, CI);
2889 case Instruction::LShr:
2891 // We know that the AND will not produce any of the bits shifted in, so if
2892 // the anded constant includes them, clear them now! This only applies to
2893 // unsigned shifts, because a signed shr may bring in set bits!
2895 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2896 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2897 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2899 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2900 return ReplaceInstUsesWith(TheAnd, Op);
2901 } else if (CI != AndRHS) {
2902 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2907 case Instruction::AShr:
2909 // See if this is shifting in some sign extension, then masking it out
2911 if (Op->hasOneUse()) {
2912 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2913 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2914 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2915 if (C == AndRHS) { // Masking out bits shifted in.
2916 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2917 // Make the argument unsigned.
2918 Value *ShVal = Op->getOperand(0);
2919 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2920 OpRHS, Op->getName()), TheAnd);
2921 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2930 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2931 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2932 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2933 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2934 /// insert new instructions.
2935 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2936 bool isSigned, bool Inside,
2938 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2939 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2940 "Lo is not <= Hi in range emission code!");
2943 if (Lo == Hi) // Trivially false.
2944 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2946 // V >= Min && V < Hi --> V < Hi
2947 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2948 ICmpInst::Predicate pred = (isSigned ?
2949 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2950 return new ICmpInst(pred, V, Hi);
2953 // Emit V-Lo <u Hi-Lo
2954 Constant *NegLo = ConstantExpr::getNeg(Lo);
2955 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2956 InsertNewInstBefore(Add, IB);
2957 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2958 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2961 if (Lo == Hi) // Trivially true.
2962 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2964 // V < Min || V >= Hi ->'V > Hi-1'
2965 Hi = SubOne(cast<ConstantInt>(Hi));
2966 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2967 ICmpInst::Predicate pred = (isSigned ?
2968 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2969 return new ICmpInst(pred, V, Hi);
2972 // Emit V-Lo > Hi-1-Lo
2973 Constant *NegLo = ConstantExpr::getNeg(Lo);
2974 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2975 InsertNewInstBefore(Add, IB);
2976 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
2977 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
2980 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2981 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2982 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2983 // not, since all 1s are not contiguous.
2984 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
2985 uint64_t V = Val->getZExtValue();
2986 if (!isShiftedMask_64(V)) return false;
2988 // look for the first zero bit after the run of ones
2989 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2990 // look for the first non-zero bit
2991 ME = 64-CountLeadingZeros_64(V);
2997 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2998 /// where isSub determines whether the operator is a sub. If we can fold one of
2999 /// the following xforms:
3001 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3002 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3003 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3005 /// return (A +/- B).
3007 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3008 ConstantInt *Mask, bool isSub,
3010 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3011 if (!LHSI || LHSI->getNumOperands() != 2 ||
3012 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3014 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3016 switch (LHSI->getOpcode()) {
3018 case Instruction::And:
3019 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3020 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3021 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3024 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3025 // part, we don't need any explicit masks to take them out of A. If that
3026 // is all N is, ignore it.
3028 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3029 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3031 if (MaskedValueIsZero(RHS, Mask))
3036 case Instruction::Or:
3037 case Instruction::Xor:
3038 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3039 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3040 ConstantExpr::getAnd(N, Mask)->isNullValue())
3047 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3049 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3050 return InsertNewInstBefore(New, I);
3053 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3054 bool Changed = SimplifyCommutative(I);
3055 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3057 if (isa<UndefValue>(Op1)) // X & undef -> 0
3058 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3062 return ReplaceInstUsesWith(I, Op1);
3064 // See if we can simplify any instructions used by the instruction whose sole
3065 // purpose is to compute bits we don't care about.
3066 uint64_t KnownZero, KnownOne;
3067 if (!isa<PackedType>(I.getType())) {
3068 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3069 KnownZero, KnownOne))
3072 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Op1)) {
3073 if (CP->isAllOnesValue())
3074 return ReplaceInstUsesWith(I, I.getOperand(0));
3078 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3079 uint64_t AndRHSMask = AndRHS->getZExtValue();
3080 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3081 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3083 // Optimize a variety of ((val OP C1) & C2) combinations...
3084 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3085 Instruction *Op0I = cast<Instruction>(Op0);
3086 Value *Op0LHS = Op0I->getOperand(0);
3087 Value *Op0RHS = Op0I->getOperand(1);
3088 switch (Op0I->getOpcode()) {
3089 case Instruction::Xor:
3090 case Instruction::Or:
3091 // If the mask is only needed on one incoming arm, push it up.
3092 if (Op0I->hasOneUse()) {
3093 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3094 // Not masking anything out for the LHS, move to RHS.
3095 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3096 Op0RHS->getName()+".masked");
3097 InsertNewInstBefore(NewRHS, I);
3098 return BinaryOperator::create(
3099 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3101 if (!isa<Constant>(Op0RHS) &&
3102 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3103 // Not masking anything out for the RHS, move to LHS.
3104 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3105 Op0LHS->getName()+".masked");
3106 InsertNewInstBefore(NewLHS, I);
3107 return BinaryOperator::create(
3108 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3113 case Instruction::Add:
3114 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3115 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3116 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3117 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3118 return BinaryOperator::createAnd(V, AndRHS);
3119 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3120 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3123 case Instruction::Sub:
3124 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3125 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3126 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3127 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3128 return BinaryOperator::createAnd(V, AndRHS);
3132 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3133 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3135 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3136 // If this is an integer truncation or change from signed-to-unsigned, and
3137 // if the source is an and/or with immediate, transform it. This
3138 // frequently occurs for bitfield accesses.
3139 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3140 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3141 CastOp->getNumOperands() == 2)
3142 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3143 if (CastOp->getOpcode() == Instruction::And) {
3144 // Change: and (cast (and X, C1) to T), C2
3145 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3146 // This will fold the two constants together, which may allow
3147 // other simplifications.
3148 Instruction *NewCast = CastInst::createTruncOrBitCast(
3149 CastOp->getOperand(0), I.getType(),
3150 CastOp->getName()+".shrunk");
3151 NewCast = InsertNewInstBefore(NewCast, I);
3152 // trunc_or_bitcast(C1)&C2
3153 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3154 C3 = ConstantExpr::getAnd(C3, AndRHS);
3155 return BinaryOperator::createAnd(NewCast, C3);
3156 } else if (CastOp->getOpcode() == Instruction::Or) {
3157 // Change: and (cast (or X, C1) to T), C2
3158 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3159 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3160 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3161 return ReplaceInstUsesWith(I, AndRHS);
3166 // Try to fold constant and into select arguments.
3167 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3168 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3170 if (isa<PHINode>(Op0))
3171 if (Instruction *NV = FoldOpIntoPhi(I))
3175 Value *Op0NotVal = dyn_castNotVal(Op0);
3176 Value *Op1NotVal = dyn_castNotVal(Op1);
3178 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3179 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3181 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3182 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3183 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3184 I.getName()+".demorgan");
3185 InsertNewInstBefore(Or, I);
3186 return BinaryOperator::createNot(Or);
3190 Value *A = 0, *B = 0;
3191 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3192 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3193 return ReplaceInstUsesWith(I, Op1);
3194 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3195 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3196 return ReplaceInstUsesWith(I, Op0);
3198 if (Op0->hasOneUse() &&
3199 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3200 if (A == Op1) { // (A^B)&A -> A&(A^B)
3201 I.swapOperands(); // Simplify below
3202 std::swap(Op0, Op1);
3203 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3204 cast<BinaryOperator>(Op0)->swapOperands();
3205 I.swapOperands(); // Simplify below
3206 std::swap(Op0, Op1);
3209 if (Op1->hasOneUse() &&
3210 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3211 if (B == Op0) { // B&(A^B) -> B&(B^A)
3212 cast<BinaryOperator>(Op1)->swapOperands();
3215 if (A == Op0) { // A&(A^B) -> A & ~B
3216 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3217 InsertNewInstBefore(NotB, I);
3218 return BinaryOperator::createAnd(A, NotB);
3223 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3224 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3225 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3228 Value *LHSVal, *RHSVal;
3229 ConstantInt *LHSCst, *RHSCst;
3230 ICmpInst::Predicate LHSCC, RHSCC;
3231 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3232 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3233 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3234 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3235 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3236 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3237 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3238 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3239 // Ensure that the larger constant is on the RHS.
3240 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3241 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3242 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3243 ICmpInst *LHS = cast<ICmpInst>(Op0);
3244 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3245 std::swap(LHS, RHS);
3246 std::swap(LHSCst, RHSCst);
3247 std::swap(LHSCC, RHSCC);
3250 // At this point, we know we have have two icmp instructions
3251 // comparing a value against two constants and and'ing the result
3252 // together. Because of the above check, we know that we only have
3253 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3254 // (from the FoldICmpLogical check above), that the two constants
3255 // are not equal and that the larger constant is on the RHS
3256 assert(LHSCst != RHSCst && "Compares not folded above?");
3259 default: assert(0 && "Unknown integer condition code!");
3260 case ICmpInst::ICMP_EQ:
3262 default: assert(0 && "Unknown integer condition code!");
3263 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3264 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3265 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3266 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3267 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3268 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3269 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3270 return ReplaceInstUsesWith(I, LHS);
3272 case ICmpInst::ICMP_NE:
3274 default: assert(0 && "Unknown integer condition code!");
3275 case ICmpInst::ICMP_ULT:
3276 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3277 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3278 break; // (X != 13 & X u< 15) -> no change
3279 case ICmpInst::ICMP_SLT:
3280 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3281 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3282 break; // (X != 13 & X s< 15) -> no change
3283 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3284 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3285 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3286 return ReplaceInstUsesWith(I, RHS);
3287 case ICmpInst::ICMP_NE:
3288 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3289 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3290 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3291 LHSVal->getName()+".off");
3292 InsertNewInstBefore(Add, I);
3293 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3294 ConstantInt::get(Add->getType(), 1));
3296 break; // (X != 13 & X != 15) -> no change
3299 case ICmpInst::ICMP_ULT:
3301 default: assert(0 && "Unknown integer condition code!");
3302 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3303 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3304 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3305 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3307 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3308 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3309 return ReplaceInstUsesWith(I, LHS);
3310 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3314 case ICmpInst::ICMP_SLT:
3316 default: assert(0 && "Unknown integer condition code!");
3317 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3318 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3319 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3320 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3322 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3323 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3324 return ReplaceInstUsesWith(I, LHS);
3325 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3329 case ICmpInst::ICMP_UGT:
3331 default: assert(0 && "Unknown integer condition code!");
3332 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3333 return ReplaceInstUsesWith(I, LHS);
3334 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3335 return ReplaceInstUsesWith(I, RHS);
3336 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3338 case ICmpInst::ICMP_NE:
3339 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3340 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3341 break; // (X u> 13 & X != 15) -> no change
3342 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3343 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3345 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3349 case ICmpInst::ICMP_SGT:
3351 default: assert(0 && "Unknown integer condition code!");
3352 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3353 return ReplaceInstUsesWith(I, LHS);
3354 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3355 return ReplaceInstUsesWith(I, RHS);
3356 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3358 case ICmpInst::ICMP_NE:
3359 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3360 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3361 break; // (X s> 13 & X != 15) -> no change
3362 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3363 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3365 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3373 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3374 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3375 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3376 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3377 const Type *SrcTy = Op0C->getOperand(0)->getType();
3378 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3379 // Only do this if the casts both really cause code to be generated.
3380 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3382 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3384 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3385 Op1C->getOperand(0),
3387 InsertNewInstBefore(NewOp, I);
3388 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3392 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3393 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3394 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3395 if (SI0->getOpcode() == SI1->getOpcode() &&
3396 SI0->getOperand(1) == SI1->getOperand(1) &&
3397 (SI0->hasOneUse() || SI1->hasOneUse())) {
3398 Instruction *NewOp =
3399 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3401 SI0->getName()), I);
3402 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3406 return Changed ? &I : 0;
3409 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3410 /// in the result. If it does, and if the specified byte hasn't been filled in
3411 /// yet, fill it in and return false.
3412 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3413 Instruction *I = dyn_cast<Instruction>(V);
3414 if (I == 0) return true;
3416 // If this is an or instruction, it is an inner node of the bswap.
3417 if (I->getOpcode() == Instruction::Or)
3418 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3419 CollectBSwapParts(I->getOperand(1), ByteValues);
3421 // If this is a shift by a constant int, and it is "24", then its operand
3422 // defines a byte. We only handle unsigned types here.
3423 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3424 // Not shifting the entire input by N-1 bytes?
3425 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3426 8*(ByteValues.size()-1))
3430 if (I->getOpcode() == Instruction::Shl) {
3431 // X << 24 defines the top byte with the lowest of the input bytes.
3432 DestNo = ByteValues.size()-1;
3434 // X >>u 24 defines the low byte with the highest of the input bytes.
3438 // If the destination byte value is already defined, the values are or'd
3439 // together, which isn't a bswap (unless it's an or of the same bits).
3440 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3442 ByteValues[DestNo] = I->getOperand(0);
3446 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3448 Value *Shift = 0, *ShiftLHS = 0;
3449 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3450 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3451 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3453 Instruction *SI = cast<Instruction>(Shift);
3455 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3456 if (ShiftAmt->getZExtValue() & 7 ||
3457 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3460 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3462 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3463 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3465 // Unknown mask for bswap.
3466 if (DestByte == ByteValues.size()) return true;
3468 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3470 if (SI->getOpcode() == Instruction::Shl)
3471 SrcByte = DestByte - ShiftBytes;
3473 SrcByte = DestByte + ShiftBytes;
3475 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3476 if (SrcByte != ByteValues.size()-DestByte-1)
3479 // If the destination byte value is already defined, the values are or'd
3480 // together, which isn't a bswap (unless it's an or of the same bits).
3481 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3483 ByteValues[DestByte] = SI->getOperand(0);
3487 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3488 /// If so, insert the new bswap intrinsic and return it.
3489 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3490 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3491 if (I.getType() == Type::Int8Ty)
3494 /// ByteValues - For each byte of the result, we keep track of which value
3495 /// defines each byte.
3496 std::vector<Value*> ByteValues;
3497 ByteValues.resize(TD->getTypeSize(I.getType()));
3499 // Try to find all the pieces corresponding to the bswap.
3500 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3501 CollectBSwapParts(I.getOperand(1), ByteValues))
3504 // Check to see if all of the bytes come from the same value.
3505 Value *V = ByteValues[0];
3506 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3508 // Check to make sure that all of the bytes come from the same value.
3509 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3510 if (ByteValues[i] != V)
3513 // If they do then *success* we can turn this into a bswap. Figure out what
3514 // bswap to make it into.
3515 Module *M = I.getParent()->getParent()->getParent();
3516 const char *FnName = 0;
3517 if (I.getType() == Type::Int16Ty)
3518 FnName = "llvm.bswap.i16";
3519 else if (I.getType() == Type::Int32Ty)
3520 FnName = "llvm.bswap.i32";
3521 else if (I.getType() == Type::Int64Ty)
3522 FnName = "llvm.bswap.i64";
3524 assert(0 && "Unknown integer type!");
3525 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3526 return new CallInst(F, V);
3530 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3531 bool Changed = SimplifyCommutative(I);
3532 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3534 if (isa<UndefValue>(Op1))
3535 return ReplaceInstUsesWith(I, // X | undef -> -1
3536 ConstantInt::getAllOnesValue(I.getType()));
3540 return ReplaceInstUsesWith(I, Op0);
3542 // See if we can simplify any instructions used by the instruction whose sole
3543 // purpose is to compute bits we don't care about.
3544 uint64_t KnownZero, KnownOne;
3545 if (!isa<PackedType>(I.getType()) &&
3546 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3547 KnownZero, KnownOne))
3551 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3552 ConstantInt *C1 = 0; Value *X = 0;
3553 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3554 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3555 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3557 InsertNewInstBefore(Or, I);
3558 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3561 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3562 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3563 std::string Op0Name = Op0->getName(); Op0->setName("");
3564 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3565 InsertNewInstBefore(Or, I);
3566 return BinaryOperator::createXor(Or,
3567 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3570 // Try to fold constant and into select arguments.
3571 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3572 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3574 if (isa<PHINode>(Op0))
3575 if (Instruction *NV = FoldOpIntoPhi(I))
3579 Value *A = 0, *B = 0;
3580 ConstantInt *C1 = 0, *C2 = 0;
3582 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3583 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3584 return ReplaceInstUsesWith(I, Op1);
3585 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3586 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3587 return ReplaceInstUsesWith(I, Op0);
3589 // (A | B) | C and A | (B | C) -> bswap if possible.
3590 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3591 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3592 match(Op1, m_Or(m_Value(), m_Value())) ||
3593 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3594 match(Op1, m_Shift(m_Value(), m_Value())))) {
3595 if (Instruction *BSwap = MatchBSwap(I))
3599 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3600 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3601 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3602 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3604 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3607 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3608 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3609 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3610 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3612 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3615 // (A & C1)|(B & C2)
3616 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3617 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3619 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3620 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3623 // If we have: ((V + N) & C1) | (V & C2)
3624 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3625 // replace with V+N.
3626 if (C1 == ConstantExpr::getNot(C2)) {
3627 Value *V1 = 0, *V2 = 0;
3628 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3629 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3630 // Add commutes, try both ways.
3631 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3632 return ReplaceInstUsesWith(I, A);
3633 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3634 return ReplaceInstUsesWith(I, A);
3636 // Or commutes, try both ways.
3637 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3638 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3639 // Add commutes, try both ways.
3640 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3641 return ReplaceInstUsesWith(I, B);
3642 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3643 return ReplaceInstUsesWith(I, B);
3648 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3649 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3650 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3651 if (SI0->getOpcode() == SI1->getOpcode() &&
3652 SI0->getOperand(1) == SI1->getOperand(1) &&
3653 (SI0->hasOneUse() || SI1->hasOneUse())) {
3654 Instruction *NewOp =
3655 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3657 SI0->getName()), I);
3658 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3662 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3663 if (A == Op1) // ~A | A == -1
3664 return ReplaceInstUsesWith(I,
3665 ConstantInt::getAllOnesValue(I.getType()));
3669 // Note, A is still live here!
3670 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3672 return ReplaceInstUsesWith(I,
3673 ConstantInt::getAllOnesValue(I.getType()));
3675 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3676 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3677 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3678 I.getName()+".demorgan"), I);
3679 return BinaryOperator::createNot(And);
3683 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3684 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3685 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3688 Value *LHSVal, *RHSVal;
3689 ConstantInt *LHSCst, *RHSCst;
3690 ICmpInst::Predicate LHSCC, RHSCC;
3691 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3692 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3693 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3694 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3695 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3696 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3697 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3698 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3699 // Ensure that the larger constant is on the RHS.
3700 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3701 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3702 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3703 ICmpInst *LHS = cast<ICmpInst>(Op0);
3704 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3705 std::swap(LHS, RHS);
3706 std::swap(LHSCst, RHSCst);
3707 std::swap(LHSCC, RHSCC);
3710 // At this point, we know we have have two icmp instructions
3711 // comparing a value against two constants and or'ing the result
3712 // together. Because of the above check, we know that we only have
3713 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3714 // FoldICmpLogical check above), that the two constants are not
3716 assert(LHSCst != RHSCst && "Compares not folded above?");
3719 default: assert(0 && "Unknown integer condition code!");
3720 case ICmpInst::ICMP_EQ:
3722 default: assert(0 && "Unknown integer condition code!");
3723 case ICmpInst::ICMP_EQ:
3724 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3725 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3726 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3727 LHSVal->getName()+".off");
3728 InsertNewInstBefore(Add, I);
3729 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3730 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3732 break; // (X == 13 | X == 15) -> no change
3733 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3734 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3736 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3737 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3738 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3739 return ReplaceInstUsesWith(I, RHS);
3742 case ICmpInst::ICMP_NE:
3744 default: assert(0 && "Unknown integer condition code!");
3745 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3746 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3747 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3748 return ReplaceInstUsesWith(I, LHS);
3749 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3750 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3751 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3752 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3755 case ICmpInst::ICMP_ULT:
3757 default: assert(0 && "Unknown integer condition code!");
3758 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3760 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3761 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3763 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3765 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3766 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3767 return ReplaceInstUsesWith(I, RHS);
3768 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3772 case ICmpInst::ICMP_SLT:
3774 default: assert(0 && "Unknown integer condition code!");
3775 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3777 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3778 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3780 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3782 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3783 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3784 return ReplaceInstUsesWith(I, RHS);
3785 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3789 case ICmpInst::ICMP_UGT:
3791 default: assert(0 && "Unknown integer condition code!");
3792 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3793 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3794 return ReplaceInstUsesWith(I, LHS);
3795 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3797 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3798 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3799 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3800 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3804 case ICmpInst::ICMP_SGT:
3806 default: assert(0 && "Unknown integer condition code!");
3807 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3808 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3809 return ReplaceInstUsesWith(I, LHS);
3810 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3812 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3813 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3814 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3815 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3823 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3824 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3825 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3826 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3827 const Type *SrcTy = Op0C->getOperand(0)->getType();
3828 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3829 // Only do this if the casts both really cause code to be generated.
3830 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3832 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3834 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3835 Op1C->getOperand(0),
3837 InsertNewInstBefore(NewOp, I);
3838 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3843 return Changed ? &I : 0;
3846 // XorSelf - Implements: X ^ X --> 0
3849 XorSelf(Value *rhs) : RHS(rhs) {}
3850 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3851 Instruction *apply(BinaryOperator &Xor) const {
3857 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3858 bool Changed = SimplifyCommutative(I);
3859 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3861 if (isa<UndefValue>(Op1))
3862 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3864 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3865 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3866 assert(Result == &I && "AssociativeOpt didn't work?");
3867 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3870 // See if we can simplify any instructions used by the instruction whose sole
3871 // purpose is to compute bits we don't care about.
3872 uint64_t KnownZero, KnownOne;
3873 if (!isa<PackedType>(I.getType()) &&
3874 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3875 KnownZero, KnownOne))
3878 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3879 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3880 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3881 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3882 return new ICmpInst(ICI->getInversePredicate(),
3883 ICI->getOperand(0), ICI->getOperand(1));
3885 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3886 // ~(c-X) == X-c-1 == X+(-c-1)
3887 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3888 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3889 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3890 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3891 ConstantInt::get(I.getType(), 1));
3892 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3895 // ~(~X & Y) --> (X | ~Y)
3896 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3897 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3898 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3900 BinaryOperator::createNot(Op0I->getOperand(1),
3901 Op0I->getOperand(1)->getName()+".not");
3902 InsertNewInstBefore(NotY, I);
3903 return BinaryOperator::createOr(Op0NotVal, NotY);
3907 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3908 if (Op0I->getOpcode() == Instruction::Add) {
3909 // ~(X-c) --> (-c-1)-X
3910 if (RHS->isAllOnesValue()) {
3911 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3912 return BinaryOperator::createSub(
3913 ConstantExpr::getSub(NegOp0CI,
3914 ConstantInt::get(I.getType(), 1)),
3915 Op0I->getOperand(0));
3917 } else if (Op0I->getOpcode() == Instruction::Or) {
3918 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3919 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3920 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3921 // Anything in both C1 and C2 is known to be zero, remove it from
3923 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3924 NewRHS = ConstantExpr::getAnd(NewRHS,
3925 ConstantExpr::getNot(CommonBits));
3926 WorkList.push_back(Op0I);
3927 I.setOperand(0, Op0I->getOperand(0));
3928 I.setOperand(1, NewRHS);
3934 // Try to fold constant and into select arguments.
3935 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3936 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3938 if (isa<PHINode>(Op0))
3939 if (Instruction *NV = FoldOpIntoPhi(I))
3943 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3945 return ReplaceInstUsesWith(I,
3946 ConstantInt::getAllOnesValue(I.getType()));
3948 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3950 return ReplaceInstUsesWith(I,
3951 ConstantInt::getAllOnesValue(I.getType()));
3953 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3954 if (Op1I->getOpcode() == Instruction::Or) {
3955 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3956 Op1I->swapOperands();
3958 std::swap(Op0, Op1);
3959 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3960 I.swapOperands(); // Simplified below.
3961 std::swap(Op0, Op1);
3963 } else if (Op1I->getOpcode() == Instruction::Xor) {
3964 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3965 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3966 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3967 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3968 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3969 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3970 Op1I->swapOperands();
3971 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3972 I.swapOperands(); // Simplified below.
3973 std::swap(Op0, Op1);
3977 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3978 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3979 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3980 Op0I->swapOperands();
3981 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3982 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3983 InsertNewInstBefore(NotB, I);
3984 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3986 } else if (Op0I->getOpcode() == Instruction::Xor) {
3987 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3988 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3989 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3990 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3991 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3992 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3993 Op0I->swapOperands();
3994 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3995 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3996 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3997 InsertNewInstBefore(N, I);
3998 return BinaryOperator::createAnd(N, Op1);
4002 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4003 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4004 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4007 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4008 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4009 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4010 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4011 const Type *SrcTy = Op0C->getOperand(0)->getType();
4012 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4013 // Only do this if the casts both really cause code to be generated.
4014 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4016 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4018 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4019 Op1C->getOperand(0),
4021 InsertNewInstBefore(NewOp, I);
4022 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4026 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4027 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
4028 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
4029 if (SI0->getOpcode() == SI1->getOpcode() &&
4030 SI0->getOperand(1) == SI1->getOperand(1) &&
4031 (SI0->hasOneUse() || SI1->hasOneUse())) {
4032 Instruction *NewOp =
4033 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4035 SI0->getName()), I);
4036 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
4040 return Changed ? &I : 0;
4043 static bool isPositive(ConstantInt *C) {
4044 return C->getSExtValue() >= 0;
4047 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4048 /// overflowed for this type.
4049 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4051 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4053 return cast<ConstantInt>(Result)->getZExtValue() <
4054 cast<ConstantInt>(In1)->getZExtValue();
4057 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4058 /// code necessary to compute the offset from the base pointer (without adding
4059 /// in the base pointer). Return the result as a signed integer of intptr size.
4060 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4061 TargetData &TD = IC.getTargetData();
4062 gep_type_iterator GTI = gep_type_begin(GEP);
4063 const Type *IntPtrTy = TD.getIntPtrType();
4064 Value *Result = Constant::getNullValue(IntPtrTy);
4066 // Build a mask for high order bits.
4067 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4069 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4070 Value *Op = GEP->getOperand(i);
4071 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4072 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4073 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4074 if (!OpC->isNullValue()) {
4075 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4076 Scale = ConstantExpr::getMul(OpC, Scale);
4077 if (Constant *RC = dyn_cast<Constant>(Result))
4078 Result = ConstantExpr::getAdd(RC, Scale);
4080 // Emit an add instruction.
4081 Result = IC.InsertNewInstBefore(
4082 BinaryOperator::createAdd(Result, Scale,
4083 GEP->getName()+".offs"), I);
4087 // Convert to correct type.
4088 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4089 Op->getName()+".c"), I);
4091 // We'll let instcombine(mul) convert this to a shl if possible.
4092 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4093 GEP->getName()+".idx"), I);
4095 // Emit an add instruction.
4096 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4097 GEP->getName()+".offs"), I);
4103 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4104 /// else. At this point we know that the GEP is on the LHS of the comparison.
4105 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4106 ICmpInst::Predicate Cond,
4108 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4110 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4111 if (isa<PointerType>(CI->getOperand(0)->getType()))
4112 RHS = CI->getOperand(0);
4114 Value *PtrBase = GEPLHS->getOperand(0);
4115 if (PtrBase == RHS) {
4116 // As an optimization, we don't actually have to compute the actual value of
4117 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4118 // each index is zero or not.
4119 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4120 Instruction *InVal = 0;
4121 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4122 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4124 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4125 if (isa<UndefValue>(C)) // undef index -> undef.
4126 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4127 if (C->isNullValue())
4129 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4130 EmitIt = false; // This is indexing into a zero sized array?
4131 } else if (isa<ConstantInt>(C))
4132 return ReplaceInstUsesWith(I, // No comparison is needed here.
4133 ConstantInt::get(Type::Int1Ty,
4134 Cond == ICmpInst::ICMP_NE));
4139 new ICmpInst(Cond, GEPLHS->getOperand(i),
4140 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4144 InVal = InsertNewInstBefore(InVal, I);
4145 InsertNewInstBefore(Comp, I);
4146 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4147 InVal = BinaryOperator::createOr(InVal, Comp);
4148 else // True if all are equal
4149 InVal = BinaryOperator::createAnd(InVal, Comp);
4157 // No comparison is needed here, all indexes = 0
4158 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4159 Cond == ICmpInst::ICMP_EQ));
4162 // Only lower this if the icmp is the only user of the GEP or if we expect
4163 // the result to fold to a constant!
4164 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4165 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4166 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4167 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4168 Constant::getNullValue(Offset->getType()));
4170 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4171 // If the base pointers are different, but the indices are the same, just
4172 // compare the base pointer.
4173 if (PtrBase != GEPRHS->getOperand(0)) {
4174 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4175 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4176 GEPRHS->getOperand(0)->getType();
4178 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4179 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4180 IndicesTheSame = false;
4184 // If all indices are the same, just compare the base pointers.
4186 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4187 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4189 // Otherwise, the base pointers are different and the indices are
4190 // different, bail out.
4194 // If one of the GEPs has all zero indices, recurse.
4195 bool AllZeros = true;
4196 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4197 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4198 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4203 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4204 ICmpInst::getSwappedPredicate(Cond), I);
4206 // If the other GEP has all zero indices, recurse.
4208 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4209 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4210 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4215 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4217 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4218 // If the GEPs only differ by one index, compare it.
4219 unsigned NumDifferences = 0; // Keep track of # differences.
4220 unsigned DiffOperand = 0; // The operand that differs.
4221 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4222 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4223 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4224 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4225 // Irreconcilable differences.
4229 if (NumDifferences++) break;
4234 if (NumDifferences == 0) // SAME GEP?
4235 return ReplaceInstUsesWith(I, // No comparison is needed here.
4236 ConstantInt::get(Type::Int1Ty,
4237 Cond == ICmpInst::ICMP_EQ));
4238 else if (NumDifferences == 1) {
4239 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4240 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4241 // Make sure we do a signed comparison here.
4242 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4246 // Only lower this if the icmp is the only user of the GEP or if we expect
4247 // the result to fold to a constant!
4248 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4249 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4250 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4251 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4252 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4253 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4259 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4260 bool Changed = SimplifyCompare(I);
4261 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4263 // Fold trivial predicates.
4264 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4265 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4266 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4267 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4269 // Simplify 'fcmp pred X, X'
4271 switch (I.getPredicate()) {
4272 default: assert(0 && "Unknown predicate!");
4273 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4274 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4275 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4276 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4277 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4278 case FCmpInst::FCMP_OLT: // True if ordered and less than
4279 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4280 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4282 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4283 case FCmpInst::FCMP_ULT: // True if unordered or less than
4284 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4285 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4286 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4287 I.setPredicate(FCmpInst::FCMP_UNO);
4288 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4291 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4292 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4293 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4294 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4295 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4296 I.setPredicate(FCmpInst::FCMP_ORD);
4297 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4302 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4303 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4305 // Handle fcmp with constant RHS
4306 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4307 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4308 switch (LHSI->getOpcode()) {
4309 case Instruction::PHI:
4310 if (Instruction *NV = FoldOpIntoPhi(I))
4313 case Instruction::Select:
4314 // If either operand of the select is a constant, we can fold the
4315 // comparison into the select arms, which will cause one to be
4316 // constant folded and the select turned into a bitwise or.
4317 Value *Op1 = 0, *Op2 = 0;
4318 if (LHSI->hasOneUse()) {
4319 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4320 // Fold the known value into the constant operand.
4321 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4322 // Insert a new FCmp of the other select operand.
4323 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4324 LHSI->getOperand(2), RHSC,
4326 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4327 // Fold the known value into the constant operand.
4328 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4329 // Insert a new FCmp of the other select operand.
4330 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4331 LHSI->getOperand(1), RHSC,
4337 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4342 return Changed ? &I : 0;
4345 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4346 bool Changed = SimplifyCompare(I);
4347 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4348 const Type *Ty = Op0->getType();
4352 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4353 isTrueWhenEqual(I)));
4355 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4356 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4358 // icmp of GlobalValues can never equal each other as long as they aren't
4359 // external weak linkage type.
4360 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4361 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4362 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4363 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4364 !isTrueWhenEqual(I)));
4366 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4367 // addresses never equal each other! We already know that Op0 != Op1.
4368 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4369 isa<ConstantPointerNull>(Op0)) &&
4370 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4371 isa<ConstantPointerNull>(Op1)))
4372 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4373 !isTrueWhenEqual(I)));
4375 // icmp's with boolean values can always be turned into bitwise operations
4376 if (Ty == Type::Int1Ty) {
4377 switch (I.getPredicate()) {
4378 default: assert(0 && "Invalid icmp instruction!");
4379 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4380 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4381 InsertNewInstBefore(Xor, I);
4382 return BinaryOperator::createNot(Xor);
4384 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4385 return BinaryOperator::createXor(Op0, Op1);
4387 case ICmpInst::ICMP_UGT:
4388 case ICmpInst::ICMP_SGT:
4389 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4391 case ICmpInst::ICMP_ULT:
4392 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4393 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4394 InsertNewInstBefore(Not, I);
4395 return BinaryOperator::createAnd(Not, Op1);
4397 case ICmpInst::ICMP_UGE:
4398 case ICmpInst::ICMP_SGE:
4399 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4401 case ICmpInst::ICMP_ULE:
4402 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4403 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4404 InsertNewInstBefore(Not, I);
4405 return BinaryOperator::createOr(Not, Op1);
4410 // See if we are doing a comparison between a constant and an instruction that
4411 // can be folded into the comparison.
4412 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4413 switch (I.getPredicate()) {
4415 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4416 if (CI->isMinValue(false))
4417 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4418 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4419 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4420 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4421 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4424 case ICmpInst::ICMP_SLT:
4425 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4426 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4427 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4428 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4429 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4430 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4433 case ICmpInst::ICMP_UGT:
4434 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4435 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4436 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4437 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4438 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4439 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4442 case ICmpInst::ICMP_SGT:
4443 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4444 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4445 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4446 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4447 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4448 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4451 case ICmpInst::ICMP_ULE:
4452 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4453 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4454 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4455 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4456 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4457 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4460 case ICmpInst::ICMP_SLE:
4461 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4462 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4463 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4464 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4465 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4466 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4469 case ICmpInst::ICMP_UGE:
4470 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4471 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4472 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4473 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4474 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4475 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4478 case ICmpInst::ICMP_SGE:
4479 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4480 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4481 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4482 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4483 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4484 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4488 // If we still have a icmp le or icmp ge instruction, turn it into the
4489 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4490 // already been handled above, this requires little checking.
4492 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4493 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4494 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4495 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4496 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4497 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4498 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4499 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4501 // See if we can fold the comparison based on bits known to be zero or one
4503 uint64_t KnownZero, KnownOne;
4504 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
4505 KnownZero, KnownOne, 0))
4508 // Given the known and unknown bits, compute a range that the LHS could be
4510 if (KnownOne | KnownZero) {
4511 // Compute the Min, Max and RHS values based on the known bits. For the
4512 // EQ and NE we use unsigned values.
4513 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4514 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4515 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4516 SRHSVal = CI->getSExtValue();
4517 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4520 URHSVal = CI->getZExtValue();
4521 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4524 switch (I.getPredicate()) { // LE/GE have been folded already.
4525 default: assert(0 && "Unknown icmp opcode!");
4526 case ICmpInst::ICMP_EQ:
4527 if (UMax < URHSVal || UMin > URHSVal)
4528 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4530 case ICmpInst::ICMP_NE:
4531 if (UMax < URHSVal || UMin > URHSVal)
4532 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4534 case ICmpInst::ICMP_ULT:
4536 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4538 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4540 case ICmpInst::ICMP_UGT:
4542 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4544 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4546 case ICmpInst::ICMP_SLT:
4548 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4550 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4552 case ICmpInst::ICMP_SGT:
4554 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4556 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4561 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4562 // instruction, see if that instruction also has constants so that the
4563 // instruction can be folded into the icmp
4564 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4565 switch (LHSI->getOpcode()) {
4566 case Instruction::And:
4567 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4568 LHSI->getOperand(0)->hasOneUse()) {
4569 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4571 // If the LHS is an AND of a truncating cast, we can widen the
4572 // and/compare to be the input width without changing the value
4573 // produced, eliminating a cast.
4574 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4575 // We can do this transformation if either the AND constant does not
4576 // have its sign bit set or if it is an equality comparison.
4577 // Extending a relational comparison when we're checking the sign
4578 // bit would not work.
4579 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4581 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4582 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4583 ConstantInt *NewCST;
4585 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4586 AndCST->getZExtValue());
4587 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4588 CI->getZExtValue());
4589 Instruction *NewAnd =
4590 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4592 InsertNewInstBefore(NewAnd, I);
4593 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4597 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4598 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4599 // happens a LOT in code produced by the C front-end, for bitfield
4601 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4603 // Check to see if there is a noop-cast between the shift and the and.
4605 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4606 if (CI->getOpcode() == Instruction::BitCast)
4607 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4611 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4612 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4613 const Type *AndTy = AndCST->getType(); // Type of the and.
4615 // We can fold this as long as we can't shift unknown bits
4616 // into the mask. This can only happen with signed shift
4617 // rights, as they sign-extend.
4619 bool CanFold = Shift->isLogicalShift();
4621 // To test for the bad case of the signed shr, see if any
4622 // of the bits shifted in could be tested after the mask.
4623 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4624 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4626 Constant *OShAmt = ConstantInt::get(Type::Int8Ty, ShAmtVal);
4628 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4630 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4636 if (Shift->getOpcode() == Instruction::Shl)
4637 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4639 NewCst = ConstantExpr::getShl(CI, ShAmt);
4641 // Check to see if we are shifting out any of the bits being
4643 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4644 // If we shifted bits out, the fold is not going to work out.
4645 // As a special case, check to see if this means that the
4646 // result is always true or false now.
4647 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4648 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4649 if (I.getPredicate() == ICmpInst::ICMP_NE)
4650 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4652 I.setOperand(1, NewCst);
4653 Constant *NewAndCST;
4654 if (Shift->getOpcode() == Instruction::Shl)
4655 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4657 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4658 LHSI->setOperand(1, NewAndCST);
4659 LHSI->setOperand(0, Shift->getOperand(0));
4660 WorkList.push_back(Shift); // Shift is dead.
4661 AddUsesToWorkList(I);
4667 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4668 // preferable because it allows the C<<Y expression to be hoisted out
4669 // of a loop if Y is invariant and X is not.
4670 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4671 I.isEquality() && !Shift->isArithmeticShift() &&
4672 isa<Instruction>(Shift->getOperand(0))) {
4675 if (Shift->getOpcode() == Instruction::LShr) {
4676 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4679 // Insert a logical shift.
4680 NS = new ShiftInst(Instruction::LShr, AndCST,
4681 Shift->getOperand(1), "tmp");
4683 InsertNewInstBefore(cast<Instruction>(NS), I);
4685 // Compute X & (C << Y).
4686 Instruction *NewAnd = BinaryOperator::createAnd(
4687 Shift->getOperand(0), NS, LHSI->getName());
4688 InsertNewInstBefore(NewAnd, I);
4690 I.setOperand(0, NewAnd);
4696 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4697 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4698 if (I.isEquality()) {
4699 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4701 // Check that the shift amount is in range. If not, don't perform
4702 // undefined shifts. When the shift is visited it will be
4704 if (ShAmt->getZExtValue() >= TypeBits)
4707 // If we are comparing against bits always shifted out, the
4708 // comparison cannot succeed.
4710 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4711 if (Comp != CI) {// Comparing against a bit that we know is zero.
4712 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4713 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4714 return ReplaceInstUsesWith(I, Cst);
4717 if (LHSI->hasOneUse()) {
4718 // Otherwise strength reduce the shift into an and.
4719 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4720 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4721 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4724 BinaryOperator::createAnd(LHSI->getOperand(0),
4725 Mask, LHSI->getName()+".mask");
4726 Value *And = InsertNewInstBefore(AndI, I);
4727 return new ICmpInst(I.getPredicate(), And,
4728 ConstantExpr::getLShr(CI, ShAmt));
4734 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4735 case Instruction::AShr:
4736 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4737 if (I.isEquality()) {
4738 // Check that the shift amount is in range. If not, don't perform
4739 // undefined shifts. When the shift is visited it will be
4741 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4742 if (ShAmt->getZExtValue() >= TypeBits)
4745 // If we are comparing against bits always shifted out, the
4746 // comparison cannot succeed.
4748 if (LHSI->getOpcode() == Instruction::LShr)
4749 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4752 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4755 if (Comp != CI) {// Comparing against a bit that we know is zero.
4756 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4757 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4758 return ReplaceInstUsesWith(I, Cst);
4761 if (LHSI->hasOneUse() || CI->isNullValue()) {
4762 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4764 // Otherwise strength reduce the shift into an and.
4765 uint64_t Val = ~0ULL; // All ones.
4766 Val <<= ShAmtVal; // Shift over to the right spot.
4767 Val &= ~0ULL >> (64-TypeBits);
4768 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4771 BinaryOperator::createAnd(LHSI->getOperand(0),
4772 Mask, LHSI->getName()+".mask");
4773 Value *And = InsertNewInstBefore(AndI, I);
4774 return new ICmpInst(I.getPredicate(), And,
4775 ConstantExpr::getShl(CI, ShAmt));
4781 case Instruction::SDiv:
4782 case Instruction::UDiv:
4783 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4784 // Fold this div into the comparison, producing a range check.
4785 // Determine, based on the divide type, what the range is being
4786 // checked. If there is an overflow on the low or high side, remember
4787 // it, otherwise compute the range [low, hi) bounding the new value.
4788 // See: InsertRangeTest above for the kinds of replacements possible.
4789 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4790 // FIXME: If the operand types don't match the type of the divide
4791 // then don't attempt this transform. The code below doesn't have the
4792 // logic to deal with a signed divide and an unsigned compare (and
4793 // vice versa). This is because (x /s C1) <s C2 produces different
4794 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4795 // (x /u C1) <u C2. Simply casting the operands and result won't
4796 // work. :( The if statement below tests that condition and bails
4798 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4799 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4802 // Initialize the variables that will indicate the nature of the
4804 bool LoOverflow = false, HiOverflow = false;
4805 ConstantInt *LoBound = 0, *HiBound = 0;
4807 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4808 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4809 // C2 (CI). By solving for X we can turn this into a range check
4810 // instead of computing a divide.
4812 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4814 // Determine if the product overflows by seeing if the product is
4815 // not equal to the divide. Make sure we do the same kind of divide
4816 // as in the LHS instruction that we're folding.
4817 bool ProdOV = !DivRHS->isNullValue() &&
4818 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4819 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4821 // Get the ICmp opcode
4822 ICmpInst::Predicate predicate = I.getPredicate();
4824 if (DivRHS->isNullValue()) {
4825 // Don't hack on divide by zeros!
4826 } else if (!DivIsSigned) { // udiv
4828 LoOverflow = ProdOV;
4829 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4830 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4831 if (CI->isNullValue()) { // (X / pos) op 0
4833 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4835 } else if (isPositive(CI)) { // (X / pos) op pos
4837 LoOverflow = ProdOV;
4838 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4839 } else { // (X / pos) op neg
4840 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4841 LoOverflow = AddWithOverflow(LoBound, Prod,
4842 cast<ConstantInt>(DivRHSH));
4844 HiOverflow = ProdOV;
4846 } else { // Divisor is < 0.
4847 if (CI->isNullValue()) { // (X / neg) op 0
4848 LoBound = AddOne(DivRHS);
4849 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4850 if (HiBound == DivRHS)
4851 LoBound = 0; // - INTMIN = INTMIN
4852 } else if (isPositive(CI)) { // (X / neg) op pos
4853 HiOverflow = LoOverflow = ProdOV;
4855 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4856 HiBound = AddOne(Prod);
4857 } else { // (X / neg) op neg
4859 LoOverflow = HiOverflow = ProdOV;
4860 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4863 // Dividing by a negate swaps the condition.
4864 predicate = ICmpInst::getSwappedPredicate(predicate);
4868 Value *X = LHSI->getOperand(0);
4869 switch (predicate) {
4870 default: assert(0 && "Unhandled icmp opcode!");
4871 case ICmpInst::ICMP_EQ:
4872 if (LoOverflow && HiOverflow)
4873 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4874 else if (HiOverflow)
4875 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4876 ICmpInst::ICMP_UGE, X, LoBound);
4877 else if (LoOverflow)
4878 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4879 ICmpInst::ICMP_ULT, X, HiBound);
4881 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4883 case ICmpInst::ICMP_NE:
4884 if (LoOverflow && HiOverflow)
4885 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4886 else if (HiOverflow)
4887 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4888 ICmpInst::ICMP_ULT, X, LoBound);
4889 else if (LoOverflow)
4890 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4891 ICmpInst::ICMP_UGE, X, HiBound);
4893 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4895 case ICmpInst::ICMP_ULT:
4896 case ICmpInst::ICMP_SLT:
4898 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4899 return new ICmpInst(predicate, X, LoBound);
4900 case ICmpInst::ICMP_UGT:
4901 case ICmpInst::ICMP_SGT:
4903 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4904 if (predicate == ICmpInst::ICMP_UGT)
4905 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4907 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4914 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4915 if (I.isEquality()) {
4916 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4918 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4919 // the second operand is a constant, simplify a bit.
4920 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4921 switch (BO->getOpcode()) {
4922 case Instruction::SRem:
4923 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4924 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4926 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4927 if (V > 1 && isPowerOf2_64(V)) {
4928 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4929 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4930 return new ICmpInst(I.getPredicate(), NewRem,
4931 Constant::getNullValue(BO->getType()));
4935 case Instruction::Add:
4936 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4937 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4938 if (BO->hasOneUse())
4939 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4940 ConstantExpr::getSub(CI, BOp1C));
4941 } else if (CI->isNullValue()) {
4942 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4943 // efficiently invertible, or if the add has just this one use.
4944 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4946 if (Value *NegVal = dyn_castNegVal(BOp1))
4947 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4948 else if (Value *NegVal = dyn_castNegVal(BOp0))
4949 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4950 else if (BO->hasOneUse()) {
4951 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4953 InsertNewInstBefore(Neg, I);
4954 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4958 case Instruction::Xor:
4959 // For the xor case, we can xor two constants together, eliminating
4960 // the explicit xor.
4961 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4962 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4963 ConstantExpr::getXor(CI, BOC));
4966 case Instruction::Sub:
4967 // Replace (([sub|xor] A, B) != 0) with (A != B)
4968 if (CI->isNullValue())
4969 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4973 case Instruction::Or:
4974 // If bits are being or'd in that are not present in the constant we
4975 // are comparing against, then the comparison could never succeed!
4976 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4977 Constant *NotCI = ConstantExpr::getNot(CI);
4978 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4979 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4984 case Instruction::And:
4985 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4986 // If bits are being compared against that are and'd out, then the
4987 // comparison can never succeed!
4988 if (!ConstantExpr::getAnd(CI,
4989 ConstantExpr::getNot(BOC))->isNullValue())
4990 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4993 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4994 if (CI == BOC && isOneBitSet(CI))
4995 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
4996 ICmpInst::ICMP_NE, Op0,
4997 Constant::getNullValue(CI->getType()));
4999 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5000 if (isSignBit(BOC)) {
5001 Value *X = BO->getOperand(0);
5002 Constant *Zero = Constant::getNullValue(X->getType());
5003 ICmpInst::Predicate pred = isICMP_NE ?
5004 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5005 return new ICmpInst(pred, X, Zero);
5008 // ((X & ~7) == 0) --> X < 8
5009 if (CI->isNullValue() && isHighOnes(BOC)) {
5010 Value *X = BO->getOperand(0);
5011 Constant *NegX = ConstantExpr::getNeg(BOC);
5012 ICmpInst::Predicate pred = isICMP_NE ?
5013 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5014 return new ICmpInst(pred, X, NegX);
5020 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5021 // Handle set{eq|ne} <intrinsic>, intcst.
5022 switch (II->getIntrinsicID()) {
5024 case Intrinsic::bswap_i16:
5025 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5026 WorkList.push_back(II); // Dead?
5027 I.setOperand(0, II->getOperand(1));
5028 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5029 ByteSwap_16(CI->getZExtValue())));
5031 case Intrinsic::bswap_i32:
5032 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5033 WorkList.push_back(II); // Dead?
5034 I.setOperand(0, II->getOperand(1));
5035 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5036 ByteSwap_32(CI->getZExtValue())));
5038 case Intrinsic::bswap_i64:
5039 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5040 WorkList.push_back(II); // Dead?
5041 I.setOperand(0, II->getOperand(1));
5042 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5043 ByteSwap_64(CI->getZExtValue())));
5047 } else { // Not a ICMP_EQ/ICMP_NE
5048 // If the LHS is a cast from an integral value of the same size, then
5049 // since we know the RHS is a constant, try to simlify.
5050 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5051 Value *CastOp = Cast->getOperand(0);
5052 const Type *SrcTy = CastOp->getType();
5053 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5054 if (SrcTy->isInteger() &&
5055 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5056 // If this is an unsigned comparison, try to make the comparison use
5057 // smaller constant values.
5058 switch (I.getPredicate()) {
5060 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5061 ConstantInt *CUI = cast<ConstantInt>(CI);
5062 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5063 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5064 ConstantInt::get(SrcTy, -1));
5067 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5068 ConstantInt *CUI = cast<ConstantInt>(CI);
5069 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5070 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5071 Constant::getNullValue(SrcTy));
5081 // Handle icmp with constant RHS
5082 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5083 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5084 switch (LHSI->getOpcode()) {
5085 case Instruction::GetElementPtr:
5086 if (RHSC->isNullValue()) {
5087 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5088 bool isAllZeros = true;
5089 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5090 if (!isa<Constant>(LHSI->getOperand(i)) ||
5091 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5096 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5097 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5101 case Instruction::PHI:
5102 if (Instruction *NV = FoldOpIntoPhi(I))
5105 case Instruction::Select:
5106 // If either operand of the select is a constant, we can fold the
5107 // comparison into the select arms, which will cause one to be
5108 // constant folded and the select turned into a bitwise or.
5109 Value *Op1 = 0, *Op2 = 0;
5110 if (LHSI->hasOneUse()) {
5111 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5112 // Fold the known value into the constant operand.
5113 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5114 // Insert a new ICmp of the other select operand.
5115 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5116 LHSI->getOperand(2), RHSC,
5118 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5119 // Fold the known value into the constant operand.
5120 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5121 // Insert a new ICmp of the other select operand.
5122 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5123 LHSI->getOperand(1), RHSC,
5129 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5134 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5135 if (User *GEP = dyn_castGetElementPtr(Op0))
5136 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5138 if (User *GEP = dyn_castGetElementPtr(Op1))
5139 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5140 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5143 // Test to see if the operands of the icmp are casted versions of other
5144 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5146 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5147 if (isa<PointerType>(Op0->getType()) &&
5148 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5149 // We keep moving the cast from the left operand over to the right
5150 // operand, where it can often be eliminated completely.
5151 Op0 = CI->getOperand(0);
5153 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5154 // so eliminate it as well.
5155 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5156 Op1 = CI2->getOperand(0);
5158 // If Op1 is a constant, we can fold the cast into the constant.
5159 if (Op0->getType() != Op1->getType())
5160 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5161 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5163 // Otherwise, cast the RHS right before the icmp
5164 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5166 return new ICmpInst(I.getPredicate(), Op0, Op1);
5170 if (isa<CastInst>(Op0)) {
5171 // Handle the special case of: icmp (cast bool to X), <cst>
5172 // This comes up when you have code like
5175 // For generality, we handle any zero-extension of any operand comparison
5176 // with a constant or another cast from the same type.
5177 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5178 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5182 if (I.isEquality()) {
5183 Value *A, *B, *C, *D;
5184 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5185 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5186 Value *OtherVal = A == Op1 ? B : A;
5187 return new ICmpInst(I.getPredicate(), OtherVal,
5188 Constant::getNullValue(A->getType()));
5191 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5192 // A^c1 == C^c2 --> A == C^(c1^c2)
5193 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5194 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5195 if (Op1->hasOneUse()) {
5196 Constant *NC = ConstantExpr::getXor(C1, C2);
5197 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5198 return new ICmpInst(I.getPredicate(), A,
5199 InsertNewInstBefore(Xor, I));
5202 // A^B == A^D -> B == D
5203 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5204 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5205 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5206 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5210 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5211 (A == Op0 || B == Op0)) {
5212 // A == (A^B) -> B == 0
5213 Value *OtherVal = A == Op0 ? B : A;
5214 return new ICmpInst(I.getPredicate(), OtherVal,
5215 Constant::getNullValue(A->getType()));
5217 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5218 // (A-B) == A -> B == 0
5219 return new ICmpInst(I.getPredicate(), B,
5220 Constant::getNullValue(B->getType()));
5222 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5223 // A == (A-B) -> B == 0
5224 return new ICmpInst(I.getPredicate(), B,
5225 Constant::getNullValue(B->getType()));
5228 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5229 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5230 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5231 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5232 Value *X = 0, *Y = 0, *Z = 0;
5235 X = B; Y = D; Z = A;
5236 } else if (A == D) {
5237 X = B; Y = C; Z = A;
5238 } else if (B == C) {
5239 X = A; Y = D; Z = B;
5240 } else if (B == D) {
5241 X = A; Y = C; Z = B;
5244 if (X) { // Build (X^Y) & Z
5245 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5246 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5247 I.setOperand(0, Op1);
5248 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5253 return Changed ? &I : 0;
5256 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5257 // We only handle extending casts so far.
5259 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5260 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5261 Value *LHSCIOp = LHSCI->getOperand(0);
5262 const Type *SrcTy = LHSCIOp->getType();
5263 const Type *DestTy = LHSCI->getType();
5266 // We only handle extension cast instructions, so far. Enforce this.
5267 if (LHSCI->getOpcode() != Instruction::ZExt &&
5268 LHSCI->getOpcode() != Instruction::SExt)
5271 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5272 bool isSignedCmp = ICI.isSignedPredicate();
5274 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5275 // Not an extension from the same type?
5276 RHSCIOp = CI->getOperand(0);
5277 if (RHSCIOp->getType() != LHSCIOp->getType())
5280 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5281 // and the other is a zext), then we can't handle this.
5282 if (CI->getOpcode() != LHSCI->getOpcode())
5285 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5286 // then we can't handle this.
5287 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5290 // Okay, just insert a compare of the reduced operands now!
5291 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5294 // If we aren't dealing with a constant on the RHS, exit early
5295 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5299 // Compute the constant that would happen if we truncated to SrcTy then
5300 // reextended to DestTy.
5301 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5302 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5304 // If the re-extended constant didn't change...
5306 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5307 // For example, we might have:
5308 // %A = sext short %X to uint
5309 // %B = icmp ugt uint %A, 1330
5310 // It is incorrect to transform this into
5311 // %B = icmp ugt short %X, 1330
5312 // because %A may have negative value.
5314 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5315 // OR operation is EQ/NE.
5316 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5317 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5322 // The re-extended constant changed so the constant cannot be represented
5323 // in the shorter type. Consequently, we cannot emit a simple comparison.
5325 // First, handle some easy cases. We know the result cannot be equal at this
5326 // point so handle the ICI.isEquality() cases
5327 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5328 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5329 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5330 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5332 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5333 // should have been folded away previously and not enter in here.
5336 // We're performing a signed comparison.
5337 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5338 Result = ConstantInt::getFalse(); // X < (small) --> false
5340 Result = ConstantInt::getTrue(); // X < (large) --> true
5342 // We're performing an unsigned comparison.
5344 // We're performing an unsigned comp with a sign extended value.
5345 // This is true if the input is >= 0. [aka >s -1]
5346 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5347 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5348 NegOne, ICI.getName()), ICI);
5350 // Unsigned extend & unsigned compare -> always true.
5351 Result = ConstantInt::getTrue();
5355 // Finally, return the value computed.
5356 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5357 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5358 return ReplaceInstUsesWith(ICI, Result);
5360 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5361 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5362 "ICmp should be folded!");
5363 if (Constant *CI = dyn_cast<Constant>(Result))
5364 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5366 return BinaryOperator::createNot(Result);
5370 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5371 assert(I.getOperand(1)->getType() == Type::Int8Ty);
5372 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5374 // shl X, 0 == X and shr X, 0 == X
5375 // shl 0, X == 0 and shr 0, X == 0
5376 if (Op1 == Constant::getNullValue(Type::Int8Ty) ||
5377 Op0 == Constant::getNullValue(Op0->getType()))
5378 return ReplaceInstUsesWith(I, Op0);
5380 if (isa<UndefValue>(Op0)) {
5381 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5382 return ReplaceInstUsesWith(I, Op0);
5383 else // undef << X -> 0, undef >>u X -> 0
5384 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5386 if (isa<UndefValue>(Op1)) {
5387 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5388 return ReplaceInstUsesWith(I, Op0);
5389 else // X << undef, X >>u undef -> 0
5390 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5393 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5394 if (I.getOpcode() == Instruction::AShr)
5395 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5396 if (CSI->isAllOnesValue())
5397 return ReplaceInstUsesWith(I, CSI);
5399 // Try to fold constant and into select arguments.
5400 if (isa<Constant>(Op0))
5401 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5402 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5405 // See if we can turn a signed shr into an unsigned shr.
5406 if (I.isArithmeticShift()) {
5407 if (MaskedValueIsZero(Op0,
5408 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5409 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5413 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5414 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5419 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5421 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5422 bool isSignedShift = I.getOpcode() == Instruction::AShr;
5423 bool isUnsignedShift = !isSignedShift;
5425 // See if we can simplify any instructions used by the instruction whose sole
5426 // purpose is to compute bits we don't care about.
5427 uint64_t KnownZero, KnownOne;
5428 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
5429 KnownZero, KnownOne))
5432 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5433 // of a signed value.
5435 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5436 if (Op1->getZExtValue() >= TypeBits) {
5437 if (isUnsignedShift || isLeftShift)
5438 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5440 I.setOperand(1, ConstantInt::get(Type::Int8Ty, TypeBits-1));
5445 // ((X*C1) << C2) == (X * (C1 << C2))
5446 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5447 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5448 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5449 return BinaryOperator::createMul(BO->getOperand(0),
5450 ConstantExpr::getShl(BOOp, Op1));
5452 // Try to fold constant and into select arguments.
5453 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5454 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5456 if (isa<PHINode>(Op0))
5457 if (Instruction *NV = FoldOpIntoPhi(I))
5460 if (Op0->hasOneUse()) {
5461 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5462 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5465 switch (Op0BO->getOpcode()) {
5467 case Instruction::Add:
5468 case Instruction::And:
5469 case Instruction::Or:
5470 case Instruction::Xor:
5471 // These operators commute.
5472 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5473 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5474 match(Op0BO->getOperand(1),
5475 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5476 Instruction *YS = new ShiftInst(Instruction::Shl,
5477 Op0BO->getOperand(0), Op1,
5479 InsertNewInstBefore(YS, I); // (Y << C)
5481 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5482 Op0BO->getOperand(1)->getName());
5483 InsertNewInstBefore(X, I); // (X + (Y << C))
5484 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5485 C2 = ConstantExpr::getShl(C2, Op1);
5486 return BinaryOperator::createAnd(X, C2);
5489 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5490 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5491 match(Op0BO->getOperand(1),
5492 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5493 m_ConstantInt(CC))) && V2 == Op1 &&
5494 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5495 Instruction *YS = new ShiftInst(Instruction::Shl,
5496 Op0BO->getOperand(0), Op1,
5498 InsertNewInstBefore(YS, I); // (Y << C)
5500 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5501 V1->getName()+".mask");
5502 InsertNewInstBefore(XM, I); // X & (CC << C)
5504 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5508 case Instruction::Sub:
5509 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5510 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5511 match(Op0BO->getOperand(0),
5512 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5513 Instruction *YS = new ShiftInst(Instruction::Shl,
5514 Op0BO->getOperand(1), Op1,
5516 InsertNewInstBefore(YS, I); // (Y << C)
5518 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5519 Op0BO->getOperand(0)->getName());
5520 InsertNewInstBefore(X, I); // (X + (Y << C))
5521 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5522 C2 = ConstantExpr::getShl(C2, Op1);
5523 return BinaryOperator::createAnd(X, C2);
5526 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5527 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5528 match(Op0BO->getOperand(0),
5529 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5530 m_ConstantInt(CC))) && V2 == Op1 &&
5531 cast<BinaryOperator>(Op0BO->getOperand(0))
5532 ->getOperand(0)->hasOneUse()) {
5533 Instruction *YS = new ShiftInst(Instruction::Shl,
5534 Op0BO->getOperand(1), Op1,
5536 InsertNewInstBefore(YS, I); // (Y << C)
5538 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5539 V1->getName()+".mask");
5540 InsertNewInstBefore(XM, I); // X & (CC << C)
5542 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5549 // If the operand is an bitwise operator with a constant RHS, and the
5550 // shift is the only use, we can pull it out of the shift.
5551 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5552 bool isValid = true; // Valid only for And, Or, Xor
5553 bool highBitSet = false; // Transform if high bit of constant set?
5555 switch (Op0BO->getOpcode()) {
5556 default: isValid = false; break; // Do not perform transform!
5557 case Instruction::Add:
5558 isValid = isLeftShift;
5560 case Instruction::Or:
5561 case Instruction::Xor:
5564 case Instruction::And:
5569 // If this is a signed shift right, and the high bit is modified
5570 // by the logical operation, do not perform the transformation.
5571 // The highBitSet boolean indicates the value of the high bit of
5572 // the constant which would cause it to be modified for this
5575 if (isValid && !isLeftShift && isSignedShift) {
5576 uint64_t Val = Op0C->getZExtValue();
5577 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5581 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5583 Instruction *NewShift =
5584 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5587 InsertNewInstBefore(NewShift, I);
5589 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5596 // Find out if this is a shift of a shift by a constant.
5597 ShiftInst *ShiftOp = 0;
5598 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5600 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5601 // If this is a noop-integer cast of a shift instruction, use the shift.
5602 if (isa<ShiftInst>(CI->getOperand(0))) {
5603 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5607 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5608 // Find the operands and properties of the input shift. Note that the
5609 // signedness of the input shift may differ from the current shift if there
5610 // is a noop cast between the two.
5611 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5612 bool isShiftOfSignedShift = ShiftOp->getOpcode() == Instruction::AShr;
5613 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5615 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5617 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5618 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5620 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5621 if (isLeftShift == isShiftOfLeftShift) {
5622 // Do not fold these shifts if the first one is signed and the second one
5623 // is unsigned and this is a right shift. Further, don't do any folding
5625 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5628 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5629 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5630 Amt = Op0->getType()->getPrimitiveSizeInBits();
5632 Value *Op = ShiftOp->getOperand(0);
5633 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5634 ConstantInt::get(Type::Int8Ty, Amt));
5635 if (I.getType() == ShiftResult->getType())
5637 InsertNewInstBefore(ShiftResult, I);
5638 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5641 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5642 // signed types, we can only support the (A >> c1) << c2 configuration,
5643 // because it can not turn an arbitrary bit of A into a sign bit.
5644 if (isUnsignedShift || isLeftShift) {
5645 // Calculate bitmask for what gets shifted off the edge.
5646 Constant *C = ConstantInt::getAllOnesValue(I.getType());
5648 C = ConstantExpr::getShl(C, ShiftAmt1C);
5650 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5652 Value *Op = ShiftOp->getOperand(0);
5655 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5656 InsertNewInstBefore(Mask, I);
5658 // Figure out what flavor of shift we should use...
5659 if (ShiftAmt1 == ShiftAmt2) {
5660 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5661 } else if (ShiftAmt1 < ShiftAmt2) {
5662 return new ShiftInst(I.getOpcode(), Mask,
5663 ConstantInt::get(Type::Int8Ty, ShiftAmt2-ShiftAmt1));
5664 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5665 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5666 return new ShiftInst(Instruction::LShr, Mask,
5667 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5669 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5670 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5673 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5674 Instruction *Shift =
5675 new ShiftInst(ShiftOp->getOpcode(), Mask,
5676 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5677 InsertNewInstBefore(Shift, I);
5679 C = ConstantInt::getAllOnesValue(Shift->getType());
5680 C = ConstantExpr::getShl(C, Op1);
5681 return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5684 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5685 // this case, C1 == C2 and C1 is 8, 16, or 32.
5686 if (ShiftAmt1 == ShiftAmt2) {
5687 const Type *SExtType = 0;
5688 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5689 case 8 : SExtType = Type::Int8Ty; break;
5690 case 16: SExtType = Type::Int16Ty; break;
5691 case 32: SExtType = Type::Int32Ty; break;
5695 Instruction *NewTrunc =
5696 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5697 InsertNewInstBefore(NewTrunc, I);
5698 return new SExtInst(NewTrunc, I.getType());
5707 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5708 /// expression. If so, decompose it, returning some value X, such that Val is
5711 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5713 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5714 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5715 Offset = CI->getZExtValue();
5717 return ConstantInt::get(Type::Int32Ty, 0);
5718 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5719 if (I->getNumOperands() == 2) {
5720 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5721 if (I->getOpcode() == Instruction::Shl) {
5722 // This is a value scaled by '1 << the shift amt'.
5723 Scale = 1U << CUI->getZExtValue();
5725 return I->getOperand(0);
5726 } else if (I->getOpcode() == Instruction::Mul) {
5727 // This value is scaled by 'CUI'.
5728 Scale = CUI->getZExtValue();
5730 return I->getOperand(0);
5731 } else if (I->getOpcode() == Instruction::Add) {
5732 // We have X+C. Check to see if we really have (X*C2)+C1,
5733 // where C1 is divisible by C2.
5736 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5737 Offset += CUI->getZExtValue();
5738 if (SubScale > 1 && (Offset % SubScale == 0)) {
5747 // Otherwise, we can't look past this.
5754 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5755 /// try to eliminate the cast by moving the type information into the alloc.
5756 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5757 AllocationInst &AI) {
5758 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5759 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5761 // Remove any uses of AI that are dead.
5762 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5763 std::vector<Instruction*> DeadUsers;
5764 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5765 Instruction *User = cast<Instruction>(*UI++);
5766 if (isInstructionTriviallyDead(User)) {
5767 while (UI != E && *UI == User)
5768 ++UI; // If this instruction uses AI more than once, don't break UI.
5770 // Add operands to the worklist.
5771 AddUsesToWorkList(*User);
5773 DOUT << "IC: DCE: " << *User;
5775 User->eraseFromParent();
5776 removeFromWorkList(User);
5780 // Get the type really allocated and the type casted to.
5781 const Type *AllocElTy = AI.getAllocatedType();
5782 const Type *CastElTy = PTy->getElementType();
5783 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5785 unsigned AllocElTyAlign = TD->getTypeAlignmentABI(AllocElTy);
5786 unsigned CastElTyAlign = TD->getTypeAlignmentABI(CastElTy);
5787 if (CastElTyAlign < AllocElTyAlign) return 0;
5789 // If the allocation has multiple uses, only promote it if we are strictly
5790 // increasing the alignment of the resultant allocation. If we keep it the
5791 // same, we open the door to infinite loops of various kinds.
5792 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5794 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5795 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5796 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5798 // See if we can satisfy the modulus by pulling a scale out of the array
5800 unsigned ArraySizeScale, ArrayOffset;
5801 Value *NumElements = // See if the array size is a decomposable linear expr.
5802 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5804 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5806 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5807 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5809 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5814 // If the allocation size is constant, form a constant mul expression
5815 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5816 if (isa<ConstantInt>(NumElements))
5817 Amt = ConstantExpr::getMul(
5818 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5819 // otherwise multiply the amount and the number of elements
5820 else if (Scale != 1) {
5821 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5822 Amt = InsertNewInstBefore(Tmp, AI);
5826 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5827 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5828 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5829 Amt = InsertNewInstBefore(Tmp, AI);
5832 std::string Name = AI.getName(); AI.setName("");
5833 AllocationInst *New;
5834 if (isa<MallocInst>(AI))
5835 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5837 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5838 InsertNewInstBefore(New, AI);
5840 // If the allocation has multiple uses, insert a cast and change all things
5841 // that used it to use the new cast. This will also hack on CI, but it will
5843 if (!AI.hasOneUse()) {
5844 AddUsesToWorkList(AI);
5845 // New is the allocation instruction, pointer typed. AI is the original
5846 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5847 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5848 InsertNewInstBefore(NewCast, AI);
5849 AI.replaceAllUsesWith(NewCast);
5851 return ReplaceInstUsesWith(CI, New);
5854 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5855 /// and return it without inserting any new casts. This is used by code that
5856 /// tries to decide whether promoting or shrinking integer operations to wider
5857 /// or smaller types will allow us to eliminate a truncate or extend.
5858 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5859 int &NumCastsRemoved) {
5860 if (isa<Constant>(V)) return true;
5862 Instruction *I = dyn_cast<Instruction>(V);
5863 if (!I || !I->hasOneUse()) return false;
5865 switch (I->getOpcode()) {
5866 case Instruction::And:
5867 case Instruction::Or:
5868 case Instruction::Xor:
5869 // These operators can all arbitrarily be extended or truncated.
5870 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5871 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5872 case Instruction::AShr:
5873 case Instruction::LShr:
5874 case Instruction::Shl:
5875 // If this is just a bitcast changing the sign of the operation, we can
5876 // convert if the operand can be converted.
5877 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5878 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5880 case Instruction::Trunc:
5881 case Instruction::ZExt:
5882 case Instruction::SExt:
5883 case Instruction::BitCast:
5884 // If this is a cast from the destination type, we can trivially eliminate
5885 // it, and this will remove a cast overall.
5886 if (I->getOperand(0)->getType() == Ty) {
5887 // If the first operand is itself a cast, and is eliminable, do not count
5888 // this as an eliminable cast. We would prefer to eliminate those two
5890 if (isa<CastInst>(I->getOperand(0)))
5898 // TODO: Can handle more cases here.
5905 /// EvaluateInDifferentType - Given an expression that
5906 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5907 /// evaluate the expression.
5908 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5910 if (Constant *C = dyn_cast<Constant>(V))
5911 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5913 // Otherwise, it must be an instruction.
5914 Instruction *I = cast<Instruction>(V);
5915 Instruction *Res = 0;
5916 switch (I->getOpcode()) {
5917 case Instruction::And:
5918 case Instruction::Or:
5919 case Instruction::Xor: {
5920 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5921 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5922 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5923 LHS, RHS, I->getName());
5926 case Instruction::AShr:
5927 case Instruction::LShr:
5928 case Instruction::Shl: {
5929 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5930 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5931 I->getOperand(1), I->getName());
5934 case Instruction::Trunc:
5935 case Instruction::ZExt:
5936 case Instruction::SExt:
5937 case Instruction::BitCast:
5938 // If the source type of the cast is the type we're trying for then we can
5939 // just return the source. There's no need to insert it because its not new.
5940 if (I->getOperand(0)->getType() == Ty)
5941 return I->getOperand(0);
5943 // Some other kind of cast, which shouldn't happen, so just ..
5946 // TODO: Can handle more cases here.
5947 assert(0 && "Unreachable!");
5951 return InsertNewInstBefore(Res, *I);
5954 /// @brief Implement the transforms common to all CastInst visitors.
5955 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5956 Value *Src = CI.getOperand(0);
5958 // Casting undef to anything results in undef so might as just replace it and
5959 // get rid of the cast.
5960 if (isa<UndefValue>(Src)) // cast undef -> undef
5961 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5963 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5964 // eliminate it now.
5965 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5966 if (Instruction::CastOps opc =
5967 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5968 // The first cast (CSrc) is eliminable so we need to fix up or replace
5969 // the second cast (CI). CSrc will then have a good chance of being dead.
5970 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
5974 // If casting the result of a getelementptr instruction with no offset, turn
5975 // this into a cast of the original pointer!
5977 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5978 bool AllZeroOperands = true;
5979 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5980 if (!isa<Constant>(GEP->getOperand(i)) ||
5981 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5982 AllZeroOperands = false;
5985 if (AllZeroOperands) {
5986 // Changing the cast operand is usually not a good idea but it is safe
5987 // here because the pointer operand is being replaced with another
5988 // pointer operand so the opcode doesn't need to change.
5989 CI.setOperand(0, GEP->getOperand(0));
5994 // If we are casting a malloc or alloca to a pointer to a type of the same
5995 // size, rewrite the allocation instruction to allocate the "right" type.
5996 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5997 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6000 // If we are casting a select then fold the cast into the select
6001 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6002 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6005 // If we are casting a PHI then fold the cast into the PHI
6006 if (isa<PHINode>(Src))
6007 if (Instruction *NV = FoldOpIntoPhi(CI))
6013 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
6014 /// integers. This function implements the common transforms for all those
6016 /// @brief Implement the transforms common to CastInst with integer operands
6017 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6018 if (Instruction *Result = commonCastTransforms(CI))
6021 Value *Src = CI.getOperand(0);
6022 const Type *SrcTy = Src->getType();
6023 const Type *DestTy = CI.getType();
6024 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6025 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6027 // See if we can simplify any instructions used by the LHS whose sole
6028 // purpose is to compute bits we don't care about.
6029 uint64_t KnownZero = 0, KnownOne = 0;
6030 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
6031 KnownZero, KnownOne))
6034 // If the source isn't an instruction or has more than one use then we
6035 // can't do anything more.
6036 Instruction *SrcI = dyn_cast<Instruction>(Src);
6037 if (!SrcI || !Src->hasOneUse())
6040 // Attempt to propagate the cast into the instruction.
6041 int NumCastsRemoved = 0;
6042 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6043 // If this cast is a truncate, evaluting in a different type always
6044 // eliminates the cast, so it is always a win. If this is a noop-cast
6045 // this just removes a noop cast which isn't pointful, but simplifies
6046 // the code. If this is a zero-extension, we need to do an AND to
6047 // maintain the clear top-part of the computation, so we require that
6048 // the input have eliminated at least one cast. If this is a sign
6049 // extension, we insert two new casts (to do the extension) so we
6050 // require that two casts have been eliminated.
6051 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6053 switch (CI.getOpcode()) {
6054 case Instruction::Trunc:
6057 case Instruction::ZExt:
6058 DoXForm = NumCastsRemoved >= 1;
6060 case Instruction::SExt:
6061 DoXForm = NumCastsRemoved >= 2;
6063 case Instruction::BitCast:
6067 // All the others use floating point so we shouldn't actually
6068 // get here because of the check above.
6069 assert(!"Unknown cast type .. unreachable");
6075 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6076 CI.getOpcode() == Instruction::SExt);
6077 assert(Res->getType() == DestTy);
6078 switch (CI.getOpcode()) {
6079 default: assert(0 && "Unknown cast type!");
6080 case Instruction::Trunc:
6081 case Instruction::BitCast:
6082 // Just replace this cast with the result.
6083 return ReplaceInstUsesWith(CI, Res);
6084 case Instruction::ZExt: {
6085 // We need to emit an AND to clear the high bits.
6086 assert(SrcBitSize < DestBitSize && "Not a zext?");
6088 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6089 if (DestBitSize < 64)
6090 C = ConstantExpr::getTrunc(C, DestTy);
6091 return BinaryOperator::createAnd(Res, C);
6093 case Instruction::SExt:
6094 // We need to emit a cast to truncate, then a cast to sext.
6095 return CastInst::create(Instruction::SExt,
6096 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6102 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6103 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6105 switch (SrcI->getOpcode()) {
6106 case Instruction::Add:
6107 case Instruction::Mul:
6108 case Instruction::And:
6109 case Instruction::Or:
6110 case Instruction::Xor:
6111 // If we are discarding information, or just changing the sign,
6113 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6114 // Don't insert two casts if they cannot be eliminated. We allow
6115 // two casts to be inserted if the sizes are the same. This could
6116 // only be converting signedness, which is a noop.
6117 if (DestBitSize == SrcBitSize ||
6118 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6119 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6120 Instruction::CastOps opcode = CI.getOpcode();
6121 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6122 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6123 return BinaryOperator::create(
6124 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6128 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6129 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6130 SrcI->getOpcode() == Instruction::Xor &&
6131 Op1 == ConstantInt::getTrue() &&
6132 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6133 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6134 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6137 case Instruction::SDiv:
6138 case Instruction::UDiv:
6139 case Instruction::SRem:
6140 case Instruction::URem:
6141 // If we are just changing the sign, rewrite.
6142 if (DestBitSize == SrcBitSize) {
6143 // Don't insert two casts if they cannot be eliminated. We allow
6144 // two casts to be inserted if the sizes are the same. This could
6145 // only be converting signedness, which is a noop.
6146 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6147 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6148 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6150 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6152 return BinaryOperator::create(
6153 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6158 case Instruction::Shl:
6159 // Allow changing the sign of the source operand. Do not allow
6160 // changing the size of the shift, UNLESS the shift amount is a
6161 // constant. We must not change variable sized shifts to a smaller
6162 // size, because it is undefined to shift more bits out than exist
6164 if (DestBitSize == SrcBitSize ||
6165 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6166 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6167 Instruction::BitCast : Instruction::Trunc);
6168 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6169 return new ShiftInst(Instruction::Shl, Op0c, Op1);
6172 case Instruction::AShr:
6173 // If this is a signed shr, and if all bits shifted in are about to be
6174 // truncated off, turn it into an unsigned shr to allow greater
6176 if (DestBitSize < SrcBitSize &&
6177 isa<ConstantInt>(Op1)) {
6178 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6179 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6180 // Insert the new logical shift right.
6181 return new ShiftInst(Instruction::LShr, Op0, Op1);
6186 case Instruction::ICmp:
6187 // If we are just checking for a icmp eq of a single bit and casting it
6188 // to an integer, then shift the bit to the appropriate place and then
6189 // cast to integer to avoid the comparison.
6190 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6191 uint64_t Op1CV = Op1C->getZExtValue();
6192 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6193 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6194 // cast (X == 1) to int --> X iff X has only the low bit set.
6195 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6196 // cast (X != 0) to int --> X iff X has only the low bit set.
6197 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6198 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6199 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6200 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6201 // If Op1C some other power of two, convert:
6202 uint64_t KnownZero, KnownOne;
6203 uint64_t TypeMask = Op1C->getType()->getBitMask();
6204 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6206 // This only works for EQ and NE
6207 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6208 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6211 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6212 bool isNE = pred == ICmpInst::ICMP_NE;
6213 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6214 // (X&4) == 2 --> false
6215 // (X&4) != 2 --> true
6216 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6217 Res = ConstantExpr::getZExt(Res, CI.getType());
6218 return ReplaceInstUsesWith(CI, Res);
6221 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6224 // Perform a logical shr by shiftamt.
6225 // Insert the shift to put the result in the low bit.
6226 In = InsertNewInstBefore(
6227 new ShiftInst(Instruction::LShr, In,
6228 ConstantInt::get(Type::Int8Ty, ShiftAmt),
6229 In->getName()+".lobit"), CI);
6232 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6233 Constant *One = ConstantInt::get(In->getType(), 1);
6234 In = BinaryOperator::createXor(In, One, "tmp");
6235 InsertNewInstBefore(cast<Instruction>(In), CI);
6238 if (CI.getType() == In->getType())
6239 return ReplaceInstUsesWith(CI, In);
6241 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6250 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6251 if (Instruction *Result = commonIntCastTransforms(CI))
6254 Value *Src = CI.getOperand(0);
6255 const Type *Ty = CI.getType();
6256 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6258 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6259 switch (SrcI->getOpcode()) {
6261 case Instruction::LShr:
6262 // We can shrink lshr to something smaller if we know the bits shifted in
6263 // are already zeros.
6264 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6265 unsigned ShAmt = ShAmtV->getZExtValue();
6267 // Get a mask for the bits shifting in.
6268 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6269 Value* SrcIOp0 = SrcI->getOperand(0);
6270 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6271 if (ShAmt >= DestBitWidth) // All zeros.
6272 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6274 // Okay, we can shrink this. Truncate the input, then return a new
6276 Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6277 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6279 } else { // This is a variable shr.
6281 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6282 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6283 // loop-invariant and CSE'd.
6284 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6285 Value *One = ConstantInt::get(SrcI->getType(), 1);
6287 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6288 SrcI->getOperand(1),
6290 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6291 SrcI->getOperand(0),
6293 Value *Zero = Constant::getNullValue(V->getType());
6294 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6304 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6305 // If one of the common conversion will work ..
6306 if (Instruction *Result = commonIntCastTransforms(CI))
6309 Value *Src = CI.getOperand(0);
6311 // If this is a cast of a cast
6312 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6313 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6314 // types and if the sizes are just right we can convert this into a logical
6315 // 'and' which will be much cheaper than the pair of casts.
6316 if (isa<TruncInst>(CSrc)) {
6317 // Get the sizes of the types involved
6318 Value *A = CSrc->getOperand(0);
6319 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6320 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6321 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6322 // If we're actually extending zero bits and the trunc is a no-op
6323 if (MidSize < DstSize && SrcSize == DstSize) {
6324 // Replace both of the casts with an And of the type mask.
6325 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
6326 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6328 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6329 // Unfortunately, if the type changed, we need to cast it back.
6330 if (And->getType() != CI.getType()) {
6331 And->setName(CSrc->getName()+".mask");
6332 InsertNewInstBefore(And, CI);
6333 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6343 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6344 return commonIntCastTransforms(CI);
6347 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6348 return commonCastTransforms(CI);
6351 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6352 return commonCastTransforms(CI);
6355 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6356 return commonCastTransforms(CI);
6359 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6360 return commonCastTransforms(CI);
6363 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6364 return commonCastTransforms(CI);
6367 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6368 return commonCastTransforms(CI);
6371 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6372 return commonCastTransforms(CI);
6375 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6376 return commonCastTransforms(CI);
6379 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6381 // If the operands are integer typed then apply the integer transforms,
6382 // otherwise just apply the common ones.
6383 Value *Src = CI.getOperand(0);
6384 const Type *SrcTy = Src->getType();
6385 const Type *DestTy = CI.getType();
6387 if (SrcTy->isInteger() && DestTy->isInteger()) {
6388 if (Instruction *Result = commonIntCastTransforms(CI))
6391 if (Instruction *Result = commonCastTransforms(CI))
6396 // Get rid of casts from one type to the same type. These are useless and can
6397 // be replaced by the operand.
6398 if (DestTy == Src->getType())
6399 return ReplaceInstUsesWith(CI, Src);
6401 // If the source and destination are pointers, and this cast is equivalent to
6402 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6403 // This can enhance SROA and other transforms that want type-safe pointers.
6404 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6405 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6406 const Type *DstElTy = DstPTy->getElementType();
6407 const Type *SrcElTy = SrcPTy->getElementType();
6409 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6410 unsigned NumZeros = 0;
6411 while (SrcElTy != DstElTy &&
6412 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6413 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6414 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6418 // If we found a path from the src to dest, create the getelementptr now.
6419 if (SrcElTy == DstElTy) {
6420 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6421 return new GetElementPtrInst(Src, Idxs);
6426 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6427 if (SVI->hasOneUse()) {
6428 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6429 // a bitconvert to a vector with the same # elts.
6430 if (isa<PackedType>(DestTy) &&
6431 cast<PackedType>(DestTy)->getNumElements() ==
6432 SVI->getType()->getNumElements()) {
6434 // If either of the operands is a cast from CI.getType(), then
6435 // evaluating the shuffle in the casted destination's type will allow
6436 // us to eliminate at least one cast.
6437 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6438 Tmp->getOperand(0)->getType() == DestTy) ||
6439 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6440 Tmp->getOperand(0)->getType() == DestTy)) {
6441 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6442 SVI->getOperand(0), DestTy, &CI);
6443 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6444 SVI->getOperand(1), DestTy, &CI);
6445 // Return a new shuffle vector. Use the same element ID's, as we
6446 // know the vector types match #elts.
6447 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6455 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6457 /// %D = select %cond, %C, %A
6459 /// %C = select %cond, %B, 0
6462 /// Assuming that the specified instruction is an operand to the select, return
6463 /// a bitmask indicating which operands of this instruction are foldable if they
6464 /// equal the other incoming value of the select.
6466 static unsigned GetSelectFoldableOperands(Instruction *I) {
6467 switch (I->getOpcode()) {
6468 case Instruction::Add:
6469 case Instruction::Mul:
6470 case Instruction::And:
6471 case Instruction::Or:
6472 case Instruction::Xor:
6473 return 3; // Can fold through either operand.
6474 case Instruction::Sub: // Can only fold on the amount subtracted.
6475 case Instruction::Shl: // Can only fold on the shift amount.
6476 case Instruction::LShr:
6477 case Instruction::AShr:
6480 return 0; // Cannot fold
6484 /// GetSelectFoldableConstant - For the same transformation as the previous
6485 /// function, return the identity constant that goes into the select.
6486 static Constant *GetSelectFoldableConstant(Instruction *I) {
6487 switch (I->getOpcode()) {
6488 default: assert(0 && "This cannot happen!"); abort();
6489 case Instruction::Add:
6490 case Instruction::Sub:
6491 case Instruction::Or:
6492 case Instruction::Xor:
6493 return Constant::getNullValue(I->getType());
6494 case Instruction::Shl:
6495 case Instruction::LShr:
6496 case Instruction::AShr:
6497 return Constant::getNullValue(Type::Int8Ty);
6498 case Instruction::And:
6499 return ConstantInt::getAllOnesValue(I->getType());
6500 case Instruction::Mul:
6501 return ConstantInt::get(I->getType(), 1);
6505 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6506 /// have the same opcode and only one use each. Try to simplify this.
6507 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6509 if (TI->getNumOperands() == 1) {
6510 // If this is a non-volatile load or a cast from the same type,
6513 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6516 return 0; // unknown unary op.
6519 // Fold this by inserting a select from the input values.
6520 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6521 FI->getOperand(0), SI.getName()+".v");
6522 InsertNewInstBefore(NewSI, SI);
6523 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6527 // Only handle binary, compare and shift operators here.
6528 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
6531 // Figure out if the operations have any operands in common.
6532 Value *MatchOp, *OtherOpT, *OtherOpF;
6534 if (TI->getOperand(0) == FI->getOperand(0)) {
6535 MatchOp = TI->getOperand(0);
6536 OtherOpT = TI->getOperand(1);
6537 OtherOpF = FI->getOperand(1);
6538 MatchIsOpZero = true;
6539 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6540 MatchOp = TI->getOperand(1);
6541 OtherOpT = TI->getOperand(0);
6542 OtherOpF = FI->getOperand(0);
6543 MatchIsOpZero = false;
6544 } else if (!TI->isCommutative()) {
6546 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6547 MatchOp = TI->getOperand(0);
6548 OtherOpT = TI->getOperand(1);
6549 OtherOpF = FI->getOperand(0);
6550 MatchIsOpZero = true;
6551 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6552 MatchOp = TI->getOperand(1);
6553 OtherOpT = TI->getOperand(0);
6554 OtherOpF = FI->getOperand(1);
6555 MatchIsOpZero = true;
6560 // If we reach here, they do have operations in common.
6561 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6562 OtherOpF, SI.getName()+".v");
6563 InsertNewInstBefore(NewSI, SI);
6565 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6567 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6569 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6572 assert(isa<ShiftInst>(TI) && "Should only have Shift here");
6574 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6576 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6579 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6580 Value *CondVal = SI.getCondition();
6581 Value *TrueVal = SI.getTrueValue();
6582 Value *FalseVal = SI.getFalseValue();
6584 // select true, X, Y -> X
6585 // select false, X, Y -> Y
6586 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6587 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6589 // select C, X, X -> X
6590 if (TrueVal == FalseVal)
6591 return ReplaceInstUsesWith(SI, TrueVal);
6593 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6594 return ReplaceInstUsesWith(SI, FalseVal);
6595 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6596 return ReplaceInstUsesWith(SI, TrueVal);
6597 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6598 if (isa<Constant>(TrueVal))
6599 return ReplaceInstUsesWith(SI, TrueVal);
6601 return ReplaceInstUsesWith(SI, FalseVal);
6604 if (SI.getType() == Type::Int1Ty) {
6605 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6606 if (C->getZExtValue()) {
6607 // Change: A = select B, true, C --> A = or B, C
6608 return BinaryOperator::createOr(CondVal, FalseVal);
6610 // Change: A = select B, false, C --> A = and !B, C
6612 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6613 "not."+CondVal->getName()), SI);
6614 return BinaryOperator::createAnd(NotCond, FalseVal);
6616 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6617 if (C->getZExtValue() == false) {
6618 // Change: A = select B, C, false --> A = and B, C
6619 return BinaryOperator::createAnd(CondVal, TrueVal);
6621 // Change: A = select B, C, true --> A = or !B, C
6623 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6624 "not."+CondVal->getName()), SI);
6625 return BinaryOperator::createOr(NotCond, TrueVal);
6630 // Selecting between two integer constants?
6631 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6632 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6633 // select C, 1, 0 -> cast C to int
6634 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6635 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6636 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6637 // select C, 0, 1 -> cast !C to int
6639 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6640 "not."+CondVal->getName()), SI);
6641 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6644 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6646 // (x <s 0) ? -1 : 0 -> ashr x, 31
6647 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6648 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6649 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6650 bool CanXForm = false;
6651 if (IC->isSignedPredicate())
6652 CanXForm = CmpCst->isNullValue() &&
6653 IC->getPredicate() == ICmpInst::ICMP_SLT;
6655 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6656 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6657 IC->getPredicate() == ICmpInst::ICMP_UGT;
6661 // The comparison constant and the result are not neccessarily the
6662 // same width. Make an all-ones value by inserting a AShr.
6663 Value *X = IC->getOperand(0);
6664 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6665 Constant *ShAmt = ConstantInt::get(Type::Int8Ty, Bits-1);
6666 Instruction *SRA = new ShiftInst(Instruction::AShr, X,
6668 InsertNewInstBefore(SRA, SI);
6670 // Finally, convert to the type of the select RHS. We figure out
6671 // if this requires a SExt, Trunc or BitCast based on the sizes.
6672 Instruction::CastOps opc = Instruction::BitCast;
6673 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6674 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6675 if (SRASize < SISize)
6676 opc = Instruction::SExt;
6677 else if (SRASize > SISize)
6678 opc = Instruction::Trunc;
6679 return CastInst::create(opc, SRA, SI.getType());
6684 // If one of the constants is zero (we know they can't both be) and we
6685 // have a fcmp instruction with zero, and we have an 'and' with the
6686 // non-constant value, eliminate this whole mess. This corresponds to
6687 // cases like this: ((X & 27) ? 27 : 0)
6688 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6689 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6690 cast<Constant>(IC->getOperand(1))->isNullValue())
6691 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6692 if (ICA->getOpcode() == Instruction::And &&
6693 isa<ConstantInt>(ICA->getOperand(1)) &&
6694 (ICA->getOperand(1) == TrueValC ||
6695 ICA->getOperand(1) == FalseValC) &&
6696 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6697 // Okay, now we know that everything is set up, we just don't
6698 // know whether we have a icmp_ne or icmp_eq and whether the
6699 // true or false val is the zero.
6700 bool ShouldNotVal = !TrueValC->isNullValue();
6701 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6704 V = InsertNewInstBefore(BinaryOperator::create(
6705 Instruction::Xor, V, ICA->getOperand(1)), SI);
6706 return ReplaceInstUsesWith(SI, V);
6711 // See if we are selecting two values based on a comparison of the two values.
6712 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6713 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6714 // Transform (X == Y) ? X : Y -> Y
6715 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6716 return ReplaceInstUsesWith(SI, FalseVal);
6717 // Transform (X != Y) ? X : Y -> X
6718 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6719 return ReplaceInstUsesWith(SI, TrueVal);
6720 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6722 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6723 // Transform (X == Y) ? Y : X -> X
6724 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6725 return ReplaceInstUsesWith(SI, FalseVal);
6726 // Transform (X != Y) ? Y : X -> Y
6727 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6728 return ReplaceInstUsesWith(SI, TrueVal);
6729 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6733 // See if we are selecting two values based on a comparison of the two values.
6734 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6735 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6736 // Transform (X == Y) ? X : Y -> Y
6737 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6738 return ReplaceInstUsesWith(SI, FalseVal);
6739 // Transform (X != Y) ? X : Y -> X
6740 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6741 return ReplaceInstUsesWith(SI, TrueVal);
6742 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6744 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6745 // Transform (X == Y) ? Y : X -> X
6746 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6747 return ReplaceInstUsesWith(SI, FalseVal);
6748 // Transform (X != Y) ? Y : X -> Y
6749 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6750 return ReplaceInstUsesWith(SI, TrueVal);
6751 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6755 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6756 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6757 if (TI->hasOneUse() && FI->hasOneUse()) {
6758 Instruction *AddOp = 0, *SubOp = 0;
6760 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6761 if (TI->getOpcode() == FI->getOpcode())
6762 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6765 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6766 // even legal for FP.
6767 if (TI->getOpcode() == Instruction::Sub &&
6768 FI->getOpcode() == Instruction::Add) {
6769 AddOp = FI; SubOp = TI;
6770 } else if (FI->getOpcode() == Instruction::Sub &&
6771 TI->getOpcode() == Instruction::Add) {
6772 AddOp = TI; SubOp = FI;
6776 Value *OtherAddOp = 0;
6777 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6778 OtherAddOp = AddOp->getOperand(1);
6779 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6780 OtherAddOp = AddOp->getOperand(0);
6784 // So at this point we know we have (Y -> OtherAddOp):
6785 // select C, (add X, Y), (sub X, Z)
6786 Value *NegVal; // Compute -Z
6787 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6788 NegVal = ConstantExpr::getNeg(C);
6790 NegVal = InsertNewInstBefore(
6791 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6794 Value *NewTrueOp = OtherAddOp;
6795 Value *NewFalseOp = NegVal;
6797 std::swap(NewTrueOp, NewFalseOp);
6798 Instruction *NewSel =
6799 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6801 NewSel = InsertNewInstBefore(NewSel, SI);
6802 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6807 // See if we can fold the select into one of our operands.
6808 if (SI.getType()->isInteger()) {
6809 // See the comment above GetSelectFoldableOperands for a description of the
6810 // transformation we are doing here.
6811 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6812 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6813 !isa<Constant>(FalseVal))
6814 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6815 unsigned OpToFold = 0;
6816 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6818 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6823 Constant *C = GetSelectFoldableConstant(TVI);
6824 std::string Name = TVI->getName(); TVI->setName("");
6825 Instruction *NewSel =
6826 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6828 InsertNewInstBefore(NewSel, SI);
6829 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6830 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6831 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6832 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6834 assert(0 && "Unknown instruction!!");
6839 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6840 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6841 !isa<Constant>(TrueVal))
6842 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6843 unsigned OpToFold = 0;
6844 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6846 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6851 Constant *C = GetSelectFoldableConstant(FVI);
6852 std::string Name = FVI->getName(); FVI->setName("");
6853 Instruction *NewSel =
6854 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6856 InsertNewInstBefore(NewSel, SI);
6857 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6858 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6859 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6860 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6862 assert(0 && "Unknown instruction!!");
6868 if (BinaryOperator::isNot(CondVal)) {
6869 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6870 SI.setOperand(1, FalseVal);
6871 SI.setOperand(2, TrueVal);
6878 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6879 /// determine, return it, otherwise return 0.
6880 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6881 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6882 unsigned Align = GV->getAlignment();
6883 if (Align == 0 && TD)
6884 Align = TD->getTypeAlignmentPref(GV->getType()->getElementType());
6886 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6887 unsigned Align = AI->getAlignment();
6888 if (Align == 0 && TD) {
6889 if (isa<AllocaInst>(AI))
6890 Align = TD->getTypeAlignmentPref(AI->getType()->getElementType());
6891 else if (isa<MallocInst>(AI)) {
6892 // Malloc returns maximally aligned memory.
6893 Align = TD->getTypeAlignmentABI(AI->getType()->getElementType());
6896 (unsigned)TD->getTypeAlignmentABI(Type::DoubleTy));
6899 (unsigned)TD->getTypeAlignmentABI(Type::Int64Ty));
6903 } else if (isa<BitCastInst>(V) ||
6904 (isa<ConstantExpr>(V) &&
6905 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6906 User *CI = cast<User>(V);
6907 if (isa<PointerType>(CI->getOperand(0)->getType()))
6908 return GetKnownAlignment(CI->getOperand(0), TD);
6910 } else if (isa<GetElementPtrInst>(V) ||
6911 (isa<ConstantExpr>(V) &&
6912 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6913 User *GEPI = cast<User>(V);
6914 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6915 if (BaseAlignment == 0) return 0;
6917 // If all indexes are zero, it is just the alignment of the base pointer.
6918 bool AllZeroOperands = true;
6919 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6920 if (!isa<Constant>(GEPI->getOperand(i)) ||
6921 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6922 AllZeroOperands = false;
6925 if (AllZeroOperands)
6926 return BaseAlignment;
6928 // Otherwise, if the base alignment is >= the alignment we expect for the
6929 // base pointer type, then we know that the resultant pointer is aligned at
6930 // least as much as its type requires.
6933 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6934 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
6935 if (TD->getTypeAlignmentABI(PtrTy->getElementType())
6937 const Type *GEPTy = GEPI->getType();
6938 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
6939 return TD->getTypeAlignmentABI(GEPPtrTy->getElementType());
6947 /// visitCallInst - CallInst simplification. This mostly only handles folding
6948 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6949 /// the heavy lifting.
6951 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6952 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6953 if (!II) return visitCallSite(&CI);
6955 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6957 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6958 bool Changed = false;
6960 // memmove/cpy/set of zero bytes is a noop.
6961 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6962 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6964 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6965 if (CI->getZExtValue() == 1) {
6966 // Replace the instruction with just byte operations. We would
6967 // transform other cases to loads/stores, but we don't know if
6968 // alignment is sufficient.
6972 // If we have a memmove and the source operation is a constant global,
6973 // then the source and dest pointers can't alias, so we can change this
6974 // into a call to memcpy.
6975 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6976 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6977 if (GVSrc->isConstant()) {
6978 Module *M = CI.getParent()->getParent()->getParent();
6980 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
6982 Name = "llvm.memcpy.i32";
6984 Name = "llvm.memcpy.i64";
6985 Constant *MemCpy = M->getOrInsertFunction(Name,
6986 CI.getCalledFunction()->getFunctionType());
6987 CI.setOperand(0, MemCpy);
6992 // If we can determine a pointer alignment that is bigger than currently
6993 // set, update the alignment.
6994 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6995 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6996 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6997 unsigned Align = std::min(Alignment1, Alignment2);
6998 if (MI->getAlignment()->getZExtValue() < Align) {
6999 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7002 } else if (isa<MemSetInst>(MI)) {
7003 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7004 if (MI->getAlignment()->getZExtValue() < Alignment) {
7005 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7010 if (Changed) return II;
7012 switch (II->getIntrinsicID()) {
7014 case Intrinsic::ppc_altivec_lvx:
7015 case Intrinsic::ppc_altivec_lvxl:
7016 case Intrinsic::x86_sse_loadu_ps:
7017 case Intrinsic::x86_sse2_loadu_pd:
7018 case Intrinsic::x86_sse2_loadu_dq:
7019 // Turn PPC lvx -> load if the pointer is known aligned.
7020 // Turn X86 loadups -> load if the pointer is known aligned.
7021 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7022 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7023 PointerType::get(II->getType()), CI);
7024 return new LoadInst(Ptr);
7027 case Intrinsic::ppc_altivec_stvx:
7028 case Intrinsic::ppc_altivec_stvxl:
7029 // Turn stvx -> store if the pointer is known aligned.
7030 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7031 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7032 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7034 return new StoreInst(II->getOperand(1), Ptr);
7037 case Intrinsic::x86_sse_storeu_ps:
7038 case Intrinsic::x86_sse2_storeu_pd:
7039 case Intrinsic::x86_sse2_storeu_dq:
7040 case Intrinsic::x86_sse2_storel_dq:
7041 // Turn X86 storeu -> store if the pointer is known aligned.
7042 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7043 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7044 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7046 return new StoreInst(II->getOperand(2), Ptr);
7050 case Intrinsic::x86_sse_cvttss2si: {
7051 // These intrinsics only demands the 0th element of its input vector. If
7052 // we can simplify the input based on that, do so now.
7054 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7056 II->setOperand(1, V);
7062 case Intrinsic::ppc_altivec_vperm:
7063 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7064 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
7065 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7067 // Check that all of the elements are integer constants or undefs.
7068 bool AllEltsOk = true;
7069 for (unsigned i = 0; i != 16; ++i) {
7070 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7071 !isa<UndefValue>(Mask->getOperand(i))) {
7078 // Cast the input vectors to byte vectors.
7079 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7080 II->getOperand(1), Mask->getType(), CI);
7081 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7082 II->getOperand(2), Mask->getType(), CI);
7083 Value *Result = UndefValue::get(Op0->getType());
7085 // Only extract each element once.
7086 Value *ExtractedElts[32];
7087 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7089 for (unsigned i = 0; i != 16; ++i) {
7090 if (isa<UndefValue>(Mask->getOperand(i)))
7092 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7093 Idx &= 31; // Match the hardware behavior.
7095 if (ExtractedElts[Idx] == 0) {
7097 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7098 InsertNewInstBefore(Elt, CI);
7099 ExtractedElts[Idx] = Elt;
7102 // Insert this value into the result vector.
7103 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7104 InsertNewInstBefore(cast<Instruction>(Result), CI);
7106 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7111 case Intrinsic::stackrestore: {
7112 // If the save is right next to the restore, remove the restore. This can
7113 // happen when variable allocas are DCE'd.
7114 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7115 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7116 BasicBlock::iterator BI = SS;
7118 return EraseInstFromFunction(CI);
7122 // If the stack restore is in a return/unwind block and if there are no
7123 // allocas or calls between the restore and the return, nuke the restore.
7124 TerminatorInst *TI = II->getParent()->getTerminator();
7125 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7126 BasicBlock::iterator BI = II;
7127 bool CannotRemove = false;
7128 for (++BI; &*BI != TI; ++BI) {
7129 if (isa<AllocaInst>(BI) ||
7130 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7131 CannotRemove = true;
7136 return EraseInstFromFunction(CI);
7143 return visitCallSite(II);
7146 // InvokeInst simplification
7148 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7149 return visitCallSite(&II);
7152 // visitCallSite - Improvements for call and invoke instructions.
7154 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7155 bool Changed = false;
7157 // If the callee is a constexpr cast of a function, attempt to move the cast
7158 // to the arguments of the call/invoke.
7159 if (transformConstExprCastCall(CS)) return 0;
7161 Value *Callee = CS.getCalledValue();
7163 if (Function *CalleeF = dyn_cast<Function>(Callee))
7164 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7165 Instruction *OldCall = CS.getInstruction();
7166 // If the call and callee calling conventions don't match, this call must
7167 // be unreachable, as the call is undefined.
7168 new StoreInst(ConstantInt::getTrue(),
7169 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7170 if (!OldCall->use_empty())
7171 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7172 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7173 return EraseInstFromFunction(*OldCall);
7177 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7178 // This instruction is not reachable, just remove it. We insert a store to
7179 // undef so that we know that this code is not reachable, despite the fact
7180 // that we can't modify the CFG here.
7181 new StoreInst(ConstantInt::getTrue(),
7182 UndefValue::get(PointerType::get(Type::Int1Ty)),
7183 CS.getInstruction());
7185 if (!CS.getInstruction()->use_empty())
7186 CS.getInstruction()->
7187 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7189 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7190 // Don't break the CFG, insert a dummy cond branch.
7191 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7192 ConstantInt::getTrue(), II);
7194 return EraseInstFromFunction(*CS.getInstruction());
7197 const PointerType *PTy = cast<PointerType>(Callee->getType());
7198 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7199 if (FTy->isVarArg()) {
7200 // See if we can optimize any arguments passed through the varargs area of
7202 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7203 E = CS.arg_end(); I != E; ++I)
7204 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7205 // If this cast does not effect the value passed through the varargs
7206 // area, we can eliminate the use of the cast.
7207 Value *Op = CI->getOperand(0);
7208 if (CI->isLosslessCast()) {
7215 return Changed ? CS.getInstruction() : 0;
7218 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7219 // attempt to move the cast to the arguments of the call/invoke.
7221 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7222 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7223 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7224 if (CE->getOpcode() != Instruction::BitCast ||
7225 !isa<Function>(CE->getOperand(0)))
7227 Function *Callee = cast<Function>(CE->getOperand(0));
7228 Instruction *Caller = CS.getInstruction();
7230 // Okay, this is a cast from a function to a different type. Unless doing so
7231 // would cause a type conversion of one of our arguments, change this call to
7232 // be a direct call with arguments casted to the appropriate types.
7234 const FunctionType *FT = Callee->getFunctionType();
7235 const Type *OldRetTy = Caller->getType();
7237 // Check to see if we are changing the return type...
7238 if (OldRetTy != FT->getReturnType()) {
7239 if (Callee->isDeclaration() && !Caller->use_empty() &&
7240 OldRetTy != FT->getReturnType() &&
7241 // Conversion is ok if changing from pointer to int of same size.
7242 !(isa<PointerType>(FT->getReturnType()) &&
7243 TD->getIntPtrType() == OldRetTy))
7244 return false; // Cannot transform this return value.
7246 // If the callsite is an invoke instruction, and the return value is used by
7247 // a PHI node in a successor, we cannot change the return type of the call
7248 // because there is no place to put the cast instruction (without breaking
7249 // the critical edge). Bail out in this case.
7250 if (!Caller->use_empty())
7251 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7252 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7254 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7255 if (PN->getParent() == II->getNormalDest() ||
7256 PN->getParent() == II->getUnwindDest())
7260 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7261 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7263 CallSite::arg_iterator AI = CS.arg_begin();
7264 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7265 const Type *ParamTy = FT->getParamType(i);
7266 const Type *ActTy = (*AI)->getType();
7267 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7268 //Either we can cast directly, or we can upconvert the argument
7269 bool isConvertible = ActTy == ParamTy ||
7270 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7271 (ParamTy->isInteger() && ActTy->isInteger() &&
7272 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7273 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7274 && c->getSExtValue() > 0);
7275 if (Callee->isDeclaration() && !isConvertible) return false;
7278 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7279 Callee->isDeclaration())
7280 return false; // Do not delete arguments unless we have a function body...
7282 // Okay, we decided that this is a safe thing to do: go ahead and start
7283 // inserting cast instructions as necessary...
7284 std::vector<Value*> Args;
7285 Args.reserve(NumActualArgs);
7287 AI = CS.arg_begin();
7288 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7289 const Type *ParamTy = FT->getParamType(i);
7290 if ((*AI)->getType() == ParamTy) {
7291 Args.push_back(*AI);
7293 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7294 false, ParamTy, false);
7295 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7296 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7300 // If the function takes more arguments than the call was taking, add them
7302 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7303 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7305 // If we are removing arguments to the function, emit an obnoxious warning...
7306 if (FT->getNumParams() < NumActualArgs)
7307 if (!FT->isVarArg()) {
7308 cerr << "WARNING: While resolving call to function '"
7309 << Callee->getName() << "' arguments were dropped!\n";
7311 // Add all of the arguments in their promoted form to the arg list...
7312 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7313 const Type *PTy = getPromotedType((*AI)->getType());
7314 if (PTy != (*AI)->getType()) {
7315 // Must promote to pass through va_arg area!
7316 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7318 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7319 InsertNewInstBefore(Cast, *Caller);
7320 Args.push_back(Cast);
7322 Args.push_back(*AI);
7327 if (FT->getReturnType() == Type::VoidTy)
7328 Caller->setName(""); // Void type should not have a name...
7331 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7332 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7333 Args, Caller->getName(), Caller);
7334 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7336 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7337 if (cast<CallInst>(Caller)->isTailCall())
7338 cast<CallInst>(NC)->setTailCall();
7339 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7342 // Insert a cast of the return type as necessary...
7344 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7345 if (NV->getType() != Type::VoidTy) {
7346 const Type *CallerTy = Caller->getType();
7347 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7349 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7351 // If this is an invoke instruction, we should insert it after the first
7352 // non-phi, instruction in the normal successor block.
7353 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7354 BasicBlock::iterator I = II->getNormalDest()->begin();
7355 while (isa<PHINode>(I)) ++I;
7356 InsertNewInstBefore(NC, *I);
7358 // Otherwise, it's a call, just insert cast right after the call instr
7359 InsertNewInstBefore(NC, *Caller);
7361 AddUsersToWorkList(*Caller);
7363 NV = UndefValue::get(Caller->getType());
7367 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7368 Caller->replaceAllUsesWith(NV);
7369 Caller->getParent()->getInstList().erase(Caller);
7370 removeFromWorkList(Caller);
7374 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7375 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7376 /// and a single binop.
7377 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7378 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7379 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7380 isa<GetElementPtrInst>(FirstInst) || isa<CmpInst>(FirstInst));
7381 unsigned Opc = FirstInst->getOpcode();
7382 Value *LHSVal = FirstInst->getOperand(0);
7383 Value *RHSVal = FirstInst->getOperand(1);
7385 const Type *LHSType = LHSVal->getType();
7386 const Type *RHSType = RHSVal->getType();
7388 // Scan to see if all operands are the same opcode, all have one use, and all
7389 // kill their operands (i.e. the operands have one use).
7390 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7391 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7392 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7393 // Verify type of the LHS matches so we don't fold cmp's of different
7394 // types or GEP's with different index types.
7395 I->getOperand(0)->getType() != LHSType ||
7396 I->getOperand(1)->getType() != RHSType)
7399 // If they are CmpInst instructions, check their predicates
7400 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7401 if (cast<CmpInst>(I)->getPredicate() !=
7402 cast<CmpInst>(FirstInst)->getPredicate())
7405 // Keep track of which operand needs a phi node.
7406 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7407 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7410 // Otherwise, this is safe to transform, determine if it is profitable.
7412 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7413 // Indexes are often folded into load/store instructions, so we don't want to
7414 // hide them behind a phi.
7415 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7418 Value *InLHS = FirstInst->getOperand(0);
7419 Value *InRHS = FirstInst->getOperand(1);
7420 PHINode *NewLHS = 0, *NewRHS = 0;
7422 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7423 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7424 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7425 InsertNewInstBefore(NewLHS, PN);
7430 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7431 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7432 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7433 InsertNewInstBefore(NewRHS, PN);
7437 // Add all operands to the new PHIs.
7438 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7440 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7441 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7444 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7445 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7449 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7450 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7451 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7452 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7454 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7455 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7457 assert(isa<GetElementPtrInst>(FirstInst));
7458 return new GetElementPtrInst(LHSVal, RHSVal);
7462 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7463 /// of the block that defines it. This means that it must be obvious the value
7464 /// of the load is not changed from the point of the load to the end of the
7466 static bool isSafeToSinkLoad(LoadInst *L) {
7467 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7469 for (++BBI; BBI != E; ++BBI)
7470 if (BBI->mayWriteToMemory())
7476 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7477 // operator and they all are only used by the PHI, PHI together their
7478 // inputs, and do the operation once, to the result of the PHI.
7479 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7480 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7482 // Scan the instruction, looking for input operations that can be folded away.
7483 // If all input operands to the phi are the same instruction (e.g. a cast from
7484 // the same type or "+42") we can pull the operation through the PHI, reducing
7485 // code size and simplifying code.
7486 Constant *ConstantOp = 0;
7487 const Type *CastSrcTy = 0;
7488 bool isVolatile = false;
7489 if (isa<CastInst>(FirstInst)) {
7490 CastSrcTy = FirstInst->getOperand(0)->getType();
7491 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7492 isa<CmpInst>(FirstInst)) {
7493 // Can fold binop, compare or shift here if the RHS is a constant,
7494 // otherwise call FoldPHIArgBinOpIntoPHI.
7495 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7496 if (ConstantOp == 0)
7497 return FoldPHIArgBinOpIntoPHI(PN);
7498 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7499 isVolatile = LI->isVolatile();
7500 // We can't sink the load if the loaded value could be modified between the
7501 // load and the PHI.
7502 if (LI->getParent() != PN.getIncomingBlock(0) ||
7503 !isSafeToSinkLoad(LI))
7505 } else if (isa<GetElementPtrInst>(FirstInst)) {
7506 if (FirstInst->getNumOperands() == 2)
7507 return FoldPHIArgBinOpIntoPHI(PN);
7508 // Can't handle general GEPs yet.
7511 return 0; // Cannot fold this operation.
7514 // Check to see if all arguments are the same operation.
7515 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7516 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7517 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7518 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7521 if (I->getOperand(0)->getType() != CastSrcTy)
7522 return 0; // Cast operation must match.
7523 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7524 // We can't sink the load if the loaded value could be modified between
7525 // the load and the PHI.
7526 if (LI->isVolatile() != isVolatile ||
7527 LI->getParent() != PN.getIncomingBlock(i) ||
7528 !isSafeToSinkLoad(LI))
7530 } else if (I->getOperand(1) != ConstantOp) {
7535 // Okay, they are all the same operation. Create a new PHI node of the
7536 // correct type, and PHI together all of the LHS's of the instructions.
7537 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7538 PN.getName()+".in");
7539 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7541 Value *InVal = FirstInst->getOperand(0);
7542 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7544 // Add all operands to the new PHI.
7545 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7546 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7547 if (NewInVal != InVal)
7549 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7554 // The new PHI unions all of the same values together. This is really
7555 // common, so we handle it intelligently here for compile-time speed.
7559 InsertNewInstBefore(NewPN, PN);
7563 // Insert and return the new operation.
7564 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7565 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7566 else if (isa<LoadInst>(FirstInst))
7567 return new LoadInst(PhiVal, "", isVolatile);
7568 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7569 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7570 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7571 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7572 PhiVal, ConstantOp);
7574 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7575 PhiVal, ConstantOp);
7578 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7580 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7581 if (PN->use_empty()) return true;
7582 if (!PN->hasOneUse()) return false;
7584 // Remember this node, and if we find the cycle, return.
7585 if (!PotentiallyDeadPHIs.insert(PN).second)
7588 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7589 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7594 // PHINode simplification
7596 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7597 // If LCSSA is around, don't mess with Phi nodes
7598 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7600 if (Value *V = PN.hasConstantValue())
7601 return ReplaceInstUsesWith(PN, V);
7603 // If all PHI operands are the same operation, pull them through the PHI,
7604 // reducing code size.
7605 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7606 PN.getIncomingValue(0)->hasOneUse())
7607 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7610 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7611 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7612 // PHI)... break the cycle.
7613 if (PN.hasOneUse()) {
7614 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7615 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7616 std::set<PHINode*> PotentiallyDeadPHIs;
7617 PotentiallyDeadPHIs.insert(&PN);
7618 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7619 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7622 // If this phi has a single use, and if that use just computes a value for
7623 // the next iteration of a loop, delete the phi. This occurs with unused
7624 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7625 // common case here is good because the only other things that catch this
7626 // are induction variable analysis (sometimes) and ADCE, which is only run
7628 if (PHIUser->hasOneUse() &&
7629 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7630 PHIUser->use_back() == &PN) {
7631 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7638 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7639 Instruction *InsertPoint,
7641 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7642 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7643 // We must cast correctly to the pointer type. Ensure that we
7644 // sign extend the integer value if it is smaller as this is
7645 // used for address computation.
7646 Instruction::CastOps opcode =
7647 (VTySize < PtrSize ? Instruction::SExt :
7648 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7649 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7653 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7654 Value *PtrOp = GEP.getOperand(0);
7655 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7656 // If so, eliminate the noop.
7657 if (GEP.getNumOperands() == 1)
7658 return ReplaceInstUsesWith(GEP, PtrOp);
7660 if (isa<UndefValue>(GEP.getOperand(0)))
7661 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7663 bool HasZeroPointerIndex = false;
7664 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7665 HasZeroPointerIndex = C->isNullValue();
7667 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7668 return ReplaceInstUsesWith(GEP, PtrOp);
7670 // Eliminate unneeded casts for indices.
7671 bool MadeChange = false;
7672 gep_type_iterator GTI = gep_type_begin(GEP);
7673 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7674 if (isa<SequentialType>(*GTI)) {
7675 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7676 if (CI->getOpcode() == Instruction::ZExt ||
7677 CI->getOpcode() == Instruction::SExt) {
7678 const Type *SrcTy = CI->getOperand(0)->getType();
7679 // We can eliminate a cast from i32 to i64 iff the target
7680 // is a 32-bit pointer target.
7681 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7683 GEP.setOperand(i, CI->getOperand(0));
7687 // If we are using a wider index than needed for this platform, shrink it
7688 // to what we need. If the incoming value needs a cast instruction,
7689 // insert it. This explicit cast can make subsequent optimizations more
7691 Value *Op = GEP.getOperand(i);
7692 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7693 if (Constant *C = dyn_cast<Constant>(Op)) {
7694 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7697 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7699 GEP.setOperand(i, Op);
7703 if (MadeChange) return &GEP;
7705 // Combine Indices - If the source pointer to this getelementptr instruction
7706 // is a getelementptr instruction, combine the indices of the two
7707 // getelementptr instructions into a single instruction.
7709 std::vector<Value*> SrcGEPOperands;
7710 if (User *Src = dyn_castGetElementPtr(PtrOp))
7711 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7713 if (!SrcGEPOperands.empty()) {
7714 // Note that if our source is a gep chain itself that we wait for that
7715 // chain to be resolved before we perform this transformation. This
7716 // avoids us creating a TON of code in some cases.
7718 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7719 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7720 return 0; // Wait until our source is folded to completion.
7722 std::vector<Value *> Indices;
7724 // Find out whether the last index in the source GEP is a sequential idx.
7725 bool EndsWithSequential = false;
7726 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7727 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7728 EndsWithSequential = !isa<StructType>(*I);
7730 // Can we combine the two pointer arithmetics offsets?
7731 if (EndsWithSequential) {
7732 // Replace: gep (gep %P, long B), long A, ...
7733 // With: T = long A+B; gep %P, T, ...
7735 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7736 if (SO1 == Constant::getNullValue(SO1->getType())) {
7738 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7741 // If they aren't the same type, convert both to an integer of the
7742 // target's pointer size.
7743 if (SO1->getType() != GO1->getType()) {
7744 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7745 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7746 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7747 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7749 unsigned PS = TD->getPointerSize();
7750 if (TD->getTypeSize(SO1->getType()) == PS) {
7751 // Convert GO1 to SO1's type.
7752 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7754 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7755 // Convert SO1 to GO1's type.
7756 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7758 const Type *PT = TD->getIntPtrType();
7759 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7760 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7764 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7765 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7767 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7768 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7772 // Recycle the GEP we already have if possible.
7773 if (SrcGEPOperands.size() == 2) {
7774 GEP.setOperand(0, SrcGEPOperands[0]);
7775 GEP.setOperand(1, Sum);
7778 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7779 SrcGEPOperands.end()-1);
7780 Indices.push_back(Sum);
7781 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7783 } else if (isa<Constant>(*GEP.idx_begin()) &&
7784 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7785 SrcGEPOperands.size() != 1) {
7786 // Otherwise we can do the fold if the first index of the GEP is a zero
7787 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7788 SrcGEPOperands.end());
7789 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7792 if (!Indices.empty())
7793 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7795 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7796 // GEP of global variable. If all of the indices for this GEP are
7797 // constants, we can promote this to a constexpr instead of an instruction.
7799 // Scan for nonconstants...
7800 std::vector<Constant*> Indices;
7801 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7802 for (; I != E && isa<Constant>(*I); ++I)
7803 Indices.push_back(cast<Constant>(*I));
7805 if (I == E) { // If they are all constants...
7806 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7808 // Replace all uses of the GEP with the new constexpr...
7809 return ReplaceInstUsesWith(GEP, CE);
7811 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7812 if (!isa<PointerType>(X->getType())) {
7813 // Not interesting. Source pointer must be a cast from pointer.
7814 } else if (HasZeroPointerIndex) {
7815 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7816 // into : GEP [10 x ubyte]* X, long 0, ...
7818 // This occurs when the program declares an array extern like "int X[];"
7820 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7821 const PointerType *XTy = cast<PointerType>(X->getType());
7822 if (const ArrayType *XATy =
7823 dyn_cast<ArrayType>(XTy->getElementType()))
7824 if (const ArrayType *CATy =
7825 dyn_cast<ArrayType>(CPTy->getElementType()))
7826 if (CATy->getElementType() == XATy->getElementType()) {
7827 // At this point, we know that the cast source type is a pointer
7828 // to an array of the same type as the destination pointer
7829 // array. Because the array type is never stepped over (there
7830 // is a leading zero) we can fold the cast into this GEP.
7831 GEP.setOperand(0, X);
7834 } else if (GEP.getNumOperands() == 2) {
7835 // Transform things like:
7836 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7837 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7838 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7839 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7840 if (isa<ArrayType>(SrcElTy) &&
7841 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7842 TD->getTypeSize(ResElTy)) {
7843 Value *V = InsertNewInstBefore(
7844 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7845 GEP.getOperand(1), GEP.getName()), GEP);
7846 // V and GEP are both pointer types --> BitCast
7847 return new BitCastInst(V, GEP.getType());
7850 // Transform things like:
7851 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7852 // (where tmp = 8*tmp2) into:
7853 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7855 if (isa<ArrayType>(SrcElTy) &&
7856 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7857 uint64_t ArrayEltSize =
7858 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7860 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7861 // allow either a mul, shift, or constant here.
7863 ConstantInt *Scale = 0;
7864 if (ArrayEltSize == 1) {
7865 NewIdx = GEP.getOperand(1);
7866 Scale = ConstantInt::get(NewIdx->getType(), 1);
7867 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7868 NewIdx = ConstantInt::get(CI->getType(), 1);
7870 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7871 if (Inst->getOpcode() == Instruction::Shl &&
7872 isa<ConstantInt>(Inst->getOperand(1))) {
7874 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7875 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7876 NewIdx = Inst->getOperand(0);
7877 } else if (Inst->getOpcode() == Instruction::Mul &&
7878 isa<ConstantInt>(Inst->getOperand(1))) {
7879 Scale = cast<ConstantInt>(Inst->getOperand(1));
7880 NewIdx = Inst->getOperand(0);
7884 // If the index will be to exactly the right offset with the scale taken
7885 // out, perform the transformation.
7886 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7887 if (isa<ConstantInt>(Scale))
7888 Scale = ConstantInt::get(Scale->getType(),
7889 Scale->getZExtValue() / ArrayEltSize);
7890 if (Scale->getZExtValue() != 1) {
7891 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7893 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7894 NewIdx = InsertNewInstBefore(Sc, GEP);
7897 // Insert the new GEP instruction.
7898 Instruction *NewGEP =
7899 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7900 NewIdx, GEP.getName());
7901 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7902 // The NewGEP must be pointer typed, so must the old one -> BitCast
7903 return new BitCastInst(NewGEP, GEP.getType());
7912 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7913 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7914 if (AI.isArrayAllocation()) // Check C != 1
7915 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7917 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7918 AllocationInst *New = 0;
7920 // Create and insert the replacement instruction...
7921 if (isa<MallocInst>(AI))
7922 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7924 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7925 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7928 InsertNewInstBefore(New, AI);
7930 // Scan to the end of the allocation instructions, to skip over a block of
7931 // allocas if possible...
7933 BasicBlock::iterator It = New;
7934 while (isa<AllocationInst>(*It)) ++It;
7936 // Now that I is pointing to the first non-allocation-inst in the block,
7937 // insert our getelementptr instruction...
7939 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
7940 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7941 New->getName()+".sub", It);
7943 // Now make everything use the getelementptr instead of the original
7945 return ReplaceInstUsesWith(AI, V);
7946 } else if (isa<UndefValue>(AI.getArraySize())) {
7947 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7950 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7951 // Note that we only do this for alloca's, because malloc should allocate and
7952 // return a unique pointer, even for a zero byte allocation.
7953 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7954 TD->getTypeSize(AI.getAllocatedType()) == 0)
7955 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7960 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7961 Value *Op = FI.getOperand(0);
7963 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7964 if (CastInst *CI = dyn_cast<CastInst>(Op))
7965 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7966 FI.setOperand(0, CI->getOperand(0));
7970 // free undef -> unreachable.
7971 if (isa<UndefValue>(Op)) {
7972 // Insert a new store to null because we cannot modify the CFG here.
7973 new StoreInst(ConstantInt::getTrue(),
7974 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
7975 return EraseInstFromFunction(FI);
7978 // If we have 'free null' delete the instruction. This can happen in stl code
7979 // when lots of inlining happens.
7980 if (isa<ConstantPointerNull>(Op))
7981 return EraseInstFromFunction(FI);
7987 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7988 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7989 User *CI = cast<User>(LI.getOperand(0));
7990 Value *CastOp = CI->getOperand(0);
7992 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7993 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7994 const Type *SrcPTy = SrcTy->getElementType();
7996 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7997 isa<PackedType>(DestPTy)) {
7998 // If the source is an array, the code below will not succeed. Check to
7999 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8001 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8002 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8003 if (ASrcTy->getNumElements() != 0) {
8004 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
8005 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8006 SrcTy = cast<PointerType>(CastOp->getType());
8007 SrcPTy = SrcTy->getElementType();
8010 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8011 isa<PackedType>(SrcPTy)) &&
8012 // Do not allow turning this into a load of an integer, which is then
8013 // casted to a pointer, this pessimizes pointer analysis a lot.
8014 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8015 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8016 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8018 // Okay, we are casting from one integer or pointer type to another of
8019 // the same size. Instead of casting the pointer before the load, cast
8020 // the result of the loaded value.
8021 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8023 LI.isVolatile()),LI);
8024 // Now cast the result of the load.
8025 return new BitCastInst(NewLoad, LI.getType());
8032 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8033 /// from this value cannot trap. If it is not obviously safe to load from the
8034 /// specified pointer, we do a quick local scan of the basic block containing
8035 /// ScanFrom, to determine if the address is already accessed.
8036 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8037 // If it is an alloca or global variable, it is always safe to load from.
8038 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8040 // Otherwise, be a little bit agressive by scanning the local block where we
8041 // want to check to see if the pointer is already being loaded or stored
8042 // from/to. If so, the previous load or store would have already trapped,
8043 // so there is no harm doing an extra load (also, CSE will later eliminate
8044 // the load entirely).
8045 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8050 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8051 if (LI->getOperand(0) == V) return true;
8052 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8053 if (SI->getOperand(1) == V) return true;
8059 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8060 Value *Op = LI.getOperand(0);
8062 // load (cast X) --> cast (load X) iff safe
8063 if (isa<CastInst>(Op))
8064 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8067 // None of the following transforms are legal for volatile loads.
8068 if (LI.isVolatile()) return 0;
8070 if (&LI.getParent()->front() != &LI) {
8071 BasicBlock::iterator BBI = &LI; --BBI;
8072 // If the instruction immediately before this is a store to the same
8073 // address, do a simple form of store->load forwarding.
8074 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8075 if (SI->getOperand(1) == LI.getOperand(0))
8076 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8077 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8078 if (LIB->getOperand(0) == LI.getOperand(0))
8079 return ReplaceInstUsesWith(LI, LIB);
8082 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8083 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8084 isa<UndefValue>(GEPI->getOperand(0))) {
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 if (Constant *C = dyn_cast<Constant>(Op)) {
8095 // load null/undef -> undef
8096 if ((C->isNullValue() || isa<UndefValue>(C))) {
8097 // Insert a new store to null instruction before the load to indicate that
8098 // this code is not reachable. We do this instead of inserting an
8099 // unreachable instruction directly because we cannot modify the CFG.
8100 new StoreInst(UndefValue::get(LI.getType()),
8101 Constant::getNullValue(Op->getType()), &LI);
8102 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8105 // Instcombine load (constant global) into the value loaded.
8106 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8107 if (GV->isConstant() && !GV->isDeclaration())
8108 return ReplaceInstUsesWith(LI, GV->getInitializer());
8110 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8111 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8112 if (CE->getOpcode() == Instruction::GetElementPtr) {
8113 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8114 if (GV->isConstant() && !GV->isDeclaration())
8116 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8117 return ReplaceInstUsesWith(LI, V);
8118 if (CE->getOperand(0)->isNullValue()) {
8119 // Insert a new store to null instruction before the load to indicate
8120 // that this code is not reachable. We do this instead of inserting
8121 // an unreachable instruction directly because we cannot modify the
8123 new StoreInst(UndefValue::get(LI.getType()),
8124 Constant::getNullValue(Op->getType()), &LI);
8125 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8128 } else if (CE->isCast()) {
8129 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8134 if (Op->hasOneUse()) {
8135 // Change select and PHI nodes to select values instead of addresses: this
8136 // helps alias analysis out a lot, allows many others simplifications, and
8137 // exposes redundancy in the code.
8139 // Note that we cannot do the transformation unless we know that the
8140 // introduced loads cannot trap! Something like this is valid as long as
8141 // the condition is always false: load (select bool %C, int* null, int* %G),
8142 // but it would not be valid if we transformed it to load from null
8145 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8146 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8147 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8148 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8149 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8150 SI->getOperand(1)->getName()+".val"), LI);
8151 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8152 SI->getOperand(2)->getName()+".val"), LI);
8153 return new SelectInst(SI->getCondition(), V1, V2);
8156 // load (select (cond, null, P)) -> load P
8157 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8158 if (C->isNullValue()) {
8159 LI.setOperand(0, SI->getOperand(2));
8163 // load (select (cond, P, null)) -> load P
8164 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8165 if (C->isNullValue()) {
8166 LI.setOperand(0, SI->getOperand(1));
8174 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8176 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8177 User *CI = cast<User>(SI.getOperand(1));
8178 Value *CastOp = CI->getOperand(0);
8180 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8181 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8182 const Type *SrcPTy = SrcTy->getElementType();
8184 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8185 // If the source is an array, the code below will not succeed. Check to
8186 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8188 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8189 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8190 if (ASrcTy->getNumElements() != 0) {
8191 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
8192 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8193 SrcTy = cast<PointerType>(CastOp->getType());
8194 SrcPTy = SrcTy->getElementType();
8197 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8198 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8199 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8201 // Okay, we are casting from one integer or pointer type to another of
8202 // the same size. Instead of casting the pointer before
8203 // the store, cast the value to be stored.
8205 Value *SIOp0 = SI.getOperand(0);
8206 Instruction::CastOps opcode = Instruction::BitCast;
8207 const Type* CastSrcTy = SIOp0->getType();
8208 const Type* CastDstTy = SrcPTy;
8209 if (isa<PointerType>(CastDstTy)) {
8210 if (CastSrcTy->isInteger())
8211 opcode = Instruction::IntToPtr;
8212 } else if (isa<IntegerType>(CastDstTy)) {
8213 if (isa<PointerType>(SIOp0->getType()))
8214 opcode = Instruction::PtrToInt;
8216 if (Constant *C = dyn_cast<Constant>(SIOp0))
8217 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8219 NewCast = IC.InsertNewInstBefore(
8220 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8222 return new StoreInst(NewCast, CastOp);
8229 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8230 Value *Val = SI.getOperand(0);
8231 Value *Ptr = SI.getOperand(1);
8233 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8234 EraseInstFromFunction(SI);
8239 // If the RHS is an alloca with a single use, zapify the store, making the
8241 if (Ptr->hasOneUse()) {
8242 if (isa<AllocaInst>(Ptr)) {
8243 EraseInstFromFunction(SI);
8248 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8249 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8250 GEP->getOperand(0)->hasOneUse()) {
8251 EraseInstFromFunction(SI);
8257 // Do really simple DSE, to catch cases where there are several consequtive
8258 // stores to the same location, separated by a few arithmetic operations. This
8259 // situation often occurs with bitfield accesses.
8260 BasicBlock::iterator BBI = &SI;
8261 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8265 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8266 // Prev store isn't volatile, and stores to the same location?
8267 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8270 EraseInstFromFunction(*PrevSI);
8276 // If this is a load, we have to stop. However, if the loaded value is from
8277 // the pointer we're loading and is producing the pointer we're storing,
8278 // then *this* store is dead (X = load P; store X -> P).
8279 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8280 if (LI == Val && LI->getOperand(0) == Ptr) {
8281 EraseInstFromFunction(SI);
8285 // Otherwise, this is a load from some other location. Stores before it
8290 // Don't skip over loads or things that can modify memory.
8291 if (BBI->mayWriteToMemory())
8296 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8298 // store X, null -> turns into 'unreachable' in SimplifyCFG
8299 if (isa<ConstantPointerNull>(Ptr)) {
8300 if (!isa<UndefValue>(Val)) {
8301 SI.setOperand(0, UndefValue::get(Val->getType()));
8302 if (Instruction *U = dyn_cast<Instruction>(Val))
8303 WorkList.push_back(U); // Dropped a use.
8306 return 0; // Do not modify these!
8309 // store undef, Ptr -> noop
8310 if (isa<UndefValue>(Val)) {
8311 EraseInstFromFunction(SI);
8316 // If the pointer destination is a cast, see if we can fold the cast into the
8318 if (isa<CastInst>(Ptr))
8319 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8321 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8323 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8327 // If this store is the last instruction in the basic block, and if the block
8328 // ends with an unconditional branch, try to move it to the successor block.
8330 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8331 if (BI->isUnconditional()) {
8332 // Check to see if the successor block has exactly two incoming edges. If
8333 // so, see if the other predecessor contains a store to the same location.
8334 // if so, insert a PHI node (if needed) and move the stores down.
8335 BasicBlock *Dest = BI->getSuccessor(0);
8337 pred_iterator PI = pred_begin(Dest);
8338 BasicBlock *Other = 0;
8339 if (*PI != BI->getParent())
8342 if (PI != pred_end(Dest)) {
8343 if (*PI != BI->getParent())
8348 if (++PI != pred_end(Dest))
8351 if (Other) { // If only one other pred...
8352 BBI = Other->getTerminator();
8353 // Make sure this other block ends in an unconditional branch and that
8354 // there is an instruction before the branch.
8355 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8356 BBI != Other->begin()) {
8358 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8360 // If this instruction is a store to the same location.
8361 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8362 // Okay, we know we can perform this transformation. Insert a PHI
8363 // node now if we need it.
8364 Value *MergedVal = OtherStore->getOperand(0);
8365 if (MergedVal != SI.getOperand(0)) {
8366 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8367 PN->reserveOperandSpace(2);
8368 PN->addIncoming(SI.getOperand(0), SI.getParent());
8369 PN->addIncoming(OtherStore->getOperand(0), Other);
8370 MergedVal = InsertNewInstBefore(PN, Dest->front());
8373 // Advance to a place where it is safe to insert the new store and
8375 BBI = Dest->begin();
8376 while (isa<PHINode>(BBI)) ++BBI;
8377 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8378 OtherStore->isVolatile()), *BBI);
8380 // Nuke the old stores.
8381 EraseInstFromFunction(SI);
8382 EraseInstFromFunction(*OtherStore);
8394 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8395 // Change br (not X), label True, label False to: br X, label False, True
8397 BasicBlock *TrueDest;
8398 BasicBlock *FalseDest;
8399 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8400 !isa<Constant>(X)) {
8401 // Swap Destinations and condition...
8403 BI.setSuccessor(0, FalseDest);
8404 BI.setSuccessor(1, TrueDest);
8408 // Cannonicalize fcmp_one -> fcmp_oeq
8409 FCmpInst::Predicate FPred; Value *Y;
8410 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8411 TrueDest, FalseDest)))
8412 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8413 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8414 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8415 std::string Name = I->getName(); I->setName("");
8416 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8417 Value *NewSCC = new FCmpInst(NewPred, X, Y, Name, I);
8418 // Swap Destinations and condition...
8419 BI.setCondition(NewSCC);
8420 BI.setSuccessor(0, FalseDest);
8421 BI.setSuccessor(1, TrueDest);
8422 removeFromWorkList(I);
8423 I->getParent()->getInstList().erase(I);
8424 WorkList.push_back(cast<Instruction>(NewSCC));
8428 // Cannonicalize icmp_ne -> icmp_eq
8429 ICmpInst::Predicate IPred;
8430 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8431 TrueDest, FalseDest)))
8432 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8433 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8434 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8435 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8436 std::string Name = I->getName(); I->setName("");
8437 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8438 Value *NewSCC = new ICmpInst(NewPred, X, Y, Name, I);
8439 // Swap Destinations and condition...
8440 BI.setCondition(NewSCC);
8441 BI.setSuccessor(0, FalseDest);
8442 BI.setSuccessor(1, TrueDest);
8443 removeFromWorkList(I);
8444 I->getParent()->getInstList().erase(I);
8445 WorkList.push_back(cast<Instruction>(NewSCC));
8452 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8453 Value *Cond = SI.getCondition();
8454 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8455 if (I->getOpcode() == Instruction::Add)
8456 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8457 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8458 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8459 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8461 SI.setOperand(0, I->getOperand(0));
8462 WorkList.push_back(I);
8469 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8470 /// is to leave as a vector operation.
8471 static bool CheapToScalarize(Value *V, bool isConstant) {
8472 if (isa<ConstantAggregateZero>(V))
8474 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8475 if (isConstant) return true;
8476 // If all elts are the same, we can extract.
8477 Constant *Op0 = C->getOperand(0);
8478 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8479 if (C->getOperand(i) != Op0)
8483 Instruction *I = dyn_cast<Instruction>(V);
8484 if (!I) return false;
8486 // Insert element gets simplified to the inserted element or is deleted if
8487 // this is constant idx extract element and its a constant idx insertelt.
8488 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8489 isa<ConstantInt>(I->getOperand(2)))
8491 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8493 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8494 if (BO->hasOneUse() &&
8495 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8496 CheapToScalarize(BO->getOperand(1), isConstant)))
8498 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8499 if (CI->hasOneUse() &&
8500 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8501 CheapToScalarize(CI->getOperand(1), isConstant)))
8507 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8508 /// elements into values that are larger than the #elts in the input.
8509 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8510 unsigned NElts = SVI->getType()->getNumElements();
8511 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8512 return std::vector<unsigned>(NElts, 0);
8513 if (isa<UndefValue>(SVI->getOperand(2)))
8514 return std::vector<unsigned>(NElts, 2*NElts);
8516 std::vector<unsigned> Result;
8517 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8518 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8519 if (isa<UndefValue>(CP->getOperand(i)))
8520 Result.push_back(NElts*2); // undef -> 8
8522 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8526 /// FindScalarElement - Given a vector and an element number, see if the scalar
8527 /// value is already around as a register, for example if it were inserted then
8528 /// extracted from the vector.
8529 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8530 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8531 const PackedType *PTy = cast<PackedType>(V->getType());
8532 unsigned Width = PTy->getNumElements();
8533 if (EltNo >= Width) // Out of range access.
8534 return UndefValue::get(PTy->getElementType());
8536 if (isa<UndefValue>(V))
8537 return UndefValue::get(PTy->getElementType());
8538 else if (isa<ConstantAggregateZero>(V))
8539 return Constant::getNullValue(PTy->getElementType());
8540 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8541 return CP->getOperand(EltNo);
8542 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8543 // If this is an insert to a variable element, we don't know what it is.
8544 if (!isa<ConstantInt>(III->getOperand(2)))
8546 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8548 // If this is an insert to the element we are looking for, return the
8551 return III->getOperand(1);
8553 // Otherwise, the insertelement doesn't modify the value, recurse on its
8555 return FindScalarElement(III->getOperand(0), EltNo);
8556 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8557 unsigned InEl = getShuffleMask(SVI)[EltNo];
8559 return FindScalarElement(SVI->getOperand(0), InEl);
8560 else if (InEl < Width*2)
8561 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8563 return UndefValue::get(PTy->getElementType());
8566 // Otherwise, we don't know.
8570 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8572 // If packed val is undef, replace extract with scalar undef.
8573 if (isa<UndefValue>(EI.getOperand(0)))
8574 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8576 // If packed val is constant 0, replace extract with scalar 0.
8577 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8578 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8580 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8581 // If packed val is constant with uniform operands, replace EI
8582 // with that operand
8583 Constant *op0 = C->getOperand(0);
8584 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8585 if (C->getOperand(i) != op0) {
8590 return ReplaceInstUsesWith(EI, op0);
8593 // If extracting a specified index from the vector, see if we can recursively
8594 // find a previously computed scalar that was inserted into the vector.
8595 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8596 // This instruction only demands the single element from the input vector.
8597 // If the input vector has a single use, simplify it based on this use
8599 uint64_t IndexVal = IdxC->getZExtValue();
8600 if (EI.getOperand(0)->hasOneUse()) {
8602 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8605 EI.setOperand(0, V);
8610 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8611 return ReplaceInstUsesWith(EI, Elt);
8614 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8615 if (I->hasOneUse()) {
8616 // Push extractelement into predecessor operation if legal and
8617 // profitable to do so
8618 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8619 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8620 if (CheapToScalarize(BO, isConstantElt)) {
8621 ExtractElementInst *newEI0 =
8622 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8623 EI.getName()+".lhs");
8624 ExtractElementInst *newEI1 =
8625 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8626 EI.getName()+".rhs");
8627 InsertNewInstBefore(newEI0, EI);
8628 InsertNewInstBefore(newEI1, EI);
8629 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8631 } else if (isa<LoadInst>(I)) {
8632 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8633 PointerType::get(EI.getType()), EI);
8634 GetElementPtrInst *GEP =
8635 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8636 InsertNewInstBefore(GEP, EI);
8637 return new LoadInst(GEP);
8640 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8641 // Extracting the inserted element?
8642 if (IE->getOperand(2) == EI.getOperand(1))
8643 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8644 // If the inserted and extracted elements are constants, they must not
8645 // be the same value, extract from the pre-inserted value instead.
8646 if (isa<Constant>(IE->getOperand(2)) &&
8647 isa<Constant>(EI.getOperand(1))) {
8648 AddUsesToWorkList(EI);
8649 EI.setOperand(0, IE->getOperand(0));
8652 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8653 // If this is extracting an element from a shufflevector, figure out where
8654 // it came from and extract from the appropriate input element instead.
8655 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8656 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8658 if (SrcIdx < SVI->getType()->getNumElements())
8659 Src = SVI->getOperand(0);
8660 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8661 SrcIdx -= SVI->getType()->getNumElements();
8662 Src = SVI->getOperand(1);
8664 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8666 return new ExtractElementInst(Src, SrcIdx);
8673 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8674 /// elements from either LHS or RHS, return the shuffle mask and true.
8675 /// Otherwise, return false.
8676 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8677 std::vector<Constant*> &Mask) {
8678 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8679 "Invalid CollectSingleShuffleElements");
8680 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8682 if (isa<UndefValue>(V)) {
8683 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8685 } else if (V == LHS) {
8686 for (unsigned i = 0; i != NumElts; ++i)
8687 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8689 } else if (V == RHS) {
8690 for (unsigned i = 0; i != NumElts; ++i)
8691 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8693 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8694 // If this is an insert of an extract from some other vector, include it.
8695 Value *VecOp = IEI->getOperand(0);
8696 Value *ScalarOp = IEI->getOperand(1);
8697 Value *IdxOp = IEI->getOperand(2);
8699 if (!isa<ConstantInt>(IdxOp))
8701 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8703 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8704 // Okay, we can handle this if the vector we are insertinting into is
8706 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8707 // If so, update the mask to reflect the inserted undef.
8708 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8711 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8712 if (isa<ConstantInt>(EI->getOperand(1)) &&
8713 EI->getOperand(0)->getType() == V->getType()) {
8714 unsigned ExtractedIdx =
8715 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8717 // This must be extracting from either LHS or RHS.
8718 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8719 // Okay, we can handle this if the vector we are insertinting into is
8721 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8722 // If so, update the mask to reflect the inserted value.
8723 if (EI->getOperand(0) == LHS) {
8724 Mask[InsertedIdx & (NumElts-1)] =
8725 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8727 assert(EI->getOperand(0) == RHS);
8728 Mask[InsertedIdx & (NumElts-1)] =
8729 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8738 // TODO: Handle shufflevector here!
8743 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8744 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8745 /// that computes V and the LHS value of the shuffle.
8746 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8748 assert(isa<PackedType>(V->getType()) &&
8749 (RHS == 0 || V->getType() == RHS->getType()) &&
8750 "Invalid shuffle!");
8751 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8753 if (isa<UndefValue>(V)) {
8754 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8756 } else if (isa<ConstantAggregateZero>(V)) {
8757 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8759 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8760 // If this is an insert of an extract from some other vector, include it.
8761 Value *VecOp = IEI->getOperand(0);
8762 Value *ScalarOp = IEI->getOperand(1);
8763 Value *IdxOp = IEI->getOperand(2);
8765 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8766 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8767 EI->getOperand(0)->getType() == V->getType()) {
8768 unsigned ExtractedIdx =
8769 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8770 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8772 // Either the extracted from or inserted into vector must be RHSVec,
8773 // otherwise we'd end up with a shuffle of three inputs.
8774 if (EI->getOperand(0) == RHS || RHS == 0) {
8775 RHS = EI->getOperand(0);
8776 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8777 Mask[InsertedIdx & (NumElts-1)] =
8778 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8783 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8784 // Everything but the extracted element is replaced with the RHS.
8785 for (unsigned i = 0; i != NumElts; ++i) {
8786 if (i != InsertedIdx)
8787 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8792 // If this insertelement is a chain that comes from exactly these two
8793 // vectors, return the vector and the effective shuffle.
8794 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8795 return EI->getOperand(0);
8800 // TODO: Handle shufflevector here!
8802 // Otherwise, can't do anything fancy. Return an identity vector.
8803 for (unsigned i = 0; i != NumElts; ++i)
8804 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8808 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8809 Value *VecOp = IE.getOperand(0);
8810 Value *ScalarOp = IE.getOperand(1);
8811 Value *IdxOp = IE.getOperand(2);
8813 // If the inserted element was extracted from some other vector, and if the
8814 // indexes are constant, try to turn this into a shufflevector operation.
8815 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8816 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8817 EI->getOperand(0)->getType() == IE.getType()) {
8818 unsigned NumVectorElts = IE.getType()->getNumElements();
8819 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8820 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8822 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8823 return ReplaceInstUsesWith(IE, VecOp);
8825 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8826 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8828 // If we are extracting a value from a vector, then inserting it right
8829 // back into the same place, just use the input vector.
8830 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8831 return ReplaceInstUsesWith(IE, VecOp);
8833 // We could theoretically do this for ANY input. However, doing so could
8834 // turn chains of insertelement instructions into a chain of shufflevector
8835 // instructions, and right now we do not merge shufflevectors. As such,
8836 // only do this in a situation where it is clear that there is benefit.
8837 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8838 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8839 // the values of VecOp, except then one read from EIOp0.
8840 // Build a new shuffle mask.
8841 std::vector<Constant*> Mask;
8842 if (isa<UndefValue>(VecOp))
8843 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8845 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8846 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8849 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8850 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8851 ConstantPacked::get(Mask));
8854 // If this insertelement isn't used by some other insertelement, turn it
8855 // (and any insertelements it points to), into one big shuffle.
8856 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8857 std::vector<Constant*> Mask;
8859 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8860 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8861 // We now have a shuffle of LHS, RHS, Mask.
8862 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8871 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8872 Value *LHS = SVI.getOperand(0);
8873 Value *RHS = SVI.getOperand(1);
8874 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8876 bool MadeChange = false;
8878 // Undefined shuffle mask -> undefined value.
8879 if (isa<UndefValue>(SVI.getOperand(2)))
8880 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8882 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8883 // the undef, change them to undefs.
8884 if (isa<UndefValue>(SVI.getOperand(1))) {
8885 // Scan to see if there are any references to the RHS. If so, replace them
8886 // with undef element refs and set MadeChange to true.
8887 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8888 if (Mask[i] >= e && Mask[i] != 2*e) {
8895 // Remap any references to RHS to use LHS.
8896 std::vector<Constant*> Elts;
8897 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8899 Elts.push_back(UndefValue::get(Type::Int32Ty));
8901 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8903 SVI.setOperand(2, ConstantPacked::get(Elts));
8907 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8908 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8909 if (LHS == RHS || isa<UndefValue>(LHS)) {
8910 if (isa<UndefValue>(LHS) && LHS == RHS) {
8911 // shuffle(undef,undef,mask) -> undef.
8912 return ReplaceInstUsesWith(SVI, LHS);
8915 // Remap any references to RHS to use LHS.
8916 std::vector<Constant*> Elts;
8917 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8919 Elts.push_back(UndefValue::get(Type::Int32Ty));
8921 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8922 (Mask[i] < e && isa<UndefValue>(LHS)))
8923 Mask[i] = 2*e; // Turn into undef.
8925 Mask[i] &= (e-1); // Force to LHS.
8926 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8929 SVI.setOperand(0, SVI.getOperand(1));
8930 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8931 SVI.setOperand(2, ConstantPacked::get(Elts));
8932 LHS = SVI.getOperand(0);
8933 RHS = SVI.getOperand(1);
8937 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8938 bool isLHSID = true, isRHSID = true;
8940 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8941 if (Mask[i] >= e*2) continue; // Ignore undef values.
8942 // Is this an identity shuffle of the LHS value?
8943 isLHSID &= (Mask[i] == i);
8945 // Is this an identity shuffle of the RHS value?
8946 isRHSID &= (Mask[i]-e == i);
8949 // Eliminate identity shuffles.
8950 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8951 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8953 // If the LHS is a shufflevector itself, see if we can combine it with this
8954 // one without producing an unusual shuffle. Here we are really conservative:
8955 // we are absolutely afraid of producing a shuffle mask not in the input
8956 // program, because the code gen may not be smart enough to turn a merged
8957 // shuffle into two specific shuffles: it may produce worse code. As such,
8958 // we only merge two shuffles if the result is one of the two input shuffle
8959 // masks. In this case, merging the shuffles just removes one instruction,
8960 // which we know is safe. This is good for things like turning:
8961 // (splat(splat)) -> splat.
8962 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8963 if (isa<UndefValue>(RHS)) {
8964 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8966 std::vector<unsigned> NewMask;
8967 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8969 NewMask.push_back(2*e);
8971 NewMask.push_back(LHSMask[Mask[i]]);
8973 // If the result mask is equal to the src shuffle or this shuffle mask, do
8975 if (NewMask == LHSMask || NewMask == Mask) {
8976 std::vector<Constant*> Elts;
8977 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8978 if (NewMask[i] >= e*2) {
8979 Elts.push_back(UndefValue::get(Type::Int32Ty));
8981 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
8984 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8985 LHSSVI->getOperand(1),
8986 ConstantPacked::get(Elts));
8991 // See if SimplifyDemandedVectorElts can simplify based on this shuffle. For
8992 // example, if this is a splat, then we only demand from one input element.
8994 if (Value *V = SimplifyDemandedVectorElts(&SVI, (1ULL << Mask.size())-1,
8996 return ReplaceInstUsesWith(SVI, V);
8998 return MadeChange ? &SVI : 0;
9003 void InstCombiner::removeFromWorkList(Instruction *I) {
9004 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
9009 /// TryToSinkInstruction - Try to move the specified instruction from its
9010 /// current block into the beginning of DestBlock, which can only happen if it's
9011 /// safe to move the instruction past all of the instructions between it and the
9012 /// end of its block.
9013 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9014 assert(I->hasOneUse() && "Invariants didn't hold!");
9016 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9017 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9019 // Do not sink alloca instructions out of the entry block.
9020 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9023 // We can only sink load instructions if there is nothing between the load and
9024 // the end of block that could change the value.
9025 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9026 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9028 if (Scan->mayWriteToMemory())
9032 BasicBlock::iterator InsertPos = DestBlock->begin();
9033 while (isa<PHINode>(InsertPos)) ++InsertPos;
9035 I->moveBefore(InsertPos);
9040 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
9041 /// from a global, return the global and the constant. Because of
9042 /// constantexprs, this function is recursive.
9043 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
9044 int64_t &Offset, const TargetData &TD) {
9045 // Trivial case, constant is the global.
9046 if ((GV = dyn_cast<GlobalValue>(C))) {
9051 // Otherwise, if this isn't a constant expr, bail out.
9052 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
9053 if (!CE) return false;
9055 // Look through ptr->int and ptr->ptr casts.
9056 if (CE->getOpcode() == Instruction::PtrToInt ||
9057 CE->getOpcode() == Instruction::BitCast)
9058 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
9060 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
9061 if (CE->getOpcode() == Instruction::GetElementPtr) {
9062 // Cannot compute this if the element type of the pointer is missing size
9064 if (!cast<PointerType>(CE->getOperand(0)->getType())->getElementType()->isSized())
9067 // If the base isn't a global+constant, we aren't either.
9068 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
9071 // Otherwise, add any offset that our operands provide.
9072 gep_type_iterator GTI = gep_type_begin(CE);
9073 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i, ++GTI) {
9074 ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(i));
9075 if (!CI) return false; // Index isn't a simple constant?
9076 if (CI->getZExtValue() == 0) continue; // Not adding anything.
9078 if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
9080 Offset += TD.getStructLayout(ST)->MemberOffsets[CI->getZExtValue()];
9082 const SequentialType *ST = cast<SequentialType>(*GTI);
9083 Offset += TD.getTypeSize(ST->getElementType())*CI->getSExtValue();
9092 /// OptimizeConstantExpr - Given a constant expression and target data layout
9093 /// information, symbolically evaluate the constant expr to something simpler
9095 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
9098 Constant *Ptr = CE->getOperand(0);
9099 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
9100 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
9101 // If this is a constant expr gep that is effectively computing an
9102 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
9103 bool isFoldableGEP = true;
9104 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
9105 if (!isa<ConstantInt>(CE->getOperand(i)))
9106 isFoldableGEP = false;
9107 if (isFoldableGEP) {
9108 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
9109 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
9110 Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
9111 return ConstantExpr::getIntToPtr(C, CE->getType());
9118 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
9119 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
9123 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
9124 // constant. This happens frequently when iterating over a global array.
9125 if (CE->getOpcode() == Instruction::Sub) {
9126 GlobalValue *GV1, *GV2;
9127 int64_t Offs1, Offs2;
9129 if (IsConstantOffsetFromGlobal(CE->getOperand(0), GV1, Offs1, *TD))
9130 if (IsConstantOffsetFromGlobal(CE->getOperand(1), GV2, Offs2, *TD) &&
9132 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
9133 return ConstantInt::get(CE->getType(), Offs1-Offs2);
9137 // TODO: Fold icmp setne/seteq as well.
9143 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9144 /// all reachable code to the worklist.
9146 /// This has a couple of tricks to make the code faster and more powerful. In
9147 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9148 /// them to the worklist (this significantly speeds up instcombine on code where
9149 /// many instructions are dead or constant). Additionally, if we find a branch
9150 /// whose condition is a known constant, we only visit the reachable successors.
9152 static void AddReachableCodeToWorklist(BasicBlock *BB,
9153 std::set<BasicBlock*> &Visited,
9154 std::vector<Instruction*> &WorkList,
9155 const TargetData *TD) {
9156 // We have now visited this block! If we've already been here, bail out.
9157 if (!Visited.insert(BB).second) return;
9159 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9160 Instruction *Inst = BBI++;
9162 // DCE instruction if trivially dead.
9163 if (isInstructionTriviallyDead(Inst)) {
9165 DOUT << "IC: DCE: " << *Inst;
9166 Inst->eraseFromParent();
9170 // ConstantProp instruction if trivially constant.
9171 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9172 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9173 C = OptimizeConstantExpr(CE, TD);
9174 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9175 Inst->replaceAllUsesWith(C);
9177 Inst->eraseFromParent();
9181 WorkList.push_back(Inst);
9184 // Recursively visit successors. If this is a branch or switch on a constant,
9185 // only visit the reachable successor.
9186 TerminatorInst *TI = BB->getTerminator();
9187 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9188 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9189 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9190 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9194 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9195 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9196 // See if this is an explicit destination.
9197 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9198 if (SI->getCaseValue(i) == Cond) {
9199 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9203 // Otherwise it is the default destination.
9204 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9209 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9210 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9213 bool InstCombiner::runOnFunction(Function &F) {
9214 bool Changed = false;
9215 TD = &getAnalysis<TargetData>();
9218 // Do a depth-first traversal of the function, populate the worklist with
9219 // the reachable instructions. Ignore blocks that are not reachable. Keep
9220 // track of which blocks we visit.
9221 std::set<BasicBlock*> Visited;
9222 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9224 // Do a quick scan over the function. If we find any blocks that are
9225 // unreachable, remove any instructions inside of them. This prevents
9226 // the instcombine code from having to deal with some bad special cases.
9227 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9228 if (!Visited.count(BB)) {
9229 Instruction *Term = BB->getTerminator();
9230 while (Term != BB->begin()) { // Remove instrs bottom-up
9231 BasicBlock::iterator I = Term; --I;
9233 DOUT << "IC: DCE: " << *I;
9236 if (!I->use_empty())
9237 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9238 I->eraseFromParent();
9243 while (!WorkList.empty()) {
9244 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9245 WorkList.pop_back();
9247 // Check to see if we can DCE the instruction.
9248 if (isInstructionTriviallyDead(I)) {
9249 // Add operands to the worklist.
9250 if (I->getNumOperands() < 4)
9251 AddUsesToWorkList(*I);
9254 DOUT << "IC: DCE: " << *I;
9256 I->eraseFromParent();
9257 removeFromWorkList(I);
9261 // Instruction isn't dead, see if we can constant propagate it.
9262 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9263 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9264 C = OptimizeConstantExpr(CE, TD);
9265 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9267 // Add operands to the worklist.
9268 AddUsesToWorkList(*I);
9269 ReplaceInstUsesWith(*I, C);
9272 I->eraseFromParent();
9273 removeFromWorkList(I);
9277 // See if we can trivially sink this instruction to a successor basic block.
9278 if (I->hasOneUse()) {
9279 BasicBlock *BB = I->getParent();
9280 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9281 if (UserParent != BB) {
9282 bool UserIsSuccessor = false;
9283 // See if the user is one of our successors.
9284 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9285 if (*SI == UserParent) {
9286 UserIsSuccessor = true;
9290 // If the user is one of our immediate successors, and if that successor
9291 // only has us as a predecessors (we'd have to split the critical edge
9292 // otherwise), we can keep going.
9293 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9294 next(pred_begin(UserParent)) == pred_end(UserParent))
9295 // Okay, the CFG is simple enough, try to sink this instruction.
9296 Changed |= TryToSinkInstruction(I, UserParent);
9300 // Now that we have an instruction, try combining it to simplify it...
9301 if (Instruction *Result = visit(*I)) {
9303 // Should we replace the old instruction with a new one?
9305 DOUT << "IC: Old = " << *I
9306 << " New = " << *Result;
9308 // Everything uses the new instruction now.
9309 I->replaceAllUsesWith(Result);
9311 // Push the new instruction and any users onto the worklist.
9312 WorkList.push_back(Result);
9313 AddUsersToWorkList(*Result);
9315 // Move the name to the new instruction first...
9316 std::string OldName = I->getName(); I->setName("");
9317 Result->setName(OldName);
9319 // Insert the new instruction into the basic block...
9320 BasicBlock *InstParent = I->getParent();
9321 BasicBlock::iterator InsertPos = I;
9323 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9324 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9327 InstParent->getInstList().insert(InsertPos, Result);
9329 // Make sure that we reprocess all operands now that we reduced their
9331 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9332 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9333 WorkList.push_back(OpI);
9335 // Instructions can end up on the worklist more than once. Make sure
9336 // we do not process an instruction that has been deleted.
9337 removeFromWorkList(I);
9339 // Erase the old instruction.
9340 InstParent->getInstList().erase(I);
9342 DOUT << "IC: MOD = " << *I;
9344 // If the instruction was modified, it's possible that it is now dead.
9345 // if so, remove it.
9346 if (isInstructionTriviallyDead(I)) {
9347 // Make sure we process all operands now that we are reducing their
9349 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9350 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9351 WorkList.push_back(OpI);
9353 // Instructions may end up in the worklist more than once. Erase all
9354 // occurrences of this instruction.
9355 removeFromWorkList(I);
9356 I->eraseFromParent();
9358 WorkList.push_back(Result);
9359 AddUsersToWorkList(*Result);
9369 FunctionPass *llvm::createInstructionCombiningPass() {
9370 return new InstCombiner();