1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
62 using namespace llvm::PatternMatch;
64 STATISTIC(NumCombined , "Number of insts combined");
65 STATISTIC(NumConstProp, "Number of constant folds");
66 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
67 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 class VISIBILITY_HIDDEN InstCombiner
72 : public FunctionPass,
73 public InstVisitor<InstCombiner, Instruction*> {
74 // Worklist of all of the instructions that need to be simplified.
75 std::vector<Instruction*> Worklist;
76 DenseMap<Instruction*, unsigned> WorklistMap;
78 bool MustPreserveLCSSA;
80 static char ID; // Pass identification, replacement for typeid
81 InstCombiner() : FunctionPass((intptr_t)&ID) {}
83 /// AddToWorkList - Add the specified instruction to the worklist if it
84 /// isn't already in it.
85 void AddToWorkList(Instruction *I) {
86 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
87 Worklist.push_back(I);
90 // RemoveFromWorkList - remove I from the worklist if it exists.
91 void RemoveFromWorkList(Instruction *I) {
92 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
93 if (It == WorklistMap.end()) return; // Not in worklist.
95 // Don't bother moving everything down, just null out the slot.
96 Worklist[It->second] = 0;
98 WorklistMap.erase(It);
101 Instruction *RemoveOneFromWorkList() {
102 Instruction *I = Worklist.back();
104 WorklistMap.erase(I);
109 /// AddUsersToWorkList - When an instruction is simplified, add all users of
110 /// the instruction to the work lists because they might get more simplified
113 void AddUsersToWorkList(Value &I) {
114 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
116 AddToWorkList(cast<Instruction>(*UI));
119 /// AddUsesToWorkList - When an instruction is simplified, add operands to
120 /// the work lists because they might get more simplified now.
122 void AddUsesToWorkList(Instruction &I) {
123 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
124 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
128 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
129 /// dead. Add all of its operands to the worklist, turning them into
130 /// undef's to reduce the number of uses of those instructions.
132 /// Return the specified operand before it is turned into an undef.
134 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
135 Value *R = I.getOperand(op);
137 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
138 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
140 // Set the operand to undef to drop the use.
141 I.setOperand(i, UndefValue::get(Op->getType()));
148 virtual bool runOnFunction(Function &F);
150 bool DoOneIteration(Function &F, unsigned ItNum);
152 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
153 AU.addRequired<TargetData>();
154 AU.addPreservedID(LCSSAID);
155 AU.setPreservesCFG();
158 TargetData &getTargetData() const { return *TD; }
160 // Visitation implementation - Implement instruction combining for different
161 // instruction types. The semantics are as follows:
163 // null - No change was made
164 // I - Change was made, I is still valid, I may be dead though
165 // otherwise - Change was made, replace I with returned instruction
167 Instruction *visitAdd(BinaryOperator &I);
168 Instruction *visitSub(BinaryOperator &I);
169 Instruction *visitMul(BinaryOperator &I);
170 Instruction *visitURem(BinaryOperator &I);
171 Instruction *visitSRem(BinaryOperator &I);
172 Instruction *visitFRem(BinaryOperator &I);
173 Instruction *commonRemTransforms(BinaryOperator &I);
174 Instruction *commonIRemTransforms(BinaryOperator &I);
175 Instruction *commonDivTransforms(BinaryOperator &I);
176 Instruction *commonIDivTransforms(BinaryOperator &I);
177 Instruction *visitUDiv(BinaryOperator &I);
178 Instruction *visitSDiv(BinaryOperator &I);
179 Instruction *visitFDiv(BinaryOperator &I);
180 Instruction *visitAnd(BinaryOperator &I);
181 Instruction *visitOr (BinaryOperator &I);
182 Instruction *visitXor(BinaryOperator &I);
183 Instruction *visitShl(BinaryOperator &I);
184 Instruction *visitAShr(BinaryOperator &I);
185 Instruction *visitLShr(BinaryOperator &I);
186 Instruction *commonShiftTransforms(BinaryOperator &I);
187 Instruction *visitFCmpInst(FCmpInst &I);
188 Instruction *visitICmpInst(ICmpInst &I);
189 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
190 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
193 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
194 ConstantInt *DivRHS);
196 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
197 ICmpInst::Predicate Cond, Instruction &I);
198 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
200 Instruction *commonCastTransforms(CastInst &CI);
201 Instruction *commonIntCastTransforms(CastInst &CI);
202 Instruction *commonPointerCastTransforms(CastInst &CI);
203 Instruction *visitTrunc(TruncInst &CI);
204 Instruction *visitZExt(ZExtInst &CI);
205 Instruction *visitSExt(SExtInst &CI);
206 Instruction *visitFPTrunc(FPTruncInst &CI);
207 Instruction *visitFPExt(CastInst &CI);
208 Instruction *visitFPToUI(CastInst &CI);
209 Instruction *visitFPToSI(CastInst &CI);
210 Instruction *visitUIToFP(CastInst &CI);
211 Instruction *visitSIToFP(CastInst &CI);
212 Instruction *visitPtrToInt(CastInst &CI);
213 Instruction *visitIntToPtr(IntToPtrInst &CI);
214 Instruction *visitBitCast(BitCastInst &CI);
215 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
217 Instruction *visitSelectInst(SelectInst &CI);
218 Instruction *visitCallInst(CallInst &CI);
219 Instruction *visitInvokeInst(InvokeInst &II);
220 Instruction *visitPHINode(PHINode &PN);
221 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
222 Instruction *visitAllocationInst(AllocationInst &AI);
223 Instruction *visitFreeInst(FreeInst &FI);
224 Instruction *visitLoadInst(LoadInst &LI);
225 Instruction *visitStoreInst(StoreInst &SI);
226 Instruction *visitBranchInst(BranchInst &BI);
227 Instruction *visitSwitchInst(SwitchInst &SI);
228 Instruction *visitInsertElementInst(InsertElementInst &IE);
229 Instruction *visitExtractElementInst(ExtractElementInst &EI);
230 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
232 // visitInstruction - Specify what to return for unhandled instructions...
233 Instruction *visitInstruction(Instruction &I) { return 0; }
236 Instruction *visitCallSite(CallSite CS);
237 bool transformConstExprCastCall(CallSite CS);
238 Instruction *transformCallThroughTrampoline(CallSite CS);
241 // InsertNewInstBefore - insert an instruction New before instruction Old
242 // in the program. Add the new instruction to the worklist.
244 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
245 assert(New && New->getParent() == 0 &&
246 "New instruction already inserted into a basic block!");
247 BasicBlock *BB = Old.getParent();
248 BB->getInstList().insert(&Old, New); // Insert inst
253 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
254 /// This also adds the cast to the worklist. Finally, this returns the
256 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
258 if (V->getType() == Ty) return V;
260 if (Constant *CV = dyn_cast<Constant>(V))
261 return ConstantExpr::getCast(opc, CV, Ty);
263 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
268 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
269 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
273 // ReplaceInstUsesWith - This method is to be used when an instruction is
274 // found to be dead, replacable with another preexisting expression. Here
275 // we add all uses of I to the worklist, replace all uses of I with the new
276 // value, then return I, so that the inst combiner will know that I was
279 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
280 AddUsersToWorkList(I); // Add all modified instrs to worklist
282 I.replaceAllUsesWith(V);
285 // If we are replacing the instruction with itself, this must be in a
286 // segment of unreachable code, so just clobber the instruction.
287 I.replaceAllUsesWith(UndefValue::get(I.getType()));
292 // UpdateValueUsesWith - This method is to be used when an value is
293 // found to be replacable with another preexisting expression or was
294 // updated. Here we add all uses of I to the worklist, replace all uses of
295 // I with the new value (unless the instruction was just updated), then
296 // return true, so that the inst combiner will know that I was modified.
298 bool UpdateValueUsesWith(Value *Old, Value *New) {
299 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
301 Old->replaceAllUsesWith(New);
302 if (Instruction *I = dyn_cast<Instruction>(Old))
304 if (Instruction *I = dyn_cast<Instruction>(New))
309 // EraseInstFromFunction - When dealing with an instruction that has side
310 // effects or produces a void value, we can't rely on DCE to delete the
311 // instruction. Instead, visit methods should return the value returned by
313 Instruction *EraseInstFromFunction(Instruction &I) {
314 assert(I.use_empty() && "Cannot erase instruction that is used!");
315 AddUsesToWorkList(I);
316 RemoveFromWorkList(&I);
318 return 0; // Don't do anything with FI
322 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
323 /// InsertBefore instruction. This is specialized a bit to avoid inserting
324 /// casts that are known to not do anything...
326 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
327 Value *V, const Type *DestTy,
328 Instruction *InsertBefore);
330 /// SimplifyCommutative - This performs a few simplifications for
331 /// commutative operators.
332 bool SimplifyCommutative(BinaryOperator &I);
334 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
335 /// most-complex to least-complex order.
336 bool SimplifyCompare(CmpInst &I);
338 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
339 /// on the demanded bits.
340 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
341 APInt& KnownZero, APInt& KnownOne,
344 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
345 uint64_t &UndefElts, unsigned Depth = 0);
347 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
348 // PHI node as operand #0, see if we can fold the instruction into the PHI
349 // (which is only possible if all operands to the PHI are constants).
350 Instruction *FoldOpIntoPhi(Instruction &I);
352 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
353 // operator and they all are only used by the PHI, PHI together their
354 // inputs, and do the operation once, to the result of the PHI.
355 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
356 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
359 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
360 ConstantInt *AndRHS, BinaryOperator &TheAnd);
362 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
363 bool isSub, Instruction &I);
364 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
365 bool isSigned, bool Inside, Instruction &IB);
366 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
367 Instruction *MatchBSwap(BinaryOperator &I);
368 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
369 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
372 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
375 char InstCombiner::ID = 0;
376 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
379 // getComplexity: Assign a complexity or rank value to LLVM Values...
380 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
381 static unsigned getComplexity(Value *V) {
382 if (isa<Instruction>(V)) {
383 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
387 if (isa<Argument>(V)) return 3;
388 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
391 // isOnlyUse - Return true if this instruction will be deleted if we stop using
393 static bool isOnlyUse(Value *V) {
394 return V->hasOneUse() || isa<Constant>(V);
397 // getPromotedType - Return the specified type promoted as it would be to pass
398 // though a va_arg area...
399 static const Type *getPromotedType(const Type *Ty) {
400 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
401 if (ITy->getBitWidth() < 32)
402 return Type::Int32Ty;
407 /// getBitCastOperand - If the specified operand is a CastInst or a constant
408 /// expression bitcast, return the operand value, otherwise return null.
409 static Value *getBitCastOperand(Value *V) {
410 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
411 return I->getOperand(0);
412 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
413 if (CE->getOpcode() == Instruction::BitCast)
414 return CE->getOperand(0);
418 /// This function is a wrapper around CastInst::isEliminableCastPair. It
419 /// simply extracts arguments and returns what that function returns.
420 static Instruction::CastOps
421 isEliminableCastPair(
422 const CastInst *CI, ///< The first cast instruction
423 unsigned opcode, ///< The opcode of the second cast instruction
424 const Type *DstTy, ///< The target type for the second cast instruction
425 TargetData *TD ///< The target data for pointer size
428 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
429 const Type *MidTy = CI->getType(); // B from above
431 // Get the opcodes of the two Cast instructions
432 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
433 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
435 return Instruction::CastOps(
436 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
437 DstTy, TD->getIntPtrType()));
440 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
441 /// in any code being generated. It does not require codegen if V is simple
442 /// enough or if the cast can be folded into other casts.
443 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
444 const Type *Ty, TargetData *TD) {
445 if (V->getType() == Ty || isa<Constant>(V)) return false;
447 // If this is another cast that can be eliminated, it isn't codegen either.
448 if (const CastInst *CI = dyn_cast<CastInst>(V))
449 if (isEliminableCastPair(CI, opcode, Ty, TD))
454 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
455 /// InsertBefore instruction. This is specialized a bit to avoid inserting
456 /// casts that are known to not do anything...
458 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
459 Value *V, const Type *DestTy,
460 Instruction *InsertBefore) {
461 if (V->getType() == DestTy) return V;
462 if (Constant *C = dyn_cast<Constant>(V))
463 return ConstantExpr::getCast(opcode, C, DestTy);
465 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
468 // SimplifyCommutative - This performs a few simplifications for commutative
471 // 1. Order operands such that they are listed from right (least complex) to
472 // left (most complex). This puts constants before unary operators before
475 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
476 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
478 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
479 bool Changed = false;
480 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
481 Changed = !I.swapOperands();
483 if (!I.isAssociative()) return Changed;
484 Instruction::BinaryOps Opcode = I.getOpcode();
485 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
486 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
487 if (isa<Constant>(I.getOperand(1))) {
488 Constant *Folded = ConstantExpr::get(I.getOpcode(),
489 cast<Constant>(I.getOperand(1)),
490 cast<Constant>(Op->getOperand(1)));
491 I.setOperand(0, Op->getOperand(0));
492 I.setOperand(1, Folded);
494 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
495 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
496 isOnlyUse(Op) && isOnlyUse(Op1)) {
497 Constant *C1 = cast<Constant>(Op->getOperand(1));
498 Constant *C2 = cast<Constant>(Op1->getOperand(1));
500 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
501 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
502 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
506 I.setOperand(0, New);
507 I.setOperand(1, Folded);
514 /// SimplifyCompare - For a CmpInst this function just orders the operands
515 /// so that theyare listed from right (least complex) to left (most complex).
516 /// This puts constants before unary operators before binary operators.
517 bool InstCombiner::SimplifyCompare(CmpInst &I) {
518 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
521 // Compare instructions are not associative so there's nothing else we can do.
525 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
526 // if the LHS is a constant zero (which is the 'negate' form).
528 static inline Value *dyn_castNegVal(Value *V) {
529 if (BinaryOperator::isNeg(V))
530 return BinaryOperator::getNegArgument(V);
532 // Constants can be considered to be negated values if they can be folded.
533 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
534 return ConstantExpr::getNeg(C);
538 static inline Value *dyn_castNotVal(Value *V) {
539 if (BinaryOperator::isNot(V))
540 return BinaryOperator::getNotArgument(V);
542 // Constants can be considered to be not'ed values...
543 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
544 return ConstantInt::get(~C->getValue());
548 // dyn_castFoldableMul - If this value is a multiply that can be folded into
549 // other computations (because it has a constant operand), return the
550 // non-constant operand of the multiply, and set CST to point to the multiplier.
551 // Otherwise, return null.
553 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
554 if (V->hasOneUse() && V->getType()->isInteger())
555 if (Instruction *I = dyn_cast<Instruction>(V)) {
556 if (I->getOpcode() == Instruction::Mul)
557 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
558 return I->getOperand(0);
559 if (I->getOpcode() == Instruction::Shl)
560 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
561 // The multiplier is really 1 << CST.
562 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
563 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
564 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
565 return I->getOperand(0);
571 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
572 /// expression, return it.
573 static User *dyn_castGetElementPtr(Value *V) {
574 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
575 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
576 if (CE->getOpcode() == Instruction::GetElementPtr)
577 return cast<User>(V);
581 /// AddOne - Add one to a ConstantInt
582 static ConstantInt *AddOne(ConstantInt *C) {
583 APInt Val(C->getValue());
584 return ConstantInt::get(++Val);
586 /// SubOne - Subtract one from a ConstantInt
587 static ConstantInt *SubOne(ConstantInt *C) {
588 APInt Val(C->getValue());
589 return ConstantInt::get(--Val);
591 /// Add - Add two ConstantInts together
592 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
593 return ConstantInt::get(C1->getValue() + C2->getValue());
595 /// And - Bitwise AND two ConstantInts together
596 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
597 return ConstantInt::get(C1->getValue() & C2->getValue());
599 /// Subtract - Subtract one ConstantInt from another
600 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
601 return ConstantInt::get(C1->getValue() - C2->getValue());
603 /// Multiply - Multiply two ConstantInts together
604 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
605 return ConstantInt::get(C1->getValue() * C2->getValue());
608 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
609 /// known to be either zero or one and return them in the KnownZero/KnownOne
610 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
612 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
613 /// we cannot optimize based on the assumption that it is zero without changing
614 /// it to be an explicit zero. If we don't change it to zero, other code could
615 /// optimized based on the contradictory assumption that it is non-zero.
616 /// Because instcombine aggressively folds operations with undef args anyway,
617 /// this won't lose us code quality.
618 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
619 APInt& KnownOne, unsigned Depth = 0) {
620 assert(V && "No Value?");
621 assert(Depth <= 6 && "Limit Search Depth");
622 uint32_t BitWidth = Mask.getBitWidth();
623 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
624 KnownZero.getBitWidth() == BitWidth &&
625 KnownOne.getBitWidth() == BitWidth &&
626 "V, Mask, KnownOne and KnownZero should have same BitWidth");
627 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
628 // We know all of the bits for a constant!
629 KnownOne = CI->getValue() & Mask;
630 KnownZero = ~KnownOne & Mask;
634 if (Depth == 6 || Mask == 0)
635 return; // Limit search depth.
637 Instruction *I = dyn_cast<Instruction>(V);
640 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
641 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
643 switch (I->getOpcode()) {
644 case Instruction::And: {
645 // If either the LHS or the RHS are Zero, the result is zero.
646 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
647 APInt Mask2(Mask & ~KnownZero);
648 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
649 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
650 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
652 // Output known-1 bits are only known if set in both the LHS & RHS.
653 KnownOne &= KnownOne2;
654 // Output known-0 are known to be clear if zero in either the LHS | RHS.
655 KnownZero |= KnownZero2;
658 case Instruction::Or: {
659 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
660 APInt Mask2(Mask & ~KnownOne);
661 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
662 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
663 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
665 // Output known-0 bits are only known if clear in both the LHS & RHS.
666 KnownZero &= KnownZero2;
667 // Output known-1 are known to be set if set in either the LHS | RHS.
668 KnownOne |= KnownOne2;
671 case Instruction::Xor: {
672 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
673 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
674 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
675 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
677 // Output known-0 bits are known if clear or set in both the LHS & RHS.
678 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
679 // Output known-1 are known to be set if set in only one of the LHS, RHS.
680 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
681 KnownZero = KnownZeroOut;
684 case Instruction::Select:
685 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
686 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
687 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
688 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
690 // Only known if known in both the LHS and RHS.
691 KnownOne &= KnownOne2;
692 KnownZero &= KnownZero2;
694 case Instruction::FPTrunc:
695 case Instruction::FPExt:
696 case Instruction::FPToUI:
697 case Instruction::FPToSI:
698 case Instruction::SIToFP:
699 case Instruction::PtrToInt:
700 case Instruction::UIToFP:
701 case Instruction::IntToPtr:
702 return; // Can't work with floating point or pointers
703 case Instruction::Trunc: {
704 // All these have integer operands
705 uint32_t SrcBitWidth =
706 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
708 MaskIn.zext(SrcBitWidth);
709 KnownZero.zext(SrcBitWidth);
710 KnownOne.zext(SrcBitWidth);
711 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
712 KnownZero.trunc(BitWidth);
713 KnownOne.trunc(BitWidth);
716 case Instruction::BitCast: {
717 const Type *SrcTy = I->getOperand(0)->getType();
718 if (SrcTy->isInteger()) {
719 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
724 case Instruction::ZExt: {
725 // Compute the bits in the result that are not present in the input.
726 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
727 uint32_t SrcBitWidth = SrcTy->getBitWidth();
730 MaskIn.trunc(SrcBitWidth);
731 KnownZero.trunc(SrcBitWidth);
732 KnownOne.trunc(SrcBitWidth);
733 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
734 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
735 // The top bits are known to be zero.
736 KnownZero.zext(BitWidth);
737 KnownOne.zext(BitWidth);
738 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
741 case Instruction::SExt: {
742 // Compute the bits in the result that are not present in the input.
743 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
744 uint32_t SrcBitWidth = SrcTy->getBitWidth();
747 MaskIn.trunc(SrcBitWidth);
748 KnownZero.trunc(SrcBitWidth);
749 KnownOne.trunc(SrcBitWidth);
750 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
751 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
752 KnownZero.zext(BitWidth);
753 KnownOne.zext(BitWidth);
755 // If the sign bit of the input is known set or clear, then we know the
756 // top bits of the result.
757 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
758 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
759 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
760 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
763 case Instruction::Shl:
764 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
765 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
766 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
767 APInt Mask2(Mask.lshr(ShiftAmt));
768 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
769 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
770 KnownZero <<= ShiftAmt;
771 KnownOne <<= ShiftAmt;
772 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
776 case Instruction::LShr:
777 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
778 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
779 // Compute the new bits that are at the top now.
780 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
782 // Unsigned shift right.
783 APInt Mask2(Mask.shl(ShiftAmt));
784 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
785 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
786 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
787 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
788 // high bits known zero.
789 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
793 case Instruction::AShr:
794 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
795 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
796 // Compute the new bits that are at the top now.
797 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
799 // Signed shift right.
800 APInt Mask2(Mask.shl(ShiftAmt));
801 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
802 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
803 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
804 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
806 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
807 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
808 KnownZero |= HighBits;
809 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
810 KnownOne |= HighBits;
817 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
818 /// this predicate to simplify operations downstream. Mask is known to be zero
819 /// for bits that V cannot have.
820 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
821 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
822 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
823 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
824 return (KnownZero & Mask) == Mask;
827 /// ShrinkDemandedConstant - Check to see if the specified operand of the
828 /// specified instruction is a constant integer. If so, check to see if there
829 /// are any bits set in the constant that are not demanded. If so, shrink the
830 /// constant and return true.
831 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
833 assert(I && "No instruction?");
834 assert(OpNo < I->getNumOperands() && "Operand index too large");
836 // If the operand is not a constant integer, nothing to do.
837 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
838 if (!OpC) return false;
840 // If there are no bits set that aren't demanded, nothing to do.
841 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
842 if ((~Demanded & OpC->getValue()) == 0)
845 // This instruction is producing bits that are not demanded. Shrink the RHS.
846 Demanded &= OpC->getValue();
847 I->setOperand(OpNo, ConstantInt::get(Demanded));
851 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
852 // set of known zero and one bits, compute the maximum and minimum values that
853 // could have the specified known zero and known one bits, returning them in
855 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
856 const APInt& KnownZero,
857 const APInt& KnownOne,
858 APInt& Min, APInt& Max) {
859 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
860 assert(KnownZero.getBitWidth() == BitWidth &&
861 KnownOne.getBitWidth() == BitWidth &&
862 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
863 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
864 APInt UnknownBits = ~(KnownZero|KnownOne);
866 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
867 // bit if it is unknown.
869 Max = KnownOne|UnknownBits;
871 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
873 Max.clear(BitWidth-1);
877 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
878 // a set of known zero and one bits, compute the maximum and minimum values that
879 // could have the specified known zero and known one bits, returning them in
881 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
882 const APInt &KnownZero,
883 const APInt &KnownOne,
884 APInt &Min, APInt &Max) {
885 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
886 assert(KnownZero.getBitWidth() == BitWidth &&
887 KnownOne.getBitWidth() == BitWidth &&
888 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
889 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
890 APInt UnknownBits = ~(KnownZero|KnownOne);
892 // The minimum value is when the unknown bits are all zeros.
894 // The maximum value is when the unknown bits are all ones.
895 Max = KnownOne|UnknownBits;
898 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
899 /// value based on the demanded bits. When this function is called, it is known
900 /// that only the bits set in DemandedMask of the result of V are ever used
901 /// downstream. Consequently, depending on the mask and V, it may be possible
902 /// to replace V with a constant or one of its operands. In such cases, this
903 /// function does the replacement and returns true. In all other cases, it
904 /// returns false after analyzing the expression and setting KnownOne and known
905 /// to be one in the expression. KnownZero contains all the bits that are known
906 /// to be zero in the expression. These are provided to potentially allow the
907 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
908 /// the expression. KnownOne and KnownZero always follow the invariant that
909 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
910 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
911 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
912 /// and KnownOne must all be the same.
913 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
914 APInt& KnownZero, APInt& KnownOne,
916 assert(V != 0 && "Null pointer of Value???");
917 assert(Depth <= 6 && "Limit Search Depth");
918 uint32_t BitWidth = DemandedMask.getBitWidth();
919 const IntegerType *VTy = cast<IntegerType>(V->getType());
920 assert(VTy->getBitWidth() == BitWidth &&
921 KnownZero.getBitWidth() == BitWidth &&
922 KnownOne.getBitWidth() == BitWidth &&
923 "Value *V, DemandedMask, KnownZero and KnownOne \
924 must have same BitWidth");
925 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
926 // We know all of the bits for a constant!
927 KnownOne = CI->getValue() & DemandedMask;
928 KnownZero = ~KnownOne & DemandedMask;
934 if (!V->hasOneUse()) { // Other users may use these bits.
935 if (Depth != 0) { // Not at the root.
936 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
937 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
940 // If this is the root being simplified, allow it to have multiple uses,
941 // just set the DemandedMask to all bits.
942 DemandedMask = APInt::getAllOnesValue(BitWidth);
943 } else if (DemandedMask == 0) { // Not demanding any bits from V.
944 if (V != UndefValue::get(VTy))
945 return UpdateValueUsesWith(V, UndefValue::get(VTy));
947 } else if (Depth == 6) { // Limit search depth.
951 Instruction *I = dyn_cast<Instruction>(V);
952 if (!I) return false; // Only analyze instructions.
954 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
955 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
956 switch (I->getOpcode()) {
958 case Instruction::And:
959 // If either the LHS or the RHS are Zero, the result is zero.
960 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
961 RHSKnownZero, RHSKnownOne, Depth+1))
963 assert((RHSKnownZero & RHSKnownOne) == 0 &&
964 "Bits known to be one AND zero?");
966 // If something is known zero on the RHS, the bits aren't demanded on the
968 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
969 LHSKnownZero, LHSKnownOne, Depth+1))
971 assert((LHSKnownZero & LHSKnownOne) == 0 &&
972 "Bits known to be one AND zero?");
974 // If all of the demanded bits are known 1 on one side, return the other.
975 // These bits cannot contribute to the result of the 'and'.
976 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
977 (DemandedMask & ~LHSKnownZero))
978 return UpdateValueUsesWith(I, I->getOperand(0));
979 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
980 (DemandedMask & ~RHSKnownZero))
981 return UpdateValueUsesWith(I, I->getOperand(1));
983 // If all of the demanded bits in the inputs are known zeros, return zero.
984 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
985 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
987 // If the RHS is a constant, see if we can simplify it.
988 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
989 return UpdateValueUsesWith(I, I);
991 // Output known-1 bits are only known if set in both the LHS & RHS.
992 RHSKnownOne &= LHSKnownOne;
993 // Output known-0 are known to be clear if zero in either the LHS | RHS.
994 RHSKnownZero |= LHSKnownZero;
996 case Instruction::Or:
997 // If either the LHS or the RHS are One, the result is One.
998 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
999 RHSKnownZero, RHSKnownOne, Depth+1))
1001 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1002 "Bits known to be one AND zero?");
1003 // If something is known one on the RHS, the bits aren't demanded on the
1005 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1006 LHSKnownZero, LHSKnownOne, Depth+1))
1008 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1009 "Bits known to be one AND zero?");
1011 // If all of the demanded bits are known zero on one side, return the other.
1012 // These bits cannot contribute to the result of the 'or'.
1013 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1014 (DemandedMask & ~LHSKnownOne))
1015 return UpdateValueUsesWith(I, I->getOperand(0));
1016 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1017 (DemandedMask & ~RHSKnownOne))
1018 return UpdateValueUsesWith(I, I->getOperand(1));
1020 // If all of the potentially set bits on one side are known to be set on
1021 // the other side, just use the 'other' side.
1022 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1023 (DemandedMask & (~RHSKnownZero)))
1024 return UpdateValueUsesWith(I, I->getOperand(0));
1025 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1026 (DemandedMask & (~LHSKnownZero)))
1027 return UpdateValueUsesWith(I, I->getOperand(1));
1029 // If the RHS is a constant, see if we can simplify it.
1030 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1031 return UpdateValueUsesWith(I, I);
1033 // Output known-0 bits are only known if clear in both the LHS & RHS.
1034 RHSKnownZero &= LHSKnownZero;
1035 // Output known-1 are known to be set if set in either the LHS | RHS.
1036 RHSKnownOne |= LHSKnownOne;
1038 case Instruction::Xor: {
1039 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1040 RHSKnownZero, RHSKnownOne, Depth+1))
1042 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1043 "Bits known to be one AND zero?");
1044 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1045 LHSKnownZero, LHSKnownOne, Depth+1))
1047 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1048 "Bits known to be one AND zero?");
1050 // If all of the demanded bits are known zero on one side, return the other.
1051 // These bits cannot contribute to the result of the 'xor'.
1052 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1053 return UpdateValueUsesWith(I, I->getOperand(0));
1054 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1055 return UpdateValueUsesWith(I, I->getOperand(1));
1057 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1058 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1059 (RHSKnownOne & LHSKnownOne);
1060 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1061 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1062 (RHSKnownOne & LHSKnownZero);
1064 // If all of the demanded bits are known to be zero on one side or the
1065 // other, turn this into an *inclusive* or.
1066 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1067 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1069 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1071 InsertNewInstBefore(Or, *I);
1072 return UpdateValueUsesWith(I, Or);
1075 // If all of the demanded bits on one side are known, and all of the set
1076 // bits on that side are also known to be set on the other side, turn this
1077 // into an AND, as we know the bits will be cleared.
1078 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1079 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1081 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1082 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1084 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1085 InsertNewInstBefore(And, *I);
1086 return UpdateValueUsesWith(I, And);
1090 // If the RHS is a constant, see if we can simplify it.
1091 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1092 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1093 return UpdateValueUsesWith(I, I);
1095 RHSKnownZero = KnownZeroOut;
1096 RHSKnownOne = KnownOneOut;
1099 case Instruction::Select:
1100 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1101 RHSKnownZero, RHSKnownOne, Depth+1))
1103 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1104 LHSKnownZero, LHSKnownOne, Depth+1))
1106 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1107 "Bits known to be one AND zero?");
1108 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1109 "Bits known to be one AND zero?");
1111 // If the operands are constants, see if we can simplify them.
1112 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1113 return UpdateValueUsesWith(I, I);
1114 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1115 return UpdateValueUsesWith(I, I);
1117 // Only known if known in both the LHS and RHS.
1118 RHSKnownOne &= LHSKnownOne;
1119 RHSKnownZero &= LHSKnownZero;
1121 case Instruction::Trunc: {
1123 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1124 DemandedMask.zext(truncBf);
1125 RHSKnownZero.zext(truncBf);
1126 RHSKnownOne.zext(truncBf);
1127 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1128 RHSKnownZero, RHSKnownOne, Depth+1))
1130 DemandedMask.trunc(BitWidth);
1131 RHSKnownZero.trunc(BitWidth);
1132 RHSKnownOne.trunc(BitWidth);
1133 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1134 "Bits known to be one AND zero?");
1137 case Instruction::BitCast:
1138 if (!I->getOperand(0)->getType()->isInteger())
1141 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1142 RHSKnownZero, RHSKnownOne, Depth+1))
1144 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1145 "Bits known to be one AND zero?");
1147 case Instruction::ZExt: {
1148 // Compute the bits in the result that are not present in the input.
1149 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1150 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1152 DemandedMask.trunc(SrcBitWidth);
1153 RHSKnownZero.trunc(SrcBitWidth);
1154 RHSKnownOne.trunc(SrcBitWidth);
1155 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1156 RHSKnownZero, RHSKnownOne, Depth+1))
1158 DemandedMask.zext(BitWidth);
1159 RHSKnownZero.zext(BitWidth);
1160 RHSKnownOne.zext(BitWidth);
1161 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1162 "Bits known to be one AND zero?");
1163 // The top bits are known to be zero.
1164 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1167 case Instruction::SExt: {
1168 // Compute the bits in the result that are not present in the input.
1169 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1170 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1172 APInt InputDemandedBits = DemandedMask &
1173 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1175 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1176 // If any of the sign extended bits are demanded, we know that the sign
1178 if ((NewBits & DemandedMask) != 0)
1179 InputDemandedBits.set(SrcBitWidth-1);
1181 InputDemandedBits.trunc(SrcBitWidth);
1182 RHSKnownZero.trunc(SrcBitWidth);
1183 RHSKnownOne.trunc(SrcBitWidth);
1184 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1185 RHSKnownZero, RHSKnownOne, Depth+1))
1187 InputDemandedBits.zext(BitWidth);
1188 RHSKnownZero.zext(BitWidth);
1189 RHSKnownOne.zext(BitWidth);
1190 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1191 "Bits known to be one AND zero?");
1193 // If the sign bit of the input is known set or clear, then we know the
1194 // top bits of the result.
1196 // If the input sign bit is known zero, or if the NewBits are not demanded
1197 // convert this into a zero extension.
1198 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1200 // Convert to ZExt cast
1201 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1202 return UpdateValueUsesWith(I, NewCast);
1203 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1204 RHSKnownOne |= NewBits;
1208 case Instruction::Add: {
1209 // Figure out what the input bits are. If the top bits of the and result
1210 // are not demanded, then the add doesn't demand them from its input
1212 uint32_t NLZ = DemandedMask.countLeadingZeros();
1214 // If there is a constant on the RHS, there are a variety of xformations
1216 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1217 // If null, this should be simplified elsewhere. Some of the xforms here
1218 // won't work if the RHS is zero.
1222 // If the top bit of the output is demanded, demand everything from the
1223 // input. Otherwise, we demand all the input bits except NLZ top bits.
1224 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1226 // Find information about known zero/one bits in the input.
1227 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1228 LHSKnownZero, LHSKnownOne, Depth+1))
1231 // If the RHS of the add has bits set that can't affect the input, reduce
1233 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1234 return UpdateValueUsesWith(I, I);
1236 // Avoid excess work.
1237 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1240 // Turn it into OR if input bits are zero.
1241 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1243 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1245 InsertNewInstBefore(Or, *I);
1246 return UpdateValueUsesWith(I, Or);
1249 // We can say something about the output known-zero and known-one bits,
1250 // depending on potential carries from the input constant and the
1251 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1252 // bits set and the RHS constant is 0x01001, then we know we have a known
1253 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1255 // To compute this, we first compute the potential carry bits. These are
1256 // the bits which may be modified. I'm not aware of a better way to do
1258 const APInt& RHSVal = RHS->getValue();
1259 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1261 // Now that we know which bits have carries, compute the known-1/0 sets.
1263 // Bits are known one if they are known zero in one operand and one in the
1264 // other, and there is no input carry.
1265 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1266 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1268 // Bits are known zero if they are known zero in both operands and there
1269 // is no input carry.
1270 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1272 // If the high-bits of this ADD are not demanded, then it does not demand
1273 // the high bits of its LHS or RHS.
1274 if (DemandedMask[BitWidth-1] == 0) {
1275 // Right fill the mask of bits for this ADD to demand the most
1276 // significant bit and all those below it.
1277 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1278 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1279 LHSKnownZero, LHSKnownOne, Depth+1))
1281 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1282 LHSKnownZero, LHSKnownOne, Depth+1))
1288 case Instruction::Sub:
1289 // If the high-bits of this SUB are not demanded, then it does not demand
1290 // the high bits of its LHS or RHS.
1291 if (DemandedMask[BitWidth-1] == 0) {
1292 // Right fill the mask of bits for this SUB to demand the most
1293 // significant bit and all those below it.
1294 uint32_t NLZ = DemandedMask.countLeadingZeros();
1295 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1296 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1297 LHSKnownZero, LHSKnownOne, Depth+1))
1299 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1300 LHSKnownZero, LHSKnownOne, Depth+1))
1304 case Instruction::Shl:
1305 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1306 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1307 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1308 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1309 RHSKnownZero, RHSKnownOne, Depth+1))
1311 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1312 "Bits known to be one AND zero?");
1313 RHSKnownZero <<= ShiftAmt;
1314 RHSKnownOne <<= ShiftAmt;
1315 // low bits known zero.
1317 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1320 case Instruction::LShr:
1321 // For a logical shift right
1322 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1323 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1325 // Unsigned shift right.
1326 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1327 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1328 RHSKnownZero, RHSKnownOne, Depth+1))
1330 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1331 "Bits known to be one AND zero?");
1332 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1333 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1335 // Compute the new bits that are at the top now.
1336 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1337 RHSKnownZero |= HighBits; // high bits known zero.
1341 case Instruction::AShr:
1342 // If this is an arithmetic shift right and only the low-bit is set, we can
1343 // always convert this into a logical shr, even if the shift amount is
1344 // variable. The low bit of the shift cannot be an input sign bit unless
1345 // the shift amount is >= the size of the datatype, which is undefined.
1346 if (DemandedMask == 1) {
1347 // Perform the logical shift right.
1348 Value *NewVal = BinaryOperator::createLShr(
1349 I->getOperand(0), I->getOperand(1), I->getName());
1350 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1351 return UpdateValueUsesWith(I, NewVal);
1354 // If the sign bit is the only bit demanded by this ashr, then there is no
1355 // need to do it, the shift doesn't change the high bit.
1356 if (DemandedMask.isSignBit())
1357 return UpdateValueUsesWith(I, I->getOperand(0));
1359 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1360 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1362 // Signed shift right.
1363 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1364 // If any of the "high bits" are demanded, we should set the sign bit as
1366 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1367 DemandedMaskIn.set(BitWidth-1);
1368 if (SimplifyDemandedBits(I->getOperand(0),
1370 RHSKnownZero, RHSKnownOne, Depth+1))
1372 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1373 "Bits known to be one AND zero?");
1374 // Compute the new bits that are at the top now.
1375 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1376 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1377 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1379 // Handle the sign bits.
1380 APInt SignBit(APInt::getSignBit(BitWidth));
1381 // Adjust to where it is now in the mask.
1382 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1384 // If the input sign bit is known to be zero, or if none of the top bits
1385 // are demanded, turn this into an unsigned shift right.
1386 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1387 (HighBits & ~DemandedMask) == HighBits) {
1388 // Perform the logical shift right.
1389 Value *NewVal = BinaryOperator::createLShr(
1390 I->getOperand(0), SA, I->getName());
1391 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1392 return UpdateValueUsesWith(I, NewVal);
1393 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1394 RHSKnownOne |= HighBits;
1400 // If the client is only demanding bits that we know, return the known
1402 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1403 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1408 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1409 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1410 /// actually used by the caller. This method analyzes which elements of the
1411 /// operand are undef and returns that information in UndefElts.
1413 /// If the information about demanded elements can be used to simplify the
1414 /// operation, the operation is simplified, then the resultant value is
1415 /// returned. This returns null if no change was made.
1416 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1417 uint64_t &UndefElts,
1419 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1420 assert(VWidth <= 64 && "Vector too wide to analyze!");
1421 uint64_t EltMask = ~0ULL >> (64-VWidth);
1422 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1423 "Invalid DemandedElts!");
1425 if (isa<UndefValue>(V)) {
1426 // If the entire vector is undefined, just return this info.
1427 UndefElts = EltMask;
1429 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1430 UndefElts = EltMask;
1431 return UndefValue::get(V->getType());
1435 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1436 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1437 Constant *Undef = UndefValue::get(EltTy);
1439 std::vector<Constant*> Elts;
1440 for (unsigned i = 0; i != VWidth; ++i)
1441 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1442 Elts.push_back(Undef);
1443 UndefElts |= (1ULL << i);
1444 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1445 Elts.push_back(Undef);
1446 UndefElts |= (1ULL << i);
1447 } else { // Otherwise, defined.
1448 Elts.push_back(CP->getOperand(i));
1451 // If we changed the constant, return it.
1452 Constant *NewCP = ConstantVector::get(Elts);
1453 return NewCP != CP ? NewCP : 0;
1454 } else if (isa<ConstantAggregateZero>(V)) {
1455 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1457 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1458 Constant *Zero = Constant::getNullValue(EltTy);
1459 Constant *Undef = UndefValue::get(EltTy);
1460 std::vector<Constant*> Elts;
1461 for (unsigned i = 0; i != VWidth; ++i)
1462 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1463 UndefElts = DemandedElts ^ EltMask;
1464 return ConstantVector::get(Elts);
1467 if (!V->hasOneUse()) { // Other users may use these bits.
1468 if (Depth != 0) { // Not at the root.
1469 // TODO: Just compute the UndefElts information recursively.
1473 } else if (Depth == 10) { // Limit search depth.
1477 Instruction *I = dyn_cast<Instruction>(V);
1478 if (!I) return false; // Only analyze instructions.
1480 bool MadeChange = false;
1481 uint64_t UndefElts2;
1483 switch (I->getOpcode()) {
1486 case Instruction::InsertElement: {
1487 // If this is a variable index, we don't know which element it overwrites.
1488 // demand exactly the same input as we produce.
1489 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1491 // Note that we can't propagate undef elt info, because we don't know
1492 // which elt is getting updated.
1493 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1494 UndefElts2, Depth+1);
1495 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1499 // If this is inserting an element that isn't demanded, remove this
1501 unsigned IdxNo = Idx->getZExtValue();
1502 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1503 return AddSoonDeadInstToWorklist(*I, 0);
1505 // Otherwise, the element inserted overwrites whatever was there, so the
1506 // input demanded set is simpler than the output set.
1507 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1508 DemandedElts & ~(1ULL << IdxNo),
1509 UndefElts, Depth+1);
1510 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1512 // The inserted element is defined.
1513 UndefElts |= 1ULL << IdxNo;
1516 case Instruction::BitCast: {
1517 // Vector->vector casts only.
1518 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1520 unsigned InVWidth = VTy->getNumElements();
1521 uint64_t InputDemandedElts = 0;
1524 if (VWidth == InVWidth) {
1525 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1526 // elements as are demanded of us.
1528 InputDemandedElts = DemandedElts;
1529 } else if (VWidth > InVWidth) {
1533 // If there are more elements in the result than there are in the source,
1534 // then an input element is live if any of the corresponding output
1535 // elements are live.
1536 Ratio = VWidth/InVWidth;
1537 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1538 if (DemandedElts & (1ULL << OutIdx))
1539 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1545 // If there are more elements in the source than there are in the result,
1546 // then an input element is live if the corresponding output element is
1548 Ratio = InVWidth/VWidth;
1549 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1550 if (DemandedElts & (1ULL << InIdx/Ratio))
1551 InputDemandedElts |= 1ULL << InIdx;
1554 // div/rem demand all inputs, because they don't want divide by zero.
1555 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1556 UndefElts2, Depth+1);
1558 I->setOperand(0, TmpV);
1562 UndefElts = UndefElts2;
1563 if (VWidth > InVWidth) {
1564 assert(0 && "Unimp");
1565 // If there are more elements in the result than there are in the source,
1566 // then an output element is undef if the corresponding input element is
1568 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1569 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1570 UndefElts |= 1ULL << OutIdx;
1571 } else if (VWidth < InVWidth) {
1572 assert(0 && "Unimp");
1573 // If there are more elements in the source than there are in the result,
1574 // then a result element is undef if all of the corresponding input
1575 // elements are undef.
1576 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1577 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1578 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1579 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1583 case Instruction::And:
1584 case Instruction::Or:
1585 case Instruction::Xor:
1586 case Instruction::Add:
1587 case Instruction::Sub:
1588 case Instruction::Mul:
1589 // div/rem demand all inputs, because they don't want divide by zero.
1590 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1591 UndefElts, Depth+1);
1592 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1593 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1594 UndefElts2, Depth+1);
1595 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1597 // Output elements are undefined if both are undefined. Consider things
1598 // like undef&0. The result is known zero, not undef.
1599 UndefElts &= UndefElts2;
1602 case Instruction::Call: {
1603 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1605 switch (II->getIntrinsicID()) {
1608 // Binary vector operations that work column-wise. A dest element is a
1609 // function of the corresponding input elements from the two inputs.
1610 case Intrinsic::x86_sse_sub_ss:
1611 case Intrinsic::x86_sse_mul_ss:
1612 case Intrinsic::x86_sse_min_ss:
1613 case Intrinsic::x86_sse_max_ss:
1614 case Intrinsic::x86_sse2_sub_sd:
1615 case Intrinsic::x86_sse2_mul_sd:
1616 case Intrinsic::x86_sse2_min_sd:
1617 case Intrinsic::x86_sse2_max_sd:
1618 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1619 UndefElts, Depth+1);
1620 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1621 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1622 UndefElts2, Depth+1);
1623 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1625 // If only the low elt is demanded and this is a scalarizable intrinsic,
1626 // scalarize it now.
1627 if (DemandedElts == 1) {
1628 switch (II->getIntrinsicID()) {
1630 case Intrinsic::x86_sse_sub_ss:
1631 case Intrinsic::x86_sse_mul_ss:
1632 case Intrinsic::x86_sse2_sub_sd:
1633 case Intrinsic::x86_sse2_mul_sd:
1634 // TODO: Lower MIN/MAX/ABS/etc
1635 Value *LHS = II->getOperand(1);
1636 Value *RHS = II->getOperand(2);
1637 // Extract the element as scalars.
1638 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1639 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1641 switch (II->getIntrinsicID()) {
1642 default: assert(0 && "Case stmts out of sync!");
1643 case Intrinsic::x86_sse_sub_ss:
1644 case Intrinsic::x86_sse2_sub_sd:
1645 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1646 II->getName()), *II);
1648 case Intrinsic::x86_sse_mul_ss:
1649 case Intrinsic::x86_sse2_mul_sd:
1650 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1651 II->getName()), *II);
1656 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1658 InsertNewInstBefore(New, *II);
1659 AddSoonDeadInstToWorklist(*II, 0);
1664 // Output elements are undefined if both are undefined. Consider things
1665 // like undef&0. The result is known zero, not undef.
1666 UndefElts &= UndefElts2;
1672 return MadeChange ? I : 0;
1675 /// @returns true if the specified compare predicate is
1676 /// true when both operands are equal...
1677 /// @brief Determine if the icmp Predicate is true when both operands are equal
1678 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1679 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1680 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1681 pred == ICmpInst::ICMP_SLE;
1684 /// @returns true if the specified compare instruction is
1685 /// true when both operands are equal...
1686 /// @brief Determine if the ICmpInst returns true when both operands are equal
1687 static bool isTrueWhenEqual(ICmpInst &ICI) {
1688 return isTrueWhenEqual(ICI.getPredicate());
1691 /// AssociativeOpt - Perform an optimization on an associative operator. This
1692 /// function is designed to check a chain of associative operators for a
1693 /// potential to apply a certain optimization. Since the optimization may be
1694 /// applicable if the expression was reassociated, this checks the chain, then
1695 /// reassociates the expression as necessary to expose the optimization
1696 /// opportunity. This makes use of a special Functor, which must define
1697 /// 'shouldApply' and 'apply' methods.
1699 template<typename Functor>
1700 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1701 unsigned Opcode = Root.getOpcode();
1702 Value *LHS = Root.getOperand(0);
1704 // Quick check, see if the immediate LHS matches...
1705 if (F.shouldApply(LHS))
1706 return F.apply(Root);
1708 // Otherwise, if the LHS is not of the same opcode as the root, return.
1709 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1710 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1711 // Should we apply this transform to the RHS?
1712 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1714 // If not to the RHS, check to see if we should apply to the LHS...
1715 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1716 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1720 // If the functor wants to apply the optimization to the RHS of LHSI,
1721 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1723 BasicBlock *BB = Root.getParent();
1725 // Now all of the instructions are in the current basic block, go ahead
1726 // and perform the reassociation.
1727 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1729 // First move the selected RHS to the LHS of the root...
1730 Root.setOperand(0, LHSI->getOperand(1));
1732 // Make what used to be the LHS of the root be the user of the root...
1733 Value *ExtraOperand = TmpLHSI->getOperand(1);
1734 if (&Root == TmpLHSI) {
1735 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1738 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1739 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1740 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1741 BasicBlock::iterator ARI = &Root; ++ARI;
1742 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1745 // Now propagate the ExtraOperand down the chain of instructions until we
1747 while (TmpLHSI != LHSI) {
1748 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1749 // Move the instruction to immediately before the chain we are
1750 // constructing to avoid breaking dominance properties.
1751 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1752 BB->getInstList().insert(ARI, NextLHSI);
1755 Value *NextOp = NextLHSI->getOperand(1);
1756 NextLHSI->setOperand(1, ExtraOperand);
1758 ExtraOperand = NextOp;
1761 // Now that the instructions are reassociated, have the functor perform
1762 // the transformation...
1763 return F.apply(Root);
1766 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1772 // AddRHS - Implements: X + X --> X << 1
1775 AddRHS(Value *rhs) : RHS(rhs) {}
1776 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1777 Instruction *apply(BinaryOperator &Add) const {
1778 return BinaryOperator::createShl(Add.getOperand(0),
1779 ConstantInt::get(Add.getType(), 1));
1783 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1785 struct AddMaskingAnd {
1787 AddMaskingAnd(Constant *c) : C2(c) {}
1788 bool shouldApply(Value *LHS) const {
1790 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1791 ConstantExpr::getAnd(C1, C2)->isNullValue();
1793 Instruction *apply(BinaryOperator &Add) const {
1794 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1798 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1800 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1801 if (Constant *SOC = dyn_cast<Constant>(SO))
1802 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1804 return IC->InsertNewInstBefore(CastInst::create(
1805 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1808 // Figure out if the constant is the left or the right argument.
1809 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1810 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1812 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1814 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1815 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1818 Value *Op0 = SO, *Op1 = ConstOperand;
1820 std::swap(Op0, Op1);
1822 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1823 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1824 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1825 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1826 SO->getName()+".cmp");
1828 assert(0 && "Unknown binary instruction type!");
1831 return IC->InsertNewInstBefore(New, I);
1834 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1835 // constant as the other operand, try to fold the binary operator into the
1836 // select arguments. This also works for Cast instructions, which obviously do
1837 // not have a second operand.
1838 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1840 // Don't modify shared select instructions
1841 if (!SI->hasOneUse()) return 0;
1842 Value *TV = SI->getOperand(1);
1843 Value *FV = SI->getOperand(2);
1845 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1846 // Bool selects with constant operands can be folded to logical ops.
1847 if (SI->getType() == Type::Int1Ty) return 0;
1849 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1850 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1852 return new SelectInst(SI->getCondition(), SelectTrueVal,
1859 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1860 /// node as operand #0, see if we can fold the instruction into the PHI (which
1861 /// is only possible if all operands to the PHI are constants).
1862 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1863 PHINode *PN = cast<PHINode>(I.getOperand(0));
1864 unsigned NumPHIValues = PN->getNumIncomingValues();
1865 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1867 // Check to see if all of the operands of the PHI are constants. If there is
1868 // one non-constant value, remember the BB it is. If there is more than one
1869 // or if *it* is a PHI, bail out.
1870 BasicBlock *NonConstBB = 0;
1871 for (unsigned i = 0; i != NumPHIValues; ++i)
1872 if (!isa<Constant>(PN->getIncomingValue(i))) {
1873 if (NonConstBB) return 0; // More than one non-const value.
1874 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1875 NonConstBB = PN->getIncomingBlock(i);
1877 // If the incoming non-constant value is in I's block, we have an infinite
1879 if (NonConstBB == I.getParent())
1883 // If there is exactly one non-constant value, we can insert a copy of the
1884 // operation in that block. However, if this is a critical edge, we would be
1885 // inserting the computation one some other paths (e.g. inside a loop). Only
1886 // do this if the pred block is unconditionally branching into the phi block.
1888 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1889 if (!BI || !BI->isUnconditional()) return 0;
1892 // Okay, we can do the transformation: create the new PHI node.
1893 PHINode *NewPN = new PHINode(I.getType(), "");
1894 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1895 InsertNewInstBefore(NewPN, *PN);
1896 NewPN->takeName(PN);
1898 // Next, add all of the operands to the PHI.
1899 if (I.getNumOperands() == 2) {
1900 Constant *C = cast<Constant>(I.getOperand(1));
1901 for (unsigned i = 0; i != NumPHIValues; ++i) {
1903 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1904 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1905 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1907 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1909 assert(PN->getIncomingBlock(i) == NonConstBB);
1910 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1911 InV = BinaryOperator::create(BO->getOpcode(),
1912 PN->getIncomingValue(i), C, "phitmp",
1913 NonConstBB->getTerminator());
1914 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1915 InV = CmpInst::create(CI->getOpcode(),
1917 PN->getIncomingValue(i), C, "phitmp",
1918 NonConstBB->getTerminator());
1920 assert(0 && "Unknown binop!");
1922 AddToWorkList(cast<Instruction>(InV));
1924 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1927 CastInst *CI = cast<CastInst>(&I);
1928 const Type *RetTy = CI->getType();
1929 for (unsigned i = 0; i != NumPHIValues; ++i) {
1931 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1932 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1934 assert(PN->getIncomingBlock(i) == NonConstBB);
1935 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1936 I.getType(), "phitmp",
1937 NonConstBB->getTerminator());
1938 AddToWorkList(cast<Instruction>(InV));
1940 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1943 return ReplaceInstUsesWith(I, NewPN);
1947 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1948 /// value is never equal to -0.0.
1950 /// Note that this function will need to be revisited when we support nondefault
1953 static bool CannotBeNegativeZero(const Value *V) {
1954 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1955 return !CFP->getValueAPF().isNegZero();
1957 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1958 if (const Instruction *I = dyn_cast<Instruction>(V)) {
1959 if (I->getOpcode() == Instruction::Add &&
1960 isa<ConstantFP>(I->getOperand(1)) &&
1961 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1964 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1965 if (II->getIntrinsicID() == Intrinsic::sqrt)
1966 return CannotBeNegativeZero(II->getOperand(1));
1968 if (const CallInst *CI = dyn_cast<CallInst>(I))
1969 if (const Function *F = CI->getCalledFunction()) {
1970 if (F->isDeclaration()) {
1971 switch (F->getNameLen()) {
1972 case 3: // abs(x) != -0.0
1973 if (!strcmp(F->getNameStart(), "abs")) return true;
1975 case 4: // abs[lf](x) != -0.0
1976 if (!strcmp(F->getNameStart(), "absf")) return true;
1977 if (!strcmp(F->getNameStart(), "absl")) return true;
1988 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1989 bool Changed = SimplifyCommutative(I);
1990 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1992 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1993 // X + undef -> undef
1994 if (isa<UndefValue>(RHS))
1995 return ReplaceInstUsesWith(I, RHS);
1998 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1999 if (RHSC->isNullValue())
2000 return ReplaceInstUsesWith(I, LHS);
2001 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2002 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2003 (I.getType())->getValueAPF()))
2004 return ReplaceInstUsesWith(I, LHS);
2007 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2008 // X + (signbit) --> X ^ signbit
2009 const APInt& Val = CI->getValue();
2010 uint32_t BitWidth = Val.getBitWidth();
2011 if (Val == APInt::getSignBit(BitWidth))
2012 return BinaryOperator::createXor(LHS, RHS);
2014 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2015 // (X & 254)+1 -> (X&254)|1
2016 if (!isa<VectorType>(I.getType())) {
2017 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2018 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2019 KnownZero, KnownOne))
2024 if (isa<PHINode>(LHS))
2025 if (Instruction *NV = FoldOpIntoPhi(I))
2028 ConstantInt *XorRHS = 0;
2030 if (isa<ConstantInt>(RHSC) &&
2031 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2032 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2033 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2035 uint32_t Size = TySizeBits / 2;
2036 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2037 APInt CFF80Val(-C0080Val);
2039 if (TySizeBits > Size) {
2040 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2041 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2042 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2043 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2044 // This is a sign extend if the top bits are known zero.
2045 if (!MaskedValueIsZero(XorLHS,
2046 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2047 Size = 0; // Not a sign ext, but can't be any others either.
2052 C0080Val = APIntOps::lshr(C0080Val, Size);
2053 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2054 } while (Size >= 1);
2056 // FIXME: This shouldn't be necessary. When the backends can handle types
2057 // with funny bit widths then this whole cascade of if statements should
2058 // be removed. It is just here to get the size of the "middle" type back
2059 // up to something that the back ends can handle.
2060 const Type *MiddleType = 0;
2063 case 32: MiddleType = Type::Int32Ty; break;
2064 case 16: MiddleType = Type::Int16Ty; break;
2065 case 8: MiddleType = Type::Int8Ty; break;
2068 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2069 InsertNewInstBefore(NewTrunc, I);
2070 return new SExtInst(NewTrunc, I.getType(), I.getName());
2076 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2077 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2079 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2080 if (RHSI->getOpcode() == Instruction::Sub)
2081 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2082 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2084 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2085 if (LHSI->getOpcode() == Instruction::Sub)
2086 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2087 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2092 if (Value *V = dyn_castNegVal(LHS))
2093 return BinaryOperator::createSub(RHS, V);
2096 if (!isa<Constant>(RHS))
2097 if (Value *V = dyn_castNegVal(RHS))
2098 return BinaryOperator::createSub(LHS, V);
2102 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2103 if (X == RHS) // X*C + X --> X * (C+1)
2104 return BinaryOperator::createMul(RHS, AddOne(C2));
2106 // X*C1 + X*C2 --> X * (C1+C2)
2108 if (X == dyn_castFoldableMul(RHS, C1))
2109 return BinaryOperator::createMul(X, Add(C1, C2));
2112 // X + X*C --> X * (C+1)
2113 if (dyn_castFoldableMul(RHS, C2) == LHS)
2114 return BinaryOperator::createMul(LHS, AddOne(C2));
2116 // X + ~X --> -1 since ~X = -X-1
2117 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2118 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2121 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2122 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2123 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2126 // W*X + Y*Z --> W * (X+Z) iff W == Y
2127 if (I.getType()->isIntOrIntVector()) {
2128 Value *W, *X, *Y, *Z;
2129 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2130 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2134 } else if (Y == X) {
2136 } else if (X == Z) {
2143 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2144 LHS->getName()), I);
2145 return BinaryOperator::createMul(W, NewAdd);
2150 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2152 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2153 return BinaryOperator::createSub(SubOne(CRHS), X);
2155 // (X & FF00) + xx00 -> (X+xx00) & FF00
2156 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2157 Constant *Anded = And(CRHS, C2);
2158 if (Anded == CRHS) {
2159 // See if all bits from the first bit set in the Add RHS up are included
2160 // in the mask. First, get the rightmost bit.
2161 const APInt& AddRHSV = CRHS->getValue();
2163 // Form a mask of all bits from the lowest bit added through the top.
2164 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2166 // See if the and mask includes all of these bits.
2167 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2169 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2170 // Okay, the xform is safe. Insert the new add pronto.
2171 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2172 LHS->getName()), I);
2173 return BinaryOperator::createAnd(NewAdd, C2);
2178 // Try to fold constant add into select arguments.
2179 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2180 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2184 // add (cast *A to intptrtype) B ->
2185 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2187 CastInst *CI = dyn_cast<CastInst>(LHS);
2190 CI = dyn_cast<CastInst>(RHS);
2193 if (CI && CI->getType()->isSized() &&
2194 (CI->getType()->getPrimitiveSizeInBits() ==
2195 TD->getIntPtrType()->getPrimitiveSizeInBits())
2196 && isa<PointerType>(CI->getOperand(0)->getType())) {
2198 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2199 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2200 PointerType::get(Type::Int8Ty, AS), I);
2201 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2202 return new PtrToIntInst(I2, CI->getType());
2206 // add (select X 0 (sub n A)) A --> select X A n
2208 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2211 SI = dyn_cast<SelectInst>(RHS);
2214 if (SI && SI->hasOneUse()) {
2215 Value *TV = SI->getTrueValue();
2216 Value *FV = SI->getFalseValue();
2219 // Can we fold the add into the argument of the select?
2220 // We check both true and false select arguments for a matching subtract.
2221 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2222 A == Other) // Fold the add into the true select value.
2223 return new SelectInst(SI->getCondition(), N, A);
2224 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2225 A == Other) // Fold the add into the false select value.
2226 return new SelectInst(SI->getCondition(), A, N);
2230 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2231 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2232 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2233 return ReplaceInstUsesWith(I, LHS);
2235 return Changed ? &I : 0;
2238 // isSignBit - Return true if the value represented by the constant only has the
2239 // highest order bit set.
2240 static bool isSignBit(ConstantInt *CI) {
2241 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2242 return CI->getValue() == APInt::getSignBit(NumBits);
2245 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2246 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2248 if (Op0 == Op1) // sub X, X -> 0
2249 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2251 // If this is a 'B = x-(-A)', change to B = x+A...
2252 if (Value *V = dyn_castNegVal(Op1))
2253 return BinaryOperator::createAdd(Op0, V);
2255 if (isa<UndefValue>(Op0))
2256 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2257 if (isa<UndefValue>(Op1))
2258 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2260 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2261 // Replace (-1 - A) with (~A)...
2262 if (C->isAllOnesValue())
2263 return BinaryOperator::createNot(Op1);
2265 // C - ~X == X + (1+C)
2267 if (match(Op1, m_Not(m_Value(X))))
2268 return BinaryOperator::createAdd(X, AddOne(C));
2270 // -(X >>u 31) -> (X >>s 31)
2271 // -(X >>s 31) -> (X >>u 31)
2273 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2274 if (SI->getOpcode() == Instruction::LShr) {
2275 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2276 // Check to see if we are shifting out everything but the sign bit.
2277 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2278 SI->getType()->getPrimitiveSizeInBits()-1) {
2279 // Ok, the transformation is safe. Insert AShr.
2280 return BinaryOperator::create(Instruction::AShr,
2281 SI->getOperand(0), CU, SI->getName());
2285 else if (SI->getOpcode() == Instruction::AShr) {
2286 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2287 // Check to see if we are shifting out everything but the sign bit.
2288 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2289 SI->getType()->getPrimitiveSizeInBits()-1) {
2290 // Ok, the transformation is safe. Insert LShr.
2291 return BinaryOperator::createLShr(
2292 SI->getOperand(0), CU, SI->getName());
2298 // Try to fold constant sub into select arguments.
2299 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2300 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2303 if (isa<PHINode>(Op0))
2304 if (Instruction *NV = FoldOpIntoPhi(I))
2308 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2309 if (Op1I->getOpcode() == Instruction::Add &&
2310 !Op0->getType()->isFPOrFPVector()) {
2311 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2312 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2313 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2314 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2315 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2316 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2317 // C1-(X+C2) --> (C1-C2)-X
2318 return BinaryOperator::createSub(Subtract(CI1, CI2),
2319 Op1I->getOperand(0));
2323 if (Op1I->hasOneUse()) {
2324 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2325 // is not used by anyone else...
2327 if (Op1I->getOpcode() == Instruction::Sub &&
2328 !Op1I->getType()->isFPOrFPVector()) {
2329 // Swap the two operands of the subexpr...
2330 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2331 Op1I->setOperand(0, IIOp1);
2332 Op1I->setOperand(1, IIOp0);
2334 // Create the new top level add instruction...
2335 return BinaryOperator::createAdd(Op0, Op1);
2338 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2340 if (Op1I->getOpcode() == Instruction::And &&
2341 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2342 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2345 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2346 return BinaryOperator::createAnd(Op0, NewNot);
2349 // 0 - (X sdiv C) -> (X sdiv -C)
2350 if (Op1I->getOpcode() == Instruction::SDiv)
2351 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2353 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2354 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2355 ConstantExpr::getNeg(DivRHS));
2357 // X - X*C --> X * (1-C)
2358 ConstantInt *C2 = 0;
2359 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2360 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2361 return BinaryOperator::createMul(Op0, CP1);
2364 // X - ((X / Y) * Y) --> X % Y
2365 if (Op1I->getOpcode() == Instruction::Mul)
2366 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2367 if (Op0 == I->getOperand(0) &&
2368 Op1I->getOperand(1) == I->getOperand(1)) {
2369 if (I->getOpcode() == Instruction::SDiv)
2370 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2371 if (I->getOpcode() == Instruction::UDiv)
2372 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2377 if (!Op0->getType()->isFPOrFPVector())
2378 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2379 if (Op0I->getOpcode() == Instruction::Add) {
2380 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2381 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2382 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2383 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2384 } else if (Op0I->getOpcode() == Instruction::Sub) {
2385 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2386 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2390 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2391 if (X == Op1) // X*C - X --> X * (C-1)
2392 return BinaryOperator::createMul(Op1, SubOne(C1));
2394 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2395 if (X == dyn_castFoldableMul(Op1, C2))
2396 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2401 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2402 /// comparison only checks the sign bit. If it only checks the sign bit, set
2403 /// TrueIfSigned if the result of the comparison is true when the input value is
2405 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2406 bool &TrueIfSigned) {
2408 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2409 TrueIfSigned = true;
2410 return RHS->isZero();
2411 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2412 TrueIfSigned = true;
2413 return RHS->isAllOnesValue();
2414 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2415 TrueIfSigned = false;
2416 return RHS->isAllOnesValue();
2417 case ICmpInst::ICMP_UGT:
2418 // True if LHS u> RHS and RHS == high-bit-mask - 1
2419 TrueIfSigned = true;
2420 return RHS->getValue() ==
2421 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2422 case ICmpInst::ICMP_UGE:
2423 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2424 TrueIfSigned = true;
2425 return RHS->getValue() ==
2426 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2432 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2433 bool Changed = SimplifyCommutative(I);
2434 Value *Op0 = I.getOperand(0);
2436 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2437 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2439 // Simplify mul instructions with a constant RHS...
2440 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2441 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2443 // ((X << C1)*C2) == (X * (C2 << C1))
2444 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2445 if (SI->getOpcode() == Instruction::Shl)
2446 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2447 return BinaryOperator::createMul(SI->getOperand(0),
2448 ConstantExpr::getShl(CI, ShOp));
2451 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2452 if (CI->equalsInt(1)) // X * 1 == X
2453 return ReplaceInstUsesWith(I, Op0);
2454 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2455 return BinaryOperator::createNeg(Op0, I.getName());
2457 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2458 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2459 return BinaryOperator::createShl(Op0,
2460 ConstantInt::get(Op0->getType(), Val.logBase2()));
2462 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2463 if (Op1F->isNullValue())
2464 return ReplaceInstUsesWith(I, Op1);
2466 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2467 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2468 // We need a better interface for long double here.
2469 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2470 if (Op1F->isExactlyValue(1.0))
2471 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2474 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2475 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2476 isa<ConstantInt>(Op0I->getOperand(1))) {
2477 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2478 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2480 InsertNewInstBefore(Add, I);
2481 Value *C1C2 = ConstantExpr::getMul(Op1,
2482 cast<Constant>(Op0I->getOperand(1)));
2483 return BinaryOperator::createAdd(Add, C1C2);
2487 // Try to fold constant mul into select arguments.
2488 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2489 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2492 if (isa<PHINode>(Op0))
2493 if (Instruction *NV = FoldOpIntoPhi(I))
2497 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2498 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2499 return BinaryOperator::createMul(Op0v, Op1v);
2501 // If one of the operands of the multiply is a cast from a boolean value, then
2502 // we know the bool is either zero or one, so this is a 'masking' multiply.
2503 // See if we can simplify things based on how the boolean was originally
2505 CastInst *BoolCast = 0;
2506 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2507 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2510 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2511 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2514 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2515 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2516 const Type *SCOpTy = SCIOp0->getType();
2519 // If the icmp is true iff the sign bit of X is set, then convert this
2520 // multiply into a shift/and combination.
2521 if (isa<ConstantInt>(SCIOp1) &&
2522 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2524 // Shift the X value right to turn it into "all signbits".
2525 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2526 SCOpTy->getPrimitiveSizeInBits()-1);
2528 InsertNewInstBefore(
2529 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2530 BoolCast->getOperand(0)->getName()+
2533 // If the multiply type is not the same as the source type, sign extend
2534 // or truncate to the multiply type.
2535 if (I.getType() != V->getType()) {
2536 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2537 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2538 Instruction::CastOps opcode =
2539 (SrcBits == DstBits ? Instruction::BitCast :
2540 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2541 V = InsertCastBefore(opcode, V, I.getType(), I);
2544 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2545 return BinaryOperator::createAnd(V, OtherOp);
2550 return Changed ? &I : 0;
2553 /// This function implements the transforms on div instructions that work
2554 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2555 /// used by the visitors to those instructions.
2556 /// @brief Transforms common to all three div instructions
2557 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2558 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2561 if (isa<UndefValue>(Op0))
2562 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2564 // X / undef -> undef
2565 if (isa<UndefValue>(Op1))
2566 return ReplaceInstUsesWith(I, Op1);
2568 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2569 // This does not apply for fdiv.
2570 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2571 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2572 // the same basic block, then we replace the select with Y, and the
2573 // condition of the select with false (if the cond value is in the same BB).
2574 // If the select has uses other than the div, this allows them to be
2575 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2576 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2577 if (ST->isNullValue()) {
2578 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2579 if (CondI && CondI->getParent() == I.getParent())
2580 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2581 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2582 I.setOperand(1, SI->getOperand(2));
2584 UpdateValueUsesWith(SI, SI->getOperand(2));
2588 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2589 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2590 if (ST->isNullValue()) {
2591 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2592 if (CondI && CondI->getParent() == I.getParent())
2593 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2594 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2595 I.setOperand(1, SI->getOperand(1));
2597 UpdateValueUsesWith(SI, SI->getOperand(1));
2605 /// This function implements the transforms common to both integer division
2606 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2607 /// division instructions.
2608 /// @brief Common integer divide transforms
2609 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2610 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2612 if (Instruction *Common = commonDivTransforms(I))
2615 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2617 if (RHS->equalsInt(1))
2618 return ReplaceInstUsesWith(I, Op0);
2620 // (X / C1) / C2 -> X / (C1*C2)
2621 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2622 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2623 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2624 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2625 Multiply(RHS, LHSRHS));
2628 if (!RHS->isZero()) { // avoid X udiv 0
2629 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2630 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2632 if (isa<PHINode>(Op0))
2633 if (Instruction *NV = FoldOpIntoPhi(I))
2638 // 0 / X == 0, we don't need to preserve faults!
2639 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2640 if (LHS->equalsInt(0))
2641 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2646 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2647 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2649 // Handle the integer div common cases
2650 if (Instruction *Common = commonIDivTransforms(I))
2653 // X udiv C^2 -> X >> C
2654 // Check to see if this is an unsigned division with an exact power of 2,
2655 // if so, convert to a right shift.
2656 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2657 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2658 return BinaryOperator::createLShr(Op0,
2659 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2662 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2663 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2664 if (RHSI->getOpcode() == Instruction::Shl &&
2665 isa<ConstantInt>(RHSI->getOperand(0))) {
2666 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2667 if (C1.isPowerOf2()) {
2668 Value *N = RHSI->getOperand(1);
2669 const Type *NTy = N->getType();
2670 if (uint32_t C2 = C1.logBase2()) {
2671 Constant *C2V = ConstantInt::get(NTy, C2);
2672 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2674 return BinaryOperator::createLShr(Op0, N);
2679 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2680 // where C1&C2 are powers of two.
2681 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2682 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2683 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2684 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2685 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2686 // Compute the shift amounts
2687 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2688 // Construct the "on true" case of the select
2689 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2690 Instruction *TSI = BinaryOperator::createLShr(
2691 Op0, TC, SI->getName()+".t");
2692 TSI = InsertNewInstBefore(TSI, I);
2694 // Construct the "on false" case of the select
2695 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2696 Instruction *FSI = BinaryOperator::createLShr(
2697 Op0, FC, SI->getName()+".f");
2698 FSI = InsertNewInstBefore(FSI, I);
2700 // construct the select instruction and return it.
2701 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2707 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2708 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2710 // Handle the integer div common cases
2711 if (Instruction *Common = commonIDivTransforms(I))
2714 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2716 if (RHS->isAllOnesValue())
2717 return BinaryOperator::createNeg(Op0);
2720 if (Value *LHSNeg = dyn_castNegVal(Op0))
2721 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2724 // If the sign bits of both operands are zero (i.e. we can prove they are
2725 // unsigned inputs), turn this into a udiv.
2726 if (I.getType()->isInteger()) {
2727 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2728 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2729 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2730 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2737 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2738 return commonDivTransforms(I);
2741 /// GetFactor - If we can prove that the specified value is at least a multiple
2742 /// of some factor, return that factor.
2743 static Constant *GetFactor(Value *V) {
2744 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2747 // Unless we can be tricky, we know this is a multiple of 1.
2748 Constant *Result = ConstantInt::get(V->getType(), 1);
2750 Instruction *I = dyn_cast<Instruction>(V);
2751 if (!I) return Result;
2753 if (I->getOpcode() == Instruction::Mul) {
2754 // Handle multiplies by a constant, etc.
2755 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2756 GetFactor(I->getOperand(1)));
2757 } else if (I->getOpcode() == Instruction::Shl) {
2758 // (X<<C) -> X * (1 << C)
2759 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2760 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2761 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2763 } else if (I->getOpcode() == Instruction::And) {
2764 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2765 // X & 0xFFF0 is known to be a multiple of 16.
2766 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2767 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2768 return ConstantExpr::getShl(Result,
2769 ConstantInt::get(Result->getType(), Zeros));
2771 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2772 // Only handle int->int casts.
2773 if (!CI->isIntegerCast())
2775 Value *Op = CI->getOperand(0);
2776 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2781 /// This function implements the transforms on rem instructions that work
2782 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2783 /// is used by the visitors to those instructions.
2784 /// @brief Transforms common to all three rem instructions
2785 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2786 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2788 // 0 % X == 0, we don't need to preserve faults!
2789 if (Constant *LHS = dyn_cast<Constant>(Op0))
2790 if (LHS->isNullValue())
2791 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2793 if (isa<UndefValue>(Op0)) // undef % X -> 0
2794 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2795 if (isa<UndefValue>(Op1))
2796 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2798 // Handle cases involving: rem X, (select Cond, Y, Z)
2799 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2800 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2801 // the same basic block, then we replace the select with Y, and the
2802 // condition of the select with false (if the cond value is in the same
2803 // BB). If the select has uses other than the div, this allows them to be
2805 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2806 if (ST->isNullValue()) {
2807 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2808 if (CondI && CondI->getParent() == I.getParent())
2809 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2810 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2811 I.setOperand(1, SI->getOperand(2));
2813 UpdateValueUsesWith(SI, SI->getOperand(2));
2816 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2817 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2818 if (ST->isNullValue()) {
2819 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2820 if (CondI && CondI->getParent() == I.getParent())
2821 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2822 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2823 I.setOperand(1, SI->getOperand(1));
2825 UpdateValueUsesWith(SI, SI->getOperand(1));
2833 /// This function implements the transforms common to both integer remainder
2834 /// instructions (urem and srem). It is called by the visitors to those integer
2835 /// remainder instructions.
2836 /// @brief Common integer remainder transforms
2837 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2838 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2840 if (Instruction *common = commonRemTransforms(I))
2843 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2844 // X % 0 == undef, we don't need to preserve faults!
2845 if (RHS->equalsInt(0))
2846 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2848 if (RHS->equalsInt(1)) // X % 1 == 0
2849 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2851 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2852 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2853 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2855 } else if (isa<PHINode>(Op0I)) {
2856 if (Instruction *NV = FoldOpIntoPhi(I))
2859 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2860 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2861 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2868 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2869 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2871 if (Instruction *common = commonIRemTransforms(I))
2874 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2875 // X urem C^2 -> X and C
2876 // Check to see if this is an unsigned remainder with an exact power of 2,
2877 // if so, convert to a bitwise and.
2878 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2879 if (C->getValue().isPowerOf2())
2880 return BinaryOperator::createAnd(Op0, SubOne(C));
2883 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2884 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2885 if (RHSI->getOpcode() == Instruction::Shl &&
2886 isa<ConstantInt>(RHSI->getOperand(0))) {
2887 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2888 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2889 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2891 return BinaryOperator::createAnd(Op0, Add);
2896 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2897 // where C1&C2 are powers of two.
2898 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2899 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2900 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2901 // STO == 0 and SFO == 0 handled above.
2902 if ((STO->getValue().isPowerOf2()) &&
2903 (SFO->getValue().isPowerOf2())) {
2904 Value *TrueAnd = InsertNewInstBefore(
2905 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2906 Value *FalseAnd = InsertNewInstBefore(
2907 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2908 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2916 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2917 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2919 // Handle the integer rem common cases
2920 if (Instruction *common = commonIRemTransforms(I))
2923 if (Value *RHSNeg = dyn_castNegVal(Op1))
2924 if (!isa<ConstantInt>(RHSNeg) ||
2925 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2927 AddUsesToWorkList(I);
2928 I.setOperand(1, RHSNeg);
2932 // If the sign bits of both operands are zero (i.e. we can prove they are
2933 // unsigned inputs), turn this into a urem.
2934 if (I.getType()->isInteger()) {
2935 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2936 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2937 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2938 return BinaryOperator::createURem(Op0, Op1, I.getName());
2945 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2946 return commonRemTransforms(I);
2949 // isMaxValueMinusOne - return true if this is Max-1
2950 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2951 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2953 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2954 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2957 // isMinValuePlusOne - return true if this is Min+1
2958 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2960 return C->getValue() == 1; // unsigned
2962 // Calculate 1111111111000000000000
2963 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2964 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2967 // isOneBitSet - Return true if there is exactly one bit set in the specified
2969 static bool isOneBitSet(const ConstantInt *CI) {
2970 return CI->getValue().isPowerOf2();
2973 // isHighOnes - Return true if the constant is of the form 1+0+.
2974 // This is the same as lowones(~X).
2975 static bool isHighOnes(const ConstantInt *CI) {
2976 return (~CI->getValue() + 1).isPowerOf2();
2979 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2980 /// are carefully arranged to allow folding of expressions such as:
2982 /// (A < B) | (A > B) --> (A != B)
2984 /// Note that this is only valid if the first and second predicates have the
2985 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2987 /// Three bits are used to represent the condition, as follows:
2992 /// <=> Value Definition
2993 /// 000 0 Always false
3000 /// 111 7 Always true
3002 static unsigned getICmpCode(const ICmpInst *ICI) {
3003 switch (ICI->getPredicate()) {
3005 case ICmpInst::ICMP_UGT: return 1; // 001
3006 case ICmpInst::ICMP_SGT: return 1; // 001
3007 case ICmpInst::ICMP_EQ: return 2; // 010
3008 case ICmpInst::ICMP_UGE: return 3; // 011
3009 case ICmpInst::ICMP_SGE: return 3; // 011
3010 case ICmpInst::ICMP_ULT: return 4; // 100
3011 case ICmpInst::ICMP_SLT: return 4; // 100
3012 case ICmpInst::ICMP_NE: return 5; // 101
3013 case ICmpInst::ICMP_ULE: return 6; // 110
3014 case ICmpInst::ICMP_SLE: return 6; // 110
3017 assert(0 && "Invalid ICmp predicate!");
3022 /// getICmpValue - This is the complement of getICmpCode, which turns an
3023 /// opcode and two operands into either a constant true or false, or a brand
3024 /// new ICmp instruction. The sign is passed in to determine which kind
3025 /// of predicate to use in new icmp instructions.
3026 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3028 default: assert(0 && "Illegal ICmp code!");
3029 case 0: return ConstantInt::getFalse();
3032 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3034 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3035 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3038 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3040 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3043 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3045 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3046 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3049 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3051 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3052 case 7: return ConstantInt::getTrue();
3056 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3057 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3058 (ICmpInst::isSignedPredicate(p1) &&
3059 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3060 (ICmpInst::isSignedPredicate(p2) &&
3061 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3065 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3066 struct FoldICmpLogical {
3069 ICmpInst::Predicate pred;
3070 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3071 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3072 pred(ICI->getPredicate()) {}
3073 bool shouldApply(Value *V) const {
3074 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3075 if (PredicatesFoldable(pred, ICI->getPredicate()))
3076 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3077 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3080 Instruction *apply(Instruction &Log) const {
3081 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3082 if (ICI->getOperand(0) != LHS) {
3083 assert(ICI->getOperand(1) == LHS);
3084 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3087 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3088 unsigned LHSCode = getICmpCode(ICI);
3089 unsigned RHSCode = getICmpCode(RHSICI);
3091 switch (Log.getOpcode()) {
3092 case Instruction::And: Code = LHSCode & RHSCode; break;
3093 case Instruction::Or: Code = LHSCode | RHSCode; break;
3094 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3095 default: assert(0 && "Illegal logical opcode!"); return 0;
3098 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3099 ICmpInst::isSignedPredicate(ICI->getPredicate());
3101 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3102 if (Instruction *I = dyn_cast<Instruction>(RV))
3104 // Otherwise, it's a constant boolean value...
3105 return IC.ReplaceInstUsesWith(Log, RV);
3108 } // end anonymous namespace
3110 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3111 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3112 // guaranteed to be a binary operator.
3113 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3115 ConstantInt *AndRHS,
3116 BinaryOperator &TheAnd) {
3117 Value *X = Op->getOperand(0);
3118 Constant *Together = 0;
3120 Together = And(AndRHS, OpRHS);
3122 switch (Op->getOpcode()) {
3123 case Instruction::Xor:
3124 if (Op->hasOneUse()) {
3125 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3126 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3127 InsertNewInstBefore(And, TheAnd);
3129 return BinaryOperator::createXor(And, Together);
3132 case Instruction::Or:
3133 if (Together == AndRHS) // (X | C) & C --> C
3134 return ReplaceInstUsesWith(TheAnd, AndRHS);
3136 if (Op->hasOneUse() && Together != OpRHS) {
3137 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3138 Instruction *Or = BinaryOperator::createOr(X, Together);
3139 InsertNewInstBefore(Or, TheAnd);
3141 return BinaryOperator::createAnd(Or, AndRHS);
3144 case Instruction::Add:
3145 if (Op->hasOneUse()) {
3146 // Adding a one to a single bit bit-field should be turned into an XOR
3147 // of the bit. First thing to check is to see if this AND is with a
3148 // single bit constant.
3149 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3151 // If there is only one bit set...
3152 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3153 // Ok, at this point, we know that we are masking the result of the
3154 // ADD down to exactly one bit. If the constant we are adding has
3155 // no bits set below this bit, then we can eliminate the ADD.
3156 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3158 // Check to see if any bits below the one bit set in AndRHSV are set.
3159 if ((AddRHS & (AndRHSV-1)) == 0) {
3160 // If not, the only thing that can effect the output of the AND is
3161 // the bit specified by AndRHSV. If that bit is set, the effect of
3162 // the XOR is to toggle the bit. If it is clear, then the ADD has
3164 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3165 TheAnd.setOperand(0, X);
3168 // Pull the XOR out of the AND.
3169 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3170 InsertNewInstBefore(NewAnd, TheAnd);
3171 NewAnd->takeName(Op);
3172 return BinaryOperator::createXor(NewAnd, AndRHS);
3179 case Instruction::Shl: {
3180 // We know that the AND will not produce any of the bits shifted in, so if
3181 // the anded constant includes them, clear them now!
3183 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3184 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3185 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3186 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3188 if (CI->getValue() == ShlMask) {
3189 // Masking out bits that the shift already masks
3190 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3191 } else if (CI != AndRHS) { // Reducing bits set in and.
3192 TheAnd.setOperand(1, CI);
3197 case Instruction::LShr:
3199 // We know that the AND will not produce any of the bits shifted in, so if
3200 // the anded constant includes them, clear them now! This only applies to
3201 // unsigned shifts, because a signed shr may bring in set bits!
3203 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3204 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3205 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3206 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3208 if (CI->getValue() == ShrMask) {
3209 // Masking out bits that the shift already masks.
3210 return ReplaceInstUsesWith(TheAnd, Op);
3211 } else if (CI != AndRHS) {
3212 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3217 case Instruction::AShr:
3219 // See if this is shifting in some sign extension, then masking it out
3221 if (Op->hasOneUse()) {
3222 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3223 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3224 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3225 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3226 if (C == AndRHS) { // Masking out bits shifted in.
3227 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3228 // Make the argument unsigned.
3229 Value *ShVal = Op->getOperand(0);
3230 ShVal = InsertNewInstBefore(
3231 BinaryOperator::createLShr(ShVal, OpRHS,
3232 Op->getName()), TheAnd);
3233 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3242 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3243 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3244 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3245 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3246 /// insert new instructions.
3247 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3248 bool isSigned, bool Inside,
3250 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3251 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3252 "Lo is not <= Hi in range emission code!");
3255 if (Lo == Hi) // Trivially false.
3256 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3258 // V >= Min && V < Hi --> V < Hi
3259 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3260 ICmpInst::Predicate pred = (isSigned ?
3261 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3262 return new ICmpInst(pred, V, Hi);
3265 // Emit V-Lo <u Hi-Lo
3266 Constant *NegLo = ConstantExpr::getNeg(Lo);
3267 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3268 InsertNewInstBefore(Add, IB);
3269 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3270 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3273 if (Lo == Hi) // Trivially true.
3274 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3276 // V < Min || V >= Hi -> V > Hi-1
3277 Hi = SubOne(cast<ConstantInt>(Hi));
3278 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3279 ICmpInst::Predicate pred = (isSigned ?
3280 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3281 return new ICmpInst(pred, V, Hi);
3284 // Emit V-Lo >u Hi-1-Lo
3285 // Note that Hi has already had one subtracted from it, above.
3286 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3287 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3288 InsertNewInstBefore(Add, IB);
3289 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3290 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3293 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3294 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3295 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3296 // not, since all 1s are not contiguous.
3297 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3298 const APInt& V = Val->getValue();
3299 uint32_t BitWidth = Val->getType()->getBitWidth();
3300 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3302 // look for the first zero bit after the run of ones
3303 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3304 // look for the first non-zero bit
3305 ME = V.getActiveBits();
3309 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3310 /// where isSub determines whether the operator is a sub. If we can fold one of
3311 /// the following xforms:
3313 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3314 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3315 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3317 /// return (A +/- B).
3319 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3320 ConstantInt *Mask, bool isSub,
3322 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3323 if (!LHSI || LHSI->getNumOperands() != 2 ||
3324 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3326 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3328 switch (LHSI->getOpcode()) {
3330 case Instruction::And:
3331 if (And(N, Mask) == Mask) {
3332 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3333 if ((Mask->getValue().countLeadingZeros() +
3334 Mask->getValue().countPopulation()) ==
3335 Mask->getValue().getBitWidth())
3338 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3339 // part, we don't need any explicit masks to take them out of A. If that
3340 // is all N is, ignore it.
3341 uint32_t MB = 0, ME = 0;
3342 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3343 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3344 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3345 if (MaskedValueIsZero(RHS, Mask))
3350 case Instruction::Or:
3351 case Instruction::Xor:
3352 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3353 if ((Mask->getValue().countLeadingZeros() +
3354 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3355 && And(N, Mask)->isZero())
3362 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3364 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3365 return InsertNewInstBefore(New, I);
3368 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3369 bool Changed = SimplifyCommutative(I);
3370 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3372 if (isa<UndefValue>(Op1)) // X & undef -> 0
3373 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3377 return ReplaceInstUsesWith(I, Op1);
3379 // See if we can simplify any instructions used by the instruction whose sole
3380 // purpose is to compute bits we don't care about.
3381 if (!isa<VectorType>(I.getType())) {
3382 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3383 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3384 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3385 KnownZero, KnownOne))
3388 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3389 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3390 return ReplaceInstUsesWith(I, I.getOperand(0));
3391 } else if (isa<ConstantAggregateZero>(Op1)) {
3392 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3396 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3397 const APInt& AndRHSMask = AndRHS->getValue();
3398 APInt NotAndRHS(~AndRHSMask);
3400 // Optimize a variety of ((val OP C1) & C2) combinations...
3401 if (isa<BinaryOperator>(Op0)) {
3402 Instruction *Op0I = cast<Instruction>(Op0);
3403 Value *Op0LHS = Op0I->getOperand(0);
3404 Value *Op0RHS = Op0I->getOperand(1);
3405 switch (Op0I->getOpcode()) {
3406 case Instruction::Xor:
3407 case Instruction::Or:
3408 // If the mask is only needed on one incoming arm, push it up.
3409 if (Op0I->hasOneUse()) {
3410 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3411 // Not masking anything out for the LHS, move to RHS.
3412 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3413 Op0RHS->getName()+".masked");
3414 InsertNewInstBefore(NewRHS, I);
3415 return BinaryOperator::create(
3416 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3418 if (!isa<Constant>(Op0RHS) &&
3419 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3420 // Not masking anything out for the RHS, move to LHS.
3421 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3422 Op0LHS->getName()+".masked");
3423 InsertNewInstBefore(NewLHS, I);
3424 return BinaryOperator::create(
3425 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3430 case Instruction::Add:
3431 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3432 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3433 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3434 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3435 return BinaryOperator::createAnd(V, AndRHS);
3436 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3437 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3440 case Instruction::Sub:
3441 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3442 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3443 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3444 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3445 return BinaryOperator::createAnd(V, AndRHS);
3449 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3450 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3452 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3453 // If this is an integer truncation or change from signed-to-unsigned, and
3454 // if the source is an and/or with immediate, transform it. This
3455 // frequently occurs for bitfield accesses.
3456 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3457 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3458 CastOp->getNumOperands() == 2)
3459 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3460 if (CastOp->getOpcode() == Instruction::And) {
3461 // Change: and (cast (and X, C1) to T), C2
3462 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3463 // This will fold the two constants together, which may allow
3464 // other simplifications.
3465 Instruction *NewCast = CastInst::createTruncOrBitCast(
3466 CastOp->getOperand(0), I.getType(),
3467 CastOp->getName()+".shrunk");
3468 NewCast = InsertNewInstBefore(NewCast, I);
3469 // trunc_or_bitcast(C1)&C2
3470 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3471 C3 = ConstantExpr::getAnd(C3, AndRHS);
3472 return BinaryOperator::createAnd(NewCast, C3);
3473 } else if (CastOp->getOpcode() == Instruction::Or) {
3474 // Change: and (cast (or X, C1) to T), C2
3475 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3476 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3477 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3478 return ReplaceInstUsesWith(I, AndRHS);
3483 // Try to fold constant and into select arguments.
3484 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3485 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3487 if (isa<PHINode>(Op0))
3488 if (Instruction *NV = FoldOpIntoPhi(I))
3492 Value *Op0NotVal = dyn_castNotVal(Op0);
3493 Value *Op1NotVal = dyn_castNotVal(Op1);
3495 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3496 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3498 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3499 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3500 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3501 I.getName()+".demorgan");
3502 InsertNewInstBefore(Or, I);
3503 return BinaryOperator::createNot(Or);
3507 Value *A = 0, *B = 0, *C = 0, *D = 0;
3508 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3509 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3510 return ReplaceInstUsesWith(I, Op1);
3512 // (A|B) & ~(A&B) -> A^B
3513 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3514 if ((A == C && B == D) || (A == D && B == C))
3515 return BinaryOperator::createXor(A, B);
3519 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3520 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3521 return ReplaceInstUsesWith(I, Op0);
3523 // ~(A&B) & (A|B) -> A^B
3524 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3525 if ((A == C && B == D) || (A == D && B == C))
3526 return BinaryOperator::createXor(A, B);
3530 if (Op0->hasOneUse() &&
3531 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3532 if (A == Op1) { // (A^B)&A -> A&(A^B)
3533 I.swapOperands(); // Simplify below
3534 std::swap(Op0, Op1);
3535 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3536 cast<BinaryOperator>(Op0)->swapOperands();
3537 I.swapOperands(); // Simplify below
3538 std::swap(Op0, Op1);
3541 if (Op1->hasOneUse() &&
3542 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3543 if (B == Op0) { // B&(A^B) -> B&(B^A)
3544 cast<BinaryOperator>(Op1)->swapOperands();
3547 if (A == Op0) { // A&(A^B) -> A & ~B
3548 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3549 InsertNewInstBefore(NotB, I);
3550 return BinaryOperator::createAnd(A, NotB);
3555 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3556 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3557 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3560 Value *LHSVal, *RHSVal;
3561 ConstantInt *LHSCst, *RHSCst;
3562 ICmpInst::Predicate LHSCC, RHSCC;
3563 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3564 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3565 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3566 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3567 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3568 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3569 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3570 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3572 // Don't try to fold ICMP_SLT + ICMP_ULT.
3573 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3574 ICmpInst::isSignedPredicate(LHSCC) ==
3575 ICmpInst::isSignedPredicate(RHSCC))) {
3576 // Ensure that the larger constant is on the RHS.
3577 ICmpInst::Predicate GT;
3578 if (ICmpInst::isSignedPredicate(LHSCC) ||
3579 (ICmpInst::isEquality(LHSCC) &&
3580 ICmpInst::isSignedPredicate(RHSCC)))
3581 GT = ICmpInst::ICMP_SGT;
3583 GT = ICmpInst::ICMP_UGT;
3585 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3586 ICmpInst *LHS = cast<ICmpInst>(Op0);
3587 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3588 std::swap(LHS, RHS);
3589 std::swap(LHSCst, RHSCst);
3590 std::swap(LHSCC, RHSCC);
3593 // At this point, we know we have have two icmp instructions
3594 // comparing a value against two constants and and'ing the result
3595 // together. Because of the above check, we know that we only have
3596 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3597 // (from the FoldICmpLogical check above), that the two constants
3598 // are not equal and that the larger constant is on the RHS
3599 assert(LHSCst != RHSCst && "Compares not folded above?");
3602 default: assert(0 && "Unknown integer condition code!");
3603 case ICmpInst::ICMP_EQ:
3605 default: assert(0 && "Unknown integer condition code!");
3606 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3607 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3608 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3609 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3610 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3611 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3612 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3613 return ReplaceInstUsesWith(I, LHS);
3615 case ICmpInst::ICMP_NE:
3617 default: assert(0 && "Unknown integer condition code!");
3618 case ICmpInst::ICMP_ULT:
3619 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3620 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3621 break; // (X != 13 & X u< 15) -> no change
3622 case ICmpInst::ICMP_SLT:
3623 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3624 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3625 break; // (X != 13 & X s< 15) -> no change
3626 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3627 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3628 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3629 return ReplaceInstUsesWith(I, RHS);
3630 case ICmpInst::ICMP_NE:
3631 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3632 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3633 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3634 LHSVal->getName()+".off");
3635 InsertNewInstBefore(Add, I);
3636 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3637 ConstantInt::get(Add->getType(), 1));
3639 break; // (X != 13 & X != 15) -> no change
3642 case ICmpInst::ICMP_ULT:
3644 default: assert(0 && "Unknown integer condition code!");
3645 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3646 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3647 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3648 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3650 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3651 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3652 return ReplaceInstUsesWith(I, LHS);
3653 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3657 case ICmpInst::ICMP_SLT:
3659 default: assert(0 && "Unknown integer condition code!");
3660 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3661 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3662 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3663 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3665 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3666 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3667 return ReplaceInstUsesWith(I, LHS);
3668 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3672 case ICmpInst::ICMP_UGT:
3674 default: assert(0 && "Unknown integer condition code!");
3675 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3676 return ReplaceInstUsesWith(I, LHS);
3677 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3678 return ReplaceInstUsesWith(I, RHS);
3679 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3681 case ICmpInst::ICMP_NE:
3682 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3683 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3684 break; // (X u> 13 & X != 15) -> no change
3685 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3686 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3688 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3692 case ICmpInst::ICMP_SGT:
3694 default: assert(0 && "Unknown integer condition code!");
3695 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3696 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3697 return ReplaceInstUsesWith(I, RHS);
3698 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3700 case ICmpInst::ICMP_NE:
3701 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3702 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3703 break; // (X s> 13 & X != 15) -> no change
3704 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3705 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3707 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3715 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3716 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3717 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3718 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3719 const Type *SrcTy = Op0C->getOperand(0)->getType();
3720 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3721 // Only do this if the casts both really cause code to be generated.
3722 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3724 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3726 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3727 Op1C->getOperand(0),
3729 InsertNewInstBefore(NewOp, I);
3730 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3734 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3735 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3736 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3737 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3738 SI0->getOperand(1) == SI1->getOperand(1) &&
3739 (SI0->hasOneUse() || SI1->hasOneUse())) {
3740 Instruction *NewOp =
3741 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3743 SI0->getName()), I);
3744 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3745 SI1->getOperand(1));
3749 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3750 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3751 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3752 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3753 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3754 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3755 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3756 // If either of the constants are nans, then the whole thing returns
3758 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3759 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3760 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3761 RHS->getOperand(0));
3766 return Changed ? &I : 0;
3769 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3770 /// in the result. If it does, and if the specified byte hasn't been filled in
3771 /// yet, fill it in and return false.
3772 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3773 Instruction *I = dyn_cast<Instruction>(V);
3774 if (I == 0) return true;
3776 // If this is an or instruction, it is an inner node of the bswap.
3777 if (I->getOpcode() == Instruction::Or)
3778 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3779 CollectBSwapParts(I->getOperand(1), ByteValues);
3781 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3782 // If this is a shift by a constant int, and it is "24", then its operand
3783 // defines a byte. We only handle unsigned types here.
3784 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3785 // Not shifting the entire input by N-1 bytes?
3786 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3787 8*(ByteValues.size()-1))
3791 if (I->getOpcode() == Instruction::Shl) {
3792 // X << 24 defines the top byte with the lowest of the input bytes.
3793 DestNo = ByteValues.size()-1;
3795 // X >>u 24 defines the low byte with the highest of the input bytes.
3799 // If the destination byte value is already defined, the values are or'd
3800 // together, which isn't a bswap (unless it's an or of the same bits).
3801 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3803 ByteValues[DestNo] = I->getOperand(0);
3807 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3809 Value *Shift = 0, *ShiftLHS = 0;
3810 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3811 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3812 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3814 Instruction *SI = cast<Instruction>(Shift);
3816 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3817 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3818 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3821 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3823 if (AndAmt->getValue().getActiveBits() > 64)
3825 uint64_t AndAmtVal = AndAmt->getZExtValue();
3826 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3827 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3829 // Unknown mask for bswap.
3830 if (DestByte == ByteValues.size()) return true;
3832 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3834 if (SI->getOpcode() == Instruction::Shl)
3835 SrcByte = DestByte - ShiftBytes;
3837 SrcByte = DestByte + ShiftBytes;
3839 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3840 if (SrcByte != ByteValues.size()-DestByte-1)
3843 // If the destination byte value is already defined, the values are or'd
3844 // together, which isn't a bswap (unless it's an or of the same bits).
3845 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3847 ByteValues[DestByte] = SI->getOperand(0);
3851 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3852 /// If so, insert the new bswap intrinsic and return it.
3853 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3854 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3855 if (!ITy || ITy->getBitWidth() % 16)
3856 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3858 /// ByteValues - For each byte of the result, we keep track of which value
3859 /// defines each byte.
3860 SmallVector<Value*, 8> ByteValues;
3861 ByteValues.resize(ITy->getBitWidth()/8);
3863 // Try to find all the pieces corresponding to the bswap.
3864 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3865 CollectBSwapParts(I.getOperand(1), ByteValues))
3868 // Check to see if all of the bytes come from the same value.
3869 Value *V = ByteValues[0];
3870 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3872 // Check to make sure that all of the bytes come from the same value.
3873 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3874 if (ByteValues[i] != V)
3876 const Type *Tys[] = { ITy };
3877 Module *M = I.getParent()->getParent()->getParent();
3878 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3879 return new CallInst(F, V);
3883 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3884 bool Changed = SimplifyCommutative(I);
3885 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3887 if (isa<UndefValue>(Op1)) // X | undef -> -1
3888 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3892 return ReplaceInstUsesWith(I, Op0);
3894 // See if we can simplify any instructions used by the instruction whose sole
3895 // purpose is to compute bits we don't care about.
3896 if (!isa<VectorType>(I.getType())) {
3897 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3898 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3899 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3900 KnownZero, KnownOne))
3902 } else if (isa<ConstantAggregateZero>(Op1)) {
3903 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3904 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3905 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3906 return ReplaceInstUsesWith(I, I.getOperand(1));
3912 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3913 ConstantInt *C1 = 0; Value *X = 0;
3914 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3915 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3916 Instruction *Or = BinaryOperator::createOr(X, RHS);
3917 InsertNewInstBefore(Or, I);
3919 return BinaryOperator::createAnd(Or,
3920 ConstantInt::get(RHS->getValue() | C1->getValue()));
3923 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3924 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3925 Instruction *Or = BinaryOperator::createOr(X, RHS);
3926 InsertNewInstBefore(Or, I);
3928 return BinaryOperator::createXor(Or,
3929 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3932 // Try to fold constant and into select arguments.
3933 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3934 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3936 if (isa<PHINode>(Op0))
3937 if (Instruction *NV = FoldOpIntoPhi(I))
3941 Value *A = 0, *B = 0;
3942 ConstantInt *C1 = 0, *C2 = 0;
3944 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3945 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3946 return ReplaceInstUsesWith(I, Op1);
3947 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3948 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3949 return ReplaceInstUsesWith(I, Op0);
3951 // (A | B) | C and A | (B | C) -> bswap if possible.
3952 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3953 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3954 match(Op1, m_Or(m_Value(), m_Value())) ||
3955 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3956 match(Op1, m_Shift(m_Value(), m_Value())))) {
3957 if (Instruction *BSwap = MatchBSwap(I))
3961 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3962 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3963 MaskedValueIsZero(Op1, C1->getValue())) {
3964 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3965 InsertNewInstBefore(NOr, I);
3967 return BinaryOperator::createXor(NOr, C1);
3970 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3971 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3972 MaskedValueIsZero(Op0, C1->getValue())) {
3973 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3974 InsertNewInstBefore(NOr, I);
3976 return BinaryOperator::createXor(NOr, C1);
3980 Value *C = 0, *D = 0;
3981 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3982 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3983 Value *V1 = 0, *V2 = 0, *V3 = 0;
3984 C1 = dyn_cast<ConstantInt>(C);
3985 C2 = dyn_cast<ConstantInt>(D);
3986 if (C1 && C2) { // (A & C1)|(B & C2)
3987 // If we have: ((V + N) & C1) | (V & C2)
3988 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3989 // replace with V+N.
3990 if (C1->getValue() == ~C2->getValue()) {
3991 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3992 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3993 // Add commutes, try both ways.
3994 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3995 return ReplaceInstUsesWith(I, A);
3996 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3997 return ReplaceInstUsesWith(I, A);
3999 // Or commutes, try both ways.
4000 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4001 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4002 // Add commutes, try both ways.
4003 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4004 return ReplaceInstUsesWith(I, B);
4005 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4006 return ReplaceInstUsesWith(I, B);
4009 V1 = 0; V2 = 0; V3 = 0;
4012 // Check to see if we have any common things being and'ed. If so, find the
4013 // terms for V1 & (V2|V3).
4014 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4015 if (A == B) // (A & C)|(A & D) == A & (C|D)
4016 V1 = A, V2 = C, V3 = D;
4017 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4018 V1 = A, V2 = B, V3 = C;
4019 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4020 V1 = C, V2 = A, V3 = D;
4021 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4022 V1 = C, V2 = A, V3 = B;
4026 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4027 return BinaryOperator::createAnd(V1, Or);
4032 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4033 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4034 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4035 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4036 SI0->getOperand(1) == SI1->getOperand(1) &&
4037 (SI0->hasOneUse() || SI1->hasOneUse())) {
4038 Instruction *NewOp =
4039 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4041 SI0->getName()), I);
4042 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4043 SI1->getOperand(1));
4047 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4048 if (A == Op1) // ~A | A == -1
4049 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4053 // Note, A is still live here!
4054 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4056 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4058 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4059 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4060 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4061 I.getName()+".demorgan"), I);
4062 return BinaryOperator::createNot(And);
4066 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4067 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4068 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4071 Value *LHSVal, *RHSVal;
4072 ConstantInt *LHSCst, *RHSCst;
4073 ICmpInst::Predicate LHSCC, RHSCC;
4074 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4075 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4076 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4077 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4078 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4079 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4080 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4081 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4082 // We can't fold (ugt x, C) | (sgt x, C2).
4083 PredicatesFoldable(LHSCC, RHSCC)) {
4084 // Ensure that the larger constant is on the RHS.
4085 ICmpInst *LHS = cast<ICmpInst>(Op0);
4087 if (ICmpInst::isSignedPredicate(LHSCC))
4088 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4090 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4093 std::swap(LHS, RHS);
4094 std::swap(LHSCst, RHSCst);
4095 std::swap(LHSCC, RHSCC);
4098 // At this point, we know we have have two icmp instructions
4099 // comparing a value against two constants and or'ing the result
4100 // together. Because of the above check, we know that we only have
4101 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4102 // FoldICmpLogical check above), that the two constants are not
4104 assert(LHSCst != RHSCst && "Compares not folded above?");
4107 default: assert(0 && "Unknown integer condition code!");
4108 case ICmpInst::ICMP_EQ:
4110 default: assert(0 && "Unknown integer condition code!");
4111 case ICmpInst::ICMP_EQ:
4112 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4113 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4114 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4115 LHSVal->getName()+".off");
4116 InsertNewInstBefore(Add, I);
4117 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4118 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4120 break; // (X == 13 | X == 15) -> no change
4121 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4122 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4124 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4125 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4126 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4127 return ReplaceInstUsesWith(I, RHS);
4130 case ICmpInst::ICMP_NE:
4132 default: assert(0 && "Unknown integer condition code!");
4133 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4134 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4135 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4136 return ReplaceInstUsesWith(I, LHS);
4137 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4138 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4139 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4140 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4143 case ICmpInst::ICMP_ULT:
4145 default: assert(0 && "Unknown integer condition code!");
4146 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4148 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4149 // If RHSCst is [us]MAXINT, it is always false. Not handling
4150 // this can cause overflow.
4151 if (RHSCst->isMaxValue(false))
4152 return ReplaceInstUsesWith(I, LHS);
4153 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4155 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4157 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4158 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4159 return ReplaceInstUsesWith(I, RHS);
4160 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4164 case ICmpInst::ICMP_SLT:
4166 default: assert(0 && "Unknown integer condition code!");
4167 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4169 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4170 // If RHSCst is [us]MAXINT, it is always false. Not handling
4171 // this can cause overflow.
4172 if (RHSCst->isMaxValue(true))
4173 return ReplaceInstUsesWith(I, LHS);
4174 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4176 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4178 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4179 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4180 return ReplaceInstUsesWith(I, RHS);
4181 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4185 case ICmpInst::ICMP_UGT:
4187 default: assert(0 && "Unknown integer condition code!");
4188 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4189 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4190 return ReplaceInstUsesWith(I, LHS);
4191 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4193 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4194 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4195 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4196 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4200 case ICmpInst::ICMP_SGT:
4202 default: assert(0 && "Unknown integer condition code!");
4203 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4204 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4205 return ReplaceInstUsesWith(I, LHS);
4206 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4208 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4209 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4210 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4211 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4219 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4220 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4221 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4222 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4223 const Type *SrcTy = Op0C->getOperand(0)->getType();
4224 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4225 // Only do this if the casts both really cause code to be generated.
4226 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4228 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4230 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4231 Op1C->getOperand(0),
4233 InsertNewInstBefore(NewOp, I);
4234 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4240 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4241 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4242 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4243 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4244 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4245 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4246 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4247 // If either of the constants are nans, then the whole thing returns
4249 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4250 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4252 // Otherwise, no need to compare the two constants, compare the
4254 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4255 RHS->getOperand(0));
4260 return Changed ? &I : 0;
4263 // XorSelf - Implements: X ^ X --> 0
4266 XorSelf(Value *rhs) : RHS(rhs) {}
4267 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4268 Instruction *apply(BinaryOperator &Xor) const {
4274 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4275 bool Changed = SimplifyCommutative(I);
4276 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4278 if (isa<UndefValue>(Op1))
4279 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4281 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4282 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4283 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4284 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4287 // See if we can simplify any instructions used by the instruction whose sole
4288 // purpose is to compute bits we don't care about.
4289 if (!isa<VectorType>(I.getType())) {
4290 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4291 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4292 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4293 KnownZero, KnownOne))
4295 } else if (isa<ConstantAggregateZero>(Op1)) {
4296 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4299 // Is this a ~ operation?
4300 if (Value *NotOp = dyn_castNotVal(&I)) {
4301 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4302 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4303 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4304 if (Op0I->getOpcode() == Instruction::And ||
4305 Op0I->getOpcode() == Instruction::Or) {
4306 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4307 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4309 BinaryOperator::createNot(Op0I->getOperand(1),
4310 Op0I->getOperand(1)->getName()+".not");
4311 InsertNewInstBefore(NotY, I);
4312 if (Op0I->getOpcode() == Instruction::And)
4313 return BinaryOperator::createOr(Op0NotVal, NotY);
4315 return BinaryOperator::createAnd(Op0NotVal, NotY);
4322 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4323 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4324 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4325 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4326 return new ICmpInst(ICI->getInversePredicate(),
4327 ICI->getOperand(0), ICI->getOperand(1));
4329 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4330 return new FCmpInst(FCI->getInversePredicate(),
4331 FCI->getOperand(0), FCI->getOperand(1));
4334 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4335 // ~(c-X) == X-c-1 == X+(-c-1)
4336 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4337 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4338 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4339 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4340 ConstantInt::get(I.getType(), 1));
4341 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4344 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4345 if (Op0I->getOpcode() == Instruction::Add) {
4346 // ~(X-c) --> (-c-1)-X
4347 if (RHS->isAllOnesValue()) {
4348 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4349 return BinaryOperator::createSub(
4350 ConstantExpr::getSub(NegOp0CI,
4351 ConstantInt::get(I.getType(), 1)),
4352 Op0I->getOperand(0));
4353 } else if (RHS->getValue().isSignBit()) {
4354 // (X + C) ^ signbit -> (X + C + signbit)
4355 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4356 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4359 } else if (Op0I->getOpcode() == Instruction::Or) {
4360 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4361 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4362 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4363 // Anything in both C1 and C2 is known to be zero, remove it from
4365 Constant *CommonBits = And(Op0CI, RHS);
4366 NewRHS = ConstantExpr::getAnd(NewRHS,
4367 ConstantExpr::getNot(CommonBits));
4368 AddToWorkList(Op0I);
4369 I.setOperand(0, Op0I->getOperand(0));
4370 I.setOperand(1, NewRHS);
4376 // Try to fold constant and into select arguments.
4377 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4378 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4380 if (isa<PHINode>(Op0))
4381 if (Instruction *NV = FoldOpIntoPhi(I))
4385 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4387 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4389 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4391 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4394 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4397 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4398 if (A == Op0) { // B^(B|A) == (A|B)^B
4399 Op1I->swapOperands();
4401 std::swap(Op0, Op1);
4402 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4403 I.swapOperands(); // Simplified below.
4404 std::swap(Op0, Op1);
4406 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4407 if (Op0 == A) // A^(A^B) == B
4408 return ReplaceInstUsesWith(I, B);
4409 else if (Op0 == B) // A^(B^A) == B
4410 return ReplaceInstUsesWith(I, A);
4411 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4412 if (A == Op0) { // A^(A&B) -> A^(B&A)
4413 Op1I->swapOperands();
4416 if (B == Op0) { // A^(B&A) -> (B&A)^A
4417 I.swapOperands(); // Simplified below.
4418 std::swap(Op0, Op1);
4423 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4426 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4427 if (A == Op1) // (B|A)^B == (A|B)^B
4429 if (B == Op1) { // (A|B)^B == A & ~B
4431 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4432 return BinaryOperator::createAnd(A, NotB);
4434 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4435 if (Op1 == A) // (A^B)^A == B
4436 return ReplaceInstUsesWith(I, B);
4437 else if (Op1 == B) // (B^A)^A == B
4438 return ReplaceInstUsesWith(I, A);
4439 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4440 if (A == Op1) // (A&B)^A -> (B&A)^A
4442 if (B == Op1 && // (B&A)^A == ~B & A
4443 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4445 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4446 return BinaryOperator::createAnd(N, Op1);
4451 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4452 if (Op0I && Op1I && Op0I->isShift() &&
4453 Op0I->getOpcode() == Op1I->getOpcode() &&
4454 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4455 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4456 Instruction *NewOp =
4457 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4458 Op1I->getOperand(0),
4459 Op0I->getName()), I);
4460 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4461 Op1I->getOperand(1));
4465 Value *A, *B, *C, *D;
4466 // (A & B)^(A | B) -> A ^ B
4467 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4468 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4469 if ((A == C && B == D) || (A == D && B == C))
4470 return BinaryOperator::createXor(A, B);
4472 // (A | B)^(A & B) -> A ^ B
4473 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4474 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4475 if ((A == C && B == D) || (A == D && B == C))
4476 return BinaryOperator::createXor(A, B);
4480 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4481 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4482 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4483 // (X & Y)^(X & Y) -> (Y^Z) & X
4484 Value *X = 0, *Y = 0, *Z = 0;
4486 X = A, Y = B, Z = D;
4488 X = A, Y = B, Z = C;
4490 X = B, Y = A, Z = D;
4492 X = B, Y = A, Z = C;
4495 Instruction *NewOp =
4496 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4497 return BinaryOperator::createAnd(NewOp, X);
4502 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4503 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4504 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4507 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4508 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4509 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4510 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4511 const Type *SrcTy = Op0C->getOperand(0)->getType();
4512 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4513 // Only do this if the casts both really cause code to be generated.
4514 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4516 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4518 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4519 Op1C->getOperand(0),
4521 InsertNewInstBefore(NewOp, I);
4522 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4526 return Changed ? &I : 0;
4529 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4530 /// overflowed for this type.
4531 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4532 ConstantInt *In2, bool IsSigned = false) {
4533 Result = cast<ConstantInt>(Add(In1, In2));
4536 if (In2->getValue().isNegative())
4537 return Result->getValue().sgt(In1->getValue());
4539 return Result->getValue().slt(In1->getValue());
4541 return Result->getValue().ult(In1->getValue());
4544 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4545 /// code necessary to compute the offset from the base pointer (without adding
4546 /// in the base pointer). Return the result as a signed integer of intptr size.
4547 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4548 TargetData &TD = IC.getTargetData();
4549 gep_type_iterator GTI = gep_type_begin(GEP);
4550 const Type *IntPtrTy = TD.getIntPtrType();
4551 Value *Result = Constant::getNullValue(IntPtrTy);
4553 // Build a mask for high order bits.
4554 unsigned IntPtrWidth = TD.getPointerSize()*8;
4555 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4557 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4558 Value *Op = GEP->getOperand(i);
4559 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4560 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4561 if (OpC->isZero()) continue;
4563 // Handle a struct index, which adds its field offset to the pointer.
4564 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4565 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4567 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4568 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4570 Result = IC.InsertNewInstBefore(
4571 BinaryOperator::createAdd(Result,
4572 ConstantInt::get(IntPtrTy, Size),
4573 GEP->getName()+".offs"), I);
4577 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4578 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4579 Scale = ConstantExpr::getMul(OC, Scale);
4580 if (Constant *RC = dyn_cast<Constant>(Result))
4581 Result = ConstantExpr::getAdd(RC, Scale);
4583 // Emit an add instruction.
4584 Result = IC.InsertNewInstBefore(
4585 BinaryOperator::createAdd(Result, Scale,
4586 GEP->getName()+".offs"), I);
4590 // Convert to correct type.
4591 if (Op->getType() != IntPtrTy) {
4592 if (Constant *OpC = dyn_cast<Constant>(Op))
4593 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4595 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4596 Op->getName()+".c"), I);
4599 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4600 if (Constant *OpC = dyn_cast<Constant>(Op))
4601 Op = ConstantExpr::getMul(OpC, Scale);
4602 else // We'll let instcombine(mul) convert this to a shl if possible.
4603 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4604 GEP->getName()+".idx"), I);
4607 // Emit an add instruction.
4608 if (isa<Constant>(Op) && isa<Constant>(Result))
4609 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4610 cast<Constant>(Result));
4612 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4613 GEP->getName()+".offs"), I);
4618 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4619 /// else. At this point we know that the GEP is on the LHS of the comparison.
4620 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4621 ICmpInst::Predicate Cond,
4623 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4625 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4626 if (isa<PointerType>(CI->getOperand(0)->getType()))
4627 RHS = CI->getOperand(0);
4629 Value *PtrBase = GEPLHS->getOperand(0);
4630 if (PtrBase == RHS) {
4631 // As an optimization, we don't actually have to compute the actual value of
4632 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4633 // each index is zero or not.
4634 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4635 Instruction *InVal = 0;
4636 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4637 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4639 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4640 if (isa<UndefValue>(C)) // undef index -> undef.
4641 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4642 if (C->isNullValue())
4644 else if (TD->getABITypeSize(GTI.getIndexedType()) == 0) {
4645 EmitIt = false; // This is indexing into a zero sized array?
4646 } else if (isa<ConstantInt>(C))
4647 return ReplaceInstUsesWith(I, // No comparison is needed here.
4648 ConstantInt::get(Type::Int1Ty,
4649 Cond == ICmpInst::ICMP_NE));
4654 new ICmpInst(Cond, GEPLHS->getOperand(i),
4655 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4659 InVal = InsertNewInstBefore(InVal, I);
4660 InsertNewInstBefore(Comp, I);
4661 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4662 InVal = BinaryOperator::createOr(InVal, Comp);
4663 else // True if all are equal
4664 InVal = BinaryOperator::createAnd(InVal, Comp);
4672 // No comparison is needed here, all indexes = 0
4673 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4674 Cond == ICmpInst::ICMP_EQ));
4677 // Only lower this if the icmp is the only user of the GEP or if we expect
4678 // the result to fold to a constant!
4679 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4680 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4681 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4682 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4683 Constant::getNullValue(Offset->getType()));
4685 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4686 // If the base pointers are different, but the indices are the same, just
4687 // compare the base pointer.
4688 if (PtrBase != GEPRHS->getOperand(0)) {
4689 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4690 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4691 GEPRHS->getOperand(0)->getType();
4693 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4694 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4695 IndicesTheSame = false;
4699 // If all indices are the same, just compare the base pointers.
4701 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4702 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4704 // Otherwise, the base pointers are different and the indices are
4705 // different, bail out.
4709 // If one of the GEPs has all zero indices, recurse.
4710 bool AllZeros = true;
4711 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4712 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4713 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4718 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4719 ICmpInst::getSwappedPredicate(Cond), I);
4721 // If the other GEP has all zero indices, recurse.
4723 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4724 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4725 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4730 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4732 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4733 // If the GEPs only differ by one index, compare it.
4734 unsigned NumDifferences = 0; // Keep track of # differences.
4735 unsigned DiffOperand = 0; // The operand that differs.
4736 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4737 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4738 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4739 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4740 // Irreconcilable differences.
4744 if (NumDifferences++) break;
4749 if (NumDifferences == 0) // SAME GEP?
4750 return ReplaceInstUsesWith(I, // No comparison is needed here.
4751 ConstantInt::get(Type::Int1Ty,
4752 isTrueWhenEqual(Cond)));
4754 else if (NumDifferences == 1) {
4755 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4756 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4757 // Make sure we do a signed comparison here.
4758 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4762 // Only lower this if the icmp is the only user of the GEP or if we expect
4763 // the result to fold to a constant!
4764 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4765 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4766 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4767 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4768 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4769 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4775 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4776 bool Changed = SimplifyCompare(I);
4777 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4779 // Fold trivial predicates.
4780 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4781 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4782 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4783 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4785 // Simplify 'fcmp pred X, X'
4787 switch (I.getPredicate()) {
4788 default: assert(0 && "Unknown predicate!");
4789 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4790 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4791 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4792 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4793 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4794 case FCmpInst::FCMP_OLT: // True if ordered and less than
4795 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4796 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4798 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4799 case FCmpInst::FCMP_ULT: // True if unordered or less than
4800 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4801 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4802 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4803 I.setPredicate(FCmpInst::FCMP_UNO);
4804 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4807 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4808 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4809 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4810 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4811 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4812 I.setPredicate(FCmpInst::FCMP_ORD);
4813 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4818 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4819 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4821 // Handle fcmp with constant RHS
4822 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4823 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4824 switch (LHSI->getOpcode()) {
4825 case Instruction::PHI:
4826 if (Instruction *NV = FoldOpIntoPhi(I))
4829 case Instruction::Select:
4830 // If either operand of the select is a constant, we can fold the
4831 // comparison into the select arms, which will cause one to be
4832 // constant folded and the select turned into a bitwise or.
4833 Value *Op1 = 0, *Op2 = 0;
4834 if (LHSI->hasOneUse()) {
4835 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4836 // Fold the known value into the constant operand.
4837 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4838 // Insert a new FCmp of the other select operand.
4839 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4840 LHSI->getOperand(2), RHSC,
4842 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4843 // Fold the known value into the constant operand.
4844 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4845 // Insert a new FCmp of the other select operand.
4846 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4847 LHSI->getOperand(1), RHSC,
4853 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4858 return Changed ? &I : 0;
4861 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4862 bool Changed = SimplifyCompare(I);
4863 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4864 const Type *Ty = Op0->getType();
4868 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4869 isTrueWhenEqual(I)));
4871 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4872 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4874 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4875 // addresses never equal each other! We already know that Op0 != Op1.
4876 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4877 isa<ConstantPointerNull>(Op0)) &&
4878 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4879 isa<ConstantPointerNull>(Op1)))
4880 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4881 !isTrueWhenEqual(I)));
4883 // icmp's with boolean values can always be turned into bitwise operations
4884 if (Ty == Type::Int1Ty) {
4885 switch (I.getPredicate()) {
4886 default: assert(0 && "Invalid icmp instruction!");
4887 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4888 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4889 InsertNewInstBefore(Xor, I);
4890 return BinaryOperator::createNot(Xor);
4892 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4893 return BinaryOperator::createXor(Op0, Op1);
4895 case ICmpInst::ICMP_UGT:
4896 case ICmpInst::ICMP_SGT:
4897 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4899 case ICmpInst::ICMP_ULT:
4900 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4901 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4902 InsertNewInstBefore(Not, I);
4903 return BinaryOperator::createAnd(Not, Op1);
4905 case ICmpInst::ICMP_UGE:
4906 case ICmpInst::ICMP_SGE:
4907 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4909 case ICmpInst::ICMP_ULE:
4910 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4911 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4912 InsertNewInstBefore(Not, I);
4913 return BinaryOperator::createOr(Not, Op1);
4918 // See if we are doing a comparison between a constant and an instruction that
4919 // can be folded into the comparison.
4920 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4923 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4924 if (I.isEquality() && CI->isNullValue() &&
4925 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4926 // (icmp cond A B) if cond is equality
4927 return new ICmpInst(I.getPredicate(), A, B);
4930 switch (I.getPredicate()) {
4932 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4933 if (CI->isMinValue(false))
4934 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4935 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4936 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4937 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4938 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4939 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4940 if (CI->isMinValue(true))
4941 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4942 ConstantInt::getAllOnesValue(Op0->getType()));
4946 case ICmpInst::ICMP_SLT:
4947 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4948 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4949 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4950 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4951 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4952 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4955 case ICmpInst::ICMP_UGT:
4956 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4957 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4958 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4959 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4960 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4961 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4963 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4964 if (CI->isMaxValue(true))
4965 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4966 ConstantInt::getNullValue(Op0->getType()));
4969 case ICmpInst::ICMP_SGT:
4970 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4971 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4972 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4973 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4974 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4975 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4978 case ICmpInst::ICMP_ULE:
4979 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4980 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4981 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4982 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4983 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4984 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4987 case ICmpInst::ICMP_SLE:
4988 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4989 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4990 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4991 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4992 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4993 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4996 case ICmpInst::ICMP_UGE:
4997 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4998 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4999 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5000 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5001 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5002 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5005 case ICmpInst::ICMP_SGE:
5006 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5007 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5008 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5009 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5010 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5011 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5015 // If we still have a icmp le or icmp ge instruction, turn it into the
5016 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5017 // already been handled above, this requires little checking.
5019 switch (I.getPredicate()) {
5021 case ICmpInst::ICMP_ULE:
5022 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5023 case ICmpInst::ICMP_SLE:
5024 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5025 case ICmpInst::ICMP_UGE:
5026 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5027 case ICmpInst::ICMP_SGE:
5028 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5031 // See if we can fold the comparison based on bits known to be zero or one
5032 // in the input. If this comparison is a normal comparison, it demands all
5033 // bits, if it is a sign bit comparison, it only demands the sign bit.
5036 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5038 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5039 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5040 if (SimplifyDemandedBits(Op0,
5041 isSignBit ? APInt::getSignBit(BitWidth)
5042 : APInt::getAllOnesValue(BitWidth),
5043 KnownZero, KnownOne, 0))
5046 // Given the known and unknown bits, compute a range that the LHS could be
5048 if ((KnownOne | KnownZero) != 0) {
5049 // Compute the Min, Max and RHS values based on the known bits. For the
5050 // EQ and NE we use unsigned values.
5051 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5052 const APInt& RHSVal = CI->getValue();
5053 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5054 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5057 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5060 switch (I.getPredicate()) { // LE/GE have been folded already.
5061 default: assert(0 && "Unknown icmp opcode!");
5062 case ICmpInst::ICMP_EQ:
5063 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5064 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5066 case ICmpInst::ICMP_NE:
5067 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5068 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5070 case ICmpInst::ICMP_ULT:
5071 if (Max.ult(RHSVal))
5072 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5073 if (Min.uge(RHSVal))
5074 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5076 case ICmpInst::ICMP_UGT:
5077 if (Min.ugt(RHSVal))
5078 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5079 if (Max.ule(RHSVal))
5080 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5082 case ICmpInst::ICMP_SLT:
5083 if (Max.slt(RHSVal))
5084 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5085 if (Min.sgt(RHSVal))
5086 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5088 case ICmpInst::ICMP_SGT:
5089 if (Min.sgt(RHSVal))
5090 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5091 if (Max.sle(RHSVal))
5092 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5097 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5098 // instruction, see if that instruction also has constants so that the
5099 // instruction can be folded into the icmp
5100 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5101 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5105 // Handle icmp with constant (but not simple integer constant) RHS
5106 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5107 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5108 switch (LHSI->getOpcode()) {
5109 case Instruction::GetElementPtr:
5110 if (RHSC->isNullValue()) {
5111 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5112 bool isAllZeros = true;
5113 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5114 if (!isa<Constant>(LHSI->getOperand(i)) ||
5115 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5120 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5121 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5125 case Instruction::PHI:
5126 if (Instruction *NV = FoldOpIntoPhi(I))
5129 case Instruction::Select: {
5130 // If either operand of the select is a constant, we can fold the
5131 // comparison into the select arms, which will cause one to be
5132 // constant folded and the select turned into a bitwise or.
5133 Value *Op1 = 0, *Op2 = 0;
5134 if (LHSI->hasOneUse()) {
5135 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5136 // Fold the known value into the constant operand.
5137 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5138 // Insert a new ICmp of the other select operand.
5139 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5140 LHSI->getOperand(2), RHSC,
5142 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5143 // Fold the known value into the constant operand.
5144 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5145 // Insert a new ICmp of the other select operand.
5146 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5147 LHSI->getOperand(1), RHSC,
5153 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5156 case Instruction::Malloc:
5157 // If we have (malloc != null), and if the malloc has a single use, we
5158 // can assume it is successful and remove the malloc.
5159 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5160 AddToWorkList(LHSI);
5161 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5162 !isTrueWhenEqual(I)));
5168 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5169 if (User *GEP = dyn_castGetElementPtr(Op0))
5170 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5172 if (User *GEP = dyn_castGetElementPtr(Op1))
5173 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5174 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5177 // Test to see if the operands of the icmp are casted versions of other
5178 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5180 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5181 if (isa<PointerType>(Op0->getType()) &&
5182 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5183 // We keep moving the cast from the left operand over to the right
5184 // operand, where it can often be eliminated completely.
5185 Op0 = CI->getOperand(0);
5187 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5188 // so eliminate it as well.
5189 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5190 Op1 = CI2->getOperand(0);
5192 // If Op1 is a constant, we can fold the cast into the constant.
5193 if (Op0->getType() != Op1->getType())
5194 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5195 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5197 // Otherwise, cast the RHS right before the icmp
5198 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5200 return new ICmpInst(I.getPredicate(), Op0, Op1);
5204 if (isa<CastInst>(Op0)) {
5205 // Handle the special case of: icmp (cast bool to X), <cst>
5206 // This comes up when you have code like
5209 // For generality, we handle any zero-extension of any operand comparison
5210 // with a constant or another cast from the same type.
5211 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5212 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5216 if (I.isEquality()) {
5217 Value *A, *B, *C, *D;
5218 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5219 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5220 Value *OtherVal = A == Op1 ? B : A;
5221 return new ICmpInst(I.getPredicate(), OtherVal,
5222 Constant::getNullValue(A->getType()));
5225 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5226 // A^c1 == C^c2 --> A == C^(c1^c2)
5227 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5228 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5229 if (Op1->hasOneUse()) {
5230 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5231 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5232 return new ICmpInst(I.getPredicate(), A,
5233 InsertNewInstBefore(Xor, I));
5236 // A^B == A^D -> B == D
5237 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5238 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5239 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5240 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5244 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5245 (A == Op0 || B == Op0)) {
5246 // A == (A^B) -> B == 0
5247 Value *OtherVal = A == Op0 ? B : A;
5248 return new ICmpInst(I.getPredicate(), OtherVal,
5249 Constant::getNullValue(A->getType()));
5251 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5252 // (A-B) == A -> B == 0
5253 return new ICmpInst(I.getPredicate(), B,
5254 Constant::getNullValue(B->getType()));
5256 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5257 // A == (A-B) -> B == 0
5258 return new ICmpInst(I.getPredicate(), B,
5259 Constant::getNullValue(B->getType()));
5262 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5263 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5264 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5265 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5266 Value *X = 0, *Y = 0, *Z = 0;
5269 X = B; Y = D; Z = A;
5270 } else if (A == D) {
5271 X = B; Y = C; Z = A;
5272 } else if (B == C) {
5273 X = A; Y = D; Z = B;
5274 } else if (B == D) {
5275 X = A; Y = C; Z = B;
5278 if (X) { // Build (X^Y) & Z
5279 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5280 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5281 I.setOperand(0, Op1);
5282 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5287 return Changed ? &I : 0;
5291 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5292 /// and CmpRHS are both known to be integer constants.
5293 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5294 ConstantInt *DivRHS) {
5295 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5296 const APInt &CmpRHSV = CmpRHS->getValue();
5298 // FIXME: If the operand types don't match the type of the divide
5299 // then don't attempt this transform. The code below doesn't have the
5300 // logic to deal with a signed divide and an unsigned compare (and
5301 // vice versa). This is because (x /s C1) <s C2 produces different
5302 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5303 // (x /u C1) <u C2. Simply casting the operands and result won't
5304 // work. :( The if statement below tests that condition and bails
5306 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5307 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5309 if (DivRHS->isZero())
5310 return 0; // The ProdOV computation fails on divide by zero.
5312 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5313 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5314 // C2 (CI). By solving for X we can turn this into a range check
5315 // instead of computing a divide.
5316 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5318 // Determine if the product overflows by seeing if the product is
5319 // not equal to the divide. Make sure we do the same kind of divide
5320 // as in the LHS instruction that we're folding.
5321 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5322 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5324 // Get the ICmp opcode
5325 ICmpInst::Predicate Pred = ICI.getPredicate();
5327 // Figure out the interval that is being checked. For example, a comparison
5328 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5329 // Compute this interval based on the constants involved and the signedness of
5330 // the compare/divide. This computes a half-open interval, keeping track of
5331 // whether either value in the interval overflows. After analysis each
5332 // overflow variable is set to 0 if it's corresponding bound variable is valid
5333 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5334 int LoOverflow = 0, HiOverflow = 0;
5335 ConstantInt *LoBound = 0, *HiBound = 0;
5338 if (!DivIsSigned) { // udiv
5339 // e.g. X/5 op 3 --> [15, 20)
5341 HiOverflow = LoOverflow = ProdOV;
5343 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5344 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5345 if (CmpRHSV == 0) { // (X / pos) op 0
5346 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5347 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5349 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5350 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5351 HiOverflow = LoOverflow = ProdOV;
5353 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5354 } else { // (X / pos) op neg
5355 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5356 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5357 LoOverflow = AddWithOverflow(LoBound, Prod,
5358 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5359 HiBound = AddOne(Prod);
5360 HiOverflow = ProdOV ? -1 : 0;
5362 } else { // Divisor is < 0.
5363 if (CmpRHSV == 0) { // (X / neg) op 0
5364 // e.g. X/-5 op 0 --> [-4, 5)
5365 LoBound = AddOne(DivRHS);
5366 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5367 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5368 HiOverflow = 1; // [INTMIN+1, overflow)
5369 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5371 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5372 // e.g. X/-5 op 3 --> [-19, -14)
5373 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5375 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5376 HiBound = AddOne(Prod);
5377 } else { // (X / neg) op neg
5378 // e.g. X/-5 op -3 --> [15, 20)
5380 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5381 HiBound = Subtract(Prod, DivRHS);
5384 // Dividing by a negative swaps the condition. LT <-> GT
5385 Pred = ICmpInst::getSwappedPredicate(Pred);
5388 Value *X = DivI->getOperand(0);
5390 default: assert(0 && "Unhandled icmp opcode!");
5391 case ICmpInst::ICMP_EQ:
5392 if (LoOverflow && HiOverflow)
5393 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5394 else if (HiOverflow)
5395 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5396 ICmpInst::ICMP_UGE, X, LoBound);
5397 else if (LoOverflow)
5398 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5399 ICmpInst::ICMP_ULT, X, HiBound);
5401 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5402 case ICmpInst::ICMP_NE:
5403 if (LoOverflow && HiOverflow)
5404 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5405 else if (HiOverflow)
5406 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5407 ICmpInst::ICMP_ULT, X, LoBound);
5408 else if (LoOverflow)
5409 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5410 ICmpInst::ICMP_UGE, X, HiBound);
5412 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5413 case ICmpInst::ICMP_ULT:
5414 case ICmpInst::ICMP_SLT:
5415 if (LoOverflow == +1) // Low bound is greater than input range.
5416 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5417 if (LoOverflow == -1) // Low bound is less than input range.
5418 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5419 return new ICmpInst(Pred, X, LoBound);
5420 case ICmpInst::ICMP_UGT:
5421 case ICmpInst::ICMP_SGT:
5422 if (HiOverflow == +1) // High bound greater than input range.
5423 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5424 else if (HiOverflow == -1) // High bound less than input range.
5425 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5426 if (Pred == ICmpInst::ICMP_UGT)
5427 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5429 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5434 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5436 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5439 const APInt &RHSV = RHS->getValue();
5441 switch (LHSI->getOpcode()) {
5442 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5443 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5444 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5446 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5447 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5448 Value *CompareVal = LHSI->getOperand(0);
5450 // If the sign bit of the XorCST is not set, there is no change to
5451 // the operation, just stop using the Xor.
5452 if (!XorCST->getValue().isNegative()) {
5453 ICI.setOperand(0, CompareVal);
5454 AddToWorkList(LHSI);
5458 // Was the old condition true if the operand is positive?
5459 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5461 // If so, the new one isn't.
5462 isTrueIfPositive ^= true;
5464 if (isTrueIfPositive)
5465 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5467 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5471 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5472 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5473 LHSI->getOperand(0)->hasOneUse()) {
5474 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5476 // If the LHS is an AND of a truncating cast, we can widen the
5477 // and/compare to be the input width without changing the value
5478 // produced, eliminating a cast.
5479 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5480 // We can do this transformation if either the AND constant does not
5481 // have its sign bit set or if it is an equality comparison.
5482 // Extending a relational comparison when we're checking the sign
5483 // bit would not work.
5484 if (Cast->hasOneUse() &&
5485 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5486 RHSV.isPositive())) {
5488 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5489 APInt NewCST = AndCST->getValue();
5490 NewCST.zext(BitWidth);
5492 NewCI.zext(BitWidth);
5493 Instruction *NewAnd =
5494 BinaryOperator::createAnd(Cast->getOperand(0),
5495 ConstantInt::get(NewCST),LHSI->getName());
5496 InsertNewInstBefore(NewAnd, ICI);
5497 return new ICmpInst(ICI.getPredicate(), NewAnd,
5498 ConstantInt::get(NewCI));
5502 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5503 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5504 // happens a LOT in code produced by the C front-end, for bitfield
5506 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5507 if (Shift && !Shift->isShift())
5511 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5512 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5513 const Type *AndTy = AndCST->getType(); // Type of the and.
5515 // We can fold this as long as we can't shift unknown bits
5516 // into the mask. This can only happen with signed shift
5517 // rights, as they sign-extend.
5519 bool CanFold = Shift->isLogicalShift();
5521 // To test for the bad case of the signed shr, see if any
5522 // of the bits shifted in could be tested after the mask.
5523 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5524 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5526 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5527 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5528 AndCST->getValue()) == 0)
5534 if (Shift->getOpcode() == Instruction::Shl)
5535 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5537 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5539 // Check to see if we are shifting out any of the bits being
5541 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5542 // If we shifted bits out, the fold is not going to work out.
5543 // As a special case, check to see if this means that the
5544 // result is always true or false now.
5545 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5546 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5547 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5548 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5550 ICI.setOperand(1, NewCst);
5551 Constant *NewAndCST;
5552 if (Shift->getOpcode() == Instruction::Shl)
5553 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5555 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5556 LHSI->setOperand(1, NewAndCST);
5557 LHSI->setOperand(0, Shift->getOperand(0));
5558 AddToWorkList(Shift); // Shift is dead.
5559 AddUsesToWorkList(ICI);
5565 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5566 // preferable because it allows the C<<Y expression to be hoisted out
5567 // of a loop if Y is invariant and X is not.
5568 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5569 ICI.isEquality() && !Shift->isArithmeticShift() &&
5570 isa<Instruction>(Shift->getOperand(0))) {
5573 if (Shift->getOpcode() == Instruction::LShr) {
5574 NS = BinaryOperator::createShl(AndCST,
5575 Shift->getOperand(1), "tmp");
5577 // Insert a logical shift.
5578 NS = BinaryOperator::createLShr(AndCST,
5579 Shift->getOperand(1), "tmp");
5581 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5583 // Compute X & (C << Y).
5584 Instruction *NewAnd =
5585 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5586 InsertNewInstBefore(NewAnd, ICI);
5588 ICI.setOperand(0, NewAnd);
5594 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5595 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5598 uint32_t TypeBits = RHSV.getBitWidth();
5600 // Check that the shift amount is in range. If not, don't perform
5601 // undefined shifts. When the shift is visited it will be
5603 if (ShAmt->uge(TypeBits))
5606 if (ICI.isEquality()) {
5607 // If we are comparing against bits always shifted out, the
5608 // comparison cannot succeed.
5610 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5611 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5612 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5613 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5614 return ReplaceInstUsesWith(ICI, Cst);
5617 if (LHSI->hasOneUse()) {
5618 // Otherwise strength reduce the shift into an and.
5619 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5621 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5624 BinaryOperator::createAnd(LHSI->getOperand(0),
5625 Mask, LHSI->getName()+".mask");
5626 Value *And = InsertNewInstBefore(AndI, ICI);
5627 return new ICmpInst(ICI.getPredicate(), And,
5628 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5632 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5633 bool TrueIfSigned = false;
5634 if (LHSI->hasOneUse() &&
5635 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5636 // (X << 31) <s 0 --> (X&1) != 0
5637 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5638 (TypeBits-ShAmt->getZExtValue()-1));
5640 BinaryOperator::createAnd(LHSI->getOperand(0),
5641 Mask, LHSI->getName()+".mask");
5642 Value *And = InsertNewInstBefore(AndI, ICI);
5644 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5645 And, Constant::getNullValue(And->getType()));
5650 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5651 case Instruction::AShr: {
5652 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5655 if (ICI.isEquality()) {
5656 // Check that the shift amount is in range. If not, don't perform
5657 // undefined shifts. When the shift is visited it will be
5659 uint32_t TypeBits = RHSV.getBitWidth();
5660 if (ShAmt->uge(TypeBits))
5662 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5664 // If we are comparing against bits always shifted out, the
5665 // comparison cannot succeed.
5666 APInt Comp = RHSV << ShAmtVal;
5667 if (LHSI->getOpcode() == Instruction::LShr)
5668 Comp = Comp.lshr(ShAmtVal);
5670 Comp = Comp.ashr(ShAmtVal);
5672 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5673 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5674 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5675 return ReplaceInstUsesWith(ICI, Cst);
5678 if (LHSI->hasOneUse() || RHSV == 0) {
5679 // Otherwise strength reduce the shift into an and.
5680 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5681 Constant *Mask = ConstantInt::get(Val);
5684 BinaryOperator::createAnd(LHSI->getOperand(0),
5685 Mask, LHSI->getName()+".mask");
5686 Value *And = InsertNewInstBefore(AndI, ICI);
5687 return new ICmpInst(ICI.getPredicate(), And,
5688 ConstantExpr::getShl(RHS, ShAmt));
5694 case Instruction::SDiv:
5695 case Instruction::UDiv:
5696 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5697 // Fold this div into the comparison, producing a range check.
5698 // Determine, based on the divide type, what the range is being
5699 // checked. If there is an overflow on the low or high side, remember
5700 // it, otherwise compute the range [low, hi) bounding the new value.
5701 // See: InsertRangeTest above for the kinds of replacements possible.
5702 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5703 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5709 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5710 if (ICI.isEquality()) {
5711 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5713 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5714 // the second operand is a constant, simplify a bit.
5715 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5716 switch (BO->getOpcode()) {
5717 case Instruction::SRem:
5718 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5719 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5720 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5721 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5722 Instruction *NewRem =
5723 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5725 InsertNewInstBefore(NewRem, ICI);
5726 return new ICmpInst(ICI.getPredicate(), NewRem,
5727 Constant::getNullValue(BO->getType()));
5731 case Instruction::Add:
5732 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5733 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5734 if (BO->hasOneUse())
5735 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5736 Subtract(RHS, BOp1C));
5737 } else if (RHSV == 0) {
5738 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5739 // efficiently invertible, or if the add has just this one use.
5740 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5742 if (Value *NegVal = dyn_castNegVal(BOp1))
5743 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5744 else if (Value *NegVal = dyn_castNegVal(BOp0))
5745 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5746 else if (BO->hasOneUse()) {
5747 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5748 InsertNewInstBefore(Neg, ICI);
5750 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5754 case Instruction::Xor:
5755 // For the xor case, we can xor two constants together, eliminating
5756 // the explicit xor.
5757 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5758 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5759 ConstantExpr::getXor(RHS, BOC));
5762 case Instruction::Sub:
5763 // Replace (([sub|xor] A, B) != 0) with (A != B)
5765 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5769 case Instruction::Or:
5770 // If bits are being or'd in that are not present in the constant we
5771 // are comparing against, then the comparison could never succeed!
5772 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5773 Constant *NotCI = ConstantExpr::getNot(RHS);
5774 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5775 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5780 case Instruction::And:
5781 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5782 // If bits are being compared against that are and'd out, then the
5783 // comparison can never succeed!
5784 if ((RHSV & ~BOC->getValue()) != 0)
5785 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5788 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5789 if (RHS == BOC && RHSV.isPowerOf2())
5790 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5791 ICmpInst::ICMP_NE, LHSI,
5792 Constant::getNullValue(RHS->getType()));
5794 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5795 if (isSignBit(BOC)) {
5796 Value *X = BO->getOperand(0);
5797 Constant *Zero = Constant::getNullValue(X->getType());
5798 ICmpInst::Predicate pred = isICMP_NE ?
5799 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5800 return new ICmpInst(pred, X, Zero);
5803 // ((X & ~7) == 0) --> X < 8
5804 if (RHSV == 0 && isHighOnes(BOC)) {
5805 Value *X = BO->getOperand(0);
5806 Constant *NegX = ConstantExpr::getNeg(BOC);
5807 ICmpInst::Predicate pred = isICMP_NE ?
5808 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5809 return new ICmpInst(pred, X, NegX);
5814 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5815 // Handle icmp {eq|ne} <intrinsic>, intcst.
5816 if (II->getIntrinsicID() == Intrinsic::bswap) {
5818 ICI.setOperand(0, II->getOperand(1));
5819 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5823 } else { // Not a ICMP_EQ/ICMP_NE
5824 // If the LHS is a cast from an integral value of the same size,
5825 // then since we know the RHS is a constant, try to simlify.
5826 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5827 Value *CastOp = Cast->getOperand(0);
5828 const Type *SrcTy = CastOp->getType();
5829 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5830 if (SrcTy->isInteger() &&
5831 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5832 // If this is an unsigned comparison, try to make the comparison use
5833 // smaller constant values.
5834 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5835 // X u< 128 => X s> -1
5836 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5837 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5838 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5839 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5840 // X u> 127 => X s< 0
5841 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5842 Constant::getNullValue(SrcTy));
5850 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5851 /// We only handle extending casts so far.
5853 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5854 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5855 Value *LHSCIOp = LHSCI->getOperand(0);
5856 const Type *SrcTy = LHSCIOp->getType();
5857 const Type *DestTy = LHSCI->getType();
5860 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5861 // integer type is the same size as the pointer type.
5862 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5863 getTargetData().getPointerSizeInBits() ==
5864 cast<IntegerType>(DestTy)->getBitWidth()) {
5866 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5867 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5868 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5869 RHSOp = RHSC->getOperand(0);
5870 // If the pointer types don't match, insert a bitcast.
5871 if (LHSCIOp->getType() != RHSOp->getType())
5872 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5876 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5879 // The code below only handles extension cast instructions, so far.
5881 if (LHSCI->getOpcode() != Instruction::ZExt &&
5882 LHSCI->getOpcode() != Instruction::SExt)
5885 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5886 bool isSignedCmp = ICI.isSignedPredicate();
5888 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5889 // Not an extension from the same type?
5890 RHSCIOp = CI->getOperand(0);
5891 if (RHSCIOp->getType() != LHSCIOp->getType())
5894 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5895 // and the other is a zext), then we can't handle this.
5896 if (CI->getOpcode() != LHSCI->getOpcode())
5899 // Deal with equality cases early.
5900 if (ICI.isEquality())
5901 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5903 // A signed comparison of sign extended values simplifies into a
5904 // signed comparison.
5905 if (isSignedCmp && isSignedExt)
5906 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5908 // The other three cases all fold into an unsigned comparison.
5909 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5912 // If we aren't dealing with a constant on the RHS, exit early
5913 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5917 // Compute the constant that would happen if we truncated to SrcTy then
5918 // reextended to DestTy.
5919 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5920 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5922 // If the re-extended constant didn't change...
5924 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5925 // For example, we might have:
5926 // %A = sext short %X to uint
5927 // %B = icmp ugt uint %A, 1330
5928 // It is incorrect to transform this into
5929 // %B = icmp ugt short %X, 1330
5930 // because %A may have negative value.
5932 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5933 // OR operation is EQ/NE.
5934 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5935 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5940 // The re-extended constant changed so the constant cannot be represented
5941 // in the shorter type. Consequently, we cannot emit a simple comparison.
5943 // First, handle some easy cases. We know the result cannot be equal at this
5944 // point so handle the ICI.isEquality() cases
5945 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5946 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5947 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5948 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5950 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5951 // should have been folded away previously and not enter in here.
5954 // We're performing a signed comparison.
5955 if (cast<ConstantInt>(CI)->getValue().isNegative())
5956 Result = ConstantInt::getFalse(); // X < (small) --> false
5958 Result = ConstantInt::getTrue(); // X < (large) --> true
5960 // We're performing an unsigned comparison.
5962 // We're performing an unsigned comp with a sign extended value.
5963 // This is true if the input is >= 0. [aka >s -1]
5964 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5965 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5966 NegOne, ICI.getName()), ICI);
5968 // Unsigned extend & unsigned compare -> always true.
5969 Result = ConstantInt::getTrue();
5973 // Finally, return the value computed.
5974 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5975 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5976 return ReplaceInstUsesWith(ICI, Result);
5978 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5979 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5980 "ICmp should be folded!");
5981 if (Constant *CI = dyn_cast<Constant>(Result))
5982 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5984 return BinaryOperator::createNot(Result);
5988 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5989 return commonShiftTransforms(I);
5992 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5993 return commonShiftTransforms(I);
5996 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5997 if (Instruction *R = commonShiftTransforms(I))
6000 Value *Op0 = I.getOperand(0);
6002 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6003 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6004 if (CSI->isAllOnesValue())
6005 return ReplaceInstUsesWith(I, CSI);
6007 // See if we can turn a signed shr into an unsigned shr.
6008 if (MaskedValueIsZero(Op0,
6009 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6010 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6015 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6016 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6017 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6019 // shl X, 0 == X and shr X, 0 == X
6020 // shl 0, X == 0 and shr 0, X == 0
6021 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6022 Op0 == Constant::getNullValue(Op0->getType()))
6023 return ReplaceInstUsesWith(I, Op0);
6025 if (isa<UndefValue>(Op0)) {
6026 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6027 return ReplaceInstUsesWith(I, Op0);
6028 else // undef << X -> 0, undef >>u X -> 0
6029 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6031 if (isa<UndefValue>(Op1)) {
6032 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6033 return ReplaceInstUsesWith(I, Op0);
6034 else // X << undef, X >>u undef -> 0
6035 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6038 // Try to fold constant and into select arguments.
6039 if (isa<Constant>(Op0))
6040 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6041 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6044 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6045 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6050 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6051 BinaryOperator &I) {
6052 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6054 // See if we can simplify any instructions used by the instruction whose sole
6055 // purpose is to compute bits we don't care about.
6056 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6057 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6058 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6059 KnownZero, KnownOne))
6062 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6063 // of a signed value.
6065 if (Op1->uge(TypeBits)) {
6066 if (I.getOpcode() != Instruction::AShr)
6067 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6069 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6074 // ((X*C1) << C2) == (X * (C1 << C2))
6075 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6076 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6077 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6078 return BinaryOperator::createMul(BO->getOperand(0),
6079 ConstantExpr::getShl(BOOp, Op1));
6081 // Try to fold constant and into select arguments.
6082 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6083 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6085 if (isa<PHINode>(Op0))
6086 if (Instruction *NV = FoldOpIntoPhi(I))
6089 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6090 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6091 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6092 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6093 // place. Don't try to do this transformation in this case. Also, we
6094 // require that the input operand is a shift-by-constant so that we have
6095 // confidence that the shifts will get folded together. We could do this
6096 // xform in more cases, but it is unlikely to be profitable.
6097 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6098 isa<ConstantInt>(TrOp->getOperand(1))) {
6099 // Okay, we'll do this xform. Make the shift of shift.
6100 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6101 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6103 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6105 // For logical shifts, the truncation has the effect of making the high
6106 // part of the register be zeros. Emulate this by inserting an AND to
6107 // clear the top bits as needed. This 'and' will usually be zapped by
6108 // other xforms later if dead.
6109 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6110 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6111 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6113 // The mask we constructed says what the trunc would do if occurring
6114 // between the shifts. We want to know the effect *after* the second
6115 // shift. We know that it is a logical shift by a constant, so adjust the
6116 // mask as appropriate.
6117 if (I.getOpcode() == Instruction::Shl)
6118 MaskV <<= Op1->getZExtValue();
6120 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6121 MaskV = MaskV.lshr(Op1->getZExtValue());
6124 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6126 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6128 // Return the value truncated to the interesting size.
6129 return new TruncInst(And, I.getType());
6133 if (Op0->hasOneUse()) {
6134 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6135 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6138 switch (Op0BO->getOpcode()) {
6140 case Instruction::Add:
6141 case Instruction::And:
6142 case Instruction::Or:
6143 case Instruction::Xor: {
6144 // These operators commute.
6145 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6146 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6147 match(Op0BO->getOperand(1),
6148 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6149 Instruction *YS = BinaryOperator::createShl(
6150 Op0BO->getOperand(0), Op1,
6152 InsertNewInstBefore(YS, I); // (Y << C)
6154 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6155 Op0BO->getOperand(1)->getName());
6156 InsertNewInstBefore(X, I); // (X + (Y << C))
6157 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6158 return BinaryOperator::createAnd(X, ConstantInt::get(
6159 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6162 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6163 Value *Op0BOOp1 = Op0BO->getOperand(1);
6164 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6166 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6167 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6169 Instruction *YS = BinaryOperator::createShl(
6170 Op0BO->getOperand(0), Op1,
6172 InsertNewInstBefore(YS, I); // (Y << C)
6174 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6175 V1->getName()+".mask");
6176 InsertNewInstBefore(XM, I); // X & (CC << C)
6178 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6183 case Instruction::Sub: {
6184 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6185 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6186 match(Op0BO->getOperand(0),
6187 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6188 Instruction *YS = BinaryOperator::createShl(
6189 Op0BO->getOperand(1), Op1,
6191 InsertNewInstBefore(YS, I); // (Y << C)
6193 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6194 Op0BO->getOperand(0)->getName());
6195 InsertNewInstBefore(X, I); // (X + (Y << C))
6196 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6197 return BinaryOperator::createAnd(X, ConstantInt::get(
6198 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6201 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6202 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6203 match(Op0BO->getOperand(0),
6204 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6205 m_ConstantInt(CC))) && V2 == Op1 &&
6206 cast<BinaryOperator>(Op0BO->getOperand(0))
6207 ->getOperand(0)->hasOneUse()) {
6208 Instruction *YS = BinaryOperator::createShl(
6209 Op0BO->getOperand(1), Op1,
6211 InsertNewInstBefore(YS, I); // (Y << C)
6213 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6214 V1->getName()+".mask");
6215 InsertNewInstBefore(XM, I); // X & (CC << C)
6217 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6225 // If the operand is an bitwise operator with a constant RHS, and the
6226 // shift is the only use, we can pull it out of the shift.
6227 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6228 bool isValid = true; // Valid only for And, Or, Xor
6229 bool highBitSet = false; // Transform if high bit of constant set?
6231 switch (Op0BO->getOpcode()) {
6232 default: isValid = false; break; // Do not perform transform!
6233 case Instruction::Add:
6234 isValid = isLeftShift;
6236 case Instruction::Or:
6237 case Instruction::Xor:
6240 case Instruction::And:
6245 // If this is a signed shift right, and the high bit is modified
6246 // by the logical operation, do not perform the transformation.
6247 // The highBitSet boolean indicates the value of the high bit of
6248 // the constant which would cause it to be modified for this
6251 if (isValid && I.getOpcode() == Instruction::AShr)
6252 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6255 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6257 Instruction *NewShift =
6258 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6259 InsertNewInstBefore(NewShift, I);
6260 NewShift->takeName(Op0BO);
6262 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6269 // Find out if this is a shift of a shift by a constant.
6270 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6271 if (ShiftOp && !ShiftOp->isShift())
6274 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6275 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6276 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6277 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6278 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6279 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6280 Value *X = ShiftOp->getOperand(0);
6282 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6283 if (AmtSum > TypeBits)
6286 const IntegerType *Ty = cast<IntegerType>(I.getType());
6288 // Check for (X << c1) << c2 and (X >> c1) >> c2
6289 if (I.getOpcode() == ShiftOp->getOpcode()) {
6290 return BinaryOperator::create(I.getOpcode(), X,
6291 ConstantInt::get(Ty, AmtSum));
6292 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6293 I.getOpcode() == Instruction::AShr) {
6294 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6295 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6296 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6297 I.getOpcode() == Instruction::LShr) {
6298 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6299 Instruction *Shift =
6300 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6301 InsertNewInstBefore(Shift, I);
6303 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6304 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6307 // Okay, if we get here, one shift must be left, and the other shift must be
6308 // right. See if the amounts are equal.
6309 if (ShiftAmt1 == ShiftAmt2) {
6310 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6311 if (I.getOpcode() == Instruction::Shl) {
6312 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6313 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6315 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6316 if (I.getOpcode() == Instruction::LShr) {
6317 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6318 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6320 // We can simplify ((X << C) >>s C) into a trunc + sext.
6321 // NOTE: we could do this for any C, but that would make 'unusual' integer
6322 // types. For now, just stick to ones well-supported by the code
6324 const Type *SExtType = 0;
6325 switch (Ty->getBitWidth() - ShiftAmt1) {
6332 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6337 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6338 InsertNewInstBefore(NewTrunc, I);
6339 return new SExtInst(NewTrunc, Ty);
6341 // Otherwise, we can't handle it yet.
6342 } else if (ShiftAmt1 < ShiftAmt2) {
6343 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6345 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6346 if (I.getOpcode() == Instruction::Shl) {
6347 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6348 ShiftOp->getOpcode() == Instruction::AShr);
6349 Instruction *Shift =
6350 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6351 InsertNewInstBefore(Shift, I);
6353 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6354 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6357 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6358 if (I.getOpcode() == Instruction::LShr) {
6359 assert(ShiftOp->getOpcode() == Instruction::Shl);
6360 Instruction *Shift =
6361 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6362 InsertNewInstBefore(Shift, I);
6364 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6365 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6368 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6370 assert(ShiftAmt2 < ShiftAmt1);
6371 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6373 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6374 if (I.getOpcode() == Instruction::Shl) {
6375 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6376 ShiftOp->getOpcode() == Instruction::AShr);
6377 Instruction *Shift =
6378 BinaryOperator::create(ShiftOp->getOpcode(), X,
6379 ConstantInt::get(Ty, ShiftDiff));
6380 InsertNewInstBefore(Shift, I);
6382 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6383 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6386 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6387 if (I.getOpcode() == Instruction::LShr) {
6388 assert(ShiftOp->getOpcode() == Instruction::Shl);
6389 Instruction *Shift =
6390 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6391 InsertNewInstBefore(Shift, I);
6393 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6394 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6397 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6404 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6405 /// expression. If so, decompose it, returning some value X, such that Val is
6408 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6410 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6411 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6412 Offset = CI->getZExtValue();
6414 return ConstantInt::get(Type::Int32Ty, 0);
6415 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6416 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6417 if (I->getOpcode() == Instruction::Shl) {
6418 // This is a value scaled by '1 << the shift amt'.
6419 Scale = 1U << RHS->getZExtValue();
6421 return I->getOperand(0);
6422 } else if (I->getOpcode() == Instruction::Mul) {
6423 // This value is scaled by 'RHS'.
6424 Scale = RHS->getZExtValue();
6426 return I->getOperand(0);
6427 } else if (I->getOpcode() == Instruction::Add) {
6428 // We have X+C. Check to see if we really have (X*C2)+C1,
6429 // where C1 is divisible by C2.
6432 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6433 Offset += RHS->getZExtValue();
6440 // Otherwise, we can't look past this.
6447 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6448 /// try to eliminate the cast by moving the type information into the alloc.
6449 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6450 AllocationInst &AI) {
6451 const PointerType *PTy = cast<PointerType>(CI.getType());
6453 // Remove any uses of AI that are dead.
6454 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6456 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6457 Instruction *User = cast<Instruction>(*UI++);
6458 if (isInstructionTriviallyDead(User)) {
6459 while (UI != E && *UI == User)
6460 ++UI; // If this instruction uses AI more than once, don't break UI.
6463 DOUT << "IC: DCE: " << *User;
6464 EraseInstFromFunction(*User);
6468 // Get the type really allocated and the type casted to.
6469 const Type *AllocElTy = AI.getAllocatedType();
6470 const Type *CastElTy = PTy->getElementType();
6471 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6473 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6474 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6475 if (CastElTyAlign < AllocElTyAlign) return 0;
6477 // If the allocation has multiple uses, only promote it if we are strictly
6478 // increasing the alignment of the resultant allocation. If we keep it the
6479 // same, we open the door to infinite loops of various kinds.
6480 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6482 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6483 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6484 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6486 // See if we can satisfy the modulus by pulling a scale out of the array
6488 unsigned ArraySizeScale;
6490 Value *NumElements = // See if the array size is a decomposable linear expr.
6491 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6493 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6495 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6496 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6498 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6503 // If the allocation size is constant, form a constant mul expression
6504 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6505 if (isa<ConstantInt>(NumElements))
6506 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6507 // otherwise multiply the amount and the number of elements
6508 else if (Scale != 1) {
6509 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6510 Amt = InsertNewInstBefore(Tmp, AI);
6514 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6515 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6516 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6517 Amt = InsertNewInstBefore(Tmp, AI);
6520 AllocationInst *New;
6521 if (isa<MallocInst>(AI))
6522 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6524 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6525 InsertNewInstBefore(New, AI);
6528 // If the allocation has multiple uses, insert a cast and change all things
6529 // that used it to use the new cast. This will also hack on CI, but it will
6531 if (!AI.hasOneUse()) {
6532 AddUsesToWorkList(AI);
6533 // New is the allocation instruction, pointer typed. AI is the original
6534 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6535 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6536 InsertNewInstBefore(NewCast, AI);
6537 AI.replaceAllUsesWith(NewCast);
6539 return ReplaceInstUsesWith(CI, New);
6542 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6543 /// and return it as type Ty without inserting any new casts and without
6544 /// changing the computed value. This is used by code that tries to decide
6545 /// whether promoting or shrinking integer operations to wider or smaller types
6546 /// will allow us to eliminate a truncate or extend.
6548 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6549 /// extension operation if Ty is larger.
6550 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6551 unsigned CastOpc, int &NumCastsRemoved) {
6552 // We can always evaluate constants in another type.
6553 if (isa<ConstantInt>(V))
6556 Instruction *I = dyn_cast<Instruction>(V);
6557 if (!I) return false;
6559 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6561 // If this is an extension or truncate, we can often eliminate it.
6562 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6563 // If this is a cast from the destination type, we can trivially eliminate
6564 // it, and this will remove a cast overall.
6565 if (I->getOperand(0)->getType() == Ty) {
6566 // If the first operand is itself a cast, and is eliminable, do not count
6567 // this as an eliminable cast. We would prefer to eliminate those two
6569 if (!isa<CastInst>(I->getOperand(0)))
6575 // We can't extend or shrink something that has multiple uses: doing so would
6576 // require duplicating the instruction in general, which isn't profitable.
6577 if (!I->hasOneUse()) return false;
6579 switch (I->getOpcode()) {
6580 case Instruction::Add:
6581 case Instruction::Sub:
6582 case Instruction::And:
6583 case Instruction::Or:
6584 case Instruction::Xor:
6585 // These operators can all arbitrarily be extended or truncated.
6586 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6588 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6591 case Instruction::Mul:
6592 // A multiply can be truncated by truncating its operands.
6593 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6594 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6596 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6599 case Instruction::Shl:
6600 // If we are truncating the result of this SHL, and if it's a shift of a
6601 // constant amount, we can always perform a SHL in a smaller type.
6602 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6603 uint32_t BitWidth = Ty->getBitWidth();
6604 if (BitWidth < OrigTy->getBitWidth() &&
6605 CI->getLimitedValue(BitWidth) < BitWidth)
6606 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6610 case Instruction::LShr:
6611 // If this is a truncate of a logical shr, we can truncate it to a smaller
6612 // lshr iff we know that the bits we would otherwise be shifting in are
6614 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6615 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6616 uint32_t BitWidth = Ty->getBitWidth();
6617 if (BitWidth < OrigBitWidth &&
6618 MaskedValueIsZero(I->getOperand(0),
6619 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6620 CI->getLimitedValue(BitWidth) < BitWidth) {
6621 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6626 case Instruction::ZExt:
6627 case Instruction::SExt:
6628 case Instruction::Trunc:
6629 // If this is the same kind of case as our original (e.g. zext+zext), we
6630 // can safely replace it. Note that replacing it does not reduce the number
6631 // of casts in the input.
6632 if (I->getOpcode() == CastOpc)
6637 // TODO: Can handle more cases here.
6644 /// EvaluateInDifferentType - Given an expression that
6645 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6646 /// evaluate the expression.
6647 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6649 if (Constant *C = dyn_cast<Constant>(V))
6650 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6652 // Otherwise, it must be an instruction.
6653 Instruction *I = cast<Instruction>(V);
6654 Instruction *Res = 0;
6655 switch (I->getOpcode()) {
6656 case Instruction::Add:
6657 case Instruction::Sub:
6658 case Instruction::Mul:
6659 case Instruction::And:
6660 case Instruction::Or:
6661 case Instruction::Xor:
6662 case Instruction::AShr:
6663 case Instruction::LShr:
6664 case Instruction::Shl: {
6665 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6666 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6667 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6668 LHS, RHS, I->getName());
6671 case Instruction::Trunc:
6672 case Instruction::ZExt:
6673 case Instruction::SExt:
6674 // If the source type of the cast is the type we're trying for then we can
6675 // just return the source. There's no need to insert it because it is not
6677 if (I->getOperand(0)->getType() == Ty)
6678 return I->getOperand(0);
6680 // Otherwise, must be the same type of case, so just reinsert a new one.
6681 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6685 // TODO: Can handle more cases here.
6686 assert(0 && "Unreachable!");
6690 return InsertNewInstBefore(Res, *I);
6693 /// @brief Implement the transforms common to all CastInst visitors.
6694 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6695 Value *Src = CI.getOperand(0);
6697 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6698 // eliminate it now.
6699 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6700 if (Instruction::CastOps opc =
6701 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6702 // The first cast (CSrc) is eliminable so we need to fix up or replace
6703 // the second cast (CI). CSrc will then have a good chance of being dead.
6704 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6708 // If we are casting a select then fold the cast into the select
6709 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6710 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6713 // If we are casting a PHI then fold the cast into the PHI
6714 if (isa<PHINode>(Src))
6715 if (Instruction *NV = FoldOpIntoPhi(CI))
6721 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6722 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6723 Value *Src = CI.getOperand(0);
6725 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6726 // If casting the result of a getelementptr instruction with no offset, turn
6727 // this into a cast of the original pointer!
6728 if (GEP->hasAllZeroIndices()) {
6729 // Changing the cast operand is usually not a good idea but it is safe
6730 // here because the pointer operand is being replaced with another
6731 // pointer operand so the opcode doesn't need to change.
6733 CI.setOperand(0, GEP->getOperand(0));
6737 // If the GEP has a single use, and the base pointer is a bitcast, and the
6738 // GEP computes a constant offset, see if we can convert these three
6739 // instructions into fewer. This typically happens with unions and other
6740 // non-type-safe code.
6741 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6742 if (GEP->hasAllConstantIndices()) {
6743 // We are guaranteed to get a constant from EmitGEPOffset.
6744 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6745 int64_t Offset = OffsetV->getSExtValue();
6747 // Get the base pointer input of the bitcast, and the type it points to.
6748 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6749 const Type *GEPIdxTy =
6750 cast<PointerType>(OrigBase->getType())->getElementType();
6751 if (GEPIdxTy->isSized()) {
6752 SmallVector<Value*, 8> NewIndices;
6754 // Start with the index over the outer type. Note that the type size
6755 // might be zero (even if the offset isn't zero) if the indexed type
6756 // is something like [0 x {int, int}]
6757 const Type *IntPtrTy = TD->getIntPtrType();
6758 int64_t FirstIdx = 0;
6759 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6760 FirstIdx = Offset/TySize;
6763 // Handle silly modulus not returning values values [0..TySize).
6767 assert(Offset >= 0);
6769 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6772 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6774 // Index into the types. If we fail, set OrigBase to null.
6776 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6777 const StructLayout *SL = TD->getStructLayout(STy);
6778 if (Offset < (int64_t)SL->getSizeInBytes()) {
6779 unsigned Elt = SL->getElementContainingOffset(Offset);
6780 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6782 Offset -= SL->getElementOffset(Elt);
6783 GEPIdxTy = STy->getElementType(Elt);
6785 // Otherwise, we can't index into this, bail out.
6789 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6790 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6791 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6792 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6795 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6797 GEPIdxTy = STy->getElementType();
6799 // Otherwise, we can't index into this, bail out.
6805 // If we were able to index down into an element, create the GEP
6806 // and bitcast the result. This eliminates one bitcast, potentially
6808 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6810 NewIndices.end(), "");
6811 InsertNewInstBefore(NGEP, CI);
6812 NGEP->takeName(GEP);
6814 if (isa<BitCastInst>(CI))
6815 return new BitCastInst(NGEP, CI.getType());
6816 assert(isa<PtrToIntInst>(CI));
6817 return new PtrToIntInst(NGEP, CI.getType());
6824 return commonCastTransforms(CI);
6829 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6830 /// integer types. This function implements the common transforms for all those
6832 /// @brief Implement the transforms common to CastInst with integer operands
6833 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6834 if (Instruction *Result = commonCastTransforms(CI))
6837 Value *Src = CI.getOperand(0);
6838 const Type *SrcTy = Src->getType();
6839 const Type *DestTy = CI.getType();
6840 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6841 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6843 // See if we can simplify any instructions used by the LHS whose sole
6844 // purpose is to compute bits we don't care about.
6845 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6846 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6847 KnownZero, KnownOne))
6850 // If the source isn't an instruction or has more than one use then we
6851 // can't do anything more.
6852 Instruction *SrcI = dyn_cast<Instruction>(Src);
6853 if (!SrcI || !Src->hasOneUse())
6856 // Attempt to propagate the cast into the instruction for int->int casts.
6857 int NumCastsRemoved = 0;
6858 if (!isa<BitCastInst>(CI) &&
6859 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6860 CI.getOpcode(), NumCastsRemoved)) {
6861 // If this cast is a truncate, evaluting in a different type always
6862 // eliminates the cast, so it is always a win. If this is a zero-extension,
6863 // we need to do an AND to maintain the clear top-part of the computation,
6864 // so we require that the input have eliminated at least one cast. If this
6865 // is a sign extension, we insert two new casts (to do the extension) so we
6866 // require that two casts have been eliminated.
6868 switch (CI.getOpcode()) {
6870 // All the others use floating point so we shouldn't actually
6871 // get here because of the check above.
6872 assert(0 && "Unknown cast type");
6873 case Instruction::Trunc:
6876 case Instruction::ZExt:
6877 DoXForm = NumCastsRemoved >= 1;
6879 case Instruction::SExt:
6880 DoXForm = NumCastsRemoved >= 2;
6885 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6886 CI.getOpcode() == Instruction::SExt);
6887 assert(Res->getType() == DestTy);
6888 switch (CI.getOpcode()) {
6889 default: assert(0 && "Unknown cast type!");
6890 case Instruction::Trunc:
6891 case Instruction::BitCast:
6892 // Just replace this cast with the result.
6893 return ReplaceInstUsesWith(CI, Res);
6894 case Instruction::ZExt: {
6895 // We need to emit an AND to clear the high bits.
6896 assert(SrcBitSize < DestBitSize && "Not a zext?");
6897 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6899 return BinaryOperator::createAnd(Res, C);
6901 case Instruction::SExt:
6902 // We need to emit a cast to truncate, then a cast to sext.
6903 return CastInst::create(Instruction::SExt,
6904 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6910 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6911 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6913 switch (SrcI->getOpcode()) {
6914 case Instruction::Add:
6915 case Instruction::Mul:
6916 case Instruction::And:
6917 case Instruction::Or:
6918 case Instruction::Xor:
6919 // If we are discarding information, rewrite.
6920 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6921 // Don't insert two casts if they cannot be eliminated. We allow
6922 // two casts to be inserted if the sizes are the same. This could
6923 // only be converting signedness, which is a noop.
6924 if (DestBitSize == SrcBitSize ||
6925 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6926 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6927 Instruction::CastOps opcode = CI.getOpcode();
6928 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6929 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6930 return BinaryOperator::create(
6931 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6935 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6936 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6937 SrcI->getOpcode() == Instruction::Xor &&
6938 Op1 == ConstantInt::getTrue() &&
6939 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6940 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6941 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6944 case Instruction::SDiv:
6945 case Instruction::UDiv:
6946 case Instruction::SRem:
6947 case Instruction::URem:
6948 // If we are just changing the sign, rewrite.
6949 if (DestBitSize == SrcBitSize) {
6950 // Don't insert two casts if they cannot be eliminated. We allow
6951 // two casts to be inserted if the sizes are the same. This could
6952 // only be converting signedness, which is a noop.
6953 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6954 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6955 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6957 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6959 return BinaryOperator::create(
6960 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6965 case Instruction::Shl:
6966 // Allow changing the sign of the source operand. Do not allow
6967 // changing the size of the shift, UNLESS the shift amount is a
6968 // constant. We must not change variable sized shifts to a smaller
6969 // size, because it is undefined to shift more bits out than exist
6971 if (DestBitSize == SrcBitSize ||
6972 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6973 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6974 Instruction::BitCast : Instruction::Trunc);
6975 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6976 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6977 return BinaryOperator::createShl(Op0c, Op1c);
6980 case Instruction::AShr:
6981 // If this is a signed shr, and if all bits shifted in are about to be
6982 // truncated off, turn it into an unsigned shr to allow greater
6984 if (DestBitSize < SrcBitSize &&
6985 isa<ConstantInt>(Op1)) {
6986 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6987 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6988 // Insert the new logical shift right.
6989 return BinaryOperator::createLShr(Op0, Op1);
6997 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6998 if (Instruction *Result = commonIntCastTransforms(CI))
7001 Value *Src = CI.getOperand(0);
7002 const Type *Ty = CI.getType();
7003 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7004 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7006 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7007 switch (SrcI->getOpcode()) {
7009 case Instruction::LShr:
7010 // We can shrink lshr to something smaller if we know the bits shifted in
7011 // are already zeros.
7012 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7013 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7015 // Get a mask for the bits shifting in.
7016 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7017 Value* SrcIOp0 = SrcI->getOperand(0);
7018 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7019 if (ShAmt >= DestBitWidth) // All zeros.
7020 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7022 // Okay, we can shrink this. Truncate the input, then return a new
7024 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7025 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7027 return BinaryOperator::createLShr(V1, V2);
7029 } else { // This is a variable shr.
7031 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7032 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7033 // loop-invariant and CSE'd.
7034 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7035 Value *One = ConstantInt::get(SrcI->getType(), 1);
7037 Value *V = InsertNewInstBefore(
7038 BinaryOperator::createShl(One, SrcI->getOperand(1),
7040 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7041 SrcI->getOperand(0),
7043 Value *Zero = Constant::getNullValue(V->getType());
7044 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7054 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7055 // If one of the common conversion will work ..
7056 if (Instruction *Result = commonIntCastTransforms(CI))
7059 Value *Src = CI.getOperand(0);
7061 // If this is a cast of a cast
7062 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7063 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7064 // types and if the sizes are just right we can convert this into a logical
7065 // 'and' which will be much cheaper than the pair of casts.
7066 if (isa<TruncInst>(CSrc)) {
7067 // Get the sizes of the types involved
7068 Value *A = CSrc->getOperand(0);
7069 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7070 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7071 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7072 // If we're actually extending zero bits and the trunc is a no-op
7073 if (MidSize < DstSize && SrcSize == DstSize) {
7074 // Replace both of the casts with an And of the type mask.
7075 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7076 Constant *AndConst = ConstantInt::get(AndValue);
7078 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7079 // Unfortunately, if the type changed, we need to cast it back.
7080 if (And->getType() != CI.getType()) {
7081 And->setName(CSrc->getName()+".mask");
7082 InsertNewInstBefore(And, CI);
7083 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7090 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7091 // If we are just checking for a icmp eq of a single bit and zext'ing it
7092 // to an integer, then shift the bit to the appropriate place and then
7093 // cast to integer to avoid the comparison.
7094 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7095 const APInt &Op1CV = Op1C->getValue();
7097 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7098 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7099 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7100 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7101 Value *In = ICI->getOperand(0);
7102 Value *Sh = ConstantInt::get(In->getType(),
7103 In->getType()->getPrimitiveSizeInBits()-1);
7104 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7105 In->getName()+".lobit"),
7107 if (In->getType() != CI.getType())
7108 In = CastInst::createIntegerCast(In, CI.getType(),
7109 false/*ZExt*/, "tmp", &CI);
7111 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7112 Constant *One = ConstantInt::get(In->getType(), 1);
7113 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7114 In->getName()+".not"),
7118 return ReplaceInstUsesWith(CI, In);
7123 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7124 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7125 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7126 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7127 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7128 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7129 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7130 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7131 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7132 // This only works for EQ and NE
7133 ICI->isEquality()) {
7134 // If Op1C some other power of two, convert:
7135 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7136 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7137 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7138 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7140 APInt KnownZeroMask(~KnownZero);
7141 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7142 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7143 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7144 // (X&4) == 2 --> false
7145 // (X&4) != 2 --> true
7146 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7147 Res = ConstantExpr::getZExt(Res, CI.getType());
7148 return ReplaceInstUsesWith(CI, Res);
7151 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7152 Value *In = ICI->getOperand(0);
7154 // Perform a logical shr by shiftamt.
7155 // Insert the shift to put the result in the low bit.
7156 In = InsertNewInstBefore(
7157 BinaryOperator::createLShr(In,
7158 ConstantInt::get(In->getType(), ShiftAmt),
7159 In->getName()+".lobit"), CI);
7162 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7163 Constant *One = ConstantInt::get(In->getType(), 1);
7164 In = BinaryOperator::createXor(In, One, "tmp");
7165 InsertNewInstBefore(cast<Instruction>(In), CI);
7168 if (CI.getType() == In->getType())
7169 return ReplaceInstUsesWith(CI, In);
7171 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7179 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7180 if (Instruction *I = commonIntCastTransforms(CI))
7183 Value *Src = CI.getOperand(0);
7185 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7186 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7187 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7188 // If we are just checking for a icmp eq of a single bit and zext'ing it
7189 // to an integer, then shift the bit to the appropriate place and then
7190 // cast to integer to avoid the comparison.
7191 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7192 const APInt &Op1CV = Op1C->getValue();
7194 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7195 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7196 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7197 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7198 Value *In = ICI->getOperand(0);
7199 Value *Sh = ConstantInt::get(In->getType(),
7200 In->getType()->getPrimitiveSizeInBits()-1);
7201 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7202 In->getName()+".lobit"),
7204 if (In->getType() != CI.getType())
7205 In = CastInst::createIntegerCast(In, CI.getType(),
7206 true/*SExt*/, "tmp", &CI);
7208 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7209 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7210 In->getName()+".not"), CI);
7212 return ReplaceInstUsesWith(CI, In);
7220 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7221 /// in the specified FP type without changing its value.
7222 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7223 const fltSemantics &Sem) {
7224 APFloat F = CFP->getValueAPF();
7225 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7226 return ConstantFP::get(FPTy, F);
7230 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7231 /// through it until we get the source value.
7232 static Value *LookThroughFPExtensions(Value *V) {
7233 if (Instruction *I = dyn_cast<Instruction>(V))
7234 if (I->getOpcode() == Instruction::FPExt)
7235 return LookThroughFPExtensions(I->getOperand(0));
7237 // If this value is a constant, return the constant in the smallest FP type
7238 // that can accurately represent it. This allows us to turn
7239 // (float)((double)X+2.0) into x+2.0f.
7240 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7241 if (CFP->getType() == Type::PPC_FP128Ty)
7242 return V; // No constant folding of this.
7243 // See if the value can be truncated to float and then reextended.
7244 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7246 if (CFP->getType() == Type::DoubleTy)
7247 return V; // Won't shrink.
7248 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7250 // Don't try to shrink to various long double types.
7256 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7257 if (Instruction *I = commonCastTransforms(CI))
7260 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7261 // smaller than the destination type, we can eliminate the truncate by doing
7262 // the add as the smaller type. This applies to add/sub/mul/div as well as
7263 // many builtins (sqrt, etc).
7264 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7265 if (OpI && OpI->hasOneUse()) {
7266 switch (OpI->getOpcode()) {
7268 case Instruction::Add:
7269 case Instruction::Sub:
7270 case Instruction::Mul:
7271 case Instruction::FDiv:
7272 case Instruction::FRem:
7273 const Type *SrcTy = OpI->getType();
7274 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7275 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7276 if (LHSTrunc->getType() != SrcTy &&
7277 RHSTrunc->getType() != SrcTy) {
7278 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7279 // If the source types were both smaller than the destination type of
7280 // the cast, do this xform.
7281 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7282 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7283 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7285 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7287 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7296 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7297 return commonCastTransforms(CI);
7300 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7301 return commonCastTransforms(CI);
7304 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7305 return commonCastTransforms(CI);
7308 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7309 return commonCastTransforms(CI);
7312 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7313 return commonCastTransforms(CI);
7316 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7317 return commonPointerCastTransforms(CI);
7320 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7321 if (Instruction *I = commonCastTransforms(CI))
7324 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7325 if (!DestPointee->isSized()) return 0;
7327 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7330 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7331 m_ConstantInt(Cst)))) {
7332 // If the source and destination operands have the same type, see if this
7333 // is a single-index GEP.
7334 if (X->getType() == CI.getType()) {
7335 // Get the size of the pointee type.
7336 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7338 // Convert the constant to intptr type.
7339 APInt Offset = Cst->getValue();
7340 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7342 // If Offset is evenly divisible by Size, we can do this xform.
7343 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7344 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7345 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7348 // TODO: Could handle other cases, e.g. where add is indexing into field of
7350 } else if (CI.getOperand(0)->hasOneUse() &&
7351 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7352 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7353 // "inttoptr+GEP" instead of "add+intptr".
7355 // Get the size of the pointee type.
7356 uint64_t Size = TD->getABITypeSize(DestPointee);
7358 // Convert the constant to intptr type.
7359 APInt Offset = Cst->getValue();
7360 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7362 // If Offset is evenly divisible by Size, we can do this xform.
7363 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7364 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7366 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7368 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7374 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7375 // If the operands are integer typed then apply the integer transforms,
7376 // otherwise just apply the common ones.
7377 Value *Src = CI.getOperand(0);
7378 const Type *SrcTy = Src->getType();
7379 const Type *DestTy = CI.getType();
7381 if (SrcTy->isInteger() && DestTy->isInteger()) {
7382 if (Instruction *Result = commonIntCastTransforms(CI))
7384 } else if (isa<PointerType>(SrcTy)) {
7385 if (Instruction *I = commonPointerCastTransforms(CI))
7388 if (Instruction *Result = commonCastTransforms(CI))
7393 // Get rid of casts from one type to the same type. These are useless and can
7394 // be replaced by the operand.
7395 if (DestTy == Src->getType())
7396 return ReplaceInstUsesWith(CI, Src);
7398 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7399 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7400 const Type *DstElTy = DstPTy->getElementType();
7401 const Type *SrcElTy = SrcPTy->getElementType();
7403 // If we are casting a malloc or alloca to a pointer to a type of the same
7404 // size, rewrite the allocation instruction to allocate the "right" type.
7405 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7406 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7409 // If the source and destination are pointers, and this cast is equivalent
7410 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7411 // This can enhance SROA and other transforms that want type-safe pointers.
7412 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7413 unsigned NumZeros = 0;
7414 while (SrcElTy != DstElTy &&
7415 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7416 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7417 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7421 // If we found a path from the src to dest, create the getelementptr now.
7422 if (SrcElTy == DstElTy) {
7423 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7424 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7425 ((Instruction*) NULL));
7429 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7430 if (SVI->hasOneUse()) {
7431 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7432 // a bitconvert to a vector with the same # elts.
7433 if (isa<VectorType>(DestTy) &&
7434 cast<VectorType>(DestTy)->getNumElements() ==
7435 SVI->getType()->getNumElements()) {
7437 // If either of the operands is a cast from CI.getType(), then
7438 // evaluating the shuffle in the casted destination's type will allow
7439 // us to eliminate at least one cast.
7440 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7441 Tmp->getOperand(0)->getType() == DestTy) ||
7442 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7443 Tmp->getOperand(0)->getType() == DestTy)) {
7444 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7445 SVI->getOperand(0), DestTy, &CI);
7446 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7447 SVI->getOperand(1), DestTy, &CI);
7448 // Return a new shuffle vector. Use the same element ID's, as we
7449 // know the vector types match #elts.
7450 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7458 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7460 /// %D = select %cond, %C, %A
7462 /// %C = select %cond, %B, 0
7465 /// Assuming that the specified instruction is an operand to the select, return
7466 /// a bitmask indicating which operands of this instruction are foldable if they
7467 /// equal the other incoming value of the select.
7469 static unsigned GetSelectFoldableOperands(Instruction *I) {
7470 switch (I->getOpcode()) {
7471 case Instruction::Add:
7472 case Instruction::Mul:
7473 case Instruction::And:
7474 case Instruction::Or:
7475 case Instruction::Xor:
7476 return 3; // Can fold through either operand.
7477 case Instruction::Sub: // Can only fold on the amount subtracted.
7478 case Instruction::Shl: // Can only fold on the shift amount.
7479 case Instruction::LShr:
7480 case Instruction::AShr:
7483 return 0; // Cannot fold
7487 /// GetSelectFoldableConstant - For the same transformation as the previous
7488 /// function, return the identity constant that goes into the select.
7489 static Constant *GetSelectFoldableConstant(Instruction *I) {
7490 switch (I->getOpcode()) {
7491 default: assert(0 && "This cannot happen!"); abort();
7492 case Instruction::Add:
7493 case Instruction::Sub:
7494 case Instruction::Or:
7495 case Instruction::Xor:
7496 case Instruction::Shl:
7497 case Instruction::LShr:
7498 case Instruction::AShr:
7499 return Constant::getNullValue(I->getType());
7500 case Instruction::And:
7501 return Constant::getAllOnesValue(I->getType());
7502 case Instruction::Mul:
7503 return ConstantInt::get(I->getType(), 1);
7507 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7508 /// have the same opcode and only one use each. Try to simplify this.
7509 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7511 if (TI->getNumOperands() == 1) {
7512 // If this is a non-volatile load or a cast from the same type,
7515 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7518 return 0; // unknown unary op.
7521 // Fold this by inserting a select from the input values.
7522 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7523 FI->getOperand(0), SI.getName()+".v");
7524 InsertNewInstBefore(NewSI, SI);
7525 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7529 // Only handle binary operators here.
7530 if (!isa<BinaryOperator>(TI))
7533 // Figure out if the operations have any operands in common.
7534 Value *MatchOp, *OtherOpT, *OtherOpF;
7536 if (TI->getOperand(0) == FI->getOperand(0)) {
7537 MatchOp = TI->getOperand(0);
7538 OtherOpT = TI->getOperand(1);
7539 OtherOpF = FI->getOperand(1);
7540 MatchIsOpZero = true;
7541 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7542 MatchOp = TI->getOperand(1);
7543 OtherOpT = TI->getOperand(0);
7544 OtherOpF = FI->getOperand(0);
7545 MatchIsOpZero = false;
7546 } else if (!TI->isCommutative()) {
7548 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7549 MatchOp = TI->getOperand(0);
7550 OtherOpT = TI->getOperand(1);
7551 OtherOpF = FI->getOperand(0);
7552 MatchIsOpZero = true;
7553 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7554 MatchOp = TI->getOperand(1);
7555 OtherOpT = TI->getOperand(0);
7556 OtherOpF = FI->getOperand(1);
7557 MatchIsOpZero = true;
7562 // If we reach here, they do have operations in common.
7563 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7564 OtherOpF, SI.getName()+".v");
7565 InsertNewInstBefore(NewSI, SI);
7567 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7569 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7571 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7573 assert(0 && "Shouldn't get here");
7577 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7578 Value *CondVal = SI.getCondition();
7579 Value *TrueVal = SI.getTrueValue();
7580 Value *FalseVal = SI.getFalseValue();
7582 // select true, X, Y -> X
7583 // select false, X, Y -> Y
7584 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7585 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7587 // select C, X, X -> X
7588 if (TrueVal == FalseVal)
7589 return ReplaceInstUsesWith(SI, TrueVal);
7591 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7592 return ReplaceInstUsesWith(SI, FalseVal);
7593 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7594 return ReplaceInstUsesWith(SI, TrueVal);
7595 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7596 if (isa<Constant>(TrueVal))
7597 return ReplaceInstUsesWith(SI, TrueVal);
7599 return ReplaceInstUsesWith(SI, FalseVal);
7602 if (SI.getType() == Type::Int1Ty) {
7603 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7604 if (C->getZExtValue()) {
7605 // Change: A = select B, true, C --> A = or B, C
7606 return BinaryOperator::createOr(CondVal, FalseVal);
7608 // Change: A = select B, false, C --> A = and !B, C
7610 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7611 "not."+CondVal->getName()), SI);
7612 return BinaryOperator::createAnd(NotCond, FalseVal);
7614 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7615 if (C->getZExtValue() == false) {
7616 // Change: A = select B, C, false --> A = and B, C
7617 return BinaryOperator::createAnd(CondVal, TrueVal);
7619 // Change: A = select B, C, true --> A = or !B, C
7621 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7622 "not."+CondVal->getName()), SI);
7623 return BinaryOperator::createOr(NotCond, TrueVal);
7627 // select a, b, a -> a&b
7628 // select a, a, b -> a|b
7629 if (CondVal == TrueVal)
7630 return BinaryOperator::createOr(CondVal, FalseVal);
7631 else if (CondVal == FalseVal)
7632 return BinaryOperator::createAnd(CondVal, TrueVal);
7635 // Selecting between two integer constants?
7636 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7637 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7638 // select C, 1, 0 -> zext C to int
7639 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7640 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7641 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7642 // select C, 0, 1 -> zext !C to int
7644 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7645 "not."+CondVal->getName()), SI);
7646 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7649 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7651 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7653 // (x <s 0) ? -1 : 0 -> ashr x, 31
7654 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7655 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7656 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7657 // The comparison constant and the result are not neccessarily the
7658 // same width. Make an all-ones value by inserting a AShr.
7659 Value *X = IC->getOperand(0);
7660 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7661 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7662 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7664 InsertNewInstBefore(SRA, SI);
7666 // Finally, convert to the type of the select RHS. We figure out
7667 // if this requires a SExt, Trunc or BitCast based on the sizes.
7668 Instruction::CastOps opc = Instruction::BitCast;
7669 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7670 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7671 if (SRASize < SISize)
7672 opc = Instruction::SExt;
7673 else if (SRASize > SISize)
7674 opc = Instruction::Trunc;
7675 return CastInst::create(opc, SRA, SI.getType());
7680 // If one of the constants is zero (we know they can't both be) and we
7681 // have an icmp instruction with zero, and we have an 'and' with the
7682 // non-constant value, eliminate this whole mess. This corresponds to
7683 // cases like this: ((X & 27) ? 27 : 0)
7684 if (TrueValC->isZero() || FalseValC->isZero())
7685 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7686 cast<Constant>(IC->getOperand(1))->isNullValue())
7687 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7688 if (ICA->getOpcode() == Instruction::And &&
7689 isa<ConstantInt>(ICA->getOperand(1)) &&
7690 (ICA->getOperand(1) == TrueValC ||
7691 ICA->getOperand(1) == FalseValC) &&
7692 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7693 // Okay, now we know that everything is set up, we just don't
7694 // know whether we have a icmp_ne or icmp_eq and whether the
7695 // true or false val is the zero.
7696 bool ShouldNotVal = !TrueValC->isZero();
7697 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7700 V = InsertNewInstBefore(BinaryOperator::create(
7701 Instruction::Xor, V, ICA->getOperand(1)), SI);
7702 return ReplaceInstUsesWith(SI, V);
7707 // See if we are selecting two values based on a comparison of the two values.
7708 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7709 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7710 // Transform (X == Y) ? X : Y -> Y
7711 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7712 // This is not safe in general for floating point:
7713 // consider X== -0, Y== +0.
7714 // It becomes safe if either operand is a nonzero constant.
7715 ConstantFP *CFPt, *CFPf;
7716 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7717 !CFPt->getValueAPF().isZero()) ||
7718 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7719 !CFPf->getValueAPF().isZero()))
7720 return ReplaceInstUsesWith(SI, FalseVal);
7722 // Transform (X != Y) ? X : Y -> X
7723 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7724 return ReplaceInstUsesWith(SI, TrueVal);
7725 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7727 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7728 // Transform (X == Y) ? Y : X -> X
7729 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7730 // This is not safe in general for floating point:
7731 // consider X== -0, Y== +0.
7732 // It becomes safe if either operand is a nonzero constant.
7733 ConstantFP *CFPt, *CFPf;
7734 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7735 !CFPt->getValueAPF().isZero()) ||
7736 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7737 !CFPf->getValueAPF().isZero()))
7738 return ReplaceInstUsesWith(SI, FalseVal);
7740 // Transform (X != Y) ? Y : X -> Y
7741 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7742 return ReplaceInstUsesWith(SI, TrueVal);
7743 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7747 // See if we are selecting two values based on a comparison of the two values.
7748 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7749 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7750 // Transform (X == Y) ? X : Y -> Y
7751 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7752 return ReplaceInstUsesWith(SI, FalseVal);
7753 // Transform (X != Y) ? X : Y -> X
7754 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7755 return ReplaceInstUsesWith(SI, TrueVal);
7756 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7758 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7759 // Transform (X == Y) ? Y : X -> X
7760 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7761 return ReplaceInstUsesWith(SI, FalseVal);
7762 // Transform (X != Y) ? Y : X -> Y
7763 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7764 return ReplaceInstUsesWith(SI, TrueVal);
7765 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7769 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7770 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7771 if (TI->hasOneUse() && FI->hasOneUse()) {
7772 Instruction *AddOp = 0, *SubOp = 0;
7774 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7775 if (TI->getOpcode() == FI->getOpcode())
7776 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7779 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7780 // even legal for FP.
7781 if (TI->getOpcode() == Instruction::Sub &&
7782 FI->getOpcode() == Instruction::Add) {
7783 AddOp = FI; SubOp = TI;
7784 } else if (FI->getOpcode() == Instruction::Sub &&
7785 TI->getOpcode() == Instruction::Add) {
7786 AddOp = TI; SubOp = FI;
7790 Value *OtherAddOp = 0;
7791 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7792 OtherAddOp = AddOp->getOperand(1);
7793 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7794 OtherAddOp = AddOp->getOperand(0);
7798 // So at this point we know we have (Y -> OtherAddOp):
7799 // select C, (add X, Y), (sub X, Z)
7800 Value *NegVal; // Compute -Z
7801 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7802 NegVal = ConstantExpr::getNeg(C);
7804 NegVal = InsertNewInstBefore(
7805 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7808 Value *NewTrueOp = OtherAddOp;
7809 Value *NewFalseOp = NegVal;
7811 std::swap(NewTrueOp, NewFalseOp);
7812 Instruction *NewSel =
7813 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7815 NewSel = InsertNewInstBefore(NewSel, SI);
7816 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7821 // See if we can fold the select into one of our operands.
7822 if (SI.getType()->isInteger()) {
7823 // See the comment above GetSelectFoldableOperands for a description of the
7824 // transformation we are doing here.
7825 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7826 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7827 !isa<Constant>(FalseVal))
7828 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7829 unsigned OpToFold = 0;
7830 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7832 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7837 Constant *C = GetSelectFoldableConstant(TVI);
7838 Instruction *NewSel =
7839 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7840 InsertNewInstBefore(NewSel, SI);
7841 NewSel->takeName(TVI);
7842 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7843 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7845 assert(0 && "Unknown instruction!!");
7850 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7851 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7852 !isa<Constant>(TrueVal))
7853 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7854 unsigned OpToFold = 0;
7855 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7857 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7862 Constant *C = GetSelectFoldableConstant(FVI);
7863 Instruction *NewSel =
7864 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7865 InsertNewInstBefore(NewSel, SI);
7866 NewSel->takeName(FVI);
7867 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7868 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7870 assert(0 && "Unknown instruction!!");
7875 if (BinaryOperator::isNot(CondVal)) {
7876 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7877 SI.setOperand(1, FalseVal);
7878 SI.setOperand(2, TrueVal);
7885 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7886 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7887 /// and it is more than the alignment of the ultimate object, see if we can
7888 /// increase the alignment of the ultimate object, making this check succeed.
7889 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7890 unsigned PrefAlign = 0) {
7891 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7892 unsigned Align = GV->getAlignment();
7893 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7894 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7896 // If there is a large requested alignment and we can, bump up the alignment
7898 if (PrefAlign > Align && GV->hasInitializer()) {
7899 GV->setAlignment(PrefAlign);
7903 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7904 unsigned Align = AI->getAlignment();
7905 if (Align == 0 && TD) {
7906 if (isa<AllocaInst>(AI))
7907 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7908 else if (isa<MallocInst>(AI)) {
7909 // Malloc returns maximally aligned memory.
7910 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7913 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7916 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7920 // If there is a requested alignment and if this is an alloca, round up. We
7921 // don't do this for malloc, because some systems can't respect the request.
7922 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7923 AI->setAlignment(PrefAlign);
7927 } else if (isa<BitCastInst>(V) ||
7928 (isa<ConstantExpr>(V) &&
7929 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7930 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7932 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7933 // If all indexes are zero, it is just the alignment of the base pointer.
7934 bool AllZeroOperands = true;
7935 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7936 if (!isa<Constant>(GEPI->getOperand(i)) ||
7937 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7938 AllZeroOperands = false;
7942 if (AllZeroOperands) {
7943 // Treat this like a bitcast.
7944 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7947 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7948 if (BaseAlignment == 0) return 0;
7950 // Otherwise, if the base alignment is >= the alignment we expect for the
7951 // base pointer type, then we know that the resultant pointer is aligned at
7952 // least as much as its type requires.
7955 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7956 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7957 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7958 if (Align <= BaseAlignment) {
7959 const Type *GEPTy = GEPI->getType();
7960 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7961 Align = std::min(Align, (unsigned)
7962 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7970 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
7971 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7972 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7973 unsigned MinAlign = std::min(DstAlign, SrcAlign);
7974 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
7976 if (CopyAlign < MinAlign) {
7977 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
7981 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7983 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
7984 if (MemOpLength == 0) return 0;
7986 // Source and destination pointer types are always "i8*" for intrinsic. See
7987 // if the size is something we can handle with a single primitive load/store.
7988 // A single load+store correctly handles overlapping memory in the memmove
7990 unsigned Size = MemOpLength->getZExtValue();
7991 if (Size == 0 || Size > 8 || (Size&(Size-1)))
7992 return 0; // If not 1/2/4/8 bytes, exit.
7994 // Use an integer load+store unless we can find something better.
7995 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7997 // Memcpy forces the use of i8* for the source and destination. That means
7998 // that if you're using memcpy to move one double around, you'll get a cast
7999 // from double* to i8*. We'd much rather use a double load+store rather than
8000 // an i64 load+store, here because this improves the odds that the source or
8001 // dest address will be promotable. See if we can find a better type than the
8002 // integer datatype.
8003 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8004 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8005 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8006 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8007 // down through these levels if so.
8008 while (!SrcETy->isFirstClassType()) {
8009 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8010 if (STy->getNumElements() == 1)
8011 SrcETy = STy->getElementType(0);
8014 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8015 if (ATy->getNumElements() == 1)
8016 SrcETy = ATy->getElementType();
8023 if (SrcETy->isFirstClassType())
8024 NewPtrTy = PointerType::getUnqual(SrcETy);
8029 // If the memcpy/memmove provides better alignment info than we can
8031 SrcAlign = std::max(SrcAlign, CopyAlign);
8032 DstAlign = std::max(DstAlign, CopyAlign);
8034 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8035 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8036 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8037 InsertNewInstBefore(L, *MI);
8038 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8040 // Set the size of the copy to 0, it will be deleted on the next iteration.
8041 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8045 /// visitCallInst - CallInst simplification. This mostly only handles folding
8046 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8047 /// the heavy lifting.
8049 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8050 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8051 if (!II) return visitCallSite(&CI);
8053 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8055 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8056 bool Changed = false;
8058 // memmove/cpy/set of zero bytes is a noop.
8059 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8060 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8062 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8063 if (CI->getZExtValue() == 1) {
8064 // Replace the instruction with just byte operations. We would
8065 // transform other cases to loads/stores, but we don't know if
8066 // alignment is sufficient.
8070 // If we have a memmove and the source operation is a constant global,
8071 // then the source and dest pointers can't alias, so we can change this
8072 // into a call to memcpy.
8073 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8074 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8075 if (GVSrc->isConstant()) {
8076 Module *M = CI.getParent()->getParent()->getParent();
8077 Intrinsic::ID MemCpyID;
8078 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8079 MemCpyID = Intrinsic::memcpy_i32;
8081 MemCpyID = Intrinsic::memcpy_i64;
8082 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8087 // If we can determine a pointer alignment that is bigger than currently
8088 // set, update the alignment.
8089 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8090 if (Instruction *I = SimplifyMemTransfer(MI))
8092 } else if (isa<MemSetInst>(MI)) {
8093 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8094 if (MI->getAlignment()->getZExtValue() < Alignment) {
8095 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8100 if (Changed) return II;
8102 switch (II->getIntrinsicID()) {
8104 case Intrinsic::ppc_altivec_lvx:
8105 case Intrinsic::ppc_altivec_lvxl:
8106 case Intrinsic::x86_sse_loadu_ps:
8107 case Intrinsic::x86_sse2_loadu_pd:
8108 case Intrinsic::x86_sse2_loadu_dq:
8109 // Turn PPC lvx -> load if the pointer is known aligned.
8110 // Turn X86 loadups -> load if the pointer is known aligned.
8111 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8112 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8113 PointerType::getUnqual(II->getType()),
8115 return new LoadInst(Ptr);
8118 case Intrinsic::ppc_altivec_stvx:
8119 case Intrinsic::ppc_altivec_stvxl:
8120 // Turn stvx -> store if the pointer is known aligned.
8121 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8122 const Type *OpPtrTy =
8123 PointerType::getUnqual(II->getOperand(1)->getType());
8124 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8125 return new StoreInst(II->getOperand(1), Ptr);
8128 case Intrinsic::x86_sse_storeu_ps:
8129 case Intrinsic::x86_sse2_storeu_pd:
8130 case Intrinsic::x86_sse2_storeu_dq:
8131 case Intrinsic::x86_sse2_storel_dq:
8132 // Turn X86 storeu -> store if the pointer is known aligned.
8133 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8134 const Type *OpPtrTy =
8135 PointerType::getUnqual(II->getOperand(2)->getType());
8136 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8137 return new StoreInst(II->getOperand(2), Ptr);
8141 case Intrinsic::x86_sse_cvttss2si: {
8142 // These intrinsics only demands the 0th element of its input vector. If
8143 // we can simplify the input based on that, do so now.
8145 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8147 II->setOperand(1, V);
8153 case Intrinsic::ppc_altivec_vperm:
8154 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8155 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8156 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8158 // Check that all of the elements are integer constants or undefs.
8159 bool AllEltsOk = true;
8160 for (unsigned i = 0; i != 16; ++i) {
8161 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8162 !isa<UndefValue>(Mask->getOperand(i))) {
8169 // Cast the input vectors to byte vectors.
8170 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8171 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8172 Value *Result = UndefValue::get(Op0->getType());
8174 // Only extract each element once.
8175 Value *ExtractedElts[32];
8176 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8178 for (unsigned i = 0; i != 16; ++i) {
8179 if (isa<UndefValue>(Mask->getOperand(i)))
8181 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8182 Idx &= 31; // Match the hardware behavior.
8184 if (ExtractedElts[Idx] == 0) {
8186 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8187 InsertNewInstBefore(Elt, CI);
8188 ExtractedElts[Idx] = Elt;
8191 // Insert this value into the result vector.
8192 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8193 InsertNewInstBefore(cast<Instruction>(Result), CI);
8195 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8200 case Intrinsic::stackrestore: {
8201 // If the save is right next to the restore, remove the restore. This can
8202 // happen when variable allocas are DCE'd.
8203 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8204 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8205 BasicBlock::iterator BI = SS;
8207 return EraseInstFromFunction(CI);
8211 // If the stack restore is in a return/unwind block and if there are no
8212 // allocas or calls between the restore and the return, nuke the restore.
8213 TerminatorInst *TI = II->getParent()->getTerminator();
8214 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
8215 BasicBlock::iterator BI = II;
8216 bool CannotRemove = false;
8217 for (++BI; &*BI != TI; ++BI) {
8218 if (isa<AllocaInst>(BI) ||
8219 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
8220 CannotRemove = true;
8225 return EraseInstFromFunction(CI);
8232 return visitCallSite(II);
8235 // InvokeInst simplification
8237 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8238 return visitCallSite(&II);
8241 // visitCallSite - Improvements for call and invoke instructions.
8243 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8244 bool Changed = false;
8246 // If the callee is a constexpr cast of a function, attempt to move the cast
8247 // to the arguments of the call/invoke.
8248 if (transformConstExprCastCall(CS)) return 0;
8250 Value *Callee = CS.getCalledValue();
8252 if (Function *CalleeF = dyn_cast<Function>(Callee))
8253 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8254 Instruction *OldCall = CS.getInstruction();
8255 // If the call and callee calling conventions don't match, this call must
8256 // be unreachable, as the call is undefined.
8257 new StoreInst(ConstantInt::getTrue(),
8258 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8260 if (!OldCall->use_empty())
8261 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8262 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8263 return EraseInstFromFunction(*OldCall);
8267 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8268 // This instruction is not reachable, just remove it. We insert a store to
8269 // undef so that we know that this code is not reachable, despite the fact
8270 // that we can't modify the CFG here.
8271 new StoreInst(ConstantInt::getTrue(),
8272 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8273 CS.getInstruction());
8275 if (!CS.getInstruction()->use_empty())
8276 CS.getInstruction()->
8277 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8279 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8280 // Don't break the CFG, insert a dummy cond branch.
8281 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8282 ConstantInt::getTrue(), II);
8284 return EraseInstFromFunction(*CS.getInstruction());
8287 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8288 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8289 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8290 return transformCallThroughTrampoline(CS);
8292 const PointerType *PTy = cast<PointerType>(Callee->getType());
8293 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8294 if (FTy->isVarArg()) {
8295 // See if we can optimize any arguments passed through the varargs area of
8297 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8298 E = CS.arg_end(); I != E; ++I)
8299 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8300 // If this cast does not effect the value passed through the varargs
8301 // area, we can eliminate the use of the cast.
8302 Value *Op = CI->getOperand(0);
8303 if (CI->isLosslessCast()) {
8310 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8311 // Inline asm calls cannot throw - mark them 'nounwind'.
8312 CS.setDoesNotThrow();
8316 return Changed ? CS.getInstruction() : 0;
8319 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8320 // attempt to move the cast to the arguments of the call/invoke.
8322 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8323 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8324 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8325 if (CE->getOpcode() != Instruction::BitCast ||
8326 !isa<Function>(CE->getOperand(0)))
8328 Function *Callee = cast<Function>(CE->getOperand(0));
8329 Instruction *Caller = CS.getInstruction();
8330 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8332 // Okay, this is a cast from a function to a different type. Unless doing so
8333 // would cause a type conversion of one of our arguments, change this call to
8334 // be a direct call with arguments casted to the appropriate types.
8336 const FunctionType *FT = Callee->getFunctionType();
8337 const Type *OldRetTy = Caller->getType();
8339 // Check to see if we are changing the return type...
8340 if (OldRetTy != FT->getReturnType()) {
8341 if (Callee->isDeclaration() && !Caller->use_empty() &&
8342 // Conversion is ok if changing from pointer to int of same size.
8343 !(isa<PointerType>(FT->getReturnType()) &&
8344 TD->getIntPtrType() == OldRetTy))
8345 return false; // Cannot transform this return value.
8347 if (!Caller->use_empty() &&
8348 // void -> non-void is handled specially
8349 FT->getReturnType() != Type::VoidTy &&
8350 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8351 return false; // Cannot transform this return value.
8353 if (CallerPAL && !Caller->use_empty()) {
8354 uint16_t RAttrs = CallerPAL->getParamAttrs(0);
8355 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8356 return false; // Attribute not compatible with transformed value.
8359 // If the callsite is an invoke instruction, and the return value is used by
8360 // a PHI node in a successor, we cannot change the return type of the call
8361 // because there is no place to put the cast instruction (without breaking
8362 // the critical edge). Bail out in this case.
8363 if (!Caller->use_empty())
8364 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8365 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8367 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8368 if (PN->getParent() == II->getNormalDest() ||
8369 PN->getParent() == II->getUnwindDest())
8373 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8374 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8376 CallSite::arg_iterator AI = CS.arg_begin();
8377 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8378 const Type *ParamTy = FT->getParamType(i);
8379 const Type *ActTy = (*AI)->getType();
8381 if (!CastInst::isCastable(ActTy, ParamTy))
8382 return false; // Cannot transform this parameter value.
8385 uint16_t PAttrs = CallerPAL->getParamAttrs(i + 1);
8386 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8387 return false; // Attribute not compatible with transformed value.
8390 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8391 // Some conversions are safe even if we do not have a body.
8392 // Either we can cast directly, or we can upconvert the argument
8393 bool isConvertible = ActTy == ParamTy ||
8394 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8395 (ParamTy->isInteger() && ActTy->isInteger() &&
8396 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8397 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8398 && c->getValue().isStrictlyPositive());
8399 if (Callee->isDeclaration() && !isConvertible) return false;
8402 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8403 Callee->isDeclaration())
8404 return false; // Do not delete arguments unless we have a function body...
8406 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8407 // In this case we have more arguments than the new function type, but we
8408 // won't be dropping them. Check that these extra arguments have attributes
8409 // that are compatible with being a vararg call argument.
8410 for (unsigned i = CallerPAL->size(); i; --i) {
8411 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8413 uint16_t PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8414 if (PAttrs & ParamAttr::VarArgsIncompatible)
8418 // Okay, we decided that this is a safe thing to do: go ahead and start
8419 // inserting cast instructions as necessary...
8420 std::vector<Value*> Args;
8421 Args.reserve(NumActualArgs);
8422 ParamAttrsVector attrVec;
8423 attrVec.reserve(NumCommonArgs);
8425 // Get any return attributes.
8426 uint16_t RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) : 0;
8428 // If the return value is not being used, the type may not be compatible
8429 // with the existing attributes. Wipe out any problematic attributes.
8430 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8432 // Add the new return attributes.
8434 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8436 AI = CS.arg_begin();
8437 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8438 const Type *ParamTy = FT->getParamType(i);
8439 if ((*AI)->getType() == ParamTy) {
8440 Args.push_back(*AI);
8442 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8443 false, ParamTy, false);
8444 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8445 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8448 // Add any parameter attributes.
8449 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8451 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8454 // If the function takes more arguments than the call was taking, add them
8456 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8457 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8459 // If we are removing arguments to the function, emit an obnoxious warning...
8460 if (FT->getNumParams() < NumActualArgs)
8461 if (!FT->isVarArg()) {
8462 cerr << "WARNING: While resolving call to function '"
8463 << Callee->getName() << "' arguments were dropped!\n";
8465 // Add all of the arguments in their promoted form to the arg list...
8466 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8467 const Type *PTy = getPromotedType((*AI)->getType());
8468 if (PTy != (*AI)->getType()) {
8469 // Must promote to pass through va_arg area!
8470 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8472 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8473 InsertNewInstBefore(Cast, *Caller);
8474 Args.push_back(Cast);
8476 Args.push_back(*AI);
8479 // Add any parameter attributes.
8480 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8482 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8486 if (FT->getReturnType() == Type::VoidTy)
8487 Caller->setName(""); // Void type should not have a name.
8489 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8492 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8493 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8494 Args.begin(), Args.end(), Caller->getName(), Caller);
8495 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8496 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8498 NC = new CallInst(Callee, Args.begin(), Args.end(),
8499 Caller->getName(), Caller);
8500 CallInst *CI = cast<CallInst>(Caller);
8501 if (CI->isTailCall())
8502 cast<CallInst>(NC)->setTailCall();
8503 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8504 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8507 // Insert a cast of the return type as necessary.
8509 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8510 if (NV->getType() != Type::VoidTy) {
8511 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8513 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8515 // If this is an invoke instruction, we should insert it after the first
8516 // non-phi, instruction in the normal successor block.
8517 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8518 BasicBlock::iterator I = II->getNormalDest()->begin();
8519 while (isa<PHINode>(I)) ++I;
8520 InsertNewInstBefore(NC, *I);
8522 // Otherwise, it's a call, just insert cast right after the call instr
8523 InsertNewInstBefore(NC, *Caller);
8525 AddUsersToWorkList(*Caller);
8527 NV = UndefValue::get(Caller->getType());
8531 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8532 Caller->replaceAllUsesWith(NV);
8533 Caller->eraseFromParent();
8534 RemoveFromWorkList(Caller);
8538 // transformCallThroughTrampoline - Turn a call to a function created by the
8539 // init_trampoline intrinsic into a direct call to the underlying function.
8541 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8542 Value *Callee = CS.getCalledValue();
8543 const PointerType *PTy = cast<PointerType>(Callee->getType());
8544 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8545 const ParamAttrsList *Attrs = CS.getParamAttrs();
8547 // If the call already has the 'nest' attribute somewhere then give up -
8548 // otherwise 'nest' would occur twice after splicing in the chain.
8549 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8552 IntrinsicInst *Tramp =
8553 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8556 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8557 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8558 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8560 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8561 unsigned NestIdx = 1;
8562 const Type *NestTy = 0;
8563 uint16_t NestAttr = 0;
8565 // Look for a parameter marked with the 'nest' attribute.
8566 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8567 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8568 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8569 // Record the parameter type and any other attributes.
8571 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8576 Instruction *Caller = CS.getInstruction();
8577 std::vector<Value*> NewArgs;
8578 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8580 ParamAttrsVector NewAttrs;
8581 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8583 // Insert the nest argument into the call argument list, which may
8584 // mean appending it. Likewise for attributes.
8586 // Add any function result attributes.
8587 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8589 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8593 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8595 if (Idx == NestIdx) {
8596 // Add the chain argument and attributes.
8597 Value *NestVal = Tramp->getOperand(3);
8598 if (NestVal->getType() != NestTy)
8599 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8600 NewArgs.push_back(NestVal);
8601 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8607 // Add the original argument and attributes.
8608 NewArgs.push_back(*I);
8609 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8612 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8618 // The trampoline may have been bitcast to a bogus type (FTy).
8619 // Handle this by synthesizing a new function type, equal to FTy
8620 // with the chain parameter inserted.
8622 std::vector<const Type*> NewTypes;
8623 NewTypes.reserve(FTy->getNumParams()+1);
8625 // Insert the chain's type into the list of parameter types, which may
8626 // mean appending it.
8629 FunctionType::param_iterator I = FTy->param_begin(),
8630 E = FTy->param_end();
8634 // Add the chain's type.
8635 NewTypes.push_back(NestTy);
8640 // Add the original type.
8641 NewTypes.push_back(*I);
8647 // Replace the trampoline call with a direct call. Let the generic
8648 // code sort out any function type mismatches.
8649 FunctionType *NewFTy =
8650 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8651 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8652 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8653 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8655 Instruction *NewCaller;
8656 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8657 NewCaller = new InvokeInst(NewCallee,
8658 II->getNormalDest(), II->getUnwindDest(),
8659 NewArgs.begin(), NewArgs.end(),
8660 Caller->getName(), Caller);
8661 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8662 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8664 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8665 Caller->getName(), Caller);
8666 if (cast<CallInst>(Caller)->isTailCall())
8667 cast<CallInst>(NewCaller)->setTailCall();
8668 cast<CallInst>(NewCaller)->
8669 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8670 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8672 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8673 Caller->replaceAllUsesWith(NewCaller);
8674 Caller->eraseFromParent();
8675 RemoveFromWorkList(Caller);
8680 // Replace the trampoline call with a direct call. Since there is no 'nest'
8681 // parameter, there is no need to adjust the argument list. Let the generic
8682 // code sort out any function type mismatches.
8683 Constant *NewCallee =
8684 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8685 CS.setCalledFunction(NewCallee);
8686 return CS.getInstruction();
8689 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8690 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8691 /// and a single binop.
8692 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8693 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8694 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8695 isa<CmpInst>(FirstInst));
8696 unsigned Opc = FirstInst->getOpcode();
8697 Value *LHSVal = FirstInst->getOperand(0);
8698 Value *RHSVal = FirstInst->getOperand(1);
8700 const Type *LHSType = LHSVal->getType();
8701 const Type *RHSType = RHSVal->getType();
8703 // Scan to see if all operands are the same opcode, all have one use, and all
8704 // kill their operands (i.e. the operands have one use).
8705 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8706 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8707 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8708 // Verify type of the LHS matches so we don't fold cmp's of different
8709 // types or GEP's with different index types.
8710 I->getOperand(0)->getType() != LHSType ||
8711 I->getOperand(1)->getType() != RHSType)
8714 // If they are CmpInst instructions, check their predicates
8715 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8716 if (cast<CmpInst>(I)->getPredicate() !=
8717 cast<CmpInst>(FirstInst)->getPredicate())
8720 // Keep track of which operand needs a phi node.
8721 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8722 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8725 // Otherwise, this is safe to transform, determine if it is profitable.
8727 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8728 // Indexes are often folded into load/store instructions, so we don't want to
8729 // hide them behind a phi.
8730 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8733 Value *InLHS = FirstInst->getOperand(0);
8734 Value *InRHS = FirstInst->getOperand(1);
8735 PHINode *NewLHS = 0, *NewRHS = 0;
8737 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8738 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8739 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8740 InsertNewInstBefore(NewLHS, PN);
8745 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8746 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8747 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8748 InsertNewInstBefore(NewRHS, PN);
8752 // Add all operands to the new PHIs.
8753 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8755 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8756 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8759 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8760 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8764 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8765 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8766 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8767 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8770 assert(isa<GetElementPtrInst>(FirstInst));
8771 return new GetElementPtrInst(LHSVal, RHSVal);
8775 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8776 /// of the block that defines it. This means that it must be obvious the value
8777 /// of the load is not changed from the point of the load to the end of the
8780 /// Finally, it is safe, but not profitable, to sink a load targetting a
8781 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8783 static bool isSafeToSinkLoad(LoadInst *L) {
8784 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8786 for (++BBI; BBI != E; ++BBI)
8787 if (BBI->mayWriteToMemory())
8790 // Check for non-address taken alloca. If not address-taken already, it isn't
8791 // profitable to do this xform.
8792 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8793 bool isAddressTaken = false;
8794 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8796 if (isa<LoadInst>(UI)) continue;
8797 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8798 // If storing TO the alloca, then the address isn't taken.
8799 if (SI->getOperand(1) == AI) continue;
8801 isAddressTaken = true;
8805 if (!isAddressTaken)
8813 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8814 // operator and they all are only used by the PHI, PHI together their
8815 // inputs, and do the operation once, to the result of the PHI.
8816 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8817 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8819 // Scan the instruction, looking for input operations that can be folded away.
8820 // If all input operands to the phi are the same instruction (e.g. a cast from
8821 // the same type or "+42") we can pull the operation through the PHI, reducing
8822 // code size and simplifying code.
8823 Constant *ConstantOp = 0;
8824 const Type *CastSrcTy = 0;
8825 bool isVolatile = false;
8826 if (isa<CastInst>(FirstInst)) {
8827 CastSrcTy = FirstInst->getOperand(0)->getType();
8828 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8829 // Can fold binop, compare or shift here if the RHS is a constant,
8830 // otherwise call FoldPHIArgBinOpIntoPHI.
8831 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8832 if (ConstantOp == 0)
8833 return FoldPHIArgBinOpIntoPHI(PN);
8834 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8835 isVolatile = LI->isVolatile();
8836 // We can't sink the load if the loaded value could be modified between the
8837 // load and the PHI.
8838 if (LI->getParent() != PN.getIncomingBlock(0) ||
8839 !isSafeToSinkLoad(LI))
8841 } else if (isa<GetElementPtrInst>(FirstInst)) {
8842 if (FirstInst->getNumOperands() == 2)
8843 return FoldPHIArgBinOpIntoPHI(PN);
8844 // Can't handle general GEPs yet.
8847 return 0; // Cannot fold this operation.
8850 // Check to see if all arguments are the same operation.
8851 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8852 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8853 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8854 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8857 if (I->getOperand(0)->getType() != CastSrcTy)
8858 return 0; // Cast operation must match.
8859 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8860 // We can't sink the load if the loaded value could be modified between
8861 // the load and the PHI.
8862 if (LI->isVolatile() != isVolatile ||
8863 LI->getParent() != PN.getIncomingBlock(i) ||
8864 !isSafeToSinkLoad(LI))
8866 } else if (I->getOperand(1) != ConstantOp) {
8871 // Okay, they are all the same operation. Create a new PHI node of the
8872 // correct type, and PHI together all of the LHS's of the instructions.
8873 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8874 PN.getName()+".in");
8875 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8877 Value *InVal = FirstInst->getOperand(0);
8878 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8880 // Add all operands to the new PHI.
8881 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8882 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8883 if (NewInVal != InVal)
8885 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8890 // The new PHI unions all of the same values together. This is really
8891 // common, so we handle it intelligently here for compile-time speed.
8895 InsertNewInstBefore(NewPN, PN);
8899 // Insert and return the new operation.
8900 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8901 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8902 else if (isa<LoadInst>(FirstInst))
8903 return new LoadInst(PhiVal, "", isVolatile);
8904 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8905 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8906 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8907 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8908 PhiVal, ConstantOp);
8910 assert(0 && "Unknown operation");
8914 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8916 static bool DeadPHICycle(PHINode *PN,
8917 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8918 if (PN->use_empty()) return true;
8919 if (!PN->hasOneUse()) return false;
8921 // Remember this node, and if we find the cycle, return.
8922 if (!PotentiallyDeadPHIs.insert(PN))
8925 // Don't scan crazily complex things.
8926 if (PotentiallyDeadPHIs.size() == 16)
8929 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8930 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8935 /// PHIsEqualValue - Return true if this phi node is always equal to
8936 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8937 /// z = some value; x = phi (y, z); y = phi (x, z)
8938 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8939 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8940 // See if we already saw this PHI node.
8941 if (!ValueEqualPHIs.insert(PN))
8944 // Don't scan crazily complex things.
8945 if (ValueEqualPHIs.size() == 16)
8948 // Scan the operands to see if they are either phi nodes or are equal to
8950 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8951 Value *Op = PN->getIncomingValue(i);
8952 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8953 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8955 } else if (Op != NonPhiInVal)
8963 // PHINode simplification
8965 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8966 // If LCSSA is around, don't mess with Phi nodes
8967 if (MustPreserveLCSSA) return 0;
8969 if (Value *V = PN.hasConstantValue())
8970 return ReplaceInstUsesWith(PN, V);
8972 // If all PHI operands are the same operation, pull them through the PHI,
8973 // reducing code size.
8974 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8975 PN.getIncomingValue(0)->hasOneUse())
8976 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8979 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8980 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8981 // PHI)... break the cycle.
8982 if (PN.hasOneUse()) {
8983 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8984 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8985 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8986 PotentiallyDeadPHIs.insert(&PN);
8987 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8988 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8991 // If this phi has a single use, and if that use just computes a value for
8992 // the next iteration of a loop, delete the phi. This occurs with unused
8993 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8994 // common case here is good because the only other things that catch this
8995 // are induction variable analysis (sometimes) and ADCE, which is only run
8997 if (PHIUser->hasOneUse() &&
8998 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8999 PHIUser->use_back() == &PN) {
9000 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9004 // We sometimes end up with phi cycles that non-obviously end up being the
9005 // same value, for example:
9006 // z = some value; x = phi (y, z); y = phi (x, z)
9007 // where the phi nodes don't necessarily need to be in the same block. Do a
9008 // quick check to see if the PHI node only contains a single non-phi value, if
9009 // so, scan to see if the phi cycle is actually equal to that value.
9011 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9012 // Scan for the first non-phi operand.
9013 while (InValNo != NumOperandVals &&
9014 isa<PHINode>(PN.getIncomingValue(InValNo)))
9017 if (InValNo != NumOperandVals) {
9018 Value *NonPhiInVal = PN.getOperand(InValNo);
9020 // Scan the rest of the operands to see if there are any conflicts, if so
9021 // there is no need to recursively scan other phis.
9022 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9023 Value *OpVal = PN.getIncomingValue(InValNo);
9024 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9028 // If we scanned over all operands, then we have one unique value plus
9029 // phi values. Scan PHI nodes to see if they all merge in each other or
9031 if (InValNo == NumOperandVals) {
9032 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9033 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9034 return ReplaceInstUsesWith(PN, NonPhiInVal);
9041 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9042 Instruction *InsertPoint,
9044 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9045 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9046 // We must cast correctly to the pointer type. Ensure that we
9047 // sign extend the integer value if it is smaller as this is
9048 // used for address computation.
9049 Instruction::CastOps opcode =
9050 (VTySize < PtrSize ? Instruction::SExt :
9051 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9052 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9056 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9057 Value *PtrOp = GEP.getOperand(0);
9058 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9059 // If so, eliminate the noop.
9060 if (GEP.getNumOperands() == 1)
9061 return ReplaceInstUsesWith(GEP, PtrOp);
9063 if (isa<UndefValue>(GEP.getOperand(0)))
9064 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9066 bool HasZeroPointerIndex = false;
9067 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9068 HasZeroPointerIndex = C->isNullValue();
9070 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9071 return ReplaceInstUsesWith(GEP, PtrOp);
9073 // Eliminate unneeded casts for indices.
9074 bool MadeChange = false;
9076 gep_type_iterator GTI = gep_type_begin(GEP);
9077 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9078 if (isa<SequentialType>(*GTI)) {
9079 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9080 if (CI->getOpcode() == Instruction::ZExt ||
9081 CI->getOpcode() == Instruction::SExt) {
9082 const Type *SrcTy = CI->getOperand(0)->getType();
9083 // We can eliminate a cast from i32 to i64 iff the target
9084 // is a 32-bit pointer target.
9085 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9087 GEP.setOperand(i, CI->getOperand(0));
9091 // If we are using a wider index than needed for this platform, shrink it
9092 // to what we need. If the incoming value needs a cast instruction,
9093 // insert it. This explicit cast can make subsequent optimizations more
9095 Value *Op = GEP.getOperand(i);
9096 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
9097 if (Constant *C = dyn_cast<Constant>(Op)) {
9098 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9101 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9103 GEP.setOperand(i, Op);
9108 if (MadeChange) return &GEP;
9110 // If this GEP instruction doesn't move the pointer, and if the input operand
9111 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9112 // real input to the dest type.
9113 if (GEP.hasAllZeroIndices()) {
9114 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9115 // If the bitcast is of an allocation, and the allocation will be
9116 // converted to match the type of the cast, don't touch this.
9117 if (isa<AllocationInst>(BCI->getOperand(0))) {
9118 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9119 if (Instruction *I = visitBitCast(*BCI)) {
9122 BCI->getParent()->getInstList().insert(BCI, I);
9123 ReplaceInstUsesWith(*BCI, I);
9128 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9132 // Combine Indices - If the source pointer to this getelementptr instruction
9133 // is a getelementptr instruction, combine the indices of the two
9134 // getelementptr instructions into a single instruction.
9136 SmallVector<Value*, 8> SrcGEPOperands;
9137 if (User *Src = dyn_castGetElementPtr(PtrOp))
9138 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9140 if (!SrcGEPOperands.empty()) {
9141 // Note that if our source is a gep chain itself that we wait for that
9142 // chain to be resolved before we perform this transformation. This
9143 // avoids us creating a TON of code in some cases.
9145 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9146 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9147 return 0; // Wait until our source is folded to completion.
9149 SmallVector<Value*, 8> Indices;
9151 // Find out whether the last index in the source GEP is a sequential idx.
9152 bool EndsWithSequential = false;
9153 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9154 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9155 EndsWithSequential = !isa<StructType>(*I);
9157 // Can we combine the two pointer arithmetics offsets?
9158 if (EndsWithSequential) {
9159 // Replace: gep (gep %P, long B), long A, ...
9160 // With: T = long A+B; gep %P, T, ...
9162 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9163 if (SO1 == Constant::getNullValue(SO1->getType())) {
9165 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9168 // If they aren't the same type, convert both to an integer of the
9169 // target's pointer size.
9170 if (SO1->getType() != GO1->getType()) {
9171 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9172 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9173 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9174 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9176 unsigned PS = TD->getPointerSizeInBits();
9177 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9178 // Convert GO1 to SO1's type.
9179 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9181 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9182 // Convert SO1 to GO1's type.
9183 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9185 const Type *PT = TD->getIntPtrType();
9186 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9187 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9191 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9192 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9194 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9195 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9199 // Recycle the GEP we already have if possible.
9200 if (SrcGEPOperands.size() == 2) {
9201 GEP.setOperand(0, SrcGEPOperands[0]);
9202 GEP.setOperand(1, Sum);
9205 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9206 SrcGEPOperands.end()-1);
9207 Indices.push_back(Sum);
9208 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9210 } else if (isa<Constant>(*GEP.idx_begin()) &&
9211 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9212 SrcGEPOperands.size() != 1) {
9213 // Otherwise we can do the fold if the first index of the GEP is a zero
9214 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9215 SrcGEPOperands.end());
9216 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9219 if (!Indices.empty())
9220 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9221 Indices.end(), GEP.getName());
9223 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9224 // GEP of global variable. If all of the indices for this GEP are
9225 // constants, we can promote this to a constexpr instead of an instruction.
9227 // Scan for nonconstants...
9228 SmallVector<Constant*, 8> Indices;
9229 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9230 for (; I != E && isa<Constant>(*I); ++I)
9231 Indices.push_back(cast<Constant>(*I));
9233 if (I == E) { // If they are all constants...
9234 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9235 &Indices[0],Indices.size());
9237 // Replace all uses of the GEP with the new constexpr...
9238 return ReplaceInstUsesWith(GEP, CE);
9240 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9241 if (!isa<PointerType>(X->getType())) {
9242 // Not interesting. Source pointer must be a cast from pointer.
9243 } else if (HasZeroPointerIndex) {
9244 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9245 // into : GEP [10 x i8]* X, i32 0, ...
9247 // This occurs when the program declares an array extern like "int X[];"
9249 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9250 const PointerType *XTy = cast<PointerType>(X->getType());
9251 if (const ArrayType *XATy =
9252 dyn_cast<ArrayType>(XTy->getElementType()))
9253 if (const ArrayType *CATy =
9254 dyn_cast<ArrayType>(CPTy->getElementType()))
9255 if (CATy->getElementType() == XATy->getElementType()) {
9256 // At this point, we know that the cast source type is a pointer
9257 // to an array of the same type as the destination pointer
9258 // array. Because the array type is never stepped over (there
9259 // is a leading zero) we can fold the cast into this GEP.
9260 GEP.setOperand(0, X);
9263 } else if (GEP.getNumOperands() == 2) {
9264 // Transform things like:
9265 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9266 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9267 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9268 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9269 if (isa<ArrayType>(SrcElTy) &&
9270 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9271 TD->getABITypeSize(ResElTy)) {
9273 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9274 Idx[1] = GEP.getOperand(1);
9275 Value *V = InsertNewInstBefore(
9276 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9277 // V and GEP are both pointer types --> BitCast
9278 return new BitCastInst(V, GEP.getType());
9281 // Transform things like:
9282 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9283 // (where tmp = 8*tmp2) into:
9284 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9286 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9287 uint64_t ArrayEltSize =
9288 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9290 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9291 // allow either a mul, shift, or constant here.
9293 ConstantInt *Scale = 0;
9294 if (ArrayEltSize == 1) {
9295 NewIdx = GEP.getOperand(1);
9296 Scale = ConstantInt::get(NewIdx->getType(), 1);
9297 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9298 NewIdx = ConstantInt::get(CI->getType(), 1);
9300 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9301 if (Inst->getOpcode() == Instruction::Shl &&
9302 isa<ConstantInt>(Inst->getOperand(1))) {
9303 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9304 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9305 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9306 NewIdx = Inst->getOperand(0);
9307 } else if (Inst->getOpcode() == Instruction::Mul &&
9308 isa<ConstantInt>(Inst->getOperand(1))) {
9309 Scale = cast<ConstantInt>(Inst->getOperand(1));
9310 NewIdx = Inst->getOperand(0);
9314 // If the index will be to exactly the right offset with the scale taken
9315 // out, perform the transformation. Note, we don't know whether Scale is
9316 // signed or not. We'll use unsigned version of division/modulo
9317 // operation after making sure Scale doesn't have the sign bit set.
9318 if (Scale && Scale->getSExtValue() >= 0LL &&
9319 Scale->getZExtValue() % ArrayEltSize == 0) {
9320 Scale = ConstantInt::get(Scale->getType(),
9321 Scale->getZExtValue() / ArrayEltSize);
9322 if (Scale->getZExtValue() != 1) {
9323 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9325 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9326 NewIdx = InsertNewInstBefore(Sc, GEP);
9329 // Insert the new GEP instruction.
9331 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9333 Instruction *NewGEP =
9334 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9335 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9336 // The NewGEP must be pointer typed, so must the old one -> BitCast
9337 return new BitCastInst(NewGEP, GEP.getType());
9346 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9347 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9348 if (AI.isArrayAllocation()) // Check C != 1
9349 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9351 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9352 AllocationInst *New = 0;
9354 // Create and insert the replacement instruction...
9355 if (isa<MallocInst>(AI))
9356 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9358 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9359 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9362 InsertNewInstBefore(New, AI);
9364 // Scan to the end of the allocation instructions, to skip over a block of
9365 // allocas if possible...
9367 BasicBlock::iterator It = New;
9368 while (isa<AllocationInst>(*It)) ++It;
9370 // Now that I is pointing to the first non-allocation-inst in the block,
9371 // insert our getelementptr instruction...
9373 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9377 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9378 New->getName()+".sub", It);
9380 // Now make everything use the getelementptr instead of the original
9382 return ReplaceInstUsesWith(AI, V);
9383 } else if (isa<UndefValue>(AI.getArraySize())) {
9384 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9387 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9388 // Note that we only do this for alloca's, because malloc should allocate and
9389 // return a unique pointer, even for a zero byte allocation.
9390 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9391 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9392 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9397 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9398 Value *Op = FI.getOperand(0);
9400 // free undef -> unreachable.
9401 if (isa<UndefValue>(Op)) {
9402 // Insert a new store to null because we cannot modify the CFG here.
9403 new StoreInst(ConstantInt::getTrue(),
9404 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9405 return EraseInstFromFunction(FI);
9408 // If we have 'free null' delete the instruction. This can happen in stl code
9409 // when lots of inlining happens.
9410 if (isa<ConstantPointerNull>(Op))
9411 return EraseInstFromFunction(FI);
9413 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9414 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9415 FI.setOperand(0, CI->getOperand(0));
9419 // Change free (gep X, 0,0,0,0) into free(X)
9420 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9421 if (GEPI->hasAllZeroIndices()) {
9422 AddToWorkList(GEPI);
9423 FI.setOperand(0, GEPI->getOperand(0));
9428 // Change free(malloc) into nothing, if the malloc has a single use.
9429 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9430 if (MI->hasOneUse()) {
9431 EraseInstFromFunction(FI);
9432 return EraseInstFromFunction(*MI);
9439 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9440 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9441 const TargetData *TD) {
9442 User *CI = cast<User>(LI.getOperand(0));
9443 Value *CastOp = CI->getOperand(0);
9445 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9446 // Instead of loading constant c string, use corresponding integer value
9447 // directly if string length is small enough.
9448 const std::string &Str = CE->getOperand(0)->getStringValue();
9450 unsigned len = Str.length();
9451 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9452 unsigned numBits = Ty->getPrimitiveSizeInBits();
9453 // Replace LI with immediate integer store.
9454 if ((numBits >> 3) == len + 1) {
9455 APInt StrVal(numBits, 0);
9456 APInt SingleChar(numBits, 0);
9457 if (TD->isLittleEndian()) {
9458 for (signed i = len-1; i >= 0; i--) {
9459 SingleChar = (uint64_t) Str[i];
9460 StrVal = (StrVal << 8) | SingleChar;
9463 for (unsigned i = 0; i < len; i++) {
9464 SingleChar = (uint64_t) Str[i];
9465 StrVal = (StrVal << 8) | SingleChar;
9467 // Append NULL at the end.
9469 StrVal = (StrVal << 8) | SingleChar;
9471 Value *NL = ConstantInt::get(StrVal);
9472 return IC.ReplaceInstUsesWith(LI, NL);
9477 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9478 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9479 const Type *SrcPTy = SrcTy->getElementType();
9481 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9482 isa<VectorType>(DestPTy)) {
9483 // If the source is an array, the code below will not succeed. Check to
9484 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9486 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9487 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9488 if (ASrcTy->getNumElements() != 0) {
9490 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9491 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9492 SrcTy = cast<PointerType>(CastOp->getType());
9493 SrcPTy = SrcTy->getElementType();
9496 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9497 isa<VectorType>(SrcPTy)) &&
9498 // Do not allow turning this into a load of an integer, which is then
9499 // casted to a pointer, this pessimizes pointer analysis a lot.
9500 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9501 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9502 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9504 // Okay, we are casting from one integer or pointer type to another of
9505 // the same size. Instead of casting the pointer before the load, cast
9506 // the result of the loaded value.
9507 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9509 LI.isVolatile()),LI);
9510 // Now cast the result of the load.
9511 return new BitCastInst(NewLoad, LI.getType());
9518 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9519 /// from this value cannot trap. If it is not obviously safe to load from the
9520 /// specified pointer, we do a quick local scan of the basic block containing
9521 /// ScanFrom, to determine if the address is already accessed.
9522 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9523 // If it is an alloca it is always safe to load from.
9524 if (isa<AllocaInst>(V)) return true;
9526 // If it is a global variable it is mostly safe to load from.
9527 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9528 // Don't try to evaluate aliases. External weak GV can be null.
9529 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9531 // Otherwise, be a little bit agressive by scanning the local block where we
9532 // want to check to see if the pointer is already being loaded or stored
9533 // from/to. If so, the previous load or store would have already trapped,
9534 // so there is no harm doing an extra load (also, CSE will later eliminate
9535 // the load entirely).
9536 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9541 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9542 if (LI->getOperand(0) == V) return true;
9543 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9544 if (SI->getOperand(1) == V) return true;
9550 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9551 /// until we find the underlying object a pointer is referring to or something
9552 /// we don't understand. Note that the returned pointer may be offset from the
9553 /// input, because we ignore GEP indices.
9554 static Value *GetUnderlyingObject(Value *Ptr) {
9556 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9557 if (CE->getOpcode() == Instruction::BitCast ||
9558 CE->getOpcode() == Instruction::GetElementPtr)
9559 Ptr = CE->getOperand(0);
9562 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9563 Ptr = BCI->getOperand(0);
9564 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9565 Ptr = GEP->getOperand(0);
9572 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9573 Value *Op = LI.getOperand(0);
9575 // Attempt to improve the alignment.
9576 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9577 if (KnownAlign > LI.getAlignment())
9578 LI.setAlignment(KnownAlign);
9580 // load (cast X) --> cast (load X) iff safe
9581 if (isa<CastInst>(Op))
9582 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9585 // None of the following transforms are legal for volatile loads.
9586 if (LI.isVolatile()) return 0;
9588 if (&LI.getParent()->front() != &LI) {
9589 BasicBlock::iterator BBI = &LI; --BBI;
9590 // If the instruction immediately before this is a store to the same
9591 // address, do a simple form of store->load forwarding.
9592 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9593 if (SI->getOperand(1) == LI.getOperand(0))
9594 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9595 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9596 if (LIB->getOperand(0) == LI.getOperand(0))
9597 return ReplaceInstUsesWith(LI, LIB);
9600 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9601 const Value *GEPI0 = GEPI->getOperand(0);
9602 // TODO: Consider a target hook for valid address spaces for this xform.
9603 if (isa<ConstantPointerNull>(GEPI0) &&
9604 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9605 // Insert a new store to null instruction before the load to indicate
9606 // that this code is not reachable. We do this instead of inserting
9607 // an unreachable instruction directly because we cannot modify the
9609 new StoreInst(UndefValue::get(LI.getType()),
9610 Constant::getNullValue(Op->getType()), &LI);
9611 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9615 if (Constant *C = dyn_cast<Constant>(Op)) {
9616 // load null/undef -> undef
9617 // TODO: Consider a target hook for valid address spaces for this xform.
9618 if (isa<UndefValue>(C) || (C->isNullValue() &&
9619 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9620 // Insert a new store to null instruction before the load to indicate that
9621 // this code is not reachable. We do this instead of inserting an
9622 // unreachable instruction directly because we cannot modify the CFG.
9623 new StoreInst(UndefValue::get(LI.getType()),
9624 Constant::getNullValue(Op->getType()), &LI);
9625 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9628 // Instcombine load (constant global) into the value loaded.
9629 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9630 if (GV->isConstant() && !GV->isDeclaration())
9631 return ReplaceInstUsesWith(LI, GV->getInitializer());
9633 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9634 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9635 if (CE->getOpcode() == Instruction::GetElementPtr) {
9636 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9637 if (GV->isConstant() && !GV->isDeclaration())
9639 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9640 return ReplaceInstUsesWith(LI, V);
9641 if (CE->getOperand(0)->isNullValue()) {
9642 // Insert a new store to null instruction before the load to indicate
9643 // that this code is not reachable. We do this instead of inserting
9644 // an unreachable instruction directly because we cannot modify the
9646 new StoreInst(UndefValue::get(LI.getType()),
9647 Constant::getNullValue(Op->getType()), &LI);
9648 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9651 } else if (CE->isCast()) {
9652 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9657 // If this load comes from anywhere in a constant global, and if the global
9658 // is all undef or zero, we know what it loads.
9659 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9660 if (GV->isConstant() && GV->hasInitializer()) {
9661 if (GV->getInitializer()->isNullValue())
9662 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9663 else if (isa<UndefValue>(GV->getInitializer()))
9664 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9668 if (Op->hasOneUse()) {
9669 // Change select and PHI nodes to select values instead of addresses: this
9670 // helps alias analysis out a lot, allows many others simplifications, and
9671 // exposes redundancy in the code.
9673 // Note that we cannot do the transformation unless we know that the
9674 // introduced loads cannot trap! Something like this is valid as long as
9675 // the condition is always false: load (select bool %C, int* null, int* %G),
9676 // but it would not be valid if we transformed it to load from null
9679 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9680 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9681 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9682 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9683 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9684 SI->getOperand(1)->getName()+".val"), LI);
9685 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9686 SI->getOperand(2)->getName()+".val"), LI);
9687 return new SelectInst(SI->getCondition(), V1, V2);
9690 // load (select (cond, null, P)) -> load P
9691 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9692 if (C->isNullValue()) {
9693 LI.setOperand(0, SI->getOperand(2));
9697 // load (select (cond, P, null)) -> load P
9698 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9699 if (C->isNullValue()) {
9700 LI.setOperand(0, SI->getOperand(1));
9708 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9710 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9711 User *CI = cast<User>(SI.getOperand(1));
9712 Value *CastOp = CI->getOperand(0);
9714 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9715 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9716 const Type *SrcPTy = SrcTy->getElementType();
9718 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9719 // If the source is an array, the code below will not succeed. Check to
9720 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9722 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9723 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9724 if (ASrcTy->getNumElements() != 0) {
9726 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9727 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9728 SrcTy = cast<PointerType>(CastOp->getType());
9729 SrcPTy = SrcTy->getElementType();
9732 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9733 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9734 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9736 // Okay, we are casting from one integer or pointer type to another of
9737 // the same size. Instead of casting the pointer before
9738 // the store, cast the value to be stored.
9740 Value *SIOp0 = SI.getOperand(0);
9741 Instruction::CastOps opcode = Instruction::BitCast;
9742 const Type* CastSrcTy = SIOp0->getType();
9743 const Type* CastDstTy = SrcPTy;
9744 if (isa<PointerType>(CastDstTy)) {
9745 if (CastSrcTy->isInteger())
9746 opcode = Instruction::IntToPtr;
9747 } else if (isa<IntegerType>(CastDstTy)) {
9748 if (isa<PointerType>(SIOp0->getType()))
9749 opcode = Instruction::PtrToInt;
9751 if (Constant *C = dyn_cast<Constant>(SIOp0))
9752 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9754 NewCast = IC.InsertNewInstBefore(
9755 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9757 return new StoreInst(NewCast, CastOp);
9764 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9765 Value *Val = SI.getOperand(0);
9766 Value *Ptr = SI.getOperand(1);
9768 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9769 EraseInstFromFunction(SI);
9774 // If the RHS is an alloca with a single use, zapify the store, making the
9776 if (Ptr->hasOneUse()) {
9777 if (isa<AllocaInst>(Ptr)) {
9778 EraseInstFromFunction(SI);
9783 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9784 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9785 GEP->getOperand(0)->hasOneUse()) {
9786 EraseInstFromFunction(SI);
9792 // Attempt to improve the alignment.
9793 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9794 if (KnownAlign > SI.getAlignment())
9795 SI.setAlignment(KnownAlign);
9797 // Do really simple DSE, to catch cases where there are several consequtive
9798 // stores to the same location, separated by a few arithmetic operations. This
9799 // situation often occurs with bitfield accesses.
9800 BasicBlock::iterator BBI = &SI;
9801 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9805 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9806 // Prev store isn't volatile, and stores to the same location?
9807 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9810 EraseInstFromFunction(*PrevSI);
9816 // If this is a load, we have to stop. However, if the loaded value is from
9817 // the pointer we're loading and is producing the pointer we're storing,
9818 // then *this* store is dead (X = load P; store X -> P).
9819 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9820 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9821 EraseInstFromFunction(SI);
9825 // Otherwise, this is a load from some other location. Stores before it
9830 // Don't skip over loads or things that can modify memory.
9831 if (BBI->mayWriteToMemory())
9836 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9838 // store X, null -> turns into 'unreachable' in SimplifyCFG
9839 if (isa<ConstantPointerNull>(Ptr)) {
9840 if (!isa<UndefValue>(Val)) {
9841 SI.setOperand(0, UndefValue::get(Val->getType()));
9842 if (Instruction *U = dyn_cast<Instruction>(Val))
9843 AddToWorkList(U); // Dropped a use.
9846 return 0; // Do not modify these!
9849 // store undef, Ptr -> noop
9850 if (isa<UndefValue>(Val)) {
9851 EraseInstFromFunction(SI);
9856 // If the pointer destination is a cast, see if we can fold the cast into the
9858 if (isa<CastInst>(Ptr))
9859 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9861 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9863 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9867 // If this store is the last instruction in the basic block, and if the block
9868 // ends with an unconditional branch, try to move it to the successor block.
9870 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9871 if (BI->isUnconditional())
9872 if (SimplifyStoreAtEndOfBlock(SI))
9873 return 0; // xform done!
9878 /// SimplifyStoreAtEndOfBlock - Turn things like:
9879 /// if () { *P = v1; } else { *P = v2 }
9880 /// into a phi node with a store in the successor.
9882 /// Simplify things like:
9883 /// *P = v1; if () { *P = v2; }
9884 /// into a phi node with a store in the successor.
9886 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9887 BasicBlock *StoreBB = SI.getParent();
9889 // Check to see if the successor block has exactly two incoming edges. If
9890 // so, see if the other predecessor contains a store to the same location.
9891 // if so, insert a PHI node (if needed) and move the stores down.
9892 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9894 // Determine whether Dest has exactly two predecessors and, if so, compute
9895 // the other predecessor.
9896 pred_iterator PI = pred_begin(DestBB);
9897 BasicBlock *OtherBB = 0;
9901 if (PI == pred_end(DestBB))
9904 if (*PI != StoreBB) {
9909 if (++PI != pred_end(DestBB))
9913 // Verify that the other block ends in a branch and is not otherwise empty.
9914 BasicBlock::iterator BBI = OtherBB->getTerminator();
9915 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9916 if (!OtherBr || BBI == OtherBB->begin())
9919 // If the other block ends in an unconditional branch, check for the 'if then
9920 // else' case. there is an instruction before the branch.
9921 StoreInst *OtherStore = 0;
9922 if (OtherBr->isUnconditional()) {
9923 // If this isn't a store, or isn't a store to the same location, bail out.
9925 OtherStore = dyn_cast<StoreInst>(BBI);
9926 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9929 // Otherwise, the other block ended with a conditional branch. If one of the
9930 // destinations is StoreBB, then we have the if/then case.
9931 if (OtherBr->getSuccessor(0) != StoreBB &&
9932 OtherBr->getSuccessor(1) != StoreBB)
9935 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9936 // if/then triangle. See if there is a store to the same ptr as SI that
9937 // lives in OtherBB.
9939 // Check to see if we find the matching store.
9940 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9941 if (OtherStore->getOperand(1) != SI.getOperand(1))
9945 // If we find something that may be using the stored value, or if we run
9946 // out of instructions, we can't do the xform.
9947 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9948 BBI == OtherBB->begin())
9952 // In order to eliminate the store in OtherBr, we have to
9953 // make sure nothing reads the stored value in StoreBB.
9954 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9955 // FIXME: This should really be AA driven.
9956 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9961 // Insert a PHI node now if we need it.
9962 Value *MergedVal = OtherStore->getOperand(0);
9963 if (MergedVal != SI.getOperand(0)) {
9964 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9965 PN->reserveOperandSpace(2);
9966 PN->addIncoming(SI.getOperand(0), SI.getParent());
9967 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9968 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9971 // Advance to a place where it is safe to insert the new store and
9973 BBI = DestBB->begin();
9974 while (isa<PHINode>(BBI)) ++BBI;
9975 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9976 OtherStore->isVolatile()), *BBI);
9978 // Nuke the old stores.
9979 EraseInstFromFunction(SI);
9980 EraseInstFromFunction(*OtherStore);
9986 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9987 // Change br (not X), label True, label False to: br X, label False, True
9989 BasicBlock *TrueDest;
9990 BasicBlock *FalseDest;
9991 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9992 !isa<Constant>(X)) {
9993 // Swap Destinations and condition...
9995 BI.setSuccessor(0, FalseDest);
9996 BI.setSuccessor(1, TrueDest);
10000 // Cannonicalize fcmp_one -> fcmp_oeq
10001 FCmpInst::Predicate FPred; Value *Y;
10002 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10003 TrueDest, FalseDest)))
10004 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10005 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10006 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10007 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10008 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10009 NewSCC->takeName(I);
10010 // Swap Destinations and condition...
10011 BI.setCondition(NewSCC);
10012 BI.setSuccessor(0, FalseDest);
10013 BI.setSuccessor(1, TrueDest);
10014 RemoveFromWorkList(I);
10015 I->eraseFromParent();
10016 AddToWorkList(NewSCC);
10020 // Cannonicalize icmp_ne -> icmp_eq
10021 ICmpInst::Predicate IPred;
10022 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10023 TrueDest, FalseDest)))
10024 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10025 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10026 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10027 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10028 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10029 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10030 NewSCC->takeName(I);
10031 // Swap Destinations and condition...
10032 BI.setCondition(NewSCC);
10033 BI.setSuccessor(0, FalseDest);
10034 BI.setSuccessor(1, TrueDest);
10035 RemoveFromWorkList(I);
10036 I->eraseFromParent();;
10037 AddToWorkList(NewSCC);
10044 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10045 Value *Cond = SI.getCondition();
10046 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10047 if (I->getOpcode() == Instruction::Add)
10048 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10049 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10050 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10051 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10053 SI.setOperand(0, I->getOperand(0));
10061 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10062 /// is to leave as a vector operation.
10063 static bool CheapToScalarize(Value *V, bool isConstant) {
10064 if (isa<ConstantAggregateZero>(V))
10066 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10067 if (isConstant) return true;
10068 // If all elts are the same, we can extract.
10069 Constant *Op0 = C->getOperand(0);
10070 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10071 if (C->getOperand(i) != Op0)
10075 Instruction *I = dyn_cast<Instruction>(V);
10076 if (!I) return false;
10078 // Insert element gets simplified to the inserted element or is deleted if
10079 // this is constant idx extract element and its a constant idx insertelt.
10080 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10081 isa<ConstantInt>(I->getOperand(2)))
10083 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10085 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10086 if (BO->hasOneUse() &&
10087 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10088 CheapToScalarize(BO->getOperand(1), isConstant)))
10090 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10091 if (CI->hasOneUse() &&
10092 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10093 CheapToScalarize(CI->getOperand(1), isConstant)))
10099 /// Read and decode a shufflevector mask.
10101 /// It turns undef elements into values that are larger than the number of
10102 /// elements in the input.
10103 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10104 unsigned NElts = SVI->getType()->getNumElements();
10105 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10106 return std::vector<unsigned>(NElts, 0);
10107 if (isa<UndefValue>(SVI->getOperand(2)))
10108 return std::vector<unsigned>(NElts, 2*NElts);
10110 std::vector<unsigned> Result;
10111 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10112 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10113 if (isa<UndefValue>(CP->getOperand(i)))
10114 Result.push_back(NElts*2); // undef -> 8
10116 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10120 /// FindScalarElement - Given a vector and an element number, see if the scalar
10121 /// value is already around as a register, for example if it were inserted then
10122 /// extracted from the vector.
10123 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10124 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10125 const VectorType *PTy = cast<VectorType>(V->getType());
10126 unsigned Width = PTy->getNumElements();
10127 if (EltNo >= Width) // Out of range access.
10128 return UndefValue::get(PTy->getElementType());
10130 if (isa<UndefValue>(V))
10131 return UndefValue::get(PTy->getElementType());
10132 else if (isa<ConstantAggregateZero>(V))
10133 return Constant::getNullValue(PTy->getElementType());
10134 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10135 return CP->getOperand(EltNo);
10136 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10137 // If this is an insert to a variable element, we don't know what it is.
10138 if (!isa<ConstantInt>(III->getOperand(2)))
10140 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10142 // If this is an insert to the element we are looking for, return the
10144 if (EltNo == IIElt)
10145 return III->getOperand(1);
10147 // Otherwise, the insertelement doesn't modify the value, recurse on its
10149 return FindScalarElement(III->getOperand(0), EltNo);
10150 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10151 unsigned InEl = getShuffleMask(SVI)[EltNo];
10153 return FindScalarElement(SVI->getOperand(0), InEl);
10154 else if (InEl < Width*2)
10155 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10157 return UndefValue::get(PTy->getElementType());
10160 // Otherwise, we don't know.
10164 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10166 // If vector val is undef, replace extract with scalar undef.
10167 if (isa<UndefValue>(EI.getOperand(0)))
10168 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10170 // If vector val is constant 0, replace extract with scalar 0.
10171 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10172 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10174 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10175 // If vector val is constant with uniform operands, replace EI
10176 // with that operand
10177 Constant *op0 = C->getOperand(0);
10178 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10179 if (C->getOperand(i) != op0) {
10184 return ReplaceInstUsesWith(EI, op0);
10187 // If extracting a specified index from the vector, see if we can recursively
10188 // find a previously computed scalar that was inserted into the vector.
10189 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10190 unsigned IndexVal = IdxC->getZExtValue();
10191 unsigned VectorWidth =
10192 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10194 // If this is extracting an invalid index, turn this into undef, to avoid
10195 // crashing the code below.
10196 if (IndexVal >= VectorWidth)
10197 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10199 // This instruction only demands the single element from the input vector.
10200 // If the input vector has a single use, simplify it based on this use
10202 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10203 uint64_t UndefElts;
10204 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10207 EI.setOperand(0, V);
10212 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10213 return ReplaceInstUsesWith(EI, Elt);
10215 // If the this extractelement is directly using a bitcast from a vector of
10216 // the same number of elements, see if we can find the source element from
10217 // it. In this case, we will end up needing to bitcast the scalars.
10218 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10219 if (const VectorType *VT =
10220 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10221 if (VT->getNumElements() == VectorWidth)
10222 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10223 return new BitCastInst(Elt, EI.getType());
10227 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10228 if (I->hasOneUse()) {
10229 // Push extractelement into predecessor operation if legal and
10230 // profitable to do so
10231 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10232 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10233 if (CheapToScalarize(BO, isConstantElt)) {
10234 ExtractElementInst *newEI0 =
10235 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10236 EI.getName()+".lhs");
10237 ExtractElementInst *newEI1 =
10238 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10239 EI.getName()+".rhs");
10240 InsertNewInstBefore(newEI0, EI);
10241 InsertNewInstBefore(newEI1, EI);
10242 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10244 } else if (isa<LoadInst>(I)) {
10246 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10247 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10248 PointerType::get(EI.getType(), AS),EI);
10249 GetElementPtrInst *GEP =
10250 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10251 InsertNewInstBefore(GEP, EI);
10252 return new LoadInst(GEP);
10255 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10256 // Extracting the inserted element?
10257 if (IE->getOperand(2) == EI.getOperand(1))
10258 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10259 // If the inserted and extracted elements are constants, they must not
10260 // be the same value, extract from the pre-inserted value instead.
10261 if (isa<Constant>(IE->getOperand(2)) &&
10262 isa<Constant>(EI.getOperand(1))) {
10263 AddUsesToWorkList(EI);
10264 EI.setOperand(0, IE->getOperand(0));
10267 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10268 // If this is extracting an element from a shufflevector, figure out where
10269 // it came from and extract from the appropriate input element instead.
10270 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10271 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10273 if (SrcIdx < SVI->getType()->getNumElements())
10274 Src = SVI->getOperand(0);
10275 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10276 SrcIdx -= SVI->getType()->getNumElements();
10277 Src = SVI->getOperand(1);
10279 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10281 return new ExtractElementInst(Src, SrcIdx);
10288 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10289 /// elements from either LHS or RHS, return the shuffle mask and true.
10290 /// Otherwise, return false.
10291 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10292 std::vector<Constant*> &Mask) {
10293 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10294 "Invalid CollectSingleShuffleElements");
10295 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10297 if (isa<UndefValue>(V)) {
10298 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10300 } else if (V == LHS) {
10301 for (unsigned i = 0; i != NumElts; ++i)
10302 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10304 } else if (V == RHS) {
10305 for (unsigned i = 0; i != NumElts; ++i)
10306 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10308 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10309 // If this is an insert of an extract from some other vector, include it.
10310 Value *VecOp = IEI->getOperand(0);
10311 Value *ScalarOp = IEI->getOperand(1);
10312 Value *IdxOp = IEI->getOperand(2);
10314 if (!isa<ConstantInt>(IdxOp))
10316 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10318 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10319 // Okay, we can handle this if the vector we are insertinting into is
10320 // transitively ok.
10321 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10322 // If so, update the mask to reflect the inserted undef.
10323 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10326 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10327 if (isa<ConstantInt>(EI->getOperand(1)) &&
10328 EI->getOperand(0)->getType() == V->getType()) {
10329 unsigned ExtractedIdx =
10330 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10332 // This must be extracting from either LHS or RHS.
10333 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10334 // Okay, we can handle this if the vector we are insertinting into is
10335 // transitively ok.
10336 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10337 // If so, update the mask to reflect the inserted value.
10338 if (EI->getOperand(0) == LHS) {
10339 Mask[InsertedIdx & (NumElts-1)] =
10340 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10342 assert(EI->getOperand(0) == RHS);
10343 Mask[InsertedIdx & (NumElts-1)] =
10344 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10353 // TODO: Handle shufflevector here!
10358 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10359 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10360 /// that computes V and the LHS value of the shuffle.
10361 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10363 assert(isa<VectorType>(V->getType()) &&
10364 (RHS == 0 || V->getType() == RHS->getType()) &&
10365 "Invalid shuffle!");
10366 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10368 if (isa<UndefValue>(V)) {
10369 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10371 } else if (isa<ConstantAggregateZero>(V)) {
10372 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10374 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10375 // If this is an insert of an extract from some other vector, include it.
10376 Value *VecOp = IEI->getOperand(0);
10377 Value *ScalarOp = IEI->getOperand(1);
10378 Value *IdxOp = IEI->getOperand(2);
10380 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10381 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10382 EI->getOperand(0)->getType() == V->getType()) {
10383 unsigned ExtractedIdx =
10384 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10385 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10387 // Either the extracted from or inserted into vector must be RHSVec,
10388 // otherwise we'd end up with a shuffle of three inputs.
10389 if (EI->getOperand(0) == RHS || RHS == 0) {
10390 RHS = EI->getOperand(0);
10391 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10392 Mask[InsertedIdx & (NumElts-1)] =
10393 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10397 if (VecOp == RHS) {
10398 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10399 // Everything but the extracted element is replaced with the RHS.
10400 for (unsigned i = 0; i != NumElts; ++i) {
10401 if (i != InsertedIdx)
10402 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10407 // If this insertelement is a chain that comes from exactly these two
10408 // vectors, return the vector and the effective shuffle.
10409 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10410 return EI->getOperand(0);
10415 // TODO: Handle shufflevector here!
10417 // Otherwise, can't do anything fancy. Return an identity vector.
10418 for (unsigned i = 0; i != NumElts; ++i)
10419 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10423 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10424 Value *VecOp = IE.getOperand(0);
10425 Value *ScalarOp = IE.getOperand(1);
10426 Value *IdxOp = IE.getOperand(2);
10428 // Inserting an undef or into an undefined place, remove this.
10429 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10430 ReplaceInstUsesWith(IE, VecOp);
10432 // If the inserted element was extracted from some other vector, and if the
10433 // indexes are constant, try to turn this into a shufflevector operation.
10434 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10435 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10436 EI->getOperand(0)->getType() == IE.getType()) {
10437 unsigned NumVectorElts = IE.getType()->getNumElements();
10438 unsigned ExtractedIdx =
10439 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10440 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10442 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10443 return ReplaceInstUsesWith(IE, VecOp);
10445 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10446 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10448 // If we are extracting a value from a vector, then inserting it right
10449 // back into the same place, just use the input vector.
10450 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10451 return ReplaceInstUsesWith(IE, VecOp);
10453 // We could theoretically do this for ANY input. However, doing so could
10454 // turn chains of insertelement instructions into a chain of shufflevector
10455 // instructions, and right now we do not merge shufflevectors. As such,
10456 // only do this in a situation where it is clear that there is benefit.
10457 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10458 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10459 // the values of VecOp, except then one read from EIOp0.
10460 // Build a new shuffle mask.
10461 std::vector<Constant*> Mask;
10462 if (isa<UndefValue>(VecOp))
10463 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10465 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10466 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10469 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10470 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10471 ConstantVector::get(Mask));
10474 // If this insertelement isn't used by some other insertelement, turn it
10475 // (and any insertelements it points to), into one big shuffle.
10476 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10477 std::vector<Constant*> Mask;
10479 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10480 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10481 // We now have a shuffle of LHS, RHS, Mask.
10482 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10491 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10492 Value *LHS = SVI.getOperand(0);
10493 Value *RHS = SVI.getOperand(1);
10494 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10496 bool MadeChange = false;
10498 // Undefined shuffle mask -> undefined value.
10499 if (isa<UndefValue>(SVI.getOperand(2)))
10500 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10502 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10503 // the undef, change them to undefs.
10504 if (isa<UndefValue>(SVI.getOperand(1))) {
10505 // Scan to see if there are any references to the RHS. If so, replace them
10506 // with undef element refs and set MadeChange to true.
10507 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10508 if (Mask[i] >= e && Mask[i] != 2*e) {
10515 // Remap any references to RHS to use LHS.
10516 std::vector<Constant*> Elts;
10517 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10518 if (Mask[i] == 2*e)
10519 Elts.push_back(UndefValue::get(Type::Int32Ty));
10521 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10523 SVI.setOperand(2, ConstantVector::get(Elts));
10527 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10528 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10529 if (LHS == RHS || isa<UndefValue>(LHS)) {
10530 if (isa<UndefValue>(LHS) && LHS == RHS) {
10531 // shuffle(undef,undef,mask) -> undef.
10532 return ReplaceInstUsesWith(SVI, LHS);
10535 // Remap any references to RHS to use LHS.
10536 std::vector<Constant*> Elts;
10537 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10538 if (Mask[i] >= 2*e)
10539 Elts.push_back(UndefValue::get(Type::Int32Ty));
10541 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10542 (Mask[i] < e && isa<UndefValue>(LHS)))
10543 Mask[i] = 2*e; // Turn into undef.
10545 Mask[i] &= (e-1); // Force to LHS.
10546 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10549 SVI.setOperand(0, SVI.getOperand(1));
10550 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10551 SVI.setOperand(2, ConstantVector::get(Elts));
10552 LHS = SVI.getOperand(0);
10553 RHS = SVI.getOperand(1);
10557 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10558 bool isLHSID = true, isRHSID = true;
10560 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10561 if (Mask[i] >= e*2) continue; // Ignore undef values.
10562 // Is this an identity shuffle of the LHS value?
10563 isLHSID &= (Mask[i] == i);
10565 // Is this an identity shuffle of the RHS value?
10566 isRHSID &= (Mask[i]-e == i);
10569 // Eliminate identity shuffles.
10570 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10571 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10573 // If the LHS is a shufflevector itself, see if we can combine it with this
10574 // one without producing an unusual shuffle. Here we are really conservative:
10575 // we are absolutely afraid of producing a shuffle mask not in the input
10576 // program, because the code gen may not be smart enough to turn a merged
10577 // shuffle into two specific shuffles: it may produce worse code. As such,
10578 // we only merge two shuffles if the result is one of the two input shuffle
10579 // masks. In this case, merging the shuffles just removes one instruction,
10580 // which we know is safe. This is good for things like turning:
10581 // (splat(splat)) -> splat.
10582 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10583 if (isa<UndefValue>(RHS)) {
10584 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10586 std::vector<unsigned> NewMask;
10587 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10588 if (Mask[i] >= 2*e)
10589 NewMask.push_back(2*e);
10591 NewMask.push_back(LHSMask[Mask[i]]);
10593 // If the result mask is equal to the src shuffle or this shuffle mask, do
10594 // the replacement.
10595 if (NewMask == LHSMask || NewMask == Mask) {
10596 std::vector<Constant*> Elts;
10597 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10598 if (NewMask[i] >= e*2) {
10599 Elts.push_back(UndefValue::get(Type::Int32Ty));
10601 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10604 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10605 LHSSVI->getOperand(1),
10606 ConstantVector::get(Elts));
10611 return MadeChange ? &SVI : 0;
10617 /// TryToSinkInstruction - Try to move the specified instruction from its
10618 /// current block into the beginning of DestBlock, which can only happen if it's
10619 /// safe to move the instruction past all of the instructions between it and the
10620 /// end of its block.
10621 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10622 assert(I->hasOneUse() && "Invariants didn't hold!");
10624 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10625 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10627 // Do not sink alloca instructions out of the entry block.
10628 if (isa<AllocaInst>(I) && I->getParent() ==
10629 &DestBlock->getParent()->getEntryBlock())
10632 // We can only sink load instructions if there is nothing between the load and
10633 // the end of block that could change the value.
10634 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10635 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10637 if (Scan->mayWriteToMemory())
10641 BasicBlock::iterator InsertPos = DestBlock->begin();
10642 while (isa<PHINode>(InsertPos)) ++InsertPos;
10644 I->moveBefore(InsertPos);
10650 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10651 /// all reachable code to the worklist.
10653 /// This has a couple of tricks to make the code faster and more powerful. In
10654 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10655 /// them to the worklist (this significantly speeds up instcombine on code where
10656 /// many instructions are dead or constant). Additionally, if we find a branch
10657 /// whose condition is a known constant, we only visit the reachable successors.
10659 static void AddReachableCodeToWorklist(BasicBlock *BB,
10660 SmallPtrSet<BasicBlock*, 64> &Visited,
10662 const TargetData *TD) {
10663 std::vector<BasicBlock*> Worklist;
10664 Worklist.push_back(BB);
10666 while (!Worklist.empty()) {
10667 BB = Worklist.back();
10668 Worklist.pop_back();
10670 // We have now visited this block! If we've already been here, ignore it.
10671 if (!Visited.insert(BB)) continue;
10673 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10674 Instruction *Inst = BBI++;
10676 // DCE instruction if trivially dead.
10677 if (isInstructionTriviallyDead(Inst)) {
10679 DOUT << "IC: DCE: " << *Inst;
10680 Inst->eraseFromParent();
10684 // ConstantProp instruction if trivially constant.
10685 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10686 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10687 Inst->replaceAllUsesWith(C);
10689 Inst->eraseFromParent();
10693 IC.AddToWorkList(Inst);
10696 // Recursively visit successors. If this is a branch or switch on a
10697 // constant, only visit the reachable successor.
10698 TerminatorInst *TI = BB->getTerminator();
10699 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10700 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10701 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10702 Worklist.push_back(BI->getSuccessor(!CondVal));
10705 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10706 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10707 // See if this is an explicit destination.
10708 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10709 if (SI->getCaseValue(i) == Cond) {
10710 Worklist.push_back(SI->getSuccessor(i));
10714 // Otherwise it is the default destination.
10715 Worklist.push_back(SI->getSuccessor(0));
10720 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10721 Worklist.push_back(TI->getSuccessor(i));
10725 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10726 bool Changed = false;
10727 TD = &getAnalysis<TargetData>();
10729 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10730 << F.getNameStr() << "\n");
10733 // Do a depth-first traversal of the function, populate the worklist with
10734 // the reachable instructions. Ignore blocks that are not reachable. Keep
10735 // track of which blocks we visit.
10736 SmallPtrSet<BasicBlock*, 64> Visited;
10737 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10739 // Do a quick scan over the function. If we find any blocks that are
10740 // unreachable, remove any instructions inside of them. This prevents
10741 // the instcombine code from having to deal with some bad special cases.
10742 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10743 if (!Visited.count(BB)) {
10744 Instruction *Term = BB->getTerminator();
10745 while (Term != BB->begin()) { // Remove instrs bottom-up
10746 BasicBlock::iterator I = Term; --I;
10748 DOUT << "IC: DCE: " << *I;
10751 if (!I->use_empty())
10752 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10753 I->eraseFromParent();
10758 while (!Worklist.empty()) {
10759 Instruction *I = RemoveOneFromWorkList();
10760 if (I == 0) continue; // skip null values.
10762 // Check to see if we can DCE the instruction.
10763 if (isInstructionTriviallyDead(I)) {
10764 // Add operands to the worklist.
10765 if (I->getNumOperands() < 4)
10766 AddUsesToWorkList(*I);
10769 DOUT << "IC: DCE: " << *I;
10771 I->eraseFromParent();
10772 RemoveFromWorkList(I);
10776 // Instruction isn't dead, see if we can constant propagate it.
10777 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10778 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10780 // Add operands to the worklist.
10781 AddUsesToWorkList(*I);
10782 ReplaceInstUsesWith(*I, C);
10785 I->eraseFromParent();
10786 RemoveFromWorkList(I);
10790 // See if we can trivially sink this instruction to a successor basic block.
10791 if (I->hasOneUse()) {
10792 BasicBlock *BB = I->getParent();
10793 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10794 if (UserParent != BB) {
10795 bool UserIsSuccessor = false;
10796 // See if the user is one of our successors.
10797 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10798 if (*SI == UserParent) {
10799 UserIsSuccessor = true;
10803 // If the user is one of our immediate successors, and if that successor
10804 // only has us as a predecessors (we'd have to split the critical edge
10805 // otherwise), we can keep going.
10806 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10807 next(pred_begin(UserParent)) == pred_end(UserParent))
10808 // Okay, the CFG is simple enough, try to sink this instruction.
10809 Changed |= TryToSinkInstruction(I, UserParent);
10813 // Now that we have an instruction, try combining it to simplify it...
10817 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10818 if (Instruction *Result = visit(*I)) {
10820 // Should we replace the old instruction with a new one?
10822 DOUT << "IC: Old = " << *I
10823 << " New = " << *Result;
10825 // Everything uses the new instruction now.
10826 I->replaceAllUsesWith(Result);
10828 // Push the new instruction and any users onto the worklist.
10829 AddToWorkList(Result);
10830 AddUsersToWorkList(*Result);
10832 // Move the name to the new instruction first.
10833 Result->takeName(I);
10835 // Insert the new instruction into the basic block...
10836 BasicBlock *InstParent = I->getParent();
10837 BasicBlock::iterator InsertPos = I;
10839 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10840 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10843 InstParent->getInstList().insert(InsertPos, Result);
10845 // Make sure that we reprocess all operands now that we reduced their
10847 AddUsesToWorkList(*I);
10849 // Instructions can end up on the worklist more than once. Make sure
10850 // we do not process an instruction that has been deleted.
10851 RemoveFromWorkList(I);
10853 // Erase the old instruction.
10854 InstParent->getInstList().erase(I);
10857 DOUT << "IC: Mod = " << OrigI
10858 << " New = " << *I;
10861 // If the instruction was modified, it's possible that it is now dead.
10862 // if so, remove it.
10863 if (isInstructionTriviallyDead(I)) {
10864 // Make sure we process all operands now that we are reducing their
10866 AddUsesToWorkList(*I);
10868 // Instructions may end up in the worklist more than once. Erase all
10869 // occurrences of this instruction.
10870 RemoveFromWorkList(I);
10871 I->eraseFromParent();
10874 AddUsersToWorkList(*I);
10881 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10883 // Do an explicit clear, this shrinks the map if needed.
10884 WorklistMap.clear();
10889 bool InstCombiner::runOnFunction(Function &F) {
10890 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10892 bool EverMadeChange = false;
10894 // Iterate while there is work to do.
10895 unsigned Iteration = 0;
10896 while (DoOneIteration(F, Iteration++))
10897 EverMadeChange = true;
10898 return EverMadeChange;
10901 FunctionPass *llvm::createInstructionCombiningPass() {
10902 return new InstCombiner();