1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/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(CastInst &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(CastInst &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 // ReplaceInstUsesWith - This method is to be used when an instruction is
269 // found to be dead, replacable with another preexisting expression. Here
270 // we add all uses of I to the worklist, replace all uses of I with the new
271 // value, then return I, so that the inst combiner will know that I was
274 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
275 AddUsersToWorkList(I); // Add all modified instrs to worklist
277 I.replaceAllUsesWith(V);
280 // If we are replacing the instruction with itself, this must be in a
281 // segment of unreachable code, so just clobber the instruction.
282 I.replaceAllUsesWith(UndefValue::get(I.getType()));
287 // UpdateValueUsesWith - This method is to be used when an value is
288 // found to be replacable with another preexisting expression or was
289 // updated. Here we add all uses of I to the worklist, replace all uses of
290 // I with the new value (unless the instruction was just updated), then
291 // return true, so that the inst combiner will know that I was modified.
293 bool UpdateValueUsesWith(Value *Old, Value *New) {
294 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
296 Old->replaceAllUsesWith(New);
297 if (Instruction *I = dyn_cast<Instruction>(Old))
299 if (Instruction *I = dyn_cast<Instruction>(New))
304 // EraseInstFromFunction - When dealing with an instruction that has side
305 // effects or produces a void value, we can't rely on DCE to delete the
306 // instruction. Instead, visit methods should return the value returned by
308 Instruction *EraseInstFromFunction(Instruction &I) {
309 assert(I.use_empty() && "Cannot erase instruction that is used!");
310 AddUsesToWorkList(I);
311 RemoveFromWorkList(&I);
313 return 0; // Don't do anything with FI
317 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
318 /// InsertBefore instruction. This is specialized a bit to avoid inserting
319 /// casts that are known to not do anything...
321 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
322 Value *V, const Type *DestTy,
323 Instruction *InsertBefore);
325 /// SimplifyCommutative - This performs a few simplifications for
326 /// commutative operators.
327 bool SimplifyCommutative(BinaryOperator &I);
329 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
330 /// most-complex to least-complex order.
331 bool SimplifyCompare(CmpInst &I);
333 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
334 /// on the demanded bits.
335 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
336 APInt& KnownZero, APInt& KnownOne,
339 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
340 uint64_t &UndefElts, unsigned Depth = 0);
342 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
343 // PHI node as operand #0, see if we can fold the instruction into the PHI
344 // (which is only possible if all operands to the PHI are constants).
345 Instruction *FoldOpIntoPhi(Instruction &I);
347 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
348 // operator and they all are only used by the PHI, PHI together their
349 // inputs, and do the operation once, to the result of the PHI.
350 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
351 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
354 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
355 ConstantInt *AndRHS, BinaryOperator &TheAnd);
357 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
358 bool isSub, Instruction &I);
359 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
360 bool isSigned, bool Inside, Instruction &IB);
361 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
362 Instruction *MatchBSwap(BinaryOperator &I);
363 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
365 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
368 char InstCombiner::ID = 0;
369 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
372 // getComplexity: Assign a complexity or rank value to LLVM Values...
373 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
374 static unsigned getComplexity(Value *V) {
375 if (isa<Instruction>(V)) {
376 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
380 if (isa<Argument>(V)) return 3;
381 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
384 // isOnlyUse - Return true if this instruction will be deleted if we stop using
386 static bool isOnlyUse(Value *V) {
387 return V->hasOneUse() || isa<Constant>(V);
390 // getPromotedType - Return the specified type promoted as it would be to pass
391 // though a va_arg area...
392 static const Type *getPromotedType(const Type *Ty) {
393 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
394 if (ITy->getBitWidth() < 32)
395 return Type::Int32Ty;
400 /// getBitCastOperand - If the specified operand is a CastInst or a constant
401 /// expression bitcast, return the operand value, otherwise return null.
402 static Value *getBitCastOperand(Value *V) {
403 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
404 return I->getOperand(0);
405 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
406 if (CE->getOpcode() == Instruction::BitCast)
407 return CE->getOperand(0);
411 /// This function is a wrapper around CastInst::isEliminableCastPair. It
412 /// simply extracts arguments and returns what that function returns.
413 static Instruction::CastOps
414 isEliminableCastPair(
415 const CastInst *CI, ///< The first cast instruction
416 unsigned opcode, ///< The opcode of the second cast instruction
417 const Type *DstTy, ///< The target type for the second cast instruction
418 TargetData *TD ///< The target data for pointer size
421 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
422 const Type *MidTy = CI->getType(); // B from above
424 // Get the opcodes of the two Cast instructions
425 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
426 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
428 return Instruction::CastOps(
429 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
430 DstTy, TD->getIntPtrType()));
433 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
434 /// in any code being generated. It does not require codegen if V is simple
435 /// enough or if the cast can be folded into other casts.
436 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
437 const Type *Ty, TargetData *TD) {
438 if (V->getType() == Ty || isa<Constant>(V)) return false;
440 // If this is another cast that can be eliminated, it isn't codegen either.
441 if (const CastInst *CI = dyn_cast<CastInst>(V))
442 if (isEliminableCastPair(CI, opcode, Ty, TD))
447 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
448 /// InsertBefore instruction. This is specialized a bit to avoid inserting
449 /// casts that are known to not do anything...
451 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
452 Value *V, const Type *DestTy,
453 Instruction *InsertBefore) {
454 if (V->getType() == DestTy) return V;
455 if (Constant *C = dyn_cast<Constant>(V))
456 return ConstantExpr::getCast(opcode, C, DestTy);
458 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
461 // SimplifyCommutative - This performs a few simplifications for commutative
464 // 1. Order operands such that they are listed from right (least complex) to
465 // left (most complex). This puts constants before unary operators before
468 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
469 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
471 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
472 bool Changed = false;
473 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
474 Changed = !I.swapOperands();
476 if (!I.isAssociative()) return Changed;
477 Instruction::BinaryOps Opcode = I.getOpcode();
478 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
479 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
480 if (isa<Constant>(I.getOperand(1))) {
481 Constant *Folded = ConstantExpr::get(I.getOpcode(),
482 cast<Constant>(I.getOperand(1)),
483 cast<Constant>(Op->getOperand(1)));
484 I.setOperand(0, Op->getOperand(0));
485 I.setOperand(1, Folded);
487 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
488 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
489 isOnlyUse(Op) && isOnlyUse(Op1)) {
490 Constant *C1 = cast<Constant>(Op->getOperand(1));
491 Constant *C2 = cast<Constant>(Op1->getOperand(1));
493 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
494 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
495 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
499 I.setOperand(0, New);
500 I.setOperand(1, Folded);
507 /// SimplifyCompare - For a CmpInst this function just orders the operands
508 /// so that theyare listed from right (least complex) to left (most complex).
509 /// This puts constants before unary operators before binary operators.
510 bool InstCombiner::SimplifyCompare(CmpInst &I) {
511 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
514 // Compare instructions are not associative so there's nothing else we can do.
518 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
519 // if the LHS is a constant zero (which is the 'negate' form).
521 static inline Value *dyn_castNegVal(Value *V) {
522 if (BinaryOperator::isNeg(V))
523 return BinaryOperator::getNegArgument(V);
525 // Constants can be considered to be negated values if they can be folded.
526 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
527 return ConstantExpr::getNeg(C);
531 static inline Value *dyn_castNotVal(Value *V) {
532 if (BinaryOperator::isNot(V))
533 return BinaryOperator::getNotArgument(V);
535 // Constants can be considered to be not'ed values...
536 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
537 return ConstantInt::get(~C->getValue());
541 // dyn_castFoldableMul - If this value is a multiply that can be folded into
542 // other computations (because it has a constant operand), return the
543 // non-constant operand of the multiply, and set CST to point to the multiplier.
544 // Otherwise, return null.
546 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
547 if (V->hasOneUse() && V->getType()->isInteger())
548 if (Instruction *I = dyn_cast<Instruction>(V)) {
549 if (I->getOpcode() == Instruction::Mul)
550 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
551 return I->getOperand(0);
552 if (I->getOpcode() == Instruction::Shl)
553 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
554 // The multiplier is really 1 << CST.
555 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
556 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
557 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
558 return I->getOperand(0);
564 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
565 /// expression, return it.
566 static User *dyn_castGetElementPtr(Value *V) {
567 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
568 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
569 if (CE->getOpcode() == Instruction::GetElementPtr)
570 return cast<User>(V);
574 /// AddOne - Add one to a ConstantInt
575 static ConstantInt *AddOne(ConstantInt *C) {
576 APInt Val(C->getValue());
577 return ConstantInt::get(++Val);
579 /// SubOne - Subtract one from a ConstantInt
580 static ConstantInt *SubOne(ConstantInt *C) {
581 APInt Val(C->getValue());
582 return ConstantInt::get(--Val);
584 /// Add - Add two ConstantInts together
585 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
586 return ConstantInt::get(C1->getValue() + C2->getValue());
588 /// And - Bitwise AND two ConstantInts together
589 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
590 return ConstantInt::get(C1->getValue() & C2->getValue());
592 /// Subtract - Subtract one ConstantInt from another
593 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
594 return ConstantInt::get(C1->getValue() - C2->getValue());
596 /// Multiply - Multiply two ConstantInts together
597 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
598 return ConstantInt::get(C1->getValue() * C2->getValue());
601 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
602 /// known to be either zero or one and return them in the KnownZero/KnownOne
603 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
605 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
606 /// we cannot optimize based on the assumption that it is zero without changing
607 /// it to be an explicit zero. If we don't change it to zero, other code could
608 /// optimized based on the contradictory assumption that it is non-zero.
609 /// Because instcombine aggressively folds operations with undef args anyway,
610 /// this won't lose us code quality.
611 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
612 APInt& KnownOne, unsigned Depth = 0) {
613 assert(V && "No Value?");
614 assert(Depth <= 6 && "Limit Search Depth");
615 uint32_t BitWidth = Mask.getBitWidth();
616 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
617 KnownZero.getBitWidth() == BitWidth &&
618 KnownOne.getBitWidth() == BitWidth &&
619 "V, Mask, KnownOne and KnownZero should have same BitWidth");
620 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
621 // We know all of the bits for a constant!
622 KnownOne = CI->getValue() & Mask;
623 KnownZero = ~KnownOne & Mask;
627 if (Depth == 6 || Mask == 0)
628 return; // Limit search depth.
630 Instruction *I = dyn_cast<Instruction>(V);
633 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
634 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
636 switch (I->getOpcode()) {
637 case Instruction::And: {
638 // If either the LHS or the RHS are Zero, the result is zero.
639 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
640 APInt Mask2(Mask & ~KnownZero);
641 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
642 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
643 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
645 // Output known-1 bits are only known if set in both the LHS & RHS.
646 KnownOne &= KnownOne2;
647 // Output known-0 are known to be clear if zero in either the LHS | RHS.
648 KnownZero |= KnownZero2;
651 case Instruction::Or: {
652 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
653 APInt Mask2(Mask & ~KnownOne);
654 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
655 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
656 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
658 // Output known-0 bits are only known if clear in both the LHS & RHS.
659 KnownZero &= KnownZero2;
660 // Output known-1 are known to be set if set in either the LHS | RHS.
661 KnownOne |= KnownOne2;
664 case Instruction::Xor: {
665 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
666 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
667 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
668 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
670 // Output known-0 bits are known if clear or set in both the LHS & RHS.
671 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
672 // Output known-1 are known to be set if set in only one of the LHS, RHS.
673 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
674 KnownZero = KnownZeroOut;
677 case Instruction::Select:
678 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
679 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
680 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
681 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
683 // Only known if known in both the LHS and RHS.
684 KnownOne &= KnownOne2;
685 KnownZero &= KnownZero2;
687 case Instruction::FPTrunc:
688 case Instruction::FPExt:
689 case Instruction::FPToUI:
690 case Instruction::FPToSI:
691 case Instruction::SIToFP:
692 case Instruction::PtrToInt:
693 case Instruction::UIToFP:
694 case Instruction::IntToPtr:
695 return; // Can't work with floating point or pointers
696 case Instruction::Trunc: {
697 // All these have integer operands
698 uint32_t SrcBitWidth =
699 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
701 MaskIn.zext(SrcBitWidth);
702 KnownZero.zext(SrcBitWidth);
703 KnownOne.zext(SrcBitWidth);
704 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
705 KnownZero.trunc(BitWidth);
706 KnownOne.trunc(BitWidth);
709 case Instruction::BitCast: {
710 const Type *SrcTy = I->getOperand(0)->getType();
711 if (SrcTy->isInteger()) {
712 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
717 case Instruction::ZExt: {
718 // Compute the bits in the result that are not present in the input.
719 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
720 uint32_t SrcBitWidth = SrcTy->getBitWidth();
723 MaskIn.trunc(SrcBitWidth);
724 KnownZero.trunc(SrcBitWidth);
725 KnownOne.trunc(SrcBitWidth);
726 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
727 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
728 // The top bits are known to be zero.
729 KnownZero.zext(BitWidth);
730 KnownOne.zext(BitWidth);
731 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
734 case Instruction::SExt: {
735 // Compute the bits in the result that are not present in the input.
736 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
737 uint32_t SrcBitWidth = SrcTy->getBitWidth();
740 MaskIn.trunc(SrcBitWidth);
741 KnownZero.trunc(SrcBitWidth);
742 KnownOne.trunc(SrcBitWidth);
743 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 KnownZero.zext(BitWidth);
746 KnownOne.zext(BitWidth);
748 // If the sign bit of the input is known set or clear, then we know the
749 // top bits of the result.
750 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
751 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
752 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
753 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
756 case Instruction::Shl:
757 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
758 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
759 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
760 APInt Mask2(Mask.lshr(ShiftAmt));
761 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
762 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
763 KnownZero <<= ShiftAmt;
764 KnownOne <<= ShiftAmt;
765 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
769 case Instruction::LShr:
770 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
771 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
772 // Compute the new bits that are at the top now.
773 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
775 // Unsigned shift right.
776 APInt Mask2(Mask.shl(ShiftAmt));
777 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
778 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
779 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
780 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
781 // high bits known zero.
782 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
786 case Instruction::AShr:
787 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
788 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
789 // Compute the new bits that are at the top now.
790 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
792 // Signed shift right.
793 APInt Mask2(Mask.shl(ShiftAmt));
794 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
795 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
796 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
797 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
799 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
800 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
801 KnownZero |= HighBits;
802 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
803 KnownOne |= HighBits;
810 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
811 /// this predicate to simplify operations downstream. Mask is known to be zero
812 /// for bits that V cannot have.
813 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
814 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
815 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
816 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
817 return (KnownZero & Mask) == Mask;
820 /// ShrinkDemandedConstant - Check to see if the specified operand of the
821 /// specified instruction is a constant integer. If so, check to see if there
822 /// are any bits set in the constant that are not demanded. If so, shrink the
823 /// constant and return true.
824 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
826 assert(I && "No instruction?");
827 assert(OpNo < I->getNumOperands() && "Operand index too large");
829 // If the operand is not a constant integer, nothing to do.
830 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
831 if (!OpC) return false;
833 // If there are no bits set that aren't demanded, nothing to do.
834 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
835 if ((~Demanded & OpC->getValue()) == 0)
838 // This instruction is producing bits that are not demanded. Shrink the RHS.
839 Demanded &= OpC->getValue();
840 I->setOperand(OpNo, ConstantInt::get(Demanded));
844 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
845 // set of known zero and one bits, compute the maximum and minimum values that
846 // could have the specified known zero and known one bits, returning them in
848 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
849 const APInt& KnownZero,
850 const APInt& KnownOne,
851 APInt& Min, APInt& Max) {
852 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
853 assert(KnownZero.getBitWidth() == BitWidth &&
854 KnownOne.getBitWidth() == BitWidth &&
855 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
856 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
857 APInt UnknownBits = ~(KnownZero|KnownOne);
859 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
860 // bit if it is unknown.
862 Max = KnownOne|UnknownBits;
864 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
866 Max.clear(BitWidth-1);
870 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
871 // a set of known zero and one bits, compute the maximum and minimum values that
872 // could have the specified known zero and known one bits, returning them in
874 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
875 const APInt &KnownZero,
876 const APInt &KnownOne,
877 APInt &Min, APInt &Max) {
878 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
879 assert(KnownZero.getBitWidth() == BitWidth &&
880 KnownOne.getBitWidth() == BitWidth &&
881 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
882 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
883 APInt UnknownBits = ~(KnownZero|KnownOne);
885 // The minimum value is when the unknown bits are all zeros.
887 // The maximum value is when the unknown bits are all ones.
888 Max = KnownOne|UnknownBits;
891 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
892 /// value based on the demanded bits. When this function is called, it is known
893 /// that only the bits set in DemandedMask of the result of V are ever used
894 /// downstream. Consequently, depending on the mask and V, it may be possible
895 /// to replace V with a constant or one of its operands. In such cases, this
896 /// function does the replacement and returns true. In all other cases, it
897 /// returns false after analyzing the expression and setting KnownOne and known
898 /// to be one in the expression. KnownZero contains all the bits that are known
899 /// to be zero in the expression. These are provided to potentially allow the
900 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
901 /// the expression. KnownOne and KnownZero always follow the invariant that
902 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
903 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
904 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
905 /// and KnownOne must all be the same.
906 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
907 APInt& KnownZero, APInt& KnownOne,
909 assert(V != 0 && "Null pointer of Value???");
910 assert(Depth <= 6 && "Limit Search Depth");
911 uint32_t BitWidth = DemandedMask.getBitWidth();
912 const IntegerType *VTy = cast<IntegerType>(V->getType());
913 assert(VTy->getBitWidth() == BitWidth &&
914 KnownZero.getBitWidth() == BitWidth &&
915 KnownOne.getBitWidth() == BitWidth &&
916 "Value *V, DemandedMask, KnownZero and KnownOne \
917 must have same BitWidth");
918 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
919 // We know all of the bits for a constant!
920 KnownOne = CI->getValue() & DemandedMask;
921 KnownZero = ~KnownOne & DemandedMask;
927 if (!V->hasOneUse()) { // Other users may use these bits.
928 if (Depth != 0) { // Not at the root.
929 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
930 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
933 // If this is the root being simplified, allow it to have multiple uses,
934 // just set the DemandedMask to all bits.
935 DemandedMask = APInt::getAllOnesValue(BitWidth);
936 } else if (DemandedMask == 0) { // Not demanding any bits from V.
937 if (V != UndefValue::get(VTy))
938 return UpdateValueUsesWith(V, UndefValue::get(VTy));
940 } else if (Depth == 6) { // Limit search depth.
944 Instruction *I = dyn_cast<Instruction>(V);
945 if (!I) return false; // Only analyze instructions.
947 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
948 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
949 switch (I->getOpcode()) {
951 case Instruction::And:
952 // If either the LHS or the RHS are Zero, the result is zero.
953 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
954 RHSKnownZero, RHSKnownOne, Depth+1))
956 assert((RHSKnownZero & RHSKnownOne) == 0 &&
957 "Bits known to be one AND zero?");
959 // If something is known zero on the RHS, the bits aren't demanded on the
961 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
962 LHSKnownZero, LHSKnownOne, Depth+1))
964 assert((LHSKnownZero & LHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
967 // If all of the demanded bits are known 1 on one side, return the other.
968 // These bits cannot contribute to the result of the 'and'.
969 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
970 (DemandedMask & ~LHSKnownZero))
971 return UpdateValueUsesWith(I, I->getOperand(0));
972 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
973 (DemandedMask & ~RHSKnownZero))
974 return UpdateValueUsesWith(I, I->getOperand(1));
976 // If all of the demanded bits in the inputs are known zeros, return zero.
977 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
978 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
980 // If the RHS is a constant, see if we can simplify it.
981 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
982 return UpdateValueUsesWith(I, I);
984 // Output known-1 bits are only known if set in both the LHS & RHS.
985 RHSKnownOne &= LHSKnownOne;
986 // Output known-0 are known to be clear if zero in either the LHS | RHS.
987 RHSKnownZero |= LHSKnownZero;
989 case Instruction::Or:
990 // If either the LHS or the RHS are One, the result is One.
991 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
992 RHSKnownZero, RHSKnownOne, Depth+1))
994 assert((RHSKnownZero & RHSKnownOne) == 0 &&
995 "Bits known to be one AND zero?");
996 // If something is known one on the RHS, the bits aren't demanded on the
998 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
999 LHSKnownZero, LHSKnownOne, Depth+1))
1001 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1002 "Bits known to be one AND zero?");
1004 // If all of the demanded bits are known zero on one side, return the other.
1005 // These bits cannot contribute to the result of the 'or'.
1006 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1007 (DemandedMask & ~LHSKnownOne))
1008 return UpdateValueUsesWith(I, I->getOperand(0));
1009 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1010 (DemandedMask & ~RHSKnownOne))
1011 return UpdateValueUsesWith(I, I->getOperand(1));
1013 // If all of the potentially set bits on one side are known to be set on
1014 // the other side, just use the 'other' side.
1015 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1016 (DemandedMask & (~RHSKnownZero)))
1017 return UpdateValueUsesWith(I, I->getOperand(0));
1018 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1019 (DemandedMask & (~LHSKnownZero)))
1020 return UpdateValueUsesWith(I, I->getOperand(1));
1022 // If the RHS is a constant, see if we can simplify it.
1023 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1024 return UpdateValueUsesWith(I, I);
1026 // Output known-0 bits are only known if clear in both the LHS & RHS.
1027 RHSKnownZero &= LHSKnownZero;
1028 // Output known-1 are known to be set if set in either the LHS | RHS.
1029 RHSKnownOne |= LHSKnownOne;
1031 case Instruction::Xor: {
1032 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1033 RHSKnownZero, RHSKnownOne, Depth+1))
1035 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1036 "Bits known to be one AND zero?");
1037 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1038 LHSKnownZero, LHSKnownOne, Depth+1))
1040 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1041 "Bits known to be one AND zero?");
1043 // If all of the demanded bits are known zero on one side, return the other.
1044 // These bits cannot contribute to the result of the 'xor'.
1045 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1046 return UpdateValueUsesWith(I, I->getOperand(0));
1047 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1048 return UpdateValueUsesWith(I, I->getOperand(1));
1050 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1051 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1052 (RHSKnownOne & LHSKnownOne);
1053 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1054 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1055 (RHSKnownOne & LHSKnownZero);
1057 // If all of the demanded bits are known to be zero on one side or the
1058 // other, turn this into an *inclusive* or.
1059 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1060 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1062 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1064 InsertNewInstBefore(Or, *I);
1065 return UpdateValueUsesWith(I, Or);
1068 // If all of the demanded bits on one side are known, and all of the set
1069 // bits on that side are also known to be set on the other side, turn this
1070 // into an AND, as we know the bits will be cleared.
1071 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1072 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1074 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1075 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1077 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1078 InsertNewInstBefore(And, *I);
1079 return UpdateValueUsesWith(I, And);
1083 // If the RHS is a constant, see if we can simplify it.
1084 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1085 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1086 return UpdateValueUsesWith(I, I);
1088 RHSKnownZero = KnownZeroOut;
1089 RHSKnownOne = KnownOneOut;
1092 case Instruction::Select:
1093 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1094 RHSKnownZero, RHSKnownOne, Depth+1))
1096 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1097 LHSKnownZero, LHSKnownOne, Depth+1))
1099 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1100 "Bits known to be one AND zero?");
1101 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1102 "Bits known to be one AND zero?");
1104 // If the operands are constants, see if we can simplify them.
1105 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1106 return UpdateValueUsesWith(I, I);
1107 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1108 return UpdateValueUsesWith(I, I);
1110 // Only known if known in both the LHS and RHS.
1111 RHSKnownOne &= LHSKnownOne;
1112 RHSKnownZero &= LHSKnownZero;
1114 case Instruction::Trunc: {
1116 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1117 DemandedMask.zext(truncBf);
1118 RHSKnownZero.zext(truncBf);
1119 RHSKnownOne.zext(truncBf);
1120 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1121 RHSKnownZero, RHSKnownOne, Depth+1))
1123 DemandedMask.trunc(BitWidth);
1124 RHSKnownZero.trunc(BitWidth);
1125 RHSKnownOne.trunc(BitWidth);
1126 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1127 "Bits known to be one AND zero?");
1130 case Instruction::BitCast:
1131 if (!I->getOperand(0)->getType()->isInteger())
1134 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1135 RHSKnownZero, RHSKnownOne, Depth+1))
1137 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1138 "Bits known to be one AND zero?");
1140 case Instruction::ZExt: {
1141 // Compute the bits in the result that are not present in the input.
1142 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1143 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1145 DemandedMask.trunc(SrcBitWidth);
1146 RHSKnownZero.trunc(SrcBitWidth);
1147 RHSKnownOne.trunc(SrcBitWidth);
1148 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1149 RHSKnownZero, RHSKnownOne, Depth+1))
1151 DemandedMask.zext(BitWidth);
1152 RHSKnownZero.zext(BitWidth);
1153 RHSKnownOne.zext(BitWidth);
1154 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1155 "Bits known to be one AND zero?");
1156 // The top bits are known to be zero.
1157 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1160 case Instruction::SExt: {
1161 // Compute the bits in the result that are not present in the input.
1162 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1163 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1165 APInt InputDemandedBits = DemandedMask &
1166 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1168 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1169 // If any of the sign extended bits are demanded, we know that the sign
1171 if ((NewBits & DemandedMask) != 0)
1172 InputDemandedBits.set(SrcBitWidth-1);
1174 InputDemandedBits.trunc(SrcBitWidth);
1175 RHSKnownZero.trunc(SrcBitWidth);
1176 RHSKnownOne.trunc(SrcBitWidth);
1177 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1178 RHSKnownZero, RHSKnownOne, Depth+1))
1180 InputDemandedBits.zext(BitWidth);
1181 RHSKnownZero.zext(BitWidth);
1182 RHSKnownOne.zext(BitWidth);
1183 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1184 "Bits known to be one AND zero?");
1186 // If the sign bit of the input is known set or clear, then we know the
1187 // top bits of the result.
1189 // If the input sign bit is known zero, or if the NewBits are not demanded
1190 // convert this into a zero extension.
1191 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1193 // Convert to ZExt cast
1194 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1195 return UpdateValueUsesWith(I, NewCast);
1196 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1197 RHSKnownOne |= NewBits;
1201 case Instruction::Add: {
1202 // Figure out what the input bits are. If the top bits of the and result
1203 // are not demanded, then the add doesn't demand them from its input
1205 uint32_t NLZ = DemandedMask.countLeadingZeros();
1207 // If there is a constant on the RHS, there are a variety of xformations
1209 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1210 // If null, this should be simplified elsewhere. Some of the xforms here
1211 // won't work if the RHS is zero.
1215 // If the top bit of the output is demanded, demand everything from the
1216 // input. Otherwise, we demand all the input bits except NLZ top bits.
1217 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1219 // Find information about known zero/one bits in the input.
1220 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1221 LHSKnownZero, LHSKnownOne, Depth+1))
1224 // If the RHS of the add has bits set that can't affect the input, reduce
1226 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1227 return UpdateValueUsesWith(I, I);
1229 // Avoid excess work.
1230 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1233 // Turn it into OR if input bits are zero.
1234 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1236 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1238 InsertNewInstBefore(Or, *I);
1239 return UpdateValueUsesWith(I, Or);
1242 // We can say something about the output known-zero and known-one bits,
1243 // depending on potential carries from the input constant and the
1244 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1245 // bits set and the RHS constant is 0x01001, then we know we have a known
1246 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1248 // To compute this, we first compute the potential carry bits. These are
1249 // the bits which may be modified. I'm not aware of a better way to do
1251 const APInt& RHSVal = RHS->getValue();
1252 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1254 // Now that we know which bits have carries, compute the known-1/0 sets.
1256 // Bits are known one if they are known zero in one operand and one in the
1257 // other, and there is no input carry.
1258 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1259 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1261 // Bits are known zero if they are known zero in both operands and there
1262 // is no input carry.
1263 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1265 // If the high-bits of this ADD are not demanded, then it does not demand
1266 // the high bits of its LHS or RHS.
1267 if (DemandedMask[BitWidth-1] == 0) {
1268 // Right fill the mask of bits for this ADD to demand the most
1269 // significant bit and all those below it.
1270 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1271 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1272 LHSKnownZero, LHSKnownOne, Depth+1))
1274 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1275 LHSKnownZero, LHSKnownOne, Depth+1))
1281 case Instruction::Sub:
1282 // If the high-bits of this SUB are not demanded, then it does not demand
1283 // the high bits of its LHS or RHS.
1284 if (DemandedMask[BitWidth-1] == 0) {
1285 // Right fill the mask of bits for this SUB to demand the most
1286 // significant bit and all those below it.
1287 uint32_t NLZ = DemandedMask.countLeadingZeros();
1288 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1289 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1290 LHSKnownZero, LHSKnownOne, Depth+1))
1292 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1293 LHSKnownZero, LHSKnownOne, Depth+1))
1297 case Instruction::Shl:
1298 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1299 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1300 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1302 RHSKnownZero, RHSKnownOne, Depth+1))
1304 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1305 "Bits known to be one AND zero?");
1306 RHSKnownZero <<= ShiftAmt;
1307 RHSKnownOne <<= ShiftAmt;
1308 // low bits known zero.
1310 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1313 case Instruction::LShr:
1314 // For a logical shift right
1315 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1316 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1318 // Unsigned shift right.
1319 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1320 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1321 RHSKnownZero, RHSKnownOne, Depth+1))
1323 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1324 "Bits known to be one AND zero?");
1325 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1326 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1328 // Compute the new bits that are at the top now.
1329 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1330 RHSKnownZero |= HighBits; // high bits known zero.
1334 case Instruction::AShr:
1335 // If this is an arithmetic shift right and only the low-bit is set, we can
1336 // always convert this into a logical shr, even if the shift amount is
1337 // variable. The low bit of the shift cannot be an input sign bit unless
1338 // the shift amount is >= the size of the datatype, which is undefined.
1339 if (DemandedMask == 1) {
1340 // Perform the logical shift right.
1341 Value *NewVal = BinaryOperator::createLShr(
1342 I->getOperand(0), I->getOperand(1), I->getName());
1343 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1344 return UpdateValueUsesWith(I, NewVal);
1347 // If the sign bit is the only bit demanded by this ashr, then there is no
1348 // need to do it, the shift doesn't change the high bit.
1349 if (DemandedMask.isSignBit())
1350 return UpdateValueUsesWith(I, I->getOperand(0));
1352 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1353 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1355 // Signed shift right.
1356 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1357 // If any of the "high bits" are demanded, we should set the sign bit as
1359 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1360 DemandedMaskIn.set(BitWidth-1);
1361 if (SimplifyDemandedBits(I->getOperand(0),
1363 RHSKnownZero, RHSKnownOne, Depth+1))
1365 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1366 "Bits known to be one AND zero?");
1367 // Compute the new bits that are at the top now.
1368 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1369 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1370 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1372 // Handle the sign bits.
1373 APInt SignBit(APInt::getSignBit(BitWidth));
1374 // Adjust to where it is now in the mask.
1375 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1377 // If the input sign bit is known to be zero, or if none of the top bits
1378 // are demanded, turn this into an unsigned shift right.
1379 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1380 (HighBits & ~DemandedMask) == HighBits) {
1381 // Perform the logical shift right.
1382 Value *NewVal = BinaryOperator::createLShr(
1383 I->getOperand(0), SA, I->getName());
1384 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1385 return UpdateValueUsesWith(I, NewVal);
1386 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1387 RHSKnownOne |= HighBits;
1393 // If the client is only demanding bits that we know, return the known
1395 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1396 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1401 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1402 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1403 /// actually used by the caller. This method analyzes which elements of the
1404 /// operand are undef and returns that information in UndefElts.
1406 /// If the information about demanded elements can be used to simplify the
1407 /// operation, the operation is simplified, then the resultant value is
1408 /// returned. This returns null if no change was made.
1409 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1410 uint64_t &UndefElts,
1412 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1413 assert(VWidth <= 64 && "Vector too wide to analyze!");
1414 uint64_t EltMask = ~0ULL >> (64-VWidth);
1415 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1416 "Invalid DemandedElts!");
1418 if (isa<UndefValue>(V)) {
1419 // If the entire vector is undefined, just return this info.
1420 UndefElts = EltMask;
1422 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1423 UndefElts = EltMask;
1424 return UndefValue::get(V->getType());
1428 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1429 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1430 Constant *Undef = UndefValue::get(EltTy);
1432 std::vector<Constant*> Elts;
1433 for (unsigned i = 0; i != VWidth; ++i)
1434 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1435 Elts.push_back(Undef);
1436 UndefElts |= (1ULL << i);
1437 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1438 Elts.push_back(Undef);
1439 UndefElts |= (1ULL << i);
1440 } else { // Otherwise, defined.
1441 Elts.push_back(CP->getOperand(i));
1444 // If we changed the constant, return it.
1445 Constant *NewCP = ConstantVector::get(Elts);
1446 return NewCP != CP ? NewCP : 0;
1447 } else if (isa<ConstantAggregateZero>(V)) {
1448 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1450 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1451 Constant *Zero = Constant::getNullValue(EltTy);
1452 Constant *Undef = UndefValue::get(EltTy);
1453 std::vector<Constant*> Elts;
1454 for (unsigned i = 0; i != VWidth; ++i)
1455 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1456 UndefElts = DemandedElts ^ EltMask;
1457 return ConstantVector::get(Elts);
1460 if (!V->hasOneUse()) { // Other users may use these bits.
1461 if (Depth != 0) { // Not at the root.
1462 // TODO: Just compute the UndefElts information recursively.
1466 } else if (Depth == 10) { // Limit search depth.
1470 Instruction *I = dyn_cast<Instruction>(V);
1471 if (!I) return false; // Only analyze instructions.
1473 bool MadeChange = false;
1474 uint64_t UndefElts2;
1476 switch (I->getOpcode()) {
1479 case Instruction::InsertElement: {
1480 // If this is a variable index, we don't know which element it overwrites.
1481 // demand exactly the same input as we produce.
1482 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1484 // Note that we can't propagate undef elt info, because we don't know
1485 // which elt is getting updated.
1486 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1487 UndefElts2, Depth+1);
1488 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1492 // If this is inserting an element that isn't demanded, remove this
1494 unsigned IdxNo = Idx->getZExtValue();
1495 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1496 return AddSoonDeadInstToWorklist(*I, 0);
1498 // Otherwise, the element inserted overwrites whatever was there, so the
1499 // input demanded set is simpler than the output set.
1500 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1501 DemandedElts & ~(1ULL << IdxNo),
1502 UndefElts, Depth+1);
1503 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1505 // The inserted element is defined.
1506 UndefElts |= 1ULL << IdxNo;
1509 case Instruction::BitCast: {
1510 // Vector->vector casts only.
1511 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1513 unsigned InVWidth = VTy->getNumElements();
1514 uint64_t InputDemandedElts = 0;
1517 if (VWidth == InVWidth) {
1518 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1519 // elements as are demanded of us.
1521 InputDemandedElts = DemandedElts;
1522 } else if (VWidth > InVWidth) {
1526 // If there are more elements in the result than there are in the source,
1527 // then an input element is live if any of the corresponding output
1528 // elements are live.
1529 Ratio = VWidth/InVWidth;
1530 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1531 if (DemandedElts & (1ULL << OutIdx))
1532 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1538 // If there are more elements in the source than there are in the result,
1539 // then an input element is live if the corresponding output element is
1541 Ratio = InVWidth/VWidth;
1542 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1543 if (DemandedElts & (1ULL << InIdx/Ratio))
1544 InputDemandedElts |= 1ULL << InIdx;
1547 // div/rem demand all inputs, because they don't want divide by zero.
1548 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1549 UndefElts2, Depth+1);
1551 I->setOperand(0, TmpV);
1555 UndefElts = UndefElts2;
1556 if (VWidth > InVWidth) {
1557 assert(0 && "Unimp");
1558 // If there are more elements in the result than there are in the source,
1559 // then an output element is undef if the corresponding input element is
1561 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1562 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1563 UndefElts |= 1ULL << OutIdx;
1564 } else if (VWidth < InVWidth) {
1565 assert(0 && "Unimp");
1566 // If there are more elements in the source than there are in the result,
1567 // then a result element is undef if all of the corresponding input
1568 // elements are undef.
1569 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1570 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1571 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1572 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1576 case Instruction::And:
1577 case Instruction::Or:
1578 case Instruction::Xor:
1579 case Instruction::Add:
1580 case Instruction::Sub:
1581 case Instruction::Mul:
1582 // div/rem demand all inputs, because they don't want divide by zero.
1583 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1584 UndefElts, Depth+1);
1585 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1586 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1587 UndefElts2, Depth+1);
1588 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1590 // Output elements are undefined if both are undefined. Consider things
1591 // like undef&0. The result is known zero, not undef.
1592 UndefElts &= UndefElts2;
1595 case Instruction::Call: {
1596 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1598 switch (II->getIntrinsicID()) {
1601 // Binary vector operations that work column-wise. A dest element is a
1602 // function of the corresponding input elements from the two inputs.
1603 case Intrinsic::x86_sse_sub_ss:
1604 case Intrinsic::x86_sse_mul_ss:
1605 case Intrinsic::x86_sse_min_ss:
1606 case Intrinsic::x86_sse_max_ss:
1607 case Intrinsic::x86_sse2_sub_sd:
1608 case Intrinsic::x86_sse2_mul_sd:
1609 case Intrinsic::x86_sse2_min_sd:
1610 case Intrinsic::x86_sse2_max_sd:
1611 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1612 UndefElts, Depth+1);
1613 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1614 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1615 UndefElts2, Depth+1);
1616 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1618 // If only the low elt is demanded and this is a scalarizable intrinsic,
1619 // scalarize it now.
1620 if (DemandedElts == 1) {
1621 switch (II->getIntrinsicID()) {
1623 case Intrinsic::x86_sse_sub_ss:
1624 case Intrinsic::x86_sse_mul_ss:
1625 case Intrinsic::x86_sse2_sub_sd:
1626 case Intrinsic::x86_sse2_mul_sd:
1627 // TODO: Lower MIN/MAX/ABS/etc
1628 Value *LHS = II->getOperand(1);
1629 Value *RHS = II->getOperand(2);
1630 // Extract the element as scalars.
1631 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1632 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1634 switch (II->getIntrinsicID()) {
1635 default: assert(0 && "Case stmts out of sync!");
1636 case Intrinsic::x86_sse_sub_ss:
1637 case Intrinsic::x86_sse2_sub_sd:
1638 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1639 II->getName()), *II);
1641 case Intrinsic::x86_sse_mul_ss:
1642 case Intrinsic::x86_sse2_mul_sd:
1643 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1644 II->getName()), *II);
1649 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1651 InsertNewInstBefore(New, *II);
1652 AddSoonDeadInstToWorklist(*II, 0);
1657 // Output elements are undefined if both are undefined. Consider things
1658 // like undef&0. The result is known zero, not undef.
1659 UndefElts &= UndefElts2;
1665 return MadeChange ? I : 0;
1668 /// @returns true if the specified compare predicate is
1669 /// true when both operands are equal...
1670 /// @brief Determine if the icmp Predicate is true when both operands are equal
1671 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1672 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1673 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1674 pred == ICmpInst::ICMP_SLE;
1677 /// @returns true if the specified compare instruction is
1678 /// true when both operands are equal...
1679 /// @brief Determine if the ICmpInst returns true when both operands are equal
1680 static bool isTrueWhenEqual(ICmpInst &ICI) {
1681 return isTrueWhenEqual(ICI.getPredicate());
1684 /// AssociativeOpt - Perform an optimization on an associative operator. This
1685 /// function is designed to check a chain of associative operators for a
1686 /// potential to apply a certain optimization. Since the optimization may be
1687 /// applicable if the expression was reassociated, this checks the chain, then
1688 /// reassociates the expression as necessary to expose the optimization
1689 /// opportunity. This makes use of a special Functor, which must define
1690 /// 'shouldApply' and 'apply' methods.
1692 template<typename Functor>
1693 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1694 unsigned Opcode = Root.getOpcode();
1695 Value *LHS = Root.getOperand(0);
1697 // Quick check, see if the immediate LHS matches...
1698 if (F.shouldApply(LHS))
1699 return F.apply(Root);
1701 // Otherwise, if the LHS is not of the same opcode as the root, return.
1702 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1703 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1704 // Should we apply this transform to the RHS?
1705 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1707 // If not to the RHS, check to see if we should apply to the LHS...
1708 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1709 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1713 // If the functor wants to apply the optimization to the RHS of LHSI,
1714 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1716 BasicBlock *BB = Root.getParent();
1718 // Now all of the instructions are in the current basic block, go ahead
1719 // and perform the reassociation.
1720 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1722 // First move the selected RHS to the LHS of the root...
1723 Root.setOperand(0, LHSI->getOperand(1));
1725 // Make what used to be the LHS of the root be the user of the root...
1726 Value *ExtraOperand = TmpLHSI->getOperand(1);
1727 if (&Root == TmpLHSI) {
1728 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1731 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1732 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1733 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1734 BasicBlock::iterator ARI = &Root; ++ARI;
1735 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1738 // Now propagate the ExtraOperand down the chain of instructions until we
1740 while (TmpLHSI != LHSI) {
1741 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1742 // Move the instruction to immediately before the chain we are
1743 // constructing to avoid breaking dominance properties.
1744 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1745 BB->getInstList().insert(ARI, NextLHSI);
1748 Value *NextOp = NextLHSI->getOperand(1);
1749 NextLHSI->setOperand(1, ExtraOperand);
1751 ExtraOperand = NextOp;
1754 // Now that the instructions are reassociated, have the functor perform
1755 // the transformation...
1756 return F.apply(Root);
1759 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1765 // AddRHS - Implements: X + X --> X << 1
1768 AddRHS(Value *rhs) : RHS(rhs) {}
1769 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1770 Instruction *apply(BinaryOperator &Add) const {
1771 return BinaryOperator::createShl(Add.getOperand(0),
1772 ConstantInt::get(Add.getType(), 1));
1776 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1778 struct AddMaskingAnd {
1780 AddMaskingAnd(Constant *c) : C2(c) {}
1781 bool shouldApply(Value *LHS) const {
1783 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1784 ConstantExpr::getAnd(C1, C2)->isNullValue();
1786 Instruction *apply(BinaryOperator &Add) const {
1787 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1791 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1793 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1794 if (Constant *SOC = dyn_cast<Constant>(SO))
1795 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1797 return IC->InsertNewInstBefore(CastInst::create(
1798 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1801 // Figure out if the constant is the left or the right argument.
1802 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1803 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1805 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1807 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1808 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1811 Value *Op0 = SO, *Op1 = ConstOperand;
1813 std::swap(Op0, Op1);
1815 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1816 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1817 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1818 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1819 SO->getName()+".cmp");
1821 assert(0 && "Unknown binary instruction type!");
1824 return IC->InsertNewInstBefore(New, I);
1827 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1828 // constant as the other operand, try to fold the binary operator into the
1829 // select arguments. This also works for Cast instructions, which obviously do
1830 // not have a second operand.
1831 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1833 // Don't modify shared select instructions
1834 if (!SI->hasOneUse()) return 0;
1835 Value *TV = SI->getOperand(1);
1836 Value *FV = SI->getOperand(2);
1838 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1839 // Bool selects with constant operands can be folded to logical ops.
1840 if (SI->getType() == Type::Int1Ty) return 0;
1842 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1843 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1845 return new SelectInst(SI->getCondition(), SelectTrueVal,
1852 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1853 /// node as operand #0, see if we can fold the instruction into the PHI (which
1854 /// is only possible if all operands to the PHI are constants).
1855 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1856 PHINode *PN = cast<PHINode>(I.getOperand(0));
1857 unsigned NumPHIValues = PN->getNumIncomingValues();
1858 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1860 // Check to see if all of the operands of the PHI are constants. If there is
1861 // one non-constant value, remember the BB it is. If there is more than one
1862 // or if *it* is a PHI, bail out.
1863 BasicBlock *NonConstBB = 0;
1864 for (unsigned i = 0; i != NumPHIValues; ++i)
1865 if (!isa<Constant>(PN->getIncomingValue(i))) {
1866 if (NonConstBB) return 0; // More than one non-const value.
1867 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1868 NonConstBB = PN->getIncomingBlock(i);
1870 // If the incoming non-constant value is in I's block, we have an infinite
1872 if (NonConstBB == I.getParent())
1876 // If there is exactly one non-constant value, we can insert a copy of the
1877 // operation in that block. However, if this is a critical edge, we would be
1878 // inserting the computation one some other paths (e.g. inside a loop). Only
1879 // do this if the pred block is unconditionally branching into the phi block.
1881 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1882 if (!BI || !BI->isUnconditional()) return 0;
1885 // Okay, we can do the transformation: create the new PHI node.
1886 PHINode *NewPN = new PHINode(I.getType(), "");
1887 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1888 InsertNewInstBefore(NewPN, *PN);
1889 NewPN->takeName(PN);
1891 // Next, add all of the operands to the PHI.
1892 if (I.getNumOperands() == 2) {
1893 Constant *C = cast<Constant>(I.getOperand(1));
1894 for (unsigned i = 0; i != NumPHIValues; ++i) {
1896 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1897 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1898 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1900 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1902 assert(PN->getIncomingBlock(i) == NonConstBB);
1903 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1904 InV = BinaryOperator::create(BO->getOpcode(),
1905 PN->getIncomingValue(i), C, "phitmp",
1906 NonConstBB->getTerminator());
1907 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1908 InV = CmpInst::create(CI->getOpcode(),
1910 PN->getIncomingValue(i), C, "phitmp",
1911 NonConstBB->getTerminator());
1913 assert(0 && "Unknown binop!");
1915 AddToWorkList(cast<Instruction>(InV));
1917 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1920 CastInst *CI = cast<CastInst>(&I);
1921 const Type *RetTy = CI->getType();
1922 for (unsigned i = 0; i != NumPHIValues; ++i) {
1924 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1925 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1927 assert(PN->getIncomingBlock(i) == NonConstBB);
1928 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1929 I.getType(), "phitmp",
1930 NonConstBB->getTerminator());
1931 AddToWorkList(cast<Instruction>(InV));
1933 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1936 return ReplaceInstUsesWith(I, NewPN);
1939 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1940 bool Changed = SimplifyCommutative(I);
1941 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1943 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1944 // X + undef -> undef
1945 if (isa<UndefValue>(RHS))
1946 return ReplaceInstUsesWith(I, RHS);
1949 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1950 if (RHSC->isNullValue())
1951 return ReplaceInstUsesWith(I, LHS);
1952 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1953 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1954 (I.getType())->getValueAPF()))
1955 return ReplaceInstUsesWith(I, LHS);
1958 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1959 // X + (signbit) --> X ^ signbit
1960 const APInt& Val = CI->getValue();
1961 uint32_t BitWidth = Val.getBitWidth();
1962 if (Val == APInt::getSignBit(BitWidth))
1963 return BinaryOperator::createXor(LHS, RHS);
1965 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1966 // (X & 254)+1 -> (X&254)|1
1967 if (!isa<VectorType>(I.getType())) {
1968 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1969 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1970 KnownZero, KnownOne))
1975 if (isa<PHINode>(LHS))
1976 if (Instruction *NV = FoldOpIntoPhi(I))
1979 ConstantInt *XorRHS = 0;
1981 if (isa<ConstantInt>(RHSC) &&
1982 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1983 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1984 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1986 uint32_t Size = TySizeBits / 2;
1987 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1988 APInt CFF80Val(-C0080Val);
1990 if (TySizeBits > Size) {
1991 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1992 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1993 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1994 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1995 // This is a sign extend if the top bits are known zero.
1996 if (!MaskedValueIsZero(XorLHS,
1997 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1998 Size = 0; // Not a sign ext, but can't be any others either.
2003 C0080Val = APIntOps::lshr(C0080Val, Size);
2004 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2005 } while (Size >= 1);
2007 // FIXME: This shouldn't be necessary. When the backends can handle types
2008 // with funny bit widths then this whole cascade of if statements should
2009 // be removed. It is just here to get the size of the "middle" type back
2010 // up to something that the back ends can handle.
2011 const Type *MiddleType = 0;
2014 case 32: MiddleType = Type::Int32Ty; break;
2015 case 16: MiddleType = Type::Int16Ty; break;
2016 case 8: MiddleType = Type::Int8Ty; break;
2019 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2020 InsertNewInstBefore(NewTrunc, I);
2021 return new SExtInst(NewTrunc, I.getType(), I.getName());
2027 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2028 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2030 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2031 if (RHSI->getOpcode() == Instruction::Sub)
2032 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2033 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2035 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2036 if (LHSI->getOpcode() == Instruction::Sub)
2037 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2038 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2043 if (Value *V = dyn_castNegVal(LHS))
2044 return BinaryOperator::createSub(RHS, V);
2047 if (!isa<Constant>(RHS))
2048 if (Value *V = dyn_castNegVal(RHS))
2049 return BinaryOperator::createSub(LHS, V);
2053 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2054 if (X == RHS) // X*C + X --> X * (C+1)
2055 return BinaryOperator::createMul(RHS, AddOne(C2));
2057 // X*C1 + X*C2 --> X * (C1+C2)
2059 if (X == dyn_castFoldableMul(RHS, C1))
2060 return BinaryOperator::createMul(X, Add(C1, C2));
2063 // X + X*C --> X * (C+1)
2064 if (dyn_castFoldableMul(RHS, C2) == LHS)
2065 return BinaryOperator::createMul(LHS, AddOne(C2));
2067 // X + ~X --> -1 since ~X = -X-1
2068 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2069 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2072 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2073 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2074 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2077 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2079 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2080 return BinaryOperator::createSub(SubOne(CRHS), X);
2082 // (X & FF00) + xx00 -> (X+xx00) & FF00
2083 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2084 Constant *Anded = And(CRHS, C2);
2085 if (Anded == CRHS) {
2086 // See if all bits from the first bit set in the Add RHS up are included
2087 // in the mask. First, get the rightmost bit.
2088 const APInt& AddRHSV = CRHS->getValue();
2090 // Form a mask of all bits from the lowest bit added through the top.
2091 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2093 // See if the and mask includes all of these bits.
2094 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2096 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2097 // Okay, the xform is safe. Insert the new add pronto.
2098 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2099 LHS->getName()), I);
2100 return BinaryOperator::createAnd(NewAdd, C2);
2105 // Try to fold constant add into select arguments.
2106 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2107 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2111 // add (cast *A to intptrtype) B ->
2112 // cast (GEP (cast *A to sbyte*) B) ->
2115 CastInst *CI = dyn_cast<CastInst>(LHS);
2118 CI = dyn_cast<CastInst>(RHS);
2121 if (CI && CI->getType()->isSized() &&
2122 (CI->getType()->getPrimitiveSizeInBits() ==
2123 TD->getIntPtrType()->getPrimitiveSizeInBits())
2124 && isa<PointerType>(CI->getOperand(0)->getType())) {
2125 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2126 PointerType::get(Type::Int8Ty), I);
2127 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2128 return new PtrToIntInst(I2, CI->getType());
2132 return Changed ? &I : 0;
2135 // isSignBit - Return true if the value represented by the constant only has the
2136 // highest order bit set.
2137 static bool isSignBit(ConstantInt *CI) {
2138 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2139 return CI->getValue() == APInt::getSignBit(NumBits);
2142 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2143 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2145 if (Op0 == Op1) // sub X, X -> 0
2146 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2148 // If this is a 'B = x-(-A)', change to B = x+A...
2149 if (Value *V = dyn_castNegVal(Op1))
2150 return BinaryOperator::createAdd(Op0, V);
2152 if (isa<UndefValue>(Op0))
2153 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2154 if (isa<UndefValue>(Op1))
2155 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2157 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2158 // Replace (-1 - A) with (~A)...
2159 if (C->isAllOnesValue())
2160 return BinaryOperator::createNot(Op1);
2162 // C - ~X == X + (1+C)
2164 if (match(Op1, m_Not(m_Value(X))))
2165 return BinaryOperator::createAdd(X, AddOne(C));
2167 // -(X >>u 31) -> (X >>s 31)
2168 // -(X >>s 31) -> (X >>u 31)
2170 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2171 if (SI->getOpcode() == Instruction::LShr) {
2172 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2173 // Check to see if we are shifting out everything but the sign bit.
2174 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2175 SI->getType()->getPrimitiveSizeInBits()-1) {
2176 // Ok, the transformation is safe. Insert AShr.
2177 return BinaryOperator::create(Instruction::AShr,
2178 SI->getOperand(0), CU, SI->getName());
2182 else if (SI->getOpcode() == Instruction::AShr) {
2183 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2184 // Check to see if we are shifting out everything but the sign bit.
2185 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2186 SI->getType()->getPrimitiveSizeInBits()-1) {
2187 // Ok, the transformation is safe. Insert LShr.
2188 return BinaryOperator::createLShr(
2189 SI->getOperand(0), CU, SI->getName());
2195 // Try to fold constant sub into select arguments.
2196 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2200 if (isa<PHINode>(Op0))
2201 if (Instruction *NV = FoldOpIntoPhi(I))
2205 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2206 if (Op1I->getOpcode() == Instruction::Add &&
2207 !Op0->getType()->isFPOrFPVector()) {
2208 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2209 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2210 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2211 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2212 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2213 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2214 // C1-(X+C2) --> (C1-C2)-X
2215 return BinaryOperator::createSub(Subtract(CI1, CI2),
2216 Op1I->getOperand(0));
2220 if (Op1I->hasOneUse()) {
2221 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2222 // is not used by anyone else...
2224 if (Op1I->getOpcode() == Instruction::Sub &&
2225 !Op1I->getType()->isFPOrFPVector()) {
2226 // Swap the two operands of the subexpr...
2227 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2228 Op1I->setOperand(0, IIOp1);
2229 Op1I->setOperand(1, IIOp0);
2231 // Create the new top level add instruction...
2232 return BinaryOperator::createAdd(Op0, Op1);
2235 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2237 if (Op1I->getOpcode() == Instruction::And &&
2238 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2239 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2242 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2243 return BinaryOperator::createAnd(Op0, NewNot);
2246 // 0 - (X sdiv C) -> (X sdiv -C)
2247 if (Op1I->getOpcode() == Instruction::SDiv)
2248 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2250 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2251 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2252 ConstantExpr::getNeg(DivRHS));
2254 // X - X*C --> X * (1-C)
2255 ConstantInt *C2 = 0;
2256 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2257 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2258 return BinaryOperator::createMul(Op0, CP1);
2261 // X - ((X / Y) * Y) --> X % Y
2262 if (Op1I->getOpcode() == Instruction::Mul)
2263 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2264 if (Op0 == I->getOperand(0) &&
2265 Op1I->getOperand(1) == I->getOperand(1)) {
2266 if (I->getOpcode() == Instruction::SDiv)
2267 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2268 if (I->getOpcode() == Instruction::UDiv)
2269 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2274 if (!Op0->getType()->isFPOrFPVector())
2275 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2276 if (Op0I->getOpcode() == Instruction::Add) {
2277 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2278 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2279 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2280 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2281 } else if (Op0I->getOpcode() == Instruction::Sub) {
2282 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2283 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2287 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2288 if (X == Op1) // X*C - X --> X * (C-1)
2289 return BinaryOperator::createMul(Op1, SubOne(C1));
2291 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2292 if (X == dyn_castFoldableMul(Op1, C2))
2293 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2298 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2299 /// comparison only checks the sign bit. If it only checks the sign bit, set
2300 /// TrueIfSigned if the result of the comparison is true when the input value is
2302 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2303 bool &TrueIfSigned) {
2305 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2306 TrueIfSigned = true;
2307 return RHS->isZero();
2308 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2309 TrueIfSigned = true;
2310 return RHS->isAllOnesValue();
2311 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2312 TrueIfSigned = false;
2313 return RHS->isAllOnesValue();
2314 case ICmpInst::ICMP_UGT:
2315 // True if LHS u> RHS and RHS == high-bit-mask - 1
2316 TrueIfSigned = true;
2317 return RHS->getValue() ==
2318 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2319 case ICmpInst::ICMP_UGE:
2320 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2321 TrueIfSigned = true;
2322 return RHS->getValue() ==
2323 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2329 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2330 bool Changed = SimplifyCommutative(I);
2331 Value *Op0 = I.getOperand(0);
2333 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2334 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2336 // Simplify mul instructions with a constant RHS...
2337 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2338 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2340 // ((X << C1)*C2) == (X * (C2 << C1))
2341 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2342 if (SI->getOpcode() == Instruction::Shl)
2343 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2344 return BinaryOperator::createMul(SI->getOperand(0),
2345 ConstantExpr::getShl(CI, ShOp));
2348 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2349 if (CI->equalsInt(1)) // X * 1 == X
2350 return ReplaceInstUsesWith(I, Op0);
2351 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2352 return BinaryOperator::createNeg(Op0, I.getName());
2354 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2355 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2356 return BinaryOperator::createShl(Op0,
2357 ConstantInt::get(Op0->getType(), Val.logBase2()));
2359 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2360 if (Op1F->isNullValue())
2361 return ReplaceInstUsesWith(I, Op1);
2363 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2364 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2365 // We need a better interface for long double here.
2366 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2367 if (Op1F->isExactlyValue(1.0))
2368 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2371 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2372 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2373 isa<ConstantInt>(Op0I->getOperand(1))) {
2374 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2375 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2377 InsertNewInstBefore(Add, I);
2378 Value *C1C2 = ConstantExpr::getMul(Op1,
2379 cast<Constant>(Op0I->getOperand(1)));
2380 return BinaryOperator::createAdd(Add, C1C2);
2384 // Try to fold constant mul into select arguments.
2385 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2386 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2389 if (isa<PHINode>(Op0))
2390 if (Instruction *NV = FoldOpIntoPhi(I))
2394 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2395 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2396 return BinaryOperator::createMul(Op0v, Op1v);
2398 // If one of the operands of the multiply is a cast from a boolean value, then
2399 // we know the bool is either zero or one, so this is a 'masking' multiply.
2400 // See if we can simplify things based on how the boolean was originally
2402 CastInst *BoolCast = 0;
2403 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2404 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2407 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2408 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2411 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2412 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2413 const Type *SCOpTy = SCIOp0->getType();
2416 // If the icmp is true iff the sign bit of X is set, then convert this
2417 // multiply into a shift/and combination.
2418 if (isa<ConstantInt>(SCIOp1) &&
2419 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2421 // Shift the X value right to turn it into "all signbits".
2422 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2423 SCOpTy->getPrimitiveSizeInBits()-1);
2425 InsertNewInstBefore(
2426 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2427 BoolCast->getOperand(0)->getName()+
2430 // If the multiply type is not the same as the source type, sign extend
2431 // or truncate to the multiply type.
2432 if (I.getType() != V->getType()) {
2433 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2434 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2435 Instruction::CastOps opcode =
2436 (SrcBits == DstBits ? Instruction::BitCast :
2437 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2438 V = InsertCastBefore(opcode, V, I.getType(), I);
2441 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2442 return BinaryOperator::createAnd(V, OtherOp);
2447 return Changed ? &I : 0;
2450 /// This function implements the transforms on div instructions that work
2451 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2452 /// used by the visitors to those instructions.
2453 /// @brief Transforms common to all three div instructions
2454 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2455 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2458 if (isa<UndefValue>(Op0))
2459 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2461 // X / undef -> undef
2462 if (isa<UndefValue>(Op1))
2463 return ReplaceInstUsesWith(I, Op1);
2465 // Handle cases involving: div X, (select Cond, Y, Z)
2466 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2467 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2468 // same basic block, then we replace the select with Y, and the condition
2469 // of the select with false (if the cond value is in the same BB). If the
2470 // select has uses other than the div, this allows them to be simplified
2471 // also. Note that div X, Y is just as good as div X, 0 (undef)
2472 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2473 if (ST->isNullValue()) {
2474 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2475 if (CondI && CondI->getParent() == I.getParent())
2476 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2477 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2478 I.setOperand(1, SI->getOperand(2));
2480 UpdateValueUsesWith(SI, SI->getOperand(2));
2484 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2485 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2486 if (ST->isNullValue()) {
2487 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2488 if (CondI && CondI->getParent() == I.getParent())
2489 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2490 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2491 I.setOperand(1, SI->getOperand(1));
2493 UpdateValueUsesWith(SI, SI->getOperand(1));
2501 /// This function implements the transforms common to both integer division
2502 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2503 /// division instructions.
2504 /// @brief Common integer divide transforms
2505 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2506 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2508 if (Instruction *Common = commonDivTransforms(I))
2511 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2513 if (RHS->equalsInt(1))
2514 return ReplaceInstUsesWith(I, Op0);
2516 // (X / C1) / C2 -> X / (C1*C2)
2517 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2518 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2519 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2520 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2521 Multiply(RHS, LHSRHS));
2524 if (!RHS->isZero()) { // avoid X udiv 0
2525 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2526 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2528 if (isa<PHINode>(Op0))
2529 if (Instruction *NV = FoldOpIntoPhi(I))
2534 // 0 / X == 0, we don't need to preserve faults!
2535 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2536 if (LHS->equalsInt(0))
2537 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2542 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2543 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2545 // Handle the integer div common cases
2546 if (Instruction *Common = commonIDivTransforms(I))
2549 // X udiv C^2 -> X >> C
2550 // Check to see if this is an unsigned division with an exact power of 2,
2551 // if so, convert to a right shift.
2552 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2553 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2554 return BinaryOperator::createLShr(Op0,
2555 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2558 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2559 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2560 if (RHSI->getOpcode() == Instruction::Shl &&
2561 isa<ConstantInt>(RHSI->getOperand(0))) {
2562 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2563 if (C1.isPowerOf2()) {
2564 Value *N = RHSI->getOperand(1);
2565 const Type *NTy = N->getType();
2566 if (uint32_t C2 = C1.logBase2()) {
2567 Constant *C2V = ConstantInt::get(NTy, C2);
2568 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2570 return BinaryOperator::createLShr(Op0, N);
2575 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2576 // where C1&C2 are powers of two.
2577 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2578 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2579 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2580 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2581 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2582 // Compute the shift amounts
2583 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2584 // Construct the "on true" case of the select
2585 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2586 Instruction *TSI = BinaryOperator::createLShr(
2587 Op0, TC, SI->getName()+".t");
2588 TSI = InsertNewInstBefore(TSI, I);
2590 // Construct the "on false" case of the select
2591 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2592 Instruction *FSI = BinaryOperator::createLShr(
2593 Op0, FC, SI->getName()+".f");
2594 FSI = InsertNewInstBefore(FSI, I);
2596 // construct the select instruction and return it.
2597 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2603 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2604 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2606 // Handle the integer div common cases
2607 if (Instruction *Common = commonIDivTransforms(I))
2610 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2612 if (RHS->isAllOnesValue())
2613 return BinaryOperator::createNeg(Op0);
2616 if (Value *LHSNeg = dyn_castNegVal(Op0))
2617 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2620 // If the sign bits of both operands are zero (i.e. we can prove they are
2621 // unsigned inputs), turn this into a udiv.
2622 if (I.getType()->isInteger()) {
2623 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2624 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2625 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2626 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2633 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2634 return commonDivTransforms(I);
2637 /// GetFactor - If we can prove that the specified value is at least a multiple
2638 /// of some factor, return that factor.
2639 static Constant *GetFactor(Value *V) {
2640 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2643 // Unless we can be tricky, we know this is a multiple of 1.
2644 Constant *Result = ConstantInt::get(V->getType(), 1);
2646 Instruction *I = dyn_cast<Instruction>(V);
2647 if (!I) return Result;
2649 if (I->getOpcode() == Instruction::Mul) {
2650 // Handle multiplies by a constant, etc.
2651 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2652 GetFactor(I->getOperand(1)));
2653 } else if (I->getOpcode() == Instruction::Shl) {
2654 // (X<<C) -> X * (1 << C)
2655 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2656 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2657 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2659 } else if (I->getOpcode() == Instruction::And) {
2660 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2661 // X & 0xFFF0 is known to be a multiple of 16.
2662 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2663 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2664 return ConstantExpr::getShl(Result,
2665 ConstantInt::get(Result->getType(), Zeros));
2667 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2668 // Only handle int->int casts.
2669 if (!CI->isIntegerCast())
2671 Value *Op = CI->getOperand(0);
2672 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2677 /// This function implements the transforms on rem instructions that work
2678 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2679 /// is used by the visitors to those instructions.
2680 /// @brief Transforms common to all three rem instructions
2681 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2682 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2684 // 0 % X == 0, we don't need to preserve faults!
2685 if (Constant *LHS = dyn_cast<Constant>(Op0))
2686 if (LHS->isNullValue())
2687 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2689 if (isa<UndefValue>(Op0)) // undef % X -> 0
2690 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2691 if (isa<UndefValue>(Op1))
2692 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2694 // Handle cases involving: rem X, (select Cond, Y, Z)
2695 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2696 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2697 // the same basic block, then we replace the select with Y, and the
2698 // condition of the select with false (if the cond value is in the same
2699 // BB). If the select has uses other than the div, this allows them to be
2701 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2702 if (ST->isNullValue()) {
2703 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2704 if (CondI && CondI->getParent() == I.getParent())
2705 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2706 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2707 I.setOperand(1, SI->getOperand(2));
2709 UpdateValueUsesWith(SI, SI->getOperand(2));
2712 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2713 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2714 if (ST->isNullValue()) {
2715 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2716 if (CondI && CondI->getParent() == I.getParent())
2717 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2718 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2719 I.setOperand(1, SI->getOperand(1));
2721 UpdateValueUsesWith(SI, SI->getOperand(1));
2729 /// This function implements the transforms common to both integer remainder
2730 /// instructions (urem and srem). It is called by the visitors to those integer
2731 /// remainder instructions.
2732 /// @brief Common integer remainder transforms
2733 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2734 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2736 if (Instruction *common = commonRemTransforms(I))
2739 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2740 // X % 0 == undef, we don't need to preserve faults!
2741 if (RHS->equalsInt(0))
2742 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2744 if (RHS->equalsInt(1)) // X % 1 == 0
2745 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2747 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2748 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2749 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2751 } else if (isa<PHINode>(Op0I)) {
2752 if (Instruction *NV = FoldOpIntoPhi(I))
2755 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2756 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2757 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2764 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2765 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2767 if (Instruction *common = commonIRemTransforms(I))
2770 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2771 // X urem C^2 -> X and C
2772 // Check to see if this is an unsigned remainder with an exact power of 2,
2773 // if so, convert to a bitwise and.
2774 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2775 if (C->getValue().isPowerOf2())
2776 return BinaryOperator::createAnd(Op0, SubOne(C));
2779 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2780 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2781 if (RHSI->getOpcode() == Instruction::Shl &&
2782 isa<ConstantInt>(RHSI->getOperand(0))) {
2783 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2784 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2785 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2787 return BinaryOperator::createAnd(Op0, Add);
2792 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2793 // where C1&C2 are powers of two.
2794 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2795 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2796 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2797 // STO == 0 and SFO == 0 handled above.
2798 if ((STO->getValue().isPowerOf2()) &&
2799 (SFO->getValue().isPowerOf2())) {
2800 Value *TrueAnd = InsertNewInstBefore(
2801 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2802 Value *FalseAnd = InsertNewInstBefore(
2803 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2804 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2812 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2813 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2815 // Handle the integer rem common cases
2816 if (Instruction *common = commonIRemTransforms(I))
2819 if (Value *RHSNeg = dyn_castNegVal(Op1))
2820 if (!isa<ConstantInt>(RHSNeg) ||
2821 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2823 AddUsesToWorkList(I);
2824 I.setOperand(1, RHSNeg);
2828 // If the sign bits of both operands are zero (i.e. we can prove they are
2829 // unsigned inputs), turn this into a urem.
2830 if (I.getType()->isInteger()) {
2831 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2832 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2833 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2834 return BinaryOperator::createURem(Op0, Op1, I.getName());
2841 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2842 return commonRemTransforms(I);
2845 // isMaxValueMinusOne - return true if this is Max-1
2846 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2847 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2849 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2850 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2853 // isMinValuePlusOne - return true if this is Min+1
2854 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2856 return C->getValue() == 1; // unsigned
2858 // Calculate 1111111111000000000000
2859 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2860 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2863 // isOneBitSet - Return true if there is exactly one bit set in the specified
2865 static bool isOneBitSet(const ConstantInt *CI) {
2866 return CI->getValue().isPowerOf2();
2869 // isHighOnes - Return true if the constant is of the form 1+0+.
2870 // This is the same as lowones(~X).
2871 static bool isHighOnes(const ConstantInt *CI) {
2872 return (~CI->getValue() + 1).isPowerOf2();
2875 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2876 /// are carefully arranged to allow folding of expressions such as:
2878 /// (A < B) | (A > B) --> (A != B)
2880 /// Note that this is only valid if the first and second predicates have the
2881 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2883 /// Three bits are used to represent the condition, as follows:
2888 /// <=> Value Definition
2889 /// 000 0 Always false
2896 /// 111 7 Always true
2898 static unsigned getICmpCode(const ICmpInst *ICI) {
2899 switch (ICI->getPredicate()) {
2901 case ICmpInst::ICMP_UGT: return 1; // 001
2902 case ICmpInst::ICMP_SGT: return 1; // 001
2903 case ICmpInst::ICMP_EQ: return 2; // 010
2904 case ICmpInst::ICMP_UGE: return 3; // 011
2905 case ICmpInst::ICMP_SGE: return 3; // 011
2906 case ICmpInst::ICMP_ULT: return 4; // 100
2907 case ICmpInst::ICMP_SLT: return 4; // 100
2908 case ICmpInst::ICMP_NE: return 5; // 101
2909 case ICmpInst::ICMP_ULE: return 6; // 110
2910 case ICmpInst::ICMP_SLE: return 6; // 110
2913 assert(0 && "Invalid ICmp predicate!");
2918 /// getICmpValue - This is the complement of getICmpCode, which turns an
2919 /// opcode and two operands into either a constant true or false, or a brand
2920 /// new ICmp instruction. The sign is passed in to determine which kind
2921 /// of predicate to use in new icmp instructions.
2922 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2924 default: assert(0 && "Illegal ICmp code!");
2925 case 0: return ConstantInt::getFalse();
2928 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2930 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2931 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2934 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2936 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2939 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2941 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2942 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2945 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2947 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2948 case 7: return ConstantInt::getTrue();
2952 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2953 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2954 (ICmpInst::isSignedPredicate(p1) &&
2955 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2956 (ICmpInst::isSignedPredicate(p2) &&
2957 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2961 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2962 struct FoldICmpLogical {
2965 ICmpInst::Predicate pred;
2966 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2967 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2968 pred(ICI->getPredicate()) {}
2969 bool shouldApply(Value *V) const {
2970 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2971 if (PredicatesFoldable(pred, ICI->getPredicate()))
2972 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2973 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2976 Instruction *apply(Instruction &Log) const {
2977 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2978 if (ICI->getOperand(0) != LHS) {
2979 assert(ICI->getOperand(1) == LHS);
2980 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2983 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2984 unsigned LHSCode = getICmpCode(ICI);
2985 unsigned RHSCode = getICmpCode(RHSICI);
2987 switch (Log.getOpcode()) {
2988 case Instruction::And: Code = LHSCode & RHSCode; break;
2989 case Instruction::Or: Code = LHSCode | RHSCode; break;
2990 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2991 default: assert(0 && "Illegal logical opcode!"); return 0;
2994 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2995 ICmpInst::isSignedPredicate(ICI->getPredicate());
2997 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2998 if (Instruction *I = dyn_cast<Instruction>(RV))
3000 // Otherwise, it's a constant boolean value...
3001 return IC.ReplaceInstUsesWith(Log, RV);
3004 } // end anonymous namespace
3006 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3007 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3008 // guaranteed to be a binary operator.
3009 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3011 ConstantInt *AndRHS,
3012 BinaryOperator &TheAnd) {
3013 Value *X = Op->getOperand(0);
3014 Constant *Together = 0;
3016 Together = And(AndRHS, OpRHS);
3018 switch (Op->getOpcode()) {
3019 case Instruction::Xor:
3020 if (Op->hasOneUse()) {
3021 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3022 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3023 InsertNewInstBefore(And, TheAnd);
3025 return BinaryOperator::createXor(And, Together);
3028 case Instruction::Or:
3029 if (Together == AndRHS) // (X | C) & C --> C
3030 return ReplaceInstUsesWith(TheAnd, AndRHS);
3032 if (Op->hasOneUse() && Together != OpRHS) {
3033 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3034 Instruction *Or = BinaryOperator::createOr(X, Together);
3035 InsertNewInstBefore(Or, TheAnd);
3037 return BinaryOperator::createAnd(Or, AndRHS);
3040 case Instruction::Add:
3041 if (Op->hasOneUse()) {
3042 // Adding a one to a single bit bit-field should be turned into an XOR
3043 // of the bit. First thing to check is to see if this AND is with a
3044 // single bit constant.
3045 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3047 // If there is only one bit set...
3048 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3049 // Ok, at this point, we know that we are masking the result of the
3050 // ADD down to exactly one bit. If the constant we are adding has
3051 // no bits set below this bit, then we can eliminate the ADD.
3052 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3054 // Check to see if any bits below the one bit set in AndRHSV are set.
3055 if ((AddRHS & (AndRHSV-1)) == 0) {
3056 // If not, the only thing that can effect the output of the AND is
3057 // the bit specified by AndRHSV. If that bit is set, the effect of
3058 // the XOR is to toggle the bit. If it is clear, then the ADD has
3060 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3061 TheAnd.setOperand(0, X);
3064 // Pull the XOR out of the AND.
3065 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3066 InsertNewInstBefore(NewAnd, TheAnd);
3067 NewAnd->takeName(Op);
3068 return BinaryOperator::createXor(NewAnd, AndRHS);
3075 case Instruction::Shl: {
3076 // We know that the AND will not produce any of the bits shifted in, so if
3077 // the anded constant includes them, clear them now!
3079 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3080 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3081 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3082 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3084 if (CI->getValue() == ShlMask) {
3085 // Masking out bits that the shift already masks
3086 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3087 } else if (CI != AndRHS) { // Reducing bits set in and.
3088 TheAnd.setOperand(1, CI);
3093 case Instruction::LShr:
3095 // We know that the AND will not produce any of the bits shifted in, so if
3096 // the anded constant includes them, clear them now! This only applies to
3097 // unsigned shifts, because a signed shr may bring in set bits!
3099 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3100 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3101 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3102 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3104 if (CI->getValue() == ShrMask) {
3105 // Masking out bits that the shift already masks.
3106 return ReplaceInstUsesWith(TheAnd, Op);
3107 } else if (CI != AndRHS) {
3108 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3113 case Instruction::AShr:
3115 // See if this is shifting in some sign extension, then masking it out
3117 if (Op->hasOneUse()) {
3118 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3119 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3120 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3121 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3122 if (C == AndRHS) { // Masking out bits shifted in.
3123 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3124 // Make the argument unsigned.
3125 Value *ShVal = Op->getOperand(0);
3126 ShVal = InsertNewInstBefore(
3127 BinaryOperator::createLShr(ShVal, OpRHS,
3128 Op->getName()), TheAnd);
3129 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3138 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3139 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3140 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3141 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3142 /// insert new instructions.
3143 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3144 bool isSigned, bool Inside,
3146 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3147 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3148 "Lo is not <= Hi in range emission code!");
3151 if (Lo == Hi) // Trivially false.
3152 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3154 // V >= Min && V < Hi --> V < Hi
3155 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3156 ICmpInst::Predicate pred = (isSigned ?
3157 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3158 return new ICmpInst(pred, V, Hi);
3161 // Emit V-Lo <u Hi-Lo
3162 Constant *NegLo = ConstantExpr::getNeg(Lo);
3163 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3164 InsertNewInstBefore(Add, IB);
3165 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3166 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3169 if (Lo == Hi) // Trivially true.
3170 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3172 // V < Min || V >= Hi -> V > Hi-1
3173 Hi = SubOne(cast<ConstantInt>(Hi));
3174 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3175 ICmpInst::Predicate pred = (isSigned ?
3176 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3177 return new ICmpInst(pred, V, Hi);
3180 // Emit V-Lo >u Hi-1-Lo
3181 // Note that Hi has already had one subtracted from it, above.
3182 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3183 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3184 InsertNewInstBefore(Add, IB);
3185 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3186 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3189 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3190 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3191 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3192 // not, since all 1s are not contiguous.
3193 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3194 const APInt& V = Val->getValue();
3195 uint32_t BitWidth = Val->getType()->getBitWidth();
3196 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3198 // look for the first zero bit after the run of ones
3199 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3200 // look for the first non-zero bit
3201 ME = V.getActiveBits();
3205 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3206 /// where isSub determines whether the operator is a sub. If we can fold one of
3207 /// the following xforms:
3209 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3210 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3211 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3213 /// return (A +/- B).
3215 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3216 ConstantInt *Mask, bool isSub,
3218 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3219 if (!LHSI || LHSI->getNumOperands() != 2 ||
3220 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3222 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3224 switch (LHSI->getOpcode()) {
3226 case Instruction::And:
3227 if (And(N, Mask) == Mask) {
3228 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3229 if ((Mask->getValue().countLeadingZeros() +
3230 Mask->getValue().countPopulation()) ==
3231 Mask->getValue().getBitWidth())
3234 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3235 // part, we don't need any explicit masks to take them out of A. If that
3236 // is all N is, ignore it.
3237 uint32_t MB = 0, ME = 0;
3238 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3239 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3240 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3241 if (MaskedValueIsZero(RHS, Mask))
3246 case Instruction::Or:
3247 case Instruction::Xor:
3248 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3249 if ((Mask->getValue().countLeadingZeros() +
3250 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3251 && And(N, Mask)->isZero())
3258 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3260 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3261 return InsertNewInstBefore(New, I);
3264 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3265 bool Changed = SimplifyCommutative(I);
3266 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3268 if (isa<UndefValue>(Op1)) // X & undef -> 0
3269 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3273 return ReplaceInstUsesWith(I, Op1);
3275 // See if we can simplify any instructions used by the instruction whose sole
3276 // purpose is to compute bits we don't care about.
3277 if (!isa<VectorType>(I.getType())) {
3278 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3279 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3280 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3281 KnownZero, KnownOne))
3284 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3285 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3286 return ReplaceInstUsesWith(I, I.getOperand(0));
3287 } else if (isa<ConstantAggregateZero>(Op1)) {
3288 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3292 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3293 const APInt& AndRHSMask = AndRHS->getValue();
3294 APInt NotAndRHS(~AndRHSMask);
3296 // Optimize a variety of ((val OP C1) & C2) combinations...
3297 if (isa<BinaryOperator>(Op0)) {
3298 Instruction *Op0I = cast<Instruction>(Op0);
3299 Value *Op0LHS = Op0I->getOperand(0);
3300 Value *Op0RHS = Op0I->getOperand(1);
3301 switch (Op0I->getOpcode()) {
3302 case Instruction::Xor:
3303 case Instruction::Or:
3304 // If the mask is only needed on one incoming arm, push it up.
3305 if (Op0I->hasOneUse()) {
3306 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3307 // Not masking anything out for the LHS, move to RHS.
3308 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3309 Op0RHS->getName()+".masked");
3310 InsertNewInstBefore(NewRHS, I);
3311 return BinaryOperator::create(
3312 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3314 if (!isa<Constant>(Op0RHS) &&
3315 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3316 // Not masking anything out for the RHS, move to LHS.
3317 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3318 Op0LHS->getName()+".masked");
3319 InsertNewInstBefore(NewLHS, I);
3320 return BinaryOperator::create(
3321 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3326 case Instruction::Add:
3327 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3328 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3329 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3330 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3331 return BinaryOperator::createAnd(V, AndRHS);
3332 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3333 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3336 case Instruction::Sub:
3337 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3338 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3339 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3340 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3341 return BinaryOperator::createAnd(V, AndRHS);
3345 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3346 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3348 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3349 // If this is an integer truncation or change from signed-to-unsigned, and
3350 // if the source is an and/or with immediate, transform it. This
3351 // frequently occurs for bitfield accesses.
3352 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3353 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3354 CastOp->getNumOperands() == 2)
3355 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3356 if (CastOp->getOpcode() == Instruction::And) {
3357 // Change: and (cast (and X, C1) to T), C2
3358 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3359 // This will fold the two constants together, which may allow
3360 // other simplifications.
3361 Instruction *NewCast = CastInst::createTruncOrBitCast(
3362 CastOp->getOperand(0), I.getType(),
3363 CastOp->getName()+".shrunk");
3364 NewCast = InsertNewInstBefore(NewCast, I);
3365 // trunc_or_bitcast(C1)&C2
3366 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3367 C3 = ConstantExpr::getAnd(C3, AndRHS);
3368 return BinaryOperator::createAnd(NewCast, C3);
3369 } else if (CastOp->getOpcode() == Instruction::Or) {
3370 // Change: and (cast (or X, C1) to T), C2
3371 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3372 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3373 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3374 return ReplaceInstUsesWith(I, AndRHS);
3379 // Try to fold constant and into select arguments.
3380 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3381 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3383 if (isa<PHINode>(Op0))
3384 if (Instruction *NV = FoldOpIntoPhi(I))
3388 Value *Op0NotVal = dyn_castNotVal(Op0);
3389 Value *Op1NotVal = dyn_castNotVal(Op1);
3391 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3392 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3394 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3395 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3396 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3397 I.getName()+".demorgan");
3398 InsertNewInstBefore(Or, I);
3399 return BinaryOperator::createNot(Or);
3403 Value *A = 0, *B = 0, *C = 0, *D = 0;
3404 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3405 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3406 return ReplaceInstUsesWith(I, Op1);
3408 // (A|B) & ~(A&B) -> A^B
3409 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3410 if ((A == C && B == D) || (A == D && B == C))
3411 return BinaryOperator::createXor(A, B);
3415 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3416 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3417 return ReplaceInstUsesWith(I, Op0);
3419 // ~(A&B) & (A|B) -> A^B
3420 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3421 if ((A == C && B == D) || (A == D && B == C))
3422 return BinaryOperator::createXor(A, B);
3426 if (Op0->hasOneUse() &&
3427 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3428 if (A == Op1) { // (A^B)&A -> A&(A^B)
3429 I.swapOperands(); // Simplify below
3430 std::swap(Op0, Op1);
3431 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3432 cast<BinaryOperator>(Op0)->swapOperands();
3433 I.swapOperands(); // Simplify below
3434 std::swap(Op0, Op1);
3437 if (Op1->hasOneUse() &&
3438 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3439 if (B == Op0) { // B&(A^B) -> B&(B^A)
3440 cast<BinaryOperator>(Op1)->swapOperands();
3443 if (A == Op0) { // A&(A^B) -> A & ~B
3444 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3445 InsertNewInstBefore(NotB, I);
3446 return BinaryOperator::createAnd(A, NotB);
3451 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3452 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3453 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3456 Value *LHSVal, *RHSVal;
3457 ConstantInt *LHSCst, *RHSCst;
3458 ICmpInst::Predicate LHSCC, RHSCC;
3459 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3460 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3461 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3462 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3463 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3464 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3465 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3466 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3468 // Don't try to fold ICMP_SLT + ICMP_ULT.
3469 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3470 ICmpInst::isSignedPredicate(LHSCC) ==
3471 ICmpInst::isSignedPredicate(RHSCC))) {
3472 // Ensure that the larger constant is on the RHS.
3473 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3474 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3475 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3476 ICmpInst *LHS = cast<ICmpInst>(Op0);
3477 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3478 std::swap(LHS, RHS);
3479 std::swap(LHSCst, RHSCst);
3480 std::swap(LHSCC, RHSCC);
3483 // At this point, we know we have have two icmp instructions
3484 // comparing a value against two constants and and'ing the result
3485 // together. Because of the above check, we know that we only have
3486 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3487 // (from the FoldICmpLogical check above), that the two constants
3488 // are not equal and that the larger constant is on the RHS
3489 assert(LHSCst != RHSCst && "Compares not folded above?");
3492 default: assert(0 && "Unknown integer condition code!");
3493 case ICmpInst::ICMP_EQ:
3495 default: assert(0 && "Unknown integer condition code!");
3496 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3497 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3498 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3499 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3500 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3501 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3502 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3503 return ReplaceInstUsesWith(I, LHS);
3505 case ICmpInst::ICMP_NE:
3507 default: assert(0 && "Unknown integer condition code!");
3508 case ICmpInst::ICMP_ULT:
3509 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3510 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3511 break; // (X != 13 & X u< 15) -> no change
3512 case ICmpInst::ICMP_SLT:
3513 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3514 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3515 break; // (X != 13 & X s< 15) -> no change
3516 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3517 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3518 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3519 return ReplaceInstUsesWith(I, RHS);
3520 case ICmpInst::ICMP_NE:
3521 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3522 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3523 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3524 LHSVal->getName()+".off");
3525 InsertNewInstBefore(Add, I);
3526 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3527 ConstantInt::get(Add->getType(), 1));
3529 break; // (X != 13 & X != 15) -> no change
3532 case ICmpInst::ICMP_ULT:
3534 default: assert(0 && "Unknown integer condition code!");
3535 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3536 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3537 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3538 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3540 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3541 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3542 return ReplaceInstUsesWith(I, LHS);
3543 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3547 case ICmpInst::ICMP_SLT:
3549 default: assert(0 && "Unknown integer condition code!");
3550 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3551 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3552 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3553 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3555 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3556 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3557 return ReplaceInstUsesWith(I, LHS);
3558 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3562 case ICmpInst::ICMP_UGT:
3564 default: assert(0 && "Unknown integer condition code!");
3565 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3566 return ReplaceInstUsesWith(I, LHS);
3567 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3568 return ReplaceInstUsesWith(I, RHS);
3569 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3571 case ICmpInst::ICMP_NE:
3572 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3573 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3574 break; // (X u> 13 & X != 15) -> no change
3575 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3576 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3578 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3582 case ICmpInst::ICMP_SGT:
3584 default: assert(0 && "Unknown integer condition code!");
3585 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3586 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3587 return ReplaceInstUsesWith(I, RHS);
3588 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3590 case ICmpInst::ICMP_NE:
3591 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3592 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3593 break; // (X s> 13 & X != 15) -> no change
3594 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3595 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3597 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3605 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3606 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3607 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3608 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3609 const Type *SrcTy = Op0C->getOperand(0)->getType();
3610 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3611 // Only do this if the casts both really cause code to be generated.
3612 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3614 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3616 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3617 Op1C->getOperand(0),
3619 InsertNewInstBefore(NewOp, I);
3620 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3624 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3625 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3626 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3627 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3628 SI0->getOperand(1) == SI1->getOperand(1) &&
3629 (SI0->hasOneUse() || SI1->hasOneUse())) {
3630 Instruction *NewOp =
3631 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3633 SI0->getName()), I);
3634 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3635 SI1->getOperand(1));
3639 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3640 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3641 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3642 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3643 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3644 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3645 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3646 // If either of the constants are nans, then the whole thing returns
3648 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3649 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3650 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3651 RHS->getOperand(0));
3656 return Changed ? &I : 0;
3659 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3660 /// in the result. If it does, and if the specified byte hasn't been filled in
3661 /// yet, fill it in and return false.
3662 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3663 Instruction *I = dyn_cast<Instruction>(V);
3664 if (I == 0) return true;
3666 // If this is an or instruction, it is an inner node of the bswap.
3667 if (I->getOpcode() == Instruction::Or)
3668 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3669 CollectBSwapParts(I->getOperand(1), ByteValues);
3671 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3672 // If this is a shift by a constant int, and it is "24", then its operand
3673 // defines a byte. We only handle unsigned types here.
3674 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3675 // Not shifting the entire input by N-1 bytes?
3676 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3677 8*(ByteValues.size()-1))
3681 if (I->getOpcode() == Instruction::Shl) {
3682 // X << 24 defines the top byte with the lowest of the input bytes.
3683 DestNo = ByteValues.size()-1;
3685 // X >>u 24 defines the low byte with the highest of the input bytes.
3689 // If the destination byte value is already defined, the values are or'd
3690 // together, which isn't a bswap (unless it's an or of the same bits).
3691 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3693 ByteValues[DestNo] = I->getOperand(0);
3697 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3699 Value *Shift = 0, *ShiftLHS = 0;
3700 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3701 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3702 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3704 Instruction *SI = cast<Instruction>(Shift);
3706 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3707 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3708 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3711 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3713 if (AndAmt->getValue().getActiveBits() > 64)
3715 uint64_t AndAmtVal = AndAmt->getZExtValue();
3716 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3717 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3719 // Unknown mask for bswap.
3720 if (DestByte == ByteValues.size()) return true;
3722 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3724 if (SI->getOpcode() == Instruction::Shl)
3725 SrcByte = DestByte - ShiftBytes;
3727 SrcByte = DestByte + ShiftBytes;
3729 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3730 if (SrcByte != ByteValues.size()-DestByte-1)
3733 // If the destination byte value is already defined, the values are or'd
3734 // together, which isn't a bswap (unless it's an or of the same bits).
3735 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3737 ByteValues[DestByte] = SI->getOperand(0);
3741 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3742 /// If so, insert the new bswap intrinsic and return it.
3743 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3744 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3745 if (!ITy || ITy->getBitWidth() % 16)
3746 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3748 /// ByteValues - For each byte of the result, we keep track of which value
3749 /// defines each byte.
3750 SmallVector<Value*, 8> ByteValues;
3751 ByteValues.resize(ITy->getBitWidth()/8);
3753 // Try to find all the pieces corresponding to the bswap.
3754 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3755 CollectBSwapParts(I.getOperand(1), ByteValues))
3758 // Check to see if all of the bytes come from the same value.
3759 Value *V = ByteValues[0];
3760 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3762 // Check to make sure that all of the bytes come from the same value.
3763 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3764 if (ByteValues[i] != V)
3766 const Type *Tys[] = { ITy };
3767 Module *M = I.getParent()->getParent()->getParent();
3768 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3769 return new CallInst(F, V);
3773 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3774 bool Changed = SimplifyCommutative(I);
3775 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3777 if (isa<UndefValue>(Op1)) // X | undef -> -1
3778 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3782 return ReplaceInstUsesWith(I, Op0);
3784 // See if we can simplify any instructions used by the instruction whose sole
3785 // purpose is to compute bits we don't care about.
3786 if (!isa<VectorType>(I.getType())) {
3787 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3788 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3789 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3790 KnownZero, KnownOne))
3792 } else if (isa<ConstantAggregateZero>(Op1)) {
3793 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3794 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3795 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3796 return ReplaceInstUsesWith(I, I.getOperand(1));
3802 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3803 ConstantInt *C1 = 0; Value *X = 0;
3804 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3805 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3806 Instruction *Or = BinaryOperator::createOr(X, RHS);
3807 InsertNewInstBefore(Or, I);
3809 return BinaryOperator::createAnd(Or,
3810 ConstantInt::get(RHS->getValue() | C1->getValue()));
3813 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3814 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3815 Instruction *Or = BinaryOperator::createOr(X, RHS);
3816 InsertNewInstBefore(Or, I);
3818 return BinaryOperator::createXor(Or,
3819 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3822 // Try to fold constant and into select arguments.
3823 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3824 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3826 if (isa<PHINode>(Op0))
3827 if (Instruction *NV = FoldOpIntoPhi(I))
3831 Value *A = 0, *B = 0;
3832 ConstantInt *C1 = 0, *C2 = 0;
3834 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3835 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3836 return ReplaceInstUsesWith(I, Op1);
3837 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3838 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3839 return ReplaceInstUsesWith(I, Op0);
3841 // (A | B) | C and A | (B | C) -> bswap if possible.
3842 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3843 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3844 match(Op1, m_Or(m_Value(), m_Value())) ||
3845 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3846 match(Op1, m_Shift(m_Value(), m_Value())))) {
3847 if (Instruction *BSwap = MatchBSwap(I))
3851 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3852 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3853 MaskedValueIsZero(Op1, C1->getValue())) {
3854 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3855 InsertNewInstBefore(NOr, I);
3857 return BinaryOperator::createXor(NOr, C1);
3860 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3861 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3862 MaskedValueIsZero(Op0, C1->getValue())) {
3863 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3864 InsertNewInstBefore(NOr, I);
3866 return BinaryOperator::createXor(NOr, C1);
3870 Value *C = 0, *D = 0;
3871 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3872 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3873 Value *V1 = 0, *V2 = 0, *V3 = 0;
3874 C1 = dyn_cast<ConstantInt>(C);
3875 C2 = dyn_cast<ConstantInt>(D);
3876 if (C1 && C2) { // (A & C1)|(B & C2)
3877 // If we have: ((V + N) & C1) | (V & C2)
3878 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3879 // replace with V+N.
3880 if (C1->getValue() == ~C2->getValue()) {
3881 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3882 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3883 // Add commutes, try both ways.
3884 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3885 return ReplaceInstUsesWith(I, A);
3886 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3887 return ReplaceInstUsesWith(I, A);
3889 // Or commutes, try both ways.
3890 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3891 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3892 // Add commutes, try both ways.
3893 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3894 return ReplaceInstUsesWith(I, B);
3895 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3896 return ReplaceInstUsesWith(I, B);
3899 V1 = 0; V2 = 0; V3 = 0;
3902 // Check to see if we have any common things being and'ed. If so, find the
3903 // terms for V1 & (V2|V3).
3904 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3905 if (A == B) // (A & C)|(A & D) == A & (C|D)
3906 V1 = A, V2 = C, V3 = D;
3907 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3908 V1 = A, V2 = B, V3 = C;
3909 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3910 V1 = C, V2 = A, V3 = D;
3911 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3912 V1 = C, V2 = A, V3 = B;
3916 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3917 return BinaryOperator::createAnd(V1, Or);
3922 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3923 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3924 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3925 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3926 SI0->getOperand(1) == SI1->getOperand(1) &&
3927 (SI0->hasOneUse() || SI1->hasOneUse())) {
3928 Instruction *NewOp =
3929 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3931 SI0->getName()), I);
3932 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3933 SI1->getOperand(1));
3937 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3938 if (A == Op1) // ~A | A == -1
3939 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3943 // Note, A is still live here!
3944 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3946 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3948 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3949 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3950 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3951 I.getName()+".demorgan"), I);
3952 return BinaryOperator::createNot(And);
3956 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3957 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3958 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3961 Value *LHSVal, *RHSVal;
3962 ConstantInt *LHSCst, *RHSCst;
3963 ICmpInst::Predicate LHSCC, RHSCC;
3964 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3965 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3966 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3967 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3968 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3969 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3970 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3971 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3972 // We can't fold (ugt x, C) | (sgt x, C2).
3973 PredicatesFoldable(LHSCC, RHSCC)) {
3974 // Ensure that the larger constant is on the RHS.
3975 ICmpInst *LHS = cast<ICmpInst>(Op0);
3977 if (ICmpInst::isSignedPredicate(LHSCC))
3978 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3980 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3983 std::swap(LHS, RHS);
3984 std::swap(LHSCst, RHSCst);
3985 std::swap(LHSCC, RHSCC);
3988 // At this point, we know we have have two icmp instructions
3989 // comparing a value against two constants and or'ing the result
3990 // together. Because of the above check, we know that we only have
3991 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3992 // FoldICmpLogical check above), that the two constants are not
3994 assert(LHSCst != RHSCst && "Compares not folded above?");
3997 default: assert(0 && "Unknown integer condition code!");
3998 case ICmpInst::ICMP_EQ:
4000 default: assert(0 && "Unknown integer condition code!");
4001 case ICmpInst::ICMP_EQ:
4002 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4003 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4004 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4005 LHSVal->getName()+".off");
4006 InsertNewInstBefore(Add, I);
4007 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4008 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4010 break; // (X == 13 | X == 15) -> no change
4011 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4012 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4014 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4015 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4016 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4017 return ReplaceInstUsesWith(I, RHS);
4020 case ICmpInst::ICMP_NE:
4022 default: assert(0 && "Unknown integer condition code!");
4023 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4024 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4025 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4026 return ReplaceInstUsesWith(I, LHS);
4027 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4028 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4029 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4030 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4033 case ICmpInst::ICMP_ULT:
4035 default: assert(0 && "Unknown integer condition code!");
4036 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4038 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4039 // If RHSCst is [us]MAXINT, it is always false. Not handling
4040 // this can cause overflow.
4041 if (RHSCst->isMaxValue(false))
4042 return ReplaceInstUsesWith(I, LHS);
4043 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4045 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4047 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4048 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4049 return ReplaceInstUsesWith(I, RHS);
4050 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4054 case ICmpInst::ICMP_SLT:
4056 default: assert(0 && "Unknown integer condition code!");
4057 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4059 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4060 // If RHSCst is [us]MAXINT, it is always false. Not handling
4061 // this can cause overflow.
4062 if (RHSCst->isMaxValue(true))
4063 return ReplaceInstUsesWith(I, LHS);
4064 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4066 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4068 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4069 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4070 return ReplaceInstUsesWith(I, RHS);
4071 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4075 case ICmpInst::ICMP_UGT:
4077 default: assert(0 && "Unknown integer condition code!");
4078 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4079 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4080 return ReplaceInstUsesWith(I, LHS);
4081 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4083 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4084 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4085 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4086 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4090 case ICmpInst::ICMP_SGT:
4092 default: assert(0 && "Unknown integer condition code!");
4093 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4094 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4095 return ReplaceInstUsesWith(I, LHS);
4096 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4098 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4099 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4100 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4101 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4109 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4110 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4111 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4112 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4113 const Type *SrcTy = Op0C->getOperand(0)->getType();
4114 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4115 // Only do this if the casts both really cause code to be generated.
4116 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4118 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4120 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4121 Op1C->getOperand(0),
4123 InsertNewInstBefore(NewOp, I);
4124 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4130 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4131 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4132 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4133 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4134 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4135 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4136 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4137 // If either of the constants are nans, then the whole thing returns
4139 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4140 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4142 // Otherwise, no need to compare the two constants, compare the
4144 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4145 RHS->getOperand(0));
4150 return Changed ? &I : 0;
4153 // XorSelf - Implements: X ^ X --> 0
4156 XorSelf(Value *rhs) : RHS(rhs) {}
4157 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4158 Instruction *apply(BinaryOperator &Xor) const {
4164 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4165 bool Changed = SimplifyCommutative(I);
4166 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4168 if (isa<UndefValue>(Op1))
4169 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4171 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4172 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4173 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4174 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4177 // See if we can simplify any instructions used by the instruction whose sole
4178 // purpose is to compute bits we don't care about.
4179 if (!isa<VectorType>(I.getType())) {
4180 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4181 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4182 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4183 KnownZero, KnownOne))
4185 } else if (isa<ConstantAggregateZero>(Op1)) {
4186 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4189 // Is this a ~ operation?
4190 if (Value *NotOp = dyn_castNotVal(&I)) {
4191 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4192 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4193 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4194 if (Op0I->getOpcode() == Instruction::And ||
4195 Op0I->getOpcode() == Instruction::Or) {
4196 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4197 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4199 BinaryOperator::createNot(Op0I->getOperand(1),
4200 Op0I->getOperand(1)->getName()+".not");
4201 InsertNewInstBefore(NotY, I);
4202 if (Op0I->getOpcode() == Instruction::And)
4203 return BinaryOperator::createOr(Op0NotVal, NotY);
4205 return BinaryOperator::createAnd(Op0NotVal, NotY);
4212 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4213 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4214 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4215 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4216 return new ICmpInst(ICI->getInversePredicate(),
4217 ICI->getOperand(0), ICI->getOperand(1));
4219 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4220 return new FCmpInst(FCI->getInversePredicate(),
4221 FCI->getOperand(0), FCI->getOperand(1));
4224 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4225 // ~(c-X) == X-c-1 == X+(-c-1)
4226 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4227 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4228 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4229 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4230 ConstantInt::get(I.getType(), 1));
4231 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4234 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4235 if (Op0I->getOpcode() == Instruction::Add) {
4236 // ~(X-c) --> (-c-1)-X
4237 if (RHS->isAllOnesValue()) {
4238 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4239 return BinaryOperator::createSub(
4240 ConstantExpr::getSub(NegOp0CI,
4241 ConstantInt::get(I.getType(), 1)),
4242 Op0I->getOperand(0));
4243 } else if (RHS->getValue().isSignBit()) {
4244 // (X + C) ^ signbit -> (X + C + signbit)
4245 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4246 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4249 } else if (Op0I->getOpcode() == Instruction::Or) {
4250 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4251 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4252 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4253 // Anything in both C1 and C2 is known to be zero, remove it from
4255 Constant *CommonBits = And(Op0CI, RHS);
4256 NewRHS = ConstantExpr::getAnd(NewRHS,
4257 ConstantExpr::getNot(CommonBits));
4258 AddToWorkList(Op0I);
4259 I.setOperand(0, Op0I->getOperand(0));
4260 I.setOperand(1, NewRHS);
4266 // Try to fold constant and into select arguments.
4267 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4268 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4270 if (isa<PHINode>(Op0))
4271 if (Instruction *NV = FoldOpIntoPhi(I))
4275 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4277 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4279 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4281 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4284 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4287 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4288 if (A == Op0) { // B^(B|A) == (A|B)^B
4289 Op1I->swapOperands();
4291 std::swap(Op0, Op1);
4292 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4293 I.swapOperands(); // Simplified below.
4294 std::swap(Op0, Op1);
4296 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4297 if (Op0 == A) // A^(A^B) == B
4298 return ReplaceInstUsesWith(I, B);
4299 else if (Op0 == B) // A^(B^A) == B
4300 return ReplaceInstUsesWith(I, A);
4301 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4302 if (A == Op0) { // A^(A&B) -> A^(B&A)
4303 Op1I->swapOperands();
4306 if (B == Op0) { // A^(B&A) -> (B&A)^A
4307 I.swapOperands(); // Simplified below.
4308 std::swap(Op0, Op1);
4313 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4316 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4317 if (A == Op1) // (B|A)^B == (A|B)^B
4319 if (B == Op1) { // (A|B)^B == A & ~B
4321 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4322 return BinaryOperator::createAnd(A, NotB);
4324 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4325 if (Op1 == A) // (A^B)^A == B
4326 return ReplaceInstUsesWith(I, B);
4327 else if (Op1 == B) // (B^A)^A == B
4328 return ReplaceInstUsesWith(I, A);
4329 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4330 if (A == Op1) // (A&B)^A -> (B&A)^A
4332 if (B == Op1 && // (B&A)^A == ~B & A
4333 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4335 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4336 return BinaryOperator::createAnd(N, Op1);
4341 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4342 if (Op0I && Op1I && Op0I->isShift() &&
4343 Op0I->getOpcode() == Op1I->getOpcode() &&
4344 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4345 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4346 Instruction *NewOp =
4347 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4348 Op1I->getOperand(0),
4349 Op0I->getName()), I);
4350 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4351 Op1I->getOperand(1));
4355 Value *A, *B, *C, *D;
4356 // (A & B)^(A | B) -> A ^ B
4357 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4358 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4359 if ((A == C && B == D) || (A == D && B == C))
4360 return BinaryOperator::createXor(A, B);
4362 // (A | B)^(A & B) -> A ^ B
4363 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4364 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4365 if ((A == C && B == D) || (A == D && B == C))
4366 return BinaryOperator::createXor(A, B);
4370 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4371 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4372 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4373 // (X & Y)^(X & Y) -> (Y^Z) & X
4374 Value *X = 0, *Y = 0, *Z = 0;
4376 X = A, Y = B, Z = D;
4378 X = A, Y = B, Z = C;
4380 X = B, Y = A, Z = D;
4382 X = B, Y = A, Z = C;
4385 Instruction *NewOp =
4386 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4387 return BinaryOperator::createAnd(NewOp, X);
4392 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4393 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4394 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4397 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4398 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4399 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4400 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4401 const Type *SrcTy = Op0C->getOperand(0)->getType();
4402 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4403 // Only do this if the casts both really cause code to be generated.
4404 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4406 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4408 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4409 Op1C->getOperand(0),
4411 InsertNewInstBefore(NewOp, I);
4412 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4416 return Changed ? &I : 0;
4419 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4420 /// overflowed for this type.
4421 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4422 ConstantInt *In2, bool IsSigned = false) {
4423 Result = cast<ConstantInt>(Add(In1, In2));
4426 if (In2->getValue().isNegative())
4427 return Result->getValue().sgt(In1->getValue());
4429 return Result->getValue().slt(In1->getValue());
4431 return Result->getValue().ult(In1->getValue());
4434 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4435 /// code necessary to compute the offset from the base pointer (without adding
4436 /// in the base pointer). Return the result as a signed integer of intptr size.
4437 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4438 TargetData &TD = IC.getTargetData();
4439 gep_type_iterator GTI = gep_type_begin(GEP);
4440 const Type *IntPtrTy = TD.getIntPtrType();
4441 Value *Result = Constant::getNullValue(IntPtrTy);
4443 // Build a mask for high order bits.
4444 unsigned IntPtrWidth = TD.getPointerSize()*8;
4445 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4447 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4448 Value *Op = GEP->getOperand(i);
4449 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4450 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4451 if (OpC->isZero()) continue;
4453 // Handle a struct index, which adds its field offset to the pointer.
4454 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4455 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4457 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4458 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4460 Result = IC.InsertNewInstBefore(
4461 BinaryOperator::createAdd(Result,
4462 ConstantInt::get(IntPtrTy, Size),
4463 GEP->getName()+".offs"), I);
4467 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4468 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4469 Scale = ConstantExpr::getMul(OC, Scale);
4470 if (Constant *RC = dyn_cast<Constant>(Result))
4471 Result = ConstantExpr::getAdd(RC, Scale);
4473 // Emit an add instruction.
4474 Result = IC.InsertNewInstBefore(
4475 BinaryOperator::createAdd(Result, Scale,
4476 GEP->getName()+".offs"), I);
4480 // Convert to correct type.
4481 if (Op->getType() != IntPtrTy) {
4482 if (Constant *OpC = dyn_cast<Constant>(Op))
4483 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4485 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4486 Op->getName()+".c"), I);
4489 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4490 if (Constant *OpC = dyn_cast<Constant>(Op))
4491 Op = ConstantExpr::getMul(OpC, Scale);
4492 else // We'll let instcombine(mul) convert this to a shl if possible.
4493 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4494 GEP->getName()+".idx"), I);
4497 // Emit an add instruction.
4498 if (isa<Constant>(Op) && isa<Constant>(Result))
4499 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4500 cast<Constant>(Result));
4502 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4503 GEP->getName()+".offs"), I);
4508 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4509 /// else. At this point we know that the GEP is on the LHS of the comparison.
4510 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4511 ICmpInst::Predicate Cond,
4513 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4515 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4516 if (isa<PointerType>(CI->getOperand(0)->getType()))
4517 RHS = CI->getOperand(0);
4519 Value *PtrBase = GEPLHS->getOperand(0);
4520 if (PtrBase == RHS) {
4521 // As an optimization, we don't actually have to compute the actual value of
4522 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4523 // each index is zero or not.
4524 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4525 Instruction *InVal = 0;
4526 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4527 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4529 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4530 if (isa<UndefValue>(C)) // undef index -> undef.
4531 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4532 if (C->isNullValue())
4534 else if (TD->getABITypeSize(GTI.getIndexedType()) == 0) {
4535 EmitIt = false; // This is indexing into a zero sized array?
4536 } else if (isa<ConstantInt>(C))
4537 return ReplaceInstUsesWith(I, // No comparison is needed here.
4538 ConstantInt::get(Type::Int1Ty,
4539 Cond == ICmpInst::ICMP_NE));
4544 new ICmpInst(Cond, GEPLHS->getOperand(i),
4545 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4549 InVal = InsertNewInstBefore(InVal, I);
4550 InsertNewInstBefore(Comp, I);
4551 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4552 InVal = BinaryOperator::createOr(InVal, Comp);
4553 else // True if all are equal
4554 InVal = BinaryOperator::createAnd(InVal, Comp);
4562 // No comparison is needed here, all indexes = 0
4563 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4564 Cond == ICmpInst::ICMP_EQ));
4567 // Only lower this if the icmp is the only user of the GEP or if we expect
4568 // the result to fold to a constant!
4569 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4570 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4571 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4572 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4573 Constant::getNullValue(Offset->getType()));
4575 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4576 // If the base pointers are different, but the indices are the same, just
4577 // compare the base pointer.
4578 if (PtrBase != GEPRHS->getOperand(0)) {
4579 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4580 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4581 GEPRHS->getOperand(0)->getType();
4583 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4584 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4585 IndicesTheSame = false;
4589 // If all indices are the same, just compare the base pointers.
4591 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4592 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4594 // Otherwise, the base pointers are different and the indices are
4595 // different, bail out.
4599 // If one of the GEPs has all zero indices, recurse.
4600 bool AllZeros = true;
4601 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4602 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4603 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4608 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4609 ICmpInst::getSwappedPredicate(Cond), I);
4611 // If the other GEP has all zero indices, recurse.
4613 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4614 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4615 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4620 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4622 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4623 // If the GEPs only differ by one index, compare it.
4624 unsigned NumDifferences = 0; // Keep track of # differences.
4625 unsigned DiffOperand = 0; // The operand that differs.
4626 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4627 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4628 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4629 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4630 // Irreconcilable differences.
4634 if (NumDifferences++) break;
4639 if (NumDifferences == 0) // SAME GEP?
4640 return ReplaceInstUsesWith(I, // No comparison is needed here.
4641 ConstantInt::get(Type::Int1Ty,
4642 isTrueWhenEqual(Cond)));
4644 else if (NumDifferences == 1) {
4645 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4646 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4647 // Make sure we do a signed comparison here.
4648 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4652 // Only lower this if the icmp is the only user of the GEP or if we expect
4653 // the result to fold to a constant!
4654 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4655 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4656 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4657 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4658 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4659 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4665 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4666 bool Changed = SimplifyCompare(I);
4667 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4669 // Fold trivial predicates.
4670 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4671 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4672 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4673 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4675 // Simplify 'fcmp pred X, X'
4677 switch (I.getPredicate()) {
4678 default: assert(0 && "Unknown predicate!");
4679 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4680 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4681 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4682 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4683 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4684 case FCmpInst::FCMP_OLT: // True if ordered and less than
4685 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4686 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4688 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4689 case FCmpInst::FCMP_ULT: // True if unordered or less than
4690 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4691 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4692 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4693 I.setPredicate(FCmpInst::FCMP_UNO);
4694 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4697 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4698 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4699 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4700 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4701 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4702 I.setPredicate(FCmpInst::FCMP_ORD);
4703 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4708 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4709 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4711 // Handle fcmp with constant RHS
4712 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4713 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4714 switch (LHSI->getOpcode()) {
4715 case Instruction::PHI:
4716 if (Instruction *NV = FoldOpIntoPhi(I))
4719 case Instruction::Select:
4720 // If either operand of the select is a constant, we can fold the
4721 // comparison into the select arms, which will cause one to be
4722 // constant folded and the select turned into a bitwise or.
4723 Value *Op1 = 0, *Op2 = 0;
4724 if (LHSI->hasOneUse()) {
4725 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4726 // Fold the known value into the constant operand.
4727 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4728 // Insert a new FCmp of the other select operand.
4729 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4730 LHSI->getOperand(2), RHSC,
4732 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4733 // Fold the known value into the constant operand.
4734 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4735 // Insert a new FCmp of the other select operand.
4736 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4737 LHSI->getOperand(1), RHSC,
4743 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4748 return Changed ? &I : 0;
4751 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4752 bool Changed = SimplifyCompare(I);
4753 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4754 const Type *Ty = Op0->getType();
4758 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4759 isTrueWhenEqual(I)));
4761 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4762 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4764 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4765 // addresses never equal each other! We already know that Op0 != Op1.
4766 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4767 isa<ConstantPointerNull>(Op0)) &&
4768 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4769 isa<ConstantPointerNull>(Op1)))
4770 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4771 !isTrueWhenEqual(I)));
4773 // icmp's with boolean values can always be turned into bitwise operations
4774 if (Ty == Type::Int1Ty) {
4775 switch (I.getPredicate()) {
4776 default: assert(0 && "Invalid icmp instruction!");
4777 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4778 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4779 InsertNewInstBefore(Xor, I);
4780 return BinaryOperator::createNot(Xor);
4782 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4783 return BinaryOperator::createXor(Op0, Op1);
4785 case ICmpInst::ICMP_UGT:
4786 case ICmpInst::ICMP_SGT:
4787 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4789 case ICmpInst::ICMP_ULT:
4790 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4791 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4792 InsertNewInstBefore(Not, I);
4793 return BinaryOperator::createAnd(Not, Op1);
4795 case ICmpInst::ICMP_UGE:
4796 case ICmpInst::ICMP_SGE:
4797 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4799 case ICmpInst::ICMP_ULE:
4800 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4801 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4802 InsertNewInstBefore(Not, I);
4803 return BinaryOperator::createOr(Not, Op1);
4808 // See if we are doing a comparison between a constant and an instruction that
4809 // can be folded into the comparison.
4810 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4811 switch (I.getPredicate()) {
4813 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4814 if (CI->isMinValue(false))
4815 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4816 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4817 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4818 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4819 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4820 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4821 if (CI->isMinValue(true))
4822 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4823 ConstantInt::getAllOnesValue(Op0->getType()));
4827 case ICmpInst::ICMP_SLT:
4828 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4829 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4830 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4831 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4832 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4833 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4836 case ICmpInst::ICMP_UGT:
4837 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4838 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4839 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4840 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4841 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4842 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4844 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4845 if (CI->isMaxValue(true))
4846 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4847 ConstantInt::getNullValue(Op0->getType()));
4850 case ICmpInst::ICMP_SGT:
4851 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4852 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4853 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4854 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4855 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4856 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4859 case ICmpInst::ICMP_ULE:
4860 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4861 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4862 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4863 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4864 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4865 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4868 case ICmpInst::ICMP_SLE:
4869 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4870 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4871 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4872 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4873 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4874 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4877 case ICmpInst::ICMP_UGE:
4878 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4879 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4880 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4881 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4882 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4883 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4886 case ICmpInst::ICMP_SGE:
4887 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4888 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4889 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4890 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4891 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4892 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4896 // If we still have a icmp le or icmp ge instruction, turn it into the
4897 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4898 // already been handled above, this requires little checking.
4900 switch (I.getPredicate()) {
4902 case ICmpInst::ICMP_ULE:
4903 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4904 case ICmpInst::ICMP_SLE:
4905 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4906 case ICmpInst::ICMP_UGE:
4907 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4908 case ICmpInst::ICMP_SGE:
4909 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4912 // See if we can fold the comparison based on bits known to be zero or one
4913 // in the input. If this comparison is a normal comparison, it demands all
4914 // bits, if it is a sign bit comparison, it only demands the sign bit.
4917 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4919 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4920 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4921 if (SimplifyDemandedBits(Op0,
4922 isSignBit ? APInt::getSignBit(BitWidth)
4923 : APInt::getAllOnesValue(BitWidth),
4924 KnownZero, KnownOne, 0))
4927 // Given the known and unknown bits, compute a range that the LHS could be
4929 if ((KnownOne | KnownZero) != 0) {
4930 // Compute the Min, Max and RHS values based on the known bits. For the
4931 // EQ and NE we use unsigned values.
4932 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4933 const APInt& RHSVal = CI->getValue();
4934 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4935 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4938 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4941 switch (I.getPredicate()) { // LE/GE have been folded already.
4942 default: assert(0 && "Unknown icmp opcode!");
4943 case ICmpInst::ICMP_EQ:
4944 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4945 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4947 case ICmpInst::ICMP_NE:
4948 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4949 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4951 case ICmpInst::ICMP_ULT:
4952 if (Max.ult(RHSVal))
4953 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4954 if (Min.uge(RHSVal))
4955 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4957 case ICmpInst::ICMP_UGT:
4958 if (Min.ugt(RHSVal))
4959 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4960 if (Max.ule(RHSVal))
4961 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4963 case ICmpInst::ICMP_SLT:
4964 if (Max.slt(RHSVal))
4965 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4966 if (Min.sgt(RHSVal))
4967 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4969 case ICmpInst::ICMP_SGT:
4970 if (Min.sgt(RHSVal))
4971 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4972 if (Max.sle(RHSVal))
4973 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4978 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4979 // instruction, see if that instruction also has constants so that the
4980 // instruction can be folded into the icmp
4981 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4982 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4986 // Handle icmp with constant (but not simple integer constant) RHS
4987 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4988 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4989 switch (LHSI->getOpcode()) {
4990 case Instruction::GetElementPtr:
4991 if (RHSC->isNullValue()) {
4992 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4993 bool isAllZeros = true;
4994 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4995 if (!isa<Constant>(LHSI->getOperand(i)) ||
4996 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5001 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5002 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5006 case Instruction::PHI:
5007 if (Instruction *NV = FoldOpIntoPhi(I))
5010 case Instruction::Select: {
5011 // If either operand of the select is a constant, we can fold the
5012 // comparison into the select arms, which will cause one to be
5013 // constant folded and the select turned into a bitwise or.
5014 Value *Op1 = 0, *Op2 = 0;
5015 if (LHSI->hasOneUse()) {
5016 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5017 // Fold the known value into the constant operand.
5018 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5019 // Insert a new ICmp of the other select operand.
5020 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5021 LHSI->getOperand(2), RHSC,
5023 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5024 // Fold the known value into the constant operand.
5025 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5026 // Insert a new ICmp of the other select operand.
5027 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5028 LHSI->getOperand(1), RHSC,
5034 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5037 case Instruction::Malloc:
5038 // If we have (malloc != null), and if the malloc has a single use, we
5039 // can assume it is successful and remove the malloc.
5040 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5041 AddToWorkList(LHSI);
5042 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5043 !isTrueWhenEqual(I)));
5049 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5050 if (User *GEP = dyn_castGetElementPtr(Op0))
5051 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5053 if (User *GEP = dyn_castGetElementPtr(Op1))
5054 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5055 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5058 // Test to see if the operands of the icmp are casted versions of other
5059 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5061 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5062 if (isa<PointerType>(Op0->getType()) &&
5063 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5064 // We keep moving the cast from the left operand over to the right
5065 // operand, where it can often be eliminated completely.
5066 Op0 = CI->getOperand(0);
5068 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5069 // so eliminate it as well.
5070 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5071 Op1 = CI2->getOperand(0);
5073 // If Op1 is a constant, we can fold the cast into the constant.
5074 if (Op0->getType() != Op1->getType())
5075 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5076 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5078 // Otherwise, cast the RHS right before the icmp
5079 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5081 return new ICmpInst(I.getPredicate(), Op0, Op1);
5085 if (isa<CastInst>(Op0)) {
5086 // Handle the special case of: icmp (cast bool to X), <cst>
5087 // This comes up when you have code like
5090 // For generality, we handle any zero-extension of any operand comparison
5091 // with a constant or another cast from the same type.
5092 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5093 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5097 if (I.isEquality()) {
5098 Value *A, *B, *C, *D;
5099 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5100 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5101 Value *OtherVal = A == Op1 ? B : A;
5102 return new ICmpInst(I.getPredicate(), OtherVal,
5103 Constant::getNullValue(A->getType()));
5106 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5107 // A^c1 == C^c2 --> A == C^(c1^c2)
5108 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5109 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5110 if (Op1->hasOneUse()) {
5111 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5112 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5113 return new ICmpInst(I.getPredicate(), A,
5114 InsertNewInstBefore(Xor, I));
5117 // A^B == A^D -> B == D
5118 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5119 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5120 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5121 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5125 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5126 (A == Op0 || B == Op0)) {
5127 // A == (A^B) -> B == 0
5128 Value *OtherVal = A == Op0 ? B : A;
5129 return new ICmpInst(I.getPredicate(), OtherVal,
5130 Constant::getNullValue(A->getType()));
5132 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5133 // (A-B) == A -> B == 0
5134 return new ICmpInst(I.getPredicate(), B,
5135 Constant::getNullValue(B->getType()));
5137 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5138 // A == (A-B) -> B == 0
5139 return new ICmpInst(I.getPredicate(), B,
5140 Constant::getNullValue(B->getType()));
5143 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5144 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5145 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5146 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5147 Value *X = 0, *Y = 0, *Z = 0;
5150 X = B; Y = D; Z = A;
5151 } else if (A == D) {
5152 X = B; Y = C; Z = A;
5153 } else if (B == C) {
5154 X = A; Y = D; Z = B;
5155 } else if (B == D) {
5156 X = A; Y = C; Z = B;
5159 if (X) { // Build (X^Y) & Z
5160 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5161 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5162 I.setOperand(0, Op1);
5163 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5168 return Changed ? &I : 0;
5172 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5173 /// and CmpRHS are both known to be integer constants.
5174 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5175 ConstantInt *DivRHS) {
5176 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5177 const APInt &CmpRHSV = CmpRHS->getValue();
5179 // FIXME: If the operand types don't match the type of the divide
5180 // then don't attempt this transform. The code below doesn't have the
5181 // logic to deal with a signed divide and an unsigned compare (and
5182 // vice versa). This is because (x /s C1) <s C2 produces different
5183 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5184 // (x /u C1) <u C2. Simply casting the operands and result won't
5185 // work. :( The if statement below tests that condition and bails
5187 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5188 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5190 if (DivRHS->isZero())
5191 return 0; // The ProdOV computation fails on divide by zero.
5193 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5194 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5195 // C2 (CI). By solving for X we can turn this into a range check
5196 // instead of computing a divide.
5197 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5199 // Determine if the product overflows by seeing if the product is
5200 // not equal to the divide. Make sure we do the same kind of divide
5201 // as in the LHS instruction that we're folding.
5202 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5203 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5205 // Get the ICmp opcode
5206 ICmpInst::Predicate Pred = ICI.getPredicate();
5208 // Figure out the interval that is being checked. For example, a comparison
5209 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5210 // Compute this interval based on the constants involved and the signedness of
5211 // the compare/divide. This computes a half-open interval, keeping track of
5212 // whether either value in the interval overflows. After analysis each
5213 // overflow variable is set to 0 if it's corresponding bound variable is valid
5214 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5215 int LoOverflow = 0, HiOverflow = 0;
5216 ConstantInt *LoBound = 0, *HiBound = 0;
5219 if (!DivIsSigned) { // udiv
5220 // e.g. X/5 op 3 --> [15, 20)
5222 HiOverflow = LoOverflow = ProdOV;
5224 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5225 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5226 if (CmpRHSV == 0) { // (X / pos) op 0
5227 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5228 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5230 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5231 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5232 HiOverflow = LoOverflow = ProdOV;
5234 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5235 } else { // (X / pos) op neg
5236 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5237 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5238 LoOverflow = AddWithOverflow(LoBound, Prod,
5239 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5240 HiBound = AddOne(Prod);
5241 HiOverflow = ProdOV ? -1 : 0;
5243 } else { // Divisor is < 0.
5244 if (CmpRHSV == 0) { // (X / neg) op 0
5245 // e.g. X/-5 op 0 --> [-4, 5)
5246 LoBound = AddOne(DivRHS);
5247 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5248 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5249 HiOverflow = 1; // [INTMIN+1, overflow)
5250 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5252 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5253 // e.g. X/-5 op 3 --> [-19, -14)
5254 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5256 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5257 HiBound = AddOne(Prod);
5258 } else { // (X / neg) op neg
5259 // e.g. X/-5 op -3 --> [15, 20)
5261 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5262 HiBound = Subtract(Prod, DivRHS);
5265 // Dividing by a negative swaps the condition. LT <-> GT
5266 Pred = ICmpInst::getSwappedPredicate(Pred);
5269 Value *X = DivI->getOperand(0);
5271 default: assert(0 && "Unhandled icmp opcode!");
5272 case ICmpInst::ICMP_EQ:
5273 if (LoOverflow && HiOverflow)
5274 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5275 else if (HiOverflow)
5276 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5277 ICmpInst::ICMP_UGE, X, LoBound);
5278 else if (LoOverflow)
5279 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5280 ICmpInst::ICMP_ULT, X, HiBound);
5282 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5283 case ICmpInst::ICMP_NE:
5284 if (LoOverflow && HiOverflow)
5285 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5286 else if (HiOverflow)
5287 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5288 ICmpInst::ICMP_ULT, X, LoBound);
5289 else if (LoOverflow)
5290 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5291 ICmpInst::ICMP_UGE, X, HiBound);
5293 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5294 case ICmpInst::ICMP_ULT:
5295 case ICmpInst::ICMP_SLT:
5296 if (LoOverflow == +1) // Low bound is greater than input range.
5297 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5298 if (LoOverflow == -1) // Low bound is less than input range.
5299 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5300 return new ICmpInst(Pred, X, LoBound);
5301 case ICmpInst::ICMP_UGT:
5302 case ICmpInst::ICMP_SGT:
5303 if (HiOverflow == +1) // High bound greater than input range.
5304 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5305 else if (HiOverflow == -1) // High bound less than input range.
5306 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5307 if (Pred == ICmpInst::ICMP_UGT)
5308 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5310 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5315 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5317 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5320 const APInt &RHSV = RHS->getValue();
5322 switch (LHSI->getOpcode()) {
5323 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5324 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5325 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5327 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5328 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5329 Value *CompareVal = LHSI->getOperand(0);
5331 // If the sign bit of the XorCST is not set, there is no change to
5332 // the operation, just stop using the Xor.
5333 if (!XorCST->getValue().isNegative()) {
5334 ICI.setOperand(0, CompareVal);
5335 AddToWorkList(LHSI);
5339 // Was the old condition true if the operand is positive?
5340 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5342 // If so, the new one isn't.
5343 isTrueIfPositive ^= true;
5345 if (isTrueIfPositive)
5346 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5348 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5352 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5353 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5354 LHSI->getOperand(0)->hasOneUse()) {
5355 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5357 // If the LHS is an AND of a truncating cast, we can widen the
5358 // and/compare to be the input width without changing the value
5359 // produced, eliminating a cast.
5360 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5361 // We can do this transformation if either the AND constant does not
5362 // have its sign bit set or if it is an equality comparison.
5363 // Extending a relational comparison when we're checking the sign
5364 // bit would not work.
5365 if (Cast->hasOneUse() &&
5366 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5367 RHSV.isPositive())) {
5369 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5370 APInt NewCST = AndCST->getValue();
5371 NewCST.zext(BitWidth);
5373 NewCI.zext(BitWidth);
5374 Instruction *NewAnd =
5375 BinaryOperator::createAnd(Cast->getOperand(0),
5376 ConstantInt::get(NewCST),LHSI->getName());
5377 InsertNewInstBefore(NewAnd, ICI);
5378 return new ICmpInst(ICI.getPredicate(), NewAnd,
5379 ConstantInt::get(NewCI));
5383 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5384 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5385 // happens a LOT in code produced by the C front-end, for bitfield
5387 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5388 if (Shift && !Shift->isShift())
5392 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5393 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5394 const Type *AndTy = AndCST->getType(); // Type of the and.
5396 // We can fold this as long as we can't shift unknown bits
5397 // into the mask. This can only happen with signed shift
5398 // rights, as they sign-extend.
5400 bool CanFold = Shift->isLogicalShift();
5402 // To test for the bad case of the signed shr, see if any
5403 // of the bits shifted in could be tested after the mask.
5404 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5405 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5407 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5408 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5409 AndCST->getValue()) == 0)
5415 if (Shift->getOpcode() == Instruction::Shl)
5416 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5418 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5420 // Check to see if we are shifting out any of the bits being
5422 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5423 // If we shifted bits out, the fold is not going to work out.
5424 // As a special case, check to see if this means that the
5425 // result is always true or false now.
5426 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5427 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5428 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5429 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5431 ICI.setOperand(1, NewCst);
5432 Constant *NewAndCST;
5433 if (Shift->getOpcode() == Instruction::Shl)
5434 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5436 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5437 LHSI->setOperand(1, NewAndCST);
5438 LHSI->setOperand(0, Shift->getOperand(0));
5439 AddToWorkList(Shift); // Shift is dead.
5440 AddUsesToWorkList(ICI);
5446 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5447 // preferable because it allows the C<<Y expression to be hoisted out
5448 // of a loop if Y is invariant and X is not.
5449 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5450 ICI.isEquality() && !Shift->isArithmeticShift() &&
5451 isa<Instruction>(Shift->getOperand(0))) {
5454 if (Shift->getOpcode() == Instruction::LShr) {
5455 NS = BinaryOperator::createShl(AndCST,
5456 Shift->getOperand(1), "tmp");
5458 // Insert a logical shift.
5459 NS = BinaryOperator::createLShr(AndCST,
5460 Shift->getOperand(1), "tmp");
5462 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5464 // Compute X & (C << Y).
5465 Instruction *NewAnd =
5466 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5467 InsertNewInstBefore(NewAnd, ICI);
5469 ICI.setOperand(0, NewAnd);
5475 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5476 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5479 uint32_t TypeBits = RHSV.getBitWidth();
5481 // Check that the shift amount is in range. If not, don't perform
5482 // undefined shifts. When the shift is visited it will be
5484 if (ShAmt->uge(TypeBits))
5487 if (ICI.isEquality()) {
5488 // If we are comparing against bits always shifted out, the
5489 // comparison cannot succeed.
5491 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5492 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5493 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5494 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5495 return ReplaceInstUsesWith(ICI, Cst);
5498 if (LHSI->hasOneUse()) {
5499 // Otherwise strength reduce the shift into an and.
5500 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5502 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5505 BinaryOperator::createAnd(LHSI->getOperand(0),
5506 Mask, LHSI->getName()+".mask");
5507 Value *And = InsertNewInstBefore(AndI, ICI);
5508 return new ICmpInst(ICI.getPredicate(), And,
5509 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5513 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5514 bool TrueIfSigned = false;
5515 if (LHSI->hasOneUse() &&
5516 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5517 // (X << 31) <s 0 --> (X&1) != 0
5518 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5519 (TypeBits-ShAmt->getZExtValue()-1));
5521 BinaryOperator::createAnd(LHSI->getOperand(0),
5522 Mask, LHSI->getName()+".mask");
5523 Value *And = InsertNewInstBefore(AndI, ICI);
5525 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5526 And, Constant::getNullValue(And->getType()));
5531 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5532 case Instruction::AShr: {
5533 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5536 if (ICI.isEquality()) {
5537 // Check that the shift amount is in range. If not, don't perform
5538 // undefined shifts. When the shift is visited it will be
5540 uint32_t TypeBits = RHSV.getBitWidth();
5541 if (ShAmt->uge(TypeBits))
5543 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5545 // If we are comparing against bits always shifted out, the
5546 // comparison cannot succeed.
5547 APInt Comp = RHSV << ShAmtVal;
5548 if (LHSI->getOpcode() == Instruction::LShr)
5549 Comp = Comp.lshr(ShAmtVal);
5551 Comp = Comp.ashr(ShAmtVal);
5553 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5554 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5555 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5556 return ReplaceInstUsesWith(ICI, Cst);
5559 if (LHSI->hasOneUse() || RHSV == 0) {
5560 // Otherwise strength reduce the shift into an and.
5561 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5562 Constant *Mask = ConstantInt::get(Val);
5565 BinaryOperator::createAnd(LHSI->getOperand(0),
5566 Mask, LHSI->getName()+".mask");
5567 Value *And = InsertNewInstBefore(AndI, ICI);
5568 return new ICmpInst(ICI.getPredicate(), And,
5569 ConstantExpr::getShl(RHS, ShAmt));
5575 case Instruction::SDiv:
5576 case Instruction::UDiv:
5577 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5578 // Fold this div into the comparison, producing a range check.
5579 // Determine, based on the divide type, what the range is being
5580 // checked. If there is an overflow on the low or high side, remember
5581 // it, otherwise compute the range [low, hi) bounding the new value.
5582 // See: InsertRangeTest above for the kinds of replacements possible.
5583 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5584 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5590 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5591 if (ICI.isEquality()) {
5592 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5594 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5595 // the second operand is a constant, simplify a bit.
5596 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5597 switch (BO->getOpcode()) {
5598 case Instruction::SRem:
5599 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5600 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5601 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5602 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5603 Instruction *NewRem =
5604 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5606 InsertNewInstBefore(NewRem, ICI);
5607 return new ICmpInst(ICI.getPredicate(), NewRem,
5608 Constant::getNullValue(BO->getType()));
5612 case Instruction::Add:
5613 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5614 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5615 if (BO->hasOneUse())
5616 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5617 Subtract(RHS, BOp1C));
5618 } else if (RHSV == 0) {
5619 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5620 // efficiently invertible, or if the add has just this one use.
5621 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5623 if (Value *NegVal = dyn_castNegVal(BOp1))
5624 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5625 else if (Value *NegVal = dyn_castNegVal(BOp0))
5626 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5627 else if (BO->hasOneUse()) {
5628 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5629 InsertNewInstBefore(Neg, ICI);
5631 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5635 case Instruction::Xor:
5636 // For the xor case, we can xor two constants together, eliminating
5637 // the explicit xor.
5638 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5639 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5640 ConstantExpr::getXor(RHS, BOC));
5643 case Instruction::Sub:
5644 // Replace (([sub|xor] A, B) != 0) with (A != B)
5646 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5650 case Instruction::Or:
5651 // If bits are being or'd in that are not present in the constant we
5652 // are comparing against, then the comparison could never succeed!
5653 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5654 Constant *NotCI = ConstantExpr::getNot(RHS);
5655 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5656 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5661 case Instruction::And:
5662 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5663 // If bits are being compared against that are and'd out, then the
5664 // comparison can never succeed!
5665 if ((RHSV & ~BOC->getValue()) != 0)
5666 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5669 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5670 if (RHS == BOC && RHSV.isPowerOf2())
5671 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5672 ICmpInst::ICMP_NE, LHSI,
5673 Constant::getNullValue(RHS->getType()));
5675 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5676 if (isSignBit(BOC)) {
5677 Value *X = BO->getOperand(0);
5678 Constant *Zero = Constant::getNullValue(X->getType());
5679 ICmpInst::Predicate pred = isICMP_NE ?
5680 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5681 return new ICmpInst(pred, X, Zero);
5684 // ((X & ~7) == 0) --> X < 8
5685 if (RHSV == 0 && isHighOnes(BOC)) {
5686 Value *X = BO->getOperand(0);
5687 Constant *NegX = ConstantExpr::getNeg(BOC);
5688 ICmpInst::Predicate pred = isICMP_NE ?
5689 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5690 return new ICmpInst(pred, X, NegX);
5695 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5696 // Handle icmp {eq|ne} <intrinsic>, intcst.
5697 if (II->getIntrinsicID() == Intrinsic::bswap) {
5699 ICI.setOperand(0, II->getOperand(1));
5700 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5704 } else { // Not a ICMP_EQ/ICMP_NE
5705 // If the LHS is a cast from an integral value of the same size,
5706 // then since we know the RHS is a constant, try to simlify.
5707 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5708 Value *CastOp = Cast->getOperand(0);
5709 const Type *SrcTy = CastOp->getType();
5710 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5711 if (SrcTy->isInteger() &&
5712 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5713 // If this is an unsigned comparison, try to make the comparison use
5714 // smaller constant values.
5715 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5716 // X u< 128 => X s> -1
5717 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5718 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5719 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5720 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5721 // X u> 127 => X s< 0
5722 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5723 Constant::getNullValue(SrcTy));
5731 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5732 /// We only handle extending casts so far.
5734 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5735 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5736 Value *LHSCIOp = LHSCI->getOperand(0);
5737 const Type *SrcTy = LHSCIOp->getType();
5738 const Type *DestTy = LHSCI->getType();
5741 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5742 // integer type is the same size as the pointer type.
5743 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5744 getTargetData().getPointerSizeInBits() ==
5745 cast<IntegerType>(DestTy)->getBitWidth()) {
5747 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5748 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5749 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5750 RHSOp = RHSC->getOperand(0);
5751 // If the pointer types don't match, insert a bitcast.
5752 if (LHSCIOp->getType() != RHSOp->getType())
5753 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5754 LHSCIOp->getType(), ICI);
5758 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5761 // The code below only handles extension cast instructions, so far.
5763 if (LHSCI->getOpcode() != Instruction::ZExt &&
5764 LHSCI->getOpcode() != Instruction::SExt)
5767 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5768 bool isSignedCmp = ICI.isSignedPredicate();
5770 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5771 // Not an extension from the same type?
5772 RHSCIOp = CI->getOperand(0);
5773 if (RHSCIOp->getType() != LHSCIOp->getType())
5776 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5777 // and the other is a zext), then we can't handle this.
5778 if (CI->getOpcode() != LHSCI->getOpcode())
5781 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5782 // then we can't handle this.
5783 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5786 // Okay, just insert a compare of the reduced operands now!
5787 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5790 // If we aren't dealing with a constant on the RHS, exit early
5791 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5795 // Compute the constant that would happen if we truncated to SrcTy then
5796 // reextended to DestTy.
5797 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5798 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5800 // If the re-extended constant didn't change...
5802 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5803 // For example, we might have:
5804 // %A = sext short %X to uint
5805 // %B = icmp ugt uint %A, 1330
5806 // It is incorrect to transform this into
5807 // %B = icmp ugt short %X, 1330
5808 // because %A may have negative value.
5810 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5811 // OR operation is EQ/NE.
5812 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5813 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5818 // The re-extended constant changed so the constant cannot be represented
5819 // in the shorter type. Consequently, we cannot emit a simple comparison.
5821 // First, handle some easy cases. We know the result cannot be equal at this
5822 // point so handle the ICI.isEquality() cases
5823 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5824 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5825 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5826 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5828 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5829 // should have been folded away previously and not enter in here.
5832 // We're performing a signed comparison.
5833 if (cast<ConstantInt>(CI)->getValue().isNegative())
5834 Result = ConstantInt::getFalse(); // X < (small) --> false
5836 Result = ConstantInt::getTrue(); // X < (large) --> true
5838 // We're performing an unsigned comparison.
5840 // We're performing an unsigned comp with a sign extended value.
5841 // This is true if the input is >= 0. [aka >s -1]
5842 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5843 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5844 NegOne, ICI.getName()), ICI);
5846 // Unsigned extend & unsigned compare -> always true.
5847 Result = ConstantInt::getTrue();
5851 // Finally, return the value computed.
5852 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5853 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5854 return ReplaceInstUsesWith(ICI, Result);
5856 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5857 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5858 "ICmp should be folded!");
5859 if (Constant *CI = dyn_cast<Constant>(Result))
5860 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5862 return BinaryOperator::createNot(Result);
5866 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5867 return commonShiftTransforms(I);
5870 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5871 return commonShiftTransforms(I);
5874 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5875 return commonShiftTransforms(I);
5878 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5879 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5880 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5882 // shl X, 0 == X and shr X, 0 == X
5883 // shl 0, X == 0 and shr 0, X == 0
5884 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5885 Op0 == Constant::getNullValue(Op0->getType()))
5886 return ReplaceInstUsesWith(I, Op0);
5888 if (isa<UndefValue>(Op0)) {
5889 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5890 return ReplaceInstUsesWith(I, Op0);
5891 else // undef << X -> 0, undef >>u X -> 0
5892 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5894 if (isa<UndefValue>(Op1)) {
5895 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5896 return ReplaceInstUsesWith(I, Op0);
5897 else // X << undef, X >>u undef -> 0
5898 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5901 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5902 if (I.getOpcode() == Instruction::AShr)
5903 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5904 if (CSI->isAllOnesValue())
5905 return ReplaceInstUsesWith(I, CSI);
5907 // Try to fold constant and into select arguments.
5908 if (isa<Constant>(Op0))
5909 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5910 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5913 // See if we can turn a signed shr into an unsigned shr.
5914 if (I.isArithmeticShift()) {
5915 if (MaskedValueIsZero(Op0,
5916 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5917 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5921 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5922 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5927 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5928 BinaryOperator &I) {
5929 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5931 // See if we can simplify any instructions used by the instruction whose sole
5932 // purpose is to compute bits we don't care about.
5933 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5934 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5935 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5936 KnownZero, KnownOne))
5939 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5940 // of a signed value.
5942 if (Op1->uge(TypeBits)) {
5943 if (I.getOpcode() != Instruction::AShr)
5944 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5946 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5951 // ((X*C1) << C2) == (X * (C1 << C2))
5952 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5953 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5954 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5955 return BinaryOperator::createMul(BO->getOperand(0),
5956 ConstantExpr::getShl(BOOp, Op1));
5958 // Try to fold constant and into select arguments.
5959 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5960 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5962 if (isa<PHINode>(Op0))
5963 if (Instruction *NV = FoldOpIntoPhi(I))
5966 if (Op0->hasOneUse()) {
5967 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5968 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5971 switch (Op0BO->getOpcode()) {
5973 case Instruction::Add:
5974 case Instruction::And:
5975 case Instruction::Or:
5976 case Instruction::Xor: {
5977 // These operators commute.
5978 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5979 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5980 match(Op0BO->getOperand(1),
5981 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5982 Instruction *YS = BinaryOperator::createShl(
5983 Op0BO->getOperand(0), Op1,
5985 InsertNewInstBefore(YS, I); // (Y << C)
5987 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5988 Op0BO->getOperand(1)->getName());
5989 InsertNewInstBefore(X, I); // (X + (Y << C))
5990 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5991 return BinaryOperator::createAnd(X, ConstantInt::get(
5992 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5995 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5996 Value *Op0BOOp1 = Op0BO->getOperand(1);
5997 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5999 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6000 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6002 Instruction *YS = BinaryOperator::createShl(
6003 Op0BO->getOperand(0), Op1,
6005 InsertNewInstBefore(YS, I); // (Y << C)
6007 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6008 V1->getName()+".mask");
6009 InsertNewInstBefore(XM, I); // X & (CC << C)
6011 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6016 case Instruction::Sub: {
6017 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6018 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6019 match(Op0BO->getOperand(0),
6020 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6021 Instruction *YS = BinaryOperator::createShl(
6022 Op0BO->getOperand(1), Op1,
6024 InsertNewInstBefore(YS, I); // (Y << C)
6026 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6027 Op0BO->getOperand(0)->getName());
6028 InsertNewInstBefore(X, I); // (X + (Y << C))
6029 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6030 return BinaryOperator::createAnd(X, ConstantInt::get(
6031 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6034 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6035 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6036 match(Op0BO->getOperand(0),
6037 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6038 m_ConstantInt(CC))) && V2 == Op1 &&
6039 cast<BinaryOperator>(Op0BO->getOperand(0))
6040 ->getOperand(0)->hasOneUse()) {
6041 Instruction *YS = BinaryOperator::createShl(
6042 Op0BO->getOperand(1), Op1,
6044 InsertNewInstBefore(YS, I); // (Y << C)
6046 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6047 V1->getName()+".mask");
6048 InsertNewInstBefore(XM, I); // X & (CC << C)
6050 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6058 // If the operand is an bitwise operator with a constant RHS, and the
6059 // shift is the only use, we can pull it out of the shift.
6060 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6061 bool isValid = true; // Valid only for And, Or, Xor
6062 bool highBitSet = false; // Transform if high bit of constant set?
6064 switch (Op0BO->getOpcode()) {
6065 default: isValid = false; break; // Do not perform transform!
6066 case Instruction::Add:
6067 isValid = isLeftShift;
6069 case Instruction::Or:
6070 case Instruction::Xor:
6073 case Instruction::And:
6078 // If this is a signed shift right, and the high bit is modified
6079 // by the logical operation, do not perform the transformation.
6080 // The highBitSet boolean indicates the value of the high bit of
6081 // the constant which would cause it to be modified for this
6084 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6085 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6089 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6091 Instruction *NewShift =
6092 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6093 InsertNewInstBefore(NewShift, I);
6094 NewShift->takeName(Op0BO);
6096 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6103 // Find out if this is a shift of a shift by a constant.
6104 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6105 if (ShiftOp && !ShiftOp->isShift())
6108 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6109 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6110 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6111 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6112 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6113 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6114 Value *X = ShiftOp->getOperand(0);
6116 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6117 if (AmtSum > TypeBits)
6120 const IntegerType *Ty = cast<IntegerType>(I.getType());
6122 // Check for (X << c1) << c2 and (X >> c1) >> c2
6123 if (I.getOpcode() == ShiftOp->getOpcode()) {
6124 return BinaryOperator::create(I.getOpcode(), X,
6125 ConstantInt::get(Ty, AmtSum));
6126 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6127 I.getOpcode() == Instruction::AShr) {
6128 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6129 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6130 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6131 I.getOpcode() == Instruction::LShr) {
6132 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6133 Instruction *Shift =
6134 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6135 InsertNewInstBefore(Shift, I);
6137 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6138 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6141 // Okay, if we get here, one shift must be left, and the other shift must be
6142 // right. See if the amounts are equal.
6143 if (ShiftAmt1 == ShiftAmt2) {
6144 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6145 if (I.getOpcode() == Instruction::Shl) {
6146 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6147 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6149 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6150 if (I.getOpcode() == Instruction::LShr) {
6151 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6152 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6154 // We can simplify ((X << C) >>s C) into a trunc + sext.
6155 // NOTE: we could do this for any C, but that would make 'unusual' integer
6156 // types. For now, just stick to ones well-supported by the code
6158 const Type *SExtType = 0;
6159 switch (Ty->getBitWidth() - ShiftAmt1) {
6166 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6171 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6172 InsertNewInstBefore(NewTrunc, I);
6173 return new SExtInst(NewTrunc, Ty);
6175 // Otherwise, we can't handle it yet.
6176 } else if (ShiftAmt1 < ShiftAmt2) {
6177 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6179 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6180 if (I.getOpcode() == Instruction::Shl) {
6181 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6182 ShiftOp->getOpcode() == Instruction::AShr);
6183 Instruction *Shift =
6184 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6185 InsertNewInstBefore(Shift, I);
6187 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6188 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6191 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6192 if (I.getOpcode() == Instruction::LShr) {
6193 assert(ShiftOp->getOpcode() == Instruction::Shl);
6194 Instruction *Shift =
6195 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6196 InsertNewInstBefore(Shift, I);
6198 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6199 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6202 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6204 assert(ShiftAmt2 < ShiftAmt1);
6205 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6207 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6208 if (I.getOpcode() == Instruction::Shl) {
6209 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6210 ShiftOp->getOpcode() == Instruction::AShr);
6211 Instruction *Shift =
6212 BinaryOperator::create(ShiftOp->getOpcode(), X,
6213 ConstantInt::get(Ty, ShiftDiff));
6214 InsertNewInstBefore(Shift, I);
6216 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6217 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6220 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6221 if (I.getOpcode() == Instruction::LShr) {
6222 assert(ShiftOp->getOpcode() == Instruction::Shl);
6223 Instruction *Shift =
6224 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6225 InsertNewInstBefore(Shift, I);
6227 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6228 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6231 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6238 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6239 /// expression. If so, decompose it, returning some value X, such that Val is
6242 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6244 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6245 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6246 Offset = CI->getZExtValue();
6248 return ConstantInt::get(Type::Int32Ty, 0);
6249 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6250 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6251 if (I->getOpcode() == Instruction::Shl) {
6252 // This is a value scaled by '1 << the shift amt'.
6253 Scale = 1U << RHS->getZExtValue();
6255 return I->getOperand(0);
6256 } else if (I->getOpcode() == Instruction::Mul) {
6257 // This value is scaled by 'RHS'.
6258 Scale = RHS->getZExtValue();
6260 return I->getOperand(0);
6261 } else if (I->getOpcode() == Instruction::Add) {
6262 // We have X+C. Check to see if we really have (X*C2)+C1,
6263 // where C1 is divisible by C2.
6266 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6267 Offset += RHS->getZExtValue();
6274 // Otherwise, we can't look past this.
6281 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6282 /// try to eliminate the cast by moving the type information into the alloc.
6283 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6284 AllocationInst &AI) {
6285 const PointerType *PTy = cast<PointerType>(CI.getType());
6287 // Remove any uses of AI that are dead.
6288 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6290 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6291 Instruction *User = cast<Instruction>(*UI++);
6292 if (isInstructionTriviallyDead(User)) {
6293 while (UI != E && *UI == User)
6294 ++UI; // If this instruction uses AI more than once, don't break UI.
6297 DOUT << "IC: DCE: " << *User;
6298 EraseInstFromFunction(*User);
6302 // Get the type really allocated and the type casted to.
6303 const Type *AllocElTy = AI.getAllocatedType();
6304 const Type *CastElTy = PTy->getElementType();
6305 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6307 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6308 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6309 if (CastElTyAlign < AllocElTyAlign) return 0;
6311 // If the allocation has multiple uses, only promote it if we are strictly
6312 // increasing the alignment of the resultant allocation. If we keep it the
6313 // same, we open the door to infinite loops of various kinds.
6314 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6316 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6317 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6318 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6320 // See if we can satisfy the modulus by pulling a scale out of the array
6322 unsigned ArraySizeScale;
6324 Value *NumElements = // See if the array size is a decomposable linear expr.
6325 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6327 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6329 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6330 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6332 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6337 // If the allocation size is constant, form a constant mul expression
6338 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6339 if (isa<ConstantInt>(NumElements))
6340 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6341 // otherwise multiply the amount and the number of elements
6342 else if (Scale != 1) {
6343 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6344 Amt = InsertNewInstBefore(Tmp, AI);
6348 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6349 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6350 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6351 Amt = InsertNewInstBefore(Tmp, AI);
6354 AllocationInst *New;
6355 if (isa<MallocInst>(AI))
6356 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6358 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6359 InsertNewInstBefore(New, AI);
6362 // If the allocation has multiple uses, insert a cast and change all things
6363 // that used it to use the new cast. This will also hack on CI, but it will
6365 if (!AI.hasOneUse()) {
6366 AddUsesToWorkList(AI);
6367 // New is the allocation instruction, pointer typed. AI is the original
6368 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6369 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6370 InsertNewInstBefore(NewCast, AI);
6371 AI.replaceAllUsesWith(NewCast);
6373 return ReplaceInstUsesWith(CI, New);
6376 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6377 /// and return it as type Ty without inserting any new casts and without
6378 /// changing the computed value. This is used by code that tries to decide
6379 /// whether promoting or shrinking integer operations to wider or smaller types
6380 /// will allow us to eliminate a truncate or extend.
6382 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6383 /// extension operation if Ty is larger.
6384 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6385 unsigned CastOpc, int &NumCastsRemoved) {
6386 // We can always evaluate constants in another type.
6387 if (isa<ConstantInt>(V))
6390 Instruction *I = dyn_cast<Instruction>(V);
6391 if (!I) return false;
6393 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6395 // If this is an extension or truncate, we can often eliminate it.
6396 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6397 // If this is a cast from the destination type, we can trivially eliminate
6398 // it, and this will remove a cast overall.
6399 if (I->getOperand(0)->getType() == Ty) {
6400 // If the first operand is itself a cast, and is eliminable, do not count
6401 // this as an eliminable cast. We would prefer to eliminate those two
6403 if (!isa<CastInst>(I->getOperand(0)))
6409 // We can't extend or shrink something that has multiple uses: doing so would
6410 // require duplicating the instruction in general, which isn't profitable.
6411 if (!I->hasOneUse()) return false;
6413 switch (I->getOpcode()) {
6414 case Instruction::Add:
6415 case Instruction::Sub:
6416 case Instruction::And:
6417 case Instruction::Or:
6418 case Instruction::Xor:
6419 // These operators can all arbitrarily be extended or truncated.
6420 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6422 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6425 case Instruction::Shl:
6426 // If we are truncating the result of this SHL, and if it's a shift of a
6427 // constant amount, we can always perform a SHL in a smaller type.
6428 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6429 uint32_t BitWidth = Ty->getBitWidth();
6430 if (BitWidth < OrigTy->getBitWidth() &&
6431 CI->getLimitedValue(BitWidth) < BitWidth)
6432 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6436 case Instruction::LShr:
6437 // If this is a truncate of a logical shr, we can truncate it to a smaller
6438 // lshr iff we know that the bits we would otherwise be shifting in are
6440 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6441 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6442 uint32_t BitWidth = Ty->getBitWidth();
6443 if (BitWidth < OrigBitWidth &&
6444 MaskedValueIsZero(I->getOperand(0),
6445 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6446 CI->getLimitedValue(BitWidth) < BitWidth) {
6447 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6452 case Instruction::ZExt:
6453 case Instruction::SExt:
6454 case Instruction::Trunc:
6455 // If this is the same kind of case as our original (e.g. zext+zext), we
6456 // can safely replace it. Note that replacing it does not reduce the number
6457 // of casts in the input.
6458 if (I->getOpcode() == CastOpc)
6463 // TODO: Can handle more cases here.
6470 /// EvaluateInDifferentType - Given an expression that
6471 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6472 /// evaluate the expression.
6473 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6475 if (Constant *C = dyn_cast<Constant>(V))
6476 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6478 // Otherwise, it must be an instruction.
6479 Instruction *I = cast<Instruction>(V);
6480 Instruction *Res = 0;
6481 switch (I->getOpcode()) {
6482 case Instruction::Add:
6483 case Instruction::Sub:
6484 case Instruction::And:
6485 case Instruction::Or:
6486 case Instruction::Xor:
6487 case Instruction::AShr:
6488 case Instruction::LShr:
6489 case Instruction::Shl: {
6490 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6491 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6492 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6493 LHS, RHS, I->getName());
6496 case Instruction::Trunc:
6497 case Instruction::ZExt:
6498 case Instruction::SExt:
6499 // If the source type of the cast is the type we're trying for then we can
6500 // just return the source. There's no need to insert it because it is not
6502 if (I->getOperand(0)->getType() == Ty)
6503 return I->getOperand(0);
6505 // Otherwise, must be the same type of case, so just reinsert a new one.
6506 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6510 // TODO: Can handle more cases here.
6511 assert(0 && "Unreachable!");
6515 return InsertNewInstBefore(Res, *I);
6518 /// @brief Implement the transforms common to all CastInst visitors.
6519 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6520 Value *Src = CI.getOperand(0);
6522 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6523 // eliminate it now.
6524 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6525 if (Instruction::CastOps opc =
6526 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6527 // The first cast (CSrc) is eliminable so we need to fix up or replace
6528 // the second cast (CI). CSrc will then have a good chance of being dead.
6529 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6533 // If we are casting a select then fold the cast into the select
6534 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6535 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6538 // If we are casting a PHI then fold the cast into the PHI
6539 if (isa<PHINode>(Src))
6540 if (Instruction *NV = FoldOpIntoPhi(CI))
6546 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6547 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6548 Value *Src = CI.getOperand(0);
6550 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6551 // If casting the result of a getelementptr instruction with no offset, turn
6552 // this into a cast of the original pointer!
6553 if (GEP->hasAllZeroIndices()) {
6554 // Changing the cast operand is usually not a good idea but it is safe
6555 // here because the pointer operand is being replaced with another
6556 // pointer operand so the opcode doesn't need to change.
6558 CI.setOperand(0, GEP->getOperand(0));
6562 // If the GEP has a single use, and the base pointer is a bitcast, and the
6563 // GEP computes a constant offset, see if we can convert these three
6564 // instructions into fewer. This typically happens with unions and other
6565 // non-type-safe code.
6566 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6567 if (GEP->hasAllConstantIndices()) {
6568 // We are guaranteed to get a constant from EmitGEPOffset.
6569 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6570 int64_t Offset = OffsetV->getSExtValue();
6572 // Get the base pointer input of the bitcast, and the type it points to.
6573 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6574 const Type *GEPIdxTy =
6575 cast<PointerType>(OrigBase->getType())->getElementType();
6576 if (GEPIdxTy->isSized()) {
6577 SmallVector<Value*, 8> NewIndices;
6579 // Start with the index over the outer type. Note that the type size
6580 // might be zero (even if the offset isn't zero) if the indexed type
6581 // is something like [0 x {int, int}]
6582 const Type *IntPtrTy = TD->getIntPtrType();
6583 int64_t FirstIdx = 0;
6584 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6585 FirstIdx = Offset/TySize;
6588 // Handle silly modulus not returning values values [0..TySize).
6592 assert(Offset >= 0);
6594 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6597 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6599 // Index into the types. If we fail, set OrigBase to null.
6601 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6602 const StructLayout *SL = TD->getStructLayout(STy);
6603 if (Offset < (int64_t)SL->getSizeInBytes()) {
6604 unsigned Elt = SL->getElementContainingOffset(Offset);
6605 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6607 Offset -= SL->getElementOffset(Elt);
6608 GEPIdxTy = STy->getElementType(Elt);
6610 // Otherwise, we can't index into this, bail out.
6614 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6615 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6616 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6617 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6620 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6622 GEPIdxTy = STy->getElementType();
6624 // Otherwise, we can't index into this, bail out.
6630 // If we were able to index down into an element, create the GEP
6631 // and bitcast the result. This eliminates one bitcast, potentially
6633 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6635 NewIndices.end(), "");
6636 InsertNewInstBefore(NGEP, CI);
6637 NGEP->takeName(GEP);
6639 if (isa<BitCastInst>(CI))
6640 return new BitCastInst(NGEP, CI.getType());
6641 assert(isa<PtrToIntInst>(CI));
6642 return new PtrToIntInst(NGEP, CI.getType());
6649 return commonCastTransforms(CI);
6654 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6655 /// integer types. This function implements the common transforms for all those
6657 /// @brief Implement the transforms common to CastInst with integer operands
6658 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6659 if (Instruction *Result = commonCastTransforms(CI))
6662 Value *Src = CI.getOperand(0);
6663 const Type *SrcTy = Src->getType();
6664 const Type *DestTy = CI.getType();
6665 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6666 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6668 // See if we can simplify any instructions used by the LHS whose sole
6669 // purpose is to compute bits we don't care about.
6670 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6671 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6672 KnownZero, KnownOne))
6675 // If the source isn't an instruction or has more than one use then we
6676 // can't do anything more.
6677 Instruction *SrcI = dyn_cast<Instruction>(Src);
6678 if (!SrcI || !Src->hasOneUse())
6681 // Attempt to propagate the cast into the instruction for int->int casts.
6682 int NumCastsRemoved = 0;
6683 if (!isa<BitCastInst>(CI) &&
6684 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6685 CI.getOpcode(), NumCastsRemoved)) {
6686 // If this cast is a truncate, evaluting in a different type always
6687 // eliminates the cast, so it is always a win. If this is a zero-extension,
6688 // we need to do an AND to maintain the clear top-part of the computation,
6689 // so we require that the input have eliminated at least one cast. If this
6690 // is a sign extension, we insert two new casts (to do the extension) so we
6691 // require that two casts have been eliminated.
6693 switch (CI.getOpcode()) {
6695 // All the others use floating point so we shouldn't actually
6696 // get here because of the check above.
6697 assert(0 && "Unknown cast type");
6698 case Instruction::Trunc:
6701 case Instruction::ZExt:
6702 DoXForm = NumCastsRemoved >= 1;
6704 case Instruction::SExt:
6705 DoXForm = NumCastsRemoved >= 2;
6710 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6711 CI.getOpcode() == Instruction::SExt);
6712 assert(Res->getType() == DestTy);
6713 switch (CI.getOpcode()) {
6714 default: assert(0 && "Unknown cast type!");
6715 case Instruction::Trunc:
6716 case Instruction::BitCast:
6717 // Just replace this cast with the result.
6718 return ReplaceInstUsesWith(CI, Res);
6719 case Instruction::ZExt: {
6720 // We need to emit an AND to clear the high bits.
6721 assert(SrcBitSize < DestBitSize && "Not a zext?");
6722 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6724 return BinaryOperator::createAnd(Res, C);
6726 case Instruction::SExt:
6727 // We need to emit a cast to truncate, then a cast to sext.
6728 return CastInst::create(Instruction::SExt,
6729 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6735 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6736 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6738 switch (SrcI->getOpcode()) {
6739 case Instruction::Add:
6740 case Instruction::Mul:
6741 case Instruction::And:
6742 case Instruction::Or:
6743 case Instruction::Xor:
6744 // If we are discarding information, rewrite.
6745 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6746 // Don't insert two casts if they cannot be eliminated. We allow
6747 // two casts to be inserted if the sizes are the same. This could
6748 // only be converting signedness, which is a noop.
6749 if (DestBitSize == SrcBitSize ||
6750 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6751 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6752 Instruction::CastOps opcode = CI.getOpcode();
6753 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6754 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6755 return BinaryOperator::create(
6756 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6760 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6761 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6762 SrcI->getOpcode() == Instruction::Xor &&
6763 Op1 == ConstantInt::getTrue() &&
6764 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6765 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6766 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6769 case Instruction::SDiv:
6770 case Instruction::UDiv:
6771 case Instruction::SRem:
6772 case Instruction::URem:
6773 // If we are just changing the sign, rewrite.
6774 if (DestBitSize == SrcBitSize) {
6775 // Don't insert two casts if they cannot be eliminated. We allow
6776 // two casts to be inserted if the sizes are the same. This could
6777 // only be converting signedness, which is a noop.
6778 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6779 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6780 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6782 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6784 return BinaryOperator::create(
6785 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6790 case Instruction::Shl:
6791 // Allow changing the sign of the source operand. Do not allow
6792 // changing the size of the shift, UNLESS the shift amount is a
6793 // constant. We must not change variable sized shifts to a smaller
6794 // size, because it is undefined to shift more bits out than exist
6796 if (DestBitSize == SrcBitSize ||
6797 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6798 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6799 Instruction::BitCast : Instruction::Trunc);
6800 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6801 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6802 return BinaryOperator::createShl(Op0c, Op1c);
6805 case Instruction::AShr:
6806 // If this is a signed shr, and if all bits shifted in are about to be
6807 // truncated off, turn it into an unsigned shr to allow greater
6809 if (DestBitSize < SrcBitSize &&
6810 isa<ConstantInt>(Op1)) {
6811 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6812 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6813 // Insert the new logical shift right.
6814 return BinaryOperator::createLShr(Op0, Op1);
6822 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6823 if (Instruction *Result = commonIntCastTransforms(CI))
6826 Value *Src = CI.getOperand(0);
6827 const Type *Ty = CI.getType();
6828 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6829 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6831 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6832 switch (SrcI->getOpcode()) {
6834 case Instruction::LShr:
6835 // We can shrink lshr to something smaller if we know the bits shifted in
6836 // are already zeros.
6837 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6838 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6840 // Get a mask for the bits shifting in.
6841 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6842 Value* SrcIOp0 = SrcI->getOperand(0);
6843 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6844 if (ShAmt >= DestBitWidth) // All zeros.
6845 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6847 // Okay, we can shrink this. Truncate the input, then return a new
6849 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6850 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6852 return BinaryOperator::createLShr(V1, V2);
6854 } else { // This is a variable shr.
6856 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6857 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6858 // loop-invariant and CSE'd.
6859 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6860 Value *One = ConstantInt::get(SrcI->getType(), 1);
6862 Value *V = InsertNewInstBefore(
6863 BinaryOperator::createShl(One, SrcI->getOperand(1),
6865 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6866 SrcI->getOperand(0),
6868 Value *Zero = Constant::getNullValue(V->getType());
6869 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6879 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6880 // If one of the common conversion will work ..
6881 if (Instruction *Result = commonIntCastTransforms(CI))
6884 Value *Src = CI.getOperand(0);
6886 // If this is a cast of a cast
6887 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6888 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6889 // types and if the sizes are just right we can convert this into a logical
6890 // 'and' which will be much cheaper than the pair of casts.
6891 if (isa<TruncInst>(CSrc)) {
6892 // Get the sizes of the types involved
6893 Value *A = CSrc->getOperand(0);
6894 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6895 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6896 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6897 // If we're actually extending zero bits and the trunc is a no-op
6898 if (MidSize < DstSize && SrcSize == DstSize) {
6899 // Replace both of the casts with an And of the type mask.
6900 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6901 Constant *AndConst = ConstantInt::get(AndValue);
6903 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6904 // Unfortunately, if the type changed, we need to cast it back.
6905 if (And->getType() != CI.getType()) {
6906 And->setName(CSrc->getName()+".mask");
6907 InsertNewInstBefore(And, CI);
6908 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6915 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6916 // If we are just checking for a icmp eq of a single bit and zext'ing it
6917 // to an integer, then shift the bit to the appropriate place and then
6918 // cast to integer to avoid the comparison.
6919 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6920 const APInt &Op1CV = Op1C->getValue();
6922 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6923 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6924 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6925 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6926 Value *In = ICI->getOperand(0);
6927 Value *Sh = ConstantInt::get(In->getType(),
6928 In->getType()->getPrimitiveSizeInBits()-1);
6929 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6930 In->getName()+".lobit"),
6932 if (In->getType() != CI.getType())
6933 In = CastInst::createIntegerCast(In, CI.getType(),
6934 false/*ZExt*/, "tmp", &CI);
6936 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6937 Constant *One = ConstantInt::get(In->getType(), 1);
6938 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6939 In->getName()+".not"),
6943 return ReplaceInstUsesWith(CI, In);
6948 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6949 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6950 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6951 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6952 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6953 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6954 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6955 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6956 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6957 // This only works for EQ and NE
6958 ICI->isEquality()) {
6959 // If Op1C some other power of two, convert:
6960 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6961 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6962 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6963 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6965 APInt KnownZeroMask(~KnownZero);
6966 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6967 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6968 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6969 // (X&4) == 2 --> false
6970 // (X&4) != 2 --> true
6971 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6972 Res = ConstantExpr::getZExt(Res, CI.getType());
6973 return ReplaceInstUsesWith(CI, Res);
6976 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6977 Value *In = ICI->getOperand(0);
6979 // Perform a logical shr by shiftamt.
6980 // Insert the shift to put the result in the low bit.
6981 In = InsertNewInstBefore(
6982 BinaryOperator::createLShr(In,
6983 ConstantInt::get(In->getType(), ShiftAmt),
6984 In->getName()+".lobit"), CI);
6987 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6988 Constant *One = ConstantInt::get(In->getType(), 1);
6989 In = BinaryOperator::createXor(In, One, "tmp");
6990 InsertNewInstBefore(cast<Instruction>(In), CI);
6993 if (CI.getType() == In->getType())
6994 return ReplaceInstUsesWith(CI, In);
6996 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7004 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7005 if (Instruction *I = commonIntCastTransforms(CI))
7008 Value *Src = CI.getOperand(0);
7010 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7011 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7012 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7013 // If we are just checking for a icmp eq of a single bit and zext'ing it
7014 // to an integer, then shift the bit to the appropriate place and then
7015 // cast to integer to avoid the comparison.
7016 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7017 const APInt &Op1CV = Op1C->getValue();
7019 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7020 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7021 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7022 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7023 Value *In = ICI->getOperand(0);
7024 Value *Sh = ConstantInt::get(In->getType(),
7025 In->getType()->getPrimitiveSizeInBits()-1);
7026 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7027 In->getName()+".lobit"),
7029 if (In->getType() != CI.getType())
7030 In = CastInst::createIntegerCast(In, CI.getType(),
7031 true/*SExt*/, "tmp", &CI);
7033 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7034 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7035 In->getName()+".not"), CI);
7037 return ReplaceInstUsesWith(CI, In);
7045 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7046 return commonCastTransforms(CI);
7049 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7050 return commonCastTransforms(CI);
7053 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7054 return commonCastTransforms(CI);
7057 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7058 return commonCastTransforms(CI);
7061 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7062 return commonCastTransforms(CI);
7065 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7066 return commonCastTransforms(CI);
7069 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7070 return commonPointerCastTransforms(CI);
7073 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7074 return commonCastTransforms(CI);
7077 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7078 // If the operands are integer typed then apply the integer transforms,
7079 // otherwise just apply the common ones.
7080 Value *Src = CI.getOperand(0);
7081 const Type *SrcTy = Src->getType();
7082 const Type *DestTy = CI.getType();
7084 if (SrcTy->isInteger() && DestTy->isInteger()) {
7085 if (Instruction *Result = commonIntCastTransforms(CI))
7087 } else if (isa<PointerType>(SrcTy)) {
7088 if (Instruction *I = commonPointerCastTransforms(CI))
7091 if (Instruction *Result = commonCastTransforms(CI))
7096 // Get rid of casts from one type to the same type. These are useless and can
7097 // be replaced by the operand.
7098 if (DestTy == Src->getType())
7099 return ReplaceInstUsesWith(CI, Src);
7101 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7102 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7103 const Type *DstElTy = DstPTy->getElementType();
7104 const Type *SrcElTy = SrcPTy->getElementType();
7106 // If we are casting a malloc or alloca to a pointer to a type of the same
7107 // size, rewrite the allocation instruction to allocate the "right" type.
7108 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7109 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7112 // If the source and destination are pointers, and this cast is equivalent
7113 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7114 // This can enhance SROA and other transforms that want type-safe pointers.
7115 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7116 unsigned NumZeros = 0;
7117 while (SrcElTy != DstElTy &&
7118 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7119 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7120 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7124 // If we found a path from the src to dest, create the getelementptr now.
7125 if (SrcElTy == DstElTy) {
7126 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7127 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7128 ((Instruction*) NULL));
7132 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7133 if (SVI->hasOneUse()) {
7134 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7135 // a bitconvert to a vector with the same # elts.
7136 if (isa<VectorType>(DestTy) &&
7137 cast<VectorType>(DestTy)->getNumElements() ==
7138 SVI->getType()->getNumElements()) {
7140 // If either of the operands is a cast from CI.getType(), then
7141 // evaluating the shuffle in the casted destination's type will allow
7142 // us to eliminate at least one cast.
7143 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7144 Tmp->getOperand(0)->getType() == DestTy) ||
7145 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7146 Tmp->getOperand(0)->getType() == DestTy)) {
7147 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7148 SVI->getOperand(0), DestTy, &CI);
7149 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7150 SVI->getOperand(1), DestTy, &CI);
7151 // Return a new shuffle vector. Use the same element ID's, as we
7152 // know the vector types match #elts.
7153 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7161 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7163 /// %D = select %cond, %C, %A
7165 /// %C = select %cond, %B, 0
7168 /// Assuming that the specified instruction is an operand to the select, return
7169 /// a bitmask indicating which operands of this instruction are foldable if they
7170 /// equal the other incoming value of the select.
7172 static unsigned GetSelectFoldableOperands(Instruction *I) {
7173 switch (I->getOpcode()) {
7174 case Instruction::Add:
7175 case Instruction::Mul:
7176 case Instruction::And:
7177 case Instruction::Or:
7178 case Instruction::Xor:
7179 return 3; // Can fold through either operand.
7180 case Instruction::Sub: // Can only fold on the amount subtracted.
7181 case Instruction::Shl: // Can only fold on the shift amount.
7182 case Instruction::LShr:
7183 case Instruction::AShr:
7186 return 0; // Cannot fold
7190 /// GetSelectFoldableConstant - For the same transformation as the previous
7191 /// function, return the identity constant that goes into the select.
7192 static Constant *GetSelectFoldableConstant(Instruction *I) {
7193 switch (I->getOpcode()) {
7194 default: assert(0 && "This cannot happen!"); abort();
7195 case Instruction::Add:
7196 case Instruction::Sub:
7197 case Instruction::Or:
7198 case Instruction::Xor:
7199 case Instruction::Shl:
7200 case Instruction::LShr:
7201 case Instruction::AShr:
7202 return Constant::getNullValue(I->getType());
7203 case Instruction::And:
7204 return Constant::getAllOnesValue(I->getType());
7205 case Instruction::Mul:
7206 return ConstantInt::get(I->getType(), 1);
7210 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7211 /// have the same opcode and only one use each. Try to simplify this.
7212 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7214 if (TI->getNumOperands() == 1) {
7215 // If this is a non-volatile load or a cast from the same type,
7218 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7221 return 0; // unknown unary op.
7224 // Fold this by inserting a select from the input values.
7225 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7226 FI->getOperand(0), SI.getName()+".v");
7227 InsertNewInstBefore(NewSI, SI);
7228 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7232 // Only handle binary operators here.
7233 if (!isa<BinaryOperator>(TI))
7236 // Figure out if the operations have any operands in common.
7237 Value *MatchOp, *OtherOpT, *OtherOpF;
7239 if (TI->getOperand(0) == FI->getOperand(0)) {
7240 MatchOp = TI->getOperand(0);
7241 OtherOpT = TI->getOperand(1);
7242 OtherOpF = FI->getOperand(1);
7243 MatchIsOpZero = true;
7244 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7245 MatchOp = TI->getOperand(1);
7246 OtherOpT = TI->getOperand(0);
7247 OtherOpF = FI->getOperand(0);
7248 MatchIsOpZero = false;
7249 } else if (!TI->isCommutative()) {
7251 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7252 MatchOp = TI->getOperand(0);
7253 OtherOpT = TI->getOperand(1);
7254 OtherOpF = FI->getOperand(0);
7255 MatchIsOpZero = true;
7256 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7257 MatchOp = TI->getOperand(1);
7258 OtherOpT = TI->getOperand(0);
7259 OtherOpF = FI->getOperand(1);
7260 MatchIsOpZero = true;
7265 // If we reach here, they do have operations in common.
7266 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7267 OtherOpF, SI.getName()+".v");
7268 InsertNewInstBefore(NewSI, SI);
7270 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7272 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7274 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7276 assert(0 && "Shouldn't get here");
7280 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7281 Value *CondVal = SI.getCondition();
7282 Value *TrueVal = SI.getTrueValue();
7283 Value *FalseVal = SI.getFalseValue();
7285 // select true, X, Y -> X
7286 // select false, X, Y -> Y
7287 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7288 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7290 // select C, X, X -> X
7291 if (TrueVal == FalseVal)
7292 return ReplaceInstUsesWith(SI, TrueVal);
7294 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7295 return ReplaceInstUsesWith(SI, FalseVal);
7296 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7297 return ReplaceInstUsesWith(SI, TrueVal);
7298 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7299 if (isa<Constant>(TrueVal))
7300 return ReplaceInstUsesWith(SI, TrueVal);
7302 return ReplaceInstUsesWith(SI, FalseVal);
7305 if (SI.getType() == Type::Int1Ty) {
7306 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7307 if (C->getZExtValue()) {
7308 // Change: A = select B, true, C --> A = or B, C
7309 return BinaryOperator::createOr(CondVal, FalseVal);
7311 // Change: A = select B, false, C --> A = and !B, C
7313 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7314 "not."+CondVal->getName()), SI);
7315 return BinaryOperator::createAnd(NotCond, FalseVal);
7317 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7318 if (C->getZExtValue() == false) {
7319 // Change: A = select B, C, false --> A = and B, C
7320 return BinaryOperator::createAnd(CondVal, TrueVal);
7322 // Change: A = select B, C, true --> A = or !B, C
7324 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7325 "not."+CondVal->getName()), SI);
7326 return BinaryOperator::createOr(NotCond, TrueVal);
7331 // Selecting between two integer constants?
7332 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7333 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7334 // select C, 1, 0 -> zext C to int
7335 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7336 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7337 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7338 // select C, 0, 1 -> zext !C to int
7340 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7341 "not."+CondVal->getName()), SI);
7342 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7345 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7347 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7349 // (x <s 0) ? -1 : 0 -> ashr x, 31
7350 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7351 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7352 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7353 // The comparison constant and the result are not neccessarily the
7354 // same width. Make an all-ones value by inserting a AShr.
7355 Value *X = IC->getOperand(0);
7356 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7357 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7358 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7360 InsertNewInstBefore(SRA, SI);
7362 // Finally, convert to the type of the select RHS. We figure out
7363 // if this requires a SExt, Trunc or BitCast based on the sizes.
7364 Instruction::CastOps opc = Instruction::BitCast;
7365 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7366 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7367 if (SRASize < SISize)
7368 opc = Instruction::SExt;
7369 else if (SRASize > SISize)
7370 opc = Instruction::Trunc;
7371 return CastInst::create(opc, SRA, SI.getType());
7376 // If one of the constants is zero (we know they can't both be) and we
7377 // have an icmp instruction with zero, and we have an 'and' with the
7378 // non-constant value, eliminate this whole mess. This corresponds to
7379 // cases like this: ((X & 27) ? 27 : 0)
7380 if (TrueValC->isZero() || FalseValC->isZero())
7381 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7382 cast<Constant>(IC->getOperand(1))->isNullValue())
7383 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7384 if (ICA->getOpcode() == Instruction::And &&
7385 isa<ConstantInt>(ICA->getOperand(1)) &&
7386 (ICA->getOperand(1) == TrueValC ||
7387 ICA->getOperand(1) == FalseValC) &&
7388 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7389 // Okay, now we know that everything is set up, we just don't
7390 // know whether we have a icmp_ne or icmp_eq and whether the
7391 // true or false val is the zero.
7392 bool ShouldNotVal = !TrueValC->isZero();
7393 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7396 V = InsertNewInstBefore(BinaryOperator::create(
7397 Instruction::Xor, V, ICA->getOperand(1)), SI);
7398 return ReplaceInstUsesWith(SI, V);
7403 // See if we are selecting two values based on a comparison of the two values.
7404 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7405 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7406 // Transform (X == Y) ? X : Y -> Y
7407 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7408 // This is not safe in general for floating point:
7409 // consider X== -0, Y== +0.
7410 // It becomes safe if either operand is a nonzero constant.
7411 ConstantFP *CFPt, *CFPf;
7412 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7413 !CFPt->getValueAPF().isZero()) ||
7414 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7415 !CFPf->getValueAPF().isZero()))
7416 return ReplaceInstUsesWith(SI, FalseVal);
7418 // Transform (X != Y) ? X : Y -> X
7419 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7420 return ReplaceInstUsesWith(SI, TrueVal);
7421 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7423 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7424 // Transform (X == Y) ? Y : X -> X
7425 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7426 // This is not safe in general for floating point:
7427 // consider X== -0, Y== +0.
7428 // It becomes safe if either operand is a nonzero constant.
7429 ConstantFP *CFPt, *CFPf;
7430 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7431 !CFPt->getValueAPF().isZero()) ||
7432 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7433 !CFPf->getValueAPF().isZero()))
7434 return ReplaceInstUsesWith(SI, FalseVal);
7436 // Transform (X != Y) ? Y : X -> Y
7437 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7438 return ReplaceInstUsesWith(SI, TrueVal);
7439 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7443 // See if we are selecting two values based on a comparison of the two values.
7444 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7445 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7446 // Transform (X == Y) ? X : Y -> Y
7447 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7448 return ReplaceInstUsesWith(SI, FalseVal);
7449 // Transform (X != Y) ? X : Y -> X
7450 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7451 return ReplaceInstUsesWith(SI, TrueVal);
7452 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7454 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7455 // Transform (X == Y) ? Y : X -> X
7456 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7457 return ReplaceInstUsesWith(SI, FalseVal);
7458 // Transform (X != Y) ? Y : X -> Y
7459 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7460 return ReplaceInstUsesWith(SI, TrueVal);
7461 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7465 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7466 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7467 if (TI->hasOneUse() && FI->hasOneUse()) {
7468 Instruction *AddOp = 0, *SubOp = 0;
7470 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7471 if (TI->getOpcode() == FI->getOpcode())
7472 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7475 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7476 // even legal for FP.
7477 if (TI->getOpcode() == Instruction::Sub &&
7478 FI->getOpcode() == Instruction::Add) {
7479 AddOp = FI; SubOp = TI;
7480 } else if (FI->getOpcode() == Instruction::Sub &&
7481 TI->getOpcode() == Instruction::Add) {
7482 AddOp = TI; SubOp = FI;
7486 Value *OtherAddOp = 0;
7487 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7488 OtherAddOp = AddOp->getOperand(1);
7489 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7490 OtherAddOp = AddOp->getOperand(0);
7494 // So at this point we know we have (Y -> OtherAddOp):
7495 // select C, (add X, Y), (sub X, Z)
7496 Value *NegVal; // Compute -Z
7497 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7498 NegVal = ConstantExpr::getNeg(C);
7500 NegVal = InsertNewInstBefore(
7501 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7504 Value *NewTrueOp = OtherAddOp;
7505 Value *NewFalseOp = NegVal;
7507 std::swap(NewTrueOp, NewFalseOp);
7508 Instruction *NewSel =
7509 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7511 NewSel = InsertNewInstBefore(NewSel, SI);
7512 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7517 // See if we can fold the select into one of our operands.
7518 if (SI.getType()->isInteger()) {
7519 // See the comment above GetSelectFoldableOperands for a description of the
7520 // transformation we are doing here.
7521 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7522 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7523 !isa<Constant>(FalseVal))
7524 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7525 unsigned OpToFold = 0;
7526 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7528 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7533 Constant *C = GetSelectFoldableConstant(TVI);
7534 Instruction *NewSel =
7535 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7536 InsertNewInstBefore(NewSel, SI);
7537 NewSel->takeName(TVI);
7538 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7539 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7541 assert(0 && "Unknown instruction!!");
7546 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7547 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7548 !isa<Constant>(TrueVal))
7549 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7550 unsigned OpToFold = 0;
7551 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7553 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7558 Constant *C = GetSelectFoldableConstant(FVI);
7559 Instruction *NewSel =
7560 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7561 InsertNewInstBefore(NewSel, SI);
7562 NewSel->takeName(FVI);
7563 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7564 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7566 assert(0 && "Unknown instruction!!");
7571 if (BinaryOperator::isNot(CondVal)) {
7572 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7573 SI.setOperand(1, FalseVal);
7574 SI.setOperand(2, TrueVal);
7581 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7582 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7583 /// and it is more than the alignment of the ultimate object, see if we can
7584 /// increase the alignment of the ultimate object, making this check succeed.
7585 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7586 unsigned PrefAlign = 0) {
7587 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7588 unsigned Align = GV->getAlignment();
7589 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7590 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7592 // If there is a large requested alignment and we can, bump up the alignment
7594 if (PrefAlign > Align && GV->hasInitializer()) {
7595 GV->setAlignment(PrefAlign);
7599 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7600 unsigned Align = AI->getAlignment();
7601 if (Align == 0 && TD) {
7602 if (isa<AllocaInst>(AI))
7603 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7604 else if (isa<MallocInst>(AI)) {
7605 // Malloc returns maximally aligned memory.
7606 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7609 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7612 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7616 // If there is a requested alignment and if this is an alloca, round up. We
7617 // don't do this for malloc, because some systems can't respect the request.
7618 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7619 AI->setAlignment(PrefAlign);
7623 } else if (isa<BitCastInst>(V) ||
7624 (isa<ConstantExpr>(V) &&
7625 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7626 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7628 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7629 // If all indexes are zero, it is just the alignment of the base pointer.
7630 bool AllZeroOperands = true;
7631 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7632 if (!isa<Constant>(GEPI->getOperand(i)) ||
7633 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7634 AllZeroOperands = false;
7638 if (AllZeroOperands) {
7639 // Treat this like a bitcast.
7640 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7643 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7644 if (BaseAlignment == 0) return 0;
7646 // Otherwise, if the base alignment is >= the alignment we expect for the
7647 // base pointer type, then we know that the resultant pointer is aligned at
7648 // least as much as its type requires.
7651 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7652 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7653 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7654 if (Align <= BaseAlignment) {
7655 const Type *GEPTy = GEPI->getType();
7656 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7657 Align = std::min(Align, (unsigned)
7658 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7667 /// visitCallInst - CallInst simplification. This mostly only handles folding
7668 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7669 /// the heavy lifting.
7671 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7672 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7673 if (!II) return visitCallSite(&CI);
7675 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7677 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7678 bool Changed = false;
7680 // memmove/cpy/set of zero bytes is a noop.
7681 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7682 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7684 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7685 if (CI->getZExtValue() == 1) {
7686 // Replace the instruction with just byte operations. We would
7687 // transform other cases to loads/stores, but we don't know if
7688 // alignment is sufficient.
7692 // If we have a memmove and the source operation is a constant global,
7693 // then the source and dest pointers can't alias, so we can change this
7694 // into a call to memcpy.
7695 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7696 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7697 if (GVSrc->isConstant()) {
7698 Module *M = CI.getParent()->getParent()->getParent();
7700 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7702 Name = "llvm.memcpy.i32";
7704 Name = "llvm.memcpy.i64";
7705 Constant *MemCpy = M->getOrInsertFunction(Name,
7706 CI.getCalledFunction()->getFunctionType());
7707 CI.setOperand(0, MemCpy);
7712 // If we can determine a pointer alignment that is bigger than currently
7713 // set, update the alignment.
7714 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7715 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7716 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7717 unsigned Align = std::min(Alignment1, Alignment2);
7718 if (MI->getAlignment()->getZExtValue() < Align) {
7719 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7723 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7725 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(CI.getOperand(3));
7727 unsigned Size = MemOpLength->getZExtValue();
7728 unsigned Align = cast<ConstantInt>(CI.getOperand(4))->getZExtValue();
7729 PointerType *NewPtrTy = NULL;
7730 // Destination pointer type is always i8 *
7731 // If Size is 8 then use Int64Ty
7732 // If Size is 4 then use Int32Ty
7733 // If Size is 2 then use Int16Ty
7734 // If Size is 1 then use Int8Ty
7735 if (Size && Size <=8 && !(Size&(Size-1)))
7736 NewPtrTy = PointerType::get(IntegerType::get(Size<<3));
7739 Value *Src = InsertCastBefore(Instruction::BitCast, CI.getOperand(2),
7741 Value *Dest = InsertCastBefore(Instruction::BitCast, CI.getOperand(1),
7743 Value *L = new LoadInst(Src, "tmp", false, Align, &CI);
7744 Value *NS = new StoreInst(L, Dest, false, Align, &CI);
7745 CI.replaceAllUsesWith(NS);
7747 return EraseInstFromFunction(CI);
7750 } else if (isa<MemSetInst>(MI)) {
7751 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7752 if (MI->getAlignment()->getZExtValue() < Alignment) {
7753 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7758 if (Changed) return II;
7760 switch (II->getIntrinsicID()) {
7762 case Intrinsic::ppc_altivec_lvx:
7763 case Intrinsic::ppc_altivec_lvxl:
7764 case Intrinsic::x86_sse_loadu_ps:
7765 case Intrinsic::x86_sse2_loadu_pd:
7766 case Intrinsic::x86_sse2_loadu_dq:
7767 // Turn PPC lvx -> load if the pointer is known aligned.
7768 // Turn X86 loadups -> load if the pointer is known aligned.
7769 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7770 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7771 PointerType::get(II->getType()), CI);
7772 return new LoadInst(Ptr);
7775 case Intrinsic::ppc_altivec_stvx:
7776 case Intrinsic::ppc_altivec_stvxl:
7777 // Turn stvx -> store if the pointer is known aligned.
7778 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7779 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7780 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7782 return new StoreInst(II->getOperand(1), Ptr);
7785 case Intrinsic::x86_sse_storeu_ps:
7786 case Intrinsic::x86_sse2_storeu_pd:
7787 case Intrinsic::x86_sse2_storeu_dq:
7788 case Intrinsic::x86_sse2_storel_dq:
7789 // Turn X86 storeu -> store if the pointer is known aligned.
7790 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7791 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7792 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7794 return new StoreInst(II->getOperand(2), Ptr);
7798 case Intrinsic::x86_sse_cvttss2si: {
7799 // These intrinsics only demands the 0th element of its input vector. If
7800 // we can simplify the input based on that, do so now.
7802 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7804 II->setOperand(1, V);
7810 case Intrinsic::ppc_altivec_vperm:
7811 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7812 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7813 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7815 // Check that all of the elements are integer constants or undefs.
7816 bool AllEltsOk = true;
7817 for (unsigned i = 0; i != 16; ++i) {
7818 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7819 !isa<UndefValue>(Mask->getOperand(i))) {
7826 // Cast the input vectors to byte vectors.
7827 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7828 II->getOperand(1), Mask->getType(), CI);
7829 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7830 II->getOperand(2), Mask->getType(), CI);
7831 Value *Result = UndefValue::get(Op0->getType());
7833 // Only extract each element once.
7834 Value *ExtractedElts[32];
7835 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7837 for (unsigned i = 0; i != 16; ++i) {
7838 if (isa<UndefValue>(Mask->getOperand(i)))
7840 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7841 Idx &= 31; // Match the hardware behavior.
7843 if (ExtractedElts[Idx] == 0) {
7845 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7846 InsertNewInstBefore(Elt, CI);
7847 ExtractedElts[Idx] = Elt;
7850 // Insert this value into the result vector.
7851 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7852 InsertNewInstBefore(cast<Instruction>(Result), CI);
7854 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7859 case Intrinsic::stackrestore: {
7860 // If the save is right next to the restore, remove the restore. This can
7861 // happen when variable allocas are DCE'd.
7862 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7863 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7864 BasicBlock::iterator BI = SS;
7866 return EraseInstFromFunction(CI);
7870 // If the stack restore is in a return/unwind block and if there are no
7871 // allocas or calls between the restore and the return, nuke the restore.
7872 TerminatorInst *TI = II->getParent()->getTerminator();
7873 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7874 BasicBlock::iterator BI = II;
7875 bool CannotRemove = false;
7876 for (++BI; &*BI != TI; ++BI) {
7877 if (isa<AllocaInst>(BI) ||
7878 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7879 CannotRemove = true;
7884 return EraseInstFromFunction(CI);
7891 return visitCallSite(II);
7894 // InvokeInst simplification
7896 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7897 return visitCallSite(&II);
7900 // visitCallSite - Improvements for call and invoke instructions.
7902 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7903 bool Changed = false;
7905 // If the callee is a constexpr cast of a function, attempt to move the cast
7906 // to the arguments of the call/invoke.
7907 if (transformConstExprCastCall(CS)) return 0;
7909 Value *Callee = CS.getCalledValue();
7911 if (Function *CalleeF = dyn_cast<Function>(Callee))
7912 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7913 Instruction *OldCall = CS.getInstruction();
7914 // If the call and callee calling conventions don't match, this call must
7915 // be unreachable, as the call is undefined.
7916 new StoreInst(ConstantInt::getTrue(),
7917 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7918 if (!OldCall->use_empty())
7919 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7920 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7921 return EraseInstFromFunction(*OldCall);
7925 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7926 // This instruction is not reachable, just remove it. We insert a store to
7927 // undef so that we know that this code is not reachable, despite the fact
7928 // that we can't modify the CFG here.
7929 new StoreInst(ConstantInt::getTrue(),
7930 UndefValue::get(PointerType::get(Type::Int1Ty)),
7931 CS.getInstruction());
7933 if (!CS.getInstruction()->use_empty())
7934 CS.getInstruction()->
7935 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7937 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7938 // Don't break the CFG, insert a dummy cond branch.
7939 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7940 ConstantInt::getTrue(), II);
7942 return EraseInstFromFunction(*CS.getInstruction());
7945 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
7946 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
7947 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
7948 return transformCallThroughTrampoline(CS);
7950 const PointerType *PTy = cast<PointerType>(Callee->getType());
7951 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7952 if (FTy->isVarArg()) {
7953 // See if we can optimize any arguments passed through the varargs area of
7955 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7956 E = CS.arg_end(); I != E; ++I)
7957 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7958 // If this cast does not effect the value passed through the varargs
7959 // area, we can eliminate the use of the cast.
7960 Value *Op = CI->getOperand(0);
7961 if (CI->isLosslessCast()) {
7968 return Changed ? CS.getInstruction() : 0;
7971 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7972 // attempt to move the cast to the arguments of the call/invoke.
7974 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7975 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7976 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7977 if (CE->getOpcode() != Instruction::BitCast ||
7978 !isa<Function>(CE->getOperand(0)))
7980 Function *Callee = cast<Function>(CE->getOperand(0));
7981 Instruction *Caller = CS.getInstruction();
7983 // Okay, this is a cast from a function to a different type. Unless doing so
7984 // would cause a type conversion of one of our arguments, change this call to
7985 // be a direct call with arguments casted to the appropriate types.
7987 const FunctionType *FT = Callee->getFunctionType();
7988 const Type *OldRetTy = Caller->getType();
7990 const FunctionType *ActualFT =
7991 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7993 // If the parameter attributes are not compatible, don't do the xform. We
7994 // don't want to lose an sret attribute or something.
7995 if (!ParamAttrsList::areCompatible(FT->getParamAttrs(),
7996 ActualFT->getParamAttrs()))
7999 // Check to see if we are changing the return type...
8000 if (OldRetTy != FT->getReturnType()) {
8001 if (Callee->isDeclaration() && !Caller->use_empty() &&
8002 // Conversion is ok if changing from pointer to int of same size.
8003 !(isa<PointerType>(FT->getReturnType()) &&
8004 TD->getIntPtrType() == OldRetTy))
8005 return false; // Cannot transform this return value.
8007 // If the callsite is an invoke instruction, and the return value is used by
8008 // a PHI node in a successor, we cannot change the return type of the call
8009 // because there is no place to put the cast instruction (without breaking
8010 // the critical edge). Bail out in this case.
8011 if (!Caller->use_empty())
8012 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8013 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8015 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8016 if (PN->getParent() == II->getNormalDest() ||
8017 PN->getParent() == II->getUnwindDest())
8021 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8022 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8024 CallSite::arg_iterator AI = CS.arg_begin();
8025 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8026 const Type *ParamTy = FT->getParamType(i);
8027 const Type *ActTy = (*AI)->getType();
8028 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8029 //Some conversions are safe even if we do not have a body.
8030 //Either we can cast directly, or we can upconvert the argument
8031 bool isConvertible = ActTy == ParamTy ||
8032 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8033 (ParamTy->isInteger() && ActTy->isInteger() &&
8034 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8035 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8036 && c->getValue().isStrictlyPositive());
8037 if (Callee->isDeclaration() && !isConvertible) return false;
8039 // Most other conversions can be done if we have a body, even if these
8040 // lose information, e.g. int->short.
8041 // Some conversions cannot be done at all, e.g. float to pointer.
8042 // Logic here parallels CastInst::getCastOpcode (the design there
8043 // requires legality checks like this be done before calling it).
8044 if (ParamTy->isInteger()) {
8045 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8046 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8049 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
8050 !isa<PointerType>(ActTy))
8052 } else if (ParamTy->isFloatingPoint()) {
8053 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8054 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8057 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
8059 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
8060 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8061 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
8064 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
8066 } else if (isa<PointerType>(ParamTy)) {
8067 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
8074 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8075 Callee->isDeclaration())
8076 return false; // Do not delete arguments unless we have a function body...
8078 // Okay, we decided that this is a safe thing to do: go ahead and start
8079 // inserting cast instructions as necessary...
8080 std::vector<Value*> Args;
8081 Args.reserve(NumActualArgs);
8083 AI = CS.arg_begin();
8084 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8085 const Type *ParamTy = FT->getParamType(i);
8086 if ((*AI)->getType() == ParamTy) {
8087 Args.push_back(*AI);
8089 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8090 false, ParamTy, false);
8091 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8092 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8096 // If the function takes more arguments than the call was taking, add them
8098 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8099 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8101 // If we are removing arguments to the function, emit an obnoxious warning...
8102 if (FT->getNumParams() < NumActualArgs)
8103 if (!FT->isVarArg()) {
8104 cerr << "WARNING: While resolving call to function '"
8105 << Callee->getName() << "' arguments were dropped!\n";
8107 // Add all of the arguments in their promoted form to the arg list...
8108 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8109 const Type *PTy = getPromotedType((*AI)->getType());
8110 if (PTy != (*AI)->getType()) {
8111 // Must promote to pass through va_arg area!
8112 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8114 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8115 InsertNewInstBefore(Cast, *Caller);
8116 Args.push_back(Cast);
8118 Args.push_back(*AI);
8123 if (FT->getReturnType() == Type::VoidTy)
8124 Caller->setName(""); // Void type should not have a name.
8127 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8128 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8129 Args.begin(), Args.end(), Caller->getName(), Caller);
8130 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8132 NC = new CallInst(Callee, Args.begin(), Args.end(),
8133 Caller->getName(), Caller);
8134 if (cast<CallInst>(Caller)->isTailCall())
8135 cast<CallInst>(NC)->setTailCall();
8136 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8139 // Insert a cast of the return type as necessary.
8141 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8142 if (NV->getType() != Type::VoidTy) {
8143 const Type *CallerTy = Caller->getType();
8144 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8146 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8148 // If this is an invoke instruction, we should insert it after the first
8149 // non-phi, instruction in the normal successor block.
8150 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8151 BasicBlock::iterator I = II->getNormalDest()->begin();
8152 while (isa<PHINode>(I)) ++I;
8153 InsertNewInstBefore(NC, *I);
8155 // Otherwise, it's a call, just insert cast right after the call instr
8156 InsertNewInstBefore(NC, *Caller);
8158 AddUsersToWorkList(*Caller);
8160 NV = UndefValue::get(Caller->getType());
8164 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8165 Caller->replaceAllUsesWith(NV);
8166 Caller->eraseFromParent();
8167 RemoveFromWorkList(Caller);
8171 // transformCallThroughTrampoline - Turn a call to a function created by the
8172 // init_trampoline intrinsic into a direct call to the underlying function.
8174 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8175 Value *Callee = CS.getCalledValue();
8176 const PointerType *PTy = cast<PointerType>(Callee->getType());
8177 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8179 IntrinsicInst *Tramp =
8180 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8183 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8184 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8185 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8187 if (const ParamAttrsList *NestAttrs = NestFTy->getParamAttrs()) {
8188 unsigned NestIdx = 1;
8189 const Type *NestTy = 0;
8190 uint16_t NestAttr = 0;
8192 // Look for a parameter marked with the 'nest' attribute.
8193 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8194 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8195 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8196 // Record the parameter type and any other attributes.
8198 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8203 Instruction *Caller = CS.getInstruction();
8204 std::vector<Value*> NewArgs;
8205 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8207 // Insert the nest argument into the call argument list, which may
8208 // mean appending it.
8211 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8213 if (Idx == NestIdx) {
8214 // Add the chain argument.
8215 Value *NestVal = Tramp->getOperand(3);
8216 if (NestVal->getType() != NestTy)
8217 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8218 NewArgs.push_back(NestVal);
8224 // Add the original argument.
8225 NewArgs.push_back(*I);
8231 // The trampoline may have been bitcast to a bogus type (FTy).
8232 // Handle this by synthesizing a new function type, equal to FTy
8233 // with the chain parameter inserted. Likewise for attributes.
8235 const ParamAttrsList *Attrs = FTy->getParamAttrs();
8236 std::vector<const Type*> NewTypes;
8237 ParamAttrsVector NewAttrs;
8238 NewTypes.reserve(FTy->getNumParams()+1);
8240 // Add any function result attributes.
8241 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8243 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8245 // Insert the chain's type into the list of parameter types, which may
8246 // mean appending it. Likewise for the chain's attributes.
8249 FunctionType::param_iterator I = FTy->param_begin(),
8250 E = FTy->param_end();
8253 if (Idx == NestIdx) {
8254 // Add the chain's type and attributes.
8255 NewTypes.push_back(NestTy);
8256 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8262 // Add the original type and attributes.
8263 NewTypes.push_back(*I);
8264 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8267 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8273 // Replace the trampoline call with a direct call. Let the generic
8274 // code sort out any function type mismatches.
8275 FunctionType *NewFTy =
8276 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg(),
8277 ParamAttrsList::get(NewAttrs));
8278 Constant *NewCallee = NestF->getType() == PointerType::get(NewFTy) ?
8279 NestF : ConstantExpr::getBitCast(NestF, PointerType::get(NewFTy));
8281 Instruction *NewCaller;
8282 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8283 NewCaller = new InvokeInst(NewCallee,
8284 II->getNormalDest(), II->getUnwindDest(),
8285 NewArgs.begin(), NewArgs.end(),
8286 Caller->getName(), Caller);
8287 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8289 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8290 Caller->getName(), Caller);
8291 if (cast<CallInst>(Caller)->isTailCall())
8292 cast<CallInst>(NewCaller)->setTailCall();
8293 cast<CallInst>(NewCaller)->
8294 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8296 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8297 Caller->replaceAllUsesWith(NewCaller);
8298 Caller->eraseFromParent();
8299 RemoveFromWorkList(Caller);
8304 // Replace the trampoline call with a direct call. Since there is no 'nest'
8305 // parameter, there is no need to adjust the argument list. Let the generic
8306 // code sort out any function type mismatches.
8307 Constant *NewCallee =
8308 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8309 CS.setCalledFunction(NewCallee);
8310 return CS.getInstruction();
8313 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8314 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8315 /// and a single binop.
8316 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8317 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8318 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8319 isa<CmpInst>(FirstInst));
8320 unsigned Opc = FirstInst->getOpcode();
8321 Value *LHSVal = FirstInst->getOperand(0);
8322 Value *RHSVal = FirstInst->getOperand(1);
8324 const Type *LHSType = LHSVal->getType();
8325 const Type *RHSType = RHSVal->getType();
8327 // Scan to see if all operands are the same opcode, all have one use, and all
8328 // kill their operands (i.e. the operands have one use).
8329 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8330 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8331 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8332 // Verify type of the LHS matches so we don't fold cmp's of different
8333 // types or GEP's with different index types.
8334 I->getOperand(0)->getType() != LHSType ||
8335 I->getOperand(1)->getType() != RHSType)
8338 // If they are CmpInst instructions, check their predicates
8339 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8340 if (cast<CmpInst>(I)->getPredicate() !=
8341 cast<CmpInst>(FirstInst)->getPredicate())
8344 // Keep track of which operand needs a phi node.
8345 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8346 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8349 // Otherwise, this is safe to transform, determine if it is profitable.
8351 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8352 // Indexes are often folded into load/store instructions, so we don't want to
8353 // hide them behind a phi.
8354 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8357 Value *InLHS = FirstInst->getOperand(0);
8358 Value *InRHS = FirstInst->getOperand(1);
8359 PHINode *NewLHS = 0, *NewRHS = 0;
8361 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8362 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8363 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8364 InsertNewInstBefore(NewLHS, PN);
8369 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8370 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8371 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8372 InsertNewInstBefore(NewRHS, PN);
8376 // Add all operands to the new PHIs.
8377 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8379 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8380 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8383 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8384 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8388 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8389 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8390 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8391 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8394 assert(isa<GetElementPtrInst>(FirstInst));
8395 return new GetElementPtrInst(LHSVal, RHSVal);
8399 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8400 /// of the block that defines it. This means that it must be obvious the value
8401 /// of the load is not changed from the point of the load to the end of the
8404 /// Finally, it is safe, but not profitable, to sink a load targetting a
8405 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8407 static bool isSafeToSinkLoad(LoadInst *L) {
8408 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8410 for (++BBI; BBI != E; ++BBI)
8411 if (BBI->mayWriteToMemory())
8414 // Check for non-address taken alloca. If not address-taken already, it isn't
8415 // profitable to do this xform.
8416 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8417 bool isAddressTaken = false;
8418 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8420 if (isa<LoadInst>(UI)) continue;
8421 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8422 // If storing TO the alloca, then the address isn't taken.
8423 if (SI->getOperand(1) == AI) continue;
8425 isAddressTaken = true;
8429 if (!isAddressTaken)
8437 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8438 // operator and they all are only used by the PHI, PHI together their
8439 // inputs, and do the operation once, to the result of the PHI.
8440 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8441 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8443 // Scan the instruction, looking for input operations that can be folded away.
8444 // If all input operands to the phi are the same instruction (e.g. a cast from
8445 // the same type or "+42") we can pull the operation through the PHI, reducing
8446 // code size and simplifying code.
8447 Constant *ConstantOp = 0;
8448 const Type *CastSrcTy = 0;
8449 bool isVolatile = false;
8450 if (isa<CastInst>(FirstInst)) {
8451 CastSrcTy = FirstInst->getOperand(0)->getType();
8452 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8453 // Can fold binop, compare or shift here if the RHS is a constant,
8454 // otherwise call FoldPHIArgBinOpIntoPHI.
8455 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8456 if (ConstantOp == 0)
8457 return FoldPHIArgBinOpIntoPHI(PN);
8458 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8459 isVolatile = LI->isVolatile();
8460 // We can't sink the load if the loaded value could be modified between the
8461 // load and the PHI.
8462 if (LI->getParent() != PN.getIncomingBlock(0) ||
8463 !isSafeToSinkLoad(LI))
8465 } else if (isa<GetElementPtrInst>(FirstInst)) {
8466 if (FirstInst->getNumOperands() == 2)
8467 return FoldPHIArgBinOpIntoPHI(PN);
8468 // Can't handle general GEPs yet.
8471 return 0; // Cannot fold this operation.
8474 // Check to see if all arguments are the same operation.
8475 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8476 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8477 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8478 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8481 if (I->getOperand(0)->getType() != CastSrcTy)
8482 return 0; // Cast operation must match.
8483 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8484 // We can't sink the load if the loaded value could be modified between
8485 // the load and the PHI.
8486 if (LI->isVolatile() != isVolatile ||
8487 LI->getParent() != PN.getIncomingBlock(i) ||
8488 !isSafeToSinkLoad(LI))
8490 } else if (I->getOperand(1) != ConstantOp) {
8495 // Okay, they are all the same operation. Create a new PHI node of the
8496 // correct type, and PHI together all of the LHS's of the instructions.
8497 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8498 PN.getName()+".in");
8499 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8501 Value *InVal = FirstInst->getOperand(0);
8502 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8504 // Add all operands to the new PHI.
8505 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8506 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8507 if (NewInVal != InVal)
8509 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8514 // The new PHI unions all of the same values together. This is really
8515 // common, so we handle it intelligently here for compile-time speed.
8519 InsertNewInstBefore(NewPN, PN);
8523 // Insert and return the new operation.
8524 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8525 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8526 else if (isa<LoadInst>(FirstInst))
8527 return new LoadInst(PhiVal, "", isVolatile);
8528 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8529 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8530 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8531 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8532 PhiVal, ConstantOp);
8534 assert(0 && "Unknown operation");
8538 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8540 static bool DeadPHICycle(PHINode *PN,
8541 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8542 if (PN->use_empty()) return true;
8543 if (!PN->hasOneUse()) return false;
8545 // Remember this node, and if we find the cycle, return.
8546 if (!PotentiallyDeadPHIs.insert(PN))
8549 // Don't scan crazily complex things.
8550 if (PotentiallyDeadPHIs.size() == 16)
8553 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8554 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8559 /// PHIsEqualValue - Return true if this phi node is always equal to
8560 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8561 /// z = some value; x = phi (y, z); y = phi (x, z)
8562 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8563 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8564 // See if we already saw this PHI node.
8565 if (!ValueEqualPHIs.insert(PN))
8568 // Don't scan crazily complex things.
8569 if (ValueEqualPHIs.size() == 16)
8572 // Scan the operands to see if they are either phi nodes or are equal to
8574 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8575 Value *Op = PN->getIncomingValue(i);
8576 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8577 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8579 } else if (Op != NonPhiInVal)
8587 // PHINode simplification
8589 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8590 // If LCSSA is around, don't mess with Phi nodes
8591 if (MustPreserveLCSSA) return 0;
8593 if (Value *V = PN.hasConstantValue())
8594 return ReplaceInstUsesWith(PN, V);
8596 // If all PHI operands are the same operation, pull them through the PHI,
8597 // reducing code size.
8598 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8599 PN.getIncomingValue(0)->hasOneUse())
8600 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8603 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8604 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8605 // PHI)... break the cycle.
8606 if (PN.hasOneUse()) {
8607 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8608 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8609 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8610 PotentiallyDeadPHIs.insert(&PN);
8611 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8612 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8615 // If this phi has a single use, and if that use just computes a value for
8616 // the next iteration of a loop, delete the phi. This occurs with unused
8617 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8618 // common case here is good because the only other things that catch this
8619 // are induction variable analysis (sometimes) and ADCE, which is only run
8621 if (PHIUser->hasOneUse() &&
8622 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8623 PHIUser->use_back() == &PN) {
8624 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8628 // We sometimes end up with phi cycles that non-obviously end up being the
8629 // same value, for example:
8630 // z = some value; x = phi (y, z); y = phi (x, z)
8631 // where the phi nodes don't necessarily need to be in the same block. Do a
8632 // quick check to see if the PHI node only contains a single non-phi value, if
8633 // so, scan to see if the phi cycle is actually equal to that value.
8635 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
8636 // Scan for the first non-phi operand.
8637 while (InValNo != NumOperandVals &&
8638 isa<PHINode>(PN.getIncomingValue(InValNo)))
8641 if (InValNo != NumOperandVals) {
8642 Value *NonPhiInVal = PN.getOperand(InValNo);
8644 // Scan the rest of the operands to see if there are any conflicts, if so
8645 // there is no need to recursively scan other phis.
8646 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
8647 Value *OpVal = PN.getIncomingValue(InValNo);
8648 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
8652 // If we scanned over all operands, then we have one unique value plus
8653 // phi values. Scan PHI nodes to see if they all merge in each other or
8655 if (InValNo == NumOperandVals) {
8656 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
8657 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
8658 return ReplaceInstUsesWith(PN, NonPhiInVal);
8665 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8666 Instruction *InsertPoint,
8668 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8669 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8670 // We must cast correctly to the pointer type. Ensure that we
8671 // sign extend the integer value if it is smaller as this is
8672 // used for address computation.
8673 Instruction::CastOps opcode =
8674 (VTySize < PtrSize ? Instruction::SExt :
8675 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8676 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8680 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8681 Value *PtrOp = GEP.getOperand(0);
8682 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8683 // If so, eliminate the noop.
8684 if (GEP.getNumOperands() == 1)
8685 return ReplaceInstUsesWith(GEP, PtrOp);
8687 if (isa<UndefValue>(GEP.getOperand(0)))
8688 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8690 bool HasZeroPointerIndex = false;
8691 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8692 HasZeroPointerIndex = C->isNullValue();
8694 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8695 return ReplaceInstUsesWith(GEP, PtrOp);
8697 // Eliminate unneeded casts for indices.
8698 bool MadeChange = false;
8700 gep_type_iterator GTI = gep_type_begin(GEP);
8701 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8702 if (isa<SequentialType>(*GTI)) {
8703 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8704 if (CI->getOpcode() == Instruction::ZExt ||
8705 CI->getOpcode() == Instruction::SExt) {
8706 const Type *SrcTy = CI->getOperand(0)->getType();
8707 // We can eliminate a cast from i32 to i64 iff the target
8708 // is a 32-bit pointer target.
8709 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8711 GEP.setOperand(i, CI->getOperand(0));
8715 // If we are using a wider index than needed for this platform, shrink it
8716 // to what we need. If the incoming value needs a cast instruction,
8717 // insert it. This explicit cast can make subsequent optimizations more
8719 Value *Op = GEP.getOperand(i);
8720 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
8721 if (Constant *C = dyn_cast<Constant>(Op)) {
8722 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8725 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8727 GEP.setOperand(i, Op);
8732 if (MadeChange) return &GEP;
8734 // If this GEP instruction doesn't move the pointer, and if the input operand
8735 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8736 // real input to the dest type.
8737 if (GEP.hasAllZeroIndices()) {
8738 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
8739 // If the bitcast is of an allocation, and the allocation will be
8740 // converted to match the type of the cast, don't touch this.
8741 if (isa<AllocationInst>(BCI->getOperand(0))) {
8742 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
8743 if (Instruction *I = visitBitCast(*BCI)) {
8746 BCI->getParent()->getInstList().insert(BCI, I);
8747 ReplaceInstUsesWith(*BCI, I);
8752 return new BitCastInst(BCI->getOperand(0), GEP.getType());
8756 // Combine Indices - If the source pointer to this getelementptr instruction
8757 // is a getelementptr instruction, combine the indices of the two
8758 // getelementptr instructions into a single instruction.
8760 SmallVector<Value*, 8> SrcGEPOperands;
8761 if (User *Src = dyn_castGetElementPtr(PtrOp))
8762 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8764 if (!SrcGEPOperands.empty()) {
8765 // Note that if our source is a gep chain itself that we wait for that
8766 // chain to be resolved before we perform this transformation. This
8767 // avoids us creating a TON of code in some cases.
8769 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8770 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8771 return 0; // Wait until our source is folded to completion.
8773 SmallVector<Value*, 8> Indices;
8775 // Find out whether the last index in the source GEP is a sequential idx.
8776 bool EndsWithSequential = false;
8777 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8778 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8779 EndsWithSequential = !isa<StructType>(*I);
8781 // Can we combine the two pointer arithmetics offsets?
8782 if (EndsWithSequential) {
8783 // Replace: gep (gep %P, long B), long A, ...
8784 // With: T = long A+B; gep %P, T, ...
8786 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8787 if (SO1 == Constant::getNullValue(SO1->getType())) {
8789 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8792 // If they aren't the same type, convert both to an integer of the
8793 // target's pointer size.
8794 if (SO1->getType() != GO1->getType()) {
8795 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8796 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8797 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8798 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8800 unsigned PS = TD->getPointerSizeInBits();
8801 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
8802 // Convert GO1 to SO1's type.
8803 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8805 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
8806 // Convert SO1 to GO1's type.
8807 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8809 const Type *PT = TD->getIntPtrType();
8810 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8811 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8815 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8816 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8818 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8819 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8823 // Recycle the GEP we already have if possible.
8824 if (SrcGEPOperands.size() == 2) {
8825 GEP.setOperand(0, SrcGEPOperands[0]);
8826 GEP.setOperand(1, Sum);
8829 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8830 SrcGEPOperands.end()-1);
8831 Indices.push_back(Sum);
8832 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8834 } else if (isa<Constant>(*GEP.idx_begin()) &&
8835 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8836 SrcGEPOperands.size() != 1) {
8837 // Otherwise we can do the fold if the first index of the GEP is a zero
8838 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8839 SrcGEPOperands.end());
8840 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8843 if (!Indices.empty())
8844 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8845 Indices.end(), GEP.getName());
8847 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8848 // GEP of global variable. If all of the indices for this GEP are
8849 // constants, we can promote this to a constexpr instead of an instruction.
8851 // Scan for nonconstants...
8852 SmallVector<Constant*, 8> Indices;
8853 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8854 for (; I != E && isa<Constant>(*I); ++I)
8855 Indices.push_back(cast<Constant>(*I));
8857 if (I == E) { // If they are all constants...
8858 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8859 &Indices[0],Indices.size());
8861 // Replace all uses of the GEP with the new constexpr...
8862 return ReplaceInstUsesWith(GEP, CE);
8864 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8865 if (!isa<PointerType>(X->getType())) {
8866 // Not interesting. Source pointer must be a cast from pointer.
8867 } else if (HasZeroPointerIndex) {
8868 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8869 // into : GEP [10 x ubyte]* X, long 0, ...
8871 // This occurs when the program declares an array extern like "int X[];"
8873 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8874 const PointerType *XTy = cast<PointerType>(X->getType());
8875 if (const ArrayType *XATy =
8876 dyn_cast<ArrayType>(XTy->getElementType()))
8877 if (const ArrayType *CATy =
8878 dyn_cast<ArrayType>(CPTy->getElementType()))
8879 if (CATy->getElementType() == XATy->getElementType()) {
8880 // At this point, we know that the cast source type is a pointer
8881 // to an array of the same type as the destination pointer
8882 // array. Because the array type is never stepped over (there
8883 // is a leading zero) we can fold the cast into this GEP.
8884 GEP.setOperand(0, X);
8887 } else if (GEP.getNumOperands() == 2) {
8888 // Transform things like:
8889 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8890 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8891 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8892 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8893 if (isa<ArrayType>(SrcElTy) &&
8894 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8895 TD->getABITypeSize(ResElTy)) {
8897 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8898 Idx[1] = GEP.getOperand(1);
8899 Value *V = InsertNewInstBefore(
8900 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8901 // V and GEP are both pointer types --> BitCast
8902 return new BitCastInst(V, GEP.getType());
8905 // Transform things like:
8906 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8907 // (where tmp = 8*tmp2) into:
8908 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8910 if (isa<ArrayType>(SrcElTy) &&
8911 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8912 uint64_t ArrayEltSize =
8913 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8915 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8916 // allow either a mul, shift, or constant here.
8918 ConstantInt *Scale = 0;
8919 if (ArrayEltSize == 1) {
8920 NewIdx = GEP.getOperand(1);
8921 Scale = ConstantInt::get(NewIdx->getType(), 1);
8922 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8923 NewIdx = ConstantInt::get(CI->getType(), 1);
8925 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8926 if (Inst->getOpcode() == Instruction::Shl &&
8927 isa<ConstantInt>(Inst->getOperand(1))) {
8928 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8929 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8930 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8931 NewIdx = Inst->getOperand(0);
8932 } else if (Inst->getOpcode() == Instruction::Mul &&
8933 isa<ConstantInt>(Inst->getOperand(1))) {
8934 Scale = cast<ConstantInt>(Inst->getOperand(1));
8935 NewIdx = Inst->getOperand(0);
8939 // If the index will be to exactly the right offset with the scale taken
8940 // out, perform the transformation.
8941 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8942 if (isa<ConstantInt>(Scale))
8943 Scale = ConstantInt::get(Scale->getType(),
8944 Scale->getZExtValue() / ArrayEltSize);
8945 if (Scale->getZExtValue() != 1) {
8946 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8948 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8949 NewIdx = InsertNewInstBefore(Sc, GEP);
8952 // Insert the new GEP instruction.
8954 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8956 Instruction *NewGEP =
8957 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
8958 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8959 // The NewGEP must be pointer typed, so must the old one -> BitCast
8960 return new BitCastInst(NewGEP, GEP.getType());
8969 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8970 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8971 if (AI.isArrayAllocation()) // Check C != 1
8972 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8974 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8975 AllocationInst *New = 0;
8977 // Create and insert the replacement instruction...
8978 if (isa<MallocInst>(AI))
8979 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8981 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8982 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8985 InsertNewInstBefore(New, AI);
8987 // Scan to the end of the allocation instructions, to skip over a block of
8988 // allocas if possible...
8990 BasicBlock::iterator It = New;
8991 while (isa<AllocationInst>(*It)) ++It;
8993 // Now that I is pointing to the first non-allocation-inst in the block,
8994 // insert our getelementptr instruction...
8996 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9000 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9001 New->getName()+".sub", It);
9003 // Now make everything use the getelementptr instead of the original
9005 return ReplaceInstUsesWith(AI, V);
9006 } else if (isa<UndefValue>(AI.getArraySize())) {
9007 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9010 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9011 // Note that we only do this for alloca's, because malloc should allocate and
9012 // return a unique pointer, even for a zero byte allocation.
9013 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9014 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9015 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9020 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9021 Value *Op = FI.getOperand(0);
9023 // free undef -> unreachable.
9024 if (isa<UndefValue>(Op)) {
9025 // Insert a new store to null because we cannot modify the CFG here.
9026 new StoreInst(ConstantInt::getTrue(),
9027 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
9028 return EraseInstFromFunction(FI);
9031 // If we have 'free null' delete the instruction. This can happen in stl code
9032 // when lots of inlining happens.
9033 if (isa<ConstantPointerNull>(Op))
9034 return EraseInstFromFunction(FI);
9036 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9037 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9038 FI.setOperand(0, CI->getOperand(0));
9042 // Change free (gep X, 0,0,0,0) into free(X)
9043 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9044 if (GEPI->hasAllZeroIndices()) {
9045 AddToWorkList(GEPI);
9046 FI.setOperand(0, GEPI->getOperand(0));
9051 // Change free(malloc) into nothing, if the malloc has a single use.
9052 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9053 if (MI->hasOneUse()) {
9054 EraseInstFromFunction(FI);
9055 return EraseInstFromFunction(*MI);
9062 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9063 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9064 const TargetData *TD) {
9065 User *CI = cast<User>(LI.getOperand(0));
9066 Value *CastOp = CI->getOperand(0);
9068 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9069 // Instead of loading constant c string, use corresponding integer value
9070 // directly if string length is small enough.
9071 const std::string &Str = CE->getOperand(0)->getStringValue();
9073 unsigned len = Str.length();
9074 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9075 unsigned numBits = Ty->getPrimitiveSizeInBits();
9076 // Replace LI with immediate integer store.
9077 if ((numBits >> 3) == len + 1) {
9078 APInt StrVal(numBits, 0);
9079 APInt SingleChar(numBits, 0);
9080 if (TD->isLittleEndian()) {
9081 for (signed i = len-1; i >= 0; i--) {
9082 SingleChar = (uint64_t) Str[i];
9083 StrVal = (StrVal << 8) | SingleChar;
9086 for (unsigned i = 0; i < len; i++) {
9087 SingleChar = (uint64_t) Str[i];
9088 StrVal = (StrVal << 8) | SingleChar;
9090 // Append NULL at the end.
9092 StrVal = (StrVal << 8) | SingleChar;
9094 Value *NL = ConstantInt::get(StrVal);
9095 return IC.ReplaceInstUsesWith(LI, NL);
9100 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9101 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9102 const Type *SrcPTy = SrcTy->getElementType();
9104 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9105 isa<VectorType>(DestPTy)) {
9106 // If the source is an array, the code below will not succeed. Check to
9107 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9109 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9110 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9111 if (ASrcTy->getNumElements() != 0) {
9113 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9114 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9115 SrcTy = cast<PointerType>(CastOp->getType());
9116 SrcPTy = SrcTy->getElementType();
9119 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9120 isa<VectorType>(SrcPTy)) &&
9121 // Do not allow turning this into a load of an integer, which is then
9122 // casted to a pointer, this pessimizes pointer analysis a lot.
9123 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9124 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9125 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9127 // Okay, we are casting from one integer or pointer type to another of
9128 // the same size. Instead of casting the pointer before the load, cast
9129 // the result of the loaded value.
9130 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9132 LI.isVolatile()),LI);
9133 // Now cast the result of the load.
9134 return new BitCastInst(NewLoad, LI.getType());
9141 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9142 /// from this value cannot trap. If it is not obviously safe to load from the
9143 /// specified pointer, we do a quick local scan of the basic block containing
9144 /// ScanFrom, to determine if the address is already accessed.
9145 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9146 // If it is an alloca it is always safe to load from.
9147 if (isa<AllocaInst>(V)) return true;
9149 // If it is a global variable it is mostly safe to load from.
9150 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9151 // Don't try to evaluate aliases. External weak GV can be null.
9152 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9154 // Otherwise, be a little bit agressive by scanning the local block where we
9155 // want to check to see if the pointer is already being loaded or stored
9156 // from/to. If so, the previous load or store would have already trapped,
9157 // so there is no harm doing an extra load (also, CSE will later eliminate
9158 // the load entirely).
9159 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9164 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9165 if (LI->getOperand(0) == V) return true;
9166 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9167 if (SI->getOperand(1) == V) return true;
9173 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9174 /// until we find the underlying object a pointer is referring to or something
9175 /// we don't understand. Note that the returned pointer may be offset from the
9176 /// input, because we ignore GEP indices.
9177 static Value *GetUnderlyingObject(Value *Ptr) {
9179 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9180 if (CE->getOpcode() == Instruction::BitCast ||
9181 CE->getOpcode() == Instruction::GetElementPtr)
9182 Ptr = CE->getOperand(0);
9185 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9186 Ptr = BCI->getOperand(0);
9187 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9188 Ptr = GEP->getOperand(0);
9195 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9196 Value *Op = LI.getOperand(0);
9198 // Attempt to improve the alignment.
9199 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9200 if (KnownAlign > LI.getAlignment())
9201 LI.setAlignment(KnownAlign);
9203 // load (cast X) --> cast (load X) iff safe
9204 if (isa<CastInst>(Op))
9205 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9208 // None of the following transforms are legal for volatile loads.
9209 if (LI.isVolatile()) return 0;
9211 if (&LI.getParent()->front() != &LI) {
9212 BasicBlock::iterator BBI = &LI; --BBI;
9213 // If the instruction immediately before this is a store to the same
9214 // address, do a simple form of store->load forwarding.
9215 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9216 if (SI->getOperand(1) == LI.getOperand(0))
9217 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9218 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9219 if (LIB->getOperand(0) == LI.getOperand(0))
9220 return ReplaceInstUsesWith(LI, LIB);
9223 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
9224 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
9225 // Insert a new store to null instruction before the load to indicate
9226 // that this code is not reachable. We do this instead of inserting
9227 // an unreachable instruction directly because we cannot modify the
9229 new StoreInst(UndefValue::get(LI.getType()),
9230 Constant::getNullValue(Op->getType()), &LI);
9231 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9234 if (Constant *C = dyn_cast<Constant>(Op)) {
9235 // load null/undef -> undef
9236 if ((C->isNullValue() || isa<UndefValue>(C))) {
9237 // Insert a new store to null instruction before the load to indicate that
9238 // this code is not reachable. We do this instead of inserting an
9239 // unreachable instruction directly because we cannot modify the CFG.
9240 new StoreInst(UndefValue::get(LI.getType()),
9241 Constant::getNullValue(Op->getType()), &LI);
9242 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9245 // Instcombine load (constant global) into the value loaded.
9246 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9247 if (GV->isConstant() && !GV->isDeclaration())
9248 return ReplaceInstUsesWith(LI, GV->getInitializer());
9250 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9251 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9252 if (CE->getOpcode() == Instruction::GetElementPtr) {
9253 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9254 if (GV->isConstant() && !GV->isDeclaration())
9256 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9257 return ReplaceInstUsesWith(LI, V);
9258 if (CE->getOperand(0)->isNullValue()) {
9259 // Insert a new store to null instruction before the load to indicate
9260 // that this code is not reachable. We do this instead of inserting
9261 // an unreachable instruction directly because we cannot modify the
9263 new StoreInst(UndefValue::get(LI.getType()),
9264 Constant::getNullValue(Op->getType()), &LI);
9265 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9268 } else if (CE->isCast()) {
9269 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9274 // If this load comes from anywhere in a constant global, and if the global
9275 // is all undef or zero, we know what it loads.
9276 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9277 if (GV->isConstant() && GV->hasInitializer()) {
9278 if (GV->getInitializer()->isNullValue())
9279 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9280 else if (isa<UndefValue>(GV->getInitializer()))
9281 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9285 if (Op->hasOneUse()) {
9286 // Change select and PHI nodes to select values instead of addresses: this
9287 // helps alias analysis out a lot, allows many others simplifications, and
9288 // exposes redundancy in the code.
9290 // Note that we cannot do the transformation unless we know that the
9291 // introduced loads cannot trap! Something like this is valid as long as
9292 // the condition is always false: load (select bool %C, int* null, int* %G),
9293 // but it would not be valid if we transformed it to load from null
9296 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9297 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9298 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9299 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9300 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9301 SI->getOperand(1)->getName()+".val"), LI);
9302 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9303 SI->getOperand(2)->getName()+".val"), LI);
9304 return new SelectInst(SI->getCondition(), V1, V2);
9307 // load (select (cond, null, P)) -> load P
9308 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9309 if (C->isNullValue()) {
9310 LI.setOperand(0, SI->getOperand(2));
9314 // load (select (cond, P, null)) -> load P
9315 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9316 if (C->isNullValue()) {
9317 LI.setOperand(0, SI->getOperand(1));
9325 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9327 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9328 User *CI = cast<User>(SI.getOperand(1));
9329 Value *CastOp = CI->getOperand(0);
9331 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9332 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9333 const Type *SrcPTy = SrcTy->getElementType();
9335 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9336 // If the source is an array, the code below will not succeed. Check to
9337 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9339 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9340 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9341 if (ASrcTy->getNumElements() != 0) {
9343 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9344 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9345 SrcTy = cast<PointerType>(CastOp->getType());
9346 SrcPTy = SrcTy->getElementType();
9349 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9350 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9351 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9353 // Okay, we are casting from one integer or pointer type to another of
9354 // the same size. Instead of casting the pointer before
9355 // the store, cast the value to be stored.
9357 Value *SIOp0 = SI.getOperand(0);
9358 Instruction::CastOps opcode = Instruction::BitCast;
9359 const Type* CastSrcTy = SIOp0->getType();
9360 const Type* CastDstTy = SrcPTy;
9361 if (isa<PointerType>(CastDstTy)) {
9362 if (CastSrcTy->isInteger())
9363 opcode = Instruction::IntToPtr;
9364 } else if (isa<IntegerType>(CastDstTy)) {
9365 if (isa<PointerType>(SIOp0->getType()))
9366 opcode = Instruction::PtrToInt;
9368 if (Constant *C = dyn_cast<Constant>(SIOp0))
9369 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9371 NewCast = IC.InsertNewInstBefore(
9372 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9374 return new StoreInst(NewCast, CastOp);
9381 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9382 Value *Val = SI.getOperand(0);
9383 Value *Ptr = SI.getOperand(1);
9385 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9386 EraseInstFromFunction(SI);
9391 // If the RHS is an alloca with a single use, zapify the store, making the
9393 if (Ptr->hasOneUse()) {
9394 if (isa<AllocaInst>(Ptr)) {
9395 EraseInstFromFunction(SI);
9400 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9401 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9402 GEP->getOperand(0)->hasOneUse()) {
9403 EraseInstFromFunction(SI);
9409 // Attempt to improve the alignment.
9410 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9411 if (KnownAlign > SI.getAlignment())
9412 SI.setAlignment(KnownAlign);
9414 // Do really simple DSE, to catch cases where there are several consequtive
9415 // stores to the same location, separated by a few arithmetic operations. This
9416 // situation often occurs with bitfield accesses.
9417 BasicBlock::iterator BBI = &SI;
9418 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9422 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9423 // Prev store isn't volatile, and stores to the same location?
9424 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9427 EraseInstFromFunction(*PrevSI);
9433 // If this is a load, we have to stop. However, if the loaded value is from
9434 // the pointer we're loading and is producing the pointer we're storing,
9435 // then *this* store is dead (X = load P; store X -> P).
9436 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9437 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9438 EraseInstFromFunction(SI);
9442 // Otherwise, this is a load from some other location. Stores before it
9447 // Don't skip over loads or things that can modify memory.
9448 if (BBI->mayWriteToMemory())
9453 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9455 // store X, null -> turns into 'unreachable' in SimplifyCFG
9456 if (isa<ConstantPointerNull>(Ptr)) {
9457 if (!isa<UndefValue>(Val)) {
9458 SI.setOperand(0, UndefValue::get(Val->getType()));
9459 if (Instruction *U = dyn_cast<Instruction>(Val))
9460 AddToWorkList(U); // Dropped a use.
9463 return 0; // Do not modify these!
9466 // store undef, Ptr -> noop
9467 if (isa<UndefValue>(Val)) {
9468 EraseInstFromFunction(SI);
9473 // If the pointer destination is a cast, see if we can fold the cast into the
9475 if (isa<CastInst>(Ptr))
9476 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9478 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9480 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9484 // If this store is the last instruction in the basic block, and if the block
9485 // ends with an unconditional branch, try to move it to the successor block.
9487 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9488 if (BI->isUnconditional())
9489 if (SimplifyStoreAtEndOfBlock(SI))
9490 return 0; // xform done!
9495 /// SimplifyStoreAtEndOfBlock - Turn things like:
9496 /// if () { *P = v1; } else { *P = v2 }
9497 /// into a phi node with a store in the successor.
9499 /// Simplify things like:
9500 /// *P = v1; if () { *P = v2; }
9501 /// into a phi node with a store in the successor.
9503 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9504 BasicBlock *StoreBB = SI.getParent();
9506 // Check to see if the successor block has exactly two incoming edges. If
9507 // so, see if the other predecessor contains a store to the same location.
9508 // if so, insert a PHI node (if needed) and move the stores down.
9509 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9511 // Determine whether Dest has exactly two predecessors and, if so, compute
9512 // the other predecessor.
9513 pred_iterator PI = pred_begin(DestBB);
9514 BasicBlock *OtherBB = 0;
9518 if (PI == pred_end(DestBB))
9521 if (*PI != StoreBB) {
9526 if (++PI != pred_end(DestBB))
9530 // Verify that the other block ends in a branch and is not otherwise empty.
9531 BasicBlock::iterator BBI = OtherBB->getTerminator();
9532 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9533 if (!OtherBr || BBI == OtherBB->begin())
9536 // If the other block ends in an unconditional branch, check for the 'if then
9537 // else' case. there is an instruction before the branch.
9538 StoreInst *OtherStore = 0;
9539 if (OtherBr->isUnconditional()) {
9540 // If this isn't a store, or isn't a store to the same location, bail out.
9542 OtherStore = dyn_cast<StoreInst>(BBI);
9543 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9546 // Otherwise, the other block ended with a conditional branch. If one of the
9547 // destinations is StoreBB, then we have the if/then case.
9548 if (OtherBr->getSuccessor(0) != StoreBB &&
9549 OtherBr->getSuccessor(1) != StoreBB)
9552 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9553 // if/then triangle. See if there is a store to the same ptr as SI that
9554 // lives in OtherBB.
9556 // Check to see if we find the matching store.
9557 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9558 if (OtherStore->getOperand(1) != SI.getOperand(1))
9562 // If we find something that may be using the stored value, or if we run
9563 // out of instructions, we can't do the xform.
9564 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9565 BBI == OtherBB->begin())
9569 // In order to eliminate the store in OtherBr, we have to
9570 // make sure nothing reads the stored value in StoreBB.
9571 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9572 // FIXME: This should really be AA driven.
9573 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9578 // Insert a PHI node now if we need it.
9579 Value *MergedVal = OtherStore->getOperand(0);
9580 if (MergedVal != SI.getOperand(0)) {
9581 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9582 PN->reserveOperandSpace(2);
9583 PN->addIncoming(SI.getOperand(0), SI.getParent());
9584 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9585 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9588 // Advance to a place where it is safe to insert the new store and
9590 BBI = DestBB->begin();
9591 while (isa<PHINode>(BBI)) ++BBI;
9592 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9593 OtherStore->isVolatile()), *BBI);
9595 // Nuke the old stores.
9596 EraseInstFromFunction(SI);
9597 EraseInstFromFunction(*OtherStore);
9603 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9604 // Change br (not X), label True, label False to: br X, label False, True
9606 BasicBlock *TrueDest;
9607 BasicBlock *FalseDest;
9608 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9609 !isa<Constant>(X)) {
9610 // Swap Destinations and condition...
9612 BI.setSuccessor(0, FalseDest);
9613 BI.setSuccessor(1, TrueDest);
9617 // Cannonicalize fcmp_one -> fcmp_oeq
9618 FCmpInst::Predicate FPred; Value *Y;
9619 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9620 TrueDest, FalseDest)))
9621 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9622 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9623 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9624 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9625 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9626 NewSCC->takeName(I);
9627 // Swap Destinations and condition...
9628 BI.setCondition(NewSCC);
9629 BI.setSuccessor(0, FalseDest);
9630 BI.setSuccessor(1, TrueDest);
9631 RemoveFromWorkList(I);
9632 I->eraseFromParent();
9633 AddToWorkList(NewSCC);
9637 // Cannonicalize icmp_ne -> icmp_eq
9638 ICmpInst::Predicate IPred;
9639 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9640 TrueDest, FalseDest)))
9641 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9642 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9643 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9644 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9645 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9646 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9647 NewSCC->takeName(I);
9648 // Swap Destinations and condition...
9649 BI.setCondition(NewSCC);
9650 BI.setSuccessor(0, FalseDest);
9651 BI.setSuccessor(1, TrueDest);
9652 RemoveFromWorkList(I);
9653 I->eraseFromParent();;
9654 AddToWorkList(NewSCC);
9661 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9662 Value *Cond = SI.getCondition();
9663 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9664 if (I->getOpcode() == Instruction::Add)
9665 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9666 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9667 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9668 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9670 SI.setOperand(0, I->getOperand(0));
9678 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9679 /// is to leave as a vector operation.
9680 static bool CheapToScalarize(Value *V, bool isConstant) {
9681 if (isa<ConstantAggregateZero>(V))
9683 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9684 if (isConstant) return true;
9685 // If all elts are the same, we can extract.
9686 Constant *Op0 = C->getOperand(0);
9687 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9688 if (C->getOperand(i) != Op0)
9692 Instruction *I = dyn_cast<Instruction>(V);
9693 if (!I) return false;
9695 // Insert element gets simplified to the inserted element or is deleted if
9696 // this is constant idx extract element and its a constant idx insertelt.
9697 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9698 isa<ConstantInt>(I->getOperand(2)))
9700 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9702 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9703 if (BO->hasOneUse() &&
9704 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9705 CheapToScalarize(BO->getOperand(1), isConstant)))
9707 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9708 if (CI->hasOneUse() &&
9709 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9710 CheapToScalarize(CI->getOperand(1), isConstant)))
9716 /// Read and decode a shufflevector mask.
9718 /// It turns undef elements into values that are larger than the number of
9719 /// elements in the input.
9720 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9721 unsigned NElts = SVI->getType()->getNumElements();
9722 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9723 return std::vector<unsigned>(NElts, 0);
9724 if (isa<UndefValue>(SVI->getOperand(2)))
9725 return std::vector<unsigned>(NElts, 2*NElts);
9727 std::vector<unsigned> Result;
9728 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9729 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9730 if (isa<UndefValue>(CP->getOperand(i)))
9731 Result.push_back(NElts*2); // undef -> 8
9733 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9737 /// FindScalarElement - Given a vector and an element number, see if the scalar
9738 /// value is already around as a register, for example if it were inserted then
9739 /// extracted from the vector.
9740 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9741 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9742 const VectorType *PTy = cast<VectorType>(V->getType());
9743 unsigned Width = PTy->getNumElements();
9744 if (EltNo >= Width) // Out of range access.
9745 return UndefValue::get(PTy->getElementType());
9747 if (isa<UndefValue>(V))
9748 return UndefValue::get(PTy->getElementType());
9749 else if (isa<ConstantAggregateZero>(V))
9750 return Constant::getNullValue(PTy->getElementType());
9751 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9752 return CP->getOperand(EltNo);
9753 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9754 // If this is an insert to a variable element, we don't know what it is.
9755 if (!isa<ConstantInt>(III->getOperand(2)))
9757 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9759 // If this is an insert to the element we are looking for, return the
9762 return III->getOperand(1);
9764 // Otherwise, the insertelement doesn't modify the value, recurse on its
9766 return FindScalarElement(III->getOperand(0), EltNo);
9767 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9768 unsigned InEl = getShuffleMask(SVI)[EltNo];
9770 return FindScalarElement(SVI->getOperand(0), InEl);
9771 else if (InEl < Width*2)
9772 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9774 return UndefValue::get(PTy->getElementType());
9777 // Otherwise, we don't know.
9781 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9783 // If vector val is undef, replace extract with scalar undef.
9784 if (isa<UndefValue>(EI.getOperand(0)))
9785 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9787 // If vector val is constant 0, replace extract with scalar 0.
9788 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9789 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9791 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9792 // If vector val is constant with uniform operands, replace EI
9793 // with that operand
9794 Constant *op0 = C->getOperand(0);
9795 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9796 if (C->getOperand(i) != op0) {
9801 return ReplaceInstUsesWith(EI, op0);
9804 // If extracting a specified index from the vector, see if we can recursively
9805 // find a previously computed scalar that was inserted into the vector.
9806 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9807 unsigned IndexVal = IdxC->getZExtValue();
9808 unsigned VectorWidth =
9809 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9811 // If this is extracting an invalid index, turn this into undef, to avoid
9812 // crashing the code below.
9813 if (IndexVal >= VectorWidth)
9814 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9816 // This instruction only demands the single element from the input vector.
9817 // If the input vector has a single use, simplify it based on this use
9819 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9821 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9824 EI.setOperand(0, V);
9829 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9830 return ReplaceInstUsesWith(EI, Elt);
9832 // If the this extractelement is directly using a bitcast from a vector of
9833 // the same number of elements, see if we can find the source element from
9834 // it. In this case, we will end up needing to bitcast the scalars.
9835 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9836 if (const VectorType *VT =
9837 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9838 if (VT->getNumElements() == VectorWidth)
9839 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9840 return new BitCastInst(Elt, EI.getType());
9844 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9845 if (I->hasOneUse()) {
9846 // Push extractelement into predecessor operation if legal and
9847 // profitable to do so
9848 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9849 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9850 if (CheapToScalarize(BO, isConstantElt)) {
9851 ExtractElementInst *newEI0 =
9852 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9853 EI.getName()+".lhs");
9854 ExtractElementInst *newEI1 =
9855 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9856 EI.getName()+".rhs");
9857 InsertNewInstBefore(newEI0, EI);
9858 InsertNewInstBefore(newEI1, EI);
9859 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9861 } else if (isa<LoadInst>(I)) {
9862 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9863 PointerType::get(EI.getType()), EI);
9864 GetElementPtrInst *GEP =
9865 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9866 InsertNewInstBefore(GEP, EI);
9867 return new LoadInst(GEP);
9870 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9871 // Extracting the inserted element?
9872 if (IE->getOperand(2) == EI.getOperand(1))
9873 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9874 // If the inserted and extracted elements are constants, they must not
9875 // be the same value, extract from the pre-inserted value instead.
9876 if (isa<Constant>(IE->getOperand(2)) &&
9877 isa<Constant>(EI.getOperand(1))) {
9878 AddUsesToWorkList(EI);
9879 EI.setOperand(0, IE->getOperand(0));
9882 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9883 // If this is extracting an element from a shufflevector, figure out where
9884 // it came from and extract from the appropriate input element instead.
9885 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9886 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9888 if (SrcIdx < SVI->getType()->getNumElements())
9889 Src = SVI->getOperand(0);
9890 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9891 SrcIdx -= SVI->getType()->getNumElements();
9892 Src = SVI->getOperand(1);
9894 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9896 return new ExtractElementInst(Src, SrcIdx);
9903 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9904 /// elements from either LHS or RHS, return the shuffle mask and true.
9905 /// Otherwise, return false.
9906 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9907 std::vector<Constant*> &Mask) {
9908 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9909 "Invalid CollectSingleShuffleElements");
9910 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9912 if (isa<UndefValue>(V)) {
9913 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9915 } else if (V == LHS) {
9916 for (unsigned i = 0; i != NumElts; ++i)
9917 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9919 } else if (V == RHS) {
9920 for (unsigned i = 0; i != NumElts; ++i)
9921 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9923 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9924 // If this is an insert of an extract from some other vector, include it.
9925 Value *VecOp = IEI->getOperand(0);
9926 Value *ScalarOp = IEI->getOperand(1);
9927 Value *IdxOp = IEI->getOperand(2);
9929 if (!isa<ConstantInt>(IdxOp))
9931 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9933 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9934 // Okay, we can handle this if the vector we are insertinting into is
9936 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9937 // If so, update the mask to reflect the inserted undef.
9938 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9941 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9942 if (isa<ConstantInt>(EI->getOperand(1)) &&
9943 EI->getOperand(0)->getType() == V->getType()) {
9944 unsigned ExtractedIdx =
9945 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9947 // This must be extracting from either LHS or RHS.
9948 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9949 // Okay, we can handle this if the vector we are insertinting into is
9951 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9952 // If so, update the mask to reflect the inserted value.
9953 if (EI->getOperand(0) == LHS) {
9954 Mask[InsertedIdx & (NumElts-1)] =
9955 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9957 assert(EI->getOperand(0) == RHS);
9958 Mask[InsertedIdx & (NumElts-1)] =
9959 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9968 // TODO: Handle shufflevector here!
9973 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9974 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9975 /// that computes V and the LHS value of the shuffle.
9976 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9978 assert(isa<VectorType>(V->getType()) &&
9979 (RHS == 0 || V->getType() == RHS->getType()) &&
9980 "Invalid shuffle!");
9981 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9983 if (isa<UndefValue>(V)) {
9984 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9986 } else if (isa<ConstantAggregateZero>(V)) {
9987 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9989 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9990 // If this is an insert of an extract from some other vector, include it.
9991 Value *VecOp = IEI->getOperand(0);
9992 Value *ScalarOp = IEI->getOperand(1);
9993 Value *IdxOp = IEI->getOperand(2);
9995 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9996 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9997 EI->getOperand(0)->getType() == V->getType()) {
9998 unsigned ExtractedIdx =
9999 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10000 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10002 // Either the extracted from or inserted into vector must be RHSVec,
10003 // otherwise we'd end up with a shuffle of three inputs.
10004 if (EI->getOperand(0) == RHS || RHS == 0) {
10005 RHS = EI->getOperand(0);
10006 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10007 Mask[InsertedIdx & (NumElts-1)] =
10008 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10012 if (VecOp == RHS) {
10013 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10014 // Everything but the extracted element is replaced with the RHS.
10015 for (unsigned i = 0; i != NumElts; ++i) {
10016 if (i != InsertedIdx)
10017 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10022 // If this insertelement is a chain that comes from exactly these two
10023 // vectors, return the vector and the effective shuffle.
10024 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10025 return EI->getOperand(0);
10030 // TODO: Handle shufflevector here!
10032 // Otherwise, can't do anything fancy. Return an identity vector.
10033 for (unsigned i = 0; i != NumElts; ++i)
10034 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10038 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10039 Value *VecOp = IE.getOperand(0);
10040 Value *ScalarOp = IE.getOperand(1);
10041 Value *IdxOp = IE.getOperand(2);
10043 // Inserting an undef or into an undefined place, remove this.
10044 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10045 ReplaceInstUsesWith(IE, VecOp);
10047 // If the inserted element was extracted from some other vector, and if the
10048 // indexes are constant, try to turn this into a shufflevector operation.
10049 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10050 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10051 EI->getOperand(0)->getType() == IE.getType()) {
10052 unsigned NumVectorElts = IE.getType()->getNumElements();
10053 unsigned ExtractedIdx =
10054 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10055 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10057 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10058 return ReplaceInstUsesWith(IE, VecOp);
10060 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10061 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10063 // If we are extracting a value from a vector, then inserting it right
10064 // back into the same place, just use the input vector.
10065 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10066 return ReplaceInstUsesWith(IE, VecOp);
10068 // We could theoretically do this for ANY input. However, doing so could
10069 // turn chains of insertelement instructions into a chain of shufflevector
10070 // instructions, and right now we do not merge shufflevectors. As such,
10071 // only do this in a situation where it is clear that there is benefit.
10072 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10073 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10074 // the values of VecOp, except then one read from EIOp0.
10075 // Build a new shuffle mask.
10076 std::vector<Constant*> Mask;
10077 if (isa<UndefValue>(VecOp))
10078 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10080 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10081 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10084 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10085 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10086 ConstantVector::get(Mask));
10089 // If this insertelement isn't used by some other insertelement, turn it
10090 // (and any insertelements it points to), into one big shuffle.
10091 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10092 std::vector<Constant*> Mask;
10094 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10095 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10096 // We now have a shuffle of LHS, RHS, Mask.
10097 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10106 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10107 Value *LHS = SVI.getOperand(0);
10108 Value *RHS = SVI.getOperand(1);
10109 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10111 bool MadeChange = false;
10113 // Undefined shuffle mask -> undefined value.
10114 if (isa<UndefValue>(SVI.getOperand(2)))
10115 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10117 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10118 // the undef, change them to undefs.
10119 if (isa<UndefValue>(SVI.getOperand(1))) {
10120 // Scan to see if there are any references to the RHS. If so, replace them
10121 // with undef element refs and set MadeChange to true.
10122 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10123 if (Mask[i] >= e && Mask[i] != 2*e) {
10130 // Remap any references to RHS to use LHS.
10131 std::vector<Constant*> Elts;
10132 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10133 if (Mask[i] == 2*e)
10134 Elts.push_back(UndefValue::get(Type::Int32Ty));
10136 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10138 SVI.setOperand(2, ConstantVector::get(Elts));
10142 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10143 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10144 if (LHS == RHS || isa<UndefValue>(LHS)) {
10145 if (isa<UndefValue>(LHS) && LHS == RHS) {
10146 // shuffle(undef,undef,mask) -> undef.
10147 return ReplaceInstUsesWith(SVI, LHS);
10150 // Remap any references to RHS to use LHS.
10151 std::vector<Constant*> Elts;
10152 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10153 if (Mask[i] >= 2*e)
10154 Elts.push_back(UndefValue::get(Type::Int32Ty));
10156 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10157 (Mask[i] < e && isa<UndefValue>(LHS)))
10158 Mask[i] = 2*e; // Turn into undef.
10160 Mask[i] &= (e-1); // Force to LHS.
10161 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10164 SVI.setOperand(0, SVI.getOperand(1));
10165 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10166 SVI.setOperand(2, ConstantVector::get(Elts));
10167 LHS = SVI.getOperand(0);
10168 RHS = SVI.getOperand(1);
10172 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10173 bool isLHSID = true, isRHSID = true;
10175 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10176 if (Mask[i] >= e*2) continue; // Ignore undef values.
10177 // Is this an identity shuffle of the LHS value?
10178 isLHSID &= (Mask[i] == i);
10180 // Is this an identity shuffle of the RHS value?
10181 isRHSID &= (Mask[i]-e == i);
10184 // Eliminate identity shuffles.
10185 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10186 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10188 // If the LHS is a shufflevector itself, see if we can combine it with this
10189 // one without producing an unusual shuffle. Here we are really conservative:
10190 // we are absolutely afraid of producing a shuffle mask not in the input
10191 // program, because the code gen may not be smart enough to turn a merged
10192 // shuffle into two specific shuffles: it may produce worse code. As such,
10193 // we only merge two shuffles if the result is one of the two input shuffle
10194 // masks. In this case, merging the shuffles just removes one instruction,
10195 // which we know is safe. This is good for things like turning:
10196 // (splat(splat)) -> splat.
10197 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10198 if (isa<UndefValue>(RHS)) {
10199 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10201 std::vector<unsigned> NewMask;
10202 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10203 if (Mask[i] >= 2*e)
10204 NewMask.push_back(2*e);
10206 NewMask.push_back(LHSMask[Mask[i]]);
10208 // If the result mask is equal to the src shuffle or this shuffle mask, do
10209 // the replacement.
10210 if (NewMask == LHSMask || NewMask == Mask) {
10211 std::vector<Constant*> Elts;
10212 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10213 if (NewMask[i] >= e*2) {
10214 Elts.push_back(UndefValue::get(Type::Int32Ty));
10216 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10219 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10220 LHSSVI->getOperand(1),
10221 ConstantVector::get(Elts));
10226 return MadeChange ? &SVI : 0;
10232 /// TryToSinkInstruction - Try to move the specified instruction from its
10233 /// current block into the beginning of DestBlock, which can only happen if it's
10234 /// safe to move the instruction past all of the instructions between it and the
10235 /// end of its block.
10236 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10237 assert(I->hasOneUse() && "Invariants didn't hold!");
10239 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10240 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10242 // Do not sink alloca instructions out of the entry block.
10243 if (isa<AllocaInst>(I) && I->getParent() ==
10244 &DestBlock->getParent()->getEntryBlock())
10247 // We can only sink load instructions if there is nothing between the load and
10248 // the end of block that could change the value.
10249 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10250 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10252 if (Scan->mayWriteToMemory())
10256 BasicBlock::iterator InsertPos = DestBlock->begin();
10257 while (isa<PHINode>(InsertPos)) ++InsertPos;
10259 I->moveBefore(InsertPos);
10265 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10266 /// all reachable code to the worklist.
10268 /// This has a couple of tricks to make the code faster and more powerful. In
10269 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10270 /// them to the worklist (this significantly speeds up instcombine on code where
10271 /// many instructions are dead or constant). Additionally, if we find a branch
10272 /// whose condition is a known constant, we only visit the reachable successors.
10274 static void AddReachableCodeToWorklist(BasicBlock *BB,
10275 SmallPtrSet<BasicBlock*, 64> &Visited,
10277 const TargetData *TD) {
10278 std::vector<BasicBlock*> Worklist;
10279 Worklist.push_back(BB);
10281 while (!Worklist.empty()) {
10282 BB = Worklist.back();
10283 Worklist.pop_back();
10285 // We have now visited this block! If we've already been here, ignore it.
10286 if (!Visited.insert(BB)) continue;
10288 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10289 Instruction *Inst = BBI++;
10291 // DCE instruction if trivially dead.
10292 if (isInstructionTriviallyDead(Inst)) {
10294 DOUT << "IC: DCE: " << *Inst;
10295 Inst->eraseFromParent();
10299 // ConstantProp instruction if trivially constant.
10300 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10301 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10302 Inst->replaceAllUsesWith(C);
10304 Inst->eraseFromParent();
10308 IC.AddToWorkList(Inst);
10311 // Recursively visit successors. If this is a branch or switch on a
10312 // constant, only visit the reachable successor.
10313 TerminatorInst *TI = BB->getTerminator();
10314 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10315 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10316 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10317 Worklist.push_back(BI->getSuccessor(!CondVal));
10320 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10321 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10322 // See if this is an explicit destination.
10323 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10324 if (SI->getCaseValue(i) == Cond) {
10325 Worklist.push_back(SI->getSuccessor(i));
10329 // Otherwise it is the default destination.
10330 Worklist.push_back(SI->getSuccessor(0));
10335 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10336 Worklist.push_back(TI->getSuccessor(i));
10340 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10341 bool Changed = false;
10342 TD = &getAnalysis<TargetData>();
10344 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10345 << F.getNameStr() << "\n");
10348 // Do a depth-first traversal of the function, populate the worklist with
10349 // the reachable instructions. Ignore blocks that are not reachable. Keep
10350 // track of which blocks we visit.
10351 SmallPtrSet<BasicBlock*, 64> Visited;
10352 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10354 // Do a quick scan over the function. If we find any blocks that are
10355 // unreachable, remove any instructions inside of them. This prevents
10356 // the instcombine code from having to deal with some bad special cases.
10357 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10358 if (!Visited.count(BB)) {
10359 Instruction *Term = BB->getTerminator();
10360 while (Term != BB->begin()) { // Remove instrs bottom-up
10361 BasicBlock::iterator I = Term; --I;
10363 DOUT << "IC: DCE: " << *I;
10366 if (!I->use_empty())
10367 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10368 I->eraseFromParent();
10373 while (!Worklist.empty()) {
10374 Instruction *I = RemoveOneFromWorkList();
10375 if (I == 0) continue; // skip null values.
10377 // Check to see if we can DCE the instruction.
10378 if (isInstructionTriviallyDead(I)) {
10379 // Add operands to the worklist.
10380 if (I->getNumOperands() < 4)
10381 AddUsesToWorkList(*I);
10384 DOUT << "IC: DCE: " << *I;
10386 I->eraseFromParent();
10387 RemoveFromWorkList(I);
10391 // Instruction isn't dead, see if we can constant propagate it.
10392 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10393 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10395 // Add operands to the worklist.
10396 AddUsesToWorkList(*I);
10397 ReplaceInstUsesWith(*I, C);
10400 I->eraseFromParent();
10401 RemoveFromWorkList(I);
10405 // See if we can trivially sink this instruction to a successor basic block.
10406 if (I->hasOneUse()) {
10407 BasicBlock *BB = I->getParent();
10408 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10409 if (UserParent != BB) {
10410 bool UserIsSuccessor = false;
10411 // See if the user is one of our successors.
10412 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10413 if (*SI == UserParent) {
10414 UserIsSuccessor = true;
10418 // If the user is one of our immediate successors, and if that successor
10419 // only has us as a predecessors (we'd have to split the critical edge
10420 // otherwise), we can keep going.
10421 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10422 next(pred_begin(UserParent)) == pred_end(UserParent))
10423 // Okay, the CFG is simple enough, try to sink this instruction.
10424 Changed |= TryToSinkInstruction(I, UserParent);
10428 // Now that we have an instruction, try combining it to simplify it...
10432 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10433 if (Instruction *Result = visit(*I)) {
10435 // Should we replace the old instruction with a new one?
10437 DOUT << "IC: Old = " << *I
10438 << " New = " << *Result;
10440 // Everything uses the new instruction now.
10441 I->replaceAllUsesWith(Result);
10443 // Push the new instruction and any users onto the worklist.
10444 AddToWorkList(Result);
10445 AddUsersToWorkList(*Result);
10447 // Move the name to the new instruction first.
10448 Result->takeName(I);
10450 // Insert the new instruction into the basic block...
10451 BasicBlock *InstParent = I->getParent();
10452 BasicBlock::iterator InsertPos = I;
10454 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10455 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10458 InstParent->getInstList().insert(InsertPos, Result);
10460 // Make sure that we reprocess all operands now that we reduced their
10462 AddUsesToWorkList(*I);
10464 // Instructions can end up on the worklist more than once. Make sure
10465 // we do not process an instruction that has been deleted.
10466 RemoveFromWorkList(I);
10468 // Erase the old instruction.
10469 InstParent->getInstList().erase(I);
10472 DOUT << "IC: Mod = " << OrigI
10473 << " New = " << *I;
10476 // If the instruction was modified, it's possible that it is now dead.
10477 // if so, remove it.
10478 if (isInstructionTriviallyDead(I)) {
10479 // Make sure we process all operands now that we are reducing their
10481 AddUsesToWorkList(*I);
10483 // Instructions may end up in the worklist more than once. Erase all
10484 // occurrences of this instruction.
10485 RemoveFromWorkList(I);
10486 I->eraseFromParent();
10489 AddUsersToWorkList(*I);
10496 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10498 // Do an explicit clear, this shrinks the map if needed.
10499 WorklistMap.clear();
10504 bool InstCombiner::runOnFunction(Function &F) {
10505 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10507 bool EverMadeChange = false;
10509 // Iterate while there is work to do.
10510 unsigned Iteration = 0;
10511 while (DoOneIteration(F, Iteration++))
10512 EverMadeChange = true;
10513 return EverMadeChange;
10516 FunctionPass *llvm::createInstructionCombiningPass() {
10517 return new InstCombiner();