1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
67 STATISTIC(NumSunkInst , "Number of instructions sunk");
70 class VISIBILITY_HIDDEN InstCombiner
71 : public FunctionPass,
72 public InstVisitor<InstCombiner, Instruction*> {
73 // Worklist of all of the instructions that need to be simplified.
74 std::vector<Instruction*> Worklist;
75 DenseMap<Instruction*, unsigned> WorklistMap;
77 bool MustPreserveLCSSA;
79 /// AddToWorkList - Add the specified instruction to the worklist if it
80 /// isn't already in it.
81 void AddToWorkList(Instruction *I) {
82 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
83 Worklist.push_back(I);
86 // RemoveFromWorkList - remove I from the worklist if it exists.
87 void RemoveFromWorkList(Instruction *I) {
88 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
89 if (It == WorklistMap.end()) return; // Not in worklist.
91 // Don't bother moving everything down, just null out the slot.
92 Worklist[It->second] = 0;
94 WorklistMap.erase(It);
97 Instruction *RemoveOneFromWorkList() {
98 Instruction *I = Worklist.back();
100 WorklistMap.erase(I);
105 /// AddUsersToWorkList - When an instruction is simplified, add all users of
106 /// the instruction to the work lists because they might get more simplified
109 void AddUsersToWorkList(Value &I) {
110 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
112 AddToWorkList(cast<Instruction>(*UI));
115 /// AddUsesToWorkList - When an instruction is simplified, add operands to
116 /// the work lists because they might get more simplified now.
118 void AddUsesToWorkList(Instruction &I) {
119 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
120 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
124 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
125 /// dead. Add all of its operands to the worklist, turning them into
126 /// undef's to reduce the number of uses of those instructions.
128 /// Return the specified operand before it is turned into an undef.
130 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
131 Value *R = I.getOperand(op);
133 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
134 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
136 // Set the operand to undef to drop the use.
137 I.setOperand(i, UndefValue::get(Op->getType()));
144 virtual bool runOnFunction(Function &F);
146 bool DoOneIteration(Function &F, unsigned ItNum);
148 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
149 AU.addRequired<TargetData>();
150 AU.addPreservedID(LCSSAID);
151 AU.setPreservesCFG();
154 TargetData &getTargetData() const { return *TD; }
156 // Visitation implementation - Implement instruction combining for different
157 // instruction types. The semantics are as follows:
159 // null - No change was made
160 // I - Change was made, I is still valid, I may be dead though
161 // otherwise - Change was made, replace I with returned instruction
163 Instruction *visitAdd(BinaryOperator &I);
164 Instruction *visitSub(BinaryOperator &I);
165 Instruction *visitMul(BinaryOperator &I);
166 Instruction *visitURem(BinaryOperator &I);
167 Instruction *visitSRem(BinaryOperator &I);
168 Instruction *visitFRem(BinaryOperator &I);
169 Instruction *commonRemTransforms(BinaryOperator &I);
170 Instruction *commonIRemTransforms(BinaryOperator &I);
171 Instruction *commonDivTransforms(BinaryOperator &I);
172 Instruction *commonIDivTransforms(BinaryOperator &I);
173 Instruction *visitUDiv(BinaryOperator &I);
174 Instruction *visitSDiv(BinaryOperator &I);
175 Instruction *visitFDiv(BinaryOperator &I);
176 Instruction *visitAnd(BinaryOperator &I);
177 Instruction *visitOr (BinaryOperator &I);
178 Instruction *visitXor(BinaryOperator &I);
179 Instruction *visitShl(BinaryOperator &I);
180 Instruction *visitAShr(BinaryOperator &I);
181 Instruction *visitLShr(BinaryOperator &I);
182 Instruction *commonShiftTransforms(BinaryOperator &I);
183 Instruction *visitFCmpInst(FCmpInst &I);
184 Instruction *visitICmpInst(ICmpInst &I);
185 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
187 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
188 ICmpInst::Predicate Cond, Instruction &I);
189 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
191 Instruction *commonCastTransforms(CastInst &CI);
192 Instruction *commonIntCastTransforms(CastInst &CI);
193 Instruction *visitTrunc(CastInst &CI);
194 Instruction *visitZExt(CastInst &CI);
195 Instruction *visitSExt(CastInst &CI);
196 Instruction *visitFPTrunc(CastInst &CI);
197 Instruction *visitFPExt(CastInst &CI);
198 Instruction *visitFPToUI(CastInst &CI);
199 Instruction *visitFPToSI(CastInst &CI);
200 Instruction *visitUIToFP(CastInst &CI);
201 Instruction *visitSIToFP(CastInst &CI);
202 Instruction *visitPtrToInt(CastInst &CI);
203 Instruction *visitIntToPtr(CastInst &CI);
204 Instruction *visitBitCast(CastInst &CI);
205 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
207 Instruction *visitSelectInst(SelectInst &CI);
208 Instruction *visitCallInst(CallInst &CI);
209 Instruction *visitInvokeInst(InvokeInst &II);
210 Instruction *visitPHINode(PHINode &PN);
211 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
212 Instruction *visitAllocationInst(AllocationInst &AI);
213 Instruction *visitFreeInst(FreeInst &FI);
214 Instruction *visitLoadInst(LoadInst &LI);
215 Instruction *visitStoreInst(StoreInst &SI);
216 Instruction *visitBranchInst(BranchInst &BI);
217 Instruction *visitSwitchInst(SwitchInst &SI);
218 Instruction *visitInsertElementInst(InsertElementInst &IE);
219 Instruction *visitExtractElementInst(ExtractElementInst &EI);
220 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
222 // visitInstruction - Specify what to return for unhandled instructions...
223 Instruction *visitInstruction(Instruction &I) { return 0; }
226 Instruction *visitCallSite(CallSite CS);
227 bool transformConstExprCastCall(CallSite CS);
230 // InsertNewInstBefore - insert an instruction New before instruction Old
231 // in the program. Add the new instruction to the worklist.
233 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
234 assert(New && New->getParent() == 0 &&
235 "New instruction already inserted into a basic block!");
236 BasicBlock *BB = Old.getParent();
237 BB->getInstList().insert(&Old, New); // Insert inst
242 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
243 /// This also adds the cast to the worklist. Finally, this returns the
245 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
247 if (V->getType() == Ty) return V;
249 if (Constant *CV = dyn_cast<Constant>(V))
250 return ConstantExpr::getCast(opc, CV, Ty);
252 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
257 // ReplaceInstUsesWith - This method is to be used when an instruction is
258 // found to be dead, replacable with another preexisting expression. Here
259 // we add all uses of I to the worklist, replace all uses of I with the new
260 // value, then return I, so that the inst combiner will know that I was
263 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
264 AddUsersToWorkList(I); // Add all modified instrs to worklist
266 I.replaceAllUsesWith(V);
269 // If we are replacing the instruction with itself, this must be in a
270 // segment of unreachable code, so just clobber the instruction.
271 I.replaceAllUsesWith(UndefValue::get(I.getType()));
276 // UpdateValueUsesWith - This method is to be used when an value is
277 // found to be replacable with another preexisting expression or was
278 // updated. Here we add all uses of I to the worklist, replace all uses of
279 // I with the new value (unless the instruction was just updated), then
280 // return true, so that the inst combiner will know that I was modified.
282 bool UpdateValueUsesWith(Value *Old, Value *New) {
283 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
285 Old->replaceAllUsesWith(New);
286 if (Instruction *I = dyn_cast<Instruction>(Old))
288 if (Instruction *I = dyn_cast<Instruction>(New))
293 // EraseInstFromFunction - When dealing with an instruction that has side
294 // effects or produces a void value, we can't rely on DCE to delete the
295 // instruction. Instead, visit methods should return the value returned by
297 Instruction *EraseInstFromFunction(Instruction &I) {
298 assert(I.use_empty() && "Cannot erase instruction that is used!");
299 AddUsesToWorkList(I);
300 RemoveFromWorkList(&I);
302 return 0; // Don't do anything with FI
306 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
307 /// InsertBefore instruction. This is specialized a bit to avoid inserting
308 /// casts that are known to not do anything...
310 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
311 Value *V, const Type *DestTy,
312 Instruction *InsertBefore);
314 /// SimplifyCommutative - This performs a few simplifications for
315 /// commutative operators.
316 bool SimplifyCommutative(BinaryOperator &I);
318 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
319 /// most-complex to least-complex order.
320 bool SimplifyCompare(CmpInst &I);
322 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
323 /// on the demanded bits.
324 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
325 APInt& KnownZero, APInt& KnownOne,
328 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
329 uint64_t &UndefElts, unsigned Depth = 0);
331 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
332 // PHI node as operand #0, see if we can fold the instruction into the PHI
333 // (which is only possible if all operands to the PHI are constants).
334 Instruction *FoldOpIntoPhi(Instruction &I);
336 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
337 // operator and they all are only used by the PHI, PHI together their
338 // inputs, and do the operation once, to the result of the PHI.
339 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
340 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
343 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
344 ConstantInt *AndRHS, BinaryOperator &TheAnd);
346 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
347 bool isSub, Instruction &I);
348 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
349 bool isSigned, bool Inside, Instruction &IB);
350 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
351 Instruction *MatchBSwap(BinaryOperator &I);
353 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
356 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
359 // getComplexity: Assign a complexity or rank value to LLVM Values...
360 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
361 static unsigned getComplexity(Value *V) {
362 if (isa<Instruction>(V)) {
363 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
367 if (isa<Argument>(V)) return 3;
368 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
371 // isOnlyUse - Return true if this instruction will be deleted if we stop using
373 static bool isOnlyUse(Value *V) {
374 return V->hasOneUse() || isa<Constant>(V);
377 // getPromotedType - Return the specified type promoted as it would be to pass
378 // though a va_arg area...
379 static const Type *getPromotedType(const Type *Ty) {
380 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
381 if (ITy->getBitWidth() < 32)
382 return Type::Int32Ty;
383 } else if (Ty == Type::FloatTy)
384 return Type::DoubleTy;
388 /// getBitCastOperand - If the specified operand is a CastInst or a constant
389 /// expression bitcast, return the operand value, otherwise return null.
390 static Value *getBitCastOperand(Value *V) {
391 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
392 return I->getOperand(0);
393 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
394 if (CE->getOpcode() == Instruction::BitCast)
395 return CE->getOperand(0);
399 /// This function is a wrapper around CastInst::isEliminableCastPair. It
400 /// simply extracts arguments and returns what that function returns.
401 static Instruction::CastOps
402 isEliminableCastPair(
403 const CastInst *CI, ///< The first cast instruction
404 unsigned opcode, ///< The opcode of the second cast instruction
405 const Type *DstTy, ///< The target type for the second cast instruction
406 TargetData *TD ///< The target data for pointer size
409 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
410 const Type *MidTy = CI->getType(); // B from above
412 // Get the opcodes of the two Cast instructions
413 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
414 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
416 return Instruction::CastOps(
417 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
418 DstTy, TD->getIntPtrType()));
421 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
422 /// in any code being generated. It does not require codegen if V is simple
423 /// enough or if the cast can be folded into other casts.
424 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
425 const Type *Ty, TargetData *TD) {
426 if (V->getType() == Ty || isa<Constant>(V)) return false;
428 // If this is another cast that can be eliminated, it isn't codegen either.
429 if (const CastInst *CI = dyn_cast<CastInst>(V))
430 if (isEliminableCastPair(CI, opcode, Ty, TD))
435 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
436 /// InsertBefore instruction. This is specialized a bit to avoid inserting
437 /// casts that are known to not do anything...
439 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
440 Value *V, const Type *DestTy,
441 Instruction *InsertBefore) {
442 if (V->getType() == DestTy) return V;
443 if (Constant *C = dyn_cast<Constant>(V))
444 return ConstantExpr::getCast(opcode, C, DestTy);
446 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
449 // SimplifyCommutative - This performs a few simplifications for commutative
452 // 1. Order operands such that they are listed from right (least complex) to
453 // left (most complex). This puts constants before unary operators before
456 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
457 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
459 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
460 bool Changed = false;
461 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
462 Changed = !I.swapOperands();
464 if (!I.isAssociative()) return Changed;
465 Instruction::BinaryOps Opcode = I.getOpcode();
466 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
467 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
468 if (isa<Constant>(I.getOperand(1))) {
469 Constant *Folded = ConstantExpr::get(I.getOpcode(),
470 cast<Constant>(I.getOperand(1)),
471 cast<Constant>(Op->getOperand(1)));
472 I.setOperand(0, Op->getOperand(0));
473 I.setOperand(1, Folded);
475 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
476 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
477 isOnlyUse(Op) && isOnlyUse(Op1)) {
478 Constant *C1 = cast<Constant>(Op->getOperand(1));
479 Constant *C2 = cast<Constant>(Op1->getOperand(1));
481 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
482 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
483 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
487 I.setOperand(0, New);
488 I.setOperand(1, Folded);
495 /// SimplifyCompare - For a CmpInst this function just orders the operands
496 /// so that theyare listed from right (least complex) to left (most complex).
497 /// This puts constants before unary operators before binary operators.
498 bool InstCombiner::SimplifyCompare(CmpInst &I) {
499 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
502 // Compare instructions are not associative so there's nothing else we can do.
506 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
507 // if the LHS is a constant zero (which is the 'negate' form).
509 static inline Value *dyn_castNegVal(Value *V) {
510 if (BinaryOperator::isNeg(V))
511 return BinaryOperator::getNegArgument(V);
513 // Constants can be considered to be negated values if they can be folded.
514 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
515 return ConstantExpr::getNeg(C);
519 static inline Value *dyn_castNotVal(Value *V) {
520 if (BinaryOperator::isNot(V))
521 return BinaryOperator::getNotArgument(V);
523 // Constants can be considered to be not'ed values...
524 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
525 return ConstantExpr::getNot(C);
529 // dyn_castFoldableMul - If this value is a multiply that can be folded into
530 // other computations (because it has a constant operand), return the
531 // non-constant operand of the multiply, and set CST to point to the multiplier.
532 // Otherwise, return null.
534 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
535 if (V->hasOneUse() && V->getType()->isInteger())
536 if (Instruction *I = dyn_cast<Instruction>(V)) {
537 if (I->getOpcode() == Instruction::Mul)
538 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
539 return I->getOperand(0);
540 if (I->getOpcode() == Instruction::Shl)
541 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
542 // The multiplier is really 1 << CST.
543 Constant *One = ConstantInt::get(V->getType(), 1);
544 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
545 return I->getOperand(0);
551 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
552 /// expression, return it.
553 static User *dyn_castGetElementPtr(Value *V) {
554 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
555 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
556 if (CE->getOpcode() == Instruction::GetElementPtr)
557 return cast<User>(V);
561 /// AddOne - Add one to a ConstantInt
562 static ConstantInt *AddOne(ConstantInt *C) {
563 APInt Val(C->getValue());
564 return ConstantInt::get(++Val);
566 /// SubOne - Subtract one from a ConstantInt
567 static ConstantInt *SubOne(ConstantInt *C) {
568 APInt Val(C->getValue());
569 return ConstantInt::get(--Val);
571 /// Add - Add two ConstantInts together
572 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
573 return ConstantInt::get(C1->getValue() + C2->getValue());
575 /// And - Bitwise AND two ConstantInts together
576 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
577 return ConstantInt::get(C1->getValue() & C2->getValue());
579 /// Subtract - Subtract one ConstantInt from another
580 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
581 return ConstantInt::get(C1->getValue() - C2->getValue());
583 /// Multiply - Multiply two ConstantInts together
584 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
585 return ConstantInt::get(C1->getValue() * C2->getValue());
588 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
589 /// known to be either zero or one and return them in the KnownZero/KnownOne
590 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
592 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
593 /// we cannot optimize based on the assumption that it is zero without changing
594 /// it to be an explicit zero. If we don't change it to zero, other code could
595 /// optimized based on the contradictory assumption that it is non-zero.
596 /// Because instcombine aggressively folds operations with undef args anyway,
597 /// this won't lose us code quality.
598 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
599 APInt& KnownOne, unsigned Depth = 0) {
600 assert(V && "No Value?");
601 assert(Depth <= 6 && "Limit Search Depth");
602 uint32_t BitWidth = Mask.getBitWidth();
603 const IntegerType *VTy = cast<IntegerType>(V->getType());
604 assert(VTy->getBitWidth() == BitWidth &&
605 KnownZero.getBitWidth() == BitWidth &&
606 KnownOne.getBitWidth() == BitWidth &&
607 "VTy, Mask, KnownOne and KnownZero should have same BitWidth");
608 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
609 // We know all of the bits for a constant!
610 KnownOne = CI->getValue() & Mask;
611 KnownZero = ~KnownOne & Mask;
615 if (Depth == 6 || Mask == 0)
616 return; // Limit search depth.
618 Instruction *I = dyn_cast<Instruction>(V);
621 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
622 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
624 switch (I->getOpcode()) {
625 case Instruction::And: {
626 // If either the LHS or the RHS are Zero, the result is zero.
627 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
628 APInt Mask2(Mask & ~KnownZero);
629 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
630 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
631 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
633 // Output known-1 bits are only known if set in both the LHS & RHS.
634 KnownOne &= KnownOne2;
635 // Output known-0 are known to be clear if zero in either the LHS | RHS.
636 KnownZero |= KnownZero2;
639 case Instruction::Or: {
640 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
641 APInt Mask2(Mask & ~KnownOne);
642 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
643 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
644 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
646 // Output known-0 bits are only known if clear in both the LHS & RHS.
647 KnownZero &= KnownZero2;
648 // Output known-1 are known to be set if set in either the LHS | RHS.
649 KnownOne |= KnownOne2;
652 case Instruction::Xor: {
653 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
654 ComputeMaskedBits(I->getOperand(0), Mask, 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 known if clear or set in both the LHS & RHS.
659 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
660 // Output known-1 are known to be set if set in only one of the LHS, RHS.
661 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
662 KnownZero = KnownZeroOut;
665 case Instruction::Select:
666 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
667 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
668 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
669 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
671 // Only known if known in both the LHS and RHS.
672 KnownOne &= KnownOne2;
673 KnownZero &= KnownZero2;
675 case Instruction::FPTrunc:
676 case Instruction::FPExt:
677 case Instruction::FPToUI:
678 case Instruction::FPToSI:
679 case Instruction::SIToFP:
680 case Instruction::PtrToInt:
681 case Instruction::UIToFP:
682 case Instruction::IntToPtr:
683 return; // Can't work with floating point or pointers
684 case Instruction::Trunc: {
685 // All these have integer operands
686 uint32_t SrcBitWidth =
687 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
688 ComputeMaskedBits(I->getOperand(0), APInt(Mask).zext(SrcBitWidth),
689 KnownZero.zext(SrcBitWidth), KnownOne.zext(SrcBitWidth), Depth+1);
690 KnownZero.trunc(BitWidth);
691 KnownOne.trunc(BitWidth);
694 case Instruction::BitCast: {
695 const Type *SrcTy = I->getOperand(0)->getType();
696 if (SrcTy->isInteger()) {
697 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
702 case Instruction::ZExt: {
703 // Compute the bits in the result that are not present in the input.
704 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
705 uint32_t SrcBitWidth = SrcTy->getBitWidth();
706 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
708 ComputeMaskedBits(I->getOperand(0), APInt(Mask).trunc(SrcBitWidth),
709 KnownZero.trunc(SrcBitWidth), KnownOne.trunc(SrcBitWidth), Depth+1);
710 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
711 // The top bits are known to be zero.
712 KnownZero.zext(BitWidth);
713 KnownOne.zext(BitWidth);
714 KnownZero |= NewBits;
717 case Instruction::SExt: {
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();
721 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
723 ComputeMaskedBits(I->getOperand(0), APInt(Mask).trunc(SrcBitWidth),
724 KnownZero.trunc(SrcBitWidth), KnownOne.trunc(SrcBitWidth), Depth+1);
725 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
726 KnownZero.zext(BitWidth);
727 KnownOne.zext(BitWidth);
729 // If the sign bit of the input is known set or clear, then we know the
730 // top bits of the result.
731 APInt InSignBit(APInt::getSignBit(SrcTy->getBitWidth()));
732 InSignBit.zext(BitWidth);
733 if ((KnownZero & InSignBit) != 0) { // Input sign bit known zero
734 KnownZero |= NewBits;
735 KnownOne &= ~NewBits;
736 } else if ((KnownOne & InSignBit) != 0) { // Input sign bit known set
738 KnownZero &= ~NewBits;
739 } else { // Input sign bit unknown
740 KnownZero &= ~NewBits;
741 KnownOne &= ~NewBits;
745 case Instruction::Shl:
746 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
747 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
748 uint64_t ShiftAmt = SA->getZExtValue();
749 APInt Mask2(Mask.lshr(ShiftAmt));
750 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
751 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
752 KnownZero <<= ShiftAmt;
753 KnownOne <<= ShiftAmt;
754 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
758 case Instruction::LShr:
759 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
760 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
761 // Compute the new bits that are at the top now.
762 uint64_t ShiftAmt = SA->getZExtValue();
763 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
765 // Unsigned shift right.
766 APInt Mask2(Mask.shl(ShiftAmt));
767 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
768 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
769 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
770 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
771 KnownZero |= HighBits; // high bits known zero.
775 case Instruction::AShr:
776 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
777 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
778 // Compute the new bits that are at the top now.
779 uint64_t ShiftAmt = SA->getZExtValue();
780 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
782 // Signed shift right.
783 APInt Mask2(Mask.shl(ShiftAmt));
784 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
785 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
786 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
787 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
789 // Handle the sign bits and adjust to where it is now in the mask.
790 APInt SignBit(APInt::getSignBit(BitWidth).lshr(ShiftAmt));
792 if ((KnownZero & SignBit) != 0) { // New bits are known zero.
793 KnownZero |= HighBits;
794 } else if ((KnownOne & SignBit) != 0) { // New bits are known one.
795 KnownOne |= HighBits;
803 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
804 /// this predicate to simplify operations downstream. Mask is known to be zero
805 /// for bits that V cannot have.
806 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
807 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
808 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
809 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
810 return (KnownZero & Mask) == Mask;
813 /// ShrinkDemandedConstant - Check to see if the specified operand of the
814 /// specified instruction is a constant integer. If so, check to see if there
815 /// are any bits set in the constant that are not demanded. If so, shrink the
816 /// constant and return true.
817 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
819 assert(I && "No instruction?");
820 assert(OpNo < I->getNumOperands() && "Operand index too large");
822 // If the operand is not a constant integer, nothing to do.
823 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
824 if (!OpC) return false;
826 // If there are no bits set that aren't demanded, nothing to do.
827 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
828 if ((~Demanded & OpC->getValue()) == 0)
831 // This instruction is producing bits that are not demanded. Shrink the RHS.
832 Demanded &= OpC->getValue();
833 I->setOperand(OpNo, ConstantInt::get(Demanded));
837 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
838 // set of known zero and one bits, compute the maximum and minimum values that
839 // could have the specified known zero and known one bits, returning them in
841 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
842 const APInt& KnownZero,
843 const APInt& KnownOne,
844 APInt& Min, APInt& Max) {
845 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
846 assert(KnownZero.getBitWidth() == BitWidth &&
847 KnownOne.getBitWidth() == BitWidth &&
848 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
849 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
850 APInt UnknownBits = ~(KnownZero|KnownOne);
852 APInt SignBit(APInt::getSignBit(BitWidth));
854 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
855 // bit if it is unknown.
857 Max = KnownOne|UnknownBits;
859 if ((SignBit & UnknownBits) != 0) { // Sign bit is unknown
865 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
866 // a set of known zero and one bits, compute the maximum and minimum values that
867 // could have the specified known zero and known one bits, returning them in
869 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
870 const APInt& KnownZero,
871 const APInt& KnownOne,
874 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
875 assert(KnownZero.getBitWidth() == BitWidth &&
876 KnownOne.getBitWidth() == BitWidth &&
877 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
878 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
879 APInt UnknownBits = ~(KnownZero|KnownOne);
881 // The minimum value is when the unknown bits are all zeros.
883 // The maximum value is when the unknown bits are all ones.
884 Max = KnownOne|UnknownBits;
887 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
888 /// value based on the demanded bits. When this function is called, it is known
889 /// that only the bits set in DemandedMask of the result of V are ever used
890 /// downstream. Consequently, depending on the mask and V, it may be possible
891 /// to replace V with a constant or one of its operands. In such cases, this
892 /// function does the replacement and returns true. In all other cases, it
893 /// returns false after analyzing the expression and setting KnownOne and known
894 /// to be one in the expression. KnownZero contains all the bits that are known
895 /// to be zero in the expression. These are provided to potentially allow the
896 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
897 /// the expression. KnownOne and KnownZero always follow the invariant that
898 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
899 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
900 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
901 /// and KnownOne must all be the same.
902 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
903 APInt& KnownZero, APInt& KnownOne,
905 assert(V != 0 && "Null pointer of Value???");
906 assert(Depth <= 6 && "Limit Search Depth");
907 uint32_t BitWidth = DemandedMask.getBitWidth();
908 const IntegerType *VTy = cast<IntegerType>(V->getType());
909 assert(VTy->getBitWidth() == BitWidth &&
910 KnownZero.getBitWidth() == BitWidth &&
911 KnownOne.getBitWidth() == BitWidth &&
912 "Value *V, DemandedMask, KnownZero and KnownOne \
913 must have same BitWidth");
914 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
915 // We know all of the bits for a constant!
916 KnownOne = CI->getValue() & DemandedMask;
917 KnownZero = ~KnownOne & DemandedMask;
923 if (!V->hasOneUse()) { // Other users may use these bits.
924 if (Depth != 0) { // Not at the root.
925 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
926 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
929 // If this is the root being simplified, allow it to have multiple uses,
930 // just set the DemandedMask to all bits.
931 DemandedMask = APInt::getAllOnesValue(BitWidth);
932 } else if (DemandedMask == 0) { // Not demanding any bits from V.
933 if (V != UndefValue::get(VTy))
934 return UpdateValueUsesWith(V, UndefValue::get(VTy));
936 } else if (Depth == 6) { // Limit search depth.
940 Instruction *I = dyn_cast<Instruction>(V);
941 if (!I) return false; // Only analyze instructions.
943 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
944 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
945 switch (I->getOpcode()) {
947 case Instruction::And:
948 // If either the LHS or the RHS are Zero, the result is zero.
949 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
950 RHSKnownZero, RHSKnownOne, Depth+1))
952 assert((RHSKnownZero & RHSKnownOne) == 0 &&
953 "Bits known to be one AND zero?");
955 // If something is known zero on the RHS, the bits aren't demanded on the
957 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
958 LHSKnownZero, LHSKnownOne, Depth+1))
960 assert((LHSKnownZero & LHSKnownOne) == 0 &&
961 "Bits known to be one AND zero?");
963 // If all of the demanded bits are known 1 on one side, return the other.
964 // These bits cannot contribute to the result of the 'and'.
965 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
966 (DemandedMask & ~LHSKnownZero))
967 return UpdateValueUsesWith(I, I->getOperand(0));
968 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
969 (DemandedMask & ~RHSKnownZero))
970 return UpdateValueUsesWith(I, I->getOperand(1));
972 // If all of the demanded bits in the inputs are known zeros, return zero.
973 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
974 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
976 // If the RHS is a constant, see if we can simplify it.
977 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
978 return UpdateValueUsesWith(I, I);
980 // Output known-1 bits are only known if set in both the LHS & RHS.
981 RHSKnownOne &= LHSKnownOne;
982 // Output known-0 are known to be clear if zero in either the LHS | RHS.
983 RHSKnownZero |= LHSKnownZero;
985 case Instruction::Or:
986 // If either the LHS or the RHS are One, the result is One.
987 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
988 RHSKnownZero, RHSKnownOne, Depth+1))
990 assert((RHSKnownZero & RHSKnownOne) == 0 &&
991 "Bits known to be one AND zero?");
992 // If something is known one on the RHS, the bits aren't demanded on the
994 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
995 LHSKnownZero, LHSKnownOne, Depth+1))
997 assert((LHSKnownZero & LHSKnownOne) == 0 &&
998 "Bits known to be one AND zero?");
1000 // If all of the demanded bits are known zero on one side, return the other.
1001 // These bits cannot contribute to the result of the 'or'.
1002 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1003 (DemandedMask & ~LHSKnownOne))
1004 return UpdateValueUsesWith(I, I->getOperand(0));
1005 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1006 (DemandedMask & ~RHSKnownOne))
1007 return UpdateValueUsesWith(I, I->getOperand(1));
1009 // If all of the potentially set bits on one side are known to be set on
1010 // the other side, just use the 'other' side.
1011 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1012 (DemandedMask & (~RHSKnownZero)))
1013 return UpdateValueUsesWith(I, I->getOperand(0));
1014 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1015 (DemandedMask & (~LHSKnownZero)))
1016 return UpdateValueUsesWith(I, I->getOperand(1));
1018 // If the RHS is a constant, see if we can simplify it.
1019 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1020 return UpdateValueUsesWith(I, I);
1022 // Output known-0 bits are only known if clear in both the LHS & RHS.
1023 RHSKnownZero &= LHSKnownZero;
1024 // Output known-1 are known to be set if set in either the LHS | RHS.
1025 RHSKnownOne |= LHSKnownOne;
1027 case Instruction::Xor: {
1028 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1029 RHSKnownZero, RHSKnownOne, Depth+1))
1031 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1032 "Bits known to be one AND zero?");
1033 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1034 LHSKnownZero, LHSKnownOne, Depth+1))
1036 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1037 "Bits known to be one AND zero?");
1039 // If all of the demanded bits are known zero on one side, return the other.
1040 // These bits cannot contribute to the result of the 'xor'.
1041 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1042 return UpdateValueUsesWith(I, I->getOperand(0));
1043 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1044 return UpdateValueUsesWith(I, I->getOperand(1));
1046 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1047 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1048 (RHSKnownOne & LHSKnownOne);
1049 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1050 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1051 (RHSKnownOne & LHSKnownZero);
1053 // If all of the demanded bits are known to be zero on one side or the
1054 // other, turn this into an *inclusive* or.
1055 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1056 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1058 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1060 InsertNewInstBefore(Or, *I);
1061 return UpdateValueUsesWith(I, Or);
1064 // If all of the demanded bits on one side are known, and all of the set
1065 // bits on that side are also known to be set on the other side, turn this
1066 // into an AND, as we know the bits will be cleared.
1067 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1068 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1070 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1071 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1073 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1074 InsertNewInstBefore(And, *I);
1075 return UpdateValueUsesWith(I, And);
1079 // If the RHS is a constant, see if we can simplify it.
1080 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1081 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1082 return UpdateValueUsesWith(I, I);
1084 RHSKnownZero = KnownZeroOut;
1085 RHSKnownOne = KnownOneOut;
1088 case Instruction::Select:
1089 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1090 RHSKnownZero, RHSKnownOne, Depth+1))
1092 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1093 LHSKnownZero, LHSKnownOne, Depth+1))
1095 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1096 "Bits known to be one AND zero?");
1097 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1098 "Bits known to be one AND zero?");
1100 // If the operands are constants, see if we can simplify them.
1101 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1102 return UpdateValueUsesWith(I, I);
1103 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1104 return UpdateValueUsesWith(I, I);
1106 // Only known if known in both the LHS and RHS.
1107 RHSKnownOne &= LHSKnownOne;
1108 RHSKnownZero &= LHSKnownZero;
1110 case Instruction::Trunc: {
1112 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1113 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.zext(truncBf),
1114 RHSKnownZero.zext(truncBf), RHSKnownOne.zext(truncBf), Depth+1))
1116 DemandedMask.trunc(BitWidth);
1117 RHSKnownZero.trunc(BitWidth);
1118 RHSKnownOne.trunc(BitWidth);
1119 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1120 "Bits known to be one AND zero?");
1123 case Instruction::BitCast:
1124 if (!I->getOperand(0)->getType()->isInteger())
1127 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1128 RHSKnownZero, RHSKnownOne, Depth+1))
1130 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1131 "Bits known to be one AND zero?");
1133 case Instruction::ZExt: {
1134 // Compute the bits in the result that are not present in the input.
1135 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1136 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1137 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1139 DemandedMask &= SrcTy->getMask().zext(BitWidth);
1140 uint32_t zextBf = SrcTy->getBitWidth();
1141 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.trunc(zextBf),
1142 RHSKnownZero.trunc(zextBf), RHSKnownOne.trunc(zextBf), Depth+1))
1144 DemandedMask.zext(BitWidth);
1145 RHSKnownZero.zext(BitWidth);
1146 RHSKnownOne.zext(BitWidth);
1147 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1148 "Bits known to be one AND zero?");
1149 // The top bits are known to be zero.
1150 RHSKnownZero |= NewBits;
1153 case Instruction::SExt: {
1154 // Compute the bits in the result that are not present in the input.
1155 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1156 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1157 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1159 // Get the sign bit for the source type
1160 APInt InSignBit(APInt::getSignBit(SrcTy->getPrimitiveSizeInBits()));
1161 InSignBit.zext(BitWidth);
1162 APInt InputDemandedBits = DemandedMask &
1163 SrcTy->getMask().zext(BitWidth);
1165 // If any of the sign extended bits are demanded, we know that the sign
1167 if ((NewBits & DemandedMask) != 0)
1168 InputDemandedBits |= InSignBit;
1170 uint32_t sextBf = SrcTy->getBitWidth();
1171 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits.trunc(sextBf),
1172 RHSKnownZero.trunc(sextBf), RHSKnownOne.trunc(sextBf), Depth+1))
1174 InputDemandedBits.zext(BitWidth);
1175 RHSKnownZero.zext(BitWidth);
1176 RHSKnownOne.zext(BitWidth);
1177 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1178 "Bits known to be one AND zero?");
1180 // If the sign bit of the input is known set or clear, then we know the
1181 // top bits of the result.
1183 // If the input sign bit is known zero, or if the NewBits are not demanded
1184 // convert this into a zero extension.
1185 if ((RHSKnownZero & InSignBit) != 0 || (NewBits & ~DemandedMask) == NewBits)
1187 // Convert to ZExt cast
1188 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1189 return UpdateValueUsesWith(I, NewCast);
1190 } else if ((RHSKnownOne & InSignBit) != 0) { // Input sign bit known set
1191 RHSKnownOne |= NewBits;
1192 RHSKnownZero &= ~NewBits;
1193 } else { // Input sign bit unknown
1194 RHSKnownZero &= ~NewBits;
1195 RHSKnownOne &= ~NewBits;
1199 case Instruction::Add: {
1200 // Figure out what the input bits are. If the top bits of the and result
1201 // are not demanded, then the add doesn't demand them from its input
1203 uint32_t NLZ = DemandedMask.countLeadingZeros();
1205 // If there is a constant on the RHS, there are a variety of xformations
1207 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1208 // If null, this should be simplified elsewhere. Some of the xforms here
1209 // won't work if the RHS is zero.
1213 // If the top bit of the output is demanded, demand everything from the
1214 // input. Otherwise, we demand all the input bits except NLZ top bits.
1215 APInt InDemandedBits(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1217 // Find information about known zero/one bits in the input.
1218 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1219 LHSKnownZero, LHSKnownOne, Depth+1))
1222 // If the RHS of the add has bits set that can't affect the input, reduce
1224 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1225 return UpdateValueUsesWith(I, I);
1227 // Avoid excess work.
1228 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1231 // Turn it into OR if input bits are zero.
1232 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1234 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1236 InsertNewInstBefore(Or, *I);
1237 return UpdateValueUsesWith(I, Or);
1240 // We can say something about the output known-zero and known-one bits,
1241 // depending on potential carries from the input constant and the
1242 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1243 // bits set and the RHS constant is 0x01001, then we know we have a known
1244 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1246 // To compute this, we first compute the potential carry bits. These are
1247 // the bits which may be modified. I'm not aware of a better way to do
1249 APInt RHSVal(RHS->getValue());
1251 bool CarryIn = false;
1252 APInt CarryBits(BitWidth, 0);
1253 const uint64_t *LHSKnownZeroRawVal = LHSKnownZero.getRawData(),
1254 *RHSRawVal = RHSVal.getRawData();
1255 for (uint32_t i = 0; i != RHSVal.getNumWords(); ++i) {
1256 uint64_t AddVal = ~LHSKnownZeroRawVal[i] + RHSRawVal[i],
1257 XorVal = ~LHSKnownZeroRawVal[i] ^ RHSRawVal[i];
1258 uint64_t WordCarryBits = AddVal ^ XorVal + CarryIn;
1259 if (AddVal < RHSRawVal[i])
1263 CarryBits.setWordToValue(i, WordCarryBits);
1266 // Now that we know which bits have carries, compute the known-1/0 sets.
1268 // Bits are known one if they are known zero in one operand and one in the
1269 // other, and there is no input carry.
1270 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1271 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1273 // Bits are known zero if they are known zero in both operands and there
1274 // is no input carry.
1275 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1277 // If the high-bits of this ADD are not demanded, then it does not demand
1278 // the high bits of its LHS or RHS.
1279 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1280 // Right fill the mask of bits for this ADD to demand the most
1281 // significant bit and all those below it.
1282 APInt DemandedFromOps = APInt::getAllOnesValue(BitWidth).lshr(NLZ);
1283 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1284 LHSKnownZero, LHSKnownOne, Depth+1))
1286 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1287 LHSKnownZero, LHSKnownOne, Depth+1))
1293 case Instruction::Sub:
1294 // If the high-bits of this SUB are not demanded, then it does not demand
1295 // the high bits of its LHS or RHS.
1296 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1297 // Right fill the mask of bits for this SUB to demand the most
1298 // significant bit and all those below it.
1299 unsigned NLZ = DemandedMask.countLeadingZeros();
1300 APInt DemandedFromOps(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1302 LHSKnownZero, LHSKnownOne, Depth+1))
1304 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1305 LHSKnownZero, LHSKnownOne, Depth+1))
1309 case Instruction::Shl:
1310 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1311 uint64_t ShiftAmt = SA->getZExtValue();
1312 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.lshr(ShiftAmt),
1313 RHSKnownZero, RHSKnownOne, Depth+1))
1315 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1316 "Bits known to be one AND zero?");
1317 RHSKnownZero <<= ShiftAmt;
1318 RHSKnownOne <<= ShiftAmt;
1319 // low bits known zero.
1321 RHSKnownZero |= APInt::getAllOnesValue(ShiftAmt).zextOrCopy(BitWidth);
1324 case Instruction::LShr:
1325 // For a logical shift right
1326 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1327 unsigned ShiftAmt = SA->getZExtValue();
1329 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
1330 // Unsigned shift right.
1331 if (SimplifyDemandedBits(I->getOperand(0),
1332 (DemandedMask.shl(ShiftAmt)) & TypeMask,
1333 RHSKnownZero, RHSKnownOne, Depth+1))
1335 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1336 "Bits known to be one AND zero?");
1337 RHSKnownZero &= TypeMask;
1338 RHSKnownOne &= TypeMask;
1339 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1340 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1342 // Compute the new bits that are at the top now.
1343 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(
1344 BitWidth - ShiftAmt));
1345 RHSKnownZero |= HighBits; // high bits known zero.
1349 case Instruction::AShr:
1350 // If this is an arithmetic shift right and only the low-bit is set, we can
1351 // always convert this into a logical shr, even if the shift amount is
1352 // variable. The low bit of the shift cannot be an input sign bit unless
1353 // the shift amount is >= the size of the datatype, which is undefined.
1354 if (DemandedMask == 1) {
1355 // Perform the logical shift right.
1356 Value *NewVal = BinaryOperator::createLShr(
1357 I->getOperand(0), I->getOperand(1), I->getName());
1358 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1359 return UpdateValueUsesWith(I, NewVal);
1362 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1363 unsigned ShiftAmt = SA->getZExtValue();
1365 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
1366 // Signed shift right.
1367 if (SimplifyDemandedBits(I->getOperand(0),
1368 (DemandedMask.shl(ShiftAmt)) & TypeMask,
1369 RHSKnownZero, RHSKnownOne, Depth+1))
1371 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1372 "Bits known to be one AND zero?");
1373 // Compute the new bits that are at the top now.
1374 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(BitWidth - ShiftAmt));
1375 RHSKnownZero &= TypeMask;
1376 RHSKnownOne &= TypeMask;
1377 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1378 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1380 // Handle the sign bits.
1381 APInt SignBit(APInt::getSignBit(BitWidth));
1382 // Adjust to where it is now in the mask.
1383 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1385 // If the input sign bit is known to be zero, or if none of the top bits
1386 // are demanded, turn this into an unsigned shift right.
1387 if ((RHSKnownZero & SignBit) != 0 ||
1388 (HighBits & ~DemandedMask) == HighBits) {
1389 // Perform the logical shift right.
1390 Value *NewVal = BinaryOperator::createLShr(
1391 I->getOperand(0), SA, I->getName());
1392 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1393 return UpdateValueUsesWith(I, NewVal);
1394 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1395 RHSKnownOne |= HighBits;
1401 // If the client is only demanding bits that we know, return the known
1403 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1404 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1409 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1410 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1411 /// actually used by the caller. This method analyzes which elements of the
1412 /// operand are undef and returns that information in UndefElts.
1414 /// If the information about demanded elements can be used to simplify the
1415 /// operation, the operation is simplified, then the resultant value is
1416 /// returned. This returns null if no change was made.
1417 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1418 uint64_t &UndefElts,
1420 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1421 assert(VWidth <= 64 && "Vector too wide to analyze!");
1422 uint64_t EltMask = ~0ULL >> (64-VWidth);
1423 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1424 "Invalid DemandedElts!");
1426 if (isa<UndefValue>(V)) {
1427 // If the entire vector is undefined, just return this info.
1428 UndefElts = EltMask;
1430 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1431 UndefElts = EltMask;
1432 return UndefValue::get(V->getType());
1436 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1437 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1438 Constant *Undef = UndefValue::get(EltTy);
1440 std::vector<Constant*> Elts;
1441 for (unsigned i = 0; i != VWidth; ++i)
1442 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1443 Elts.push_back(Undef);
1444 UndefElts |= (1ULL << i);
1445 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1446 Elts.push_back(Undef);
1447 UndefElts |= (1ULL << i);
1448 } else { // Otherwise, defined.
1449 Elts.push_back(CP->getOperand(i));
1452 // If we changed the constant, return it.
1453 Constant *NewCP = ConstantVector::get(Elts);
1454 return NewCP != CP ? NewCP : 0;
1455 } else if (isa<ConstantAggregateZero>(V)) {
1456 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1458 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1459 Constant *Zero = Constant::getNullValue(EltTy);
1460 Constant *Undef = UndefValue::get(EltTy);
1461 std::vector<Constant*> Elts;
1462 for (unsigned i = 0; i != VWidth; ++i)
1463 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1464 UndefElts = DemandedElts ^ EltMask;
1465 return ConstantVector::get(Elts);
1468 if (!V->hasOneUse()) { // Other users may use these bits.
1469 if (Depth != 0) { // Not at the root.
1470 // TODO: Just compute the UndefElts information recursively.
1474 } else if (Depth == 10) { // Limit search depth.
1478 Instruction *I = dyn_cast<Instruction>(V);
1479 if (!I) return false; // Only analyze instructions.
1481 bool MadeChange = false;
1482 uint64_t UndefElts2;
1484 switch (I->getOpcode()) {
1487 case Instruction::InsertElement: {
1488 // If this is a variable index, we don't know which element it overwrites.
1489 // demand exactly the same input as we produce.
1490 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1492 // Note that we can't propagate undef elt info, because we don't know
1493 // which elt is getting updated.
1494 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1495 UndefElts2, Depth+1);
1496 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1500 // If this is inserting an element that isn't demanded, remove this
1502 unsigned IdxNo = Idx->getZExtValue();
1503 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1504 return AddSoonDeadInstToWorklist(*I, 0);
1506 // Otherwise, the element inserted overwrites whatever was there, so the
1507 // input demanded set is simpler than the output set.
1508 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1509 DemandedElts & ~(1ULL << IdxNo),
1510 UndefElts, Depth+1);
1511 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1513 // The inserted element is defined.
1514 UndefElts |= 1ULL << IdxNo;
1518 case Instruction::And:
1519 case Instruction::Or:
1520 case Instruction::Xor:
1521 case Instruction::Add:
1522 case Instruction::Sub:
1523 case Instruction::Mul:
1524 // div/rem demand all inputs, because they don't want divide by zero.
1525 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1526 UndefElts, Depth+1);
1527 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1528 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1529 UndefElts2, Depth+1);
1530 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1532 // Output elements are undefined if both are undefined. Consider things
1533 // like undef&0. The result is known zero, not undef.
1534 UndefElts &= UndefElts2;
1537 case Instruction::Call: {
1538 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1540 switch (II->getIntrinsicID()) {
1543 // Binary vector operations that work column-wise. A dest element is a
1544 // function of the corresponding input elements from the two inputs.
1545 case Intrinsic::x86_sse_sub_ss:
1546 case Intrinsic::x86_sse_mul_ss:
1547 case Intrinsic::x86_sse_min_ss:
1548 case Intrinsic::x86_sse_max_ss:
1549 case Intrinsic::x86_sse2_sub_sd:
1550 case Intrinsic::x86_sse2_mul_sd:
1551 case Intrinsic::x86_sse2_min_sd:
1552 case Intrinsic::x86_sse2_max_sd:
1553 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1554 UndefElts, Depth+1);
1555 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1556 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1557 UndefElts2, Depth+1);
1558 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1560 // If only the low elt is demanded and this is a scalarizable intrinsic,
1561 // scalarize it now.
1562 if (DemandedElts == 1) {
1563 switch (II->getIntrinsicID()) {
1565 case Intrinsic::x86_sse_sub_ss:
1566 case Intrinsic::x86_sse_mul_ss:
1567 case Intrinsic::x86_sse2_sub_sd:
1568 case Intrinsic::x86_sse2_mul_sd:
1569 // TODO: Lower MIN/MAX/ABS/etc
1570 Value *LHS = II->getOperand(1);
1571 Value *RHS = II->getOperand(2);
1572 // Extract the element as scalars.
1573 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1574 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1576 switch (II->getIntrinsicID()) {
1577 default: assert(0 && "Case stmts out of sync!");
1578 case Intrinsic::x86_sse_sub_ss:
1579 case Intrinsic::x86_sse2_sub_sd:
1580 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1581 II->getName()), *II);
1583 case Intrinsic::x86_sse_mul_ss:
1584 case Intrinsic::x86_sse2_mul_sd:
1585 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1586 II->getName()), *II);
1591 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1593 InsertNewInstBefore(New, *II);
1594 AddSoonDeadInstToWorklist(*II, 0);
1599 // Output elements are undefined if both are undefined. Consider things
1600 // like undef&0. The result is known zero, not undef.
1601 UndefElts &= UndefElts2;
1607 return MadeChange ? I : 0;
1610 /// @returns true if the specified compare instruction is
1611 /// true when both operands are equal...
1612 /// @brief Determine if the ICmpInst returns true if both operands are equal
1613 static bool isTrueWhenEqual(ICmpInst &ICI) {
1614 ICmpInst::Predicate pred = ICI.getPredicate();
1615 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1616 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1617 pred == ICmpInst::ICMP_SLE;
1620 /// AssociativeOpt - Perform an optimization on an associative operator. This
1621 /// function is designed to check a chain of associative operators for a
1622 /// potential to apply a certain optimization. Since the optimization may be
1623 /// applicable if the expression was reassociated, this checks the chain, then
1624 /// reassociates the expression as necessary to expose the optimization
1625 /// opportunity. This makes use of a special Functor, which must define
1626 /// 'shouldApply' and 'apply' methods.
1628 template<typename Functor>
1629 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1630 unsigned Opcode = Root.getOpcode();
1631 Value *LHS = Root.getOperand(0);
1633 // Quick check, see if the immediate LHS matches...
1634 if (F.shouldApply(LHS))
1635 return F.apply(Root);
1637 // Otherwise, if the LHS is not of the same opcode as the root, return.
1638 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1639 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1640 // Should we apply this transform to the RHS?
1641 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1643 // If not to the RHS, check to see if we should apply to the LHS...
1644 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1645 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1649 // If the functor wants to apply the optimization to the RHS of LHSI,
1650 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1652 BasicBlock *BB = Root.getParent();
1654 // Now all of the instructions are in the current basic block, go ahead
1655 // and perform the reassociation.
1656 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1658 // First move the selected RHS to the LHS of the root...
1659 Root.setOperand(0, LHSI->getOperand(1));
1661 // Make what used to be the LHS of the root be the user of the root...
1662 Value *ExtraOperand = TmpLHSI->getOperand(1);
1663 if (&Root == TmpLHSI) {
1664 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1667 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1668 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1669 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1670 BasicBlock::iterator ARI = &Root; ++ARI;
1671 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1674 // Now propagate the ExtraOperand down the chain of instructions until we
1676 while (TmpLHSI != LHSI) {
1677 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1678 // Move the instruction to immediately before the chain we are
1679 // constructing to avoid breaking dominance properties.
1680 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1681 BB->getInstList().insert(ARI, NextLHSI);
1684 Value *NextOp = NextLHSI->getOperand(1);
1685 NextLHSI->setOperand(1, ExtraOperand);
1687 ExtraOperand = NextOp;
1690 // Now that the instructions are reassociated, have the functor perform
1691 // the transformation...
1692 return F.apply(Root);
1695 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1701 // AddRHS - Implements: X + X --> X << 1
1704 AddRHS(Value *rhs) : RHS(rhs) {}
1705 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1706 Instruction *apply(BinaryOperator &Add) const {
1707 return BinaryOperator::createShl(Add.getOperand(0),
1708 ConstantInt::get(Add.getType(), 1));
1712 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1714 struct AddMaskingAnd {
1716 AddMaskingAnd(Constant *c) : C2(c) {}
1717 bool shouldApply(Value *LHS) const {
1719 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1720 ConstantExpr::getAnd(C1, C2)->isNullValue();
1722 Instruction *apply(BinaryOperator &Add) const {
1723 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1727 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1729 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1730 if (Constant *SOC = dyn_cast<Constant>(SO))
1731 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1733 return IC->InsertNewInstBefore(CastInst::create(
1734 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1737 // Figure out if the constant is the left or the right argument.
1738 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1739 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1741 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1743 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1744 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1747 Value *Op0 = SO, *Op1 = ConstOperand;
1749 std::swap(Op0, Op1);
1751 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1752 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1753 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1754 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1755 SO->getName()+".cmp");
1757 assert(0 && "Unknown binary instruction type!");
1760 return IC->InsertNewInstBefore(New, I);
1763 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1764 // constant as the other operand, try to fold the binary operator into the
1765 // select arguments. This also works for Cast instructions, which obviously do
1766 // not have a second operand.
1767 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1769 // Don't modify shared select instructions
1770 if (!SI->hasOneUse()) return 0;
1771 Value *TV = SI->getOperand(1);
1772 Value *FV = SI->getOperand(2);
1774 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1775 // Bool selects with constant operands can be folded to logical ops.
1776 if (SI->getType() == Type::Int1Ty) return 0;
1778 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1779 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1781 return new SelectInst(SI->getCondition(), SelectTrueVal,
1788 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1789 /// node as operand #0, see if we can fold the instruction into the PHI (which
1790 /// is only possible if all operands to the PHI are constants).
1791 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1792 PHINode *PN = cast<PHINode>(I.getOperand(0));
1793 unsigned NumPHIValues = PN->getNumIncomingValues();
1794 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1796 // Check to see if all of the operands of the PHI are constants. If there is
1797 // one non-constant value, remember the BB it is. If there is more than one
1798 // or if *it* is a PHI, bail out.
1799 BasicBlock *NonConstBB = 0;
1800 for (unsigned i = 0; i != NumPHIValues; ++i)
1801 if (!isa<Constant>(PN->getIncomingValue(i))) {
1802 if (NonConstBB) return 0; // More than one non-const value.
1803 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1804 NonConstBB = PN->getIncomingBlock(i);
1806 // If the incoming non-constant value is in I's block, we have an infinite
1808 if (NonConstBB == I.getParent())
1812 // If there is exactly one non-constant value, we can insert a copy of the
1813 // operation in that block. However, if this is a critical edge, we would be
1814 // inserting the computation one some other paths (e.g. inside a loop). Only
1815 // do this if the pred block is unconditionally branching into the phi block.
1817 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1818 if (!BI || !BI->isUnconditional()) return 0;
1821 // Okay, we can do the transformation: create the new PHI node.
1822 PHINode *NewPN = new PHINode(I.getType(), "");
1823 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1824 InsertNewInstBefore(NewPN, *PN);
1825 NewPN->takeName(PN);
1827 // Next, add all of the operands to the PHI.
1828 if (I.getNumOperands() == 2) {
1829 Constant *C = cast<Constant>(I.getOperand(1));
1830 for (unsigned i = 0; i != NumPHIValues; ++i) {
1832 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1833 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1834 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1836 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1838 assert(PN->getIncomingBlock(i) == NonConstBB);
1839 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1840 InV = BinaryOperator::create(BO->getOpcode(),
1841 PN->getIncomingValue(i), C, "phitmp",
1842 NonConstBB->getTerminator());
1843 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1844 InV = CmpInst::create(CI->getOpcode(),
1846 PN->getIncomingValue(i), C, "phitmp",
1847 NonConstBB->getTerminator());
1849 assert(0 && "Unknown binop!");
1851 AddToWorkList(cast<Instruction>(InV));
1853 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1856 CastInst *CI = cast<CastInst>(&I);
1857 const Type *RetTy = CI->getType();
1858 for (unsigned i = 0; i != NumPHIValues; ++i) {
1860 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1861 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1863 assert(PN->getIncomingBlock(i) == NonConstBB);
1864 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1865 I.getType(), "phitmp",
1866 NonConstBB->getTerminator());
1867 AddToWorkList(cast<Instruction>(InV));
1869 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1872 return ReplaceInstUsesWith(I, NewPN);
1875 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1876 bool Changed = SimplifyCommutative(I);
1877 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1879 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1880 // X + undef -> undef
1881 if (isa<UndefValue>(RHS))
1882 return ReplaceInstUsesWith(I, RHS);
1885 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1886 if (RHSC->isNullValue())
1887 return ReplaceInstUsesWith(I, LHS);
1888 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1889 if (CFP->isExactlyValue(-0.0))
1890 return ReplaceInstUsesWith(I, LHS);
1893 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1894 // X + (signbit) --> X ^ signbit
1895 APInt Val(CI->getValue());
1896 unsigned BitWidth = Val.getBitWidth();
1897 if (Val == APInt::getSignBit(BitWidth))
1898 return BinaryOperator::createXor(LHS, RHS);
1900 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1901 // (X & 254)+1 -> (X&254)|1
1902 if (!isa<VectorType>(I.getType())) {
1903 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1904 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1905 KnownZero, KnownOne))
1910 if (isa<PHINode>(LHS))
1911 if (Instruction *NV = FoldOpIntoPhi(I))
1914 ConstantInt *XorRHS = 0;
1916 if (isa<ConstantInt>(RHSC) &&
1917 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1918 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1919 APInt RHSVal(cast<ConstantInt>(RHSC)->getValue());
1921 unsigned Size = TySizeBits / 2;
1922 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1923 APInt CFF80Val(-C0080Val);
1925 if (TySizeBits > Size) {
1926 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1927 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1928 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1929 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1930 // This is a sign extend if the top bits are known zero.
1931 APInt Mask(APInt::getAllOnesValue(TySizeBits));
1933 if (!MaskedValueIsZero(XorLHS, Mask))
1934 Size = 0; // Not a sign ext, but can't be any others either.
1939 C0080Val = APIntOps::lshr(C0080Val, Size);
1940 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1941 } while (Size >= 1);
1944 const Type *MiddleType = IntegerType::get(Size);
1945 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1946 InsertNewInstBefore(NewTrunc, I);
1947 return new SExtInst(NewTrunc, I.getType());
1953 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1954 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1956 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1957 if (RHSI->getOpcode() == Instruction::Sub)
1958 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1959 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1961 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1962 if (LHSI->getOpcode() == Instruction::Sub)
1963 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1964 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1969 if (Value *V = dyn_castNegVal(LHS))
1970 return BinaryOperator::createSub(RHS, V);
1973 if (!isa<Constant>(RHS))
1974 if (Value *V = dyn_castNegVal(RHS))
1975 return BinaryOperator::createSub(LHS, V);
1979 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1980 if (X == RHS) // X*C + X --> X * (C+1)
1981 return BinaryOperator::createMul(RHS, AddOne(C2));
1983 // X*C1 + X*C2 --> X * (C1+C2)
1985 if (X == dyn_castFoldableMul(RHS, C1))
1986 return BinaryOperator::createMul(X, Add(C1, C2));
1989 // X + X*C --> X * (C+1)
1990 if (dyn_castFoldableMul(RHS, C2) == LHS)
1991 return BinaryOperator::createMul(LHS, AddOne(C2));
1993 // X + ~X --> -1 since ~X = -X-1
1994 if (dyn_castNotVal(LHS) == RHS ||
1995 dyn_castNotVal(RHS) == LHS)
1996 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1999 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2000 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2001 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2004 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2006 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2007 return BinaryOperator::createSub(SubOne(CRHS), X);
2009 // (X & FF00) + xx00 -> (X+xx00) & FF00
2010 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2011 Constant *Anded = And(CRHS, C2);
2012 if (Anded == CRHS) {
2013 // See if all bits from the first bit set in the Add RHS up are included
2014 // in the mask. First, get the rightmost bit.
2015 APInt AddRHSV(CRHS->getValue());
2017 // Form a mask of all bits from the lowest bit added through the top.
2018 APInt AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
2019 AddRHSHighBits &= C2->getType()->getMask();
2021 // See if the and mask includes all of these bits.
2022 APInt AddRHSHighBitsAnd = AddRHSHighBits & C2->getValue();
2024 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2025 // Okay, the xform is safe. Insert the new add pronto.
2026 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2027 LHS->getName()), I);
2028 return BinaryOperator::createAnd(NewAdd, C2);
2033 // Try to fold constant add into select arguments.
2034 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2035 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2039 // add (cast *A to intptrtype) B ->
2040 // cast (GEP (cast *A to sbyte*) B) ->
2043 CastInst *CI = dyn_cast<CastInst>(LHS);
2046 CI = dyn_cast<CastInst>(RHS);
2049 if (CI && CI->getType()->isSized() &&
2050 (CI->getType()->getPrimitiveSizeInBits() ==
2051 TD->getIntPtrType()->getPrimitiveSizeInBits())
2052 && isa<PointerType>(CI->getOperand(0)->getType())) {
2053 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2054 PointerType::get(Type::Int8Ty), I);
2055 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2056 return new PtrToIntInst(I2, CI->getType());
2060 return Changed ? &I : 0;
2063 // isSignBit - Return true if the value represented by the constant only has the
2064 // highest order bit set.
2065 static bool isSignBit(ConstantInt *CI) {
2066 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
2067 return CI->getValue() == APInt::getSignBit(NumBits);
2070 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2071 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2073 if (Op0 == Op1) // sub X, X -> 0
2074 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2076 // If this is a 'B = x-(-A)', change to B = x+A...
2077 if (Value *V = dyn_castNegVal(Op1))
2078 return BinaryOperator::createAdd(Op0, V);
2080 if (isa<UndefValue>(Op0))
2081 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2082 if (isa<UndefValue>(Op1))
2083 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2085 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2086 // Replace (-1 - A) with (~A)...
2087 if (C->isAllOnesValue())
2088 return BinaryOperator::createNot(Op1);
2090 // C - ~X == X + (1+C)
2092 if (match(Op1, m_Not(m_Value(X))))
2093 return BinaryOperator::createAdd(X, AddOne(C));
2095 // -(X >>u 31) -> (X >>s 31)
2096 // -(X >>s 31) -> (X >>u 31)
2097 if (C->isNullValue()) {
2098 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2099 if (SI->getOpcode() == Instruction::LShr) {
2100 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2101 // Check to see if we are shifting out everything but the sign bit.
2102 if (CU->getZExtValue() ==
2103 SI->getType()->getPrimitiveSizeInBits()-1) {
2104 // Ok, the transformation is safe. Insert AShr.
2105 return BinaryOperator::create(Instruction::AShr,
2106 SI->getOperand(0), CU, SI->getName());
2110 else if (SI->getOpcode() == Instruction::AShr) {
2111 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2112 // Check to see if we are shifting out everything but the sign bit.
2113 if (CU->getZExtValue() ==
2114 SI->getType()->getPrimitiveSizeInBits()-1) {
2115 // Ok, the transformation is safe. Insert LShr.
2116 return BinaryOperator::createLShr(
2117 SI->getOperand(0), CU, SI->getName());
2123 // Try to fold constant sub into select arguments.
2124 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2125 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2128 if (isa<PHINode>(Op0))
2129 if (Instruction *NV = FoldOpIntoPhi(I))
2133 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2134 if (Op1I->getOpcode() == Instruction::Add &&
2135 !Op0->getType()->isFPOrFPVector()) {
2136 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2137 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2138 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2139 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2140 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2141 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2142 // C1-(X+C2) --> (C1-C2)-X
2143 return BinaryOperator::createSub(Subtract(CI1, CI2),
2144 Op1I->getOperand(0));
2148 if (Op1I->hasOneUse()) {
2149 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2150 // is not used by anyone else...
2152 if (Op1I->getOpcode() == Instruction::Sub &&
2153 !Op1I->getType()->isFPOrFPVector()) {
2154 // Swap the two operands of the subexpr...
2155 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2156 Op1I->setOperand(0, IIOp1);
2157 Op1I->setOperand(1, IIOp0);
2159 // Create the new top level add instruction...
2160 return BinaryOperator::createAdd(Op0, Op1);
2163 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2165 if (Op1I->getOpcode() == Instruction::And &&
2166 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2167 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2170 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2171 return BinaryOperator::createAnd(Op0, NewNot);
2174 // 0 - (X sdiv C) -> (X sdiv -C)
2175 if (Op1I->getOpcode() == Instruction::SDiv)
2176 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2177 if (CSI->isNullValue())
2178 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2179 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2180 ConstantExpr::getNeg(DivRHS));
2182 // X - X*C --> X * (1-C)
2183 ConstantInt *C2 = 0;
2184 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2185 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2186 return BinaryOperator::createMul(Op0, CP1);
2191 if (!Op0->getType()->isFPOrFPVector())
2192 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2193 if (Op0I->getOpcode() == Instruction::Add) {
2194 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2195 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2196 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2197 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2198 } else if (Op0I->getOpcode() == Instruction::Sub) {
2199 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2200 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2204 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2205 if (X == Op1) // X*C - X --> X * (C-1)
2206 return BinaryOperator::createMul(Op1, SubOne(C1));
2208 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2209 if (X == dyn_castFoldableMul(Op1, C2))
2210 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2215 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2216 /// really just returns true if the most significant (sign) bit is set.
2217 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2219 case ICmpInst::ICMP_SLT:
2220 // True if LHS s< RHS and RHS == 0
2221 return RHS->isNullValue();
2222 case ICmpInst::ICMP_SLE:
2223 // True if LHS s<= RHS and RHS == -1
2224 return RHS->isAllOnesValue();
2225 case ICmpInst::ICMP_UGE:
2226 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2227 return RHS->getValue() ==
2228 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2229 case ICmpInst::ICMP_UGT:
2230 // True if LHS u> RHS and RHS == high-bit-mask - 1
2231 return RHS->getValue() ==
2232 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2238 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2239 bool Changed = SimplifyCommutative(I);
2240 Value *Op0 = I.getOperand(0);
2242 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2243 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2245 // Simplify mul instructions with a constant RHS...
2246 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2247 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2249 // ((X << C1)*C2) == (X * (C2 << C1))
2250 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2251 if (SI->getOpcode() == Instruction::Shl)
2252 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2253 return BinaryOperator::createMul(SI->getOperand(0),
2254 ConstantExpr::getShl(CI, ShOp));
2256 if (CI->isNullValue())
2257 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2258 if (CI->equalsInt(1)) // X * 1 == X
2259 return ReplaceInstUsesWith(I, Op0);
2260 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2261 return BinaryOperator::createNeg(Op0, I.getName());
2263 APInt Val(cast<ConstantInt>(CI)->getValue());
2264 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2265 return BinaryOperator::createShl(Op0,
2266 ConstantInt::get(Op0->getType(), Val.logBase2()));
2268 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2269 if (Op1F->isNullValue())
2270 return ReplaceInstUsesWith(I, Op1);
2272 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2273 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2274 if (Op1F->getValue() == 1.0)
2275 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2278 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2279 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2280 isa<ConstantInt>(Op0I->getOperand(1))) {
2281 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2282 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2284 InsertNewInstBefore(Add, I);
2285 Value *C1C2 = ConstantExpr::getMul(Op1,
2286 cast<Constant>(Op0I->getOperand(1)));
2287 return BinaryOperator::createAdd(Add, C1C2);
2291 // Try to fold constant mul into select arguments.
2292 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2293 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2296 if (isa<PHINode>(Op0))
2297 if (Instruction *NV = FoldOpIntoPhi(I))
2301 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2302 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2303 return BinaryOperator::createMul(Op0v, Op1v);
2305 // If one of the operands of the multiply is a cast from a boolean value, then
2306 // we know the bool is either zero or one, so this is a 'masking' multiply.
2307 // See if we can simplify things based on how the boolean was originally
2309 CastInst *BoolCast = 0;
2310 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2311 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2314 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2315 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2318 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2319 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2320 const Type *SCOpTy = SCIOp0->getType();
2322 // If the icmp is true iff the sign bit of X is set, then convert this
2323 // multiply into a shift/and combination.
2324 if (isa<ConstantInt>(SCIOp1) &&
2325 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2326 // Shift the X value right to turn it into "all signbits".
2327 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2328 SCOpTy->getPrimitiveSizeInBits()-1);
2330 InsertNewInstBefore(
2331 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2332 BoolCast->getOperand(0)->getName()+
2335 // If the multiply type is not the same as the source type, sign extend
2336 // or truncate to the multiply type.
2337 if (I.getType() != V->getType()) {
2338 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2339 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2340 Instruction::CastOps opcode =
2341 (SrcBits == DstBits ? Instruction::BitCast :
2342 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2343 V = InsertCastBefore(opcode, V, I.getType(), I);
2346 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2347 return BinaryOperator::createAnd(V, OtherOp);
2352 return Changed ? &I : 0;
2355 /// This function implements the transforms on div instructions that work
2356 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2357 /// used by the visitors to those instructions.
2358 /// @brief Transforms common to all three div instructions
2359 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2360 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2363 if (isa<UndefValue>(Op0))
2364 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2366 // X / undef -> undef
2367 if (isa<UndefValue>(Op1))
2368 return ReplaceInstUsesWith(I, Op1);
2370 // Handle cases involving: div X, (select Cond, Y, Z)
2371 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2372 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2373 // same basic block, then we replace the select with Y, and the condition
2374 // of the select with false (if the cond value is in the same BB). If the
2375 // select has uses other than the div, this allows them to be simplified
2376 // also. Note that div X, Y is just as good as div X, 0 (undef)
2377 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2378 if (ST->isNullValue()) {
2379 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2380 if (CondI && CondI->getParent() == I.getParent())
2381 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2382 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2383 I.setOperand(1, SI->getOperand(2));
2385 UpdateValueUsesWith(SI, SI->getOperand(2));
2389 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2390 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2391 if (ST->isNullValue()) {
2392 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2393 if (CondI && CondI->getParent() == I.getParent())
2394 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2395 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2396 I.setOperand(1, SI->getOperand(1));
2398 UpdateValueUsesWith(SI, SI->getOperand(1));
2406 /// This function implements the transforms common to both integer division
2407 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2408 /// division instructions.
2409 /// @brief Common integer divide transforms
2410 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2411 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2413 if (Instruction *Common = commonDivTransforms(I))
2416 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2418 if (RHS->equalsInt(1))
2419 return ReplaceInstUsesWith(I, Op0);
2421 // (X / C1) / C2 -> X / (C1*C2)
2422 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2423 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2424 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2425 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2426 Multiply(RHS, LHSRHS));
2429 if (!RHS->isZero()) { // avoid X udiv 0
2430 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2431 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2433 if (isa<PHINode>(Op0))
2434 if (Instruction *NV = FoldOpIntoPhi(I))
2439 // 0 / X == 0, we don't need to preserve faults!
2440 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2441 if (LHS->equalsInt(0))
2442 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2447 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2448 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2450 // Handle the integer div common cases
2451 if (Instruction *Common = commonIDivTransforms(I))
2454 // X udiv C^2 -> X >> C
2455 // Check to see if this is an unsigned division with an exact power of 2,
2456 // if so, convert to a right shift.
2457 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2458 if (!C->isZero() && C->getValue().isPowerOf2()) // Don't break X / 0
2459 return BinaryOperator::createLShr(Op0,
2460 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2463 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2464 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2465 if (RHSI->getOpcode() == Instruction::Shl &&
2466 isa<ConstantInt>(RHSI->getOperand(0))) {
2467 APInt C1(cast<ConstantInt>(RHSI->getOperand(0))->getValue());
2468 if (C1.isPowerOf2()) {
2469 Value *N = RHSI->getOperand(1);
2470 const Type *NTy = N->getType();
2471 if (uint32_t C2 = C1.logBase2()) {
2472 Constant *C2V = ConstantInt::get(NTy, C2);
2473 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2475 return BinaryOperator::createLShr(Op0, N);
2480 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2481 // where C1&C2 are powers of two.
2482 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2483 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2484 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2485 APInt TVA(STO->getValue()), FVA(SFO->getValue());
2486 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2487 // Compute the shift amounts
2488 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2489 // Construct the "on true" case of the select
2490 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2491 Instruction *TSI = BinaryOperator::createLShr(
2492 Op0, TC, SI->getName()+".t");
2493 TSI = InsertNewInstBefore(TSI, I);
2495 // Construct the "on false" case of the select
2496 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2497 Instruction *FSI = BinaryOperator::createLShr(
2498 Op0, FC, SI->getName()+".f");
2499 FSI = InsertNewInstBefore(FSI, I);
2501 // construct the select instruction and return it.
2502 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2508 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2509 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2511 // Handle the integer div common cases
2512 if (Instruction *Common = commonIDivTransforms(I))
2515 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2517 if (RHS->isAllOnesValue())
2518 return BinaryOperator::createNeg(Op0);
2521 if (Value *LHSNeg = dyn_castNegVal(Op0))
2522 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2525 // If the sign bits of both operands are zero (i.e. we can prove they are
2526 // unsigned inputs), turn this into a udiv.
2527 if (I.getType()->isInteger()) {
2528 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2529 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2530 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2537 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2538 return commonDivTransforms(I);
2541 /// GetFactor - If we can prove that the specified value is at least a multiple
2542 /// of some factor, return that factor.
2543 static Constant *GetFactor(Value *V) {
2544 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2547 // Unless we can be tricky, we know this is a multiple of 1.
2548 Constant *Result = ConstantInt::get(V->getType(), 1);
2550 Instruction *I = dyn_cast<Instruction>(V);
2551 if (!I) return Result;
2553 if (I->getOpcode() == Instruction::Mul) {
2554 // Handle multiplies by a constant, etc.
2555 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2556 GetFactor(I->getOperand(1)));
2557 } else if (I->getOpcode() == Instruction::Shl) {
2558 // (X<<C) -> X * (1 << C)
2559 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2560 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2561 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2563 } else if (I->getOpcode() == Instruction::And) {
2564 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2565 // X & 0xFFF0 is known to be a multiple of 16.
2566 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2567 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2568 return ConstantExpr::getShl(Result,
2569 ConstantInt::get(Result->getType(), Zeros));
2571 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2572 // Only handle int->int casts.
2573 if (!CI->isIntegerCast())
2575 Value *Op = CI->getOperand(0);
2576 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2581 /// This function implements the transforms on rem instructions that work
2582 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2583 /// is used by the visitors to those instructions.
2584 /// @brief Transforms common to all three rem instructions
2585 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2586 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2588 // 0 % X == 0, we don't need to preserve faults!
2589 if (Constant *LHS = dyn_cast<Constant>(Op0))
2590 if (LHS->isNullValue())
2591 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2593 if (isa<UndefValue>(Op0)) // undef % X -> 0
2594 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2595 if (isa<UndefValue>(Op1))
2596 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2598 // Handle cases involving: rem X, (select Cond, Y, Z)
2599 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2600 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2601 // the same basic block, then we replace the select with Y, and the
2602 // condition of the select with false (if the cond value is in the same
2603 // BB). If the select has uses other than the div, this allows them to be
2605 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2606 if (ST->isNullValue()) {
2607 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2608 if (CondI && CondI->getParent() == I.getParent())
2609 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2610 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2611 I.setOperand(1, SI->getOperand(2));
2613 UpdateValueUsesWith(SI, SI->getOperand(2));
2616 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2617 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2618 if (ST->isNullValue()) {
2619 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2620 if (CondI && CondI->getParent() == I.getParent())
2621 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2622 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2623 I.setOperand(1, SI->getOperand(1));
2625 UpdateValueUsesWith(SI, SI->getOperand(1));
2633 /// This function implements the transforms common to both integer remainder
2634 /// instructions (urem and srem). It is called by the visitors to those integer
2635 /// remainder instructions.
2636 /// @brief Common integer remainder transforms
2637 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2638 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2640 if (Instruction *common = commonRemTransforms(I))
2643 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2644 // X % 0 == undef, we don't need to preserve faults!
2645 if (RHS->equalsInt(0))
2646 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2648 if (RHS->equalsInt(1)) // X % 1 == 0
2649 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2651 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2652 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2653 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2655 } else if (isa<PHINode>(Op0I)) {
2656 if (Instruction *NV = FoldOpIntoPhi(I))
2659 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2660 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2661 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2668 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2669 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2671 if (Instruction *common = commonIRemTransforms(I))
2674 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2675 // X urem C^2 -> X and C
2676 // Check to see if this is an unsigned remainder with an exact power of 2,
2677 // if so, convert to a bitwise and.
2678 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2679 if (C->getValue().isPowerOf2())
2680 return BinaryOperator::createAnd(Op0, SubOne(C));
2683 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2684 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2685 if (RHSI->getOpcode() == Instruction::Shl &&
2686 isa<ConstantInt>(RHSI->getOperand(0))) {
2687 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2688 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2689 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2691 return BinaryOperator::createAnd(Op0, Add);
2696 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2697 // where C1&C2 are powers of two.
2698 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2699 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2700 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2701 // STO == 0 and SFO == 0 handled above.
2702 if ((STO->getValue().isPowerOf2()) &&
2703 (SFO->getValue().isPowerOf2())) {
2704 Value *TrueAnd = InsertNewInstBefore(
2705 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2706 Value *FalseAnd = InsertNewInstBefore(
2707 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2708 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2716 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2717 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2719 if (Instruction *common = commonIRemTransforms(I))
2722 if (Value *RHSNeg = dyn_castNegVal(Op1))
2723 if (!isa<ConstantInt>(RHSNeg) ||
2724 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2726 AddUsesToWorkList(I);
2727 I.setOperand(1, RHSNeg);
2731 // If the top bits of both operands are zero (i.e. we can prove they are
2732 // unsigned inputs), turn this into a urem.
2733 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2734 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2735 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2736 return BinaryOperator::createURem(Op0, Op1, I.getName());
2742 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2743 return commonRemTransforms(I);
2746 // isMaxValueMinusOne - return true if this is Max-1
2747 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2748 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2750 // Calculate 0111111111..11111
2751 APInt Val(APInt::getSignedMaxValue(TypeBits));
2752 return C->getValue() == Val-1;
2754 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2757 // isMinValuePlusOne - return true if this is Min+1
2758 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2760 // Calculate 1111111111000000000000
2761 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2762 APInt Val(APInt::getSignedMinValue(TypeBits));
2763 return C->getValue() == Val+1;
2765 return C->getValue() == 1; // unsigned
2768 // isOneBitSet - Return true if there is exactly one bit set in the specified
2770 static bool isOneBitSet(const ConstantInt *CI) {
2771 return CI->getValue().isPowerOf2();
2774 // isHighOnes - Return true if the constant is of the form 1+0+.
2775 // This is the same as lowones(~X).
2776 static bool isHighOnes(const ConstantInt *CI) {
2777 return (~CI->getValue() + 1).isPowerOf2();
2780 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2781 /// are carefully arranged to allow folding of expressions such as:
2783 /// (A < B) | (A > B) --> (A != B)
2785 /// Note that this is only valid if the first and second predicates have the
2786 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2788 /// Three bits are used to represent the condition, as follows:
2793 /// <=> Value Definition
2794 /// 000 0 Always false
2801 /// 111 7 Always true
2803 static unsigned getICmpCode(const ICmpInst *ICI) {
2804 switch (ICI->getPredicate()) {
2806 case ICmpInst::ICMP_UGT: return 1; // 001
2807 case ICmpInst::ICMP_SGT: return 1; // 001
2808 case ICmpInst::ICMP_EQ: return 2; // 010
2809 case ICmpInst::ICMP_UGE: return 3; // 011
2810 case ICmpInst::ICMP_SGE: return 3; // 011
2811 case ICmpInst::ICMP_ULT: return 4; // 100
2812 case ICmpInst::ICMP_SLT: return 4; // 100
2813 case ICmpInst::ICMP_NE: return 5; // 101
2814 case ICmpInst::ICMP_ULE: return 6; // 110
2815 case ICmpInst::ICMP_SLE: return 6; // 110
2818 assert(0 && "Invalid ICmp predicate!");
2823 /// getICmpValue - This is the complement of getICmpCode, which turns an
2824 /// opcode and two operands into either a constant true or false, or a brand
2825 /// new /// ICmp instruction. The sign is passed in to determine which kind
2826 /// of predicate to use in new icmp instructions.
2827 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2829 default: assert(0 && "Illegal ICmp code!");
2830 case 0: return ConstantInt::getFalse();
2833 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2835 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2836 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2839 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2841 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2844 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2846 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2847 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2850 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2852 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2853 case 7: return ConstantInt::getTrue();
2857 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2858 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2859 (ICmpInst::isSignedPredicate(p1) &&
2860 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2861 (ICmpInst::isSignedPredicate(p2) &&
2862 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2866 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2867 struct FoldICmpLogical {
2870 ICmpInst::Predicate pred;
2871 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2872 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2873 pred(ICI->getPredicate()) {}
2874 bool shouldApply(Value *V) const {
2875 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2876 if (PredicatesFoldable(pred, ICI->getPredicate()))
2877 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2878 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2881 Instruction *apply(Instruction &Log) const {
2882 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2883 if (ICI->getOperand(0) != LHS) {
2884 assert(ICI->getOperand(1) == LHS);
2885 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2888 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2889 unsigned LHSCode = getICmpCode(ICI);
2890 unsigned RHSCode = getICmpCode(RHSICI);
2892 switch (Log.getOpcode()) {
2893 case Instruction::And: Code = LHSCode & RHSCode; break;
2894 case Instruction::Or: Code = LHSCode | RHSCode; break;
2895 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2896 default: assert(0 && "Illegal logical opcode!"); return 0;
2899 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2900 ICmpInst::isSignedPredicate(ICI->getPredicate());
2902 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2903 if (Instruction *I = dyn_cast<Instruction>(RV))
2905 // Otherwise, it's a constant boolean value...
2906 return IC.ReplaceInstUsesWith(Log, RV);
2909 } // end anonymous namespace
2911 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2912 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2913 // guaranteed to be a binary operator.
2914 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2916 ConstantInt *AndRHS,
2917 BinaryOperator &TheAnd) {
2918 Value *X = Op->getOperand(0);
2919 Constant *Together = 0;
2921 Together = And(AndRHS, OpRHS);
2923 switch (Op->getOpcode()) {
2924 case Instruction::Xor:
2925 if (Op->hasOneUse()) {
2926 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2927 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2928 InsertNewInstBefore(And, TheAnd);
2930 return BinaryOperator::createXor(And, Together);
2933 case Instruction::Or:
2934 if (Together == AndRHS) // (X | C) & C --> C
2935 return ReplaceInstUsesWith(TheAnd, AndRHS);
2937 if (Op->hasOneUse() && Together != OpRHS) {
2938 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2939 Instruction *Or = BinaryOperator::createOr(X, Together);
2940 InsertNewInstBefore(Or, TheAnd);
2942 return BinaryOperator::createAnd(Or, AndRHS);
2945 case Instruction::Add:
2946 if (Op->hasOneUse()) {
2947 // Adding a one to a single bit bit-field should be turned into an XOR
2948 // of the bit. First thing to check is to see if this AND is with a
2949 // single bit constant.
2950 APInt AndRHSV(cast<ConstantInt>(AndRHS)->getValue());
2952 // If there is only one bit set...
2953 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2954 // Ok, at this point, we know that we are masking the result of the
2955 // ADD down to exactly one bit. If the constant we are adding has
2956 // no bits set below this bit, then we can eliminate the ADD.
2957 APInt AddRHS(cast<ConstantInt>(OpRHS)->getValue());
2959 // Check to see if any bits below the one bit set in AndRHSV are set.
2960 if ((AddRHS & (AndRHSV-1)) == 0) {
2961 // If not, the only thing that can effect the output of the AND is
2962 // the bit specified by AndRHSV. If that bit is set, the effect of
2963 // the XOR is to toggle the bit. If it is clear, then the ADD has
2965 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2966 TheAnd.setOperand(0, X);
2969 // Pull the XOR out of the AND.
2970 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2971 InsertNewInstBefore(NewAnd, TheAnd);
2972 NewAnd->takeName(Op);
2973 return BinaryOperator::createXor(NewAnd, AndRHS);
2980 case Instruction::Shl: {
2981 // We know that the AND will not produce any of the bits shifted in, so if
2982 // the anded constant includes them, clear them now!
2984 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2985 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2986 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2988 if (CI == ShlMask) { // Masking out bits that the shift already masks
2989 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2990 } else if (CI != AndRHS) { // Reducing bits set in and.
2991 TheAnd.setOperand(1, CI);
2996 case Instruction::LShr:
2998 // We know that the AND will not produce any of the bits shifted in, so if
2999 // the anded constant includes them, clear them now! This only applies to
3000 // unsigned shifts, because a signed shr may bring in set bits!
3002 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3003 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3004 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
3006 if (CI == ShrMask) { // Masking out bits that the shift already masks.
3007 return ReplaceInstUsesWith(TheAnd, Op);
3008 } else if (CI != AndRHS) {
3009 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3014 case Instruction::AShr:
3016 // See if this is shifting in some sign extension, then masking it out
3018 if (Op->hasOneUse()) {
3019 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3020 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3021 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
3022 if (C == AndRHS) { // Masking out bits shifted in.
3023 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3024 // Make the argument unsigned.
3025 Value *ShVal = Op->getOperand(0);
3026 ShVal = InsertNewInstBefore(
3027 BinaryOperator::createLShr(ShVal, OpRHS,
3028 Op->getName()), TheAnd);
3029 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3038 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3039 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3040 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3041 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3042 /// insert new instructions.
3043 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3044 bool isSigned, bool Inside,
3046 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3047 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3048 "Lo is not <= Hi in range emission code!");
3051 if (Lo == Hi) // Trivially false.
3052 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3054 // V >= Min && V < Hi --> V < Hi
3055 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3056 ICmpInst::Predicate pred = (isSigned ?
3057 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3058 return new ICmpInst(pred, V, Hi);
3061 // Emit V-Lo <u Hi-Lo
3062 Constant *NegLo = ConstantExpr::getNeg(Lo);
3063 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3064 InsertNewInstBefore(Add, IB);
3065 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3066 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3069 if (Lo == Hi) // Trivially true.
3070 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3072 // V < Min || V >= Hi -> V > Hi-1
3073 Hi = SubOne(cast<ConstantInt>(Hi));
3074 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3075 ICmpInst::Predicate pred = (isSigned ?
3076 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3077 return new ICmpInst(pred, V, Hi);
3080 // Emit V-Lo >u Hi-1-Lo
3081 // Note that Hi has already had one subtracted from it, above.
3082 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3083 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3084 InsertNewInstBefore(Add, IB);
3085 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3086 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3089 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3090 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3091 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3092 // not, since all 1s are not contiguous.
3093 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3094 APInt V = Val->getValue();
3095 uint32_t BitWidth = Val->getType()->getBitWidth();
3096 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3098 // look for the first zero bit after the run of ones
3099 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3100 // look for the first non-zero bit
3101 ME = V.getActiveBits();
3105 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3106 /// where isSub determines whether the operator is a sub. If we can fold one of
3107 /// the following xforms:
3109 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3110 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3111 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3113 /// return (A +/- B).
3115 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3116 ConstantInt *Mask, bool isSub,
3118 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3119 if (!LHSI || LHSI->getNumOperands() != 2 ||
3120 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3122 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3124 switch (LHSI->getOpcode()) {
3126 case Instruction::And:
3127 if (And(N, Mask) == Mask) {
3128 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3129 if ((Mask->getValue().countLeadingZeros() +
3130 Mask->getValue().countPopulation()) ==
3131 Mask->getValue().getBitWidth())
3134 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3135 // part, we don't need any explicit masks to take them out of A. If that
3136 // is all N is, ignore it.
3138 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3139 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3140 APInt Mask(APInt::getAllOnesValue(BitWidth));
3141 Mask = Mask.lshr(BitWidth-MB+1);
3142 if (MaskedValueIsZero(RHS, Mask))
3147 case Instruction::Or:
3148 case Instruction::Xor:
3149 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3150 if ((Mask->getValue().countLeadingZeros() +
3151 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3152 && And(N, Mask)->isNullValue())
3159 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3161 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3162 return InsertNewInstBefore(New, I);
3165 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3166 bool Changed = SimplifyCommutative(I);
3167 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3169 if (isa<UndefValue>(Op1)) // X & undef -> 0
3170 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3174 return ReplaceInstUsesWith(I, Op1);
3176 // See if we can simplify any instructions used by the instruction whose sole
3177 // purpose is to compute bits we don't care about.
3178 if (!isa<VectorType>(I.getType())) {
3179 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3180 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3181 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3182 KnownZero, KnownOne))
3185 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3186 if (CP->isAllOnesValue())
3187 return ReplaceInstUsesWith(I, I.getOperand(0));
3191 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3192 APInt AndRHSMask(AndRHS->getValue());
3193 APInt TypeMask(cast<IntegerType>(Op0->getType())->getMask());
3194 APInt NotAndRHS = AndRHSMask^TypeMask;
3196 // Optimize a variety of ((val OP C1) & C2) combinations...
3197 if (isa<BinaryOperator>(Op0)) {
3198 Instruction *Op0I = cast<Instruction>(Op0);
3199 Value *Op0LHS = Op0I->getOperand(0);
3200 Value *Op0RHS = Op0I->getOperand(1);
3201 switch (Op0I->getOpcode()) {
3202 case Instruction::Xor:
3203 case Instruction::Or:
3204 // If the mask is only needed on one incoming arm, push it up.
3205 if (Op0I->hasOneUse()) {
3206 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3207 // Not masking anything out for the LHS, move to RHS.
3208 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3209 Op0RHS->getName()+".masked");
3210 InsertNewInstBefore(NewRHS, I);
3211 return BinaryOperator::create(
3212 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3214 if (!isa<Constant>(Op0RHS) &&
3215 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3216 // Not masking anything out for the RHS, move to LHS.
3217 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3218 Op0LHS->getName()+".masked");
3219 InsertNewInstBefore(NewLHS, I);
3220 return BinaryOperator::create(
3221 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3226 case Instruction::Add:
3227 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3228 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3229 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3230 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3231 return BinaryOperator::createAnd(V, AndRHS);
3232 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3233 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3236 case Instruction::Sub:
3237 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3238 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3239 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3240 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3241 return BinaryOperator::createAnd(V, AndRHS);
3245 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3246 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3248 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3249 // If this is an integer truncation or change from signed-to-unsigned, and
3250 // if the source is an and/or with immediate, transform it. This
3251 // frequently occurs for bitfield accesses.
3252 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3253 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3254 CastOp->getNumOperands() == 2)
3255 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3256 if (CastOp->getOpcode() == Instruction::And) {
3257 // Change: and (cast (and X, C1) to T), C2
3258 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3259 // This will fold the two constants together, which may allow
3260 // other simplifications.
3261 Instruction *NewCast = CastInst::createTruncOrBitCast(
3262 CastOp->getOperand(0), I.getType(),
3263 CastOp->getName()+".shrunk");
3264 NewCast = InsertNewInstBefore(NewCast, I);
3265 // trunc_or_bitcast(C1)&C2
3266 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3267 C3 = ConstantExpr::getAnd(C3, AndRHS);
3268 return BinaryOperator::createAnd(NewCast, C3);
3269 } else if (CastOp->getOpcode() == Instruction::Or) {
3270 // Change: and (cast (or X, C1) to T), C2
3271 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3272 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3273 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3274 return ReplaceInstUsesWith(I, AndRHS);
3279 // Try to fold constant and into select arguments.
3280 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3281 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3283 if (isa<PHINode>(Op0))
3284 if (Instruction *NV = FoldOpIntoPhi(I))
3288 Value *Op0NotVal = dyn_castNotVal(Op0);
3289 Value *Op1NotVal = dyn_castNotVal(Op1);
3291 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3292 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3294 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3295 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3296 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3297 I.getName()+".demorgan");
3298 InsertNewInstBefore(Or, I);
3299 return BinaryOperator::createNot(Or);
3303 Value *A = 0, *B = 0;
3304 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3305 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3306 return ReplaceInstUsesWith(I, Op1);
3307 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3308 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3309 return ReplaceInstUsesWith(I, Op0);
3311 if (Op0->hasOneUse() &&
3312 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3313 if (A == Op1) { // (A^B)&A -> A&(A^B)
3314 I.swapOperands(); // Simplify below
3315 std::swap(Op0, Op1);
3316 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3317 cast<BinaryOperator>(Op0)->swapOperands();
3318 I.swapOperands(); // Simplify below
3319 std::swap(Op0, Op1);
3322 if (Op1->hasOneUse() &&
3323 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3324 if (B == Op0) { // B&(A^B) -> B&(B^A)
3325 cast<BinaryOperator>(Op1)->swapOperands();
3328 if (A == Op0) { // A&(A^B) -> A & ~B
3329 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3330 InsertNewInstBefore(NotB, I);
3331 return BinaryOperator::createAnd(A, NotB);
3336 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3337 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3338 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3341 Value *LHSVal, *RHSVal;
3342 ConstantInt *LHSCst, *RHSCst;
3343 ICmpInst::Predicate LHSCC, RHSCC;
3344 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3345 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3346 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3347 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3348 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3349 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3350 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3351 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3352 // Ensure that the larger constant is on the RHS.
3353 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3354 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3355 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3356 ICmpInst *LHS = cast<ICmpInst>(Op0);
3357 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3358 std::swap(LHS, RHS);
3359 std::swap(LHSCst, RHSCst);
3360 std::swap(LHSCC, RHSCC);
3363 // At this point, we know we have have two icmp instructions
3364 // comparing a value against two constants and and'ing the result
3365 // together. Because of the above check, we know that we only have
3366 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3367 // (from the FoldICmpLogical check above), that the two constants
3368 // are not equal and that the larger constant is on the RHS
3369 assert(LHSCst != RHSCst && "Compares not folded above?");
3372 default: assert(0 && "Unknown integer condition code!");
3373 case ICmpInst::ICMP_EQ:
3375 default: assert(0 && "Unknown integer condition code!");
3376 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3377 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3378 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3379 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3380 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3381 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3382 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3383 return ReplaceInstUsesWith(I, LHS);
3385 case ICmpInst::ICMP_NE:
3387 default: assert(0 && "Unknown integer condition code!");
3388 case ICmpInst::ICMP_ULT:
3389 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3390 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3391 break; // (X != 13 & X u< 15) -> no change
3392 case ICmpInst::ICMP_SLT:
3393 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3394 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3395 break; // (X != 13 & X s< 15) -> no change
3396 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3397 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3398 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3399 return ReplaceInstUsesWith(I, RHS);
3400 case ICmpInst::ICMP_NE:
3401 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3402 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3403 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3404 LHSVal->getName()+".off");
3405 InsertNewInstBefore(Add, I);
3406 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3407 ConstantInt::get(Add->getType(), 1));
3409 break; // (X != 13 & X != 15) -> no change
3412 case ICmpInst::ICMP_ULT:
3414 default: assert(0 && "Unknown integer condition code!");
3415 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3416 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3417 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3418 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3420 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3421 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3422 return ReplaceInstUsesWith(I, LHS);
3423 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3427 case ICmpInst::ICMP_SLT:
3429 default: assert(0 && "Unknown integer condition code!");
3430 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3431 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3432 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3433 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3435 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3436 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3437 return ReplaceInstUsesWith(I, LHS);
3438 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3442 case ICmpInst::ICMP_UGT:
3444 default: assert(0 && "Unknown integer condition code!");
3445 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3446 return ReplaceInstUsesWith(I, LHS);
3447 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3448 return ReplaceInstUsesWith(I, RHS);
3449 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3451 case ICmpInst::ICMP_NE:
3452 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3453 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3454 break; // (X u> 13 & X != 15) -> no change
3455 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3456 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3458 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3462 case ICmpInst::ICMP_SGT:
3464 default: assert(0 && "Unknown integer condition code!");
3465 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3466 return ReplaceInstUsesWith(I, LHS);
3467 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3468 return ReplaceInstUsesWith(I, RHS);
3469 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3471 case ICmpInst::ICMP_NE:
3472 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3473 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3474 break; // (X s> 13 & X != 15) -> no change
3475 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3476 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3478 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3486 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3487 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3488 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3489 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3490 const Type *SrcTy = Op0C->getOperand(0)->getType();
3491 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3492 // Only do this if the casts both really cause code to be generated.
3493 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3495 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3497 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3498 Op1C->getOperand(0),
3500 InsertNewInstBefore(NewOp, I);
3501 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3505 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3506 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3507 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3508 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3509 SI0->getOperand(1) == SI1->getOperand(1) &&
3510 (SI0->hasOneUse() || SI1->hasOneUse())) {
3511 Instruction *NewOp =
3512 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3514 SI0->getName()), I);
3515 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3516 SI1->getOperand(1));
3520 return Changed ? &I : 0;
3523 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3524 /// in the result. If it does, and if the specified byte hasn't been filled in
3525 /// yet, fill it in and return false.
3526 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3527 Instruction *I = dyn_cast<Instruction>(V);
3528 if (I == 0) return true;
3530 // If this is an or instruction, it is an inner node of the bswap.
3531 if (I->getOpcode() == Instruction::Or)
3532 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3533 CollectBSwapParts(I->getOperand(1), ByteValues);
3535 // If this is a shift by a constant int, and it is "24", then its operand
3536 // defines a byte. We only handle unsigned types here.
3537 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3538 // Not shifting the entire input by N-1 bytes?
3539 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3540 8*(ByteValues.size()-1))
3544 if (I->getOpcode() == Instruction::Shl) {
3545 // X << 24 defines the top byte with the lowest of the input bytes.
3546 DestNo = ByteValues.size()-1;
3548 // X >>u 24 defines the low byte with the highest of the input bytes.
3552 // If the destination byte value is already defined, the values are or'd
3553 // together, which isn't a bswap (unless it's an or of the same bits).
3554 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3556 ByteValues[DestNo] = I->getOperand(0);
3560 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3562 Value *Shift = 0, *ShiftLHS = 0;
3563 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3564 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3565 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3567 Instruction *SI = cast<Instruction>(Shift);
3569 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3570 if (ShiftAmt->getZExtValue() & 7 ||
3571 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3574 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3576 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3577 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3579 // Unknown mask for bswap.
3580 if (DestByte == ByteValues.size()) return true;
3582 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3584 if (SI->getOpcode() == Instruction::Shl)
3585 SrcByte = DestByte - ShiftBytes;
3587 SrcByte = DestByte + ShiftBytes;
3589 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3590 if (SrcByte != ByteValues.size()-DestByte-1)
3593 // If the destination byte value is already defined, the values are or'd
3594 // together, which isn't a bswap (unless it's an or of the same bits).
3595 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3597 ByteValues[DestByte] = SI->getOperand(0);
3601 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3602 /// If so, insert the new bswap intrinsic and return it.
3603 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3604 // We cannot bswap one byte.
3605 if (I.getType() == Type::Int8Ty)
3608 /// ByteValues - For each byte of the result, we keep track of which value
3609 /// defines each byte.
3610 SmallVector<Value*, 8> ByteValues;
3611 ByteValues.resize(TD->getTypeSize(I.getType()));
3613 // Try to find all the pieces corresponding to the bswap.
3614 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3615 CollectBSwapParts(I.getOperand(1), ByteValues))
3618 // Check to see if all of the bytes come from the same value.
3619 Value *V = ByteValues[0];
3620 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3622 // Check to make sure that all of the bytes come from the same value.
3623 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3624 if (ByteValues[i] != V)
3627 // If they do then *success* we can turn this into a bswap. Figure out what
3628 // bswap to make it into.
3629 Module *M = I.getParent()->getParent()->getParent();
3630 const char *FnName = 0;
3631 if (I.getType() == Type::Int16Ty)
3632 FnName = "llvm.bswap.i16";
3633 else if (I.getType() == Type::Int32Ty)
3634 FnName = "llvm.bswap.i32";
3635 else if (I.getType() == Type::Int64Ty)
3636 FnName = "llvm.bswap.i64";
3638 assert(0 && "Unknown integer type!");
3639 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3640 return new CallInst(F, V);
3644 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3645 bool Changed = SimplifyCommutative(I);
3646 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3648 if (isa<UndefValue>(Op1)) // X | undef -> -1
3649 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
3653 return ReplaceInstUsesWith(I, Op0);
3655 // See if we can simplify any instructions used by the instruction whose sole
3656 // purpose is to compute bits we don't care about.
3657 if (!isa<VectorType>(I.getType())) {
3658 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3659 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3660 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3661 KnownZero, KnownOne))
3666 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3667 ConstantInt *C1 = 0; Value *X = 0;
3668 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3669 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3670 Instruction *Or = BinaryOperator::createOr(X, RHS);
3671 InsertNewInstBefore(Or, I);
3673 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3676 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3677 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3678 Instruction *Or = BinaryOperator::createOr(X, RHS);
3679 InsertNewInstBefore(Or, I);
3681 return BinaryOperator::createXor(Or,
3682 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3685 // Try to fold constant and into select arguments.
3686 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3687 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3689 if (isa<PHINode>(Op0))
3690 if (Instruction *NV = FoldOpIntoPhi(I))
3694 Value *A = 0, *B = 0;
3695 ConstantInt *C1 = 0, *C2 = 0;
3697 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3698 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3699 return ReplaceInstUsesWith(I, Op1);
3700 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3701 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3702 return ReplaceInstUsesWith(I, Op0);
3704 // (A | B) | C and A | (B | C) -> bswap if possible.
3705 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3706 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3707 match(Op1, m_Or(m_Value(), m_Value())) ||
3708 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3709 match(Op1, m_Shift(m_Value(), m_Value())))) {
3710 if (Instruction *BSwap = MatchBSwap(I))
3714 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3715 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3716 MaskedValueIsZero(Op1, C1->getValue())) {
3717 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3718 InsertNewInstBefore(NOr, I);
3720 return BinaryOperator::createXor(NOr, C1);
3723 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3724 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3725 MaskedValueIsZero(Op0, C1->getValue())) {
3726 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3727 InsertNewInstBefore(NOr, I);
3729 return BinaryOperator::createXor(NOr, C1);
3732 // (A & C1)|(B & C2)
3733 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3734 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3736 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3737 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3740 // If we have: ((V + N) & C1) | (V & C2)
3741 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3742 // replace with V+N.
3743 if (C1 == ConstantExpr::getNot(C2)) {
3744 Value *V1 = 0, *V2 = 0;
3745 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3746 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3747 // Add commutes, try both ways.
3748 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3749 return ReplaceInstUsesWith(I, A);
3750 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3751 return ReplaceInstUsesWith(I, A);
3753 // Or commutes, try both ways.
3754 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3755 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3756 // Add commutes, try both ways.
3757 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3758 return ReplaceInstUsesWith(I, B);
3759 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3760 return ReplaceInstUsesWith(I, B);
3765 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3766 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3767 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3768 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3769 SI0->getOperand(1) == SI1->getOperand(1) &&
3770 (SI0->hasOneUse() || SI1->hasOneUse())) {
3771 Instruction *NewOp =
3772 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3774 SI0->getName()), I);
3775 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3776 SI1->getOperand(1));
3780 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3781 if (A == Op1) // ~A | A == -1
3782 return ReplaceInstUsesWith(I,
3783 ConstantInt::getAllOnesValue(I.getType()));
3787 // Note, A is still live here!
3788 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3790 return ReplaceInstUsesWith(I,
3791 ConstantInt::getAllOnesValue(I.getType()));
3793 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3794 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3795 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3796 I.getName()+".demorgan"), I);
3797 return BinaryOperator::createNot(And);
3801 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3802 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3803 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3806 Value *LHSVal, *RHSVal;
3807 ConstantInt *LHSCst, *RHSCst;
3808 ICmpInst::Predicate LHSCC, RHSCC;
3809 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3810 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3811 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3812 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3813 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3814 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3815 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3816 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3817 // Ensure that the larger constant is on the RHS.
3818 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3819 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3820 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3821 ICmpInst *LHS = cast<ICmpInst>(Op0);
3822 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3823 std::swap(LHS, RHS);
3824 std::swap(LHSCst, RHSCst);
3825 std::swap(LHSCC, RHSCC);
3828 // At this point, we know we have have two icmp instructions
3829 // comparing a value against two constants and or'ing the result
3830 // together. Because of the above check, we know that we only have
3831 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3832 // FoldICmpLogical check above), that the two constants are not
3834 assert(LHSCst != RHSCst && "Compares not folded above?");
3837 default: assert(0 && "Unknown integer condition code!");
3838 case ICmpInst::ICMP_EQ:
3840 default: assert(0 && "Unknown integer condition code!");
3841 case ICmpInst::ICMP_EQ:
3842 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3843 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3844 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3845 LHSVal->getName()+".off");
3846 InsertNewInstBefore(Add, I);
3847 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3848 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3850 break; // (X == 13 | X == 15) -> no change
3851 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3852 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3854 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3855 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3856 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3857 return ReplaceInstUsesWith(I, RHS);
3860 case ICmpInst::ICMP_NE:
3862 default: assert(0 && "Unknown integer condition code!");
3863 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3864 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3865 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3866 return ReplaceInstUsesWith(I, LHS);
3867 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3868 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3869 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3870 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3873 case ICmpInst::ICMP_ULT:
3875 default: assert(0 && "Unknown integer condition code!");
3876 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3878 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3879 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3881 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3883 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3884 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3885 return ReplaceInstUsesWith(I, RHS);
3886 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3890 case ICmpInst::ICMP_SLT:
3892 default: assert(0 && "Unknown integer condition code!");
3893 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3895 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3896 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3898 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3900 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3901 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3902 return ReplaceInstUsesWith(I, RHS);
3903 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3907 case ICmpInst::ICMP_UGT:
3909 default: assert(0 && "Unknown integer condition code!");
3910 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3911 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3912 return ReplaceInstUsesWith(I, LHS);
3913 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3915 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3916 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3917 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3918 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3922 case ICmpInst::ICMP_SGT:
3924 default: assert(0 && "Unknown integer condition code!");
3925 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3926 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3927 return ReplaceInstUsesWith(I, LHS);
3928 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3930 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3931 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3932 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3933 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3941 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3942 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3943 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3944 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3945 const Type *SrcTy = Op0C->getOperand(0)->getType();
3946 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3947 // Only do this if the casts both really cause code to be generated.
3948 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3950 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3952 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3953 Op1C->getOperand(0),
3955 InsertNewInstBefore(NewOp, I);
3956 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3961 return Changed ? &I : 0;
3964 // XorSelf - Implements: X ^ X --> 0
3967 XorSelf(Value *rhs) : RHS(rhs) {}
3968 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3969 Instruction *apply(BinaryOperator &Xor) const {
3975 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3976 bool Changed = SimplifyCommutative(I);
3977 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3979 if (isa<UndefValue>(Op1))
3980 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3982 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3983 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3984 assert(Result == &I && "AssociativeOpt didn't work?");
3985 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3988 // See if we can simplify any instructions used by the instruction whose sole
3989 // purpose is to compute bits we don't care about.
3990 if (!isa<VectorType>(I.getType())) {
3991 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3992 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3993 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3994 KnownZero, KnownOne))
3998 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3999 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4000 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4001 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4002 return new ICmpInst(ICI->getInversePredicate(),
4003 ICI->getOperand(0), ICI->getOperand(1));
4005 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4006 // ~(c-X) == X-c-1 == X+(-c-1)
4007 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4008 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4009 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4010 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4011 ConstantInt::get(I.getType(), 1));
4012 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4015 // ~(~X & Y) --> (X | ~Y)
4016 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
4017 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4018 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4020 BinaryOperator::createNot(Op0I->getOperand(1),
4021 Op0I->getOperand(1)->getName()+".not");
4022 InsertNewInstBefore(NotY, I);
4023 return BinaryOperator::createOr(Op0NotVal, NotY);
4027 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4028 if (Op0I->getOpcode() == Instruction::Add) {
4029 // ~(X-c) --> (-c-1)-X
4030 if (RHS->isAllOnesValue()) {
4031 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4032 return BinaryOperator::createSub(
4033 ConstantExpr::getSub(NegOp0CI,
4034 ConstantInt::get(I.getType(), 1)),
4035 Op0I->getOperand(0));
4037 } else if (Op0I->getOpcode() == Instruction::Or) {
4038 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4039 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4040 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4041 // Anything in both C1 and C2 is known to be zero, remove it from
4043 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
4044 NewRHS = ConstantExpr::getAnd(NewRHS,
4045 ConstantExpr::getNot(CommonBits));
4046 AddToWorkList(Op0I);
4047 I.setOperand(0, Op0I->getOperand(0));
4048 I.setOperand(1, NewRHS);
4054 // Try to fold constant and into select arguments.
4055 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4056 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4058 if (isa<PHINode>(Op0))
4059 if (Instruction *NV = FoldOpIntoPhi(I))
4063 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4065 return ReplaceInstUsesWith(I,
4066 ConstantInt::getAllOnesValue(I.getType()));
4068 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4070 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4073 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4076 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4077 if (A == Op0) { // B^(B|A) == (A|B)^B
4078 Op1I->swapOperands();
4080 std::swap(Op0, Op1);
4081 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4082 I.swapOperands(); // Simplified below.
4083 std::swap(Op0, Op1);
4085 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4086 if (Op0 == A) // A^(A^B) == B
4087 return ReplaceInstUsesWith(I, B);
4088 else if (Op0 == B) // A^(B^A) == B
4089 return ReplaceInstUsesWith(I, A);
4090 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4091 if (A == Op0) // A^(A&B) -> A^(B&A)
4092 Op1I->swapOperands();
4093 if (B == Op0) { // A^(B&A) -> (B&A)^A
4094 I.swapOperands(); // Simplified below.
4095 std::swap(Op0, Op1);
4100 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4103 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4104 if (A == Op1) // (B|A)^B == (A|B)^B
4106 if (B == Op1) { // (A|B)^B == A & ~B
4108 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4109 return BinaryOperator::createAnd(A, NotB);
4111 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4112 if (Op1 == A) // (A^B)^A == B
4113 return ReplaceInstUsesWith(I, B);
4114 else if (Op1 == B) // (B^A)^A == B
4115 return ReplaceInstUsesWith(I, A);
4116 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4117 if (A == Op1) // (A&B)^A -> (B&A)^A
4119 if (B == Op1 && // (B&A)^A == ~B & A
4120 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4122 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4123 return BinaryOperator::createAnd(N, Op1);
4128 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4129 if (Op0I && Op1I && Op0I->isShift() &&
4130 Op0I->getOpcode() == Op1I->getOpcode() &&
4131 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4132 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4133 Instruction *NewOp =
4134 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4135 Op1I->getOperand(0),
4136 Op0I->getName()), I);
4137 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4138 Op1I->getOperand(1));
4142 Value *A, *B, *C, *D;
4143 // (A & B)^(A | B) -> A ^ B
4144 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4145 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4146 if ((A == C && B == D) || (A == D && B == C))
4147 return BinaryOperator::createXor(A, B);
4149 // (A | B)^(A & B) -> A ^ B
4150 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4151 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4152 if ((A == C && B == D) || (A == D && B == C))
4153 return BinaryOperator::createXor(A, B);
4157 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4158 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4159 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4160 // (X & Y)^(X & Y) -> (Y^Z) & X
4161 Value *X = 0, *Y = 0, *Z = 0;
4163 X = A, Y = B, Z = D;
4165 X = A, Y = B, Z = C;
4167 X = B, Y = A, Z = D;
4169 X = B, Y = A, Z = C;
4172 Instruction *NewOp =
4173 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4174 return BinaryOperator::createAnd(NewOp, X);
4179 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4180 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4181 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4184 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4185 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4186 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4187 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4188 const Type *SrcTy = Op0C->getOperand(0)->getType();
4189 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4190 // Only do this if the casts both really cause code to be generated.
4191 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4193 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4195 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4196 Op1C->getOperand(0),
4198 InsertNewInstBefore(NewOp, I);
4199 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4203 return Changed ? &I : 0;
4206 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4207 /// overflowed for this type.
4208 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4209 ConstantInt *In2, bool IsSigned = false) {
4210 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4213 if (In2->getValue().isNegative())
4214 return Result->getValue().sgt(In1->getValue());
4216 return Result->getValue().slt(In1->getValue());
4218 return Result->getValue().ult(In1->getValue());
4221 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4222 /// code necessary to compute the offset from the base pointer (without adding
4223 /// in the base pointer). Return the result as a signed integer of intptr size.
4224 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4225 TargetData &TD = IC.getTargetData();
4226 gep_type_iterator GTI = gep_type_begin(GEP);
4227 const Type *IntPtrTy = TD.getIntPtrType();
4228 Value *Result = Constant::getNullValue(IntPtrTy);
4230 // Build a mask for high order bits.
4231 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4233 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4234 Value *Op = GEP->getOperand(i);
4235 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4236 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4237 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4238 if (!OpC->isNullValue()) {
4239 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4240 Scale = ConstantExpr::getMul(OpC, Scale);
4241 if (Constant *RC = dyn_cast<Constant>(Result))
4242 Result = ConstantExpr::getAdd(RC, Scale);
4244 // Emit an add instruction.
4245 Result = IC.InsertNewInstBefore(
4246 BinaryOperator::createAdd(Result, Scale,
4247 GEP->getName()+".offs"), I);
4251 // Convert to correct type.
4252 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4253 Op->getName()+".c"), I);
4255 // We'll let instcombine(mul) convert this to a shl if possible.
4256 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4257 GEP->getName()+".idx"), I);
4259 // Emit an add instruction.
4260 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4261 GEP->getName()+".offs"), I);
4267 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4268 /// else. At this point we know that the GEP is on the LHS of the comparison.
4269 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4270 ICmpInst::Predicate Cond,
4272 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4274 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4275 if (isa<PointerType>(CI->getOperand(0)->getType()))
4276 RHS = CI->getOperand(0);
4278 Value *PtrBase = GEPLHS->getOperand(0);
4279 if (PtrBase == RHS) {
4280 // As an optimization, we don't actually have to compute the actual value of
4281 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4282 // each index is zero or not.
4283 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4284 Instruction *InVal = 0;
4285 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4286 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4288 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4289 if (isa<UndefValue>(C)) // undef index -> undef.
4290 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4291 if (C->isNullValue())
4293 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4294 EmitIt = false; // This is indexing into a zero sized array?
4295 } else if (isa<ConstantInt>(C))
4296 return ReplaceInstUsesWith(I, // No comparison is needed here.
4297 ConstantInt::get(Type::Int1Ty,
4298 Cond == ICmpInst::ICMP_NE));
4303 new ICmpInst(Cond, GEPLHS->getOperand(i),
4304 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4308 InVal = InsertNewInstBefore(InVal, I);
4309 InsertNewInstBefore(Comp, I);
4310 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4311 InVal = BinaryOperator::createOr(InVal, Comp);
4312 else // True if all are equal
4313 InVal = BinaryOperator::createAnd(InVal, Comp);
4321 // No comparison is needed here, all indexes = 0
4322 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4323 Cond == ICmpInst::ICMP_EQ));
4326 // Only lower this if the icmp is the only user of the GEP or if we expect
4327 // the result to fold to a constant!
4328 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4329 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4330 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4331 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4332 Constant::getNullValue(Offset->getType()));
4334 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4335 // If the base pointers are different, but the indices are the same, just
4336 // compare the base pointer.
4337 if (PtrBase != GEPRHS->getOperand(0)) {
4338 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4339 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4340 GEPRHS->getOperand(0)->getType();
4342 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4343 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4344 IndicesTheSame = false;
4348 // If all indices are the same, just compare the base pointers.
4350 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4351 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4353 // Otherwise, the base pointers are different and the indices are
4354 // different, bail out.
4358 // If one of the GEPs has all zero indices, recurse.
4359 bool AllZeros = true;
4360 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4361 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4362 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4367 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4368 ICmpInst::getSwappedPredicate(Cond), I);
4370 // If the other GEP has all zero indices, recurse.
4372 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4373 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4374 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4379 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4381 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4382 // If the GEPs only differ by one index, compare it.
4383 unsigned NumDifferences = 0; // Keep track of # differences.
4384 unsigned DiffOperand = 0; // The operand that differs.
4385 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4386 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4387 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4388 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4389 // Irreconcilable differences.
4393 if (NumDifferences++) break;
4398 if (NumDifferences == 0) // SAME GEP?
4399 return ReplaceInstUsesWith(I, // No comparison is needed here.
4400 ConstantInt::get(Type::Int1Ty,
4401 Cond == ICmpInst::ICMP_EQ));
4402 else if (NumDifferences == 1) {
4403 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4404 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4405 // Make sure we do a signed comparison here.
4406 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4410 // Only lower this if the icmp is the only user of the GEP or if we expect
4411 // the result to fold to a constant!
4412 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4413 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4414 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4415 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4416 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4417 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4423 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4424 bool Changed = SimplifyCompare(I);
4425 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4427 // Fold trivial predicates.
4428 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4429 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4430 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4431 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4433 // Simplify 'fcmp pred X, X'
4435 switch (I.getPredicate()) {
4436 default: assert(0 && "Unknown predicate!");
4437 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4438 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4439 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4440 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4441 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4442 case FCmpInst::FCMP_OLT: // True if ordered and less than
4443 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4444 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4446 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4447 case FCmpInst::FCMP_ULT: // True if unordered or less than
4448 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4449 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4450 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4451 I.setPredicate(FCmpInst::FCMP_UNO);
4452 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4455 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4456 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4457 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4458 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4459 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4460 I.setPredicate(FCmpInst::FCMP_ORD);
4461 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4466 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4467 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4469 // Handle fcmp with constant RHS
4470 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4471 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4472 switch (LHSI->getOpcode()) {
4473 case Instruction::PHI:
4474 if (Instruction *NV = FoldOpIntoPhi(I))
4477 case Instruction::Select:
4478 // If either operand of the select is a constant, we can fold the
4479 // comparison into the select arms, which will cause one to be
4480 // constant folded and the select turned into a bitwise or.
4481 Value *Op1 = 0, *Op2 = 0;
4482 if (LHSI->hasOneUse()) {
4483 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4484 // Fold the known value into the constant operand.
4485 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4486 // Insert a new FCmp of the other select operand.
4487 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4488 LHSI->getOperand(2), RHSC,
4490 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4491 // Fold the known value into the constant operand.
4492 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4493 // Insert a new FCmp of the other select operand.
4494 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4495 LHSI->getOperand(1), RHSC,
4501 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4506 return Changed ? &I : 0;
4509 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4510 bool Changed = SimplifyCompare(I);
4511 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4512 const Type *Ty = Op0->getType();
4516 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4517 isTrueWhenEqual(I)));
4519 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4520 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4522 // icmp of GlobalValues can never equal each other as long as they aren't
4523 // external weak linkage type.
4524 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4525 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4526 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4527 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4528 !isTrueWhenEqual(I)));
4530 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4531 // addresses never equal each other! We already know that Op0 != Op1.
4532 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4533 isa<ConstantPointerNull>(Op0)) &&
4534 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4535 isa<ConstantPointerNull>(Op1)))
4536 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4537 !isTrueWhenEqual(I)));
4539 // icmp's with boolean values can always be turned into bitwise operations
4540 if (Ty == Type::Int1Ty) {
4541 switch (I.getPredicate()) {
4542 default: assert(0 && "Invalid icmp instruction!");
4543 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4544 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4545 InsertNewInstBefore(Xor, I);
4546 return BinaryOperator::createNot(Xor);
4548 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4549 return BinaryOperator::createXor(Op0, Op1);
4551 case ICmpInst::ICMP_UGT:
4552 case ICmpInst::ICMP_SGT:
4553 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4555 case ICmpInst::ICMP_ULT:
4556 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4557 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4558 InsertNewInstBefore(Not, I);
4559 return BinaryOperator::createAnd(Not, Op1);
4561 case ICmpInst::ICMP_UGE:
4562 case ICmpInst::ICMP_SGE:
4563 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4565 case ICmpInst::ICMP_ULE:
4566 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4567 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4568 InsertNewInstBefore(Not, I);
4569 return BinaryOperator::createOr(Not, Op1);
4574 // See if we are doing a comparison between a constant and an instruction that
4575 // can be folded into the comparison.
4576 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4577 switch (I.getPredicate()) {
4579 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4580 if (CI->isMinValue(false))
4581 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4582 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4583 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4584 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4585 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4588 case ICmpInst::ICMP_SLT:
4589 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4590 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4591 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4592 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4593 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4594 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4597 case ICmpInst::ICMP_UGT:
4598 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4599 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4600 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4601 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4602 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4603 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4606 case ICmpInst::ICMP_SGT:
4607 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4608 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4609 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4610 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4611 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4612 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4615 case ICmpInst::ICMP_ULE:
4616 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4617 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4618 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4619 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4620 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4621 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4624 case ICmpInst::ICMP_SLE:
4625 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4626 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4627 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4628 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4629 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4630 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4633 case ICmpInst::ICMP_UGE:
4634 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4635 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4636 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4637 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4638 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4639 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4642 case ICmpInst::ICMP_SGE:
4643 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4644 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4645 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4646 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4647 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4648 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4652 // If we still have a icmp le or icmp ge instruction, turn it into the
4653 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4654 // already been handled above, this requires little checking.
4656 switch (I.getPredicate()) {
4658 case ICmpInst::ICMP_ULE:
4659 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4660 case ICmpInst::ICMP_SLE:
4661 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4662 case ICmpInst::ICMP_UGE:
4663 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4664 case ICmpInst::ICMP_SGE:
4665 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4668 // See if we can fold the comparison based on bits known to be zero or one
4670 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4671 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4672 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
4673 KnownZero, KnownOne, 0))
4676 // Given the known and unknown bits, compute a range that the LHS could be
4678 if ((KnownOne | KnownZero) != 0) {
4679 // Compute the Min, Max and RHS values based on the known bits. For the
4680 // EQ and NE we use unsigned values.
4681 APInt Min(BitWidth, 0), Max(BitWidth, 0), RHSVal(CI->getValue());
4682 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4683 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4686 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4689 switch (I.getPredicate()) { // LE/GE have been folded already.
4690 default: assert(0 && "Unknown icmp opcode!");
4691 case ICmpInst::ICMP_EQ:
4692 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4693 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4695 case ICmpInst::ICMP_NE:
4696 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4697 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4699 case ICmpInst::ICMP_ULT:
4700 if (Max.ult(RHSVal))
4701 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4702 if (Min.ugt(RHSVal))
4703 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4705 case ICmpInst::ICMP_UGT:
4706 if (Min.ugt(RHSVal))
4707 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4708 if (Max.ult(RHSVal))
4709 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4711 case ICmpInst::ICMP_SLT:
4712 if (Max.slt(RHSVal))
4713 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4714 if (Min.sgt(RHSVal))
4715 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4717 case ICmpInst::ICMP_SGT:
4718 if (Min.sgt(RHSVal))
4719 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4720 if (Max.slt(RHSVal))
4721 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4726 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4727 // instruction, see if that instruction also has constants so that the
4728 // instruction can be folded into the icmp
4729 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4730 switch (LHSI->getOpcode()) {
4731 case Instruction::And:
4732 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4733 LHSI->getOperand(0)->hasOneUse()) {
4734 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4736 // If the LHS is an AND of a truncating cast, we can widen the
4737 // and/compare to be the input width without changing the value
4738 // produced, eliminating a cast.
4739 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4740 // We can do this transformation if either the AND constant does not
4741 // have its sign bit set or if it is an equality comparison.
4742 // Extending a relational comparison when we're checking the sign
4743 // bit would not work.
4744 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4745 (I.isEquality() || AndCST->getValue().isPositive() &&
4746 CI->getValue().isPositive())) {
4747 ConstantInt *NewCST;
4749 APInt NewCSTVal(AndCST->getValue()), NewCIVal(CI->getValue());
4750 uint32_t BitWidth = cast<IntegerType>(
4751 Cast->getOperand(0)->getType())->getBitWidth();
4752 NewCST = ConstantInt::get(NewCSTVal.zext(BitWidth));
4753 NewCI = ConstantInt::get(NewCIVal.zext(BitWidth));
4754 Instruction *NewAnd =
4755 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4757 InsertNewInstBefore(NewAnd, I);
4758 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4762 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4763 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4764 // happens a LOT in code produced by the C front-end, for bitfield
4766 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4767 if (Shift && !Shift->isShift())
4771 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4772 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4773 const Type *AndTy = AndCST->getType(); // Type of the and.
4775 // We can fold this as long as we can't shift unknown bits
4776 // into the mask. This can only happen with signed shift
4777 // rights, as they sign-extend.
4779 bool CanFold = Shift->isLogicalShift();
4781 // To test for the bad case of the signed shr, see if any
4782 // of the bits shifted in could be tested after the mask.
4783 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4784 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4786 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4788 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4790 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4796 if (Shift->getOpcode() == Instruction::Shl)
4797 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4799 NewCst = ConstantExpr::getShl(CI, ShAmt);
4801 // Check to see if we are shifting out any of the bits being
4803 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4804 // If we shifted bits out, the fold is not going to work out.
4805 // As a special case, check to see if this means that the
4806 // result is always true or false now.
4807 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4808 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4809 if (I.getPredicate() == ICmpInst::ICMP_NE)
4810 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4812 I.setOperand(1, NewCst);
4813 Constant *NewAndCST;
4814 if (Shift->getOpcode() == Instruction::Shl)
4815 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4817 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4818 LHSI->setOperand(1, NewAndCST);
4819 LHSI->setOperand(0, Shift->getOperand(0));
4820 AddToWorkList(Shift); // Shift is dead.
4821 AddUsesToWorkList(I);
4827 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4828 // preferable because it allows the C<<Y expression to be hoisted out
4829 // of a loop if Y is invariant and X is not.
4830 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4831 I.isEquality() && !Shift->isArithmeticShift() &&
4832 isa<Instruction>(Shift->getOperand(0))) {
4835 if (Shift->getOpcode() == Instruction::LShr) {
4836 NS = BinaryOperator::createShl(AndCST,
4837 Shift->getOperand(1), "tmp");
4839 // Insert a logical shift.
4840 NS = BinaryOperator::createLShr(AndCST,
4841 Shift->getOperand(1), "tmp");
4843 InsertNewInstBefore(cast<Instruction>(NS), I);
4845 // Compute X & (C << Y).
4846 Instruction *NewAnd = BinaryOperator::createAnd(
4847 Shift->getOperand(0), NS, LHSI->getName());
4848 InsertNewInstBefore(NewAnd, I);
4850 I.setOperand(0, NewAnd);
4856 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4857 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4858 if (I.isEquality()) {
4859 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4861 // Check that the shift amount is in range. If not, don't perform
4862 // undefined shifts. When the shift is visited it will be
4864 if (ShAmt->getZExtValue() >= TypeBits)
4867 // If we are comparing against bits always shifted out, the
4868 // comparison cannot succeed.
4870 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4871 if (Comp != CI) {// Comparing against a bit that we know is zero.
4872 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4873 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4874 return ReplaceInstUsesWith(I, Cst);
4877 if (LHSI->hasOneUse()) {
4878 // Otherwise strength reduce the shift into an and.
4879 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4880 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4881 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4884 BinaryOperator::createAnd(LHSI->getOperand(0),
4885 Mask, LHSI->getName()+".mask");
4886 Value *And = InsertNewInstBefore(AndI, I);
4887 return new ICmpInst(I.getPredicate(), And,
4888 ConstantExpr::getLShr(CI, ShAmt));
4894 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4895 case Instruction::AShr:
4896 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4897 if (I.isEquality()) {
4898 // Check that the shift amount is in range. If not, don't perform
4899 // undefined shifts. When the shift is visited it will be
4901 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4902 if (ShAmt->getZExtValue() >= TypeBits)
4905 // If we are comparing against bits always shifted out, the
4906 // comparison cannot succeed.
4908 if (LHSI->getOpcode() == Instruction::LShr)
4909 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4912 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4915 if (Comp != CI) {// Comparing against a bit that we know is zero.
4916 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4917 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4918 return ReplaceInstUsesWith(I, Cst);
4921 if (LHSI->hasOneUse() || CI->isNullValue()) {
4922 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4924 // Otherwise strength reduce the shift into an and.
4925 APInt Val(APInt::getAllOnesValue(TypeBits).shl(ShAmtVal));
4926 Constant *Mask = ConstantInt::get(Val);
4929 BinaryOperator::createAnd(LHSI->getOperand(0),
4930 Mask, LHSI->getName()+".mask");
4931 Value *And = InsertNewInstBefore(AndI, I);
4932 return new ICmpInst(I.getPredicate(), And,
4933 ConstantExpr::getShl(CI, ShAmt));
4939 case Instruction::SDiv:
4940 case Instruction::UDiv:
4941 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4942 // Fold this div into the comparison, producing a range check.
4943 // Determine, based on the divide type, what the range is being
4944 // checked. If there is an overflow on the low or high side, remember
4945 // it, otherwise compute the range [low, hi) bounding the new value.
4946 // See: InsertRangeTest above for the kinds of replacements possible.
4947 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4948 // FIXME: If the operand types don't match the type of the divide
4949 // then don't attempt this transform. The code below doesn't have the
4950 // logic to deal with a signed divide and an unsigned compare (and
4951 // vice versa). This is because (x /s C1) <s C2 produces different
4952 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4953 // (x /u C1) <u C2. Simply casting the operands and result won't
4954 // work. :( The if statement below tests that condition and bails
4956 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4957 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4959 if (DivRHS->isZero())
4960 break; // Don't hack on div by zero
4962 // Initialize the variables that will indicate the nature of the
4964 bool LoOverflow = false, HiOverflow = false;
4965 ConstantInt *LoBound = 0, *HiBound = 0;
4967 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4968 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4969 // C2 (CI). By solving for X we can turn this into a range check
4970 // instead of computing a divide.
4972 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4974 // Determine if the product overflows by seeing if the product is
4975 // not equal to the divide. Make sure we do the same kind of divide
4976 // as in the LHS instruction that we're folding.
4977 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4978 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4980 // Get the ICmp opcode
4981 ICmpInst::Predicate predicate = I.getPredicate();
4983 if (!DivIsSigned) { // udiv
4985 LoOverflow = ProdOV;
4986 HiOverflow = ProdOV ||
4987 AddWithOverflow(HiBound, LoBound, DivRHS, false);
4988 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
4989 if (CI->isNullValue()) { // (X / pos) op 0
4991 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4993 } else if (CI->getValue().isPositive()) { // (X / pos) op pos
4995 LoOverflow = ProdOV;
4996 HiOverflow = ProdOV ||
4997 AddWithOverflow(HiBound, Prod, DivRHS, true);
4998 } else { // (X / pos) op neg
4999 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5000 LoOverflow = AddWithOverflow(LoBound, Prod,
5001 cast<ConstantInt>(DivRHSH), true);
5002 HiBound = AddOne(Prod);
5003 HiOverflow = ProdOV;
5005 } else { // Divisor is < 0.
5006 if (CI->isNullValue()) { // (X / neg) op 0
5007 LoBound = AddOne(DivRHS);
5008 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5009 if (HiBound == DivRHS)
5010 LoBound = 0; // - INTMIN = INTMIN
5011 } else if (CI->getValue().isPositive()) { // (X / neg) op pos
5012 HiOverflow = LoOverflow = ProdOV;
5014 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5016 HiBound = AddOne(Prod);
5017 } else { // (X / neg) op neg
5019 LoOverflow = HiOverflow = ProdOV;
5020 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
5023 // Dividing by a negate swaps the condition.
5024 predicate = ICmpInst::getSwappedPredicate(predicate);
5028 Value *X = LHSI->getOperand(0);
5029 switch (predicate) {
5030 default: assert(0 && "Unhandled icmp opcode!");
5031 case ICmpInst::ICMP_EQ:
5032 if (LoOverflow && HiOverflow)
5033 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5034 else if (HiOverflow)
5035 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5036 ICmpInst::ICMP_UGE, X, LoBound);
5037 else if (LoOverflow)
5038 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5039 ICmpInst::ICMP_ULT, X, HiBound);
5041 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5043 case ICmpInst::ICMP_NE:
5044 if (LoOverflow && HiOverflow)
5045 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5046 else if (HiOverflow)
5047 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5048 ICmpInst::ICMP_ULT, X, LoBound);
5049 else if (LoOverflow)
5050 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5051 ICmpInst::ICMP_UGE, X, HiBound);
5053 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5055 case ICmpInst::ICMP_ULT:
5056 case ICmpInst::ICMP_SLT:
5058 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5059 return new ICmpInst(predicate, X, LoBound);
5060 case ICmpInst::ICMP_UGT:
5061 case ICmpInst::ICMP_SGT:
5063 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5064 if (predicate == ICmpInst::ICMP_UGT)
5065 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5067 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5074 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5075 if (I.isEquality()) {
5076 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5078 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5079 // the second operand is a constant, simplify a bit.
5080 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
5081 switch (BO->getOpcode()) {
5082 case Instruction::SRem:
5083 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5084 if (CI->isZero() && isa<ConstantInt>(BO->getOperand(1)) &&
5086 APInt V(cast<ConstantInt>(BO->getOperand(1))->getValue());
5087 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5088 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
5089 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
5090 return new ICmpInst(I.getPredicate(), NewRem,
5091 Constant::getNullValue(BO->getType()));
5095 case Instruction::Add:
5096 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5097 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5098 if (BO->hasOneUse())
5099 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5100 ConstantExpr::getSub(CI, BOp1C));
5101 } else if (CI->isNullValue()) {
5102 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5103 // efficiently invertible, or if the add has just this one use.
5104 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5106 if (Value *NegVal = dyn_castNegVal(BOp1))
5107 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
5108 else if (Value *NegVal = dyn_castNegVal(BOp0))
5109 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
5110 else if (BO->hasOneUse()) {
5111 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5112 InsertNewInstBefore(Neg, I);
5114 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5118 case Instruction::Xor:
5119 // For the xor case, we can xor two constants together, eliminating
5120 // the explicit xor.
5121 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5122 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5123 ConstantExpr::getXor(CI, BOC));
5126 case Instruction::Sub:
5127 // Replace (([sub|xor] A, B) != 0) with (A != B)
5129 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5133 case Instruction::Or:
5134 // If bits are being or'd in that are not present in the constant we
5135 // are comparing against, then the comparison could never succeed!
5136 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5137 Constant *NotCI = ConstantExpr::getNot(CI);
5138 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5139 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5144 case Instruction::And:
5145 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5146 // If bits are being compared against that are and'd out, then the
5147 // comparison can never succeed!
5148 if (!ConstantExpr::getAnd(CI,
5149 ConstantExpr::getNot(BOC))->isNullValue())
5150 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5153 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5154 if (CI == BOC && isOneBitSet(CI))
5155 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5156 ICmpInst::ICMP_NE, Op0,
5157 Constant::getNullValue(CI->getType()));
5159 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5160 if (isSignBit(BOC)) {
5161 Value *X = BO->getOperand(0);
5162 Constant *Zero = Constant::getNullValue(X->getType());
5163 ICmpInst::Predicate pred = isICMP_NE ?
5164 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5165 return new ICmpInst(pred, X, Zero);
5168 // ((X & ~7) == 0) --> X < 8
5169 if (CI->isNullValue() && isHighOnes(BOC)) {
5170 Value *X = BO->getOperand(0);
5171 Constant *NegX = ConstantExpr::getNeg(BOC);
5172 ICmpInst::Predicate pred = isICMP_NE ?
5173 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5174 return new ICmpInst(pred, X, NegX);
5180 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5181 // Handle set{eq|ne} <intrinsic>, intcst.
5182 switch (II->getIntrinsicID()) {
5184 case Intrinsic::bswap_i16:
5185 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5186 AddToWorkList(II); // Dead?
5187 I.setOperand(0, II->getOperand(1));
5188 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5189 ByteSwap_16(CI->getZExtValue())));
5191 case Intrinsic::bswap_i32:
5192 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5193 AddToWorkList(II); // Dead?
5194 I.setOperand(0, II->getOperand(1));
5195 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5196 ByteSwap_32(CI->getZExtValue())));
5198 case Intrinsic::bswap_i64:
5199 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5200 AddToWorkList(II); // Dead?
5201 I.setOperand(0, II->getOperand(1));
5202 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5203 ByteSwap_64(CI->getZExtValue())));
5207 } else { // Not a ICMP_EQ/ICMP_NE
5208 // If the LHS is a cast from an integral value of the same size, then
5209 // since we know the RHS is a constant, try to simlify.
5210 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5211 Value *CastOp = Cast->getOperand(0);
5212 const Type *SrcTy = CastOp->getType();
5213 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5214 if (SrcTy->isInteger() &&
5215 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5216 // If this is an unsigned comparison, try to make the comparison use
5217 // smaller constant values.
5218 switch (I.getPredicate()) {
5220 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5221 ConstantInt *CUI = cast<ConstantInt>(CI);
5222 if (CUI->getValue() == APInt::getSignBit(SrcTySize))
5223 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5224 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5227 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5228 ConstantInt *CUI = cast<ConstantInt>(CI);
5229 if (CUI->getValue() == APInt::getSignedMaxValue(SrcTySize))
5230 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5231 Constant::getNullValue(SrcTy));
5241 // Handle icmp with constant RHS
5242 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5243 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5244 switch (LHSI->getOpcode()) {
5245 case Instruction::GetElementPtr:
5246 if (RHSC->isNullValue()) {
5247 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5248 bool isAllZeros = true;
5249 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5250 if (!isa<Constant>(LHSI->getOperand(i)) ||
5251 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5256 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5257 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5261 case Instruction::PHI:
5262 if (Instruction *NV = FoldOpIntoPhi(I))
5265 case Instruction::Select:
5266 // If either operand of the select is a constant, we can fold the
5267 // comparison into the select arms, which will cause one to be
5268 // constant folded and the select turned into a bitwise or.
5269 Value *Op1 = 0, *Op2 = 0;
5270 if (LHSI->hasOneUse()) {
5271 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5272 // Fold the known value into the constant operand.
5273 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5274 // Insert a new ICmp of the other select operand.
5275 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5276 LHSI->getOperand(2), RHSC,
5278 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5279 // Fold the known value into the constant operand.
5280 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5281 // Insert a new ICmp of the other select operand.
5282 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5283 LHSI->getOperand(1), RHSC,
5289 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5294 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5295 if (User *GEP = dyn_castGetElementPtr(Op0))
5296 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5298 if (User *GEP = dyn_castGetElementPtr(Op1))
5299 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5300 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5303 // Test to see if the operands of the icmp are casted versions of other
5304 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5306 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5307 if (isa<PointerType>(Op0->getType()) &&
5308 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5309 // We keep moving the cast from the left operand over to the right
5310 // operand, where it can often be eliminated completely.
5311 Op0 = CI->getOperand(0);
5313 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5314 // so eliminate it as well.
5315 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5316 Op1 = CI2->getOperand(0);
5318 // If Op1 is a constant, we can fold the cast into the constant.
5319 if (Op0->getType() != Op1->getType())
5320 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5321 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5323 // Otherwise, cast the RHS right before the icmp
5324 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5326 return new ICmpInst(I.getPredicate(), Op0, Op1);
5330 if (isa<CastInst>(Op0)) {
5331 // Handle the special case of: icmp (cast bool to X), <cst>
5332 // This comes up when you have code like
5335 // For generality, we handle any zero-extension of any operand comparison
5336 // with a constant or another cast from the same type.
5337 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5338 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5342 if (I.isEquality()) {
5343 Value *A, *B, *C, *D;
5344 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5345 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5346 Value *OtherVal = A == Op1 ? B : A;
5347 return new ICmpInst(I.getPredicate(), OtherVal,
5348 Constant::getNullValue(A->getType()));
5351 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5352 // A^c1 == C^c2 --> A == C^(c1^c2)
5353 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5354 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5355 if (Op1->hasOneUse()) {
5356 Constant *NC = ConstantExpr::getXor(C1, C2);
5357 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5358 return new ICmpInst(I.getPredicate(), A,
5359 InsertNewInstBefore(Xor, I));
5362 // A^B == A^D -> B == D
5363 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5364 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5365 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5366 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5370 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5371 (A == Op0 || B == Op0)) {
5372 // A == (A^B) -> B == 0
5373 Value *OtherVal = A == Op0 ? B : A;
5374 return new ICmpInst(I.getPredicate(), OtherVal,
5375 Constant::getNullValue(A->getType()));
5377 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5378 // (A-B) == A -> B == 0
5379 return new ICmpInst(I.getPredicate(), B,
5380 Constant::getNullValue(B->getType()));
5382 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5383 // A == (A-B) -> B == 0
5384 return new ICmpInst(I.getPredicate(), B,
5385 Constant::getNullValue(B->getType()));
5388 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5389 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5390 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5391 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5392 Value *X = 0, *Y = 0, *Z = 0;
5395 X = B; Y = D; Z = A;
5396 } else if (A == D) {
5397 X = B; Y = C; Z = A;
5398 } else if (B == C) {
5399 X = A; Y = D; Z = B;
5400 } else if (B == D) {
5401 X = A; Y = C; Z = B;
5404 if (X) { // Build (X^Y) & Z
5405 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5406 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5407 I.setOperand(0, Op1);
5408 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5413 return Changed ? &I : 0;
5416 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5417 // We only handle extending casts so far.
5419 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5420 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5421 Value *LHSCIOp = LHSCI->getOperand(0);
5422 const Type *SrcTy = LHSCIOp->getType();
5423 const Type *DestTy = LHSCI->getType();
5426 // We only handle extension cast instructions, so far. Enforce this.
5427 if (LHSCI->getOpcode() != Instruction::ZExt &&
5428 LHSCI->getOpcode() != Instruction::SExt)
5431 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5432 bool isSignedCmp = ICI.isSignedPredicate();
5434 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5435 // Not an extension from the same type?
5436 RHSCIOp = CI->getOperand(0);
5437 if (RHSCIOp->getType() != LHSCIOp->getType())
5440 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5441 // and the other is a zext), then we can't handle this.
5442 if (CI->getOpcode() != LHSCI->getOpcode())
5445 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5446 // then we can't handle this.
5447 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5450 // Okay, just insert a compare of the reduced operands now!
5451 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5454 // If we aren't dealing with a constant on the RHS, exit early
5455 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5459 // Compute the constant that would happen if we truncated to SrcTy then
5460 // reextended to DestTy.
5461 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5462 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5464 // If the re-extended constant didn't change...
5466 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5467 // For example, we might have:
5468 // %A = sext short %X to uint
5469 // %B = icmp ugt uint %A, 1330
5470 // It is incorrect to transform this into
5471 // %B = icmp ugt short %X, 1330
5472 // because %A may have negative value.
5474 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5475 // OR operation is EQ/NE.
5476 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5477 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5482 // The re-extended constant changed so the constant cannot be represented
5483 // in the shorter type. Consequently, we cannot emit a simple comparison.
5485 // First, handle some easy cases. We know the result cannot be equal at this
5486 // point so handle the ICI.isEquality() cases
5487 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5488 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5489 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5490 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5492 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5493 // should have been folded away previously and not enter in here.
5496 // We're performing a signed comparison.
5497 if (cast<ConstantInt>(CI)->getValue().isNegative())
5498 Result = ConstantInt::getFalse(); // X < (small) --> false
5500 Result = ConstantInt::getTrue(); // X < (large) --> true
5502 // We're performing an unsigned comparison.
5504 // We're performing an unsigned comp with a sign extended value.
5505 // This is true if the input is >= 0. [aka >s -1]
5506 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5507 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5508 NegOne, ICI.getName()), ICI);
5510 // Unsigned extend & unsigned compare -> always true.
5511 Result = ConstantInt::getTrue();
5515 // Finally, return the value computed.
5516 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5517 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5518 return ReplaceInstUsesWith(ICI, Result);
5520 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5521 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5522 "ICmp should be folded!");
5523 if (Constant *CI = dyn_cast<Constant>(Result))
5524 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5526 return BinaryOperator::createNot(Result);
5530 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5531 return commonShiftTransforms(I);
5534 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5535 return commonShiftTransforms(I);
5538 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5539 return commonShiftTransforms(I);
5542 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5543 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5544 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5546 // shl X, 0 == X and shr X, 0 == X
5547 // shl 0, X == 0 and shr 0, X == 0
5548 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5549 Op0 == Constant::getNullValue(Op0->getType()))
5550 return ReplaceInstUsesWith(I, Op0);
5552 if (isa<UndefValue>(Op0)) {
5553 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5554 return ReplaceInstUsesWith(I, Op0);
5555 else // undef << X -> 0, undef >>u X -> 0
5556 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5558 if (isa<UndefValue>(Op1)) {
5559 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5560 return ReplaceInstUsesWith(I, Op0);
5561 else // X << undef, X >>u undef -> 0
5562 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5565 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5566 if (I.getOpcode() == Instruction::AShr)
5567 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5568 if (CSI->isAllOnesValue())
5569 return ReplaceInstUsesWith(I, CSI);
5571 // Try to fold constant and into select arguments.
5572 if (isa<Constant>(Op0))
5573 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5574 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5577 // See if we can turn a signed shr into an unsigned shr.
5578 if (I.isArithmeticShift()) {
5579 if (MaskedValueIsZero(Op0,
5580 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5581 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5585 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5586 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5591 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5592 BinaryOperator &I) {
5593 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5595 // See if we can simplify any instructions used by the instruction whose sole
5596 // purpose is to compute bits we don't care about.
5597 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5598 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5599 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5600 KnownZero, KnownOne))
5603 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5604 // of a signed value.
5606 if (Op1->getZExtValue() >= TypeBits) { // shift amount always <= 32 bits
5607 if (I.getOpcode() != Instruction::AShr)
5608 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5610 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5615 // ((X*C1) << C2) == (X * (C1 << C2))
5616 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5617 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5618 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5619 return BinaryOperator::createMul(BO->getOperand(0),
5620 ConstantExpr::getShl(BOOp, Op1));
5622 // Try to fold constant and into select arguments.
5623 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5624 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5626 if (isa<PHINode>(Op0))
5627 if (Instruction *NV = FoldOpIntoPhi(I))
5630 if (Op0->hasOneUse()) {
5631 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5632 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5635 switch (Op0BO->getOpcode()) {
5637 case Instruction::Add:
5638 case Instruction::And:
5639 case Instruction::Or:
5640 case Instruction::Xor: {
5641 // These operators commute.
5642 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5643 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5644 match(Op0BO->getOperand(1),
5645 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5646 Instruction *YS = BinaryOperator::createShl(
5647 Op0BO->getOperand(0), Op1,
5649 InsertNewInstBefore(YS, I); // (Y << C)
5651 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5652 Op0BO->getOperand(1)->getName());
5653 InsertNewInstBefore(X, I); // (X + (Y << C))
5654 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5655 C2 = ConstantExpr::getShl(C2, Op1);
5656 return BinaryOperator::createAnd(X, C2);
5659 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5660 Value *Op0BOOp1 = Op0BO->getOperand(1);
5661 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5663 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5664 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5666 Instruction *YS = BinaryOperator::createShl(
5667 Op0BO->getOperand(0), Op1,
5669 InsertNewInstBefore(YS, I); // (Y << C)
5671 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5672 V1->getName()+".mask");
5673 InsertNewInstBefore(XM, I); // X & (CC << C)
5675 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5680 case Instruction::Sub: {
5681 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5682 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5683 match(Op0BO->getOperand(0),
5684 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5685 Instruction *YS = BinaryOperator::createShl(
5686 Op0BO->getOperand(1), Op1,
5688 InsertNewInstBefore(YS, I); // (Y << C)
5690 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5691 Op0BO->getOperand(0)->getName());
5692 InsertNewInstBefore(X, I); // (X + (Y << C))
5693 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5694 C2 = ConstantExpr::getShl(C2, Op1);
5695 return BinaryOperator::createAnd(X, C2);
5698 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5699 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5700 match(Op0BO->getOperand(0),
5701 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5702 m_ConstantInt(CC))) && V2 == Op1 &&
5703 cast<BinaryOperator>(Op0BO->getOperand(0))
5704 ->getOperand(0)->hasOneUse()) {
5705 Instruction *YS = BinaryOperator::createShl(
5706 Op0BO->getOperand(1), Op1,
5708 InsertNewInstBefore(YS, I); // (Y << C)
5710 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5711 V1->getName()+".mask");
5712 InsertNewInstBefore(XM, I); // X & (CC << C)
5714 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5722 // If the operand is an bitwise operator with a constant RHS, and the
5723 // shift is the only use, we can pull it out of the shift.
5724 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5725 bool isValid = true; // Valid only for And, Or, Xor
5726 bool highBitSet = false; // Transform if high bit of constant set?
5728 switch (Op0BO->getOpcode()) {
5729 default: isValid = false; break; // Do not perform transform!
5730 case Instruction::Add:
5731 isValid = isLeftShift;
5733 case Instruction::Or:
5734 case Instruction::Xor:
5737 case Instruction::And:
5742 // If this is a signed shift right, and the high bit is modified
5743 // by the logical operation, do not perform the transformation.
5744 // The highBitSet boolean indicates the value of the high bit of
5745 // the constant which would cause it to be modified for this
5748 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5749 isValid = ((Op0C->getValue() & APInt::getSignBit(TypeBits)) != 0) ==
5754 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5756 Instruction *NewShift =
5757 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5758 InsertNewInstBefore(NewShift, I);
5759 NewShift->takeName(Op0BO);
5761 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5768 // Find out if this is a shift of a shift by a constant.
5769 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5770 if (ShiftOp && !ShiftOp->isShift())
5773 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5774 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5775 // These shift amounts are always <= 32 bits.
5776 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5777 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5778 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5779 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5780 Value *X = ShiftOp->getOperand(0);
5782 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5783 if (AmtSum > TypeBits)
5786 const IntegerType *Ty = cast<IntegerType>(I.getType());
5788 // Check for (X << c1) << c2 and (X >> c1) >> c2
5789 if (I.getOpcode() == ShiftOp->getOpcode()) {
5790 return BinaryOperator::create(I.getOpcode(), X,
5791 ConstantInt::get(Ty, AmtSum));
5792 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5793 I.getOpcode() == Instruction::AShr) {
5794 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5795 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5796 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5797 I.getOpcode() == Instruction::LShr) {
5798 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5799 Instruction *Shift =
5800 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5801 InsertNewInstBefore(Shift, I);
5803 APInt Mask(Ty->getMask().lshr(ShiftAmt2));
5804 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5807 // Okay, if we get here, one shift must be left, and the other shift must be
5808 // right. See if the amounts are equal.
5809 if (ShiftAmt1 == ShiftAmt2) {
5810 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5811 if (I.getOpcode() == Instruction::Shl) {
5812 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
5813 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5815 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5816 if (I.getOpcode() == Instruction::LShr) {
5817 APInt Mask(Ty->getMask().lshr(ShiftAmt1));
5818 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5820 // We can simplify ((X << C) >>s C) into a trunc + sext.
5821 // NOTE: we could do this for any C, but that would make 'unusual' integer
5822 // types. For now, just stick to ones well-supported by the code
5824 const Type *SExtType = 0;
5825 switch (Ty->getBitWidth() - ShiftAmt1) {
5826 case 1 : SExtType = Type::Int1Ty; break;
5827 case 8 : SExtType = Type::Int8Ty; break;
5828 case 16 : SExtType = Type::Int16Ty; break;
5829 case 32 : SExtType = Type::Int32Ty; break;
5830 case 64 : SExtType = Type::Int64Ty; break;
5834 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5835 InsertNewInstBefore(NewTrunc, I);
5836 return new SExtInst(NewTrunc, Ty);
5838 // Otherwise, we can't handle it yet.
5839 } else if (ShiftAmt1 < ShiftAmt2) {
5840 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5842 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5843 if (I.getOpcode() == Instruction::Shl) {
5844 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5845 ShiftOp->getOpcode() == Instruction::AShr);
5846 Instruction *Shift =
5847 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5848 InsertNewInstBefore(Shift, I);
5850 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5851 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5854 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5855 if (I.getOpcode() == Instruction::LShr) {
5856 assert(ShiftOp->getOpcode() == Instruction::Shl);
5857 Instruction *Shift =
5858 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5859 InsertNewInstBefore(Shift, I);
5861 APInt Mask(APInt::getLowBitsSet(TypeBits, ShiftAmt2));
5862 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5865 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5867 assert(ShiftAmt2 < ShiftAmt1);
5868 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5870 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5871 if (I.getOpcode() == Instruction::Shl) {
5872 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5873 ShiftOp->getOpcode() == Instruction::AShr);
5874 Instruction *Shift =
5875 BinaryOperator::create(ShiftOp->getOpcode(), X,
5876 ConstantInt::get(Ty, ShiftDiff));
5877 InsertNewInstBefore(Shift, I);
5879 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5880 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5883 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5884 if (I.getOpcode() == Instruction::LShr) {
5885 assert(ShiftOp->getOpcode() == Instruction::Shl);
5886 Instruction *Shift =
5887 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5888 InsertNewInstBefore(Shift, I);
5890 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5891 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5894 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5901 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5902 /// expression. If so, decompose it, returning some value X, such that Val is
5905 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5907 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5908 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5909 Offset = CI->getZExtValue();
5911 return ConstantInt::get(Type::Int32Ty, 0);
5912 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5913 if (I->getNumOperands() == 2) {
5914 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5915 if (I->getOpcode() == Instruction::Shl) {
5916 // This is a value scaled by '1 << the shift amt'.
5917 Scale = 1U << CUI->getZExtValue();
5919 return I->getOperand(0);
5920 } else if (I->getOpcode() == Instruction::Mul) {
5921 // This value is scaled by 'CUI'.
5922 Scale = CUI->getZExtValue();
5924 return I->getOperand(0);
5925 } else if (I->getOpcode() == Instruction::Add) {
5926 // We have X+C. Check to see if we really have (X*C2)+C1,
5927 // where C1 is divisible by C2.
5930 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5931 Offset += CUI->getZExtValue();
5932 if (SubScale > 1 && (Offset % SubScale == 0)) {
5941 // Otherwise, we can't look past this.
5948 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5949 /// try to eliminate the cast by moving the type information into the alloc.
5950 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5951 AllocationInst &AI) {
5952 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5953 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5955 // Remove any uses of AI that are dead.
5956 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5958 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5959 Instruction *User = cast<Instruction>(*UI++);
5960 if (isInstructionTriviallyDead(User)) {
5961 while (UI != E && *UI == User)
5962 ++UI; // If this instruction uses AI more than once, don't break UI.
5965 DOUT << "IC: DCE: " << *User;
5966 EraseInstFromFunction(*User);
5970 // Get the type really allocated and the type casted to.
5971 const Type *AllocElTy = AI.getAllocatedType();
5972 const Type *CastElTy = PTy->getElementType();
5973 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5975 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5976 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5977 if (CastElTyAlign < AllocElTyAlign) return 0;
5979 // If the allocation has multiple uses, only promote it if we are strictly
5980 // increasing the alignment of the resultant allocation. If we keep it the
5981 // same, we open the door to infinite loops of various kinds.
5982 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5984 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5985 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5986 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5988 // See if we can satisfy the modulus by pulling a scale out of the array
5990 unsigned ArraySizeScale, ArrayOffset;
5991 Value *NumElements = // See if the array size is a decomposable linear expr.
5992 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5994 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5996 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5997 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5999 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6004 // If the allocation size is constant, form a constant mul expression
6005 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6006 if (isa<ConstantInt>(NumElements))
6007 Amt = ConstantExpr::getMul(
6008 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6009 // otherwise multiply the amount and the number of elements
6010 else if (Scale != 1) {
6011 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6012 Amt = InsertNewInstBefore(Tmp, AI);
6016 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6017 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
6018 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6019 Amt = InsertNewInstBefore(Tmp, AI);
6022 AllocationInst *New;
6023 if (isa<MallocInst>(AI))
6024 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6026 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6027 InsertNewInstBefore(New, AI);
6030 // If the allocation has multiple uses, insert a cast and change all things
6031 // that used it to use the new cast. This will also hack on CI, but it will
6033 if (!AI.hasOneUse()) {
6034 AddUsesToWorkList(AI);
6035 // New is the allocation instruction, pointer typed. AI is the original
6036 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6037 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6038 InsertNewInstBefore(NewCast, AI);
6039 AI.replaceAllUsesWith(NewCast);
6041 return ReplaceInstUsesWith(CI, New);
6044 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6045 /// and return it as type Ty without inserting any new casts and without
6046 /// changing the computed value. This is used by code that tries to decide
6047 /// whether promoting or shrinking integer operations to wider or smaller types
6048 /// will allow us to eliminate a truncate or extend.
6050 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6051 /// extension operation if Ty is larger.
6052 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6053 int &NumCastsRemoved) {
6054 // We can always evaluate constants in another type.
6055 if (isa<ConstantInt>(V))
6058 Instruction *I = dyn_cast<Instruction>(V);
6059 if (!I) return false;
6061 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6063 switch (I->getOpcode()) {
6064 case Instruction::Add:
6065 case Instruction::Sub:
6066 case Instruction::And:
6067 case Instruction::Or:
6068 case Instruction::Xor:
6069 if (!I->hasOneUse()) return false;
6070 // These operators can all arbitrarily be extended or truncated.
6071 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6072 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6074 case Instruction::Shl:
6075 if (!I->hasOneUse()) return false;
6076 // If we are truncating the result of this SHL, and if it's a shift of a
6077 // constant amount, we can always perform a SHL in a smaller type.
6078 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6079 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6080 CI->getZExtValue() < Ty->getBitWidth())
6081 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6084 case Instruction::LShr:
6085 if (!I->hasOneUse()) return false;
6086 // If this is a truncate of a logical shr, we can truncate it to a smaller
6087 // lshr iff we know that the bits we would otherwise be shifting in are
6089 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6090 uint32_t BitWidth = OrigTy->getBitWidth();
6091 if (Ty->getBitWidth() < BitWidth &&
6092 MaskedValueIsZero(I->getOperand(0),
6093 APInt::getAllOnesValue(BitWidth) &
6094 APInt::getAllOnesValue(Ty->getBitWidth()).zextOrTrunc(BitWidth).flip())
6095 && CI->getZExtValue() < Ty->getBitWidth()) {
6096 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
6100 case Instruction::Trunc:
6101 case Instruction::ZExt:
6102 case Instruction::SExt:
6103 // If this is a cast from the destination type, we can trivially eliminate
6104 // it, and this will remove a cast overall.
6105 if (I->getOperand(0)->getType() == Ty) {
6106 // If the first operand is itself a cast, and is eliminable, do not count
6107 // this as an eliminable cast. We would prefer to eliminate those two
6109 if (isa<CastInst>(I->getOperand(0)))
6117 // TODO: Can handle more cases here.
6124 /// EvaluateInDifferentType - Given an expression that
6125 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6126 /// evaluate the expression.
6127 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6129 if (Constant *C = dyn_cast<Constant>(V))
6130 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6132 // Otherwise, it must be an instruction.
6133 Instruction *I = cast<Instruction>(V);
6134 Instruction *Res = 0;
6135 switch (I->getOpcode()) {
6136 case Instruction::Add:
6137 case Instruction::Sub:
6138 case Instruction::And:
6139 case Instruction::Or:
6140 case Instruction::Xor:
6141 case Instruction::AShr:
6142 case Instruction::LShr:
6143 case Instruction::Shl: {
6144 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6145 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6146 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6147 LHS, RHS, I->getName());
6150 case Instruction::Trunc:
6151 case Instruction::ZExt:
6152 case Instruction::SExt:
6153 case Instruction::BitCast:
6154 // If the source type of the cast is the type we're trying for then we can
6155 // just return the source. There's no need to insert it because its not new.
6156 if (I->getOperand(0)->getType() == Ty)
6157 return I->getOperand(0);
6159 // Some other kind of cast, which shouldn't happen, so just ..
6162 // TODO: Can handle more cases here.
6163 assert(0 && "Unreachable!");
6167 return InsertNewInstBefore(Res, *I);
6170 /// @brief Implement the transforms common to all CastInst visitors.
6171 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6172 Value *Src = CI.getOperand(0);
6174 // Casting undef to anything results in undef so might as just replace it and
6175 // get rid of the cast.
6176 if (isa<UndefValue>(Src)) // cast undef -> undef
6177 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6179 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6180 // eliminate it now.
6181 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6182 if (Instruction::CastOps opc =
6183 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6184 // The first cast (CSrc) is eliminable so we need to fix up or replace
6185 // the second cast (CI). CSrc will then have a good chance of being dead.
6186 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6190 // If casting the result of a getelementptr instruction with no offset, turn
6191 // this into a cast of the original pointer!
6193 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6194 bool AllZeroOperands = true;
6195 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6196 if (!isa<Constant>(GEP->getOperand(i)) ||
6197 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6198 AllZeroOperands = false;
6201 if (AllZeroOperands) {
6202 // Changing the cast operand is usually not a good idea but it is safe
6203 // here because the pointer operand is being replaced with another
6204 // pointer operand so the opcode doesn't need to change.
6205 CI.setOperand(0, GEP->getOperand(0));
6210 // If we are casting a malloc or alloca to a pointer to a type of the same
6211 // size, rewrite the allocation instruction to allocate the "right" type.
6212 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6213 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6216 // If we are casting a select then fold the cast into the select
6217 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6218 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6221 // If we are casting a PHI then fold the cast into the PHI
6222 if (isa<PHINode>(Src))
6223 if (Instruction *NV = FoldOpIntoPhi(CI))
6229 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6230 /// integer types. This function implements the common transforms for all those
6232 /// @brief Implement the transforms common to CastInst with integer operands
6233 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6234 if (Instruction *Result = commonCastTransforms(CI))
6237 Value *Src = CI.getOperand(0);
6238 const Type *SrcTy = Src->getType();
6239 const Type *DestTy = CI.getType();
6240 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6241 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6243 // See if we can simplify any instructions used by the LHS whose sole
6244 // purpose is to compute bits we don't care about.
6245 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6246 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6247 KnownZero, KnownOne))
6250 // If the source isn't an instruction or has more than one use then we
6251 // can't do anything more.
6252 Instruction *SrcI = dyn_cast<Instruction>(Src);
6253 if (!SrcI || !Src->hasOneUse())
6256 // Attempt to propagate the cast into the instruction for int->int casts.
6257 int NumCastsRemoved = 0;
6258 if (!isa<BitCastInst>(CI) &&
6259 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6261 // If this cast is a truncate, evaluting in a different type always
6262 // eliminates the cast, so it is always a win. If this is a noop-cast
6263 // this just removes a noop cast which isn't pointful, but simplifies
6264 // the code. If this is a zero-extension, we need to do an AND to
6265 // maintain the clear top-part of the computation, so we require that
6266 // the input have eliminated at least one cast. If this is a sign
6267 // extension, we insert two new casts (to do the extension) so we
6268 // require that two casts have been eliminated.
6270 switch (CI.getOpcode()) {
6272 // All the others use floating point so we shouldn't actually
6273 // get here because of the check above.
6274 assert(0 && "Unknown cast type");
6275 case Instruction::Trunc:
6278 case Instruction::ZExt:
6279 DoXForm = NumCastsRemoved >= 1;
6281 case Instruction::SExt:
6282 DoXForm = NumCastsRemoved >= 2;
6284 case Instruction::BitCast:
6290 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6291 CI.getOpcode() == Instruction::SExt);
6292 assert(Res->getType() == DestTy);
6293 switch (CI.getOpcode()) {
6294 default: assert(0 && "Unknown cast type!");
6295 case Instruction::Trunc:
6296 case Instruction::BitCast:
6297 // Just replace this cast with the result.
6298 return ReplaceInstUsesWith(CI, Res);
6299 case Instruction::ZExt: {
6300 // We need to emit an AND to clear the high bits.
6301 assert(SrcBitSize < DestBitSize && "Not a zext?");
6302 Constant *C = ConstantInt::get(APInt::getAllOnesValue(SrcBitSize));
6303 C = ConstantExpr::getZExt(C, DestTy);
6304 return BinaryOperator::createAnd(Res, C);
6306 case Instruction::SExt:
6307 // We need to emit a cast to truncate, then a cast to sext.
6308 return CastInst::create(Instruction::SExt,
6309 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6315 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6316 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6318 switch (SrcI->getOpcode()) {
6319 case Instruction::Add:
6320 case Instruction::Mul:
6321 case Instruction::And:
6322 case Instruction::Or:
6323 case Instruction::Xor:
6324 // If we are discarding information, or just changing the sign,
6326 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6327 // Don't insert two casts if they cannot be eliminated. We allow
6328 // two casts to be inserted if the sizes are the same. This could
6329 // only be converting signedness, which is a noop.
6330 if (DestBitSize == SrcBitSize ||
6331 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6332 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6333 Instruction::CastOps opcode = CI.getOpcode();
6334 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6335 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6336 return BinaryOperator::create(
6337 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6341 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6342 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6343 SrcI->getOpcode() == Instruction::Xor &&
6344 Op1 == ConstantInt::getTrue() &&
6345 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6346 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6347 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6350 case Instruction::SDiv:
6351 case Instruction::UDiv:
6352 case Instruction::SRem:
6353 case Instruction::URem:
6354 // If we are just changing the sign, rewrite.
6355 if (DestBitSize == SrcBitSize) {
6356 // Don't insert two casts if they cannot be eliminated. We allow
6357 // two casts to be inserted if the sizes are the same. This could
6358 // only be converting signedness, which is a noop.
6359 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6360 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6361 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6363 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6365 return BinaryOperator::create(
6366 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6371 case Instruction::Shl:
6372 // Allow changing the sign of the source operand. Do not allow
6373 // changing the size of the shift, UNLESS the shift amount is a
6374 // constant. We must not change variable sized shifts to a smaller
6375 // size, because it is undefined to shift more bits out than exist
6377 if (DestBitSize == SrcBitSize ||
6378 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6379 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6380 Instruction::BitCast : Instruction::Trunc);
6381 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6382 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6383 return BinaryOperator::createShl(Op0c, Op1c);
6386 case Instruction::AShr:
6387 // If this is a signed shr, and if all bits shifted in are about to be
6388 // truncated off, turn it into an unsigned shr to allow greater
6390 if (DestBitSize < SrcBitSize &&
6391 isa<ConstantInt>(Op1)) {
6392 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6393 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6394 // Insert the new logical shift right.
6395 return BinaryOperator::createLShr(Op0, Op1);
6400 case Instruction::ICmp:
6401 // If we are just checking for a icmp eq of a single bit and casting it
6402 // to an integer, then shift the bit to the appropriate place and then
6403 // cast to integer to avoid the comparison.
6404 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6405 APInt Op1CV(Op1C->getValue());
6406 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6407 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6408 // cast (X == 1) to int --> X iff X has only the low bit set.
6409 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6410 // cast (X != 0) to int --> X iff X has only the low bit set.
6411 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6412 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6413 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6414 if (Op1CV == 0 || Op1CV.isPowerOf2()) {
6415 // If Op1C some other power of two, convert:
6416 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6417 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6418 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6419 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6421 // This only works for EQ and NE
6422 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6423 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6426 APInt KnownZeroMask(KnownZero ^ TypeMask);
6427 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6428 bool isNE = pred == ICmpInst::ICMP_NE;
6429 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6430 // (X&4) == 2 --> false
6431 // (X&4) != 2 --> true
6432 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6433 Res = ConstantExpr::getZExt(Res, CI.getType());
6434 return ReplaceInstUsesWith(CI, Res);
6437 unsigned ShiftAmt = KnownZeroMask.logBase2();
6440 // Perform a logical shr by shiftamt.
6441 // Insert the shift to put the result in the low bit.
6442 In = InsertNewInstBefore(
6443 BinaryOperator::createLShr(In,
6444 ConstantInt::get(In->getType(), ShiftAmt),
6445 In->getName()+".lobit"), CI);
6448 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6449 Constant *One = ConstantInt::get(In->getType(), 1);
6450 In = BinaryOperator::createXor(In, One, "tmp");
6451 InsertNewInstBefore(cast<Instruction>(In), CI);
6454 if (CI.getType() == In->getType())
6455 return ReplaceInstUsesWith(CI, In);
6457 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6466 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6467 if (Instruction *Result = commonIntCastTransforms(CI))
6470 Value *Src = CI.getOperand(0);
6471 const Type *Ty = CI.getType();
6472 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6473 unsigned SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6475 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6476 switch (SrcI->getOpcode()) {
6478 case Instruction::LShr:
6479 // We can shrink lshr to something smaller if we know the bits shifted in
6480 // are already zeros.
6481 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6482 unsigned ShAmt = ShAmtV->getZExtValue();
6484 // Get a mask for the bits shifting in.
6485 APInt Mask(APInt::getAllOnesValue(SrcBitWidth).lshr(
6486 SrcBitWidth-ShAmt).shl(DestBitWidth));
6487 Value* SrcIOp0 = SrcI->getOperand(0);
6488 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6489 if (ShAmt >= DestBitWidth) // All zeros.
6490 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6492 // Okay, we can shrink this. Truncate the input, then return a new
6494 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6495 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6497 return BinaryOperator::createLShr(V1, V2);
6499 } else { // This is a variable shr.
6501 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6502 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6503 // loop-invariant and CSE'd.
6504 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6505 Value *One = ConstantInt::get(SrcI->getType(), 1);
6507 Value *V = InsertNewInstBefore(
6508 BinaryOperator::createShl(One, SrcI->getOperand(1),
6510 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6511 SrcI->getOperand(0),
6513 Value *Zero = Constant::getNullValue(V->getType());
6514 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6524 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6525 // If one of the common conversion will work ..
6526 if (Instruction *Result = commonIntCastTransforms(CI))
6529 Value *Src = CI.getOperand(0);
6531 // If this is a cast of a cast
6532 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6533 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6534 // types and if the sizes are just right we can convert this into a logical
6535 // 'and' which will be much cheaper than the pair of casts.
6536 if (isa<TruncInst>(CSrc)) {
6537 // Get the sizes of the types involved
6538 Value *A = CSrc->getOperand(0);
6539 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6540 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6541 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6542 // If we're actually extending zero bits and the trunc is a no-op
6543 if (MidSize < DstSize && SrcSize == DstSize) {
6544 // Replace both of the casts with an And of the type mask.
6545 APInt AndValue(APInt::getAllOnesValue(MidSize).zext(SrcSize));
6546 Constant *AndConst = ConstantInt::get(AndValue);
6548 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6549 // Unfortunately, if the type changed, we need to cast it back.
6550 if (And->getType() != CI.getType()) {
6551 And->setName(CSrc->getName()+".mask");
6552 InsertNewInstBefore(And, CI);
6553 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6563 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6564 return commonIntCastTransforms(CI);
6567 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6568 return commonCastTransforms(CI);
6571 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6572 return commonCastTransforms(CI);
6575 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6576 return commonCastTransforms(CI);
6579 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6580 return commonCastTransforms(CI);
6583 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6584 return commonCastTransforms(CI);
6587 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6588 return commonCastTransforms(CI);
6591 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6592 return commonCastTransforms(CI);
6595 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6596 return commonCastTransforms(CI);
6599 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6601 // If the operands are integer typed then apply the integer transforms,
6602 // otherwise just apply the common ones.
6603 Value *Src = CI.getOperand(0);
6604 const Type *SrcTy = Src->getType();
6605 const Type *DestTy = CI.getType();
6607 if (SrcTy->isInteger() && DestTy->isInteger()) {
6608 if (Instruction *Result = commonIntCastTransforms(CI))
6611 if (Instruction *Result = commonCastTransforms(CI))
6616 // Get rid of casts from one type to the same type. These are useless and can
6617 // be replaced by the operand.
6618 if (DestTy == Src->getType())
6619 return ReplaceInstUsesWith(CI, Src);
6621 // If the source and destination are pointers, and this cast is equivalent to
6622 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6623 // This can enhance SROA and other transforms that want type-safe pointers.
6624 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6625 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6626 const Type *DstElTy = DstPTy->getElementType();
6627 const Type *SrcElTy = SrcPTy->getElementType();
6629 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6630 unsigned NumZeros = 0;
6631 while (SrcElTy != DstElTy &&
6632 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6633 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6634 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6638 // If we found a path from the src to dest, create the getelementptr now.
6639 if (SrcElTy == DstElTy) {
6640 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6641 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6646 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6647 if (SVI->hasOneUse()) {
6648 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6649 // a bitconvert to a vector with the same # elts.
6650 if (isa<VectorType>(DestTy) &&
6651 cast<VectorType>(DestTy)->getNumElements() ==
6652 SVI->getType()->getNumElements()) {
6654 // If either of the operands is a cast from CI.getType(), then
6655 // evaluating the shuffle in the casted destination's type will allow
6656 // us to eliminate at least one cast.
6657 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6658 Tmp->getOperand(0)->getType() == DestTy) ||
6659 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6660 Tmp->getOperand(0)->getType() == DestTy)) {
6661 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6662 SVI->getOperand(0), DestTy, &CI);
6663 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6664 SVI->getOperand(1), DestTy, &CI);
6665 // Return a new shuffle vector. Use the same element ID's, as we
6666 // know the vector types match #elts.
6667 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6675 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6677 /// %D = select %cond, %C, %A
6679 /// %C = select %cond, %B, 0
6682 /// Assuming that the specified instruction is an operand to the select, return
6683 /// a bitmask indicating which operands of this instruction are foldable if they
6684 /// equal the other incoming value of the select.
6686 static unsigned GetSelectFoldableOperands(Instruction *I) {
6687 switch (I->getOpcode()) {
6688 case Instruction::Add:
6689 case Instruction::Mul:
6690 case Instruction::And:
6691 case Instruction::Or:
6692 case Instruction::Xor:
6693 return 3; // Can fold through either operand.
6694 case Instruction::Sub: // Can only fold on the amount subtracted.
6695 case Instruction::Shl: // Can only fold on the shift amount.
6696 case Instruction::LShr:
6697 case Instruction::AShr:
6700 return 0; // Cannot fold
6704 /// GetSelectFoldableConstant - For the same transformation as the previous
6705 /// function, return the identity constant that goes into the select.
6706 static Constant *GetSelectFoldableConstant(Instruction *I) {
6707 switch (I->getOpcode()) {
6708 default: assert(0 && "This cannot happen!"); abort();
6709 case Instruction::Add:
6710 case Instruction::Sub:
6711 case Instruction::Or:
6712 case Instruction::Xor:
6713 case Instruction::Shl:
6714 case Instruction::LShr:
6715 case Instruction::AShr:
6716 return Constant::getNullValue(I->getType());
6717 case Instruction::And:
6718 return ConstantInt::getAllOnesValue(I->getType());
6719 case Instruction::Mul:
6720 return ConstantInt::get(I->getType(), 1);
6724 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6725 /// have the same opcode and only one use each. Try to simplify this.
6726 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6728 if (TI->getNumOperands() == 1) {
6729 // If this is a non-volatile load or a cast from the same type,
6732 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6735 return 0; // unknown unary op.
6738 // Fold this by inserting a select from the input values.
6739 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6740 FI->getOperand(0), SI.getName()+".v");
6741 InsertNewInstBefore(NewSI, SI);
6742 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6746 // Only handle binary operators here.
6747 if (!isa<BinaryOperator>(TI))
6750 // Figure out if the operations have any operands in common.
6751 Value *MatchOp, *OtherOpT, *OtherOpF;
6753 if (TI->getOperand(0) == FI->getOperand(0)) {
6754 MatchOp = TI->getOperand(0);
6755 OtherOpT = TI->getOperand(1);
6756 OtherOpF = FI->getOperand(1);
6757 MatchIsOpZero = true;
6758 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6759 MatchOp = TI->getOperand(1);
6760 OtherOpT = TI->getOperand(0);
6761 OtherOpF = FI->getOperand(0);
6762 MatchIsOpZero = false;
6763 } else if (!TI->isCommutative()) {
6765 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6766 MatchOp = TI->getOperand(0);
6767 OtherOpT = TI->getOperand(1);
6768 OtherOpF = FI->getOperand(0);
6769 MatchIsOpZero = true;
6770 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6771 MatchOp = TI->getOperand(1);
6772 OtherOpT = TI->getOperand(0);
6773 OtherOpF = FI->getOperand(1);
6774 MatchIsOpZero = true;
6779 // If we reach here, they do have operations in common.
6780 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6781 OtherOpF, SI.getName()+".v");
6782 InsertNewInstBefore(NewSI, SI);
6784 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6786 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6788 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6790 assert(0 && "Shouldn't get here");
6794 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6795 Value *CondVal = SI.getCondition();
6796 Value *TrueVal = SI.getTrueValue();
6797 Value *FalseVal = SI.getFalseValue();
6799 // select true, X, Y -> X
6800 // select false, X, Y -> Y
6801 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6802 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6804 // select C, X, X -> X
6805 if (TrueVal == FalseVal)
6806 return ReplaceInstUsesWith(SI, TrueVal);
6808 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6809 return ReplaceInstUsesWith(SI, FalseVal);
6810 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6811 return ReplaceInstUsesWith(SI, TrueVal);
6812 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6813 if (isa<Constant>(TrueVal))
6814 return ReplaceInstUsesWith(SI, TrueVal);
6816 return ReplaceInstUsesWith(SI, FalseVal);
6819 if (SI.getType() == Type::Int1Ty) {
6820 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6821 if (C->getZExtValue()) {
6822 // Change: A = select B, true, C --> A = or B, C
6823 return BinaryOperator::createOr(CondVal, FalseVal);
6825 // Change: A = select B, false, C --> A = and !B, C
6827 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6828 "not."+CondVal->getName()), SI);
6829 return BinaryOperator::createAnd(NotCond, FalseVal);
6831 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6832 if (C->getZExtValue() == false) {
6833 // Change: A = select B, C, false --> A = and B, C
6834 return BinaryOperator::createAnd(CondVal, TrueVal);
6836 // Change: A = select B, C, true --> A = or !B, C
6838 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6839 "not."+CondVal->getName()), SI);
6840 return BinaryOperator::createOr(NotCond, TrueVal);
6845 // Selecting between two integer constants?
6846 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6847 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6848 // select C, 1, 0 -> cast C to int
6849 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
6850 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6851 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
6852 // select C, 0, 1 -> cast !C to int
6854 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6855 "not."+CondVal->getName()), SI);
6856 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6859 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6861 // (x <s 0) ? -1 : 0 -> ashr x, 31
6862 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6863 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
6864 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6865 bool CanXForm = false;
6866 if (IC->isSignedPredicate())
6867 CanXForm = CmpCst->isZero() &&
6868 IC->getPredicate() == ICmpInst::ICMP_SLT;
6870 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6871 CanXForm = CmpCst->getValue() == APInt::getSignedMaxValue(Bits) &&
6872 IC->getPredicate() == ICmpInst::ICMP_UGT;
6876 // The comparison constant and the result are not neccessarily the
6877 // same width. Make an all-ones value by inserting a AShr.
6878 Value *X = IC->getOperand(0);
6879 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6880 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6881 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6883 InsertNewInstBefore(SRA, SI);
6885 // Finally, convert to the type of the select RHS. We figure out
6886 // if this requires a SExt, Trunc or BitCast based on the sizes.
6887 Instruction::CastOps opc = Instruction::BitCast;
6888 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6889 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6890 if (SRASize < SISize)
6891 opc = Instruction::SExt;
6892 else if (SRASize > SISize)
6893 opc = Instruction::Trunc;
6894 return CastInst::create(opc, SRA, SI.getType());
6899 // If one of the constants is zero (we know they can't both be) and we
6900 // have a fcmp instruction with zero, and we have an 'and' with the
6901 // non-constant value, eliminate this whole mess. This corresponds to
6902 // cases like this: ((X & 27) ? 27 : 0)
6903 if (TrueValC->isZero() || FalseValC->isZero())
6904 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6905 cast<Constant>(IC->getOperand(1))->isNullValue())
6906 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6907 if (ICA->getOpcode() == Instruction::And &&
6908 isa<ConstantInt>(ICA->getOperand(1)) &&
6909 (ICA->getOperand(1) == TrueValC ||
6910 ICA->getOperand(1) == FalseValC) &&
6911 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6912 // Okay, now we know that everything is set up, we just don't
6913 // know whether we have a icmp_ne or icmp_eq and whether the
6914 // true or false val is the zero.
6915 bool ShouldNotVal = !TrueValC->isZero();
6916 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6919 V = InsertNewInstBefore(BinaryOperator::create(
6920 Instruction::Xor, V, ICA->getOperand(1)), SI);
6921 return ReplaceInstUsesWith(SI, V);
6926 // See if we are selecting two values based on a comparison of the two values.
6927 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6928 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6929 // Transform (X == Y) ? X : Y -> Y
6930 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6931 return ReplaceInstUsesWith(SI, FalseVal);
6932 // Transform (X != Y) ? X : Y -> X
6933 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6934 return ReplaceInstUsesWith(SI, TrueVal);
6935 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6937 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6938 // Transform (X == Y) ? Y : X -> X
6939 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6940 return ReplaceInstUsesWith(SI, FalseVal);
6941 // Transform (X != Y) ? Y : X -> Y
6942 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6943 return ReplaceInstUsesWith(SI, TrueVal);
6944 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6948 // See if we are selecting two values based on a comparison of the two values.
6949 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6950 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6951 // Transform (X == Y) ? X : Y -> Y
6952 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6953 return ReplaceInstUsesWith(SI, FalseVal);
6954 // Transform (X != Y) ? X : Y -> X
6955 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6956 return ReplaceInstUsesWith(SI, TrueVal);
6957 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6959 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6960 // Transform (X == Y) ? Y : X -> X
6961 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6962 return ReplaceInstUsesWith(SI, FalseVal);
6963 // Transform (X != Y) ? Y : X -> Y
6964 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6965 return ReplaceInstUsesWith(SI, TrueVal);
6966 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6970 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6971 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6972 if (TI->hasOneUse() && FI->hasOneUse()) {
6973 Instruction *AddOp = 0, *SubOp = 0;
6975 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6976 if (TI->getOpcode() == FI->getOpcode())
6977 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6980 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6981 // even legal for FP.
6982 if (TI->getOpcode() == Instruction::Sub &&
6983 FI->getOpcode() == Instruction::Add) {
6984 AddOp = FI; SubOp = TI;
6985 } else if (FI->getOpcode() == Instruction::Sub &&
6986 TI->getOpcode() == Instruction::Add) {
6987 AddOp = TI; SubOp = FI;
6991 Value *OtherAddOp = 0;
6992 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6993 OtherAddOp = AddOp->getOperand(1);
6994 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6995 OtherAddOp = AddOp->getOperand(0);
6999 // So at this point we know we have (Y -> OtherAddOp):
7000 // select C, (add X, Y), (sub X, Z)
7001 Value *NegVal; // Compute -Z
7002 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7003 NegVal = ConstantExpr::getNeg(C);
7005 NegVal = InsertNewInstBefore(
7006 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7009 Value *NewTrueOp = OtherAddOp;
7010 Value *NewFalseOp = NegVal;
7012 std::swap(NewTrueOp, NewFalseOp);
7013 Instruction *NewSel =
7014 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7016 NewSel = InsertNewInstBefore(NewSel, SI);
7017 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7022 // See if we can fold the select into one of our operands.
7023 if (SI.getType()->isInteger()) {
7024 // See the comment above GetSelectFoldableOperands for a description of the
7025 // transformation we are doing here.
7026 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7027 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7028 !isa<Constant>(FalseVal))
7029 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7030 unsigned OpToFold = 0;
7031 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7033 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7038 Constant *C = GetSelectFoldableConstant(TVI);
7039 Instruction *NewSel =
7040 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7041 InsertNewInstBefore(NewSel, SI);
7042 NewSel->takeName(TVI);
7043 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7044 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7046 assert(0 && "Unknown instruction!!");
7051 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7052 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7053 !isa<Constant>(TrueVal))
7054 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7055 unsigned OpToFold = 0;
7056 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7058 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7063 Constant *C = GetSelectFoldableConstant(FVI);
7064 Instruction *NewSel =
7065 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7066 InsertNewInstBefore(NewSel, SI);
7067 NewSel->takeName(FVI);
7068 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7069 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7071 assert(0 && "Unknown instruction!!");
7076 if (BinaryOperator::isNot(CondVal)) {
7077 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7078 SI.setOperand(1, FalseVal);
7079 SI.setOperand(2, TrueVal);
7086 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7087 /// determine, return it, otherwise return 0.
7088 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7089 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7090 unsigned Align = GV->getAlignment();
7091 if (Align == 0 && TD)
7092 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7094 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7095 unsigned Align = AI->getAlignment();
7096 if (Align == 0 && TD) {
7097 if (isa<AllocaInst>(AI))
7098 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7099 else if (isa<MallocInst>(AI)) {
7100 // Malloc returns maximally aligned memory.
7101 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7104 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7107 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7111 } else if (isa<BitCastInst>(V) ||
7112 (isa<ConstantExpr>(V) &&
7113 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7114 User *CI = cast<User>(V);
7115 if (isa<PointerType>(CI->getOperand(0)->getType()))
7116 return GetKnownAlignment(CI->getOperand(0), TD);
7118 } else if (isa<GetElementPtrInst>(V) ||
7119 (isa<ConstantExpr>(V) &&
7120 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7121 User *GEPI = cast<User>(V);
7122 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7123 if (BaseAlignment == 0) return 0;
7125 // If all indexes are zero, it is just the alignment of the base pointer.
7126 bool AllZeroOperands = true;
7127 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7128 if (!isa<Constant>(GEPI->getOperand(i)) ||
7129 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7130 AllZeroOperands = false;
7133 if (AllZeroOperands)
7134 return BaseAlignment;
7136 // Otherwise, if the base alignment is >= the alignment we expect for the
7137 // base pointer type, then we know that the resultant pointer is aligned at
7138 // least as much as its type requires.
7141 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7142 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7143 if (TD->getABITypeAlignment(PtrTy->getElementType())
7145 const Type *GEPTy = GEPI->getType();
7146 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7147 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7155 /// visitCallInst - CallInst simplification. This mostly only handles folding
7156 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7157 /// the heavy lifting.
7159 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7160 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7161 if (!II) return visitCallSite(&CI);
7163 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7165 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7166 bool Changed = false;
7168 // memmove/cpy/set of zero bytes is a noop.
7169 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7170 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7172 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7173 if (CI->getZExtValue() == 1) {
7174 // Replace the instruction with just byte operations. We would
7175 // transform other cases to loads/stores, but we don't know if
7176 // alignment is sufficient.
7180 // If we have a memmove and the source operation is a constant global,
7181 // then the source and dest pointers can't alias, so we can change this
7182 // into a call to memcpy.
7183 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7184 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7185 if (GVSrc->isConstant()) {
7186 Module *M = CI.getParent()->getParent()->getParent();
7188 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7190 Name = "llvm.memcpy.i32";
7192 Name = "llvm.memcpy.i64";
7193 Constant *MemCpy = M->getOrInsertFunction(Name,
7194 CI.getCalledFunction()->getFunctionType());
7195 CI.setOperand(0, MemCpy);
7200 // If we can determine a pointer alignment that is bigger than currently
7201 // set, update the alignment.
7202 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7203 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7204 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7205 unsigned Align = std::min(Alignment1, Alignment2);
7206 if (MI->getAlignment()->getZExtValue() < Align) {
7207 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7210 } else if (isa<MemSetInst>(MI)) {
7211 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7212 if (MI->getAlignment()->getZExtValue() < Alignment) {
7213 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7218 if (Changed) return II;
7220 switch (II->getIntrinsicID()) {
7222 case Intrinsic::ppc_altivec_lvx:
7223 case Intrinsic::ppc_altivec_lvxl:
7224 case Intrinsic::x86_sse_loadu_ps:
7225 case Intrinsic::x86_sse2_loadu_pd:
7226 case Intrinsic::x86_sse2_loadu_dq:
7227 // Turn PPC lvx -> load if the pointer is known aligned.
7228 // Turn X86 loadups -> load if the pointer is known aligned.
7229 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7230 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7231 PointerType::get(II->getType()), CI);
7232 return new LoadInst(Ptr);
7235 case Intrinsic::ppc_altivec_stvx:
7236 case Intrinsic::ppc_altivec_stvxl:
7237 // Turn stvx -> store if the pointer is known aligned.
7238 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7239 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7240 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7242 return new StoreInst(II->getOperand(1), Ptr);
7245 case Intrinsic::x86_sse_storeu_ps:
7246 case Intrinsic::x86_sse2_storeu_pd:
7247 case Intrinsic::x86_sse2_storeu_dq:
7248 case Intrinsic::x86_sse2_storel_dq:
7249 // Turn X86 storeu -> store if the pointer is known aligned.
7250 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7251 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7252 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7254 return new StoreInst(II->getOperand(2), Ptr);
7258 case Intrinsic::x86_sse_cvttss2si: {
7259 // These intrinsics only demands the 0th element of its input vector. If
7260 // we can simplify the input based on that, do so now.
7262 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7264 II->setOperand(1, V);
7270 case Intrinsic::ppc_altivec_vperm:
7271 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7272 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7273 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7275 // Check that all of the elements are integer constants or undefs.
7276 bool AllEltsOk = true;
7277 for (unsigned i = 0; i != 16; ++i) {
7278 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7279 !isa<UndefValue>(Mask->getOperand(i))) {
7286 // Cast the input vectors to byte vectors.
7287 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7288 II->getOperand(1), Mask->getType(), CI);
7289 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7290 II->getOperand(2), Mask->getType(), CI);
7291 Value *Result = UndefValue::get(Op0->getType());
7293 // Only extract each element once.
7294 Value *ExtractedElts[32];
7295 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7297 for (unsigned i = 0; i != 16; ++i) {
7298 if (isa<UndefValue>(Mask->getOperand(i)))
7300 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7301 Idx &= 31; // Match the hardware behavior.
7303 if (ExtractedElts[Idx] == 0) {
7305 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7306 InsertNewInstBefore(Elt, CI);
7307 ExtractedElts[Idx] = Elt;
7310 // Insert this value into the result vector.
7311 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7312 InsertNewInstBefore(cast<Instruction>(Result), CI);
7314 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7319 case Intrinsic::stackrestore: {
7320 // If the save is right next to the restore, remove the restore. This can
7321 // happen when variable allocas are DCE'd.
7322 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7323 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7324 BasicBlock::iterator BI = SS;
7326 return EraseInstFromFunction(CI);
7330 // If the stack restore is in a return/unwind block and if there are no
7331 // allocas or calls between the restore and the return, nuke the restore.
7332 TerminatorInst *TI = II->getParent()->getTerminator();
7333 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7334 BasicBlock::iterator BI = II;
7335 bool CannotRemove = false;
7336 for (++BI; &*BI != TI; ++BI) {
7337 if (isa<AllocaInst>(BI) ||
7338 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7339 CannotRemove = true;
7344 return EraseInstFromFunction(CI);
7351 return visitCallSite(II);
7354 // InvokeInst simplification
7356 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7357 return visitCallSite(&II);
7360 // visitCallSite - Improvements for call and invoke instructions.
7362 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7363 bool Changed = false;
7365 // If the callee is a constexpr cast of a function, attempt to move the cast
7366 // to the arguments of the call/invoke.
7367 if (transformConstExprCastCall(CS)) return 0;
7369 Value *Callee = CS.getCalledValue();
7371 if (Function *CalleeF = dyn_cast<Function>(Callee))
7372 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7373 Instruction *OldCall = CS.getInstruction();
7374 // If the call and callee calling conventions don't match, this call must
7375 // be unreachable, as the call is undefined.
7376 new StoreInst(ConstantInt::getTrue(),
7377 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7378 if (!OldCall->use_empty())
7379 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7380 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7381 return EraseInstFromFunction(*OldCall);
7385 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7386 // This instruction is not reachable, just remove it. We insert a store to
7387 // undef so that we know that this code is not reachable, despite the fact
7388 // that we can't modify the CFG here.
7389 new StoreInst(ConstantInt::getTrue(),
7390 UndefValue::get(PointerType::get(Type::Int1Ty)),
7391 CS.getInstruction());
7393 if (!CS.getInstruction()->use_empty())
7394 CS.getInstruction()->
7395 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7397 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7398 // Don't break the CFG, insert a dummy cond branch.
7399 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7400 ConstantInt::getTrue(), II);
7402 return EraseInstFromFunction(*CS.getInstruction());
7405 const PointerType *PTy = cast<PointerType>(Callee->getType());
7406 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7407 if (FTy->isVarArg()) {
7408 // See if we can optimize any arguments passed through the varargs area of
7410 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7411 E = CS.arg_end(); I != E; ++I)
7412 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7413 // If this cast does not effect the value passed through the varargs
7414 // area, we can eliminate the use of the cast.
7415 Value *Op = CI->getOperand(0);
7416 if (CI->isLosslessCast()) {
7423 return Changed ? CS.getInstruction() : 0;
7426 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7427 // attempt to move the cast to the arguments of the call/invoke.
7429 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7430 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7431 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7432 if (CE->getOpcode() != Instruction::BitCast ||
7433 !isa<Function>(CE->getOperand(0)))
7435 Function *Callee = cast<Function>(CE->getOperand(0));
7436 Instruction *Caller = CS.getInstruction();
7438 // Okay, this is a cast from a function to a different type. Unless doing so
7439 // would cause a type conversion of one of our arguments, change this call to
7440 // be a direct call with arguments casted to the appropriate types.
7442 const FunctionType *FT = Callee->getFunctionType();
7443 const Type *OldRetTy = Caller->getType();
7445 // Check to see if we are changing the return type...
7446 if (OldRetTy != FT->getReturnType()) {
7447 if (Callee->isDeclaration() && !Caller->use_empty() &&
7448 // Conversion is ok if changing from pointer to int of same size.
7449 !(isa<PointerType>(FT->getReturnType()) &&
7450 TD->getIntPtrType() == OldRetTy))
7451 return false; // Cannot transform this return value.
7453 // If the callsite is an invoke instruction, and the return value is used by
7454 // a PHI node in a successor, we cannot change the return type of the call
7455 // because there is no place to put the cast instruction (without breaking
7456 // the critical edge). Bail out in this case.
7457 if (!Caller->use_empty())
7458 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7459 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7461 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7462 if (PN->getParent() == II->getNormalDest() ||
7463 PN->getParent() == II->getUnwindDest())
7467 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7468 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7470 CallSite::arg_iterator AI = CS.arg_begin();
7471 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7472 const Type *ParamTy = FT->getParamType(i);
7473 const Type *ActTy = (*AI)->getType();
7474 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7475 //Either we can cast directly, or we can upconvert the argument
7476 bool isConvertible = ActTy == ParamTy ||
7477 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7478 (ParamTy->isInteger() && ActTy->isInteger() &&
7479 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7480 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7481 && c->getValue().isStrictlyPositive());
7482 if (Callee->isDeclaration() && !isConvertible) return false;
7485 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7486 Callee->isDeclaration())
7487 return false; // Do not delete arguments unless we have a function body...
7489 // Okay, we decided that this is a safe thing to do: go ahead and start
7490 // inserting cast instructions as necessary...
7491 std::vector<Value*> Args;
7492 Args.reserve(NumActualArgs);
7494 AI = CS.arg_begin();
7495 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7496 const Type *ParamTy = FT->getParamType(i);
7497 if ((*AI)->getType() == ParamTy) {
7498 Args.push_back(*AI);
7500 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7501 false, ParamTy, false);
7502 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7503 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7507 // If the function takes more arguments than the call was taking, add them
7509 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7510 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7512 // If we are removing arguments to the function, emit an obnoxious warning...
7513 if (FT->getNumParams() < NumActualArgs)
7514 if (!FT->isVarArg()) {
7515 cerr << "WARNING: While resolving call to function '"
7516 << Callee->getName() << "' arguments were dropped!\n";
7518 // Add all of the arguments in their promoted form to the arg list...
7519 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7520 const Type *PTy = getPromotedType((*AI)->getType());
7521 if (PTy != (*AI)->getType()) {
7522 // Must promote to pass through va_arg area!
7523 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7525 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7526 InsertNewInstBefore(Cast, *Caller);
7527 Args.push_back(Cast);
7529 Args.push_back(*AI);
7534 if (FT->getReturnType() == Type::VoidTy)
7535 Caller->setName(""); // Void type should not have a name.
7538 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7539 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7540 &Args[0], Args.size(), Caller->getName(), Caller);
7541 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7543 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7544 if (cast<CallInst>(Caller)->isTailCall())
7545 cast<CallInst>(NC)->setTailCall();
7546 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7549 // Insert a cast of the return type as necessary.
7551 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7552 if (NV->getType() != Type::VoidTy) {
7553 const Type *CallerTy = Caller->getType();
7554 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7556 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7558 // If this is an invoke instruction, we should insert it after the first
7559 // non-phi, instruction in the normal successor block.
7560 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7561 BasicBlock::iterator I = II->getNormalDest()->begin();
7562 while (isa<PHINode>(I)) ++I;
7563 InsertNewInstBefore(NC, *I);
7565 // Otherwise, it's a call, just insert cast right after the call instr
7566 InsertNewInstBefore(NC, *Caller);
7568 AddUsersToWorkList(*Caller);
7570 NV = UndefValue::get(Caller->getType());
7574 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7575 Caller->replaceAllUsesWith(NV);
7576 Caller->eraseFromParent();
7577 RemoveFromWorkList(Caller);
7581 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7582 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7583 /// and a single binop.
7584 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7585 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7586 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7587 isa<CmpInst>(FirstInst));
7588 unsigned Opc = FirstInst->getOpcode();
7589 Value *LHSVal = FirstInst->getOperand(0);
7590 Value *RHSVal = FirstInst->getOperand(1);
7592 const Type *LHSType = LHSVal->getType();
7593 const Type *RHSType = RHSVal->getType();
7595 // Scan to see if all operands are the same opcode, all have one use, and all
7596 // kill their operands (i.e. the operands have one use).
7597 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7598 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7599 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7600 // Verify type of the LHS matches so we don't fold cmp's of different
7601 // types or GEP's with different index types.
7602 I->getOperand(0)->getType() != LHSType ||
7603 I->getOperand(1)->getType() != RHSType)
7606 // If they are CmpInst instructions, check their predicates
7607 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7608 if (cast<CmpInst>(I)->getPredicate() !=
7609 cast<CmpInst>(FirstInst)->getPredicate())
7612 // Keep track of which operand needs a phi node.
7613 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7614 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7617 // Otherwise, this is safe to transform, determine if it is profitable.
7619 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7620 // Indexes are often folded into load/store instructions, so we don't want to
7621 // hide them behind a phi.
7622 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7625 Value *InLHS = FirstInst->getOperand(0);
7626 Value *InRHS = FirstInst->getOperand(1);
7627 PHINode *NewLHS = 0, *NewRHS = 0;
7629 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7630 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7631 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7632 InsertNewInstBefore(NewLHS, PN);
7637 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7638 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7639 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7640 InsertNewInstBefore(NewRHS, PN);
7644 // Add all operands to the new PHIs.
7645 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7647 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7648 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7651 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7652 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7656 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7657 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7658 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7659 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7662 assert(isa<GetElementPtrInst>(FirstInst));
7663 return new GetElementPtrInst(LHSVal, RHSVal);
7667 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7668 /// of the block that defines it. This means that it must be obvious the value
7669 /// of the load is not changed from the point of the load to the end of the
7672 /// Finally, it is safe, but not profitable, to sink a load targetting a
7673 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7675 static bool isSafeToSinkLoad(LoadInst *L) {
7676 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7678 for (++BBI; BBI != E; ++BBI)
7679 if (BBI->mayWriteToMemory())
7682 // Check for non-address taken alloca. If not address-taken already, it isn't
7683 // profitable to do this xform.
7684 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7685 bool isAddressTaken = false;
7686 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7688 if (isa<LoadInst>(UI)) continue;
7689 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7690 // If storing TO the alloca, then the address isn't taken.
7691 if (SI->getOperand(1) == AI) continue;
7693 isAddressTaken = true;
7697 if (!isAddressTaken)
7705 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7706 // operator and they all are only used by the PHI, PHI together their
7707 // inputs, and do the operation once, to the result of the PHI.
7708 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7709 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7711 // Scan the instruction, looking for input operations that can be folded away.
7712 // If all input operands to the phi are the same instruction (e.g. a cast from
7713 // the same type or "+42") we can pull the operation through the PHI, reducing
7714 // code size and simplifying code.
7715 Constant *ConstantOp = 0;
7716 const Type *CastSrcTy = 0;
7717 bool isVolatile = false;
7718 if (isa<CastInst>(FirstInst)) {
7719 CastSrcTy = FirstInst->getOperand(0)->getType();
7720 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7721 // Can fold binop, compare or shift here if the RHS is a constant,
7722 // otherwise call FoldPHIArgBinOpIntoPHI.
7723 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7724 if (ConstantOp == 0)
7725 return FoldPHIArgBinOpIntoPHI(PN);
7726 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7727 isVolatile = LI->isVolatile();
7728 // We can't sink the load if the loaded value could be modified between the
7729 // load and the PHI.
7730 if (LI->getParent() != PN.getIncomingBlock(0) ||
7731 !isSafeToSinkLoad(LI))
7733 } else if (isa<GetElementPtrInst>(FirstInst)) {
7734 if (FirstInst->getNumOperands() == 2)
7735 return FoldPHIArgBinOpIntoPHI(PN);
7736 // Can't handle general GEPs yet.
7739 return 0; // Cannot fold this operation.
7742 // Check to see if all arguments are the same operation.
7743 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7744 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7745 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7746 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7749 if (I->getOperand(0)->getType() != CastSrcTy)
7750 return 0; // Cast operation must match.
7751 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7752 // We can't sink the load if the loaded value could be modified between
7753 // the load and the PHI.
7754 if (LI->isVolatile() != isVolatile ||
7755 LI->getParent() != PN.getIncomingBlock(i) ||
7756 !isSafeToSinkLoad(LI))
7758 } else if (I->getOperand(1) != ConstantOp) {
7763 // Okay, they are all the same operation. Create a new PHI node of the
7764 // correct type, and PHI together all of the LHS's of the instructions.
7765 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7766 PN.getName()+".in");
7767 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7769 Value *InVal = FirstInst->getOperand(0);
7770 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7772 // Add all operands to the new PHI.
7773 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7774 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7775 if (NewInVal != InVal)
7777 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7782 // The new PHI unions all of the same values together. This is really
7783 // common, so we handle it intelligently here for compile-time speed.
7787 InsertNewInstBefore(NewPN, PN);
7791 // Insert and return the new operation.
7792 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7793 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7794 else if (isa<LoadInst>(FirstInst))
7795 return new LoadInst(PhiVal, "", isVolatile);
7796 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7797 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7798 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7799 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7800 PhiVal, ConstantOp);
7802 assert(0 && "Unknown operation");
7806 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7808 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7809 if (PN->use_empty()) return true;
7810 if (!PN->hasOneUse()) return false;
7812 // Remember this node, and if we find the cycle, return.
7813 if (!PotentiallyDeadPHIs.insert(PN).second)
7816 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7817 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7822 // PHINode simplification
7824 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7825 // If LCSSA is around, don't mess with Phi nodes
7826 if (MustPreserveLCSSA) return 0;
7828 if (Value *V = PN.hasConstantValue())
7829 return ReplaceInstUsesWith(PN, V);
7831 // If all PHI operands are the same operation, pull them through the PHI,
7832 // reducing code size.
7833 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7834 PN.getIncomingValue(0)->hasOneUse())
7835 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7838 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7839 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7840 // PHI)... break the cycle.
7841 if (PN.hasOneUse()) {
7842 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7843 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7844 std::set<PHINode*> PotentiallyDeadPHIs;
7845 PotentiallyDeadPHIs.insert(&PN);
7846 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7847 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7850 // If this phi has a single use, and if that use just computes a value for
7851 // the next iteration of a loop, delete the phi. This occurs with unused
7852 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7853 // common case here is good because the only other things that catch this
7854 // are induction variable analysis (sometimes) and ADCE, which is only run
7856 if (PHIUser->hasOneUse() &&
7857 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7858 PHIUser->use_back() == &PN) {
7859 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7866 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7867 Instruction *InsertPoint,
7869 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7870 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7871 // We must cast correctly to the pointer type. Ensure that we
7872 // sign extend the integer value if it is smaller as this is
7873 // used for address computation.
7874 Instruction::CastOps opcode =
7875 (VTySize < PtrSize ? Instruction::SExt :
7876 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7877 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7881 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7882 Value *PtrOp = GEP.getOperand(0);
7883 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7884 // If so, eliminate the noop.
7885 if (GEP.getNumOperands() == 1)
7886 return ReplaceInstUsesWith(GEP, PtrOp);
7888 if (isa<UndefValue>(GEP.getOperand(0)))
7889 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7891 bool HasZeroPointerIndex = false;
7892 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7893 HasZeroPointerIndex = C->isNullValue();
7895 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7896 return ReplaceInstUsesWith(GEP, PtrOp);
7898 // Keep track of whether all indices are zero constants integers.
7899 bool AllZeroIndices = true;
7901 // Eliminate unneeded casts for indices.
7902 bool MadeChange = false;
7904 gep_type_iterator GTI = gep_type_begin(GEP);
7905 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
7906 // Track whether this GEP has all zero indices, if so, it doesn't move the
7907 // input pointer, it just changes its type.
7908 if (AllZeroIndices) {
7909 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(i)))
7910 AllZeroIndices = CI->isNullValue();
7912 AllZeroIndices = false;
7914 if (isa<SequentialType>(*GTI)) {
7915 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7916 if (CI->getOpcode() == Instruction::ZExt ||
7917 CI->getOpcode() == Instruction::SExt) {
7918 const Type *SrcTy = CI->getOperand(0)->getType();
7919 // We can eliminate a cast from i32 to i64 iff the target
7920 // is a 32-bit pointer target.
7921 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7923 GEP.setOperand(i, CI->getOperand(0));
7927 // If we are using a wider index than needed for this platform, shrink it
7928 // to what we need. If the incoming value needs a cast instruction,
7929 // insert it. This explicit cast can make subsequent optimizations more
7931 Value *Op = GEP.getOperand(i);
7932 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7933 if (Constant *C = dyn_cast<Constant>(Op)) {
7934 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7937 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7939 GEP.setOperand(i, Op);
7944 if (MadeChange) return &GEP;
7946 // If this GEP instruction doesn't move the pointer, and if the input operand
7947 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
7948 // real input to the dest type.
7949 if (AllZeroIndices && isa<BitCastInst>(GEP.getOperand(0)))
7950 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
7953 // Combine Indices - If the source pointer to this getelementptr instruction
7954 // is a getelementptr instruction, combine the indices of the two
7955 // getelementptr instructions into a single instruction.
7957 SmallVector<Value*, 8> SrcGEPOperands;
7958 if (User *Src = dyn_castGetElementPtr(PtrOp))
7959 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
7961 if (!SrcGEPOperands.empty()) {
7962 // Note that if our source is a gep chain itself that we wait for that
7963 // chain to be resolved before we perform this transformation. This
7964 // avoids us creating a TON of code in some cases.
7966 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7967 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7968 return 0; // Wait until our source is folded to completion.
7970 SmallVector<Value*, 8> Indices;
7972 // Find out whether the last index in the source GEP is a sequential idx.
7973 bool EndsWithSequential = false;
7974 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7975 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7976 EndsWithSequential = !isa<StructType>(*I);
7978 // Can we combine the two pointer arithmetics offsets?
7979 if (EndsWithSequential) {
7980 // Replace: gep (gep %P, long B), long A, ...
7981 // With: T = long A+B; gep %P, T, ...
7983 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7984 if (SO1 == Constant::getNullValue(SO1->getType())) {
7986 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7989 // If they aren't the same type, convert both to an integer of the
7990 // target's pointer size.
7991 if (SO1->getType() != GO1->getType()) {
7992 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7993 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7994 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7995 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7997 unsigned PS = TD->getPointerSize();
7998 if (TD->getTypeSize(SO1->getType()) == PS) {
7999 // Convert GO1 to SO1's type.
8000 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8002 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8003 // Convert SO1 to GO1's type.
8004 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8006 const Type *PT = TD->getIntPtrType();
8007 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8008 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8012 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8013 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8015 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8016 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8020 // Recycle the GEP we already have if possible.
8021 if (SrcGEPOperands.size() == 2) {
8022 GEP.setOperand(0, SrcGEPOperands[0]);
8023 GEP.setOperand(1, Sum);
8026 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8027 SrcGEPOperands.end()-1);
8028 Indices.push_back(Sum);
8029 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8031 } else if (isa<Constant>(*GEP.idx_begin()) &&
8032 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8033 SrcGEPOperands.size() != 1) {
8034 // Otherwise we can do the fold if the first index of the GEP is a zero
8035 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8036 SrcGEPOperands.end());
8037 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8040 if (!Indices.empty())
8041 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8042 Indices.size(), GEP.getName());
8044 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8045 // GEP of global variable. If all of the indices for this GEP are
8046 // constants, we can promote this to a constexpr instead of an instruction.
8048 // Scan for nonconstants...
8049 SmallVector<Constant*, 8> Indices;
8050 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8051 for (; I != E && isa<Constant>(*I); ++I)
8052 Indices.push_back(cast<Constant>(*I));
8054 if (I == E) { // If they are all constants...
8055 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8056 &Indices[0],Indices.size());
8058 // Replace all uses of the GEP with the new constexpr...
8059 return ReplaceInstUsesWith(GEP, CE);
8061 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8062 if (!isa<PointerType>(X->getType())) {
8063 // Not interesting. Source pointer must be a cast from pointer.
8064 } else if (HasZeroPointerIndex) {
8065 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8066 // into : GEP [10 x ubyte]* X, long 0, ...
8068 // This occurs when the program declares an array extern like "int X[];"
8070 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8071 const PointerType *XTy = cast<PointerType>(X->getType());
8072 if (const ArrayType *XATy =
8073 dyn_cast<ArrayType>(XTy->getElementType()))
8074 if (const ArrayType *CATy =
8075 dyn_cast<ArrayType>(CPTy->getElementType()))
8076 if (CATy->getElementType() == XATy->getElementType()) {
8077 // At this point, we know that the cast source type is a pointer
8078 // to an array of the same type as the destination pointer
8079 // array. Because the array type is never stepped over (there
8080 // is a leading zero) we can fold the cast into this GEP.
8081 GEP.setOperand(0, X);
8084 } else if (GEP.getNumOperands() == 2) {
8085 // Transform things like:
8086 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8087 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8088 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8089 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8090 if (isa<ArrayType>(SrcElTy) &&
8091 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8092 TD->getTypeSize(ResElTy)) {
8093 Value *V = InsertNewInstBefore(
8094 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8095 GEP.getOperand(1), GEP.getName()), GEP);
8096 // V and GEP are both pointer types --> BitCast
8097 return new BitCastInst(V, GEP.getType());
8100 // Transform things like:
8101 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8102 // (where tmp = 8*tmp2) into:
8103 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8105 if (isa<ArrayType>(SrcElTy) &&
8106 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8107 uint64_t ArrayEltSize =
8108 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8110 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8111 // allow either a mul, shift, or constant here.
8113 ConstantInt *Scale = 0;
8114 if (ArrayEltSize == 1) {
8115 NewIdx = GEP.getOperand(1);
8116 Scale = ConstantInt::get(NewIdx->getType(), 1);
8117 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8118 NewIdx = ConstantInt::get(CI->getType(), 1);
8120 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8121 if (Inst->getOpcode() == Instruction::Shl &&
8122 isa<ConstantInt>(Inst->getOperand(1))) {
8124 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
8125 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
8126 NewIdx = Inst->getOperand(0);
8127 } else if (Inst->getOpcode() == Instruction::Mul &&
8128 isa<ConstantInt>(Inst->getOperand(1))) {
8129 Scale = cast<ConstantInt>(Inst->getOperand(1));
8130 NewIdx = Inst->getOperand(0);
8134 // If the index will be to exactly the right offset with the scale taken
8135 // out, perform the transformation.
8136 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8137 if (isa<ConstantInt>(Scale))
8138 Scale = ConstantInt::get(Scale->getType(),
8139 Scale->getZExtValue() / ArrayEltSize);
8140 if (Scale->getZExtValue() != 1) {
8141 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8143 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8144 NewIdx = InsertNewInstBefore(Sc, GEP);
8147 // Insert the new GEP instruction.
8148 Instruction *NewGEP =
8149 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8150 NewIdx, GEP.getName());
8151 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8152 // The NewGEP must be pointer typed, so must the old one -> BitCast
8153 return new BitCastInst(NewGEP, GEP.getType());
8162 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8163 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8164 if (AI.isArrayAllocation()) // Check C != 1
8165 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8167 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8168 AllocationInst *New = 0;
8170 // Create and insert the replacement instruction...
8171 if (isa<MallocInst>(AI))
8172 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8174 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8175 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8178 InsertNewInstBefore(New, AI);
8180 // Scan to the end of the allocation instructions, to skip over a block of
8181 // allocas if possible...
8183 BasicBlock::iterator It = New;
8184 while (isa<AllocationInst>(*It)) ++It;
8186 // Now that I is pointing to the first non-allocation-inst in the block,
8187 // insert our getelementptr instruction...
8189 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8190 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8191 New->getName()+".sub", It);
8193 // Now make everything use the getelementptr instead of the original
8195 return ReplaceInstUsesWith(AI, V);
8196 } else if (isa<UndefValue>(AI.getArraySize())) {
8197 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8200 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8201 // Note that we only do this for alloca's, because malloc should allocate and
8202 // return a unique pointer, even for a zero byte allocation.
8203 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8204 TD->getTypeSize(AI.getAllocatedType()) == 0)
8205 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8210 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8211 Value *Op = FI.getOperand(0);
8213 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8214 if (CastInst *CI = dyn_cast<CastInst>(Op))
8215 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8216 FI.setOperand(0, CI->getOperand(0));
8220 // free undef -> unreachable.
8221 if (isa<UndefValue>(Op)) {
8222 // Insert a new store to null because we cannot modify the CFG here.
8223 new StoreInst(ConstantInt::getTrue(),
8224 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8225 return EraseInstFromFunction(FI);
8228 // If we have 'free null' delete the instruction. This can happen in stl code
8229 // when lots of inlining happens.
8230 if (isa<ConstantPointerNull>(Op))
8231 return EraseInstFromFunction(FI);
8237 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8238 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8239 User *CI = cast<User>(LI.getOperand(0));
8240 Value *CastOp = CI->getOperand(0);
8242 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8243 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8244 const Type *SrcPTy = SrcTy->getElementType();
8246 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8247 isa<VectorType>(DestPTy)) {
8248 // If the source is an array, the code below will not succeed. Check to
8249 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8251 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8252 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8253 if (ASrcTy->getNumElements() != 0) {
8255 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8256 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8257 SrcTy = cast<PointerType>(CastOp->getType());
8258 SrcPTy = SrcTy->getElementType();
8261 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8262 isa<VectorType>(SrcPTy)) &&
8263 // Do not allow turning this into a load of an integer, which is then
8264 // casted to a pointer, this pessimizes pointer analysis a lot.
8265 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8266 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8267 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8269 // Okay, we are casting from one integer or pointer type to another of
8270 // the same size. Instead of casting the pointer before the load, cast
8271 // the result of the loaded value.
8272 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8274 LI.isVolatile()),LI);
8275 // Now cast the result of the load.
8276 return new BitCastInst(NewLoad, LI.getType());
8283 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8284 /// from this value cannot trap. If it is not obviously safe to load from the
8285 /// specified pointer, we do a quick local scan of the basic block containing
8286 /// ScanFrom, to determine if the address is already accessed.
8287 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8288 // If it is an alloca or global variable, it is always safe to load from.
8289 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8291 // Otherwise, be a little bit agressive by scanning the local block where we
8292 // want to check to see if the pointer is already being loaded or stored
8293 // from/to. If so, the previous load or store would have already trapped,
8294 // so there is no harm doing an extra load (also, CSE will later eliminate
8295 // the load entirely).
8296 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8301 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8302 if (LI->getOperand(0) == V) return true;
8303 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8304 if (SI->getOperand(1) == V) return true;
8310 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8311 Value *Op = LI.getOperand(0);
8313 // load (cast X) --> cast (load X) iff safe
8314 if (isa<CastInst>(Op))
8315 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8318 // None of the following transforms are legal for volatile loads.
8319 if (LI.isVolatile()) return 0;
8321 if (&LI.getParent()->front() != &LI) {
8322 BasicBlock::iterator BBI = &LI; --BBI;
8323 // If the instruction immediately before this is a store to the same
8324 // address, do a simple form of store->load forwarding.
8325 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8326 if (SI->getOperand(1) == LI.getOperand(0))
8327 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8328 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8329 if (LIB->getOperand(0) == LI.getOperand(0))
8330 return ReplaceInstUsesWith(LI, LIB);
8333 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8334 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8335 isa<UndefValue>(GEPI->getOperand(0))) {
8336 // Insert a new store to null instruction before the load to indicate
8337 // that this code is not reachable. We do this instead of inserting
8338 // an unreachable instruction directly because we cannot modify the
8340 new StoreInst(UndefValue::get(LI.getType()),
8341 Constant::getNullValue(Op->getType()), &LI);
8342 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8345 if (Constant *C = dyn_cast<Constant>(Op)) {
8346 // load null/undef -> undef
8347 if ((C->isNullValue() || isa<UndefValue>(C))) {
8348 // Insert a new store to null instruction before the load to indicate that
8349 // this code is not reachable. We do this instead of inserting an
8350 // unreachable instruction directly because we cannot modify the CFG.
8351 new StoreInst(UndefValue::get(LI.getType()),
8352 Constant::getNullValue(Op->getType()), &LI);
8353 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8356 // Instcombine load (constant global) into the value loaded.
8357 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8358 if (GV->isConstant() && !GV->isDeclaration())
8359 return ReplaceInstUsesWith(LI, GV->getInitializer());
8361 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8362 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8363 if (CE->getOpcode() == Instruction::GetElementPtr) {
8364 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8365 if (GV->isConstant() && !GV->isDeclaration())
8367 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8368 return ReplaceInstUsesWith(LI, V);
8369 if (CE->getOperand(0)->isNullValue()) {
8370 // Insert a new store to null instruction before the load to indicate
8371 // that this code is not reachable. We do this instead of inserting
8372 // an unreachable instruction directly because we cannot modify the
8374 new StoreInst(UndefValue::get(LI.getType()),
8375 Constant::getNullValue(Op->getType()), &LI);
8376 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8379 } else if (CE->isCast()) {
8380 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8385 if (Op->hasOneUse()) {
8386 // Change select and PHI nodes to select values instead of addresses: this
8387 // helps alias analysis out a lot, allows many others simplifications, and
8388 // exposes redundancy in the code.
8390 // Note that we cannot do the transformation unless we know that the
8391 // introduced loads cannot trap! Something like this is valid as long as
8392 // the condition is always false: load (select bool %C, int* null, int* %G),
8393 // but it would not be valid if we transformed it to load from null
8396 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8397 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8398 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8399 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8400 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8401 SI->getOperand(1)->getName()+".val"), LI);
8402 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8403 SI->getOperand(2)->getName()+".val"), LI);
8404 return new SelectInst(SI->getCondition(), V1, V2);
8407 // load (select (cond, null, P)) -> load P
8408 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8409 if (C->isNullValue()) {
8410 LI.setOperand(0, SI->getOperand(2));
8414 // load (select (cond, P, null)) -> load P
8415 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8416 if (C->isNullValue()) {
8417 LI.setOperand(0, SI->getOperand(1));
8425 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8427 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8428 User *CI = cast<User>(SI.getOperand(1));
8429 Value *CastOp = CI->getOperand(0);
8431 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8432 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8433 const Type *SrcPTy = SrcTy->getElementType();
8435 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8436 // If the source is an array, the code below will not succeed. Check to
8437 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8439 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8440 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8441 if (ASrcTy->getNumElements() != 0) {
8443 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8444 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8445 SrcTy = cast<PointerType>(CastOp->getType());
8446 SrcPTy = SrcTy->getElementType();
8449 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8450 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8451 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8453 // Okay, we are casting from one integer or pointer type to another of
8454 // the same size. Instead of casting the pointer before
8455 // the store, cast the value to be stored.
8457 Value *SIOp0 = SI.getOperand(0);
8458 Instruction::CastOps opcode = Instruction::BitCast;
8459 const Type* CastSrcTy = SIOp0->getType();
8460 const Type* CastDstTy = SrcPTy;
8461 if (isa<PointerType>(CastDstTy)) {
8462 if (CastSrcTy->isInteger())
8463 opcode = Instruction::IntToPtr;
8464 } else if (isa<IntegerType>(CastDstTy)) {
8465 if (isa<PointerType>(SIOp0->getType()))
8466 opcode = Instruction::PtrToInt;
8468 if (Constant *C = dyn_cast<Constant>(SIOp0))
8469 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8471 NewCast = IC.InsertNewInstBefore(
8472 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8474 return new StoreInst(NewCast, CastOp);
8481 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8482 Value *Val = SI.getOperand(0);
8483 Value *Ptr = SI.getOperand(1);
8485 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8486 EraseInstFromFunction(SI);
8491 // If the RHS is an alloca with a single use, zapify the store, making the
8493 if (Ptr->hasOneUse()) {
8494 if (isa<AllocaInst>(Ptr)) {
8495 EraseInstFromFunction(SI);
8500 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8501 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8502 GEP->getOperand(0)->hasOneUse()) {
8503 EraseInstFromFunction(SI);
8509 // Do really simple DSE, to catch cases where there are several consequtive
8510 // stores to the same location, separated by a few arithmetic operations. This
8511 // situation often occurs with bitfield accesses.
8512 BasicBlock::iterator BBI = &SI;
8513 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8517 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8518 // Prev store isn't volatile, and stores to the same location?
8519 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8522 EraseInstFromFunction(*PrevSI);
8528 // If this is a load, we have to stop. However, if the loaded value is from
8529 // the pointer we're loading and is producing the pointer we're storing,
8530 // then *this* store is dead (X = load P; store X -> P).
8531 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8532 if (LI == Val && LI->getOperand(0) == Ptr) {
8533 EraseInstFromFunction(SI);
8537 // Otherwise, this is a load from some other location. Stores before it
8542 // Don't skip over loads or things that can modify memory.
8543 if (BBI->mayWriteToMemory())
8548 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8550 // store X, null -> turns into 'unreachable' in SimplifyCFG
8551 if (isa<ConstantPointerNull>(Ptr)) {
8552 if (!isa<UndefValue>(Val)) {
8553 SI.setOperand(0, UndefValue::get(Val->getType()));
8554 if (Instruction *U = dyn_cast<Instruction>(Val))
8555 AddToWorkList(U); // Dropped a use.
8558 return 0; // Do not modify these!
8561 // store undef, Ptr -> noop
8562 if (isa<UndefValue>(Val)) {
8563 EraseInstFromFunction(SI);
8568 // If the pointer destination is a cast, see if we can fold the cast into the
8570 if (isa<CastInst>(Ptr))
8571 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8573 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8575 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8579 // If this store is the last instruction in the basic block, and if the block
8580 // ends with an unconditional branch, try to move it to the successor block.
8582 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8583 if (BI->isUnconditional()) {
8584 // Check to see if the successor block has exactly two incoming edges. If
8585 // so, see if the other predecessor contains a store to the same location.
8586 // if so, insert a PHI node (if needed) and move the stores down.
8587 BasicBlock *Dest = BI->getSuccessor(0);
8589 pred_iterator PI = pred_begin(Dest);
8590 BasicBlock *Other = 0;
8591 if (*PI != BI->getParent())
8594 if (PI != pred_end(Dest)) {
8595 if (*PI != BI->getParent())
8600 if (++PI != pred_end(Dest))
8603 if (Other) { // If only one other pred...
8604 BBI = Other->getTerminator();
8605 // Make sure this other block ends in an unconditional branch and that
8606 // there is an instruction before the branch.
8607 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8608 BBI != Other->begin()) {
8610 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8612 // If this instruction is a store to the same location.
8613 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8614 // Okay, we know we can perform this transformation. Insert a PHI
8615 // node now if we need it.
8616 Value *MergedVal = OtherStore->getOperand(0);
8617 if (MergedVal != SI.getOperand(0)) {
8618 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8619 PN->reserveOperandSpace(2);
8620 PN->addIncoming(SI.getOperand(0), SI.getParent());
8621 PN->addIncoming(OtherStore->getOperand(0), Other);
8622 MergedVal = InsertNewInstBefore(PN, Dest->front());
8625 // Advance to a place where it is safe to insert the new store and
8627 BBI = Dest->begin();
8628 while (isa<PHINode>(BBI)) ++BBI;
8629 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8630 OtherStore->isVolatile()), *BBI);
8632 // Nuke the old stores.
8633 EraseInstFromFunction(SI);
8634 EraseInstFromFunction(*OtherStore);
8646 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8647 // Change br (not X), label True, label False to: br X, label False, True
8649 BasicBlock *TrueDest;
8650 BasicBlock *FalseDest;
8651 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8652 !isa<Constant>(X)) {
8653 // Swap Destinations and condition...
8655 BI.setSuccessor(0, FalseDest);
8656 BI.setSuccessor(1, TrueDest);
8660 // Cannonicalize fcmp_one -> fcmp_oeq
8661 FCmpInst::Predicate FPred; Value *Y;
8662 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8663 TrueDest, FalseDest)))
8664 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8665 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8666 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8667 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8668 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8669 NewSCC->takeName(I);
8670 // Swap Destinations and condition...
8671 BI.setCondition(NewSCC);
8672 BI.setSuccessor(0, FalseDest);
8673 BI.setSuccessor(1, TrueDest);
8674 RemoveFromWorkList(I);
8675 I->eraseFromParent();
8676 AddToWorkList(NewSCC);
8680 // Cannonicalize icmp_ne -> icmp_eq
8681 ICmpInst::Predicate IPred;
8682 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8683 TrueDest, FalseDest)))
8684 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8685 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8686 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8687 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8688 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8689 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8690 NewSCC->takeName(I);
8691 // Swap Destinations and condition...
8692 BI.setCondition(NewSCC);
8693 BI.setSuccessor(0, FalseDest);
8694 BI.setSuccessor(1, TrueDest);
8695 RemoveFromWorkList(I);
8696 I->eraseFromParent();;
8697 AddToWorkList(NewSCC);
8704 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8705 Value *Cond = SI.getCondition();
8706 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8707 if (I->getOpcode() == Instruction::Add)
8708 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8709 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8710 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8711 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8713 SI.setOperand(0, I->getOperand(0));
8721 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8722 /// is to leave as a vector operation.
8723 static bool CheapToScalarize(Value *V, bool isConstant) {
8724 if (isa<ConstantAggregateZero>(V))
8726 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8727 if (isConstant) return true;
8728 // If all elts are the same, we can extract.
8729 Constant *Op0 = C->getOperand(0);
8730 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8731 if (C->getOperand(i) != Op0)
8735 Instruction *I = dyn_cast<Instruction>(V);
8736 if (!I) return false;
8738 // Insert element gets simplified to the inserted element or is deleted if
8739 // this is constant idx extract element and its a constant idx insertelt.
8740 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8741 isa<ConstantInt>(I->getOperand(2)))
8743 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8745 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8746 if (BO->hasOneUse() &&
8747 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8748 CheapToScalarize(BO->getOperand(1), isConstant)))
8750 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8751 if (CI->hasOneUse() &&
8752 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8753 CheapToScalarize(CI->getOperand(1), isConstant)))
8759 /// Read and decode a shufflevector mask.
8761 /// It turns undef elements into values that are larger than the number of
8762 /// elements in the input.
8763 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8764 unsigned NElts = SVI->getType()->getNumElements();
8765 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8766 return std::vector<unsigned>(NElts, 0);
8767 if (isa<UndefValue>(SVI->getOperand(2)))
8768 return std::vector<unsigned>(NElts, 2*NElts);
8770 std::vector<unsigned> Result;
8771 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8772 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8773 if (isa<UndefValue>(CP->getOperand(i)))
8774 Result.push_back(NElts*2); // undef -> 8
8776 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8780 /// FindScalarElement - Given a vector and an element number, see if the scalar
8781 /// value is already around as a register, for example if it were inserted then
8782 /// extracted from the vector.
8783 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8784 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8785 const VectorType *PTy = cast<VectorType>(V->getType());
8786 unsigned Width = PTy->getNumElements();
8787 if (EltNo >= Width) // Out of range access.
8788 return UndefValue::get(PTy->getElementType());
8790 if (isa<UndefValue>(V))
8791 return UndefValue::get(PTy->getElementType());
8792 else if (isa<ConstantAggregateZero>(V))
8793 return Constant::getNullValue(PTy->getElementType());
8794 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8795 return CP->getOperand(EltNo);
8796 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8797 // If this is an insert to a variable element, we don't know what it is.
8798 if (!isa<ConstantInt>(III->getOperand(2)))
8800 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8802 // If this is an insert to the element we are looking for, return the
8805 return III->getOperand(1);
8807 // Otherwise, the insertelement doesn't modify the value, recurse on its
8809 return FindScalarElement(III->getOperand(0), EltNo);
8810 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8811 unsigned InEl = getShuffleMask(SVI)[EltNo];
8813 return FindScalarElement(SVI->getOperand(0), InEl);
8814 else if (InEl < Width*2)
8815 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8817 return UndefValue::get(PTy->getElementType());
8820 // Otherwise, we don't know.
8824 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8826 // If packed val is undef, replace extract with scalar undef.
8827 if (isa<UndefValue>(EI.getOperand(0)))
8828 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8830 // If packed val is constant 0, replace extract with scalar 0.
8831 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8832 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8834 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8835 // If packed val is constant with uniform operands, replace EI
8836 // with that operand
8837 Constant *op0 = C->getOperand(0);
8838 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8839 if (C->getOperand(i) != op0) {
8844 return ReplaceInstUsesWith(EI, op0);
8847 // If extracting a specified index from the vector, see if we can recursively
8848 // find a previously computed scalar that was inserted into the vector.
8849 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8850 // This instruction only demands the single element from the input vector.
8851 // If the input vector has a single use, simplify it based on this use
8853 uint64_t IndexVal = IdxC->getZExtValue();
8854 if (EI.getOperand(0)->hasOneUse()) {
8856 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8859 EI.setOperand(0, V);
8864 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8865 return ReplaceInstUsesWith(EI, Elt);
8868 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8869 if (I->hasOneUse()) {
8870 // Push extractelement into predecessor operation if legal and
8871 // profitable to do so
8872 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8873 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8874 if (CheapToScalarize(BO, isConstantElt)) {
8875 ExtractElementInst *newEI0 =
8876 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8877 EI.getName()+".lhs");
8878 ExtractElementInst *newEI1 =
8879 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8880 EI.getName()+".rhs");
8881 InsertNewInstBefore(newEI0, EI);
8882 InsertNewInstBefore(newEI1, EI);
8883 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8885 } else if (isa<LoadInst>(I)) {
8886 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8887 PointerType::get(EI.getType()), EI);
8888 GetElementPtrInst *GEP =
8889 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8890 InsertNewInstBefore(GEP, EI);
8891 return new LoadInst(GEP);
8894 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8895 // Extracting the inserted element?
8896 if (IE->getOperand(2) == EI.getOperand(1))
8897 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8898 // If the inserted and extracted elements are constants, they must not
8899 // be the same value, extract from the pre-inserted value instead.
8900 if (isa<Constant>(IE->getOperand(2)) &&
8901 isa<Constant>(EI.getOperand(1))) {
8902 AddUsesToWorkList(EI);
8903 EI.setOperand(0, IE->getOperand(0));
8906 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8907 // If this is extracting an element from a shufflevector, figure out where
8908 // it came from and extract from the appropriate input element instead.
8909 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8910 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8912 if (SrcIdx < SVI->getType()->getNumElements())
8913 Src = SVI->getOperand(0);
8914 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8915 SrcIdx -= SVI->getType()->getNumElements();
8916 Src = SVI->getOperand(1);
8918 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8920 return new ExtractElementInst(Src, SrcIdx);
8927 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8928 /// elements from either LHS or RHS, return the shuffle mask and true.
8929 /// Otherwise, return false.
8930 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8931 std::vector<Constant*> &Mask) {
8932 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8933 "Invalid CollectSingleShuffleElements");
8934 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8936 if (isa<UndefValue>(V)) {
8937 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8939 } else if (V == LHS) {
8940 for (unsigned i = 0; i != NumElts; ++i)
8941 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8943 } else if (V == RHS) {
8944 for (unsigned i = 0; i != NumElts; ++i)
8945 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8947 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8948 // If this is an insert of an extract from some other vector, include it.
8949 Value *VecOp = IEI->getOperand(0);
8950 Value *ScalarOp = IEI->getOperand(1);
8951 Value *IdxOp = IEI->getOperand(2);
8953 if (!isa<ConstantInt>(IdxOp))
8955 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8957 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8958 // Okay, we can handle this if the vector we are insertinting into is
8960 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8961 // If so, update the mask to reflect the inserted undef.
8962 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8965 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8966 if (isa<ConstantInt>(EI->getOperand(1)) &&
8967 EI->getOperand(0)->getType() == V->getType()) {
8968 unsigned ExtractedIdx =
8969 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8971 // This must be extracting from either LHS or RHS.
8972 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8973 // Okay, we can handle this if the vector we are insertinting into is
8975 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8976 // If so, update the mask to reflect the inserted value.
8977 if (EI->getOperand(0) == LHS) {
8978 Mask[InsertedIdx & (NumElts-1)] =
8979 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8981 assert(EI->getOperand(0) == RHS);
8982 Mask[InsertedIdx & (NumElts-1)] =
8983 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8992 // TODO: Handle shufflevector here!
8997 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8998 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8999 /// that computes V and the LHS value of the shuffle.
9000 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9002 assert(isa<VectorType>(V->getType()) &&
9003 (RHS == 0 || V->getType() == RHS->getType()) &&
9004 "Invalid shuffle!");
9005 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9007 if (isa<UndefValue>(V)) {
9008 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9010 } else if (isa<ConstantAggregateZero>(V)) {
9011 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9013 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9014 // If this is an insert of an extract from some other vector, include it.
9015 Value *VecOp = IEI->getOperand(0);
9016 Value *ScalarOp = IEI->getOperand(1);
9017 Value *IdxOp = IEI->getOperand(2);
9019 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9020 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9021 EI->getOperand(0)->getType() == V->getType()) {
9022 unsigned ExtractedIdx =
9023 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9024 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9026 // Either the extracted from or inserted into vector must be RHSVec,
9027 // otherwise we'd end up with a shuffle of three inputs.
9028 if (EI->getOperand(0) == RHS || RHS == 0) {
9029 RHS = EI->getOperand(0);
9030 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9031 Mask[InsertedIdx & (NumElts-1)] =
9032 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9037 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9038 // Everything but the extracted element is replaced with the RHS.
9039 for (unsigned i = 0; i != NumElts; ++i) {
9040 if (i != InsertedIdx)
9041 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9046 // If this insertelement is a chain that comes from exactly these two
9047 // vectors, return the vector and the effective shuffle.
9048 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9049 return EI->getOperand(0);
9054 // TODO: Handle shufflevector here!
9056 // Otherwise, can't do anything fancy. Return an identity vector.
9057 for (unsigned i = 0; i != NumElts; ++i)
9058 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9062 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9063 Value *VecOp = IE.getOperand(0);
9064 Value *ScalarOp = IE.getOperand(1);
9065 Value *IdxOp = IE.getOperand(2);
9067 // If the inserted element was extracted from some other vector, and if the
9068 // indexes are constant, try to turn this into a shufflevector operation.
9069 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9070 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9071 EI->getOperand(0)->getType() == IE.getType()) {
9072 unsigned NumVectorElts = IE.getType()->getNumElements();
9073 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9074 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9076 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9077 return ReplaceInstUsesWith(IE, VecOp);
9079 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9080 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9082 // If we are extracting a value from a vector, then inserting it right
9083 // back into the same place, just use the input vector.
9084 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9085 return ReplaceInstUsesWith(IE, VecOp);
9087 // We could theoretically do this for ANY input. However, doing so could
9088 // turn chains of insertelement instructions into a chain of shufflevector
9089 // instructions, and right now we do not merge shufflevectors. As such,
9090 // only do this in a situation where it is clear that there is benefit.
9091 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9092 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9093 // the values of VecOp, except then one read from EIOp0.
9094 // Build a new shuffle mask.
9095 std::vector<Constant*> Mask;
9096 if (isa<UndefValue>(VecOp))
9097 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9099 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9100 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9103 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9104 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9105 ConstantVector::get(Mask));
9108 // If this insertelement isn't used by some other insertelement, turn it
9109 // (and any insertelements it points to), into one big shuffle.
9110 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9111 std::vector<Constant*> Mask;
9113 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9114 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9115 // We now have a shuffle of LHS, RHS, Mask.
9116 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9125 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9126 Value *LHS = SVI.getOperand(0);
9127 Value *RHS = SVI.getOperand(1);
9128 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9130 bool MadeChange = false;
9132 // Undefined shuffle mask -> undefined value.
9133 if (isa<UndefValue>(SVI.getOperand(2)))
9134 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9136 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9137 // the undef, change them to undefs.
9138 if (isa<UndefValue>(SVI.getOperand(1))) {
9139 // Scan to see if there are any references to the RHS. If so, replace them
9140 // with undef element refs and set MadeChange to true.
9141 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9142 if (Mask[i] >= e && Mask[i] != 2*e) {
9149 // Remap any references to RHS to use LHS.
9150 std::vector<Constant*> Elts;
9151 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9153 Elts.push_back(UndefValue::get(Type::Int32Ty));
9155 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9157 SVI.setOperand(2, ConstantVector::get(Elts));
9161 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9162 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9163 if (LHS == RHS || isa<UndefValue>(LHS)) {
9164 if (isa<UndefValue>(LHS) && LHS == RHS) {
9165 // shuffle(undef,undef,mask) -> undef.
9166 return ReplaceInstUsesWith(SVI, LHS);
9169 // Remap any references to RHS to use LHS.
9170 std::vector<Constant*> Elts;
9171 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9173 Elts.push_back(UndefValue::get(Type::Int32Ty));
9175 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9176 (Mask[i] < e && isa<UndefValue>(LHS)))
9177 Mask[i] = 2*e; // Turn into undef.
9179 Mask[i] &= (e-1); // Force to LHS.
9180 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9183 SVI.setOperand(0, SVI.getOperand(1));
9184 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9185 SVI.setOperand(2, ConstantVector::get(Elts));
9186 LHS = SVI.getOperand(0);
9187 RHS = SVI.getOperand(1);
9191 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9192 bool isLHSID = true, isRHSID = true;
9194 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9195 if (Mask[i] >= e*2) continue; // Ignore undef values.
9196 // Is this an identity shuffle of the LHS value?
9197 isLHSID &= (Mask[i] == i);
9199 // Is this an identity shuffle of the RHS value?
9200 isRHSID &= (Mask[i]-e == i);
9203 // Eliminate identity shuffles.
9204 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9205 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9207 // If the LHS is a shufflevector itself, see if we can combine it with this
9208 // one without producing an unusual shuffle. Here we are really conservative:
9209 // we are absolutely afraid of producing a shuffle mask not in the input
9210 // program, because the code gen may not be smart enough to turn a merged
9211 // shuffle into two specific shuffles: it may produce worse code. As such,
9212 // we only merge two shuffles if the result is one of the two input shuffle
9213 // masks. In this case, merging the shuffles just removes one instruction,
9214 // which we know is safe. This is good for things like turning:
9215 // (splat(splat)) -> splat.
9216 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9217 if (isa<UndefValue>(RHS)) {
9218 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9220 std::vector<unsigned> NewMask;
9221 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9223 NewMask.push_back(2*e);
9225 NewMask.push_back(LHSMask[Mask[i]]);
9227 // If the result mask is equal to the src shuffle or this shuffle mask, do
9229 if (NewMask == LHSMask || NewMask == Mask) {
9230 std::vector<Constant*> Elts;
9231 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9232 if (NewMask[i] >= e*2) {
9233 Elts.push_back(UndefValue::get(Type::Int32Ty));
9235 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9238 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9239 LHSSVI->getOperand(1),
9240 ConstantVector::get(Elts));
9245 return MadeChange ? &SVI : 0;
9251 /// TryToSinkInstruction - Try to move the specified instruction from its
9252 /// current block into the beginning of DestBlock, which can only happen if it's
9253 /// safe to move the instruction past all of the instructions between it and the
9254 /// end of its block.
9255 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9256 assert(I->hasOneUse() && "Invariants didn't hold!");
9258 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9259 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9261 // Do not sink alloca instructions out of the entry block.
9262 if (isa<AllocaInst>(I) && I->getParent() ==
9263 &DestBlock->getParent()->getEntryBlock())
9266 // We can only sink load instructions if there is nothing between the load and
9267 // the end of block that could change the value.
9268 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9269 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9271 if (Scan->mayWriteToMemory())
9275 BasicBlock::iterator InsertPos = DestBlock->begin();
9276 while (isa<PHINode>(InsertPos)) ++InsertPos;
9278 I->moveBefore(InsertPos);
9284 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9285 /// all reachable code to the worklist.
9287 /// This has a couple of tricks to make the code faster and more powerful. In
9288 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9289 /// them to the worklist (this significantly speeds up instcombine on code where
9290 /// many instructions are dead or constant). Additionally, if we find a branch
9291 /// whose condition is a known constant, we only visit the reachable successors.
9293 static void AddReachableCodeToWorklist(BasicBlock *BB,
9294 SmallPtrSet<BasicBlock*, 64> &Visited,
9296 const TargetData *TD) {
9297 std::vector<BasicBlock*> Worklist;
9298 Worklist.push_back(BB);
9300 while (!Worklist.empty()) {
9301 BB = Worklist.back();
9302 Worklist.pop_back();
9304 // We have now visited this block! If we've already been here, ignore it.
9305 if (!Visited.insert(BB)) continue;
9307 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9308 Instruction *Inst = BBI++;
9310 // DCE instruction if trivially dead.
9311 if (isInstructionTriviallyDead(Inst)) {
9313 DOUT << "IC: DCE: " << *Inst;
9314 Inst->eraseFromParent();
9318 // ConstantProp instruction if trivially constant.
9319 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9320 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9321 Inst->replaceAllUsesWith(C);
9323 Inst->eraseFromParent();
9327 IC.AddToWorkList(Inst);
9330 // Recursively visit successors. If this is a branch or switch on a
9331 // constant, only visit the reachable successor.
9332 TerminatorInst *TI = BB->getTerminator();
9333 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9334 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9335 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9336 Worklist.push_back(BI->getSuccessor(!CondVal));
9339 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9340 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9341 // See if this is an explicit destination.
9342 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9343 if (SI->getCaseValue(i) == Cond) {
9344 Worklist.push_back(SI->getSuccessor(i));
9348 // Otherwise it is the default destination.
9349 Worklist.push_back(SI->getSuccessor(0));
9354 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9355 Worklist.push_back(TI->getSuccessor(i));
9359 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9360 bool Changed = false;
9361 TD = &getAnalysis<TargetData>();
9363 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9364 << F.getNameStr() << "\n");
9367 // Do a depth-first traversal of the function, populate the worklist with
9368 // the reachable instructions. Ignore blocks that are not reachable. Keep
9369 // track of which blocks we visit.
9370 SmallPtrSet<BasicBlock*, 64> Visited;
9371 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9373 // Do a quick scan over the function. If we find any blocks that are
9374 // unreachable, remove any instructions inside of them. This prevents
9375 // the instcombine code from having to deal with some bad special cases.
9376 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9377 if (!Visited.count(BB)) {
9378 Instruction *Term = BB->getTerminator();
9379 while (Term != BB->begin()) { // Remove instrs bottom-up
9380 BasicBlock::iterator I = Term; --I;
9382 DOUT << "IC: DCE: " << *I;
9385 if (!I->use_empty())
9386 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9387 I->eraseFromParent();
9392 while (!Worklist.empty()) {
9393 Instruction *I = RemoveOneFromWorkList();
9394 if (I == 0) continue; // skip null values.
9396 // Check to see if we can DCE the instruction.
9397 if (isInstructionTriviallyDead(I)) {
9398 // Add operands to the worklist.
9399 if (I->getNumOperands() < 4)
9400 AddUsesToWorkList(*I);
9403 DOUT << "IC: DCE: " << *I;
9405 I->eraseFromParent();
9406 RemoveFromWorkList(I);
9410 // Instruction isn't dead, see if we can constant propagate it.
9411 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9412 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9414 // Add operands to the worklist.
9415 AddUsesToWorkList(*I);
9416 ReplaceInstUsesWith(*I, C);
9419 I->eraseFromParent();
9420 RemoveFromWorkList(I);
9424 // See if we can trivially sink this instruction to a successor basic block.
9425 if (I->hasOneUse()) {
9426 BasicBlock *BB = I->getParent();
9427 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9428 if (UserParent != BB) {
9429 bool UserIsSuccessor = false;
9430 // See if the user is one of our successors.
9431 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9432 if (*SI == UserParent) {
9433 UserIsSuccessor = true;
9437 // If the user is one of our immediate successors, and if that successor
9438 // only has us as a predecessors (we'd have to split the critical edge
9439 // otherwise), we can keep going.
9440 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9441 next(pred_begin(UserParent)) == pred_end(UserParent))
9442 // Okay, the CFG is simple enough, try to sink this instruction.
9443 Changed |= TryToSinkInstruction(I, UserParent);
9447 // Now that we have an instruction, try combining it to simplify it...
9448 if (Instruction *Result = visit(*I)) {
9450 // Should we replace the old instruction with a new one?
9452 DOUT << "IC: Old = " << *I
9453 << " New = " << *Result;
9455 // Everything uses the new instruction now.
9456 I->replaceAllUsesWith(Result);
9458 // Push the new instruction and any users onto the worklist.
9459 AddToWorkList(Result);
9460 AddUsersToWorkList(*Result);
9462 // Move the name to the new instruction first.
9463 Result->takeName(I);
9465 // Insert the new instruction into the basic block...
9466 BasicBlock *InstParent = I->getParent();
9467 BasicBlock::iterator InsertPos = I;
9469 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9470 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9473 InstParent->getInstList().insert(InsertPos, Result);
9475 // Make sure that we reprocess all operands now that we reduced their
9477 AddUsesToWorkList(*I);
9479 // Instructions can end up on the worklist more than once. Make sure
9480 // we do not process an instruction that has been deleted.
9481 RemoveFromWorkList(I);
9483 // Erase the old instruction.
9484 InstParent->getInstList().erase(I);
9486 DOUT << "IC: MOD = " << *I;
9488 // If the instruction was modified, it's possible that it is now dead.
9489 // if so, remove it.
9490 if (isInstructionTriviallyDead(I)) {
9491 // Make sure we process all operands now that we are reducing their
9493 AddUsesToWorkList(*I);
9495 // Instructions may end up in the worklist more than once. Erase all
9496 // occurrences of this instruction.
9497 RemoveFromWorkList(I);
9498 I->eraseFromParent();
9501 AddUsersToWorkList(*I);
9508 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9513 bool InstCombiner::runOnFunction(Function &F) {
9514 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
9516 bool EverMadeChange = false;
9518 // Iterate while there is work to do.
9519 unsigned Iteration = 0;
9520 while (DoOneIteration(F, Iteration++))
9521 EverMadeChange = true;
9522 return EverMadeChange;
9525 FunctionPass *llvm::createInstructionCombiningPass() {
9526 return new InstCombiner();