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())) {
2126 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2127 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2128 PointerType::get(Type::Int8Ty, AS), I);
2129 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2130 return new PtrToIntInst(I2, CI->getType());
2134 return Changed ? &I : 0;
2137 // isSignBit - Return true if the value represented by the constant only has the
2138 // highest order bit set.
2139 static bool isSignBit(ConstantInt *CI) {
2140 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2141 return CI->getValue() == APInt::getSignBit(NumBits);
2144 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2145 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2147 if (Op0 == Op1) // sub X, X -> 0
2148 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2150 // If this is a 'B = x-(-A)', change to B = x+A...
2151 if (Value *V = dyn_castNegVal(Op1))
2152 return BinaryOperator::createAdd(Op0, V);
2154 if (isa<UndefValue>(Op0))
2155 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2156 if (isa<UndefValue>(Op1))
2157 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2159 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2160 // Replace (-1 - A) with (~A)...
2161 if (C->isAllOnesValue())
2162 return BinaryOperator::createNot(Op1);
2164 // C - ~X == X + (1+C)
2166 if (match(Op1, m_Not(m_Value(X))))
2167 return BinaryOperator::createAdd(X, AddOne(C));
2169 // -(X >>u 31) -> (X >>s 31)
2170 // -(X >>s 31) -> (X >>u 31)
2172 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2173 if (SI->getOpcode() == Instruction::LShr) {
2174 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2175 // Check to see if we are shifting out everything but the sign bit.
2176 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2177 SI->getType()->getPrimitiveSizeInBits()-1) {
2178 // Ok, the transformation is safe. Insert AShr.
2179 return BinaryOperator::create(Instruction::AShr,
2180 SI->getOperand(0), CU, SI->getName());
2184 else if (SI->getOpcode() == Instruction::AShr) {
2185 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2186 // Check to see if we are shifting out everything but the sign bit.
2187 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2188 SI->getType()->getPrimitiveSizeInBits()-1) {
2189 // Ok, the transformation is safe. Insert LShr.
2190 return BinaryOperator::createLShr(
2191 SI->getOperand(0), CU, SI->getName());
2197 // Try to fold constant sub into select arguments.
2198 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2199 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2202 if (isa<PHINode>(Op0))
2203 if (Instruction *NV = FoldOpIntoPhi(I))
2207 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2208 if (Op1I->getOpcode() == Instruction::Add &&
2209 !Op0->getType()->isFPOrFPVector()) {
2210 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2211 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2212 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2213 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2214 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2215 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2216 // C1-(X+C2) --> (C1-C2)-X
2217 return BinaryOperator::createSub(Subtract(CI1, CI2),
2218 Op1I->getOperand(0));
2222 if (Op1I->hasOneUse()) {
2223 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2224 // is not used by anyone else...
2226 if (Op1I->getOpcode() == Instruction::Sub &&
2227 !Op1I->getType()->isFPOrFPVector()) {
2228 // Swap the two operands of the subexpr...
2229 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2230 Op1I->setOperand(0, IIOp1);
2231 Op1I->setOperand(1, IIOp0);
2233 // Create the new top level add instruction...
2234 return BinaryOperator::createAdd(Op0, Op1);
2237 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2239 if (Op1I->getOpcode() == Instruction::And &&
2240 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2241 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2244 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2245 return BinaryOperator::createAnd(Op0, NewNot);
2248 // 0 - (X sdiv C) -> (X sdiv -C)
2249 if (Op1I->getOpcode() == Instruction::SDiv)
2250 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2252 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2253 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2254 ConstantExpr::getNeg(DivRHS));
2256 // X - X*C --> X * (1-C)
2257 ConstantInt *C2 = 0;
2258 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2259 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2260 return BinaryOperator::createMul(Op0, CP1);
2263 // X - ((X / Y) * Y) --> X % Y
2264 if (Op1I->getOpcode() == Instruction::Mul)
2265 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2266 if (Op0 == I->getOperand(0) &&
2267 Op1I->getOperand(1) == I->getOperand(1)) {
2268 if (I->getOpcode() == Instruction::SDiv)
2269 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2270 if (I->getOpcode() == Instruction::UDiv)
2271 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2276 if (!Op0->getType()->isFPOrFPVector())
2277 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2278 if (Op0I->getOpcode() == Instruction::Add) {
2279 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2280 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2281 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2282 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2283 } else if (Op0I->getOpcode() == Instruction::Sub) {
2284 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2285 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2289 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2290 if (X == Op1) // X*C - X --> X * (C-1)
2291 return BinaryOperator::createMul(Op1, SubOne(C1));
2293 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2294 if (X == dyn_castFoldableMul(Op1, C2))
2295 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2300 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2301 /// comparison only checks the sign bit. If it only checks the sign bit, set
2302 /// TrueIfSigned if the result of the comparison is true when the input value is
2304 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2305 bool &TrueIfSigned) {
2307 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2308 TrueIfSigned = true;
2309 return RHS->isZero();
2310 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2311 TrueIfSigned = true;
2312 return RHS->isAllOnesValue();
2313 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2314 TrueIfSigned = false;
2315 return RHS->isAllOnesValue();
2316 case ICmpInst::ICMP_UGT:
2317 // True if LHS u> RHS and RHS == high-bit-mask - 1
2318 TrueIfSigned = true;
2319 return RHS->getValue() ==
2320 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2321 case ICmpInst::ICMP_UGE:
2322 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2323 TrueIfSigned = true;
2324 return RHS->getValue() ==
2325 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2331 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2332 bool Changed = SimplifyCommutative(I);
2333 Value *Op0 = I.getOperand(0);
2335 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2336 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2338 // Simplify mul instructions with a constant RHS...
2339 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2340 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2342 // ((X << C1)*C2) == (X * (C2 << C1))
2343 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2344 if (SI->getOpcode() == Instruction::Shl)
2345 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2346 return BinaryOperator::createMul(SI->getOperand(0),
2347 ConstantExpr::getShl(CI, ShOp));
2350 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2351 if (CI->equalsInt(1)) // X * 1 == X
2352 return ReplaceInstUsesWith(I, Op0);
2353 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2354 return BinaryOperator::createNeg(Op0, I.getName());
2356 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2357 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2358 return BinaryOperator::createShl(Op0,
2359 ConstantInt::get(Op0->getType(), Val.logBase2()));
2361 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2362 if (Op1F->isNullValue())
2363 return ReplaceInstUsesWith(I, Op1);
2365 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2366 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2367 // We need a better interface for long double here.
2368 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2369 if (Op1F->isExactlyValue(1.0))
2370 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2373 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2374 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2375 isa<ConstantInt>(Op0I->getOperand(1))) {
2376 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2377 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2379 InsertNewInstBefore(Add, I);
2380 Value *C1C2 = ConstantExpr::getMul(Op1,
2381 cast<Constant>(Op0I->getOperand(1)));
2382 return BinaryOperator::createAdd(Add, C1C2);
2386 // Try to fold constant mul into select arguments.
2387 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2388 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2391 if (isa<PHINode>(Op0))
2392 if (Instruction *NV = FoldOpIntoPhi(I))
2396 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2397 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2398 return BinaryOperator::createMul(Op0v, Op1v);
2400 // If one of the operands of the multiply is a cast from a boolean value, then
2401 // we know the bool is either zero or one, so this is a 'masking' multiply.
2402 // See if we can simplify things based on how the boolean was originally
2404 CastInst *BoolCast = 0;
2405 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2406 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2409 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2410 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2413 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2414 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2415 const Type *SCOpTy = SCIOp0->getType();
2418 // If the icmp is true iff the sign bit of X is set, then convert this
2419 // multiply into a shift/and combination.
2420 if (isa<ConstantInt>(SCIOp1) &&
2421 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2423 // Shift the X value right to turn it into "all signbits".
2424 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2425 SCOpTy->getPrimitiveSizeInBits()-1);
2427 InsertNewInstBefore(
2428 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2429 BoolCast->getOperand(0)->getName()+
2432 // If the multiply type is not the same as the source type, sign extend
2433 // or truncate to the multiply type.
2434 if (I.getType() != V->getType()) {
2435 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2436 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2437 Instruction::CastOps opcode =
2438 (SrcBits == DstBits ? Instruction::BitCast :
2439 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2440 V = InsertCastBefore(opcode, V, I.getType(), I);
2443 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2444 return BinaryOperator::createAnd(V, OtherOp);
2449 return Changed ? &I : 0;
2452 /// This function implements the transforms on div instructions that work
2453 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2454 /// used by the visitors to those instructions.
2455 /// @brief Transforms common to all three div instructions
2456 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2457 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2460 if (isa<UndefValue>(Op0))
2461 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2463 // X / undef -> undef
2464 if (isa<UndefValue>(Op1))
2465 return ReplaceInstUsesWith(I, Op1);
2467 // Handle cases involving: div X, (select Cond, Y, Z)
2468 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2469 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2470 // same basic block, then we replace the select with Y, and the condition
2471 // of the select with false (if the cond value is in the same BB). If the
2472 // select has uses other than the div, this allows them to be simplified
2473 // also. Note that div X, Y is just as good as div X, 0 (undef)
2474 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2475 if (ST->isNullValue()) {
2476 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2477 if (CondI && CondI->getParent() == I.getParent())
2478 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2479 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2480 I.setOperand(1, SI->getOperand(2));
2482 UpdateValueUsesWith(SI, SI->getOperand(2));
2486 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2487 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2488 if (ST->isNullValue()) {
2489 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2490 if (CondI && CondI->getParent() == I.getParent())
2491 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2492 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2493 I.setOperand(1, SI->getOperand(1));
2495 UpdateValueUsesWith(SI, SI->getOperand(1));
2503 /// This function implements the transforms common to both integer division
2504 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2505 /// division instructions.
2506 /// @brief Common integer divide transforms
2507 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2508 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2510 if (Instruction *Common = commonDivTransforms(I))
2513 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2515 if (RHS->equalsInt(1))
2516 return ReplaceInstUsesWith(I, Op0);
2518 // (X / C1) / C2 -> X / (C1*C2)
2519 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2520 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2521 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2522 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2523 Multiply(RHS, LHSRHS));
2526 if (!RHS->isZero()) { // avoid X udiv 0
2527 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2528 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2530 if (isa<PHINode>(Op0))
2531 if (Instruction *NV = FoldOpIntoPhi(I))
2536 // 0 / X == 0, we don't need to preserve faults!
2537 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2538 if (LHS->equalsInt(0))
2539 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2544 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2545 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2547 // Handle the integer div common cases
2548 if (Instruction *Common = commonIDivTransforms(I))
2551 // X udiv C^2 -> X >> C
2552 // Check to see if this is an unsigned division with an exact power of 2,
2553 // if so, convert to a right shift.
2554 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2555 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2556 return BinaryOperator::createLShr(Op0,
2557 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2560 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2561 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2562 if (RHSI->getOpcode() == Instruction::Shl &&
2563 isa<ConstantInt>(RHSI->getOperand(0))) {
2564 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2565 if (C1.isPowerOf2()) {
2566 Value *N = RHSI->getOperand(1);
2567 const Type *NTy = N->getType();
2568 if (uint32_t C2 = C1.logBase2()) {
2569 Constant *C2V = ConstantInt::get(NTy, C2);
2570 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2572 return BinaryOperator::createLShr(Op0, N);
2577 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2578 // where C1&C2 are powers of two.
2579 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2580 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2581 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2582 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2583 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2584 // Compute the shift amounts
2585 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2586 // Construct the "on true" case of the select
2587 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2588 Instruction *TSI = BinaryOperator::createLShr(
2589 Op0, TC, SI->getName()+".t");
2590 TSI = InsertNewInstBefore(TSI, I);
2592 // Construct the "on false" case of the select
2593 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2594 Instruction *FSI = BinaryOperator::createLShr(
2595 Op0, FC, SI->getName()+".f");
2596 FSI = InsertNewInstBefore(FSI, I);
2598 // construct the select instruction and return it.
2599 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2605 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2606 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2608 // Handle the integer div common cases
2609 if (Instruction *Common = commonIDivTransforms(I))
2612 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2614 if (RHS->isAllOnesValue())
2615 return BinaryOperator::createNeg(Op0);
2618 if (Value *LHSNeg = dyn_castNegVal(Op0))
2619 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2622 // If the sign bits of both operands are zero (i.e. we can prove they are
2623 // unsigned inputs), turn this into a udiv.
2624 if (I.getType()->isInteger()) {
2625 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2626 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2627 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2628 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2635 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2636 return commonDivTransforms(I);
2639 /// GetFactor - If we can prove that the specified value is at least a multiple
2640 /// of some factor, return that factor.
2641 static Constant *GetFactor(Value *V) {
2642 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2645 // Unless we can be tricky, we know this is a multiple of 1.
2646 Constant *Result = ConstantInt::get(V->getType(), 1);
2648 Instruction *I = dyn_cast<Instruction>(V);
2649 if (!I) return Result;
2651 if (I->getOpcode() == Instruction::Mul) {
2652 // Handle multiplies by a constant, etc.
2653 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2654 GetFactor(I->getOperand(1)));
2655 } else if (I->getOpcode() == Instruction::Shl) {
2656 // (X<<C) -> X * (1 << C)
2657 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2658 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2659 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2661 } else if (I->getOpcode() == Instruction::And) {
2662 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2663 // X & 0xFFF0 is known to be a multiple of 16.
2664 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2665 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2666 return ConstantExpr::getShl(Result,
2667 ConstantInt::get(Result->getType(), Zeros));
2669 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2670 // Only handle int->int casts.
2671 if (!CI->isIntegerCast())
2673 Value *Op = CI->getOperand(0);
2674 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2679 /// This function implements the transforms on rem instructions that work
2680 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2681 /// is used by the visitors to those instructions.
2682 /// @brief Transforms common to all three rem instructions
2683 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2684 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2686 // 0 % X == 0, we don't need to preserve faults!
2687 if (Constant *LHS = dyn_cast<Constant>(Op0))
2688 if (LHS->isNullValue())
2689 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2691 if (isa<UndefValue>(Op0)) // undef % X -> 0
2692 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2693 if (isa<UndefValue>(Op1))
2694 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2696 // Handle cases involving: rem X, (select Cond, Y, Z)
2697 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2698 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2699 // the same basic block, then we replace the select with Y, and the
2700 // condition of the select with false (if the cond value is in the same
2701 // BB). If the select has uses other than the div, this allows them to be
2703 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2704 if (ST->isNullValue()) {
2705 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2706 if (CondI && CondI->getParent() == I.getParent())
2707 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2708 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2709 I.setOperand(1, SI->getOperand(2));
2711 UpdateValueUsesWith(SI, SI->getOperand(2));
2714 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2715 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2716 if (ST->isNullValue()) {
2717 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2718 if (CondI && CondI->getParent() == I.getParent())
2719 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2720 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2721 I.setOperand(1, SI->getOperand(1));
2723 UpdateValueUsesWith(SI, SI->getOperand(1));
2731 /// This function implements the transforms common to both integer remainder
2732 /// instructions (urem and srem). It is called by the visitors to those integer
2733 /// remainder instructions.
2734 /// @brief Common integer remainder transforms
2735 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2736 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2738 if (Instruction *common = commonRemTransforms(I))
2741 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2742 // X % 0 == undef, we don't need to preserve faults!
2743 if (RHS->equalsInt(0))
2744 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2746 if (RHS->equalsInt(1)) // X % 1 == 0
2747 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2749 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2750 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2751 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2753 } else if (isa<PHINode>(Op0I)) {
2754 if (Instruction *NV = FoldOpIntoPhi(I))
2757 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2758 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2759 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2766 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2767 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2769 if (Instruction *common = commonIRemTransforms(I))
2772 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2773 // X urem C^2 -> X and C
2774 // Check to see if this is an unsigned remainder with an exact power of 2,
2775 // if so, convert to a bitwise and.
2776 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2777 if (C->getValue().isPowerOf2())
2778 return BinaryOperator::createAnd(Op0, SubOne(C));
2781 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2782 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2783 if (RHSI->getOpcode() == Instruction::Shl &&
2784 isa<ConstantInt>(RHSI->getOperand(0))) {
2785 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2786 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2787 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2789 return BinaryOperator::createAnd(Op0, Add);
2794 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2795 // where C1&C2 are powers of two.
2796 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2797 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2798 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2799 // STO == 0 and SFO == 0 handled above.
2800 if ((STO->getValue().isPowerOf2()) &&
2801 (SFO->getValue().isPowerOf2())) {
2802 Value *TrueAnd = InsertNewInstBefore(
2803 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2804 Value *FalseAnd = InsertNewInstBefore(
2805 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2806 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2814 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2815 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2817 // Handle the integer rem common cases
2818 if (Instruction *common = commonIRemTransforms(I))
2821 if (Value *RHSNeg = dyn_castNegVal(Op1))
2822 if (!isa<ConstantInt>(RHSNeg) ||
2823 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2825 AddUsesToWorkList(I);
2826 I.setOperand(1, RHSNeg);
2830 // If the sign bits of both operands are zero (i.e. we can prove they are
2831 // unsigned inputs), turn this into a urem.
2832 if (I.getType()->isInteger()) {
2833 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2834 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2835 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2836 return BinaryOperator::createURem(Op0, Op1, I.getName());
2843 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2844 return commonRemTransforms(I);
2847 // isMaxValueMinusOne - return true if this is Max-1
2848 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2849 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2851 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2852 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2855 // isMinValuePlusOne - return true if this is Min+1
2856 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2858 return C->getValue() == 1; // unsigned
2860 // Calculate 1111111111000000000000
2861 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2862 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2865 // isOneBitSet - Return true if there is exactly one bit set in the specified
2867 static bool isOneBitSet(const ConstantInt *CI) {
2868 return CI->getValue().isPowerOf2();
2871 // isHighOnes - Return true if the constant is of the form 1+0+.
2872 // This is the same as lowones(~X).
2873 static bool isHighOnes(const ConstantInt *CI) {
2874 return (~CI->getValue() + 1).isPowerOf2();
2877 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2878 /// are carefully arranged to allow folding of expressions such as:
2880 /// (A < B) | (A > B) --> (A != B)
2882 /// Note that this is only valid if the first and second predicates have the
2883 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2885 /// Three bits are used to represent the condition, as follows:
2890 /// <=> Value Definition
2891 /// 000 0 Always false
2898 /// 111 7 Always true
2900 static unsigned getICmpCode(const ICmpInst *ICI) {
2901 switch (ICI->getPredicate()) {
2903 case ICmpInst::ICMP_UGT: return 1; // 001
2904 case ICmpInst::ICMP_SGT: return 1; // 001
2905 case ICmpInst::ICMP_EQ: return 2; // 010
2906 case ICmpInst::ICMP_UGE: return 3; // 011
2907 case ICmpInst::ICMP_SGE: return 3; // 011
2908 case ICmpInst::ICMP_ULT: return 4; // 100
2909 case ICmpInst::ICMP_SLT: return 4; // 100
2910 case ICmpInst::ICMP_NE: return 5; // 101
2911 case ICmpInst::ICMP_ULE: return 6; // 110
2912 case ICmpInst::ICMP_SLE: return 6; // 110
2915 assert(0 && "Invalid ICmp predicate!");
2920 /// getICmpValue - This is the complement of getICmpCode, which turns an
2921 /// opcode and two operands into either a constant true or false, or a brand
2922 /// new ICmp instruction. The sign is passed in to determine which kind
2923 /// of predicate to use in new icmp instructions.
2924 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2926 default: assert(0 && "Illegal ICmp code!");
2927 case 0: return ConstantInt::getFalse();
2930 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2932 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2933 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2936 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2938 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2941 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2943 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2944 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2947 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2949 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2950 case 7: return ConstantInt::getTrue();
2954 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2955 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2956 (ICmpInst::isSignedPredicate(p1) &&
2957 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2958 (ICmpInst::isSignedPredicate(p2) &&
2959 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2963 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2964 struct FoldICmpLogical {
2967 ICmpInst::Predicate pred;
2968 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2969 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2970 pred(ICI->getPredicate()) {}
2971 bool shouldApply(Value *V) const {
2972 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2973 if (PredicatesFoldable(pred, ICI->getPredicate()))
2974 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2975 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2978 Instruction *apply(Instruction &Log) const {
2979 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2980 if (ICI->getOperand(0) != LHS) {
2981 assert(ICI->getOperand(1) == LHS);
2982 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2985 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2986 unsigned LHSCode = getICmpCode(ICI);
2987 unsigned RHSCode = getICmpCode(RHSICI);
2989 switch (Log.getOpcode()) {
2990 case Instruction::And: Code = LHSCode & RHSCode; break;
2991 case Instruction::Or: Code = LHSCode | RHSCode; break;
2992 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2993 default: assert(0 && "Illegal logical opcode!"); return 0;
2996 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2997 ICmpInst::isSignedPredicate(ICI->getPredicate());
2999 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3000 if (Instruction *I = dyn_cast<Instruction>(RV))
3002 // Otherwise, it's a constant boolean value...
3003 return IC.ReplaceInstUsesWith(Log, RV);
3006 } // end anonymous namespace
3008 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3009 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3010 // guaranteed to be a binary operator.
3011 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3013 ConstantInt *AndRHS,
3014 BinaryOperator &TheAnd) {
3015 Value *X = Op->getOperand(0);
3016 Constant *Together = 0;
3018 Together = And(AndRHS, OpRHS);
3020 switch (Op->getOpcode()) {
3021 case Instruction::Xor:
3022 if (Op->hasOneUse()) {
3023 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3024 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3025 InsertNewInstBefore(And, TheAnd);
3027 return BinaryOperator::createXor(And, Together);
3030 case Instruction::Or:
3031 if (Together == AndRHS) // (X | C) & C --> C
3032 return ReplaceInstUsesWith(TheAnd, AndRHS);
3034 if (Op->hasOneUse() && Together != OpRHS) {
3035 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3036 Instruction *Or = BinaryOperator::createOr(X, Together);
3037 InsertNewInstBefore(Or, TheAnd);
3039 return BinaryOperator::createAnd(Or, AndRHS);
3042 case Instruction::Add:
3043 if (Op->hasOneUse()) {
3044 // Adding a one to a single bit bit-field should be turned into an XOR
3045 // of the bit. First thing to check is to see if this AND is with a
3046 // single bit constant.
3047 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3049 // If there is only one bit set...
3050 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3051 // Ok, at this point, we know that we are masking the result of the
3052 // ADD down to exactly one bit. If the constant we are adding has
3053 // no bits set below this bit, then we can eliminate the ADD.
3054 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3056 // Check to see if any bits below the one bit set in AndRHSV are set.
3057 if ((AddRHS & (AndRHSV-1)) == 0) {
3058 // If not, the only thing that can effect the output of the AND is
3059 // the bit specified by AndRHSV. If that bit is set, the effect of
3060 // the XOR is to toggle the bit. If it is clear, then the ADD has
3062 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3063 TheAnd.setOperand(0, X);
3066 // Pull the XOR out of the AND.
3067 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3068 InsertNewInstBefore(NewAnd, TheAnd);
3069 NewAnd->takeName(Op);
3070 return BinaryOperator::createXor(NewAnd, AndRHS);
3077 case Instruction::Shl: {
3078 // We know that the AND will not produce any of the bits shifted in, so if
3079 // the anded constant includes them, clear them now!
3081 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3082 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3083 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3084 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3086 if (CI->getValue() == ShlMask) {
3087 // Masking out bits that the shift already masks
3088 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3089 } else if (CI != AndRHS) { // Reducing bits set in and.
3090 TheAnd.setOperand(1, CI);
3095 case Instruction::LShr:
3097 // We know that the AND will not produce any of the bits shifted in, so if
3098 // the anded constant includes them, clear them now! This only applies to
3099 // unsigned shifts, because a signed shr may bring in set bits!
3101 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3102 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3103 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3104 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3106 if (CI->getValue() == ShrMask) {
3107 // Masking out bits that the shift already masks.
3108 return ReplaceInstUsesWith(TheAnd, Op);
3109 } else if (CI != AndRHS) {
3110 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3115 case Instruction::AShr:
3117 // See if this is shifting in some sign extension, then masking it out
3119 if (Op->hasOneUse()) {
3120 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3121 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3122 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3123 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3124 if (C == AndRHS) { // Masking out bits shifted in.
3125 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3126 // Make the argument unsigned.
3127 Value *ShVal = Op->getOperand(0);
3128 ShVal = InsertNewInstBefore(
3129 BinaryOperator::createLShr(ShVal, OpRHS,
3130 Op->getName()), TheAnd);
3131 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3140 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3141 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3142 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3143 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3144 /// insert new instructions.
3145 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3146 bool isSigned, bool Inside,
3148 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3149 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3150 "Lo is not <= Hi in range emission code!");
3153 if (Lo == Hi) // Trivially false.
3154 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3156 // V >= Min && V < Hi --> V < Hi
3157 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3158 ICmpInst::Predicate pred = (isSigned ?
3159 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3160 return new ICmpInst(pred, V, Hi);
3163 // Emit V-Lo <u Hi-Lo
3164 Constant *NegLo = ConstantExpr::getNeg(Lo);
3165 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3166 InsertNewInstBefore(Add, IB);
3167 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3168 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3171 if (Lo == Hi) // Trivially true.
3172 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3174 // V < Min || V >= Hi -> V > Hi-1
3175 Hi = SubOne(cast<ConstantInt>(Hi));
3176 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3177 ICmpInst::Predicate pred = (isSigned ?
3178 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3179 return new ICmpInst(pred, V, Hi);
3182 // Emit V-Lo >u Hi-1-Lo
3183 // Note that Hi has already had one subtracted from it, above.
3184 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3185 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3186 InsertNewInstBefore(Add, IB);
3187 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3188 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3191 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3192 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3193 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3194 // not, since all 1s are not contiguous.
3195 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3196 const APInt& V = Val->getValue();
3197 uint32_t BitWidth = Val->getType()->getBitWidth();
3198 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3200 // look for the first zero bit after the run of ones
3201 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3202 // look for the first non-zero bit
3203 ME = V.getActiveBits();
3207 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3208 /// where isSub determines whether the operator is a sub. If we can fold one of
3209 /// the following xforms:
3211 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3212 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3213 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3215 /// return (A +/- B).
3217 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3218 ConstantInt *Mask, bool isSub,
3220 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3221 if (!LHSI || LHSI->getNumOperands() != 2 ||
3222 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3224 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3226 switch (LHSI->getOpcode()) {
3228 case Instruction::And:
3229 if (And(N, Mask) == Mask) {
3230 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3231 if ((Mask->getValue().countLeadingZeros() +
3232 Mask->getValue().countPopulation()) ==
3233 Mask->getValue().getBitWidth())
3236 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3237 // part, we don't need any explicit masks to take them out of A. If that
3238 // is all N is, ignore it.
3239 uint32_t MB = 0, ME = 0;
3240 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3241 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3242 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3243 if (MaskedValueIsZero(RHS, Mask))
3248 case Instruction::Or:
3249 case Instruction::Xor:
3250 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3251 if ((Mask->getValue().countLeadingZeros() +
3252 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3253 && And(N, Mask)->isZero())
3260 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3262 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3263 return InsertNewInstBefore(New, I);
3266 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3267 bool Changed = SimplifyCommutative(I);
3268 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3270 if (isa<UndefValue>(Op1)) // X & undef -> 0
3271 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3275 return ReplaceInstUsesWith(I, Op1);
3277 // See if we can simplify any instructions used by the instruction whose sole
3278 // purpose is to compute bits we don't care about.
3279 if (!isa<VectorType>(I.getType())) {
3280 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3281 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3282 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3283 KnownZero, KnownOne))
3286 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3287 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3288 return ReplaceInstUsesWith(I, I.getOperand(0));
3289 } else if (isa<ConstantAggregateZero>(Op1)) {
3290 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3294 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3295 const APInt& AndRHSMask = AndRHS->getValue();
3296 APInt NotAndRHS(~AndRHSMask);
3298 // Optimize a variety of ((val OP C1) & C2) combinations...
3299 if (isa<BinaryOperator>(Op0)) {
3300 Instruction *Op0I = cast<Instruction>(Op0);
3301 Value *Op0LHS = Op0I->getOperand(0);
3302 Value *Op0RHS = Op0I->getOperand(1);
3303 switch (Op0I->getOpcode()) {
3304 case Instruction::Xor:
3305 case Instruction::Or:
3306 // If the mask is only needed on one incoming arm, push it up.
3307 if (Op0I->hasOneUse()) {
3308 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3309 // Not masking anything out for the LHS, move to RHS.
3310 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3311 Op0RHS->getName()+".masked");
3312 InsertNewInstBefore(NewRHS, I);
3313 return BinaryOperator::create(
3314 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3316 if (!isa<Constant>(Op0RHS) &&
3317 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3318 // Not masking anything out for the RHS, move to LHS.
3319 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3320 Op0LHS->getName()+".masked");
3321 InsertNewInstBefore(NewLHS, I);
3322 return BinaryOperator::create(
3323 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3328 case Instruction::Add:
3329 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3330 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3331 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3332 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3333 return BinaryOperator::createAnd(V, AndRHS);
3334 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3335 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3338 case Instruction::Sub:
3339 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3340 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3341 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3342 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3343 return BinaryOperator::createAnd(V, AndRHS);
3347 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3348 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3350 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3351 // If this is an integer truncation or change from signed-to-unsigned, and
3352 // if the source is an and/or with immediate, transform it. This
3353 // frequently occurs for bitfield accesses.
3354 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3355 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3356 CastOp->getNumOperands() == 2)
3357 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3358 if (CastOp->getOpcode() == Instruction::And) {
3359 // Change: and (cast (and X, C1) to T), C2
3360 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3361 // This will fold the two constants together, which may allow
3362 // other simplifications.
3363 Instruction *NewCast = CastInst::createTruncOrBitCast(
3364 CastOp->getOperand(0), I.getType(),
3365 CastOp->getName()+".shrunk");
3366 NewCast = InsertNewInstBefore(NewCast, I);
3367 // trunc_or_bitcast(C1)&C2
3368 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3369 C3 = ConstantExpr::getAnd(C3, AndRHS);
3370 return BinaryOperator::createAnd(NewCast, C3);
3371 } else if (CastOp->getOpcode() == Instruction::Or) {
3372 // Change: and (cast (or X, C1) to T), C2
3373 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3374 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3375 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3376 return ReplaceInstUsesWith(I, AndRHS);
3381 // Try to fold constant and into select arguments.
3382 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3383 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3385 if (isa<PHINode>(Op0))
3386 if (Instruction *NV = FoldOpIntoPhi(I))
3390 Value *Op0NotVal = dyn_castNotVal(Op0);
3391 Value *Op1NotVal = dyn_castNotVal(Op1);
3393 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3394 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3396 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3397 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3398 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3399 I.getName()+".demorgan");
3400 InsertNewInstBefore(Or, I);
3401 return BinaryOperator::createNot(Or);
3405 Value *A = 0, *B = 0, *C = 0, *D = 0;
3406 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3407 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3408 return ReplaceInstUsesWith(I, Op1);
3410 // (A|B) & ~(A&B) -> A^B
3411 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3412 if ((A == C && B == D) || (A == D && B == C))
3413 return BinaryOperator::createXor(A, B);
3417 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3418 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3419 return ReplaceInstUsesWith(I, Op0);
3421 // ~(A&B) & (A|B) -> A^B
3422 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3423 if ((A == C && B == D) || (A == D && B == C))
3424 return BinaryOperator::createXor(A, B);
3428 if (Op0->hasOneUse() &&
3429 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3430 if (A == Op1) { // (A^B)&A -> A&(A^B)
3431 I.swapOperands(); // Simplify below
3432 std::swap(Op0, Op1);
3433 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3434 cast<BinaryOperator>(Op0)->swapOperands();
3435 I.swapOperands(); // Simplify below
3436 std::swap(Op0, Op1);
3439 if (Op1->hasOneUse() &&
3440 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3441 if (B == Op0) { // B&(A^B) -> B&(B^A)
3442 cast<BinaryOperator>(Op1)->swapOperands();
3445 if (A == Op0) { // A&(A^B) -> A & ~B
3446 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3447 InsertNewInstBefore(NotB, I);
3448 return BinaryOperator::createAnd(A, NotB);
3453 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3454 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3455 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3458 Value *LHSVal, *RHSVal;
3459 ConstantInt *LHSCst, *RHSCst;
3460 ICmpInst::Predicate LHSCC, RHSCC;
3461 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3462 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3463 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3464 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3465 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3466 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3467 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3468 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3470 // Don't try to fold ICMP_SLT + ICMP_ULT.
3471 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3472 ICmpInst::isSignedPredicate(LHSCC) ==
3473 ICmpInst::isSignedPredicate(RHSCC))) {
3474 // Ensure that the larger constant is on the RHS.
3475 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3476 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3477 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3478 ICmpInst *LHS = cast<ICmpInst>(Op0);
3479 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3480 std::swap(LHS, RHS);
3481 std::swap(LHSCst, RHSCst);
3482 std::swap(LHSCC, RHSCC);
3485 // At this point, we know we have have two icmp instructions
3486 // comparing a value against two constants and and'ing the result
3487 // together. Because of the above check, we know that we only have
3488 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3489 // (from the FoldICmpLogical check above), that the two constants
3490 // are not equal and that the larger constant is on the RHS
3491 assert(LHSCst != RHSCst && "Compares not folded above?");
3494 default: assert(0 && "Unknown integer condition code!");
3495 case ICmpInst::ICMP_EQ:
3497 default: assert(0 && "Unknown integer condition code!");
3498 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3499 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3500 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3501 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3502 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3503 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3504 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3505 return ReplaceInstUsesWith(I, LHS);
3507 case ICmpInst::ICMP_NE:
3509 default: assert(0 && "Unknown integer condition code!");
3510 case ICmpInst::ICMP_ULT:
3511 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3512 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3513 break; // (X != 13 & X u< 15) -> no change
3514 case ICmpInst::ICMP_SLT:
3515 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3516 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3517 break; // (X != 13 & X s< 15) -> no change
3518 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3519 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3520 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3521 return ReplaceInstUsesWith(I, RHS);
3522 case ICmpInst::ICMP_NE:
3523 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3524 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3525 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3526 LHSVal->getName()+".off");
3527 InsertNewInstBefore(Add, I);
3528 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3529 ConstantInt::get(Add->getType(), 1));
3531 break; // (X != 13 & X != 15) -> no change
3534 case ICmpInst::ICMP_ULT:
3536 default: assert(0 && "Unknown integer condition code!");
3537 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3538 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3539 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3540 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3542 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3543 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3544 return ReplaceInstUsesWith(I, LHS);
3545 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3549 case ICmpInst::ICMP_SLT:
3551 default: assert(0 && "Unknown integer condition code!");
3552 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3553 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3554 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3555 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3557 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3558 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3559 return ReplaceInstUsesWith(I, LHS);
3560 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3564 case ICmpInst::ICMP_UGT:
3566 default: assert(0 && "Unknown integer condition code!");
3567 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3568 return ReplaceInstUsesWith(I, LHS);
3569 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3570 return ReplaceInstUsesWith(I, RHS);
3571 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3573 case ICmpInst::ICMP_NE:
3574 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3575 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3576 break; // (X u> 13 & X != 15) -> no change
3577 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3578 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3580 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3584 case ICmpInst::ICMP_SGT:
3586 default: assert(0 && "Unknown integer condition code!");
3587 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3588 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3589 return ReplaceInstUsesWith(I, RHS);
3590 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3592 case ICmpInst::ICMP_NE:
3593 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3594 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3595 break; // (X s> 13 & X != 15) -> no change
3596 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3597 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3599 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3607 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3608 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3609 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3610 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3611 const Type *SrcTy = Op0C->getOperand(0)->getType();
3612 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3613 // Only do this if the casts both really cause code to be generated.
3614 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3616 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3618 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3619 Op1C->getOperand(0),
3621 InsertNewInstBefore(NewOp, I);
3622 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3626 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3627 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3628 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3629 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3630 SI0->getOperand(1) == SI1->getOperand(1) &&
3631 (SI0->hasOneUse() || SI1->hasOneUse())) {
3632 Instruction *NewOp =
3633 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3635 SI0->getName()), I);
3636 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3637 SI1->getOperand(1));
3641 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3642 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3643 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3644 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3645 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3646 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3647 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3648 // If either of the constants are nans, then the whole thing returns
3650 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3651 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3652 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3653 RHS->getOperand(0));
3658 return Changed ? &I : 0;
3661 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3662 /// in the result. If it does, and if the specified byte hasn't been filled in
3663 /// yet, fill it in and return false.
3664 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3665 Instruction *I = dyn_cast<Instruction>(V);
3666 if (I == 0) return true;
3668 // If this is an or instruction, it is an inner node of the bswap.
3669 if (I->getOpcode() == Instruction::Or)
3670 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3671 CollectBSwapParts(I->getOperand(1), ByteValues);
3673 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3674 // If this is a shift by a constant int, and it is "24", then its operand
3675 // defines a byte. We only handle unsigned types here.
3676 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3677 // Not shifting the entire input by N-1 bytes?
3678 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3679 8*(ByteValues.size()-1))
3683 if (I->getOpcode() == Instruction::Shl) {
3684 // X << 24 defines the top byte with the lowest of the input bytes.
3685 DestNo = ByteValues.size()-1;
3687 // X >>u 24 defines the low byte with the highest of the input bytes.
3691 // If the destination byte value is already defined, the values are or'd
3692 // together, which isn't a bswap (unless it's an or of the same bits).
3693 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3695 ByteValues[DestNo] = I->getOperand(0);
3699 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3701 Value *Shift = 0, *ShiftLHS = 0;
3702 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3703 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3704 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3706 Instruction *SI = cast<Instruction>(Shift);
3708 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3709 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3710 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3713 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3715 if (AndAmt->getValue().getActiveBits() > 64)
3717 uint64_t AndAmtVal = AndAmt->getZExtValue();
3718 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3719 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3721 // Unknown mask for bswap.
3722 if (DestByte == ByteValues.size()) return true;
3724 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3726 if (SI->getOpcode() == Instruction::Shl)
3727 SrcByte = DestByte - ShiftBytes;
3729 SrcByte = DestByte + ShiftBytes;
3731 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3732 if (SrcByte != ByteValues.size()-DestByte-1)
3735 // If the destination byte value is already defined, the values are or'd
3736 // together, which isn't a bswap (unless it's an or of the same bits).
3737 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3739 ByteValues[DestByte] = SI->getOperand(0);
3743 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3744 /// If so, insert the new bswap intrinsic and return it.
3745 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3746 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3747 if (!ITy || ITy->getBitWidth() % 16)
3748 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3750 /// ByteValues - For each byte of the result, we keep track of which value
3751 /// defines each byte.
3752 SmallVector<Value*, 8> ByteValues;
3753 ByteValues.resize(ITy->getBitWidth()/8);
3755 // Try to find all the pieces corresponding to the bswap.
3756 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3757 CollectBSwapParts(I.getOperand(1), ByteValues))
3760 // Check to see if all of the bytes come from the same value.
3761 Value *V = ByteValues[0];
3762 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3764 // Check to make sure that all of the bytes come from the same value.
3765 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3766 if (ByteValues[i] != V)
3768 const Type *Tys[] = { ITy };
3769 Module *M = I.getParent()->getParent()->getParent();
3770 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3771 return new CallInst(F, V);
3775 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3776 bool Changed = SimplifyCommutative(I);
3777 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3779 if (isa<UndefValue>(Op1)) // X | undef -> -1
3780 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3784 return ReplaceInstUsesWith(I, Op0);
3786 // See if we can simplify any instructions used by the instruction whose sole
3787 // purpose is to compute bits we don't care about.
3788 if (!isa<VectorType>(I.getType())) {
3789 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3790 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3791 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3792 KnownZero, KnownOne))
3794 } else if (isa<ConstantAggregateZero>(Op1)) {
3795 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3796 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3797 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3798 return ReplaceInstUsesWith(I, I.getOperand(1));
3804 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3805 ConstantInt *C1 = 0; Value *X = 0;
3806 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3807 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3808 Instruction *Or = BinaryOperator::createOr(X, RHS);
3809 InsertNewInstBefore(Or, I);
3811 return BinaryOperator::createAnd(Or,
3812 ConstantInt::get(RHS->getValue() | C1->getValue()));
3815 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3816 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3817 Instruction *Or = BinaryOperator::createOr(X, RHS);
3818 InsertNewInstBefore(Or, I);
3820 return BinaryOperator::createXor(Or,
3821 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3824 // Try to fold constant and into select arguments.
3825 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3826 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3828 if (isa<PHINode>(Op0))
3829 if (Instruction *NV = FoldOpIntoPhi(I))
3833 Value *A = 0, *B = 0;
3834 ConstantInt *C1 = 0, *C2 = 0;
3836 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3837 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3838 return ReplaceInstUsesWith(I, Op1);
3839 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3840 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3841 return ReplaceInstUsesWith(I, Op0);
3843 // (A | B) | C and A | (B | C) -> bswap if possible.
3844 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3845 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3846 match(Op1, m_Or(m_Value(), m_Value())) ||
3847 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3848 match(Op1, m_Shift(m_Value(), m_Value())))) {
3849 if (Instruction *BSwap = MatchBSwap(I))
3853 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3854 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3855 MaskedValueIsZero(Op1, C1->getValue())) {
3856 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3857 InsertNewInstBefore(NOr, I);
3859 return BinaryOperator::createXor(NOr, C1);
3862 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3863 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3864 MaskedValueIsZero(Op0, C1->getValue())) {
3865 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3866 InsertNewInstBefore(NOr, I);
3868 return BinaryOperator::createXor(NOr, C1);
3872 Value *C = 0, *D = 0;
3873 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3874 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3875 Value *V1 = 0, *V2 = 0, *V3 = 0;
3876 C1 = dyn_cast<ConstantInt>(C);
3877 C2 = dyn_cast<ConstantInt>(D);
3878 if (C1 && C2) { // (A & C1)|(B & C2)
3879 // If we have: ((V + N) & C1) | (V & C2)
3880 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3881 // replace with V+N.
3882 if (C1->getValue() == ~C2->getValue()) {
3883 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3884 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3885 // Add commutes, try both ways.
3886 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3887 return ReplaceInstUsesWith(I, A);
3888 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3889 return ReplaceInstUsesWith(I, A);
3891 // Or commutes, try both ways.
3892 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3893 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3894 // Add commutes, try both ways.
3895 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3896 return ReplaceInstUsesWith(I, B);
3897 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3898 return ReplaceInstUsesWith(I, B);
3901 V1 = 0; V2 = 0; V3 = 0;
3904 // Check to see if we have any common things being and'ed. If so, find the
3905 // terms for V1 & (V2|V3).
3906 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3907 if (A == B) // (A & C)|(A & D) == A & (C|D)
3908 V1 = A, V2 = C, V3 = D;
3909 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3910 V1 = A, V2 = B, V3 = C;
3911 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3912 V1 = C, V2 = A, V3 = D;
3913 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3914 V1 = C, V2 = A, V3 = B;
3918 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3919 return BinaryOperator::createAnd(V1, Or);
3924 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3925 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3926 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3927 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3928 SI0->getOperand(1) == SI1->getOperand(1) &&
3929 (SI0->hasOneUse() || SI1->hasOneUse())) {
3930 Instruction *NewOp =
3931 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3933 SI0->getName()), I);
3934 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3935 SI1->getOperand(1));
3939 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3940 if (A == Op1) // ~A | A == -1
3941 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3945 // Note, A is still live here!
3946 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3948 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3950 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3951 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3952 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3953 I.getName()+".demorgan"), I);
3954 return BinaryOperator::createNot(And);
3958 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3959 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3960 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3963 Value *LHSVal, *RHSVal;
3964 ConstantInt *LHSCst, *RHSCst;
3965 ICmpInst::Predicate LHSCC, RHSCC;
3966 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3967 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3968 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3969 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3970 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3971 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3972 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3973 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3974 // We can't fold (ugt x, C) | (sgt x, C2).
3975 PredicatesFoldable(LHSCC, RHSCC)) {
3976 // Ensure that the larger constant is on the RHS.
3977 ICmpInst *LHS = cast<ICmpInst>(Op0);
3979 if (ICmpInst::isSignedPredicate(LHSCC))
3980 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3982 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3985 std::swap(LHS, RHS);
3986 std::swap(LHSCst, RHSCst);
3987 std::swap(LHSCC, RHSCC);
3990 // At this point, we know we have have two icmp instructions
3991 // comparing a value against two constants and or'ing the result
3992 // together. Because of the above check, we know that we only have
3993 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3994 // FoldICmpLogical check above), that the two constants are not
3996 assert(LHSCst != RHSCst && "Compares not folded above?");
3999 default: assert(0 && "Unknown integer condition code!");
4000 case ICmpInst::ICMP_EQ:
4002 default: assert(0 && "Unknown integer condition code!");
4003 case ICmpInst::ICMP_EQ:
4004 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4005 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4006 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4007 LHSVal->getName()+".off");
4008 InsertNewInstBefore(Add, I);
4009 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4010 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4012 break; // (X == 13 | X == 15) -> no change
4013 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4014 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4016 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4017 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4018 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4019 return ReplaceInstUsesWith(I, RHS);
4022 case ICmpInst::ICMP_NE:
4024 default: assert(0 && "Unknown integer condition code!");
4025 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4026 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4027 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4028 return ReplaceInstUsesWith(I, LHS);
4029 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4030 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4031 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4032 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4035 case ICmpInst::ICMP_ULT:
4037 default: assert(0 && "Unknown integer condition code!");
4038 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4040 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4041 // If RHSCst is [us]MAXINT, it is always false. Not handling
4042 // this can cause overflow.
4043 if (RHSCst->isMaxValue(false))
4044 return ReplaceInstUsesWith(I, LHS);
4045 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4047 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4049 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4050 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4051 return ReplaceInstUsesWith(I, RHS);
4052 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4056 case ICmpInst::ICMP_SLT:
4058 default: assert(0 && "Unknown integer condition code!");
4059 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4061 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4062 // If RHSCst is [us]MAXINT, it is always false. Not handling
4063 // this can cause overflow.
4064 if (RHSCst->isMaxValue(true))
4065 return ReplaceInstUsesWith(I, LHS);
4066 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4068 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4070 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4071 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4072 return ReplaceInstUsesWith(I, RHS);
4073 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4077 case ICmpInst::ICMP_UGT:
4079 default: assert(0 && "Unknown integer condition code!");
4080 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4081 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4082 return ReplaceInstUsesWith(I, LHS);
4083 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4085 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4086 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4087 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4088 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4092 case ICmpInst::ICMP_SGT:
4094 default: assert(0 && "Unknown integer condition code!");
4095 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4096 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4097 return ReplaceInstUsesWith(I, LHS);
4098 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4100 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4101 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4102 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4103 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4111 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4112 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4113 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4114 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4115 const Type *SrcTy = Op0C->getOperand(0)->getType();
4116 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4117 // Only do this if the casts both really cause code to be generated.
4118 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4120 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4122 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4123 Op1C->getOperand(0),
4125 InsertNewInstBefore(NewOp, I);
4126 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4132 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4133 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4134 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4135 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4136 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4137 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4138 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4139 // If either of the constants are nans, then the whole thing returns
4141 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4142 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4144 // Otherwise, no need to compare the two constants, compare the
4146 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4147 RHS->getOperand(0));
4152 return Changed ? &I : 0;
4155 // XorSelf - Implements: X ^ X --> 0
4158 XorSelf(Value *rhs) : RHS(rhs) {}
4159 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4160 Instruction *apply(BinaryOperator &Xor) const {
4166 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4167 bool Changed = SimplifyCommutative(I);
4168 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4170 if (isa<UndefValue>(Op1))
4171 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4173 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4174 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4175 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4176 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4179 // See if we can simplify any instructions used by the instruction whose sole
4180 // purpose is to compute bits we don't care about.
4181 if (!isa<VectorType>(I.getType())) {
4182 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4183 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4184 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4185 KnownZero, KnownOne))
4187 } else if (isa<ConstantAggregateZero>(Op1)) {
4188 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4191 // Is this a ~ operation?
4192 if (Value *NotOp = dyn_castNotVal(&I)) {
4193 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4194 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4195 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4196 if (Op0I->getOpcode() == Instruction::And ||
4197 Op0I->getOpcode() == Instruction::Or) {
4198 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4199 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4201 BinaryOperator::createNot(Op0I->getOperand(1),
4202 Op0I->getOperand(1)->getName()+".not");
4203 InsertNewInstBefore(NotY, I);
4204 if (Op0I->getOpcode() == Instruction::And)
4205 return BinaryOperator::createOr(Op0NotVal, NotY);
4207 return BinaryOperator::createAnd(Op0NotVal, NotY);
4214 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4215 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4216 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4217 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4218 return new ICmpInst(ICI->getInversePredicate(),
4219 ICI->getOperand(0), ICI->getOperand(1));
4221 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4222 return new FCmpInst(FCI->getInversePredicate(),
4223 FCI->getOperand(0), FCI->getOperand(1));
4226 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4227 // ~(c-X) == X-c-1 == X+(-c-1)
4228 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4229 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4230 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4231 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4232 ConstantInt::get(I.getType(), 1));
4233 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4236 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4237 if (Op0I->getOpcode() == Instruction::Add) {
4238 // ~(X-c) --> (-c-1)-X
4239 if (RHS->isAllOnesValue()) {
4240 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4241 return BinaryOperator::createSub(
4242 ConstantExpr::getSub(NegOp0CI,
4243 ConstantInt::get(I.getType(), 1)),
4244 Op0I->getOperand(0));
4245 } else if (RHS->getValue().isSignBit()) {
4246 // (X + C) ^ signbit -> (X + C + signbit)
4247 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4248 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4251 } else if (Op0I->getOpcode() == Instruction::Or) {
4252 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4253 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4254 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4255 // Anything in both C1 and C2 is known to be zero, remove it from
4257 Constant *CommonBits = And(Op0CI, RHS);
4258 NewRHS = ConstantExpr::getAnd(NewRHS,
4259 ConstantExpr::getNot(CommonBits));
4260 AddToWorkList(Op0I);
4261 I.setOperand(0, Op0I->getOperand(0));
4262 I.setOperand(1, NewRHS);
4268 // Try to fold constant and into select arguments.
4269 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4270 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4272 if (isa<PHINode>(Op0))
4273 if (Instruction *NV = FoldOpIntoPhi(I))
4277 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4279 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4281 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4283 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4286 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4289 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4290 if (A == Op0) { // B^(B|A) == (A|B)^B
4291 Op1I->swapOperands();
4293 std::swap(Op0, Op1);
4294 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4295 I.swapOperands(); // Simplified below.
4296 std::swap(Op0, Op1);
4298 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4299 if (Op0 == A) // A^(A^B) == B
4300 return ReplaceInstUsesWith(I, B);
4301 else if (Op0 == B) // A^(B^A) == B
4302 return ReplaceInstUsesWith(I, A);
4303 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4304 if (A == Op0) { // A^(A&B) -> A^(B&A)
4305 Op1I->swapOperands();
4308 if (B == Op0) { // A^(B&A) -> (B&A)^A
4309 I.swapOperands(); // Simplified below.
4310 std::swap(Op0, Op1);
4315 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4318 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4319 if (A == Op1) // (B|A)^B == (A|B)^B
4321 if (B == Op1) { // (A|B)^B == A & ~B
4323 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4324 return BinaryOperator::createAnd(A, NotB);
4326 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4327 if (Op1 == A) // (A^B)^A == B
4328 return ReplaceInstUsesWith(I, B);
4329 else if (Op1 == B) // (B^A)^A == B
4330 return ReplaceInstUsesWith(I, A);
4331 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4332 if (A == Op1) // (A&B)^A -> (B&A)^A
4334 if (B == Op1 && // (B&A)^A == ~B & A
4335 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4337 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4338 return BinaryOperator::createAnd(N, Op1);
4343 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4344 if (Op0I && Op1I && Op0I->isShift() &&
4345 Op0I->getOpcode() == Op1I->getOpcode() &&
4346 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4347 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4348 Instruction *NewOp =
4349 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4350 Op1I->getOperand(0),
4351 Op0I->getName()), I);
4352 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4353 Op1I->getOperand(1));
4357 Value *A, *B, *C, *D;
4358 // (A & B)^(A | B) -> A ^ B
4359 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4360 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4361 if ((A == C && B == D) || (A == D && B == C))
4362 return BinaryOperator::createXor(A, B);
4364 // (A | B)^(A & B) -> A ^ B
4365 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4366 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4367 if ((A == C && B == D) || (A == D && B == C))
4368 return BinaryOperator::createXor(A, B);
4372 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4373 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4374 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4375 // (X & Y)^(X & Y) -> (Y^Z) & X
4376 Value *X = 0, *Y = 0, *Z = 0;
4378 X = A, Y = B, Z = D;
4380 X = A, Y = B, Z = C;
4382 X = B, Y = A, Z = D;
4384 X = B, Y = A, Z = C;
4387 Instruction *NewOp =
4388 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4389 return BinaryOperator::createAnd(NewOp, X);
4394 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4395 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4396 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4399 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4400 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4401 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4402 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4403 const Type *SrcTy = Op0C->getOperand(0)->getType();
4404 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4405 // Only do this if the casts both really cause code to be generated.
4406 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4408 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4410 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4411 Op1C->getOperand(0),
4413 InsertNewInstBefore(NewOp, I);
4414 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4418 return Changed ? &I : 0;
4421 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4422 /// overflowed for this type.
4423 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4424 ConstantInt *In2, bool IsSigned = false) {
4425 Result = cast<ConstantInt>(Add(In1, In2));
4428 if (In2->getValue().isNegative())
4429 return Result->getValue().sgt(In1->getValue());
4431 return Result->getValue().slt(In1->getValue());
4433 return Result->getValue().ult(In1->getValue());
4436 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4437 /// code necessary to compute the offset from the base pointer (without adding
4438 /// in the base pointer). Return the result as a signed integer of intptr size.
4439 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4440 TargetData &TD = IC.getTargetData();
4441 gep_type_iterator GTI = gep_type_begin(GEP);
4442 const Type *IntPtrTy = TD.getIntPtrType();
4443 Value *Result = Constant::getNullValue(IntPtrTy);
4445 // Build a mask for high order bits.
4446 unsigned IntPtrWidth = TD.getPointerSize()*8;
4447 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4449 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4450 Value *Op = GEP->getOperand(i);
4451 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4452 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4453 if (OpC->isZero()) continue;
4455 // Handle a struct index, which adds its field offset to the pointer.
4456 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4457 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4459 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4460 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4462 Result = IC.InsertNewInstBefore(
4463 BinaryOperator::createAdd(Result,
4464 ConstantInt::get(IntPtrTy, Size),
4465 GEP->getName()+".offs"), I);
4469 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4470 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4471 Scale = ConstantExpr::getMul(OC, Scale);
4472 if (Constant *RC = dyn_cast<Constant>(Result))
4473 Result = ConstantExpr::getAdd(RC, Scale);
4475 // Emit an add instruction.
4476 Result = IC.InsertNewInstBefore(
4477 BinaryOperator::createAdd(Result, Scale,
4478 GEP->getName()+".offs"), I);
4482 // Convert to correct type.
4483 if (Op->getType() != IntPtrTy) {
4484 if (Constant *OpC = dyn_cast<Constant>(Op))
4485 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4487 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4488 Op->getName()+".c"), I);
4491 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4492 if (Constant *OpC = dyn_cast<Constant>(Op))
4493 Op = ConstantExpr::getMul(OpC, Scale);
4494 else // We'll let instcombine(mul) convert this to a shl if possible.
4495 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4496 GEP->getName()+".idx"), I);
4499 // Emit an add instruction.
4500 if (isa<Constant>(Op) && isa<Constant>(Result))
4501 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4502 cast<Constant>(Result));
4504 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4505 GEP->getName()+".offs"), I);
4510 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4511 /// else. At this point we know that the GEP is on the LHS of the comparison.
4512 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4513 ICmpInst::Predicate Cond,
4515 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4517 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4518 if (isa<PointerType>(CI->getOperand(0)->getType()))
4519 RHS = CI->getOperand(0);
4521 Value *PtrBase = GEPLHS->getOperand(0);
4522 if (PtrBase == RHS) {
4523 // As an optimization, we don't actually have to compute the actual value of
4524 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4525 // each index is zero or not.
4526 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4527 Instruction *InVal = 0;
4528 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4529 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4531 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4532 if (isa<UndefValue>(C)) // undef index -> undef.
4533 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4534 if (C->isNullValue())
4536 else if (TD->getABITypeSize(GTI.getIndexedType()) == 0) {
4537 EmitIt = false; // This is indexing into a zero sized array?
4538 } else if (isa<ConstantInt>(C))
4539 return ReplaceInstUsesWith(I, // No comparison is needed here.
4540 ConstantInt::get(Type::Int1Ty,
4541 Cond == ICmpInst::ICMP_NE));
4546 new ICmpInst(Cond, GEPLHS->getOperand(i),
4547 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4551 InVal = InsertNewInstBefore(InVal, I);
4552 InsertNewInstBefore(Comp, I);
4553 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4554 InVal = BinaryOperator::createOr(InVal, Comp);
4555 else // True if all are equal
4556 InVal = BinaryOperator::createAnd(InVal, Comp);
4564 // No comparison is needed here, all indexes = 0
4565 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4566 Cond == ICmpInst::ICMP_EQ));
4569 // Only lower this if the icmp is the only user of the GEP or if we expect
4570 // the result to fold to a constant!
4571 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4572 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4573 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4574 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4575 Constant::getNullValue(Offset->getType()));
4577 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4578 // If the base pointers are different, but the indices are the same, just
4579 // compare the base pointer.
4580 if (PtrBase != GEPRHS->getOperand(0)) {
4581 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4582 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4583 GEPRHS->getOperand(0)->getType();
4585 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4586 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4587 IndicesTheSame = false;
4591 // If all indices are the same, just compare the base pointers.
4593 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4594 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4596 // Otherwise, the base pointers are different and the indices are
4597 // different, bail out.
4601 // If one of the GEPs has all zero indices, recurse.
4602 bool AllZeros = true;
4603 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4604 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4605 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4610 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4611 ICmpInst::getSwappedPredicate(Cond), I);
4613 // If the other GEP has all zero indices, recurse.
4615 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4616 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4617 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4622 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4624 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4625 // If the GEPs only differ by one index, compare it.
4626 unsigned NumDifferences = 0; // Keep track of # differences.
4627 unsigned DiffOperand = 0; // The operand that differs.
4628 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4629 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4630 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4631 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4632 // Irreconcilable differences.
4636 if (NumDifferences++) break;
4641 if (NumDifferences == 0) // SAME GEP?
4642 return ReplaceInstUsesWith(I, // No comparison is needed here.
4643 ConstantInt::get(Type::Int1Ty,
4644 isTrueWhenEqual(Cond)));
4646 else if (NumDifferences == 1) {
4647 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4648 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4649 // Make sure we do a signed comparison here.
4650 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4654 // Only lower this if the icmp is the only user of the GEP or if we expect
4655 // the result to fold to a constant!
4656 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4657 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4658 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4659 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4660 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4661 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4667 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4668 bool Changed = SimplifyCompare(I);
4669 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4671 // Fold trivial predicates.
4672 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4673 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4674 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4675 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4677 // Simplify 'fcmp pred X, X'
4679 switch (I.getPredicate()) {
4680 default: assert(0 && "Unknown predicate!");
4681 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4682 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4683 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4684 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4685 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4686 case FCmpInst::FCMP_OLT: // True if ordered and less than
4687 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4688 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4690 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4691 case FCmpInst::FCMP_ULT: // True if unordered or less than
4692 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4693 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4694 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4695 I.setPredicate(FCmpInst::FCMP_UNO);
4696 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4699 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4700 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4701 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4702 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4703 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4704 I.setPredicate(FCmpInst::FCMP_ORD);
4705 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4710 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4711 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4713 // Handle fcmp with constant RHS
4714 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4715 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4716 switch (LHSI->getOpcode()) {
4717 case Instruction::PHI:
4718 if (Instruction *NV = FoldOpIntoPhi(I))
4721 case Instruction::Select:
4722 // If either operand of the select is a constant, we can fold the
4723 // comparison into the select arms, which will cause one to be
4724 // constant folded and the select turned into a bitwise or.
4725 Value *Op1 = 0, *Op2 = 0;
4726 if (LHSI->hasOneUse()) {
4727 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4728 // Fold the known value into the constant operand.
4729 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4730 // Insert a new FCmp of the other select operand.
4731 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4732 LHSI->getOperand(2), RHSC,
4734 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4735 // Fold the known value into the constant operand.
4736 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4737 // Insert a new FCmp of the other select operand.
4738 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4739 LHSI->getOperand(1), RHSC,
4745 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4750 return Changed ? &I : 0;
4753 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4754 bool Changed = SimplifyCompare(I);
4755 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4756 const Type *Ty = Op0->getType();
4760 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4761 isTrueWhenEqual(I)));
4763 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4764 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4766 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4767 // addresses never equal each other! We already know that Op0 != Op1.
4768 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4769 isa<ConstantPointerNull>(Op0)) &&
4770 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4771 isa<ConstantPointerNull>(Op1)))
4772 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4773 !isTrueWhenEqual(I)));
4775 // icmp's with boolean values can always be turned into bitwise operations
4776 if (Ty == Type::Int1Ty) {
4777 switch (I.getPredicate()) {
4778 default: assert(0 && "Invalid icmp instruction!");
4779 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4780 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4781 InsertNewInstBefore(Xor, I);
4782 return BinaryOperator::createNot(Xor);
4784 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4785 return BinaryOperator::createXor(Op0, Op1);
4787 case ICmpInst::ICMP_UGT:
4788 case ICmpInst::ICMP_SGT:
4789 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4791 case ICmpInst::ICMP_ULT:
4792 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4793 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4794 InsertNewInstBefore(Not, I);
4795 return BinaryOperator::createAnd(Not, Op1);
4797 case ICmpInst::ICMP_UGE:
4798 case ICmpInst::ICMP_SGE:
4799 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4801 case ICmpInst::ICMP_ULE:
4802 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4803 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4804 InsertNewInstBefore(Not, I);
4805 return BinaryOperator::createOr(Not, Op1);
4810 // See if we are doing a comparison between a constant and an instruction that
4811 // can be folded into the comparison.
4812 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4813 switch (I.getPredicate()) {
4815 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4816 if (CI->isMinValue(false))
4817 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4818 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4819 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4820 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4821 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4822 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4823 if (CI->isMinValue(true))
4824 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4825 ConstantInt::getAllOnesValue(Op0->getType()));
4829 case ICmpInst::ICMP_SLT:
4830 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4831 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4832 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4833 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4834 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4835 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4838 case ICmpInst::ICMP_UGT:
4839 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4840 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4841 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4842 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4843 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4844 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4846 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4847 if (CI->isMaxValue(true))
4848 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4849 ConstantInt::getNullValue(Op0->getType()));
4852 case ICmpInst::ICMP_SGT:
4853 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4854 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4855 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4856 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4857 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4858 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4861 case ICmpInst::ICMP_ULE:
4862 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4863 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4864 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4865 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4866 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4867 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4870 case ICmpInst::ICMP_SLE:
4871 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4872 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4873 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4874 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4875 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4876 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4879 case ICmpInst::ICMP_UGE:
4880 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4881 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4882 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4883 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4884 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4885 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4888 case ICmpInst::ICMP_SGE:
4889 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4890 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4891 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4892 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4893 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4894 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4898 // If we still have a icmp le or icmp ge instruction, turn it into the
4899 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4900 // already been handled above, this requires little checking.
4902 switch (I.getPredicate()) {
4904 case ICmpInst::ICMP_ULE:
4905 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4906 case ICmpInst::ICMP_SLE:
4907 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4908 case ICmpInst::ICMP_UGE:
4909 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4910 case ICmpInst::ICMP_SGE:
4911 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4914 // See if we can fold the comparison based on bits known to be zero or one
4915 // in the input. If this comparison is a normal comparison, it demands all
4916 // bits, if it is a sign bit comparison, it only demands the sign bit.
4919 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4921 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4922 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4923 if (SimplifyDemandedBits(Op0,
4924 isSignBit ? APInt::getSignBit(BitWidth)
4925 : APInt::getAllOnesValue(BitWidth),
4926 KnownZero, KnownOne, 0))
4929 // Given the known and unknown bits, compute a range that the LHS could be
4931 if ((KnownOne | KnownZero) != 0) {
4932 // Compute the Min, Max and RHS values based on the known bits. For the
4933 // EQ and NE we use unsigned values.
4934 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4935 const APInt& RHSVal = CI->getValue();
4936 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4937 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4940 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4943 switch (I.getPredicate()) { // LE/GE have been folded already.
4944 default: assert(0 && "Unknown icmp opcode!");
4945 case ICmpInst::ICMP_EQ:
4946 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4947 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4949 case ICmpInst::ICMP_NE:
4950 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4951 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4953 case ICmpInst::ICMP_ULT:
4954 if (Max.ult(RHSVal))
4955 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4956 if (Min.uge(RHSVal))
4957 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4959 case ICmpInst::ICMP_UGT:
4960 if (Min.ugt(RHSVal))
4961 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4962 if (Max.ule(RHSVal))
4963 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4965 case ICmpInst::ICMP_SLT:
4966 if (Max.slt(RHSVal))
4967 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4968 if (Min.sgt(RHSVal))
4969 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4971 case ICmpInst::ICMP_SGT:
4972 if (Min.sgt(RHSVal))
4973 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4974 if (Max.sle(RHSVal))
4975 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4980 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4981 // instruction, see if that instruction also has constants so that the
4982 // instruction can be folded into the icmp
4983 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4984 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4988 // Handle icmp with constant (but not simple integer constant) RHS
4989 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4990 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4991 switch (LHSI->getOpcode()) {
4992 case Instruction::GetElementPtr:
4993 if (RHSC->isNullValue()) {
4994 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4995 bool isAllZeros = true;
4996 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4997 if (!isa<Constant>(LHSI->getOperand(i)) ||
4998 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5003 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5004 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5008 case Instruction::PHI:
5009 if (Instruction *NV = FoldOpIntoPhi(I))
5012 case Instruction::Select: {
5013 // If either operand of the select is a constant, we can fold the
5014 // comparison into the select arms, which will cause one to be
5015 // constant folded and the select turned into a bitwise or.
5016 Value *Op1 = 0, *Op2 = 0;
5017 if (LHSI->hasOneUse()) {
5018 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5019 // Fold the known value into the constant operand.
5020 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5021 // Insert a new ICmp of the other select operand.
5022 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5023 LHSI->getOperand(2), RHSC,
5025 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5026 // Fold the known value into the constant operand.
5027 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5028 // Insert a new ICmp of the other select operand.
5029 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5030 LHSI->getOperand(1), RHSC,
5036 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5039 case Instruction::Malloc:
5040 // If we have (malloc != null), and if the malloc has a single use, we
5041 // can assume it is successful and remove the malloc.
5042 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5043 AddToWorkList(LHSI);
5044 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5045 !isTrueWhenEqual(I)));
5051 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5052 if (User *GEP = dyn_castGetElementPtr(Op0))
5053 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5055 if (User *GEP = dyn_castGetElementPtr(Op1))
5056 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5057 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5060 // Test to see if the operands of the icmp are casted versions of other
5061 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5063 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5064 if (isa<PointerType>(Op0->getType()) &&
5065 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5066 // We keep moving the cast from the left operand over to the right
5067 // operand, where it can often be eliminated completely.
5068 Op0 = CI->getOperand(0);
5070 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5071 // so eliminate it as well.
5072 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5073 Op1 = CI2->getOperand(0);
5075 // If Op1 is a constant, we can fold the cast into the constant.
5076 if (Op0->getType() != Op1->getType())
5077 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5078 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5080 // Otherwise, cast the RHS right before the icmp
5081 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5083 return new ICmpInst(I.getPredicate(), Op0, Op1);
5087 if (isa<CastInst>(Op0)) {
5088 // Handle the special case of: icmp (cast bool to X), <cst>
5089 // This comes up when you have code like
5092 // For generality, we handle any zero-extension of any operand comparison
5093 // with a constant or another cast from the same type.
5094 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5095 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5099 if (I.isEquality()) {
5100 Value *A, *B, *C, *D;
5101 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5102 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5103 Value *OtherVal = A == Op1 ? B : A;
5104 return new ICmpInst(I.getPredicate(), OtherVal,
5105 Constant::getNullValue(A->getType()));
5108 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5109 // A^c1 == C^c2 --> A == C^(c1^c2)
5110 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5111 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5112 if (Op1->hasOneUse()) {
5113 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5114 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5115 return new ICmpInst(I.getPredicate(), A,
5116 InsertNewInstBefore(Xor, I));
5119 // A^B == A^D -> B == D
5120 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5121 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5122 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5123 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5127 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5128 (A == Op0 || B == Op0)) {
5129 // A == (A^B) -> B == 0
5130 Value *OtherVal = A == Op0 ? B : A;
5131 return new ICmpInst(I.getPredicate(), OtherVal,
5132 Constant::getNullValue(A->getType()));
5134 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5135 // (A-B) == A -> B == 0
5136 return new ICmpInst(I.getPredicate(), B,
5137 Constant::getNullValue(B->getType()));
5139 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5140 // A == (A-B) -> B == 0
5141 return new ICmpInst(I.getPredicate(), B,
5142 Constant::getNullValue(B->getType()));
5145 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5146 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5147 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5148 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5149 Value *X = 0, *Y = 0, *Z = 0;
5152 X = B; Y = D; Z = A;
5153 } else if (A == D) {
5154 X = B; Y = C; Z = A;
5155 } else if (B == C) {
5156 X = A; Y = D; Z = B;
5157 } else if (B == D) {
5158 X = A; Y = C; Z = B;
5161 if (X) { // Build (X^Y) & Z
5162 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5163 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5164 I.setOperand(0, Op1);
5165 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5170 return Changed ? &I : 0;
5174 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5175 /// and CmpRHS are both known to be integer constants.
5176 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5177 ConstantInt *DivRHS) {
5178 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5179 const APInt &CmpRHSV = CmpRHS->getValue();
5181 // FIXME: If the operand types don't match the type of the divide
5182 // then don't attempt this transform. The code below doesn't have the
5183 // logic to deal with a signed divide and an unsigned compare (and
5184 // vice versa). This is because (x /s C1) <s C2 produces different
5185 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5186 // (x /u C1) <u C2. Simply casting the operands and result won't
5187 // work. :( The if statement below tests that condition and bails
5189 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5190 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5192 if (DivRHS->isZero())
5193 return 0; // The ProdOV computation fails on divide by zero.
5195 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5196 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5197 // C2 (CI). By solving for X we can turn this into a range check
5198 // instead of computing a divide.
5199 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5201 // Determine if the product overflows by seeing if the product is
5202 // not equal to the divide. Make sure we do the same kind of divide
5203 // as in the LHS instruction that we're folding.
5204 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5205 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5207 // Get the ICmp opcode
5208 ICmpInst::Predicate Pred = ICI.getPredicate();
5210 // Figure out the interval that is being checked. For example, a comparison
5211 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5212 // Compute this interval based on the constants involved and the signedness of
5213 // the compare/divide. This computes a half-open interval, keeping track of
5214 // whether either value in the interval overflows. After analysis each
5215 // overflow variable is set to 0 if it's corresponding bound variable is valid
5216 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5217 int LoOverflow = 0, HiOverflow = 0;
5218 ConstantInt *LoBound = 0, *HiBound = 0;
5221 if (!DivIsSigned) { // udiv
5222 // e.g. X/5 op 3 --> [15, 20)
5224 HiOverflow = LoOverflow = ProdOV;
5226 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5227 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5228 if (CmpRHSV == 0) { // (X / pos) op 0
5229 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5230 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5232 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5233 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5234 HiOverflow = LoOverflow = ProdOV;
5236 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5237 } else { // (X / pos) op neg
5238 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5239 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5240 LoOverflow = AddWithOverflow(LoBound, Prod,
5241 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5242 HiBound = AddOne(Prod);
5243 HiOverflow = ProdOV ? -1 : 0;
5245 } else { // Divisor is < 0.
5246 if (CmpRHSV == 0) { // (X / neg) op 0
5247 // e.g. X/-5 op 0 --> [-4, 5)
5248 LoBound = AddOne(DivRHS);
5249 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5250 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5251 HiOverflow = 1; // [INTMIN+1, overflow)
5252 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5254 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5255 // e.g. X/-5 op 3 --> [-19, -14)
5256 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5258 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5259 HiBound = AddOne(Prod);
5260 } else { // (X / neg) op neg
5261 // e.g. X/-5 op -3 --> [15, 20)
5263 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5264 HiBound = Subtract(Prod, DivRHS);
5267 // Dividing by a negative swaps the condition. LT <-> GT
5268 Pred = ICmpInst::getSwappedPredicate(Pred);
5271 Value *X = DivI->getOperand(0);
5273 default: assert(0 && "Unhandled icmp opcode!");
5274 case ICmpInst::ICMP_EQ:
5275 if (LoOverflow && HiOverflow)
5276 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5277 else if (HiOverflow)
5278 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5279 ICmpInst::ICMP_UGE, X, LoBound);
5280 else if (LoOverflow)
5281 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5282 ICmpInst::ICMP_ULT, X, HiBound);
5284 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5285 case ICmpInst::ICMP_NE:
5286 if (LoOverflow && HiOverflow)
5287 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5288 else if (HiOverflow)
5289 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5290 ICmpInst::ICMP_ULT, X, LoBound);
5291 else if (LoOverflow)
5292 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5293 ICmpInst::ICMP_UGE, X, HiBound);
5295 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5296 case ICmpInst::ICMP_ULT:
5297 case ICmpInst::ICMP_SLT:
5298 if (LoOverflow == +1) // Low bound is greater than input range.
5299 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5300 if (LoOverflow == -1) // Low bound is less than input range.
5301 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5302 return new ICmpInst(Pred, X, LoBound);
5303 case ICmpInst::ICMP_UGT:
5304 case ICmpInst::ICMP_SGT:
5305 if (HiOverflow == +1) // High bound greater than input range.
5306 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5307 else if (HiOverflow == -1) // High bound less than input range.
5308 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5309 if (Pred == ICmpInst::ICMP_UGT)
5310 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5312 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5317 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5319 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5322 const APInt &RHSV = RHS->getValue();
5324 switch (LHSI->getOpcode()) {
5325 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5326 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5327 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5329 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5330 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5331 Value *CompareVal = LHSI->getOperand(0);
5333 // If the sign bit of the XorCST is not set, there is no change to
5334 // the operation, just stop using the Xor.
5335 if (!XorCST->getValue().isNegative()) {
5336 ICI.setOperand(0, CompareVal);
5337 AddToWorkList(LHSI);
5341 // Was the old condition true if the operand is positive?
5342 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5344 // If so, the new one isn't.
5345 isTrueIfPositive ^= true;
5347 if (isTrueIfPositive)
5348 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5350 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5354 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5355 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5356 LHSI->getOperand(0)->hasOneUse()) {
5357 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5359 // If the LHS is an AND of a truncating cast, we can widen the
5360 // and/compare to be the input width without changing the value
5361 // produced, eliminating a cast.
5362 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5363 // We can do this transformation if either the AND constant does not
5364 // have its sign bit set or if it is an equality comparison.
5365 // Extending a relational comparison when we're checking the sign
5366 // bit would not work.
5367 if (Cast->hasOneUse() &&
5368 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5369 RHSV.isPositive())) {
5371 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5372 APInt NewCST = AndCST->getValue();
5373 NewCST.zext(BitWidth);
5375 NewCI.zext(BitWidth);
5376 Instruction *NewAnd =
5377 BinaryOperator::createAnd(Cast->getOperand(0),
5378 ConstantInt::get(NewCST),LHSI->getName());
5379 InsertNewInstBefore(NewAnd, ICI);
5380 return new ICmpInst(ICI.getPredicate(), NewAnd,
5381 ConstantInt::get(NewCI));
5385 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5386 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5387 // happens a LOT in code produced by the C front-end, for bitfield
5389 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5390 if (Shift && !Shift->isShift())
5394 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5395 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5396 const Type *AndTy = AndCST->getType(); // Type of the and.
5398 // We can fold this as long as we can't shift unknown bits
5399 // into the mask. This can only happen with signed shift
5400 // rights, as they sign-extend.
5402 bool CanFold = Shift->isLogicalShift();
5404 // To test for the bad case of the signed shr, see if any
5405 // of the bits shifted in could be tested after the mask.
5406 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5407 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5409 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5410 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5411 AndCST->getValue()) == 0)
5417 if (Shift->getOpcode() == Instruction::Shl)
5418 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5420 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5422 // Check to see if we are shifting out any of the bits being
5424 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5425 // If we shifted bits out, the fold is not going to work out.
5426 // As a special case, check to see if this means that the
5427 // result is always true or false now.
5428 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5429 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5430 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5431 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5433 ICI.setOperand(1, NewCst);
5434 Constant *NewAndCST;
5435 if (Shift->getOpcode() == Instruction::Shl)
5436 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5438 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5439 LHSI->setOperand(1, NewAndCST);
5440 LHSI->setOperand(0, Shift->getOperand(0));
5441 AddToWorkList(Shift); // Shift is dead.
5442 AddUsesToWorkList(ICI);
5448 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5449 // preferable because it allows the C<<Y expression to be hoisted out
5450 // of a loop if Y is invariant and X is not.
5451 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5452 ICI.isEquality() && !Shift->isArithmeticShift() &&
5453 isa<Instruction>(Shift->getOperand(0))) {
5456 if (Shift->getOpcode() == Instruction::LShr) {
5457 NS = BinaryOperator::createShl(AndCST,
5458 Shift->getOperand(1), "tmp");
5460 // Insert a logical shift.
5461 NS = BinaryOperator::createLShr(AndCST,
5462 Shift->getOperand(1), "tmp");
5464 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5466 // Compute X & (C << Y).
5467 Instruction *NewAnd =
5468 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5469 InsertNewInstBefore(NewAnd, ICI);
5471 ICI.setOperand(0, NewAnd);
5477 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5478 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5481 uint32_t TypeBits = RHSV.getBitWidth();
5483 // Check that the shift amount is in range. If not, don't perform
5484 // undefined shifts. When the shift is visited it will be
5486 if (ShAmt->uge(TypeBits))
5489 if (ICI.isEquality()) {
5490 // If we are comparing against bits always shifted out, the
5491 // comparison cannot succeed.
5493 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5494 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5495 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5496 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5497 return ReplaceInstUsesWith(ICI, Cst);
5500 if (LHSI->hasOneUse()) {
5501 // Otherwise strength reduce the shift into an and.
5502 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5504 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5507 BinaryOperator::createAnd(LHSI->getOperand(0),
5508 Mask, LHSI->getName()+".mask");
5509 Value *And = InsertNewInstBefore(AndI, ICI);
5510 return new ICmpInst(ICI.getPredicate(), And,
5511 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5515 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5516 bool TrueIfSigned = false;
5517 if (LHSI->hasOneUse() &&
5518 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5519 // (X << 31) <s 0 --> (X&1) != 0
5520 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5521 (TypeBits-ShAmt->getZExtValue()-1));
5523 BinaryOperator::createAnd(LHSI->getOperand(0),
5524 Mask, LHSI->getName()+".mask");
5525 Value *And = InsertNewInstBefore(AndI, ICI);
5527 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5528 And, Constant::getNullValue(And->getType()));
5533 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5534 case Instruction::AShr: {
5535 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5538 if (ICI.isEquality()) {
5539 // Check that the shift amount is in range. If not, don't perform
5540 // undefined shifts. When the shift is visited it will be
5542 uint32_t TypeBits = RHSV.getBitWidth();
5543 if (ShAmt->uge(TypeBits))
5545 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5547 // If we are comparing against bits always shifted out, the
5548 // comparison cannot succeed.
5549 APInt Comp = RHSV << ShAmtVal;
5550 if (LHSI->getOpcode() == Instruction::LShr)
5551 Comp = Comp.lshr(ShAmtVal);
5553 Comp = Comp.ashr(ShAmtVal);
5555 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5556 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5557 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5558 return ReplaceInstUsesWith(ICI, Cst);
5561 if (LHSI->hasOneUse() || RHSV == 0) {
5562 // Otherwise strength reduce the shift into an and.
5563 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5564 Constant *Mask = ConstantInt::get(Val);
5567 BinaryOperator::createAnd(LHSI->getOperand(0),
5568 Mask, LHSI->getName()+".mask");
5569 Value *And = InsertNewInstBefore(AndI, ICI);
5570 return new ICmpInst(ICI.getPredicate(), And,
5571 ConstantExpr::getShl(RHS, ShAmt));
5577 case Instruction::SDiv:
5578 case Instruction::UDiv:
5579 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5580 // Fold this div into the comparison, producing a range check.
5581 // Determine, based on the divide type, what the range is being
5582 // checked. If there is an overflow on the low or high side, remember
5583 // it, otherwise compute the range [low, hi) bounding the new value.
5584 // See: InsertRangeTest above for the kinds of replacements possible.
5585 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5586 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5592 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5593 if (ICI.isEquality()) {
5594 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5596 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5597 // the second operand is a constant, simplify a bit.
5598 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5599 switch (BO->getOpcode()) {
5600 case Instruction::SRem:
5601 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5602 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5603 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5604 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5605 Instruction *NewRem =
5606 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5608 InsertNewInstBefore(NewRem, ICI);
5609 return new ICmpInst(ICI.getPredicate(), NewRem,
5610 Constant::getNullValue(BO->getType()));
5614 case Instruction::Add:
5615 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5616 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5617 if (BO->hasOneUse())
5618 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5619 Subtract(RHS, BOp1C));
5620 } else if (RHSV == 0) {
5621 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5622 // efficiently invertible, or if the add has just this one use.
5623 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5625 if (Value *NegVal = dyn_castNegVal(BOp1))
5626 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5627 else if (Value *NegVal = dyn_castNegVal(BOp0))
5628 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5629 else if (BO->hasOneUse()) {
5630 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5631 InsertNewInstBefore(Neg, ICI);
5633 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5637 case Instruction::Xor:
5638 // For the xor case, we can xor two constants together, eliminating
5639 // the explicit xor.
5640 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5641 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5642 ConstantExpr::getXor(RHS, BOC));
5645 case Instruction::Sub:
5646 // Replace (([sub|xor] A, B) != 0) with (A != B)
5648 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5652 case Instruction::Or:
5653 // If bits are being or'd in that are not present in the constant we
5654 // are comparing against, then the comparison could never succeed!
5655 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5656 Constant *NotCI = ConstantExpr::getNot(RHS);
5657 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5658 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5663 case Instruction::And:
5664 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5665 // If bits are being compared against that are and'd out, then the
5666 // comparison can never succeed!
5667 if ((RHSV & ~BOC->getValue()) != 0)
5668 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5671 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5672 if (RHS == BOC && RHSV.isPowerOf2())
5673 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5674 ICmpInst::ICMP_NE, LHSI,
5675 Constant::getNullValue(RHS->getType()));
5677 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5678 if (isSignBit(BOC)) {
5679 Value *X = BO->getOperand(0);
5680 Constant *Zero = Constant::getNullValue(X->getType());
5681 ICmpInst::Predicate pred = isICMP_NE ?
5682 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5683 return new ICmpInst(pred, X, Zero);
5686 // ((X & ~7) == 0) --> X < 8
5687 if (RHSV == 0 && isHighOnes(BOC)) {
5688 Value *X = BO->getOperand(0);
5689 Constant *NegX = ConstantExpr::getNeg(BOC);
5690 ICmpInst::Predicate pred = isICMP_NE ?
5691 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5692 return new ICmpInst(pred, X, NegX);
5697 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5698 // Handle icmp {eq|ne} <intrinsic>, intcst.
5699 if (II->getIntrinsicID() == Intrinsic::bswap) {
5701 ICI.setOperand(0, II->getOperand(1));
5702 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5706 } else { // Not a ICMP_EQ/ICMP_NE
5707 // If the LHS is a cast from an integral value of the same size,
5708 // then since we know the RHS is a constant, try to simlify.
5709 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5710 Value *CastOp = Cast->getOperand(0);
5711 const Type *SrcTy = CastOp->getType();
5712 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5713 if (SrcTy->isInteger() &&
5714 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5715 // If this is an unsigned comparison, try to make the comparison use
5716 // smaller constant values.
5717 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5718 // X u< 128 => X s> -1
5719 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5720 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5721 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5722 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5723 // X u> 127 => X s< 0
5724 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5725 Constant::getNullValue(SrcTy));
5733 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5734 /// We only handle extending casts so far.
5736 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5737 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5738 Value *LHSCIOp = LHSCI->getOperand(0);
5739 const Type *SrcTy = LHSCIOp->getType();
5740 const Type *DestTy = LHSCI->getType();
5743 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5744 // integer type is the same size as the pointer type.
5745 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5746 getTargetData().getPointerSizeInBits() ==
5747 cast<IntegerType>(DestTy)->getBitWidth()) {
5749 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5750 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5751 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5752 RHSOp = RHSC->getOperand(0);
5753 // If the pointer types don't match, insert a bitcast.
5754 if (LHSCIOp->getType() != RHSOp->getType())
5755 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5756 LHSCIOp->getType(), ICI);
5760 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5763 // The code below only handles extension cast instructions, so far.
5765 if (LHSCI->getOpcode() != Instruction::ZExt &&
5766 LHSCI->getOpcode() != Instruction::SExt)
5769 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5770 bool isSignedCmp = ICI.isSignedPredicate();
5772 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5773 // Not an extension from the same type?
5774 RHSCIOp = CI->getOperand(0);
5775 if (RHSCIOp->getType() != LHSCIOp->getType())
5778 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5779 // and the other is a zext), then we can't handle this.
5780 if (CI->getOpcode() != LHSCI->getOpcode())
5783 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5784 // then we can't handle this.
5785 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5788 // Okay, just insert a compare of the reduced operands now!
5789 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5792 // If we aren't dealing with a constant on the RHS, exit early
5793 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5797 // Compute the constant that would happen if we truncated to SrcTy then
5798 // reextended to DestTy.
5799 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5800 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5802 // If the re-extended constant didn't change...
5804 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5805 // For example, we might have:
5806 // %A = sext short %X to uint
5807 // %B = icmp ugt uint %A, 1330
5808 // It is incorrect to transform this into
5809 // %B = icmp ugt short %X, 1330
5810 // because %A may have negative value.
5812 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5813 // OR operation is EQ/NE.
5814 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5815 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5820 // The re-extended constant changed so the constant cannot be represented
5821 // in the shorter type. Consequently, we cannot emit a simple comparison.
5823 // First, handle some easy cases. We know the result cannot be equal at this
5824 // point so handle the ICI.isEquality() cases
5825 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5826 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5827 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5828 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5830 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5831 // should have been folded away previously and not enter in here.
5834 // We're performing a signed comparison.
5835 if (cast<ConstantInt>(CI)->getValue().isNegative())
5836 Result = ConstantInt::getFalse(); // X < (small) --> false
5838 Result = ConstantInt::getTrue(); // X < (large) --> true
5840 // We're performing an unsigned comparison.
5842 // We're performing an unsigned comp with a sign extended value.
5843 // This is true if the input is >= 0. [aka >s -1]
5844 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5845 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5846 NegOne, ICI.getName()), ICI);
5848 // Unsigned extend & unsigned compare -> always true.
5849 Result = ConstantInt::getTrue();
5853 // Finally, return the value computed.
5854 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5855 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5856 return ReplaceInstUsesWith(ICI, Result);
5858 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5859 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5860 "ICmp should be folded!");
5861 if (Constant *CI = dyn_cast<Constant>(Result))
5862 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5864 return BinaryOperator::createNot(Result);
5868 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5869 return commonShiftTransforms(I);
5872 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5873 return commonShiftTransforms(I);
5876 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5877 if (Instruction *R = commonShiftTransforms(I))
5880 Value *Op0 = I.getOperand(0);
5882 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5883 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5884 if (CSI->isAllOnesValue())
5885 return ReplaceInstUsesWith(I, CSI);
5887 // See if we can turn a signed shr into an unsigned shr.
5888 if (MaskedValueIsZero(Op0,
5889 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
5890 return BinaryOperator::createLShr(Op0, I.getOperand(1));
5895 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5896 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5897 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5899 // shl X, 0 == X and shr X, 0 == X
5900 // shl 0, X == 0 and shr 0, X == 0
5901 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5902 Op0 == Constant::getNullValue(Op0->getType()))
5903 return ReplaceInstUsesWith(I, Op0);
5905 if (isa<UndefValue>(Op0)) {
5906 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5907 return ReplaceInstUsesWith(I, Op0);
5908 else // undef << X -> 0, undef >>u X -> 0
5909 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5911 if (isa<UndefValue>(Op1)) {
5912 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5913 return ReplaceInstUsesWith(I, Op0);
5914 else // X << undef, X >>u undef -> 0
5915 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5918 // Try to fold constant and into select arguments.
5919 if (isa<Constant>(Op0))
5920 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5921 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5924 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5925 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5930 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5931 BinaryOperator &I) {
5932 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5934 // See if we can simplify any instructions used by the instruction whose sole
5935 // purpose is to compute bits we don't care about.
5936 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5937 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5938 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5939 KnownZero, KnownOne))
5942 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5943 // of a signed value.
5945 if (Op1->uge(TypeBits)) {
5946 if (I.getOpcode() != Instruction::AShr)
5947 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5949 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5954 // ((X*C1) << C2) == (X * (C1 << C2))
5955 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5956 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5957 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5958 return BinaryOperator::createMul(BO->getOperand(0),
5959 ConstantExpr::getShl(BOOp, Op1));
5961 // Try to fold constant and into select arguments.
5962 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5963 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5965 if (isa<PHINode>(Op0))
5966 if (Instruction *NV = FoldOpIntoPhi(I))
5969 if (Op0->hasOneUse()) {
5970 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5971 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5974 switch (Op0BO->getOpcode()) {
5976 case Instruction::Add:
5977 case Instruction::And:
5978 case Instruction::Or:
5979 case Instruction::Xor: {
5980 // These operators commute.
5981 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5982 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5983 match(Op0BO->getOperand(1),
5984 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5985 Instruction *YS = BinaryOperator::createShl(
5986 Op0BO->getOperand(0), Op1,
5988 InsertNewInstBefore(YS, I); // (Y << C)
5990 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5991 Op0BO->getOperand(1)->getName());
5992 InsertNewInstBefore(X, I); // (X + (Y << C))
5993 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5994 return BinaryOperator::createAnd(X, ConstantInt::get(
5995 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5998 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5999 Value *Op0BOOp1 = Op0BO->getOperand(1);
6000 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6002 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6003 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6005 Instruction *YS = BinaryOperator::createShl(
6006 Op0BO->getOperand(0), Op1,
6008 InsertNewInstBefore(YS, I); // (Y << C)
6010 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6011 V1->getName()+".mask");
6012 InsertNewInstBefore(XM, I); // X & (CC << C)
6014 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6019 case Instruction::Sub: {
6020 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6021 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6022 match(Op0BO->getOperand(0),
6023 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6024 Instruction *YS = BinaryOperator::createShl(
6025 Op0BO->getOperand(1), Op1,
6027 InsertNewInstBefore(YS, I); // (Y << C)
6029 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6030 Op0BO->getOperand(0)->getName());
6031 InsertNewInstBefore(X, I); // (X + (Y << C))
6032 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6033 return BinaryOperator::createAnd(X, ConstantInt::get(
6034 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6037 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6038 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6039 match(Op0BO->getOperand(0),
6040 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6041 m_ConstantInt(CC))) && V2 == Op1 &&
6042 cast<BinaryOperator>(Op0BO->getOperand(0))
6043 ->getOperand(0)->hasOneUse()) {
6044 Instruction *YS = BinaryOperator::createShl(
6045 Op0BO->getOperand(1), Op1,
6047 InsertNewInstBefore(YS, I); // (Y << C)
6049 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6050 V1->getName()+".mask");
6051 InsertNewInstBefore(XM, I); // X & (CC << C)
6053 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6061 // If the operand is an bitwise operator with a constant RHS, and the
6062 // shift is the only use, we can pull it out of the shift.
6063 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6064 bool isValid = true; // Valid only for And, Or, Xor
6065 bool highBitSet = false; // Transform if high bit of constant set?
6067 switch (Op0BO->getOpcode()) {
6068 default: isValid = false; break; // Do not perform transform!
6069 case Instruction::Add:
6070 isValid = isLeftShift;
6072 case Instruction::Or:
6073 case Instruction::Xor:
6076 case Instruction::And:
6081 // If this is a signed shift right, and the high bit is modified
6082 // by the logical operation, do not perform the transformation.
6083 // The highBitSet boolean indicates the value of the high bit of
6084 // the constant which would cause it to be modified for this
6087 if (isValid && I.getOpcode() == Instruction::AShr)
6088 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6091 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6093 Instruction *NewShift =
6094 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6095 InsertNewInstBefore(NewShift, I);
6096 NewShift->takeName(Op0BO);
6098 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6105 // Find out if this is a shift of a shift by a constant.
6106 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6107 if (ShiftOp && !ShiftOp->isShift())
6110 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6111 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6112 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6113 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6114 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6115 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6116 Value *X = ShiftOp->getOperand(0);
6118 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6119 if (AmtSum > TypeBits)
6122 const IntegerType *Ty = cast<IntegerType>(I.getType());
6124 // Check for (X << c1) << c2 and (X >> c1) >> c2
6125 if (I.getOpcode() == ShiftOp->getOpcode()) {
6126 return BinaryOperator::create(I.getOpcode(), X,
6127 ConstantInt::get(Ty, AmtSum));
6128 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6129 I.getOpcode() == Instruction::AShr) {
6130 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6131 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6132 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6133 I.getOpcode() == Instruction::LShr) {
6134 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6135 Instruction *Shift =
6136 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6137 InsertNewInstBefore(Shift, I);
6139 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6140 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6143 // Okay, if we get here, one shift must be left, and the other shift must be
6144 // right. See if the amounts are equal.
6145 if (ShiftAmt1 == ShiftAmt2) {
6146 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6147 if (I.getOpcode() == Instruction::Shl) {
6148 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6149 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6151 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6152 if (I.getOpcode() == Instruction::LShr) {
6153 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6154 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6156 // We can simplify ((X << C) >>s C) into a trunc + sext.
6157 // NOTE: we could do this for any C, but that would make 'unusual' integer
6158 // types. For now, just stick to ones well-supported by the code
6160 const Type *SExtType = 0;
6161 switch (Ty->getBitWidth() - ShiftAmt1) {
6168 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6173 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6174 InsertNewInstBefore(NewTrunc, I);
6175 return new SExtInst(NewTrunc, Ty);
6177 // Otherwise, we can't handle it yet.
6178 } else if (ShiftAmt1 < ShiftAmt2) {
6179 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6181 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6182 if (I.getOpcode() == Instruction::Shl) {
6183 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6184 ShiftOp->getOpcode() == Instruction::AShr);
6185 Instruction *Shift =
6186 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6187 InsertNewInstBefore(Shift, I);
6189 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6190 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6193 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6194 if (I.getOpcode() == Instruction::LShr) {
6195 assert(ShiftOp->getOpcode() == Instruction::Shl);
6196 Instruction *Shift =
6197 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6198 InsertNewInstBefore(Shift, I);
6200 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6201 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6204 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6206 assert(ShiftAmt2 < ShiftAmt1);
6207 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6209 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6210 if (I.getOpcode() == Instruction::Shl) {
6211 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6212 ShiftOp->getOpcode() == Instruction::AShr);
6213 Instruction *Shift =
6214 BinaryOperator::create(ShiftOp->getOpcode(), X,
6215 ConstantInt::get(Ty, ShiftDiff));
6216 InsertNewInstBefore(Shift, I);
6218 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6219 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6222 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6223 if (I.getOpcode() == Instruction::LShr) {
6224 assert(ShiftOp->getOpcode() == Instruction::Shl);
6225 Instruction *Shift =
6226 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6227 InsertNewInstBefore(Shift, I);
6229 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6230 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6233 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6240 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6241 /// expression. If so, decompose it, returning some value X, such that Val is
6244 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6246 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6247 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6248 Offset = CI->getZExtValue();
6250 return ConstantInt::get(Type::Int32Ty, 0);
6251 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6252 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6253 if (I->getOpcode() == Instruction::Shl) {
6254 // This is a value scaled by '1 << the shift amt'.
6255 Scale = 1U << RHS->getZExtValue();
6257 return I->getOperand(0);
6258 } else if (I->getOpcode() == Instruction::Mul) {
6259 // This value is scaled by 'RHS'.
6260 Scale = RHS->getZExtValue();
6262 return I->getOperand(0);
6263 } else if (I->getOpcode() == Instruction::Add) {
6264 // We have X+C. Check to see if we really have (X*C2)+C1,
6265 // where C1 is divisible by C2.
6268 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6269 Offset += RHS->getZExtValue();
6276 // Otherwise, we can't look past this.
6283 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6284 /// try to eliminate the cast by moving the type information into the alloc.
6285 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6286 AllocationInst &AI) {
6287 const PointerType *PTy = cast<PointerType>(CI.getType());
6289 // Remove any uses of AI that are dead.
6290 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6292 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6293 Instruction *User = cast<Instruction>(*UI++);
6294 if (isInstructionTriviallyDead(User)) {
6295 while (UI != E && *UI == User)
6296 ++UI; // If this instruction uses AI more than once, don't break UI.
6299 DOUT << "IC: DCE: " << *User;
6300 EraseInstFromFunction(*User);
6304 // Get the type really allocated and the type casted to.
6305 const Type *AllocElTy = AI.getAllocatedType();
6306 const Type *CastElTy = PTy->getElementType();
6307 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6309 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6310 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6311 if (CastElTyAlign < AllocElTyAlign) return 0;
6313 // If the allocation has multiple uses, only promote it if we are strictly
6314 // increasing the alignment of the resultant allocation. If we keep it the
6315 // same, we open the door to infinite loops of various kinds.
6316 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6318 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6319 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6320 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6322 // See if we can satisfy the modulus by pulling a scale out of the array
6324 unsigned ArraySizeScale;
6326 Value *NumElements = // See if the array size is a decomposable linear expr.
6327 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6329 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6331 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6332 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6334 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6339 // If the allocation size is constant, form a constant mul expression
6340 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6341 if (isa<ConstantInt>(NumElements))
6342 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6343 // otherwise multiply the amount and the number of elements
6344 else if (Scale != 1) {
6345 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6346 Amt = InsertNewInstBefore(Tmp, AI);
6350 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6351 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6352 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6353 Amt = InsertNewInstBefore(Tmp, AI);
6356 AllocationInst *New;
6357 if (isa<MallocInst>(AI))
6358 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6360 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6361 InsertNewInstBefore(New, AI);
6364 // If the allocation has multiple uses, insert a cast and change all things
6365 // that used it to use the new cast. This will also hack on CI, but it will
6367 if (!AI.hasOneUse()) {
6368 AddUsesToWorkList(AI);
6369 // New is the allocation instruction, pointer typed. AI is the original
6370 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6371 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6372 InsertNewInstBefore(NewCast, AI);
6373 AI.replaceAllUsesWith(NewCast);
6375 return ReplaceInstUsesWith(CI, New);
6378 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6379 /// and return it as type Ty without inserting any new casts and without
6380 /// changing the computed value. This is used by code that tries to decide
6381 /// whether promoting or shrinking integer operations to wider or smaller types
6382 /// will allow us to eliminate a truncate or extend.
6384 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6385 /// extension operation if Ty is larger.
6386 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6387 unsigned CastOpc, int &NumCastsRemoved) {
6388 // We can always evaluate constants in another type.
6389 if (isa<ConstantInt>(V))
6392 Instruction *I = dyn_cast<Instruction>(V);
6393 if (!I) return false;
6395 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6397 // If this is an extension or truncate, we can often eliminate it.
6398 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6399 // If this is a cast from the destination type, we can trivially eliminate
6400 // it, and this will remove a cast overall.
6401 if (I->getOperand(0)->getType() == Ty) {
6402 // If the first operand is itself a cast, and is eliminable, do not count
6403 // this as an eliminable cast. We would prefer to eliminate those two
6405 if (!isa<CastInst>(I->getOperand(0)))
6411 // We can't extend or shrink something that has multiple uses: doing so would
6412 // require duplicating the instruction in general, which isn't profitable.
6413 if (!I->hasOneUse()) return false;
6415 switch (I->getOpcode()) {
6416 case Instruction::Add:
6417 case Instruction::Sub:
6418 case Instruction::And:
6419 case Instruction::Or:
6420 case Instruction::Xor:
6421 // These operators can all arbitrarily be extended or truncated.
6422 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6424 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6427 case Instruction::Shl:
6428 // If we are truncating the result of this SHL, and if it's a shift of a
6429 // constant amount, we can always perform a SHL in a smaller type.
6430 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6431 uint32_t BitWidth = Ty->getBitWidth();
6432 if (BitWidth < OrigTy->getBitWidth() &&
6433 CI->getLimitedValue(BitWidth) < BitWidth)
6434 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6438 case Instruction::LShr:
6439 // If this is a truncate of a logical shr, we can truncate it to a smaller
6440 // lshr iff we know that the bits we would otherwise be shifting in are
6442 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6443 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6444 uint32_t BitWidth = Ty->getBitWidth();
6445 if (BitWidth < OrigBitWidth &&
6446 MaskedValueIsZero(I->getOperand(0),
6447 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6448 CI->getLimitedValue(BitWidth) < BitWidth) {
6449 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6454 case Instruction::ZExt:
6455 case Instruction::SExt:
6456 case Instruction::Trunc:
6457 // If this is the same kind of case as our original (e.g. zext+zext), we
6458 // can safely replace it. Note that replacing it does not reduce the number
6459 // of casts in the input.
6460 if (I->getOpcode() == CastOpc)
6465 // TODO: Can handle more cases here.
6472 /// EvaluateInDifferentType - Given an expression that
6473 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6474 /// evaluate the expression.
6475 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6477 if (Constant *C = dyn_cast<Constant>(V))
6478 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6480 // Otherwise, it must be an instruction.
6481 Instruction *I = cast<Instruction>(V);
6482 Instruction *Res = 0;
6483 switch (I->getOpcode()) {
6484 case Instruction::Add:
6485 case Instruction::Sub:
6486 case Instruction::And:
6487 case Instruction::Or:
6488 case Instruction::Xor:
6489 case Instruction::AShr:
6490 case Instruction::LShr:
6491 case Instruction::Shl: {
6492 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6493 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6494 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6495 LHS, RHS, I->getName());
6498 case Instruction::Trunc:
6499 case Instruction::ZExt:
6500 case Instruction::SExt:
6501 // If the source type of the cast is the type we're trying for then we can
6502 // just return the source. There's no need to insert it because it is not
6504 if (I->getOperand(0)->getType() == Ty)
6505 return I->getOperand(0);
6507 // Otherwise, must be the same type of case, so just reinsert a new one.
6508 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6512 // TODO: Can handle more cases here.
6513 assert(0 && "Unreachable!");
6517 return InsertNewInstBefore(Res, *I);
6520 /// @brief Implement the transforms common to all CastInst visitors.
6521 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6522 Value *Src = CI.getOperand(0);
6524 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6525 // eliminate it now.
6526 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6527 if (Instruction::CastOps opc =
6528 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6529 // The first cast (CSrc) is eliminable so we need to fix up or replace
6530 // the second cast (CI). CSrc will then have a good chance of being dead.
6531 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6535 // If we are casting a select then fold the cast into the select
6536 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6537 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6540 // If we are casting a PHI then fold the cast into the PHI
6541 if (isa<PHINode>(Src))
6542 if (Instruction *NV = FoldOpIntoPhi(CI))
6548 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6549 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6550 Value *Src = CI.getOperand(0);
6552 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6553 // If casting the result of a getelementptr instruction with no offset, turn
6554 // this into a cast of the original pointer!
6555 if (GEP->hasAllZeroIndices()) {
6556 // Changing the cast operand is usually not a good idea but it is safe
6557 // here because the pointer operand is being replaced with another
6558 // pointer operand so the opcode doesn't need to change.
6560 CI.setOperand(0, GEP->getOperand(0));
6564 // If the GEP has a single use, and the base pointer is a bitcast, and the
6565 // GEP computes a constant offset, see if we can convert these three
6566 // instructions into fewer. This typically happens with unions and other
6567 // non-type-safe code.
6568 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6569 if (GEP->hasAllConstantIndices()) {
6570 // We are guaranteed to get a constant from EmitGEPOffset.
6571 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6572 int64_t Offset = OffsetV->getSExtValue();
6574 // Get the base pointer input of the bitcast, and the type it points to.
6575 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6576 const Type *GEPIdxTy =
6577 cast<PointerType>(OrigBase->getType())->getElementType();
6578 if (GEPIdxTy->isSized()) {
6579 SmallVector<Value*, 8> NewIndices;
6581 // Start with the index over the outer type. Note that the type size
6582 // might be zero (even if the offset isn't zero) if the indexed type
6583 // is something like [0 x {int, int}]
6584 const Type *IntPtrTy = TD->getIntPtrType();
6585 int64_t FirstIdx = 0;
6586 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6587 FirstIdx = Offset/TySize;
6590 // Handle silly modulus not returning values values [0..TySize).
6594 assert(Offset >= 0);
6596 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6599 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6601 // Index into the types. If we fail, set OrigBase to null.
6603 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6604 const StructLayout *SL = TD->getStructLayout(STy);
6605 if (Offset < (int64_t)SL->getSizeInBytes()) {
6606 unsigned Elt = SL->getElementContainingOffset(Offset);
6607 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6609 Offset -= SL->getElementOffset(Elt);
6610 GEPIdxTy = STy->getElementType(Elt);
6612 // Otherwise, we can't index into this, bail out.
6616 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6617 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6618 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6619 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6622 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6624 GEPIdxTy = STy->getElementType();
6626 // Otherwise, we can't index into this, bail out.
6632 // If we were able to index down into an element, create the GEP
6633 // and bitcast the result. This eliminates one bitcast, potentially
6635 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6637 NewIndices.end(), "");
6638 InsertNewInstBefore(NGEP, CI);
6639 NGEP->takeName(GEP);
6641 if (isa<BitCastInst>(CI))
6642 return new BitCastInst(NGEP, CI.getType());
6643 assert(isa<PtrToIntInst>(CI));
6644 return new PtrToIntInst(NGEP, CI.getType());
6651 return commonCastTransforms(CI);
6656 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6657 /// integer types. This function implements the common transforms for all those
6659 /// @brief Implement the transforms common to CastInst with integer operands
6660 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6661 if (Instruction *Result = commonCastTransforms(CI))
6664 Value *Src = CI.getOperand(0);
6665 const Type *SrcTy = Src->getType();
6666 const Type *DestTy = CI.getType();
6667 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6668 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6670 // See if we can simplify any instructions used by the LHS whose sole
6671 // purpose is to compute bits we don't care about.
6672 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6673 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6674 KnownZero, KnownOne))
6677 // If the source isn't an instruction or has more than one use then we
6678 // can't do anything more.
6679 Instruction *SrcI = dyn_cast<Instruction>(Src);
6680 if (!SrcI || !Src->hasOneUse())
6683 // Attempt to propagate the cast into the instruction for int->int casts.
6684 int NumCastsRemoved = 0;
6685 if (!isa<BitCastInst>(CI) &&
6686 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6687 CI.getOpcode(), NumCastsRemoved)) {
6688 // If this cast is a truncate, evaluting in a different type always
6689 // eliminates the cast, so it is always a win. If this is a zero-extension,
6690 // we need to do an AND to maintain the clear top-part of the computation,
6691 // so we require that the input have eliminated at least one cast. If this
6692 // is a sign extension, we insert two new casts (to do the extension) so we
6693 // require that two casts have been eliminated.
6695 switch (CI.getOpcode()) {
6697 // All the others use floating point so we shouldn't actually
6698 // get here because of the check above.
6699 assert(0 && "Unknown cast type");
6700 case Instruction::Trunc:
6703 case Instruction::ZExt:
6704 DoXForm = NumCastsRemoved >= 1;
6706 case Instruction::SExt:
6707 DoXForm = NumCastsRemoved >= 2;
6712 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6713 CI.getOpcode() == Instruction::SExt);
6714 assert(Res->getType() == DestTy);
6715 switch (CI.getOpcode()) {
6716 default: assert(0 && "Unknown cast type!");
6717 case Instruction::Trunc:
6718 case Instruction::BitCast:
6719 // Just replace this cast with the result.
6720 return ReplaceInstUsesWith(CI, Res);
6721 case Instruction::ZExt: {
6722 // We need to emit an AND to clear the high bits.
6723 assert(SrcBitSize < DestBitSize && "Not a zext?");
6724 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6726 return BinaryOperator::createAnd(Res, C);
6728 case Instruction::SExt:
6729 // We need to emit a cast to truncate, then a cast to sext.
6730 return CastInst::create(Instruction::SExt,
6731 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6737 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6738 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6740 switch (SrcI->getOpcode()) {
6741 case Instruction::Add:
6742 case Instruction::Mul:
6743 case Instruction::And:
6744 case Instruction::Or:
6745 case Instruction::Xor:
6746 // If we are discarding information, rewrite.
6747 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6748 // Don't insert two casts if they cannot be eliminated. We allow
6749 // two casts to be inserted if the sizes are the same. This could
6750 // only be converting signedness, which is a noop.
6751 if (DestBitSize == SrcBitSize ||
6752 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6753 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6754 Instruction::CastOps opcode = CI.getOpcode();
6755 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6756 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6757 return BinaryOperator::create(
6758 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6762 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6763 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6764 SrcI->getOpcode() == Instruction::Xor &&
6765 Op1 == ConstantInt::getTrue() &&
6766 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6767 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6768 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6771 case Instruction::SDiv:
6772 case Instruction::UDiv:
6773 case Instruction::SRem:
6774 case Instruction::URem:
6775 // If we are just changing the sign, rewrite.
6776 if (DestBitSize == SrcBitSize) {
6777 // Don't insert two casts if they cannot be eliminated. We allow
6778 // two casts to be inserted if the sizes are the same. This could
6779 // only be converting signedness, which is a noop.
6780 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6781 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6782 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6784 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6786 return BinaryOperator::create(
6787 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6792 case Instruction::Shl:
6793 // Allow changing the sign of the source operand. Do not allow
6794 // changing the size of the shift, UNLESS the shift amount is a
6795 // constant. We must not change variable sized shifts to a smaller
6796 // size, because it is undefined to shift more bits out than exist
6798 if (DestBitSize == SrcBitSize ||
6799 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6800 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6801 Instruction::BitCast : Instruction::Trunc);
6802 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6803 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6804 return BinaryOperator::createShl(Op0c, Op1c);
6807 case Instruction::AShr:
6808 // If this is a signed shr, and if all bits shifted in are about to be
6809 // truncated off, turn it into an unsigned shr to allow greater
6811 if (DestBitSize < SrcBitSize &&
6812 isa<ConstantInt>(Op1)) {
6813 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6814 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6815 // Insert the new logical shift right.
6816 return BinaryOperator::createLShr(Op0, Op1);
6824 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6825 if (Instruction *Result = commonIntCastTransforms(CI))
6828 Value *Src = CI.getOperand(0);
6829 const Type *Ty = CI.getType();
6830 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6831 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6833 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6834 switch (SrcI->getOpcode()) {
6836 case Instruction::LShr:
6837 // We can shrink lshr to something smaller if we know the bits shifted in
6838 // are already zeros.
6839 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6840 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6842 // Get a mask for the bits shifting in.
6843 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6844 Value* SrcIOp0 = SrcI->getOperand(0);
6845 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6846 if (ShAmt >= DestBitWidth) // All zeros.
6847 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6849 // Okay, we can shrink this. Truncate the input, then return a new
6851 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6852 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6854 return BinaryOperator::createLShr(V1, V2);
6856 } else { // This is a variable shr.
6858 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6859 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6860 // loop-invariant and CSE'd.
6861 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6862 Value *One = ConstantInt::get(SrcI->getType(), 1);
6864 Value *V = InsertNewInstBefore(
6865 BinaryOperator::createShl(One, SrcI->getOperand(1),
6867 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6868 SrcI->getOperand(0),
6870 Value *Zero = Constant::getNullValue(V->getType());
6871 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6881 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6882 // If one of the common conversion will work ..
6883 if (Instruction *Result = commonIntCastTransforms(CI))
6886 Value *Src = CI.getOperand(0);
6888 // If this is a cast of a cast
6889 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6890 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6891 // types and if the sizes are just right we can convert this into a logical
6892 // 'and' which will be much cheaper than the pair of casts.
6893 if (isa<TruncInst>(CSrc)) {
6894 // Get the sizes of the types involved
6895 Value *A = CSrc->getOperand(0);
6896 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6897 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6898 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6899 // If we're actually extending zero bits and the trunc is a no-op
6900 if (MidSize < DstSize && SrcSize == DstSize) {
6901 // Replace both of the casts with an And of the type mask.
6902 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6903 Constant *AndConst = ConstantInt::get(AndValue);
6905 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6906 // Unfortunately, if the type changed, we need to cast it back.
6907 if (And->getType() != CI.getType()) {
6908 And->setName(CSrc->getName()+".mask");
6909 InsertNewInstBefore(And, CI);
6910 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6917 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6918 // If we are just checking for a icmp eq of a single bit and zext'ing it
6919 // to an integer, then shift the bit to the appropriate place and then
6920 // cast to integer to avoid the comparison.
6921 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6922 const APInt &Op1CV = Op1C->getValue();
6924 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6925 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6926 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6927 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6928 Value *In = ICI->getOperand(0);
6929 Value *Sh = ConstantInt::get(In->getType(),
6930 In->getType()->getPrimitiveSizeInBits()-1);
6931 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6932 In->getName()+".lobit"),
6934 if (In->getType() != CI.getType())
6935 In = CastInst::createIntegerCast(In, CI.getType(),
6936 false/*ZExt*/, "tmp", &CI);
6938 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6939 Constant *One = ConstantInt::get(In->getType(), 1);
6940 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6941 In->getName()+".not"),
6945 return ReplaceInstUsesWith(CI, In);
6950 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6951 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6952 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6953 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6954 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6955 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6956 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6957 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6958 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6959 // This only works for EQ and NE
6960 ICI->isEquality()) {
6961 // If Op1C some other power of two, convert:
6962 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6963 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6964 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6965 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6967 APInt KnownZeroMask(~KnownZero);
6968 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6969 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6970 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6971 // (X&4) == 2 --> false
6972 // (X&4) != 2 --> true
6973 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6974 Res = ConstantExpr::getZExt(Res, CI.getType());
6975 return ReplaceInstUsesWith(CI, Res);
6978 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6979 Value *In = ICI->getOperand(0);
6981 // Perform a logical shr by shiftamt.
6982 // Insert the shift to put the result in the low bit.
6983 In = InsertNewInstBefore(
6984 BinaryOperator::createLShr(In,
6985 ConstantInt::get(In->getType(), ShiftAmt),
6986 In->getName()+".lobit"), CI);
6989 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6990 Constant *One = ConstantInt::get(In->getType(), 1);
6991 In = BinaryOperator::createXor(In, One, "tmp");
6992 InsertNewInstBefore(cast<Instruction>(In), CI);
6995 if (CI.getType() == In->getType())
6996 return ReplaceInstUsesWith(CI, In);
6998 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7006 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7007 if (Instruction *I = commonIntCastTransforms(CI))
7010 Value *Src = CI.getOperand(0);
7012 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7013 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7014 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7015 // If we are just checking for a icmp eq of a single bit and zext'ing it
7016 // to an integer, then shift the bit to the appropriate place and then
7017 // cast to integer to avoid the comparison.
7018 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7019 const APInt &Op1CV = Op1C->getValue();
7021 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7022 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7023 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7024 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7025 Value *In = ICI->getOperand(0);
7026 Value *Sh = ConstantInt::get(In->getType(),
7027 In->getType()->getPrimitiveSizeInBits()-1);
7028 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7029 In->getName()+".lobit"),
7031 if (In->getType() != CI.getType())
7032 In = CastInst::createIntegerCast(In, CI.getType(),
7033 true/*SExt*/, "tmp", &CI);
7035 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7036 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7037 In->getName()+".not"), CI);
7039 return ReplaceInstUsesWith(CI, In);
7047 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7048 return commonCastTransforms(CI);
7051 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7052 return commonCastTransforms(CI);
7055 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7056 return commonCastTransforms(CI);
7059 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7060 return commonCastTransforms(CI);
7063 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7064 return commonCastTransforms(CI);
7067 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7068 return commonCastTransforms(CI);
7071 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7072 return commonPointerCastTransforms(CI);
7075 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7076 return commonCastTransforms(CI);
7079 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7080 // If the operands are integer typed then apply the integer transforms,
7081 // otherwise just apply the common ones.
7082 Value *Src = CI.getOperand(0);
7083 const Type *SrcTy = Src->getType();
7084 const Type *DestTy = CI.getType();
7086 if (SrcTy->isInteger() && DestTy->isInteger()) {
7087 if (Instruction *Result = commonIntCastTransforms(CI))
7089 } else if (isa<PointerType>(SrcTy)) {
7090 if (Instruction *I = commonPointerCastTransforms(CI))
7093 if (Instruction *Result = commonCastTransforms(CI))
7098 // Get rid of casts from one type to the same type. These are useless and can
7099 // be replaced by the operand.
7100 if (DestTy == Src->getType())
7101 return ReplaceInstUsesWith(CI, Src);
7103 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7104 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7105 const Type *DstElTy = DstPTy->getElementType();
7106 const Type *SrcElTy = SrcPTy->getElementType();
7108 // If we are casting a malloc or alloca to a pointer to a type of the same
7109 // size, rewrite the allocation instruction to allocate the "right" type.
7110 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7111 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7114 // If the source and destination are pointers, and this cast is equivalent
7115 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7116 // This can enhance SROA and other transforms that want type-safe pointers.
7117 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7118 unsigned NumZeros = 0;
7119 while (SrcElTy != DstElTy &&
7120 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7121 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7122 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7126 // If we found a path from the src to dest, create the getelementptr now.
7127 if (SrcElTy == DstElTy) {
7128 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7129 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7130 ((Instruction*) NULL));
7134 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7135 if (SVI->hasOneUse()) {
7136 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7137 // a bitconvert to a vector with the same # elts.
7138 if (isa<VectorType>(DestTy) &&
7139 cast<VectorType>(DestTy)->getNumElements() ==
7140 SVI->getType()->getNumElements()) {
7142 // If either of the operands is a cast from CI.getType(), then
7143 // evaluating the shuffle in the casted destination's type will allow
7144 // us to eliminate at least one cast.
7145 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7146 Tmp->getOperand(0)->getType() == DestTy) ||
7147 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7148 Tmp->getOperand(0)->getType() == DestTy)) {
7149 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7150 SVI->getOperand(0), DestTy, &CI);
7151 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7152 SVI->getOperand(1), DestTy, &CI);
7153 // Return a new shuffle vector. Use the same element ID's, as we
7154 // know the vector types match #elts.
7155 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7163 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7165 /// %D = select %cond, %C, %A
7167 /// %C = select %cond, %B, 0
7170 /// Assuming that the specified instruction is an operand to the select, return
7171 /// a bitmask indicating which operands of this instruction are foldable if they
7172 /// equal the other incoming value of the select.
7174 static unsigned GetSelectFoldableOperands(Instruction *I) {
7175 switch (I->getOpcode()) {
7176 case Instruction::Add:
7177 case Instruction::Mul:
7178 case Instruction::And:
7179 case Instruction::Or:
7180 case Instruction::Xor:
7181 return 3; // Can fold through either operand.
7182 case Instruction::Sub: // Can only fold on the amount subtracted.
7183 case Instruction::Shl: // Can only fold on the shift amount.
7184 case Instruction::LShr:
7185 case Instruction::AShr:
7188 return 0; // Cannot fold
7192 /// GetSelectFoldableConstant - For the same transformation as the previous
7193 /// function, return the identity constant that goes into the select.
7194 static Constant *GetSelectFoldableConstant(Instruction *I) {
7195 switch (I->getOpcode()) {
7196 default: assert(0 && "This cannot happen!"); abort();
7197 case Instruction::Add:
7198 case Instruction::Sub:
7199 case Instruction::Or:
7200 case Instruction::Xor:
7201 case Instruction::Shl:
7202 case Instruction::LShr:
7203 case Instruction::AShr:
7204 return Constant::getNullValue(I->getType());
7205 case Instruction::And:
7206 return Constant::getAllOnesValue(I->getType());
7207 case Instruction::Mul:
7208 return ConstantInt::get(I->getType(), 1);
7212 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7213 /// have the same opcode and only one use each. Try to simplify this.
7214 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7216 if (TI->getNumOperands() == 1) {
7217 // If this is a non-volatile load or a cast from the same type,
7220 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7223 return 0; // unknown unary op.
7226 // Fold this by inserting a select from the input values.
7227 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7228 FI->getOperand(0), SI.getName()+".v");
7229 InsertNewInstBefore(NewSI, SI);
7230 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7234 // Only handle binary operators here.
7235 if (!isa<BinaryOperator>(TI))
7238 // Figure out if the operations have any operands in common.
7239 Value *MatchOp, *OtherOpT, *OtherOpF;
7241 if (TI->getOperand(0) == FI->getOperand(0)) {
7242 MatchOp = TI->getOperand(0);
7243 OtherOpT = TI->getOperand(1);
7244 OtherOpF = FI->getOperand(1);
7245 MatchIsOpZero = true;
7246 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7247 MatchOp = TI->getOperand(1);
7248 OtherOpT = TI->getOperand(0);
7249 OtherOpF = FI->getOperand(0);
7250 MatchIsOpZero = false;
7251 } else if (!TI->isCommutative()) {
7253 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7254 MatchOp = TI->getOperand(0);
7255 OtherOpT = TI->getOperand(1);
7256 OtherOpF = FI->getOperand(0);
7257 MatchIsOpZero = true;
7258 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7259 MatchOp = TI->getOperand(1);
7260 OtherOpT = TI->getOperand(0);
7261 OtherOpF = FI->getOperand(1);
7262 MatchIsOpZero = true;
7267 // If we reach here, they do have operations in common.
7268 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7269 OtherOpF, SI.getName()+".v");
7270 InsertNewInstBefore(NewSI, SI);
7272 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7274 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7276 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7278 assert(0 && "Shouldn't get here");
7282 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7283 Value *CondVal = SI.getCondition();
7284 Value *TrueVal = SI.getTrueValue();
7285 Value *FalseVal = SI.getFalseValue();
7287 // select true, X, Y -> X
7288 // select false, X, Y -> Y
7289 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7290 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7292 // select C, X, X -> X
7293 if (TrueVal == FalseVal)
7294 return ReplaceInstUsesWith(SI, TrueVal);
7296 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7297 return ReplaceInstUsesWith(SI, FalseVal);
7298 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7299 return ReplaceInstUsesWith(SI, TrueVal);
7300 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7301 if (isa<Constant>(TrueVal))
7302 return ReplaceInstUsesWith(SI, TrueVal);
7304 return ReplaceInstUsesWith(SI, FalseVal);
7307 if (SI.getType() == Type::Int1Ty) {
7308 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7309 if (C->getZExtValue()) {
7310 // Change: A = select B, true, C --> A = or B, C
7311 return BinaryOperator::createOr(CondVal, FalseVal);
7313 // Change: A = select B, false, C --> A = and !B, C
7315 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7316 "not."+CondVal->getName()), SI);
7317 return BinaryOperator::createAnd(NotCond, FalseVal);
7319 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7320 if (C->getZExtValue() == false) {
7321 // Change: A = select B, C, false --> A = and B, C
7322 return BinaryOperator::createAnd(CondVal, TrueVal);
7324 // Change: A = select B, C, true --> A = or !B, C
7326 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7327 "not."+CondVal->getName()), SI);
7328 return BinaryOperator::createOr(NotCond, TrueVal);
7332 // select a, b, a -> a&b
7333 // select a, a, b -> a|b
7334 if (CondVal == TrueVal)
7335 return BinaryOperator::createOr(CondVal, FalseVal);
7336 else if (CondVal == FalseVal)
7337 return BinaryOperator::createAnd(CondVal, TrueVal);
7340 // Selecting between two integer constants?
7341 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7342 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7343 // select C, 1, 0 -> zext C to int
7344 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7345 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7346 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7347 // select C, 0, 1 -> zext !C to int
7349 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7350 "not."+CondVal->getName()), SI);
7351 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7354 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7356 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7358 // (x <s 0) ? -1 : 0 -> ashr x, 31
7359 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7360 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7361 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7362 // The comparison constant and the result are not neccessarily the
7363 // same width. Make an all-ones value by inserting a AShr.
7364 Value *X = IC->getOperand(0);
7365 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7366 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7367 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7369 InsertNewInstBefore(SRA, SI);
7371 // Finally, convert to the type of the select RHS. We figure out
7372 // if this requires a SExt, Trunc or BitCast based on the sizes.
7373 Instruction::CastOps opc = Instruction::BitCast;
7374 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7375 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7376 if (SRASize < SISize)
7377 opc = Instruction::SExt;
7378 else if (SRASize > SISize)
7379 opc = Instruction::Trunc;
7380 return CastInst::create(opc, SRA, SI.getType());
7385 // If one of the constants is zero (we know they can't both be) and we
7386 // have an icmp instruction with zero, and we have an 'and' with the
7387 // non-constant value, eliminate this whole mess. This corresponds to
7388 // cases like this: ((X & 27) ? 27 : 0)
7389 if (TrueValC->isZero() || FalseValC->isZero())
7390 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7391 cast<Constant>(IC->getOperand(1))->isNullValue())
7392 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7393 if (ICA->getOpcode() == Instruction::And &&
7394 isa<ConstantInt>(ICA->getOperand(1)) &&
7395 (ICA->getOperand(1) == TrueValC ||
7396 ICA->getOperand(1) == FalseValC) &&
7397 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7398 // Okay, now we know that everything is set up, we just don't
7399 // know whether we have a icmp_ne or icmp_eq and whether the
7400 // true or false val is the zero.
7401 bool ShouldNotVal = !TrueValC->isZero();
7402 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7405 V = InsertNewInstBefore(BinaryOperator::create(
7406 Instruction::Xor, V, ICA->getOperand(1)), SI);
7407 return ReplaceInstUsesWith(SI, V);
7412 // See if we are selecting two values based on a comparison of the two values.
7413 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7414 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7415 // Transform (X == Y) ? X : Y -> Y
7416 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7417 // This is not safe in general for floating point:
7418 // consider X== -0, Y== +0.
7419 // It becomes safe if either operand is a nonzero constant.
7420 ConstantFP *CFPt, *CFPf;
7421 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7422 !CFPt->getValueAPF().isZero()) ||
7423 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7424 !CFPf->getValueAPF().isZero()))
7425 return ReplaceInstUsesWith(SI, FalseVal);
7427 // Transform (X != Y) ? X : Y -> X
7428 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7429 return ReplaceInstUsesWith(SI, TrueVal);
7430 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7432 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7433 // Transform (X == Y) ? Y : X -> X
7434 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7435 // This is not safe in general for floating point:
7436 // consider X== -0, Y== +0.
7437 // It becomes safe if either operand is a nonzero constant.
7438 ConstantFP *CFPt, *CFPf;
7439 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7440 !CFPt->getValueAPF().isZero()) ||
7441 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7442 !CFPf->getValueAPF().isZero()))
7443 return ReplaceInstUsesWith(SI, FalseVal);
7445 // Transform (X != Y) ? Y : X -> Y
7446 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7447 return ReplaceInstUsesWith(SI, TrueVal);
7448 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7452 // See if we are selecting two values based on a comparison of the two values.
7453 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7454 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7455 // Transform (X == Y) ? X : Y -> Y
7456 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7457 return ReplaceInstUsesWith(SI, FalseVal);
7458 // Transform (X != Y) ? X : Y -> X
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.
7463 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7464 // Transform (X == Y) ? Y : X -> X
7465 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7466 return ReplaceInstUsesWith(SI, FalseVal);
7467 // Transform (X != Y) ? Y : X -> Y
7468 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7469 return ReplaceInstUsesWith(SI, TrueVal);
7470 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7474 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7475 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7476 if (TI->hasOneUse() && FI->hasOneUse()) {
7477 Instruction *AddOp = 0, *SubOp = 0;
7479 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7480 if (TI->getOpcode() == FI->getOpcode())
7481 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7484 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7485 // even legal for FP.
7486 if (TI->getOpcode() == Instruction::Sub &&
7487 FI->getOpcode() == Instruction::Add) {
7488 AddOp = FI; SubOp = TI;
7489 } else if (FI->getOpcode() == Instruction::Sub &&
7490 TI->getOpcode() == Instruction::Add) {
7491 AddOp = TI; SubOp = FI;
7495 Value *OtherAddOp = 0;
7496 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7497 OtherAddOp = AddOp->getOperand(1);
7498 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7499 OtherAddOp = AddOp->getOperand(0);
7503 // So at this point we know we have (Y -> OtherAddOp):
7504 // select C, (add X, Y), (sub X, Z)
7505 Value *NegVal; // Compute -Z
7506 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7507 NegVal = ConstantExpr::getNeg(C);
7509 NegVal = InsertNewInstBefore(
7510 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7513 Value *NewTrueOp = OtherAddOp;
7514 Value *NewFalseOp = NegVal;
7516 std::swap(NewTrueOp, NewFalseOp);
7517 Instruction *NewSel =
7518 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7520 NewSel = InsertNewInstBefore(NewSel, SI);
7521 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7526 // See if we can fold the select into one of our operands.
7527 if (SI.getType()->isInteger()) {
7528 // See the comment above GetSelectFoldableOperands for a description of the
7529 // transformation we are doing here.
7530 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7531 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7532 !isa<Constant>(FalseVal))
7533 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7534 unsigned OpToFold = 0;
7535 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7537 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7542 Constant *C = GetSelectFoldableConstant(TVI);
7543 Instruction *NewSel =
7544 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7545 InsertNewInstBefore(NewSel, SI);
7546 NewSel->takeName(TVI);
7547 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7548 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7550 assert(0 && "Unknown instruction!!");
7555 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7556 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7557 !isa<Constant>(TrueVal))
7558 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7559 unsigned OpToFold = 0;
7560 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7562 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7567 Constant *C = GetSelectFoldableConstant(FVI);
7568 Instruction *NewSel =
7569 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7570 InsertNewInstBefore(NewSel, SI);
7571 NewSel->takeName(FVI);
7572 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7573 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7575 assert(0 && "Unknown instruction!!");
7580 if (BinaryOperator::isNot(CondVal)) {
7581 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7582 SI.setOperand(1, FalseVal);
7583 SI.setOperand(2, TrueVal);
7590 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7591 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7592 /// and it is more than the alignment of the ultimate object, see if we can
7593 /// increase the alignment of the ultimate object, making this check succeed.
7594 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7595 unsigned PrefAlign = 0) {
7596 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7597 unsigned Align = GV->getAlignment();
7598 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7599 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7601 // If there is a large requested alignment and we can, bump up the alignment
7603 if (PrefAlign > Align && GV->hasInitializer()) {
7604 GV->setAlignment(PrefAlign);
7608 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7609 unsigned Align = AI->getAlignment();
7610 if (Align == 0 && TD) {
7611 if (isa<AllocaInst>(AI))
7612 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7613 else if (isa<MallocInst>(AI)) {
7614 // Malloc returns maximally aligned memory.
7615 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7618 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7621 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7625 // If there is a requested alignment and if this is an alloca, round up. We
7626 // don't do this for malloc, because some systems can't respect the request.
7627 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7628 AI->setAlignment(PrefAlign);
7632 } else if (isa<BitCastInst>(V) ||
7633 (isa<ConstantExpr>(V) &&
7634 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7635 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7637 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7638 // If all indexes are zero, it is just the alignment of the base pointer.
7639 bool AllZeroOperands = true;
7640 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7641 if (!isa<Constant>(GEPI->getOperand(i)) ||
7642 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7643 AllZeroOperands = false;
7647 if (AllZeroOperands) {
7648 // Treat this like a bitcast.
7649 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7652 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7653 if (BaseAlignment == 0) return 0;
7655 // Otherwise, if the base alignment is >= the alignment we expect for the
7656 // base pointer type, then we know that the resultant pointer is aligned at
7657 // least as much as its type requires.
7660 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7661 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7662 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7663 if (Align <= BaseAlignment) {
7664 const Type *GEPTy = GEPI->getType();
7665 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7666 Align = std::min(Align, (unsigned)
7667 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7676 /// visitCallInst - CallInst simplification. This mostly only handles folding
7677 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7678 /// the heavy lifting.
7680 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7681 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7682 if (!II) return visitCallSite(&CI);
7684 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7686 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7687 bool Changed = false;
7689 // memmove/cpy/set of zero bytes is a noop.
7690 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7691 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7693 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7694 if (CI->getZExtValue() == 1) {
7695 // Replace the instruction with just byte operations. We would
7696 // transform other cases to loads/stores, but we don't know if
7697 // alignment is sufficient.
7701 // If we have a memmove and the source operation is a constant global,
7702 // then the source and dest pointers can't alias, so we can change this
7703 // into a call to memcpy.
7704 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7705 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7706 if (GVSrc->isConstant()) {
7707 Module *M = CI.getParent()->getParent()->getParent();
7709 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7711 Name = "llvm.memcpy.i32";
7713 Name = "llvm.memcpy.i64";
7714 Constant *MemCpy = M->getOrInsertFunction(Name,
7715 CI.getCalledFunction()->getFunctionType());
7716 CI.setOperand(0, MemCpy);
7721 // If we can determine a pointer alignment that is bigger than currently
7722 // set, update the alignment.
7723 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7724 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7725 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7726 unsigned Align = std::min(Alignment1, Alignment2);
7727 if (MI->getAlignment()->getZExtValue() < Align) {
7728 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7732 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7734 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(CI.getOperand(3));
7736 unsigned Size = MemOpLength->getZExtValue();
7737 unsigned Align = cast<ConstantInt>(CI.getOperand(4))->getZExtValue();
7738 PointerType *NewPtrTy = NULL;
7739 // Destination pointer type is always i8 *
7740 // If Size is 8 then use Int64Ty
7741 // If Size is 4 then use Int32Ty
7742 // If Size is 2 then use Int16Ty
7743 // If Size is 1 then use Int8Ty
7744 if (Size && Size <=8 && !(Size&(Size-1)))
7745 NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7748 Value *Src = InsertCastBefore(Instruction::BitCast, CI.getOperand(2),
7750 Value *Dest = InsertCastBefore(Instruction::BitCast, CI.getOperand(1),
7752 Value *L = new LoadInst(Src, "tmp", false, Align, &CI);
7753 Value *NS = new StoreInst(L, Dest, false, Align, &CI);
7754 CI.replaceAllUsesWith(NS);
7756 return EraseInstFromFunction(CI);
7759 } else if (isa<MemSetInst>(MI)) {
7760 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7761 if (MI->getAlignment()->getZExtValue() < Alignment) {
7762 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7767 if (Changed) return II;
7769 switch (II->getIntrinsicID()) {
7771 case Intrinsic::ppc_altivec_lvx:
7772 case Intrinsic::ppc_altivec_lvxl:
7773 case Intrinsic::x86_sse_loadu_ps:
7774 case Intrinsic::x86_sse2_loadu_pd:
7775 case Intrinsic::x86_sse2_loadu_dq:
7776 // Turn PPC lvx -> load if the pointer is known aligned.
7777 // Turn X86 loadups -> load if the pointer is known aligned.
7778 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7780 InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7781 PointerType::getUnqual(II->getType()), CI);
7782 return new LoadInst(Ptr);
7785 case Intrinsic::ppc_altivec_stvx:
7786 case Intrinsic::ppc_altivec_stvxl:
7787 // Turn stvx -> store if the pointer is known aligned.
7788 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7789 const Type *OpPtrTy =
7790 PointerType::getUnqual(II->getOperand(1)->getType());
7791 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7793 return new StoreInst(II->getOperand(1), Ptr);
7796 case Intrinsic::x86_sse_storeu_ps:
7797 case Intrinsic::x86_sse2_storeu_pd:
7798 case Intrinsic::x86_sse2_storeu_dq:
7799 case Intrinsic::x86_sse2_storel_dq:
7800 // Turn X86 storeu -> store if the pointer is known aligned.
7801 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7802 const Type *OpPtrTy =
7803 PointerType::getUnqual(II->getOperand(2)->getType());
7804 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7806 return new StoreInst(II->getOperand(2), Ptr);
7810 case Intrinsic::x86_sse_cvttss2si: {
7811 // These intrinsics only demands the 0th element of its input vector. If
7812 // we can simplify the input based on that, do so now.
7814 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7816 II->setOperand(1, V);
7822 case Intrinsic::ppc_altivec_vperm:
7823 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7824 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7825 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7827 // Check that all of the elements are integer constants or undefs.
7828 bool AllEltsOk = true;
7829 for (unsigned i = 0; i != 16; ++i) {
7830 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7831 !isa<UndefValue>(Mask->getOperand(i))) {
7838 // Cast the input vectors to byte vectors.
7839 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7840 II->getOperand(1), Mask->getType(), CI);
7841 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7842 II->getOperand(2), Mask->getType(), CI);
7843 Value *Result = UndefValue::get(Op0->getType());
7845 // Only extract each element once.
7846 Value *ExtractedElts[32];
7847 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7849 for (unsigned i = 0; i != 16; ++i) {
7850 if (isa<UndefValue>(Mask->getOperand(i)))
7852 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7853 Idx &= 31; // Match the hardware behavior.
7855 if (ExtractedElts[Idx] == 0) {
7857 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7858 InsertNewInstBefore(Elt, CI);
7859 ExtractedElts[Idx] = Elt;
7862 // Insert this value into the result vector.
7863 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7864 InsertNewInstBefore(cast<Instruction>(Result), CI);
7866 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7871 case Intrinsic::stackrestore: {
7872 // If the save is right next to the restore, remove the restore. This can
7873 // happen when variable allocas are DCE'd.
7874 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7875 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7876 BasicBlock::iterator BI = SS;
7878 return EraseInstFromFunction(CI);
7882 // If the stack restore is in a return/unwind block and if there are no
7883 // allocas or calls between the restore and the return, nuke the restore.
7884 TerminatorInst *TI = II->getParent()->getTerminator();
7885 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7886 BasicBlock::iterator BI = II;
7887 bool CannotRemove = false;
7888 for (++BI; &*BI != TI; ++BI) {
7889 if (isa<AllocaInst>(BI) ||
7890 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7891 CannotRemove = true;
7896 return EraseInstFromFunction(CI);
7903 return visitCallSite(II);
7906 // InvokeInst simplification
7908 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7909 return visitCallSite(&II);
7912 // visitCallSite - Improvements for call and invoke instructions.
7914 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7915 bool Changed = false;
7917 // If the callee is a constexpr cast of a function, attempt to move the cast
7918 // to the arguments of the call/invoke.
7919 if (transformConstExprCastCall(CS)) return 0;
7921 Value *Callee = CS.getCalledValue();
7923 if (Function *CalleeF = dyn_cast<Function>(Callee))
7924 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7925 Instruction *OldCall = CS.getInstruction();
7926 // If the call and callee calling conventions don't match, this call must
7927 // be unreachable, as the call is undefined.
7928 new StoreInst(ConstantInt::getTrue(),
7929 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
7931 if (!OldCall->use_empty())
7932 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7933 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7934 return EraseInstFromFunction(*OldCall);
7938 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7939 // This instruction is not reachable, just remove it. We insert a store to
7940 // undef so that we know that this code is not reachable, despite the fact
7941 // that we can't modify the CFG here.
7942 new StoreInst(ConstantInt::getTrue(),
7943 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
7944 CS.getInstruction());
7946 if (!CS.getInstruction()->use_empty())
7947 CS.getInstruction()->
7948 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7950 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7951 // Don't break the CFG, insert a dummy cond branch.
7952 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7953 ConstantInt::getTrue(), II);
7955 return EraseInstFromFunction(*CS.getInstruction());
7958 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
7959 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
7960 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
7961 return transformCallThroughTrampoline(CS);
7963 const PointerType *PTy = cast<PointerType>(Callee->getType());
7964 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7965 if (FTy->isVarArg()) {
7966 // See if we can optimize any arguments passed through the varargs area of
7968 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7969 E = CS.arg_end(); I != E; ++I)
7970 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7971 // If this cast does not effect the value passed through the varargs
7972 // area, we can eliminate the use of the cast.
7973 Value *Op = CI->getOperand(0);
7974 if (CI->isLosslessCast()) {
7981 if (isa<InlineAsm>(Callee) && !CS.isNoUnwind()) {
7982 // Inline asm calls cannot throw - mark them 'nounwind'.
7983 const ParamAttrsList *PAL = CS.getParamAttrs();
7984 uint16_t RAttributes = PAL ? PAL->getParamAttrs(0) : 0;
7985 RAttributes |= ParamAttr::NoUnwind;
7987 ParamAttrsVector modVec;
7988 modVec.push_back(ParamAttrsWithIndex::get(0, RAttributes));
7989 PAL = ParamAttrsList::getModified(PAL, modVec);
7990 CS.setParamAttrs(PAL);
7994 return Changed ? CS.getInstruction() : 0;
7997 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7998 // attempt to move the cast to the arguments of the call/invoke.
8000 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8001 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8002 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8003 if (CE->getOpcode() != Instruction::BitCast ||
8004 !isa<Function>(CE->getOperand(0)))
8006 Function *Callee = cast<Function>(CE->getOperand(0));
8007 Instruction *Caller = CS.getInstruction();
8009 // Okay, this is a cast from a function to a different type. Unless doing so
8010 // would cause a type conversion of one of our arguments, change this call to
8011 // be a direct call with arguments casted to the appropriate types.
8013 const FunctionType *FT = Callee->getFunctionType();
8014 const Type *OldRetTy = Caller->getType();
8016 const ParamAttrsList* CallerPAL = 0;
8017 if (CallInst *CallerCI = dyn_cast<CallInst>(Caller))
8018 CallerPAL = CallerCI->getParamAttrs();
8019 else if (InvokeInst *CallerII = dyn_cast<InvokeInst>(Caller))
8020 CallerPAL = CallerII->getParamAttrs();
8022 // If the parameter attributes are not compatible, don't do the xform. We
8023 // don't want to lose an sret attribute or something.
8024 if (!ParamAttrsList::areCompatible(CallerPAL, Callee->getParamAttrs()))
8027 // Check to see if we are changing the return type...
8028 if (OldRetTy != FT->getReturnType()) {
8029 if (Callee->isDeclaration() && !Caller->use_empty() &&
8030 // Conversion is ok if changing from pointer to int of same size.
8031 !(isa<PointerType>(FT->getReturnType()) &&
8032 TD->getIntPtrType() == OldRetTy))
8033 return false; // Cannot transform this return value.
8035 // If the callsite is an invoke instruction, and the return value is used by
8036 // a PHI node in a successor, we cannot change the return type of the call
8037 // because there is no place to put the cast instruction (without breaking
8038 // the critical edge). Bail out in this case.
8039 if (!Caller->use_empty())
8040 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8041 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8043 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8044 if (PN->getParent() == II->getNormalDest() ||
8045 PN->getParent() == II->getUnwindDest())
8049 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8050 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8052 CallSite::arg_iterator AI = CS.arg_begin();
8053 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8054 const Type *ParamTy = FT->getParamType(i);
8055 const Type *ActTy = (*AI)->getType();
8056 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8057 //Some conversions are safe even if we do not have a body.
8058 //Either we can cast directly, or we can upconvert the argument
8059 bool isConvertible = ActTy == ParamTy ||
8060 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8061 (ParamTy->isInteger() && ActTy->isInteger() &&
8062 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8063 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8064 && c->getValue().isStrictlyPositive());
8065 if (Callee->isDeclaration() && !isConvertible) return false;
8067 // Most other conversions can be done if we have a body, even if these
8068 // lose information, e.g. int->short.
8069 // Some conversions cannot be done at all, e.g. float to pointer.
8070 // Logic here parallels CastInst::getCastOpcode (the design there
8071 // requires legality checks like this be done before calling it).
8072 if (ParamTy->isInteger()) {
8073 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8074 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8077 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
8078 !isa<PointerType>(ActTy))
8080 } else if (ParamTy->isFloatingPoint()) {
8081 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8082 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8085 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
8087 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
8088 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8089 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
8092 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
8094 } else if (isa<PointerType>(ParamTy)) {
8095 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
8102 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8103 Callee->isDeclaration())
8104 return false; // Do not delete arguments unless we have a function body...
8106 // Okay, we decided that this is a safe thing to do: go ahead and start
8107 // inserting cast instructions as necessary...
8108 std::vector<Value*> Args;
8109 Args.reserve(NumActualArgs);
8111 AI = CS.arg_begin();
8112 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8113 const Type *ParamTy = FT->getParamType(i);
8114 if ((*AI)->getType() == ParamTy) {
8115 Args.push_back(*AI);
8117 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8118 false, ParamTy, false);
8119 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8120 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8124 // If the function takes more arguments than the call was taking, add them
8126 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8127 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8129 // If we are removing arguments to the function, emit an obnoxious warning...
8130 if (FT->getNumParams() < NumActualArgs)
8131 if (!FT->isVarArg()) {
8132 cerr << "WARNING: While resolving call to function '"
8133 << Callee->getName() << "' arguments were dropped!\n";
8135 // Add all of the arguments in their promoted form to the arg list...
8136 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8137 const Type *PTy = getPromotedType((*AI)->getType());
8138 if (PTy != (*AI)->getType()) {
8139 // Must promote to pass through va_arg area!
8140 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8142 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8143 InsertNewInstBefore(Cast, *Caller);
8144 Args.push_back(Cast);
8146 Args.push_back(*AI);
8151 if (FT->getReturnType() == Type::VoidTy)
8152 Caller->setName(""); // Void type should not have a name.
8155 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8156 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8157 Args.begin(), Args.end(), Caller->getName(), Caller);
8158 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8159 cast<InvokeInst>(NC)->setParamAttrs(CallerPAL);
8161 NC = new CallInst(Callee, Args.begin(), Args.end(),
8162 Caller->getName(), Caller);
8163 CallInst *CI = cast<CallInst>(Caller);
8164 if (CI->isTailCall())
8165 cast<CallInst>(NC)->setTailCall();
8166 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8167 cast<CallInst>(NC)->setParamAttrs(CallerPAL);
8170 // Insert a cast of the return type as necessary.
8172 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8173 if (NV->getType() != Type::VoidTy) {
8174 const Type *CallerTy = Caller->getType();
8175 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8177 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8179 // If this is an invoke instruction, we should insert it after the first
8180 // non-phi, instruction in the normal successor block.
8181 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8182 BasicBlock::iterator I = II->getNormalDest()->begin();
8183 while (isa<PHINode>(I)) ++I;
8184 InsertNewInstBefore(NC, *I);
8186 // Otherwise, it's a call, just insert cast right after the call instr
8187 InsertNewInstBefore(NC, *Caller);
8189 AddUsersToWorkList(*Caller);
8191 NV = UndefValue::get(Caller->getType());
8195 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8196 Caller->replaceAllUsesWith(NV);
8197 Caller->eraseFromParent();
8198 RemoveFromWorkList(Caller);
8202 // transformCallThroughTrampoline - Turn a call to a function created by the
8203 // init_trampoline intrinsic into a direct call to the underlying function.
8205 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8206 Value *Callee = CS.getCalledValue();
8207 const PointerType *PTy = cast<PointerType>(Callee->getType());
8208 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8210 IntrinsicInst *Tramp =
8211 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8214 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8215 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8216 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8218 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8219 unsigned NestIdx = 1;
8220 const Type *NestTy = 0;
8221 uint16_t NestAttr = 0;
8223 // Look for a parameter marked with the 'nest' attribute.
8224 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8225 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8226 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8227 // Record the parameter type and any other attributes.
8229 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8234 Instruction *Caller = CS.getInstruction();
8235 std::vector<Value*> NewArgs;
8236 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8238 // Insert the nest argument into the call argument list, which may
8239 // mean appending it.
8242 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8244 if (Idx == NestIdx) {
8245 // Add the chain argument.
8246 Value *NestVal = Tramp->getOperand(3);
8247 if (NestVal->getType() != NestTy)
8248 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8249 NewArgs.push_back(NestVal);
8255 // Add the original argument.
8256 NewArgs.push_back(*I);
8262 // The trampoline may have been bitcast to a bogus type (FTy).
8263 // Handle this by synthesizing a new function type, equal to FTy
8264 // with the chain parameter inserted. Likewise for attributes.
8266 const ParamAttrsList *Attrs = CS.getParamAttrs();
8267 std::vector<const Type*> NewTypes;
8268 ParamAttrsVector NewAttrs;
8269 NewTypes.reserve(FTy->getNumParams()+1);
8271 // Add any function result attributes.
8272 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8274 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8276 // Insert the chain's type into the list of parameter types, which may
8277 // mean appending it. Likewise for the chain's attributes.
8280 FunctionType::param_iterator I = FTy->param_begin(),
8281 E = FTy->param_end();
8284 if (Idx == NestIdx) {
8285 // Add the chain's type and attributes.
8286 NewTypes.push_back(NestTy);
8287 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8293 // Add the original type and attributes.
8294 NewTypes.push_back(*I);
8295 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8298 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8304 // Replace the trampoline call with a direct call. Let the generic
8305 // code sort out any function type mismatches.
8306 FunctionType *NewFTy =
8307 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8308 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8309 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8310 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8312 Instruction *NewCaller;
8313 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8314 NewCaller = new InvokeInst(NewCallee,
8315 II->getNormalDest(), II->getUnwindDest(),
8316 NewArgs.begin(), NewArgs.end(),
8317 Caller->getName(), Caller);
8318 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8319 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8321 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8322 Caller->getName(), Caller);
8323 if (cast<CallInst>(Caller)->isTailCall())
8324 cast<CallInst>(NewCaller)->setTailCall();
8325 cast<CallInst>(NewCaller)->
8326 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8327 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8329 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8330 Caller->replaceAllUsesWith(NewCaller);
8331 Caller->eraseFromParent();
8332 RemoveFromWorkList(Caller);
8337 // Replace the trampoline call with a direct call. Since there is no 'nest'
8338 // parameter, there is no need to adjust the argument list. Let the generic
8339 // code sort out any function type mismatches.
8340 Constant *NewCallee =
8341 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8342 CS.setCalledFunction(NewCallee);
8343 return CS.getInstruction();
8346 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8347 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8348 /// and a single binop.
8349 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8350 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8351 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8352 isa<CmpInst>(FirstInst));
8353 unsigned Opc = FirstInst->getOpcode();
8354 Value *LHSVal = FirstInst->getOperand(0);
8355 Value *RHSVal = FirstInst->getOperand(1);
8357 const Type *LHSType = LHSVal->getType();
8358 const Type *RHSType = RHSVal->getType();
8360 // Scan to see if all operands are the same opcode, all have one use, and all
8361 // kill their operands (i.e. the operands have one use).
8362 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8363 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8364 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8365 // Verify type of the LHS matches so we don't fold cmp's of different
8366 // types or GEP's with different index types.
8367 I->getOperand(0)->getType() != LHSType ||
8368 I->getOperand(1)->getType() != RHSType)
8371 // If they are CmpInst instructions, check their predicates
8372 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8373 if (cast<CmpInst>(I)->getPredicate() !=
8374 cast<CmpInst>(FirstInst)->getPredicate())
8377 // Keep track of which operand needs a phi node.
8378 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8379 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8382 // Otherwise, this is safe to transform, determine if it is profitable.
8384 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8385 // Indexes are often folded into load/store instructions, so we don't want to
8386 // hide them behind a phi.
8387 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8390 Value *InLHS = FirstInst->getOperand(0);
8391 Value *InRHS = FirstInst->getOperand(1);
8392 PHINode *NewLHS = 0, *NewRHS = 0;
8394 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8395 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8396 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8397 InsertNewInstBefore(NewLHS, PN);
8402 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8403 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8404 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8405 InsertNewInstBefore(NewRHS, PN);
8409 // Add all operands to the new PHIs.
8410 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8412 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8413 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8416 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8417 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8421 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8422 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8423 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8424 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8427 assert(isa<GetElementPtrInst>(FirstInst));
8428 return new GetElementPtrInst(LHSVal, RHSVal);
8432 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8433 /// of the block that defines it. This means that it must be obvious the value
8434 /// of the load is not changed from the point of the load to the end of the
8437 /// Finally, it is safe, but not profitable, to sink a load targetting a
8438 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8440 static bool isSafeToSinkLoad(LoadInst *L) {
8441 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8443 for (++BBI; BBI != E; ++BBI)
8444 if (BBI->mayWriteToMemory())
8447 // Check for non-address taken alloca. If not address-taken already, it isn't
8448 // profitable to do this xform.
8449 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8450 bool isAddressTaken = false;
8451 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8453 if (isa<LoadInst>(UI)) continue;
8454 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8455 // If storing TO the alloca, then the address isn't taken.
8456 if (SI->getOperand(1) == AI) continue;
8458 isAddressTaken = true;
8462 if (!isAddressTaken)
8470 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8471 // operator and they all are only used by the PHI, PHI together their
8472 // inputs, and do the operation once, to the result of the PHI.
8473 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8474 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8476 // Scan the instruction, looking for input operations that can be folded away.
8477 // If all input operands to the phi are the same instruction (e.g. a cast from
8478 // the same type or "+42") we can pull the operation through the PHI, reducing
8479 // code size and simplifying code.
8480 Constant *ConstantOp = 0;
8481 const Type *CastSrcTy = 0;
8482 bool isVolatile = false;
8483 if (isa<CastInst>(FirstInst)) {
8484 CastSrcTy = FirstInst->getOperand(0)->getType();
8485 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8486 // Can fold binop, compare or shift here if the RHS is a constant,
8487 // otherwise call FoldPHIArgBinOpIntoPHI.
8488 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8489 if (ConstantOp == 0)
8490 return FoldPHIArgBinOpIntoPHI(PN);
8491 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8492 isVolatile = LI->isVolatile();
8493 // We can't sink the load if the loaded value could be modified between the
8494 // load and the PHI.
8495 if (LI->getParent() != PN.getIncomingBlock(0) ||
8496 !isSafeToSinkLoad(LI))
8498 } else if (isa<GetElementPtrInst>(FirstInst)) {
8499 if (FirstInst->getNumOperands() == 2)
8500 return FoldPHIArgBinOpIntoPHI(PN);
8501 // Can't handle general GEPs yet.
8504 return 0; // Cannot fold this operation.
8507 // Check to see if all arguments are the same operation.
8508 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8509 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8510 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8511 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8514 if (I->getOperand(0)->getType() != CastSrcTy)
8515 return 0; // Cast operation must match.
8516 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8517 // We can't sink the load if the loaded value could be modified between
8518 // the load and the PHI.
8519 if (LI->isVolatile() != isVolatile ||
8520 LI->getParent() != PN.getIncomingBlock(i) ||
8521 !isSafeToSinkLoad(LI))
8523 } else if (I->getOperand(1) != ConstantOp) {
8528 // Okay, they are all the same operation. Create a new PHI node of the
8529 // correct type, and PHI together all of the LHS's of the instructions.
8530 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8531 PN.getName()+".in");
8532 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8534 Value *InVal = FirstInst->getOperand(0);
8535 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8537 // Add all operands to the new PHI.
8538 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8539 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8540 if (NewInVal != InVal)
8542 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8547 // The new PHI unions all of the same values together. This is really
8548 // common, so we handle it intelligently here for compile-time speed.
8552 InsertNewInstBefore(NewPN, PN);
8556 // Insert and return the new operation.
8557 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8558 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8559 else if (isa<LoadInst>(FirstInst))
8560 return new LoadInst(PhiVal, "", isVolatile);
8561 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8562 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8563 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8564 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8565 PhiVal, ConstantOp);
8567 assert(0 && "Unknown operation");
8571 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8573 static bool DeadPHICycle(PHINode *PN,
8574 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8575 if (PN->use_empty()) return true;
8576 if (!PN->hasOneUse()) return false;
8578 // Remember this node, and if we find the cycle, return.
8579 if (!PotentiallyDeadPHIs.insert(PN))
8582 // Don't scan crazily complex things.
8583 if (PotentiallyDeadPHIs.size() == 16)
8586 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8587 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8592 /// PHIsEqualValue - Return true if this phi node is always equal to
8593 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8594 /// z = some value; x = phi (y, z); y = phi (x, z)
8595 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8596 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8597 // See if we already saw this PHI node.
8598 if (!ValueEqualPHIs.insert(PN))
8601 // Don't scan crazily complex things.
8602 if (ValueEqualPHIs.size() == 16)
8605 // Scan the operands to see if they are either phi nodes or are equal to
8607 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8608 Value *Op = PN->getIncomingValue(i);
8609 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8610 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8612 } else if (Op != NonPhiInVal)
8620 // PHINode simplification
8622 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8623 // If LCSSA is around, don't mess with Phi nodes
8624 if (MustPreserveLCSSA) return 0;
8626 if (Value *V = PN.hasConstantValue())
8627 return ReplaceInstUsesWith(PN, V);
8629 // If all PHI operands are the same operation, pull them through the PHI,
8630 // reducing code size.
8631 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8632 PN.getIncomingValue(0)->hasOneUse())
8633 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8636 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8637 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8638 // PHI)... break the cycle.
8639 if (PN.hasOneUse()) {
8640 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8641 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8642 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8643 PotentiallyDeadPHIs.insert(&PN);
8644 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8645 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8648 // If this phi has a single use, and if that use just computes a value for
8649 // the next iteration of a loop, delete the phi. This occurs with unused
8650 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8651 // common case here is good because the only other things that catch this
8652 // are induction variable analysis (sometimes) and ADCE, which is only run
8654 if (PHIUser->hasOneUse() &&
8655 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8656 PHIUser->use_back() == &PN) {
8657 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8661 // We sometimes end up with phi cycles that non-obviously end up being the
8662 // same value, for example:
8663 // z = some value; x = phi (y, z); y = phi (x, z)
8664 // where the phi nodes don't necessarily need to be in the same block. Do a
8665 // quick check to see if the PHI node only contains a single non-phi value, if
8666 // so, scan to see if the phi cycle is actually equal to that value.
8668 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
8669 // Scan for the first non-phi operand.
8670 while (InValNo != NumOperandVals &&
8671 isa<PHINode>(PN.getIncomingValue(InValNo)))
8674 if (InValNo != NumOperandVals) {
8675 Value *NonPhiInVal = PN.getOperand(InValNo);
8677 // Scan the rest of the operands to see if there are any conflicts, if so
8678 // there is no need to recursively scan other phis.
8679 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
8680 Value *OpVal = PN.getIncomingValue(InValNo);
8681 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
8685 // If we scanned over all operands, then we have one unique value plus
8686 // phi values. Scan PHI nodes to see if they all merge in each other or
8688 if (InValNo == NumOperandVals) {
8689 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
8690 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
8691 return ReplaceInstUsesWith(PN, NonPhiInVal);
8698 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8699 Instruction *InsertPoint,
8701 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8702 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8703 // We must cast correctly to the pointer type. Ensure that we
8704 // sign extend the integer value if it is smaller as this is
8705 // used for address computation.
8706 Instruction::CastOps opcode =
8707 (VTySize < PtrSize ? Instruction::SExt :
8708 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8709 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8713 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8714 Value *PtrOp = GEP.getOperand(0);
8715 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8716 // If so, eliminate the noop.
8717 if (GEP.getNumOperands() == 1)
8718 return ReplaceInstUsesWith(GEP, PtrOp);
8720 if (isa<UndefValue>(GEP.getOperand(0)))
8721 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8723 bool HasZeroPointerIndex = false;
8724 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8725 HasZeroPointerIndex = C->isNullValue();
8727 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8728 return ReplaceInstUsesWith(GEP, PtrOp);
8730 // Eliminate unneeded casts for indices.
8731 bool MadeChange = false;
8733 gep_type_iterator GTI = gep_type_begin(GEP);
8734 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8735 if (isa<SequentialType>(*GTI)) {
8736 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8737 if (CI->getOpcode() == Instruction::ZExt ||
8738 CI->getOpcode() == Instruction::SExt) {
8739 const Type *SrcTy = CI->getOperand(0)->getType();
8740 // We can eliminate a cast from i32 to i64 iff the target
8741 // is a 32-bit pointer target.
8742 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8744 GEP.setOperand(i, CI->getOperand(0));
8748 // If we are using a wider index than needed for this platform, shrink it
8749 // to what we need. If the incoming value needs a cast instruction,
8750 // insert it. This explicit cast can make subsequent optimizations more
8752 Value *Op = GEP.getOperand(i);
8753 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
8754 if (Constant *C = dyn_cast<Constant>(Op)) {
8755 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8758 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8760 GEP.setOperand(i, Op);
8765 if (MadeChange) return &GEP;
8767 // If this GEP instruction doesn't move the pointer, and if the input operand
8768 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8769 // real input to the dest type.
8770 if (GEP.hasAllZeroIndices()) {
8771 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
8772 // If the bitcast is of an allocation, and the allocation will be
8773 // converted to match the type of the cast, don't touch this.
8774 if (isa<AllocationInst>(BCI->getOperand(0))) {
8775 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
8776 if (Instruction *I = visitBitCast(*BCI)) {
8779 BCI->getParent()->getInstList().insert(BCI, I);
8780 ReplaceInstUsesWith(*BCI, I);
8785 return new BitCastInst(BCI->getOperand(0), GEP.getType());
8789 // Combine Indices - If the source pointer to this getelementptr instruction
8790 // is a getelementptr instruction, combine the indices of the two
8791 // getelementptr instructions into a single instruction.
8793 SmallVector<Value*, 8> SrcGEPOperands;
8794 if (User *Src = dyn_castGetElementPtr(PtrOp))
8795 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8797 if (!SrcGEPOperands.empty()) {
8798 // Note that if our source is a gep chain itself that we wait for that
8799 // chain to be resolved before we perform this transformation. This
8800 // avoids us creating a TON of code in some cases.
8802 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8803 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8804 return 0; // Wait until our source is folded to completion.
8806 SmallVector<Value*, 8> Indices;
8808 // Find out whether the last index in the source GEP is a sequential idx.
8809 bool EndsWithSequential = false;
8810 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8811 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8812 EndsWithSequential = !isa<StructType>(*I);
8814 // Can we combine the two pointer arithmetics offsets?
8815 if (EndsWithSequential) {
8816 // Replace: gep (gep %P, long B), long A, ...
8817 // With: T = long A+B; gep %P, T, ...
8819 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8820 if (SO1 == Constant::getNullValue(SO1->getType())) {
8822 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8825 // If they aren't the same type, convert both to an integer of the
8826 // target's pointer size.
8827 if (SO1->getType() != GO1->getType()) {
8828 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8829 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8830 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8831 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8833 unsigned PS = TD->getPointerSizeInBits();
8834 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
8835 // Convert GO1 to SO1's type.
8836 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8838 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
8839 // Convert SO1 to GO1's type.
8840 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8842 const Type *PT = TD->getIntPtrType();
8843 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8844 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8848 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8849 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8851 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8852 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8856 // Recycle the GEP we already have if possible.
8857 if (SrcGEPOperands.size() == 2) {
8858 GEP.setOperand(0, SrcGEPOperands[0]);
8859 GEP.setOperand(1, Sum);
8862 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8863 SrcGEPOperands.end()-1);
8864 Indices.push_back(Sum);
8865 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8867 } else if (isa<Constant>(*GEP.idx_begin()) &&
8868 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8869 SrcGEPOperands.size() != 1) {
8870 // Otherwise we can do the fold if the first index of the GEP is a zero
8871 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8872 SrcGEPOperands.end());
8873 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8876 if (!Indices.empty())
8877 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8878 Indices.end(), GEP.getName());
8880 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8881 // GEP of global variable. If all of the indices for this GEP are
8882 // constants, we can promote this to a constexpr instead of an instruction.
8884 // Scan for nonconstants...
8885 SmallVector<Constant*, 8> Indices;
8886 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8887 for (; I != E && isa<Constant>(*I); ++I)
8888 Indices.push_back(cast<Constant>(*I));
8890 if (I == E) { // If they are all constants...
8891 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8892 &Indices[0],Indices.size());
8894 // Replace all uses of the GEP with the new constexpr...
8895 return ReplaceInstUsesWith(GEP, CE);
8897 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8898 if (!isa<PointerType>(X->getType())) {
8899 // Not interesting. Source pointer must be a cast from pointer.
8900 } else if (HasZeroPointerIndex) {
8901 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
8902 // into : GEP [10 x i8]* X, i32 0, ...
8904 // This occurs when the program declares an array extern like "int X[];"
8906 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8907 const PointerType *XTy = cast<PointerType>(X->getType());
8908 if (const ArrayType *XATy =
8909 dyn_cast<ArrayType>(XTy->getElementType()))
8910 if (const ArrayType *CATy =
8911 dyn_cast<ArrayType>(CPTy->getElementType()))
8912 if (CATy->getElementType() == XATy->getElementType()) {
8913 // At this point, we know that the cast source type is a pointer
8914 // to an array of the same type as the destination pointer
8915 // array. Because the array type is never stepped over (there
8916 // is a leading zero) we can fold the cast into this GEP.
8917 GEP.setOperand(0, X);
8920 } else if (GEP.getNumOperands() == 2) {
8921 // Transform things like:
8922 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
8923 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
8924 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8925 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8926 if (isa<ArrayType>(SrcElTy) &&
8927 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8928 TD->getABITypeSize(ResElTy)) {
8930 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8931 Idx[1] = GEP.getOperand(1);
8932 Value *V = InsertNewInstBefore(
8933 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8934 // V and GEP are both pointer types --> BitCast
8935 return new BitCastInst(V, GEP.getType());
8938 // Transform things like:
8939 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
8940 // (where tmp = 8*tmp2) into:
8941 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
8943 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
8944 uint64_t ArrayEltSize =
8945 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8947 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8948 // allow either a mul, shift, or constant here.
8950 ConstantInt *Scale = 0;
8951 if (ArrayEltSize == 1) {
8952 NewIdx = GEP.getOperand(1);
8953 Scale = ConstantInt::get(NewIdx->getType(), 1);
8954 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8955 NewIdx = ConstantInt::get(CI->getType(), 1);
8957 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8958 if (Inst->getOpcode() == Instruction::Shl &&
8959 isa<ConstantInt>(Inst->getOperand(1))) {
8960 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8961 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8962 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8963 NewIdx = Inst->getOperand(0);
8964 } else if (Inst->getOpcode() == Instruction::Mul &&
8965 isa<ConstantInt>(Inst->getOperand(1))) {
8966 Scale = cast<ConstantInt>(Inst->getOperand(1));
8967 NewIdx = Inst->getOperand(0);
8971 // If the index will be to exactly the right offset with the scale taken
8972 // out, perform the transformation. Note, we don't know whether Scale is
8973 // signed or not. We'll use unsigned version of division/modulo
8974 // operation after making sure Scale doesn't have the sign bit set.
8975 if (Scale && Scale->getSExtValue() >= 0LL &&
8976 Scale->getZExtValue() % ArrayEltSize == 0) {
8977 Scale = ConstantInt::get(Scale->getType(),
8978 Scale->getZExtValue() / ArrayEltSize);
8979 if (Scale->getZExtValue() != 1) {
8980 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8982 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8983 NewIdx = InsertNewInstBefore(Sc, GEP);
8986 // Insert the new GEP instruction.
8988 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8990 Instruction *NewGEP =
8991 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
8992 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8993 // The NewGEP must be pointer typed, so must the old one -> BitCast
8994 return new BitCastInst(NewGEP, GEP.getType());
9003 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9004 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9005 if (AI.isArrayAllocation()) // Check C != 1
9006 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9008 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9009 AllocationInst *New = 0;
9011 // Create and insert the replacement instruction...
9012 if (isa<MallocInst>(AI))
9013 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9015 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9016 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9019 InsertNewInstBefore(New, AI);
9021 // Scan to the end of the allocation instructions, to skip over a block of
9022 // allocas if possible...
9024 BasicBlock::iterator It = New;
9025 while (isa<AllocationInst>(*It)) ++It;
9027 // Now that I is pointing to the first non-allocation-inst in the block,
9028 // insert our getelementptr instruction...
9030 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9034 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9035 New->getName()+".sub", It);
9037 // Now make everything use the getelementptr instead of the original
9039 return ReplaceInstUsesWith(AI, V);
9040 } else if (isa<UndefValue>(AI.getArraySize())) {
9041 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9044 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9045 // Note that we only do this for alloca's, because malloc should allocate and
9046 // return a unique pointer, even for a zero byte allocation.
9047 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9048 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9049 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9054 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9055 Value *Op = FI.getOperand(0);
9057 // free undef -> unreachable.
9058 if (isa<UndefValue>(Op)) {
9059 // Insert a new store to null because we cannot modify the CFG here.
9060 new StoreInst(ConstantInt::getTrue(),
9061 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9062 return EraseInstFromFunction(FI);
9065 // If we have 'free null' delete the instruction. This can happen in stl code
9066 // when lots of inlining happens.
9067 if (isa<ConstantPointerNull>(Op))
9068 return EraseInstFromFunction(FI);
9070 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9071 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9072 FI.setOperand(0, CI->getOperand(0));
9076 // Change free (gep X, 0,0,0,0) into free(X)
9077 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9078 if (GEPI->hasAllZeroIndices()) {
9079 AddToWorkList(GEPI);
9080 FI.setOperand(0, GEPI->getOperand(0));
9085 // Change free(malloc) into nothing, if the malloc has a single use.
9086 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9087 if (MI->hasOneUse()) {
9088 EraseInstFromFunction(FI);
9089 return EraseInstFromFunction(*MI);
9096 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9097 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9098 const TargetData *TD) {
9099 User *CI = cast<User>(LI.getOperand(0));
9100 Value *CastOp = CI->getOperand(0);
9102 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9103 // Instead of loading constant c string, use corresponding integer value
9104 // directly if string length is small enough.
9105 const std::string &Str = CE->getOperand(0)->getStringValue();
9107 unsigned len = Str.length();
9108 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9109 unsigned numBits = Ty->getPrimitiveSizeInBits();
9110 // Replace LI with immediate integer store.
9111 if ((numBits >> 3) == len + 1) {
9112 APInt StrVal(numBits, 0);
9113 APInt SingleChar(numBits, 0);
9114 if (TD->isLittleEndian()) {
9115 for (signed i = len-1; i >= 0; i--) {
9116 SingleChar = (uint64_t) Str[i];
9117 StrVal = (StrVal << 8) | SingleChar;
9120 for (unsigned i = 0; i < len; i++) {
9121 SingleChar = (uint64_t) Str[i];
9122 StrVal = (StrVal << 8) | SingleChar;
9124 // Append NULL at the end.
9126 StrVal = (StrVal << 8) | SingleChar;
9128 Value *NL = ConstantInt::get(StrVal);
9129 return IC.ReplaceInstUsesWith(LI, NL);
9134 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9135 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9136 const Type *SrcPTy = SrcTy->getElementType();
9138 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9139 isa<VectorType>(DestPTy)) {
9140 // If the source is an array, the code below will not succeed. Check to
9141 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9143 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9144 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9145 if (ASrcTy->getNumElements() != 0) {
9147 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9148 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9149 SrcTy = cast<PointerType>(CastOp->getType());
9150 SrcPTy = SrcTy->getElementType();
9153 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9154 isa<VectorType>(SrcPTy)) &&
9155 // Do not allow turning this into a load of an integer, which is then
9156 // casted to a pointer, this pessimizes pointer analysis a lot.
9157 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9158 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9159 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9161 // Okay, we are casting from one integer or pointer type to another of
9162 // the same size. Instead of casting the pointer before the load, cast
9163 // the result of the loaded value.
9164 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9166 LI.isVolatile()),LI);
9167 // Now cast the result of the load.
9168 return new BitCastInst(NewLoad, LI.getType());
9175 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9176 /// from this value cannot trap. If it is not obviously safe to load from the
9177 /// specified pointer, we do a quick local scan of the basic block containing
9178 /// ScanFrom, to determine if the address is already accessed.
9179 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9180 // If it is an alloca it is always safe to load from.
9181 if (isa<AllocaInst>(V)) return true;
9183 // If it is a global variable it is mostly safe to load from.
9184 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9185 // Don't try to evaluate aliases. External weak GV can be null.
9186 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9188 // Otherwise, be a little bit agressive by scanning the local block where we
9189 // want to check to see if the pointer is already being loaded or stored
9190 // from/to. If so, the previous load or store would have already trapped,
9191 // so there is no harm doing an extra load (also, CSE will later eliminate
9192 // the load entirely).
9193 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9198 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9199 if (LI->getOperand(0) == V) return true;
9200 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9201 if (SI->getOperand(1) == V) return true;
9207 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9208 /// until we find the underlying object a pointer is referring to or something
9209 /// we don't understand. Note that the returned pointer may be offset from the
9210 /// input, because we ignore GEP indices.
9211 static Value *GetUnderlyingObject(Value *Ptr) {
9213 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9214 if (CE->getOpcode() == Instruction::BitCast ||
9215 CE->getOpcode() == Instruction::GetElementPtr)
9216 Ptr = CE->getOperand(0);
9219 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9220 Ptr = BCI->getOperand(0);
9221 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9222 Ptr = GEP->getOperand(0);
9229 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9230 Value *Op = LI.getOperand(0);
9232 // Attempt to improve the alignment.
9233 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9234 if (KnownAlign > LI.getAlignment())
9235 LI.setAlignment(KnownAlign);
9237 // load (cast X) --> cast (load X) iff safe
9238 if (isa<CastInst>(Op))
9239 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9242 // None of the following transforms are legal for volatile loads.
9243 if (LI.isVolatile()) return 0;
9245 if (&LI.getParent()->front() != &LI) {
9246 BasicBlock::iterator BBI = &LI; --BBI;
9247 // If the instruction immediately before this is a store to the same
9248 // address, do a simple form of store->load forwarding.
9249 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9250 if (SI->getOperand(1) == LI.getOperand(0))
9251 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9252 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9253 if (LIB->getOperand(0) == LI.getOperand(0))
9254 return ReplaceInstUsesWith(LI, LIB);
9257 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
9258 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
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 if (Constant *C = dyn_cast<Constant>(Op)) {
9269 // load null/undef -> undef
9270 if ((C->isNullValue() || isa<UndefValue>(C))) {
9271 // Insert a new store to null instruction before the load to indicate that
9272 // this code is not reachable. We do this instead of inserting an
9273 // unreachable instruction directly because we cannot modify the CFG.
9274 new StoreInst(UndefValue::get(LI.getType()),
9275 Constant::getNullValue(Op->getType()), &LI);
9276 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9279 // Instcombine load (constant global) into the value loaded.
9280 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9281 if (GV->isConstant() && !GV->isDeclaration())
9282 return ReplaceInstUsesWith(LI, GV->getInitializer());
9284 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9285 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9286 if (CE->getOpcode() == Instruction::GetElementPtr) {
9287 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9288 if (GV->isConstant() && !GV->isDeclaration())
9290 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9291 return ReplaceInstUsesWith(LI, V);
9292 if (CE->getOperand(0)->isNullValue()) {
9293 // Insert a new store to null instruction before the load to indicate
9294 // that this code is not reachable. We do this instead of inserting
9295 // an unreachable instruction directly because we cannot modify the
9297 new StoreInst(UndefValue::get(LI.getType()),
9298 Constant::getNullValue(Op->getType()), &LI);
9299 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9302 } else if (CE->isCast()) {
9303 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9308 // If this load comes from anywhere in a constant global, and if the global
9309 // is all undef or zero, we know what it loads.
9310 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9311 if (GV->isConstant() && GV->hasInitializer()) {
9312 if (GV->getInitializer()->isNullValue())
9313 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9314 else if (isa<UndefValue>(GV->getInitializer()))
9315 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9319 if (Op->hasOneUse()) {
9320 // Change select and PHI nodes to select values instead of addresses: this
9321 // helps alias analysis out a lot, allows many others simplifications, and
9322 // exposes redundancy in the code.
9324 // Note that we cannot do the transformation unless we know that the
9325 // introduced loads cannot trap! Something like this is valid as long as
9326 // the condition is always false: load (select bool %C, int* null, int* %G),
9327 // but it would not be valid if we transformed it to load from null
9330 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9331 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9332 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9333 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9334 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9335 SI->getOperand(1)->getName()+".val"), LI);
9336 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9337 SI->getOperand(2)->getName()+".val"), LI);
9338 return new SelectInst(SI->getCondition(), V1, V2);
9341 // load (select (cond, null, P)) -> load P
9342 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9343 if (C->isNullValue()) {
9344 LI.setOperand(0, SI->getOperand(2));
9348 // load (select (cond, P, null)) -> load P
9349 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9350 if (C->isNullValue()) {
9351 LI.setOperand(0, SI->getOperand(1));
9359 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9361 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9362 User *CI = cast<User>(SI.getOperand(1));
9363 Value *CastOp = CI->getOperand(0);
9365 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9366 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9367 const Type *SrcPTy = SrcTy->getElementType();
9369 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9370 // If the source is an array, the code below will not succeed. Check to
9371 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9373 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9374 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9375 if (ASrcTy->getNumElements() != 0) {
9377 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9378 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9379 SrcTy = cast<PointerType>(CastOp->getType());
9380 SrcPTy = SrcTy->getElementType();
9383 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9384 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9385 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9387 // Okay, we are casting from one integer or pointer type to another of
9388 // the same size. Instead of casting the pointer before
9389 // the store, cast the value to be stored.
9391 Value *SIOp0 = SI.getOperand(0);
9392 Instruction::CastOps opcode = Instruction::BitCast;
9393 const Type* CastSrcTy = SIOp0->getType();
9394 const Type* CastDstTy = SrcPTy;
9395 if (isa<PointerType>(CastDstTy)) {
9396 if (CastSrcTy->isInteger())
9397 opcode = Instruction::IntToPtr;
9398 } else if (isa<IntegerType>(CastDstTy)) {
9399 if (isa<PointerType>(SIOp0->getType()))
9400 opcode = Instruction::PtrToInt;
9402 if (Constant *C = dyn_cast<Constant>(SIOp0))
9403 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9405 NewCast = IC.InsertNewInstBefore(
9406 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9408 return new StoreInst(NewCast, CastOp);
9415 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9416 Value *Val = SI.getOperand(0);
9417 Value *Ptr = SI.getOperand(1);
9419 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9420 EraseInstFromFunction(SI);
9425 // If the RHS is an alloca with a single use, zapify the store, making the
9427 if (Ptr->hasOneUse()) {
9428 if (isa<AllocaInst>(Ptr)) {
9429 EraseInstFromFunction(SI);
9434 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9435 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9436 GEP->getOperand(0)->hasOneUse()) {
9437 EraseInstFromFunction(SI);
9443 // Attempt to improve the alignment.
9444 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9445 if (KnownAlign > SI.getAlignment())
9446 SI.setAlignment(KnownAlign);
9448 // Do really simple DSE, to catch cases where there are several consequtive
9449 // stores to the same location, separated by a few arithmetic operations. This
9450 // situation often occurs with bitfield accesses.
9451 BasicBlock::iterator BBI = &SI;
9452 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9456 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9457 // Prev store isn't volatile, and stores to the same location?
9458 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9461 EraseInstFromFunction(*PrevSI);
9467 // If this is a load, we have to stop. However, if the loaded value is from
9468 // the pointer we're loading and is producing the pointer we're storing,
9469 // then *this* store is dead (X = load P; store X -> P).
9470 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9471 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9472 EraseInstFromFunction(SI);
9476 // Otherwise, this is a load from some other location. Stores before it
9481 // Don't skip over loads or things that can modify memory.
9482 if (BBI->mayWriteToMemory())
9487 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9489 // store X, null -> turns into 'unreachable' in SimplifyCFG
9490 if (isa<ConstantPointerNull>(Ptr)) {
9491 if (!isa<UndefValue>(Val)) {
9492 SI.setOperand(0, UndefValue::get(Val->getType()));
9493 if (Instruction *U = dyn_cast<Instruction>(Val))
9494 AddToWorkList(U); // Dropped a use.
9497 return 0; // Do not modify these!
9500 // store undef, Ptr -> noop
9501 if (isa<UndefValue>(Val)) {
9502 EraseInstFromFunction(SI);
9507 // If the pointer destination is a cast, see if we can fold the cast into the
9509 if (isa<CastInst>(Ptr))
9510 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9512 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9514 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9518 // If this store is the last instruction in the basic block, and if the block
9519 // ends with an unconditional branch, try to move it to the successor block.
9521 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9522 if (BI->isUnconditional())
9523 if (SimplifyStoreAtEndOfBlock(SI))
9524 return 0; // xform done!
9529 /// SimplifyStoreAtEndOfBlock - Turn things like:
9530 /// if () { *P = v1; } else { *P = v2 }
9531 /// into a phi node with a store in the successor.
9533 /// Simplify things like:
9534 /// *P = v1; if () { *P = v2; }
9535 /// into a phi node with a store in the successor.
9537 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9538 BasicBlock *StoreBB = SI.getParent();
9540 // Check to see if the successor block has exactly two incoming edges. If
9541 // so, see if the other predecessor contains a store to the same location.
9542 // if so, insert a PHI node (if needed) and move the stores down.
9543 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9545 // Determine whether Dest has exactly two predecessors and, if so, compute
9546 // the other predecessor.
9547 pred_iterator PI = pred_begin(DestBB);
9548 BasicBlock *OtherBB = 0;
9552 if (PI == pred_end(DestBB))
9555 if (*PI != StoreBB) {
9560 if (++PI != pred_end(DestBB))
9564 // Verify that the other block ends in a branch and is not otherwise empty.
9565 BasicBlock::iterator BBI = OtherBB->getTerminator();
9566 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9567 if (!OtherBr || BBI == OtherBB->begin())
9570 // If the other block ends in an unconditional branch, check for the 'if then
9571 // else' case. there is an instruction before the branch.
9572 StoreInst *OtherStore = 0;
9573 if (OtherBr->isUnconditional()) {
9574 // If this isn't a store, or isn't a store to the same location, bail out.
9576 OtherStore = dyn_cast<StoreInst>(BBI);
9577 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9580 // Otherwise, the other block ended with a conditional branch. If one of the
9581 // destinations is StoreBB, then we have the if/then case.
9582 if (OtherBr->getSuccessor(0) != StoreBB &&
9583 OtherBr->getSuccessor(1) != StoreBB)
9586 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9587 // if/then triangle. See if there is a store to the same ptr as SI that
9588 // lives in OtherBB.
9590 // Check to see if we find the matching store.
9591 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9592 if (OtherStore->getOperand(1) != SI.getOperand(1))
9596 // If we find something that may be using the stored value, or if we run
9597 // out of instructions, we can't do the xform.
9598 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9599 BBI == OtherBB->begin())
9603 // In order to eliminate the store in OtherBr, we have to
9604 // make sure nothing reads the stored value in StoreBB.
9605 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9606 // FIXME: This should really be AA driven.
9607 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9612 // Insert a PHI node now if we need it.
9613 Value *MergedVal = OtherStore->getOperand(0);
9614 if (MergedVal != SI.getOperand(0)) {
9615 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9616 PN->reserveOperandSpace(2);
9617 PN->addIncoming(SI.getOperand(0), SI.getParent());
9618 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9619 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9622 // Advance to a place where it is safe to insert the new store and
9624 BBI = DestBB->begin();
9625 while (isa<PHINode>(BBI)) ++BBI;
9626 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9627 OtherStore->isVolatile()), *BBI);
9629 // Nuke the old stores.
9630 EraseInstFromFunction(SI);
9631 EraseInstFromFunction(*OtherStore);
9637 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9638 // Change br (not X), label True, label False to: br X, label False, True
9640 BasicBlock *TrueDest;
9641 BasicBlock *FalseDest;
9642 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9643 !isa<Constant>(X)) {
9644 // Swap Destinations and condition...
9646 BI.setSuccessor(0, FalseDest);
9647 BI.setSuccessor(1, TrueDest);
9651 // Cannonicalize fcmp_one -> fcmp_oeq
9652 FCmpInst::Predicate FPred; Value *Y;
9653 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9654 TrueDest, FalseDest)))
9655 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9656 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9657 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9658 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9659 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9660 NewSCC->takeName(I);
9661 // Swap Destinations and condition...
9662 BI.setCondition(NewSCC);
9663 BI.setSuccessor(0, FalseDest);
9664 BI.setSuccessor(1, TrueDest);
9665 RemoveFromWorkList(I);
9666 I->eraseFromParent();
9667 AddToWorkList(NewSCC);
9671 // Cannonicalize icmp_ne -> icmp_eq
9672 ICmpInst::Predicate IPred;
9673 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9674 TrueDest, FalseDest)))
9675 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9676 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9677 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9678 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9679 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9680 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9681 NewSCC->takeName(I);
9682 // Swap Destinations and condition...
9683 BI.setCondition(NewSCC);
9684 BI.setSuccessor(0, FalseDest);
9685 BI.setSuccessor(1, TrueDest);
9686 RemoveFromWorkList(I);
9687 I->eraseFromParent();;
9688 AddToWorkList(NewSCC);
9695 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9696 Value *Cond = SI.getCondition();
9697 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9698 if (I->getOpcode() == Instruction::Add)
9699 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9700 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9701 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9702 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9704 SI.setOperand(0, I->getOperand(0));
9712 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9713 /// is to leave as a vector operation.
9714 static bool CheapToScalarize(Value *V, bool isConstant) {
9715 if (isa<ConstantAggregateZero>(V))
9717 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9718 if (isConstant) return true;
9719 // If all elts are the same, we can extract.
9720 Constant *Op0 = C->getOperand(0);
9721 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9722 if (C->getOperand(i) != Op0)
9726 Instruction *I = dyn_cast<Instruction>(V);
9727 if (!I) return false;
9729 // Insert element gets simplified to the inserted element or is deleted if
9730 // this is constant idx extract element and its a constant idx insertelt.
9731 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9732 isa<ConstantInt>(I->getOperand(2)))
9734 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9736 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9737 if (BO->hasOneUse() &&
9738 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9739 CheapToScalarize(BO->getOperand(1), isConstant)))
9741 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9742 if (CI->hasOneUse() &&
9743 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9744 CheapToScalarize(CI->getOperand(1), isConstant)))
9750 /// Read and decode a shufflevector mask.
9752 /// It turns undef elements into values that are larger than the number of
9753 /// elements in the input.
9754 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9755 unsigned NElts = SVI->getType()->getNumElements();
9756 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9757 return std::vector<unsigned>(NElts, 0);
9758 if (isa<UndefValue>(SVI->getOperand(2)))
9759 return std::vector<unsigned>(NElts, 2*NElts);
9761 std::vector<unsigned> Result;
9762 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9763 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9764 if (isa<UndefValue>(CP->getOperand(i)))
9765 Result.push_back(NElts*2); // undef -> 8
9767 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9771 /// FindScalarElement - Given a vector and an element number, see if the scalar
9772 /// value is already around as a register, for example if it were inserted then
9773 /// extracted from the vector.
9774 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9775 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9776 const VectorType *PTy = cast<VectorType>(V->getType());
9777 unsigned Width = PTy->getNumElements();
9778 if (EltNo >= Width) // Out of range access.
9779 return UndefValue::get(PTy->getElementType());
9781 if (isa<UndefValue>(V))
9782 return UndefValue::get(PTy->getElementType());
9783 else if (isa<ConstantAggregateZero>(V))
9784 return Constant::getNullValue(PTy->getElementType());
9785 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9786 return CP->getOperand(EltNo);
9787 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9788 // If this is an insert to a variable element, we don't know what it is.
9789 if (!isa<ConstantInt>(III->getOperand(2)))
9791 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9793 // If this is an insert to the element we are looking for, return the
9796 return III->getOperand(1);
9798 // Otherwise, the insertelement doesn't modify the value, recurse on its
9800 return FindScalarElement(III->getOperand(0), EltNo);
9801 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9802 unsigned InEl = getShuffleMask(SVI)[EltNo];
9804 return FindScalarElement(SVI->getOperand(0), InEl);
9805 else if (InEl < Width*2)
9806 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9808 return UndefValue::get(PTy->getElementType());
9811 // Otherwise, we don't know.
9815 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9817 // If vector val is undef, replace extract with scalar undef.
9818 if (isa<UndefValue>(EI.getOperand(0)))
9819 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9821 // If vector val is constant 0, replace extract with scalar 0.
9822 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9823 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9825 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9826 // If vector val is constant with uniform operands, replace EI
9827 // with that operand
9828 Constant *op0 = C->getOperand(0);
9829 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9830 if (C->getOperand(i) != op0) {
9835 return ReplaceInstUsesWith(EI, op0);
9838 // If extracting a specified index from the vector, see if we can recursively
9839 // find a previously computed scalar that was inserted into the vector.
9840 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9841 unsigned IndexVal = IdxC->getZExtValue();
9842 unsigned VectorWidth =
9843 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9845 // If this is extracting an invalid index, turn this into undef, to avoid
9846 // crashing the code below.
9847 if (IndexVal >= VectorWidth)
9848 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9850 // This instruction only demands the single element from the input vector.
9851 // If the input vector has a single use, simplify it based on this use
9853 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9855 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9858 EI.setOperand(0, V);
9863 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9864 return ReplaceInstUsesWith(EI, Elt);
9866 // If the this extractelement is directly using a bitcast from a vector of
9867 // the same number of elements, see if we can find the source element from
9868 // it. In this case, we will end up needing to bitcast the scalars.
9869 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9870 if (const VectorType *VT =
9871 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9872 if (VT->getNumElements() == VectorWidth)
9873 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9874 return new BitCastInst(Elt, EI.getType());
9878 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9879 if (I->hasOneUse()) {
9880 // Push extractelement into predecessor operation if legal and
9881 // profitable to do so
9882 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9883 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9884 if (CheapToScalarize(BO, isConstantElt)) {
9885 ExtractElementInst *newEI0 =
9886 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9887 EI.getName()+".lhs");
9888 ExtractElementInst *newEI1 =
9889 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9890 EI.getName()+".rhs");
9891 InsertNewInstBefore(newEI0, EI);
9892 InsertNewInstBefore(newEI1, EI);
9893 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9895 } else if (isa<LoadInst>(I)) {
9897 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
9898 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9899 PointerType::get(EI.getType(), AS), EI);
9900 GetElementPtrInst *GEP =
9901 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9902 InsertNewInstBefore(GEP, EI);
9903 return new LoadInst(GEP);
9906 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9907 // Extracting the inserted element?
9908 if (IE->getOperand(2) == EI.getOperand(1))
9909 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9910 // If the inserted and extracted elements are constants, they must not
9911 // be the same value, extract from the pre-inserted value instead.
9912 if (isa<Constant>(IE->getOperand(2)) &&
9913 isa<Constant>(EI.getOperand(1))) {
9914 AddUsesToWorkList(EI);
9915 EI.setOperand(0, IE->getOperand(0));
9918 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9919 // If this is extracting an element from a shufflevector, figure out where
9920 // it came from and extract from the appropriate input element instead.
9921 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9922 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9924 if (SrcIdx < SVI->getType()->getNumElements())
9925 Src = SVI->getOperand(0);
9926 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9927 SrcIdx -= SVI->getType()->getNumElements();
9928 Src = SVI->getOperand(1);
9930 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9932 return new ExtractElementInst(Src, SrcIdx);
9939 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9940 /// elements from either LHS or RHS, return the shuffle mask and true.
9941 /// Otherwise, return false.
9942 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9943 std::vector<Constant*> &Mask) {
9944 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9945 "Invalid CollectSingleShuffleElements");
9946 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9948 if (isa<UndefValue>(V)) {
9949 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9951 } else if (V == LHS) {
9952 for (unsigned i = 0; i != NumElts; ++i)
9953 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9955 } else if (V == RHS) {
9956 for (unsigned i = 0; i != NumElts; ++i)
9957 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9959 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9960 // If this is an insert of an extract from some other vector, include it.
9961 Value *VecOp = IEI->getOperand(0);
9962 Value *ScalarOp = IEI->getOperand(1);
9963 Value *IdxOp = IEI->getOperand(2);
9965 if (!isa<ConstantInt>(IdxOp))
9967 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9969 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9970 // Okay, we can handle this if the vector we are insertinting into is
9972 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9973 // If so, update the mask to reflect the inserted undef.
9974 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9977 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9978 if (isa<ConstantInt>(EI->getOperand(1)) &&
9979 EI->getOperand(0)->getType() == V->getType()) {
9980 unsigned ExtractedIdx =
9981 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9983 // This must be extracting from either LHS or RHS.
9984 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9985 // Okay, we can handle this if the vector we are insertinting into is
9987 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9988 // If so, update the mask to reflect the inserted value.
9989 if (EI->getOperand(0) == LHS) {
9990 Mask[InsertedIdx & (NumElts-1)] =
9991 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9993 assert(EI->getOperand(0) == RHS);
9994 Mask[InsertedIdx & (NumElts-1)] =
9995 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10004 // TODO: Handle shufflevector here!
10009 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10010 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10011 /// that computes V and the LHS value of the shuffle.
10012 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10014 assert(isa<VectorType>(V->getType()) &&
10015 (RHS == 0 || V->getType() == RHS->getType()) &&
10016 "Invalid shuffle!");
10017 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10019 if (isa<UndefValue>(V)) {
10020 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10022 } else if (isa<ConstantAggregateZero>(V)) {
10023 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10025 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10026 // If this is an insert of an extract from some other vector, include it.
10027 Value *VecOp = IEI->getOperand(0);
10028 Value *ScalarOp = IEI->getOperand(1);
10029 Value *IdxOp = IEI->getOperand(2);
10031 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10032 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10033 EI->getOperand(0)->getType() == V->getType()) {
10034 unsigned ExtractedIdx =
10035 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10036 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10038 // Either the extracted from or inserted into vector must be RHSVec,
10039 // otherwise we'd end up with a shuffle of three inputs.
10040 if (EI->getOperand(0) == RHS || RHS == 0) {
10041 RHS = EI->getOperand(0);
10042 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10043 Mask[InsertedIdx & (NumElts-1)] =
10044 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10048 if (VecOp == RHS) {
10049 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10050 // Everything but the extracted element is replaced with the RHS.
10051 for (unsigned i = 0; i != NumElts; ++i) {
10052 if (i != InsertedIdx)
10053 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10058 // If this insertelement is a chain that comes from exactly these two
10059 // vectors, return the vector and the effective shuffle.
10060 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10061 return EI->getOperand(0);
10066 // TODO: Handle shufflevector here!
10068 // Otherwise, can't do anything fancy. Return an identity vector.
10069 for (unsigned i = 0; i != NumElts; ++i)
10070 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10074 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10075 Value *VecOp = IE.getOperand(0);
10076 Value *ScalarOp = IE.getOperand(1);
10077 Value *IdxOp = IE.getOperand(2);
10079 // Inserting an undef or into an undefined place, remove this.
10080 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10081 ReplaceInstUsesWith(IE, VecOp);
10083 // If the inserted element was extracted from some other vector, and if the
10084 // indexes are constant, try to turn this into a shufflevector operation.
10085 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10086 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10087 EI->getOperand(0)->getType() == IE.getType()) {
10088 unsigned NumVectorElts = IE.getType()->getNumElements();
10089 unsigned ExtractedIdx =
10090 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10091 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10093 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10094 return ReplaceInstUsesWith(IE, VecOp);
10096 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10097 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10099 // If we are extracting a value from a vector, then inserting it right
10100 // back into the same place, just use the input vector.
10101 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10102 return ReplaceInstUsesWith(IE, VecOp);
10104 // We could theoretically do this for ANY input. However, doing so could
10105 // turn chains of insertelement instructions into a chain of shufflevector
10106 // instructions, and right now we do not merge shufflevectors. As such,
10107 // only do this in a situation where it is clear that there is benefit.
10108 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10109 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10110 // the values of VecOp, except then one read from EIOp0.
10111 // Build a new shuffle mask.
10112 std::vector<Constant*> Mask;
10113 if (isa<UndefValue>(VecOp))
10114 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10116 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10117 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10120 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10121 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10122 ConstantVector::get(Mask));
10125 // If this insertelement isn't used by some other insertelement, turn it
10126 // (and any insertelements it points to), into one big shuffle.
10127 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10128 std::vector<Constant*> Mask;
10130 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10131 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10132 // We now have a shuffle of LHS, RHS, Mask.
10133 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10142 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10143 Value *LHS = SVI.getOperand(0);
10144 Value *RHS = SVI.getOperand(1);
10145 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10147 bool MadeChange = false;
10149 // Undefined shuffle mask -> undefined value.
10150 if (isa<UndefValue>(SVI.getOperand(2)))
10151 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10153 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10154 // the undef, change them to undefs.
10155 if (isa<UndefValue>(SVI.getOperand(1))) {
10156 // Scan to see if there are any references to the RHS. If so, replace them
10157 // with undef element refs and set MadeChange to true.
10158 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10159 if (Mask[i] >= e && Mask[i] != 2*e) {
10166 // Remap any references to RHS to use LHS.
10167 std::vector<Constant*> Elts;
10168 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10169 if (Mask[i] == 2*e)
10170 Elts.push_back(UndefValue::get(Type::Int32Ty));
10172 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10174 SVI.setOperand(2, ConstantVector::get(Elts));
10178 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10179 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10180 if (LHS == RHS || isa<UndefValue>(LHS)) {
10181 if (isa<UndefValue>(LHS) && LHS == RHS) {
10182 // shuffle(undef,undef,mask) -> undef.
10183 return ReplaceInstUsesWith(SVI, LHS);
10186 // Remap any references to RHS to use LHS.
10187 std::vector<Constant*> Elts;
10188 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10189 if (Mask[i] >= 2*e)
10190 Elts.push_back(UndefValue::get(Type::Int32Ty));
10192 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10193 (Mask[i] < e && isa<UndefValue>(LHS)))
10194 Mask[i] = 2*e; // Turn into undef.
10196 Mask[i] &= (e-1); // Force to LHS.
10197 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10200 SVI.setOperand(0, SVI.getOperand(1));
10201 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10202 SVI.setOperand(2, ConstantVector::get(Elts));
10203 LHS = SVI.getOperand(0);
10204 RHS = SVI.getOperand(1);
10208 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10209 bool isLHSID = true, isRHSID = true;
10211 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10212 if (Mask[i] >= e*2) continue; // Ignore undef values.
10213 // Is this an identity shuffle of the LHS value?
10214 isLHSID &= (Mask[i] == i);
10216 // Is this an identity shuffle of the RHS value?
10217 isRHSID &= (Mask[i]-e == i);
10220 // Eliminate identity shuffles.
10221 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10222 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10224 // If the LHS is a shufflevector itself, see if we can combine it with this
10225 // one without producing an unusual shuffle. Here we are really conservative:
10226 // we are absolutely afraid of producing a shuffle mask not in the input
10227 // program, because the code gen may not be smart enough to turn a merged
10228 // shuffle into two specific shuffles: it may produce worse code. As such,
10229 // we only merge two shuffles if the result is one of the two input shuffle
10230 // masks. In this case, merging the shuffles just removes one instruction,
10231 // which we know is safe. This is good for things like turning:
10232 // (splat(splat)) -> splat.
10233 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10234 if (isa<UndefValue>(RHS)) {
10235 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10237 std::vector<unsigned> NewMask;
10238 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10239 if (Mask[i] >= 2*e)
10240 NewMask.push_back(2*e);
10242 NewMask.push_back(LHSMask[Mask[i]]);
10244 // If the result mask is equal to the src shuffle or this shuffle mask, do
10245 // the replacement.
10246 if (NewMask == LHSMask || NewMask == Mask) {
10247 std::vector<Constant*> Elts;
10248 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10249 if (NewMask[i] >= e*2) {
10250 Elts.push_back(UndefValue::get(Type::Int32Ty));
10252 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10255 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10256 LHSSVI->getOperand(1),
10257 ConstantVector::get(Elts));
10262 return MadeChange ? &SVI : 0;
10268 /// TryToSinkInstruction - Try to move the specified instruction from its
10269 /// current block into the beginning of DestBlock, which can only happen if it's
10270 /// safe to move the instruction past all of the instructions between it and the
10271 /// end of its block.
10272 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10273 assert(I->hasOneUse() && "Invariants didn't hold!");
10275 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10276 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10278 // Do not sink alloca instructions out of the entry block.
10279 if (isa<AllocaInst>(I) && I->getParent() ==
10280 &DestBlock->getParent()->getEntryBlock())
10283 // We can only sink load instructions if there is nothing between the load and
10284 // the end of block that could change the value.
10285 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10286 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10288 if (Scan->mayWriteToMemory())
10292 BasicBlock::iterator InsertPos = DestBlock->begin();
10293 while (isa<PHINode>(InsertPos)) ++InsertPos;
10295 I->moveBefore(InsertPos);
10301 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10302 /// all reachable code to the worklist.
10304 /// This has a couple of tricks to make the code faster and more powerful. In
10305 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10306 /// them to the worklist (this significantly speeds up instcombine on code where
10307 /// many instructions are dead or constant). Additionally, if we find a branch
10308 /// whose condition is a known constant, we only visit the reachable successors.
10310 static void AddReachableCodeToWorklist(BasicBlock *BB,
10311 SmallPtrSet<BasicBlock*, 64> &Visited,
10313 const TargetData *TD) {
10314 std::vector<BasicBlock*> Worklist;
10315 Worklist.push_back(BB);
10317 while (!Worklist.empty()) {
10318 BB = Worklist.back();
10319 Worklist.pop_back();
10321 // We have now visited this block! If we've already been here, ignore it.
10322 if (!Visited.insert(BB)) continue;
10324 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10325 Instruction *Inst = BBI++;
10327 // DCE instruction if trivially dead.
10328 if (isInstructionTriviallyDead(Inst)) {
10330 DOUT << "IC: DCE: " << *Inst;
10331 Inst->eraseFromParent();
10335 // ConstantProp instruction if trivially constant.
10336 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10337 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10338 Inst->replaceAllUsesWith(C);
10340 Inst->eraseFromParent();
10344 IC.AddToWorkList(Inst);
10347 // Recursively visit successors. If this is a branch or switch on a
10348 // constant, only visit the reachable successor.
10349 TerminatorInst *TI = BB->getTerminator();
10350 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10351 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10352 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10353 Worklist.push_back(BI->getSuccessor(!CondVal));
10356 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10357 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10358 // See if this is an explicit destination.
10359 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10360 if (SI->getCaseValue(i) == Cond) {
10361 Worklist.push_back(SI->getSuccessor(i));
10365 // Otherwise it is the default destination.
10366 Worklist.push_back(SI->getSuccessor(0));
10371 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10372 Worklist.push_back(TI->getSuccessor(i));
10376 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10377 bool Changed = false;
10378 TD = &getAnalysis<TargetData>();
10380 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10381 << F.getNameStr() << "\n");
10384 // Do a depth-first traversal of the function, populate the worklist with
10385 // the reachable instructions. Ignore blocks that are not reachable. Keep
10386 // track of which blocks we visit.
10387 SmallPtrSet<BasicBlock*, 64> Visited;
10388 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10390 // Do a quick scan over the function. If we find any blocks that are
10391 // unreachable, remove any instructions inside of them. This prevents
10392 // the instcombine code from having to deal with some bad special cases.
10393 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10394 if (!Visited.count(BB)) {
10395 Instruction *Term = BB->getTerminator();
10396 while (Term != BB->begin()) { // Remove instrs bottom-up
10397 BasicBlock::iterator I = Term; --I;
10399 DOUT << "IC: DCE: " << *I;
10402 if (!I->use_empty())
10403 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10404 I->eraseFromParent();
10409 while (!Worklist.empty()) {
10410 Instruction *I = RemoveOneFromWorkList();
10411 if (I == 0) continue; // skip null values.
10413 // Check to see if we can DCE the instruction.
10414 if (isInstructionTriviallyDead(I)) {
10415 // Add operands to the worklist.
10416 if (I->getNumOperands() < 4)
10417 AddUsesToWorkList(*I);
10420 DOUT << "IC: DCE: " << *I;
10422 I->eraseFromParent();
10423 RemoveFromWorkList(I);
10427 // Instruction isn't dead, see if we can constant propagate it.
10428 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10429 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10431 // Add operands to the worklist.
10432 AddUsesToWorkList(*I);
10433 ReplaceInstUsesWith(*I, C);
10436 I->eraseFromParent();
10437 RemoveFromWorkList(I);
10441 // See if we can trivially sink this instruction to a successor basic block.
10442 if (I->hasOneUse()) {
10443 BasicBlock *BB = I->getParent();
10444 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10445 if (UserParent != BB) {
10446 bool UserIsSuccessor = false;
10447 // See if the user is one of our successors.
10448 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10449 if (*SI == UserParent) {
10450 UserIsSuccessor = true;
10454 // If the user is one of our immediate successors, and if that successor
10455 // only has us as a predecessors (we'd have to split the critical edge
10456 // otherwise), we can keep going.
10457 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10458 next(pred_begin(UserParent)) == pred_end(UserParent))
10459 // Okay, the CFG is simple enough, try to sink this instruction.
10460 Changed |= TryToSinkInstruction(I, UserParent);
10464 // Now that we have an instruction, try combining it to simplify it...
10468 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10469 if (Instruction *Result = visit(*I)) {
10471 // Should we replace the old instruction with a new one?
10473 DOUT << "IC: Old = " << *I
10474 << " New = " << *Result;
10476 // Everything uses the new instruction now.
10477 I->replaceAllUsesWith(Result);
10479 // Push the new instruction and any users onto the worklist.
10480 AddToWorkList(Result);
10481 AddUsersToWorkList(*Result);
10483 // Move the name to the new instruction first.
10484 Result->takeName(I);
10486 // Insert the new instruction into the basic block...
10487 BasicBlock *InstParent = I->getParent();
10488 BasicBlock::iterator InsertPos = I;
10490 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10491 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10494 InstParent->getInstList().insert(InsertPos, Result);
10496 // Make sure that we reprocess all operands now that we reduced their
10498 AddUsesToWorkList(*I);
10500 // Instructions can end up on the worklist more than once. Make sure
10501 // we do not process an instruction that has been deleted.
10502 RemoveFromWorkList(I);
10504 // Erase the old instruction.
10505 InstParent->getInstList().erase(I);
10508 DOUT << "IC: Mod = " << OrigI
10509 << " New = " << *I;
10512 // If the instruction was modified, it's possible that it is now dead.
10513 // if so, remove it.
10514 if (isInstructionTriviallyDead(I)) {
10515 // Make sure we process all operands now that we are reducing their
10517 AddUsesToWorkList(*I);
10519 // Instructions may end up in the worklist more than once. Erase all
10520 // occurrences of this instruction.
10521 RemoveFromWorkList(I);
10522 I->eraseFromParent();
10525 AddUsersToWorkList(*I);
10532 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10534 // Do an explicit clear, this shrinks the map if needed.
10535 WorklistMap.clear();
10540 bool InstCombiner::runOnFunction(Function &F) {
10541 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10543 bool EverMadeChange = false;
10545 // Iterate while there is work to do.
10546 unsigned Iteration = 0;
10547 while (DoOneIteration(F, Iteration++))
10548 EverMadeChange = true;
10549 return EverMadeChange;
10552 FunctionPass *llvm::createInstructionCombiningPass() {
10553 return new InstCombiner();