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(IntToPtrInst &CI);
214 Instruction *visitBitCast(BitCastInst &CI);
215 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
217 Instruction *visitSelectInst(SelectInst &CI);
218 Instruction *visitCallInst(CallInst &CI);
219 Instruction *visitInvokeInst(InvokeInst &II);
220 Instruction *visitPHINode(PHINode &PN);
221 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
222 Instruction *visitAllocationInst(AllocationInst &AI);
223 Instruction *visitFreeInst(FreeInst &FI);
224 Instruction *visitLoadInst(LoadInst &LI);
225 Instruction *visitStoreInst(StoreInst &SI);
226 Instruction *visitBranchInst(BranchInst &BI);
227 Instruction *visitSwitchInst(SwitchInst &SI);
228 Instruction *visitInsertElementInst(InsertElementInst &IE);
229 Instruction *visitExtractElementInst(ExtractElementInst &EI);
230 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
232 // visitInstruction - Specify what to return for unhandled instructions...
233 Instruction *visitInstruction(Instruction &I) { return 0; }
236 Instruction *visitCallSite(CallSite CS);
237 bool transformConstExprCastCall(CallSite CS);
238 Instruction *transformCallThroughTrampoline(CallSite CS);
241 // InsertNewInstBefore - insert an instruction New before instruction Old
242 // in the program. Add the new instruction to the worklist.
244 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
245 assert(New && New->getParent() == 0 &&
246 "New instruction already inserted into a basic block!");
247 BasicBlock *BB = Old.getParent();
248 BB->getInstList().insert(&Old, New); // Insert inst
253 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
254 /// This also adds the cast to the worklist. Finally, this returns the
256 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
258 if (V->getType() == Ty) return V;
260 if (Constant *CV = dyn_cast<Constant>(V))
261 return ConstantExpr::getCast(opc, CV, Ty);
263 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
268 // 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;
3499 if (ICmpInst::isSignedPredicate(LHSCC) ||
3500 (ICmpInst::isEquality(LHSCC) &&
3501 ICmpInst::isSignedPredicate(RHSCC)))
3502 GT = ICmpInst::ICMP_SGT;
3504 GT = ICmpInst::ICMP_UGT;
3506 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3507 ICmpInst *LHS = cast<ICmpInst>(Op0);
3508 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3509 std::swap(LHS, RHS);
3510 std::swap(LHSCst, RHSCst);
3511 std::swap(LHSCC, RHSCC);
3514 // At this point, we know we have have two icmp instructions
3515 // comparing a value against two constants and and'ing the result
3516 // together. Because of the above check, we know that we only have
3517 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3518 // (from the FoldICmpLogical check above), that the two constants
3519 // are not equal and that the larger constant is on the RHS
3520 assert(LHSCst != RHSCst && "Compares not folded above?");
3523 default: assert(0 && "Unknown integer condition code!");
3524 case ICmpInst::ICMP_EQ:
3526 default: assert(0 && "Unknown integer condition code!");
3527 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3528 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3529 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3530 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3531 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3532 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3533 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3534 return ReplaceInstUsesWith(I, LHS);
3536 case ICmpInst::ICMP_NE:
3538 default: assert(0 && "Unknown integer condition code!");
3539 case ICmpInst::ICMP_ULT:
3540 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3541 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3542 break; // (X != 13 & X u< 15) -> no change
3543 case ICmpInst::ICMP_SLT:
3544 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3545 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3546 break; // (X != 13 & X s< 15) -> no change
3547 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3548 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3549 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3550 return ReplaceInstUsesWith(I, RHS);
3551 case ICmpInst::ICMP_NE:
3552 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3553 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3554 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3555 LHSVal->getName()+".off");
3556 InsertNewInstBefore(Add, I);
3557 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3558 ConstantInt::get(Add->getType(), 1));
3560 break; // (X != 13 & X != 15) -> no change
3563 case ICmpInst::ICMP_ULT:
3565 default: assert(0 && "Unknown integer condition code!");
3566 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3567 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3568 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3569 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3571 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3572 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3573 return ReplaceInstUsesWith(I, LHS);
3574 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3578 case ICmpInst::ICMP_SLT:
3580 default: assert(0 && "Unknown integer condition code!");
3581 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3582 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3583 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3584 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3586 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3587 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3588 return ReplaceInstUsesWith(I, LHS);
3589 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3593 case ICmpInst::ICMP_UGT:
3595 default: assert(0 && "Unknown integer condition code!");
3596 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3597 return ReplaceInstUsesWith(I, LHS);
3598 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3599 return ReplaceInstUsesWith(I, RHS);
3600 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3602 case ICmpInst::ICMP_NE:
3603 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3604 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3605 break; // (X u> 13 & X != 15) -> no change
3606 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3607 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3609 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3613 case ICmpInst::ICMP_SGT:
3615 default: assert(0 && "Unknown integer condition code!");
3616 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3617 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3618 return ReplaceInstUsesWith(I, RHS);
3619 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3621 case ICmpInst::ICMP_NE:
3622 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3623 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3624 break; // (X s> 13 & X != 15) -> no change
3625 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3626 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3628 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3636 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3637 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3638 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3639 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3640 const Type *SrcTy = Op0C->getOperand(0)->getType();
3641 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3642 // Only do this if the casts both really cause code to be generated.
3643 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3645 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3647 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3648 Op1C->getOperand(0),
3650 InsertNewInstBefore(NewOp, I);
3651 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3655 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3656 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3657 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3658 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3659 SI0->getOperand(1) == SI1->getOperand(1) &&
3660 (SI0->hasOneUse() || SI1->hasOneUse())) {
3661 Instruction *NewOp =
3662 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3664 SI0->getName()), I);
3665 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3666 SI1->getOperand(1));
3670 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3671 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3672 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3673 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3674 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3675 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3676 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3677 // If either of the constants are nans, then the whole thing returns
3679 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3680 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3681 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3682 RHS->getOperand(0));
3687 return Changed ? &I : 0;
3690 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3691 /// in the result. If it does, and if the specified byte hasn't been filled in
3692 /// yet, fill it in and return false.
3693 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3694 Instruction *I = dyn_cast<Instruction>(V);
3695 if (I == 0) return true;
3697 // If this is an or instruction, it is an inner node of the bswap.
3698 if (I->getOpcode() == Instruction::Or)
3699 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3700 CollectBSwapParts(I->getOperand(1), ByteValues);
3702 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3703 // If this is a shift by a constant int, and it is "24", then its operand
3704 // defines a byte. We only handle unsigned types here.
3705 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3706 // Not shifting the entire input by N-1 bytes?
3707 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3708 8*(ByteValues.size()-1))
3712 if (I->getOpcode() == Instruction::Shl) {
3713 // X << 24 defines the top byte with the lowest of the input bytes.
3714 DestNo = ByteValues.size()-1;
3716 // X >>u 24 defines the low byte with the highest of the input bytes.
3720 // If the destination byte value is already defined, the values are or'd
3721 // together, which isn't a bswap (unless it's an or of the same bits).
3722 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3724 ByteValues[DestNo] = I->getOperand(0);
3728 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3730 Value *Shift = 0, *ShiftLHS = 0;
3731 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3732 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3733 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3735 Instruction *SI = cast<Instruction>(Shift);
3737 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3738 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3739 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3742 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3744 if (AndAmt->getValue().getActiveBits() > 64)
3746 uint64_t AndAmtVal = AndAmt->getZExtValue();
3747 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3748 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3750 // Unknown mask for bswap.
3751 if (DestByte == ByteValues.size()) return true;
3753 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3755 if (SI->getOpcode() == Instruction::Shl)
3756 SrcByte = DestByte - ShiftBytes;
3758 SrcByte = DestByte + ShiftBytes;
3760 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3761 if (SrcByte != ByteValues.size()-DestByte-1)
3764 // If the destination byte value is already defined, the values are or'd
3765 // together, which isn't a bswap (unless it's an or of the same bits).
3766 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3768 ByteValues[DestByte] = SI->getOperand(0);
3772 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3773 /// If so, insert the new bswap intrinsic and return it.
3774 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3775 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3776 if (!ITy || ITy->getBitWidth() % 16)
3777 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3779 /// ByteValues - For each byte of the result, we keep track of which value
3780 /// defines each byte.
3781 SmallVector<Value*, 8> ByteValues;
3782 ByteValues.resize(ITy->getBitWidth()/8);
3784 // Try to find all the pieces corresponding to the bswap.
3785 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3786 CollectBSwapParts(I.getOperand(1), ByteValues))
3789 // Check to see if all of the bytes come from the same value.
3790 Value *V = ByteValues[0];
3791 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3793 // Check to make sure that all of the bytes come from the same value.
3794 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3795 if (ByteValues[i] != V)
3797 const Type *Tys[] = { ITy };
3798 Module *M = I.getParent()->getParent()->getParent();
3799 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3800 return new CallInst(F, V);
3804 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3805 bool Changed = SimplifyCommutative(I);
3806 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3808 if (isa<UndefValue>(Op1)) // X | undef -> -1
3809 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3813 return ReplaceInstUsesWith(I, Op0);
3815 // See if we can simplify any instructions used by the instruction whose sole
3816 // purpose is to compute bits we don't care about.
3817 if (!isa<VectorType>(I.getType())) {
3818 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3819 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3820 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3821 KnownZero, KnownOne))
3823 } else if (isa<ConstantAggregateZero>(Op1)) {
3824 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3825 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3826 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3827 return ReplaceInstUsesWith(I, I.getOperand(1));
3833 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3834 ConstantInt *C1 = 0; Value *X = 0;
3835 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3836 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3837 Instruction *Or = BinaryOperator::createOr(X, RHS);
3838 InsertNewInstBefore(Or, I);
3840 return BinaryOperator::createAnd(Or,
3841 ConstantInt::get(RHS->getValue() | C1->getValue()));
3844 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3845 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3846 Instruction *Or = BinaryOperator::createOr(X, RHS);
3847 InsertNewInstBefore(Or, I);
3849 return BinaryOperator::createXor(Or,
3850 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3853 // Try to fold constant and into select arguments.
3854 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3855 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3857 if (isa<PHINode>(Op0))
3858 if (Instruction *NV = FoldOpIntoPhi(I))
3862 Value *A = 0, *B = 0;
3863 ConstantInt *C1 = 0, *C2 = 0;
3865 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3866 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3867 return ReplaceInstUsesWith(I, Op1);
3868 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3869 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3870 return ReplaceInstUsesWith(I, Op0);
3872 // (A | B) | C and A | (B | C) -> bswap if possible.
3873 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3874 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3875 match(Op1, m_Or(m_Value(), m_Value())) ||
3876 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3877 match(Op1, m_Shift(m_Value(), m_Value())))) {
3878 if (Instruction *BSwap = MatchBSwap(I))
3882 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3883 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3884 MaskedValueIsZero(Op1, C1->getValue())) {
3885 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3886 InsertNewInstBefore(NOr, I);
3888 return BinaryOperator::createXor(NOr, C1);
3891 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3892 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3893 MaskedValueIsZero(Op0, C1->getValue())) {
3894 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3895 InsertNewInstBefore(NOr, I);
3897 return BinaryOperator::createXor(NOr, C1);
3901 Value *C = 0, *D = 0;
3902 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3903 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3904 Value *V1 = 0, *V2 = 0, *V3 = 0;
3905 C1 = dyn_cast<ConstantInt>(C);
3906 C2 = dyn_cast<ConstantInt>(D);
3907 if (C1 && C2) { // (A & C1)|(B & C2)
3908 // If we have: ((V + N) & C1) | (V & C2)
3909 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3910 // replace with V+N.
3911 if (C1->getValue() == ~C2->getValue()) {
3912 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3913 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3914 // Add commutes, try both ways.
3915 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3916 return ReplaceInstUsesWith(I, A);
3917 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3918 return ReplaceInstUsesWith(I, A);
3920 // Or commutes, try both ways.
3921 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3922 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3923 // Add commutes, try both ways.
3924 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3925 return ReplaceInstUsesWith(I, B);
3926 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3927 return ReplaceInstUsesWith(I, B);
3930 V1 = 0; V2 = 0; V3 = 0;
3933 // Check to see if we have any common things being and'ed. If so, find the
3934 // terms for V1 & (V2|V3).
3935 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3936 if (A == B) // (A & C)|(A & D) == A & (C|D)
3937 V1 = A, V2 = C, V3 = D;
3938 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3939 V1 = A, V2 = B, V3 = C;
3940 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3941 V1 = C, V2 = A, V3 = D;
3942 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3943 V1 = C, V2 = A, V3 = B;
3947 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3948 return BinaryOperator::createAnd(V1, Or);
3953 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3954 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3955 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3956 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3957 SI0->getOperand(1) == SI1->getOperand(1) &&
3958 (SI0->hasOneUse() || SI1->hasOneUse())) {
3959 Instruction *NewOp =
3960 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3962 SI0->getName()), I);
3963 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3964 SI1->getOperand(1));
3968 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3969 if (A == Op1) // ~A | A == -1
3970 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3974 // Note, A is still live here!
3975 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3977 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3979 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3980 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3981 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3982 I.getName()+".demorgan"), I);
3983 return BinaryOperator::createNot(And);
3987 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3988 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3989 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3992 Value *LHSVal, *RHSVal;
3993 ConstantInt *LHSCst, *RHSCst;
3994 ICmpInst::Predicate LHSCC, RHSCC;
3995 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3996 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3997 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3998 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3999 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4000 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4001 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4002 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4003 // We can't fold (ugt x, C) | (sgt x, C2).
4004 PredicatesFoldable(LHSCC, RHSCC)) {
4005 // Ensure that the larger constant is on the RHS.
4006 ICmpInst *LHS = cast<ICmpInst>(Op0);
4008 if (ICmpInst::isSignedPredicate(LHSCC))
4009 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4011 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4014 std::swap(LHS, RHS);
4015 std::swap(LHSCst, RHSCst);
4016 std::swap(LHSCC, RHSCC);
4019 // At this point, we know we have have two icmp instructions
4020 // comparing a value against two constants and or'ing the result
4021 // together. Because of the above check, we know that we only have
4022 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4023 // FoldICmpLogical check above), that the two constants are not
4025 assert(LHSCst != RHSCst && "Compares not folded above?");
4028 default: assert(0 && "Unknown integer condition code!");
4029 case ICmpInst::ICMP_EQ:
4031 default: assert(0 && "Unknown integer condition code!");
4032 case ICmpInst::ICMP_EQ:
4033 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4034 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4035 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4036 LHSVal->getName()+".off");
4037 InsertNewInstBefore(Add, I);
4038 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4039 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4041 break; // (X == 13 | X == 15) -> no change
4042 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4043 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4045 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4046 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4047 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4048 return ReplaceInstUsesWith(I, RHS);
4051 case ICmpInst::ICMP_NE:
4053 default: assert(0 && "Unknown integer condition code!");
4054 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4055 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4056 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4057 return ReplaceInstUsesWith(I, LHS);
4058 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4059 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4060 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4061 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4064 case ICmpInst::ICMP_ULT:
4066 default: assert(0 && "Unknown integer condition code!");
4067 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4069 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4070 // If RHSCst is [us]MAXINT, it is always false. Not handling
4071 // this can cause overflow.
4072 if (RHSCst->isMaxValue(false))
4073 return ReplaceInstUsesWith(I, LHS);
4074 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4076 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4078 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4079 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4080 return ReplaceInstUsesWith(I, RHS);
4081 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4085 case ICmpInst::ICMP_SLT:
4087 default: assert(0 && "Unknown integer condition code!");
4088 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4090 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4091 // If RHSCst is [us]MAXINT, it is always false. Not handling
4092 // this can cause overflow.
4093 if (RHSCst->isMaxValue(true))
4094 return ReplaceInstUsesWith(I, LHS);
4095 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4097 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4099 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4100 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4101 return ReplaceInstUsesWith(I, RHS);
4102 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4106 case ICmpInst::ICMP_UGT:
4108 default: assert(0 && "Unknown integer condition code!");
4109 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4110 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4111 return ReplaceInstUsesWith(I, LHS);
4112 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4114 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4115 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4116 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4117 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4121 case ICmpInst::ICMP_SGT:
4123 default: assert(0 && "Unknown integer condition code!");
4124 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4125 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4126 return ReplaceInstUsesWith(I, LHS);
4127 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4129 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4130 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4131 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4132 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4140 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4141 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4142 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4143 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4144 const Type *SrcTy = Op0C->getOperand(0)->getType();
4145 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4146 // Only do this if the casts both really cause code to be generated.
4147 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4149 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4151 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4152 Op1C->getOperand(0),
4154 InsertNewInstBefore(NewOp, I);
4155 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4161 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4162 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4163 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4164 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4165 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4166 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4167 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4168 // If either of the constants are nans, then the whole thing returns
4170 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4171 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4173 // Otherwise, no need to compare the two constants, compare the
4175 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4176 RHS->getOperand(0));
4181 return Changed ? &I : 0;
4184 // XorSelf - Implements: X ^ X --> 0
4187 XorSelf(Value *rhs) : RHS(rhs) {}
4188 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4189 Instruction *apply(BinaryOperator &Xor) const {
4195 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4196 bool Changed = SimplifyCommutative(I);
4197 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4199 if (isa<UndefValue>(Op1))
4200 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4202 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4203 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4204 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4205 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4208 // See if we can simplify any instructions used by the instruction whose sole
4209 // purpose is to compute bits we don't care about.
4210 if (!isa<VectorType>(I.getType())) {
4211 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4212 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4213 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4214 KnownZero, KnownOne))
4216 } else if (isa<ConstantAggregateZero>(Op1)) {
4217 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4220 // Is this a ~ operation?
4221 if (Value *NotOp = dyn_castNotVal(&I)) {
4222 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4223 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4224 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4225 if (Op0I->getOpcode() == Instruction::And ||
4226 Op0I->getOpcode() == Instruction::Or) {
4227 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4228 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4230 BinaryOperator::createNot(Op0I->getOperand(1),
4231 Op0I->getOperand(1)->getName()+".not");
4232 InsertNewInstBefore(NotY, I);
4233 if (Op0I->getOpcode() == Instruction::And)
4234 return BinaryOperator::createOr(Op0NotVal, NotY);
4236 return BinaryOperator::createAnd(Op0NotVal, NotY);
4243 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4244 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4245 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4246 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4247 return new ICmpInst(ICI->getInversePredicate(),
4248 ICI->getOperand(0), ICI->getOperand(1));
4250 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4251 return new FCmpInst(FCI->getInversePredicate(),
4252 FCI->getOperand(0), FCI->getOperand(1));
4255 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4256 // ~(c-X) == X-c-1 == X+(-c-1)
4257 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4258 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4259 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4260 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4261 ConstantInt::get(I.getType(), 1));
4262 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4265 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4266 if (Op0I->getOpcode() == Instruction::Add) {
4267 // ~(X-c) --> (-c-1)-X
4268 if (RHS->isAllOnesValue()) {
4269 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4270 return BinaryOperator::createSub(
4271 ConstantExpr::getSub(NegOp0CI,
4272 ConstantInt::get(I.getType(), 1)),
4273 Op0I->getOperand(0));
4274 } else if (RHS->getValue().isSignBit()) {
4275 // (X + C) ^ signbit -> (X + C + signbit)
4276 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4277 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4280 } else if (Op0I->getOpcode() == Instruction::Or) {
4281 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4282 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4283 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4284 // Anything in both C1 and C2 is known to be zero, remove it from
4286 Constant *CommonBits = And(Op0CI, RHS);
4287 NewRHS = ConstantExpr::getAnd(NewRHS,
4288 ConstantExpr::getNot(CommonBits));
4289 AddToWorkList(Op0I);
4290 I.setOperand(0, Op0I->getOperand(0));
4291 I.setOperand(1, NewRHS);
4297 // Try to fold constant and into select arguments.
4298 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4299 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4301 if (isa<PHINode>(Op0))
4302 if (Instruction *NV = FoldOpIntoPhi(I))
4306 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4308 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4310 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4312 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4315 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4318 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4319 if (A == Op0) { // B^(B|A) == (A|B)^B
4320 Op1I->swapOperands();
4322 std::swap(Op0, Op1);
4323 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4324 I.swapOperands(); // Simplified below.
4325 std::swap(Op0, Op1);
4327 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4328 if (Op0 == A) // A^(A^B) == B
4329 return ReplaceInstUsesWith(I, B);
4330 else if (Op0 == B) // A^(B^A) == B
4331 return ReplaceInstUsesWith(I, A);
4332 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4333 if (A == Op0) { // A^(A&B) -> A^(B&A)
4334 Op1I->swapOperands();
4337 if (B == Op0) { // A^(B&A) -> (B&A)^A
4338 I.swapOperands(); // Simplified below.
4339 std::swap(Op0, Op1);
4344 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4347 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4348 if (A == Op1) // (B|A)^B == (A|B)^B
4350 if (B == Op1) { // (A|B)^B == A & ~B
4352 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4353 return BinaryOperator::createAnd(A, NotB);
4355 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4356 if (Op1 == A) // (A^B)^A == B
4357 return ReplaceInstUsesWith(I, B);
4358 else if (Op1 == B) // (B^A)^A == B
4359 return ReplaceInstUsesWith(I, A);
4360 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4361 if (A == Op1) // (A&B)^A -> (B&A)^A
4363 if (B == Op1 && // (B&A)^A == ~B & A
4364 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4366 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4367 return BinaryOperator::createAnd(N, Op1);
4372 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4373 if (Op0I && Op1I && Op0I->isShift() &&
4374 Op0I->getOpcode() == Op1I->getOpcode() &&
4375 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4376 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4377 Instruction *NewOp =
4378 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4379 Op1I->getOperand(0),
4380 Op0I->getName()), I);
4381 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4382 Op1I->getOperand(1));
4386 Value *A, *B, *C, *D;
4387 // (A & B)^(A | B) -> A ^ B
4388 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4389 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4390 if ((A == C && B == D) || (A == D && B == C))
4391 return BinaryOperator::createXor(A, B);
4393 // (A | B)^(A & B) -> A ^ B
4394 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4395 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4396 if ((A == C && B == D) || (A == D && B == C))
4397 return BinaryOperator::createXor(A, B);
4401 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4402 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4403 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4404 // (X & Y)^(X & Y) -> (Y^Z) & X
4405 Value *X = 0, *Y = 0, *Z = 0;
4407 X = A, Y = B, Z = D;
4409 X = A, Y = B, Z = C;
4411 X = B, Y = A, Z = D;
4413 X = B, Y = A, Z = C;
4416 Instruction *NewOp =
4417 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4418 return BinaryOperator::createAnd(NewOp, X);
4423 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4424 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4425 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4428 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4429 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4430 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4431 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4432 const Type *SrcTy = Op0C->getOperand(0)->getType();
4433 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4434 // Only do this if the casts both really cause code to be generated.
4435 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4437 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4439 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4440 Op1C->getOperand(0),
4442 InsertNewInstBefore(NewOp, I);
4443 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4447 return Changed ? &I : 0;
4450 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4451 /// overflowed for this type.
4452 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4453 ConstantInt *In2, bool IsSigned = false) {
4454 Result = cast<ConstantInt>(Add(In1, In2));
4457 if (In2->getValue().isNegative())
4458 return Result->getValue().sgt(In1->getValue());
4460 return Result->getValue().slt(In1->getValue());
4462 return Result->getValue().ult(In1->getValue());
4465 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4466 /// code necessary to compute the offset from the base pointer (without adding
4467 /// in the base pointer). Return the result as a signed integer of intptr size.
4468 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4469 TargetData &TD = IC.getTargetData();
4470 gep_type_iterator GTI = gep_type_begin(GEP);
4471 const Type *IntPtrTy = TD.getIntPtrType();
4472 Value *Result = Constant::getNullValue(IntPtrTy);
4474 // Build a mask for high order bits.
4475 unsigned IntPtrWidth = TD.getPointerSize()*8;
4476 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4478 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4479 Value *Op = GEP->getOperand(i);
4480 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4481 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4482 if (OpC->isZero()) continue;
4484 // Handle a struct index, which adds its field offset to the pointer.
4485 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4486 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4488 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4489 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4491 Result = IC.InsertNewInstBefore(
4492 BinaryOperator::createAdd(Result,
4493 ConstantInt::get(IntPtrTy, Size),
4494 GEP->getName()+".offs"), I);
4498 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4499 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4500 Scale = ConstantExpr::getMul(OC, Scale);
4501 if (Constant *RC = dyn_cast<Constant>(Result))
4502 Result = ConstantExpr::getAdd(RC, Scale);
4504 // Emit an add instruction.
4505 Result = IC.InsertNewInstBefore(
4506 BinaryOperator::createAdd(Result, Scale,
4507 GEP->getName()+".offs"), I);
4511 // Convert to correct type.
4512 if (Op->getType() != IntPtrTy) {
4513 if (Constant *OpC = dyn_cast<Constant>(Op))
4514 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4516 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4517 Op->getName()+".c"), I);
4520 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4521 if (Constant *OpC = dyn_cast<Constant>(Op))
4522 Op = ConstantExpr::getMul(OpC, Scale);
4523 else // We'll let instcombine(mul) convert this to a shl if possible.
4524 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4525 GEP->getName()+".idx"), I);
4528 // Emit an add instruction.
4529 if (isa<Constant>(Op) && isa<Constant>(Result))
4530 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4531 cast<Constant>(Result));
4533 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4534 GEP->getName()+".offs"), I);
4539 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4540 /// else. At this point we know that the GEP is on the LHS of the comparison.
4541 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4542 ICmpInst::Predicate Cond,
4544 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4546 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4547 if (isa<PointerType>(CI->getOperand(0)->getType()))
4548 RHS = CI->getOperand(0);
4550 Value *PtrBase = GEPLHS->getOperand(0);
4551 if (PtrBase == RHS) {
4552 // As an optimization, we don't actually have to compute the actual value of
4553 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4554 // each index is zero or not.
4555 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4556 Instruction *InVal = 0;
4557 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4558 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4560 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4561 if (isa<UndefValue>(C)) // undef index -> undef.
4562 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4563 if (C->isNullValue())
4565 else if (TD->getABITypeSize(GTI.getIndexedType()) == 0) {
4566 EmitIt = false; // This is indexing into a zero sized array?
4567 } else if (isa<ConstantInt>(C))
4568 return ReplaceInstUsesWith(I, // No comparison is needed here.
4569 ConstantInt::get(Type::Int1Ty,
4570 Cond == ICmpInst::ICMP_NE));
4575 new ICmpInst(Cond, GEPLHS->getOperand(i),
4576 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4580 InVal = InsertNewInstBefore(InVal, I);
4581 InsertNewInstBefore(Comp, I);
4582 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4583 InVal = BinaryOperator::createOr(InVal, Comp);
4584 else // True if all are equal
4585 InVal = BinaryOperator::createAnd(InVal, Comp);
4593 // No comparison is needed here, all indexes = 0
4594 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4595 Cond == ICmpInst::ICMP_EQ));
4598 // Only lower this if the icmp is the only user of the GEP or if we expect
4599 // the result to fold to a constant!
4600 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4601 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4602 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4603 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4604 Constant::getNullValue(Offset->getType()));
4606 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4607 // If the base pointers are different, but the indices are the same, just
4608 // compare the base pointer.
4609 if (PtrBase != GEPRHS->getOperand(0)) {
4610 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4611 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4612 GEPRHS->getOperand(0)->getType();
4614 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4615 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4616 IndicesTheSame = false;
4620 // If all indices are the same, just compare the base pointers.
4622 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4623 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4625 // Otherwise, the base pointers are different and the indices are
4626 // different, bail out.
4630 // If one of the GEPs has all zero indices, recurse.
4631 bool AllZeros = true;
4632 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4633 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4634 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4639 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4640 ICmpInst::getSwappedPredicate(Cond), I);
4642 // If the other GEP has all zero indices, recurse.
4644 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4645 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4646 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4651 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4653 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4654 // If the GEPs only differ by one index, compare it.
4655 unsigned NumDifferences = 0; // Keep track of # differences.
4656 unsigned DiffOperand = 0; // The operand that differs.
4657 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4658 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4659 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4660 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4661 // Irreconcilable differences.
4665 if (NumDifferences++) break;
4670 if (NumDifferences == 0) // SAME GEP?
4671 return ReplaceInstUsesWith(I, // No comparison is needed here.
4672 ConstantInt::get(Type::Int1Ty,
4673 isTrueWhenEqual(Cond)));
4675 else if (NumDifferences == 1) {
4676 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4677 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4678 // Make sure we do a signed comparison here.
4679 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4683 // Only lower this if the icmp is the only user of the GEP or if we expect
4684 // the result to fold to a constant!
4685 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4686 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4687 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4688 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4689 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4690 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4696 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4697 bool Changed = SimplifyCompare(I);
4698 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4700 // Fold trivial predicates.
4701 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4702 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4703 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4704 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4706 // Simplify 'fcmp pred X, X'
4708 switch (I.getPredicate()) {
4709 default: assert(0 && "Unknown predicate!");
4710 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4711 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4712 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4713 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4714 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4715 case FCmpInst::FCMP_OLT: // True if ordered and less than
4716 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4717 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4719 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4720 case FCmpInst::FCMP_ULT: // True if unordered or less than
4721 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4722 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4723 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4724 I.setPredicate(FCmpInst::FCMP_UNO);
4725 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4728 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4729 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4730 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4731 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4732 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4733 I.setPredicate(FCmpInst::FCMP_ORD);
4734 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4739 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4740 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4742 // Handle fcmp with constant RHS
4743 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4744 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4745 switch (LHSI->getOpcode()) {
4746 case Instruction::PHI:
4747 if (Instruction *NV = FoldOpIntoPhi(I))
4750 case Instruction::Select:
4751 // If either operand of the select is a constant, we can fold the
4752 // comparison into the select arms, which will cause one to be
4753 // constant folded and the select turned into a bitwise or.
4754 Value *Op1 = 0, *Op2 = 0;
4755 if (LHSI->hasOneUse()) {
4756 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4757 // Fold the known value into the constant operand.
4758 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4759 // Insert a new FCmp of the other select operand.
4760 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4761 LHSI->getOperand(2), RHSC,
4763 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4764 // Fold the known value into the constant operand.
4765 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4766 // Insert a new FCmp of the other select operand.
4767 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4768 LHSI->getOperand(1), RHSC,
4774 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4779 return Changed ? &I : 0;
4782 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4783 bool Changed = SimplifyCompare(I);
4784 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4785 const Type *Ty = Op0->getType();
4789 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4790 isTrueWhenEqual(I)));
4792 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4793 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4795 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4796 // addresses never equal each other! We already know that Op0 != Op1.
4797 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4798 isa<ConstantPointerNull>(Op0)) &&
4799 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4800 isa<ConstantPointerNull>(Op1)))
4801 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4802 !isTrueWhenEqual(I)));
4804 // icmp's with boolean values can always be turned into bitwise operations
4805 if (Ty == Type::Int1Ty) {
4806 switch (I.getPredicate()) {
4807 default: assert(0 && "Invalid icmp instruction!");
4808 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4809 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4810 InsertNewInstBefore(Xor, I);
4811 return BinaryOperator::createNot(Xor);
4813 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4814 return BinaryOperator::createXor(Op0, Op1);
4816 case ICmpInst::ICMP_UGT:
4817 case ICmpInst::ICMP_SGT:
4818 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4820 case ICmpInst::ICMP_ULT:
4821 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4822 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4823 InsertNewInstBefore(Not, I);
4824 return BinaryOperator::createAnd(Not, Op1);
4826 case ICmpInst::ICMP_UGE:
4827 case ICmpInst::ICMP_SGE:
4828 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4830 case ICmpInst::ICMP_ULE:
4831 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4832 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4833 InsertNewInstBefore(Not, I);
4834 return BinaryOperator::createOr(Not, Op1);
4839 // See if we are doing a comparison between a constant and an instruction that
4840 // can be folded into the comparison.
4841 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4844 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4845 if (I.isEquality() && CI->isNullValue() &&
4846 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4847 // (icmp cond A B) if cond is equality
4848 return new ICmpInst(I.getPredicate(), A, B);
4851 switch (I.getPredicate()) {
4853 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4854 if (CI->isMinValue(false))
4855 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4856 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4857 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4858 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4859 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4860 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4861 if (CI->isMinValue(true))
4862 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4863 ConstantInt::getAllOnesValue(Op0->getType()));
4867 case ICmpInst::ICMP_SLT:
4868 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4869 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4870 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4871 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4872 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4873 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4876 case ICmpInst::ICMP_UGT:
4877 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4878 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4879 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4880 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4881 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4882 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4884 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4885 if (CI->isMaxValue(true))
4886 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4887 ConstantInt::getNullValue(Op0->getType()));
4890 case ICmpInst::ICMP_SGT:
4891 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4892 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4893 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4894 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4895 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4896 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4899 case ICmpInst::ICMP_ULE:
4900 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4901 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4902 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4903 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4904 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4905 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4908 case ICmpInst::ICMP_SLE:
4909 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4910 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4911 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4912 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4913 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4914 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4917 case ICmpInst::ICMP_UGE:
4918 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4919 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4920 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4921 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4922 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4923 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4926 case ICmpInst::ICMP_SGE:
4927 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4928 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4929 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4930 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4931 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4932 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4936 // If we still have a icmp le or icmp ge instruction, turn it into the
4937 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4938 // already been handled above, this requires little checking.
4940 switch (I.getPredicate()) {
4942 case ICmpInst::ICMP_ULE:
4943 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4944 case ICmpInst::ICMP_SLE:
4945 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4946 case ICmpInst::ICMP_UGE:
4947 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4948 case ICmpInst::ICMP_SGE:
4949 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4952 // See if we can fold the comparison based on bits known to be zero or one
4953 // in the input. If this comparison is a normal comparison, it demands all
4954 // bits, if it is a sign bit comparison, it only demands the sign bit.
4957 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4959 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4960 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4961 if (SimplifyDemandedBits(Op0,
4962 isSignBit ? APInt::getSignBit(BitWidth)
4963 : APInt::getAllOnesValue(BitWidth),
4964 KnownZero, KnownOne, 0))
4967 // Given the known and unknown bits, compute a range that the LHS could be
4969 if ((KnownOne | KnownZero) != 0) {
4970 // Compute the Min, Max and RHS values based on the known bits. For the
4971 // EQ and NE we use unsigned values.
4972 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4973 const APInt& RHSVal = CI->getValue();
4974 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4975 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4978 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4981 switch (I.getPredicate()) { // LE/GE have been folded already.
4982 default: assert(0 && "Unknown icmp opcode!");
4983 case ICmpInst::ICMP_EQ:
4984 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4985 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4987 case ICmpInst::ICMP_NE:
4988 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4989 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4991 case ICmpInst::ICMP_ULT:
4992 if (Max.ult(RHSVal))
4993 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4994 if (Min.uge(RHSVal))
4995 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4997 case ICmpInst::ICMP_UGT:
4998 if (Min.ugt(RHSVal))
4999 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5000 if (Max.ule(RHSVal))
5001 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5003 case ICmpInst::ICMP_SLT:
5004 if (Max.slt(RHSVal))
5005 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5006 if (Min.sgt(RHSVal))
5007 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5009 case ICmpInst::ICMP_SGT:
5010 if (Min.sgt(RHSVal))
5011 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5012 if (Max.sle(RHSVal))
5013 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5018 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5019 // instruction, see if that instruction also has constants so that the
5020 // instruction can be folded into the icmp
5021 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5022 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5026 // Handle icmp with constant (but not simple integer constant) RHS
5027 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5028 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5029 switch (LHSI->getOpcode()) {
5030 case Instruction::GetElementPtr:
5031 if (RHSC->isNullValue()) {
5032 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5033 bool isAllZeros = true;
5034 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5035 if (!isa<Constant>(LHSI->getOperand(i)) ||
5036 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5041 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5042 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5046 case Instruction::PHI:
5047 if (Instruction *NV = FoldOpIntoPhi(I))
5050 case Instruction::Select: {
5051 // If either operand of the select is a constant, we can fold the
5052 // comparison into the select arms, which will cause one to be
5053 // constant folded and the select turned into a bitwise or.
5054 Value *Op1 = 0, *Op2 = 0;
5055 if (LHSI->hasOneUse()) {
5056 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5057 // Fold the known value into the constant operand.
5058 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5059 // Insert a new ICmp of the other select operand.
5060 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5061 LHSI->getOperand(2), RHSC,
5063 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5064 // Fold the known value into the constant operand.
5065 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5066 // Insert a new ICmp of the other select operand.
5067 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5068 LHSI->getOperand(1), RHSC,
5074 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5077 case Instruction::Malloc:
5078 // If we have (malloc != null), and if the malloc has a single use, we
5079 // can assume it is successful and remove the malloc.
5080 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5081 AddToWorkList(LHSI);
5082 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5083 !isTrueWhenEqual(I)));
5089 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5090 if (User *GEP = dyn_castGetElementPtr(Op0))
5091 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5093 if (User *GEP = dyn_castGetElementPtr(Op1))
5094 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5095 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5098 // Test to see if the operands of the icmp are casted versions of other
5099 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5101 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5102 if (isa<PointerType>(Op0->getType()) &&
5103 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5104 // We keep moving the cast from the left operand over to the right
5105 // operand, where it can often be eliminated completely.
5106 Op0 = CI->getOperand(0);
5108 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5109 // so eliminate it as well.
5110 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5111 Op1 = CI2->getOperand(0);
5113 // If Op1 is a constant, we can fold the cast into the constant.
5114 if (Op0->getType() != Op1->getType())
5115 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5116 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5118 // Otherwise, cast the RHS right before the icmp
5119 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5121 return new ICmpInst(I.getPredicate(), Op0, Op1);
5125 if (isa<CastInst>(Op0)) {
5126 // Handle the special case of: icmp (cast bool to X), <cst>
5127 // This comes up when you have code like
5130 // For generality, we handle any zero-extension of any operand comparison
5131 // with a constant or another cast from the same type.
5132 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5133 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5137 if (I.isEquality()) {
5138 Value *A, *B, *C, *D;
5139 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5140 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5141 Value *OtherVal = A == Op1 ? B : A;
5142 return new ICmpInst(I.getPredicate(), OtherVal,
5143 Constant::getNullValue(A->getType()));
5146 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5147 // A^c1 == C^c2 --> A == C^(c1^c2)
5148 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5149 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5150 if (Op1->hasOneUse()) {
5151 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5152 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5153 return new ICmpInst(I.getPredicate(), A,
5154 InsertNewInstBefore(Xor, I));
5157 // A^B == A^D -> B == D
5158 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5159 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5160 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5161 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5165 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5166 (A == Op0 || B == Op0)) {
5167 // A == (A^B) -> B == 0
5168 Value *OtherVal = A == Op0 ? B : A;
5169 return new ICmpInst(I.getPredicate(), OtherVal,
5170 Constant::getNullValue(A->getType()));
5172 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5173 // (A-B) == A -> B == 0
5174 return new ICmpInst(I.getPredicate(), B,
5175 Constant::getNullValue(B->getType()));
5177 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5178 // A == (A-B) -> B == 0
5179 return new ICmpInst(I.getPredicate(), B,
5180 Constant::getNullValue(B->getType()));
5183 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5184 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5185 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5186 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5187 Value *X = 0, *Y = 0, *Z = 0;
5190 X = B; Y = D; Z = A;
5191 } else if (A == D) {
5192 X = B; Y = C; Z = A;
5193 } else if (B == C) {
5194 X = A; Y = D; Z = B;
5195 } else if (B == D) {
5196 X = A; Y = C; Z = B;
5199 if (X) { // Build (X^Y) & Z
5200 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5201 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5202 I.setOperand(0, Op1);
5203 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5208 return Changed ? &I : 0;
5212 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5213 /// and CmpRHS are both known to be integer constants.
5214 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5215 ConstantInt *DivRHS) {
5216 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5217 const APInt &CmpRHSV = CmpRHS->getValue();
5219 // FIXME: If the operand types don't match the type of the divide
5220 // then don't attempt this transform. The code below doesn't have the
5221 // logic to deal with a signed divide and an unsigned compare (and
5222 // vice versa). This is because (x /s C1) <s C2 produces different
5223 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5224 // (x /u C1) <u C2. Simply casting the operands and result won't
5225 // work. :( The if statement below tests that condition and bails
5227 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5228 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5230 if (DivRHS->isZero())
5231 return 0; // The ProdOV computation fails on divide by zero.
5233 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5234 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5235 // C2 (CI). By solving for X we can turn this into a range check
5236 // instead of computing a divide.
5237 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5239 // Determine if the product overflows by seeing if the product is
5240 // not equal to the divide. Make sure we do the same kind of divide
5241 // as in the LHS instruction that we're folding.
5242 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5243 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5245 // Get the ICmp opcode
5246 ICmpInst::Predicate Pred = ICI.getPredicate();
5248 // Figure out the interval that is being checked. For example, a comparison
5249 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5250 // Compute this interval based on the constants involved and the signedness of
5251 // the compare/divide. This computes a half-open interval, keeping track of
5252 // whether either value in the interval overflows. After analysis each
5253 // overflow variable is set to 0 if it's corresponding bound variable is valid
5254 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5255 int LoOverflow = 0, HiOverflow = 0;
5256 ConstantInt *LoBound = 0, *HiBound = 0;
5259 if (!DivIsSigned) { // udiv
5260 // e.g. X/5 op 3 --> [15, 20)
5262 HiOverflow = LoOverflow = ProdOV;
5264 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5265 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5266 if (CmpRHSV == 0) { // (X / pos) op 0
5267 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5268 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5270 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5271 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5272 HiOverflow = LoOverflow = ProdOV;
5274 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5275 } else { // (X / pos) op neg
5276 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5277 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5278 LoOverflow = AddWithOverflow(LoBound, Prod,
5279 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5280 HiBound = AddOne(Prod);
5281 HiOverflow = ProdOV ? -1 : 0;
5283 } else { // Divisor is < 0.
5284 if (CmpRHSV == 0) { // (X / neg) op 0
5285 // e.g. X/-5 op 0 --> [-4, 5)
5286 LoBound = AddOne(DivRHS);
5287 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5288 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5289 HiOverflow = 1; // [INTMIN+1, overflow)
5290 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5292 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5293 // e.g. X/-5 op 3 --> [-19, -14)
5294 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5296 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5297 HiBound = AddOne(Prod);
5298 } else { // (X / neg) op neg
5299 // e.g. X/-5 op -3 --> [15, 20)
5301 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5302 HiBound = Subtract(Prod, DivRHS);
5305 // Dividing by a negative swaps the condition. LT <-> GT
5306 Pred = ICmpInst::getSwappedPredicate(Pred);
5309 Value *X = DivI->getOperand(0);
5311 default: assert(0 && "Unhandled icmp opcode!");
5312 case ICmpInst::ICMP_EQ:
5313 if (LoOverflow && HiOverflow)
5314 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5315 else if (HiOverflow)
5316 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5317 ICmpInst::ICMP_UGE, X, LoBound);
5318 else if (LoOverflow)
5319 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5320 ICmpInst::ICMP_ULT, X, HiBound);
5322 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5323 case ICmpInst::ICMP_NE:
5324 if (LoOverflow && HiOverflow)
5325 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5326 else if (HiOverflow)
5327 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5328 ICmpInst::ICMP_ULT, X, LoBound);
5329 else if (LoOverflow)
5330 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5331 ICmpInst::ICMP_UGE, X, HiBound);
5333 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5334 case ICmpInst::ICMP_ULT:
5335 case ICmpInst::ICMP_SLT:
5336 if (LoOverflow == +1) // Low bound is greater than input range.
5337 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5338 if (LoOverflow == -1) // Low bound is less than input range.
5339 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5340 return new ICmpInst(Pred, X, LoBound);
5341 case ICmpInst::ICMP_UGT:
5342 case ICmpInst::ICMP_SGT:
5343 if (HiOverflow == +1) // High bound greater than input range.
5344 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5345 else if (HiOverflow == -1) // High bound less than input range.
5346 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5347 if (Pred == ICmpInst::ICMP_UGT)
5348 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5350 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5355 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5357 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5360 const APInt &RHSV = RHS->getValue();
5362 switch (LHSI->getOpcode()) {
5363 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5364 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5365 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5367 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5368 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5369 Value *CompareVal = LHSI->getOperand(0);
5371 // If the sign bit of the XorCST is not set, there is no change to
5372 // the operation, just stop using the Xor.
5373 if (!XorCST->getValue().isNegative()) {
5374 ICI.setOperand(0, CompareVal);
5375 AddToWorkList(LHSI);
5379 // Was the old condition true if the operand is positive?
5380 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5382 // If so, the new one isn't.
5383 isTrueIfPositive ^= true;
5385 if (isTrueIfPositive)
5386 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5388 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5392 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5393 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5394 LHSI->getOperand(0)->hasOneUse()) {
5395 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5397 // If the LHS is an AND of a truncating cast, we can widen the
5398 // and/compare to be the input width without changing the value
5399 // produced, eliminating a cast.
5400 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5401 // We can do this transformation if either the AND constant does not
5402 // have its sign bit set or if it is an equality comparison.
5403 // Extending a relational comparison when we're checking the sign
5404 // bit would not work.
5405 if (Cast->hasOneUse() &&
5406 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5407 RHSV.isPositive())) {
5409 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5410 APInt NewCST = AndCST->getValue();
5411 NewCST.zext(BitWidth);
5413 NewCI.zext(BitWidth);
5414 Instruction *NewAnd =
5415 BinaryOperator::createAnd(Cast->getOperand(0),
5416 ConstantInt::get(NewCST),LHSI->getName());
5417 InsertNewInstBefore(NewAnd, ICI);
5418 return new ICmpInst(ICI.getPredicate(), NewAnd,
5419 ConstantInt::get(NewCI));
5423 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5424 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5425 // happens a LOT in code produced by the C front-end, for bitfield
5427 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5428 if (Shift && !Shift->isShift())
5432 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5433 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5434 const Type *AndTy = AndCST->getType(); // Type of the and.
5436 // We can fold this as long as we can't shift unknown bits
5437 // into the mask. This can only happen with signed shift
5438 // rights, as they sign-extend.
5440 bool CanFold = Shift->isLogicalShift();
5442 // To test for the bad case of the signed shr, see if any
5443 // of the bits shifted in could be tested after the mask.
5444 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5445 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5447 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5448 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5449 AndCST->getValue()) == 0)
5455 if (Shift->getOpcode() == Instruction::Shl)
5456 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5458 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5460 // Check to see if we are shifting out any of the bits being
5462 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5463 // If we shifted bits out, the fold is not going to work out.
5464 // As a special case, check to see if this means that the
5465 // result is always true or false now.
5466 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5467 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5468 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5469 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5471 ICI.setOperand(1, NewCst);
5472 Constant *NewAndCST;
5473 if (Shift->getOpcode() == Instruction::Shl)
5474 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5476 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5477 LHSI->setOperand(1, NewAndCST);
5478 LHSI->setOperand(0, Shift->getOperand(0));
5479 AddToWorkList(Shift); // Shift is dead.
5480 AddUsesToWorkList(ICI);
5486 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5487 // preferable because it allows the C<<Y expression to be hoisted out
5488 // of a loop if Y is invariant and X is not.
5489 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5490 ICI.isEquality() && !Shift->isArithmeticShift() &&
5491 isa<Instruction>(Shift->getOperand(0))) {
5494 if (Shift->getOpcode() == Instruction::LShr) {
5495 NS = BinaryOperator::createShl(AndCST,
5496 Shift->getOperand(1), "tmp");
5498 // Insert a logical shift.
5499 NS = BinaryOperator::createLShr(AndCST,
5500 Shift->getOperand(1), "tmp");
5502 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5504 // Compute X & (C << Y).
5505 Instruction *NewAnd =
5506 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5507 InsertNewInstBefore(NewAnd, ICI);
5509 ICI.setOperand(0, NewAnd);
5515 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5516 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5519 uint32_t TypeBits = RHSV.getBitWidth();
5521 // Check that the shift amount is in range. If not, don't perform
5522 // undefined shifts. When the shift is visited it will be
5524 if (ShAmt->uge(TypeBits))
5527 if (ICI.isEquality()) {
5528 // If we are comparing against bits always shifted out, the
5529 // comparison cannot succeed.
5531 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5532 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5533 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5534 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5535 return ReplaceInstUsesWith(ICI, Cst);
5538 if (LHSI->hasOneUse()) {
5539 // Otherwise strength reduce the shift into an and.
5540 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5542 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5545 BinaryOperator::createAnd(LHSI->getOperand(0),
5546 Mask, LHSI->getName()+".mask");
5547 Value *And = InsertNewInstBefore(AndI, ICI);
5548 return new ICmpInst(ICI.getPredicate(), And,
5549 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5553 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5554 bool TrueIfSigned = false;
5555 if (LHSI->hasOneUse() &&
5556 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5557 // (X << 31) <s 0 --> (X&1) != 0
5558 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5559 (TypeBits-ShAmt->getZExtValue()-1));
5561 BinaryOperator::createAnd(LHSI->getOperand(0),
5562 Mask, LHSI->getName()+".mask");
5563 Value *And = InsertNewInstBefore(AndI, ICI);
5565 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5566 And, Constant::getNullValue(And->getType()));
5571 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5572 case Instruction::AShr: {
5573 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5576 if (ICI.isEquality()) {
5577 // Check that the shift amount is in range. If not, don't perform
5578 // undefined shifts. When the shift is visited it will be
5580 uint32_t TypeBits = RHSV.getBitWidth();
5581 if (ShAmt->uge(TypeBits))
5583 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5585 // If we are comparing against bits always shifted out, the
5586 // comparison cannot succeed.
5587 APInt Comp = RHSV << ShAmtVal;
5588 if (LHSI->getOpcode() == Instruction::LShr)
5589 Comp = Comp.lshr(ShAmtVal);
5591 Comp = Comp.ashr(ShAmtVal);
5593 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5594 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5595 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5596 return ReplaceInstUsesWith(ICI, Cst);
5599 if (LHSI->hasOneUse() || RHSV == 0) {
5600 // Otherwise strength reduce the shift into an and.
5601 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5602 Constant *Mask = ConstantInt::get(Val);
5605 BinaryOperator::createAnd(LHSI->getOperand(0),
5606 Mask, LHSI->getName()+".mask");
5607 Value *And = InsertNewInstBefore(AndI, ICI);
5608 return new ICmpInst(ICI.getPredicate(), And,
5609 ConstantExpr::getShl(RHS, ShAmt));
5615 case Instruction::SDiv:
5616 case Instruction::UDiv:
5617 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5618 // Fold this div into the comparison, producing a range check.
5619 // Determine, based on the divide type, what the range is being
5620 // checked. If there is an overflow on the low or high side, remember
5621 // it, otherwise compute the range [low, hi) bounding the new value.
5622 // See: InsertRangeTest above for the kinds of replacements possible.
5623 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5624 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5630 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5631 if (ICI.isEquality()) {
5632 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5634 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5635 // the second operand is a constant, simplify a bit.
5636 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5637 switch (BO->getOpcode()) {
5638 case Instruction::SRem:
5639 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5640 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5641 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5642 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5643 Instruction *NewRem =
5644 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5646 InsertNewInstBefore(NewRem, ICI);
5647 return new ICmpInst(ICI.getPredicate(), NewRem,
5648 Constant::getNullValue(BO->getType()));
5652 case Instruction::Add:
5653 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5654 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5655 if (BO->hasOneUse())
5656 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5657 Subtract(RHS, BOp1C));
5658 } else if (RHSV == 0) {
5659 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5660 // efficiently invertible, or if the add has just this one use.
5661 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5663 if (Value *NegVal = dyn_castNegVal(BOp1))
5664 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5665 else if (Value *NegVal = dyn_castNegVal(BOp0))
5666 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5667 else if (BO->hasOneUse()) {
5668 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5669 InsertNewInstBefore(Neg, ICI);
5671 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5675 case Instruction::Xor:
5676 // For the xor case, we can xor two constants together, eliminating
5677 // the explicit xor.
5678 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5679 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5680 ConstantExpr::getXor(RHS, BOC));
5683 case Instruction::Sub:
5684 // Replace (([sub|xor] A, B) != 0) with (A != B)
5686 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5690 case Instruction::Or:
5691 // If bits are being or'd in that are not present in the constant we
5692 // are comparing against, then the comparison could never succeed!
5693 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5694 Constant *NotCI = ConstantExpr::getNot(RHS);
5695 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5696 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5701 case Instruction::And:
5702 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5703 // If bits are being compared against that are and'd out, then the
5704 // comparison can never succeed!
5705 if ((RHSV & ~BOC->getValue()) != 0)
5706 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5709 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5710 if (RHS == BOC && RHSV.isPowerOf2())
5711 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5712 ICmpInst::ICMP_NE, LHSI,
5713 Constant::getNullValue(RHS->getType()));
5715 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5716 if (isSignBit(BOC)) {
5717 Value *X = BO->getOperand(0);
5718 Constant *Zero = Constant::getNullValue(X->getType());
5719 ICmpInst::Predicate pred = isICMP_NE ?
5720 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5721 return new ICmpInst(pred, X, Zero);
5724 // ((X & ~7) == 0) --> X < 8
5725 if (RHSV == 0 && isHighOnes(BOC)) {
5726 Value *X = BO->getOperand(0);
5727 Constant *NegX = ConstantExpr::getNeg(BOC);
5728 ICmpInst::Predicate pred = isICMP_NE ?
5729 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5730 return new ICmpInst(pred, X, NegX);
5735 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5736 // Handle icmp {eq|ne} <intrinsic>, intcst.
5737 if (II->getIntrinsicID() == Intrinsic::bswap) {
5739 ICI.setOperand(0, II->getOperand(1));
5740 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5744 } else { // Not a ICMP_EQ/ICMP_NE
5745 // If the LHS is a cast from an integral value of the same size,
5746 // then since we know the RHS is a constant, try to simlify.
5747 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5748 Value *CastOp = Cast->getOperand(0);
5749 const Type *SrcTy = CastOp->getType();
5750 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5751 if (SrcTy->isInteger() &&
5752 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5753 // If this is an unsigned comparison, try to make the comparison use
5754 // smaller constant values.
5755 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5756 // X u< 128 => X s> -1
5757 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5758 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5759 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5760 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5761 // X u> 127 => X s< 0
5762 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5763 Constant::getNullValue(SrcTy));
5771 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5772 /// We only handle extending casts so far.
5774 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5775 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5776 Value *LHSCIOp = LHSCI->getOperand(0);
5777 const Type *SrcTy = LHSCIOp->getType();
5778 const Type *DestTy = LHSCI->getType();
5781 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5782 // integer type is the same size as the pointer type.
5783 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5784 getTargetData().getPointerSizeInBits() ==
5785 cast<IntegerType>(DestTy)->getBitWidth()) {
5787 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5788 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5789 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5790 RHSOp = RHSC->getOperand(0);
5791 // If the pointer types don't match, insert a bitcast.
5792 if (LHSCIOp->getType() != RHSOp->getType())
5793 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5794 LHSCIOp->getType(), ICI);
5798 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5801 // The code below only handles extension cast instructions, so far.
5803 if (LHSCI->getOpcode() != Instruction::ZExt &&
5804 LHSCI->getOpcode() != Instruction::SExt)
5807 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5808 bool isSignedCmp = ICI.isSignedPredicate();
5810 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5811 // Not an extension from the same type?
5812 RHSCIOp = CI->getOperand(0);
5813 if (RHSCIOp->getType() != LHSCIOp->getType())
5816 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5817 // and the other is a zext), then we can't handle this.
5818 if (CI->getOpcode() != LHSCI->getOpcode())
5821 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5822 // then we can't handle this.
5823 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5826 // Okay, just insert a compare of the reduced operands now!
5827 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5830 // If we aren't dealing with a constant on the RHS, exit early
5831 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5835 // Compute the constant that would happen if we truncated to SrcTy then
5836 // reextended to DestTy.
5837 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5838 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5840 // If the re-extended constant didn't change...
5842 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5843 // For example, we might have:
5844 // %A = sext short %X to uint
5845 // %B = icmp ugt uint %A, 1330
5846 // It is incorrect to transform this into
5847 // %B = icmp ugt short %X, 1330
5848 // because %A may have negative value.
5850 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5851 // OR operation is EQ/NE.
5852 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5853 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5858 // The re-extended constant changed so the constant cannot be represented
5859 // in the shorter type. Consequently, we cannot emit a simple comparison.
5861 // First, handle some easy cases. We know the result cannot be equal at this
5862 // point so handle the ICI.isEquality() cases
5863 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5864 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5865 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5866 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5868 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5869 // should have been folded away previously and not enter in here.
5872 // We're performing a signed comparison.
5873 if (cast<ConstantInt>(CI)->getValue().isNegative())
5874 Result = ConstantInt::getFalse(); // X < (small) --> false
5876 Result = ConstantInt::getTrue(); // X < (large) --> true
5878 // We're performing an unsigned comparison.
5880 // We're performing an unsigned comp with a sign extended value.
5881 // This is true if the input is >= 0. [aka >s -1]
5882 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5883 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5884 NegOne, ICI.getName()), ICI);
5886 // Unsigned extend & unsigned compare -> always true.
5887 Result = ConstantInt::getTrue();
5891 // Finally, return the value computed.
5892 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5893 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5894 return ReplaceInstUsesWith(ICI, Result);
5896 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5897 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5898 "ICmp should be folded!");
5899 if (Constant *CI = dyn_cast<Constant>(Result))
5900 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5902 return BinaryOperator::createNot(Result);
5906 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5907 return commonShiftTransforms(I);
5910 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5911 return commonShiftTransforms(I);
5914 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5915 if (Instruction *R = commonShiftTransforms(I))
5918 Value *Op0 = I.getOperand(0);
5920 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5921 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5922 if (CSI->isAllOnesValue())
5923 return ReplaceInstUsesWith(I, CSI);
5925 // See if we can turn a signed shr into an unsigned shr.
5926 if (MaskedValueIsZero(Op0,
5927 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
5928 return BinaryOperator::createLShr(Op0, I.getOperand(1));
5933 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5934 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5935 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5937 // shl X, 0 == X and shr X, 0 == X
5938 // shl 0, X == 0 and shr 0, X == 0
5939 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5940 Op0 == Constant::getNullValue(Op0->getType()))
5941 return ReplaceInstUsesWith(I, Op0);
5943 if (isa<UndefValue>(Op0)) {
5944 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5945 return ReplaceInstUsesWith(I, Op0);
5946 else // undef << X -> 0, undef >>u X -> 0
5947 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5949 if (isa<UndefValue>(Op1)) {
5950 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5951 return ReplaceInstUsesWith(I, Op0);
5952 else // X << undef, X >>u undef -> 0
5953 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5956 // Try to fold constant and into select arguments.
5957 if (isa<Constant>(Op0))
5958 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5959 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5962 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5963 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5968 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5969 BinaryOperator &I) {
5970 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5972 // See if we can simplify any instructions used by the instruction whose sole
5973 // purpose is to compute bits we don't care about.
5974 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5975 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5976 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5977 KnownZero, KnownOne))
5980 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5981 // of a signed value.
5983 if (Op1->uge(TypeBits)) {
5984 if (I.getOpcode() != Instruction::AShr)
5985 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5987 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5992 // ((X*C1) << C2) == (X * (C1 << C2))
5993 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5994 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5995 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5996 return BinaryOperator::createMul(BO->getOperand(0),
5997 ConstantExpr::getShl(BOOp, Op1));
5999 // Try to fold constant and into select arguments.
6000 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6001 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6003 if (isa<PHINode>(Op0))
6004 if (Instruction *NV = FoldOpIntoPhi(I))
6007 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6008 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6009 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6010 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6011 // place. Don't try to do this transformation in this case. Also, we
6012 // require that the input operand is a shift-by-constant so that we have
6013 // confidence that the shifts will get folded together. We could do this
6014 // xform in more cases, but it is unlikely to be profitable.
6015 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6016 isa<ConstantInt>(TrOp->getOperand(1))) {
6017 // Okay, we'll do this xform. Make the shift of shift.
6018 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6019 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6021 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6023 // For logical shifts, the truncation has the effect of making the high
6024 // part of the register be zeros. Emulate this by inserting an AND to
6025 // clear the top bits as needed. This 'and' will usually be zapped by
6026 // other xforms later if dead.
6027 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6028 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6029 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6031 // The mask we constructed says what the trunc would do if occurring
6032 // between the shifts. We want to know the effect *after* the second
6033 // shift. We know that it is a logical shift by a constant, so adjust the
6034 // mask as appropriate.
6035 if (I.getOpcode() == Instruction::Shl)
6036 MaskV <<= Op1->getZExtValue();
6038 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6039 MaskV = MaskV.lshr(Op1->getZExtValue());
6042 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6044 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6046 // Return the value truncated to the interesting size.
6047 return new TruncInst(And, I.getType());
6051 if (Op0->hasOneUse()) {
6052 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6053 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6056 switch (Op0BO->getOpcode()) {
6058 case Instruction::Add:
6059 case Instruction::And:
6060 case Instruction::Or:
6061 case Instruction::Xor: {
6062 // These operators commute.
6063 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6064 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6065 match(Op0BO->getOperand(1),
6066 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6067 Instruction *YS = BinaryOperator::createShl(
6068 Op0BO->getOperand(0), Op1,
6070 InsertNewInstBefore(YS, I); // (Y << C)
6072 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6073 Op0BO->getOperand(1)->getName());
6074 InsertNewInstBefore(X, I); // (X + (Y << C))
6075 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6076 return BinaryOperator::createAnd(X, ConstantInt::get(
6077 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6080 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6081 Value *Op0BOOp1 = Op0BO->getOperand(1);
6082 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6084 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6085 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6087 Instruction *YS = BinaryOperator::createShl(
6088 Op0BO->getOperand(0), Op1,
6090 InsertNewInstBefore(YS, I); // (Y << C)
6092 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6093 V1->getName()+".mask");
6094 InsertNewInstBefore(XM, I); // X & (CC << C)
6096 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6101 case Instruction::Sub: {
6102 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6103 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6104 match(Op0BO->getOperand(0),
6105 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6106 Instruction *YS = BinaryOperator::createShl(
6107 Op0BO->getOperand(1), Op1,
6109 InsertNewInstBefore(YS, I); // (Y << C)
6111 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6112 Op0BO->getOperand(0)->getName());
6113 InsertNewInstBefore(X, I); // (X + (Y << C))
6114 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6115 return BinaryOperator::createAnd(X, ConstantInt::get(
6116 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6119 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6120 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6121 match(Op0BO->getOperand(0),
6122 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6123 m_ConstantInt(CC))) && V2 == Op1 &&
6124 cast<BinaryOperator>(Op0BO->getOperand(0))
6125 ->getOperand(0)->hasOneUse()) {
6126 Instruction *YS = BinaryOperator::createShl(
6127 Op0BO->getOperand(1), Op1,
6129 InsertNewInstBefore(YS, I); // (Y << C)
6131 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6132 V1->getName()+".mask");
6133 InsertNewInstBefore(XM, I); // X & (CC << C)
6135 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6143 // If the operand is an bitwise operator with a constant RHS, and the
6144 // shift is the only use, we can pull it out of the shift.
6145 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6146 bool isValid = true; // Valid only for And, Or, Xor
6147 bool highBitSet = false; // Transform if high bit of constant set?
6149 switch (Op0BO->getOpcode()) {
6150 default: isValid = false; break; // Do not perform transform!
6151 case Instruction::Add:
6152 isValid = isLeftShift;
6154 case Instruction::Or:
6155 case Instruction::Xor:
6158 case Instruction::And:
6163 // If this is a signed shift right, and the high bit is modified
6164 // by the logical operation, do not perform the transformation.
6165 // The highBitSet boolean indicates the value of the high bit of
6166 // the constant which would cause it to be modified for this
6169 if (isValid && I.getOpcode() == Instruction::AShr)
6170 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6173 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6175 Instruction *NewShift =
6176 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6177 InsertNewInstBefore(NewShift, I);
6178 NewShift->takeName(Op0BO);
6180 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6187 // Find out if this is a shift of a shift by a constant.
6188 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6189 if (ShiftOp && !ShiftOp->isShift())
6192 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6193 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6194 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6195 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6196 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6197 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6198 Value *X = ShiftOp->getOperand(0);
6200 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6201 if (AmtSum > TypeBits)
6204 const IntegerType *Ty = cast<IntegerType>(I.getType());
6206 // Check for (X << c1) << c2 and (X >> c1) >> c2
6207 if (I.getOpcode() == ShiftOp->getOpcode()) {
6208 return BinaryOperator::create(I.getOpcode(), X,
6209 ConstantInt::get(Ty, AmtSum));
6210 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6211 I.getOpcode() == Instruction::AShr) {
6212 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6213 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6214 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6215 I.getOpcode() == Instruction::LShr) {
6216 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6217 Instruction *Shift =
6218 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6219 InsertNewInstBefore(Shift, I);
6221 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6222 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6225 // Okay, if we get here, one shift must be left, and the other shift must be
6226 // right. See if the amounts are equal.
6227 if (ShiftAmt1 == ShiftAmt2) {
6228 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6229 if (I.getOpcode() == Instruction::Shl) {
6230 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6231 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6233 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6234 if (I.getOpcode() == Instruction::LShr) {
6235 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6236 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6238 // We can simplify ((X << C) >>s C) into a trunc + sext.
6239 // NOTE: we could do this for any C, but that would make 'unusual' integer
6240 // types. For now, just stick to ones well-supported by the code
6242 const Type *SExtType = 0;
6243 switch (Ty->getBitWidth() - ShiftAmt1) {
6250 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6255 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6256 InsertNewInstBefore(NewTrunc, I);
6257 return new SExtInst(NewTrunc, Ty);
6259 // Otherwise, we can't handle it yet.
6260 } else if (ShiftAmt1 < ShiftAmt2) {
6261 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6263 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6264 if (I.getOpcode() == Instruction::Shl) {
6265 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6266 ShiftOp->getOpcode() == Instruction::AShr);
6267 Instruction *Shift =
6268 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6269 InsertNewInstBefore(Shift, I);
6271 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6272 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6275 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6276 if (I.getOpcode() == Instruction::LShr) {
6277 assert(ShiftOp->getOpcode() == Instruction::Shl);
6278 Instruction *Shift =
6279 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6280 InsertNewInstBefore(Shift, I);
6282 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6283 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6286 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6288 assert(ShiftAmt2 < ShiftAmt1);
6289 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6291 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6292 if (I.getOpcode() == Instruction::Shl) {
6293 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6294 ShiftOp->getOpcode() == Instruction::AShr);
6295 Instruction *Shift =
6296 BinaryOperator::create(ShiftOp->getOpcode(), X,
6297 ConstantInt::get(Ty, ShiftDiff));
6298 InsertNewInstBefore(Shift, I);
6300 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6301 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6304 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6305 if (I.getOpcode() == Instruction::LShr) {
6306 assert(ShiftOp->getOpcode() == Instruction::Shl);
6307 Instruction *Shift =
6308 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6309 InsertNewInstBefore(Shift, I);
6311 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6312 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6315 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6322 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6323 /// expression. If so, decompose it, returning some value X, such that Val is
6326 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6328 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6329 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6330 Offset = CI->getZExtValue();
6332 return ConstantInt::get(Type::Int32Ty, 0);
6333 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6334 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6335 if (I->getOpcode() == Instruction::Shl) {
6336 // This is a value scaled by '1 << the shift amt'.
6337 Scale = 1U << RHS->getZExtValue();
6339 return I->getOperand(0);
6340 } else if (I->getOpcode() == Instruction::Mul) {
6341 // This value is scaled by 'RHS'.
6342 Scale = RHS->getZExtValue();
6344 return I->getOperand(0);
6345 } else if (I->getOpcode() == Instruction::Add) {
6346 // We have X+C. Check to see if we really have (X*C2)+C1,
6347 // where C1 is divisible by C2.
6350 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6351 Offset += RHS->getZExtValue();
6358 // Otherwise, we can't look past this.
6365 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6366 /// try to eliminate the cast by moving the type information into the alloc.
6367 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6368 AllocationInst &AI) {
6369 const PointerType *PTy = cast<PointerType>(CI.getType());
6371 // Remove any uses of AI that are dead.
6372 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6374 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6375 Instruction *User = cast<Instruction>(*UI++);
6376 if (isInstructionTriviallyDead(User)) {
6377 while (UI != E && *UI == User)
6378 ++UI; // If this instruction uses AI more than once, don't break UI.
6381 DOUT << "IC: DCE: " << *User;
6382 EraseInstFromFunction(*User);
6386 // Get the type really allocated and the type casted to.
6387 const Type *AllocElTy = AI.getAllocatedType();
6388 const Type *CastElTy = PTy->getElementType();
6389 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6391 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6392 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6393 if (CastElTyAlign < AllocElTyAlign) return 0;
6395 // If the allocation has multiple uses, only promote it if we are strictly
6396 // increasing the alignment of the resultant allocation. If we keep it the
6397 // same, we open the door to infinite loops of various kinds.
6398 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6400 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6401 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6402 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6404 // See if we can satisfy the modulus by pulling a scale out of the array
6406 unsigned ArraySizeScale;
6408 Value *NumElements = // See if the array size is a decomposable linear expr.
6409 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6411 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6413 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6414 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6416 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6421 // If the allocation size is constant, form a constant mul expression
6422 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6423 if (isa<ConstantInt>(NumElements))
6424 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6425 // otherwise multiply the amount and the number of elements
6426 else if (Scale != 1) {
6427 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6428 Amt = InsertNewInstBefore(Tmp, AI);
6432 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6433 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6434 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6435 Amt = InsertNewInstBefore(Tmp, AI);
6438 AllocationInst *New;
6439 if (isa<MallocInst>(AI))
6440 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6442 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6443 InsertNewInstBefore(New, AI);
6446 // If the allocation has multiple uses, insert a cast and change all things
6447 // that used it to use the new cast. This will also hack on CI, but it will
6449 if (!AI.hasOneUse()) {
6450 AddUsesToWorkList(AI);
6451 // New is the allocation instruction, pointer typed. AI is the original
6452 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6453 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6454 InsertNewInstBefore(NewCast, AI);
6455 AI.replaceAllUsesWith(NewCast);
6457 return ReplaceInstUsesWith(CI, New);
6460 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6461 /// and return it as type Ty without inserting any new casts and without
6462 /// changing the computed value. This is used by code that tries to decide
6463 /// whether promoting or shrinking integer operations to wider or smaller types
6464 /// will allow us to eliminate a truncate or extend.
6466 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6467 /// extension operation if Ty is larger.
6468 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6469 unsigned CastOpc, int &NumCastsRemoved) {
6470 // We can always evaluate constants in another type.
6471 if (isa<ConstantInt>(V))
6474 Instruction *I = dyn_cast<Instruction>(V);
6475 if (!I) return false;
6477 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6479 // If this is an extension or truncate, we can often eliminate it.
6480 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6481 // If this is a cast from the destination type, we can trivially eliminate
6482 // it, and this will remove a cast overall.
6483 if (I->getOperand(0)->getType() == Ty) {
6484 // If the first operand is itself a cast, and is eliminable, do not count
6485 // this as an eliminable cast. We would prefer to eliminate those two
6487 if (!isa<CastInst>(I->getOperand(0)))
6493 // We can't extend or shrink something that has multiple uses: doing so would
6494 // require duplicating the instruction in general, which isn't profitable.
6495 if (!I->hasOneUse()) return false;
6497 switch (I->getOpcode()) {
6498 case Instruction::Add:
6499 case Instruction::Sub:
6500 case Instruction::And:
6501 case Instruction::Or:
6502 case Instruction::Xor:
6503 // These operators can all arbitrarily be extended or truncated.
6504 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6506 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6509 case Instruction::Shl:
6510 // If we are truncating the result of this SHL, and if it's a shift of a
6511 // constant amount, we can always perform a SHL in a smaller type.
6512 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6513 uint32_t BitWidth = Ty->getBitWidth();
6514 if (BitWidth < OrigTy->getBitWidth() &&
6515 CI->getLimitedValue(BitWidth) < BitWidth)
6516 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6520 case Instruction::LShr:
6521 // If this is a truncate of a logical shr, we can truncate it to a smaller
6522 // lshr iff we know that the bits we would otherwise be shifting in are
6524 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6525 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6526 uint32_t BitWidth = Ty->getBitWidth();
6527 if (BitWidth < OrigBitWidth &&
6528 MaskedValueIsZero(I->getOperand(0),
6529 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6530 CI->getLimitedValue(BitWidth) < BitWidth) {
6531 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6536 case Instruction::ZExt:
6537 case Instruction::SExt:
6538 case Instruction::Trunc:
6539 // If this is the same kind of case as our original (e.g. zext+zext), we
6540 // can safely replace it. Note that replacing it does not reduce the number
6541 // of casts in the input.
6542 if (I->getOpcode() == CastOpc)
6547 // TODO: Can handle more cases here.
6554 /// EvaluateInDifferentType - Given an expression that
6555 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6556 /// evaluate the expression.
6557 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6559 if (Constant *C = dyn_cast<Constant>(V))
6560 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6562 // Otherwise, it must be an instruction.
6563 Instruction *I = cast<Instruction>(V);
6564 Instruction *Res = 0;
6565 switch (I->getOpcode()) {
6566 case Instruction::Add:
6567 case Instruction::Sub:
6568 case Instruction::And:
6569 case Instruction::Or:
6570 case Instruction::Xor:
6571 case Instruction::AShr:
6572 case Instruction::LShr:
6573 case Instruction::Shl: {
6574 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6575 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6576 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6577 LHS, RHS, I->getName());
6580 case Instruction::Trunc:
6581 case Instruction::ZExt:
6582 case Instruction::SExt:
6583 // If the source type of the cast is the type we're trying for then we can
6584 // just return the source. There's no need to insert it because it is not
6586 if (I->getOperand(0)->getType() == Ty)
6587 return I->getOperand(0);
6589 // Otherwise, must be the same type of case, so just reinsert a new one.
6590 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6594 // TODO: Can handle more cases here.
6595 assert(0 && "Unreachable!");
6599 return InsertNewInstBefore(Res, *I);
6602 /// @brief Implement the transforms common to all CastInst visitors.
6603 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6604 Value *Src = CI.getOperand(0);
6606 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6607 // eliminate it now.
6608 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6609 if (Instruction::CastOps opc =
6610 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6611 // The first cast (CSrc) is eliminable so we need to fix up or replace
6612 // the second cast (CI). CSrc will then have a good chance of being dead.
6613 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6617 // If we are casting a select then fold the cast into the select
6618 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6619 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6622 // If we are casting a PHI then fold the cast into the PHI
6623 if (isa<PHINode>(Src))
6624 if (Instruction *NV = FoldOpIntoPhi(CI))
6630 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6631 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6632 Value *Src = CI.getOperand(0);
6634 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6635 // If casting the result of a getelementptr instruction with no offset, turn
6636 // this into a cast of the original pointer!
6637 if (GEP->hasAllZeroIndices()) {
6638 // Changing the cast operand is usually not a good idea but it is safe
6639 // here because the pointer operand is being replaced with another
6640 // pointer operand so the opcode doesn't need to change.
6642 CI.setOperand(0, GEP->getOperand(0));
6646 // If the GEP has a single use, and the base pointer is a bitcast, and the
6647 // GEP computes a constant offset, see if we can convert these three
6648 // instructions into fewer. This typically happens with unions and other
6649 // non-type-safe code.
6650 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6651 if (GEP->hasAllConstantIndices()) {
6652 // We are guaranteed to get a constant from EmitGEPOffset.
6653 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6654 int64_t Offset = OffsetV->getSExtValue();
6656 // Get the base pointer input of the bitcast, and the type it points to.
6657 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6658 const Type *GEPIdxTy =
6659 cast<PointerType>(OrigBase->getType())->getElementType();
6660 if (GEPIdxTy->isSized()) {
6661 SmallVector<Value*, 8> NewIndices;
6663 // Start with the index over the outer type. Note that the type size
6664 // might be zero (even if the offset isn't zero) if the indexed type
6665 // is something like [0 x {int, int}]
6666 const Type *IntPtrTy = TD->getIntPtrType();
6667 int64_t FirstIdx = 0;
6668 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6669 FirstIdx = Offset/TySize;
6672 // Handle silly modulus not returning values values [0..TySize).
6676 assert(Offset >= 0);
6678 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6681 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6683 // Index into the types. If we fail, set OrigBase to null.
6685 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6686 const StructLayout *SL = TD->getStructLayout(STy);
6687 if (Offset < (int64_t)SL->getSizeInBytes()) {
6688 unsigned Elt = SL->getElementContainingOffset(Offset);
6689 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6691 Offset -= SL->getElementOffset(Elt);
6692 GEPIdxTy = STy->getElementType(Elt);
6694 // Otherwise, we can't index into this, bail out.
6698 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6699 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6700 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6701 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6704 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6706 GEPIdxTy = STy->getElementType();
6708 // Otherwise, we can't index into this, bail out.
6714 // If we were able to index down into an element, create the GEP
6715 // and bitcast the result. This eliminates one bitcast, potentially
6717 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6719 NewIndices.end(), "");
6720 InsertNewInstBefore(NGEP, CI);
6721 NGEP->takeName(GEP);
6723 if (isa<BitCastInst>(CI))
6724 return new BitCastInst(NGEP, CI.getType());
6725 assert(isa<PtrToIntInst>(CI));
6726 return new PtrToIntInst(NGEP, CI.getType());
6733 return commonCastTransforms(CI);
6738 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6739 /// integer types. This function implements the common transforms for all those
6741 /// @brief Implement the transforms common to CastInst with integer operands
6742 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6743 if (Instruction *Result = commonCastTransforms(CI))
6746 Value *Src = CI.getOperand(0);
6747 const Type *SrcTy = Src->getType();
6748 const Type *DestTy = CI.getType();
6749 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6750 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6752 // See if we can simplify any instructions used by the LHS whose sole
6753 // purpose is to compute bits we don't care about.
6754 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6755 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6756 KnownZero, KnownOne))
6759 // If the source isn't an instruction or has more than one use then we
6760 // can't do anything more.
6761 Instruction *SrcI = dyn_cast<Instruction>(Src);
6762 if (!SrcI || !Src->hasOneUse())
6765 // Attempt to propagate the cast into the instruction for int->int casts.
6766 int NumCastsRemoved = 0;
6767 if (!isa<BitCastInst>(CI) &&
6768 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6769 CI.getOpcode(), NumCastsRemoved)) {
6770 // If this cast is a truncate, evaluting in a different type always
6771 // eliminates the cast, so it is always a win. If this is a zero-extension,
6772 // we need to do an AND to maintain the clear top-part of the computation,
6773 // so we require that the input have eliminated at least one cast. If this
6774 // is a sign extension, we insert two new casts (to do the extension) so we
6775 // require that two casts have been eliminated.
6777 switch (CI.getOpcode()) {
6779 // All the others use floating point so we shouldn't actually
6780 // get here because of the check above.
6781 assert(0 && "Unknown cast type");
6782 case Instruction::Trunc:
6785 case Instruction::ZExt:
6786 DoXForm = NumCastsRemoved >= 1;
6788 case Instruction::SExt:
6789 DoXForm = NumCastsRemoved >= 2;
6794 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6795 CI.getOpcode() == Instruction::SExt);
6796 assert(Res->getType() == DestTy);
6797 switch (CI.getOpcode()) {
6798 default: assert(0 && "Unknown cast type!");
6799 case Instruction::Trunc:
6800 case Instruction::BitCast:
6801 // Just replace this cast with the result.
6802 return ReplaceInstUsesWith(CI, Res);
6803 case Instruction::ZExt: {
6804 // We need to emit an AND to clear the high bits.
6805 assert(SrcBitSize < DestBitSize && "Not a zext?");
6806 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6808 return BinaryOperator::createAnd(Res, C);
6810 case Instruction::SExt:
6811 // We need to emit a cast to truncate, then a cast to sext.
6812 return CastInst::create(Instruction::SExt,
6813 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6819 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6820 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6822 switch (SrcI->getOpcode()) {
6823 case Instruction::Add:
6824 case Instruction::Mul:
6825 case Instruction::And:
6826 case Instruction::Or:
6827 case Instruction::Xor:
6828 // If we are discarding information, rewrite.
6829 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6830 // Don't insert two casts if they cannot be eliminated. We allow
6831 // two casts to be inserted if the sizes are the same. This could
6832 // only be converting signedness, which is a noop.
6833 if (DestBitSize == SrcBitSize ||
6834 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6835 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6836 Instruction::CastOps opcode = CI.getOpcode();
6837 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6838 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6839 return BinaryOperator::create(
6840 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6844 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6845 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6846 SrcI->getOpcode() == Instruction::Xor &&
6847 Op1 == ConstantInt::getTrue() &&
6848 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6849 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6850 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6853 case Instruction::SDiv:
6854 case Instruction::UDiv:
6855 case Instruction::SRem:
6856 case Instruction::URem:
6857 // If we are just changing the sign, rewrite.
6858 if (DestBitSize == SrcBitSize) {
6859 // Don't insert two casts if they cannot be eliminated. We allow
6860 // two casts to be inserted if the sizes are the same. This could
6861 // only be converting signedness, which is a noop.
6862 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6863 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6864 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6866 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6868 return BinaryOperator::create(
6869 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6874 case Instruction::Shl:
6875 // Allow changing the sign of the source operand. Do not allow
6876 // changing the size of the shift, UNLESS the shift amount is a
6877 // constant. We must not change variable sized shifts to a smaller
6878 // size, because it is undefined to shift more bits out than exist
6880 if (DestBitSize == SrcBitSize ||
6881 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6882 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6883 Instruction::BitCast : Instruction::Trunc);
6884 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6885 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6886 return BinaryOperator::createShl(Op0c, Op1c);
6889 case Instruction::AShr:
6890 // If this is a signed shr, and if all bits shifted in are about to be
6891 // truncated off, turn it into an unsigned shr to allow greater
6893 if (DestBitSize < SrcBitSize &&
6894 isa<ConstantInt>(Op1)) {
6895 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6896 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6897 // Insert the new logical shift right.
6898 return BinaryOperator::createLShr(Op0, Op1);
6906 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6907 if (Instruction *Result = commonIntCastTransforms(CI))
6910 Value *Src = CI.getOperand(0);
6911 const Type *Ty = CI.getType();
6912 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6913 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6915 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6916 switch (SrcI->getOpcode()) {
6918 case Instruction::LShr:
6919 // We can shrink lshr to something smaller if we know the bits shifted in
6920 // are already zeros.
6921 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6922 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6924 // Get a mask for the bits shifting in.
6925 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6926 Value* SrcIOp0 = SrcI->getOperand(0);
6927 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6928 if (ShAmt >= DestBitWidth) // All zeros.
6929 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6931 // Okay, we can shrink this. Truncate the input, then return a new
6933 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6934 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6936 return BinaryOperator::createLShr(V1, V2);
6938 } else { // This is a variable shr.
6940 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6941 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6942 // loop-invariant and CSE'd.
6943 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6944 Value *One = ConstantInt::get(SrcI->getType(), 1);
6946 Value *V = InsertNewInstBefore(
6947 BinaryOperator::createShl(One, SrcI->getOperand(1),
6949 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6950 SrcI->getOperand(0),
6952 Value *Zero = Constant::getNullValue(V->getType());
6953 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6963 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6964 // If one of the common conversion will work ..
6965 if (Instruction *Result = commonIntCastTransforms(CI))
6968 Value *Src = CI.getOperand(0);
6970 // If this is a cast of a cast
6971 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6972 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6973 // types and if the sizes are just right we can convert this into a logical
6974 // 'and' which will be much cheaper than the pair of casts.
6975 if (isa<TruncInst>(CSrc)) {
6976 // Get the sizes of the types involved
6977 Value *A = CSrc->getOperand(0);
6978 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6979 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6980 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6981 // If we're actually extending zero bits and the trunc is a no-op
6982 if (MidSize < DstSize && SrcSize == DstSize) {
6983 // Replace both of the casts with an And of the type mask.
6984 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6985 Constant *AndConst = ConstantInt::get(AndValue);
6987 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6988 // Unfortunately, if the type changed, we need to cast it back.
6989 if (And->getType() != CI.getType()) {
6990 And->setName(CSrc->getName()+".mask");
6991 InsertNewInstBefore(And, CI);
6992 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6999 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7000 // If we are just checking for a icmp eq of a single bit and zext'ing it
7001 // to an integer, then shift the bit to the appropriate place and then
7002 // cast to integer to avoid the comparison.
7003 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7004 const APInt &Op1CV = Op1C->getValue();
7006 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7007 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7008 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7009 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7010 Value *In = ICI->getOperand(0);
7011 Value *Sh = ConstantInt::get(In->getType(),
7012 In->getType()->getPrimitiveSizeInBits()-1);
7013 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7014 In->getName()+".lobit"),
7016 if (In->getType() != CI.getType())
7017 In = CastInst::createIntegerCast(In, CI.getType(),
7018 false/*ZExt*/, "tmp", &CI);
7020 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7021 Constant *One = ConstantInt::get(In->getType(), 1);
7022 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7023 In->getName()+".not"),
7027 return ReplaceInstUsesWith(CI, In);
7032 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7033 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7034 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7035 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7036 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7037 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7038 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7039 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7040 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7041 // This only works for EQ and NE
7042 ICI->isEquality()) {
7043 // If Op1C some other power of two, convert:
7044 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7045 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7046 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7047 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7049 APInt KnownZeroMask(~KnownZero);
7050 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7051 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7052 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7053 // (X&4) == 2 --> false
7054 // (X&4) != 2 --> true
7055 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7056 Res = ConstantExpr::getZExt(Res, CI.getType());
7057 return ReplaceInstUsesWith(CI, Res);
7060 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7061 Value *In = ICI->getOperand(0);
7063 // Perform a logical shr by shiftamt.
7064 // Insert the shift to put the result in the low bit.
7065 In = InsertNewInstBefore(
7066 BinaryOperator::createLShr(In,
7067 ConstantInt::get(In->getType(), ShiftAmt),
7068 In->getName()+".lobit"), CI);
7071 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7072 Constant *One = ConstantInt::get(In->getType(), 1);
7073 In = BinaryOperator::createXor(In, One, "tmp");
7074 InsertNewInstBefore(cast<Instruction>(In), CI);
7077 if (CI.getType() == In->getType())
7078 return ReplaceInstUsesWith(CI, In);
7080 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7088 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7089 if (Instruction *I = commonIntCastTransforms(CI))
7092 Value *Src = CI.getOperand(0);
7094 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7095 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7096 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7097 // If we are just checking for a icmp eq of a single bit and zext'ing it
7098 // to an integer, then shift the bit to the appropriate place and then
7099 // cast to integer to avoid the comparison.
7100 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7101 const APInt &Op1CV = Op1C->getValue();
7103 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7104 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7105 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7106 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7107 Value *In = ICI->getOperand(0);
7108 Value *Sh = ConstantInt::get(In->getType(),
7109 In->getType()->getPrimitiveSizeInBits()-1);
7110 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7111 In->getName()+".lobit"),
7113 if (In->getType() != CI.getType())
7114 In = CastInst::createIntegerCast(In, CI.getType(),
7115 true/*SExt*/, "tmp", &CI);
7117 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7118 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7119 In->getName()+".not"), CI);
7121 return ReplaceInstUsesWith(CI, In);
7129 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7130 return commonCastTransforms(CI);
7133 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7134 return commonCastTransforms(CI);
7137 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7138 return commonCastTransforms(CI);
7141 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7142 return commonCastTransforms(CI);
7145 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7146 return commonCastTransforms(CI);
7149 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7150 return commonCastTransforms(CI);
7153 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7154 return commonPointerCastTransforms(CI);
7157 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7158 if (Instruction *I = commonCastTransforms(CI))
7161 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7162 if (!DestPointee->isSized()) return 0;
7164 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7167 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7168 m_ConstantInt(Cst)))) {
7169 // If the source and destination operands have the same type, see if this
7170 // is a single-index GEP.
7171 if (X->getType() == CI.getType()) {
7172 // Get the size of the pointee type.
7173 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7175 // Convert the constant to intptr type.
7176 APInt Offset = Cst->getValue();
7177 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7179 // If Offset is evenly divisible by Size, we can do this xform.
7180 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7181 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7182 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7185 // TODO: Could handle other cases, e.g. where add is indexing into field of
7187 } else if (CI.getOperand(0)->hasOneUse() &&
7188 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7189 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7190 // "inttoptr+GEP" instead of "add+intptr".
7192 // Get the size of the pointee type.
7193 uint64_t Size = TD->getABITypeSize(DestPointee);
7195 // Convert the constant to intptr type.
7196 APInt Offset = Cst->getValue();
7197 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7199 // If Offset is evenly divisible by Size, we can do this xform.
7200 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7201 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7203 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7205 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7211 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7212 // If the operands are integer typed then apply the integer transforms,
7213 // otherwise just apply the common ones.
7214 Value *Src = CI.getOperand(0);
7215 const Type *SrcTy = Src->getType();
7216 const Type *DestTy = CI.getType();
7218 if (SrcTy->isInteger() && DestTy->isInteger()) {
7219 if (Instruction *Result = commonIntCastTransforms(CI))
7221 } else if (isa<PointerType>(SrcTy)) {
7222 if (Instruction *I = commonPointerCastTransforms(CI))
7225 if (Instruction *Result = commonCastTransforms(CI))
7230 // Get rid of casts from one type to the same type. These are useless and can
7231 // be replaced by the operand.
7232 if (DestTy == Src->getType())
7233 return ReplaceInstUsesWith(CI, Src);
7235 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7236 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7237 const Type *DstElTy = DstPTy->getElementType();
7238 const Type *SrcElTy = SrcPTy->getElementType();
7240 // If we are casting a malloc or alloca to a pointer to a type of the same
7241 // size, rewrite the allocation instruction to allocate the "right" type.
7242 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7243 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7246 // If the source and destination are pointers, and this cast is equivalent
7247 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7248 // This can enhance SROA and other transforms that want type-safe pointers.
7249 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7250 unsigned NumZeros = 0;
7251 while (SrcElTy != DstElTy &&
7252 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7253 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7254 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7258 // If we found a path from the src to dest, create the getelementptr now.
7259 if (SrcElTy == DstElTy) {
7260 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7261 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7262 ((Instruction*) NULL));
7266 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7267 if (SVI->hasOneUse()) {
7268 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7269 // a bitconvert to a vector with the same # elts.
7270 if (isa<VectorType>(DestTy) &&
7271 cast<VectorType>(DestTy)->getNumElements() ==
7272 SVI->getType()->getNumElements()) {
7274 // If either of the operands is a cast from CI.getType(), then
7275 // evaluating the shuffle in the casted destination's type will allow
7276 // us to eliminate at least one cast.
7277 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7278 Tmp->getOperand(0)->getType() == DestTy) ||
7279 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7280 Tmp->getOperand(0)->getType() == DestTy)) {
7281 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7282 SVI->getOperand(0), DestTy, &CI);
7283 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7284 SVI->getOperand(1), DestTy, &CI);
7285 // Return a new shuffle vector. Use the same element ID's, as we
7286 // know the vector types match #elts.
7287 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7295 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7297 /// %D = select %cond, %C, %A
7299 /// %C = select %cond, %B, 0
7302 /// Assuming that the specified instruction is an operand to the select, return
7303 /// a bitmask indicating which operands of this instruction are foldable if they
7304 /// equal the other incoming value of the select.
7306 static unsigned GetSelectFoldableOperands(Instruction *I) {
7307 switch (I->getOpcode()) {
7308 case Instruction::Add:
7309 case Instruction::Mul:
7310 case Instruction::And:
7311 case Instruction::Or:
7312 case Instruction::Xor:
7313 return 3; // Can fold through either operand.
7314 case Instruction::Sub: // Can only fold on the amount subtracted.
7315 case Instruction::Shl: // Can only fold on the shift amount.
7316 case Instruction::LShr:
7317 case Instruction::AShr:
7320 return 0; // Cannot fold
7324 /// GetSelectFoldableConstant - For the same transformation as the previous
7325 /// function, return the identity constant that goes into the select.
7326 static Constant *GetSelectFoldableConstant(Instruction *I) {
7327 switch (I->getOpcode()) {
7328 default: assert(0 && "This cannot happen!"); abort();
7329 case Instruction::Add:
7330 case Instruction::Sub:
7331 case Instruction::Or:
7332 case Instruction::Xor:
7333 case Instruction::Shl:
7334 case Instruction::LShr:
7335 case Instruction::AShr:
7336 return Constant::getNullValue(I->getType());
7337 case Instruction::And:
7338 return Constant::getAllOnesValue(I->getType());
7339 case Instruction::Mul:
7340 return ConstantInt::get(I->getType(), 1);
7344 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7345 /// have the same opcode and only one use each. Try to simplify this.
7346 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7348 if (TI->getNumOperands() == 1) {
7349 // If this is a non-volatile load or a cast from the same type,
7352 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7355 return 0; // unknown unary op.
7358 // Fold this by inserting a select from the input values.
7359 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7360 FI->getOperand(0), SI.getName()+".v");
7361 InsertNewInstBefore(NewSI, SI);
7362 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7366 // Only handle binary operators here.
7367 if (!isa<BinaryOperator>(TI))
7370 // Figure out if the operations have any operands in common.
7371 Value *MatchOp, *OtherOpT, *OtherOpF;
7373 if (TI->getOperand(0) == FI->getOperand(0)) {
7374 MatchOp = TI->getOperand(0);
7375 OtherOpT = TI->getOperand(1);
7376 OtherOpF = FI->getOperand(1);
7377 MatchIsOpZero = true;
7378 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7379 MatchOp = TI->getOperand(1);
7380 OtherOpT = TI->getOperand(0);
7381 OtherOpF = FI->getOperand(0);
7382 MatchIsOpZero = false;
7383 } else if (!TI->isCommutative()) {
7385 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7386 MatchOp = TI->getOperand(0);
7387 OtherOpT = TI->getOperand(1);
7388 OtherOpF = FI->getOperand(0);
7389 MatchIsOpZero = true;
7390 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7391 MatchOp = TI->getOperand(1);
7392 OtherOpT = TI->getOperand(0);
7393 OtherOpF = FI->getOperand(1);
7394 MatchIsOpZero = true;
7399 // If we reach here, they do have operations in common.
7400 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7401 OtherOpF, SI.getName()+".v");
7402 InsertNewInstBefore(NewSI, SI);
7404 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7406 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7408 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7410 assert(0 && "Shouldn't get here");
7414 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7415 Value *CondVal = SI.getCondition();
7416 Value *TrueVal = SI.getTrueValue();
7417 Value *FalseVal = SI.getFalseValue();
7419 // select true, X, Y -> X
7420 // select false, X, Y -> Y
7421 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7422 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7424 // select C, X, X -> X
7425 if (TrueVal == FalseVal)
7426 return ReplaceInstUsesWith(SI, TrueVal);
7428 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7429 return ReplaceInstUsesWith(SI, FalseVal);
7430 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7431 return ReplaceInstUsesWith(SI, TrueVal);
7432 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7433 if (isa<Constant>(TrueVal))
7434 return ReplaceInstUsesWith(SI, TrueVal);
7436 return ReplaceInstUsesWith(SI, FalseVal);
7439 if (SI.getType() == Type::Int1Ty) {
7440 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7441 if (C->getZExtValue()) {
7442 // Change: A = select B, true, C --> A = or B, C
7443 return BinaryOperator::createOr(CondVal, FalseVal);
7445 // Change: A = select B, false, C --> A = and !B, C
7447 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7448 "not."+CondVal->getName()), SI);
7449 return BinaryOperator::createAnd(NotCond, FalseVal);
7451 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7452 if (C->getZExtValue() == false) {
7453 // Change: A = select B, C, false --> A = and B, C
7454 return BinaryOperator::createAnd(CondVal, TrueVal);
7456 // Change: A = select B, C, true --> A = or !B, C
7458 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7459 "not."+CondVal->getName()), SI);
7460 return BinaryOperator::createOr(NotCond, TrueVal);
7464 // select a, b, a -> a&b
7465 // select a, a, b -> a|b
7466 if (CondVal == TrueVal)
7467 return BinaryOperator::createOr(CondVal, FalseVal);
7468 else if (CondVal == FalseVal)
7469 return BinaryOperator::createAnd(CondVal, TrueVal);
7472 // Selecting between two integer constants?
7473 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7474 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7475 // select C, 1, 0 -> zext C to int
7476 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7477 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7478 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7479 // select C, 0, 1 -> zext !C to int
7481 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7482 "not."+CondVal->getName()), SI);
7483 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7486 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7488 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7490 // (x <s 0) ? -1 : 0 -> ashr x, 31
7491 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7492 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7493 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7494 // The comparison constant and the result are not neccessarily the
7495 // same width. Make an all-ones value by inserting a AShr.
7496 Value *X = IC->getOperand(0);
7497 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7498 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7499 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7501 InsertNewInstBefore(SRA, SI);
7503 // Finally, convert to the type of the select RHS. We figure out
7504 // if this requires a SExt, Trunc or BitCast based on the sizes.
7505 Instruction::CastOps opc = Instruction::BitCast;
7506 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7507 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7508 if (SRASize < SISize)
7509 opc = Instruction::SExt;
7510 else if (SRASize > SISize)
7511 opc = Instruction::Trunc;
7512 return CastInst::create(opc, SRA, SI.getType());
7517 // If one of the constants is zero (we know they can't both be) and we
7518 // have an icmp instruction with zero, and we have an 'and' with the
7519 // non-constant value, eliminate this whole mess. This corresponds to
7520 // cases like this: ((X & 27) ? 27 : 0)
7521 if (TrueValC->isZero() || FalseValC->isZero())
7522 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7523 cast<Constant>(IC->getOperand(1))->isNullValue())
7524 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7525 if (ICA->getOpcode() == Instruction::And &&
7526 isa<ConstantInt>(ICA->getOperand(1)) &&
7527 (ICA->getOperand(1) == TrueValC ||
7528 ICA->getOperand(1) == FalseValC) &&
7529 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7530 // Okay, now we know that everything is set up, we just don't
7531 // know whether we have a icmp_ne or icmp_eq and whether the
7532 // true or false val is the zero.
7533 bool ShouldNotVal = !TrueValC->isZero();
7534 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7537 V = InsertNewInstBefore(BinaryOperator::create(
7538 Instruction::Xor, V, ICA->getOperand(1)), SI);
7539 return ReplaceInstUsesWith(SI, V);
7544 // See if we are selecting two values based on a comparison of the two values.
7545 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7546 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7547 // Transform (X == Y) ? X : Y -> Y
7548 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7549 // This is not safe in general for floating point:
7550 // consider X== -0, Y== +0.
7551 // It becomes safe if either operand is a nonzero constant.
7552 ConstantFP *CFPt, *CFPf;
7553 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7554 !CFPt->getValueAPF().isZero()) ||
7555 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7556 !CFPf->getValueAPF().isZero()))
7557 return ReplaceInstUsesWith(SI, FalseVal);
7559 // Transform (X != Y) ? X : Y -> X
7560 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7561 return ReplaceInstUsesWith(SI, TrueVal);
7562 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7564 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7565 // Transform (X == Y) ? Y : X -> X
7566 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7567 // This is not safe in general for floating point:
7568 // consider X== -0, Y== +0.
7569 // It becomes safe if either operand is a nonzero constant.
7570 ConstantFP *CFPt, *CFPf;
7571 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7572 !CFPt->getValueAPF().isZero()) ||
7573 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7574 !CFPf->getValueAPF().isZero()))
7575 return ReplaceInstUsesWith(SI, FalseVal);
7577 // Transform (X != Y) ? Y : X -> Y
7578 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7579 return ReplaceInstUsesWith(SI, TrueVal);
7580 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7584 // See if we are selecting two values based on a comparison of the two values.
7585 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7586 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7587 // Transform (X == Y) ? X : Y -> Y
7588 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7589 return ReplaceInstUsesWith(SI, FalseVal);
7590 // Transform (X != Y) ? X : Y -> X
7591 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7592 return ReplaceInstUsesWith(SI, TrueVal);
7593 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7595 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7596 // Transform (X == Y) ? Y : X -> X
7597 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7598 return ReplaceInstUsesWith(SI, FalseVal);
7599 // Transform (X != Y) ? Y : X -> Y
7600 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7601 return ReplaceInstUsesWith(SI, TrueVal);
7602 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7606 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7607 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7608 if (TI->hasOneUse() && FI->hasOneUse()) {
7609 Instruction *AddOp = 0, *SubOp = 0;
7611 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7612 if (TI->getOpcode() == FI->getOpcode())
7613 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7616 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7617 // even legal for FP.
7618 if (TI->getOpcode() == Instruction::Sub &&
7619 FI->getOpcode() == Instruction::Add) {
7620 AddOp = FI; SubOp = TI;
7621 } else if (FI->getOpcode() == Instruction::Sub &&
7622 TI->getOpcode() == Instruction::Add) {
7623 AddOp = TI; SubOp = FI;
7627 Value *OtherAddOp = 0;
7628 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7629 OtherAddOp = AddOp->getOperand(1);
7630 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7631 OtherAddOp = AddOp->getOperand(0);
7635 // So at this point we know we have (Y -> OtherAddOp):
7636 // select C, (add X, Y), (sub X, Z)
7637 Value *NegVal; // Compute -Z
7638 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7639 NegVal = ConstantExpr::getNeg(C);
7641 NegVal = InsertNewInstBefore(
7642 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7645 Value *NewTrueOp = OtherAddOp;
7646 Value *NewFalseOp = NegVal;
7648 std::swap(NewTrueOp, NewFalseOp);
7649 Instruction *NewSel =
7650 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7652 NewSel = InsertNewInstBefore(NewSel, SI);
7653 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7658 // See if we can fold the select into one of our operands.
7659 if (SI.getType()->isInteger()) {
7660 // See the comment above GetSelectFoldableOperands for a description of the
7661 // transformation we are doing here.
7662 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7663 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7664 !isa<Constant>(FalseVal))
7665 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7666 unsigned OpToFold = 0;
7667 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7669 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7674 Constant *C = GetSelectFoldableConstant(TVI);
7675 Instruction *NewSel =
7676 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7677 InsertNewInstBefore(NewSel, SI);
7678 NewSel->takeName(TVI);
7679 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7680 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7682 assert(0 && "Unknown instruction!!");
7687 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7688 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7689 !isa<Constant>(TrueVal))
7690 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7691 unsigned OpToFold = 0;
7692 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7694 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7699 Constant *C = GetSelectFoldableConstant(FVI);
7700 Instruction *NewSel =
7701 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7702 InsertNewInstBefore(NewSel, SI);
7703 NewSel->takeName(FVI);
7704 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7705 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7707 assert(0 && "Unknown instruction!!");
7712 if (BinaryOperator::isNot(CondVal)) {
7713 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7714 SI.setOperand(1, FalseVal);
7715 SI.setOperand(2, TrueVal);
7722 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7723 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7724 /// and it is more than the alignment of the ultimate object, see if we can
7725 /// increase the alignment of the ultimate object, making this check succeed.
7726 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7727 unsigned PrefAlign = 0) {
7728 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7729 unsigned Align = GV->getAlignment();
7730 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7731 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7733 // If there is a large requested alignment and we can, bump up the alignment
7735 if (PrefAlign > Align && GV->hasInitializer()) {
7736 GV->setAlignment(PrefAlign);
7740 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7741 unsigned Align = AI->getAlignment();
7742 if (Align == 0 && TD) {
7743 if (isa<AllocaInst>(AI))
7744 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7745 else if (isa<MallocInst>(AI)) {
7746 // Malloc returns maximally aligned memory.
7747 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7750 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7753 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7757 // If there is a requested alignment and if this is an alloca, round up. We
7758 // don't do this for malloc, because some systems can't respect the request.
7759 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7760 AI->setAlignment(PrefAlign);
7764 } else if (isa<BitCastInst>(V) ||
7765 (isa<ConstantExpr>(V) &&
7766 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7767 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7769 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7770 // If all indexes are zero, it is just the alignment of the base pointer.
7771 bool AllZeroOperands = true;
7772 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7773 if (!isa<Constant>(GEPI->getOperand(i)) ||
7774 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7775 AllZeroOperands = false;
7779 if (AllZeroOperands) {
7780 // Treat this like a bitcast.
7781 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7784 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7785 if (BaseAlignment == 0) return 0;
7787 // Otherwise, if the base alignment is >= the alignment we expect for the
7788 // base pointer type, then we know that the resultant pointer is aligned at
7789 // least as much as its type requires.
7792 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7793 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7794 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7795 if (Align <= BaseAlignment) {
7796 const Type *GEPTy = GEPI->getType();
7797 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7798 Align = std::min(Align, (unsigned)
7799 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7808 /// visitCallInst - CallInst simplification. This mostly only handles folding
7809 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7810 /// the heavy lifting.
7812 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7813 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7814 if (!II) return visitCallSite(&CI);
7816 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7818 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7819 bool Changed = false;
7821 // memmove/cpy/set of zero bytes is a noop.
7822 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7823 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7825 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7826 if (CI->getZExtValue() == 1) {
7827 // Replace the instruction with just byte operations. We would
7828 // transform other cases to loads/stores, but we don't know if
7829 // alignment is sufficient.
7833 // If we have a memmove and the source operation is a constant global,
7834 // then the source and dest pointers can't alias, so we can change this
7835 // into a call to memcpy.
7836 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7837 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7838 if (GVSrc->isConstant()) {
7839 Module *M = CI.getParent()->getParent()->getParent();
7841 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7843 Name = "llvm.memcpy.i32";
7845 Name = "llvm.memcpy.i64";
7846 Constant *MemCpy = M->getOrInsertFunction(Name,
7847 CI.getCalledFunction()->getFunctionType());
7848 CI.setOperand(0, MemCpy);
7853 // If we can determine a pointer alignment that is bigger than currently
7854 // set, update the alignment.
7855 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7856 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7857 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7858 unsigned Align = std::min(Alignment1, Alignment2);
7859 if (MI->getAlignment()->getZExtValue() < Align) {
7860 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7864 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7866 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(CI.getOperand(3));
7868 unsigned Size = MemOpLength->getZExtValue();
7869 unsigned Align = cast<ConstantInt>(CI.getOperand(4))->getZExtValue();
7870 PointerType *NewPtrTy = NULL;
7871 // Destination pointer type is always i8 *
7872 // If Size is 8 then use Int64Ty
7873 // If Size is 4 then use Int32Ty
7874 // If Size is 2 then use Int16Ty
7875 // If Size is 1 then use Int8Ty
7876 if (Size && Size <=8 && !(Size&(Size-1)))
7877 NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7880 Value *Src = InsertCastBefore(Instruction::BitCast, CI.getOperand(2),
7882 Value *Dest = InsertCastBefore(Instruction::BitCast, CI.getOperand(1),
7884 Value *L = new LoadInst(Src, "tmp", false, Align, &CI);
7885 Value *NS = new StoreInst(L, Dest, false, Align, &CI);
7886 CI.replaceAllUsesWith(NS);
7888 return EraseInstFromFunction(CI);
7891 } else if (isa<MemSetInst>(MI)) {
7892 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7893 if (MI->getAlignment()->getZExtValue() < Alignment) {
7894 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7899 if (Changed) return II;
7901 switch (II->getIntrinsicID()) {
7903 case Intrinsic::ppc_altivec_lvx:
7904 case Intrinsic::ppc_altivec_lvxl:
7905 case Intrinsic::x86_sse_loadu_ps:
7906 case Intrinsic::x86_sse2_loadu_pd:
7907 case Intrinsic::x86_sse2_loadu_dq:
7908 // Turn PPC lvx -> load if the pointer is known aligned.
7909 // Turn X86 loadups -> load if the pointer is known aligned.
7910 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7912 InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7913 PointerType::getUnqual(II->getType()), CI);
7914 return new LoadInst(Ptr);
7917 case Intrinsic::ppc_altivec_stvx:
7918 case Intrinsic::ppc_altivec_stvxl:
7919 // Turn stvx -> store if the pointer is known aligned.
7920 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7921 const Type *OpPtrTy =
7922 PointerType::getUnqual(II->getOperand(1)->getType());
7923 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7925 return new StoreInst(II->getOperand(1), Ptr);
7928 case Intrinsic::x86_sse_storeu_ps:
7929 case Intrinsic::x86_sse2_storeu_pd:
7930 case Intrinsic::x86_sse2_storeu_dq:
7931 case Intrinsic::x86_sse2_storel_dq:
7932 // Turn X86 storeu -> store if the pointer is known aligned.
7933 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7934 const Type *OpPtrTy =
7935 PointerType::getUnqual(II->getOperand(2)->getType());
7936 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7938 return new StoreInst(II->getOperand(2), Ptr);
7942 case Intrinsic::x86_sse_cvttss2si: {
7943 // These intrinsics only demands the 0th element of its input vector. If
7944 // we can simplify the input based on that, do so now.
7946 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7948 II->setOperand(1, V);
7954 case Intrinsic::ppc_altivec_vperm:
7955 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7956 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7957 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7959 // Check that all of the elements are integer constants or undefs.
7960 bool AllEltsOk = true;
7961 for (unsigned i = 0; i != 16; ++i) {
7962 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7963 !isa<UndefValue>(Mask->getOperand(i))) {
7970 // Cast the input vectors to byte vectors.
7971 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7972 II->getOperand(1), Mask->getType(), CI);
7973 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7974 II->getOperand(2), Mask->getType(), CI);
7975 Value *Result = UndefValue::get(Op0->getType());
7977 // Only extract each element once.
7978 Value *ExtractedElts[32];
7979 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7981 for (unsigned i = 0; i != 16; ++i) {
7982 if (isa<UndefValue>(Mask->getOperand(i)))
7984 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7985 Idx &= 31; // Match the hardware behavior.
7987 if (ExtractedElts[Idx] == 0) {
7989 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7990 InsertNewInstBefore(Elt, CI);
7991 ExtractedElts[Idx] = Elt;
7994 // Insert this value into the result vector.
7995 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7996 InsertNewInstBefore(cast<Instruction>(Result), CI);
7998 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8003 case Intrinsic::stackrestore: {
8004 // If the save is right next to the restore, remove the restore. This can
8005 // happen when variable allocas are DCE'd.
8006 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8007 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8008 BasicBlock::iterator BI = SS;
8010 return EraseInstFromFunction(CI);
8014 // If the stack restore is in a return/unwind block and if there are no
8015 // allocas or calls between the restore and the return, nuke the restore.
8016 TerminatorInst *TI = II->getParent()->getTerminator();
8017 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
8018 BasicBlock::iterator BI = II;
8019 bool CannotRemove = false;
8020 for (++BI; &*BI != TI; ++BI) {
8021 if (isa<AllocaInst>(BI) ||
8022 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
8023 CannotRemove = true;
8028 return EraseInstFromFunction(CI);
8035 return visitCallSite(II);
8038 // InvokeInst simplification
8040 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8041 return visitCallSite(&II);
8044 // visitCallSite - Improvements for call and invoke instructions.
8046 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8047 bool Changed = false;
8049 // If the callee is a constexpr cast of a function, attempt to move the cast
8050 // to the arguments of the call/invoke.
8051 if (transformConstExprCastCall(CS)) return 0;
8053 Value *Callee = CS.getCalledValue();
8055 if (Function *CalleeF = dyn_cast<Function>(Callee))
8056 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8057 Instruction *OldCall = CS.getInstruction();
8058 // If the call and callee calling conventions don't match, this call must
8059 // be unreachable, as the call is undefined.
8060 new StoreInst(ConstantInt::getTrue(),
8061 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8063 if (!OldCall->use_empty())
8064 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8065 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8066 return EraseInstFromFunction(*OldCall);
8070 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8071 // This instruction is not reachable, just remove it. We insert a store to
8072 // undef so that we know that this code is not reachable, despite the fact
8073 // that we can't modify the CFG here.
8074 new StoreInst(ConstantInt::getTrue(),
8075 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8076 CS.getInstruction());
8078 if (!CS.getInstruction()->use_empty())
8079 CS.getInstruction()->
8080 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8082 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8083 // Don't break the CFG, insert a dummy cond branch.
8084 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8085 ConstantInt::getTrue(), II);
8087 return EraseInstFromFunction(*CS.getInstruction());
8090 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8091 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8092 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8093 return transformCallThroughTrampoline(CS);
8095 const PointerType *PTy = cast<PointerType>(Callee->getType());
8096 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8097 if (FTy->isVarArg()) {
8098 // See if we can optimize any arguments passed through the varargs area of
8100 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8101 E = CS.arg_end(); I != E; ++I)
8102 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8103 // If this cast does not effect the value passed through the varargs
8104 // area, we can eliminate the use of the cast.
8105 Value *Op = CI->getOperand(0);
8106 if (CI->isLosslessCast()) {
8113 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8114 // Inline asm calls cannot throw - mark them 'nounwind'.
8115 CS.setDoesNotThrow();
8119 return Changed ? CS.getInstruction() : 0;
8122 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8123 // attempt to move the cast to the arguments of the call/invoke.
8125 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8126 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8127 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8128 if (CE->getOpcode() != Instruction::BitCast ||
8129 !isa<Function>(CE->getOperand(0)))
8131 Function *Callee = cast<Function>(CE->getOperand(0));
8132 Instruction *Caller = CS.getInstruction();
8133 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8135 // Okay, this is a cast from a function to a different type. Unless doing so
8136 // would cause a type conversion of one of our arguments, change this call to
8137 // be a direct call with arguments casted to the appropriate types.
8139 const FunctionType *FT = Callee->getFunctionType();
8140 const Type *OldRetTy = Caller->getType();
8142 // Check to see if we are changing the return type...
8143 if (OldRetTy != FT->getReturnType()) {
8144 if (Callee->isDeclaration() && !Caller->use_empty() &&
8145 // Conversion is ok if changing from pointer to int of same size.
8146 !(isa<PointerType>(FT->getReturnType()) &&
8147 TD->getIntPtrType() == OldRetTy))
8148 return false; // Cannot transform this return value.
8150 if (!Caller->use_empty() &&
8151 // void -> non-void is handled specially
8152 FT->getReturnType() != Type::VoidTy &&
8153 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8154 return false; // Cannot transform this return value.
8156 if (CallerPAL && !Caller->use_empty()) {
8157 uint16_t RAttrs = CallerPAL->getParamAttrs(0);
8158 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8159 return false; // Attribute not compatible with transformed value.
8162 // If the callsite is an invoke instruction, and the return value is used by
8163 // a PHI node in a successor, we cannot change the return type of the call
8164 // because there is no place to put the cast instruction (without breaking
8165 // the critical edge). Bail out in this case.
8166 if (!Caller->use_empty())
8167 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8168 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8170 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8171 if (PN->getParent() == II->getNormalDest() ||
8172 PN->getParent() == II->getUnwindDest())
8176 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8177 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8179 CallSite::arg_iterator AI = CS.arg_begin();
8180 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8181 const Type *ParamTy = FT->getParamType(i);
8182 const Type *ActTy = (*AI)->getType();
8184 if (!CastInst::isCastable(ActTy, ParamTy))
8185 return false; // Cannot transform this parameter value.
8188 uint16_t PAttrs = CallerPAL->getParamAttrs(i + 1);
8189 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8190 return false; // Attribute not compatible with transformed value.
8193 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8194 // Some conversions are safe even if we do not have a body.
8195 // Either we can cast directly, or we can upconvert the argument
8196 bool isConvertible = ActTy == ParamTy ||
8197 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8198 (ParamTy->isInteger() && ActTy->isInteger() &&
8199 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8200 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8201 && c->getValue().isStrictlyPositive());
8202 if (Callee->isDeclaration() && !isConvertible) return false;
8205 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8206 Callee->isDeclaration())
8207 return false; // Do not delete arguments unless we have a function body...
8209 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8210 // In this case we have more arguments than the new function type, but we
8211 // won't be dropping them. Check that these extra arguments have attributes
8212 // that are compatible with being a vararg call argument.
8213 for (unsigned i = CallerPAL->size(); i; --i) {
8214 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8216 uint16_t PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8217 if (PAttrs & ParamAttr::VarArgsIncompatible)
8221 // Okay, we decided that this is a safe thing to do: go ahead and start
8222 // inserting cast instructions as necessary...
8223 std::vector<Value*> Args;
8224 Args.reserve(NumActualArgs);
8225 ParamAttrsVector attrVec;
8226 attrVec.reserve(NumCommonArgs);
8228 // Get any return attributes.
8229 uint16_t RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) : 0;
8231 // If the return value is not being used, the type may not be compatible
8232 // with the existing attributes. Wipe out any problematic attributes.
8233 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8235 // Add the new return attributes.
8237 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8239 AI = CS.arg_begin();
8240 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8241 const Type *ParamTy = FT->getParamType(i);
8242 if ((*AI)->getType() == ParamTy) {
8243 Args.push_back(*AI);
8245 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8246 false, ParamTy, false);
8247 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8248 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8251 // Add any parameter attributes.
8252 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8254 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8257 // If the function takes more arguments than the call was taking, add them
8259 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8260 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8262 // If we are removing arguments to the function, emit an obnoxious warning...
8263 if (FT->getNumParams() < NumActualArgs)
8264 if (!FT->isVarArg()) {
8265 cerr << "WARNING: While resolving call to function '"
8266 << Callee->getName() << "' arguments were dropped!\n";
8268 // Add all of the arguments in their promoted form to the arg list...
8269 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8270 const Type *PTy = getPromotedType((*AI)->getType());
8271 if (PTy != (*AI)->getType()) {
8272 // Must promote to pass through va_arg area!
8273 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8275 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8276 InsertNewInstBefore(Cast, *Caller);
8277 Args.push_back(Cast);
8279 Args.push_back(*AI);
8282 // Add any parameter attributes.
8283 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8285 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8289 if (FT->getReturnType() == Type::VoidTy)
8290 Caller->setName(""); // Void type should not have a name.
8292 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8295 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8296 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8297 Args.begin(), Args.end(), Caller->getName(), Caller);
8298 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8299 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8301 NC = new CallInst(Callee, Args.begin(), Args.end(),
8302 Caller->getName(), Caller);
8303 CallInst *CI = cast<CallInst>(Caller);
8304 if (CI->isTailCall())
8305 cast<CallInst>(NC)->setTailCall();
8306 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8307 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8310 // Insert a cast of the return type as necessary.
8312 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8313 if (NV->getType() != Type::VoidTy) {
8314 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8316 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8318 // If this is an invoke instruction, we should insert it after the first
8319 // non-phi, instruction in the normal successor block.
8320 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8321 BasicBlock::iterator I = II->getNormalDest()->begin();
8322 while (isa<PHINode>(I)) ++I;
8323 InsertNewInstBefore(NC, *I);
8325 // Otherwise, it's a call, just insert cast right after the call instr
8326 InsertNewInstBefore(NC, *Caller);
8328 AddUsersToWorkList(*Caller);
8330 NV = UndefValue::get(Caller->getType());
8334 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8335 Caller->replaceAllUsesWith(NV);
8336 Caller->eraseFromParent();
8337 RemoveFromWorkList(Caller);
8341 // transformCallThroughTrampoline - Turn a call to a function created by the
8342 // init_trampoline intrinsic into a direct call to the underlying function.
8344 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8345 Value *Callee = CS.getCalledValue();
8346 const PointerType *PTy = cast<PointerType>(Callee->getType());
8347 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8349 IntrinsicInst *Tramp =
8350 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8353 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8354 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8355 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8357 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8358 unsigned NestIdx = 1;
8359 const Type *NestTy = 0;
8360 uint16_t NestAttr = 0;
8362 // Look for a parameter marked with the 'nest' attribute.
8363 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8364 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8365 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8366 // Record the parameter type and any other attributes.
8368 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8373 Instruction *Caller = CS.getInstruction();
8374 std::vector<Value*> NewArgs;
8375 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8377 // Insert the nest argument into the call argument list, which may
8378 // mean appending it.
8381 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8383 if (Idx == NestIdx) {
8384 // Add the chain argument.
8385 Value *NestVal = Tramp->getOperand(3);
8386 if (NestVal->getType() != NestTy)
8387 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8388 NewArgs.push_back(NestVal);
8394 // Add the original argument.
8395 NewArgs.push_back(*I);
8401 // The trampoline may have been bitcast to a bogus type (FTy).
8402 // Handle this by synthesizing a new function type, equal to FTy
8403 // with the chain parameter inserted. Likewise for attributes.
8405 const ParamAttrsList *Attrs = CS.getParamAttrs();
8406 std::vector<const Type*> NewTypes;
8407 ParamAttrsVector NewAttrs;
8408 NewTypes.reserve(FTy->getNumParams()+1);
8410 // Add any function result attributes.
8411 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8413 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8415 // Insert the chain's type into the list of parameter types, which may
8416 // mean appending it. Likewise for the chain's attributes.
8419 FunctionType::param_iterator I = FTy->param_begin(),
8420 E = FTy->param_end();
8423 if (Idx == NestIdx) {
8424 // Add the chain's type and attributes.
8425 NewTypes.push_back(NestTy);
8426 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8432 // Add the original type and attributes.
8433 NewTypes.push_back(*I);
8434 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8437 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8443 // Replace the trampoline call with a direct call. Let the generic
8444 // code sort out any function type mismatches.
8445 FunctionType *NewFTy =
8446 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8447 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8448 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8449 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8451 Instruction *NewCaller;
8452 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8453 NewCaller = new InvokeInst(NewCallee,
8454 II->getNormalDest(), II->getUnwindDest(),
8455 NewArgs.begin(), NewArgs.end(),
8456 Caller->getName(), Caller);
8457 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8458 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8460 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8461 Caller->getName(), Caller);
8462 if (cast<CallInst>(Caller)->isTailCall())
8463 cast<CallInst>(NewCaller)->setTailCall();
8464 cast<CallInst>(NewCaller)->
8465 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8466 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8468 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8469 Caller->replaceAllUsesWith(NewCaller);
8470 Caller->eraseFromParent();
8471 RemoveFromWorkList(Caller);
8476 // Replace the trampoline call with a direct call. Since there is no 'nest'
8477 // parameter, there is no need to adjust the argument list. Let the generic
8478 // code sort out any function type mismatches.
8479 Constant *NewCallee =
8480 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8481 CS.setCalledFunction(NewCallee);
8482 return CS.getInstruction();
8485 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8486 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8487 /// and a single binop.
8488 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8489 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8490 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8491 isa<CmpInst>(FirstInst));
8492 unsigned Opc = FirstInst->getOpcode();
8493 Value *LHSVal = FirstInst->getOperand(0);
8494 Value *RHSVal = FirstInst->getOperand(1);
8496 const Type *LHSType = LHSVal->getType();
8497 const Type *RHSType = RHSVal->getType();
8499 // Scan to see if all operands are the same opcode, all have one use, and all
8500 // kill their operands (i.e. the operands have one use).
8501 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8502 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8503 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8504 // Verify type of the LHS matches so we don't fold cmp's of different
8505 // types or GEP's with different index types.
8506 I->getOperand(0)->getType() != LHSType ||
8507 I->getOperand(1)->getType() != RHSType)
8510 // If they are CmpInst instructions, check their predicates
8511 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8512 if (cast<CmpInst>(I)->getPredicate() !=
8513 cast<CmpInst>(FirstInst)->getPredicate())
8516 // Keep track of which operand needs a phi node.
8517 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8518 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8521 // Otherwise, this is safe to transform, determine if it is profitable.
8523 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8524 // Indexes are often folded into load/store instructions, so we don't want to
8525 // hide them behind a phi.
8526 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8529 Value *InLHS = FirstInst->getOperand(0);
8530 Value *InRHS = FirstInst->getOperand(1);
8531 PHINode *NewLHS = 0, *NewRHS = 0;
8533 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8534 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8535 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8536 InsertNewInstBefore(NewLHS, PN);
8541 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8542 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8543 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8544 InsertNewInstBefore(NewRHS, PN);
8548 // Add all operands to the new PHIs.
8549 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8551 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8552 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8555 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8556 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8560 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8561 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8562 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8563 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8566 assert(isa<GetElementPtrInst>(FirstInst));
8567 return new GetElementPtrInst(LHSVal, RHSVal);
8571 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8572 /// of the block that defines it. This means that it must be obvious the value
8573 /// of the load is not changed from the point of the load to the end of the
8576 /// Finally, it is safe, but not profitable, to sink a load targetting a
8577 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8579 static bool isSafeToSinkLoad(LoadInst *L) {
8580 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8582 for (++BBI; BBI != E; ++BBI)
8583 if (BBI->mayWriteToMemory())
8586 // Check for non-address taken alloca. If not address-taken already, it isn't
8587 // profitable to do this xform.
8588 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8589 bool isAddressTaken = false;
8590 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8592 if (isa<LoadInst>(UI)) continue;
8593 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8594 // If storing TO the alloca, then the address isn't taken.
8595 if (SI->getOperand(1) == AI) continue;
8597 isAddressTaken = true;
8601 if (!isAddressTaken)
8609 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8610 // operator and they all are only used by the PHI, PHI together their
8611 // inputs, and do the operation once, to the result of the PHI.
8612 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8613 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8615 // Scan the instruction, looking for input operations that can be folded away.
8616 // If all input operands to the phi are the same instruction (e.g. a cast from
8617 // the same type or "+42") we can pull the operation through the PHI, reducing
8618 // code size and simplifying code.
8619 Constant *ConstantOp = 0;
8620 const Type *CastSrcTy = 0;
8621 bool isVolatile = false;
8622 if (isa<CastInst>(FirstInst)) {
8623 CastSrcTy = FirstInst->getOperand(0)->getType();
8624 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8625 // Can fold binop, compare or shift here if the RHS is a constant,
8626 // otherwise call FoldPHIArgBinOpIntoPHI.
8627 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8628 if (ConstantOp == 0)
8629 return FoldPHIArgBinOpIntoPHI(PN);
8630 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8631 isVolatile = LI->isVolatile();
8632 // We can't sink the load if the loaded value could be modified between the
8633 // load and the PHI.
8634 if (LI->getParent() != PN.getIncomingBlock(0) ||
8635 !isSafeToSinkLoad(LI))
8637 } else if (isa<GetElementPtrInst>(FirstInst)) {
8638 if (FirstInst->getNumOperands() == 2)
8639 return FoldPHIArgBinOpIntoPHI(PN);
8640 // Can't handle general GEPs yet.
8643 return 0; // Cannot fold this operation.
8646 // Check to see if all arguments are the same operation.
8647 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8648 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8649 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8650 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8653 if (I->getOperand(0)->getType() != CastSrcTy)
8654 return 0; // Cast operation must match.
8655 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8656 // We can't sink the load if the loaded value could be modified between
8657 // the load and the PHI.
8658 if (LI->isVolatile() != isVolatile ||
8659 LI->getParent() != PN.getIncomingBlock(i) ||
8660 !isSafeToSinkLoad(LI))
8662 } else if (I->getOperand(1) != ConstantOp) {
8667 // Okay, they are all the same operation. Create a new PHI node of the
8668 // correct type, and PHI together all of the LHS's of the instructions.
8669 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8670 PN.getName()+".in");
8671 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8673 Value *InVal = FirstInst->getOperand(0);
8674 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8676 // Add all operands to the new PHI.
8677 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8678 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8679 if (NewInVal != InVal)
8681 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8686 // The new PHI unions all of the same values together. This is really
8687 // common, so we handle it intelligently here for compile-time speed.
8691 InsertNewInstBefore(NewPN, PN);
8695 // Insert and return the new operation.
8696 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8697 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8698 else if (isa<LoadInst>(FirstInst))
8699 return new LoadInst(PhiVal, "", isVolatile);
8700 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8701 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8702 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8703 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8704 PhiVal, ConstantOp);
8706 assert(0 && "Unknown operation");
8710 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8712 static bool DeadPHICycle(PHINode *PN,
8713 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8714 if (PN->use_empty()) return true;
8715 if (!PN->hasOneUse()) return false;
8717 // Remember this node, and if we find the cycle, return.
8718 if (!PotentiallyDeadPHIs.insert(PN))
8721 // Don't scan crazily complex things.
8722 if (PotentiallyDeadPHIs.size() == 16)
8725 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8726 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8731 /// PHIsEqualValue - Return true if this phi node is always equal to
8732 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8733 /// z = some value; x = phi (y, z); y = phi (x, z)
8734 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8735 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8736 // See if we already saw this PHI node.
8737 if (!ValueEqualPHIs.insert(PN))
8740 // Don't scan crazily complex things.
8741 if (ValueEqualPHIs.size() == 16)
8744 // Scan the operands to see if they are either phi nodes or are equal to
8746 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8747 Value *Op = PN->getIncomingValue(i);
8748 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8749 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8751 } else if (Op != NonPhiInVal)
8759 // PHINode simplification
8761 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8762 // If LCSSA is around, don't mess with Phi nodes
8763 if (MustPreserveLCSSA) return 0;
8765 if (Value *V = PN.hasConstantValue())
8766 return ReplaceInstUsesWith(PN, V);
8768 // If all PHI operands are the same operation, pull them through the PHI,
8769 // reducing code size.
8770 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8771 PN.getIncomingValue(0)->hasOneUse())
8772 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8775 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8776 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8777 // PHI)... break the cycle.
8778 if (PN.hasOneUse()) {
8779 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8780 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8781 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8782 PotentiallyDeadPHIs.insert(&PN);
8783 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8784 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8787 // If this phi has a single use, and if that use just computes a value for
8788 // the next iteration of a loop, delete the phi. This occurs with unused
8789 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8790 // common case here is good because the only other things that catch this
8791 // are induction variable analysis (sometimes) and ADCE, which is only run
8793 if (PHIUser->hasOneUse() &&
8794 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8795 PHIUser->use_back() == &PN) {
8796 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8800 // We sometimes end up with phi cycles that non-obviously end up being the
8801 // same value, for example:
8802 // z = some value; x = phi (y, z); y = phi (x, z)
8803 // where the phi nodes don't necessarily need to be in the same block. Do a
8804 // quick check to see if the PHI node only contains a single non-phi value, if
8805 // so, scan to see if the phi cycle is actually equal to that value.
8807 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
8808 // Scan for the first non-phi operand.
8809 while (InValNo != NumOperandVals &&
8810 isa<PHINode>(PN.getIncomingValue(InValNo)))
8813 if (InValNo != NumOperandVals) {
8814 Value *NonPhiInVal = PN.getOperand(InValNo);
8816 // Scan the rest of the operands to see if there are any conflicts, if so
8817 // there is no need to recursively scan other phis.
8818 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
8819 Value *OpVal = PN.getIncomingValue(InValNo);
8820 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
8824 // If we scanned over all operands, then we have one unique value plus
8825 // phi values. Scan PHI nodes to see if they all merge in each other or
8827 if (InValNo == NumOperandVals) {
8828 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
8829 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
8830 return ReplaceInstUsesWith(PN, NonPhiInVal);
8837 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8838 Instruction *InsertPoint,
8840 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8841 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8842 // We must cast correctly to the pointer type. Ensure that we
8843 // sign extend the integer value if it is smaller as this is
8844 // used for address computation.
8845 Instruction::CastOps opcode =
8846 (VTySize < PtrSize ? Instruction::SExt :
8847 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8848 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8852 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8853 Value *PtrOp = GEP.getOperand(0);
8854 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8855 // If so, eliminate the noop.
8856 if (GEP.getNumOperands() == 1)
8857 return ReplaceInstUsesWith(GEP, PtrOp);
8859 if (isa<UndefValue>(GEP.getOperand(0)))
8860 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8862 bool HasZeroPointerIndex = false;
8863 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8864 HasZeroPointerIndex = C->isNullValue();
8866 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8867 return ReplaceInstUsesWith(GEP, PtrOp);
8869 // Eliminate unneeded casts for indices.
8870 bool MadeChange = false;
8872 gep_type_iterator GTI = gep_type_begin(GEP);
8873 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8874 if (isa<SequentialType>(*GTI)) {
8875 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8876 if (CI->getOpcode() == Instruction::ZExt ||
8877 CI->getOpcode() == Instruction::SExt) {
8878 const Type *SrcTy = CI->getOperand(0)->getType();
8879 // We can eliminate a cast from i32 to i64 iff the target
8880 // is a 32-bit pointer target.
8881 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8883 GEP.setOperand(i, CI->getOperand(0));
8887 // If we are using a wider index than needed for this platform, shrink it
8888 // to what we need. If the incoming value needs a cast instruction,
8889 // insert it. This explicit cast can make subsequent optimizations more
8891 Value *Op = GEP.getOperand(i);
8892 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
8893 if (Constant *C = dyn_cast<Constant>(Op)) {
8894 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8897 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8899 GEP.setOperand(i, Op);
8904 if (MadeChange) return &GEP;
8906 // If this GEP instruction doesn't move the pointer, and if the input operand
8907 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8908 // real input to the dest type.
8909 if (GEP.hasAllZeroIndices()) {
8910 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
8911 // If the bitcast is of an allocation, and the allocation will be
8912 // converted to match the type of the cast, don't touch this.
8913 if (isa<AllocationInst>(BCI->getOperand(0))) {
8914 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
8915 if (Instruction *I = visitBitCast(*BCI)) {
8918 BCI->getParent()->getInstList().insert(BCI, I);
8919 ReplaceInstUsesWith(*BCI, I);
8924 return new BitCastInst(BCI->getOperand(0), GEP.getType());
8928 // Combine Indices - If the source pointer to this getelementptr instruction
8929 // is a getelementptr instruction, combine the indices of the two
8930 // getelementptr instructions into a single instruction.
8932 SmallVector<Value*, 8> SrcGEPOperands;
8933 if (User *Src = dyn_castGetElementPtr(PtrOp))
8934 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8936 if (!SrcGEPOperands.empty()) {
8937 // Note that if our source is a gep chain itself that we wait for that
8938 // chain to be resolved before we perform this transformation. This
8939 // avoids us creating a TON of code in some cases.
8941 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8942 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8943 return 0; // Wait until our source is folded to completion.
8945 SmallVector<Value*, 8> Indices;
8947 // Find out whether the last index in the source GEP is a sequential idx.
8948 bool EndsWithSequential = false;
8949 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8950 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8951 EndsWithSequential = !isa<StructType>(*I);
8953 // Can we combine the two pointer arithmetics offsets?
8954 if (EndsWithSequential) {
8955 // Replace: gep (gep %P, long B), long A, ...
8956 // With: T = long A+B; gep %P, T, ...
8958 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8959 if (SO1 == Constant::getNullValue(SO1->getType())) {
8961 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8964 // If they aren't the same type, convert both to an integer of the
8965 // target's pointer size.
8966 if (SO1->getType() != GO1->getType()) {
8967 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8968 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8969 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8970 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8972 unsigned PS = TD->getPointerSizeInBits();
8973 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
8974 // Convert GO1 to SO1's type.
8975 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8977 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
8978 // Convert SO1 to GO1's type.
8979 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8981 const Type *PT = TD->getIntPtrType();
8982 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8983 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8987 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8988 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8990 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8991 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8995 // Recycle the GEP we already have if possible.
8996 if (SrcGEPOperands.size() == 2) {
8997 GEP.setOperand(0, SrcGEPOperands[0]);
8998 GEP.setOperand(1, Sum);
9001 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9002 SrcGEPOperands.end()-1);
9003 Indices.push_back(Sum);
9004 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9006 } else if (isa<Constant>(*GEP.idx_begin()) &&
9007 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9008 SrcGEPOperands.size() != 1) {
9009 // Otherwise we can do the fold if the first index of the GEP is a zero
9010 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9011 SrcGEPOperands.end());
9012 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9015 if (!Indices.empty())
9016 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9017 Indices.end(), GEP.getName());
9019 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9020 // GEP of global variable. If all of the indices for this GEP are
9021 // constants, we can promote this to a constexpr instead of an instruction.
9023 // Scan for nonconstants...
9024 SmallVector<Constant*, 8> Indices;
9025 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9026 for (; I != E && isa<Constant>(*I); ++I)
9027 Indices.push_back(cast<Constant>(*I));
9029 if (I == E) { // If they are all constants...
9030 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9031 &Indices[0],Indices.size());
9033 // Replace all uses of the GEP with the new constexpr...
9034 return ReplaceInstUsesWith(GEP, CE);
9036 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9037 if (!isa<PointerType>(X->getType())) {
9038 // Not interesting. Source pointer must be a cast from pointer.
9039 } else if (HasZeroPointerIndex) {
9040 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9041 // into : GEP [10 x i8]* X, i32 0, ...
9043 // This occurs when the program declares an array extern like "int X[];"
9045 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9046 const PointerType *XTy = cast<PointerType>(X->getType());
9047 if (const ArrayType *XATy =
9048 dyn_cast<ArrayType>(XTy->getElementType()))
9049 if (const ArrayType *CATy =
9050 dyn_cast<ArrayType>(CPTy->getElementType()))
9051 if (CATy->getElementType() == XATy->getElementType()) {
9052 // At this point, we know that the cast source type is a pointer
9053 // to an array of the same type as the destination pointer
9054 // array. Because the array type is never stepped over (there
9055 // is a leading zero) we can fold the cast into this GEP.
9056 GEP.setOperand(0, X);
9059 } else if (GEP.getNumOperands() == 2) {
9060 // Transform things like:
9061 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9062 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9063 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9064 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9065 if (isa<ArrayType>(SrcElTy) &&
9066 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9067 TD->getABITypeSize(ResElTy)) {
9069 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9070 Idx[1] = GEP.getOperand(1);
9071 Value *V = InsertNewInstBefore(
9072 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9073 // V and GEP are both pointer types --> BitCast
9074 return new BitCastInst(V, GEP.getType());
9077 // Transform things like:
9078 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9079 // (where tmp = 8*tmp2) into:
9080 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9082 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9083 uint64_t ArrayEltSize =
9084 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9086 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9087 // allow either a mul, shift, or constant here.
9089 ConstantInt *Scale = 0;
9090 if (ArrayEltSize == 1) {
9091 NewIdx = GEP.getOperand(1);
9092 Scale = ConstantInt::get(NewIdx->getType(), 1);
9093 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9094 NewIdx = ConstantInt::get(CI->getType(), 1);
9096 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9097 if (Inst->getOpcode() == Instruction::Shl &&
9098 isa<ConstantInt>(Inst->getOperand(1))) {
9099 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9100 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9101 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9102 NewIdx = Inst->getOperand(0);
9103 } else if (Inst->getOpcode() == Instruction::Mul &&
9104 isa<ConstantInt>(Inst->getOperand(1))) {
9105 Scale = cast<ConstantInt>(Inst->getOperand(1));
9106 NewIdx = Inst->getOperand(0);
9110 // If the index will be to exactly the right offset with the scale taken
9111 // out, perform the transformation. Note, we don't know whether Scale is
9112 // signed or not. We'll use unsigned version of division/modulo
9113 // operation after making sure Scale doesn't have the sign bit set.
9114 if (Scale && Scale->getSExtValue() >= 0LL &&
9115 Scale->getZExtValue() % ArrayEltSize == 0) {
9116 Scale = ConstantInt::get(Scale->getType(),
9117 Scale->getZExtValue() / ArrayEltSize);
9118 if (Scale->getZExtValue() != 1) {
9119 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9121 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9122 NewIdx = InsertNewInstBefore(Sc, GEP);
9125 // Insert the new GEP instruction.
9127 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9129 Instruction *NewGEP =
9130 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9131 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9132 // The NewGEP must be pointer typed, so must the old one -> BitCast
9133 return new BitCastInst(NewGEP, GEP.getType());
9142 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9143 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9144 if (AI.isArrayAllocation()) // Check C != 1
9145 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9147 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9148 AllocationInst *New = 0;
9150 // Create and insert the replacement instruction...
9151 if (isa<MallocInst>(AI))
9152 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9154 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9155 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9158 InsertNewInstBefore(New, AI);
9160 // Scan to the end of the allocation instructions, to skip over a block of
9161 // allocas if possible...
9163 BasicBlock::iterator It = New;
9164 while (isa<AllocationInst>(*It)) ++It;
9166 // Now that I is pointing to the first non-allocation-inst in the block,
9167 // insert our getelementptr instruction...
9169 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9173 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9174 New->getName()+".sub", It);
9176 // Now make everything use the getelementptr instead of the original
9178 return ReplaceInstUsesWith(AI, V);
9179 } else if (isa<UndefValue>(AI.getArraySize())) {
9180 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9183 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9184 // Note that we only do this for alloca's, because malloc should allocate and
9185 // return a unique pointer, even for a zero byte allocation.
9186 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9187 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9188 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9193 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9194 Value *Op = FI.getOperand(0);
9196 // free undef -> unreachable.
9197 if (isa<UndefValue>(Op)) {
9198 // Insert a new store to null because we cannot modify the CFG here.
9199 new StoreInst(ConstantInt::getTrue(),
9200 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9201 return EraseInstFromFunction(FI);
9204 // If we have 'free null' delete the instruction. This can happen in stl code
9205 // when lots of inlining happens.
9206 if (isa<ConstantPointerNull>(Op))
9207 return EraseInstFromFunction(FI);
9209 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9210 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9211 FI.setOperand(0, CI->getOperand(0));
9215 // Change free (gep X, 0,0,0,0) into free(X)
9216 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9217 if (GEPI->hasAllZeroIndices()) {
9218 AddToWorkList(GEPI);
9219 FI.setOperand(0, GEPI->getOperand(0));
9224 // Change free(malloc) into nothing, if the malloc has a single use.
9225 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9226 if (MI->hasOneUse()) {
9227 EraseInstFromFunction(FI);
9228 return EraseInstFromFunction(*MI);
9235 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9236 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9237 const TargetData *TD) {
9238 User *CI = cast<User>(LI.getOperand(0));
9239 Value *CastOp = CI->getOperand(0);
9241 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9242 // Instead of loading constant c string, use corresponding integer value
9243 // directly if string length is small enough.
9244 const std::string &Str = CE->getOperand(0)->getStringValue();
9246 unsigned len = Str.length();
9247 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9248 unsigned numBits = Ty->getPrimitiveSizeInBits();
9249 // Replace LI with immediate integer store.
9250 if ((numBits >> 3) == len + 1) {
9251 APInt StrVal(numBits, 0);
9252 APInt SingleChar(numBits, 0);
9253 if (TD->isLittleEndian()) {
9254 for (signed i = len-1; i >= 0; i--) {
9255 SingleChar = (uint64_t) Str[i];
9256 StrVal = (StrVal << 8) | SingleChar;
9259 for (unsigned i = 0; i < len; i++) {
9260 SingleChar = (uint64_t) Str[i];
9261 StrVal = (StrVal << 8) | SingleChar;
9263 // Append NULL at the end.
9265 StrVal = (StrVal << 8) | SingleChar;
9267 Value *NL = ConstantInt::get(StrVal);
9268 return IC.ReplaceInstUsesWith(LI, NL);
9273 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9274 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9275 const Type *SrcPTy = SrcTy->getElementType();
9277 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9278 isa<VectorType>(DestPTy)) {
9279 // If the source is an array, the code below will not succeed. Check to
9280 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9282 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9283 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9284 if (ASrcTy->getNumElements() != 0) {
9286 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9287 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9288 SrcTy = cast<PointerType>(CastOp->getType());
9289 SrcPTy = SrcTy->getElementType();
9292 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9293 isa<VectorType>(SrcPTy)) &&
9294 // Do not allow turning this into a load of an integer, which is then
9295 // casted to a pointer, this pessimizes pointer analysis a lot.
9296 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9297 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9298 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9300 // Okay, we are casting from one integer or pointer type to another of
9301 // the same size. Instead of casting the pointer before the load, cast
9302 // the result of the loaded value.
9303 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9305 LI.isVolatile()),LI);
9306 // Now cast the result of the load.
9307 return new BitCastInst(NewLoad, LI.getType());
9314 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9315 /// from this value cannot trap. If it is not obviously safe to load from the
9316 /// specified pointer, we do a quick local scan of the basic block containing
9317 /// ScanFrom, to determine if the address is already accessed.
9318 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9319 // If it is an alloca it is always safe to load from.
9320 if (isa<AllocaInst>(V)) return true;
9322 // If it is a global variable it is mostly safe to load from.
9323 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9324 // Don't try to evaluate aliases. External weak GV can be null.
9325 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9327 // Otherwise, be a little bit agressive by scanning the local block where we
9328 // want to check to see if the pointer is already being loaded or stored
9329 // from/to. If so, the previous load or store would have already trapped,
9330 // so there is no harm doing an extra load (also, CSE will later eliminate
9331 // the load entirely).
9332 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9337 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9338 if (LI->getOperand(0) == V) return true;
9339 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9340 if (SI->getOperand(1) == V) return true;
9346 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9347 /// until we find the underlying object a pointer is referring to or something
9348 /// we don't understand. Note that the returned pointer may be offset from the
9349 /// input, because we ignore GEP indices.
9350 static Value *GetUnderlyingObject(Value *Ptr) {
9352 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9353 if (CE->getOpcode() == Instruction::BitCast ||
9354 CE->getOpcode() == Instruction::GetElementPtr)
9355 Ptr = CE->getOperand(0);
9358 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9359 Ptr = BCI->getOperand(0);
9360 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9361 Ptr = GEP->getOperand(0);
9368 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9369 Value *Op = LI.getOperand(0);
9371 // Attempt to improve the alignment.
9372 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9373 if (KnownAlign > LI.getAlignment())
9374 LI.setAlignment(KnownAlign);
9376 // load (cast X) --> cast (load X) iff safe
9377 if (isa<CastInst>(Op))
9378 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9381 // None of the following transforms are legal for volatile loads.
9382 if (LI.isVolatile()) return 0;
9384 if (&LI.getParent()->front() != &LI) {
9385 BasicBlock::iterator BBI = &LI; --BBI;
9386 // If the instruction immediately before this is a store to the same
9387 // address, do a simple form of store->load forwarding.
9388 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9389 if (SI->getOperand(1) == LI.getOperand(0))
9390 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9391 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9392 if (LIB->getOperand(0) == LI.getOperand(0))
9393 return ReplaceInstUsesWith(LI, LIB);
9396 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9397 const Value *GEPI0 = GEPI->getOperand(0);
9398 // TODO: Consider a target hook for valid address spaces for this xform.
9399 if (isa<ConstantPointerNull>(GEPI0) &&
9400 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9401 // Insert a new store to null instruction before the load to indicate
9402 // that this code is not reachable. We do this instead of inserting
9403 // an unreachable instruction directly because we cannot modify the
9405 new StoreInst(UndefValue::get(LI.getType()),
9406 Constant::getNullValue(Op->getType()), &LI);
9407 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9411 if (Constant *C = dyn_cast<Constant>(Op)) {
9412 // load null/undef -> undef
9413 // TODO: Consider a target hook for valid address spaces for this xform.
9414 if (isa<UndefValue>(C) || (C->isNullValue() &&
9415 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9416 // Insert a new store to null instruction before the load to indicate that
9417 // this code is not reachable. We do this instead of inserting an
9418 // unreachable instruction directly because we cannot modify the CFG.
9419 new StoreInst(UndefValue::get(LI.getType()),
9420 Constant::getNullValue(Op->getType()), &LI);
9421 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9424 // Instcombine load (constant global) into the value loaded.
9425 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9426 if (GV->isConstant() && !GV->isDeclaration())
9427 return ReplaceInstUsesWith(LI, GV->getInitializer());
9429 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9430 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9431 if (CE->getOpcode() == Instruction::GetElementPtr) {
9432 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9433 if (GV->isConstant() && !GV->isDeclaration())
9435 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9436 return ReplaceInstUsesWith(LI, V);
9437 if (CE->getOperand(0)->isNullValue()) {
9438 // Insert a new store to null instruction before the load to indicate
9439 // that this code is not reachable. We do this instead of inserting
9440 // an unreachable instruction directly because we cannot modify the
9442 new StoreInst(UndefValue::get(LI.getType()),
9443 Constant::getNullValue(Op->getType()), &LI);
9444 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9447 } else if (CE->isCast()) {
9448 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9453 // If this load comes from anywhere in a constant global, and if the global
9454 // is all undef or zero, we know what it loads.
9455 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9456 if (GV->isConstant() && GV->hasInitializer()) {
9457 if (GV->getInitializer()->isNullValue())
9458 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9459 else if (isa<UndefValue>(GV->getInitializer()))
9460 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9464 if (Op->hasOneUse()) {
9465 // Change select and PHI nodes to select values instead of addresses: this
9466 // helps alias analysis out a lot, allows many others simplifications, and
9467 // exposes redundancy in the code.
9469 // Note that we cannot do the transformation unless we know that the
9470 // introduced loads cannot trap! Something like this is valid as long as
9471 // the condition is always false: load (select bool %C, int* null, int* %G),
9472 // but it would not be valid if we transformed it to load from null
9475 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9476 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9477 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9478 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9479 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9480 SI->getOperand(1)->getName()+".val"), LI);
9481 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9482 SI->getOperand(2)->getName()+".val"), LI);
9483 return new SelectInst(SI->getCondition(), V1, V2);
9486 // load (select (cond, null, P)) -> load P
9487 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9488 if (C->isNullValue()) {
9489 LI.setOperand(0, SI->getOperand(2));
9493 // load (select (cond, P, null)) -> load P
9494 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9495 if (C->isNullValue()) {
9496 LI.setOperand(0, SI->getOperand(1));
9504 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9506 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9507 User *CI = cast<User>(SI.getOperand(1));
9508 Value *CastOp = CI->getOperand(0);
9510 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9511 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9512 const Type *SrcPTy = SrcTy->getElementType();
9514 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9515 // If the source is an array, the code below will not succeed. Check to
9516 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9518 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9519 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9520 if (ASrcTy->getNumElements() != 0) {
9522 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9523 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9524 SrcTy = cast<PointerType>(CastOp->getType());
9525 SrcPTy = SrcTy->getElementType();
9528 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9529 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9530 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9532 // Okay, we are casting from one integer or pointer type to another of
9533 // the same size. Instead of casting the pointer before
9534 // the store, cast the value to be stored.
9536 Value *SIOp0 = SI.getOperand(0);
9537 Instruction::CastOps opcode = Instruction::BitCast;
9538 const Type* CastSrcTy = SIOp0->getType();
9539 const Type* CastDstTy = SrcPTy;
9540 if (isa<PointerType>(CastDstTy)) {
9541 if (CastSrcTy->isInteger())
9542 opcode = Instruction::IntToPtr;
9543 } else if (isa<IntegerType>(CastDstTy)) {
9544 if (isa<PointerType>(SIOp0->getType()))
9545 opcode = Instruction::PtrToInt;
9547 if (Constant *C = dyn_cast<Constant>(SIOp0))
9548 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9550 NewCast = IC.InsertNewInstBefore(
9551 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9553 return new StoreInst(NewCast, CastOp);
9560 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9561 Value *Val = SI.getOperand(0);
9562 Value *Ptr = SI.getOperand(1);
9564 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9565 EraseInstFromFunction(SI);
9570 // If the RHS is an alloca with a single use, zapify the store, making the
9572 if (Ptr->hasOneUse()) {
9573 if (isa<AllocaInst>(Ptr)) {
9574 EraseInstFromFunction(SI);
9579 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9580 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9581 GEP->getOperand(0)->hasOneUse()) {
9582 EraseInstFromFunction(SI);
9588 // Attempt to improve the alignment.
9589 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9590 if (KnownAlign > SI.getAlignment())
9591 SI.setAlignment(KnownAlign);
9593 // Do really simple DSE, to catch cases where there are several consequtive
9594 // stores to the same location, separated by a few arithmetic operations. This
9595 // situation often occurs with bitfield accesses.
9596 BasicBlock::iterator BBI = &SI;
9597 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9601 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9602 // Prev store isn't volatile, and stores to the same location?
9603 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9606 EraseInstFromFunction(*PrevSI);
9612 // If this is a load, we have to stop. However, if the loaded value is from
9613 // the pointer we're loading and is producing the pointer we're storing,
9614 // then *this* store is dead (X = load P; store X -> P).
9615 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9616 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9617 EraseInstFromFunction(SI);
9621 // Otherwise, this is a load from some other location. Stores before it
9626 // Don't skip over loads or things that can modify memory.
9627 if (BBI->mayWriteToMemory())
9632 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9634 // store X, null -> turns into 'unreachable' in SimplifyCFG
9635 if (isa<ConstantPointerNull>(Ptr)) {
9636 if (!isa<UndefValue>(Val)) {
9637 SI.setOperand(0, UndefValue::get(Val->getType()));
9638 if (Instruction *U = dyn_cast<Instruction>(Val))
9639 AddToWorkList(U); // Dropped a use.
9642 return 0; // Do not modify these!
9645 // store undef, Ptr -> noop
9646 if (isa<UndefValue>(Val)) {
9647 EraseInstFromFunction(SI);
9652 // If the pointer destination is a cast, see if we can fold the cast into the
9654 if (isa<CastInst>(Ptr))
9655 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9657 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9659 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9663 // If this store is the last instruction in the basic block, and if the block
9664 // ends with an unconditional branch, try to move it to the successor block.
9666 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9667 if (BI->isUnconditional())
9668 if (SimplifyStoreAtEndOfBlock(SI))
9669 return 0; // xform done!
9674 /// SimplifyStoreAtEndOfBlock - Turn things like:
9675 /// if () { *P = v1; } else { *P = v2 }
9676 /// into a phi node with a store in the successor.
9678 /// Simplify things like:
9679 /// *P = v1; if () { *P = v2; }
9680 /// into a phi node with a store in the successor.
9682 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9683 BasicBlock *StoreBB = SI.getParent();
9685 // Check to see if the successor block has exactly two incoming edges. If
9686 // so, see if the other predecessor contains a store to the same location.
9687 // if so, insert a PHI node (if needed) and move the stores down.
9688 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9690 // Determine whether Dest has exactly two predecessors and, if so, compute
9691 // the other predecessor.
9692 pred_iterator PI = pred_begin(DestBB);
9693 BasicBlock *OtherBB = 0;
9697 if (PI == pred_end(DestBB))
9700 if (*PI != StoreBB) {
9705 if (++PI != pred_end(DestBB))
9709 // Verify that the other block ends in a branch and is not otherwise empty.
9710 BasicBlock::iterator BBI = OtherBB->getTerminator();
9711 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9712 if (!OtherBr || BBI == OtherBB->begin())
9715 // If the other block ends in an unconditional branch, check for the 'if then
9716 // else' case. there is an instruction before the branch.
9717 StoreInst *OtherStore = 0;
9718 if (OtherBr->isUnconditional()) {
9719 // If this isn't a store, or isn't a store to the same location, bail out.
9721 OtherStore = dyn_cast<StoreInst>(BBI);
9722 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9725 // Otherwise, the other block ended with a conditional branch. If one of the
9726 // destinations is StoreBB, then we have the if/then case.
9727 if (OtherBr->getSuccessor(0) != StoreBB &&
9728 OtherBr->getSuccessor(1) != StoreBB)
9731 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9732 // if/then triangle. See if there is a store to the same ptr as SI that
9733 // lives in OtherBB.
9735 // Check to see if we find the matching store.
9736 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9737 if (OtherStore->getOperand(1) != SI.getOperand(1))
9741 // If we find something that may be using the stored value, or if we run
9742 // out of instructions, we can't do the xform.
9743 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9744 BBI == OtherBB->begin())
9748 // In order to eliminate the store in OtherBr, we have to
9749 // make sure nothing reads the stored value in StoreBB.
9750 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9751 // FIXME: This should really be AA driven.
9752 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9757 // Insert a PHI node now if we need it.
9758 Value *MergedVal = OtherStore->getOperand(0);
9759 if (MergedVal != SI.getOperand(0)) {
9760 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9761 PN->reserveOperandSpace(2);
9762 PN->addIncoming(SI.getOperand(0), SI.getParent());
9763 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9764 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9767 // Advance to a place where it is safe to insert the new store and
9769 BBI = DestBB->begin();
9770 while (isa<PHINode>(BBI)) ++BBI;
9771 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9772 OtherStore->isVolatile()), *BBI);
9774 // Nuke the old stores.
9775 EraseInstFromFunction(SI);
9776 EraseInstFromFunction(*OtherStore);
9782 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9783 // Change br (not X), label True, label False to: br X, label False, True
9785 BasicBlock *TrueDest;
9786 BasicBlock *FalseDest;
9787 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9788 !isa<Constant>(X)) {
9789 // Swap Destinations and condition...
9791 BI.setSuccessor(0, FalseDest);
9792 BI.setSuccessor(1, TrueDest);
9796 // Cannonicalize fcmp_one -> fcmp_oeq
9797 FCmpInst::Predicate FPred; Value *Y;
9798 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9799 TrueDest, FalseDest)))
9800 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9801 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9802 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9803 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9804 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9805 NewSCC->takeName(I);
9806 // Swap Destinations and condition...
9807 BI.setCondition(NewSCC);
9808 BI.setSuccessor(0, FalseDest);
9809 BI.setSuccessor(1, TrueDest);
9810 RemoveFromWorkList(I);
9811 I->eraseFromParent();
9812 AddToWorkList(NewSCC);
9816 // Cannonicalize icmp_ne -> icmp_eq
9817 ICmpInst::Predicate IPred;
9818 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9819 TrueDest, FalseDest)))
9820 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9821 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9822 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9823 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9824 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9825 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9826 NewSCC->takeName(I);
9827 // Swap Destinations and condition...
9828 BI.setCondition(NewSCC);
9829 BI.setSuccessor(0, FalseDest);
9830 BI.setSuccessor(1, TrueDest);
9831 RemoveFromWorkList(I);
9832 I->eraseFromParent();;
9833 AddToWorkList(NewSCC);
9840 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9841 Value *Cond = SI.getCondition();
9842 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9843 if (I->getOpcode() == Instruction::Add)
9844 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9845 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9846 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9847 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9849 SI.setOperand(0, I->getOperand(0));
9857 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9858 /// is to leave as a vector operation.
9859 static bool CheapToScalarize(Value *V, bool isConstant) {
9860 if (isa<ConstantAggregateZero>(V))
9862 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9863 if (isConstant) return true;
9864 // If all elts are the same, we can extract.
9865 Constant *Op0 = C->getOperand(0);
9866 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9867 if (C->getOperand(i) != Op0)
9871 Instruction *I = dyn_cast<Instruction>(V);
9872 if (!I) return false;
9874 // Insert element gets simplified to the inserted element or is deleted if
9875 // this is constant idx extract element and its a constant idx insertelt.
9876 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9877 isa<ConstantInt>(I->getOperand(2)))
9879 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9881 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9882 if (BO->hasOneUse() &&
9883 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9884 CheapToScalarize(BO->getOperand(1), isConstant)))
9886 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9887 if (CI->hasOneUse() &&
9888 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9889 CheapToScalarize(CI->getOperand(1), isConstant)))
9895 /// Read and decode a shufflevector mask.
9897 /// It turns undef elements into values that are larger than the number of
9898 /// elements in the input.
9899 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9900 unsigned NElts = SVI->getType()->getNumElements();
9901 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9902 return std::vector<unsigned>(NElts, 0);
9903 if (isa<UndefValue>(SVI->getOperand(2)))
9904 return std::vector<unsigned>(NElts, 2*NElts);
9906 std::vector<unsigned> Result;
9907 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9908 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9909 if (isa<UndefValue>(CP->getOperand(i)))
9910 Result.push_back(NElts*2); // undef -> 8
9912 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9916 /// FindScalarElement - Given a vector and an element number, see if the scalar
9917 /// value is already around as a register, for example if it were inserted then
9918 /// extracted from the vector.
9919 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9920 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9921 const VectorType *PTy = cast<VectorType>(V->getType());
9922 unsigned Width = PTy->getNumElements();
9923 if (EltNo >= Width) // Out of range access.
9924 return UndefValue::get(PTy->getElementType());
9926 if (isa<UndefValue>(V))
9927 return UndefValue::get(PTy->getElementType());
9928 else if (isa<ConstantAggregateZero>(V))
9929 return Constant::getNullValue(PTy->getElementType());
9930 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9931 return CP->getOperand(EltNo);
9932 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9933 // If this is an insert to a variable element, we don't know what it is.
9934 if (!isa<ConstantInt>(III->getOperand(2)))
9936 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9938 // If this is an insert to the element we are looking for, return the
9941 return III->getOperand(1);
9943 // Otherwise, the insertelement doesn't modify the value, recurse on its
9945 return FindScalarElement(III->getOperand(0), EltNo);
9946 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9947 unsigned InEl = getShuffleMask(SVI)[EltNo];
9949 return FindScalarElement(SVI->getOperand(0), InEl);
9950 else if (InEl < Width*2)
9951 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9953 return UndefValue::get(PTy->getElementType());
9956 // Otherwise, we don't know.
9960 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9962 // If vector val is undef, replace extract with scalar undef.
9963 if (isa<UndefValue>(EI.getOperand(0)))
9964 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9966 // If vector val is constant 0, replace extract with scalar 0.
9967 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9968 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9970 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9971 // If vector val is constant with uniform operands, replace EI
9972 // with that operand
9973 Constant *op0 = C->getOperand(0);
9974 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9975 if (C->getOperand(i) != op0) {
9980 return ReplaceInstUsesWith(EI, op0);
9983 // If extracting a specified index from the vector, see if we can recursively
9984 // find a previously computed scalar that was inserted into the vector.
9985 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9986 unsigned IndexVal = IdxC->getZExtValue();
9987 unsigned VectorWidth =
9988 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9990 // If this is extracting an invalid index, turn this into undef, to avoid
9991 // crashing the code below.
9992 if (IndexVal >= VectorWidth)
9993 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9995 // This instruction only demands the single element from the input vector.
9996 // If the input vector has a single use, simplify it based on this use
9998 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10000 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10003 EI.setOperand(0, V);
10008 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10009 return ReplaceInstUsesWith(EI, Elt);
10011 // If the this extractelement is directly using a bitcast from a vector of
10012 // the same number of elements, see if we can find the source element from
10013 // it. In this case, we will end up needing to bitcast the scalars.
10014 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10015 if (const VectorType *VT =
10016 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10017 if (VT->getNumElements() == VectorWidth)
10018 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10019 return new BitCastInst(Elt, EI.getType());
10023 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10024 if (I->hasOneUse()) {
10025 // Push extractelement into predecessor operation if legal and
10026 // profitable to do so
10027 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10028 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10029 if (CheapToScalarize(BO, isConstantElt)) {
10030 ExtractElementInst *newEI0 =
10031 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10032 EI.getName()+".lhs");
10033 ExtractElementInst *newEI1 =
10034 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10035 EI.getName()+".rhs");
10036 InsertNewInstBefore(newEI0, EI);
10037 InsertNewInstBefore(newEI1, EI);
10038 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10040 } else if (isa<LoadInst>(I)) {
10042 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10043 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
10044 PointerType::get(EI.getType(), AS), EI);
10045 GetElementPtrInst *GEP =
10046 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10047 InsertNewInstBefore(GEP, EI);
10048 return new LoadInst(GEP);
10051 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10052 // Extracting the inserted element?
10053 if (IE->getOperand(2) == EI.getOperand(1))
10054 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10055 // If the inserted and extracted elements are constants, they must not
10056 // be the same value, extract from the pre-inserted value instead.
10057 if (isa<Constant>(IE->getOperand(2)) &&
10058 isa<Constant>(EI.getOperand(1))) {
10059 AddUsesToWorkList(EI);
10060 EI.setOperand(0, IE->getOperand(0));
10063 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10064 // If this is extracting an element from a shufflevector, figure out where
10065 // it came from and extract from the appropriate input element instead.
10066 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10067 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10069 if (SrcIdx < SVI->getType()->getNumElements())
10070 Src = SVI->getOperand(0);
10071 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10072 SrcIdx -= SVI->getType()->getNumElements();
10073 Src = SVI->getOperand(1);
10075 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10077 return new ExtractElementInst(Src, SrcIdx);
10084 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10085 /// elements from either LHS or RHS, return the shuffle mask and true.
10086 /// Otherwise, return false.
10087 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10088 std::vector<Constant*> &Mask) {
10089 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10090 "Invalid CollectSingleShuffleElements");
10091 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10093 if (isa<UndefValue>(V)) {
10094 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10096 } else if (V == LHS) {
10097 for (unsigned i = 0; i != NumElts; ++i)
10098 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10100 } else if (V == RHS) {
10101 for (unsigned i = 0; i != NumElts; ++i)
10102 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10104 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10105 // If this is an insert of an extract from some other vector, include it.
10106 Value *VecOp = IEI->getOperand(0);
10107 Value *ScalarOp = IEI->getOperand(1);
10108 Value *IdxOp = IEI->getOperand(2);
10110 if (!isa<ConstantInt>(IdxOp))
10112 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10114 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10115 // Okay, we can handle this if the vector we are insertinting into is
10116 // transitively ok.
10117 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10118 // If so, update the mask to reflect the inserted undef.
10119 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10122 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10123 if (isa<ConstantInt>(EI->getOperand(1)) &&
10124 EI->getOperand(0)->getType() == V->getType()) {
10125 unsigned ExtractedIdx =
10126 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10128 // This must be extracting from either LHS or RHS.
10129 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10130 // Okay, we can handle this if the vector we are insertinting into is
10131 // transitively ok.
10132 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10133 // If so, update the mask to reflect the inserted value.
10134 if (EI->getOperand(0) == LHS) {
10135 Mask[InsertedIdx & (NumElts-1)] =
10136 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10138 assert(EI->getOperand(0) == RHS);
10139 Mask[InsertedIdx & (NumElts-1)] =
10140 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10149 // TODO: Handle shufflevector here!
10154 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10155 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10156 /// that computes V and the LHS value of the shuffle.
10157 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10159 assert(isa<VectorType>(V->getType()) &&
10160 (RHS == 0 || V->getType() == RHS->getType()) &&
10161 "Invalid shuffle!");
10162 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10164 if (isa<UndefValue>(V)) {
10165 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10167 } else if (isa<ConstantAggregateZero>(V)) {
10168 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10170 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10171 // If this is an insert of an extract from some other vector, include it.
10172 Value *VecOp = IEI->getOperand(0);
10173 Value *ScalarOp = IEI->getOperand(1);
10174 Value *IdxOp = IEI->getOperand(2);
10176 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10177 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10178 EI->getOperand(0)->getType() == V->getType()) {
10179 unsigned ExtractedIdx =
10180 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10181 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10183 // Either the extracted from or inserted into vector must be RHSVec,
10184 // otherwise we'd end up with a shuffle of three inputs.
10185 if (EI->getOperand(0) == RHS || RHS == 0) {
10186 RHS = EI->getOperand(0);
10187 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10188 Mask[InsertedIdx & (NumElts-1)] =
10189 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10193 if (VecOp == RHS) {
10194 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10195 // Everything but the extracted element is replaced with the RHS.
10196 for (unsigned i = 0; i != NumElts; ++i) {
10197 if (i != InsertedIdx)
10198 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10203 // If this insertelement is a chain that comes from exactly these two
10204 // vectors, return the vector and the effective shuffle.
10205 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10206 return EI->getOperand(0);
10211 // TODO: Handle shufflevector here!
10213 // Otherwise, can't do anything fancy. Return an identity vector.
10214 for (unsigned i = 0; i != NumElts; ++i)
10215 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10219 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10220 Value *VecOp = IE.getOperand(0);
10221 Value *ScalarOp = IE.getOperand(1);
10222 Value *IdxOp = IE.getOperand(2);
10224 // Inserting an undef or into an undefined place, remove this.
10225 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10226 ReplaceInstUsesWith(IE, VecOp);
10228 // If the inserted element was extracted from some other vector, and if the
10229 // indexes are constant, try to turn this into a shufflevector operation.
10230 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10231 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10232 EI->getOperand(0)->getType() == IE.getType()) {
10233 unsigned NumVectorElts = IE.getType()->getNumElements();
10234 unsigned ExtractedIdx =
10235 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10236 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10238 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10239 return ReplaceInstUsesWith(IE, VecOp);
10241 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10242 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10244 // If we are extracting a value from a vector, then inserting it right
10245 // back into the same place, just use the input vector.
10246 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10247 return ReplaceInstUsesWith(IE, VecOp);
10249 // We could theoretically do this for ANY input. However, doing so could
10250 // turn chains of insertelement instructions into a chain of shufflevector
10251 // instructions, and right now we do not merge shufflevectors. As such,
10252 // only do this in a situation where it is clear that there is benefit.
10253 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10254 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10255 // the values of VecOp, except then one read from EIOp0.
10256 // Build a new shuffle mask.
10257 std::vector<Constant*> Mask;
10258 if (isa<UndefValue>(VecOp))
10259 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10261 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10262 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10265 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10266 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10267 ConstantVector::get(Mask));
10270 // If this insertelement isn't used by some other insertelement, turn it
10271 // (and any insertelements it points to), into one big shuffle.
10272 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10273 std::vector<Constant*> Mask;
10275 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10276 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10277 // We now have a shuffle of LHS, RHS, Mask.
10278 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10287 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10288 Value *LHS = SVI.getOperand(0);
10289 Value *RHS = SVI.getOperand(1);
10290 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10292 bool MadeChange = false;
10294 // Undefined shuffle mask -> undefined value.
10295 if (isa<UndefValue>(SVI.getOperand(2)))
10296 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10298 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10299 // the undef, change them to undefs.
10300 if (isa<UndefValue>(SVI.getOperand(1))) {
10301 // Scan to see if there are any references to the RHS. If so, replace them
10302 // with undef element refs and set MadeChange to true.
10303 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10304 if (Mask[i] >= e && Mask[i] != 2*e) {
10311 // Remap any references to RHS to use LHS.
10312 std::vector<Constant*> Elts;
10313 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10314 if (Mask[i] == 2*e)
10315 Elts.push_back(UndefValue::get(Type::Int32Ty));
10317 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10319 SVI.setOperand(2, ConstantVector::get(Elts));
10323 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10324 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10325 if (LHS == RHS || isa<UndefValue>(LHS)) {
10326 if (isa<UndefValue>(LHS) && LHS == RHS) {
10327 // shuffle(undef,undef,mask) -> undef.
10328 return ReplaceInstUsesWith(SVI, LHS);
10331 // Remap any references to RHS to use LHS.
10332 std::vector<Constant*> Elts;
10333 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10334 if (Mask[i] >= 2*e)
10335 Elts.push_back(UndefValue::get(Type::Int32Ty));
10337 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10338 (Mask[i] < e && isa<UndefValue>(LHS)))
10339 Mask[i] = 2*e; // Turn into undef.
10341 Mask[i] &= (e-1); // Force to LHS.
10342 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10345 SVI.setOperand(0, SVI.getOperand(1));
10346 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10347 SVI.setOperand(2, ConstantVector::get(Elts));
10348 LHS = SVI.getOperand(0);
10349 RHS = SVI.getOperand(1);
10353 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10354 bool isLHSID = true, isRHSID = true;
10356 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10357 if (Mask[i] >= e*2) continue; // Ignore undef values.
10358 // Is this an identity shuffle of the LHS value?
10359 isLHSID &= (Mask[i] == i);
10361 // Is this an identity shuffle of the RHS value?
10362 isRHSID &= (Mask[i]-e == i);
10365 // Eliminate identity shuffles.
10366 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10367 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10369 // If the LHS is a shufflevector itself, see if we can combine it with this
10370 // one without producing an unusual shuffle. Here we are really conservative:
10371 // we are absolutely afraid of producing a shuffle mask not in the input
10372 // program, because the code gen may not be smart enough to turn a merged
10373 // shuffle into two specific shuffles: it may produce worse code. As such,
10374 // we only merge two shuffles if the result is one of the two input shuffle
10375 // masks. In this case, merging the shuffles just removes one instruction,
10376 // which we know is safe. This is good for things like turning:
10377 // (splat(splat)) -> splat.
10378 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10379 if (isa<UndefValue>(RHS)) {
10380 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10382 std::vector<unsigned> NewMask;
10383 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10384 if (Mask[i] >= 2*e)
10385 NewMask.push_back(2*e);
10387 NewMask.push_back(LHSMask[Mask[i]]);
10389 // If the result mask is equal to the src shuffle or this shuffle mask, do
10390 // the replacement.
10391 if (NewMask == LHSMask || NewMask == Mask) {
10392 std::vector<Constant*> Elts;
10393 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10394 if (NewMask[i] >= e*2) {
10395 Elts.push_back(UndefValue::get(Type::Int32Ty));
10397 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10400 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10401 LHSSVI->getOperand(1),
10402 ConstantVector::get(Elts));
10407 return MadeChange ? &SVI : 0;
10413 /// TryToSinkInstruction - Try to move the specified instruction from its
10414 /// current block into the beginning of DestBlock, which can only happen if it's
10415 /// safe to move the instruction past all of the instructions between it and the
10416 /// end of its block.
10417 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10418 assert(I->hasOneUse() && "Invariants didn't hold!");
10420 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10421 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10423 // Do not sink alloca instructions out of the entry block.
10424 if (isa<AllocaInst>(I) && I->getParent() ==
10425 &DestBlock->getParent()->getEntryBlock())
10428 // We can only sink load instructions if there is nothing between the load and
10429 // the end of block that could change the value.
10430 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10431 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10433 if (Scan->mayWriteToMemory())
10437 BasicBlock::iterator InsertPos = DestBlock->begin();
10438 while (isa<PHINode>(InsertPos)) ++InsertPos;
10440 I->moveBefore(InsertPos);
10446 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10447 /// all reachable code to the worklist.
10449 /// This has a couple of tricks to make the code faster and more powerful. In
10450 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10451 /// them to the worklist (this significantly speeds up instcombine on code where
10452 /// many instructions are dead or constant). Additionally, if we find a branch
10453 /// whose condition is a known constant, we only visit the reachable successors.
10455 static void AddReachableCodeToWorklist(BasicBlock *BB,
10456 SmallPtrSet<BasicBlock*, 64> &Visited,
10458 const TargetData *TD) {
10459 std::vector<BasicBlock*> Worklist;
10460 Worklist.push_back(BB);
10462 while (!Worklist.empty()) {
10463 BB = Worklist.back();
10464 Worklist.pop_back();
10466 // We have now visited this block! If we've already been here, ignore it.
10467 if (!Visited.insert(BB)) continue;
10469 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10470 Instruction *Inst = BBI++;
10472 // DCE instruction if trivially dead.
10473 if (isInstructionTriviallyDead(Inst)) {
10475 DOUT << "IC: DCE: " << *Inst;
10476 Inst->eraseFromParent();
10480 // ConstantProp instruction if trivially constant.
10481 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10482 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10483 Inst->replaceAllUsesWith(C);
10485 Inst->eraseFromParent();
10489 IC.AddToWorkList(Inst);
10492 // Recursively visit successors. If this is a branch or switch on a
10493 // constant, only visit the reachable successor.
10494 TerminatorInst *TI = BB->getTerminator();
10495 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10496 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10497 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10498 Worklist.push_back(BI->getSuccessor(!CondVal));
10501 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10502 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10503 // See if this is an explicit destination.
10504 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10505 if (SI->getCaseValue(i) == Cond) {
10506 Worklist.push_back(SI->getSuccessor(i));
10510 // Otherwise it is the default destination.
10511 Worklist.push_back(SI->getSuccessor(0));
10516 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10517 Worklist.push_back(TI->getSuccessor(i));
10521 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10522 bool Changed = false;
10523 TD = &getAnalysis<TargetData>();
10525 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10526 << F.getNameStr() << "\n");
10529 // Do a depth-first traversal of the function, populate the worklist with
10530 // the reachable instructions. Ignore blocks that are not reachable. Keep
10531 // track of which blocks we visit.
10532 SmallPtrSet<BasicBlock*, 64> Visited;
10533 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10535 // Do a quick scan over the function. If we find any blocks that are
10536 // unreachable, remove any instructions inside of them. This prevents
10537 // the instcombine code from having to deal with some bad special cases.
10538 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10539 if (!Visited.count(BB)) {
10540 Instruction *Term = BB->getTerminator();
10541 while (Term != BB->begin()) { // Remove instrs bottom-up
10542 BasicBlock::iterator I = Term; --I;
10544 DOUT << "IC: DCE: " << *I;
10547 if (!I->use_empty())
10548 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10549 I->eraseFromParent();
10554 while (!Worklist.empty()) {
10555 Instruction *I = RemoveOneFromWorkList();
10556 if (I == 0) continue; // skip null values.
10558 // Check to see if we can DCE the instruction.
10559 if (isInstructionTriviallyDead(I)) {
10560 // Add operands to the worklist.
10561 if (I->getNumOperands() < 4)
10562 AddUsesToWorkList(*I);
10565 DOUT << "IC: DCE: " << *I;
10567 I->eraseFromParent();
10568 RemoveFromWorkList(I);
10572 // Instruction isn't dead, see if we can constant propagate it.
10573 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10574 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10576 // Add operands to the worklist.
10577 AddUsesToWorkList(*I);
10578 ReplaceInstUsesWith(*I, C);
10581 I->eraseFromParent();
10582 RemoveFromWorkList(I);
10586 // See if we can trivially sink this instruction to a successor basic block.
10587 if (I->hasOneUse()) {
10588 BasicBlock *BB = I->getParent();
10589 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10590 if (UserParent != BB) {
10591 bool UserIsSuccessor = false;
10592 // See if the user is one of our successors.
10593 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10594 if (*SI == UserParent) {
10595 UserIsSuccessor = true;
10599 // If the user is one of our immediate successors, and if that successor
10600 // only has us as a predecessors (we'd have to split the critical edge
10601 // otherwise), we can keep going.
10602 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10603 next(pred_begin(UserParent)) == pred_end(UserParent))
10604 // Okay, the CFG is simple enough, try to sink this instruction.
10605 Changed |= TryToSinkInstruction(I, UserParent);
10609 // Now that we have an instruction, try combining it to simplify it...
10613 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10614 if (Instruction *Result = visit(*I)) {
10616 // Should we replace the old instruction with a new one?
10618 DOUT << "IC: Old = " << *I
10619 << " New = " << *Result;
10621 // Everything uses the new instruction now.
10622 I->replaceAllUsesWith(Result);
10624 // Push the new instruction and any users onto the worklist.
10625 AddToWorkList(Result);
10626 AddUsersToWorkList(*Result);
10628 // Move the name to the new instruction first.
10629 Result->takeName(I);
10631 // Insert the new instruction into the basic block...
10632 BasicBlock *InstParent = I->getParent();
10633 BasicBlock::iterator InsertPos = I;
10635 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10636 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10639 InstParent->getInstList().insert(InsertPos, Result);
10641 // Make sure that we reprocess all operands now that we reduced their
10643 AddUsesToWorkList(*I);
10645 // Instructions can end up on the worklist more than once. Make sure
10646 // we do not process an instruction that has been deleted.
10647 RemoveFromWorkList(I);
10649 // Erase the old instruction.
10650 InstParent->getInstList().erase(I);
10653 DOUT << "IC: Mod = " << OrigI
10654 << " New = " << *I;
10657 // If the instruction was modified, it's possible that it is now dead.
10658 // if so, remove it.
10659 if (isInstructionTriviallyDead(I)) {
10660 // Make sure we process all operands now that we are reducing their
10662 AddUsesToWorkList(*I);
10664 // Instructions may end up in the worklist more than once. Erase all
10665 // occurrences of this instruction.
10666 RemoveFromWorkList(I);
10667 I->eraseFromParent();
10670 AddUsersToWorkList(*I);
10677 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10679 // Do an explicit clear, this shrinks the map if needed.
10680 WorklistMap.clear();
10685 bool InstCombiner::runOnFunction(Function &F) {
10686 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10688 bool EverMadeChange = false;
10690 // Iterate while there is work to do.
10691 unsigned Iteration = 0;
10692 while (DoOneIteration(F, Iteration++))
10693 EverMadeChange = true;
10694 return EverMadeChange;
10697 FunctionPass *llvm::createInstructionCombiningPass() {
10698 return new InstCombiner();