1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
62 using namespace llvm::PatternMatch;
64 STATISTIC(NumCombined , "Number of insts combined");
65 STATISTIC(NumConstProp, "Number of constant folds");
66 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
67 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 class VISIBILITY_HIDDEN InstCombiner
72 : public FunctionPass,
73 public InstVisitor<InstCombiner, Instruction*> {
74 // Worklist of all of the instructions that need to be simplified.
75 std::vector<Instruction*> Worklist;
76 DenseMap<Instruction*, unsigned> WorklistMap;
78 bool MustPreserveLCSSA;
80 static char ID; // Pass identification, replacement for typeid
81 InstCombiner() : FunctionPass((intptr_t)&ID) {}
83 /// AddToWorkList - Add the specified instruction to the worklist if it
84 /// isn't already in it.
85 void AddToWorkList(Instruction *I) {
86 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
87 Worklist.push_back(I);
90 // RemoveFromWorkList - remove I from the worklist if it exists.
91 void RemoveFromWorkList(Instruction *I) {
92 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
93 if (It == WorklistMap.end()) return; // Not in worklist.
95 // Don't bother moving everything down, just null out the slot.
96 Worklist[It->second] = 0;
98 WorklistMap.erase(It);
101 Instruction *RemoveOneFromWorkList() {
102 Instruction *I = Worklist.back();
104 WorklistMap.erase(I);
109 /// AddUsersToWorkList - When an instruction is simplified, add all users of
110 /// the instruction to the work lists because they might get more simplified
113 void AddUsersToWorkList(Value &I) {
114 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
116 AddToWorkList(cast<Instruction>(*UI));
119 /// AddUsesToWorkList - When an instruction is simplified, add operands to
120 /// the work lists because they might get more simplified now.
122 void AddUsesToWorkList(Instruction &I) {
123 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
124 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
128 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
129 /// dead. Add all of its operands to the worklist, turning them into
130 /// undef's to reduce the number of uses of those instructions.
132 /// Return the specified operand before it is turned into an undef.
134 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
135 Value *R = I.getOperand(op);
137 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
138 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
140 // Set the operand to undef to drop the use.
141 I.setOperand(i, UndefValue::get(Op->getType()));
148 virtual bool runOnFunction(Function &F);
150 bool DoOneIteration(Function &F, unsigned ItNum);
152 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
153 AU.addRequired<TargetData>();
154 AU.addPreservedID(LCSSAID);
155 AU.setPreservesCFG();
158 TargetData &getTargetData() const { return *TD; }
160 // Visitation implementation - Implement instruction combining for different
161 // instruction types. The semantics are as follows:
163 // null - No change was made
164 // I - Change was made, I is still valid, I may be dead though
165 // otherwise - Change was made, replace I with returned instruction
167 Instruction *visitAdd(BinaryOperator &I);
168 Instruction *visitSub(BinaryOperator &I);
169 Instruction *visitMul(BinaryOperator &I);
170 Instruction *visitURem(BinaryOperator &I);
171 Instruction *visitSRem(BinaryOperator &I);
172 Instruction *visitFRem(BinaryOperator &I);
173 Instruction *commonRemTransforms(BinaryOperator &I);
174 Instruction *commonIRemTransforms(BinaryOperator &I);
175 Instruction *commonDivTransforms(BinaryOperator &I);
176 Instruction *commonIDivTransforms(BinaryOperator &I);
177 Instruction *visitUDiv(BinaryOperator &I);
178 Instruction *visitSDiv(BinaryOperator &I);
179 Instruction *visitFDiv(BinaryOperator &I);
180 Instruction *visitAnd(BinaryOperator &I);
181 Instruction *visitOr (BinaryOperator &I);
182 Instruction *visitXor(BinaryOperator &I);
183 Instruction *visitShl(BinaryOperator &I);
184 Instruction *visitAShr(BinaryOperator &I);
185 Instruction *visitLShr(BinaryOperator &I);
186 Instruction *commonShiftTransforms(BinaryOperator &I);
187 Instruction *visitFCmpInst(FCmpInst &I);
188 Instruction *visitICmpInst(ICmpInst &I);
189 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
190 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
193 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
194 ConstantInt *DivRHS);
196 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
197 ICmpInst::Predicate Cond, Instruction &I);
198 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
200 Instruction *commonCastTransforms(CastInst &CI);
201 Instruction *commonIntCastTransforms(CastInst &CI);
202 Instruction *commonPointerCastTransforms(CastInst &CI);
203 Instruction *visitTrunc(TruncInst &CI);
204 Instruction *visitZExt(ZExtInst &CI);
205 Instruction *visitSExt(SExtInst &CI);
206 Instruction *visitFPTrunc(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) --> intptrtype
2114 CastInst *CI = dyn_cast<CastInst>(LHS);
2117 CI = dyn_cast<CastInst>(RHS);
2120 if (CI && CI->getType()->isSized() &&
2121 (CI->getType()->getPrimitiveSizeInBits() ==
2122 TD->getIntPtrType()->getPrimitiveSizeInBits())
2123 && isa<PointerType>(CI->getOperand(0)->getType())) {
2125 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2126 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2127 PointerType::get(Type::Int8Ty, AS), I);
2128 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2129 return new PtrToIntInst(I2, CI->getType());
2133 // add (select X 0 (sub n A)) A --> select X A n
2135 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2138 SI = dyn_cast<SelectInst>(RHS);
2141 if (SI && SI->hasOneUse()) {
2142 Value *TV = SI->getTrueValue();
2143 Value *FV = SI->getFalseValue();
2146 // Can we fold the add into the argument of the select?
2147 // We check both true and false select arguments for a matching subtract.
2148 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2149 A == Other) // Fold the add into the true select value.
2150 return new SelectInst(SI->getCondition(), N, A);
2151 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2152 A == Other) // Fold the add into the false select value.
2153 return new SelectInst(SI->getCondition(), A, N);
2157 return Changed ? &I : 0;
2160 // isSignBit - Return true if the value represented by the constant only has the
2161 // highest order bit set.
2162 static bool isSignBit(ConstantInt *CI) {
2163 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2164 return CI->getValue() == APInt::getSignBit(NumBits);
2167 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2168 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2170 if (Op0 == Op1) // sub X, X -> 0
2171 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2173 // If this is a 'B = x-(-A)', change to B = x+A...
2174 if (Value *V = dyn_castNegVal(Op1))
2175 return BinaryOperator::createAdd(Op0, V);
2177 if (isa<UndefValue>(Op0))
2178 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2179 if (isa<UndefValue>(Op1))
2180 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2182 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2183 // Replace (-1 - A) with (~A)...
2184 if (C->isAllOnesValue())
2185 return BinaryOperator::createNot(Op1);
2187 // C - ~X == X + (1+C)
2189 if (match(Op1, m_Not(m_Value(X))))
2190 return BinaryOperator::createAdd(X, AddOne(C));
2192 // -(X >>u 31) -> (X >>s 31)
2193 // -(X >>s 31) -> (X >>u 31)
2195 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2196 if (SI->getOpcode() == Instruction::LShr) {
2197 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2198 // Check to see if we are shifting out everything but the sign bit.
2199 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2200 SI->getType()->getPrimitiveSizeInBits()-1) {
2201 // Ok, the transformation is safe. Insert AShr.
2202 return BinaryOperator::create(Instruction::AShr,
2203 SI->getOperand(0), CU, SI->getName());
2207 else if (SI->getOpcode() == Instruction::AShr) {
2208 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2209 // Check to see if we are shifting out everything but the sign bit.
2210 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2211 SI->getType()->getPrimitiveSizeInBits()-1) {
2212 // Ok, the transformation is safe. Insert LShr.
2213 return BinaryOperator::createLShr(
2214 SI->getOperand(0), CU, SI->getName());
2220 // Try to fold constant sub into select arguments.
2221 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2222 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2225 if (isa<PHINode>(Op0))
2226 if (Instruction *NV = FoldOpIntoPhi(I))
2230 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2231 if (Op1I->getOpcode() == Instruction::Add &&
2232 !Op0->getType()->isFPOrFPVector()) {
2233 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2234 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2235 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2236 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2237 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2238 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2239 // C1-(X+C2) --> (C1-C2)-X
2240 return BinaryOperator::createSub(Subtract(CI1, CI2),
2241 Op1I->getOperand(0));
2245 if (Op1I->hasOneUse()) {
2246 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2247 // is not used by anyone else...
2249 if (Op1I->getOpcode() == Instruction::Sub &&
2250 !Op1I->getType()->isFPOrFPVector()) {
2251 // Swap the two operands of the subexpr...
2252 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2253 Op1I->setOperand(0, IIOp1);
2254 Op1I->setOperand(1, IIOp0);
2256 // Create the new top level add instruction...
2257 return BinaryOperator::createAdd(Op0, Op1);
2260 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2262 if (Op1I->getOpcode() == Instruction::And &&
2263 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2264 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2267 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2268 return BinaryOperator::createAnd(Op0, NewNot);
2271 // 0 - (X sdiv C) -> (X sdiv -C)
2272 if (Op1I->getOpcode() == Instruction::SDiv)
2273 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2275 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2276 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2277 ConstantExpr::getNeg(DivRHS));
2279 // X - X*C --> X * (1-C)
2280 ConstantInt *C2 = 0;
2281 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2282 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2283 return BinaryOperator::createMul(Op0, CP1);
2286 // X - ((X / Y) * Y) --> X % Y
2287 if (Op1I->getOpcode() == Instruction::Mul)
2288 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2289 if (Op0 == I->getOperand(0) &&
2290 Op1I->getOperand(1) == I->getOperand(1)) {
2291 if (I->getOpcode() == Instruction::SDiv)
2292 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2293 if (I->getOpcode() == Instruction::UDiv)
2294 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2299 if (!Op0->getType()->isFPOrFPVector())
2300 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2301 if (Op0I->getOpcode() == Instruction::Add) {
2302 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2303 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2304 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2305 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2306 } else if (Op0I->getOpcode() == Instruction::Sub) {
2307 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2308 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2312 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2313 if (X == Op1) // X*C - X --> X * (C-1)
2314 return BinaryOperator::createMul(Op1, SubOne(C1));
2316 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2317 if (X == dyn_castFoldableMul(Op1, C2))
2318 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2323 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2324 /// comparison only checks the sign bit. If it only checks the sign bit, set
2325 /// TrueIfSigned if the result of the comparison is true when the input value is
2327 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2328 bool &TrueIfSigned) {
2330 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2331 TrueIfSigned = true;
2332 return RHS->isZero();
2333 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2334 TrueIfSigned = true;
2335 return RHS->isAllOnesValue();
2336 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2337 TrueIfSigned = false;
2338 return RHS->isAllOnesValue();
2339 case ICmpInst::ICMP_UGT:
2340 // True if LHS u> RHS and RHS == high-bit-mask - 1
2341 TrueIfSigned = true;
2342 return RHS->getValue() ==
2343 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2344 case ICmpInst::ICMP_UGE:
2345 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2346 TrueIfSigned = true;
2347 return RHS->getValue() ==
2348 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2354 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2355 bool Changed = SimplifyCommutative(I);
2356 Value *Op0 = I.getOperand(0);
2358 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2359 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2361 // Simplify mul instructions with a constant RHS...
2362 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2363 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2365 // ((X << C1)*C2) == (X * (C2 << C1))
2366 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2367 if (SI->getOpcode() == Instruction::Shl)
2368 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2369 return BinaryOperator::createMul(SI->getOperand(0),
2370 ConstantExpr::getShl(CI, ShOp));
2373 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2374 if (CI->equalsInt(1)) // X * 1 == X
2375 return ReplaceInstUsesWith(I, Op0);
2376 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2377 return BinaryOperator::createNeg(Op0, I.getName());
2379 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2380 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2381 return BinaryOperator::createShl(Op0,
2382 ConstantInt::get(Op0->getType(), Val.logBase2()));
2384 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2385 if (Op1F->isNullValue())
2386 return ReplaceInstUsesWith(I, Op1);
2388 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2389 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2390 // We need a better interface for long double here.
2391 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2392 if (Op1F->isExactlyValue(1.0))
2393 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2396 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2397 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2398 isa<ConstantInt>(Op0I->getOperand(1))) {
2399 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2400 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2402 InsertNewInstBefore(Add, I);
2403 Value *C1C2 = ConstantExpr::getMul(Op1,
2404 cast<Constant>(Op0I->getOperand(1)));
2405 return BinaryOperator::createAdd(Add, C1C2);
2409 // Try to fold constant mul into select arguments.
2410 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2411 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2414 if (isa<PHINode>(Op0))
2415 if (Instruction *NV = FoldOpIntoPhi(I))
2419 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2420 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2421 return BinaryOperator::createMul(Op0v, Op1v);
2423 // If one of the operands of the multiply is a cast from a boolean value, then
2424 // we know the bool is either zero or one, so this is a 'masking' multiply.
2425 // See if we can simplify things based on how the boolean was originally
2427 CastInst *BoolCast = 0;
2428 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2429 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2432 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2433 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2436 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2437 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2438 const Type *SCOpTy = SCIOp0->getType();
2441 // If the icmp is true iff the sign bit of X is set, then convert this
2442 // multiply into a shift/and combination.
2443 if (isa<ConstantInt>(SCIOp1) &&
2444 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2446 // Shift the X value right to turn it into "all signbits".
2447 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2448 SCOpTy->getPrimitiveSizeInBits()-1);
2450 InsertNewInstBefore(
2451 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2452 BoolCast->getOperand(0)->getName()+
2455 // If the multiply type is not the same as the source type, sign extend
2456 // or truncate to the multiply type.
2457 if (I.getType() != V->getType()) {
2458 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2459 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2460 Instruction::CastOps opcode =
2461 (SrcBits == DstBits ? Instruction::BitCast :
2462 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2463 V = InsertCastBefore(opcode, V, I.getType(), I);
2466 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2467 return BinaryOperator::createAnd(V, OtherOp);
2472 return Changed ? &I : 0;
2475 /// This function implements the transforms on div instructions that work
2476 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2477 /// used by the visitors to those instructions.
2478 /// @brief Transforms common to all three div instructions
2479 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2480 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2483 if (isa<UndefValue>(Op0))
2484 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2486 // X / undef -> undef
2487 if (isa<UndefValue>(Op1))
2488 return ReplaceInstUsesWith(I, Op1);
2490 // Handle cases involving: div X, (select Cond, Y, Z)
2491 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2492 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2493 // same basic block, then we replace the select with Y, and the condition
2494 // of the select with false (if the cond value is in the same BB). If the
2495 // select has uses other than the div, this allows them to be simplified
2496 // also. Note that div X, Y is just as good as div X, 0 (undef)
2497 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2498 if (ST->isNullValue()) {
2499 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2500 if (CondI && CondI->getParent() == I.getParent())
2501 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2502 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2503 I.setOperand(1, SI->getOperand(2));
2505 UpdateValueUsesWith(SI, SI->getOperand(2));
2509 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2510 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2511 if (ST->isNullValue()) {
2512 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2513 if (CondI && CondI->getParent() == I.getParent())
2514 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2515 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2516 I.setOperand(1, SI->getOperand(1));
2518 UpdateValueUsesWith(SI, SI->getOperand(1));
2526 /// This function implements the transforms common to both integer division
2527 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2528 /// division instructions.
2529 /// @brief Common integer divide transforms
2530 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2531 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2533 if (Instruction *Common = commonDivTransforms(I))
2536 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2538 if (RHS->equalsInt(1))
2539 return ReplaceInstUsesWith(I, Op0);
2541 // (X / C1) / C2 -> X / (C1*C2)
2542 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2543 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2544 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2545 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2546 Multiply(RHS, LHSRHS));
2549 if (!RHS->isZero()) { // avoid X udiv 0
2550 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2551 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2553 if (isa<PHINode>(Op0))
2554 if (Instruction *NV = FoldOpIntoPhi(I))
2559 // 0 / X == 0, we don't need to preserve faults!
2560 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2561 if (LHS->equalsInt(0))
2562 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2567 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2568 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 // Handle the integer div common cases
2571 if (Instruction *Common = commonIDivTransforms(I))
2574 // X udiv C^2 -> X >> C
2575 // Check to see if this is an unsigned division with an exact power of 2,
2576 // if so, convert to a right shift.
2577 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2578 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2579 return BinaryOperator::createLShr(Op0,
2580 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2583 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2584 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2585 if (RHSI->getOpcode() == Instruction::Shl &&
2586 isa<ConstantInt>(RHSI->getOperand(0))) {
2587 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2588 if (C1.isPowerOf2()) {
2589 Value *N = RHSI->getOperand(1);
2590 const Type *NTy = N->getType();
2591 if (uint32_t C2 = C1.logBase2()) {
2592 Constant *C2V = ConstantInt::get(NTy, C2);
2593 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2595 return BinaryOperator::createLShr(Op0, N);
2600 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2601 // where C1&C2 are powers of two.
2602 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2603 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2604 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2605 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2606 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2607 // Compute the shift amounts
2608 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2609 // Construct the "on true" case of the select
2610 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2611 Instruction *TSI = BinaryOperator::createLShr(
2612 Op0, TC, SI->getName()+".t");
2613 TSI = InsertNewInstBefore(TSI, I);
2615 // Construct the "on false" case of the select
2616 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2617 Instruction *FSI = BinaryOperator::createLShr(
2618 Op0, FC, SI->getName()+".f");
2619 FSI = InsertNewInstBefore(FSI, I);
2621 // construct the select instruction and return it.
2622 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2628 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2629 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2631 // Handle the integer div common cases
2632 if (Instruction *Common = commonIDivTransforms(I))
2635 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2637 if (RHS->isAllOnesValue())
2638 return BinaryOperator::createNeg(Op0);
2641 if (Value *LHSNeg = dyn_castNegVal(Op0))
2642 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2645 // If the sign bits of both operands are zero (i.e. we can prove they are
2646 // unsigned inputs), turn this into a udiv.
2647 if (I.getType()->isInteger()) {
2648 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2649 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2650 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2651 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2658 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2659 return commonDivTransforms(I);
2662 /// GetFactor - If we can prove that the specified value is at least a multiple
2663 /// of some factor, return that factor.
2664 static Constant *GetFactor(Value *V) {
2665 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2668 // Unless we can be tricky, we know this is a multiple of 1.
2669 Constant *Result = ConstantInt::get(V->getType(), 1);
2671 Instruction *I = dyn_cast<Instruction>(V);
2672 if (!I) return Result;
2674 if (I->getOpcode() == Instruction::Mul) {
2675 // Handle multiplies by a constant, etc.
2676 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2677 GetFactor(I->getOperand(1)));
2678 } else if (I->getOpcode() == Instruction::Shl) {
2679 // (X<<C) -> X * (1 << C)
2680 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2681 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2682 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2684 } else if (I->getOpcode() == Instruction::And) {
2685 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2686 // X & 0xFFF0 is known to be a multiple of 16.
2687 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2688 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2689 return ConstantExpr::getShl(Result,
2690 ConstantInt::get(Result->getType(), Zeros));
2692 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2693 // Only handle int->int casts.
2694 if (!CI->isIntegerCast())
2696 Value *Op = CI->getOperand(0);
2697 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2702 /// This function implements the transforms on rem instructions that work
2703 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2704 /// is used by the visitors to those instructions.
2705 /// @brief Transforms common to all three rem instructions
2706 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2707 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2709 // 0 % X == 0, we don't need to preserve faults!
2710 if (Constant *LHS = dyn_cast<Constant>(Op0))
2711 if (LHS->isNullValue())
2712 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2714 if (isa<UndefValue>(Op0)) // undef % X -> 0
2715 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2716 if (isa<UndefValue>(Op1))
2717 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2719 // Handle cases involving: rem X, (select Cond, Y, Z)
2720 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2721 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2722 // the same basic block, then we replace the select with Y, and the
2723 // condition of the select with false (if the cond value is in the same
2724 // BB). If the select has uses other than the div, this allows them to be
2726 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2727 if (ST->isNullValue()) {
2728 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2729 if (CondI && CondI->getParent() == I.getParent())
2730 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2731 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2732 I.setOperand(1, SI->getOperand(2));
2734 UpdateValueUsesWith(SI, SI->getOperand(2));
2737 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2738 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2739 if (ST->isNullValue()) {
2740 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2741 if (CondI && CondI->getParent() == I.getParent())
2742 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2743 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2744 I.setOperand(1, SI->getOperand(1));
2746 UpdateValueUsesWith(SI, SI->getOperand(1));
2754 /// This function implements the transforms common to both integer remainder
2755 /// instructions (urem and srem). It is called by the visitors to those integer
2756 /// remainder instructions.
2757 /// @brief Common integer remainder transforms
2758 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2759 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2761 if (Instruction *common = commonRemTransforms(I))
2764 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2765 // X % 0 == undef, we don't need to preserve faults!
2766 if (RHS->equalsInt(0))
2767 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2769 if (RHS->equalsInt(1)) // X % 1 == 0
2770 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2772 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2773 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2774 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2776 } else if (isa<PHINode>(Op0I)) {
2777 if (Instruction *NV = FoldOpIntoPhi(I))
2780 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2781 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2782 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2789 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2790 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2792 if (Instruction *common = commonIRemTransforms(I))
2795 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2796 // X urem C^2 -> X and C
2797 // Check to see if this is an unsigned remainder with an exact power of 2,
2798 // if so, convert to a bitwise and.
2799 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2800 if (C->getValue().isPowerOf2())
2801 return BinaryOperator::createAnd(Op0, SubOne(C));
2804 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2805 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2806 if (RHSI->getOpcode() == Instruction::Shl &&
2807 isa<ConstantInt>(RHSI->getOperand(0))) {
2808 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2809 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2810 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2812 return BinaryOperator::createAnd(Op0, Add);
2817 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2818 // where C1&C2 are powers of two.
2819 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2820 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2821 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2822 // STO == 0 and SFO == 0 handled above.
2823 if ((STO->getValue().isPowerOf2()) &&
2824 (SFO->getValue().isPowerOf2())) {
2825 Value *TrueAnd = InsertNewInstBefore(
2826 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2827 Value *FalseAnd = InsertNewInstBefore(
2828 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2829 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2837 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2838 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2840 // Handle the integer rem common cases
2841 if (Instruction *common = commonIRemTransforms(I))
2844 if (Value *RHSNeg = dyn_castNegVal(Op1))
2845 if (!isa<ConstantInt>(RHSNeg) ||
2846 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2848 AddUsesToWorkList(I);
2849 I.setOperand(1, RHSNeg);
2853 // If the sign bits of both operands are zero (i.e. we can prove they are
2854 // unsigned inputs), turn this into a urem.
2855 if (I.getType()->isInteger()) {
2856 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2857 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2858 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2859 return BinaryOperator::createURem(Op0, Op1, I.getName());
2866 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2867 return commonRemTransforms(I);
2870 // isMaxValueMinusOne - return true if this is Max-1
2871 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2872 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2874 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2875 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2878 // isMinValuePlusOne - return true if this is Min+1
2879 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2881 return C->getValue() == 1; // unsigned
2883 // Calculate 1111111111000000000000
2884 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2885 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2888 // isOneBitSet - Return true if there is exactly one bit set in the specified
2890 static bool isOneBitSet(const ConstantInt *CI) {
2891 return CI->getValue().isPowerOf2();
2894 // isHighOnes - Return true if the constant is of the form 1+0+.
2895 // This is the same as lowones(~X).
2896 static bool isHighOnes(const ConstantInt *CI) {
2897 return (~CI->getValue() + 1).isPowerOf2();
2900 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2901 /// are carefully arranged to allow folding of expressions such as:
2903 /// (A < B) | (A > B) --> (A != B)
2905 /// Note that this is only valid if the first and second predicates have the
2906 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2908 /// Three bits are used to represent the condition, as follows:
2913 /// <=> Value Definition
2914 /// 000 0 Always false
2921 /// 111 7 Always true
2923 static unsigned getICmpCode(const ICmpInst *ICI) {
2924 switch (ICI->getPredicate()) {
2926 case ICmpInst::ICMP_UGT: return 1; // 001
2927 case ICmpInst::ICMP_SGT: return 1; // 001
2928 case ICmpInst::ICMP_EQ: return 2; // 010
2929 case ICmpInst::ICMP_UGE: return 3; // 011
2930 case ICmpInst::ICMP_SGE: return 3; // 011
2931 case ICmpInst::ICMP_ULT: return 4; // 100
2932 case ICmpInst::ICMP_SLT: return 4; // 100
2933 case ICmpInst::ICMP_NE: return 5; // 101
2934 case ICmpInst::ICMP_ULE: return 6; // 110
2935 case ICmpInst::ICMP_SLE: return 6; // 110
2938 assert(0 && "Invalid ICmp predicate!");
2943 /// getICmpValue - This is the complement of getICmpCode, which turns an
2944 /// opcode and two operands into either a constant true or false, or a brand
2945 /// new ICmp instruction. The sign is passed in to determine which kind
2946 /// of predicate to use in new icmp instructions.
2947 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2949 default: assert(0 && "Illegal ICmp code!");
2950 case 0: return ConstantInt::getFalse();
2953 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2955 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2956 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2959 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2961 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2964 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2966 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2967 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2970 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2972 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2973 case 7: return ConstantInt::getTrue();
2977 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2978 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2979 (ICmpInst::isSignedPredicate(p1) &&
2980 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2981 (ICmpInst::isSignedPredicate(p2) &&
2982 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2986 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2987 struct FoldICmpLogical {
2990 ICmpInst::Predicate pred;
2991 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2992 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2993 pred(ICI->getPredicate()) {}
2994 bool shouldApply(Value *V) const {
2995 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2996 if (PredicatesFoldable(pred, ICI->getPredicate()))
2997 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2998 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3001 Instruction *apply(Instruction &Log) const {
3002 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3003 if (ICI->getOperand(0) != LHS) {
3004 assert(ICI->getOperand(1) == LHS);
3005 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3008 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3009 unsigned LHSCode = getICmpCode(ICI);
3010 unsigned RHSCode = getICmpCode(RHSICI);
3012 switch (Log.getOpcode()) {
3013 case Instruction::And: Code = LHSCode & RHSCode; break;
3014 case Instruction::Or: Code = LHSCode | RHSCode; break;
3015 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3016 default: assert(0 && "Illegal logical opcode!"); return 0;
3019 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3020 ICmpInst::isSignedPredicate(ICI->getPredicate());
3022 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3023 if (Instruction *I = dyn_cast<Instruction>(RV))
3025 // Otherwise, it's a constant boolean value...
3026 return IC.ReplaceInstUsesWith(Log, RV);
3029 } // end anonymous namespace
3031 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3032 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3033 // guaranteed to be a binary operator.
3034 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3036 ConstantInt *AndRHS,
3037 BinaryOperator &TheAnd) {
3038 Value *X = Op->getOperand(0);
3039 Constant *Together = 0;
3041 Together = And(AndRHS, OpRHS);
3043 switch (Op->getOpcode()) {
3044 case Instruction::Xor:
3045 if (Op->hasOneUse()) {
3046 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3047 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3048 InsertNewInstBefore(And, TheAnd);
3050 return BinaryOperator::createXor(And, Together);
3053 case Instruction::Or:
3054 if (Together == AndRHS) // (X | C) & C --> C
3055 return ReplaceInstUsesWith(TheAnd, AndRHS);
3057 if (Op->hasOneUse() && Together != OpRHS) {
3058 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3059 Instruction *Or = BinaryOperator::createOr(X, Together);
3060 InsertNewInstBefore(Or, TheAnd);
3062 return BinaryOperator::createAnd(Or, AndRHS);
3065 case Instruction::Add:
3066 if (Op->hasOneUse()) {
3067 // Adding a one to a single bit bit-field should be turned into an XOR
3068 // of the bit. First thing to check is to see if this AND is with a
3069 // single bit constant.
3070 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3072 // If there is only one bit set...
3073 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3074 // Ok, at this point, we know that we are masking the result of the
3075 // ADD down to exactly one bit. If the constant we are adding has
3076 // no bits set below this bit, then we can eliminate the ADD.
3077 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3079 // Check to see if any bits below the one bit set in AndRHSV are set.
3080 if ((AddRHS & (AndRHSV-1)) == 0) {
3081 // If not, the only thing that can effect the output of the AND is
3082 // the bit specified by AndRHSV. If that bit is set, the effect of
3083 // the XOR is to toggle the bit. If it is clear, then the ADD has
3085 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3086 TheAnd.setOperand(0, X);
3089 // Pull the XOR out of the AND.
3090 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3091 InsertNewInstBefore(NewAnd, TheAnd);
3092 NewAnd->takeName(Op);
3093 return BinaryOperator::createXor(NewAnd, AndRHS);
3100 case Instruction::Shl: {
3101 // We know that the AND will not produce any of the bits shifted in, so if
3102 // the anded constant includes them, clear them now!
3104 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3105 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3106 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3107 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3109 if (CI->getValue() == ShlMask) {
3110 // Masking out bits that the shift already masks
3111 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3112 } else if (CI != AndRHS) { // Reducing bits set in and.
3113 TheAnd.setOperand(1, CI);
3118 case Instruction::LShr:
3120 // We know that the AND will not produce any of the bits shifted in, so if
3121 // the anded constant includes them, clear them now! This only applies to
3122 // unsigned shifts, because a signed shr may bring in set bits!
3124 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3125 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3126 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3127 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3129 if (CI->getValue() == ShrMask) {
3130 // Masking out bits that the shift already masks.
3131 return ReplaceInstUsesWith(TheAnd, Op);
3132 } else if (CI != AndRHS) {
3133 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3138 case Instruction::AShr:
3140 // See if this is shifting in some sign extension, then masking it out
3142 if (Op->hasOneUse()) {
3143 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3144 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3145 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3146 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3147 if (C == AndRHS) { // Masking out bits shifted in.
3148 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3149 // Make the argument unsigned.
3150 Value *ShVal = Op->getOperand(0);
3151 ShVal = InsertNewInstBefore(
3152 BinaryOperator::createLShr(ShVal, OpRHS,
3153 Op->getName()), TheAnd);
3154 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3163 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3164 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3165 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3166 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3167 /// insert new instructions.
3168 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3169 bool isSigned, bool Inside,
3171 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3172 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3173 "Lo is not <= Hi in range emission code!");
3176 if (Lo == Hi) // Trivially false.
3177 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3179 // V >= Min && V < Hi --> V < Hi
3180 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3181 ICmpInst::Predicate pred = (isSigned ?
3182 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3183 return new ICmpInst(pred, V, Hi);
3186 // Emit V-Lo <u Hi-Lo
3187 Constant *NegLo = ConstantExpr::getNeg(Lo);
3188 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3189 InsertNewInstBefore(Add, IB);
3190 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3191 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3194 if (Lo == Hi) // Trivially true.
3195 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3197 // V < Min || V >= Hi -> V > Hi-1
3198 Hi = SubOne(cast<ConstantInt>(Hi));
3199 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3200 ICmpInst::Predicate pred = (isSigned ?
3201 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3202 return new ICmpInst(pred, V, Hi);
3205 // Emit V-Lo >u Hi-1-Lo
3206 // Note that Hi has already had one subtracted from it, above.
3207 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3208 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3209 InsertNewInstBefore(Add, IB);
3210 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3211 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3214 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3215 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3216 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3217 // not, since all 1s are not contiguous.
3218 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3219 const APInt& V = Val->getValue();
3220 uint32_t BitWidth = Val->getType()->getBitWidth();
3221 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3223 // look for the first zero bit after the run of ones
3224 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3225 // look for the first non-zero bit
3226 ME = V.getActiveBits();
3230 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3231 /// where isSub determines whether the operator is a sub. If we can fold one of
3232 /// the following xforms:
3234 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3235 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3236 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3238 /// return (A +/- B).
3240 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3241 ConstantInt *Mask, bool isSub,
3243 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3244 if (!LHSI || LHSI->getNumOperands() != 2 ||
3245 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3247 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3249 switch (LHSI->getOpcode()) {
3251 case Instruction::And:
3252 if (And(N, Mask) == Mask) {
3253 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3254 if ((Mask->getValue().countLeadingZeros() +
3255 Mask->getValue().countPopulation()) ==
3256 Mask->getValue().getBitWidth())
3259 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3260 // part, we don't need any explicit masks to take them out of A. If that
3261 // is all N is, ignore it.
3262 uint32_t MB = 0, ME = 0;
3263 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3264 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3265 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3266 if (MaskedValueIsZero(RHS, Mask))
3271 case Instruction::Or:
3272 case Instruction::Xor:
3273 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3274 if ((Mask->getValue().countLeadingZeros() +
3275 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3276 && And(N, Mask)->isZero())
3283 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3285 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3286 return InsertNewInstBefore(New, I);
3289 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3290 bool Changed = SimplifyCommutative(I);
3291 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3293 if (isa<UndefValue>(Op1)) // X & undef -> 0
3294 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3298 return ReplaceInstUsesWith(I, Op1);
3300 // See if we can simplify any instructions used by the instruction whose sole
3301 // purpose is to compute bits we don't care about.
3302 if (!isa<VectorType>(I.getType())) {
3303 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3304 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3305 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3306 KnownZero, KnownOne))
3309 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3310 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3311 return ReplaceInstUsesWith(I, I.getOperand(0));
3312 } else if (isa<ConstantAggregateZero>(Op1)) {
3313 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3317 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3318 const APInt& AndRHSMask = AndRHS->getValue();
3319 APInt NotAndRHS(~AndRHSMask);
3321 // Optimize a variety of ((val OP C1) & C2) combinations...
3322 if (isa<BinaryOperator>(Op0)) {
3323 Instruction *Op0I = cast<Instruction>(Op0);
3324 Value *Op0LHS = Op0I->getOperand(0);
3325 Value *Op0RHS = Op0I->getOperand(1);
3326 switch (Op0I->getOpcode()) {
3327 case Instruction::Xor:
3328 case Instruction::Or:
3329 // If the mask is only needed on one incoming arm, push it up.
3330 if (Op0I->hasOneUse()) {
3331 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3332 // Not masking anything out for the LHS, move to RHS.
3333 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3334 Op0RHS->getName()+".masked");
3335 InsertNewInstBefore(NewRHS, I);
3336 return BinaryOperator::create(
3337 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3339 if (!isa<Constant>(Op0RHS) &&
3340 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3341 // Not masking anything out for the RHS, move to LHS.
3342 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3343 Op0LHS->getName()+".masked");
3344 InsertNewInstBefore(NewLHS, I);
3345 return BinaryOperator::create(
3346 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3351 case Instruction::Add:
3352 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3353 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3354 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3355 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3356 return BinaryOperator::createAnd(V, AndRHS);
3357 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3358 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3361 case Instruction::Sub:
3362 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3363 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3364 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3365 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3366 return BinaryOperator::createAnd(V, AndRHS);
3370 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3371 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3373 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3374 // If this is an integer truncation or change from signed-to-unsigned, and
3375 // if the source is an and/or with immediate, transform it. This
3376 // frequently occurs for bitfield accesses.
3377 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3378 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3379 CastOp->getNumOperands() == 2)
3380 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3381 if (CastOp->getOpcode() == Instruction::And) {
3382 // Change: and (cast (and X, C1) to T), C2
3383 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3384 // This will fold the two constants together, which may allow
3385 // other simplifications.
3386 Instruction *NewCast = CastInst::createTruncOrBitCast(
3387 CastOp->getOperand(0), I.getType(),
3388 CastOp->getName()+".shrunk");
3389 NewCast = InsertNewInstBefore(NewCast, I);
3390 // trunc_or_bitcast(C1)&C2
3391 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3392 C3 = ConstantExpr::getAnd(C3, AndRHS);
3393 return BinaryOperator::createAnd(NewCast, C3);
3394 } else if (CastOp->getOpcode() == Instruction::Or) {
3395 // Change: and (cast (or X, C1) to T), C2
3396 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3397 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3398 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3399 return ReplaceInstUsesWith(I, AndRHS);
3404 // Try to fold constant and into select arguments.
3405 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3406 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3408 if (isa<PHINode>(Op0))
3409 if (Instruction *NV = FoldOpIntoPhi(I))
3413 Value *Op0NotVal = dyn_castNotVal(Op0);
3414 Value *Op1NotVal = dyn_castNotVal(Op1);
3416 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3417 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3419 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3420 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3421 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3422 I.getName()+".demorgan");
3423 InsertNewInstBefore(Or, I);
3424 return BinaryOperator::createNot(Or);
3428 Value *A = 0, *B = 0, *C = 0, *D = 0;
3429 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3430 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3431 return ReplaceInstUsesWith(I, Op1);
3433 // (A|B) & ~(A&B) -> A^B
3434 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3435 if ((A == C && B == D) || (A == D && B == C))
3436 return BinaryOperator::createXor(A, B);
3440 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3441 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3442 return ReplaceInstUsesWith(I, Op0);
3444 // ~(A&B) & (A|B) -> A^B
3445 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3446 if ((A == C && B == D) || (A == D && B == C))
3447 return BinaryOperator::createXor(A, B);
3451 if (Op0->hasOneUse() &&
3452 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3453 if (A == Op1) { // (A^B)&A -> A&(A^B)
3454 I.swapOperands(); // Simplify below
3455 std::swap(Op0, Op1);
3456 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3457 cast<BinaryOperator>(Op0)->swapOperands();
3458 I.swapOperands(); // Simplify below
3459 std::swap(Op0, Op1);
3462 if (Op1->hasOneUse() &&
3463 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3464 if (B == Op0) { // B&(A^B) -> B&(B^A)
3465 cast<BinaryOperator>(Op1)->swapOperands();
3468 if (A == Op0) { // A&(A^B) -> A & ~B
3469 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3470 InsertNewInstBefore(NotB, I);
3471 return BinaryOperator::createAnd(A, NotB);
3476 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3477 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3478 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3481 Value *LHSVal, *RHSVal;
3482 ConstantInt *LHSCst, *RHSCst;
3483 ICmpInst::Predicate LHSCC, RHSCC;
3484 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3485 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3486 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3487 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3488 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3489 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3490 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3491 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3493 // Don't try to fold ICMP_SLT + ICMP_ULT.
3494 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3495 ICmpInst::isSignedPredicate(LHSCC) ==
3496 ICmpInst::isSignedPredicate(RHSCC))) {
3497 // Ensure that the larger constant is on the RHS.
3498 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3499 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3500 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3501 ICmpInst *LHS = cast<ICmpInst>(Op0);
3502 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3503 std::swap(LHS, RHS);
3504 std::swap(LHSCst, RHSCst);
3505 std::swap(LHSCC, RHSCC);
3508 // At this point, we know we have have two icmp instructions
3509 // comparing a value against two constants and and'ing the result
3510 // together. Because of the above check, we know that we only have
3511 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3512 // (from the FoldICmpLogical check above), that the two constants
3513 // are not equal and that the larger constant is on the RHS
3514 assert(LHSCst != RHSCst && "Compares not folded above?");
3517 default: assert(0 && "Unknown integer condition code!");
3518 case ICmpInst::ICMP_EQ:
3520 default: assert(0 && "Unknown integer condition code!");
3521 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3522 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3523 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3524 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3525 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3526 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3527 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3528 return ReplaceInstUsesWith(I, LHS);
3530 case ICmpInst::ICMP_NE:
3532 default: assert(0 && "Unknown integer condition code!");
3533 case ICmpInst::ICMP_ULT:
3534 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3535 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3536 break; // (X != 13 & X u< 15) -> no change
3537 case ICmpInst::ICMP_SLT:
3538 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3539 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3540 break; // (X != 13 & X s< 15) -> no change
3541 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3542 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3543 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3544 return ReplaceInstUsesWith(I, RHS);
3545 case ICmpInst::ICMP_NE:
3546 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3547 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3548 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3549 LHSVal->getName()+".off");
3550 InsertNewInstBefore(Add, I);
3551 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3552 ConstantInt::get(Add->getType(), 1));
3554 break; // (X != 13 & X != 15) -> no change
3557 case ICmpInst::ICMP_ULT:
3559 default: assert(0 && "Unknown integer condition code!");
3560 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3561 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3562 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3563 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3565 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3566 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3567 return ReplaceInstUsesWith(I, LHS);
3568 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3572 case ICmpInst::ICMP_SLT:
3574 default: assert(0 && "Unknown integer condition code!");
3575 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3576 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3577 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3578 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3580 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3581 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3582 return ReplaceInstUsesWith(I, LHS);
3583 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3587 case ICmpInst::ICMP_UGT:
3589 default: assert(0 && "Unknown integer condition code!");
3590 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3591 return ReplaceInstUsesWith(I, LHS);
3592 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3593 return ReplaceInstUsesWith(I, RHS);
3594 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3596 case ICmpInst::ICMP_NE:
3597 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3598 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3599 break; // (X u> 13 & X != 15) -> no change
3600 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3601 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3603 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3607 case ICmpInst::ICMP_SGT:
3609 default: assert(0 && "Unknown integer condition code!");
3610 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3611 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3612 return ReplaceInstUsesWith(I, RHS);
3613 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3615 case ICmpInst::ICMP_NE:
3616 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3617 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3618 break; // (X s> 13 & X != 15) -> no change
3619 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3620 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3622 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3630 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3631 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3632 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3633 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3634 const Type *SrcTy = Op0C->getOperand(0)->getType();
3635 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3636 // Only do this if the casts both really cause code to be generated.
3637 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3639 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3641 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3642 Op1C->getOperand(0),
3644 InsertNewInstBefore(NewOp, I);
3645 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3649 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3650 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3651 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3652 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3653 SI0->getOperand(1) == SI1->getOperand(1) &&
3654 (SI0->hasOneUse() || SI1->hasOneUse())) {
3655 Instruction *NewOp =
3656 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3658 SI0->getName()), I);
3659 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3660 SI1->getOperand(1));
3664 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3665 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3666 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3667 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3668 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3669 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3670 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3671 // If either of the constants are nans, then the whole thing returns
3673 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3674 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3675 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3676 RHS->getOperand(0));
3681 return Changed ? &I : 0;
3684 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3685 /// in the result. If it does, and if the specified byte hasn't been filled in
3686 /// yet, fill it in and return false.
3687 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3688 Instruction *I = dyn_cast<Instruction>(V);
3689 if (I == 0) return true;
3691 // If this is an or instruction, it is an inner node of the bswap.
3692 if (I->getOpcode() == Instruction::Or)
3693 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3694 CollectBSwapParts(I->getOperand(1), ByteValues);
3696 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3697 // If this is a shift by a constant int, and it is "24", then its operand
3698 // defines a byte. We only handle unsigned types here.
3699 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3700 // Not shifting the entire input by N-1 bytes?
3701 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3702 8*(ByteValues.size()-1))
3706 if (I->getOpcode() == Instruction::Shl) {
3707 // X << 24 defines the top byte with the lowest of the input bytes.
3708 DestNo = ByteValues.size()-1;
3710 // X >>u 24 defines the low byte with the highest of the input bytes.
3714 // If the destination byte value is already defined, the values are or'd
3715 // together, which isn't a bswap (unless it's an or of the same bits).
3716 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3718 ByteValues[DestNo] = I->getOperand(0);
3722 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3724 Value *Shift = 0, *ShiftLHS = 0;
3725 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3726 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3727 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3729 Instruction *SI = cast<Instruction>(Shift);
3731 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3732 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3733 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3736 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3738 if (AndAmt->getValue().getActiveBits() > 64)
3740 uint64_t AndAmtVal = AndAmt->getZExtValue();
3741 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3742 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3744 // Unknown mask for bswap.
3745 if (DestByte == ByteValues.size()) return true;
3747 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3749 if (SI->getOpcode() == Instruction::Shl)
3750 SrcByte = DestByte - ShiftBytes;
3752 SrcByte = DestByte + ShiftBytes;
3754 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3755 if (SrcByte != ByteValues.size()-DestByte-1)
3758 // If the destination byte value is already defined, the values are or'd
3759 // together, which isn't a bswap (unless it's an or of the same bits).
3760 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3762 ByteValues[DestByte] = SI->getOperand(0);
3766 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3767 /// If so, insert the new bswap intrinsic and return it.
3768 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3769 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3770 if (!ITy || ITy->getBitWidth() % 16)
3771 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3773 /// ByteValues - For each byte of the result, we keep track of which value
3774 /// defines each byte.
3775 SmallVector<Value*, 8> ByteValues;
3776 ByteValues.resize(ITy->getBitWidth()/8);
3778 // Try to find all the pieces corresponding to the bswap.
3779 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3780 CollectBSwapParts(I.getOperand(1), ByteValues))
3783 // Check to see if all of the bytes come from the same value.
3784 Value *V = ByteValues[0];
3785 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3787 // Check to make sure that all of the bytes come from the same value.
3788 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3789 if (ByteValues[i] != V)
3791 const Type *Tys[] = { ITy };
3792 Module *M = I.getParent()->getParent()->getParent();
3793 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3794 return new CallInst(F, V);
3798 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3799 bool Changed = SimplifyCommutative(I);
3800 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3802 if (isa<UndefValue>(Op1)) // X | undef -> -1
3803 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3807 return ReplaceInstUsesWith(I, Op0);
3809 // See if we can simplify any instructions used by the instruction whose sole
3810 // purpose is to compute bits we don't care about.
3811 if (!isa<VectorType>(I.getType())) {
3812 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3813 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3814 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3815 KnownZero, KnownOne))
3817 } else if (isa<ConstantAggregateZero>(Op1)) {
3818 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3819 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3820 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3821 return ReplaceInstUsesWith(I, I.getOperand(1));
3827 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3828 ConstantInt *C1 = 0; Value *X = 0;
3829 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3830 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3831 Instruction *Or = BinaryOperator::createOr(X, RHS);
3832 InsertNewInstBefore(Or, I);
3834 return BinaryOperator::createAnd(Or,
3835 ConstantInt::get(RHS->getValue() | C1->getValue()));
3838 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3839 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3840 Instruction *Or = BinaryOperator::createOr(X, RHS);
3841 InsertNewInstBefore(Or, I);
3843 return BinaryOperator::createXor(Or,
3844 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3847 // Try to fold constant and into select arguments.
3848 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3849 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3851 if (isa<PHINode>(Op0))
3852 if (Instruction *NV = FoldOpIntoPhi(I))
3856 Value *A = 0, *B = 0;
3857 ConstantInt *C1 = 0, *C2 = 0;
3859 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3860 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3861 return ReplaceInstUsesWith(I, Op1);
3862 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3863 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3864 return ReplaceInstUsesWith(I, Op0);
3866 // (A | B) | C and A | (B | C) -> bswap if possible.
3867 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3868 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3869 match(Op1, m_Or(m_Value(), m_Value())) ||
3870 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3871 match(Op1, m_Shift(m_Value(), m_Value())))) {
3872 if (Instruction *BSwap = MatchBSwap(I))
3876 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3877 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3878 MaskedValueIsZero(Op1, C1->getValue())) {
3879 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3880 InsertNewInstBefore(NOr, I);
3882 return BinaryOperator::createXor(NOr, C1);
3885 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3886 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3887 MaskedValueIsZero(Op0, C1->getValue())) {
3888 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3889 InsertNewInstBefore(NOr, I);
3891 return BinaryOperator::createXor(NOr, C1);
3895 Value *C = 0, *D = 0;
3896 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3897 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3898 Value *V1 = 0, *V2 = 0, *V3 = 0;
3899 C1 = dyn_cast<ConstantInt>(C);
3900 C2 = dyn_cast<ConstantInt>(D);
3901 if (C1 && C2) { // (A & C1)|(B & C2)
3902 // If we have: ((V + N) & C1) | (V & C2)
3903 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3904 // replace with V+N.
3905 if (C1->getValue() == ~C2->getValue()) {
3906 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3907 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3908 // Add commutes, try both ways.
3909 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3910 return ReplaceInstUsesWith(I, A);
3911 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3912 return ReplaceInstUsesWith(I, A);
3914 // Or commutes, try both ways.
3915 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3916 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3917 // Add commutes, try both ways.
3918 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3919 return ReplaceInstUsesWith(I, B);
3920 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3921 return ReplaceInstUsesWith(I, B);
3924 V1 = 0; V2 = 0; V3 = 0;
3927 // Check to see if we have any common things being and'ed. If so, find the
3928 // terms for V1 & (V2|V3).
3929 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3930 if (A == B) // (A & C)|(A & D) == A & (C|D)
3931 V1 = A, V2 = C, V3 = D;
3932 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3933 V1 = A, V2 = B, V3 = C;
3934 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3935 V1 = C, V2 = A, V3 = D;
3936 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3937 V1 = C, V2 = A, V3 = B;
3941 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3942 return BinaryOperator::createAnd(V1, Or);
3947 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3948 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3949 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3950 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3951 SI0->getOperand(1) == SI1->getOperand(1) &&
3952 (SI0->hasOneUse() || SI1->hasOneUse())) {
3953 Instruction *NewOp =
3954 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3956 SI0->getName()), I);
3957 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3958 SI1->getOperand(1));
3962 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3963 if (A == Op1) // ~A | A == -1
3964 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3968 // Note, A is still live here!
3969 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3971 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3973 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3974 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3975 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3976 I.getName()+".demorgan"), I);
3977 return BinaryOperator::createNot(And);
3981 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3982 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3983 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3986 Value *LHSVal, *RHSVal;
3987 ConstantInt *LHSCst, *RHSCst;
3988 ICmpInst::Predicate LHSCC, RHSCC;
3989 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3990 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3991 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3992 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3993 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3994 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3995 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3996 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3997 // We can't fold (ugt x, C) | (sgt x, C2).
3998 PredicatesFoldable(LHSCC, RHSCC)) {
3999 // Ensure that the larger constant is on the RHS.
4000 ICmpInst *LHS = cast<ICmpInst>(Op0);
4002 if (ICmpInst::isSignedPredicate(LHSCC))
4003 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4005 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4008 std::swap(LHS, RHS);
4009 std::swap(LHSCst, RHSCst);
4010 std::swap(LHSCC, RHSCC);
4013 // At this point, we know we have have two icmp instructions
4014 // comparing a value against two constants and or'ing the result
4015 // together. Because of the above check, we know that we only have
4016 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4017 // FoldICmpLogical check above), that the two constants are not
4019 assert(LHSCst != RHSCst && "Compares not folded above?");
4022 default: assert(0 && "Unknown integer condition code!");
4023 case ICmpInst::ICMP_EQ:
4025 default: assert(0 && "Unknown integer condition code!");
4026 case ICmpInst::ICMP_EQ:
4027 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4028 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4029 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4030 LHSVal->getName()+".off");
4031 InsertNewInstBefore(Add, I);
4032 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4033 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4035 break; // (X == 13 | X == 15) -> no change
4036 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4037 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4039 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4040 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4041 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4042 return ReplaceInstUsesWith(I, RHS);
4045 case ICmpInst::ICMP_NE:
4047 default: assert(0 && "Unknown integer condition code!");
4048 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4049 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4050 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4051 return ReplaceInstUsesWith(I, LHS);
4052 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4053 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4054 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4055 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4058 case ICmpInst::ICMP_ULT:
4060 default: assert(0 && "Unknown integer condition code!");
4061 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4063 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4064 // If RHSCst is [us]MAXINT, it is always false. Not handling
4065 // this can cause overflow.
4066 if (RHSCst->isMaxValue(false))
4067 return ReplaceInstUsesWith(I, LHS);
4068 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4070 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4072 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4073 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4074 return ReplaceInstUsesWith(I, RHS);
4075 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4079 case ICmpInst::ICMP_SLT:
4081 default: assert(0 && "Unknown integer condition code!");
4082 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4084 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4085 // If RHSCst is [us]MAXINT, it is always false. Not handling
4086 // this can cause overflow.
4087 if (RHSCst->isMaxValue(true))
4088 return ReplaceInstUsesWith(I, LHS);
4089 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4091 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4093 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4094 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4095 return ReplaceInstUsesWith(I, RHS);
4096 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4100 case ICmpInst::ICMP_UGT:
4102 default: assert(0 && "Unknown integer condition code!");
4103 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4104 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4105 return ReplaceInstUsesWith(I, LHS);
4106 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4108 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4109 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4110 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4111 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4115 case ICmpInst::ICMP_SGT:
4117 default: assert(0 && "Unknown integer condition code!");
4118 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4119 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4120 return ReplaceInstUsesWith(I, LHS);
4121 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4123 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4124 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4125 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4126 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4134 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4135 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4136 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4137 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4138 const Type *SrcTy = Op0C->getOperand(0)->getType();
4139 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4140 // Only do this if the casts both really cause code to be generated.
4141 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4143 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4145 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4146 Op1C->getOperand(0),
4148 InsertNewInstBefore(NewOp, I);
4149 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4155 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4156 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4157 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4158 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4159 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4160 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4161 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4162 // If either of the constants are nans, then the whole thing returns
4164 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4165 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4167 // Otherwise, no need to compare the two constants, compare the
4169 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4170 RHS->getOperand(0));
4175 return Changed ? &I : 0;
4178 // XorSelf - Implements: X ^ X --> 0
4181 XorSelf(Value *rhs) : RHS(rhs) {}
4182 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4183 Instruction *apply(BinaryOperator &Xor) const {
4189 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4190 bool Changed = SimplifyCommutative(I);
4191 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4193 if (isa<UndefValue>(Op1))
4194 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4196 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4197 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4198 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4199 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4202 // See if we can simplify any instructions used by the instruction whose sole
4203 // purpose is to compute bits we don't care about.
4204 if (!isa<VectorType>(I.getType())) {
4205 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4206 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4207 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4208 KnownZero, KnownOne))
4210 } else if (isa<ConstantAggregateZero>(Op1)) {
4211 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4214 // Is this a ~ operation?
4215 if (Value *NotOp = dyn_castNotVal(&I)) {
4216 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4217 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4218 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4219 if (Op0I->getOpcode() == Instruction::And ||
4220 Op0I->getOpcode() == Instruction::Or) {
4221 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4222 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4224 BinaryOperator::createNot(Op0I->getOperand(1),
4225 Op0I->getOperand(1)->getName()+".not");
4226 InsertNewInstBefore(NotY, I);
4227 if (Op0I->getOpcode() == Instruction::And)
4228 return BinaryOperator::createOr(Op0NotVal, NotY);
4230 return BinaryOperator::createAnd(Op0NotVal, NotY);
4237 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4238 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4239 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4240 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4241 return new ICmpInst(ICI->getInversePredicate(),
4242 ICI->getOperand(0), ICI->getOperand(1));
4244 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4245 return new FCmpInst(FCI->getInversePredicate(),
4246 FCI->getOperand(0), FCI->getOperand(1));
4249 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4250 // ~(c-X) == X-c-1 == X+(-c-1)
4251 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4252 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4253 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4254 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4255 ConstantInt::get(I.getType(), 1));
4256 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4259 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4260 if (Op0I->getOpcode() == Instruction::Add) {
4261 // ~(X-c) --> (-c-1)-X
4262 if (RHS->isAllOnesValue()) {
4263 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4264 return BinaryOperator::createSub(
4265 ConstantExpr::getSub(NegOp0CI,
4266 ConstantInt::get(I.getType(), 1)),
4267 Op0I->getOperand(0));
4268 } else if (RHS->getValue().isSignBit()) {
4269 // (X + C) ^ signbit -> (X + C + signbit)
4270 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4271 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4274 } else if (Op0I->getOpcode() == Instruction::Or) {
4275 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4276 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4277 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4278 // Anything in both C1 and C2 is known to be zero, remove it from
4280 Constant *CommonBits = And(Op0CI, RHS);
4281 NewRHS = ConstantExpr::getAnd(NewRHS,
4282 ConstantExpr::getNot(CommonBits));
4283 AddToWorkList(Op0I);
4284 I.setOperand(0, Op0I->getOperand(0));
4285 I.setOperand(1, NewRHS);
4291 // Try to fold constant and into select arguments.
4292 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4293 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4295 if (isa<PHINode>(Op0))
4296 if (Instruction *NV = FoldOpIntoPhi(I))
4300 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4302 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4304 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4306 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4309 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4312 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4313 if (A == Op0) { // B^(B|A) == (A|B)^B
4314 Op1I->swapOperands();
4316 std::swap(Op0, Op1);
4317 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4318 I.swapOperands(); // Simplified below.
4319 std::swap(Op0, Op1);
4321 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4322 if (Op0 == A) // A^(A^B) == B
4323 return ReplaceInstUsesWith(I, B);
4324 else if (Op0 == B) // A^(B^A) == B
4325 return ReplaceInstUsesWith(I, A);
4326 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4327 if (A == Op0) { // A^(A&B) -> A^(B&A)
4328 Op1I->swapOperands();
4331 if (B == Op0) { // A^(B&A) -> (B&A)^A
4332 I.swapOperands(); // Simplified below.
4333 std::swap(Op0, Op1);
4338 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4341 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4342 if (A == Op1) // (B|A)^B == (A|B)^B
4344 if (B == Op1) { // (A|B)^B == A & ~B
4346 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4347 return BinaryOperator::createAnd(A, NotB);
4349 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4350 if (Op1 == A) // (A^B)^A == B
4351 return ReplaceInstUsesWith(I, B);
4352 else if (Op1 == B) // (B^A)^A == B
4353 return ReplaceInstUsesWith(I, A);
4354 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4355 if (A == Op1) // (A&B)^A -> (B&A)^A
4357 if (B == Op1 && // (B&A)^A == ~B & A
4358 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4360 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4361 return BinaryOperator::createAnd(N, Op1);
4366 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4367 if (Op0I && Op1I && Op0I->isShift() &&
4368 Op0I->getOpcode() == Op1I->getOpcode() &&
4369 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4370 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4371 Instruction *NewOp =
4372 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4373 Op1I->getOperand(0),
4374 Op0I->getName()), I);
4375 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4376 Op1I->getOperand(1));
4380 Value *A, *B, *C, *D;
4381 // (A & B)^(A | B) -> A ^ B
4382 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4383 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4384 if ((A == C && B == D) || (A == D && B == C))
4385 return BinaryOperator::createXor(A, B);
4387 // (A | B)^(A & B) -> A ^ B
4388 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4389 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4390 if ((A == C && B == D) || (A == D && B == C))
4391 return BinaryOperator::createXor(A, B);
4395 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4396 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4397 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4398 // (X & Y)^(X & Y) -> (Y^Z) & X
4399 Value *X = 0, *Y = 0, *Z = 0;
4401 X = A, Y = B, Z = D;
4403 X = A, Y = B, Z = C;
4405 X = B, Y = A, Z = D;
4407 X = B, Y = A, Z = C;
4410 Instruction *NewOp =
4411 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4412 return BinaryOperator::createAnd(NewOp, X);
4417 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4418 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4419 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4422 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4423 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4424 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4425 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4426 const Type *SrcTy = Op0C->getOperand(0)->getType();
4427 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4428 // Only do this if the casts both really cause code to be generated.
4429 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4431 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4433 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4434 Op1C->getOperand(0),
4436 InsertNewInstBefore(NewOp, I);
4437 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4441 return Changed ? &I : 0;
4444 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4445 /// overflowed for this type.
4446 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4447 ConstantInt *In2, bool IsSigned = false) {
4448 Result = cast<ConstantInt>(Add(In1, In2));
4451 if (In2->getValue().isNegative())
4452 return Result->getValue().sgt(In1->getValue());
4454 return Result->getValue().slt(In1->getValue());
4456 return Result->getValue().ult(In1->getValue());
4459 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4460 /// code necessary to compute the offset from the base pointer (without adding
4461 /// in the base pointer). Return the result as a signed integer of intptr size.
4462 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4463 TargetData &TD = IC.getTargetData();
4464 gep_type_iterator GTI = gep_type_begin(GEP);
4465 const Type *IntPtrTy = TD.getIntPtrType();
4466 Value *Result = Constant::getNullValue(IntPtrTy);
4468 // Build a mask for high order bits.
4469 unsigned IntPtrWidth = TD.getPointerSize()*8;
4470 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4472 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4473 Value *Op = GEP->getOperand(i);
4474 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4475 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4476 if (OpC->isZero()) continue;
4478 // Handle a struct index, which adds its field offset to the pointer.
4479 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4480 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4482 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4483 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4485 Result = IC.InsertNewInstBefore(
4486 BinaryOperator::createAdd(Result,
4487 ConstantInt::get(IntPtrTy, Size),
4488 GEP->getName()+".offs"), I);
4492 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4493 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4494 Scale = ConstantExpr::getMul(OC, Scale);
4495 if (Constant *RC = dyn_cast<Constant>(Result))
4496 Result = ConstantExpr::getAdd(RC, Scale);
4498 // Emit an add instruction.
4499 Result = IC.InsertNewInstBefore(
4500 BinaryOperator::createAdd(Result, Scale,
4501 GEP->getName()+".offs"), I);
4505 // Convert to correct type.
4506 if (Op->getType() != IntPtrTy) {
4507 if (Constant *OpC = dyn_cast<Constant>(Op))
4508 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4510 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4511 Op->getName()+".c"), I);
4514 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4515 if (Constant *OpC = dyn_cast<Constant>(Op))
4516 Op = ConstantExpr::getMul(OpC, Scale);
4517 else // We'll let instcombine(mul) convert this to a shl if possible.
4518 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4519 GEP->getName()+".idx"), I);
4522 // Emit an add instruction.
4523 if (isa<Constant>(Op) && isa<Constant>(Result))
4524 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4525 cast<Constant>(Result));
4527 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4528 GEP->getName()+".offs"), I);
4533 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4534 /// else. At this point we know that the GEP is on the LHS of the comparison.
4535 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4536 ICmpInst::Predicate Cond,
4538 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4540 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4541 if (isa<PointerType>(CI->getOperand(0)->getType()))
4542 RHS = CI->getOperand(0);
4544 Value *PtrBase = GEPLHS->getOperand(0);
4545 if (PtrBase == RHS) {
4546 // As an optimization, we don't actually have to compute the actual value of
4547 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4548 // each index is zero or not.
4549 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4550 Instruction *InVal = 0;
4551 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4552 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4554 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4555 if (isa<UndefValue>(C)) // undef index -> undef.
4556 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4557 if (C->isNullValue())
4559 else if (TD->getABITypeSize(GTI.getIndexedType()) == 0) {
4560 EmitIt = false; // This is indexing into a zero sized array?
4561 } else if (isa<ConstantInt>(C))
4562 return ReplaceInstUsesWith(I, // No comparison is needed here.
4563 ConstantInt::get(Type::Int1Ty,
4564 Cond == ICmpInst::ICMP_NE));
4569 new ICmpInst(Cond, GEPLHS->getOperand(i),
4570 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4574 InVal = InsertNewInstBefore(InVal, I);
4575 InsertNewInstBefore(Comp, I);
4576 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4577 InVal = BinaryOperator::createOr(InVal, Comp);
4578 else // True if all are equal
4579 InVal = BinaryOperator::createAnd(InVal, Comp);
4587 // No comparison is needed here, all indexes = 0
4588 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4589 Cond == ICmpInst::ICMP_EQ));
4592 // Only lower this if the icmp is the only user of the GEP or if we expect
4593 // the result to fold to a constant!
4594 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4595 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4596 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4597 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4598 Constant::getNullValue(Offset->getType()));
4600 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4601 // If the base pointers are different, but the indices are the same, just
4602 // compare the base pointer.
4603 if (PtrBase != GEPRHS->getOperand(0)) {
4604 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4605 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4606 GEPRHS->getOperand(0)->getType();
4608 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4609 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4610 IndicesTheSame = false;
4614 // If all indices are the same, just compare the base pointers.
4616 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4617 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4619 // Otherwise, the base pointers are different and the indices are
4620 // different, bail out.
4624 // If one of the GEPs has all zero indices, recurse.
4625 bool AllZeros = true;
4626 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4627 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4628 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4633 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4634 ICmpInst::getSwappedPredicate(Cond), I);
4636 // If the other GEP has all zero indices, recurse.
4638 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4639 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4640 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4645 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4647 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4648 // If the GEPs only differ by one index, compare it.
4649 unsigned NumDifferences = 0; // Keep track of # differences.
4650 unsigned DiffOperand = 0; // The operand that differs.
4651 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4652 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4653 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4654 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4655 // Irreconcilable differences.
4659 if (NumDifferences++) break;
4664 if (NumDifferences == 0) // SAME GEP?
4665 return ReplaceInstUsesWith(I, // No comparison is needed here.
4666 ConstantInt::get(Type::Int1Ty,
4667 isTrueWhenEqual(Cond)));
4669 else if (NumDifferences == 1) {
4670 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4671 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4672 // Make sure we do a signed comparison here.
4673 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4677 // Only lower this if the icmp is the only user of the GEP or if we expect
4678 // the result to fold to a constant!
4679 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4680 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4681 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4682 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4683 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4684 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4690 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4691 bool Changed = SimplifyCompare(I);
4692 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4694 // Fold trivial predicates.
4695 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4696 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4697 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4698 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4700 // Simplify 'fcmp pred X, X'
4702 switch (I.getPredicate()) {
4703 default: assert(0 && "Unknown predicate!");
4704 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4705 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4706 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4707 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4708 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4709 case FCmpInst::FCMP_OLT: // True if ordered and less than
4710 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4711 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4713 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4714 case FCmpInst::FCMP_ULT: // True if unordered or less than
4715 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4716 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4717 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4718 I.setPredicate(FCmpInst::FCMP_UNO);
4719 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4722 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4723 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4724 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4725 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4726 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4727 I.setPredicate(FCmpInst::FCMP_ORD);
4728 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4733 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4734 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4736 // Handle fcmp with constant RHS
4737 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4738 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4739 switch (LHSI->getOpcode()) {
4740 case Instruction::PHI:
4741 if (Instruction *NV = FoldOpIntoPhi(I))
4744 case Instruction::Select:
4745 // If either operand of the select is a constant, we can fold the
4746 // comparison into the select arms, which will cause one to be
4747 // constant folded and the select turned into a bitwise or.
4748 Value *Op1 = 0, *Op2 = 0;
4749 if (LHSI->hasOneUse()) {
4750 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4751 // Fold the known value into the constant operand.
4752 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4753 // Insert a new FCmp of the other select operand.
4754 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4755 LHSI->getOperand(2), RHSC,
4757 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4758 // Fold the known value into the constant operand.
4759 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4760 // Insert a new FCmp of the other select operand.
4761 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4762 LHSI->getOperand(1), RHSC,
4768 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4773 return Changed ? &I : 0;
4776 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4777 bool Changed = SimplifyCompare(I);
4778 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4779 const Type *Ty = Op0->getType();
4783 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4784 isTrueWhenEqual(I)));
4786 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4787 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4789 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4790 // addresses never equal each other! We already know that Op0 != Op1.
4791 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4792 isa<ConstantPointerNull>(Op0)) &&
4793 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4794 isa<ConstantPointerNull>(Op1)))
4795 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4796 !isTrueWhenEqual(I)));
4798 // icmp's with boolean values can always be turned into bitwise operations
4799 if (Ty == Type::Int1Ty) {
4800 switch (I.getPredicate()) {
4801 default: assert(0 && "Invalid icmp instruction!");
4802 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4803 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4804 InsertNewInstBefore(Xor, I);
4805 return BinaryOperator::createNot(Xor);
4807 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4808 return BinaryOperator::createXor(Op0, Op1);
4810 case ICmpInst::ICMP_UGT:
4811 case ICmpInst::ICMP_SGT:
4812 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4814 case ICmpInst::ICMP_ULT:
4815 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4816 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4817 InsertNewInstBefore(Not, I);
4818 return BinaryOperator::createAnd(Not, Op1);
4820 case ICmpInst::ICMP_UGE:
4821 case ICmpInst::ICMP_SGE:
4822 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4824 case ICmpInst::ICMP_ULE:
4825 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4826 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4827 InsertNewInstBefore(Not, I);
4828 return BinaryOperator::createOr(Not, Op1);
4833 // See if we are doing a comparison between a constant and an instruction that
4834 // can be folded into the comparison.
4835 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4838 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4839 if (I.isEquality() && CI->isNullValue() &&
4840 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4841 // (icmp cond A B) if cond is equality
4842 return new ICmpInst(I.getPredicate(), A, B);
4845 switch (I.getPredicate()) {
4847 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4848 if (CI->isMinValue(false))
4849 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4850 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4851 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4852 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4853 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4854 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4855 if (CI->isMinValue(true))
4856 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4857 ConstantInt::getAllOnesValue(Op0->getType()));
4861 case ICmpInst::ICMP_SLT:
4862 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4863 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4864 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4865 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4866 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4867 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4870 case ICmpInst::ICMP_UGT:
4871 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4872 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4873 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4874 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4875 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4876 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4878 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4879 if (CI->isMaxValue(true))
4880 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4881 ConstantInt::getNullValue(Op0->getType()));
4884 case ICmpInst::ICMP_SGT:
4885 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4886 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4887 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4888 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4889 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4890 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4893 case ICmpInst::ICMP_ULE:
4894 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4895 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4896 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4897 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4898 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4899 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4902 case ICmpInst::ICMP_SLE:
4903 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4904 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4905 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4906 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4907 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4908 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4911 case ICmpInst::ICMP_UGE:
4912 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4913 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4914 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4915 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4916 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4917 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4920 case ICmpInst::ICMP_SGE:
4921 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4922 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4923 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4924 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4925 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4926 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4930 // If we still have a icmp le or icmp ge instruction, turn it into the
4931 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4932 // already been handled above, this requires little checking.
4934 switch (I.getPredicate()) {
4936 case ICmpInst::ICMP_ULE:
4937 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4938 case ICmpInst::ICMP_SLE:
4939 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4940 case ICmpInst::ICMP_UGE:
4941 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4942 case ICmpInst::ICMP_SGE:
4943 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4946 // See if we can fold the comparison based on bits known to be zero or one
4947 // in the input. If this comparison is a normal comparison, it demands all
4948 // bits, if it is a sign bit comparison, it only demands the sign bit.
4951 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4953 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4954 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4955 if (SimplifyDemandedBits(Op0,
4956 isSignBit ? APInt::getSignBit(BitWidth)
4957 : APInt::getAllOnesValue(BitWidth),
4958 KnownZero, KnownOne, 0))
4961 // Given the known and unknown bits, compute a range that the LHS could be
4963 if ((KnownOne | KnownZero) != 0) {
4964 // Compute the Min, Max and RHS values based on the known bits. For the
4965 // EQ and NE we use unsigned values.
4966 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4967 const APInt& RHSVal = CI->getValue();
4968 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4969 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4972 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4975 switch (I.getPredicate()) { // LE/GE have been folded already.
4976 default: assert(0 && "Unknown icmp opcode!");
4977 case ICmpInst::ICMP_EQ:
4978 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4979 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4981 case ICmpInst::ICMP_NE:
4982 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4983 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4985 case ICmpInst::ICMP_ULT:
4986 if (Max.ult(RHSVal))
4987 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4988 if (Min.uge(RHSVal))
4989 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4991 case ICmpInst::ICMP_UGT:
4992 if (Min.ugt(RHSVal))
4993 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4994 if (Max.ule(RHSVal))
4995 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4997 case ICmpInst::ICMP_SLT:
4998 if (Max.slt(RHSVal))
4999 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5000 if (Min.sgt(RHSVal))
5001 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5003 case ICmpInst::ICMP_SGT:
5004 if (Min.sgt(RHSVal))
5005 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5006 if (Max.sle(RHSVal))
5007 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5012 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5013 // instruction, see if that instruction also has constants so that the
5014 // instruction can be folded into the icmp
5015 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5016 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5020 // Handle icmp with constant (but not simple integer constant) RHS
5021 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5022 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5023 switch (LHSI->getOpcode()) {
5024 case Instruction::GetElementPtr:
5025 if (RHSC->isNullValue()) {
5026 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5027 bool isAllZeros = true;
5028 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5029 if (!isa<Constant>(LHSI->getOperand(i)) ||
5030 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5035 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5036 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5040 case Instruction::PHI:
5041 if (Instruction *NV = FoldOpIntoPhi(I))
5044 case Instruction::Select: {
5045 // If either operand of the select is a constant, we can fold the
5046 // comparison into the select arms, which will cause one to be
5047 // constant folded and the select turned into a bitwise or.
5048 Value *Op1 = 0, *Op2 = 0;
5049 if (LHSI->hasOneUse()) {
5050 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5051 // Fold the known value into the constant operand.
5052 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5053 // Insert a new ICmp of the other select operand.
5054 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5055 LHSI->getOperand(2), RHSC,
5057 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5058 // Fold the known value into the constant operand.
5059 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5060 // Insert a new ICmp of the other select operand.
5061 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5062 LHSI->getOperand(1), RHSC,
5068 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5071 case Instruction::Malloc:
5072 // If we have (malloc != null), and if the malloc has a single use, we
5073 // can assume it is successful and remove the malloc.
5074 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5075 AddToWorkList(LHSI);
5076 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5077 !isTrueWhenEqual(I)));
5083 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5084 if (User *GEP = dyn_castGetElementPtr(Op0))
5085 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5087 if (User *GEP = dyn_castGetElementPtr(Op1))
5088 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5089 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5092 // Test to see if the operands of the icmp are casted versions of other
5093 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5095 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5096 if (isa<PointerType>(Op0->getType()) &&
5097 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5098 // We keep moving the cast from the left operand over to the right
5099 // operand, where it can often be eliminated completely.
5100 Op0 = CI->getOperand(0);
5102 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5103 // so eliminate it as well.
5104 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5105 Op1 = CI2->getOperand(0);
5107 // If Op1 is a constant, we can fold the cast into the constant.
5108 if (Op0->getType() != Op1->getType())
5109 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5110 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5112 // Otherwise, cast the RHS right before the icmp
5113 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5115 return new ICmpInst(I.getPredicate(), Op0, Op1);
5119 if (isa<CastInst>(Op0)) {
5120 // Handle the special case of: icmp (cast bool to X), <cst>
5121 // This comes up when you have code like
5124 // For generality, we handle any zero-extension of any operand comparison
5125 // with a constant or another cast from the same type.
5126 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5127 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5131 if (I.isEquality()) {
5132 Value *A, *B, *C, *D;
5133 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5134 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5135 Value *OtherVal = A == Op1 ? B : A;
5136 return new ICmpInst(I.getPredicate(), OtherVal,
5137 Constant::getNullValue(A->getType()));
5140 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5141 // A^c1 == C^c2 --> A == C^(c1^c2)
5142 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5143 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5144 if (Op1->hasOneUse()) {
5145 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5146 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5147 return new ICmpInst(I.getPredicate(), A,
5148 InsertNewInstBefore(Xor, I));
5151 // A^B == A^D -> B == D
5152 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5153 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5154 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5155 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5159 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5160 (A == Op0 || B == Op0)) {
5161 // A == (A^B) -> B == 0
5162 Value *OtherVal = A == Op0 ? B : A;
5163 return new ICmpInst(I.getPredicate(), OtherVal,
5164 Constant::getNullValue(A->getType()));
5166 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5167 // (A-B) == A -> B == 0
5168 return new ICmpInst(I.getPredicate(), B,
5169 Constant::getNullValue(B->getType()));
5171 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5172 // A == (A-B) -> B == 0
5173 return new ICmpInst(I.getPredicate(), B,
5174 Constant::getNullValue(B->getType()));
5177 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5178 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5179 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5180 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5181 Value *X = 0, *Y = 0, *Z = 0;
5184 X = B; Y = D; Z = A;
5185 } else if (A == D) {
5186 X = B; Y = C; Z = A;
5187 } else if (B == C) {
5188 X = A; Y = D; Z = B;
5189 } else if (B == D) {
5190 X = A; Y = C; Z = B;
5193 if (X) { // Build (X^Y) & Z
5194 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5195 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5196 I.setOperand(0, Op1);
5197 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5202 return Changed ? &I : 0;
5206 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5207 /// and CmpRHS are both known to be integer constants.
5208 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5209 ConstantInt *DivRHS) {
5210 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5211 const APInt &CmpRHSV = CmpRHS->getValue();
5213 // FIXME: If the operand types don't match the type of the divide
5214 // then don't attempt this transform. The code below doesn't have the
5215 // logic to deal with a signed divide and an unsigned compare (and
5216 // vice versa). This is because (x /s C1) <s C2 produces different
5217 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5218 // (x /u C1) <u C2. Simply casting the operands and result won't
5219 // work. :( The if statement below tests that condition and bails
5221 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5222 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5224 if (DivRHS->isZero())
5225 return 0; // The ProdOV computation fails on divide by zero.
5227 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5228 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5229 // C2 (CI). By solving for X we can turn this into a range check
5230 // instead of computing a divide.
5231 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5233 // Determine if the product overflows by seeing if the product is
5234 // not equal to the divide. Make sure we do the same kind of divide
5235 // as in the LHS instruction that we're folding.
5236 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5237 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5239 // Get the ICmp opcode
5240 ICmpInst::Predicate Pred = ICI.getPredicate();
5242 // Figure out the interval that is being checked. For example, a comparison
5243 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5244 // Compute this interval based on the constants involved and the signedness of
5245 // the compare/divide. This computes a half-open interval, keeping track of
5246 // whether either value in the interval overflows. After analysis each
5247 // overflow variable is set to 0 if it's corresponding bound variable is valid
5248 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5249 int LoOverflow = 0, HiOverflow = 0;
5250 ConstantInt *LoBound = 0, *HiBound = 0;
5253 if (!DivIsSigned) { // udiv
5254 // e.g. X/5 op 3 --> [15, 20)
5256 HiOverflow = LoOverflow = ProdOV;
5258 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5259 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5260 if (CmpRHSV == 0) { // (X / pos) op 0
5261 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5262 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5264 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5265 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5266 HiOverflow = LoOverflow = ProdOV;
5268 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5269 } else { // (X / pos) op neg
5270 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5271 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5272 LoOverflow = AddWithOverflow(LoBound, Prod,
5273 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5274 HiBound = AddOne(Prod);
5275 HiOverflow = ProdOV ? -1 : 0;
5277 } else { // Divisor is < 0.
5278 if (CmpRHSV == 0) { // (X / neg) op 0
5279 // e.g. X/-5 op 0 --> [-4, 5)
5280 LoBound = AddOne(DivRHS);
5281 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5282 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5283 HiOverflow = 1; // [INTMIN+1, overflow)
5284 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5286 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5287 // e.g. X/-5 op 3 --> [-19, -14)
5288 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5290 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5291 HiBound = AddOne(Prod);
5292 } else { // (X / neg) op neg
5293 // e.g. X/-5 op -3 --> [15, 20)
5295 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5296 HiBound = Subtract(Prod, DivRHS);
5299 // Dividing by a negative swaps the condition. LT <-> GT
5300 Pred = ICmpInst::getSwappedPredicate(Pred);
5303 Value *X = DivI->getOperand(0);
5305 default: assert(0 && "Unhandled icmp opcode!");
5306 case ICmpInst::ICMP_EQ:
5307 if (LoOverflow && HiOverflow)
5308 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5309 else if (HiOverflow)
5310 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5311 ICmpInst::ICMP_UGE, X, LoBound);
5312 else if (LoOverflow)
5313 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5314 ICmpInst::ICMP_ULT, X, HiBound);
5316 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5317 case ICmpInst::ICMP_NE:
5318 if (LoOverflow && HiOverflow)
5319 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5320 else if (HiOverflow)
5321 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5322 ICmpInst::ICMP_ULT, X, LoBound);
5323 else if (LoOverflow)
5324 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5325 ICmpInst::ICMP_UGE, X, HiBound);
5327 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5328 case ICmpInst::ICMP_ULT:
5329 case ICmpInst::ICMP_SLT:
5330 if (LoOverflow == +1) // Low bound is greater than input range.
5331 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5332 if (LoOverflow == -1) // Low bound is less than input range.
5333 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5334 return new ICmpInst(Pred, X, LoBound);
5335 case ICmpInst::ICMP_UGT:
5336 case ICmpInst::ICMP_SGT:
5337 if (HiOverflow == +1) // High bound greater than input range.
5338 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5339 else if (HiOverflow == -1) // High bound less than input range.
5340 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5341 if (Pred == ICmpInst::ICMP_UGT)
5342 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5344 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5349 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5351 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5354 const APInt &RHSV = RHS->getValue();
5356 switch (LHSI->getOpcode()) {
5357 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5358 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5359 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5361 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5362 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5363 Value *CompareVal = LHSI->getOperand(0);
5365 // If the sign bit of the XorCST is not set, there is no change to
5366 // the operation, just stop using the Xor.
5367 if (!XorCST->getValue().isNegative()) {
5368 ICI.setOperand(0, CompareVal);
5369 AddToWorkList(LHSI);
5373 // Was the old condition true if the operand is positive?
5374 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5376 // If so, the new one isn't.
5377 isTrueIfPositive ^= true;
5379 if (isTrueIfPositive)
5380 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5382 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5386 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5387 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5388 LHSI->getOperand(0)->hasOneUse()) {
5389 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5391 // If the LHS is an AND of a truncating cast, we can widen the
5392 // and/compare to be the input width without changing the value
5393 // produced, eliminating a cast.
5394 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5395 // We can do this transformation if either the AND constant does not
5396 // have its sign bit set or if it is an equality comparison.
5397 // Extending a relational comparison when we're checking the sign
5398 // bit would not work.
5399 if (Cast->hasOneUse() &&
5400 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5401 RHSV.isPositive())) {
5403 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5404 APInt NewCST = AndCST->getValue();
5405 NewCST.zext(BitWidth);
5407 NewCI.zext(BitWidth);
5408 Instruction *NewAnd =
5409 BinaryOperator::createAnd(Cast->getOperand(0),
5410 ConstantInt::get(NewCST),LHSI->getName());
5411 InsertNewInstBefore(NewAnd, ICI);
5412 return new ICmpInst(ICI.getPredicate(), NewAnd,
5413 ConstantInt::get(NewCI));
5417 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5418 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5419 // happens a LOT in code produced by the C front-end, for bitfield
5421 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5422 if (Shift && !Shift->isShift())
5426 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5427 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5428 const Type *AndTy = AndCST->getType(); // Type of the and.
5430 // We can fold this as long as we can't shift unknown bits
5431 // into the mask. This can only happen with signed shift
5432 // rights, as they sign-extend.
5434 bool CanFold = Shift->isLogicalShift();
5436 // To test for the bad case of the signed shr, see if any
5437 // of the bits shifted in could be tested after the mask.
5438 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5439 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5441 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5442 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5443 AndCST->getValue()) == 0)
5449 if (Shift->getOpcode() == Instruction::Shl)
5450 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5452 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5454 // Check to see if we are shifting out any of the bits being
5456 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5457 // If we shifted bits out, the fold is not going to work out.
5458 // As a special case, check to see if this means that the
5459 // result is always true or false now.
5460 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5461 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5462 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5463 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5465 ICI.setOperand(1, NewCst);
5466 Constant *NewAndCST;
5467 if (Shift->getOpcode() == Instruction::Shl)
5468 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5470 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5471 LHSI->setOperand(1, NewAndCST);
5472 LHSI->setOperand(0, Shift->getOperand(0));
5473 AddToWorkList(Shift); // Shift is dead.
5474 AddUsesToWorkList(ICI);
5480 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5481 // preferable because it allows the C<<Y expression to be hoisted out
5482 // of a loop if Y is invariant and X is not.
5483 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5484 ICI.isEquality() && !Shift->isArithmeticShift() &&
5485 isa<Instruction>(Shift->getOperand(0))) {
5488 if (Shift->getOpcode() == Instruction::LShr) {
5489 NS = BinaryOperator::createShl(AndCST,
5490 Shift->getOperand(1), "tmp");
5492 // Insert a logical shift.
5493 NS = BinaryOperator::createLShr(AndCST,
5494 Shift->getOperand(1), "tmp");
5496 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5498 // Compute X & (C << Y).
5499 Instruction *NewAnd =
5500 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5501 InsertNewInstBefore(NewAnd, ICI);
5503 ICI.setOperand(0, NewAnd);
5509 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5510 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5513 uint32_t TypeBits = RHSV.getBitWidth();
5515 // Check that the shift amount is in range. If not, don't perform
5516 // undefined shifts. When the shift is visited it will be
5518 if (ShAmt->uge(TypeBits))
5521 if (ICI.isEquality()) {
5522 // If we are comparing against bits always shifted out, the
5523 // comparison cannot succeed.
5525 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5526 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5527 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5528 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5529 return ReplaceInstUsesWith(ICI, Cst);
5532 if (LHSI->hasOneUse()) {
5533 // Otherwise strength reduce the shift into an and.
5534 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5536 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5539 BinaryOperator::createAnd(LHSI->getOperand(0),
5540 Mask, LHSI->getName()+".mask");
5541 Value *And = InsertNewInstBefore(AndI, ICI);
5542 return new ICmpInst(ICI.getPredicate(), And,
5543 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5547 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5548 bool TrueIfSigned = false;
5549 if (LHSI->hasOneUse() &&
5550 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5551 // (X << 31) <s 0 --> (X&1) != 0
5552 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5553 (TypeBits-ShAmt->getZExtValue()-1));
5555 BinaryOperator::createAnd(LHSI->getOperand(0),
5556 Mask, LHSI->getName()+".mask");
5557 Value *And = InsertNewInstBefore(AndI, ICI);
5559 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5560 And, Constant::getNullValue(And->getType()));
5565 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5566 case Instruction::AShr: {
5567 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5570 if (ICI.isEquality()) {
5571 // Check that the shift amount is in range. If not, don't perform
5572 // undefined shifts. When the shift is visited it will be
5574 uint32_t TypeBits = RHSV.getBitWidth();
5575 if (ShAmt->uge(TypeBits))
5577 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5579 // If we are comparing against bits always shifted out, the
5580 // comparison cannot succeed.
5581 APInt Comp = RHSV << ShAmtVal;
5582 if (LHSI->getOpcode() == Instruction::LShr)
5583 Comp = Comp.lshr(ShAmtVal);
5585 Comp = Comp.ashr(ShAmtVal);
5587 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5588 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5589 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5590 return ReplaceInstUsesWith(ICI, Cst);
5593 if (LHSI->hasOneUse() || RHSV == 0) {
5594 // Otherwise strength reduce the shift into an and.
5595 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5596 Constant *Mask = ConstantInt::get(Val);
5599 BinaryOperator::createAnd(LHSI->getOperand(0),
5600 Mask, LHSI->getName()+".mask");
5601 Value *And = InsertNewInstBefore(AndI, ICI);
5602 return new ICmpInst(ICI.getPredicate(), And,
5603 ConstantExpr::getShl(RHS, ShAmt));
5609 case Instruction::SDiv:
5610 case Instruction::UDiv:
5611 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5612 // Fold this div into the comparison, producing a range check.
5613 // Determine, based on the divide type, what the range is being
5614 // checked. If there is an overflow on the low or high side, remember
5615 // it, otherwise compute the range [low, hi) bounding the new value.
5616 // See: InsertRangeTest above for the kinds of replacements possible.
5617 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5618 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5624 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5625 if (ICI.isEquality()) {
5626 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5628 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5629 // the second operand is a constant, simplify a bit.
5630 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5631 switch (BO->getOpcode()) {
5632 case Instruction::SRem:
5633 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5634 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5635 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5636 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5637 Instruction *NewRem =
5638 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5640 InsertNewInstBefore(NewRem, ICI);
5641 return new ICmpInst(ICI.getPredicate(), NewRem,
5642 Constant::getNullValue(BO->getType()));
5646 case Instruction::Add:
5647 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5648 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5649 if (BO->hasOneUse())
5650 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5651 Subtract(RHS, BOp1C));
5652 } else if (RHSV == 0) {
5653 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5654 // efficiently invertible, or if the add has just this one use.
5655 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5657 if (Value *NegVal = dyn_castNegVal(BOp1))
5658 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5659 else if (Value *NegVal = dyn_castNegVal(BOp0))
5660 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5661 else if (BO->hasOneUse()) {
5662 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5663 InsertNewInstBefore(Neg, ICI);
5665 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5669 case Instruction::Xor:
5670 // For the xor case, we can xor two constants together, eliminating
5671 // the explicit xor.
5672 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5673 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5674 ConstantExpr::getXor(RHS, BOC));
5677 case Instruction::Sub:
5678 // Replace (([sub|xor] A, B) != 0) with (A != B)
5680 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5684 case Instruction::Or:
5685 // If bits are being or'd in that are not present in the constant we
5686 // are comparing against, then the comparison could never succeed!
5687 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5688 Constant *NotCI = ConstantExpr::getNot(RHS);
5689 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5690 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5695 case Instruction::And:
5696 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5697 // If bits are being compared against that are and'd out, then the
5698 // comparison can never succeed!
5699 if ((RHSV & ~BOC->getValue()) != 0)
5700 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5703 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5704 if (RHS == BOC && RHSV.isPowerOf2())
5705 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5706 ICmpInst::ICMP_NE, LHSI,
5707 Constant::getNullValue(RHS->getType()));
5709 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5710 if (isSignBit(BOC)) {
5711 Value *X = BO->getOperand(0);
5712 Constant *Zero = Constant::getNullValue(X->getType());
5713 ICmpInst::Predicate pred = isICMP_NE ?
5714 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5715 return new ICmpInst(pred, X, Zero);
5718 // ((X & ~7) == 0) --> X < 8
5719 if (RHSV == 0 && isHighOnes(BOC)) {
5720 Value *X = BO->getOperand(0);
5721 Constant *NegX = ConstantExpr::getNeg(BOC);
5722 ICmpInst::Predicate pred = isICMP_NE ?
5723 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5724 return new ICmpInst(pred, X, NegX);
5729 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5730 // Handle icmp {eq|ne} <intrinsic>, intcst.
5731 if (II->getIntrinsicID() == Intrinsic::bswap) {
5733 ICI.setOperand(0, II->getOperand(1));
5734 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5738 } else { // Not a ICMP_EQ/ICMP_NE
5739 // If the LHS is a cast from an integral value of the same size,
5740 // then since we know the RHS is a constant, try to simlify.
5741 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5742 Value *CastOp = Cast->getOperand(0);
5743 const Type *SrcTy = CastOp->getType();
5744 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5745 if (SrcTy->isInteger() &&
5746 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5747 // If this is an unsigned comparison, try to make the comparison use
5748 // smaller constant values.
5749 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5750 // X u< 128 => X s> -1
5751 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5752 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5753 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5754 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5755 // X u> 127 => X s< 0
5756 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5757 Constant::getNullValue(SrcTy));
5765 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5766 /// We only handle extending casts so far.
5768 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5769 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5770 Value *LHSCIOp = LHSCI->getOperand(0);
5771 const Type *SrcTy = LHSCIOp->getType();
5772 const Type *DestTy = LHSCI->getType();
5775 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5776 // integer type is the same size as the pointer type.
5777 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5778 getTargetData().getPointerSizeInBits() ==
5779 cast<IntegerType>(DestTy)->getBitWidth()) {
5781 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5782 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5783 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5784 RHSOp = RHSC->getOperand(0);
5785 // If the pointer types don't match, insert a bitcast.
5786 if (LHSCIOp->getType() != RHSOp->getType())
5787 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5788 LHSCIOp->getType(), ICI);
5792 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5795 // The code below only handles extension cast instructions, so far.
5797 if (LHSCI->getOpcode() != Instruction::ZExt &&
5798 LHSCI->getOpcode() != Instruction::SExt)
5801 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5802 bool isSignedCmp = ICI.isSignedPredicate();
5804 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5805 // Not an extension from the same type?
5806 RHSCIOp = CI->getOperand(0);
5807 if (RHSCIOp->getType() != LHSCIOp->getType())
5810 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5811 // and the other is a zext), then we can't handle this.
5812 if (CI->getOpcode() != LHSCI->getOpcode())
5815 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5816 // then we can't handle this.
5817 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5820 // Okay, just insert a compare of the reduced operands now!
5821 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5824 // If we aren't dealing with a constant on the RHS, exit early
5825 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5829 // Compute the constant that would happen if we truncated to SrcTy then
5830 // reextended to DestTy.
5831 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5832 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5834 // If the re-extended constant didn't change...
5836 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5837 // For example, we might have:
5838 // %A = sext short %X to uint
5839 // %B = icmp ugt uint %A, 1330
5840 // It is incorrect to transform this into
5841 // %B = icmp ugt short %X, 1330
5842 // because %A may have negative value.
5844 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5845 // OR operation is EQ/NE.
5846 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5847 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5852 // The re-extended constant changed so the constant cannot be represented
5853 // in the shorter type. Consequently, we cannot emit a simple comparison.
5855 // First, handle some easy cases. We know the result cannot be equal at this
5856 // point so handle the ICI.isEquality() cases
5857 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5858 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5859 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5860 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5862 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5863 // should have been folded away previously and not enter in here.
5866 // We're performing a signed comparison.
5867 if (cast<ConstantInt>(CI)->getValue().isNegative())
5868 Result = ConstantInt::getFalse(); // X < (small) --> false
5870 Result = ConstantInt::getTrue(); // X < (large) --> true
5872 // We're performing an unsigned comparison.
5874 // We're performing an unsigned comp with a sign extended value.
5875 // This is true if the input is >= 0. [aka >s -1]
5876 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5877 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5878 NegOne, ICI.getName()), ICI);
5880 // Unsigned extend & unsigned compare -> always true.
5881 Result = ConstantInt::getTrue();
5885 // Finally, return the value computed.
5886 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5887 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5888 return ReplaceInstUsesWith(ICI, Result);
5890 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5891 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5892 "ICmp should be folded!");
5893 if (Constant *CI = dyn_cast<Constant>(Result))
5894 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5896 return BinaryOperator::createNot(Result);
5900 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5901 return commonShiftTransforms(I);
5904 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5905 return commonShiftTransforms(I);
5908 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5909 if (Instruction *R = commonShiftTransforms(I))
5912 Value *Op0 = I.getOperand(0);
5914 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5915 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5916 if (CSI->isAllOnesValue())
5917 return ReplaceInstUsesWith(I, CSI);
5919 // See if we can turn a signed shr into an unsigned shr.
5920 if (MaskedValueIsZero(Op0,
5921 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
5922 return BinaryOperator::createLShr(Op0, I.getOperand(1));
5927 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5928 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5929 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5931 // shl X, 0 == X and shr X, 0 == X
5932 // shl 0, X == 0 and shr 0, X == 0
5933 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5934 Op0 == Constant::getNullValue(Op0->getType()))
5935 return ReplaceInstUsesWith(I, Op0);
5937 if (isa<UndefValue>(Op0)) {
5938 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5939 return ReplaceInstUsesWith(I, Op0);
5940 else // undef << X -> 0, undef >>u X -> 0
5941 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5943 if (isa<UndefValue>(Op1)) {
5944 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5945 return ReplaceInstUsesWith(I, Op0);
5946 else // X << undef, X >>u undef -> 0
5947 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5950 // Try to fold constant and into select arguments.
5951 if (isa<Constant>(Op0))
5952 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5953 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5956 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5957 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5962 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5963 BinaryOperator &I) {
5964 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5966 // See if we can simplify any instructions used by the instruction whose sole
5967 // purpose is to compute bits we don't care about.
5968 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5969 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5970 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5971 KnownZero, KnownOne))
5974 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5975 // of a signed value.
5977 if (Op1->uge(TypeBits)) {
5978 if (I.getOpcode() != Instruction::AShr)
5979 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5981 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5986 // ((X*C1) << C2) == (X * (C1 << C2))
5987 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5988 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5989 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5990 return BinaryOperator::createMul(BO->getOperand(0),
5991 ConstantExpr::getShl(BOOp, Op1));
5993 // Try to fold constant and into select arguments.
5994 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5995 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5997 if (isa<PHINode>(Op0))
5998 if (Instruction *NV = FoldOpIntoPhi(I))
6001 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6002 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6003 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6004 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6005 // place. Don't try to do this transformation in this case. Also, we
6006 // require that the input operand is a shift-by-constant so that we have
6007 // confidence that the shifts will get folded together. We could do this
6008 // xform in more cases, but it is unlikely to be profitable.
6009 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6010 isa<ConstantInt>(TrOp->getOperand(1))) {
6011 // Okay, we'll do this xform. Make the shift of shift.
6012 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6013 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6015 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6017 // For logical shifts, the truncation has the effect of making the high
6018 // part of the register be zeros. Emulate this by inserting an AND to
6019 // clear the top bits as needed. This 'and' will usually be zapped by
6020 // other xforms later if dead.
6021 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6022 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6023 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6025 // The mask we constructed says what the trunc would do if occurring
6026 // between the shifts. We want to know the effect *after* the second
6027 // shift. We know that it is a logical shift by a constant, so adjust the
6028 // mask as appropriate.
6029 if (I.getOpcode() == Instruction::Shl)
6030 MaskV <<= Op1->getZExtValue();
6032 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6033 MaskV = MaskV.lshr(Op1->getZExtValue());
6036 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6038 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6040 // Return the value truncated to the interesting size.
6041 return new TruncInst(And, I.getType());
6045 if (Op0->hasOneUse()) {
6046 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6047 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6050 switch (Op0BO->getOpcode()) {
6052 case Instruction::Add:
6053 case Instruction::And:
6054 case Instruction::Or:
6055 case Instruction::Xor: {
6056 // These operators commute.
6057 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6058 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6059 match(Op0BO->getOperand(1),
6060 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6061 Instruction *YS = BinaryOperator::createShl(
6062 Op0BO->getOperand(0), Op1,
6064 InsertNewInstBefore(YS, I); // (Y << C)
6066 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6067 Op0BO->getOperand(1)->getName());
6068 InsertNewInstBefore(X, I); // (X + (Y << C))
6069 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6070 return BinaryOperator::createAnd(X, ConstantInt::get(
6071 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6074 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6075 Value *Op0BOOp1 = Op0BO->getOperand(1);
6076 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6078 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6079 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6081 Instruction *YS = BinaryOperator::createShl(
6082 Op0BO->getOperand(0), Op1,
6084 InsertNewInstBefore(YS, I); // (Y << C)
6086 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6087 V1->getName()+".mask");
6088 InsertNewInstBefore(XM, I); // X & (CC << C)
6090 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6095 case Instruction::Sub: {
6096 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6097 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6098 match(Op0BO->getOperand(0),
6099 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6100 Instruction *YS = BinaryOperator::createShl(
6101 Op0BO->getOperand(1), Op1,
6103 InsertNewInstBefore(YS, I); // (Y << C)
6105 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6106 Op0BO->getOperand(0)->getName());
6107 InsertNewInstBefore(X, I); // (X + (Y << C))
6108 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6109 return BinaryOperator::createAnd(X, ConstantInt::get(
6110 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6113 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6114 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6115 match(Op0BO->getOperand(0),
6116 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6117 m_ConstantInt(CC))) && V2 == Op1 &&
6118 cast<BinaryOperator>(Op0BO->getOperand(0))
6119 ->getOperand(0)->hasOneUse()) {
6120 Instruction *YS = BinaryOperator::createShl(
6121 Op0BO->getOperand(1), Op1,
6123 InsertNewInstBefore(YS, I); // (Y << C)
6125 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6126 V1->getName()+".mask");
6127 InsertNewInstBefore(XM, I); // X & (CC << C)
6129 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6137 // If the operand is an bitwise operator with a constant RHS, and the
6138 // shift is the only use, we can pull it out of the shift.
6139 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6140 bool isValid = true; // Valid only for And, Or, Xor
6141 bool highBitSet = false; // Transform if high bit of constant set?
6143 switch (Op0BO->getOpcode()) {
6144 default: isValid = false; break; // Do not perform transform!
6145 case Instruction::Add:
6146 isValid = isLeftShift;
6148 case Instruction::Or:
6149 case Instruction::Xor:
6152 case Instruction::And:
6157 // If this is a signed shift right, and the high bit is modified
6158 // by the logical operation, do not perform the transformation.
6159 // The highBitSet boolean indicates the value of the high bit of
6160 // the constant which would cause it to be modified for this
6163 if (isValid && I.getOpcode() == Instruction::AShr)
6164 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6167 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6169 Instruction *NewShift =
6170 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6171 InsertNewInstBefore(NewShift, I);
6172 NewShift->takeName(Op0BO);
6174 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6181 // Find out if this is a shift of a shift by a constant.
6182 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6183 if (ShiftOp && !ShiftOp->isShift())
6186 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6187 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6188 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6189 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6190 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6191 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6192 Value *X = ShiftOp->getOperand(0);
6194 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6195 if (AmtSum > TypeBits)
6198 const IntegerType *Ty = cast<IntegerType>(I.getType());
6200 // Check for (X << c1) << c2 and (X >> c1) >> c2
6201 if (I.getOpcode() == ShiftOp->getOpcode()) {
6202 return BinaryOperator::create(I.getOpcode(), X,
6203 ConstantInt::get(Ty, AmtSum));
6204 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6205 I.getOpcode() == Instruction::AShr) {
6206 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6207 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6208 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6209 I.getOpcode() == Instruction::LShr) {
6210 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6211 Instruction *Shift =
6212 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6213 InsertNewInstBefore(Shift, I);
6215 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6216 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6219 // Okay, if we get here, one shift must be left, and the other shift must be
6220 // right. See if the amounts are equal.
6221 if (ShiftAmt1 == ShiftAmt2) {
6222 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6223 if (I.getOpcode() == Instruction::Shl) {
6224 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6225 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6227 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6228 if (I.getOpcode() == Instruction::LShr) {
6229 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6230 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6232 // We can simplify ((X << C) >>s C) into a trunc + sext.
6233 // NOTE: we could do this for any C, but that would make 'unusual' integer
6234 // types. For now, just stick to ones well-supported by the code
6236 const Type *SExtType = 0;
6237 switch (Ty->getBitWidth() - ShiftAmt1) {
6244 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6249 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6250 InsertNewInstBefore(NewTrunc, I);
6251 return new SExtInst(NewTrunc, Ty);
6253 // Otherwise, we can't handle it yet.
6254 } else if (ShiftAmt1 < ShiftAmt2) {
6255 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6257 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6258 if (I.getOpcode() == Instruction::Shl) {
6259 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6260 ShiftOp->getOpcode() == Instruction::AShr);
6261 Instruction *Shift =
6262 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6263 InsertNewInstBefore(Shift, I);
6265 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6266 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6269 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6270 if (I.getOpcode() == Instruction::LShr) {
6271 assert(ShiftOp->getOpcode() == Instruction::Shl);
6272 Instruction *Shift =
6273 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6274 InsertNewInstBefore(Shift, I);
6276 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6277 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6280 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6282 assert(ShiftAmt2 < ShiftAmt1);
6283 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6285 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6286 if (I.getOpcode() == Instruction::Shl) {
6287 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6288 ShiftOp->getOpcode() == Instruction::AShr);
6289 Instruction *Shift =
6290 BinaryOperator::create(ShiftOp->getOpcode(), X,
6291 ConstantInt::get(Ty, ShiftDiff));
6292 InsertNewInstBefore(Shift, I);
6294 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6295 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6298 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6299 if (I.getOpcode() == Instruction::LShr) {
6300 assert(ShiftOp->getOpcode() == Instruction::Shl);
6301 Instruction *Shift =
6302 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6303 InsertNewInstBefore(Shift, I);
6305 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6306 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6309 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6316 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6317 /// expression. If so, decompose it, returning some value X, such that Val is
6320 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6322 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6323 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6324 Offset = CI->getZExtValue();
6326 return ConstantInt::get(Type::Int32Ty, 0);
6327 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6328 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6329 if (I->getOpcode() == Instruction::Shl) {
6330 // This is a value scaled by '1 << the shift amt'.
6331 Scale = 1U << RHS->getZExtValue();
6333 return I->getOperand(0);
6334 } else if (I->getOpcode() == Instruction::Mul) {
6335 // This value is scaled by 'RHS'.
6336 Scale = RHS->getZExtValue();
6338 return I->getOperand(0);
6339 } else if (I->getOpcode() == Instruction::Add) {
6340 // We have X+C. Check to see if we really have (X*C2)+C1,
6341 // where C1 is divisible by C2.
6344 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6345 Offset += RHS->getZExtValue();
6352 // Otherwise, we can't look past this.
6359 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6360 /// try to eliminate the cast by moving the type information into the alloc.
6361 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6362 AllocationInst &AI) {
6363 const PointerType *PTy = cast<PointerType>(CI.getType());
6365 // Remove any uses of AI that are dead.
6366 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6368 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6369 Instruction *User = cast<Instruction>(*UI++);
6370 if (isInstructionTriviallyDead(User)) {
6371 while (UI != E && *UI == User)
6372 ++UI; // If this instruction uses AI more than once, don't break UI.
6375 DOUT << "IC: DCE: " << *User;
6376 EraseInstFromFunction(*User);
6380 // Get the type really allocated and the type casted to.
6381 const Type *AllocElTy = AI.getAllocatedType();
6382 const Type *CastElTy = PTy->getElementType();
6383 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6385 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6386 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6387 if (CastElTyAlign < AllocElTyAlign) return 0;
6389 // If the allocation has multiple uses, only promote it if we are strictly
6390 // increasing the alignment of the resultant allocation. If we keep it the
6391 // same, we open the door to infinite loops of various kinds.
6392 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6394 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6395 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6396 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6398 // See if we can satisfy the modulus by pulling a scale out of the array
6400 unsigned ArraySizeScale;
6402 Value *NumElements = // See if the array size is a decomposable linear expr.
6403 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6405 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6407 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6408 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6410 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6415 // If the allocation size is constant, form a constant mul expression
6416 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6417 if (isa<ConstantInt>(NumElements))
6418 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6419 // otherwise multiply the amount and the number of elements
6420 else if (Scale != 1) {
6421 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6422 Amt = InsertNewInstBefore(Tmp, AI);
6426 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6427 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6428 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6429 Amt = InsertNewInstBefore(Tmp, AI);
6432 AllocationInst *New;
6433 if (isa<MallocInst>(AI))
6434 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6436 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6437 InsertNewInstBefore(New, AI);
6440 // If the allocation has multiple uses, insert a cast and change all things
6441 // that used it to use the new cast. This will also hack on CI, but it will
6443 if (!AI.hasOneUse()) {
6444 AddUsesToWorkList(AI);
6445 // New is the allocation instruction, pointer typed. AI is the original
6446 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6447 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6448 InsertNewInstBefore(NewCast, AI);
6449 AI.replaceAllUsesWith(NewCast);
6451 return ReplaceInstUsesWith(CI, New);
6454 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6455 /// and return it as type Ty without inserting any new casts and without
6456 /// changing the computed value. This is used by code that tries to decide
6457 /// whether promoting or shrinking integer operations to wider or smaller types
6458 /// will allow us to eliminate a truncate or extend.
6460 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6461 /// extension operation if Ty is larger.
6462 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6463 unsigned CastOpc, int &NumCastsRemoved) {
6464 // We can always evaluate constants in another type.
6465 if (isa<ConstantInt>(V))
6468 Instruction *I = dyn_cast<Instruction>(V);
6469 if (!I) return false;
6471 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6473 // If this is an extension or truncate, we can often eliminate it.
6474 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6475 // If this is a cast from the destination type, we can trivially eliminate
6476 // it, and this will remove a cast overall.
6477 if (I->getOperand(0)->getType() == Ty) {
6478 // If the first operand is itself a cast, and is eliminable, do not count
6479 // this as an eliminable cast. We would prefer to eliminate those two
6481 if (!isa<CastInst>(I->getOperand(0)))
6487 // We can't extend or shrink something that has multiple uses: doing so would
6488 // require duplicating the instruction in general, which isn't profitable.
6489 if (!I->hasOneUse()) return false;
6491 switch (I->getOpcode()) {
6492 case Instruction::Add:
6493 case Instruction::Sub:
6494 case Instruction::And:
6495 case Instruction::Or:
6496 case Instruction::Xor:
6497 // These operators can all arbitrarily be extended or truncated.
6498 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6500 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6503 case Instruction::Shl:
6504 // If we are truncating the result of this SHL, and if it's a shift of a
6505 // constant amount, we can always perform a SHL in a smaller type.
6506 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6507 uint32_t BitWidth = Ty->getBitWidth();
6508 if (BitWidth < OrigTy->getBitWidth() &&
6509 CI->getLimitedValue(BitWidth) < BitWidth)
6510 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6514 case Instruction::LShr:
6515 // If this is a truncate of a logical shr, we can truncate it to a smaller
6516 // lshr iff we know that the bits we would otherwise be shifting in are
6518 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6519 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6520 uint32_t BitWidth = Ty->getBitWidth();
6521 if (BitWidth < OrigBitWidth &&
6522 MaskedValueIsZero(I->getOperand(0),
6523 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6524 CI->getLimitedValue(BitWidth) < BitWidth) {
6525 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6530 case Instruction::ZExt:
6531 case Instruction::SExt:
6532 case Instruction::Trunc:
6533 // If this is the same kind of case as our original (e.g. zext+zext), we
6534 // can safely replace it. Note that replacing it does not reduce the number
6535 // of casts in the input.
6536 if (I->getOpcode() == CastOpc)
6541 // TODO: Can handle more cases here.
6548 /// EvaluateInDifferentType - Given an expression that
6549 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6550 /// evaluate the expression.
6551 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6553 if (Constant *C = dyn_cast<Constant>(V))
6554 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6556 // Otherwise, it must be an instruction.
6557 Instruction *I = cast<Instruction>(V);
6558 Instruction *Res = 0;
6559 switch (I->getOpcode()) {
6560 case Instruction::Add:
6561 case Instruction::Sub:
6562 case Instruction::And:
6563 case Instruction::Or:
6564 case Instruction::Xor:
6565 case Instruction::AShr:
6566 case Instruction::LShr:
6567 case Instruction::Shl: {
6568 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6569 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6570 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6571 LHS, RHS, I->getName());
6574 case Instruction::Trunc:
6575 case Instruction::ZExt:
6576 case Instruction::SExt:
6577 // If the source type of the cast is the type we're trying for then we can
6578 // just return the source. There's no need to insert it because it is not
6580 if (I->getOperand(0)->getType() == Ty)
6581 return I->getOperand(0);
6583 // Otherwise, must be the same type of case, so just reinsert a new one.
6584 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6588 // TODO: Can handle more cases here.
6589 assert(0 && "Unreachable!");
6593 return InsertNewInstBefore(Res, *I);
6596 /// @brief Implement the transforms common to all CastInst visitors.
6597 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6598 Value *Src = CI.getOperand(0);
6600 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6601 // eliminate it now.
6602 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6603 if (Instruction::CastOps opc =
6604 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6605 // The first cast (CSrc) is eliminable so we need to fix up or replace
6606 // the second cast (CI). CSrc will then have a good chance of being dead.
6607 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6611 // If we are casting a select then fold the cast into the select
6612 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6613 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6616 // If we are casting a PHI then fold the cast into the PHI
6617 if (isa<PHINode>(Src))
6618 if (Instruction *NV = FoldOpIntoPhi(CI))
6624 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6625 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6626 Value *Src = CI.getOperand(0);
6628 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6629 // If casting the result of a getelementptr instruction with no offset, turn
6630 // this into a cast of the original pointer!
6631 if (GEP->hasAllZeroIndices()) {
6632 // Changing the cast operand is usually not a good idea but it is safe
6633 // here because the pointer operand is being replaced with another
6634 // pointer operand so the opcode doesn't need to change.
6636 CI.setOperand(0, GEP->getOperand(0));
6640 // If the GEP has a single use, and the base pointer is a bitcast, and the
6641 // GEP computes a constant offset, see if we can convert these three
6642 // instructions into fewer. This typically happens with unions and other
6643 // non-type-safe code.
6644 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6645 if (GEP->hasAllConstantIndices()) {
6646 // We are guaranteed to get a constant from EmitGEPOffset.
6647 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6648 int64_t Offset = OffsetV->getSExtValue();
6650 // Get the base pointer input of the bitcast, and the type it points to.
6651 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6652 const Type *GEPIdxTy =
6653 cast<PointerType>(OrigBase->getType())->getElementType();
6654 if (GEPIdxTy->isSized()) {
6655 SmallVector<Value*, 8> NewIndices;
6657 // Start with the index over the outer type. Note that the type size
6658 // might be zero (even if the offset isn't zero) if the indexed type
6659 // is something like [0 x {int, int}]
6660 const Type *IntPtrTy = TD->getIntPtrType();
6661 int64_t FirstIdx = 0;
6662 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6663 FirstIdx = Offset/TySize;
6666 // Handle silly modulus not returning values values [0..TySize).
6670 assert(Offset >= 0);
6672 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6675 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6677 // Index into the types. If we fail, set OrigBase to null.
6679 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6680 const StructLayout *SL = TD->getStructLayout(STy);
6681 if (Offset < (int64_t)SL->getSizeInBytes()) {
6682 unsigned Elt = SL->getElementContainingOffset(Offset);
6683 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6685 Offset -= SL->getElementOffset(Elt);
6686 GEPIdxTy = STy->getElementType(Elt);
6688 // Otherwise, we can't index into this, bail out.
6692 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6693 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6694 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6695 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6698 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6700 GEPIdxTy = STy->getElementType();
6702 // Otherwise, we can't index into this, bail out.
6708 // If we were able to index down into an element, create the GEP
6709 // and bitcast the result. This eliminates one bitcast, potentially
6711 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6713 NewIndices.end(), "");
6714 InsertNewInstBefore(NGEP, CI);
6715 NGEP->takeName(GEP);
6717 if (isa<BitCastInst>(CI))
6718 return new BitCastInst(NGEP, CI.getType());
6719 assert(isa<PtrToIntInst>(CI));
6720 return new PtrToIntInst(NGEP, CI.getType());
6727 return commonCastTransforms(CI);
6732 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6733 /// integer types. This function implements the common transforms for all those
6735 /// @brief Implement the transforms common to CastInst with integer operands
6736 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6737 if (Instruction *Result = commonCastTransforms(CI))
6740 Value *Src = CI.getOperand(0);
6741 const Type *SrcTy = Src->getType();
6742 const Type *DestTy = CI.getType();
6743 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6744 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6746 // See if we can simplify any instructions used by the LHS whose sole
6747 // purpose is to compute bits we don't care about.
6748 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6749 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6750 KnownZero, KnownOne))
6753 // If the source isn't an instruction or has more than one use then we
6754 // can't do anything more.
6755 Instruction *SrcI = dyn_cast<Instruction>(Src);
6756 if (!SrcI || !Src->hasOneUse())
6759 // Attempt to propagate the cast into the instruction for int->int casts.
6760 int NumCastsRemoved = 0;
6761 if (!isa<BitCastInst>(CI) &&
6762 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6763 CI.getOpcode(), NumCastsRemoved)) {
6764 // If this cast is a truncate, evaluting in a different type always
6765 // eliminates the cast, so it is always a win. If this is a zero-extension,
6766 // we need to do an AND to maintain the clear top-part of the computation,
6767 // so we require that the input have eliminated at least one cast. If this
6768 // is a sign extension, we insert two new casts (to do the extension) so we
6769 // require that two casts have been eliminated.
6771 switch (CI.getOpcode()) {
6773 // All the others use floating point so we shouldn't actually
6774 // get here because of the check above.
6775 assert(0 && "Unknown cast type");
6776 case Instruction::Trunc:
6779 case Instruction::ZExt:
6780 DoXForm = NumCastsRemoved >= 1;
6782 case Instruction::SExt:
6783 DoXForm = NumCastsRemoved >= 2;
6788 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6789 CI.getOpcode() == Instruction::SExt);
6790 assert(Res->getType() == DestTy);
6791 switch (CI.getOpcode()) {
6792 default: assert(0 && "Unknown cast type!");
6793 case Instruction::Trunc:
6794 case Instruction::BitCast:
6795 // Just replace this cast with the result.
6796 return ReplaceInstUsesWith(CI, Res);
6797 case Instruction::ZExt: {
6798 // We need to emit an AND to clear the high bits.
6799 assert(SrcBitSize < DestBitSize && "Not a zext?");
6800 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6802 return BinaryOperator::createAnd(Res, C);
6804 case Instruction::SExt:
6805 // We need to emit a cast to truncate, then a cast to sext.
6806 return CastInst::create(Instruction::SExt,
6807 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6813 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6814 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6816 switch (SrcI->getOpcode()) {
6817 case Instruction::Add:
6818 case Instruction::Mul:
6819 case Instruction::And:
6820 case Instruction::Or:
6821 case Instruction::Xor:
6822 // If we are discarding information, rewrite.
6823 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6824 // Don't insert two casts if they cannot be eliminated. We allow
6825 // two casts to be inserted if the sizes are the same. This could
6826 // only be converting signedness, which is a noop.
6827 if (DestBitSize == SrcBitSize ||
6828 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6829 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6830 Instruction::CastOps opcode = CI.getOpcode();
6831 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6832 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6833 return BinaryOperator::create(
6834 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6838 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6839 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6840 SrcI->getOpcode() == Instruction::Xor &&
6841 Op1 == ConstantInt::getTrue() &&
6842 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6843 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6844 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6847 case Instruction::SDiv:
6848 case Instruction::UDiv:
6849 case Instruction::SRem:
6850 case Instruction::URem:
6851 // If we are just changing the sign, rewrite.
6852 if (DestBitSize == SrcBitSize) {
6853 // Don't insert two casts if they cannot be eliminated. We allow
6854 // two casts to be inserted if the sizes are the same. This could
6855 // only be converting signedness, which is a noop.
6856 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6857 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6858 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6860 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6862 return BinaryOperator::create(
6863 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6868 case Instruction::Shl:
6869 // Allow changing the sign of the source operand. Do not allow
6870 // changing the size of the shift, UNLESS the shift amount is a
6871 // constant. We must not change variable sized shifts to a smaller
6872 // size, because it is undefined to shift more bits out than exist
6874 if (DestBitSize == SrcBitSize ||
6875 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6876 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6877 Instruction::BitCast : Instruction::Trunc);
6878 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6879 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6880 return BinaryOperator::createShl(Op0c, Op1c);
6883 case Instruction::AShr:
6884 // If this is a signed shr, and if all bits shifted in are about to be
6885 // truncated off, turn it into an unsigned shr to allow greater
6887 if (DestBitSize < SrcBitSize &&
6888 isa<ConstantInt>(Op1)) {
6889 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6890 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6891 // Insert the new logical shift right.
6892 return BinaryOperator::createLShr(Op0, Op1);
6900 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6901 if (Instruction *Result = commonIntCastTransforms(CI))
6904 Value *Src = CI.getOperand(0);
6905 const Type *Ty = CI.getType();
6906 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6907 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6909 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6910 switch (SrcI->getOpcode()) {
6912 case Instruction::LShr:
6913 // We can shrink lshr to something smaller if we know the bits shifted in
6914 // are already zeros.
6915 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6916 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6918 // Get a mask for the bits shifting in.
6919 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6920 Value* SrcIOp0 = SrcI->getOperand(0);
6921 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6922 if (ShAmt >= DestBitWidth) // All zeros.
6923 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6925 // Okay, we can shrink this. Truncate the input, then return a new
6927 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6928 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6930 return BinaryOperator::createLShr(V1, V2);
6932 } else { // This is a variable shr.
6934 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6935 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6936 // loop-invariant and CSE'd.
6937 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6938 Value *One = ConstantInt::get(SrcI->getType(), 1);
6940 Value *V = InsertNewInstBefore(
6941 BinaryOperator::createShl(One, SrcI->getOperand(1),
6943 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6944 SrcI->getOperand(0),
6946 Value *Zero = Constant::getNullValue(V->getType());
6947 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6957 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6958 // If one of the common conversion will work ..
6959 if (Instruction *Result = commonIntCastTransforms(CI))
6962 Value *Src = CI.getOperand(0);
6964 // If this is a cast of a cast
6965 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6966 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6967 // types and if the sizes are just right we can convert this into a logical
6968 // 'and' which will be much cheaper than the pair of casts.
6969 if (isa<TruncInst>(CSrc)) {
6970 // Get the sizes of the types involved
6971 Value *A = CSrc->getOperand(0);
6972 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6973 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6974 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6975 // If we're actually extending zero bits and the trunc is a no-op
6976 if (MidSize < DstSize && SrcSize == DstSize) {
6977 // Replace both of the casts with an And of the type mask.
6978 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6979 Constant *AndConst = ConstantInt::get(AndValue);
6981 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6982 // Unfortunately, if the type changed, we need to cast it back.
6983 if (And->getType() != CI.getType()) {
6984 And->setName(CSrc->getName()+".mask");
6985 InsertNewInstBefore(And, CI);
6986 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6993 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6994 // If we are just checking for a icmp eq of a single bit and zext'ing it
6995 // to an integer, then shift the bit to the appropriate place and then
6996 // cast to integer to avoid the comparison.
6997 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6998 const APInt &Op1CV = Op1C->getValue();
7000 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7001 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7002 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7003 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7004 Value *In = ICI->getOperand(0);
7005 Value *Sh = ConstantInt::get(In->getType(),
7006 In->getType()->getPrimitiveSizeInBits()-1);
7007 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7008 In->getName()+".lobit"),
7010 if (In->getType() != CI.getType())
7011 In = CastInst::createIntegerCast(In, CI.getType(),
7012 false/*ZExt*/, "tmp", &CI);
7014 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7015 Constant *One = ConstantInt::get(In->getType(), 1);
7016 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7017 In->getName()+".not"),
7021 return ReplaceInstUsesWith(CI, In);
7026 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7027 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7028 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7029 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7030 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7031 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7032 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7033 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7034 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7035 // This only works for EQ and NE
7036 ICI->isEquality()) {
7037 // If Op1C some other power of two, convert:
7038 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7039 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7040 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7041 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7043 APInt KnownZeroMask(~KnownZero);
7044 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7045 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7046 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7047 // (X&4) == 2 --> false
7048 // (X&4) != 2 --> true
7049 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7050 Res = ConstantExpr::getZExt(Res, CI.getType());
7051 return ReplaceInstUsesWith(CI, Res);
7054 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7055 Value *In = ICI->getOperand(0);
7057 // Perform a logical shr by shiftamt.
7058 // Insert the shift to put the result in the low bit.
7059 In = InsertNewInstBefore(
7060 BinaryOperator::createLShr(In,
7061 ConstantInt::get(In->getType(), ShiftAmt),
7062 In->getName()+".lobit"), CI);
7065 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7066 Constant *One = ConstantInt::get(In->getType(), 1);
7067 In = BinaryOperator::createXor(In, One, "tmp");
7068 InsertNewInstBefore(cast<Instruction>(In), CI);
7071 if (CI.getType() == In->getType())
7072 return ReplaceInstUsesWith(CI, In);
7074 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7082 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7083 if (Instruction *I = commonIntCastTransforms(CI))
7086 Value *Src = CI.getOperand(0);
7088 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7089 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7090 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7091 // If we are just checking for a icmp eq of a single bit and zext'ing it
7092 // to an integer, then shift the bit to the appropriate place and then
7093 // cast to integer to avoid the comparison.
7094 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7095 const APInt &Op1CV = Op1C->getValue();
7097 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7098 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7099 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7100 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7101 Value *In = ICI->getOperand(0);
7102 Value *Sh = ConstantInt::get(In->getType(),
7103 In->getType()->getPrimitiveSizeInBits()-1);
7104 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7105 In->getName()+".lobit"),
7107 if (In->getType() != CI.getType())
7108 In = CastInst::createIntegerCast(In, CI.getType(),
7109 true/*SExt*/, "tmp", &CI);
7111 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7112 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7113 In->getName()+".not"), CI);
7115 return ReplaceInstUsesWith(CI, In);
7123 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7124 return commonCastTransforms(CI);
7127 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7128 return commonCastTransforms(CI);
7131 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7132 return commonCastTransforms(CI);
7135 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7136 return commonCastTransforms(CI);
7139 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7140 return commonCastTransforms(CI);
7143 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7144 return commonCastTransforms(CI);
7147 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7148 return commonPointerCastTransforms(CI);
7151 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7152 return commonCastTransforms(CI);
7155 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7156 // If the operands are integer typed then apply the integer transforms,
7157 // otherwise just apply the common ones.
7158 Value *Src = CI.getOperand(0);
7159 const Type *SrcTy = Src->getType();
7160 const Type *DestTy = CI.getType();
7162 if (SrcTy->isInteger() && DestTy->isInteger()) {
7163 if (Instruction *Result = commonIntCastTransforms(CI))
7165 } else if (isa<PointerType>(SrcTy)) {
7166 if (Instruction *I = commonPointerCastTransforms(CI))
7169 if (Instruction *Result = commonCastTransforms(CI))
7174 // Get rid of casts from one type to the same type. These are useless and can
7175 // be replaced by the operand.
7176 if (DestTy == Src->getType())
7177 return ReplaceInstUsesWith(CI, Src);
7179 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7180 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7181 const Type *DstElTy = DstPTy->getElementType();
7182 const Type *SrcElTy = SrcPTy->getElementType();
7184 // If we are casting a malloc or alloca to a pointer to a type of the same
7185 // size, rewrite the allocation instruction to allocate the "right" type.
7186 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7187 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7190 // If the source and destination are pointers, and this cast is equivalent
7191 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7192 // This can enhance SROA and other transforms that want type-safe pointers.
7193 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7194 unsigned NumZeros = 0;
7195 while (SrcElTy != DstElTy &&
7196 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7197 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7198 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7202 // If we found a path from the src to dest, create the getelementptr now.
7203 if (SrcElTy == DstElTy) {
7204 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7205 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7206 ((Instruction*) NULL));
7210 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7211 if (SVI->hasOneUse()) {
7212 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7213 // a bitconvert to a vector with the same # elts.
7214 if (isa<VectorType>(DestTy) &&
7215 cast<VectorType>(DestTy)->getNumElements() ==
7216 SVI->getType()->getNumElements()) {
7218 // If either of the operands is a cast from CI.getType(), then
7219 // evaluating the shuffle in the casted destination's type will allow
7220 // us to eliminate at least one cast.
7221 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7222 Tmp->getOperand(0)->getType() == DestTy) ||
7223 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7224 Tmp->getOperand(0)->getType() == DestTy)) {
7225 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7226 SVI->getOperand(0), DestTy, &CI);
7227 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7228 SVI->getOperand(1), DestTy, &CI);
7229 // Return a new shuffle vector. Use the same element ID's, as we
7230 // know the vector types match #elts.
7231 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7239 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7241 /// %D = select %cond, %C, %A
7243 /// %C = select %cond, %B, 0
7246 /// Assuming that the specified instruction is an operand to the select, return
7247 /// a bitmask indicating which operands of this instruction are foldable if they
7248 /// equal the other incoming value of the select.
7250 static unsigned GetSelectFoldableOperands(Instruction *I) {
7251 switch (I->getOpcode()) {
7252 case Instruction::Add:
7253 case Instruction::Mul:
7254 case Instruction::And:
7255 case Instruction::Or:
7256 case Instruction::Xor:
7257 return 3; // Can fold through either operand.
7258 case Instruction::Sub: // Can only fold on the amount subtracted.
7259 case Instruction::Shl: // Can only fold on the shift amount.
7260 case Instruction::LShr:
7261 case Instruction::AShr:
7264 return 0; // Cannot fold
7268 /// GetSelectFoldableConstant - For the same transformation as the previous
7269 /// function, return the identity constant that goes into the select.
7270 static Constant *GetSelectFoldableConstant(Instruction *I) {
7271 switch (I->getOpcode()) {
7272 default: assert(0 && "This cannot happen!"); abort();
7273 case Instruction::Add:
7274 case Instruction::Sub:
7275 case Instruction::Or:
7276 case Instruction::Xor:
7277 case Instruction::Shl:
7278 case Instruction::LShr:
7279 case Instruction::AShr:
7280 return Constant::getNullValue(I->getType());
7281 case Instruction::And:
7282 return Constant::getAllOnesValue(I->getType());
7283 case Instruction::Mul:
7284 return ConstantInt::get(I->getType(), 1);
7288 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7289 /// have the same opcode and only one use each. Try to simplify this.
7290 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7292 if (TI->getNumOperands() == 1) {
7293 // If this is a non-volatile load or a cast from the same type,
7296 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7299 return 0; // unknown unary op.
7302 // Fold this by inserting a select from the input values.
7303 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7304 FI->getOperand(0), SI.getName()+".v");
7305 InsertNewInstBefore(NewSI, SI);
7306 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7310 // Only handle binary operators here.
7311 if (!isa<BinaryOperator>(TI))
7314 // Figure out if the operations have any operands in common.
7315 Value *MatchOp, *OtherOpT, *OtherOpF;
7317 if (TI->getOperand(0) == FI->getOperand(0)) {
7318 MatchOp = TI->getOperand(0);
7319 OtherOpT = TI->getOperand(1);
7320 OtherOpF = FI->getOperand(1);
7321 MatchIsOpZero = true;
7322 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7323 MatchOp = TI->getOperand(1);
7324 OtherOpT = TI->getOperand(0);
7325 OtherOpF = FI->getOperand(0);
7326 MatchIsOpZero = false;
7327 } else if (!TI->isCommutative()) {
7329 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7330 MatchOp = TI->getOperand(0);
7331 OtherOpT = TI->getOperand(1);
7332 OtherOpF = FI->getOperand(0);
7333 MatchIsOpZero = true;
7334 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7335 MatchOp = TI->getOperand(1);
7336 OtherOpT = TI->getOperand(0);
7337 OtherOpF = FI->getOperand(1);
7338 MatchIsOpZero = true;
7343 // If we reach here, they do have operations in common.
7344 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7345 OtherOpF, SI.getName()+".v");
7346 InsertNewInstBefore(NewSI, SI);
7348 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7350 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7352 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7354 assert(0 && "Shouldn't get here");
7358 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7359 Value *CondVal = SI.getCondition();
7360 Value *TrueVal = SI.getTrueValue();
7361 Value *FalseVal = SI.getFalseValue();
7363 // select true, X, Y -> X
7364 // select false, X, Y -> Y
7365 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7366 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7368 // select C, X, X -> X
7369 if (TrueVal == FalseVal)
7370 return ReplaceInstUsesWith(SI, TrueVal);
7372 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7373 return ReplaceInstUsesWith(SI, FalseVal);
7374 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7375 return ReplaceInstUsesWith(SI, TrueVal);
7376 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7377 if (isa<Constant>(TrueVal))
7378 return ReplaceInstUsesWith(SI, TrueVal);
7380 return ReplaceInstUsesWith(SI, FalseVal);
7383 if (SI.getType() == Type::Int1Ty) {
7384 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7385 if (C->getZExtValue()) {
7386 // Change: A = select B, true, C --> A = or B, C
7387 return BinaryOperator::createOr(CondVal, FalseVal);
7389 // Change: A = select B, false, C --> A = and !B, C
7391 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7392 "not."+CondVal->getName()), SI);
7393 return BinaryOperator::createAnd(NotCond, FalseVal);
7395 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7396 if (C->getZExtValue() == false) {
7397 // Change: A = select B, C, false --> A = and B, C
7398 return BinaryOperator::createAnd(CondVal, TrueVal);
7400 // Change: A = select B, C, true --> A = or !B, C
7402 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7403 "not."+CondVal->getName()), SI);
7404 return BinaryOperator::createOr(NotCond, TrueVal);
7408 // select a, b, a -> a&b
7409 // select a, a, b -> a|b
7410 if (CondVal == TrueVal)
7411 return BinaryOperator::createOr(CondVal, FalseVal);
7412 else if (CondVal == FalseVal)
7413 return BinaryOperator::createAnd(CondVal, TrueVal);
7416 // Selecting between two integer constants?
7417 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7418 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7419 // select C, 1, 0 -> zext C to int
7420 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7421 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7422 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7423 // select C, 0, 1 -> zext !C to int
7425 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7426 "not."+CondVal->getName()), SI);
7427 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7430 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7432 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7434 // (x <s 0) ? -1 : 0 -> ashr x, 31
7435 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7436 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7437 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7438 // The comparison constant and the result are not neccessarily the
7439 // same width. Make an all-ones value by inserting a AShr.
7440 Value *X = IC->getOperand(0);
7441 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7442 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7443 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7445 InsertNewInstBefore(SRA, SI);
7447 // Finally, convert to the type of the select RHS. We figure out
7448 // if this requires a SExt, Trunc or BitCast based on the sizes.
7449 Instruction::CastOps opc = Instruction::BitCast;
7450 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7451 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7452 if (SRASize < SISize)
7453 opc = Instruction::SExt;
7454 else if (SRASize > SISize)
7455 opc = Instruction::Trunc;
7456 return CastInst::create(opc, SRA, SI.getType());
7461 // If one of the constants is zero (we know they can't both be) and we
7462 // have an icmp instruction with zero, and we have an 'and' with the
7463 // non-constant value, eliminate this whole mess. This corresponds to
7464 // cases like this: ((X & 27) ? 27 : 0)
7465 if (TrueValC->isZero() || FalseValC->isZero())
7466 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7467 cast<Constant>(IC->getOperand(1))->isNullValue())
7468 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7469 if (ICA->getOpcode() == Instruction::And &&
7470 isa<ConstantInt>(ICA->getOperand(1)) &&
7471 (ICA->getOperand(1) == TrueValC ||
7472 ICA->getOperand(1) == FalseValC) &&
7473 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7474 // Okay, now we know that everything is set up, we just don't
7475 // know whether we have a icmp_ne or icmp_eq and whether the
7476 // true or false val is the zero.
7477 bool ShouldNotVal = !TrueValC->isZero();
7478 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7481 V = InsertNewInstBefore(BinaryOperator::create(
7482 Instruction::Xor, V, ICA->getOperand(1)), SI);
7483 return ReplaceInstUsesWith(SI, V);
7488 // See if we are selecting two values based on a comparison of the two values.
7489 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7490 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7491 // Transform (X == Y) ? X : Y -> Y
7492 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7493 // This is not safe in general for floating point:
7494 // consider X== -0, Y== +0.
7495 // It becomes safe if either operand is a nonzero constant.
7496 ConstantFP *CFPt, *CFPf;
7497 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7498 !CFPt->getValueAPF().isZero()) ||
7499 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7500 !CFPf->getValueAPF().isZero()))
7501 return ReplaceInstUsesWith(SI, FalseVal);
7503 // Transform (X != Y) ? X : Y -> X
7504 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7505 return ReplaceInstUsesWith(SI, TrueVal);
7506 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7508 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7509 // Transform (X == Y) ? Y : X -> X
7510 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7511 // This is not safe in general for floating point:
7512 // consider X== -0, Y== +0.
7513 // It becomes safe if either operand is a nonzero constant.
7514 ConstantFP *CFPt, *CFPf;
7515 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7516 !CFPt->getValueAPF().isZero()) ||
7517 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7518 !CFPf->getValueAPF().isZero()))
7519 return ReplaceInstUsesWith(SI, FalseVal);
7521 // Transform (X != Y) ? Y : X -> Y
7522 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7523 return ReplaceInstUsesWith(SI, TrueVal);
7524 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7528 // See if we are selecting two values based on a comparison of the two values.
7529 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7530 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7531 // Transform (X == Y) ? X : Y -> Y
7532 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7533 return ReplaceInstUsesWith(SI, FalseVal);
7534 // Transform (X != Y) ? X : Y -> X
7535 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7536 return ReplaceInstUsesWith(SI, TrueVal);
7537 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7539 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7540 // Transform (X == Y) ? Y : X -> X
7541 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7542 return ReplaceInstUsesWith(SI, FalseVal);
7543 // Transform (X != Y) ? Y : X -> Y
7544 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7545 return ReplaceInstUsesWith(SI, TrueVal);
7546 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7550 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7551 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7552 if (TI->hasOneUse() && FI->hasOneUse()) {
7553 Instruction *AddOp = 0, *SubOp = 0;
7555 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7556 if (TI->getOpcode() == FI->getOpcode())
7557 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7560 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7561 // even legal for FP.
7562 if (TI->getOpcode() == Instruction::Sub &&
7563 FI->getOpcode() == Instruction::Add) {
7564 AddOp = FI; SubOp = TI;
7565 } else if (FI->getOpcode() == Instruction::Sub &&
7566 TI->getOpcode() == Instruction::Add) {
7567 AddOp = TI; SubOp = FI;
7571 Value *OtherAddOp = 0;
7572 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7573 OtherAddOp = AddOp->getOperand(1);
7574 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7575 OtherAddOp = AddOp->getOperand(0);
7579 // So at this point we know we have (Y -> OtherAddOp):
7580 // select C, (add X, Y), (sub X, Z)
7581 Value *NegVal; // Compute -Z
7582 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7583 NegVal = ConstantExpr::getNeg(C);
7585 NegVal = InsertNewInstBefore(
7586 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7589 Value *NewTrueOp = OtherAddOp;
7590 Value *NewFalseOp = NegVal;
7592 std::swap(NewTrueOp, NewFalseOp);
7593 Instruction *NewSel =
7594 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7596 NewSel = InsertNewInstBefore(NewSel, SI);
7597 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7602 // See if we can fold the select into one of our operands.
7603 if (SI.getType()->isInteger()) {
7604 // See the comment above GetSelectFoldableOperands for a description of the
7605 // transformation we are doing here.
7606 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7607 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7608 !isa<Constant>(FalseVal))
7609 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7610 unsigned OpToFold = 0;
7611 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7613 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7618 Constant *C = GetSelectFoldableConstant(TVI);
7619 Instruction *NewSel =
7620 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7621 InsertNewInstBefore(NewSel, SI);
7622 NewSel->takeName(TVI);
7623 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7624 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7626 assert(0 && "Unknown instruction!!");
7631 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7632 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7633 !isa<Constant>(TrueVal))
7634 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7635 unsigned OpToFold = 0;
7636 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7638 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7643 Constant *C = GetSelectFoldableConstant(FVI);
7644 Instruction *NewSel =
7645 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7646 InsertNewInstBefore(NewSel, SI);
7647 NewSel->takeName(FVI);
7648 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7649 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7651 assert(0 && "Unknown instruction!!");
7656 if (BinaryOperator::isNot(CondVal)) {
7657 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7658 SI.setOperand(1, FalseVal);
7659 SI.setOperand(2, TrueVal);
7666 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7667 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7668 /// and it is more than the alignment of the ultimate object, see if we can
7669 /// increase the alignment of the ultimate object, making this check succeed.
7670 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7671 unsigned PrefAlign = 0) {
7672 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7673 unsigned Align = GV->getAlignment();
7674 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7675 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7677 // If there is a large requested alignment and we can, bump up the alignment
7679 if (PrefAlign > Align && GV->hasInitializer()) {
7680 GV->setAlignment(PrefAlign);
7684 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7685 unsigned Align = AI->getAlignment();
7686 if (Align == 0 && TD) {
7687 if (isa<AllocaInst>(AI))
7688 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7689 else if (isa<MallocInst>(AI)) {
7690 // Malloc returns maximally aligned memory.
7691 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7694 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7697 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7701 // If there is a requested alignment and if this is an alloca, round up. We
7702 // don't do this for malloc, because some systems can't respect the request.
7703 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7704 AI->setAlignment(PrefAlign);
7708 } else if (isa<BitCastInst>(V) ||
7709 (isa<ConstantExpr>(V) &&
7710 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7711 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7713 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7714 // If all indexes are zero, it is just the alignment of the base pointer.
7715 bool AllZeroOperands = true;
7716 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7717 if (!isa<Constant>(GEPI->getOperand(i)) ||
7718 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7719 AllZeroOperands = false;
7723 if (AllZeroOperands) {
7724 // Treat this like a bitcast.
7725 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7728 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7729 if (BaseAlignment == 0) return 0;
7731 // Otherwise, if the base alignment is >= the alignment we expect for the
7732 // base pointer type, then we know that the resultant pointer is aligned at
7733 // least as much as its type requires.
7736 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7737 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7738 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7739 if (Align <= BaseAlignment) {
7740 const Type *GEPTy = GEPI->getType();
7741 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7742 Align = std::min(Align, (unsigned)
7743 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7752 /// visitCallInst - CallInst simplification. This mostly only handles folding
7753 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7754 /// the heavy lifting.
7756 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7757 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7758 if (!II) return visitCallSite(&CI);
7760 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7762 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7763 bool Changed = false;
7765 // memmove/cpy/set of zero bytes is a noop.
7766 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7767 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7769 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7770 if (CI->getZExtValue() == 1) {
7771 // Replace the instruction with just byte operations. We would
7772 // transform other cases to loads/stores, but we don't know if
7773 // alignment is sufficient.
7777 // If we have a memmove and the source operation is a constant global,
7778 // then the source and dest pointers can't alias, so we can change this
7779 // into a call to memcpy.
7780 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7781 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7782 if (GVSrc->isConstant()) {
7783 Module *M = CI.getParent()->getParent()->getParent();
7785 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7787 Name = "llvm.memcpy.i32";
7789 Name = "llvm.memcpy.i64";
7790 Constant *MemCpy = M->getOrInsertFunction(Name,
7791 CI.getCalledFunction()->getFunctionType());
7792 CI.setOperand(0, MemCpy);
7797 // If we can determine a pointer alignment that is bigger than currently
7798 // set, update the alignment.
7799 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7800 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7801 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7802 unsigned Align = std::min(Alignment1, Alignment2);
7803 if (MI->getAlignment()->getZExtValue() < Align) {
7804 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7808 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7810 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(CI.getOperand(3));
7812 unsigned Size = MemOpLength->getZExtValue();
7813 unsigned Align = cast<ConstantInt>(CI.getOperand(4))->getZExtValue();
7814 PointerType *NewPtrTy = NULL;
7815 // Destination pointer type is always i8 *
7816 // If Size is 8 then use Int64Ty
7817 // If Size is 4 then use Int32Ty
7818 // If Size is 2 then use Int16Ty
7819 // If Size is 1 then use Int8Ty
7820 if (Size && Size <=8 && !(Size&(Size-1)))
7821 NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7824 Value *Src = InsertCastBefore(Instruction::BitCast, CI.getOperand(2),
7826 Value *Dest = InsertCastBefore(Instruction::BitCast, CI.getOperand(1),
7828 Value *L = new LoadInst(Src, "tmp", false, Align, &CI);
7829 Value *NS = new StoreInst(L, Dest, false, Align, &CI);
7830 CI.replaceAllUsesWith(NS);
7832 return EraseInstFromFunction(CI);
7835 } else if (isa<MemSetInst>(MI)) {
7836 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7837 if (MI->getAlignment()->getZExtValue() < Alignment) {
7838 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7843 if (Changed) return II;
7845 switch (II->getIntrinsicID()) {
7847 case Intrinsic::ppc_altivec_lvx:
7848 case Intrinsic::ppc_altivec_lvxl:
7849 case Intrinsic::x86_sse_loadu_ps:
7850 case Intrinsic::x86_sse2_loadu_pd:
7851 case Intrinsic::x86_sse2_loadu_dq:
7852 // Turn PPC lvx -> load if the pointer is known aligned.
7853 // Turn X86 loadups -> load if the pointer is known aligned.
7854 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7856 InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7857 PointerType::getUnqual(II->getType()), CI);
7858 return new LoadInst(Ptr);
7861 case Intrinsic::ppc_altivec_stvx:
7862 case Intrinsic::ppc_altivec_stvxl:
7863 // Turn stvx -> store if the pointer is known aligned.
7864 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7865 const Type *OpPtrTy =
7866 PointerType::getUnqual(II->getOperand(1)->getType());
7867 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7869 return new StoreInst(II->getOperand(1), Ptr);
7872 case Intrinsic::x86_sse_storeu_ps:
7873 case Intrinsic::x86_sse2_storeu_pd:
7874 case Intrinsic::x86_sse2_storeu_dq:
7875 case Intrinsic::x86_sse2_storel_dq:
7876 // Turn X86 storeu -> store if the pointer is known aligned.
7877 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7878 const Type *OpPtrTy =
7879 PointerType::getUnqual(II->getOperand(2)->getType());
7880 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7882 return new StoreInst(II->getOperand(2), Ptr);
7886 case Intrinsic::x86_sse_cvttss2si: {
7887 // These intrinsics only demands the 0th element of its input vector. If
7888 // we can simplify the input based on that, do so now.
7890 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7892 II->setOperand(1, V);
7898 case Intrinsic::ppc_altivec_vperm:
7899 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7900 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7901 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7903 // Check that all of the elements are integer constants or undefs.
7904 bool AllEltsOk = true;
7905 for (unsigned i = 0; i != 16; ++i) {
7906 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7907 !isa<UndefValue>(Mask->getOperand(i))) {
7914 // Cast the input vectors to byte vectors.
7915 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7916 II->getOperand(1), Mask->getType(), CI);
7917 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7918 II->getOperand(2), Mask->getType(), CI);
7919 Value *Result = UndefValue::get(Op0->getType());
7921 // Only extract each element once.
7922 Value *ExtractedElts[32];
7923 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7925 for (unsigned i = 0; i != 16; ++i) {
7926 if (isa<UndefValue>(Mask->getOperand(i)))
7928 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7929 Idx &= 31; // Match the hardware behavior.
7931 if (ExtractedElts[Idx] == 0) {
7933 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7934 InsertNewInstBefore(Elt, CI);
7935 ExtractedElts[Idx] = Elt;
7938 // Insert this value into the result vector.
7939 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7940 InsertNewInstBefore(cast<Instruction>(Result), CI);
7942 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7947 case Intrinsic::stackrestore: {
7948 // If the save is right next to the restore, remove the restore. This can
7949 // happen when variable allocas are DCE'd.
7950 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7951 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7952 BasicBlock::iterator BI = SS;
7954 return EraseInstFromFunction(CI);
7958 // If the stack restore is in a return/unwind block and if there are no
7959 // allocas or calls between the restore and the return, nuke the restore.
7960 TerminatorInst *TI = II->getParent()->getTerminator();
7961 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7962 BasicBlock::iterator BI = II;
7963 bool CannotRemove = false;
7964 for (++BI; &*BI != TI; ++BI) {
7965 if (isa<AllocaInst>(BI) ||
7966 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7967 CannotRemove = true;
7972 return EraseInstFromFunction(CI);
7979 return visitCallSite(II);
7982 // InvokeInst simplification
7984 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7985 return visitCallSite(&II);
7988 // visitCallSite - Improvements for call and invoke instructions.
7990 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7991 bool Changed = false;
7993 // If the callee is a constexpr cast of a function, attempt to move the cast
7994 // to the arguments of the call/invoke.
7995 if (transformConstExprCastCall(CS)) return 0;
7997 Value *Callee = CS.getCalledValue();
7999 if (Function *CalleeF = dyn_cast<Function>(Callee))
8000 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8001 Instruction *OldCall = CS.getInstruction();
8002 // If the call and callee calling conventions don't match, this call must
8003 // be unreachable, as the call is undefined.
8004 new StoreInst(ConstantInt::getTrue(),
8005 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8007 if (!OldCall->use_empty())
8008 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8009 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8010 return EraseInstFromFunction(*OldCall);
8014 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8015 // This instruction is not reachable, just remove it. We insert a store to
8016 // undef so that we know that this code is not reachable, despite the fact
8017 // that we can't modify the CFG here.
8018 new StoreInst(ConstantInt::getTrue(),
8019 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8020 CS.getInstruction());
8022 if (!CS.getInstruction()->use_empty())
8023 CS.getInstruction()->
8024 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8026 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8027 // Don't break the CFG, insert a dummy cond branch.
8028 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8029 ConstantInt::getTrue(), II);
8031 return EraseInstFromFunction(*CS.getInstruction());
8034 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8035 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8036 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8037 return transformCallThroughTrampoline(CS);
8039 const PointerType *PTy = cast<PointerType>(Callee->getType());
8040 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8041 if (FTy->isVarArg()) {
8042 // See if we can optimize any arguments passed through the varargs area of
8044 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8045 E = CS.arg_end(); I != E; ++I)
8046 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8047 // If this cast does not effect the value passed through the varargs
8048 // area, we can eliminate the use of the cast.
8049 Value *Op = CI->getOperand(0);
8050 if (CI->isLosslessCast()) {
8057 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8058 // Inline asm calls cannot throw - mark them 'nounwind'.
8059 CS.setDoesNotThrow();
8063 return Changed ? CS.getInstruction() : 0;
8066 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8067 // attempt to move the cast to the arguments of the call/invoke.
8069 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8070 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8071 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8072 if (CE->getOpcode() != Instruction::BitCast ||
8073 !isa<Function>(CE->getOperand(0)))
8075 Function *Callee = cast<Function>(CE->getOperand(0));
8076 Instruction *Caller = CS.getInstruction();
8078 // Okay, this is a cast from a function to a different type. Unless doing so
8079 // would cause a type conversion of one of our arguments, change this call to
8080 // be a direct call with arguments casted to the appropriate types.
8082 const FunctionType *FT = Callee->getFunctionType();
8083 const Type *OldRetTy = Caller->getType();
8085 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8087 // If the parameter attributes are not compatible, don't do the xform. We
8088 // don't want to lose an sret attribute or something.
8089 if (!ParamAttrsList::areCompatible(CallerPAL, Callee->getParamAttrs()))
8092 // Check to see if we are changing the return type...
8093 if (OldRetTy != FT->getReturnType()) {
8094 if (Callee->isDeclaration() && !Caller->use_empty() &&
8095 // Conversion is ok if changing from pointer to int of same size.
8096 !(isa<PointerType>(FT->getReturnType()) &&
8097 TD->getIntPtrType() == OldRetTy))
8098 return false; // Cannot transform this return value.
8100 if (!Caller->use_empty() &&
8101 !CastInst::isCastable(FT->getReturnType(), OldRetTy) &&
8102 // void -> non-void is handled specially
8103 FT->getReturnType() != Type::VoidTy)
8104 return false; // Cannot transform this return value.
8106 // If the callsite is an invoke instruction, and the return value is used by
8107 // a PHI node in a successor, we cannot change the return type of the call
8108 // because there is no place to put the cast instruction (without breaking
8109 // the critical edge). Bail out in this case.
8110 if (!Caller->use_empty())
8111 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8112 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8114 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8115 if (PN->getParent() == II->getNormalDest() ||
8116 PN->getParent() == II->getUnwindDest())
8120 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8121 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8123 CallSite::arg_iterator AI = CS.arg_begin();
8124 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8125 const Type *ParamTy = FT->getParamType(i);
8126 const Type *ActTy = (*AI)->getType();
8128 if (!CastInst::isCastable(ActTy, ParamTy))
8131 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8132 // Some conversions are safe even if we do not have a body.
8133 // Either we can cast directly, or we can upconvert the argument
8134 bool isConvertible = ActTy == ParamTy ||
8135 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8136 (ParamTy->isInteger() && ActTy->isInteger() &&
8137 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8138 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8139 && c->getValue().isStrictlyPositive());
8140 if (Callee->isDeclaration() && !isConvertible) return false;
8143 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8144 Callee->isDeclaration())
8145 return false; // Do not delete arguments unless we have a function body...
8147 // Okay, we decided that this is a safe thing to do: go ahead and start
8148 // inserting cast instructions as necessary...
8149 std::vector<Value*> Args;
8150 Args.reserve(NumActualArgs);
8152 AI = CS.arg_begin();
8153 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8154 const Type *ParamTy = FT->getParamType(i);
8155 if ((*AI)->getType() == ParamTy) {
8156 Args.push_back(*AI);
8158 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8159 false, ParamTy, false);
8160 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8161 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8165 // If the function takes more arguments than the call was taking, add them
8167 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8168 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8170 // If we are removing arguments to the function, emit an obnoxious warning...
8171 if (FT->getNumParams() < NumActualArgs)
8172 if (!FT->isVarArg()) {
8173 cerr << "WARNING: While resolving call to function '"
8174 << Callee->getName() << "' arguments were dropped!\n";
8176 // Add all of the arguments in their promoted form to the arg list...
8177 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8178 const Type *PTy = getPromotedType((*AI)->getType());
8179 if (PTy != (*AI)->getType()) {
8180 // Must promote to pass through va_arg area!
8181 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8183 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8184 InsertNewInstBefore(Cast, *Caller);
8185 Args.push_back(Cast);
8187 Args.push_back(*AI);
8192 if (FT->getReturnType() == Type::VoidTy)
8193 Caller->setName(""); // Void type should not have a name.
8196 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8197 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8198 Args.begin(), Args.end(), Caller->getName(), Caller);
8199 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8200 cast<InvokeInst>(NC)->setParamAttrs(CallerPAL);
8202 NC = new CallInst(Callee, Args.begin(), Args.end(),
8203 Caller->getName(), Caller);
8204 CallInst *CI = cast<CallInst>(Caller);
8205 if (CI->isTailCall())
8206 cast<CallInst>(NC)->setTailCall();
8207 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8208 cast<CallInst>(NC)->setParamAttrs(CallerPAL);
8211 // Insert a cast of the return type as necessary.
8213 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8214 if (NV->getType() != Type::VoidTy) {
8215 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8217 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8219 // If this is an invoke instruction, we should insert it after the first
8220 // non-phi, instruction in the normal successor block.
8221 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8222 BasicBlock::iterator I = II->getNormalDest()->begin();
8223 while (isa<PHINode>(I)) ++I;
8224 InsertNewInstBefore(NC, *I);
8226 // Otherwise, it's a call, just insert cast right after the call instr
8227 InsertNewInstBefore(NC, *Caller);
8229 AddUsersToWorkList(*Caller);
8231 NV = UndefValue::get(Caller->getType());
8235 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8236 Caller->replaceAllUsesWith(NV);
8237 Caller->eraseFromParent();
8238 RemoveFromWorkList(Caller);
8242 // transformCallThroughTrampoline - Turn a call to a function created by the
8243 // init_trampoline intrinsic into a direct call to the underlying function.
8245 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8246 Value *Callee = CS.getCalledValue();
8247 const PointerType *PTy = cast<PointerType>(Callee->getType());
8248 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8250 IntrinsicInst *Tramp =
8251 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8254 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8255 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8256 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8258 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8259 unsigned NestIdx = 1;
8260 const Type *NestTy = 0;
8261 uint16_t NestAttr = 0;
8263 // Look for a parameter marked with the 'nest' attribute.
8264 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8265 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8266 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8267 // Record the parameter type and any other attributes.
8269 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8274 Instruction *Caller = CS.getInstruction();
8275 std::vector<Value*> NewArgs;
8276 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8278 // Insert the nest argument into the call argument list, which may
8279 // mean appending it.
8282 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8284 if (Idx == NestIdx) {
8285 // Add the chain argument.
8286 Value *NestVal = Tramp->getOperand(3);
8287 if (NestVal->getType() != NestTy)
8288 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8289 NewArgs.push_back(NestVal);
8295 // Add the original argument.
8296 NewArgs.push_back(*I);
8302 // The trampoline may have been bitcast to a bogus type (FTy).
8303 // Handle this by synthesizing a new function type, equal to FTy
8304 // with the chain parameter inserted. Likewise for attributes.
8306 const ParamAttrsList *Attrs = CS.getParamAttrs();
8307 std::vector<const Type*> NewTypes;
8308 ParamAttrsVector NewAttrs;
8309 NewTypes.reserve(FTy->getNumParams()+1);
8311 // Add any function result attributes.
8312 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8314 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8316 // Insert the chain's type into the list of parameter types, which may
8317 // mean appending it. Likewise for the chain's attributes.
8320 FunctionType::param_iterator I = FTy->param_begin(),
8321 E = FTy->param_end();
8324 if (Idx == NestIdx) {
8325 // Add the chain's type and attributes.
8326 NewTypes.push_back(NestTy);
8327 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8333 // Add the original type and attributes.
8334 NewTypes.push_back(*I);
8335 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8338 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8344 // Replace the trampoline call with a direct call. Let the generic
8345 // code sort out any function type mismatches.
8346 FunctionType *NewFTy =
8347 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8348 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8349 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8350 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8352 Instruction *NewCaller;
8353 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8354 NewCaller = new InvokeInst(NewCallee,
8355 II->getNormalDest(), II->getUnwindDest(),
8356 NewArgs.begin(), NewArgs.end(),
8357 Caller->getName(), Caller);
8358 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8359 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8361 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8362 Caller->getName(), Caller);
8363 if (cast<CallInst>(Caller)->isTailCall())
8364 cast<CallInst>(NewCaller)->setTailCall();
8365 cast<CallInst>(NewCaller)->
8366 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8367 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8369 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8370 Caller->replaceAllUsesWith(NewCaller);
8371 Caller->eraseFromParent();
8372 RemoveFromWorkList(Caller);
8377 // Replace the trampoline call with a direct call. Since there is no 'nest'
8378 // parameter, there is no need to adjust the argument list. Let the generic
8379 // code sort out any function type mismatches.
8380 Constant *NewCallee =
8381 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8382 CS.setCalledFunction(NewCallee);
8383 return CS.getInstruction();
8386 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8387 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8388 /// and a single binop.
8389 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8390 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8391 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8392 isa<CmpInst>(FirstInst));
8393 unsigned Opc = FirstInst->getOpcode();
8394 Value *LHSVal = FirstInst->getOperand(0);
8395 Value *RHSVal = FirstInst->getOperand(1);
8397 const Type *LHSType = LHSVal->getType();
8398 const Type *RHSType = RHSVal->getType();
8400 // Scan to see if all operands are the same opcode, all have one use, and all
8401 // kill their operands (i.e. the operands have one use).
8402 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8403 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8404 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8405 // Verify type of the LHS matches so we don't fold cmp's of different
8406 // types or GEP's with different index types.
8407 I->getOperand(0)->getType() != LHSType ||
8408 I->getOperand(1)->getType() != RHSType)
8411 // If they are CmpInst instructions, check their predicates
8412 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8413 if (cast<CmpInst>(I)->getPredicate() !=
8414 cast<CmpInst>(FirstInst)->getPredicate())
8417 // Keep track of which operand needs a phi node.
8418 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8419 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8422 // Otherwise, this is safe to transform, determine if it is profitable.
8424 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8425 // Indexes are often folded into load/store instructions, so we don't want to
8426 // hide them behind a phi.
8427 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8430 Value *InLHS = FirstInst->getOperand(0);
8431 Value *InRHS = FirstInst->getOperand(1);
8432 PHINode *NewLHS = 0, *NewRHS = 0;
8434 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8435 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8436 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8437 InsertNewInstBefore(NewLHS, PN);
8442 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8443 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8444 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8445 InsertNewInstBefore(NewRHS, PN);
8449 // Add all operands to the new PHIs.
8450 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8452 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8453 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8456 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8457 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8461 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8462 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8463 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8464 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8467 assert(isa<GetElementPtrInst>(FirstInst));
8468 return new GetElementPtrInst(LHSVal, RHSVal);
8472 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8473 /// of the block that defines it. This means that it must be obvious the value
8474 /// of the load is not changed from the point of the load to the end of the
8477 /// Finally, it is safe, but not profitable, to sink a load targetting a
8478 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8480 static bool isSafeToSinkLoad(LoadInst *L) {
8481 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8483 for (++BBI; BBI != E; ++BBI)
8484 if (BBI->mayWriteToMemory())
8487 // Check for non-address taken alloca. If not address-taken already, it isn't
8488 // profitable to do this xform.
8489 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8490 bool isAddressTaken = false;
8491 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8493 if (isa<LoadInst>(UI)) continue;
8494 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8495 // If storing TO the alloca, then the address isn't taken.
8496 if (SI->getOperand(1) == AI) continue;
8498 isAddressTaken = true;
8502 if (!isAddressTaken)
8510 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8511 // operator and they all are only used by the PHI, PHI together their
8512 // inputs, and do the operation once, to the result of the PHI.
8513 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8514 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8516 // Scan the instruction, looking for input operations that can be folded away.
8517 // If all input operands to the phi are the same instruction (e.g. a cast from
8518 // the same type or "+42") we can pull the operation through the PHI, reducing
8519 // code size and simplifying code.
8520 Constant *ConstantOp = 0;
8521 const Type *CastSrcTy = 0;
8522 bool isVolatile = false;
8523 if (isa<CastInst>(FirstInst)) {
8524 CastSrcTy = FirstInst->getOperand(0)->getType();
8525 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8526 // Can fold binop, compare or shift here if the RHS is a constant,
8527 // otherwise call FoldPHIArgBinOpIntoPHI.
8528 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8529 if (ConstantOp == 0)
8530 return FoldPHIArgBinOpIntoPHI(PN);
8531 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8532 isVolatile = LI->isVolatile();
8533 // We can't sink the load if the loaded value could be modified between the
8534 // load and the PHI.
8535 if (LI->getParent() != PN.getIncomingBlock(0) ||
8536 !isSafeToSinkLoad(LI))
8538 } else if (isa<GetElementPtrInst>(FirstInst)) {
8539 if (FirstInst->getNumOperands() == 2)
8540 return FoldPHIArgBinOpIntoPHI(PN);
8541 // Can't handle general GEPs yet.
8544 return 0; // Cannot fold this operation.
8547 // Check to see if all arguments are the same operation.
8548 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8549 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8550 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8551 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8554 if (I->getOperand(0)->getType() != CastSrcTy)
8555 return 0; // Cast operation must match.
8556 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8557 // We can't sink the load if the loaded value could be modified between
8558 // the load and the PHI.
8559 if (LI->isVolatile() != isVolatile ||
8560 LI->getParent() != PN.getIncomingBlock(i) ||
8561 !isSafeToSinkLoad(LI))
8563 } else if (I->getOperand(1) != ConstantOp) {
8568 // Okay, they are all the same operation. Create a new PHI node of the
8569 // correct type, and PHI together all of the LHS's of the instructions.
8570 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8571 PN.getName()+".in");
8572 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8574 Value *InVal = FirstInst->getOperand(0);
8575 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8577 // Add all operands to the new PHI.
8578 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8579 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8580 if (NewInVal != InVal)
8582 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8587 // The new PHI unions all of the same values together. This is really
8588 // common, so we handle it intelligently here for compile-time speed.
8592 InsertNewInstBefore(NewPN, PN);
8596 // Insert and return the new operation.
8597 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8598 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8599 else if (isa<LoadInst>(FirstInst))
8600 return new LoadInst(PhiVal, "", isVolatile);
8601 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8602 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8603 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8604 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8605 PhiVal, ConstantOp);
8607 assert(0 && "Unknown operation");
8611 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8613 static bool DeadPHICycle(PHINode *PN,
8614 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8615 if (PN->use_empty()) return true;
8616 if (!PN->hasOneUse()) return false;
8618 // Remember this node, and if we find the cycle, return.
8619 if (!PotentiallyDeadPHIs.insert(PN))
8622 // Don't scan crazily complex things.
8623 if (PotentiallyDeadPHIs.size() == 16)
8626 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8627 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8632 /// PHIsEqualValue - Return true if this phi node is always equal to
8633 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8634 /// z = some value; x = phi (y, z); y = phi (x, z)
8635 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8636 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8637 // See if we already saw this PHI node.
8638 if (!ValueEqualPHIs.insert(PN))
8641 // Don't scan crazily complex things.
8642 if (ValueEqualPHIs.size() == 16)
8645 // Scan the operands to see if they are either phi nodes or are equal to
8647 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8648 Value *Op = PN->getIncomingValue(i);
8649 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8650 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8652 } else if (Op != NonPhiInVal)
8660 // PHINode simplification
8662 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8663 // If LCSSA is around, don't mess with Phi nodes
8664 if (MustPreserveLCSSA) return 0;
8666 if (Value *V = PN.hasConstantValue())
8667 return ReplaceInstUsesWith(PN, V);
8669 // If all PHI operands are the same operation, pull them through the PHI,
8670 // reducing code size.
8671 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8672 PN.getIncomingValue(0)->hasOneUse())
8673 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8676 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8677 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8678 // PHI)... break the cycle.
8679 if (PN.hasOneUse()) {
8680 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8681 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8682 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8683 PotentiallyDeadPHIs.insert(&PN);
8684 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8685 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8688 // If this phi has a single use, and if that use just computes a value for
8689 // the next iteration of a loop, delete the phi. This occurs with unused
8690 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8691 // common case here is good because the only other things that catch this
8692 // are induction variable analysis (sometimes) and ADCE, which is only run
8694 if (PHIUser->hasOneUse() &&
8695 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8696 PHIUser->use_back() == &PN) {
8697 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8701 // We sometimes end up with phi cycles that non-obviously end up being the
8702 // same value, for example:
8703 // z = some value; x = phi (y, z); y = phi (x, z)
8704 // where the phi nodes don't necessarily need to be in the same block. Do a
8705 // quick check to see if the PHI node only contains a single non-phi value, if
8706 // so, scan to see if the phi cycle is actually equal to that value.
8708 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
8709 // Scan for the first non-phi operand.
8710 while (InValNo != NumOperandVals &&
8711 isa<PHINode>(PN.getIncomingValue(InValNo)))
8714 if (InValNo != NumOperandVals) {
8715 Value *NonPhiInVal = PN.getOperand(InValNo);
8717 // Scan the rest of the operands to see if there are any conflicts, if so
8718 // there is no need to recursively scan other phis.
8719 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
8720 Value *OpVal = PN.getIncomingValue(InValNo);
8721 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
8725 // If we scanned over all operands, then we have one unique value plus
8726 // phi values. Scan PHI nodes to see if they all merge in each other or
8728 if (InValNo == NumOperandVals) {
8729 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
8730 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
8731 return ReplaceInstUsesWith(PN, NonPhiInVal);
8738 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8739 Instruction *InsertPoint,
8741 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8742 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8743 // We must cast correctly to the pointer type. Ensure that we
8744 // sign extend the integer value if it is smaller as this is
8745 // used for address computation.
8746 Instruction::CastOps opcode =
8747 (VTySize < PtrSize ? Instruction::SExt :
8748 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8749 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8753 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8754 Value *PtrOp = GEP.getOperand(0);
8755 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8756 // If so, eliminate the noop.
8757 if (GEP.getNumOperands() == 1)
8758 return ReplaceInstUsesWith(GEP, PtrOp);
8760 if (isa<UndefValue>(GEP.getOperand(0)))
8761 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8763 bool HasZeroPointerIndex = false;
8764 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8765 HasZeroPointerIndex = C->isNullValue();
8767 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8768 return ReplaceInstUsesWith(GEP, PtrOp);
8770 // Eliminate unneeded casts for indices.
8771 bool MadeChange = false;
8773 gep_type_iterator GTI = gep_type_begin(GEP);
8774 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8775 if (isa<SequentialType>(*GTI)) {
8776 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8777 if (CI->getOpcode() == Instruction::ZExt ||
8778 CI->getOpcode() == Instruction::SExt) {
8779 const Type *SrcTy = CI->getOperand(0)->getType();
8780 // We can eliminate a cast from i32 to i64 iff the target
8781 // is a 32-bit pointer target.
8782 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8784 GEP.setOperand(i, CI->getOperand(0));
8788 // If we are using a wider index than needed for this platform, shrink it
8789 // to what we need. If the incoming value needs a cast instruction,
8790 // insert it. This explicit cast can make subsequent optimizations more
8792 Value *Op = GEP.getOperand(i);
8793 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
8794 if (Constant *C = dyn_cast<Constant>(Op)) {
8795 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8798 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8800 GEP.setOperand(i, Op);
8805 if (MadeChange) return &GEP;
8807 // If this GEP instruction doesn't move the pointer, and if the input operand
8808 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8809 // real input to the dest type.
8810 if (GEP.hasAllZeroIndices()) {
8811 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
8812 // If the bitcast is of an allocation, and the allocation will be
8813 // converted to match the type of the cast, don't touch this.
8814 if (isa<AllocationInst>(BCI->getOperand(0))) {
8815 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
8816 if (Instruction *I = visitBitCast(*BCI)) {
8819 BCI->getParent()->getInstList().insert(BCI, I);
8820 ReplaceInstUsesWith(*BCI, I);
8825 return new BitCastInst(BCI->getOperand(0), GEP.getType());
8829 // Combine Indices - If the source pointer to this getelementptr instruction
8830 // is a getelementptr instruction, combine the indices of the two
8831 // getelementptr instructions into a single instruction.
8833 SmallVector<Value*, 8> SrcGEPOperands;
8834 if (User *Src = dyn_castGetElementPtr(PtrOp))
8835 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8837 if (!SrcGEPOperands.empty()) {
8838 // Note that if our source is a gep chain itself that we wait for that
8839 // chain to be resolved before we perform this transformation. This
8840 // avoids us creating a TON of code in some cases.
8842 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8843 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8844 return 0; // Wait until our source is folded to completion.
8846 SmallVector<Value*, 8> Indices;
8848 // Find out whether the last index in the source GEP is a sequential idx.
8849 bool EndsWithSequential = false;
8850 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8851 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8852 EndsWithSequential = !isa<StructType>(*I);
8854 // Can we combine the two pointer arithmetics offsets?
8855 if (EndsWithSequential) {
8856 // Replace: gep (gep %P, long B), long A, ...
8857 // With: T = long A+B; gep %P, T, ...
8859 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8860 if (SO1 == Constant::getNullValue(SO1->getType())) {
8862 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8865 // If they aren't the same type, convert both to an integer of the
8866 // target's pointer size.
8867 if (SO1->getType() != GO1->getType()) {
8868 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8869 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8870 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8871 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8873 unsigned PS = TD->getPointerSizeInBits();
8874 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
8875 // Convert GO1 to SO1's type.
8876 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8878 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
8879 // Convert SO1 to GO1's type.
8880 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8882 const Type *PT = TD->getIntPtrType();
8883 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8884 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8888 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8889 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8891 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8892 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8896 // Recycle the GEP we already have if possible.
8897 if (SrcGEPOperands.size() == 2) {
8898 GEP.setOperand(0, SrcGEPOperands[0]);
8899 GEP.setOperand(1, Sum);
8902 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8903 SrcGEPOperands.end()-1);
8904 Indices.push_back(Sum);
8905 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8907 } else if (isa<Constant>(*GEP.idx_begin()) &&
8908 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8909 SrcGEPOperands.size() != 1) {
8910 // Otherwise we can do the fold if the first index of the GEP is a zero
8911 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8912 SrcGEPOperands.end());
8913 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8916 if (!Indices.empty())
8917 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8918 Indices.end(), GEP.getName());
8920 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8921 // GEP of global variable. If all of the indices for this GEP are
8922 // constants, we can promote this to a constexpr instead of an instruction.
8924 // Scan for nonconstants...
8925 SmallVector<Constant*, 8> Indices;
8926 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8927 for (; I != E && isa<Constant>(*I); ++I)
8928 Indices.push_back(cast<Constant>(*I));
8930 if (I == E) { // If they are all constants...
8931 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8932 &Indices[0],Indices.size());
8934 // Replace all uses of the GEP with the new constexpr...
8935 return ReplaceInstUsesWith(GEP, CE);
8937 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8938 if (!isa<PointerType>(X->getType())) {
8939 // Not interesting. Source pointer must be a cast from pointer.
8940 } else if (HasZeroPointerIndex) {
8941 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
8942 // into : GEP [10 x i8]* X, i32 0, ...
8944 // This occurs when the program declares an array extern like "int X[];"
8946 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8947 const PointerType *XTy = cast<PointerType>(X->getType());
8948 if (const ArrayType *XATy =
8949 dyn_cast<ArrayType>(XTy->getElementType()))
8950 if (const ArrayType *CATy =
8951 dyn_cast<ArrayType>(CPTy->getElementType()))
8952 if (CATy->getElementType() == XATy->getElementType()) {
8953 // At this point, we know that the cast source type is a pointer
8954 // to an array of the same type as the destination pointer
8955 // array. Because the array type is never stepped over (there
8956 // is a leading zero) we can fold the cast into this GEP.
8957 GEP.setOperand(0, X);
8960 } else if (GEP.getNumOperands() == 2) {
8961 // Transform things like:
8962 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
8963 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
8964 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8965 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8966 if (isa<ArrayType>(SrcElTy) &&
8967 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8968 TD->getABITypeSize(ResElTy)) {
8970 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8971 Idx[1] = GEP.getOperand(1);
8972 Value *V = InsertNewInstBefore(
8973 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8974 // V and GEP are both pointer types --> BitCast
8975 return new BitCastInst(V, GEP.getType());
8978 // Transform things like:
8979 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
8980 // (where tmp = 8*tmp2) into:
8981 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
8983 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
8984 uint64_t ArrayEltSize =
8985 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8987 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8988 // allow either a mul, shift, or constant here.
8990 ConstantInt *Scale = 0;
8991 if (ArrayEltSize == 1) {
8992 NewIdx = GEP.getOperand(1);
8993 Scale = ConstantInt::get(NewIdx->getType(), 1);
8994 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8995 NewIdx = ConstantInt::get(CI->getType(), 1);
8997 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8998 if (Inst->getOpcode() == Instruction::Shl &&
8999 isa<ConstantInt>(Inst->getOperand(1))) {
9000 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9001 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9002 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9003 NewIdx = Inst->getOperand(0);
9004 } else if (Inst->getOpcode() == Instruction::Mul &&
9005 isa<ConstantInt>(Inst->getOperand(1))) {
9006 Scale = cast<ConstantInt>(Inst->getOperand(1));
9007 NewIdx = Inst->getOperand(0);
9011 // If the index will be to exactly the right offset with the scale taken
9012 // out, perform the transformation. Note, we don't know whether Scale is
9013 // signed or not. We'll use unsigned version of division/modulo
9014 // operation after making sure Scale doesn't have the sign bit set.
9015 if (Scale && Scale->getSExtValue() >= 0LL &&
9016 Scale->getZExtValue() % ArrayEltSize == 0) {
9017 Scale = ConstantInt::get(Scale->getType(),
9018 Scale->getZExtValue() / ArrayEltSize);
9019 if (Scale->getZExtValue() != 1) {
9020 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9022 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9023 NewIdx = InsertNewInstBefore(Sc, GEP);
9026 // Insert the new GEP instruction.
9028 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9030 Instruction *NewGEP =
9031 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9032 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9033 // The NewGEP must be pointer typed, so must the old one -> BitCast
9034 return new BitCastInst(NewGEP, GEP.getType());
9043 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9044 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9045 if (AI.isArrayAllocation()) // Check C != 1
9046 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9048 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9049 AllocationInst *New = 0;
9051 // Create and insert the replacement instruction...
9052 if (isa<MallocInst>(AI))
9053 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9055 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9056 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9059 InsertNewInstBefore(New, AI);
9061 // Scan to the end of the allocation instructions, to skip over a block of
9062 // allocas if possible...
9064 BasicBlock::iterator It = New;
9065 while (isa<AllocationInst>(*It)) ++It;
9067 // Now that I is pointing to the first non-allocation-inst in the block,
9068 // insert our getelementptr instruction...
9070 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9074 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9075 New->getName()+".sub", It);
9077 // Now make everything use the getelementptr instead of the original
9079 return ReplaceInstUsesWith(AI, V);
9080 } else if (isa<UndefValue>(AI.getArraySize())) {
9081 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9084 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9085 // Note that we only do this for alloca's, because malloc should allocate and
9086 // return a unique pointer, even for a zero byte allocation.
9087 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9088 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9089 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9094 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9095 Value *Op = FI.getOperand(0);
9097 // free undef -> unreachable.
9098 if (isa<UndefValue>(Op)) {
9099 // Insert a new store to null because we cannot modify the CFG here.
9100 new StoreInst(ConstantInt::getTrue(),
9101 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9102 return EraseInstFromFunction(FI);
9105 // If we have 'free null' delete the instruction. This can happen in stl code
9106 // when lots of inlining happens.
9107 if (isa<ConstantPointerNull>(Op))
9108 return EraseInstFromFunction(FI);
9110 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9111 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9112 FI.setOperand(0, CI->getOperand(0));
9116 // Change free (gep X, 0,0,0,0) into free(X)
9117 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9118 if (GEPI->hasAllZeroIndices()) {
9119 AddToWorkList(GEPI);
9120 FI.setOperand(0, GEPI->getOperand(0));
9125 // Change free(malloc) into nothing, if the malloc has a single use.
9126 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9127 if (MI->hasOneUse()) {
9128 EraseInstFromFunction(FI);
9129 return EraseInstFromFunction(*MI);
9136 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9137 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9138 const TargetData *TD) {
9139 User *CI = cast<User>(LI.getOperand(0));
9140 Value *CastOp = CI->getOperand(0);
9142 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9143 // Instead of loading constant c string, use corresponding integer value
9144 // directly if string length is small enough.
9145 const std::string &Str = CE->getOperand(0)->getStringValue();
9147 unsigned len = Str.length();
9148 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9149 unsigned numBits = Ty->getPrimitiveSizeInBits();
9150 // Replace LI with immediate integer store.
9151 if ((numBits >> 3) == len + 1) {
9152 APInt StrVal(numBits, 0);
9153 APInt SingleChar(numBits, 0);
9154 if (TD->isLittleEndian()) {
9155 for (signed i = len-1; i >= 0; i--) {
9156 SingleChar = (uint64_t) Str[i];
9157 StrVal = (StrVal << 8) | SingleChar;
9160 for (unsigned i = 0; i < len; i++) {
9161 SingleChar = (uint64_t) Str[i];
9162 StrVal = (StrVal << 8) | SingleChar;
9164 // Append NULL at the end.
9166 StrVal = (StrVal << 8) | SingleChar;
9168 Value *NL = ConstantInt::get(StrVal);
9169 return IC.ReplaceInstUsesWith(LI, NL);
9174 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9175 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9176 const Type *SrcPTy = SrcTy->getElementType();
9178 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9179 isa<VectorType>(DestPTy)) {
9180 // If the source is an array, the code below will not succeed. Check to
9181 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9183 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9184 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9185 if (ASrcTy->getNumElements() != 0) {
9187 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9188 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9189 SrcTy = cast<PointerType>(CastOp->getType());
9190 SrcPTy = SrcTy->getElementType();
9193 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9194 isa<VectorType>(SrcPTy)) &&
9195 // Do not allow turning this into a load of an integer, which is then
9196 // casted to a pointer, this pessimizes pointer analysis a lot.
9197 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9198 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9199 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9201 // Okay, we are casting from one integer or pointer type to another of
9202 // the same size. Instead of casting the pointer before the load, cast
9203 // the result of the loaded value.
9204 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9206 LI.isVolatile()),LI);
9207 // Now cast the result of the load.
9208 return new BitCastInst(NewLoad, LI.getType());
9215 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9216 /// from this value cannot trap. If it is not obviously safe to load from the
9217 /// specified pointer, we do a quick local scan of the basic block containing
9218 /// ScanFrom, to determine if the address is already accessed.
9219 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9220 // If it is an alloca it is always safe to load from.
9221 if (isa<AllocaInst>(V)) return true;
9223 // If it is a global variable it is mostly safe to load from.
9224 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9225 // Don't try to evaluate aliases. External weak GV can be null.
9226 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9228 // Otherwise, be a little bit agressive by scanning the local block where we
9229 // want to check to see if the pointer is already being loaded or stored
9230 // from/to. If so, the previous load or store would have already trapped,
9231 // so there is no harm doing an extra load (also, CSE will later eliminate
9232 // the load entirely).
9233 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9238 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9239 if (LI->getOperand(0) == V) return true;
9240 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9241 if (SI->getOperand(1) == V) return true;
9247 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9248 /// until we find the underlying object a pointer is referring to or something
9249 /// we don't understand. Note that the returned pointer may be offset from the
9250 /// input, because we ignore GEP indices.
9251 static Value *GetUnderlyingObject(Value *Ptr) {
9253 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9254 if (CE->getOpcode() == Instruction::BitCast ||
9255 CE->getOpcode() == Instruction::GetElementPtr)
9256 Ptr = CE->getOperand(0);
9259 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9260 Ptr = BCI->getOperand(0);
9261 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9262 Ptr = GEP->getOperand(0);
9269 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9270 Value *Op = LI.getOperand(0);
9272 // Attempt to improve the alignment.
9273 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9274 if (KnownAlign > LI.getAlignment())
9275 LI.setAlignment(KnownAlign);
9277 // load (cast X) --> cast (load X) iff safe
9278 if (isa<CastInst>(Op))
9279 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9282 // None of the following transforms are legal for volatile loads.
9283 if (LI.isVolatile()) return 0;
9285 if (&LI.getParent()->front() != &LI) {
9286 BasicBlock::iterator BBI = &LI; --BBI;
9287 // If the instruction immediately before this is a store to the same
9288 // address, do a simple form of store->load forwarding.
9289 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9290 if (SI->getOperand(1) == LI.getOperand(0))
9291 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9292 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9293 if (LIB->getOperand(0) == LI.getOperand(0))
9294 return ReplaceInstUsesWith(LI, LIB);
9297 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9298 const Value *GEPI0 = GEPI->getOperand(0);
9299 // TODO: Consider a target hook for valid address spaces for this xform.
9300 if (isa<ConstantPointerNull>(GEPI0) &&
9301 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9302 // Insert a new store to null instruction before the load to indicate
9303 // that this code is not reachable. We do this instead of inserting
9304 // an unreachable instruction directly because we cannot modify the
9306 new StoreInst(UndefValue::get(LI.getType()),
9307 Constant::getNullValue(Op->getType()), &LI);
9308 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9312 if (Constant *C = dyn_cast<Constant>(Op)) {
9313 // load null/undef -> undef
9314 // TODO: Consider a target hook for valid address spaces for this xform.
9315 if (isa<UndefValue>(C) || (C->isNullValue() &&
9316 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9317 // Insert a new store to null instruction before the load to indicate that
9318 // this code is not reachable. We do this instead of inserting an
9319 // unreachable instruction directly because we cannot modify the CFG.
9320 new StoreInst(UndefValue::get(LI.getType()),
9321 Constant::getNullValue(Op->getType()), &LI);
9322 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9325 // Instcombine load (constant global) into the value loaded.
9326 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9327 if (GV->isConstant() && !GV->isDeclaration())
9328 return ReplaceInstUsesWith(LI, GV->getInitializer());
9330 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9331 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9332 if (CE->getOpcode() == Instruction::GetElementPtr) {
9333 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9334 if (GV->isConstant() && !GV->isDeclaration())
9336 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9337 return ReplaceInstUsesWith(LI, V);
9338 if (CE->getOperand(0)->isNullValue()) {
9339 // Insert a new store to null instruction before the load to indicate
9340 // that this code is not reachable. We do this instead of inserting
9341 // an unreachable instruction directly because we cannot modify the
9343 new StoreInst(UndefValue::get(LI.getType()),
9344 Constant::getNullValue(Op->getType()), &LI);
9345 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9348 } else if (CE->isCast()) {
9349 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9354 // If this load comes from anywhere in a constant global, and if the global
9355 // is all undef or zero, we know what it loads.
9356 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9357 if (GV->isConstant() && GV->hasInitializer()) {
9358 if (GV->getInitializer()->isNullValue())
9359 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9360 else if (isa<UndefValue>(GV->getInitializer()))
9361 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9365 if (Op->hasOneUse()) {
9366 // Change select and PHI nodes to select values instead of addresses: this
9367 // helps alias analysis out a lot, allows many others simplifications, and
9368 // exposes redundancy in the code.
9370 // Note that we cannot do the transformation unless we know that the
9371 // introduced loads cannot trap! Something like this is valid as long as
9372 // the condition is always false: load (select bool %C, int* null, int* %G),
9373 // but it would not be valid if we transformed it to load from null
9376 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9377 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9378 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9379 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9380 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9381 SI->getOperand(1)->getName()+".val"), LI);
9382 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9383 SI->getOperand(2)->getName()+".val"), LI);
9384 return new SelectInst(SI->getCondition(), V1, V2);
9387 // load (select (cond, null, P)) -> load P
9388 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9389 if (C->isNullValue()) {
9390 LI.setOperand(0, SI->getOperand(2));
9394 // load (select (cond, P, null)) -> load P
9395 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9396 if (C->isNullValue()) {
9397 LI.setOperand(0, SI->getOperand(1));
9405 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9407 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9408 User *CI = cast<User>(SI.getOperand(1));
9409 Value *CastOp = CI->getOperand(0);
9411 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9412 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9413 const Type *SrcPTy = SrcTy->getElementType();
9415 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9416 // If the source is an array, the code below will not succeed. Check to
9417 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9419 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9420 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9421 if (ASrcTy->getNumElements() != 0) {
9423 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9424 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9425 SrcTy = cast<PointerType>(CastOp->getType());
9426 SrcPTy = SrcTy->getElementType();
9429 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9430 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9431 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9433 // Okay, we are casting from one integer or pointer type to another of
9434 // the same size. Instead of casting the pointer before
9435 // the store, cast the value to be stored.
9437 Value *SIOp0 = SI.getOperand(0);
9438 Instruction::CastOps opcode = Instruction::BitCast;
9439 const Type* CastSrcTy = SIOp0->getType();
9440 const Type* CastDstTy = SrcPTy;
9441 if (isa<PointerType>(CastDstTy)) {
9442 if (CastSrcTy->isInteger())
9443 opcode = Instruction::IntToPtr;
9444 } else if (isa<IntegerType>(CastDstTy)) {
9445 if (isa<PointerType>(SIOp0->getType()))
9446 opcode = Instruction::PtrToInt;
9448 if (Constant *C = dyn_cast<Constant>(SIOp0))
9449 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9451 NewCast = IC.InsertNewInstBefore(
9452 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9454 return new StoreInst(NewCast, CastOp);
9461 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9462 Value *Val = SI.getOperand(0);
9463 Value *Ptr = SI.getOperand(1);
9465 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9466 EraseInstFromFunction(SI);
9471 // If the RHS is an alloca with a single use, zapify the store, making the
9473 if (Ptr->hasOneUse()) {
9474 if (isa<AllocaInst>(Ptr)) {
9475 EraseInstFromFunction(SI);
9480 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9481 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9482 GEP->getOperand(0)->hasOneUse()) {
9483 EraseInstFromFunction(SI);
9489 // Attempt to improve the alignment.
9490 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9491 if (KnownAlign > SI.getAlignment())
9492 SI.setAlignment(KnownAlign);
9494 // Do really simple DSE, to catch cases where there are several consequtive
9495 // stores to the same location, separated by a few arithmetic operations. This
9496 // situation often occurs with bitfield accesses.
9497 BasicBlock::iterator BBI = &SI;
9498 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9502 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9503 // Prev store isn't volatile, and stores to the same location?
9504 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9507 EraseInstFromFunction(*PrevSI);
9513 // If this is a load, we have to stop. However, if the loaded value is from
9514 // the pointer we're loading and is producing the pointer we're storing,
9515 // then *this* store is dead (X = load P; store X -> P).
9516 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9517 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9518 EraseInstFromFunction(SI);
9522 // Otherwise, this is a load from some other location. Stores before it
9527 // Don't skip over loads or things that can modify memory.
9528 if (BBI->mayWriteToMemory())
9533 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9535 // store X, null -> turns into 'unreachable' in SimplifyCFG
9536 if (isa<ConstantPointerNull>(Ptr)) {
9537 if (!isa<UndefValue>(Val)) {
9538 SI.setOperand(0, UndefValue::get(Val->getType()));
9539 if (Instruction *U = dyn_cast<Instruction>(Val))
9540 AddToWorkList(U); // Dropped a use.
9543 return 0; // Do not modify these!
9546 // store undef, Ptr -> noop
9547 if (isa<UndefValue>(Val)) {
9548 EraseInstFromFunction(SI);
9553 // If the pointer destination is a cast, see if we can fold the cast into the
9555 if (isa<CastInst>(Ptr))
9556 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9558 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9560 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9564 // If this store is the last instruction in the basic block, and if the block
9565 // ends with an unconditional branch, try to move it to the successor block.
9567 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9568 if (BI->isUnconditional())
9569 if (SimplifyStoreAtEndOfBlock(SI))
9570 return 0; // xform done!
9575 /// SimplifyStoreAtEndOfBlock - Turn things like:
9576 /// if () { *P = v1; } else { *P = v2 }
9577 /// into a phi node with a store in the successor.
9579 /// Simplify things like:
9580 /// *P = v1; if () { *P = v2; }
9581 /// into a phi node with a store in the successor.
9583 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9584 BasicBlock *StoreBB = SI.getParent();
9586 // Check to see if the successor block has exactly two incoming edges. If
9587 // so, see if the other predecessor contains a store to the same location.
9588 // if so, insert a PHI node (if needed) and move the stores down.
9589 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9591 // Determine whether Dest has exactly two predecessors and, if so, compute
9592 // the other predecessor.
9593 pred_iterator PI = pred_begin(DestBB);
9594 BasicBlock *OtherBB = 0;
9598 if (PI == pred_end(DestBB))
9601 if (*PI != StoreBB) {
9606 if (++PI != pred_end(DestBB))
9610 // Verify that the other block ends in a branch and is not otherwise empty.
9611 BasicBlock::iterator BBI = OtherBB->getTerminator();
9612 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9613 if (!OtherBr || BBI == OtherBB->begin())
9616 // If the other block ends in an unconditional branch, check for the 'if then
9617 // else' case. there is an instruction before the branch.
9618 StoreInst *OtherStore = 0;
9619 if (OtherBr->isUnconditional()) {
9620 // If this isn't a store, or isn't a store to the same location, bail out.
9622 OtherStore = dyn_cast<StoreInst>(BBI);
9623 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9626 // Otherwise, the other block ended with a conditional branch. If one of the
9627 // destinations is StoreBB, then we have the if/then case.
9628 if (OtherBr->getSuccessor(0) != StoreBB &&
9629 OtherBr->getSuccessor(1) != StoreBB)
9632 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9633 // if/then triangle. See if there is a store to the same ptr as SI that
9634 // lives in OtherBB.
9636 // Check to see if we find the matching store.
9637 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9638 if (OtherStore->getOperand(1) != SI.getOperand(1))
9642 // If we find something that may be using the stored value, or if we run
9643 // out of instructions, we can't do the xform.
9644 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9645 BBI == OtherBB->begin())
9649 // In order to eliminate the store in OtherBr, we have to
9650 // make sure nothing reads the stored value in StoreBB.
9651 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9652 // FIXME: This should really be AA driven.
9653 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9658 // Insert a PHI node now if we need it.
9659 Value *MergedVal = OtherStore->getOperand(0);
9660 if (MergedVal != SI.getOperand(0)) {
9661 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9662 PN->reserveOperandSpace(2);
9663 PN->addIncoming(SI.getOperand(0), SI.getParent());
9664 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9665 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9668 // Advance to a place where it is safe to insert the new store and
9670 BBI = DestBB->begin();
9671 while (isa<PHINode>(BBI)) ++BBI;
9672 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9673 OtherStore->isVolatile()), *BBI);
9675 // Nuke the old stores.
9676 EraseInstFromFunction(SI);
9677 EraseInstFromFunction(*OtherStore);
9683 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9684 // Change br (not X), label True, label False to: br X, label False, True
9686 BasicBlock *TrueDest;
9687 BasicBlock *FalseDest;
9688 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9689 !isa<Constant>(X)) {
9690 // Swap Destinations and condition...
9692 BI.setSuccessor(0, FalseDest);
9693 BI.setSuccessor(1, TrueDest);
9697 // Cannonicalize fcmp_one -> fcmp_oeq
9698 FCmpInst::Predicate FPred; Value *Y;
9699 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9700 TrueDest, FalseDest)))
9701 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9702 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9703 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9704 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9705 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9706 NewSCC->takeName(I);
9707 // Swap Destinations and condition...
9708 BI.setCondition(NewSCC);
9709 BI.setSuccessor(0, FalseDest);
9710 BI.setSuccessor(1, TrueDest);
9711 RemoveFromWorkList(I);
9712 I->eraseFromParent();
9713 AddToWorkList(NewSCC);
9717 // Cannonicalize icmp_ne -> icmp_eq
9718 ICmpInst::Predicate IPred;
9719 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9720 TrueDest, FalseDest)))
9721 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9722 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9723 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9724 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9725 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9726 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9727 NewSCC->takeName(I);
9728 // Swap Destinations and condition...
9729 BI.setCondition(NewSCC);
9730 BI.setSuccessor(0, FalseDest);
9731 BI.setSuccessor(1, TrueDest);
9732 RemoveFromWorkList(I);
9733 I->eraseFromParent();;
9734 AddToWorkList(NewSCC);
9741 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9742 Value *Cond = SI.getCondition();
9743 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9744 if (I->getOpcode() == Instruction::Add)
9745 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9746 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9747 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9748 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9750 SI.setOperand(0, I->getOperand(0));
9758 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9759 /// is to leave as a vector operation.
9760 static bool CheapToScalarize(Value *V, bool isConstant) {
9761 if (isa<ConstantAggregateZero>(V))
9763 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9764 if (isConstant) return true;
9765 // If all elts are the same, we can extract.
9766 Constant *Op0 = C->getOperand(0);
9767 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9768 if (C->getOperand(i) != Op0)
9772 Instruction *I = dyn_cast<Instruction>(V);
9773 if (!I) return false;
9775 // Insert element gets simplified to the inserted element or is deleted if
9776 // this is constant idx extract element and its a constant idx insertelt.
9777 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9778 isa<ConstantInt>(I->getOperand(2)))
9780 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9782 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9783 if (BO->hasOneUse() &&
9784 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9785 CheapToScalarize(BO->getOperand(1), isConstant)))
9787 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9788 if (CI->hasOneUse() &&
9789 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9790 CheapToScalarize(CI->getOperand(1), isConstant)))
9796 /// Read and decode a shufflevector mask.
9798 /// It turns undef elements into values that are larger than the number of
9799 /// elements in the input.
9800 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9801 unsigned NElts = SVI->getType()->getNumElements();
9802 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9803 return std::vector<unsigned>(NElts, 0);
9804 if (isa<UndefValue>(SVI->getOperand(2)))
9805 return std::vector<unsigned>(NElts, 2*NElts);
9807 std::vector<unsigned> Result;
9808 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9809 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9810 if (isa<UndefValue>(CP->getOperand(i)))
9811 Result.push_back(NElts*2); // undef -> 8
9813 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9817 /// FindScalarElement - Given a vector and an element number, see if the scalar
9818 /// value is already around as a register, for example if it were inserted then
9819 /// extracted from the vector.
9820 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9821 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9822 const VectorType *PTy = cast<VectorType>(V->getType());
9823 unsigned Width = PTy->getNumElements();
9824 if (EltNo >= Width) // Out of range access.
9825 return UndefValue::get(PTy->getElementType());
9827 if (isa<UndefValue>(V))
9828 return UndefValue::get(PTy->getElementType());
9829 else if (isa<ConstantAggregateZero>(V))
9830 return Constant::getNullValue(PTy->getElementType());
9831 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9832 return CP->getOperand(EltNo);
9833 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9834 // If this is an insert to a variable element, we don't know what it is.
9835 if (!isa<ConstantInt>(III->getOperand(2)))
9837 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9839 // If this is an insert to the element we are looking for, return the
9842 return III->getOperand(1);
9844 // Otherwise, the insertelement doesn't modify the value, recurse on its
9846 return FindScalarElement(III->getOperand(0), EltNo);
9847 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9848 unsigned InEl = getShuffleMask(SVI)[EltNo];
9850 return FindScalarElement(SVI->getOperand(0), InEl);
9851 else if (InEl < Width*2)
9852 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9854 return UndefValue::get(PTy->getElementType());
9857 // Otherwise, we don't know.
9861 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9863 // If vector val is undef, replace extract with scalar undef.
9864 if (isa<UndefValue>(EI.getOperand(0)))
9865 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9867 // If vector val is constant 0, replace extract with scalar 0.
9868 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9869 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9871 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9872 // If vector val is constant with uniform operands, replace EI
9873 // with that operand
9874 Constant *op0 = C->getOperand(0);
9875 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9876 if (C->getOperand(i) != op0) {
9881 return ReplaceInstUsesWith(EI, op0);
9884 // If extracting a specified index from the vector, see if we can recursively
9885 // find a previously computed scalar that was inserted into the vector.
9886 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9887 unsigned IndexVal = IdxC->getZExtValue();
9888 unsigned VectorWidth =
9889 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9891 // If this is extracting an invalid index, turn this into undef, to avoid
9892 // crashing the code below.
9893 if (IndexVal >= VectorWidth)
9894 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9896 // This instruction only demands the single element from the input vector.
9897 // If the input vector has a single use, simplify it based on this use
9899 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9901 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9904 EI.setOperand(0, V);
9909 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9910 return ReplaceInstUsesWith(EI, Elt);
9912 // If the this extractelement is directly using a bitcast from a vector of
9913 // the same number of elements, see if we can find the source element from
9914 // it. In this case, we will end up needing to bitcast the scalars.
9915 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9916 if (const VectorType *VT =
9917 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9918 if (VT->getNumElements() == VectorWidth)
9919 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9920 return new BitCastInst(Elt, EI.getType());
9924 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9925 if (I->hasOneUse()) {
9926 // Push extractelement into predecessor operation if legal and
9927 // profitable to do so
9928 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9929 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9930 if (CheapToScalarize(BO, isConstantElt)) {
9931 ExtractElementInst *newEI0 =
9932 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9933 EI.getName()+".lhs");
9934 ExtractElementInst *newEI1 =
9935 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9936 EI.getName()+".rhs");
9937 InsertNewInstBefore(newEI0, EI);
9938 InsertNewInstBefore(newEI1, EI);
9939 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9941 } else if (isa<LoadInst>(I)) {
9943 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
9944 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9945 PointerType::get(EI.getType(), AS), EI);
9946 GetElementPtrInst *GEP =
9947 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9948 InsertNewInstBefore(GEP, EI);
9949 return new LoadInst(GEP);
9952 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9953 // Extracting the inserted element?
9954 if (IE->getOperand(2) == EI.getOperand(1))
9955 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9956 // If the inserted and extracted elements are constants, they must not
9957 // be the same value, extract from the pre-inserted value instead.
9958 if (isa<Constant>(IE->getOperand(2)) &&
9959 isa<Constant>(EI.getOperand(1))) {
9960 AddUsesToWorkList(EI);
9961 EI.setOperand(0, IE->getOperand(0));
9964 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9965 // If this is extracting an element from a shufflevector, figure out where
9966 // it came from and extract from the appropriate input element instead.
9967 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9968 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9970 if (SrcIdx < SVI->getType()->getNumElements())
9971 Src = SVI->getOperand(0);
9972 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9973 SrcIdx -= SVI->getType()->getNumElements();
9974 Src = SVI->getOperand(1);
9976 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9978 return new ExtractElementInst(Src, SrcIdx);
9985 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9986 /// elements from either LHS or RHS, return the shuffle mask and true.
9987 /// Otherwise, return false.
9988 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9989 std::vector<Constant*> &Mask) {
9990 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9991 "Invalid CollectSingleShuffleElements");
9992 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9994 if (isa<UndefValue>(V)) {
9995 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9997 } else if (V == LHS) {
9998 for (unsigned i = 0; i != NumElts; ++i)
9999 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10001 } else if (V == RHS) {
10002 for (unsigned i = 0; i != NumElts; ++i)
10003 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10005 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10006 // If this is an insert of an extract from some other vector, include it.
10007 Value *VecOp = IEI->getOperand(0);
10008 Value *ScalarOp = IEI->getOperand(1);
10009 Value *IdxOp = IEI->getOperand(2);
10011 if (!isa<ConstantInt>(IdxOp))
10013 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10015 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10016 // Okay, we can handle this if the vector we are insertinting into is
10017 // transitively ok.
10018 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10019 // If so, update the mask to reflect the inserted undef.
10020 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10023 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10024 if (isa<ConstantInt>(EI->getOperand(1)) &&
10025 EI->getOperand(0)->getType() == V->getType()) {
10026 unsigned ExtractedIdx =
10027 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10029 // This must be extracting from either LHS or RHS.
10030 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10031 // Okay, we can handle this if the vector we are insertinting into is
10032 // transitively ok.
10033 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10034 // If so, update the mask to reflect the inserted value.
10035 if (EI->getOperand(0) == LHS) {
10036 Mask[InsertedIdx & (NumElts-1)] =
10037 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10039 assert(EI->getOperand(0) == RHS);
10040 Mask[InsertedIdx & (NumElts-1)] =
10041 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10050 // TODO: Handle shufflevector here!
10055 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10056 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10057 /// that computes V and the LHS value of the shuffle.
10058 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10060 assert(isa<VectorType>(V->getType()) &&
10061 (RHS == 0 || V->getType() == RHS->getType()) &&
10062 "Invalid shuffle!");
10063 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10065 if (isa<UndefValue>(V)) {
10066 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10068 } else if (isa<ConstantAggregateZero>(V)) {
10069 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10071 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10072 // If this is an insert of an extract from some other vector, include it.
10073 Value *VecOp = IEI->getOperand(0);
10074 Value *ScalarOp = IEI->getOperand(1);
10075 Value *IdxOp = IEI->getOperand(2);
10077 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10078 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10079 EI->getOperand(0)->getType() == V->getType()) {
10080 unsigned ExtractedIdx =
10081 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10082 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10084 // Either the extracted from or inserted into vector must be RHSVec,
10085 // otherwise we'd end up with a shuffle of three inputs.
10086 if (EI->getOperand(0) == RHS || RHS == 0) {
10087 RHS = EI->getOperand(0);
10088 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10089 Mask[InsertedIdx & (NumElts-1)] =
10090 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10094 if (VecOp == RHS) {
10095 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10096 // Everything but the extracted element is replaced with the RHS.
10097 for (unsigned i = 0; i != NumElts; ++i) {
10098 if (i != InsertedIdx)
10099 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10104 // If this insertelement is a chain that comes from exactly these two
10105 // vectors, return the vector and the effective shuffle.
10106 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10107 return EI->getOperand(0);
10112 // TODO: Handle shufflevector here!
10114 // Otherwise, can't do anything fancy. Return an identity vector.
10115 for (unsigned i = 0; i != NumElts; ++i)
10116 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10120 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10121 Value *VecOp = IE.getOperand(0);
10122 Value *ScalarOp = IE.getOperand(1);
10123 Value *IdxOp = IE.getOperand(2);
10125 // Inserting an undef or into an undefined place, remove this.
10126 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10127 ReplaceInstUsesWith(IE, VecOp);
10129 // If the inserted element was extracted from some other vector, and if the
10130 // indexes are constant, try to turn this into a shufflevector operation.
10131 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10132 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10133 EI->getOperand(0)->getType() == IE.getType()) {
10134 unsigned NumVectorElts = IE.getType()->getNumElements();
10135 unsigned ExtractedIdx =
10136 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10137 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10139 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10140 return ReplaceInstUsesWith(IE, VecOp);
10142 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10143 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10145 // If we are extracting a value from a vector, then inserting it right
10146 // back into the same place, just use the input vector.
10147 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10148 return ReplaceInstUsesWith(IE, VecOp);
10150 // We could theoretically do this for ANY input. However, doing so could
10151 // turn chains of insertelement instructions into a chain of shufflevector
10152 // instructions, and right now we do not merge shufflevectors. As such,
10153 // only do this in a situation where it is clear that there is benefit.
10154 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10155 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10156 // the values of VecOp, except then one read from EIOp0.
10157 // Build a new shuffle mask.
10158 std::vector<Constant*> Mask;
10159 if (isa<UndefValue>(VecOp))
10160 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10162 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10163 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10166 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10167 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10168 ConstantVector::get(Mask));
10171 // If this insertelement isn't used by some other insertelement, turn it
10172 // (and any insertelements it points to), into one big shuffle.
10173 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10174 std::vector<Constant*> Mask;
10176 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10177 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10178 // We now have a shuffle of LHS, RHS, Mask.
10179 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10188 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10189 Value *LHS = SVI.getOperand(0);
10190 Value *RHS = SVI.getOperand(1);
10191 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10193 bool MadeChange = false;
10195 // Undefined shuffle mask -> undefined value.
10196 if (isa<UndefValue>(SVI.getOperand(2)))
10197 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10199 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10200 // the undef, change them to undefs.
10201 if (isa<UndefValue>(SVI.getOperand(1))) {
10202 // Scan to see if there are any references to the RHS. If so, replace them
10203 // with undef element refs and set MadeChange to true.
10204 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10205 if (Mask[i] >= e && Mask[i] != 2*e) {
10212 // Remap any references to RHS to use LHS.
10213 std::vector<Constant*> Elts;
10214 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10215 if (Mask[i] == 2*e)
10216 Elts.push_back(UndefValue::get(Type::Int32Ty));
10218 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10220 SVI.setOperand(2, ConstantVector::get(Elts));
10224 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10225 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10226 if (LHS == RHS || isa<UndefValue>(LHS)) {
10227 if (isa<UndefValue>(LHS) && LHS == RHS) {
10228 // shuffle(undef,undef,mask) -> undef.
10229 return ReplaceInstUsesWith(SVI, LHS);
10232 // Remap any references to RHS to use LHS.
10233 std::vector<Constant*> Elts;
10234 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10235 if (Mask[i] >= 2*e)
10236 Elts.push_back(UndefValue::get(Type::Int32Ty));
10238 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10239 (Mask[i] < e && isa<UndefValue>(LHS)))
10240 Mask[i] = 2*e; // Turn into undef.
10242 Mask[i] &= (e-1); // Force to LHS.
10243 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10246 SVI.setOperand(0, SVI.getOperand(1));
10247 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10248 SVI.setOperand(2, ConstantVector::get(Elts));
10249 LHS = SVI.getOperand(0);
10250 RHS = SVI.getOperand(1);
10254 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10255 bool isLHSID = true, isRHSID = true;
10257 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10258 if (Mask[i] >= e*2) continue; // Ignore undef values.
10259 // Is this an identity shuffle of the LHS value?
10260 isLHSID &= (Mask[i] == i);
10262 // Is this an identity shuffle of the RHS value?
10263 isRHSID &= (Mask[i]-e == i);
10266 // Eliminate identity shuffles.
10267 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10268 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10270 // If the LHS is a shufflevector itself, see if we can combine it with this
10271 // one without producing an unusual shuffle. Here we are really conservative:
10272 // we are absolutely afraid of producing a shuffle mask not in the input
10273 // program, because the code gen may not be smart enough to turn a merged
10274 // shuffle into two specific shuffles: it may produce worse code. As such,
10275 // we only merge two shuffles if the result is one of the two input shuffle
10276 // masks. In this case, merging the shuffles just removes one instruction,
10277 // which we know is safe. This is good for things like turning:
10278 // (splat(splat)) -> splat.
10279 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10280 if (isa<UndefValue>(RHS)) {
10281 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10283 std::vector<unsigned> NewMask;
10284 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10285 if (Mask[i] >= 2*e)
10286 NewMask.push_back(2*e);
10288 NewMask.push_back(LHSMask[Mask[i]]);
10290 // If the result mask is equal to the src shuffle or this shuffle mask, do
10291 // the replacement.
10292 if (NewMask == LHSMask || NewMask == Mask) {
10293 std::vector<Constant*> Elts;
10294 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10295 if (NewMask[i] >= e*2) {
10296 Elts.push_back(UndefValue::get(Type::Int32Ty));
10298 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10301 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10302 LHSSVI->getOperand(1),
10303 ConstantVector::get(Elts));
10308 return MadeChange ? &SVI : 0;
10314 /// TryToSinkInstruction - Try to move the specified instruction from its
10315 /// current block into the beginning of DestBlock, which can only happen if it's
10316 /// safe to move the instruction past all of the instructions between it and the
10317 /// end of its block.
10318 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10319 assert(I->hasOneUse() && "Invariants didn't hold!");
10321 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10322 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10324 // Do not sink alloca instructions out of the entry block.
10325 if (isa<AllocaInst>(I) && I->getParent() ==
10326 &DestBlock->getParent()->getEntryBlock())
10329 // We can only sink load instructions if there is nothing between the load and
10330 // the end of block that could change the value.
10331 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10332 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10334 if (Scan->mayWriteToMemory())
10338 BasicBlock::iterator InsertPos = DestBlock->begin();
10339 while (isa<PHINode>(InsertPos)) ++InsertPos;
10341 I->moveBefore(InsertPos);
10347 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10348 /// all reachable code to the worklist.
10350 /// This has a couple of tricks to make the code faster and more powerful. In
10351 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10352 /// them to the worklist (this significantly speeds up instcombine on code where
10353 /// many instructions are dead or constant). Additionally, if we find a branch
10354 /// whose condition is a known constant, we only visit the reachable successors.
10356 static void AddReachableCodeToWorklist(BasicBlock *BB,
10357 SmallPtrSet<BasicBlock*, 64> &Visited,
10359 const TargetData *TD) {
10360 std::vector<BasicBlock*> Worklist;
10361 Worklist.push_back(BB);
10363 while (!Worklist.empty()) {
10364 BB = Worklist.back();
10365 Worklist.pop_back();
10367 // We have now visited this block! If we've already been here, ignore it.
10368 if (!Visited.insert(BB)) continue;
10370 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10371 Instruction *Inst = BBI++;
10373 // DCE instruction if trivially dead.
10374 if (isInstructionTriviallyDead(Inst)) {
10376 DOUT << "IC: DCE: " << *Inst;
10377 Inst->eraseFromParent();
10381 // ConstantProp instruction if trivially constant.
10382 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10383 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10384 Inst->replaceAllUsesWith(C);
10386 Inst->eraseFromParent();
10390 IC.AddToWorkList(Inst);
10393 // Recursively visit successors. If this is a branch or switch on a
10394 // constant, only visit the reachable successor.
10395 TerminatorInst *TI = BB->getTerminator();
10396 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10397 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10398 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10399 Worklist.push_back(BI->getSuccessor(!CondVal));
10402 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10403 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10404 // See if this is an explicit destination.
10405 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10406 if (SI->getCaseValue(i) == Cond) {
10407 Worklist.push_back(SI->getSuccessor(i));
10411 // Otherwise it is the default destination.
10412 Worklist.push_back(SI->getSuccessor(0));
10417 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10418 Worklist.push_back(TI->getSuccessor(i));
10422 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10423 bool Changed = false;
10424 TD = &getAnalysis<TargetData>();
10426 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10427 << F.getNameStr() << "\n");
10430 // Do a depth-first traversal of the function, populate the worklist with
10431 // the reachable instructions. Ignore blocks that are not reachable. Keep
10432 // track of which blocks we visit.
10433 SmallPtrSet<BasicBlock*, 64> Visited;
10434 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10436 // Do a quick scan over the function. If we find any blocks that are
10437 // unreachable, remove any instructions inside of them. This prevents
10438 // the instcombine code from having to deal with some bad special cases.
10439 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10440 if (!Visited.count(BB)) {
10441 Instruction *Term = BB->getTerminator();
10442 while (Term != BB->begin()) { // Remove instrs bottom-up
10443 BasicBlock::iterator I = Term; --I;
10445 DOUT << "IC: DCE: " << *I;
10448 if (!I->use_empty())
10449 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10450 I->eraseFromParent();
10455 while (!Worklist.empty()) {
10456 Instruction *I = RemoveOneFromWorkList();
10457 if (I == 0) continue; // skip null values.
10459 // Check to see if we can DCE the instruction.
10460 if (isInstructionTriviallyDead(I)) {
10461 // Add operands to the worklist.
10462 if (I->getNumOperands() < 4)
10463 AddUsesToWorkList(*I);
10466 DOUT << "IC: DCE: " << *I;
10468 I->eraseFromParent();
10469 RemoveFromWorkList(I);
10473 // Instruction isn't dead, see if we can constant propagate it.
10474 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10475 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10477 // Add operands to the worklist.
10478 AddUsesToWorkList(*I);
10479 ReplaceInstUsesWith(*I, C);
10482 I->eraseFromParent();
10483 RemoveFromWorkList(I);
10487 // See if we can trivially sink this instruction to a successor basic block.
10488 if (I->hasOneUse()) {
10489 BasicBlock *BB = I->getParent();
10490 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10491 if (UserParent != BB) {
10492 bool UserIsSuccessor = false;
10493 // See if the user is one of our successors.
10494 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10495 if (*SI == UserParent) {
10496 UserIsSuccessor = true;
10500 // If the user is one of our immediate successors, and if that successor
10501 // only has us as a predecessors (we'd have to split the critical edge
10502 // otherwise), we can keep going.
10503 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10504 next(pred_begin(UserParent)) == pred_end(UserParent))
10505 // Okay, the CFG is simple enough, try to sink this instruction.
10506 Changed |= TryToSinkInstruction(I, UserParent);
10510 // Now that we have an instruction, try combining it to simplify it...
10514 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10515 if (Instruction *Result = visit(*I)) {
10517 // Should we replace the old instruction with a new one?
10519 DOUT << "IC: Old = " << *I
10520 << " New = " << *Result;
10522 // Everything uses the new instruction now.
10523 I->replaceAllUsesWith(Result);
10525 // Push the new instruction and any users onto the worklist.
10526 AddToWorkList(Result);
10527 AddUsersToWorkList(*Result);
10529 // Move the name to the new instruction first.
10530 Result->takeName(I);
10532 // Insert the new instruction into the basic block...
10533 BasicBlock *InstParent = I->getParent();
10534 BasicBlock::iterator InsertPos = I;
10536 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10537 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10540 InstParent->getInstList().insert(InsertPos, Result);
10542 // Make sure that we reprocess all operands now that we reduced their
10544 AddUsesToWorkList(*I);
10546 // Instructions can end up on the worklist more than once. Make sure
10547 // we do not process an instruction that has been deleted.
10548 RemoveFromWorkList(I);
10550 // Erase the old instruction.
10551 InstParent->getInstList().erase(I);
10554 DOUT << "IC: Mod = " << OrigI
10555 << " New = " << *I;
10558 // If the instruction was modified, it's possible that it is now dead.
10559 // if so, remove it.
10560 if (isInstructionTriviallyDead(I)) {
10561 // Make sure we process all operands now that we are reducing their
10563 AddUsesToWorkList(*I);
10565 // Instructions may end up in the worklist more than once. Erase all
10566 // occurrences of this instruction.
10567 RemoveFromWorkList(I);
10568 I->eraseFromParent();
10571 AddUsersToWorkList(*I);
10578 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10580 // Do an explicit clear, this shrinks the map if needed.
10581 WorklistMap.clear();
10586 bool InstCombiner::runOnFunction(Function &F) {
10587 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10589 bool EverMadeChange = false;
10591 // Iterate while there is work to do.
10592 unsigned Iteration = 0;
10593 while (DoOneIteration(F, Iteration++))
10594 EverMadeChange = true;
10595 return EverMadeChange;
10598 FunctionPass *llvm::createInstructionCombiningPass() {
10599 return new InstCombiner();