1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
62 using namespace llvm::PatternMatch;
64 STATISTIC(NumCombined , "Number of insts combined");
65 STATISTIC(NumConstProp, "Number of constant folds");
66 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
67 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 class VISIBILITY_HIDDEN InstCombiner
72 : public FunctionPass,
73 public InstVisitor<InstCombiner, Instruction*> {
74 // Worklist of all of the instructions that need to be simplified.
75 std::vector<Instruction*> Worklist;
76 DenseMap<Instruction*, unsigned> WorklistMap;
78 bool MustPreserveLCSSA;
80 static char ID; // Pass identification, replacement for typeid
81 InstCombiner() : FunctionPass((intptr_t)&ID) {}
83 /// AddToWorkList - Add the specified instruction to the worklist if it
84 /// isn't already in it.
85 void AddToWorkList(Instruction *I) {
86 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
87 Worklist.push_back(I);
90 // RemoveFromWorkList - remove I from the worklist if it exists.
91 void RemoveFromWorkList(Instruction *I) {
92 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
93 if (It == WorklistMap.end()) return; // Not in worklist.
95 // Don't bother moving everything down, just null out the slot.
96 Worklist[It->second] = 0;
98 WorklistMap.erase(It);
101 Instruction *RemoveOneFromWorkList() {
102 Instruction *I = Worklist.back();
104 WorklistMap.erase(I);
109 /// AddUsersToWorkList - When an instruction is simplified, add all users of
110 /// the instruction to the work lists because they might get more simplified
113 void AddUsersToWorkList(Value &I) {
114 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
116 AddToWorkList(cast<Instruction>(*UI));
119 /// AddUsesToWorkList - When an instruction is simplified, add operands to
120 /// the work lists because they might get more simplified now.
122 void AddUsesToWorkList(Instruction &I) {
123 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
124 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
128 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
129 /// dead. Add all of its operands to the worklist, turning them into
130 /// undef's to reduce the number of uses of those instructions.
132 /// Return the specified operand before it is turned into an undef.
134 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
135 Value *R = I.getOperand(op);
137 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
138 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
140 // Set the operand to undef to drop the use.
141 I.setOperand(i, UndefValue::get(Op->getType()));
148 virtual bool runOnFunction(Function &F);
150 bool DoOneIteration(Function &F, unsigned ItNum);
152 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
153 AU.addRequired<TargetData>();
154 AU.addPreservedID(LCSSAID);
155 AU.setPreservesCFG();
158 TargetData &getTargetData() const { return *TD; }
160 // Visitation implementation - Implement instruction combining for different
161 // instruction types. The semantics are as follows:
163 // null - No change was made
164 // I - Change was made, I is still valid, I may be dead though
165 // otherwise - Change was made, replace I with returned instruction
167 Instruction *visitAdd(BinaryOperator &I);
168 Instruction *visitSub(BinaryOperator &I);
169 Instruction *visitMul(BinaryOperator &I);
170 Instruction *visitURem(BinaryOperator &I);
171 Instruction *visitSRem(BinaryOperator &I);
172 Instruction *visitFRem(BinaryOperator &I);
173 Instruction *commonRemTransforms(BinaryOperator &I);
174 Instruction *commonIRemTransforms(BinaryOperator &I);
175 Instruction *commonDivTransforms(BinaryOperator &I);
176 Instruction *commonIDivTransforms(BinaryOperator &I);
177 Instruction *visitUDiv(BinaryOperator &I);
178 Instruction *visitSDiv(BinaryOperator &I);
179 Instruction *visitFDiv(BinaryOperator &I);
180 Instruction *visitAnd(BinaryOperator &I);
181 Instruction *visitOr (BinaryOperator &I);
182 Instruction *visitXor(BinaryOperator &I);
183 Instruction *visitShl(BinaryOperator &I);
184 Instruction *visitAShr(BinaryOperator &I);
185 Instruction *visitLShr(BinaryOperator &I);
186 Instruction *commonShiftTransforms(BinaryOperator &I);
187 Instruction *visitFCmpInst(FCmpInst &I);
188 Instruction *visitICmpInst(ICmpInst &I);
189 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
190 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
193 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
194 ConstantInt *DivRHS);
196 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
197 ICmpInst::Predicate Cond, Instruction &I);
198 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
200 Instruction *commonCastTransforms(CastInst &CI);
201 Instruction *commonIntCastTransforms(CastInst &CI);
202 Instruction *commonPointerCastTransforms(CastInst &CI);
203 Instruction *visitTrunc(TruncInst &CI);
204 Instruction *visitZExt(ZExtInst &CI);
205 Instruction *visitSExt(SExtInst &CI);
206 Instruction *visitFPTrunc(CastInst &CI);
207 Instruction *visitFPExt(CastInst &CI);
208 Instruction *visitFPToUI(CastInst &CI);
209 Instruction *visitFPToSI(CastInst &CI);
210 Instruction *visitUIToFP(CastInst &CI);
211 Instruction *visitSIToFP(CastInst &CI);
212 Instruction *visitPtrToInt(CastInst &CI);
213 Instruction *visitIntToPtr(CastInst &CI);
214 Instruction *visitBitCast(BitCastInst &CI);
215 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
217 Instruction *visitSelectInst(SelectInst &CI);
218 Instruction *visitCallInst(CallInst &CI);
219 Instruction *visitInvokeInst(InvokeInst &II);
220 Instruction *visitPHINode(PHINode &PN);
221 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
222 Instruction *visitAllocationInst(AllocationInst &AI);
223 Instruction *visitFreeInst(FreeInst &FI);
224 Instruction *visitLoadInst(LoadInst &LI);
225 Instruction *visitStoreInst(StoreInst &SI);
226 Instruction *visitBranchInst(BranchInst &BI);
227 Instruction *visitSwitchInst(SwitchInst &SI);
228 Instruction *visitInsertElementInst(InsertElementInst &IE);
229 Instruction *visitExtractElementInst(ExtractElementInst &EI);
230 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
232 // visitInstruction - Specify what to return for unhandled instructions...
233 Instruction *visitInstruction(Instruction &I) { return 0; }
236 Instruction *visitCallSite(CallSite CS);
237 bool transformConstExprCastCall(CallSite CS);
238 Instruction *transformCallThroughTrampoline(CallSite CS);
241 // InsertNewInstBefore - insert an instruction New before instruction Old
242 // in the program. Add the new instruction to the worklist.
244 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
245 assert(New && New->getParent() == 0 &&
246 "New instruction already inserted into a basic block!");
247 BasicBlock *BB = Old.getParent();
248 BB->getInstList().insert(&Old, New); // Insert inst
253 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
254 /// This also adds the cast to the worklist. Finally, this returns the
256 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
258 if (V->getType() == Ty) return V;
260 if (Constant *CV = dyn_cast<Constant>(V))
261 return ConstantExpr::getCast(opc, CV, Ty);
263 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
268 // ReplaceInstUsesWith - This method is to be used when an instruction is
269 // found to be dead, replacable with another preexisting expression. Here
270 // we add all uses of I to the worklist, replace all uses of I with the new
271 // value, then return I, so that the inst combiner will know that I was
274 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
275 AddUsersToWorkList(I); // Add all modified instrs to worklist
277 I.replaceAllUsesWith(V);
280 // If we are replacing the instruction with itself, this must be in a
281 // segment of unreachable code, so just clobber the instruction.
282 I.replaceAllUsesWith(UndefValue::get(I.getType()));
287 // UpdateValueUsesWith - This method is to be used when an value is
288 // found to be replacable with another preexisting expression or was
289 // updated. Here we add all uses of I to the worklist, replace all uses of
290 // I with the new value (unless the instruction was just updated), then
291 // return true, so that the inst combiner will know that I was modified.
293 bool UpdateValueUsesWith(Value *Old, Value *New) {
294 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
296 Old->replaceAllUsesWith(New);
297 if (Instruction *I = dyn_cast<Instruction>(Old))
299 if (Instruction *I = dyn_cast<Instruction>(New))
304 // EraseInstFromFunction - When dealing with an instruction that has side
305 // effects or produces a void value, we can't rely on DCE to delete the
306 // instruction. Instead, visit methods should return the value returned by
308 Instruction *EraseInstFromFunction(Instruction &I) {
309 assert(I.use_empty() && "Cannot erase instruction that is used!");
310 AddUsesToWorkList(I);
311 RemoveFromWorkList(&I);
313 return 0; // Don't do anything with FI
317 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
318 /// InsertBefore instruction. This is specialized a bit to avoid inserting
319 /// casts that are known to not do anything...
321 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
322 Value *V, const Type *DestTy,
323 Instruction *InsertBefore);
325 /// SimplifyCommutative - This performs a few simplifications for
326 /// commutative operators.
327 bool SimplifyCommutative(BinaryOperator &I);
329 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
330 /// most-complex to least-complex order.
331 bool SimplifyCompare(CmpInst &I);
333 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
334 /// on the demanded bits.
335 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
336 APInt& KnownZero, APInt& KnownOne,
339 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
340 uint64_t &UndefElts, unsigned Depth = 0);
342 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
343 // PHI node as operand #0, see if we can fold the instruction into the PHI
344 // (which is only possible if all operands to the PHI are constants).
345 Instruction *FoldOpIntoPhi(Instruction &I);
347 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
348 // operator and they all are only used by the PHI, PHI together their
349 // inputs, and do the operation once, to the result of the PHI.
350 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
351 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
354 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
355 ConstantInt *AndRHS, BinaryOperator &TheAnd);
357 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
358 bool isSub, Instruction &I);
359 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
360 bool isSigned, bool Inside, Instruction &IB);
361 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
362 Instruction *MatchBSwap(BinaryOperator &I);
363 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
365 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
368 char InstCombiner::ID = 0;
369 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
372 // getComplexity: Assign a complexity or rank value to LLVM Values...
373 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
374 static unsigned getComplexity(Value *V) {
375 if (isa<Instruction>(V)) {
376 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
380 if (isa<Argument>(V)) return 3;
381 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
384 // isOnlyUse - Return true if this instruction will be deleted if we stop using
386 static bool isOnlyUse(Value *V) {
387 return V->hasOneUse() || isa<Constant>(V);
390 // getPromotedType - Return the specified type promoted as it would be to pass
391 // though a va_arg area...
392 static const Type *getPromotedType(const Type *Ty) {
393 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
394 if (ITy->getBitWidth() < 32)
395 return Type::Int32Ty;
400 /// getBitCastOperand - If the specified operand is a CastInst or a constant
401 /// expression bitcast, return the operand value, otherwise return null.
402 static Value *getBitCastOperand(Value *V) {
403 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
404 return I->getOperand(0);
405 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
406 if (CE->getOpcode() == Instruction::BitCast)
407 return CE->getOperand(0);
411 /// This function is a wrapper around CastInst::isEliminableCastPair. It
412 /// simply extracts arguments and returns what that function returns.
413 static Instruction::CastOps
414 isEliminableCastPair(
415 const CastInst *CI, ///< The first cast instruction
416 unsigned opcode, ///< The opcode of the second cast instruction
417 const Type *DstTy, ///< The target type for the second cast instruction
418 TargetData *TD ///< The target data for pointer size
421 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
422 const Type *MidTy = CI->getType(); // B from above
424 // Get the opcodes of the two Cast instructions
425 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
426 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
428 return Instruction::CastOps(
429 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
430 DstTy, TD->getIntPtrType()));
433 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
434 /// in any code being generated. It does not require codegen if V is simple
435 /// enough or if the cast can be folded into other casts.
436 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
437 const Type *Ty, TargetData *TD) {
438 if (V->getType() == Ty || isa<Constant>(V)) return false;
440 // If this is another cast that can be eliminated, it isn't codegen either.
441 if (const CastInst *CI = dyn_cast<CastInst>(V))
442 if (isEliminableCastPair(CI, opcode, Ty, TD))
447 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
448 /// InsertBefore instruction. This is specialized a bit to avoid inserting
449 /// casts that are known to not do anything...
451 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
452 Value *V, const Type *DestTy,
453 Instruction *InsertBefore) {
454 if (V->getType() == DestTy) return V;
455 if (Constant *C = dyn_cast<Constant>(V))
456 return ConstantExpr::getCast(opcode, C, DestTy);
458 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
461 // SimplifyCommutative - This performs a few simplifications for commutative
464 // 1. Order operands such that they are listed from right (least complex) to
465 // left (most complex). This puts constants before unary operators before
468 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
469 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
471 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
472 bool Changed = false;
473 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
474 Changed = !I.swapOperands();
476 if (!I.isAssociative()) return Changed;
477 Instruction::BinaryOps Opcode = I.getOpcode();
478 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
479 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
480 if (isa<Constant>(I.getOperand(1))) {
481 Constant *Folded = ConstantExpr::get(I.getOpcode(),
482 cast<Constant>(I.getOperand(1)),
483 cast<Constant>(Op->getOperand(1)));
484 I.setOperand(0, Op->getOperand(0));
485 I.setOperand(1, Folded);
487 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
488 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
489 isOnlyUse(Op) && isOnlyUse(Op1)) {
490 Constant *C1 = cast<Constant>(Op->getOperand(1));
491 Constant *C2 = cast<Constant>(Op1->getOperand(1));
493 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
494 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
495 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
499 I.setOperand(0, New);
500 I.setOperand(1, Folded);
507 /// SimplifyCompare - For a CmpInst this function just orders the operands
508 /// so that theyare listed from right (least complex) to left (most complex).
509 /// This puts constants before unary operators before binary operators.
510 bool InstCombiner::SimplifyCompare(CmpInst &I) {
511 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
514 // Compare instructions are not associative so there's nothing else we can do.
518 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
519 // if the LHS is a constant zero (which is the 'negate' form).
521 static inline Value *dyn_castNegVal(Value *V) {
522 if (BinaryOperator::isNeg(V))
523 return BinaryOperator::getNegArgument(V);
525 // Constants can be considered to be negated values if they can be folded.
526 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
527 return ConstantExpr::getNeg(C);
531 static inline Value *dyn_castNotVal(Value *V) {
532 if (BinaryOperator::isNot(V))
533 return BinaryOperator::getNotArgument(V);
535 // Constants can be considered to be not'ed values...
536 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
537 return ConstantInt::get(~C->getValue());
541 // dyn_castFoldableMul - If this value is a multiply that can be folded into
542 // other computations (because it has a constant operand), return the
543 // non-constant operand of the multiply, and set CST to point to the multiplier.
544 // Otherwise, return null.
546 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
547 if (V->hasOneUse() && V->getType()->isInteger())
548 if (Instruction *I = dyn_cast<Instruction>(V)) {
549 if (I->getOpcode() == Instruction::Mul)
550 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
551 return I->getOperand(0);
552 if (I->getOpcode() == Instruction::Shl)
553 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
554 // The multiplier is really 1 << CST.
555 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
556 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
557 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
558 return I->getOperand(0);
564 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
565 /// expression, return it.
566 static User *dyn_castGetElementPtr(Value *V) {
567 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
568 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
569 if (CE->getOpcode() == Instruction::GetElementPtr)
570 return cast<User>(V);
574 /// AddOne - Add one to a ConstantInt
575 static ConstantInt *AddOne(ConstantInt *C) {
576 APInt Val(C->getValue());
577 return ConstantInt::get(++Val);
579 /// SubOne - Subtract one from a ConstantInt
580 static ConstantInt *SubOne(ConstantInt *C) {
581 APInt Val(C->getValue());
582 return ConstantInt::get(--Val);
584 /// Add - Add two ConstantInts together
585 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
586 return ConstantInt::get(C1->getValue() + C2->getValue());
588 /// And - Bitwise AND two ConstantInts together
589 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
590 return ConstantInt::get(C1->getValue() & C2->getValue());
592 /// Subtract - Subtract one ConstantInt from another
593 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
594 return ConstantInt::get(C1->getValue() - C2->getValue());
596 /// Multiply - Multiply two ConstantInts together
597 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
598 return ConstantInt::get(C1->getValue() * C2->getValue());
601 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
602 /// known to be either zero or one and return them in the KnownZero/KnownOne
603 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
605 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
606 /// we cannot optimize based on the assumption that it is zero without changing
607 /// it to be an explicit zero. If we don't change it to zero, other code could
608 /// optimized based on the contradictory assumption that it is non-zero.
609 /// Because instcombine aggressively folds operations with undef args anyway,
610 /// this won't lose us code quality.
611 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
612 APInt& KnownOne, unsigned Depth = 0) {
613 assert(V && "No Value?");
614 assert(Depth <= 6 && "Limit Search Depth");
615 uint32_t BitWidth = Mask.getBitWidth();
616 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
617 KnownZero.getBitWidth() == BitWidth &&
618 KnownOne.getBitWidth() == BitWidth &&
619 "V, Mask, KnownOne and KnownZero should have same BitWidth");
620 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
621 // We know all of the bits for a constant!
622 KnownOne = CI->getValue() & Mask;
623 KnownZero = ~KnownOne & Mask;
627 if (Depth == 6 || Mask == 0)
628 return; // Limit search depth.
630 Instruction *I = dyn_cast<Instruction>(V);
633 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
634 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
636 switch (I->getOpcode()) {
637 case Instruction::And: {
638 // If either the LHS or the RHS are Zero, the result is zero.
639 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
640 APInt Mask2(Mask & ~KnownZero);
641 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
642 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
643 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
645 // Output known-1 bits are only known if set in both the LHS & RHS.
646 KnownOne &= KnownOne2;
647 // Output known-0 are known to be clear if zero in either the LHS | RHS.
648 KnownZero |= KnownZero2;
651 case Instruction::Or: {
652 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
653 APInt Mask2(Mask & ~KnownOne);
654 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
655 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
656 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
658 // Output known-0 bits are only known if clear in both the LHS & RHS.
659 KnownZero &= KnownZero2;
660 // Output known-1 are known to be set if set in either the LHS | RHS.
661 KnownOne |= KnownOne2;
664 case Instruction::Xor: {
665 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
666 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
667 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
668 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
670 // Output known-0 bits are known if clear or set in both the LHS & RHS.
671 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
672 // Output known-1 are known to be set if set in only one of the LHS, RHS.
673 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
674 KnownZero = KnownZeroOut;
677 case Instruction::Select:
678 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
679 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
680 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
681 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
683 // Only known if known in both the LHS and RHS.
684 KnownOne &= KnownOne2;
685 KnownZero &= KnownZero2;
687 case Instruction::FPTrunc:
688 case Instruction::FPExt:
689 case Instruction::FPToUI:
690 case Instruction::FPToSI:
691 case Instruction::SIToFP:
692 case Instruction::PtrToInt:
693 case Instruction::UIToFP:
694 case Instruction::IntToPtr:
695 return; // Can't work with floating point or pointers
696 case Instruction::Trunc: {
697 // All these have integer operands
698 uint32_t SrcBitWidth =
699 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
701 MaskIn.zext(SrcBitWidth);
702 KnownZero.zext(SrcBitWidth);
703 KnownOne.zext(SrcBitWidth);
704 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
705 KnownZero.trunc(BitWidth);
706 KnownOne.trunc(BitWidth);
709 case Instruction::BitCast: {
710 const Type *SrcTy = I->getOperand(0)->getType();
711 if (SrcTy->isInteger()) {
712 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
717 case Instruction::ZExt: {
718 // Compute the bits in the result that are not present in the input.
719 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
720 uint32_t SrcBitWidth = SrcTy->getBitWidth();
723 MaskIn.trunc(SrcBitWidth);
724 KnownZero.trunc(SrcBitWidth);
725 KnownOne.trunc(SrcBitWidth);
726 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
727 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
728 // The top bits are known to be zero.
729 KnownZero.zext(BitWidth);
730 KnownOne.zext(BitWidth);
731 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
734 case Instruction::SExt: {
735 // Compute the bits in the result that are not present in the input.
736 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
737 uint32_t SrcBitWidth = SrcTy->getBitWidth();
740 MaskIn.trunc(SrcBitWidth);
741 KnownZero.trunc(SrcBitWidth);
742 KnownOne.trunc(SrcBitWidth);
743 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 KnownZero.zext(BitWidth);
746 KnownOne.zext(BitWidth);
748 // If the sign bit of the input is known set or clear, then we know the
749 // top bits of the result.
750 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
751 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
752 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
753 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
756 case Instruction::Shl:
757 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
758 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
759 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
760 APInt Mask2(Mask.lshr(ShiftAmt));
761 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
762 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
763 KnownZero <<= ShiftAmt;
764 KnownOne <<= ShiftAmt;
765 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
769 case Instruction::LShr:
770 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
771 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
772 // Compute the new bits that are at the top now.
773 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
775 // Unsigned shift right.
776 APInt Mask2(Mask.shl(ShiftAmt));
777 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
778 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
779 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
780 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
781 // high bits known zero.
782 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
786 case Instruction::AShr:
787 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
788 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
789 // Compute the new bits that are at the top now.
790 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
792 // Signed shift right.
793 APInt Mask2(Mask.shl(ShiftAmt));
794 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
795 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
796 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
797 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
799 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
800 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
801 KnownZero |= HighBits;
802 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
803 KnownOne |= HighBits;
810 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
811 /// this predicate to simplify operations downstream. Mask is known to be zero
812 /// for bits that V cannot have.
813 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
814 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
815 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
816 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
817 return (KnownZero & Mask) == Mask;
820 /// ShrinkDemandedConstant - Check to see if the specified operand of the
821 /// specified instruction is a constant integer. If so, check to see if there
822 /// are any bits set in the constant that are not demanded. If so, shrink the
823 /// constant and return true.
824 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
826 assert(I && "No instruction?");
827 assert(OpNo < I->getNumOperands() && "Operand index too large");
829 // If the operand is not a constant integer, nothing to do.
830 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
831 if (!OpC) return false;
833 // If there are no bits set that aren't demanded, nothing to do.
834 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
835 if ((~Demanded & OpC->getValue()) == 0)
838 // This instruction is producing bits that are not demanded. Shrink the RHS.
839 Demanded &= OpC->getValue();
840 I->setOperand(OpNo, ConstantInt::get(Demanded));
844 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
845 // set of known zero and one bits, compute the maximum and minimum values that
846 // could have the specified known zero and known one bits, returning them in
848 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
849 const APInt& KnownZero,
850 const APInt& KnownOne,
851 APInt& Min, APInt& Max) {
852 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
853 assert(KnownZero.getBitWidth() == BitWidth &&
854 KnownOne.getBitWidth() == BitWidth &&
855 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
856 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
857 APInt UnknownBits = ~(KnownZero|KnownOne);
859 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
860 // bit if it is unknown.
862 Max = KnownOne|UnknownBits;
864 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
866 Max.clear(BitWidth-1);
870 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
871 // a set of known zero and one bits, compute the maximum and minimum values that
872 // could have the specified known zero and known one bits, returning them in
874 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
875 const APInt &KnownZero,
876 const APInt &KnownOne,
877 APInt &Min, APInt &Max) {
878 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
879 assert(KnownZero.getBitWidth() == BitWidth &&
880 KnownOne.getBitWidth() == BitWidth &&
881 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
882 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
883 APInt UnknownBits = ~(KnownZero|KnownOne);
885 // The minimum value is when the unknown bits are all zeros.
887 // The maximum value is when the unknown bits are all ones.
888 Max = KnownOne|UnknownBits;
891 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
892 /// value based on the demanded bits. When this function is called, it is known
893 /// that only the bits set in DemandedMask of the result of V are ever used
894 /// downstream. Consequently, depending on the mask and V, it may be possible
895 /// to replace V with a constant or one of its operands. In such cases, this
896 /// function does the replacement and returns true. In all other cases, it
897 /// returns false after analyzing the expression and setting KnownOne and known
898 /// to be one in the expression. KnownZero contains all the bits that are known
899 /// to be zero in the expression. These are provided to potentially allow the
900 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
901 /// the expression. KnownOne and KnownZero always follow the invariant that
902 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
903 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
904 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
905 /// and KnownOne must all be the same.
906 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
907 APInt& KnownZero, APInt& KnownOne,
909 assert(V != 0 && "Null pointer of Value???");
910 assert(Depth <= 6 && "Limit Search Depth");
911 uint32_t BitWidth = DemandedMask.getBitWidth();
912 const IntegerType *VTy = cast<IntegerType>(V->getType());
913 assert(VTy->getBitWidth() == BitWidth &&
914 KnownZero.getBitWidth() == BitWidth &&
915 KnownOne.getBitWidth() == BitWidth &&
916 "Value *V, DemandedMask, KnownZero and KnownOne \
917 must have same BitWidth");
918 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
919 // We know all of the bits for a constant!
920 KnownOne = CI->getValue() & DemandedMask;
921 KnownZero = ~KnownOne & DemandedMask;
927 if (!V->hasOneUse()) { // Other users may use these bits.
928 if (Depth != 0) { // Not at the root.
929 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
930 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
933 // If this is the root being simplified, allow it to have multiple uses,
934 // just set the DemandedMask to all bits.
935 DemandedMask = APInt::getAllOnesValue(BitWidth);
936 } else if (DemandedMask == 0) { // Not demanding any bits from V.
937 if (V != UndefValue::get(VTy))
938 return UpdateValueUsesWith(V, UndefValue::get(VTy));
940 } else if (Depth == 6) { // Limit search depth.
944 Instruction *I = dyn_cast<Instruction>(V);
945 if (!I) return false; // Only analyze instructions.
947 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
948 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
949 switch (I->getOpcode()) {
951 case Instruction::And:
952 // If either the LHS or the RHS are Zero, the result is zero.
953 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
954 RHSKnownZero, RHSKnownOne, Depth+1))
956 assert((RHSKnownZero & RHSKnownOne) == 0 &&
957 "Bits known to be one AND zero?");
959 // If something is known zero on the RHS, the bits aren't demanded on the
961 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
962 LHSKnownZero, LHSKnownOne, Depth+1))
964 assert((LHSKnownZero & LHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
967 // If all of the demanded bits are known 1 on one side, return the other.
968 // These bits cannot contribute to the result of the 'and'.
969 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
970 (DemandedMask & ~LHSKnownZero))
971 return UpdateValueUsesWith(I, I->getOperand(0));
972 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
973 (DemandedMask & ~RHSKnownZero))
974 return UpdateValueUsesWith(I, I->getOperand(1));
976 // If all of the demanded bits in the inputs are known zeros, return zero.
977 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
978 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
980 // If the RHS is a constant, see if we can simplify it.
981 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
982 return UpdateValueUsesWith(I, I);
984 // Output known-1 bits are only known if set in both the LHS & RHS.
985 RHSKnownOne &= LHSKnownOne;
986 // Output known-0 are known to be clear if zero in either the LHS | RHS.
987 RHSKnownZero |= LHSKnownZero;
989 case Instruction::Or:
990 // If either the LHS or the RHS are One, the result is One.
991 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
992 RHSKnownZero, RHSKnownOne, Depth+1))
994 assert((RHSKnownZero & RHSKnownOne) == 0 &&
995 "Bits known to be one AND zero?");
996 // If something is known one on the RHS, the bits aren't demanded on the
998 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
999 LHSKnownZero, LHSKnownOne, Depth+1))
1001 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1002 "Bits known to be one AND zero?");
1004 // If all of the demanded bits are known zero on one side, return the other.
1005 // These bits cannot contribute to the result of the 'or'.
1006 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1007 (DemandedMask & ~LHSKnownOne))
1008 return UpdateValueUsesWith(I, I->getOperand(0));
1009 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1010 (DemandedMask & ~RHSKnownOne))
1011 return UpdateValueUsesWith(I, I->getOperand(1));
1013 // If all of the potentially set bits on one side are known to be set on
1014 // the other side, just use the 'other' side.
1015 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1016 (DemandedMask & (~RHSKnownZero)))
1017 return UpdateValueUsesWith(I, I->getOperand(0));
1018 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1019 (DemandedMask & (~LHSKnownZero)))
1020 return UpdateValueUsesWith(I, I->getOperand(1));
1022 // If the RHS is a constant, see if we can simplify it.
1023 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1024 return UpdateValueUsesWith(I, I);
1026 // Output known-0 bits are only known if clear in both the LHS & RHS.
1027 RHSKnownZero &= LHSKnownZero;
1028 // Output known-1 are known to be set if set in either the LHS | RHS.
1029 RHSKnownOne |= LHSKnownOne;
1031 case Instruction::Xor: {
1032 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1033 RHSKnownZero, RHSKnownOne, Depth+1))
1035 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1036 "Bits known to be one AND zero?");
1037 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1038 LHSKnownZero, LHSKnownOne, Depth+1))
1040 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1041 "Bits known to be one AND zero?");
1043 // If all of the demanded bits are known zero on one side, return the other.
1044 // These bits cannot contribute to the result of the 'xor'.
1045 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1046 return UpdateValueUsesWith(I, I->getOperand(0));
1047 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1048 return UpdateValueUsesWith(I, I->getOperand(1));
1050 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1051 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1052 (RHSKnownOne & LHSKnownOne);
1053 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1054 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1055 (RHSKnownOne & LHSKnownZero);
1057 // If all of the demanded bits are known to be zero on one side or the
1058 // other, turn this into an *inclusive* or.
1059 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1060 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1062 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1064 InsertNewInstBefore(Or, *I);
1065 return UpdateValueUsesWith(I, Or);
1068 // If all of the demanded bits on one side are known, and all of the set
1069 // bits on that side are also known to be set on the other side, turn this
1070 // into an AND, as we know the bits will be cleared.
1071 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1072 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1074 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1075 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1077 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1078 InsertNewInstBefore(And, *I);
1079 return UpdateValueUsesWith(I, And);
1083 // If the RHS is a constant, see if we can simplify it.
1084 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1085 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1086 return UpdateValueUsesWith(I, I);
1088 RHSKnownZero = KnownZeroOut;
1089 RHSKnownOne = KnownOneOut;
1092 case Instruction::Select:
1093 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1094 RHSKnownZero, RHSKnownOne, Depth+1))
1096 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1097 LHSKnownZero, LHSKnownOne, Depth+1))
1099 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1100 "Bits known to be one AND zero?");
1101 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1102 "Bits known to be one AND zero?");
1104 // If the operands are constants, see if we can simplify them.
1105 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1106 return UpdateValueUsesWith(I, I);
1107 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1108 return UpdateValueUsesWith(I, I);
1110 // Only known if known in both the LHS and RHS.
1111 RHSKnownOne &= LHSKnownOne;
1112 RHSKnownZero &= LHSKnownZero;
1114 case Instruction::Trunc: {
1116 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1117 DemandedMask.zext(truncBf);
1118 RHSKnownZero.zext(truncBf);
1119 RHSKnownOne.zext(truncBf);
1120 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1121 RHSKnownZero, RHSKnownOne, Depth+1))
1123 DemandedMask.trunc(BitWidth);
1124 RHSKnownZero.trunc(BitWidth);
1125 RHSKnownOne.trunc(BitWidth);
1126 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1127 "Bits known to be one AND zero?");
1130 case Instruction::BitCast:
1131 if (!I->getOperand(0)->getType()->isInteger())
1134 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1135 RHSKnownZero, RHSKnownOne, Depth+1))
1137 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1138 "Bits known to be one AND zero?");
1140 case Instruction::ZExt: {
1141 // Compute the bits in the result that are not present in the input.
1142 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1143 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1145 DemandedMask.trunc(SrcBitWidth);
1146 RHSKnownZero.trunc(SrcBitWidth);
1147 RHSKnownOne.trunc(SrcBitWidth);
1148 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1149 RHSKnownZero, RHSKnownOne, Depth+1))
1151 DemandedMask.zext(BitWidth);
1152 RHSKnownZero.zext(BitWidth);
1153 RHSKnownOne.zext(BitWidth);
1154 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1155 "Bits known to be one AND zero?");
1156 // The top bits are known to be zero.
1157 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1160 case Instruction::SExt: {
1161 // Compute the bits in the result that are not present in the input.
1162 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1163 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1165 APInt InputDemandedBits = DemandedMask &
1166 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1168 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1169 // If any of the sign extended bits are demanded, we know that the sign
1171 if ((NewBits & DemandedMask) != 0)
1172 InputDemandedBits.set(SrcBitWidth-1);
1174 InputDemandedBits.trunc(SrcBitWidth);
1175 RHSKnownZero.trunc(SrcBitWidth);
1176 RHSKnownOne.trunc(SrcBitWidth);
1177 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1178 RHSKnownZero, RHSKnownOne, Depth+1))
1180 InputDemandedBits.zext(BitWidth);
1181 RHSKnownZero.zext(BitWidth);
1182 RHSKnownOne.zext(BitWidth);
1183 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1184 "Bits known to be one AND zero?");
1186 // If the sign bit of the input is known set or clear, then we know the
1187 // top bits of the result.
1189 // If the input sign bit is known zero, or if the NewBits are not demanded
1190 // convert this into a zero extension.
1191 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1193 // Convert to ZExt cast
1194 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1195 return UpdateValueUsesWith(I, NewCast);
1196 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1197 RHSKnownOne |= NewBits;
1201 case Instruction::Add: {
1202 // Figure out what the input bits are. If the top bits of the and result
1203 // are not demanded, then the add doesn't demand them from its input
1205 uint32_t NLZ = DemandedMask.countLeadingZeros();
1207 // If there is a constant on the RHS, there are a variety of xformations
1209 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1210 // If null, this should be simplified elsewhere. Some of the xforms here
1211 // won't work if the RHS is zero.
1215 // If the top bit of the output is demanded, demand everything from the
1216 // input. Otherwise, we demand all the input bits except NLZ top bits.
1217 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1219 // Find information about known zero/one bits in the input.
1220 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1221 LHSKnownZero, LHSKnownOne, Depth+1))
1224 // If the RHS of the add has bits set that can't affect the input, reduce
1226 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1227 return UpdateValueUsesWith(I, I);
1229 // Avoid excess work.
1230 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1233 // Turn it into OR if input bits are zero.
1234 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1236 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1238 InsertNewInstBefore(Or, *I);
1239 return UpdateValueUsesWith(I, Or);
1242 // We can say something about the output known-zero and known-one bits,
1243 // depending on potential carries from the input constant and the
1244 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1245 // bits set and the RHS constant is 0x01001, then we know we have a known
1246 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1248 // To compute this, we first compute the potential carry bits. These are
1249 // the bits which may be modified. I'm not aware of a better way to do
1251 const APInt& RHSVal = RHS->getValue();
1252 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1254 // Now that we know which bits have carries, compute the known-1/0 sets.
1256 // Bits are known one if they are known zero in one operand and one in the
1257 // other, and there is no input carry.
1258 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1259 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1261 // Bits are known zero if they are known zero in both operands and there
1262 // is no input carry.
1263 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1265 // If the high-bits of this ADD are not demanded, then it does not demand
1266 // the high bits of its LHS or RHS.
1267 if (DemandedMask[BitWidth-1] == 0) {
1268 // Right fill the mask of bits for this ADD to demand the most
1269 // significant bit and all those below it.
1270 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1271 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1272 LHSKnownZero, LHSKnownOne, Depth+1))
1274 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1275 LHSKnownZero, LHSKnownOne, Depth+1))
1281 case Instruction::Sub:
1282 // If the high-bits of this SUB are not demanded, then it does not demand
1283 // the high bits of its LHS or RHS.
1284 if (DemandedMask[BitWidth-1] == 0) {
1285 // Right fill the mask of bits for this SUB to demand the most
1286 // significant bit and all those below it.
1287 uint32_t NLZ = DemandedMask.countLeadingZeros();
1288 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1289 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1290 LHSKnownZero, LHSKnownOne, Depth+1))
1292 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1293 LHSKnownZero, LHSKnownOne, Depth+1))
1297 case Instruction::Shl:
1298 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1299 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1300 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1302 RHSKnownZero, RHSKnownOne, Depth+1))
1304 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1305 "Bits known to be one AND zero?");
1306 RHSKnownZero <<= ShiftAmt;
1307 RHSKnownOne <<= ShiftAmt;
1308 // low bits known zero.
1310 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1313 case Instruction::LShr:
1314 // For a logical shift right
1315 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1316 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1318 // Unsigned shift right.
1319 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1320 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1321 RHSKnownZero, RHSKnownOne, Depth+1))
1323 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1324 "Bits known to be one AND zero?");
1325 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1326 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1328 // Compute the new bits that are at the top now.
1329 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1330 RHSKnownZero |= HighBits; // high bits known zero.
1334 case Instruction::AShr:
1335 // If this is an arithmetic shift right and only the low-bit is set, we can
1336 // always convert this into a logical shr, even if the shift amount is
1337 // variable. The low bit of the shift cannot be an input sign bit unless
1338 // the shift amount is >= the size of the datatype, which is undefined.
1339 if (DemandedMask == 1) {
1340 // Perform the logical shift right.
1341 Value *NewVal = BinaryOperator::createLShr(
1342 I->getOperand(0), I->getOperand(1), I->getName());
1343 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1344 return UpdateValueUsesWith(I, NewVal);
1347 // If the sign bit is the only bit demanded by this ashr, then there is no
1348 // need to do it, the shift doesn't change the high bit.
1349 if (DemandedMask.isSignBit())
1350 return UpdateValueUsesWith(I, I->getOperand(0));
1352 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1353 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1355 // Signed shift right.
1356 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1357 // If any of the "high bits" are demanded, we should set the sign bit as
1359 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1360 DemandedMaskIn.set(BitWidth-1);
1361 if (SimplifyDemandedBits(I->getOperand(0),
1363 RHSKnownZero, RHSKnownOne, Depth+1))
1365 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1366 "Bits known to be one AND zero?");
1367 // Compute the new bits that are at the top now.
1368 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1369 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1370 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1372 // Handle the sign bits.
1373 APInt SignBit(APInt::getSignBit(BitWidth));
1374 // Adjust to where it is now in the mask.
1375 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1377 // If the input sign bit is known to be zero, or if none of the top bits
1378 // are demanded, turn this into an unsigned shift right.
1379 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1380 (HighBits & ~DemandedMask) == HighBits) {
1381 // Perform the logical shift right.
1382 Value *NewVal = BinaryOperator::createLShr(
1383 I->getOperand(0), SA, I->getName());
1384 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1385 return UpdateValueUsesWith(I, NewVal);
1386 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1387 RHSKnownOne |= HighBits;
1393 // If the client is only demanding bits that we know, return the known
1395 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1396 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1401 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1402 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1403 /// actually used by the caller. This method analyzes which elements of the
1404 /// operand are undef and returns that information in UndefElts.
1406 /// If the information about demanded elements can be used to simplify the
1407 /// operation, the operation is simplified, then the resultant value is
1408 /// returned. This returns null if no change was made.
1409 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1410 uint64_t &UndefElts,
1412 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1413 assert(VWidth <= 64 && "Vector too wide to analyze!");
1414 uint64_t EltMask = ~0ULL >> (64-VWidth);
1415 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1416 "Invalid DemandedElts!");
1418 if (isa<UndefValue>(V)) {
1419 // If the entire vector is undefined, just return this info.
1420 UndefElts = EltMask;
1422 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1423 UndefElts = EltMask;
1424 return UndefValue::get(V->getType());
1428 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1429 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1430 Constant *Undef = UndefValue::get(EltTy);
1432 std::vector<Constant*> Elts;
1433 for (unsigned i = 0; i != VWidth; ++i)
1434 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1435 Elts.push_back(Undef);
1436 UndefElts |= (1ULL << i);
1437 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1438 Elts.push_back(Undef);
1439 UndefElts |= (1ULL << i);
1440 } else { // Otherwise, defined.
1441 Elts.push_back(CP->getOperand(i));
1444 // If we changed the constant, return it.
1445 Constant *NewCP = ConstantVector::get(Elts);
1446 return NewCP != CP ? NewCP : 0;
1447 } else if (isa<ConstantAggregateZero>(V)) {
1448 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1450 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1451 Constant *Zero = Constant::getNullValue(EltTy);
1452 Constant *Undef = UndefValue::get(EltTy);
1453 std::vector<Constant*> Elts;
1454 for (unsigned i = 0; i != VWidth; ++i)
1455 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1456 UndefElts = DemandedElts ^ EltMask;
1457 return ConstantVector::get(Elts);
1460 if (!V->hasOneUse()) { // Other users may use these bits.
1461 if (Depth != 0) { // Not at the root.
1462 // TODO: Just compute the UndefElts information recursively.
1466 } else if (Depth == 10) { // Limit search depth.
1470 Instruction *I = dyn_cast<Instruction>(V);
1471 if (!I) return false; // Only analyze instructions.
1473 bool MadeChange = false;
1474 uint64_t UndefElts2;
1476 switch (I->getOpcode()) {
1479 case Instruction::InsertElement: {
1480 // If this is a variable index, we don't know which element it overwrites.
1481 // demand exactly the same input as we produce.
1482 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1484 // Note that we can't propagate undef elt info, because we don't know
1485 // which elt is getting updated.
1486 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1487 UndefElts2, Depth+1);
1488 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1492 // If this is inserting an element that isn't demanded, remove this
1494 unsigned IdxNo = Idx->getZExtValue();
1495 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1496 return AddSoonDeadInstToWorklist(*I, 0);
1498 // Otherwise, the element inserted overwrites whatever was there, so the
1499 // input demanded set is simpler than the output set.
1500 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1501 DemandedElts & ~(1ULL << IdxNo),
1502 UndefElts, Depth+1);
1503 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1505 // The inserted element is defined.
1506 UndefElts |= 1ULL << IdxNo;
1509 case Instruction::BitCast: {
1510 // Vector->vector casts only.
1511 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1513 unsigned InVWidth = VTy->getNumElements();
1514 uint64_t InputDemandedElts = 0;
1517 if (VWidth == InVWidth) {
1518 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1519 // elements as are demanded of us.
1521 InputDemandedElts = DemandedElts;
1522 } else if (VWidth > InVWidth) {
1526 // If there are more elements in the result than there are in the source,
1527 // then an input element is live if any of the corresponding output
1528 // elements are live.
1529 Ratio = VWidth/InVWidth;
1530 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1531 if (DemandedElts & (1ULL << OutIdx))
1532 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1538 // If there are more elements in the source than there are in the result,
1539 // then an input element is live if the corresponding output element is
1541 Ratio = InVWidth/VWidth;
1542 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1543 if (DemandedElts & (1ULL << InIdx/Ratio))
1544 InputDemandedElts |= 1ULL << InIdx;
1547 // div/rem demand all inputs, because they don't want divide by zero.
1548 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1549 UndefElts2, Depth+1);
1551 I->setOperand(0, TmpV);
1555 UndefElts = UndefElts2;
1556 if (VWidth > InVWidth) {
1557 assert(0 && "Unimp");
1558 // If there are more elements in the result than there are in the source,
1559 // then an output element is undef if the corresponding input element is
1561 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1562 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1563 UndefElts |= 1ULL << OutIdx;
1564 } else if (VWidth < InVWidth) {
1565 assert(0 && "Unimp");
1566 // If there are more elements in the source than there are in the result,
1567 // then a result element is undef if all of the corresponding input
1568 // elements are undef.
1569 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1570 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1571 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1572 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1576 case Instruction::And:
1577 case Instruction::Or:
1578 case Instruction::Xor:
1579 case Instruction::Add:
1580 case Instruction::Sub:
1581 case Instruction::Mul:
1582 // div/rem demand all inputs, because they don't want divide by zero.
1583 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1584 UndefElts, Depth+1);
1585 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1586 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1587 UndefElts2, Depth+1);
1588 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1590 // Output elements are undefined if both are undefined. Consider things
1591 // like undef&0. The result is known zero, not undef.
1592 UndefElts &= UndefElts2;
1595 case Instruction::Call: {
1596 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1598 switch (II->getIntrinsicID()) {
1601 // Binary vector operations that work column-wise. A dest element is a
1602 // function of the corresponding input elements from the two inputs.
1603 case Intrinsic::x86_sse_sub_ss:
1604 case Intrinsic::x86_sse_mul_ss:
1605 case Intrinsic::x86_sse_min_ss:
1606 case Intrinsic::x86_sse_max_ss:
1607 case Intrinsic::x86_sse2_sub_sd:
1608 case Intrinsic::x86_sse2_mul_sd:
1609 case Intrinsic::x86_sse2_min_sd:
1610 case Intrinsic::x86_sse2_max_sd:
1611 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1612 UndefElts, Depth+1);
1613 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1614 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1615 UndefElts2, Depth+1);
1616 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1618 // If only the low elt is demanded and this is a scalarizable intrinsic,
1619 // scalarize it now.
1620 if (DemandedElts == 1) {
1621 switch (II->getIntrinsicID()) {
1623 case Intrinsic::x86_sse_sub_ss:
1624 case Intrinsic::x86_sse_mul_ss:
1625 case Intrinsic::x86_sse2_sub_sd:
1626 case Intrinsic::x86_sse2_mul_sd:
1627 // TODO: Lower MIN/MAX/ABS/etc
1628 Value *LHS = II->getOperand(1);
1629 Value *RHS = II->getOperand(2);
1630 // Extract the element as scalars.
1631 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1632 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1634 switch (II->getIntrinsicID()) {
1635 default: assert(0 && "Case stmts out of sync!");
1636 case Intrinsic::x86_sse_sub_ss:
1637 case Intrinsic::x86_sse2_sub_sd:
1638 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1639 II->getName()), *II);
1641 case Intrinsic::x86_sse_mul_ss:
1642 case Intrinsic::x86_sse2_mul_sd:
1643 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1644 II->getName()), *II);
1649 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1651 InsertNewInstBefore(New, *II);
1652 AddSoonDeadInstToWorklist(*II, 0);
1657 // Output elements are undefined if both are undefined. Consider things
1658 // like undef&0. The result is known zero, not undef.
1659 UndefElts &= UndefElts2;
1665 return MadeChange ? I : 0;
1668 /// @returns true if the specified compare predicate is
1669 /// true when both operands are equal...
1670 /// @brief Determine if the icmp Predicate is true when both operands are equal
1671 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1672 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1673 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1674 pred == ICmpInst::ICMP_SLE;
1677 /// @returns true if the specified compare instruction is
1678 /// true when both operands are equal...
1679 /// @brief Determine if the ICmpInst returns true when both operands are equal
1680 static bool isTrueWhenEqual(ICmpInst &ICI) {
1681 return isTrueWhenEqual(ICI.getPredicate());
1684 /// AssociativeOpt - Perform an optimization on an associative operator. This
1685 /// function is designed to check a chain of associative operators for a
1686 /// potential to apply a certain optimization. Since the optimization may be
1687 /// applicable if the expression was reassociated, this checks the chain, then
1688 /// reassociates the expression as necessary to expose the optimization
1689 /// opportunity. This makes use of a special Functor, which must define
1690 /// 'shouldApply' and 'apply' methods.
1692 template<typename Functor>
1693 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1694 unsigned Opcode = Root.getOpcode();
1695 Value *LHS = Root.getOperand(0);
1697 // Quick check, see if the immediate LHS matches...
1698 if (F.shouldApply(LHS))
1699 return F.apply(Root);
1701 // Otherwise, if the LHS is not of the same opcode as the root, return.
1702 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1703 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1704 // Should we apply this transform to the RHS?
1705 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1707 // If not to the RHS, check to see if we should apply to the LHS...
1708 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1709 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1713 // If the functor wants to apply the optimization to the RHS of LHSI,
1714 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1716 BasicBlock *BB = Root.getParent();
1718 // Now all of the instructions are in the current basic block, go ahead
1719 // and perform the reassociation.
1720 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1722 // First move the selected RHS to the LHS of the root...
1723 Root.setOperand(0, LHSI->getOperand(1));
1725 // Make what used to be the LHS of the root be the user of the root...
1726 Value *ExtraOperand = TmpLHSI->getOperand(1);
1727 if (&Root == TmpLHSI) {
1728 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1731 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1732 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1733 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1734 BasicBlock::iterator ARI = &Root; ++ARI;
1735 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1738 // Now propagate the ExtraOperand down the chain of instructions until we
1740 while (TmpLHSI != LHSI) {
1741 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1742 // Move the instruction to immediately before the chain we are
1743 // constructing to avoid breaking dominance properties.
1744 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1745 BB->getInstList().insert(ARI, NextLHSI);
1748 Value *NextOp = NextLHSI->getOperand(1);
1749 NextLHSI->setOperand(1, ExtraOperand);
1751 ExtraOperand = NextOp;
1754 // Now that the instructions are reassociated, have the functor perform
1755 // the transformation...
1756 return F.apply(Root);
1759 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1765 // AddRHS - Implements: X + X --> X << 1
1768 AddRHS(Value *rhs) : RHS(rhs) {}
1769 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1770 Instruction *apply(BinaryOperator &Add) const {
1771 return BinaryOperator::createShl(Add.getOperand(0),
1772 ConstantInt::get(Add.getType(), 1));
1776 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1778 struct AddMaskingAnd {
1780 AddMaskingAnd(Constant *c) : C2(c) {}
1781 bool shouldApply(Value *LHS) const {
1783 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1784 ConstantExpr::getAnd(C1, C2)->isNullValue();
1786 Instruction *apply(BinaryOperator &Add) const {
1787 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1791 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1793 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1794 if (Constant *SOC = dyn_cast<Constant>(SO))
1795 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1797 return IC->InsertNewInstBefore(CastInst::create(
1798 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1801 // Figure out if the constant is the left or the right argument.
1802 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1803 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1805 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1807 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1808 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1811 Value *Op0 = SO, *Op1 = ConstOperand;
1813 std::swap(Op0, Op1);
1815 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1816 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1817 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1818 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1819 SO->getName()+".cmp");
1821 assert(0 && "Unknown binary instruction type!");
1824 return IC->InsertNewInstBefore(New, I);
1827 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1828 // constant as the other operand, try to fold the binary operator into the
1829 // select arguments. This also works for Cast instructions, which obviously do
1830 // not have a second operand.
1831 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1833 // Don't modify shared select instructions
1834 if (!SI->hasOneUse()) return 0;
1835 Value *TV = SI->getOperand(1);
1836 Value *FV = SI->getOperand(2);
1838 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1839 // Bool selects with constant operands can be folded to logical ops.
1840 if (SI->getType() == Type::Int1Ty) return 0;
1842 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1843 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1845 return new SelectInst(SI->getCondition(), SelectTrueVal,
1852 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1853 /// node as operand #0, see if we can fold the instruction into the PHI (which
1854 /// is only possible if all operands to the PHI are constants).
1855 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1856 PHINode *PN = cast<PHINode>(I.getOperand(0));
1857 unsigned NumPHIValues = PN->getNumIncomingValues();
1858 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1860 // Check to see if all of the operands of the PHI are constants. If there is
1861 // one non-constant value, remember the BB it is. If there is more than one
1862 // or if *it* is a PHI, bail out.
1863 BasicBlock *NonConstBB = 0;
1864 for (unsigned i = 0; i != NumPHIValues; ++i)
1865 if (!isa<Constant>(PN->getIncomingValue(i))) {
1866 if (NonConstBB) return 0; // More than one non-const value.
1867 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1868 NonConstBB = PN->getIncomingBlock(i);
1870 // If the incoming non-constant value is in I's block, we have an infinite
1872 if (NonConstBB == I.getParent())
1876 // If there is exactly one non-constant value, we can insert a copy of the
1877 // operation in that block. However, if this is a critical edge, we would be
1878 // inserting the computation one some other paths (e.g. inside a loop). Only
1879 // do this if the pred block is unconditionally branching into the phi block.
1881 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1882 if (!BI || !BI->isUnconditional()) return 0;
1885 // Okay, we can do the transformation: create the new PHI node.
1886 PHINode *NewPN = new PHINode(I.getType(), "");
1887 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1888 InsertNewInstBefore(NewPN, *PN);
1889 NewPN->takeName(PN);
1891 // Next, add all of the operands to the PHI.
1892 if (I.getNumOperands() == 2) {
1893 Constant *C = cast<Constant>(I.getOperand(1));
1894 for (unsigned i = 0; i != NumPHIValues; ++i) {
1896 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1897 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1898 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1900 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1902 assert(PN->getIncomingBlock(i) == NonConstBB);
1903 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1904 InV = BinaryOperator::create(BO->getOpcode(),
1905 PN->getIncomingValue(i), C, "phitmp",
1906 NonConstBB->getTerminator());
1907 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1908 InV = CmpInst::create(CI->getOpcode(),
1910 PN->getIncomingValue(i), C, "phitmp",
1911 NonConstBB->getTerminator());
1913 assert(0 && "Unknown binop!");
1915 AddToWorkList(cast<Instruction>(InV));
1917 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1920 CastInst *CI = cast<CastInst>(&I);
1921 const Type *RetTy = CI->getType();
1922 for (unsigned i = 0; i != NumPHIValues; ++i) {
1924 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1925 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1927 assert(PN->getIncomingBlock(i) == NonConstBB);
1928 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1929 I.getType(), "phitmp",
1930 NonConstBB->getTerminator());
1931 AddToWorkList(cast<Instruction>(InV));
1933 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1936 return ReplaceInstUsesWith(I, NewPN);
1939 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1940 bool Changed = SimplifyCommutative(I);
1941 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1943 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1944 // X + undef -> undef
1945 if (isa<UndefValue>(RHS))
1946 return ReplaceInstUsesWith(I, RHS);
1949 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1950 if (RHSC->isNullValue())
1951 return ReplaceInstUsesWith(I, LHS);
1952 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1953 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1954 (I.getType())->getValueAPF()))
1955 return ReplaceInstUsesWith(I, LHS);
1958 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1959 // X + (signbit) --> X ^ signbit
1960 const APInt& Val = CI->getValue();
1961 uint32_t BitWidth = Val.getBitWidth();
1962 if (Val == APInt::getSignBit(BitWidth))
1963 return BinaryOperator::createXor(LHS, RHS);
1965 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1966 // (X & 254)+1 -> (X&254)|1
1967 if (!isa<VectorType>(I.getType())) {
1968 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1969 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1970 KnownZero, KnownOne))
1975 if (isa<PHINode>(LHS))
1976 if (Instruction *NV = FoldOpIntoPhi(I))
1979 ConstantInt *XorRHS = 0;
1981 if (isa<ConstantInt>(RHSC) &&
1982 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1983 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1984 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1986 uint32_t Size = TySizeBits / 2;
1987 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1988 APInt CFF80Val(-C0080Val);
1990 if (TySizeBits > Size) {
1991 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1992 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1993 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1994 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1995 // This is a sign extend if the top bits are known zero.
1996 if (!MaskedValueIsZero(XorLHS,
1997 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1998 Size = 0; // Not a sign ext, but can't be any others either.
2003 C0080Val = APIntOps::lshr(C0080Val, Size);
2004 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2005 } while (Size >= 1);
2007 // FIXME: This shouldn't be necessary. When the backends can handle types
2008 // with funny bit widths then this whole cascade of if statements should
2009 // be removed. It is just here to get the size of the "middle" type back
2010 // up to something that the back ends can handle.
2011 const Type *MiddleType = 0;
2014 case 32: MiddleType = Type::Int32Ty; break;
2015 case 16: MiddleType = Type::Int16Ty; break;
2016 case 8: MiddleType = Type::Int8Ty; break;
2019 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2020 InsertNewInstBefore(NewTrunc, I);
2021 return new SExtInst(NewTrunc, I.getType(), I.getName());
2027 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2028 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2030 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2031 if (RHSI->getOpcode() == Instruction::Sub)
2032 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2033 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2035 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2036 if (LHSI->getOpcode() == Instruction::Sub)
2037 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2038 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2043 if (Value *V = dyn_castNegVal(LHS))
2044 return BinaryOperator::createSub(RHS, V);
2047 if (!isa<Constant>(RHS))
2048 if (Value *V = dyn_castNegVal(RHS))
2049 return BinaryOperator::createSub(LHS, V);
2053 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2054 if (X == RHS) // X*C + X --> X * (C+1)
2055 return BinaryOperator::createMul(RHS, AddOne(C2));
2057 // X*C1 + X*C2 --> X * (C1+C2)
2059 if (X == dyn_castFoldableMul(RHS, C1))
2060 return BinaryOperator::createMul(X, Add(C1, C2));
2063 // X + X*C --> X * (C+1)
2064 if (dyn_castFoldableMul(RHS, C2) == LHS)
2065 return BinaryOperator::createMul(LHS, AddOne(C2));
2067 // X + ~X --> -1 since ~X = -X-1
2068 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2069 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2072 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2073 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2074 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2077 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2079 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2080 return BinaryOperator::createSub(SubOne(CRHS), X);
2082 // (X & FF00) + xx00 -> (X+xx00) & FF00
2083 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2084 Constant *Anded = And(CRHS, C2);
2085 if (Anded == CRHS) {
2086 // See if all bits from the first bit set in the Add RHS up are included
2087 // in the mask. First, get the rightmost bit.
2088 const APInt& AddRHSV = CRHS->getValue();
2090 // Form a mask of all bits from the lowest bit added through the top.
2091 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2093 // See if the and mask includes all of these bits.
2094 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2096 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2097 // Okay, the xform is safe. Insert the new add pronto.
2098 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2099 LHS->getName()), I);
2100 return BinaryOperator::createAnd(NewAdd, C2);
2105 // Try to fold constant add into select arguments.
2106 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2107 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2111 // add (cast *A to intptrtype) B ->
2112 // cast (GEP (cast *A to sbyte*) B) ->
2115 CastInst *CI = dyn_cast<CastInst>(LHS);
2118 CI = dyn_cast<CastInst>(RHS);
2121 if (CI && CI->getType()->isSized() &&
2122 (CI->getType()->getPrimitiveSizeInBits() ==
2123 TD->getIntPtrType()->getPrimitiveSizeInBits())
2124 && isa<PointerType>(CI->getOperand(0)->getType())) {
2125 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2126 PointerType::get(Type::Int8Ty), I);
2127 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2128 return new PtrToIntInst(I2, CI->getType());
2132 return Changed ? &I : 0;
2135 // isSignBit - Return true if the value represented by the constant only has the
2136 // highest order bit set.
2137 static bool isSignBit(ConstantInt *CI) {
2138 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2139 return CI->getValue() == APInt::getSignBit(NumBits);
2142 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2143 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2145 if (Op0 == Op1) // sub X, X -> 0
2146 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2148 // If this is a 'B = x-(-A)', change to B = x+A...
2149 if (Value *V = dyn_castNegVal(Op1))
2150 return BinaryOperator::createAdd(Op0, V);
2152 if (isa<UndefValue>(Op0))
2153 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2154 if (isa<UndefValue>(Op1))
2155 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2157 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2158 // Replace (-1 - A) with (~A)...
2159 if (C->isAllOnesValue())
2160 return BinaryOperator::createNot(Op1);
2162 // C - ~X == X + (1+C)
2164 if (match(Op1, m_Not(m_Value(X))))
2165 return BinaryOperator::createAdd(X, AddOne(C));
2167 // -(X >>u 31) -> (X >>s 31)
2168 // -(X >>s 31) -> (X >>u 31)
2170 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2171 if (SI->getOpcode() == Instruction::LShr) {
2172 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2173 // Check to see if we are shifting out everything but the sign bit.
2174 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2175 SI->getType()->getPrimitiveSizeInBits()-1) {
2176 // Ok, the transformation is safe. Insert AShr.
2177 return BinaryOperator::create(Instruction::AShr,
2178 SI->getOperand(0), CU, SI->getName());
2182 else if (SI->getOpcode() == Instruction::AShr) {
2183 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2184 // Check to see if we are shifting out everything but the sign bit.
2185 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2186 SI->getType()->getPrimitiveSizeInBits()-1) {
2187 // Ok, the transformation is safe. Insert LShr.
2188 return BinaryOperator::createLShr(
2189 SI->getOperand(0), CU, SI->getName());
2195 // Try to fold constant sub into select arguments.
2196 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2200 if (isa<PHINode>(Op0))
2201 if (Instruction *NV = FoldOpIntoPhi(I))
2205 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2206 if (Op1I->getOpcode() == Instruction::Add &&
2207 !Op0->getType()->isFPOrFPVector()) {
2208 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2209 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2210 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2211 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2212 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2213 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2214 // C1-(X+C2) --> (C1-C2)-X
2215 return BinaryOperator::createSub(Subtract(CI1, CI2),
2216 Op1I->getOperand(0));
2220 if (Op1I->hasOneUse()) {
2221 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2222 // is not used by anyone else...
2224 if (Op1I->getOpcode() == Instruction::Sub &&
2225 !Op1I->getType()->isFPOrFPVector()) {
2226 // Swap the two operands of the subexpr...
2227 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2228 Op1I->setOperand(0, IIOp1);
2229 Op1I->setOperand(1, IIOp0);
2231 // Create the new top level add instruction...
2232 return BinaryOperator::createAdd(Op0, Op1);
2235 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2237 if (Op1I->getOpcode() == Instruction::And &&
2238 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2239 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2242 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2243 return BinaryOperator::createAnd(Op0, NewNot);
2246 // 0 - (X sdiv C) -> (X sdiv -C)
2247 if (Op1I->getOpcode() == Instruction::SDiv)
2248 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2250 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2251 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2252 ConstantExpr::getNeg(DivRHS));
2254 // X - X*C --> X * (1-C)
2255 ConstantInt *C2 = 0;
2256 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2257 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2258 return BinaryOperator::createMul(Op0, CP1);
2261 // X - ((X / Y) * Y) --> X % Y
2262 if (Op1I->getOpcode() == Instruction::Mul)
2263 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2264 if (Op0 == I->getOperand(0) &&
2265 Op1I->getOperand(1) == I->getOperand(1)) {
2266 if (I->getOpcode() == Instruction::SDiv)
2267 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2268 if (I->getOpcode() == Instruction::UDiv)
2269 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2274 if (!Op0->getType()->isFPOrFPVector())
2275 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2276 if (Op0I->getOpcode() == Instruction::Add) {
2277 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2278 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2279 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2280 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2281 } else if (Op0I->getOpcode() == Instruction::Sub) {
2282 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2283 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2287 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2288 if (X == Op1) // X*C - X --> X * (C-1)
2289 return BinaryOperator::createMul(Op1, SubOne(C1));
2291 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2292 if (X == dyn_castFoldableMul(Op1, C2))
2293 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2298 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2299 /// comparison only checks the sign bit. If it only checks the sign bit, set
2300 /// TrueIfSigned if the result of the comparison is true when the input value is
2302 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2303 bool &TrueIfSigned) {
2305 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2306 TrueIfSigned = true;
2307 return RHS->isZero();
2308 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2309 TrueIfSigned = true;
2310 return RHS->isAllOnesValue();
2311 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2312 TrueIfSigned = false;
2313 return RHS->isAllOnesValue();
2314 case ICmpInst::ICMP_UGT:
2315 // True if LHS u> RHS and RHS == high-bit-mask - 1
2316 TrueIfSigned = true;
2317 return RHS->getValue() ==
2318 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2319 case ICmpInst::ICMP_UGE:
2320 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2321 TrueIfSigned = true;
2322 return RHS->getValue() ==
2323 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2329 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2330 bool Changed = SimplifyCommutative(I);
2331 Value *Op0 = I.getOperand(0);
2333 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2334 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2336 // Simplify mul instructions with a constant RHS...
2337 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2338 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2340 // ((X << C1)*C2) == (X * (C2 << C1))
2341 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2342 if (SI->getOpcode() == Instruction::Shl)
2343 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2344 return BinaryOperator::createMul(SI->getOperand(0),
2345 ConstantExpr::getShl(CI, ShOp));
2348 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2349 if (CI->equalsInt(1)) // X * 1 == X
2350 return ReplaceInstUsesWith(I, Op0);
2351 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2352 return BinaryOperator::createNeg(Op0, I.getName());
2354 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2355 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2356 return BinaryOperator::createShl(Op0,
2357 ConstantInt::get(Op0->getType(), Val.logBase2()));
2359 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2360 if (Op1F->isNullValue())
2361 return ReplaceInstUsesWith(I, Op1);
2363 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2364 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2365 // We need a better interface for long double here.
2366 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2367 if (Op1F->isExactlyValue(1.0))
2368 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2371 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2372 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2373 isa<ConstantInt>(Op0I->getOperand(1))) {
2374 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2375 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2377 InsertNewInstBefore(Add, I);
2378 Value *C1C2 = ConstantExpr::getMul(Op1,
2379 cast<Constant>(Op0I->getOperand(1)));
2380 return BinaryOperator::createAdd(Add, C1C2);
2384 // Try to fold constant mul into select arguments.
2385 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2386 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2389 if (isa<PHINode>(Op0))
2390 if (Instruction *NV = FoldOpIntoPhi(I))
2394 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2395 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2396 return BinaryOperator::createMul(Op0v, Op1v);
2398 // If one of the operands of the multiply is a cast from a boolean value, then
2399 // we know the bool is either zero or one, so this is a 'masking' multiply.
2400 // See if we can simplify things based on how the boolean was originally
2402 CastInst *BoolCast = 0;
2403 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2404 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2407 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2408 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2411 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2412 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2413 const Type *SCOpTy = SCIOp0->getType();
2416 // If the icmp is true iff the sign bit of X is set, then convert this
2417 // multiply into a shift/and combination.
2418 if (isa<ConstantInt>(SCIOp1) &&
2419 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2421 // Shift the X value right to turn it into "all signbits".
2422 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2423 SCOpTy->getPrimitiveSizeInBits()-1);
2425 InsertNewInstBefore(
2426 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2427 BoolCast->getOperand(0)->getName()+
2430 // If the multiply type is not the same as the source type, sign extend
2431 // or truncate to the multiply type.
2432 if (I.getType() != V->getType()) {
2433 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2434 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2435 Instruction::CastOps opcode =
2436 (SrcBits == DstBits ? Instruction::BitCast :
2437 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2438 V = InsertCastBefore(opcode, V, I.getType(), I);
2441 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2442 return BinaryOperator::createAnd(V, OtherOp);
2447 return Changed ? &I : 0;
2450 /// This function implements the transforms on div instructions that work
2451 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2452 /// used by the visitors to those instructions.
2453 /// @brief Transforms common to all three div instructions
2454 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2455 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2458 if (isa<UndefValue>(Op0))
2459 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2461 // X / undef -> undef
2462 if (isa<UndefValue>(Op1))
2463 return ReplaceInstUsesWith(I, Op1);
2465 // Handle cases involving: div X, (select Cond, Y, Z)
2466 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2467 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2468 // same basic block, then we replace the select with Y, and the condition
2469 // of the select with false (if the cond value is in the same BB). If the
2470 // select has uses other than the div, this allows them to be simplified
2471 // also. Note that div X, Y is just as good as div X, 0 (undef)
2472 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2473 if (ST->isNullValue()) {
2474 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2475 if (CondI && CondI->getParent() == I.getParent())
2476 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2477 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2478 I.setOperand(1, SI->getOperand(2));
2480 UpdateValueUsesWith(SI, SI->getOperand(2));
2484 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2485 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2486 if (ST->isNullValue()) {
2487 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2488 if (CondI && CondI->getParent() == I.getParent())
2489 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2490 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2491 I.setOperand(1, SI->getOperand(1));
2493 UpdateValueUsesWith(SI, SI->getOperand(1));
2501 /// This function implements the transforms common to both integer division
2502 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2503 /// division instructions.
2504 /// @brief Common integer divide transforms
2505 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2506 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2508 if (Instruction *Common = commonDivTransforms(I))
2511 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2513 if (RHS->equalsInt(1))
2514 return ReplaceInstUsesWith(I, Op0);
2516 // (X / C1) / C2 -> X / (C1*C2)
2517 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2518 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2519 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2520 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2521 Multiply(RHS, LHSRHS));
2524 if (!RHS->isZero()) { // avoid X udiv 0
2525 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2526 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2528 if (isa<PHINode>(Op0))
2529 if (Instruction *NV = FoldOpIntoPhi(I))
2534 // 0 / X == 0, we don't need to preserve faults!
2535 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2536 if (LHS->equalsInt(0))
2537 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2542 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2543 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2545 // Handle the integer div common cases
2546 if (Instruction *Common = commonIDivTransforms(I))
2549 // X udiv C^2 -> X >> C
2550 // Check to see if this is an unsigned division with an exact power of 2,
2551 // if so, convert to a right shift.
2552 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2553 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2554 return BinaryOperator::createLShr(Op0,
2555 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2558 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2559 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2560 if (RHSI->getOpcode() == Instruction::Shl &&
2561 isa<ConstantInt>(RHSI->getOperand(0))) {
2562 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2563 if (C1.isPowerOf2()) {
2564 Value *N = RHSI->getOperand(1);
2565 const Type *NTy = N->getType();
2566 if (uint32_t C2 = C1.logBase2()) {
2567 Constant *C2V = ConstantInt::get(NTy, C2);
2568 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2570 return BinaryOperator::createLShr(Op0, N);
2575 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2576 // where C1&C2 are powers of two.
2577 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2578 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2579 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2580 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2581 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2582 // Compute the shift amounts
2583 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2584 // Construct the "on true" case of the select
2585 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2586 Instruction *TSI = BinaryOperator::createLShr(
2587 Op0, TC, SI->getName()+".t");
2588 TSI = InsertNewInstBefore(TSI, I);
2590 // Construct the "on false" case of the select
2591 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2592 Instruction *FSI = BinaryOperator::createLShr(
2593 Op0, FC, SI->getName()+".f");
2594 FSI = InsertNewInstBefore(FSI, I);
2596 // construct the select instruction and return it.
2597 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2603 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2604 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2606 // Handle the integer div common cases
2607 if (Instruction *Common = commonIDivTransforms(I))
2610 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2612 if (RHS->isAllOnesValue())
2613 return BinaryOperator::createNeg(Op0);
2616 if (Value *LHSNeg = dyn_castNegVal(Op0))
2617 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2620 // If the sign bits of both operands are zero (i.e. we can prove they are
2621 // unsigned inputs), turn this into a udiv.
2622 if (I.getType()->isInteger()) {
2623 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2624 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2625 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2632 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2633 return commonDivTransforms(I);
2636 /// GetFactor - If we can prove that the specified value is at least a multiple
2637 /// of some factor, return that factor.
2638 static Constant *GetFactor(Value *V) {
2639 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2642 // Unless we can be tricky, we know this is a multiple of 1.
2643 Constant *Result = ConstantInt::get(V->getType(), 1);
2645 Instruction *I = dyn_cast<Instruction>(V);
2646 if (!I) return Result;
2648 if (I->getOpcode() == Instruction::Mul) {
2649 // Handle multiplies by a constant, etc.
2650 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2651 GetFactor(I->getOperand(1)));
2652 } else if (I->getOpcode() == Instruction::Shl) {
2653 // (X<<C) -> X * (1 << C)
2654 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2655 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2656 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2658 } else if (I->getOpcode() == Instruction::And) {
2659 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2660 // X & 0xFFF0 is known to be a multiple of 16.
2661 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2662 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2663 return ConstantExpr::getShl(Result,
2664 ConstantInt::get(Result->getType(), Zeros));
2666 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2667 // Only handle int->int casts.
2668 if (!CI->isIntegerCast())
2670 Value *Op = CI->getOperand(0);
2671 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2676 /// This function implements the transforms on rem instructions that work
2677 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2678 /// is used by the visitors to those instructions.
2679 /// @brief Transforms common to all three rem instructions
2680 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2681 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2683 // 0 % X == 0, we don't need to preserve faults!
2684 if (Constant *LHS = dyn_cast<Constant>(Op0))
2685 if (LHS->isNullValue())
2686 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2688 if (isa<UndefValue>(Op0)) // undef % X -> 0
2689 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2690 if (isa<UndefValue>(Op1))
2691 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2693 // Handle cases involving: rem X, (select Cond, Y, Z)
2694 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2695 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2696 // the same basic block, then we replace the select with Y, and the
2697 // condition of the select with false (if the cond value is in the same
2698 // BB). If the select has uses other than the div, this allows them to be
2700 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2701 if (ST->isNullValue()) {
2702 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2703 if (CondI && CondI->getParent() == I.getParent())
2704 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2705 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2706 I.setOperand(1, SI->getOperand(2));
2708 UpdateValueUsesWith(SI, SI->getOperand(2));
2711 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2712 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2713 if (ST->isNullValue()) {
2714 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2715 if (CondI && CondI->getParent() == I.getParent())
2716 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2717 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2718 I.setOperand(1, SI->getOperand(1));
2720 UpdateValueUsesWith(SI, SI->getOperand(1));
2728 /// This function implements the transforms common to both integer remainder
2729 /// instructions (urem and srem). It is called by the visitors to those integer
2730 /// remainder instructions.
2731 /// @brief Common integer remainder transforms
2732 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2733 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2735 if (Instruction *common = commonRemTransforms(I))
2738 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2739 // X % 0 == undef, we don't need to preserve faults!
2740 if (RHS->equalsInt(0))
2741 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2743 if (RHS->equalsInt(1)) // X % 1 == 0
2744 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2746 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2747 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2748 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2750 } else if (isa<PHINode>(Op0I)) {
2751 if (Instruction *NV = FoldOpIntoPhi(I))
2754 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2755 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2756 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2763 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2764 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2766 if (Instruction *common = commonIRemTransforms(I))
2769 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2770 // X urem C^2 -> X and C
2771 // Check to see if this is an unsigned remainder with an exact power of 2,
2772 // if so, convert to a bitwise and.
2773 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2774 if (C->getValue().isPowerOf2())
2775 return BinaryOperator::createAnd(Op0, SubOne(C));
2778 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2779 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2780 if (RHSI->getOpcode() == Instruction::Shl &&
2781 isa<ConstantInt>(RHSI->getOperand(0))) {
2782 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2783 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2784 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2786 return BinaryOperator::createAnd(Op0, Add);
2791 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2792 // where C1&C2 are powers of two.
2793 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2794 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2795 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2796 // STO == 0 and SFO == 0 handled above.
2797 if ((STO->getValue().isPowerOf2()) &&
2798 (SFO->getValue().isPowerOf2())) {
2799 Value *TrueAnd = InsertNewInstBefore(
2800 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2801 Value *FalseAnd = InsertNewInstBefore(
2802 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2803 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2811 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2812 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2814 if (Instruction *common = commonIRemTransforms(I))
2817 if (Value *RHSNeg = dyn_castNegVal(Op1))
2818 if (!isa<ConstantInt>(RHSNeg) ||
2819 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2821 AddUsesToWorkList(I);
2822 I.setOperand(1, RHSNeg);
2826 // If the top bits of both operands are zero (i.e. we can prove they are
2827 // unsigned inputs), turn this into a urem.
2828 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2829 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2830 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2831 return BinaryOperator::createURem(Op0, Op1, I.getName());
2837 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2838 return commonRemTransforms(I);
2841 // isMaxValueMinusOne - return true if this is Max-1
2842 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2843 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2845 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2846 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2849 // isMinValuePlusOne - return true if this is Min+1
2850 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2852 return C->getValue() == 1; // unsigned
2854 // Calculate 1111111111000000000000
2855 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2856 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2859 // isOneBitSet - Return true if there is exactly one bit set in the specified
2861 static bool isOneBitSet(const ConstantInt *CI) {
2862 return CI->getValue().isPowerOf2();
2865 // isHighOnes - Return true if the constant is of the form 1+0+.
2866 // This is the same as lowones(~X).
2867 static bool isHighOnes(const ConstantInt *CI) {
2868 return (~CI->getValue() + 1).isPowerOf2();
2871 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2872 /// are carefully arranged to allow folding of expressions such as:
2874 /// (A < B) | (A > B) --> (A != B)
2876 /// Note that this is only valid if the first and second predicates have the
2877 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2879 /// Three bits are used to represent the condition, as follows:
2884 /// <=> Value Definition
2885 /// 000 0 Always false
2892 /// 111 7 Always true
2894 static unsigned getICmpCode(const ICmpInst *ICI) {
2895 switch (ICI->getPredicate()) {
2897 case ICmpInst::ICMP_UGT: return 1; // 001
2898 case ICmpInst::ICMP_SGT: return 1; // 001
2899 case ICmpInst::ICMP_EQ: return 2; // 010
2900 case ICmpInst::ICMP_UGE: return 3; // 011
2901 case ICmpInst::ICMP_SGE: return 3; // 011
2902 case ICmpInst::ICMP_ULT: return 4; // 100
2903 case ICmpInst::ICMP_SLT: return 4; // 100
2904 case ICmpInst::ICMP_NE: return 5; // 101
2905 case ICmpInst::ICMP_ULE: return 6; // 110
2906 case ICmpInst::ICMP_SLE: return 6; // 110
2909 assert(0 && "Invalid ICmp predicate!");
2914 /// getICmpValue - This is the complement of getICmpCode, which turns an
2915 /// opcode and two operands into either a constant true or false, or a brand
2916 /// new ICmp instruction. The sign is passed in to determine which kind
2917 /// of predicate to use in new icmp instructions.
2918 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2920 default: assert(0 && "Illegal ICmp code!");
2921 case 0: return ConstantInt::getFalse();
2924 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2926 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2927 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2930 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2932 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2935 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2937 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2938 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2941 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2943 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2944 case 7: return ConstantInt::getTrue();
2948 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2949 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2950 (ICmpInst::isSignedPredicate(p1) &&
2951 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2952 (ICmpInst::isSignedPredicate(p2) &&
2953 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2957 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2958 struct FoldICmpLogical {
2961 ICmpInst::Predicate pred;
2962 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2963 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2964 pred(ICI->getPredicate()) {}
2965 bool shouldApply(Value *V) const {
2966 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2967 if (PredicatesFoldable(pred, ICI->getPredicate()))
2968 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2969 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2972 Instruction *apply(Instruction &Log) const {
2973 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2974 if (ICI->getOperand(0) != LHS) {
2975 assert(ICI->getOperand(1) == LHS);
2976 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2979 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2980 unsigned LHSCode = getICmpCode(ICI);
2981 unsigned RHSCode = getICmpCode(RHSICI);
2983 switch (Log.getOpcode()) {
2984 case Instruction::And: Code = LHSCode & RHSCode; break;
2985 case Instruction::Or: Code = LHSCode | RHSCode; break;
2986 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2987 default: assert(0 && "Illegal logical opcode!"); return 0;
2990 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2991 ICmpInst::isSignedPredicate(ICI->getPredicate());
2993 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2994 if (Instruction *I = dyn_cast<Instruction>(RV))
2996 // Otherwise, it's a constant boolean value...
2997 return IC.ReplaceInstUsesWith(Log, RV);
3000 } // end anonymous namespace
3002 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3003 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3004 // guaranteed to be a binary operator.
3005 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3007 ConstantInt *AndRHS,
3008 BinaryOperator &TheAnd) {
3009 Value *X = Op->getOperand(0);
3010 Constant *Together = 0;
3012 Together = And(AndRHS, OpRHS);
3014 switch (Op->getOpcode()) {
3015 case Instruction::Xor:
3016 if (Op->hasOneUse()) {
3017 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3018 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3019 InsertNewInstBefore(And, TheAnd);
3021 return BinaryOperator::createXor(And, Together);
3024 case Instruction::Or:
3025 if (Together == AndRHS) // (X | C) & C --> C
3026 return ReplaceInstUsesWith(TheAnd, AndRHS);
3028 if (Op->hasOneUse() && Together != OpRHS) {
3029 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3030 Instruction *Or = BinaryOperator::createOr(X, Together);
3031 InsertNewInstBefore(Or, TheAnd);
3033 return BinaryOperator::createAnd(Or, AndRHS);
3036 case Instruction::Add:
3037 if (Op->hasOneUse()) {
3038 // Adding a one to a single bit bit-field should be turned into an XOR
3039 // of the bit. First thing to check is to see if this AND is with a
3040 // single bit constant.
3041 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3043 // If there is only one bit set...
3044 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3045 // Ok, at this point, we know that we are masking the result of the
3046 // ADD down to exactly one bit. If the constant we are adding has
3047 // no bits set below this bit, then we can eliminate the ADD.
3048 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3050 // Check to see if any bits below the one bit set in AndRHSV are set.
3051 if ((AddRHS & (AndRHSV-1)) == 0) {
3052 // If not, the only thing that can effect the output of the AND is
3053 // the bit specified by AndRHSV. If that bit is set, the effect of
3054 // the XOR is to toggle the bit. If it is clear, then the ADD has
3056 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3057 TheAnd.setOperand(0, X);
3060 // Pull the XOR out of the AND.
3061 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3062 InsertNewInstBefore(NewAnd, TheAnd);
3063 NewAnd->takeName(Op);
3064 return BinaryOperator::createXor(NewAnd, AndRHS);
3071 case Instruction::Shl: {
3072 // We know that the AND will not produce any of the bits shifted in, so if
3073 // the anded constant includes them, clear them now!
3075 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3076 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3077 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3078 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3080 if (CI->getValue() == ShlMask) {
3081 // Masking out bits that the shift already masks
3082 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3083 } else if (CI != AndRHS) { // Reducing bits set in and.
3084 TheAnd.setOperand(1, CI);
3089 case Instruction::LShr:
3091 // We know that the AND will not produce any of the bits shifted in, so if
3092 // the anded constant includes them, clear them now! This only applies to
3093 // unsigned shifts, because a signed shr may bring in set bits!
3095 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3096 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3097 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3098 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3100 if (CI->getValue() == ShrMask) {
3101 // Masking out bits that the shift already masks.
3102 return ReplaceInstUsesWith(TheAnd, Op);
3103 } else if (CI != AndRHS) {
3104 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3109 case Instruction::AShr:
3111 // See if this is shifting in some sign extension, then masking it out
3113 if (Op->hasOneUse()) {
3114 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3115 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3116 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3117 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3118 if (C == AndRHS) { // Masking out bits shifted in.
3119 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3120 // Make the argument unsigned.
3121 Value *ShVal = Op->getOperand(0);
3122 ShVal = InsertNewInstBefore(
3123 BinaryOperator::createLShr(ShVal, OpRHS,
3124 Op->getName()), TheAnd);
3125 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3134 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3135 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3136 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3137 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3138 /// insert new instructions.
3139 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3140 bool isSigned, bool Inside,
3142 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3143 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3144 "Lo is not <= Hi in range emission code!");
3147 if (Lo == Hi) // Trivially false.
3148 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3150 // V >= Min && V < Hi --> V < Hi
3151 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3152 ICmpInst::Predicate pred = (isSigned ?
3153 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3154 return new ICmpInst(pred, V, Hi);
3157 // Emit V-Lo <u Hi-Lo
3158 Constant *NegLo = ConstantExpr::getNeg(Lo);
3159 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3160 InsertNewInstBefore(Add, IB);
3161 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3162 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3165 if (Lo == Hi) // Trivially true.
3166 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3168 // V < Min || V >= Hi -> V > Hi-1
3169 Hi = SubOne(cast<ConstantInt>(Hi));
3170 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3171 ICmpInst::Predicate pred = (isSigned ?
3172 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3173 return new ICmpInst(pred, V, Hi);
3176 // Emit V-Lo >u Hi-1-Lo
3177 // Note that Hi has already had one subtracted from it, above.
3178 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3179 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3180 InsertNewInstBefore(Add, IB);
3181 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3182 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3185 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3186 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3187 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3188 // not, since all 1s are not contiguous.
3189 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3190 const APInt& V = Val->getValue();
3191 uint32_t BitWidth = Val->getType()->getBitWidth();
3192 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3194 // look for the first zero bit after the run of ones
3195 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3196 // look for the first non-zero bit
3197 ME = V.getActiveBits();
3201 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3202 /// where isSub determines whether the operator is a sub. If we can fold one of
3203 /// the following xforms:
3205 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3206 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3207 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3209 /// return (A +/- B).
3211 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3212 ConstantInt *Mask, bool isSub,
3214 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3215 if (!LHSI || LHSI->getNumOperands() != 2 ||
3216 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3218 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3220 switch (LHSI->getOpcode()) {
3222 case Instruction::And:
3223 if (And(N, Mask) == Mask) {
3224 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3225 if ((Mask->getValue().countLeadingZeros() +
3226 Mask->getValue().countPopulation()) ==
3227 Mask->getValue().getBitWidth())
3230 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3231 // part, we don't need any explicit masks to take them out of A. If that
3232 // is all N is, ignore it.
3233 uint32_t MB = 0, ME = 0;
3234 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3235 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3236 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3237 if (MaskedValueIsZero(RHS, Mask))
3242 case Instruction::Or:
3243 case Instruction::Xor:
3244 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3245 if ((Mask->getValue().countLeadingZeros() +
3246 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3247 && And(N, Mask)->isZero())
3254 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3256 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3257 return InsertNewInstBefore(New, I);
3260 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3261 bool Changed = SimplifyCommutative(I);
3262 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3264 if (isa<UndefValue>(Op1)) // X & undef -> 0
3265 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3269 return ReplaceInstUsesWith(I, Op1);
3271 // See if we can simplify any instructions used by the instruction whose sole
3272 // purpose is to compute bits we don't care about.
3273 if (!isa<VectorType>(I.getType())) {
3274 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3275 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3276 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3277 KnownZero, KnownOne))
3280 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3281 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3282 return ReplaceInstUsesWith(I, I.getOperand(0));
3283 } else if (isa<ConstantAggregateZero>(Op1)) {
3284 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3288 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3289 const APInt& AndRHSMask = AndRHS->getValue();
3290 APInt NotAndRHS(~AndRHSMask);
3292 // Optimize a variety of ((val OP C1) & C2) combinations...
3293 if (isa<BinaryOperator>(Op0)) {
3294 Instruction *Op0I = cast<Instruction>(Op0);
3295 Value *Op0LHS = Op0I->getOperand(0);
3296 Value *Op0RHS = Op0I->getOperand(1);
3297 switch (Op0I->getOpcode()) {
3298 case Instruction::Xor:
3299 case Instruction::Or:
3300 // If the mask is only needed on one incoming arm, push it up.
3301 if (Op0I->hasOneUse()) {
3302 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3303 // Not masking anything out for the LHS, move to RHS.
3304 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3305 Op0RHS->getName()+".masked");
3306 InsertNewInstBefore(NewRHS, I);
3307 return BinaryOperator::create(
3308 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3310 if (!isa<Constant>(Op0RHS) &&
3311 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3312 // Not masking anything out for the RHS, move to LHS.
3313 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3314 Op0LHS->getName()+".masked");
3315 InsertNewInstBefore(NewLHS, I);
3316 return BinaryOperator::create(
3317 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3322 case Instruction::Add:
3323 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3324 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3325 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3326 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3327 return BinaryOperator::createAnd(V, AndRHS);
3328 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3329 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3332 case Instruction::Sub:
3333 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3334 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3335 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3336 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3337 return BinaryOperator::createAnd(V, AndRHS);
3341 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3342 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3344 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3345 // If this is an integer truncation or change from signed-to-unsigned, and
3346 // if the source is an and/or with immediate, transform it. This
3347 // frequently occurs for bitfield accesses.
3348 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3349 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3350 CastOp->getNumOperands() == 2)
3351 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3352 if (CastOp->getOpcode() == Instruction::And) {
3353 // Change: and (cast (and X, C1) to T), C2
3354 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3355 // This will fold the two constants together, which may allow
3356 // other simplifications.
3357 Instruction *NewCast = CastInst::createTruncOrBitCast(
3358 CastOp->getOperand(0), I.getType(),
3359 CastOp->getName()+".shrunk");
3360 NewCast = InsertNewInstBefore(NewCast, I);
3361 // trunc_or_bitcast(C1)&C2
3362 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3363 C3 = ConstantExpr::getAnd(C3, AndRHS);
3364 return BinaryOperator::createAnd(NewCast, C3);
3365 } else if (CastOp->getOpcode() == Instruction::Or) {
3366 // Change: and (cast (or X, C1) to T), C2
3367 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3368 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3369 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3370 return ReplaceInstUsesWith(I, AndRHS);
3375 // Try to fold constant and into select arguments.
3376 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3377 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3379 if (isa<PHINode>(Op0))
3380 if (Instruction *NV = FoldOpIntoPhi(I))
3384 Value *Op0NotVal = dyn_castNotVal(Op0);
3385 Value *Op1NotVal = dyn_castNotVal(Op1);
3387 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3388 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3390 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3391 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3392 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3393 I.getName()+".demorgan");
3394 InsertNewInstBefore(Or, I);
3395 return BinaryOperator::createNot(Or);
3399 Value *A = 0, *B = 0, *C = 0, *D = 0;
3400 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3401 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3402 return ReplaceInstUsesWith(I, Op1);
3404 // (A|B) & ~(A&B) -> A^B
3405 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3406 if ((A == C && B == D) || (A == D && B == C))
3407 return BinaryOperator::createXor(A, B);
3411 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3412 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3413 return ReplaceInstUsesWith(I, Op0);
3415 // ~(A&B) & (A|B) -> A^B
3416 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3417 if ((A == C && B == D) || (A == D && B == C))
3418 return BinaryOperator::createXor(A, B);
3422 if (Op0->hasOneUse() &&
3423 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3424 if (A == Op1) { // (A^B)&A -> A&(A^B)
3425 I.swapOperands(); // Simplify below
3426 std::swap(Op0, Op1);
3427 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3428 cast<BinaryOperator>(Op0)->swapOperands();
3429 I.swapOperands(); // Simplify below
3430 std::swap(Op0, Op1);
3433 if (Op1->hasOneUse() &&
3434 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3435 if (B == Op0) { // B&(A^B) -> B&(B^A)
3436 cast<BinaryOperator>(Op1)->swapOperands();
3439 if (A == Op0) { // A&(A^B) -> A & ~B
3440 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3441 InsertNewInstBefore(NotB, I);
3442 return BinaryOperator::createAnd(A, NotB);
3447 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3448 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3449 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3452 Value *LHSVal, *RHSVal;
3453 ConstantInt *LHSCst, *RHSCst;
3454 ICmpInst::Predicate LHSCC, RHSCC;
3455 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3456 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3457 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3458 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3459 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3460 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3461 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3462 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3463 // Ensure that the larger constant is on the RHS.
3464 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3465 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3466 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3467 ICmpInst *LHS = cast<ICmpInst>(Op0);
3468 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3469 std::swap(LHS, RHS);
3470 std::swap(LHSCst, RHSCst);
3471 std::swap(LHSCC, RHSCC);
3474 // At this point, we know we have have two icmp instructions
3475 // comparing a value against two constants and and'ing the result
3476 // together. Because of the above check, we know that we only have
3477 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3478 // (from the FoldICmpLogical check above), that the two constants
3479 // are not equal and that the larger constant is on the RHS
3480 assert(LHSCst != RHSCst && "Compares not folded above?");
3483 default: assert(0 && "Unknown integer condition code!");
3484 case ICmpInst::ICMP_EQ:
3486 default: assert(0 && "Unknown integer condition code!");
3487 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3488 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3489 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3490 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3491 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3492 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3493 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3494 return ReplaceInstUsesWith(I, LHS);
3496 case ICmpInst::ICMP_NE:
3498 default: assert(0 && "Unknown integer condition code!");
3499 case ICmpInst::ICMP_ULT:
3500 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3501 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3502 break; // (X != 13 & X u< 15) -> no change
3503 case ICmpInst::ICMP_SLT:
3504 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3505 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3506 break; // (X != 13 & X s< 15) -> no change
3507 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3508 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3509 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3510 return ReplaceInstUsesWith(I, RHS);
3511 case ICmpInst::ICMP_NE:
3512 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3513 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3514 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3515 LHSVal->getName()+".off");
3516 InsertNewInstBefore(Add, I);
3517 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3518 ConstantInt::get(Add->getType(), 1));
3520 break; // (X != 13 & X != 15) -> no change
3523 case ICmpInst::ICMP_ULT:
3525 default: assert(0 && "Unknown integer condition code!");
3526 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3527 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3528 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3529 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3531 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3532 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3533 return ReplaceInstUsesWith(I, LHS);
3534 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3538 case ICmpInst::ICMP_SLT:
3540 default: assert(0 && "Unknown integer condition code!");
3541 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3542 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3543 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3544 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3546 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3547 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3548 return ReplaceInstUsesWith(I, LHS);
3549 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3553 case ICmpInst::ICMP_UGT:
3555 default: assert(0 && "Unknown integer condition code!");
3556 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3557 return ReplaceInstUsesWith(I, LHS);
3558 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3559 return ReplaceInstUsesWith(I, RHS);
3560 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3562 case ICmpInst::ICMP_NE:
3563 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3564 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3565 break; // (X u> 13 & X != 15) -> no change
3566 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3567 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3569 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3573 case ICmpInst::ICMP_SGT:
3575 default: assert(0 && "Unknown integer condition code!");
3576 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3577 return ReplaceInstUsesWith(I, LHS);
3578 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3579 return ReplaceInstUsesWith(I, RHS);
3580 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3582 case ICmpInst::ICMP_NE:
3583 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3584 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3585 break; // (X s> 13 & X != 15) -> no change
3586 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3587 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3589 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3597 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3598 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3599 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3600 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3601 const Type *SrcTy = Op0C->getOperand(0)->getType();
3602 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3603 // Only do this if the casts both really cause code to be generated.
3604 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3606 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3608 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3609 Op1C->getOperand(0),
3611 InsertNewInstBefore(NewOp, I);
3612 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3616 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3617 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3618 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3619 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3620 SI0->getOperand(1) == SI1->getOperand(1) &&
3621 (SI0->hasOneUse() || SI1->hasOneUse())) {
3622 Instruction *NewOp =
3623 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3625 SI0->getName()), I);
3626 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3627 SI1->getOperand(1));
3631 return Changed ? &I : 0;
3634 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3635 /// in the result. If it does, and if the specified byte hasn't been filled in
3636 /// yet, fill it in and return false.
3637 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3638 Instruction *I = dyn_cast<Instruction>(V);
3639 if (I == 0) return true;
3641 // If this is an or instruction, it is an inner node of the bswap.
3642 if (I->getOpcode() == Instruction::Or)
3643 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3644 CollectBSwapParts(I->getOperand(1), ByteValues);
3646 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3647 // If this is a shift by a constant int, and it is "24", then its operand
3648 // defines a byte. We only handle unsigned types here.
3649 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3650 // Not shifting the entire input by N-1 bytes?
3651 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3652 8*(ByteValues.size()-1))
3656 if (I->getOpcode() == Instruction::Shl) {
3657 // X << 24 defines the top byte with the lowest of the input bytes.
3658 DestNo = ByteValues.size()-1;
3660 // X >>u 24 defines the low byte with the highest of the input bytes.
3664 // If the destination byte value is already defined, the values are or'd
3665 // together, which isn't a bswap (unless it's an or of the same bits).
3666 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3668 ByteValues[DestNo] = I->getOperand(0);
3672 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3674 Value *Shift = 0, *ShiftLHS = 0;
3675 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3676 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3677 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3679 Instruction *SI = cast<Instruction>(Shift);
3681 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3682 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3683 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3686 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3688 if (AndAmt->getValue().getActiveBits() > 64)
3690 uint64_t AndAmtVal = AndAmt->getZExtValue();
3691 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3692 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3694 // Unknown mask for bswap.
3695 if (DestByte == ByteValues.size()) return true;
3697 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3699 if (SI->getOpcode() == Instruction::Shl)
3700 SrcByte = DestByte - ShiftBytes;
3702 SrcByte = DestByte + ShiftBytes;
3704 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3705 if (SrcByte != ByteValues.size()-DestByte-1)
3708 // If the destination byte value is already defined, the values are or'd
3709 // together, which isn't a bswap (unless it's an or of the same bits).
3710 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3712 ByteValues[DestByte] = SI->getOperand(0);
3716 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3717 /// If so, insert the new bswap intrinsic and return it.
3718 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3719 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3720 if (!ITy || ITy->getBitWidth() % 16)
3721 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3723 /// ByteValues - For each byte of the result, we keep track of which value
3724 /// defines each byte.
3725 SmallVector<Value*, 8> ByteValues;
3726 ByteValues.resize(ITy->getBitWidth()/8);
3728 // Try to find all the pieces corresponding to the bswap.
3729 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3730 CollectBSwapParts(I.getOperand(1), ByteValues))
3733 // Check to see if all of the bytes come from the same value.
3734 Value *V = ByteValues[0];
3735 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3737 // Check to make sure that all of the bytes come from the same value.
3738 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3739 if (ByteValues[i] != V)
3741 const Type *Tys[] = { ITy };
3742 Module *M = I.getParent()->getParent()->getParent();
3743 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3744 return new CallInst(F, V);
3748 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3749 bool Changed = SimplifyCommutative(I);
3750 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3752 if (isa<UndefValue>(Op1)) // X | undef -> -1
3753 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3757 return ReplaceInstUsesWith(I, Op0);
3759 // See if we can simplify any instructions used by the instruction whose sole
3760 // purpose is to compute bits we don't care about.
3761 if (!isa<VectorType>(I.getType())) {
3762 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3763 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3764 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3765 KnownZero, KnownOne))
3767 } else if (isa<ConstantAggregateZero>(Op1)) {
3768 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3769 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3770 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3771 return ReplaceInstUsesWith(I, I.getOperand(1));
3777 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3778 ConstantInt *C1 = 0; Value *X = 0;
3779 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3780 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3781 Instruction *Or = BinaryOperator::createOr(X, RHS);
3782 InsertNewInstBefore(Or, I);
3784 return BinaryOperator::createAnd(Or,
3785 ConstantInt::get(RHS->getValue() | C1->getValue()));
3788 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3789 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3790 Instruction *Or = BinaryOperator::createOr(X, RHS);
3791 InsertNewInstBefore(Or, I);
3793 return BinaryOperator::createXor(Or,
3794 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3797 // Try to fold constant and into select arguments.
3798 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3799 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3801 if (isa<PHINode>(Op0))
3802 if (Instruction *NV = FoldOpIntoPhi(I))
3806 Value *A = 0, *B = 0;
3807 ConstantInt *C1 = 0, *C2 = 0;
3809 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3810 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3811 return ReplaceInstUsesWith(I, Op1);
3812 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3813 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3814 return ReplaceInstUsesWith(I, Op0);
3816 // (A | B) | C and A | (B | C) -> bswap if possible.
3817 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3818 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3819 match(Op1, m_Or(m_Value(), m_Value())) ||
3820 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3821 match(Op1, m_Shift(m_Value(), m_Value())))) {
3822 if (Instruction *BSwap = MatchBSwap(I))
3826 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3827 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3828 MaskedValueIsZero(Op1, C1->getValue())) {
3829 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3830 InsertNewInstBefore(NOr, I);
3832 return BinaryOperator::createXor(NOr, C1);
3835 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3836 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3837 MaskedValueIsZero(Op0, C1->getValue())) {
3838 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3839 InsertNewInstBefore(NOr, I);
3841 return BinaryOperator::createXor(NOr, C1);
3845 Value *C = 0, *D = 0;
3846 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3847 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3848 Value *V1 = 0, *V2 = 0, *V3 = 0;
3849 C1 = dyn_cast<ConstantInt>(C);
3850 C2 = dyn_cast<ConstantInt>(D);
3851 if (C1 && C2) { // (A & C1)|(B & C2)
3852 // If we have: ((V + N) & C1) | (V & C2)
3853 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3854 // replace with V+N.
3855 if (C1->getValue() == ~C2->getValue()) {
3856 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3857 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3858 // Add commutes, try both ways.
3859 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3860 return ReplaceInstUsesWith(I, A);
3861 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3862 return ReplaceInstUsesWith(I, A);
3864 // Or commutes, try both ways.
3865 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3866 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3867 // Add commutes, try both ways.
3868 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3869 return ReplaceInstUsesWith(I, B);
3870 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3871 return ReplaceInstUsesWith(I, B);
3874 V1 = 0; V2 = 0; V3 = 0;
3877 // Check to see if we have any common things being and'ed. If so, find the
3878 // terms for V1 & (V2|V3).
3879 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3880 if (A == B) // (A & C)|(A & D) == A & (C|D)
3881 V1 = A, V2 = C, V3 = D;
3882 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3883 V1 = A, V2 = B, V3 = C;
3884 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3885 V1 = C, V2 = A, V3 = D;
3886 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3887 V1 = C, V2 = A, V3 = B;
3891 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3892 return BinaryOperator::createAnd(V1, Or);
3897 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3898 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3899 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3900 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3901 SI0->getOperand(1) == SI1->getOperand(1) &&
3902 (SI0->hasOneUse() || SI1->hasOneUse())) {
3903 Instruction *NewOp =
3904 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3906 SI0->getName()), I);
3907 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3908 SI1->getOperand(1));
3912 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3913 if (A == Op1) // ~A | A == -1
3914 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3918 // Note, A is still live here!
3919 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3921 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3923 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3924 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3925 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3926 I.getName()+".demorgan"), I);
3927 return BinaryOperator::createNot(And);
3931 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3932 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3933 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3936 Value *LHSVal, *RHSVal;
3937 ConstantInt *LHSCst, *RHSCst;
3938 ICmpInst::Predicate LHSCC, RHSCC;
3939 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3940 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3941 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3942 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3943 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3944 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3945 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3946 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3947 // We can't fold (ugt x, C) | (sgt x, C2).
3948 PredicatesFoldable(LHSCC, RHSCC)) {
3949 // Ensure that the larger constant is on the RHS.
3950 ICmpInst *LHS = cast<ICmpInst>(Op0);
3952 if (ICmpInst::isSignedPredicate(LHSCC))
3953 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3955 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3958 std::swap(LHS, RHS);
3959 std::swap(LHSCst, RHSCst);
3960 std::swap(LHSCC, RHSCC);
3963 // At this point, we know we have have two icmp instructions
3964 // comparing a value against two constants and or'ing the result
3965 // together. Because of the above check, we know that we only have
3966 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3967 // FoldICmpLogical check above), that the two constants are not
3969 assert(LHSCst != RHSCst && "Compares not folded above?");
3972 default: assert(0 && "Unknown integer condition code!");
3973 case ICmpInst::ICMP_EQ:
3975 default: assert(0 && "Unknown integer condition code!");
3976 case ICmpInst::ICMP_EQ:
3977 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3978 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3979 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3980 LHSVal->getName()+".off");
3981 InsertNewInstBefore(Add, I);
3982 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3983 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3985 break; // (X == 13 | X == 15) -> no change
3986 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3987 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3989 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3990 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3991 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3992 return ReplaceInstUsesWith(I, RHS);
3995 case ICmpInst::ICMP_NE:
3997 default: assert(0 && "Unknown integer condition code!");
3998 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3999 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4000 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4001 return ReplaceInstUsesWith(I, LHS);
4002 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4003 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4004 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4005 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4008 case ICmpInst::ICMP_ULT:
4010 default: assert(0 && "Unknown integer condition code!");
4011 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4013 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4014 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4016 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4018 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4019 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4020 return ReplaceInstUsesWith(I, RHS);
4021 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4025 case ICmpInst::ICMP_SLT:
4027 default: assert(0 && "Unknown integer condition code!");
4028 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4030 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4031 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4033 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4035 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4036 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4037 return ReplaceInstUsesWith(I, RHS);
4038 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4042 case ICmpInst::ICMP_UGT:
4044 default: assert(0 && "Unknown integer condition code!");
4045 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4046 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4047 return ReplaceInstUsesWith(I, LHS);
4048 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4050 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4051 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4052 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4053 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4057 case ICmpInst::ICMP_SGT:
4059 default: assert(0 && "Unknown integer condition code!");
4060 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4061 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4062 return ReplaceInstUsesWith(I, LHS);
4063 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4065 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4066 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4067 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4068 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4076 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4077 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4078 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4079 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4080 const Type *SrcTy = Op0C->getOperand(0)->getType();
4081 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4082 // Only do this if the casts both really cause code to be generated.
4083 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4085 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4087 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4088 Op1C->getOperand(0),
4090 InsertNewInstBefore(NewOp, I);
4091 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4096 return Changed ? &I : 0;
4099 // XorSelf - Implements: X ^ X --> 0
4102 XorSelf(Value *rhs) : RHS(rhs) {}
4103 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4104 Instruction *apply(BinaryOperator &Xor) const {
4110 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4111 bool Changed = SimplifyCommutative(I);
4112 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4114 if (isa<UndefValue>(Op1))
4115 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4117 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4118 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4119 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4120 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4123 // See if we can simplify any instructions used by the instruction whose sole
4124 // purpose is to compute bits we don't care about.
4125 if (!isa<VectorType>(I.getType())) {
4126 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4127 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4128 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4129 KnownZero, KnownOne))
4131 } else if (isa<ConstantAggregateZero>(Op1)) {
4132 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4135 // Is this a ~ operation?
4136 if (Value *NotOp = dyn_castNotVal(&I)) {
4137 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4138 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4139 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4140 if (Op0I->getOpcode() == Instruction::And ||
4141 Op0I->getOpcode() == Instruction::Or) {
4142 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4143 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4145 BinaryOperator::createNot(Op0I->getOperand(1),
4146 Op0I->getOperand(1)->getName()+".not");
4147 InsertNewInstBefore(NotY, I);
4148 if (Op0I->getOpcode() == Instruction::And)
4149 return BinaryOperator::createOr(Op0NotVal, NotY);
4151 return BinaryOperator::createAnd(Op0NotVal, NotY);
4158 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4159 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4160 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4161 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4162 return new ICmpInst(ICI->getInversePredicate(),
4163 ICI->getOperand(0), ICI->getOperand(1));
4165 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4166 return new FCmpInst(FCI->getInversePredicate(),
4167 FCI->getOperand(0), FCI->getOperand(1));
4170 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4171 // ~(c-X) == X-c-1 == X+(-c-1)
4172 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4173 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4174 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4175 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4176 ConstantInt::get(I.getType(), 1));
4177 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4180 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4181 if (Op0I->getOpcode() == Instruction::Add) {
4182 // ~(X-c) --> (-c-1)-X
4183 if (RHS->isAllOnesValue()) {
4184 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4185 return BinaryOperator::createSub(
4186 ConstantExpr::getSub(NegOp0CI,
4187 ConstantInt::get(I.getType(), 1)),
4188 Op0I->getOperand(0));
4189 } else if (RHS->getValue().isSignBit()) {
4190 // (X + C) ^ signbit -> (X + C + signbit)
4191 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4192 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4195 } else if (Op0I->getOpcode() == Instruction::Or) {
4196 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4197 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4198 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4199 // Anything in both C1 and C2 is known to be zero, remove it from
4201 Constant *CommonBits = And(Op0CI, RHS);
4202 NewRHS = ConstantExpr::getAnd(NewRHS,
4203 ConstantExpr::getNot(CommonBits));
4204 AddToWorkList(Op0I);
4205 I.setOperand(0, Op0I->getOperand(0));
4206 I.setOperand(1, NewRHS);
4212 // Try to fold constant and into select arguments.
4213 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4214 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4216 if (isa<PHINode>(Op0))
4217 if (Instruction *NV = FoldOpIntoPhi(I))
4221 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4223 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4225 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4227 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4230 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4233 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4234 if (A == Op0) { // B^(B|A) == (A|B)^B
4235 Op1I->swapOperands();
4237 std::swap(Op0, Op1);
4238 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4239 I.swapOperands(); // Simplified below.
4240 std::swap(Op0, Op1);
4242 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4243 if (Op0 == A) // A^(A^B) == B
4244 return ReplaceInstUsesWith(I, B);
4245 else if (Op0 == B) // A^(B^A) == B
4246 return ReplaceInstUsesWith(I, A);
4247 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4248 if (A == Op0) { // A^(A&B) -> A^(B&A)
4249 Op1I->swapOperands();
4252 if (B == Op0) { // A^(B&A) -> (B&A)^A
4253 I.swapOperands(); // Simplified below.
4254 std::swap(Op0, Op1);
4259 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4262 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4263 if (A == Op1) // (B|A)^B == (A|B)^B
4265 if (B == Op1) { // (A|B)^B == A & ~B
4267 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4268 return BinaryOperator::createAnd(A, NotB);
4270 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4271 if (Op1 == A) // (A^B)^A == B
4272 return ReplaceInstUsesWith(I, B);
4273 else if (Op1 == B) // (B^A)^A == B
4274 return ReplaceInstUsesWith(I, A);
4275 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4276 if (A == Op1) // (A&B)^A -> (B&A)^A
4278 if (B == Op1 && // (B&A)^A == ~B & A
4279 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4281 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4282 return BinaryOperator::createAnd(N, Op1);
4287 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4288 if (Op0I && Op1I && Op0I->isShift() &&
4289 Op0I->getOpcode() == Op1I->getOpcode() &&
4290 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4291 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4292 Instruction *NewOp =
4293 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4294 Op1I->getOperand(0),
4295 Op0I->getName()), I);
4296 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4297 Op1I->getOperand(1));
4301 Value *A, *B, *C, *D;
4302 // (A & B)^(A | B) -> A ^ B
4303 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4304 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4305 if ((A == C && B == D) || (A == D && B == C))
4306 return BinaryOperator::createXor(A, B);
4308 // (A | B)^(A & B) -> A ^ B
4309 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4310 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4311 if ((A == C && B == D) || (A == D && B == C))
4312 return BinaryOperator::createXor(A, B);
4316 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4317 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4318 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4319 // (X & Y)^(X & Y) -> (Y^Z) & X
4320 Value *X = 0, *Y = 0, *Z = 0;
4322 X = A, Y = B, Z = D;
4324 X = A, Y = B, Z = C;
4326 X = B, Y = A, Z = D;
4328 X = B, Y = A, Z = C;
4331 Instruction *NewOp =
4332 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4333 return BinaryOperator::createAnd(NewOp, X);
4338 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4339 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4340 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4343 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4344 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4345 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4346 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4347 const Type *SrcTy = Op0C->getOperand(0)->getType();
4348 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4349 // Only do this if the casts both really cause code to be generated.
4350 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4352 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4354 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4355 Op1C->getOperand(0),
4357 InsertNewInstBefore(NewOp, I);
4358 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4362 return Changed ? &I : 0;
4365 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4366 /// overflowed for this type.
4367 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4368 ConstantInt *In2, bool IsSigned = false) {
4369 Result = cast<ConstantInt>(Add(In1, In2));
4372 if (In2->getValue().isNegative())
4373 return Result->getValue().sgt(In1->getValue());
4375 return Result->getValue().slt(In1->getValue());
4377 return Result->getValue().ult(In1->getValue());
4380 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4381 /// code necessary to compute the offset from the base pointer (without adding
4382 /// in the base pointer). Return the result as a signed integer of intptr size.
4383 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4384 TargetData &TD = IC.getTargetData();
4385 gep_type_iterator GTI = gep_type_begin(GEP);
4386 const Type *IntPtrTy = TD.getIntPtrType();
4387 Value *Result = Constant::getNullValue(IntPtrTy);
4389 // Build a mask for high order bits.
4390 unsigned IntPtrWidth = TD.getPointerSize()*8;
4391 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4393 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4394 Value *Op = GEP->getOperand(i);
4395 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4396 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4397 if (OpC->isZero()) continue;
4399 // Handle a struct index, which adds its field offset to the pointer.
4400 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4401 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4403 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4404 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4406 Result = IC.InsertNewInstBefore(
4407 BinaryOperator::createAdd(Result,
4408 ConstantInt::get(IntPtrTy, Size),
4409 GEP->getName()+".offs"), I);
4413 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4414 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4415 Scale = ConstantExpr::getMul(OC, Scale);
4416 if (Constant *RC = dyn_cast<Constant>(Result))
4417 Result = ConstantExpr::getAdd(RC, Scale);
4419 // Emit an add instruction.
4420 Result = IC.InsertNewInstBefore(
4421 BinaryOperator::createAdd(Result, Scale,
4422 GEP->getName()+".offs"), I);
4426 // Convert to correct type.
4427 if (Op->getType() != IntPtrTy) {
4428 if (Constant *OpC = dyn_cast<Constant>(Op))
4429 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4431 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4432 Op->getName()+".c"), I);
4435 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4436 if (Constant *OpC = dyn_cast<Constant>(Op))
4437 Op = ConstantExpr::getMul(OpC, Scale);
4438 else // We'll let instcombine(mul) convert this to a shl if possible.
4439 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4440 GEP->getName()+".idx"), I);
4443 // Emit an add instruction.
4444 if (isa<Constant>(Op) && isa<Constant>(Result))
4445 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4446 cast<Constant>(Result));
4448 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4449 GEP->getName()+".offs"), I);
4454 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4455 /// else. At this point we know that the GEP is on the LHS of the comparison.
4456 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4457 ICmpInst::Predicate Cond,
4459 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4461 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4462 if (isa<PointerType>(CI->getOperand(0)->getType()))
4463 RHS = CI->getOperand(0);
4465 Value *PtrBase = GEPLHS->getOperand(0);
4466 if (PtrBase == RHS) {
4467 // As an optimization, we don't actually have to compute the actual value of
4468 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4469 // each index is zero or not.
4470 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4471 Instruction *InVal = 0;
4472 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4473 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4475 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4476 if (isa<UndefValue>(C)) // undef index -> undef.
4477 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4478 if (C->isNullValue())
4480 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4481 EmitIt = false; // This is indexing into a zero sized array?
4482 } else if (isa<ConstantInt>(C))
4483 return ReplaceInstUsesWith(I, // No comparison is needed here.
4484 ConstantInt::get(Type::Int1Ty,
4485 Cond == ICmpInst::ICMP_NE));
4490 new ICmpInst(Cond, GEPLHS->getOperand(i),
4491 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4495 InVal = InsertNewInstBefore(InVal, I);
4496 InsertNewInstBefore(Comp, I);
4497 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4498 InVal = BinaryOperator::createOr(InVal, Comp);
4499 else // True if all are equal
4500 InVal = BinaryOperator::createAnd(InVal, Comp);
4508 // No comparison is needed here, all indexes = 0
4509 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4510 Cond == ICmpInst::ICMP_EQ));
4513 // Only lower this if the icmp is the only user of the GEP or if we expect
4514 // the result to fold to a constant!
4515 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4516 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4517 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4518 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4519 Constant::getNullValue(Offset->getType()));
4521 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4522 // If the base pointers are different, but the indices are the same, just
4523 // compare the base pointer.
4524 if (PtrBase != GEPRHS->getOperand(0)) {
4525 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4526 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4527 GEPRHS->getOperand(0)->getType();
4529 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4530 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4531 IndicesTheSame = false;
4535 // If all indices are the same, just compare the base pointers.
4537 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4538 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4540 // Otherwise, the base pointers are different and the indices are
4541 // different, bail out.
4545 // If one of the GEPs has all zero indices, recurse.
4546 bool AllZeros = true;
4547 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4548 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4549 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4554 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4555 ICmpInst::getSwappedPredicate(Cond), I);
4557 // If the other GEP has all zero indices, recurse.
4559 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4560 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4561 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4566 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4568 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4569 // If the GEPs only differ by one index, compare it.
4570 unsigned NumDifferences = 0; // Keep track of # differences.
4571 unsigned DiffOperand = 0; // The operand that differs.
4572 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4573 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4574 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4575 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4576 // Irreconcilable differences.
4580 if (NumDifferences++) break;
4585 if (NumDifferences == 0) // SAME GEP?
4586 return ReplaceInstUsesWith(I, // No comparison is needed here.
4587 ConstantInt::get(Type::Int1Ty,
4588 isTrueWhenEqual(Cond)));
4590 else if (NumDifferences == 1) {
4591 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4592 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4593 // Make sure we do a signed comparison here.
4594 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4598 // Only lower this if the icmp is the only user of the GEP or if we expect
4599 // the result to fold to a constant!
4600 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4601 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4602 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4603 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4604 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4605 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4611 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4612 bool Changed = SimplifyCompare(I);
4613 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4615 // Fold trivial predicates.
4616 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4617 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4618 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4619 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4621 // Simplify 'fcmp pred X, X'
4623 switch (I.getPredicate()) {
4624 default: assert(0 && "Unknown predicate!");
4625 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4626 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4627 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4628 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4629 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4630 case FCmpInst::FCMP_OLT: // True if ordered and less than
4631 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4632 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4634 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4635 case FCmpInst::FCMP_ULT: // True if unordered or less than
4636 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4637 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4638 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4639 I.setPredicate(FCmpInst::FCMP_UNO);
4640 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4643 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4644 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4645 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4646 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4647 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4648 I.setPredicate(FCmpInst::FCMP_ORD);
4649 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4654 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4655 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4657 // Handle fcmp with constant RHS
4658 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4659 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4660 switch (LHSI->getOpcode()) {
4661 case Instruction::PHI:
4662 if (Instruction *NV = FoldOpIntoPhi(I))
4665 case Instruction::Select:
4666 // If either operand of the select is a constant, we can fold the
4667 // comparison into the select arms, which will cause one to be
4668 // constant folded and the select turned into a bitwise or.
4669 Value *Op1 = 0, *Op2 = 0;
4670 if (LHSI->hasOneUse()) {
4671 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4672 // Fold the known value into the constant operand.
4673 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4674 // Insert a new FCmp of the other select operand.
4675 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4676 LHSI->getOperand(2), RHSC,
4678 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4679 // Fold the known value into the constant operand.
4680 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4681 // Insert a new FCmp of the other select operand.
4682 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4683 LHSI->getOperand(1), RHSC,
4689 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4694 return Changed ? &I : 0;
4697 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4698 bool Changed = SimplifyCompare(I);
4699 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4700 const Type *Ty = Op0->getType();
4704 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4705 isTrueWhenEqual(I)));
4707 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4708 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4710 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4711 // addresses never equal each other! We already know that Op0 != Op1.
4712 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4713 isa<ConstantPointerNull>(Op0)) &&
4714 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4715 isa<ConstantPointerNull>(Op1)))
4716 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4717 !isTrueWhenEqual(I)));
4719 // icmp's with boolean values can always be turned into bitwise operations
4720 if (Ty == Type::Int1Ty) {
4721 switch (I.getPredicate()) {
4722 default: assert(0 && "Invalid icmp instruction!");
4723 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4724 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4725 InsertNewInstBefore(Xor, I);
4726 return BinaryOperator::createNot(Xor);
4728 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4729 return BinaryOperator::createXor(Op0, Op1);
4731 case ICmpInst::ICMP_UGT:
4732 case ICmpInst::ICMP_SGT:
4733 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4735 case ICmpInst::ICMP_ULT:
4736 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4737 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4738 InsertNewInstBefore(Not, I);
4739 return BinaryOperator::createAnd(Not, Op1);
4741 case ICmpInst::ICMP_UGE:
4742 case ICmpInst::ICMP_SGE:
4743 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4745 case ICmpInst::ICMP_ULE:
4746 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4747 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4748 InsertNewInstBefore(Not, I);
4749 return BinaryOperator::createOr(Not, Op1);
4754 // See if we are doing a comparison between a constant and an instruction that
4755 // can be folded into the comparison.
4756 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4757 switch (I.getPredicate()) {
4759 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4760 if (CI->isMinValue(false))
4761 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4762 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4763 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4764 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4765 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4766 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4767 if (CI->isMinValue(true))
4768 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4769 ConstantInt::getAllOnesValue(Op0->getType()));
4773 case ICmpInst::ICMP_SLT:
4774 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4775 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4776 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4777 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4778 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4779 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4782 case ICmpInst::ICMP_UGT:
4783 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4784 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4785 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4786 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4787 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4788 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4790 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4791 if (CI->isMaxValue(true))
4792 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4793 ConstantInt::getNullValue(Op0->getType()));
4796 case ICmpInst::ICMP_SGT:
4797 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4798 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4799 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4800 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4801 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4802 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4805 case ICmpInst::ICMP_ULE:
4806 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4807 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4808 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4809 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4810 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4811 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4814 case ICmpInst::ICMP_SLE:
4815 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4816 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4817 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4818 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4819 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4820 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4823 case ICmpInst::ICMP_UGE:
4824 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4825 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4826 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4827 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4828 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4829 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4832 case ICmpInst::ICMP_SGE:
4833 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4834 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4835 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4836 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4837 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4838 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4842 // If we still have a icmp le or icmp ge instruction, turn it into the
4843 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4844 // already been handled above, this requires little checking.
4846 switch (I.getPredicate()) {
4848 case ICmpInst::ICMP_ULE:
4849 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4850 case ICmpInst::ICMP_SLE:
4851 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4852 case ICmpInst::ICMP_UGE:
4853 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4854 case ICmpInst::ICMP_SGE:
4855 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4858 // See if we can fold the comparison based on bits known to be zero or one
4859 // in the input. If this comparison is a normal comparison, it demands all
4860 // bits, if it is a sign bit comparison, it only demands the sign bit.
4863 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4865 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4866 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4867 if (SimplifyDemandedBits(Op0,
4868 isSignBit ? APInt::getSignBit(BitWidth)
4869 : APInt::getAllOnesValue(BitWidth),
4870 KnownZero, KnownOne, 0))
4873 // Given the known and unknown bits, compute a range that the LHS could be
4875 if ((KnownOne | KnownZero) != 0) {
4876 // Compute the Min, Max and RHS values based on the known bits. For the
4877 // EQ and NE we use unsigned values.
4878 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4879 const APInt& RHSVal = CI->getValue();
4880 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4881 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4884 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4887 switch (I.getPredicate()) { // LE/GE have been folded already.
4888 default: assert(0 && "Unknown icmp opcode!");
4889 case ICmpInst::ICMP_EQ:
4890 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4891 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4893 case ICmpInst::ICMP_NE:
4894 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4895 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4897 case ICmpInst::ICMP_ULT:
4898 if (Max.ult(RHSVal))
4899 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4900 if (Min.uge(RHSVal))
4901 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4903 case ICmpInst::ICMP_UGT:
4904 if (Min.ugt(RHSVal))
4905 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4906 if (Max.ule(RHSVal))
4907 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4909 case ICmpInst::ICMP_SLT:
4910 if (Max.slt(RHSVal))
4911 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4912 if (Min.sgt(RHSVal))
4913 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4915 case ICmpInst::ICMP_SGT:
4916 if (Min.sgt(RHSVal))
4917 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4918 if (Max.sle(RHSVal))
4919 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4924 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4925 // instruction, see if that instruction also has constants so that the
4926 // instruction can be folded into the icmp
4927 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4928 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4932 // Handle icmp with constant (but not simple integer constant) RHS
4933 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4934 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4935 switch (LHSI->getOpcode()) {
4936 case Instruction::GetElementPtr:
4937 if (RHSC->isNullValue()) {
4938 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4939 bool isAllZeros = true;
4940 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4941 if (!isa<Constant>(LHSI->getOperand(i)) ||
4942 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4947 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4948 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4952 case Instruction::PHI:
4953 if (Instruction *NV = FoldOpIntoPhi(I))
4956 case Instruction::Select: {
4957 // If either operand of the select is a constant, we can fold the
4958 // comparison into the select arms, which will cause one to be
4959 // constant folded and the select turned into a bitwise or.
4960 Value *Op1 = 0, *Op2 = 0;
4961 if (LHSI->hasOneUse()) {
4962 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4963 // Fold the known value into the constant operand.
4964 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4965 // Insert a new ICmp of the other select operand.
4966 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4967 LHSI->getOperand(2), RHSC,
4969 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4970 // Fold the known value into the constant operand.
4971 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4972 // Insert a new ICmp of the other select operand.
4973 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4974 LHSI->getOperand(1), RHSC,
4980 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4983 case Instruction::Malloc:
4984 // If we have (malloc != null), and if the malloc has a single use, we
4985 // can assume it is successful and remove the malloc.
4986 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4987 AddToWorkList(LHSI);
4988 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4989 !isTrueWhenEqual(I)));
4995 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4996 if (User *GEP = dyn_castGetElementPtr(Op0))
4997 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4999 if (User *GEP = dyn_castGetElementPtr(Op1))
5000 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5001 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5004 // Test to see if the operands of the icmp are casted versions of other
5005 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5007 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5008 if (isa<PointerType>(Op0->getType()) &&
5009 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5010 // We keep moving the cast from the left operand over to the right
5011 // operand, where it can often be eliminated completely.
5012 Op0 = CI->getOperand(0);
5014 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5015 // so eliminate it as well.
5016 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5017 Op1 = CI2->getOperand(0);
5019 // If Op1 is a constant, we can fold the cast into the constant.
5020 if (Op0->getType() != Op1->getType())
5021 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5022 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5024 // Otherwise, cast the RHS right before the icmp
5025 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5027 return new ICmpInst(I.getPredicate(), Op0, Op1);
5031 if (isa<CastInst>(Op0)) {
5032 // Handle the special case of: icmp (cast bool to X), <cst>
5033 // This comes up when you have code like
5036 // For generality, we handle any zero-extension of any operand comparison
5037 // with a constant or another cast from the same type.
5038 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5039 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5043 if (I.isEquality()) {
5044 Value *A, *B, *C, *D;
5045 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5046 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5047 Value *OtherVal = A == Op1 ? B : A;
5048 return new ICmpInst(I.getPredicate(), OtherVal,
5049 Constant::getNullValue(A->getType()));
5052 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5053 // A^c1 == C^c2 --> A == C^(c1^c2)
5054 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5055 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5056 if (Op1->hasOneUse()) {
5057 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5058 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5059 return new ICmpInst(I.getPredicate(), A,
5060 InsertNewInstBefore(Xor, I));
5063 // A^B == A^D -> B == D
5064 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5065 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5066 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5067 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5071 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5072 (A == Op0 || B == Op0)) {
5073 // A == (A^B) -> B == 0
5074 Value *OtherVal = A == Op0 ? B : A;
5075 return new ICmpInst(I.getPredicate(), OtherVal,
5076 Constant::getNullValue(A->getType()));
5078 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5079 // (A-B) == A -> B == 0
5080 return new ICmpInst(I.getPredicate(), B,
5081 Constant::getNullValue(B->getType()));
5083 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5084 // A == (A-B) -> B == 0
5085 return new ICmpInst(I.getPredicate(), B,
5086 Constant::getNullValue(B->getType()));
5089 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5090 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5091 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5092 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5093 Value *X = 0, *Y = 0, *Z = 0;
5096 X = B; Y = D; Z = A;
5097 } else if (A == D) {
5098 X = B; Y = C; Z = A;
5099 } else if (B == C) {
5100 X = A; Y = D; Z = B;
5101 } else if (B == D) {
5102 X = A; Y = C; Z = B;
5105 if (X) { // Build (X^Y) & Z
5106 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5107 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5108 I.setOperand(0, Op1);
5109 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5114 return Changed ? &I : 0;
5118 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5119 /// and CmpRHS are both known to be integer constants.
5120 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5121 ConstantInt *DivRHS) {
5122 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5123 const APInt &CmpRHSV = CmpRHS->getValue();
5125 // FIXME: If the operand types don't match the type of the divide
5126 // then don't attempt this transform. The code below doesn't have the
5127 // logic to deal with a signed divide and an unsigned compare (and
5128 // vice versa). This is because (x /s C1) <s C2 produces different
5129 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5130 // (x /u C1) <u C2. Simply casting the operands and result won't
5131 // work. :( The if statement below tests that condition and bails
5133 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5134 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5136 if (DivRHS->isZero())
5137 return 0; // The ProdOV computation fails on divide by zero.
5139 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5140 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5141 // C2 (CI). By solving for X we can turn this into a range check
5142 // instead of computing a divide.
5143 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5145 // Determine if the product overflows by seeing if the product is
5146 // not equal to the divide. Make sure we do the same kind of divide
5147 // as in the LHS instruction that we're folding.
5148 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5149 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5151 // Get the ICmp opcode
5152 ICmpInst::Predicate Pred = ICI.getPredicate();
5154 // Figure out the interval that is being checked. For example, a comparison
5155 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5156 // Compute this interval based on the constants involved and the signedness of
5157 // the compare/divide. This computes a half-open interval, keeping track of
5158 // whether either value in the interval overflows. After analysis each
5159 // overflow variable is set to 0 if it's corresponding bound variable is valid
5160 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5161 int LoOverflow = 0, HiOverflow = 0;
5162 ConstantInt *LoBound = 0, *HiBound = 0;
5165 if (!DivIsSigned) { // udiv
5166 // e.g. X/5 op 3 --> [15, 20)
5168 HiOverflow = LoOverflow = ProdOV;
5170 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5171 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5172 if (CmpRHSV == 0) { // (X / pos) op 0
5173 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5174 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5176 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5177 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5178 HiOverflow = LoOverflow = ProdOV;
5180 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5181 } else { // (X / pos) op neg
5182 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5183 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5184 LoOverflow = AddWithOverflow(LoBound, Prod,
5185 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5186 HiBound = AddOne(Prod);
5187 HiOverflow = ProdOV ? -1 : 0;
5189 } else { // Divisor is < 0.
5190 if (CmpRHSV == 0) { // (X / neg) op 0
5191 // e.g. X/-5 op 0 --> [-4, 5)
5192 LoBound = AddOne(DivRHS);
5193 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5194 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5195 HiOverflow = 1; // [INTMIN+1, overflow)
5196 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5198 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5199 // e.g. X/-5 op 3 --> [-19, -14)
5200 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5202 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5203 HiBound = AddOne(Prod);
5204 } else { // (X / neg) op neg
5205 // e.g. X/-5 op -3 --> [15, 20)
5207 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5208 HiBound = Subtract(Prod, DivRHS);
5211 // Dividing by a negative swaps the condition. LT <-> GT
5212 Pred = ICmpInst::getSwappedPredicate(Pred);
5215 Value *X = DivI->getOperand(0);
5217 default: assert(0 && "Unhandled icmp opcode!");
5218 case ICmpInst::ICMP_EQ:
5219 if (LoOverflow && HiOverflow)
5220 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5221 else if (HiOverflow)
5222 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5223 ICmpInst::ICMP_UGE, X, LoBound);
5224 else if (LoOverflow)
5225 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5226 ICmpInst::ICMP_ULT, X, HiBound);
5228 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5229 case ICmpInst::ICMP_NE:
5230 if (LoOverflow && HiOverflow)
5231 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5232 else if (HiOverflow)
5233 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5234 ICmpInst::ICMP_ULT, X, LoBound);
5235 else if (LoOverflow)
5236 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5237 ICmpInst::ICMP_UGE, X, HiBound);
5239 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5240 case ICmpInst::ICMP_ULT:
5241 case ICmpInst::ICMP_SLT:
5242 if (LoOverflow == +1) // Low bound is greater than input range.
5243 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5244 if (LoOverflow == -1) // Low bound is less than input range.
5245 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5246 return new ICmpInst(Pred, X, LoBound);
5247 case ICmpInst::ICMP_UGT:
5248 case ICmpInst::ICMP_SGT:
5249 if (HiOverflow == +1) // High bound greater than input range.
5250 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5251 else if (HiOverflow == -1) // High bound less than input range.
5252 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5253 if (Pred == ICmpInst::ICMP_UGT)
5254 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5256 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5261 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5263 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5266 const APInt &RHSV = RHS->getValue();
5268 switch (LHSI->getOpcode()) {
5269 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5270 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5271 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5273 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5274 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5275 Value *CompareVal = LHSI->getOperand(0);
5277 // If the sign bit of the XorCST is not set, there is no change to
5278 // the operation, just stop using the Xor.
5279 if (!XorCST->getValue().isNegative()) {
5280 ICI.setOperand(0, CompareVal);
5281 AddToWorkList(LHSI);
5285 // Was the old condition true if the operand is positive?
5286 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5288 // If so, the new one isn't.
5289 isTrueIfPositive ^= true;
5291 if (isTrueIfPositive)
5292 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5294 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5298 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5299 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5300 LHSI->getOperand(0)->hasOneUse()) {
5301 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5303 // If the LHS is an AND of a truncating cast, we can widen the
5304 // and/compare to be the input width without changing the value
5305 // produced, eliminating a cast.
5306 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5307 // We can do this transformation if either the AND constant does not
5308 // have its sign bit set or if it is an equality comparison.
5309 // Extending a relational comparison when we're checking the sign
5310 // bit would not work.
5311 if (Cast->hasOneUse() &&
5312 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5313 RHSV.isPositive())) {
5315 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5316 APInt NewCST = AndCST->getValue();
5317 NewCST.zext(BitWidth);
5319 NewCI.zext(BitWidth);
5320 Instruction *NewAnd =
5321 BinaryOperator::createAnd(Cast->getOperand(0),
5322 ConstantInt::get(NewCST),LHSI->getName());
5323 InsertNewInstBefore(NewAnd, ICI);
5324 return new ICmpInst(ICI.getPredicate(), NewAnd,
5325 ConstantInt::get(NewCI));
5329 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5330 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5331 // happens a LOT in code produced by the C front-end, for bitfield
5333 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5334 if (Shift && !Shift->isShift())
5338 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5339 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5340 const Type *AndTy = AndCST->getType(); // Type of the and.
5342 // We can fold this as long as we can't shift unknown bits
5343 // into the mask. This can only happen with signed shift
5344 // rights, as they sign-extend.
5346 bool CanFold = Shift->isLogicalShift();
5348 // To test for the bad case of the signed shr, see if any
5349 // of the bits shifted in could be tested after the mask.
5350 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5351 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5353 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5354 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5355 AndCST->getValue()) == 0)
5361 if (Shift->getOpcode() == Instruction::Shl)
5362 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5364 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5366 // Check to see if we are shifting out any of the bits being
5368 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5369 // If we shifted bits out, the fold is not going to work out.
5370 // As a special case, check to see if this means that the
5371 // result is always true or false now.
5372 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5373 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5374 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5375 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5377 ICI.setOperand(1, NewCst);
5378 Constant *NewAndCST;
5379 if (Shift->getOpcode() == Instruction::Shl)
5380 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5382 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5383 LHSI->setOperand(1, NewAndCST);
5384 LHSI->setOperand(0, Shift->getOperand(0));
5385 AddToWorkList(Shift); // Shift is dead.
5386 AddUsesToWorkList(ICI);
5392 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5393 // preferable because it allows the C<<Y expression to be hoisted out
5394 // of a loop if Y is invariant and X is not.
5395 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5396 ICI.isEquality() && !Shift->isArithmeticShift() &&
5397 isa<Instruction>(Shift->getOperand(0))) {
5400 if (Shift->getOpcode() == Instruction::LShr) {
5401 NS = BinaryOperator::createShl(AndCST,
5402 Shift->getOperand(1), "tmp");
5404 // Insert a logical shift.
5405 NS = BinaryOperator::createLShr(AndCST,
5406 Shift->getOperand(1), "tmp");
5408 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5410 // Compute X & (C << Y).
5411 Instruction *NewAnd =
5412 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5413 InsertNewInstBefore(NewAnd, ICI);
5415 ICI.setOperand(0, NewAnd);
5421 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5422 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5425 uint32_t TypeBits = RHSV.getBitWidth();
5427 // Check that the shift amount is in range. If not, don't perform
5428 // undefined shifts. When the shift is visited it will be
5430 if (ShAmt->uge(TypeBits))
5433 if (ICI.isEquality()) {
5434 // If we are comparing against bits always shifted out, the
5435 // comparison cannot succeed.
5437 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5438 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5439 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5440 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5441 return ReplaceInstUsesWith(ICI, Cst);
5444 if (LHSI->hasOneUse()) {
5445 // Otherwise strength reduce the shift into an and.
5446 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5448 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5451 BinaryOperator::createAnd(LHSI->getOperand(0),
5452 Mask, LHSI->getName()+".mask");
5453 Value *And = InsertNewInstBefore(AndI, ICI);
5454 return new ICmpInst(ICI.getPredicate(), And,
5455 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5459 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5460 bool TrueIfSigned = false;
5461 if (LHSI->hasOneUse() &&
5462 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5463 // (X << 31) <s 0 --> (X&1) != 0
5464 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5465 (TypeBits-ShAmt->getZExtValue()-1));
5467 BinaryOperator::createAnd(LHSI->getOperand(0),
5468 Mask, LHSI->getName()+".mask");
5469 Value *And = InsertNewInstBefore(AndI, ICI);
5471 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5472 And, Constant::getNullValue(And->getType()));
5477 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5478 case Instruction::AShr: {
5479 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5482 if (ICI.isEquality()) {
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 uint32_t TypeBits = RHSV.getBitWidth();
5487 if (ShAmt->uge(TypeBits))
5489 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5491 // If we are comparing against bits always shifted out, the
5492 // comparison cannot succeed.
5493 APInt Comp = RHSV << ShAmtVal;
5494 if (LHSI->getOpcode() == Instruction::LShr)
5495 Comp = Comp.lshr(ShAmtVal);
5497 Comp = Comp.ashr(ShAmtVal);
5499 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5500 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5501 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5502 return ReplaceInstUsesWith(ICI, Cst);
5505 if (LHSI->hasOneUse() || RHSV == 0) {
5506 // Otherwise strength reduce the shift into an and.
5507 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5508 Constant *Mask = ConstantInt::get(Val);
5511 BinaryOperator::createAnd(LHSI->getOperand(0),
5512 Mask, LHSI->getName()+".mask");
5513 Value *And = InsertNewInstBefore(AndI, ICI);
5514 return new ICmpInst(ICI.getPredicate(), And,
5515 ConstantExpr::getShl(RHS, ShAmt));
5521 case Instruction::SDiv:
5522 case Instruction::UDiv:
5523 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5524 // Fold this div into the comparison, producing a range check.
5525 // Determine, based on the divide type, what the range is being
5526 // checked. If there is an overflow on the low or high side, remember
5527 // it, otherwise compute the range [low, hi) bounding the new value.
5528 // See: InsertRangeTest above for the kinds of replacements possible.
5529 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5530 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5536 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5537 if (ICI.isEquality()) {
5538 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5540 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5541 // the second operand is a constant, simplify a bit.
5542 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5543 switch (BO->getOpcode()) {
5544 case Instruction::SRem:
5545 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5546 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5547 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5548 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5549 Instruction *NewRem =
5550 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5552 InsertNewInstBefore(NewRem, ICI);
5553 return new ICmpInst(ICI.getPredicate(), NewRem,
5554 Constant::getNullValue(BO->getType()));
5558 case Instruction::Add:
5559 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5560 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5561 if (BO->hasOneUse())
5562 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5563 Subtract(RHS, BOp1C));
5564 } else if (RHSV == 0) {
5565 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5566 // efficiently invertible, or if the add has just this one use.
5567 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5569 if (Value *NegVal = dyn_castNegVal(BOp1))
5570 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5571 else if (Value *NegVal = dyn_castNegVal(BOp0))
5572 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5573 else if (BO->hasOneUse()) {
5574 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5575 InsertNewInstBefore(Neg, ICI);
5577 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5581 case Instruction::Xor:
5582 // For the xor case, we can xor two constants together, eliminating
5583 // the explicit xor.
5584 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5585 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5586 ConstantExpr::getXor(RHS, BOC));
5589 case Instruction::Sub:
5590 // Replace (([sub|xor] A, B) != 0) with (A != B)
5592 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5596 case Instruction::Or:
5597 // If bits are being or'd in that are not present in the constant we
5598 // are comparing against, then the comparison could never succeed!
5599 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5600 Constant *NotCI = ConstantExpr::getNot(RHS);
5601 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5602 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5607 case Instruction::And:
5608 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5609 // If bits are being compared against that are and'd out, then the
5610 // comparison can never succeed!
5611 if ((RHSV & ~BOC->getValue()) != 0)
5612 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5615 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5616 if (RHS == BOC && RHSV.isPowerOf2())
5617 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5618 ICmpInst::ICMP_NE, LHSI,
5619 Constant::getNullValue(RHS->getType()));
5621 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5622 if (isSignBit(BOC)) {
5623 Value *X = BO->getOperand(0);
5624 Constant *Zero = Constant::getNullValue(X->getType());
5625 ICmpInst::Predicate pred = isICMP_NE ?
5626 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5627 return new ICmpInst(pred, X, Zero);
5630 // ((X & ~7) == 0) --> X < 8
5631 if (RHSV == 0 && isHighOnes(BOC)) {
5632 Value *X = BO->getOperand(0);
5633 Constant *NegX = ConstantExpr::getNeg(BOC);
5634 ICmpInst::Predicate pred = isICMP_NE ?
5635 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5636 return new ICmpInst(pred, X, NegX);
5641 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5642 // Handle icmp {eq|ne} <intrinsic>, intcst.
5643 if (II->getIntrinsicID() == Intrinsic::bswap) {
5645 ICI.setOperand(0, II->getOperand(1));
5646 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5650 } else { // Not a ICMP_EQ/ICMP_NE
5651 // If the LHS is a cast from an integral value of the same size,
5652 // then since we know the RHS is a constant, try to simlify.
5653 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5654 Value *CastOp = Cast->getOperand(0);
5655 const Type *SrcTy = CastOp->getType();
5656 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5657 if (SrcTy->isInteger() &&
5658 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5659 // If this is an unsigned comparison, try to make the comparison use
5660 // smaller constant values.
5661 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5662 // X u< 128 => X s> -1
5663 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5664 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5665 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5666 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5667 // X u> 127 => X s< 0
5668 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5669 Constant::getNullValue(SrcTy));
5677 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5678 /// We only handle extending casts so far.
5680 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5681 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5682 Value *LHSCIOp = LHSCI->getOperand(0);
5683 const Type *SrcTy = LHSCIOp->getType();
5684 const Type *DestTy = LHSCI->getType();
5687 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5688 // integer type is the same size as the pointer type.
5689 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5690 getTargetData().getPointerSizeInBits() ==
5691 cast<IntegerType>(DestTy)->getBitWidth()) {
5693 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5694 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5695 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5696 RHSOp = RHSC->getOperand(0);
5697 // If the pointer types don't match, insert a bitcast.
5698 if (LHSCIOp->getType() != RHSOp->getType())
5699 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5700 LHSCIOp->getType(), ICI);
5704 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5707 // The code below only handles extension cast instructions, so far.
5709 if (LHSCI->getOpcode() != Instruction::ZExt &&
5710 LHSCI->getOpcode() != Instruction::SExt)
5713 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5714 bool isSignedCmp = ICI.isSignedPredicate();
5716 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5717 // Not an extension from the same type?
5718 RHSCIOp = CI->getOperand(0);
5719 if (RHSCIOp->getType() != LHSCIOp->getType())
5722 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5723 // and the other is a zext), then we can't handle this.
5724 if (CI->getOpcode() != LHSCI->getOpcode())
5727 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5728 // then we can't handle this.
5729 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5732 // Okay, just insert a compare of the reduced operands now!
5733 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5736 // If we aren't dealing with a constant on the RHS, exit early
5737 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5741 // Compute the constant that would happen if we truncated to SrcTy then
5742 // reextended to DestTy.
5743 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5744 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5746 // If the re-extended constant didn't change...
5748 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5749 // For example, we might have:
5750 // %A = sext short %X to uint
5751 // %B = icmp ugt uint %A, 1330
5752 // It is incorrect to transform this into
5753 // %B = icmp ugt short %X, 1330
5754 // because %A may have negative value.
5756 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5757 // OR operation is EQ/NE.
5758 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5759 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5764 // The re-extended constant changed so the constant cannot be represented
5765 // in the shorter type. Consequently, we cannot emit a simple comparison.
5767 // First, handle some easy cases. We know the result cannot be equal at this
5768 // point so handle the ICI.isEquality() cases
5769 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5770 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5771 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5772 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5774 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5775 // should have been folded away previously and not enter in here.
5778 // We're performing a signed comparison.
5779 if (cast<ConstantInt>(CI)->getValue().isNegative())
5780 Result = ConstantInt::getFalse(); // X < (small) --> false
5782 Result = ConstantInt::getTrue(); // X < (large) --> true
5784 // We're performing an unsigned comparison.
5786 // We're performing an unsigned comp with a sign extended value.
5787 // This is true if the input is >= 0. [aka >s -1]
5788 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5789 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5790 NegOne, ICI.getName()), ICI);
5792 // Unsigned extend & unsigned compare -> always true.
5793 Result = ConstantInt::getTrue();
5797 // Finally, return the value computed.
5798 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5799 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5800 return ReplaceInstUsesWith(ICI, Result);
5802 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5803 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5804 "ICmp should be folded!");
5805 if (Constant *CI = dyn_cast<Constant>(Result))
5806 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5808 return BinaryOperator::createNot(Result);
5812 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5813 return commonShiftTransforms(I);
5816 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5817 return commonShiftTransforms(I);
5820 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5821 return commonShiftTransforms(I);
5824 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5825 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5826 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5828 // shl X, 0 == X and shr X, 0 == X
5829 // shl 0, X == 0 and shr 0, X == 0
5830 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5831 Op0 == Constant::getNullValue(Op0->getType()))
5832 return ReplaceInstUsesWith(I, Op0);
5834 if (isa<UndefValue>(Op0)) {
5835 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5836 return ReplaceInstUsesWith(I, Op0);
5837 else // undef << X -> 0, undef >>u X -> 0
5838 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5840 if (isa<UndefValue>(Op1)) {
5841 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5842 return ReplaceInstUsesWith(I, Op0);
5843 else // X << undef, X >>u undef -> 0
5844 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5847 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5848 if (I.getOpcode() == Instruction::AShr)
5849 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5850 if (CSI->isAllOnesValue())
5851 return ReplaceInstUsesWith(I, CSI);
5853 // Try to fold constant and into select arguments.
5854 if (isa<Constant>(Op0))
5855 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5856 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5859 // See if we can turn a signed shr into an unsigned shr.
5860 if (I.isArithmeticShift()) {
5861 if (MaskedValueIsZero(Op0,
5862 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5863 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5867 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5868 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5873 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5874 BinaryOperator &I) {
5875 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5877 // See if we can simplify any instructions used by the instruction whose sole
5878 // purpose is to compute bits we don't care about.
5879 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5880 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5881 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5882 KnownZero, KnownOne))
5885 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5886 // of a signed value.
5888 if (Op1->uge(TypeBits)) {
5889 if (I.getOpcode() != Instruction::AShr)
5890 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5892 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5897 // ((X*C1) << C2) == (X * (C1 << C2))
5898 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5899 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5900 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5901 return BinaryOperator::createMul(BO->getOperand(0),
5902 ConstantExpr::getShl(BOOp, Op1));
5904 // Try to fold constant and into select arguments.
5905 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5906 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5908 if (isa<PHINode>(Op0))
5909 if (Instruction *NV = FoldOpIntoPhi(I))
5912 if (Op0->hasOneUse()) {
5913 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5914 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5917 switch (Op0BO->getOpcode()) {
5919 case Instruction::Add:
5920 case Instruction::And:
5921 case Instruction::Or:
5922 case Instruction::Xor: {
5923 // These operators commute.
5924 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5925 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5926 match(Op0BO->getOperand(1),
5927 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5928 Instruction *YS = BinaryOperator::createShl(
5929 Op0BO->getOperand(0), Op1,
5931 InsertNewInstBefore(YS, I); // (Y << C)
5933 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5934 Op0BO->getOperand(1)->getName());
5935 InsertNewInstBefore(X, I); // (X + (Y << C))
5936 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5937 return BinaryOperator::createAnd(X, ConstantInt::get(
5938 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5941 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5942 Value *Op0BOOp1 = Op0BO->getOperand(1);
5943 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5945 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5946 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5948 Instruction *YS = BinaryOperator::createShl(
5949 Op0BO->getOperand(0), Op1,
5951 InsertNewInstBefore(YS, I); // (Y << C)
5953 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5954 V1->getName()+".mask");
5955 InsertNewInstBefore(XM, I); // X & (CC << C)
5957 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5962 case Instruction::Sub: {
5963 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5964 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5965 match(Op0BO->getOperand(0),
5966 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5967 Instruction *YS = BinaryOperator::createShl(
5968 Op0BO->getOperand(1), Op1,
5970 InsertNewInstBefore(YS, I); // (Y << C)
5972 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5973 Op0BO->getOperand(0)->getName());
5974 InsertNewInstBefore(X, I); // (X + (Y << C))
5975 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5976 return BinaryOperator::createAnd(X, ConstantInt::get(
5977 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5980 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5981 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5982 match(Op0BO->getOperand(0),
5983 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5984 m_ConstantInt(CC))) && V2 == Op1 &&
5985 cast<BinaryOperator>(Op0BO->getOperand(0))
5986 ->getOperand(0)->hasOneUse()) {
5987 Instruction *YS = BinaryOperator::createShl(
5988 Op0BO->getOperand(1), Op1,
5990 InsertNewInstBefore(YS, I); // (Y << C)
5992 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5993 V1->getName()+".mask");
5994 InsertNewInstBefore(XM, I); // X & (CC << C)
5996 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6004 // If the operand is an bitwise operator with a constant RHS, and the
6005 // shift is the only use, we can pull it out of the shift.
6006 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6007 bool isValid = true; // Valid only for And, Or, Xor
6008 bool highBitSet = false; // Transform if high bit of constant set?
6010 switch (Op0BO->getOpcode()) {
6011 default: isValid = false; break; // Do not perform transform!
6012 case Instruction::Add:
6013 isValid = isLeftShift;
6015 case Instruction::Or:
6016 case Instruction::Xor:
6019 case Instruction::And:
6024 // If this is a signed shift right, and the high bit is modified
6025 // by the logical operation, do not perform the transformation.
6026 // The highBitSet boolean indicates the value of the high bit of
6027 // the constant which would cause it to be modified for this
6030 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6031 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6035 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6037 Instruction *NewShift =
6038 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6039 InsertNewInstBefore(NewShift, I);
6040 NewShift->takeName(Op0BO);
6042 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6049 // Find out if this is a shift of a shift by a constant.
6050 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6051 if (ShiftOp && !ShiftOp->isShift())
6054 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6055 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6056 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6057 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6058 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6059 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6060 Value *X = ShiftOp->getOperand(0);
6062 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6063 if (AmtSum > TypeBits)
6066 const IntegerType *Ty = cast<IntegerType>(I.getType());
6068 // Check for (X << c1) << c2 and (X >> c1) >> c2
6069 if (I.getOpcode() == ShiftOp->getOpcode()) {
6070 return BinaryOperator::create(I.getOpcode(), X,
6071 ConstantInt::get(Ty, AmtSum));
6072 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6073 I.getOpcode() == Instruction::AShr) {
6074 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6075 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6076 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6077 I.getOpcode() == Instruction::LShr) {
6078 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6079 Instruction *Shift =
6080 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6081 InsertNewInstBefore(Shift, I);
6083 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6084 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6087 // Okay, if we get here, one shift must be left, and the other shift must be
6088 // right. See if the amounts are equal.
6089 if (ShiftAmt1 == ShiftAmt2) {
6090 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6091 if (I.getOpcode() == Instruction::Shl) {
6092 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6093 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6095 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6096 if (I.getOpcode() == Instruction::LShr) {
6097 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6098 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6100 // We can simplify ((X << C) >>s C) into a trunc + sext.
6101 // NOTE: we could do this for any C, but that would make 'unusual' integer
6102 // types. For now, just stick to ones well-supported by the code
6104 const Type *SExtType = 0;
6105 switch (Ty->getBitWidth() - ShiftAmt1) {
6112 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6117 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6118 InsertNewInstBefore(NewTrunc, I);
6119 return new SExtInst(NewTrunc, Ty);
6121 // Otherwise, we can't handle it yet.
6122 } else if (ShiftAmt1 < ShiftAmt2) {
6123 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6125 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6126 if (I.getOpcode() == Instruction::Shl) {
6127 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6128 ShiftOp->getOpcode() == Instruction::AShr);
6129 Instruction *Shift =
6130 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6131 InsertNewInstBefore(Shift, I);
6133 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6134 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6137 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6138 if (I.getOpcode() == Instruction::LShr) {
6139 assert(ShiftOp->getOpcode() == Instruction::Shl);
6140 Instruction *Shift =
6141 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6142 InsertNewInstBefore(Shift, I);
6144 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6145 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6148 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6150 assert(ShiftAmt2 < ShiftAmt1);
6151 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6153 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6154 if (I.getOpcode() == Instruction::Shl) {
6155 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6156 ShiftOp->getOpcode() == Instruction::AShr);
6157 Instruction *Shift =
6158 BinaryOperator::create(ShiftOp->getOpcode(), X,
6159 ConstantInt::get(Ty, ShiftDiff));
6160 InsertNewInstBefore(Shift, I);
6162 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6163 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6166 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6167 if (I.getOpcode() == Instruction::LShr) {
6168 assert(ShiftOp->getOpcode() == Instruction::Shl);
6169 Instruction *Shift =
6170 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6171 InsertNewInstBefore(Shift, I);
6173 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6174 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6177 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6184 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6185 /// expression. If so, decompose it, returning some value X, such that Val is
6188 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6190 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6191 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6192 Offset = CI->getZExtValue();
6194 return ConstantInt::get(Type::Int32Ty, 0);
6195 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6196 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6197 if (I->getOpcode() == Instruction::Shl) {
6198 // This is a value scaled by '1 << the shift amt'.
6199 Scale = 1U << RHS->getZExtValue();
6201 return I->getOperand(0);
6202 } else if (I->getOpcode() == Instruction::Mul) {
6203 // This value is scaled by 'RHS'.
6204 Scale = RHS->getZExtValue();
6206 return I->getOperand(0);
6207 } else if (I->getOpcode() == Instruction::Add) {
6208 // We have X+C. Check to see if we really have (X*C2)+C1,
6209 // where C1 is divisible by C2.
6212 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6213 Offset += RHS->getZExtValue();
6220 // Otherwise, we can't look past this.
6227 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6228 /// try to eliminate the cast by moving the type information into the alloc.
6229 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6230 AllocationInst &AI) {
6231 const PointerType *PTy = cast<PointerType>(CI.getType());
6233 // Remove any uses of AI that are dead.
6234 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6236 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6237 Instruction *User = cast<Instruction>(*UI++);
6238 if (isInstructionTriviallyDead(User)) {
6239 while (UI != E && *UI == User)
6240 ++UI; // If this instruction uses AI more than once, don't break UI.
6243 DOUT << "IC: DCE: " << *User;
6244 EraseInstFromFunction(*User);
6248 // Get the type really allocated and the type casted to.
6249 const Type *AllocElTy = AI.getAllocatedType();
6250 const Type *CastElTy = PTy->getElementType();
6251 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6253 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6254 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6255 if (CastElTyAlign < AllocElTyAlign) return 0;
6257 // If the allocation has multiple uses, only promote it if we are strictly
6258 // increasing the alignment of the resultant allocation. If we keep it the
6259 // same, we open the door to infinite loops of various kinds.
6260 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6262 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6263 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6264 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6266 // See if we can satisfy the modulus by pulling a scale out of the array
6268 unsigned ArraySizeScale;
6270 Value *NumElements = // See if the array size is a decomposable linear expr.
6271 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6273 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6275 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6276 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6278 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6283 // If the allocation size is constant, form a constant mul expression
6284 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6285 if (isa<ConstantInt>(NumElements))
6286 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6287 // otherwise multiply the amount and the number of elements
6288 else if (Scale != 1) {
6289 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6290 Amt = InsertNewInstBefore(Tmp, AI);
6294 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6295 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6296 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6297 Amt = InsertNewInstBefore(Tmp, AI);
6300 AllocationInst *New;
6301 if (isa<MallocInst>(AI))
6302 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6304 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6305 InsertNewInstBefore(New, AI);
6308 // If the allocation has multiple uses, insert a cast and change all things
6309 // that used it to use the new cast. This will also hack on CI, but it will
6311 if (!AI.hasOneUse()) {
6312 AddUsesToWorkList(AI);
6313 // New is the allocation instruction, pointer typed. AI is the original
6314 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6315 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6316 InsertNewInstBefore(NewCast, AI);
6317 AI.replaceAllUsesWith(NewCast);
6319 return ReplaceInstUsesWith(CI, New);
6322 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6323 /// and return it as type Ty without inserting any new casts and without
6324 /// changing the computed value. This is used by code that tries to decide
6325 /// whether promoting or shrinking integer operations to wider or smaller types
6326 /// will allow us to eliminate a truncate or extend.
6328 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6329 /// extension operation if Ty is larger.
6330 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6331 unsigned CastOpc, int &NumCastsRemoved) {
6332 // We can always evaluate constants in another type.
6333 if (isa<ConstantInt>(V))
6336 Instruction *I = dyn_cast<Instruction>(V);
6337 if (!I) return false;
6339 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6341 // If this is an extension or truncate, we can often eliminate it.
6342 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6343 // If this is a cast from the destination type, we can trivially eliminate
6344 // it, and this will remove a cast overall.
6345 if (I->getOperand(0)->getType() == Ty) {
6346 // If the first operand is itself a cast, and is eliminable, do not count
6347 // this as an eliminable cast. We would prefer to eliminate those two
6349 if (!isa<CastInst>(I->getOperand(0)))
6355 // We can't extend or shrink something that has multiple uses: doing so would
6356 // require duplicating the instruction in general, which isn't profitable.
6357 if (!I->hasOneUse()) return false;
6359 switch (I->getOpcode()) {
6360 case Instruction::Add:
6361 case Instruction::Sub:
6362 case Instruction::And:
6363 case Instruction::Or:
6364 case Instruction::Xor:
6365 // These operators can all arbitrarily be extended or truncated.
6366 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6368 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6371 case Instruction::Shl:
6372 // If we are truncating the result of this SHL, and if it's a shift of a
6373 // constant amount, we can always perform a SHL in a smaller type.
6374 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6375 uint32_t BitWidth = Ty->getBitWidth();
6376 if (BitWidth < OrigTy->getBitWidth() &&
6377 CI->getLimitedValue(BitWidth) < BitWidth)
6378 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6382 case Instruction::LShr:
6383 // If this is a truncate of a logical shr, we can truncate it to a smaller
6384 // lshr iff we know that the bits we would otherwise be shifting in are
6386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6387 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6388 uint32_t BitWidth = Ty->getBitWidth();
6389 if (BitWidth < OrigBitWidth &&
6390 MaskedValueIsZero(I->getOperand(0),
6391 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6392 CI->getLimitedValue(BitWidth) < BitWidth) {
6393 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6398 case Instruction::ZExt:
6399 case Instruction::SExt:
6400 case Instruction::Trunc:
6401 // If this is the same kind of case as our original (e.g. zext+zext), we
6402 // can safely replace it. Note that replacing it does not reduce the number
6403 // of casts in the input.
6404 if (I->getOpcode() == CastOpc)
6409 // TODO: Can handle more cases here.
6416 /// EvaluateInDifferentType - Given an expression that
6417 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6418 /// evaluate the expression.
6419 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6421 if (Constant *C = dyn_cast<Constant>(V))
6422 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6424 // Otherwise, it must be an instruction.
6425 Instruction *I = cast<Instruction>(V);
6426 Instruction *Res = 0;
6427 switch (I->getOpcode()) {
6428 case Instruction::Add:
6429 case Instruction::Sub:
6430 case Instruction::And:
6431 case Instruction::Or:
6432 case Instruction::Xor:
6433 case Instruction::AShr:
6434 case Instruction::LShr:
6435 case Instruction::Shl: {
6436 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6437 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6438 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6439 LHS, RHS, I->getName());
6442 case Instruction::Trunc:
6443 case Instruction::ZExt:
6444 case Instruction::SExt:
6445 // If the source type of the cast is the type we're trying for then we can
6446 // just return the source. There's no need to insert it because it is not
6448 if (I->getOperand(0)->getType() == Ty)
6449 return I->getOperand(0);
6451 // Otherwise, must be the same type of case, so just reinsert a new one.
6452 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6456 // TODO: Can handle more cases here.
6457 assert(0 && "Unreachable!");
6461 return InsertNewInstBefore(Res, *I);
6464 /// @brief Implement the transforms common to all CastInst visitors.
6465 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6466 Value *Src = CI.getOperand(0);
6468 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6469 // eliminate it now.
6470 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6471 if (Instruction::CastOps opc =
6472 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6473 // The first cast (CSrc) is eliminable so we need to fix up or replace
6474 // the second cast (CI). CSrc will then have a good chance of being dead.
6475 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6479 // If we are casting a select then fold the cast into the select
6480 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6481 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6484 // If we are casting a PHI then fold the cast into the PHI
6485 if (isa<PHINode>(Src))
6486 if (Instruction *NV = FoldOpIntoPhi(CI))
6492 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6493 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6494 Value *Src = CI.getOperand(0);
6496 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6497 // If casting the result of a getelementptr instruction with no offset, turn
6498 // this into a cast of the original pointer!
6499 if (GEP->hasAllZeroIndices()) {
6500 // Changing the cast operand is usually not a good idea but it is safe
6501 // here because the pointer operand is being replaced with another
6502 // pointer operand so the opcode doesn't need to change.
6504 CI.setOperand(0, GEP->getOperand(0));
6508 // If the GEP has a single use, and the base pointer is a bitcast, and the
6509 // GEP computes a constant offset, see if we can convert these three
6510 // instructions into fewer. This typically happens with unions and other
6511 // non-type-safe code.
6512 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6513 if (GEP->hasAllConstantIndices()) {
6514 // We are guaranteed to get a constant from EmitGEPOffset.
6515 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6516 int64_t Offset = OffsetV->getSExtValue();
6518 // Get the base pointer input of the bitcast, and the type it points to.
6519 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6520 const Type *GEPIdxTy =
6521 cast<PointerType>(OrigBase->getType())->getElementType();
6522 if (GEPIdxTy->isSized()) {
6523 SmallVector<Value*, 8> NewIndices;
6525 // Start with the index over the outer type. Note that the type size
6526 // might be zero (even if the offset isn't zero) if the indexed type
6527 // is something like [0 x {int, int}]
6528 const Type *IntPtrTy = TD->getIntPtrType();
6529 int64_t FirstIdx = 0;
6530 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6531 FirstIdx = Offset/TySize;
6534 // Handle silly modulus not returning values values [0..TySize).
6538 assert(Offset >= 0);
6540 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6543 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6545 // Index into the types. If we fail, set OrigBase to null.
6547 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6548 const StructLayout *SL = TD->getStructLayout(STy);
6549 if (Offset < (int64_t)SL->getSizeInBytes()) {
6550 unsigned Elt = SL->getElementContainingOffset(Offset);
6551 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6553 Offset -= SL->getElementOffset(Elt);
6554 GEPIdxTy = STy->getElementType(Elt);
6556 // Otherwise, we can't index into this, bail out.
6560 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6561 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6562 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6563 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6566 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6568 GEPIdxTy = STy->getElementType();
6570 // Otherwise, we can't index into this, bail out.
6576 // If we were able to index down into an element, create the GEP
6577 // and bitcast the result. This eliminates one bitcast, potentially
6579 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6581 NewIndices.end(), "");
6582 InsertNewInstBefore(NGEP, CI);
6583 NGEP->takeName(GEP);
6585 if (isa<BitCastInst>(CI))
6586 return new BitCastInst(NGEP, CI.getType());
6587 assert(isa<PtrToIntInst>(CI));
6588 return new PtrToIntInst(NGEP, CI.getType());
6595 return commonCastTransforms(CI);
6600 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6601 /// integer types. This function implements the common transforms for all those
6603 /// @brief Implement the transforms common to CastInst with integer operands
6604 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6605 if (Instruction *Result = commonCastTransforms(CI))
6608 Value *Src = CI.getOperand(0);
6609 const Type *SrcTy = Src->getType();
6610 const Type *DestTy = CI.getType();
6611 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6612 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6614 // See if we can simplify any instructions used by the LHS whose sole
6615 // purpose is to compute bits we don't care about.
6616 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6617 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6618 KnownZero, KnownOne))
6621 // If the source isn't an instruction or has more than one use then we
6622 // can't do anything more.
6623 Instruction *SrcI = dyn_cast<Instruction>(Src);
6624 if (!SrcI || !Src->hasOneUse())
6627 // Attempt to propagate the cast into the instruction for int->int casts.
6628 int NumCastsRemoved = 0;
6629 if (!isa<BitCastInst>(CI) &&
6630 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6631 CI.getOpcode(), NumCastsRemoved)) {
6632 // If this cast is a truncate, evaluting in a different type always
6633 // eliminates the cast, so it is always a win. If this is a zero-extension,
6634 // we need to do an AND to maintain the clear top-part of the computation,
6635 // so we require that the input have eliminated at least one cast. If this
6636 // is a sign extension, we insert two new casts (to do the extension) so we
6637 // require that two casts have been eliminated.
6639 switch (CI.getOpcode()) {
6641 // All the others use floating point so we shouldn't actually
6642 // get here because of the check above.
6643 assert(0 && "Unknown cast type");
6644 case Instruction::Trunc:
6647 case Instruction::ZExt:
6648 DoXForm = NumCastsRemoved >= 1;
6650 case Instruction::SExt:
6651 DoXForm = NumCastsRemoved >= 2;
6656 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6657 CI.getOpcode() == Instruction::SExt);
6658 assert(Res->getType() == DestTy);
6659 switch (CI.getOpcode()) {
6660 default: assert(0 && "Unknown cast type!");
6661 case Instruction::Trunc:
6662 case Instruction::BitCast:
6663 // Just replace this cast with the result.
6664 return ReplaceInstUsesWith(CI, Res);
6665 case Instruction::ZExt: {
6666 // We need to emit an AND to clear the high bits.
6667 assert(SrcBitSize < DestBitSize && "Not a zext?");
6668 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6670 return BinaryOperator::createAnd(Res, C);
6672 case Instruction::SExt:
6673 // We need to emit a cast to truncate, then a cast to sext.
6674 return CastInst::create(Instruction::SExt,
6675 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6681 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6682 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6684 switch (SrcI->getOpcode()) {
6685 case Instruction::Add:
6686 case Instruction::Mul:
6687 case Instruction::And:
6688 case Instruction::Or:
6689 case Instruction::Xor:
6690 // If we are discarding information, rewrite.
6691 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6692 // Don't insert two casts if they cannot be eliminated. We allow
6693 // two casts to be inserted if the sizes are the same. This could
6694 // only be converting signedness, which is a noop.
6695 if (DestBitSize == SrcBitSize ||
6696 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6697 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6698 Instruction::CastOps opcode = CI.getOpcode();
6699 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6700 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6701 return BinaryOperator::create(
6702 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6706 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6707 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6708 SrcI->getOpcode() == Instruction::Xor &&
6709 Op1 == ConstantInt::getTrue() &&
6710 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6711 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6712 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6715 case Instruction::SDiv:
6716 case Instruction::UDiv:
6717 case Instruction::SRem:
6718 case Instruction::URem:
6719 // If we are just changing the sign, rewrite.
6720 if (DestBitSize == SrcBitSize) {
6721 // Don't insert two casts if they cannot be eliminated. We allow
6722 // two casts to be inserted if the sizes are the same. This could
6723 // only be converting signedness, which is a noop.
6724 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6725 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6726 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6728 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6730 return BinaryOperator::create(
6731 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6736 case Instruction::Shl:
6737 // Allow changing the sign of the source operand. Do not allow
6738 // changing the size of the shift, UNLESS the shift amount is a
6739 // constant. We must not change variable sized shifts to a smaller
6740 // size, because it is undefined to shift more bits out than exist
6742 if (DestBitSize == SrcBitSize ||
6743 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6744 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6745 Instruction::BitCast : Instruction::Trunc);
6746 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6747 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6748 return BinaryOperator::createShl(Op0c, Op1c);
6751 case Instruction::AShr:
6752 // If this is a signed shr, and if all bits shifted in are about to be
6753 // truncated off, turn it into an unsigned shr to allow greater
6755 if (DestBitSize < SrcBitSize &&
6756 isa<ConstantInt>(Op1)) {
6757 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6758 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6759 // Insert the new logical shift right.
6760 return BinaryOperator::createLShr(Op0, Op1);
6768 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6769 if (Instruction *Result = commonIntCastTransforms(CI))
6772 Value *Src = CI.getOperand(0);
6773 const Type *Ty = CI.getType();
6774 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6775 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6777 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6778 switch (SrcI->getOpcode()) {
6780 case Instruction::LShr:
6781 // We can shrink lshr to something smaller if we know the bits shifted in
6782 // are already zeros.
6783 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6784 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6786 // Get a mask for the bits shifting in.
6787 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6788 Value* SrcIOp0 = SrcI->getOperand(0);
6789 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6790 if (ShAmt >= DestBitWidth) // All zeros.
6791 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6793 // Okay, we can shrink this. Truncate the input, then return a new
6795 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6796 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6798 return BinaryOperator::createLShr(V1, V2);
6800 } else { // This is a variable shr.
6802 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6803 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6804 // loop-invariant and CSE'd.
6805 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6806 Value *One = ConstantInt::get(SrcI->getType(), 1);
6808 Value *V = InsertNewInstBefore(
6809 BinaryOperator::createShl(One, SrcI->getOperand(1),
6811 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6812 SrcI->getOperand(0),
6814 Value *Zero = Constant::getNullValue(V->getType());
6815 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6825 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6826 // If one of the common conversion will work ..
6827 if (Instruction *Result = commonIntCastTransforms(CI))
6830 Value *Src = CI.getOperand(0);
6832 // If this is a cast of a cast
6833 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6834 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6835 // types and if the sizes are just right we can convert this into a logical
6836 // 'and' which will be much cheaper than the pair of casts.
6837 if (isa<TruncInst>(CSrc)) {
6838 // Get the sizes of the types involved
6839 Value *A = CSrc->getOperand(0);
6840 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6841 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6842 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6843 // If we're actually extending zero bits and the trunc is a no-op
6844 if (MidSize < DstSize && SrcSize == DstSize) {
6845 // Replace both of the casts with an And of the type mask.
6846 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6847 Constant *AndConst = ConstantInt::get(AndValue);
6849 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6850 // Unfortunately, if the type changed, we need to cast it back.
6851 if (And->getType() != CI.getType()) {
6852 And->setName(CSrc->getName()+".mask");
6853 InsertNewInstBefore(And, CI);
6854 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6861 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6862 // If we are just checking for a icmp eq of a single bit and zext'ing it
6863 // to an integer, then shift the bit to the appropriate place and then
6864 // cast to integer to avoid the comparison.
6865 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6866 const APInt &Op1CV = Op1C->getValue();
6868 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6869 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6870 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6871 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6872 Value *In = ICI->getOperand(0);
6873 Value *Sh = ConstantInt::get(In->getType(),
6874 In->getType()->getPrimitiveSizeInBits()-1);
6875 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6876 In->getName()+".lobit"),
6878 if (In->getType() != CI.getType())
6879 In = CastInst::createIntegerCast(In, CI.getType(),
6880 false/*ZExt*/, "tmp", &CI);
6882 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6883 Constant *One = ConstantInt::get(In->getType(), 1);
6884 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6885 In->getName()+".not"),
6889 return ReplaceInstUsesWith(CI, In);
6894 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6895 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6896 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6897 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6898 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6899 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6900 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6901 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6902 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6903 // This only works for EQ and NE
6904 ICI->isEquality()) {
6905 // If Op1C some other power of two, convert:
6906 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6907 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6908 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6909 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6911 APInt KnownZeroMask(~KnownZero);
6912 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6913 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6914 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6915 // (X&4) == 2 --> false
6916 // (X&4) != 2 --> true
6917 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6918 Res = ConstantExpr::getZExt(Res, CI.getType());
6919 return ReplaceInstUsesWith(CI, Res);
6922 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6923 Value *In = ICI->getOperand(0);
6925 // Perform a logical shr by shiftamt.
6926 // Insert the shift to put the result in the low bit.
6927 In = InsertNewInstBefore(
6928 BinaryOperator::createLShr(In,
6929 ConstantInt::get(In->getType(), ShiftAmt),
6930 In->getName()+".lobit"), CI);
6933 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6934 Constant *One = ConstantInt::get(In->getType(), 1);
6935 In = BinaryOperator::createXor(In, One, "tmp");
6936 InsertNewInstBefore(cast<Instruction>(In), CI);
6939 if (CI.getType() == In->getType())
6940 return ReplaceInstUsesWith(CI, In);
6942 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6950 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6951 if (Instruction *I = commonIntCastTransforms(CI))
6954 Value *Src = CI.getOperand(0);
6956 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6957 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6958 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6959 // If we are just checking for a icmp eq of a single bit and zext'ing it
6960 // to an integer, then shift the bit to the appropriate place and then
6961 // cast to integer to avoid the comparison.
6962 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6963 const APInt &Op1CV = Op1C->getValue();
6965 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6966 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6967 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6968 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6969 Value *In = ICI->getOperand(0);
6970 Value *Sh = ConstantInt::get(In->getType(),
6971 In->getType()->getPrimitiveSizeInBits()-1);
6972 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6973 In->getName()+".lobit"),
6975 if (In->getType() != CI.getType())
6976 In = CastInst::createIntegerCast(In, CI.getType(),
6977 true/*SExt*/, "tmp", &CI);
6979 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6980 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6981 In->getName()+".not"), CI);
6983 return ReplaceInstUsesWith(CI, In);
6991 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6992 return commonCastTransforms(CI);
6995 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6996 return commonCastTransforms(CI);
6999 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7000 return commonCastTransforms(CI);
7003 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7004 return commonCastTransforms(CI);
7007 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7008 return commonCastTransforms(CI);
7011 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7012 return commonCastTransforms(CI);
7015 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7016 return commonPointerCastTransforms(CI);
7019 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7020 return commonCastTransforms(CI);
7023 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7024 // If the operands are integer typed then apply the integer transforms,
7025 // otherwise just apply the common ones.
7026 Value *Src = CI.getOperand(0);
7027 const Type *SrcTy = Src->getType();
7028 const Type *DestTy = CI.getType();
7030 if (SrcTy->isInteger() && DestTy->isInteger()) {
7031 if (Instruction *Result = commonIntCastTransforms(CI))
7033 } else if (isa<PointerType>(SrcTy)) {
7034 if (Instruction *I = commonPointerCastTransforms(CI))
7037 if (Instruction *Result = commonCastTransforms(CI))
7042 // Get rid of casts from one type to the same type. These are useless and can
7043 // be replaced by the operand.
7044 if (DestTy == Src->getType())
7045 return ReplaceInstUsesWith(CI, Src);
7047 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7048 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7049 const Type *DstElTy = DstPTy->getElementType();
7050 const Type *SrcElTy = SrcPTy->getElementType();
7052 // If we are casting a malloc or alloca to a pointer to a type of the same
7053 // size, rewrite the allocation instruction to allocate the "right" type.
7054 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7055 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7058 // If the source and destination are pointers, and this cast is equivalent
7059 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7060 // This can enhance SROA and other transforms that want type-safe pointers.
7061 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7062 unsigned NumZeros = 0;
7063 while (SrcElTy != DstElTy &&
7064 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7065 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7066 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7070 // If we found a path from the src to dest, create the getelementptr now.
7071 if (SrcElTy == DstElTy) {
7072 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7073 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7074 ((Instruction*) NULL));
7078 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7079 if (SVI->hasOneUse()) {
7080 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7081 // a bitconvert to a vector with the same # elts.
7082 if (isa<VectorType>(DestTy) &&
7083 cast<VectorType>(DestTy)->getNumElements() ==
7084 SVI->getType()->getNumElements()) {
7086 // If either of the operands is a cast from CI.getType(), then
7087 // evaluating the shuffle in the casted destination's type will allow
7088 // us to eliminate at least one cast.
7089 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7090 Tmp->getOperand(0)->getType() == DestTy) ||
7091 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7092 Tmp->getOperand(0)->getType() == DestTy)) {
7093 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7094 SVI->getOperand(0), DestTy, &CI);
7095 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7096 SVI->getOperand(1), DestTy, &CI);
7097 // Return a new shuffle vector. Use the same element ID's, as we
7098 // know the vector types match #elts.
7099 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7107 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7109 /// %D = select %cond, %C, %A
7111 /// %C = select %cond, %B, 0
7114 /// Assuming that the specified instruction is an operand to the select, return
7115 /// a bitmask indicating which operands of this instruction are foldable if they
7116 /// equal the other incoming value of the select.
7118 static unsigned GetSelectFoldableOperands(Instruction *I) {
7119 switch (I->getOpcode()) {
7120 case Instruction::Add:
7121 case Instruction::Mul:
7122 case Instruction::And:
7123 case Instruction::Or:
7124 case Instruction::Xor:
7125 return 3; // Can fold through either operand.
7126 case Instruction::Sub: // Can only fold on the amount subtracted.
7127 case Instruction::Shl: // Can only fold on the shift amount.
7128 case Instruction::LShr:
7129 case Instruction::AShr:
7132 return 0; // Cannot fold
7136 /// GetSelectFoldableConstant - For the same transformation as the previous
7137 /// function, return the identity constant that goes into the select.
7138 static Constant *GetSelectFoldableConstant(Instruction *I) {
7139 switch (I->getOpcode()) {
7140 default: assert(0 && "This cannot happen!"); abort();
7141 case Instruction::Add:
7142 case Instruction::Sub:
7143 case Instruction::Or:
7144 case Instruction::Xor:
7145 case Instruction::Shl:
7146 case Instruction::LShr:
7147 case Instruction::AShr:
7148 return Constant::getNullValue(I->getType());
7149 case Instruction::And:
7150 return Constant::getAllOnesValue(I->getType());
7151 case Instruction::Mul:
7152 return ConstantInt::get(I->getType(), 1);
7156 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7157 /// have the same opcode and only one use each. Try to simplify this.
7158 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7160 if (TI->getNumOperands() == 1) {
7161 // If this is a non-volatile load or a cast from the same type,
7164 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7167 return 0; // unknown unary op.
7170 // Fold this by inserting a select from the input values.
7171 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7172 FI->getOperand(0), SI.getName()+".v");
7173 InsertNewInstBefore(NewSI, SI);
7174 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7178 // Only handle binary operators here.
7179 if (!isa<BinaryOperator>(TI))
7182 // Figure out if the operations have any operands in common.
7183 Value *MatchOp, *OtherOpT, *OtherOpF;
7185 if (TI->getOperand(0) == FI->getOperand(0)) {
7186 MatchOp = TI->getOperand(0);
7187 OtherOpT = TI->getOperand(1);
7188 OtherOpF = FI->getOperand(1);
7189 MatchIsOpZero = true;
7190 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7191 MatchOp = TI->getOperand(1);
7192 OtherOpT = TI->getOperand(0);
7193 OtherOpF = FI->getOperand(0);
7194 MatchIsOpZero = false;
7195 } else if (!TI->isCommutative()) {
7197 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7198 MatchOp = TI->getOperand(0);
7199 OtherOpT = TI->getOperand(1);
7200 OtherOpF = FI->getOperand(0);
7201 MatchIsOpZero = true;
7202 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7203 MatchOp = TI->getOperand(1);
7204 OtherOpT = TI->getOperand(0);
7205 OtherOpF = FI->getOperand(1);
7206 MatchIsOpZero = true;
7211 // If we reach here, they do have operations in common.
7212 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7213 OtherOpF, SI.getName()+".v");
7214 InsertNewInstBefore(NewSI, SI);
7216 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7218 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7220 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7222 assert(0 && "Shouldn't get here");
7226 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7227 Value *CondVal = SI.getCondition();
7228 Value *TrueVal = SI.getTrueValue();
7229 Value *FalseVal = SI.getFalseValue();
7231 // select true, X, Y -> X
7232 // select false, X, Y -> Y
7233 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7234 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7236 // select C, X, X -> X
7237 if (TrueVal == FalseVal)
7238 return ReplaceInstUsesWith(SI, TrueVal);
7240 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7241 return ReplaceInstUsesWith(SI, FalseVal);
7242 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7243 return ReplaceInstUsesWith(SI, TrueVal);
7244 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7245 if (isa<Constant>(TrueVal))
7246 return ReplaceInstUsesWith(SI, TrueVal);
7248 return ReplaceInstUsesWith(SI, FalseVal);
7251 if (SI.getType() == Type::Int1Ty) {
7252 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7253 if (C->getZExtValue()) {
7254 // Change: A = select B, true, C --> A = or B, C
7255 return BinaryOperator::createOr(CondVal, FalseVal);
7257 // Change: A = select B, false, C --> A = and !B, C
7259 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7260 "not."+CondVal->getName()), SI);
7261 return BinaryOperator::createAnd(NotCond, FalseVal);
7263 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7264 if (C->getZExtValue() == false) {
7265 // Change: A = select B, C, false --> A = and B, C
7266 return BinaryOperator::createAnd(CondVal, TrueVal);
7268 // Change: A = select B, C, true --> A = or !B, C
7270 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7271 "not."+CondVal->getName()), SI);
7272 return BinaryOperator::createOr(NotCond, TrueVal);
7277 // Selecting between two integer constants?
7278 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7279 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7280 // select C, 1, 0 -> zext C to int
7281 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7282 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7283 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7284 // select C, 0, 1 -> zext !C to int
7286 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7287 "not."+CondVal->getName()), SI);
7288 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7291 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7293 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7295 // (x <s 0) ? -1 : 0 -> ashr x, 31
7296 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7297 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7298 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7299 // The comparison constant and the result are not neccessarily the
7300 // same width. Make an all-ones value by inserting a AShr.
7301 Value *X = IC->getOperand(0);
7302 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7303 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7304 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7306 InsertNewInstBefore(SRA, SI);
7308 // Finally, convert to the type of the select RHS. We figure out
7309 // if this requires a SExt, Trunc or BitCast based on the sizes.
7310 Instruction::CastOps opc = Instruction::BitCast;
7311 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7312 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7313 if (SRASize < SISize)
7314 opc = Instruction::SExt;
7315 else if (SRASize > SISize)
7316 opc = Instruction::Trunc;
7317 return CastInst::create(opc, SRA, SI.getType());
7322 // If one of the constants is zero (we know they can't both be) and we
7323 // have an icmp instruction with zero, and we have an 'and' with the
7324 // non-constant value, eliminate this whole mess. This corresponds to
7325 // cases like this: ((X & 27) ? 27 : 0)
7326 if (TrueValC->isZero() || FalseValC->isZero())
7327 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7328 cast<Constant>(IC->getOperand(1))->isNullValue())
7329 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7330 if (ICA->getOpcode() == Instruction::And &&
7331 isa<ConstantInt>(ICA->getOperand(1)) &&
7332 (ICA->getOperand(1) == TrueValC ||
7333 ICA->getOperand(1) == FalseValC) &&
7334 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7335 // Okay, now we know that everything is set up, we just don't
7336 // know whether we have a icmp_ne or icmp_eq and whether the
7337 // true or false val is the zero.
7338 bool ShouldNotVal = !TrueValC->isZero();
7339 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7342 V = InsertNewInstBefore(BinaryOperator::create(
7343 Instruction::Xor, V, ICA->getOperand(1)), SI);
7344 return ReplaceInstUsesWith(SI, V);
7349 // See if we are selecting two values based on a comparison of the two values.
7350 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7351 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7352 // Transform (X == Y) ? X : Y -> Y
7353 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7354 // This is not safe in general for floating point:
7355 // consider X== -0, Y== +0.
7356 // It becomes safe if either operand is a nonzero constant.
7357 ConstantFP *CFPt, *CFPf;
7358 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7359 !CFPt->getValueAPF().isZero()) ||
7360 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7361 !CFPf->getValueAPF().isZero()))
7362 return ReplaceInstUsesWith(SI, FalseVal);
7364 // Transform (X != Y) ? X : Y -> X
7365 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7366 return ReplaceInstUsesWith(SI, TrueVal);
7367 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7369 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7370 // Transform (X == Y) ? Y : X -> X
7371 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7372 // This is not safe in general for floating point:
7373 // consider X== -0, Y== +0.
7374 // It becomes safe if either operand is a nonzero constant.
7375 ConstantFP *CFPt, *CFPf;
7376 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7377 !CFPt->getValueAPF().isZero()) ||
7378 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7379 !CFPf->getValueAPF().isZero()))
7380 return ReplaceInstUsesWith(SI, FalseVal);
7382 // Transform (X != Y) ? Y : X -> Y
7383 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7384 return ReplaceInstUsesWith(SI, TrueVal);
7385 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7389 // See if we are selecting two values based on a comparison of the two values.
7390 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7391 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7392 // Transform (X == Y) ? X : Y -> Y
7393 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7394 return ReplaceInstUsesWith(SI, FalseVal);
7395 // Transform (X != Y) ? X : Y -> X
7396 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7397 return ReplaceInstUsesWith(SI, TrueVal);
7398 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7400 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7401 // Transform (X == Y) ? Y : X -> X
7402 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7403 return ReplaceInstUsesWith(SI, FalseVal);
7404 // Transform (X != Y) ? Y : X -> Y
7405 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7406 return ReplaceInstUsesWith(SI, TrueVal);
7407 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7411 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7412 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7413 if (TI->hasOneUse() && FI->hasOneUse()) {
7414 Instruction *AddOp = 0, *SubOp = 0;
7416 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7417 if (TI->getOpcode() == FI->getOpcode())
7418 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7421 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7422 // even legal for FP.
7423 if (TI->getOpcode() == Instruction::Sub &&
7424 FI->getOpcode() == Instruction::Add) {
7425 AddOp = FI; SubOp = TI;
7426 } else if (FI->getOpcode() == Instruction::Sub &&
7427 TI->getOpcode() == Instruction::Add) {
7428 AddOp = TI; SubOp = FI;
7432 Value *OtherAddOp = 0;
7433 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7434 OtherAddOp = AddOp->getOperand(1);
7435 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7436 OtherAddOp = AddOp->getOperand(0);
7440 // So at this point we know we have (Y -> OtherAddOp):
7441 // select C, (add X, Y), (sub X, Z)
7442 Value *NegVal; // Compute -Z
7443 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7444 NegVal = ConstantExpr::getNeg(C);
7446 NegVal = InsertNewInstBefore(
7447 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7450 Value *NewTrueOp = OtherAddOp;
7451 Value *NewFalseOp = NegVal;
7453 std::swap(NewTrueOp, NewFalseOp);
7454 Instruction *NewSel =
7455 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7457 NewSel = InsertNewInstBefore(NewSel, SI);
7458 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7463 // See if we can fold the select into one of our operands.
7464 if (SI.getType()->isInteger()) {
7465 // See the comment above GetSelectFoldableOperands for a description of the
7466 // transformation we are doing here.
7467 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7468 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7469 !isa<Constant>(FalseVal))
7470 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7471 unsigned OpToFold = 0;
7472 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7474 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7479 Constant *C = GetSelectFoldableConstant(TVI);
7480 Instruction *NewSel =
7481 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7482 InsertNewInstBefore(NewSel, SI);
7483 NewSel->takeName(TVI);
7484 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7485 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7487 assert(0 && "Unknown instruction!!");
7492 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7493 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7494 !isa<Constant>(TrueVal))
7495 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7496 unsigned OpToFold = 0;
7497 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7499 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7504 Constant *C = GetSelectFoldableConstant(FVI);
7505 Instruction *NewSel =
7506 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7507 InsertNewInstBefore(NewSel, SI);
7508 NewSel->takeName(FVI);
7509 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7510 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7512 assert(0 && "Unknown instruction!!");
7517 if (BinaryOperator::isNot(CondVal)) {
7518 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7519 SI.setOperand(1, FalseVal);
7520 SI.setOperand(2, TrueVal);
7527 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7528 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7529 /// and it is more than the alignment of the ultimate object, see if we can
7530 /// increase the alignment of the ultimate object, making this check succeed.
7531 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7532 unsigned PrefAlign = 0) {
7533 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7534 unsigned Align = GV->getAlignment();
7535 if (Align == 0 && TD)
7536 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7538 // If there is a large requested alignment and we can, bump up the alignment
7540 if (PrefAlign > Align && GV->hasInitializer()) {
7541 GV->setAlignment(PrefAlign);
7545 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7546 unsigned Align = AI->getAlignment();
7547 if (Align == 0 && TD) {
7548 if (isa<AllocaInst>(AI))
7549 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7550 else if (isa<MallocInst>(AI)) {
7551 // Malloc returns maximally aligned memory.
7552 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7555 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7558 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7562 // If there is a requested alignment and if this is an alloca, round up. We
7563 // don't do this for malloc, because some systems can't respect the request.
7564 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7565 AI->setAlignment(PrefAlign);
7569 } else if (isa<BitCastInst>(V) ||
7570 (isa<ConstantExpr>(V) &&
7571 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7572 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7574 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7575 // If all indexes are zero, it is just the alignment of the base pointer.
7576 bool AllZeroOperands = true;
7577 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7578 if (!isa<Constant>(GEPI->getOperand(i)) ||
7579 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7580 AllZeroOperands = false;
7584 if (AllZeroOperands) {
7585 // Treat this like a bitcast.
7586 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7589 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7590 if (BaseAlignment == 0) return 0;
7592 // Otherwise, if the base alignment is >= the alignment we expect for the
7593 // base pointer type, then we know that the resultant pointer is aligned at
7594 // least as much as its type requires.
7597 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7598 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7599 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7600 if (Align <= BaseAlignment) {
7601 const Type *GEPTy = GEPI->getType();
7602 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7603 Align = std::min(Align, (unsigned)
7604 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7613 /// visitCallInst - CallInst simplification. This mostly only handles folding
7614 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7615 /// the heavy lifting.
7617 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7618 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7619 if (!II) return visitCallSite(&CI);
7621 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7623 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7624 bool Changed = false;
7626 // memmove/cpy/set of zero bytes is a noop.
7627 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7628 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7630 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7631 if (CI->getZExtValue() == 1) {
7632 // Replace the instruction with just byte operations. We would
7633 // transform other cases to loads/stores, but we don't know if
7634 // alignment is sufficient.
7638 // If we have a memmove and the source operation is a constant global,
7639 // then the source and dest pointers can't alias, so we can change this
7640 // into a call to memcpy.
7641 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7642 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7643 if (GVSrc->isConstant()) {
7644 Module *M = CI.getParent()->getParent()->getParent();
7646 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7648 Name = "llvm.memcpy.i32";
7650 Name = "llvm.memcpy.i64";
7651 Constant *MemCpy = M->getOrInsertFunction(Name,
7652 CI.getCalledFunction()->getFunctionType());
7653 CI.setOperand(0, MemCpy);
7658 // If we can determine a pointer alignment that is bigger than currently
7659 // set, update the alignment.
7660 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7661 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7662 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7663 unsigned Align = std::min(Alignment1, Alignment2);
7664 if (MI->getAlignment()->getZExtValue() < Align) {
7665 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7669 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7671 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(CI.getOperand(3));
7672 if (isa<MemCpyInst>(MI))
7674 unsigned Size = MemOpLength->getZExtValue();
7675 unsigned Align = cast<ConstantInt>(CI.getOperand(4))->getZExtValue();
7676 const PointerType *PTy = cast<PointerType>(CI.getOperand(1)->getType());
7677 const Type *MTy = PTy->getElementType();
7678 PointerType *NewPtrTy = NULL;
7679 if (MTy == Type::Int8Ty) {
7681 NewPtrTy = PointerType::get(Type::Int64Ty);
7683 NewPtrTy = PointerType::get(Type::Int32Ty);
7685 NewPtrTy = PointerType::get(Type::Int16Ty);
7687 NewPtrTy = PointerType::get(Type::Int8Ty);
7688 } else if (MTy == Type::Int16Ty) {
7690 NewPtrTy = PointerType::get(Type::Int64Ty);
7692 NewPtrTy = PointerType::get(Type::Int32Ty);
7694 NewPtrTy = PointerType::get(Type::Int16Ty);
7695 } else if (MTy == Type::Int32Ty) {
7697 NewPtrTy = PointerType::get(Type::Int64Ty);
7699 NewPtrTy = PointerType::get(Type::Int32Ty);
7700 } else if (MTy == Type::Int64Ty) {
7702 NewPtrTy = PointerType::get(Type::Int64Ty);
7706 InsertCastBefore(Instruction::BitCast,CI.getOperand(2),NewPtrTy,CI);
7708 InsertCastBefore(Instruction::BitCast,CI.getOperand(1),NewPtrTy,CI);
7709 Value *L = new LoadInst(Src, "tmp", false, Align, &CI);
7710 Value *NS = new StoreInst(L, Dest, false, Align, &CI);
7711 CI.replaceAllUsesWith(NS);
7713 return EraseInstFromFunction(CI);
7716 } else if (isa<MemSetInst>(MI)) {
7717 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7718 if (MI->getAlignment()->getZExtValue() < Alignment) {
7719 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7724 if (Changed) return II;
7726 switch (II->getIntrinsicID()) {
7728 case Intrinsic::ppc_altivec_lvx:
7729 case Intrinsic::ppc_altivec_lvxl:
7730 case Intrinsic::x86_sse_loadu_ps:
7731 case Intrinsic::x86_sse2_loadu_pd:
7732 case Intrinsic::x86_sse2_loadu_dq:
7733 // Turn PPC lvx -> load if the pointer is known aligned.
7734 // Turn X86 loadups -> load if the pointer is known aligned.
7735 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7736 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7737 PointerType::get(II->getType()), CI);
7738 return new LoadInst(Ptr);
7741 case Intrinsic::ppc_altivec_stvx:
7742 case Intrinsic::ppc_altivec_stvxl:
7743 // Turn stvx -> store if the pointer is known aligned.
7744 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7745 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7746 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7748 return new StoreInst(II->getOperand(1), Ptr);
7751 case Intrinsic::x86_sse_storeu_ps:
7752 case Intrinsic::x86_sse2_storeu_pd:
7753 case Intrinsic::x86_sse2_storeu_dq:
7754 case Intrinsic::x86_sse2_storel_dq:
7755 // Turn X86 storeu -> store if the pointer is known aligned.
7756 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7757 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7758 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7760 return new StoreInst(II->getOperand(2), Ptr);
7764 case Intrinsic::x86_sse_cvttss2si: {
7765 // These intrinsics only demands the 0th element of its input vector. If
7766 // we can simplify the input based on that, do so now.
7768 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7770 II->setOperand(1, V);
7776 case Intrinsic::ppc_altivec_vperm:
7777 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7778 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7779 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7781 // Check that all of the elements are integer constants or undefs.
7782 bool AllEltsOk = true;
7783 for (unsigned i = 0; i != 16; ++i) {
7784 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7785 !isa<UndefValue>(Mask->getOperand(i))) {
7792 // Cast the input vectors to byte vectors.
7793 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7794 II->getOperand(1), Mask->getType(), CI);
7795 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7796 II->getOperand(2), Mask->getType(), CI);
7797 Value *Result = UndefValue::get(Op0->getType());
7799 // Only extract each element once.
7800 Value *ExtractedElts[32];
7801 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7803 for (unsigned i = 0; i != 16; ++i) {
7804 if (isa<UndefValue>(Mask->getOperand(i)))
7806 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7807 Idx &= 31; // Match the hardware behavior.
7809 if (ExtractedElts[Idx] == 0) {
7811 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7812 InsertNewInstBefore(Elt, CI);
7813 ExtractedElts[Idx] = Elt;
7816 // Insert this value into the result vector.
7817 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7818 InsertNewInstBefore(cast<Instruction>(Result), CI);
7820 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7825 case Intrinsic::stackrestore: {
7826 // If the save is right next to the restore, remove the restore. This can
7827 // happen when variable allocas are DCE'd.
7828 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7829 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7830 BasicBlock::iterator BI = SS;
7832 return EraseInstFromFunction(CI);
7836 // If the stack restore is in a return/unwind block and if there are no
7837 // allocas or calls between the restore and the return, nuke the restore.
7838 TerminatorInst *TI = II->getParent()->getTerminator();
7839 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7840 BasicBlock::iterator BI = II;
7841 bool CannotRemove = false;
7842 for (++BI; &*BI != TI; ++BI) {
7843 if (isa<AllocaInst>(BI) ||
7844 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7845 CannotRemove = true;
7850 return EraseInstFromFunction(CI);
7857 return visitCallSite(II);
7860 // InvokeInst simplification
7862 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7863 return visitCallSite(&II);
7866 // visitCallSite - Improvements for call and invoke instructions.
7868 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7869 bool Changed = false;
7871 // If the callee is a constexpr cast of a function, attempt to move the cast
7872 // to the arguments of the call/invoke.
7873 if (transformConstExprCastCall(CS)) return 0;
7875 Value *Callee = CS.getCalledValue();
7877 if (Function *CalleeF = dyn_cast<Function>(Callee))
7878 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7879 Instruction *OldCall = CS.getInstruction();
7880 // If the call and callee calling conventions don't match, this call must
7881 // be unreachable, as the call is undefined.
7882 new StoreInst(ConstantInt::getTrue(),
7883 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7884 if (!OldCall->use_empty())
7885 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7886 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7887 return EraseInstFromFunction(*OldCall);
7891 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7892 // This instruction is not reachable, just remove it. We insert a store to
7893 // undef so that we know that this code is not reachable, despite the fact
7894 // that we can't modify the CFG here.
7895 new StoreInst(ConstantInt::getTrue(),
7896 UndefValue::get(PointerType::get(Type::Int1Ty)),
7897 CS.getInstruction());
7899 if (!CS.getInstruction()->use_empty())
7900 CS.getInstruction()->
7901 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7903 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7904 // Don't break the CFG, insert a dummy cond branch.
7905 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7906 ConstantInt::getTrue(), II);
7908 return EraseInstFromFunction(*CS.getInstruction());
7911 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
7912 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
7913 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
7914 return transformCallThroughTrampoline(CS);
7916 const PointerType *PTy = cast<PointerType>(Callee->getType());
7917 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7918 if (FTy->isVarArg()) {
7919 // See if we can optimize any arguments passed through the varargs area of
7921 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7922 E = CS.arg_end(); I != E; ++I)
7923 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7924 // If this cast does not effect the value passed through the varargs
7925 // area, we can eliminate the use of the cast.
7926 Value *Op = CI->getOperand(0);
7927 if (CI->isLosslessCast()) {
7934 return Changed ? CS.getInstruction() : 0;
7937 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7938 // attempt to move the cast to the arguments of the call/invoke.
7940 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7941 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7942 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7943 if (CE->getOpcode() != Instruction::BitCast ||
7944 !isa<Function>(CE->getOperand(0)))
7946 Function *Callee = cast<Function>(CE->getOperand(0));
7947 Instruction *Caller = CS.getInstruction();
7949 // Okay, this is a cast from a function to a different type. Unless doing so
7950 // would cause a type conversion of one of our arguments, change this call to
7951 // be a direct call with arguments casted to the appropriate types.
7953 const FunctionType *FT = Callee->getFunctionType();
7954 const Type *OldRetTy = Caller->getType();
7956 const FunctionType *ActualFT =
7957 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7959 // If the parameter attributes don't match up, don't do the xform. We don't
7960 // want to lose an sret attribute or something.
7961 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
7964 // Check to see if we are changing the return type...
7965 if (OldRetTy != FT->getReturnType()) {
7966 if (Callee->isDeclaration() && !Caller->use_empty() &&
7967 // Conversion is ok if changing from pointer to int of same size.
7968 !(isa<PointerType>(FT->getReturnType()) &&
7969 TD->getIntPtrType() == OldRetTy))
7970 return false; // Cannot transform this return value.
7972 // If the callsite is an invoke instruction, and the return value is used by
7973 // a PHI node in a successor, we cannot change the return type of the call
7974 // because there is no place to put the cast instruction (without breaking
7975 // the critical edge). Bail out in this case.
7976 if (!Caller->use_empty())
7977 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7978 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7980 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7981 if (PN->getParent() == II->getNormalDest() ||
7982 PN->getParent() == II->getUnwindDest())
7986 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7987 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7989 CallSite::arg_iterator AI = CS.arg_begin();
7990 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7991 const Type *ParamTy = FT->getParamType(i);
7992 const Type *ActTy = (*AI)->getType();
7993 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7994 //Some conversions are safe even if we do not have a body.
7995 //Either we can cast directly, or we can upconvert the argument
7996 bool isConvertible = ActTy == ParamTy ||
7997 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7998 (ParamTy->isInteger() && ActTy->isInteger() &&
7999 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8000 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8001 && c->getValue().isStrictlyPositive());
8002 if (Callee->isDeclaration() && !isConvertible) return false;
8004 // Most other conversions can be done if we have a body, even if these
8005 // lose information, e.g. int->short.
8006 // Some conversions cannot be done at all, e.g. float to pointer.
8007 // Logic here parallels CastInst::getCastOpcode (the design there
8008 // requires legality checks like this be done before calling it).
8009 if (ParamTy->isInteger()) {
8010 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8011 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8014 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
8015 !isa<PointerType>(ActTy))
8017 } else if (ParamTy->isFloatingPoint()) {
8018 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8019 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8022 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
8024 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
8025 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8026 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
8029 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
8031 } else if (isa<PointerType>(ParamTy)) {
8032 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
8039 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8040 Callee->isDeclaration())
8041 return false; // Do not delete arguments unless we have a function body...
8043 // Okay, we decided that this is a safe thing to do: go ahead and start
8044 // inserting cast instructions as necessary...
8045 std::vector<Value*> Args;
8046 Args.reserve(NumActualArgs);
8048 AI = CS.arg_begin();
8049 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8050 const Type *ParamTy = FT->getParamType(i);
8051 if ((*AI)->getType() == ParamTy) {
8052 Args.push_back(*AI);
8054 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8055 false, ParamTy, false);
8056 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8057 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8061 // If the function takes more arguments than the call was taking, add them
8063 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8064 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8066 // If we are removing arguments to the function, emit an obnoxious warning...
8067 if (FT->getNumParams() < NumActualArgs)
8068 if (!FT->isVarArg()) {
8069 cerr << "WARNING: While resolving call to function '"
8070 << Callee->getName() << "' arguments were dropped!\n";
8072 // Add all of the arguments in their promoted form to the arg list...
8073 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8074 const Type *PTy = getPromotedType((*AI)->getType());
8075 if (PTy != (*AI)->getType()) {
8076 // Must promote to pass through va_arg area!
8077 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8079 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8080 InsertNewInstBefore(Cast, *Caller);
8081 Args.push_back(Cast);
8083 Args.push_back(*AI);
8088 if (FT->getReturnType() == Type::VoidTy)
8089 Caller->setName(""); // Void type should not have a name.
8092 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8093 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8094 Args.begin(), Args.end(), Caller->getName(), Caller);
8095 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8097 NC = new CallInst(Callee, Args.begin(), Args.end(),
8098 Caller->getName(), Caller);
8099 if (cast<CallInst>(Caller)->isTailCall())
8100 cast<CallInst>(NC)->setTailCall();
8101 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8104 // Insert a cast of the return type as necessary.
8106 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8107 if (NV->getType() != Type::VoidTy) {
8108 const Type *CallerTy = Caller->getType();
8109 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8111 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8113 // If this is an invoke instruction, we should insert it after the first
8114 // non-phi, instruction in the normal successor block.
8115 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8116 BasicBlock::iterator I = II->getNormalDest()->begin();
8117 while (isa<PHINode>(I)) ++I;
8118 InsertNewInstBefore(NC, *I);
8120 // Otherwise, it's a call, just insert cast right after the call instr
8121 InsertNewInstBefore(NC, *Caller);
8123 AddUsersToWorkList(*Caller);
8125 NV = UndefValue::get(Caller->getType());
8129 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8130 Caller->replaceAllUsesWith(NV);
8131 Caller->eraseFromParent();
8132 RemoveFromWorkList(Caller);
8136 // transformCallThroughTrampoline - Turn a call to a function created by the
8137 // init_trampoline intrinsic into a direct call to the underlying function.
8139 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8140 Value *Callee = CS.getCalledValue();
8141 const PointerType *PTy = cast<PointerType>(Callee->getType());
8142 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8144 IntrinsicInst *Tramp =
8145 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8148 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8149 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8150 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8152 if (const ParamAttrsList *NestAttrs = NestFTy->getParamAttrs()) {
8153 unsigned NestIdx = 1;
8154 const Type *NestTy = 0;
8155 uint16_t NestAttr = 0;
8157 // Look for a parameter marked with the 'nest' attribute.
8158 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8159 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8160 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8161 // Record the parameter type and any other attributes.
8163 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8168 Instruction *Caller = CS.getInstruction();
8169 std::vector<Value*> NewArgs;
8170 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8172 // Insert the nest argument into the call argument list, which may
8173 // mean appending it.
8176 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8178 if (Idx == NestIdx) {
8179 // Add the chain argument.
8180 Value *NestVal = Tramp->getOperand(3);
8181 if (NestVal->getType() != NestTy)
8182 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8183 NewArgs.push_back(NestVal);
8189 // Add the original argument.
8190 NewArgs.push_back(*I);
8196 // The trampoline may have been bitcast to a bogus type (FTy).
8197 // Handle this by synthesizing a new function type, equal to FTy
8198 // with the chain parameter inserted. Likewise for attributes.
8200 const ParamAttrsList *Attrs = FTy->getParamAttrs();
8201 std::vector<const Type*> NewTypes;
8202 ParamAttrsVector NewAttrs;
8203 NewTypes.reserve(FTy->getNumParams()+1);
8205 // Add any function result attributes.
8206 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8208 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8210 // Insert the chain's type into the list of parameter types, which may
8211 // mean appending it. Likewise for the chain's attributes.
8214 FunctionType::param_iterator I = FTy->param_begin(),
8215 E = FTy->param_end();
8218 if (Idx == NestIdx) {
8219 // Add the chain's type and attributes.
8220 NewTypes.push_back(NestTy);
8221 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8227 // Add the original type and attributes.
8228 NewTypes.push_back(*I);
8229 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8232 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8238 // Replace the trampoline call with a direct call. Let the generic
8239 // code sort out any function type mismatches.
8240 FunctionType *NewFTy =
8241 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg(),
8242 ParamAttrsList::get(NewAttrs));
8243 Constant *NewCallee = NestF->getType() == PointerType::get(NewFTy) ?
8244 NestF : ConstantExpr::getBitCast(NestF, PointerType::get(NewFTy));
8246 Instruction *NewCaller;
8247 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8248 NewCaller = new InvokeInst(NewCallee,
8249 II->getNormalDest(), II->getUnwindDest(),
8250 NewArgs.begin(), NewArgs.end(),
8251 Caller->getName(), Caller);
8252 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8254 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8255 Caller->getName(), Caller);
8256 if (cast<CallInst>(Caller)->isTailCall())
8257 cast<CallInst>(NewCaller)->setTailCall();
8258 cast<CallInst>(NewCaller)->
8259 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8261 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8262 Caller->replaceAllUsesWith(NewCaller);
8263 Caller->eraseFromParent();
8264 RemoveFromWorkList(Caller);
8269 // Replace the trampoline call with a direct call. Since there is no 'nest'
8270 // parameter, there is no need to adjust the argument list. Let the generic
8271 // code sort out any function type mismatches.
8272 Constant *NewCallee =
8273 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8274 CS.setCalledFunction(NewCallee);
8275 return CS.getInstruction();
8278 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8279 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8280 /// and a single binop.
8281 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8282 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8283 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8284 isa<CmpInst>(FirstInst));
8285 unsigned Opc = FirstInst->getOpcode();
8286 Value *LHSVal = FirstInst->getOperand(0);
8287 Value *RHSVal = FirstInst->getOperand(1);
8289 const Type *LHSType = LHSVal->getType();
8290 const Type *RHSType = RHSVal->getType();
8292 // Scan to see if all operands are the same opcode, all have one use, and all
8293 // kill their operands (i.e. the operands have one use).
8294 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8295 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8296 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8297 // Verify type of the LHS matches so we don't fold cmp's of different
8298 // types or GEP's with different index types.
8299 I->getOperand(0)->getType() != LHSType ||
8300 I->getOperand(1)->getType() != RHSType)
8303 // If they are CmpInst instructions, check their predicates
8304 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8305 if (cast<CmpInst>(I)->getPredicate() !=
8306 cast<CmpInst>(FirstInst)->getPredicate())
8309 // Keep track of which operand needs a phi node.
8310 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8311 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8314 // Otherwise, this is safe to transform, determine if it is profitable.
8316 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8317 // Indexes are often folded into load/store instructions, so we don't want to
8318 // hide them behind a phi.
8319 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8322 Value *InLHS = FirstInst->getOperand(0);
8323 Value *InRHS = FirstInst->getOperand(1);
8324 PHINode *NewLHS = 0, *NewRHS = 0;
8326 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8327 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8328 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8329 InsertNewInstBefore(NewLHS, PN);
8334 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8335 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8336 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8337 InsertNewInstBefore(NewRHS, PN);
8341 // Add all operands to the new PHIs.
8342 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8344 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8345 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8348 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8349 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8353 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8354 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8355 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8356 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8359 assert(isa<GetElementPtrInst>(FirstInst));
8360 return new GetElementPtrInst(LHSVal, RHSVal);
8364 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8365 /// of the block that defines it. This means that it must be obvious the value
8366 /// of the load is not changed from the point of the load to the end of the
8369 /// Finally, it is safe, but not profitable, to sink a load targetting a
8370 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8372 static bool isSafeToSinkLoad(LoadInst *L) {
8373 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8375 for (++BBI; BBI != E; ++BBI)
8376 if (BBI->mayWriteToMemory())
8379 // Check for non-address taken alloca. If not address-taken already, it isn't
8380 // profitable to do this xform.
8381 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8382 bool isAddressTaken = false;
8383 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8385 if (isa<LoadInst>(UI)) continue;
8386 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8387 // If storing TO the alloca, then the address isn't taken.
8388 if (SI->getOperand(1) == AI) continue;
8390 isAddressTaken = true;
8394 if (!isAddressTaken)
8402 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8403 // operator and they all are only used by the PHI, PHI together their
8404 // inputs, and do the operation once, to the result of the PHI.
8405 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8406 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8408 // Scan the instruction, looking for input operations that can be folded away.
8409 // If all input operands to the phi are the same instruction (e.g. a cast from
8410 // the same type or "+42") we can pull the operation through the PHI, reducing
8411 // code size and simplifying code.
8412 Constant *ConstantOp = 0;
8413 const Type *CastSrcTy = 0;
8414 bool isVolatile = false;
8415 if (isa<CastInst>(FirstInst)) {
8416 CastSrcTy = FirstInst->getOperand(0)->getType();
8417 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8418 // Can fold binop, compare or shift here if the RHS is a constant,
8419 // otherwise call FoldPHIArgBinOpIntoPHI.
8420 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8421 if (ConstantOp == 0)
8422 return FoldPHIArgBinOpIntoPHI(PN);
8423 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8424 isVolatile = LI->isVolatile();
8425 // We can't sink the load if the loaded value could be modified between the
8426 // load and the PHI.
8427 if (LI->getParent() != PN.getIncomingBlock(0) ||
8428 !isSafeToSinkLoad(LI))
8430 } else if (isa<GetElementPtrInst>(FirstInst)) {
8431 if (FirstInst->getNumOperands() == 2)
8432 return FoldPHIArgBinOpIntoPHI(PN);
8433 // Can't handle general GEPs yet.
8436 return 0; // Cannot fold this operation.
8439 // Check to see if all arguments are the same operation.
8440 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8441 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8442 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8443 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8446 if (I->getOperand(0)->getType() != CastSrcTy)
8447 return 0; // Cast operation must match.
8448 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8449 // We can't sink the load if the loaded value could be modified between
8450 // the load and the PHI.
8451 if (LI->isVolatile() != isVolatile ||
8452 LI->getParent() != PN.getIncomingBlock(i) ||
8453 !isSafeToSinkLoad(LI))
8455 } else if (I->getOperand(1) != ConstantOp) {
8460 // Okay, they are all the same operation. Create a new PHI node of the
8461 // correct type, and PHI together all of the LHS's of the instructions.
8462 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8463 PN.getName()+".in");
8464 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8466 Value *InVal = FirstInst->getOperand(0);
8467 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8469 // Add all operands to the new PHI.
8470 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8471 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8472 if (NewInVal != InVal)
8474 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8479 // The new PHI unions all of the same values together. This is really
8480 // common, so we handle it intelligently here for compile-time speed.
8484 InsertNewInstBefore(NewPN, PN);
8488 // Insert and return the new operation.
8489 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8490 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8491 else if (isa<LoadInst>(FirstInst))
8492 return new LoadInst(PhiVal, "", isVolatile);
8493 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8494 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8495 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8496 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8497 PhiVal, ConstantOp);
8499 assert(0 && "Unknown operation");
8503 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8505 static bool DeadPHICycle(PHINode *PN,
8506 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8507 if (PN->use_empty()) return true;
8508 if (!PN->hasOneUse()) return false;
8510 // Remember this node, and if we find the cycle, return.
8511 if (!PotentiallyDeadPHIs.insert(PN))
8514 // Don't scan crazily complex things.
8515 if (PotentiallyDeadPHIs.size() == 16)
8518 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8519 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8524 // PHINode simplification
8526 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8527 // If LCSSA is around, don't mess with Phi nodes
8528 if (MustPreserveLCSSA) return 0;
8530 if (Value *V = PN.hasConstantValue())
8531 return ReplaceInstUsesWith(PN, V);
8533 // If all PHI operands are the same operation, pull them through the PHI,
8534 // reducing code size.
8535 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8536 PN.getIncomingValue(0)->hasOneUse())
8537 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8540 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8541 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8542 // PHI)... break the cycle.
8543 if (PN.hasOneUse()) {
8544 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8545 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8546 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8547 PotentiallyDeadPHIs.insert(&PN);
8548 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8549 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8552 // If this phi has a single use, and if that use just computes a value for
8553 // the next iteration of a loop, delete the phi. This occurs with unused
8554 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8555 // common case here is good because the only other things that catch this
8556 // are induction variable analysis (sometimes) and ADCE, which is only run
8558 if (PHIUser->hasOneUse() &&
8559 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8560 PHIUser->use_back() == &PN) {
8561 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8568 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8569 Instruction *InsertPoint,
8571 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8572 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8573 // We must cast correctly to the pointer type. Ensure that we
8574 // sign extend the integer value if it is smaller as this is
8575 // used for address computation.
8576 Instruction::CastOps opcode =
8577 (VTySize < PtrSize ? Instruction::SExt :
8578 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8579 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8583 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8584 Value *PtrOp = GEP.getOperand(0);
8585 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8586 // If so, eliminate the noop.
8587 if (GEP.getNumOperands() == 1)
8588 return ReplaceInstUsesWith(GEP, PtrOp);
8590 if (isa<UndefValue>(GEP.getOperand(0)))
8591 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8593 bool HasZeroPointerIndex = false;
8594 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8595 HasZeroPointerIndex = C->isNullValue();
8597 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8598 return ReplaceInstUsesWith(GEP, PtrOp);
8600 // Eliminate unneeded casts for indices.
8601 bool MadeChange = false;
8603 gep_type_iterator GTI = gep_type_begin(GEP);
8604 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8605 if (isa<SequentialType>(*GTI)) {
8606 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8607 if (CI->getOpcode() == Instruction::ZExt ||
8608 CI->getOpcode() == Instruction::SExt) {
8609 const Type *SrcTy = CI->getOperand(0)->getType();
8610 // We can eliminate a cast from i32 to i64 iff the target
8611 // is a 32-bit pointer target.
8612 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8614 GEP.setOperand(i, CI->getOperand(0));
8618 // If we are using a wider index than needed for this platform, shrink it
8619 // to what we need. If the incoming value needs a cast instruction,
8620 // insert it. This explicit cast can make subsequent optimizations more
8622 Value *Op = GEP.getOperand(i);
8623 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8624 if (Constant *C = dyn_cast<Constant>(Op)) {
8625 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8628 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8630 GEP.setOperand(i, Op);
8635 if (MadeChange) return &GEP;
8637 // If this GEP instruction doesn't move the pointer, and if the input operand
8638 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8639 // real input to the dest type.
8640 if (GEP.hasAllZeroIndices()) {
8641 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
8642 // If the bitcast is of an allocation, and the allocation will be
8643 // converted to match the type of the cast, don't touch this.
8644 if (isa<AllocationInst>(BCI->getOperand(0))) {
8645 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
8646 if (Instruction *I = visitBitCast(*BCI)) {
8649 BCI->getParent()->getInstList().insert(BCI, I);
8650 ReplaceInstUsesWith(*BCI, I);
8655 return new BitCastInst(BCI->getOperand(0), GEP.getType());
8659 // Combine Indices - If the source pointer to this getelementptr instruction
8660 // is a getelementptr instruction, combine the indices of the two
8661 // getelementptr instructions into a single instruction.
8663 SmallVector<Value*, 8> SrcGEPOperands;
8664 if (User *Src = dyn_castGetElementPtr(PtrOp))
8665 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8667 if (!SrcGEPOperands.empty()) {
8668 // Note that if our source is a gep chain itself that we wait for that
8669 // chain to be resolved before we perform this transformation. This
8670 // avoids us creating a TON of code in some cases.
8672 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8673 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8674 return 0; // Wait until our source is folded to completion.
8676 SmallVector<Value*, 8> Indices;
8678 // Find out whether the last index in the source GEP is a sequential idx.
8679 bool EndsWithSequential = false;
8680 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8681 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8682 EndsWithSequential = !isa<StructType>(*I);
8684 // Can we combine the two pointer arithmetics offsets?
8685 if (EndsWithSequential) {
8686 // Replace: gep (gep %P, long B), long A, ...
8687 // With: T = long A+B; gep %P, T, ...
8689 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8690 if (SO1 == Constant::getNullValue(SO1->getType())) {
8692 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8695 // If they aren't the same type, convert both to an integer of the
8696 // target's pointer size.
8697 if (SO1->getType() != GO1->getType()) {
8698 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8699 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8700 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8701 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8703 unsigned PS = TD->getPointerSize();
8704 if (TD->getTypeSize(SO1->getType()) == PS) {
8705 // Convert GO1 to SO1's type.
8706 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8708 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8709 // Convert SO1 to GO1's type.
8710 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8712 const Type *PT = TD->getIntPtrType();
8713 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8714 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8718 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8719 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8721 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8722 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8726 // Recycle the GEP we already have if possible.
8727 if (SrcGEPOperands.size() == 2) {
8728 GEP.setOperand(0, SrcGEPOperands[0]);
8729 GEP.setOperand(1, Sum);
8732 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8733 SrcGEPOperands.end()-1);
8734 Indices.push_back(Sum);
8735 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8737 } else if (isa<Constant>(*GEP.idx_begin()) &&
8738 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8739 SrcGEPOperands.size() != 1) {
8740 // Otherwise we can do the fold if the first index of the GEP is a zero
8741 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8742 SrcGEPOperands.end());
8743 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8746 if (!Indices.empty())
8747 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8748 Indices.end(), GEP.getName());
8750 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8751 // GEP of global variable. If all of the indices for this GEP are
8752 // constants, we can promote this to a constexpr instead of an instruction.
8754 // Scan for nonconstants...
8755 SmallVector<Constant*, 8> Indices;
8756 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8757 for (; I != E && isa<Constant>(*I); ++I)
8758 Indices.push_back(cast<Constant>(*I));
8760 if (I == E) { // If they are all constants...
8761 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8762 &Indices[0],Indices.size());
8764 // Replace all uses of the GEP with the new constexpr...
8765 return ReplaceInstUsesWith(GEP, CE);
8767 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8768 if (!isa<PointerType>(X->getType())) {
8769 // Not interesting. Source pointer must be a cast from pointer.
8770 } else if (HasZeroPointerIndex) {
8771 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8772 // into : GEP [10 x ubyte]* X, long 0, ...
8774 // This occurs when the program declares an array extern like "int X[];"
8776 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8777 const PointerType *XTy = cast<PointerType>(X->getType());
8778 if (const ArrayType *XATy =
8779 dyn_cast<ArrayType>(XTy->getElementType()))
8780 if (const ArrayType *CATy =
8781 dyn_cast<ArrayType>(CPTy->getElementType()))
8782 if (CATy->getElementType() == XATy->getElementType()) {
8783 // At this point, we know that the cast source type is a pointer
8784 // to an array of the same type as the destination pointer
8785 // array. Because the array type is never stepped over (there
8786 // is a leading zero) we can fold the cast into this GEP.
8787 GEP.setOperand(0, X);
8790 } else if (GEP.getNumOperands() == 2) {
8791 // Transform things like:
8792 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8793 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8794 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8795 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8796 if (isa<ArrayType>(SrcElTy) &&
8797 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8798 TD->getTypeSize(ResElTy)) {
8800 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8801 Idx[1] = GEP.getOperand(1);
8802 Value *V = InsertNewInstBefore(
8803 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8804 // V and GEP are both pointer types --> BitCast
8805 return new BitCastInst(V, GEP.getType());
8808 // Transform things like:
8809 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8810 // (where tmp = 8*tmp2) into:
8811 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8813 if (isa<ArrayType>(SrcElTy) &&
8814 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8815 uint64_t ArrayEltSize =
8816 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8818 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8819 // allow either a mul, shift, or constant here.
8821 ConstantInt *Scale = 0;
8822 if (ArrayEltSize == 1) {
8823 NewIdx = GEP.getOperand(1);
8824 Scale = ConstantInt::get(NewIdx->getType(), 1);
8825 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8826 NewIdx = ConstantInt::get(CI->getType(), 1);
8828 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8829 if (Inst->getOpcode() == Instruction::Shl &&
8830 isa<ConstantInt>(Inst->getOperand(1))) {
8831 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8832 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8833 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8834 NewIdx = Inst->getOperand(0);
8835 } else if (Inst->getOpcode() == Instruction::Mul &&
8836 isa<ConstantInt>(Inst->getOperand(1))) {
8837 Scale = cast<ConstantInt>(Inst->getOperand(1));
8838 NewIdx = Inst->getOperand(0);
8842 // If the index will be to exactly the right offset with the scale taken
8843 // out, perform the transformation.
8844 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8845 if (isa<ConstantInt>(Scale))
8846 Scale = ConstantInt::get(Scale->getType(),
8847 Scale->getZExtValue() / ArrayEltSize);
8848 if (Scale->getZExtValue() != 1) {
8849 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8851 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8852 NewIdx = InsertNewInstBefore(Sc, GEP);
8855 // Insert the new GEP instruction.
8857 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8859 Instruction *NewGEP =
8860 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
8861 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8862 // The NewGEP must be pointer typed, so must the old one -> BitCast
8863 return new BitCastInst(NewGEP, GEP.getType());
8872 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8873 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8874 if (AI.isArrayAllocation()) // Check C != 1
8875 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8877 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8878 AllocationInst *New = 0;
8880 // Create and insert the replacement instruction...
8881 if (isa<MallocInst>(AI))
8882 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8884 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8885 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8888 InsertNewInstBefore(New, AI);
8890 // Scan to the end of the allocation instructions, to skip over a block of
8891 // allocas if possible...
8893 BasicBlock::iterator It = New;
8894 while (isa<AllocationInst>(*It)) ++It;
8896 // Now that I is pointing to the first non-allocation-inst in the block,
8897 // insert our getelementptr instruction...
8899 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8903 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
8904 New->getName()+".sub", It);
8906 // Now make everything use the getelementptr instead of the original
8908 return ReplaceInstUsesWith(AI, V);
8909 } else if (isa<UndefValue>(AI.getArraySize())) {
8910 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8913 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8914 // Note that we only do this for alloca's, because malloc should allocate and
8915 // return a unique pointer, even for a zero byte allocation.
8916 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8917 TD->getTypeSize(AI.getAllocatedType()) == 0)
8918 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8923 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8924 Value *Op = FI.getOperand(0);
8926 // free undef -> unreachable.
8927 if (isa<UndefValue>(Op)) {
8928 // Insert a new store to null because we cannot modify the CFG here.
8929 new StoreInst(ConstantInt::getTrue(),
8930 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8931 return EraseInstFromFunction(FI);
8934 // If we have 'free null' delete the instruction. This can happen in stl code
8935 // when lots of inlining happens.
8936 if (isa<ConstantPointerNull>(Op))
8937 return EraseInstFromFunction(FI);
8939 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8940 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8941 FI.setOperand(0, CI->getOperand(0));
8945 // Change free (gep X, 0,0,0,0) into free(X)
8946 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8947 if (GEPI->hasAllZeroIndices()) {
8948 AddToWorkList(GEPI);
8949 FI.setOperand(0, GEPI->getOperand(0));
8954 // Change free(malloc) into nothing, if the malloc has a single use.
8955 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8956 if (MI->hasOneUse()) {
8957 EraseInstFromFunction(FI);
8958 return EraseInstFromFunction(*MI);
8965 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8966 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8967 User *CI = cast<User>(LI.getOperand(0));
8968 Value *CastOp = CI->getOperand(0);
8970 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8971 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8972 const Type *SrcPTy = SrcTy->getElementType();
8974 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8975 isa<VectorType>(DestPTy)) {
8976 // If the source is an array, the code below will not succeed. Check to
8977 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8979 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8980 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8981 if (ASrcTy->getNumElements() != 0) {
8983 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8984 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8985 SrcTy = cast<PointerType>(CastOp->getType());
8986 SrcPTy = SrcTy->getElementType();
8989 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8990 isa<VectorType>(SrcPTy)) &&
8991 // Do not allow turning this into a load of an integer, which is then
8992 // casted to a pointer, this pessimizes pointer analysis a lot.
8993 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8994 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8995 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8997 // Okay, we are casting from one integer or pointer type to another of
8998 // the same size. Instead of casting the pointer before the load, cast
8999 // the result of the loaded value.
9000 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9002 LI.isVolatile()),LI);
9003 // Now cast the result of the load.
9004 return new BitCastInst(NewLoad, LI.getType());
9011 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9012 /// from this value cannot trap. If it is not obviously safe to load from the
9013 /// specified pointer, we do a quick local scan of the basic block containing
9014 /// ScanFrom, to determine if the address is already accessed.
9015 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9016 // If it is an alloca it is always safe to load from.
9017 if (isa<AllocaInst>(V)) return true;
9019 // If it is a global variable it is mostly safe to load from.
9020 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9021 // Don't try to evaluate aliases. External weak GV can be null.
9022 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9024 // Otherwise, be a little bit agressive by scanning the local block where we
9025 // want to check to see if the pointer is already being loaded or stored
9026 // from/to. If so, the previous load or store would have already trapped,
9027 // so there is no harm doing an extra load (also, CSE will later eliminate
9028 // the load entirely).
9029 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9034 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9035 if (LI->getOperand(0) == V) return true;
9036 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9037 if (SI->getOperand(1) == V) return true;
9043 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9044 /// until we find the underlying object a pointer is referring to or something
9045 /// we don't understand. Note that the returned pointer may be offset from the
9046 /// input, because we ignore GEP indices.
9047 static Value *GetUnderlyingObject(Value *Ptr) {
9049 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9050 if (CE->getOpcode() == Instruction::BitCast ||
9051 CE->getOpcode() == Instruction::GetElementPtr)
9052 Ptr = CE->getOperand(0);
9055 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9056 Ptr = BCI->getOperand(0);
9057 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9058 Ptr = GEP->getOperand(0);
9065 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9066 Value *Op = LI.getOperand(0);
9068 // Attempt to improve the alignment.
9069 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9070 if (KnownAlign > LI.getAlignment())
9071 LI.setAlignment(KnownAlign);
9073 // load (cast X) --> cast (load X) iff safe
9074 if (isa<CastInst>(Op))
9075 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9078 // None of the following transforms are legal for volatile loads.
9079 if (LI.isVolatile()) return 0;
9081 if (&LI.getParent()->front() != &LI) {
9082 BasicBlock::iterator BBI = &LI; --BBI;
9083 // If the instruction immediately before this is a store to the same
9084 // address, do a simple form of store->load forwarding.
9085 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9086 if (SI->getOperand(1) == LI.getOperand(0))
9087 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9088 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9089 if (LIB->getOperand(0) == LI.getOperand(0))
9090 return ReplaceInstUsesWith(LI, LIB);
9093 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
9094 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
9095 // Insert a new store to null instruction before the load to indicate
9096 // that this code is not reachable. We do this instead of inserting
9097 // an unreachable instruction directly because we cannot modify the
9099 new StoreInst(UndefValue::get(LI.getType()),
9100 Constant::getNullValue(Op->getType()), &LI);
9101 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9104 if (Constant *C = dyn_cast<Constant>(Op)) {
9105 // load null/undef -> undef
9106 if ((C->isNullValue() || isa<UndefValue>(C))) {
9107 // Insert a new store to null instruction before the load to indicate that
9108 // this code is not reachable. We do this instead of inserting an
9109 // unreachable instruction directly because we cannot modify the CFG.
9110 new StoreInst(UndefValue::get(LI.getType()),
9111 Constant::getNullValue(Op->getType()), &LI);
9112 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9115 // Instcombine load (constant global) into the value loaded.
9116 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9117 if (GV->isConstant() && !GV->isDeclaration())
9118 return ReplaceInstUsesWith(LI, GV->getInitializer());
9120 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9121 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9122 if (CE->getOpcode() == Instruction::GetElementPtr) {
9123 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9124 if (GV->isConstant() && !GV->isDeclaration())
9126 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9127 return ReplaceInstUsesWith(LI, V);
9128 if (CE->getOperand(0)->isNullValue()) {
9129 // Insert a new store to null instruction before the load to indicate
9130 // that this code is not reachable. We do this instead of inserting
9131 // an unreachable instruction directly because we cannot modify the
9133 new StoreInst(UndefValue::get(LI.getType()),
9134 Constant::getNullValue(Op->getType()), &LI);
9135 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9138 } else if (CE->isCast()) {
9139 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9144 // If this load comes from anywhere in a constant global, and if the global
9145 // is all undef or zero, we know what it loads.
9146 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9147 if (GV->isConstant() && GV->hasInitializer()) {
9148 if (GV->getInitializer()->isNullValue())
9149 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9150 else if (isa<UndefValue>(GV->getInitializer()))
9151 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9155 if (Op->hasOneUse()) {
9156 // Change select and PHI nodes to select values instead of addresses: this
9157 // helps alias analysis out a lot, allows many others simplifications, and
9158 // exposes redundancy in the code.
9160 // Note that we cannot do the transformation unless we know that the
9161 // introduced loads cannot trap! Something like this is valid as long as
9162 // the condition is always false: load (select bool %C, int* null, int* %G),
9163 // but it would not be valid if we transformed it to load from null
9166 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9167 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9168 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9169 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9170 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9171 SI->getOperand(1)->getName()+".val"), LI);
9172 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9173 SI->getOperand(2)->getName()+".val"), LI);
9174 return new SelectInst(SI->getCondition(), V1, V2);
9177 // load (select (cond, null, P)) -> load P
9178 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9179 if (C->isNullValue()) {
9180 LI.setOperand(0, SI->getOperand(2));
9184 // load (select (cond, P, null)) -> load P
9185 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9186 if (C->isNullValue()) {
9187 LI.setOperand(0, SI->getOperand(1));
9195 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9197 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9198 User *CI = cast<User>(SI.getOperand(1));
9199 Value *CastOp = CI->getOperand(0);
9201 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9202 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9203 const Type *SrcPTy = SrcTy->getElementType();
9205 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9206 // If the source is an array, the code below will not succeed. Check to
9207 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9209 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9210 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9211 if (ASrcTy->getNumElements() != 0) {
9213 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9214 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9215 SrcTy = cast<PointerType>(CastOp->getType());
9216 SrcPTy = SrcTy->getElementType();
9219 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9220 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9221 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9223 // Okay, we are casting from one integer or pointer type to another of
9224 // the same size. Instead of casting the pointer before
9225 // the store, cast the value to be stored.
9227 Value *SIOp0 = SI.getOperand(0);
9228 Instruction::CastOps opcode = Instruction::BitCast;
9229 const Type* CastSrcTy = SIOp0->getType();
9230 const Type* CastDstTy = SrcPTy;
9231 if (isa<PointerType>(CastDstTy)) {
9232 if (CastSrcTy->isInteger())
9233 opcode = Instruction::IntToPtr;
9234 } else if (isa<IntegerType>(CastDstTy)) {
9235 if (isa<PointerType>(SIOp0->getType()))
9236 opcode = Instruction::PtrToInt;
9238 if (Constant *C = dyn_cast<Constant>(SIOp0))
9239 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9241 NewCast = IC.InsertNewInstBefore(
9242 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9244 return new StoreInst(NewCast, CastOp);
9251 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9252 Value *Val = SI.getOperand(0);
9253 Value *Ptr = SI.getOperand(1);
9255 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9256 EraseInstFromFunction(SI);
9261 // If the RHS is an alloca with a single use, zapify the store, making the
9263 if (Ptr->hasOneUse()) {
9264 if (isa<AllocaInst>(Ptr)) {
9265 EraseInstFromFunction(SI);
9270 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9271 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9272 GEP->getOperand(0)->hasOneUse()) {
9273 EraseInstFromFunction(SI);
9279 // Attempt to improve the alignment.
9280 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9281 if (KnownAlign > SI.getAlignment())
9282 SI.setAlignment(KnownAlign);
9284 // Do really simple DSE, to catch cases where there are several consequtive
9285 // stores to the same location, separated by a few arithmetic operations. This
9286 // situation often occurs with bitfield accesses.
9287 BasicBlock::iterator BBI = &SI;
9288 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9292 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9293 // Prev store isn't volatile, and stores to the same location?
9294 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9297 EraseInstFromFunction(*PrevSI);
9303 // If this is a load, we have to stop. However, if the loaded value is from
9304 // the pointer we're loading and is producing the pointer we're storing,
9305 // then *this* store is dead (X = load P; store X -> P).
9306 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9307 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9308 EraseInstFromFunction(SI);
9312 // Otherwise, this is a load from some other location. Stores before it
9317 // Don't skip over loads or things that can modify memory.
9318 if (BBI->mayWriteToMemory())
9323 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9325 // store X, null -> turns into 'unreachable' in SimplifyCFG
9326 if (isa<ConstantPointerNull>(Ptr)) {
9327 if (!isa<UndefValue>(Val)) {
9328 SI.setOperand(0, UndefValue::get(Val->getType()));
9329 if (Instruction *U = dyn_cast<Instruction>(Val))
9330 AddToWorkList(U); // Dropped a use.
9333 return 0; // Do not modify these!
9336 // store undef, Ptr -> noop
9337 if (isa<UndefValue>(Val)) {
9338 EraseInstFromFunction(SI);
9343 // If the pointer destination is a cast, see if we can fold the cast into the
9345 if (isa<CastInst>(Ptr))
9346 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9348 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9350 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9354 // If this store is the last instruction in the basic block, and if the block
9355 // ends with an unconditional branch, try to move it to the successor block.
9357 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9358 if (BI->isUnconditional())
9359 if (SimplifyStoreAtEndOfBlock(SI))
9360 return 0; // xform done!
9365 /// SimplifyStoreAtEndOfBlock - Turn things like:
9366 /// if () { *P = v1; } else { *P = v2 }
9367 /// into a phi node with a store in the successor.
9369 /// Simplify things like:
9370 /// *P = v1; if () { *P = v2; }
9371 /// into a phi node with a store in the successor.
9373 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9374 BasicBlock *StoreBB = SI.getParent();
9376 // Check to see if the successor block has exactly two incoming edges. If
9377 // so, see if the other predecessor contains a store to the same location.
9378 // if so, insert a PHI node (if needed) and move the stores down.
9379 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9381 // Determine whether Dest has exactly two predecessors and, if so, compute
9382 // the other predecessor.
9383 pred_iterator PI = pred_begin(DestBB);
9384 BasicBlock *OtherBB = 0;
9388 if (PI == pred_end(DestBB))
9391 if (*PI != StoreBB) {
9396 if (++PI != pred_end(DestBB))
9400 // Verify that the other block ends in a branch and is not otherwise empty.
9401 BasicBlock::iterator BBI = OtherBB->getTerminator();
9402 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9403 if (!OtherBr || BBI == OtherBB->begin())
9406 // If the other block ends in an unconditional branch, check for the 'if then
9407 // else' case. there is an instruction before the branch.
9408 StoreInst *OtherStore = 0;
9409 if (OtherBr->isUnconditional()) {
9410 // If this isn't a store, or isn't a store to the same location, bail out.
9412 OtherStore = dyn_cast<StoreInst>(BBI);
9413 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9416 // Otherwise, the other block ended with a conditional branch. If one of the
9417 // destinations is StoreBB, then we have the if/then case.
9418 if (OtherBr->getSuccessor(0) != StoreBB &&
9419 OtherBr->getSuccessor(1) != StoreBB)
9422 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9423 // if/then triangle. See if there is a store to the same ptr as SI that
9424 // lives in OtherBB.
9426 // Check to see if we find the matching store.
9427 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9428 if (OtherStore->getOperand(1) != SI.getOperand(1))
9432 // If we find something that may be using the stored value, or if we run
9433 // out of instructions, we can't do the xform.
9434 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9435 BBI == OtherBB->begin())
9439 // In order to eliminate the store in OtherBr, we have to
9440 // make sure nothing reads the stored value in StoreBB.
9441 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9442 // FIXME: This should really be AA driven.
9443 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9448 // Insert a PHI node now if we need it.
9449 Value *MergedVal = OtherStore->getOperand(0);
9450 if (MergedVal != SI.getOperand(0)) {
9451 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9452 PN->reserveOperandSpace(2);
9453 PN->addIncoming(SI.getOperand(0), SI.getParent());
9454 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9455 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9458 // Advance to a place where it is safe to insert the new store and
9460 BBI = DestBB->begin();
9461 while (isa<PHINode>(BBI)) ++BBI;
9462 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9463 OtherStore->isVolatile()), *BBI);
9465 // Nuke the old stores.
9466 EraseInstFromFunction(SI);
9467 EraseInstFromFunction(*OtherStore);
9473 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9474 // Change br (not X), label True, label False to: br X, label False, True
9476 BasicBlock *TrueDest;
9477 BasicBlock *FalseDest;
9478 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9479 !isa<Constant>(X)) {
9480 // Swap Destinations and condition...
9482 BI.setSuccessor(0, FalseDest);
9483 BI.setSuccessor(1, TrueDest);
9487 // Cannonicalize fcmp_one -> fcmp_oeq
9488 FCmpInst::Predicate FPred; Value *Y;
9489 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9490 TrueDest, FalseDest)))
9491 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9492 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9493 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9494 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9495 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9496 NewSCC->takeName(I);
9497 // Swap Destinations and condition...
9498 BI.setCondition(NewSCC);
9499 BI.setSuccessor(0, FalseDest);
9500 BI.setSuccessor(1, TrueDest);
9501 RemoveFromWorkList(I);
9502 I->eraseFromParent();
9503 AddToWorkList(NewSCC);
9507 // Cannonicalize icmp_ne -> icmp_eq
9508 ICmpInst::Predicate IPred;
9509 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9510 TrueDest, FalseDest)))
9511 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9512 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9513 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9514 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9515 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9516 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9517 NewSCC->takeName(I);
9518 // Swap Destinations and condition...
9519 BI.setCondition(NewSCC);
9520 BI.setSuccessor(0, FalseDest);
9521 BI.setSuccessor(1, TrueDest);
9522 RemoveFromWorkList(I);
9523 I->eraseFromParent();;
9524 AddToWorkList(NewSCC);
9531 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9532 Value *Cond = SI.getCondition();
9533 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9534 if (I->getOpcode() == Instruction::Add)
9535 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9536 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9537 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9538 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9540 SI.setOperand(0, I->getOperand(0));
9548 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9549 /// is to leave as a vector operation.
9550 static bool CheapToScalarize(Value *V, bool isConstant) {
9551 if (isa<ConstantAggregateZero>(V))
9553 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9554 if (isConstant) return true;
9555 // If all elts are the same, we can extract.
9556 Constant *Op0 = C->getOperand(0);
9557 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9558 if (C->getOperand(i) != Op0)
9562 Instruction *I = dyn_cast<Instruction>(V);
9563 if (!I) return false;
9565 // Insert element gets simplified to the inserted element or is deleted if
9566 // this is constant idx extract element and its a constant idx insertelt.
9567 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9568 isa<ConstantInt>(I->getOperand(2)))
9570 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9572 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9573 if (BO->hasOneUse() &&
9574 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9575 CheapToScalarize(BO->getOperand(1), isConstant)))
9577 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9578 if (CI->hasOneUse() &&
9579 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9580 CheapToScalarize(CI->getOperand(1), isConstant)))
9586 /// Read and decode a shufflevector mask.
9588 /// It turns undef elements into values that are larger than the number of
9589 /// elements in the input.
9590 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9591 unsigned NElts = SVI->getType()->getNumElements();
9592 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9593 return std::vector<unsigned>(NElts, 0);
9594 if (isa<UndefValue>(SVI->getOperand(2)))
9595 return std::vector<unsigned>(NElts, 2*NElts);
9597 std::vector<unsigned> Result;
9598 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9599 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9600 if (isa<UndefValue>(CP->getOperand(i)))
9601 Result.push_back(NElts*2); // undef -> 8
9603 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9607 /// FindScalarElement - Given a vector and an element number, see if the scalar
9608 /// value is already around as a register, for example if it were inserted then
9609 /// extracted from the vector.
9610 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9611 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9612 const VectorType *PTy = cast<VectorType>(V->getType());
9613 unsigned Width = PTy->getNumElements();
9614 if (EltNo >= Width) // Out of range access.
9615 return UndefValue::get(PTy->getElementType());
9617 if (isa<UndefValue>(V))
9618 return UndefValue::get(PTy->getElementType());
9619 else if (isa<ConstantAggregateZero>(V))
9620 return Constant::getNullValue(PTy->getElementType());
9621 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9622 return CP->getOperand(EltNo);
9623 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9624 // If this is an insert to a variable element, we don't know what it is.
9625 if (!isa<ConstantInt>(III->getOperand(2)))
9627 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9629 // If this is an insert to the element we are looking for, return the
9632 return III->getOperand(1);
9634 // Otherwise, the insertelement doesn't modify the value, recurse on its
9636 return FindScalarElement(III->getOperand(0), EltNo);
9637 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9638 unsigned InEl = getShuffleMask(SVI)[EltNo];
9640 return FindScalarElement(SVI->getOperand(0), InEl);
9641 else if (InEl < Width*2)
9642 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9644 return UndefValue::get(PTy->getElementType());
9647 // Otherwise, we don't know.
9651 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9653 // If vector val is undef, replace extract with scalar undef.
9654 if (isa<UndefValue>(EI.getOperand(0)))
9655 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9657 // If vector val is constant 0, replace extract with scalar 0.
9658 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9659 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9661 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9662 // If vector val is constant with uniform operands, replace EI
9663 // with that operand
9664 Constant *op0 = C->getOperand(0);
9665 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9666 if (C->getOperand(i) != op0) {
9671 return ReplaceInstUsesWith(EI, op0);
9674 // If extracting a specified index from the vector, see if we can recursively
9675 // find a previously computed scalar that was inserted into the vector.
9676 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9677 unsigned IndexVal = IdxC->getZExtValue();
9678 unsigned VectorWidth =
9679 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9681 // If this is extracting an invalid index, turn this into undef, to avoid
9682 // crashing the code below.
9683 if (IndexVal >= VectorWidth)
9684 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9686 // This instruction only demands the single element from the input vector.
9687 // If the input vector has a single use, simplify it based on this use
9689 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9691 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9694 EI.setOperand(0, V);
9699 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9700 return ReplaceInstUsesWith(EI, Elt);
9702 // If the this extractelement is directly using a bitcast from a vector of
9703 // the same number of elements, see if we can find the source element from
9704 // it. In this case, we will end up needing to bitcast the scalars.
9705 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9706 if (const VectorType *VT =
9707 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9708 if (VT->getNumElements() == VectorWidth)
9709 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9710 return new BitCastInst(Elt, EI.getType());
9714 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9715 if (I->hasOneUse()) {
9716 // Push extractelement into predecessor operation if legal and
9717 // profitable to do so
9718 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9719 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9720 if (CheapToScalarize(BO, isConstantElt)) {
9721 ExtractElementInst *newEI0 =
9722 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9723 EI.getName()+".lhs");
9724 ExtractElementInst *newEI1 =
9725 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9726 EI.getName()+".rhs");
9727 InsertNewInstBefore(newEI0, EI);
9728 InsertNewInstBefore(newEI1, EI);
9729 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9731 } else if (isa<LoadInst>(I)) {
9732 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9733 PointerType::get(EI.getType()), EI);
9734 GetElementPtrInst *GEP =
9735 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9736 InsertNewInstBefore(GEP, EI);
9737 return new LoadInst(GEP);
9740 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9741 // Extracting the inserted element?
9742 if (IE->getOperand(2) == EI.getOperand(1))
9743 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9744 // If the inserted and extracted elements are constants, they must not
9745 // be the same value, extract from the pre-inserted value instead.
9746 if (isa<Constant>(IE->getOperand(2)) &&
9747 isa<Constant>(EI.getOperand(1))) {
9748 AddUsesToWorkList(EI);
9749 EI.setOperand(0, IE->getOperand(0));
9752 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9753 // If this is extracting an element from a shufflevector, figure out where
9754 // it came from and extract from the appropriate input element instead.
9755 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9756 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9758 if (SrcIdx < SVI->getType()->getNumElements())
9759 Src = SVI->getOperand(0);
9760 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9761 SrcIdx -= SVI->getType()->getNumElements();
9762 Src = SVI->getOperand(1);
9764 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9766 return new ExtractElementInst(Src, SrcIdx);
9773 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9774 /// elements from either LHS or RHS, return the shuffle mask and true.
9775 /// Otherwise, return false.
9776 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9777 std::vector<Constant*> &Mask) {
9778 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9779 "Invalid CollectSingleShuffleElements");
9780 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9782 if (isa<UndefValue>(V)) {
9783 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9785 } else if (V == LHS) {
9786 for (unsigned i = 0; i != NumElts; ++i)
9787 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9789 } else if (V == RHS) {
9790 for (unsigned i = 0; i != NumElts; ++i)
9791 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9793 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9794 // If this is an insert of an extract from some other vector, include it.
9795 Value *VecOp = IEI->getOperand(0);
9796 Value *ScalarOp = IEI->getOperand(1);
9797 Value *IdxOp = IEI->getOperand(2);
9799 if (!isa<ConstantInt>(IdxOp))
9801 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9803 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9804 // Okay, we can handle this if the vector we are insertinting into is
9806 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9807 // If so, update the mask to reflect the inserted undef.
9808 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9811 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9812 if (isa<ConstantInt>(EI->getOperand(1)) &&
9813 EI->getOperand(0)->getType() == V->getType()) {
9814 unsigned ExtractedIdx =
9815 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9817 // This must be extracting from either LHS or RHS.
9818 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9819 // Okay, we can handle this if the vector we are insertinting into is
9821 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9822 // If so, update the mask to reflect the inserted value.
9823 if (EI->getOperand(0) == LHS) {
9824 Mask[InsertedIdx & (NumElts-1)] =
9825 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9827 assert(EI->getOperand(0) == RHS);
9828 Mask[InsertedIdx & (NumElts-1)] =
9829 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9838 // TODO: Handle shufflevector here!
9843 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9844 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9845 /// that computes V and the LHS value of the shuffle.
9846 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9848 assert(isa<VectorType>(V->getType()) &&
9849 (RHS == 0 || V->getType() == RHS->getType()) &&
9850 "Invalid shuffle!");
9851 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9853 if (isa<UndefValue>(V)) {
9854 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9856 } else if (isa<ConstantAggregateZero>(V)) {
9857 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9859 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9860 // If this is an insert of an extract from some other vector, include it.
9861 Value *VecOp = IEI->getOperand(0);
9862 Value *ScalarOp = IEI->getOperand(1);
9863 Value *IdxOp = IEI->getOperand(2);
9865 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9866 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9867 EI->getOperand(0)->getType() == V->getType()) {
9868 unsigned ExtractedIdx =
9869 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9870 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9872 // Either the extracted from or inserted into vector must be RHSVec,
9873 // otherwise we'd end up with a shuffle of three inputs.
9874 if (EI->getOperand(0) == RHS || RHS == 0) {
9875 RHS = EI->getOperand(0);
9876 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9877 Mask[InsertedIdx & (NumElts-1)] =
9878 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9883 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9884 // Everything but the extracted element is replaced with the RHS.
9885 for (unsigned i = 0; i != NumElts; ++i) {
9886 if (i != InsertedIdx)
9887 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9892 // If this insertelement is a chain that comes from exactly these two
9893 // vectors, return the vector and the effective shuffle.
9894 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9895 return EI->getOperand(0);
9900 // TODO: Handle shufflevector here!
9902 // Otherwise, can't do anything fancy. Return an identity vector.
9903 for (unsigned i = 0; i != NumElts; ++i)
9904 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9908 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9909 Value *VecOp = IE.getOperand(0);
9910 Value *ScalarOp = IE.getOperand(1);
9911 Value *IdxOp = IE.getOperand(2);
9913 // Inserting an undef or into an undefined place, remove this.
9914 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9915 ReplaceInstUsesWith(IE, VecOp);
9917 // If the inserted element was extracted from some other vector, and if the
9918 // indexes are constant, try to turn this into a shufflevector operation.
9919 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9920 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9921 EI->getOperand(0)->getType() == IE.getType()) {
9922 unsigned NumVectorElts = IE.getType()->getNumElements();
9923 unsigned ExtractedIdx =
9924 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9925 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9927 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9928 return ReplaceInstUsesWith(IE, VecOp);
9930 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9931 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9933 // If we are extracting a value from a vector, then inserting it right
9934 // back into the same place, just use the input vector.
9935 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9936 return ReplaceInstUsesWith(IE, VecOp);
9938 // We could theoretically do this for ANY input. However, doing so could
9939 // turn chains of insertelement instructions into a chain of shufflevector
9940 // instructions, and right now we do not merge shufflevectors. As such,
9941 // only do this in a situation where it is clear that there is benefit.
9942 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9943 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9944 // the values of VecOp, except then one read from EIOp0.
9945 // Build a new shuffle mask.
9946 std::vector<Constant*> Mask;
9947 if (isa<UndefValue>(VecOp))
9948 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9950 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9951 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9954 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9955 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9956 ConstantVector::get(Mask));
9959 // If this insertelement isn't used by some other insertelement, turn it
9960 // (and any insertelements it points to), into one big shuffle.
9961 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9962 std::vector<Constant*> Mask;
9964 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9965 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9966 // We now have a shuffle of LHS, RHS, Mask.
9967 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9976 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9977 Value *LHS = SVI.getOperand(0);
9978 Value *RHS = SVI.getOperand(1);
9979 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9981 bool MadeChange = false;
9983 // Undefined shuffle mask -> undefined value.
9984 if (isa<UndefValue>(SVI.getOperand(2)))
9985 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9987 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9988 // the undef, change them to undefs.
9989 if (isa<UndefValue>(SVI.getOperand(1))) {
9990 // Scan to see if there are any references to the RHS. If so, replace them
9991 // with undef element refs and set MadeChange to true.
9992 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9993 if (Mask[i] >= e && Mask[i] != 2*e) {
10000 // Remap any references to RHS to use LHS.
10001 std::vector<Constant*> Elts;
10002 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10003 if (Mask[i] == 2*e)
10004 Elts.push_back(UndefValue::get(Type::Int32Ty));
10006 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10008 SVI.setOperand(2, ConstantVector::get(Elts));
10012 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10013 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10014 if (LHS == RHS || isa<UndefValue>(LHS)) {
10015 if (isa<UndefValue>(LHS) && LHS == RHS) {
10016 // shuffle(undef,undef,mask) -> undef.
10017 return ReplaceInstUsesWith(SVI, LHS);
10020 // Remap any references to RHS to use LHS.
10021 std::vector<Constant*> Elts;
10022 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10023 if (Mask[i] >= 2*e)
10024 Elts.push_back(UndefValue::get(Type::Int32Ty));
10026 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10027 (Mask[i] < e && isa<UndefValue>(LHS)))
10028 Mask[i] = 2*e; // Turn into undef.
10030 Mask[i] &= (e-1); // Force to LHS.
10031 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10034 SVI.setOperand(0, SVI.getOperand(1));
10035 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10036 SVI.setOperand(2, ConstantVector::get(Elts));
10037 LHS = SVI.getOperand(0);
10038 RHS = SVI.getOperand(1);
10042 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10043 bool isLHSID = true, isRHSID = true;
10045 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10046 if (Mask[i] >= e*2) continue; // Ignore undef values.
10047 // Is this an identity shuffle of the LHS value?
10048 isLHSID &= (Mask[i] == i);
10050 // Is this an identity shuffle of the RHS value?
10051 isRHSID &= (Mask[i]-e == i);
10054 // Eliminate identity shuffles.
10055 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10056 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10058 // If the LHS is a shufflevector itself, see if we can combine it with this
10059 // one without producing an unusual shuffle. Here we are really conservative:
10060 // we are absolutely afraid of producing a shuffle mask not in the input
10061 // program, because the code gen may not be smart enough to turn a merged
10062 // shuffle into two specific shuffles: it may produce worse code. As such,
10063 // we only merge two shuffles if the result is one of the two input shuffle
10064 // masks. In this case, merging the shuffles just removes one instruction,
10065 // which we know is safe. This is good for things like turning:
10066 // (splat(splat)) -> splat.
10067 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10068 if (isa<UndefValue>(RHS)) {
10069 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10071 std::vector<unsigned> NewMask;
10072 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10073 if (Mask[i] >= 2*e)
10074 NewMask.push_back(2*e);
10076 NewMask.push_back(LHSMask[Mask[i]]);
10078 // If the result mask is equal to the src shuffle or this shuffle mask, do
10079 // the replacement.
10080 if (NewMask == LHSMask || NewMask == Mask) {
10081 std::vector<Constant*> Elts;
10082 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10083 if (NewMask[i] >= e*2) {
10084 Elts.push_back(UndefValue::get(Type::Int32Ty));
10086 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10089 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10090 LHSSVI->getOperand(1),
10091 ConstantVector::get(Elts));
10096 return MadeChange ? &SVI : 0;
10102 /// TryToSinkInstruction - Try to move the specified instruction from its
10103 /// current block into the beginning of DestBlock, which can only happen if it's
10104 /// safe to move the instruction past all of the instructions between it and the
10105 /// end of its block.
10106 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10107 assert(I->hasOneUse() && "Invariants didn't hold!");
10109 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10110 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10112 // Do not sink alloca instructions out of the entry block.
10113 if (isa<AllocaInst>(I) && I->getParent() ==
10114 &DestBlock->getParent()->getEntryBlock())
10117 // We can only sink load instructions if there is nothing between the load and
10118 // the end of block that could change the value.
10119 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10120 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10122 if (Scan->mayWriteToMemory())
10126 BasicBlock::iterator InsertPos = DestBlock->begin();
10127 while (isa<PHINode>(InsertPos)) ++InsertPos;
10129 I->moveBefore(InsertPos);
10135 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10136 /// all reachable code to the worklist.
10138 /// This has a couple of tricks to make the code faster and more powerful. In
10139 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10140 /// them to the worklist (this significantly speeds up instcombine on code where
10141 /// many instructions are dead or constant). Additionally, if we find a branch
10142 /// whose condition is a known constant, we only visit the reachable successors.
10144 static void AddReachableCodeToWorklist(BasicBlock *BB,
10145 SmallPtrSet<BasicBlock*, 64> &Visited,
10147 const TargetData *TD) {
10148 std::vector<BasicBlock*> Worklist;
10149 Worklist.push_back(BB);
10151 while (!Worklist.empty()) {
10152 BB = Worklist.back();
10153 Worklist.pop_back();
10155 // We have now visited this block! If we've already been here, ignore it.
10156 if (!Visited.insert(BB)) continue;
10158 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10159 Instruction *Inst = BBI++;
10161 // DCE instruction if trivially dead.
10162 if (isInstructionTriviallyDead(Inst)) {
10164 DOUT << "IC: DCE: " << *Inst;
10165 Inst->eraseFromParent();
10169 // ConstantProp instruction if trivially constant.
10170 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10171 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10172 Inst->replaceAllUsesWith(C);
10174 Inst->eraseFromParent();
10178 IC.AddToWorkList(Inst);
10181 // Recursively visit successors. If this is a branch or switch on a
10182 // constant, only visit the reachable successor.
10183 TerminatorInst *TI = BB->getTerminator();
10184 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10185 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10186 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10187 Worklist.push_back(BI->getSuccessor(!CondVal));
10190 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10191 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10192 // See if this is an explicit destination.
10193 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10194 if (SI->getCaseValue(i) == Cond) {
10195 Worklist.push_back(SI->getSuccessor(i));
10199 // Otherwise it is the default destination.
10200 Worklist.push_back(SI->getSuccessor(0));
10205 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10206 Worklist.push_back(TI->getSuccessor(i));
10210 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10211 bool Changed = false;
10212 TD = &getAnalysis<TargetData>();
10214 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10215 << F.getNameStr() << "\n");
10218 // Do a depth-first traversal of the function, populate the worklist with
10219 // the reachable instructions. Ignore blocks that are not reachable. Keep
10220 // track of which blocks we visit.
10221 SmallPtrSet<BasicBlock*, 64> Visited;
10222 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10224 // Do a quick scan over the function. If we find any blocks that are
10225 // unreachable, remove any instructions inside of them. This prevents
10226 // the instcombine code from having to deal with some bad special cases.
10227 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10228 if (!Visited.count(BB)) {
10229 Instruction *Term = BB->getTerminator();
10230 while (Term != BB->begin()) { // Remove instrs bottom-up
10231 BasicBlock::iterator I = Term; --I;
10233 DOUT << "IC: DCE: " << *I;
10236 if (!I->use_empty())
10237 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10238 I->eraseFromParent();
10243 while (!Worklist.empty()) {
10244 Instruction *I = RemoveOneFromWorkList();
10245 if (I == 0) continue; // skip null values.
10247 // Check to see if we can DCE the instruction.
10248 if (isInstructionTriviallyDead(I)) {
10249 // Add operands to the worklist.
10250 if (I->getNumOperands() < 4)
10251 AddUsesToWorkList(*I);
10254 DOUT << "IC: DCE: " << *I;
10256 I->eraseFromParent();
10257 RemoveFromWorkList(I);
10261 // Instruction isn't dead, see if we can constant propagate it.
10262 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10263 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10265 // Add operands to the worklist.
10266 AddUsesToWorkList(*I);
10267 ReplaceInstUsesWith(*I, C);
10270 I->eraseFromParent();
10271 RemoveFromWorkList(I);
10275 // See if we can trivially sink this instruction to a successor basic block.
10276 if (I->hasOneUse()) {
10277 BasicBlock *BB = I->getParent();
10278 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10279 if (UserParent != BB) {
10280 bool UserIsSuccessor = false;
10281 // See if the user is one of our successors.
10282 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10283 if (*SI == UserParent) {
10284 UserIsSuccessor = true;
10288 // If the user is one of our immediate successors, and if that successor
10289 // only has us as a predecessors (we'd have to split the critical edge
10290 // otherwise), we can keep going.
10291 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10292 next(pred_begin(UserParent)) == pred_end(UserParent))
10293 // Okay, the CFG is simple enough, try to sink this instruction.
10294 Changed |= TryToSinkInstruction(I, UserParent);
10298 // Now that we have an instruction, try combining it to simplify it...
10302 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10303 if (Instruction *Result = visit(*I)) {
10305 // Should we replace the old instruction with a new one?
10307 DOUT << "IC: Old = " << *I
10308 << " New = " << *Result;
10310 // Everything uses the new instruction now.
10311 I->replaceAllUsesWith(Result);
10313 // Push the new instruction and any users onto the worklist.
10314 AddToWorkList(Result);
10315 AddUsersToWorkList(*Result);
10317 // Move the name to the new instruction first.
10318 Result->takeName(I);
10320 // Insert the new instruction into the basic block...
10321 BasicBlock *InstParent = I->getParent();
10322 BasicBlock::iterator InsertPos = I;
10324 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10325 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10328 InstParent->getInstList().insert(InsertPos, Result);
10330 // Make sure that we reprocess all operands now that we reduced their
10332 AddUsesToWorkList(*I);
10334 // Instructions can end up on the worklist more than once. Make sure
10335 // we do not process an instruction that has been deleted.
10336 RemoveFromWorkList(I);
10338 // Erase the old instruction.
10339 InstParent->getInstList().erase(I);
10342 DOUT << "IC: Mod = " << OrigI
10343 << " New = " << *I;
10346 // If the instruction was modified, it's possible that it is now dead.
10347 // if so, remove it.
10348 if (isInstructionTriviallyDead(I)) {
10349 // Make sure we process all operands now that we are reducing their
10351 AddUsesToWorkList(*I);
10353 // Instructions may end up in the worklist more than once. Erase all
10354 // occurrences of this instruction.
10355 RemoveFromWorkList(I);
10356 I->eraseFromParent();
10359 AddUsersToWorkList(*I);
10366 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10368 // Do an explicit clear, this shrinks the map if needed.
10369 WorklistMap.clear();
10374 bool InstCombiner::runOnFunction(Function &F) {
10375 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10377 bool EverMadeChange = false;
10379 // Iterate while there is work to do.
10380 unsigned Iteration = 0;
10381 while (DoOneIteration(F, Iteration++))
10382 EverMadeChange = true;
10383 return EverMadeChange;
10386 FunctionPass *llvm::createInstructionCombiningPass() {
10387 return new InstCombiner();