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);
1943 // FIXME: This shouldn't be necessary. When the backends can handle types
1944 // with funny bit widths then this whole cascade of if statements should
1945 // be removed. It is just here to get the size of the "middle" type back
1946 // up to something that the back ends can handle.
1947 const Type *MiddleType = 0;
1950 case 32: MiddleType = Type::Int32Ty; break;
1951 case 16: MiddleType = Type::Int16Ty; break;
1952 case 8: MiddleType = Type::Int8Ty; break;
1955 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1956 InsertNewInstBefore(NewTrunc, I);
1957 return new SExtInst(NewTrunc, I.getType(), I.getName());
1963 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1964 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1966 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1967 if (RHSI->getOpcode() == Instruction::Sub)
1968 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1969 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1971 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1972 if (LHSI->getOpcode() == Instruction::Sub)
1973 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1974 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1979 if (Value *V = dyn_castNegVal(LHS))
1980 return BinaryOperator::createSub(RHS, V);
1983 if (!isa<Constant>(RHS))
1984 if (Value *V = dyn_castNegVal(RHS))
1985 return BinaryOperator::createSub(LHS, V);
1989 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1990 if (X == RHS) // X*C + X --> X * (C+1)
1991 return BinaryOperator::createMul(RHS, AddOne(C2));
1993 // X*C1 + X*C2 --> X * (C1+C2)
1995 if (X == dyn_castFoldableMul(RHS, C1))
1996 return BinaryOperator::createMul(X, Add(C1, C2));
1999 // X + X*C --> X * (C+1)
2000 if (dyn_castFoldableMul(RHS, C2) == LHS)
2001 return BinaryOperator::createMul(LHS, AddOne(C2));
2003 // X + ~X --> -1 since ~X = -X-1
2004 if (dyn_castNotVal(LHS) == RHS ||
2005 dyn_castNotVal(RHS) == LHS)
2006 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
2009 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2010 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2011 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2014 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2016 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2017 return BinaryOperator::createSub(SubOne(CRHS), X);
2019 // (X & FF00) + xx00 -> (X+xx00) & FF00
2020 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2021 Constant *Anded = And(CRHS, C2);
2022 if (Anded == CRHS) {
2023 // See if all bits from the first bit set in the Add RHS up are included
2024 // in the mask. First, get the rightmost bit.
2025 APInt AddRHSV(CRHS->getValue());
2027 // Form a mask of all bits from the lowest bit added through the top.
2028 APInt AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
2029 AddRHSHighBits &= C2->getType()->getMask();
2031 // See if the and mask includes all of these bits.
2032 APInt AddRHSHighBitsAnd = AddRHSHighBits & C2->getValue();
2034 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2035 // Okay, the xform is safe. Insert the new add pronto.
2036 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2037 LHS->getName()), I);
2038 return BinaryOperator::createAnd(NewAdd, C2);
2043 // Try to fold constant add into select arguments.
2044 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2045 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2049 // add (cast *A to intptrtype) B ->
2050 // cast (GEP (cast *A to sbyte*) B) ->
2053 CastInst *CI = dyn_cast<CastInst>(LHS);
2056 CI = dyn_cast<CastInst>(RHS);
2059 if (CI && CI->getType()->isSized() &&
2060 (CI->getType()->getPrimitiveSizeInBits() ==
2061 TD->getIntPtrType()->getPrimitiveSizeInBits())
2062 && isa<PointerType>(CI->getOperand(0)->getType())) {
2063 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2064 PointerType::get(Type::Int8Ty), I);
2065 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2066 return new PtrToIntInst(I2, CI->getType());
2070 return Changed ? &I : 0;
2073 // isSignBit - Return true if the value represented by the constant only has the
2074 // highest order bit set.
2075 static bool isSignBit(ConstantInt *CI) {
2076 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
2077 return CI->getValue() == APInt::getSignBit(NumBits);
2080 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2081 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2083 if (Op0 == Op1) // sub X, X -> 0
2084 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2086 // If this is a 'B = x-(-A)', change to B = x+A...
2087 if (Value *V = dyn_castNegVal(Op1))
2088 return BinaryOperator::createAdd(Op0, V);
2090 if (isa<UndefValue>(Op0))
2091 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2092 if (isa<UndefValue>(Op1))
2093 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2095 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2096 // Replace (-1 - A) with (~A)...
2097 if (C->isAllOnesValue())
2098 return BinaryOperator::createNot(Op1);
2100 // C - ~X == X + (1+C)
2102 if (match(Op1, m_Not(m_Value(X))))
2103 return BinaryOperator::createAdd(X, AddOne(C));
2105 // -(X >>u 31) -> (X >>s 31)
2106 // -(X >>s 31) -> (X >>u 31)
2107 if (C->isNullValue()) {
2108 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2109 if (SI->getOpcode() == Instruction::LShr) {
2110 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2111 // Check to see if we are shifting out everything but the sign bit.
2112 if (CU->getZExtValue() ==
2113 SI->getType()->getPrimitiveSizeInBits()-1) {
2114 // Ok, the transformation is safe. Insert AShr.
2115 return BinaryOperator::create(Instruction::AShr,
2116 SI->getOperand(0), CU, SI->getName());
2120 else if (SI->getOpcode() == Instruction::AShr) {
2121 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2122 // Check to see if we are shifting out everything but the sign bit.
2123 if (CU->getZExtValue() ==
2124 SI->getType()->getPrimitiveSizeInBits()-1) {
2125 // Ok, the transformation is safe. Insert LShr.
2126 return BinaryOperator::createLShr(
2127 SI->getOperand(0), CU, SI->getName());
2133 // Try to fold constant sub into select arguments.
2134 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2135 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2138 if (isa<PHINode>(Op0))
2139 if (Instruction *NV = FoldOpIntoPhi(I))
2143 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2144 if (Op1I->getOpcode() == Instruction::Add &&
2145 !Op0->getType()->isFPOrFPVector()) {
2146 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2147 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2148 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2149 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2150 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2151 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2152 // C1-(X+C2) --> (C1-C2)-X
2153 return BinaryOperator::createSub(Subtract(CI1, CI2),
2154 Op1I->getOperand(0));
2158 if (Op1I->hasOneUse()) {
2159 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2160 // is not used by anyone else...
2162 if (Op1I->getOpcode() == Instruction::Sub &&
2163 !Op1I->getType()->isFPOrFPVector()) {
2164 // Swap the two operands of the subexpr...
2165 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2166 Op1I->setOperand(0, IIOp1);
2167 Op1I->setOperand(1, IIOp0);
2169 // Create the new top level add instruction...
2170 return BinaryOperator::createAdd(Op0, Op1);
2173 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2175 if (Op1I->getOpcode() == Instruction::And &&
2176 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2177 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2180 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2181 return BinaryOperator::createAnd(Op0, NewNot);
2184 // 0 - (X sdiv C) -> (X sdiv -C)
2185 if (Op1I->getOpcode() == Instruction::SDiv)
2186 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2187 if (CSI->isNullValue())
2188 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2189 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2190 ConstantExpr::getNeg(DivRHS));
2192 // X - X*C --> X * (1-C)
2193 ConstantInt *C2 = 0;
2194 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2195 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2196 return BinaryOperator::createMul(Op0, CP1);
2201 if (!Op0->getType()->isFPOrFPVector())
2202 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2203 if (Op0I->getOpcode() == Instruction::Add) {
2204 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2205 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2206 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2207 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2208 } else if (Op0I->getOpcode() == Instruction::Sub) {
2209 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2210 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2214 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2215 if (X == Op1) // X*C - X --> X * (C-1)
2216 return BinaryOperator::createMul(Op1, SubOne(C1));
2218 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2219 if (X == dyn_castFoldableMul(Op1, C2))
2220 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2225 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2226 /// really just returns true if the most significant (sign) bit is set.
2227 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2229 case ICmpInst::ICMP_SLT:
2230 // True if LHS s< RHS and RHS == 0
2231 return RHS->isNullValue();
2232 case ICmpInst::ICMP_SLE:
2233 // True if LHS s<= RHS and RHS == -1
2234 return RHS->isAllOnesValue();
2235 case ICmpInst::ICMP_UGE:
2236 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2237 return RHS->getValue() ==
2238 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2239 case ICmpInst::ICMP_UGT:
2240 // True if LHS u> RHS and RHS == high-bit-mask - 1
2241 return RHS->getValue() ==
2242 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2248 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2249 bool Changed = SimplifyCommutative(I);
2250 Value *Op0 = I.getOperand(0);
2252 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2253 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2255 // Simplify mul instructions with a constant RHS...
2256 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2257 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2259 // ((X << C1)*C2) == (X * (C2 << C1))
2260 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2261 if (SI->getOpcode() == Instruction::Shl)
2262 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2263 return BinaryOperator::createMul(SI->getOperand(0),
2264 ConstantExpr::getShl(CI, ShOp));
2266 if (CI->isNullValue())
2267 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2268 if (CI->equalsInt(1)) // X * 1 == X
2269 return ReplaceInstUsesWith(I, Op0);
2270 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2271 return BinaryOperator::createNeg(Op0, I.getName());
2273 APInt Val(cast<ConstantInt>(CI)->getValue());
2274 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2275 return BinaryOperator::createShl(Op0,
2276 ConstantInt::get(Op0->getType(), Val.logBase2()));
2278 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2279 if (Op1F->isNullValue())
2280 return ReplaceInstUsesWith(I, Op1);
2282 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2283 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2284 if (Op1F->getValue() == 1.0)
2285 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2288 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2289 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2290 isa<ConstantInt>(Op0I->getOperand(1))) {
2291 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2292 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2294 InsertNewInstBefore(Add, I);
2295 Value *C1C2 = ConstantExpr::getMul(Op1,
2296 cast<Constant>(Op0I->getOperand(1)));
2297 return BinaryOperator::createAdd(Add, C1C2);
2301 // Try to fold constant mul into select arguments.
2302 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2303 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2306 if (isa<PHINode>(Op0))
2307 if (Instruction *NV = FoldOpIntoPhi(I))
2311 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2312 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2313 return BinaryOperator::createMul(Op0v, Op1v);
2315 // If one of the operands of the multiply is a cast from a boolean value, then
2316 // we know the bool is either zero or one, so this is a 'masking' multiply.
2317 // See if we can simplify things based on how the boolean was originally
2319 CastInst *BoolCast = 0;
2320 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2321 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2324 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2325 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2328 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2329 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2330 const Type *SCOpTy = SCIOp0->getType();
2332 // If the icmp is true iff the sign bit of X is set, then convert this
2333 // multiply into a shift/and combination.
2334 if (isa<ConstantInt>(SCIOp1) &&
2335 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2336 // Shift the X value right to turn it into "all signbits".
2337 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2338 SCOpTy->getPrimitiveSizeInBits()-1);
2340 InsertNewInstBefore(
2341 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2342 BoolCast->getOperand(0)->getName()+
2345 // If the multiply type is not the same as the source type, sign extend
2346 // or truncate to the multiply type.
2347 if (I.getType() != V->getType()) {
2348 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2349 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2350 Instruction::CastOps opcode =
2351 (SrcBits == DstBits ? Instruction::BitCast :
2352 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2353 V = InsertCastBefore(opcode, V, I.getType(), I);
2356 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2357 return BinaryOperator::createAnd(V, OtherOp);
2362 return Changed ? &I : 0;
2365 /// This function implements the transforms on div instructions that work
2366 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2367 /// used by the visitors to those instructions.
2368 /// @brief Transforms common to all three div instructions
2369 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2370 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2373 if (isa<UndefValue>(Op0))
2374 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2376 // X / undef -> undef
2377 if (isa<UndefValue>(Op1))
2378 return ReplaceInstUsesWith(I, Op1);
2380 // Handle cases involving: div X, (select Cond, Y, Z)
2381 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2382 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2383 // same basic block, then we replace the select with Y, and the condition
2384 // of the select with false (if the cond value is in the same BB). If the
2385 // select has uses other than the div, this allows them to be simplified
2386 // also. Note that div X, Y is just as good as div X, 0 (undef)
2387 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2388 if (ST->isNullValue()) {
2389 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2390 if (CondI && CondI->getParent() == I.getParent())
2391 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2392 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2393 I.setOperand(1, SI->getOperand(2));
2395 UpdateValueUsesWith(SI, SI->getOperand(2));
2399 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2400 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2401 if (ST->isNullValue()) {
2402 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2403 if (CondI && CondI->getParent() == I.getParent())
2404 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2405 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2406 I.setOperand(1, SI->getOperand(1));
2408 UpdateValueUsesWith(SI, SI->getOperand(1));
2416 /// This function implements the transforms common to both integer division
2417 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2418 /// division instructions.
2419 /// @brief Common integer divide transforms
2420 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2421 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2423 if (Instruction *Common = commonDivTransforms(I))
2426 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2428 if (RHS->equalsInt(1))
2429 return ReplaceInstUsesWith(I, Op0);
2431 // (X / C1) / C2 -> X / (C1*C2)
2432 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2433 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2434 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2435 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2436 Multiply(RHS, LHSRHS));
2439 if (!RHS->isZero()) { // avoid X udiv 0
2440 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2441 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2443 if (isa<PHINode>(Op0))
2444 if (Instruction *NV = FoldOpIntoPhi(I))
2449 // 0 / X == 0, we don't need to preserve faults!
2450 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2451 if (LHS->equalsInt(0))
2452 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2457 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2458 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2460 // Handle the integer div common cases
2461 if (Instruction *Common = commonIDivTransforms(I))
2464 // X udiv C^2 -> X >> C
2465 // Check to see if this is an unsigned division with an exact power of 2,
2466 // if so, convert to a right shift.
2467 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2468 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2469 return BinaryOperator::createLShr(Op0,
2470 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2473 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2474 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2475 if (RHSI->getOpcode() == Instruction::Shl &&
2476 isa<ConstantInt>(RHSI->getOperand(0))) {
2477 APInt C1(cast<ConstantInt>(RHSI->getOperand(0))->getValue());
2478 if (C1.isPowerOf2()) {
2479 Value *N = RHSI->getOperand(1);
2480 const Type *NTy = N->getType();
2481 if (uint32_t C2 = C1.logBase2()) {
2482 Constant *C2V = ConstantInt::get(NTy, C2);
2483 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2485 return BinaryOperator::createLShr(Op0, N);
2490 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2491 // where C1&C2 are powers of two.
2492 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2493 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2494 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2495 APInt TVA(STO->getValue()), FVA(SFO->getValue());
2496 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2497 // Compute the shift amounts
2498 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2499 // Construct the "on true" case of the select
2500 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2501 Instruction *TSI = BinaryOperator::createLShr(
2502 Op0, TC, SI->getName()+".t");
2503 TSI = InsertNewInstBefore(TSI, I);
2505 // Construct the "on false" case of the select
2506 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2507 Instruction *FSI = BinaryOperator::createLShr(
2508 Op0, FC, SI->getName()+".f");
2509 FSI = InsertNewInstBefore(FSI, I);
2511 // construct the select instruction and return it.
2512 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2518 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2519 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2521 // Handle the integer div common cases
2522 if (Instruction *Common = commonIDivTransforms(I))
2525 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2527 if (RHS->isAllOnesValue())
2528 return BinaryOperator::createNeg(Op0);
2531 if (Value *LHSNeg = dyn_castNegVal(Op0))
2532 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2535 // If the sign bits of both operands are zero (i.e. we can prove they are
2536 // unsigned inputs), turn this into a udiv.
2537 if (I.getType()->isInteger()) {
2538 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2539 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2540 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2547 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2548 return commonDivTransforms(I);
2551 /// GetFactor - If we can prove that the specified value is at least a multiple
2552 /// of some factor, return that factor.
2553 static Constant *GetFactor(Value *V) {
2554 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2557 // Unless we can be tricky, we know this is a multiple of 1.
2558 Constant *Result = ConstantInt::get(V->getType(), 1);
2560 Instruction *I = dyn_cast<Instruction>(V);
2561 if (!I) return Result;
2563 if (I->getOpcode() == Instruction::Mul) {
2564 // Handle multiplies by a constant, etc.
2565 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2566 GetFactor(I->getOperand(1)));
2567 } else if (I->getOpcode() == Instruction::Shl) {
2568 // (X<<C) -> X * (1 << C)
2569 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2570 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2571 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2573 } else if (I->getOpcode() == Instruction::And) {
2574 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2575 // X & 0xFFF0 is known to be a multiple of 16.
2576 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2577 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2578 return ConstantExpr::getShl(Result,
2579 ConstantInt::get(Result->getType(), Zeros));
2581 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2582 // Only handle int->int casts.
2583 if (!CI->isIntegerCast())
2585 Value *Op = CI->getOperand(0);
2586 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2591 /// This function implements the transforms on rem instructions that work
2592 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2593 /// is used by the visitors to those instructions.
2594 /// @brief Transforms common to all three rem instructions
2595 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2596 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2598 // 0 % X == 0, we don't need to preserve faults!
2599 if (Constant *LHS = dyn_cast<Constant>(Op0))
2600 if (LHS->isNullValue())
2601 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2603 if (isa<UndefValue>(Op0)) // undef % X -> 0
2604 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2605 if (isa<UndefValue>(Op1))
2606 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2608 // Handle cases involving: rem X, (select Cond, Y, Z)
2609 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2610 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2611 // the same basic block, then we replace the select with Y, and the
2612 // condition of the select with false (if the cond value is in the same
2613 // BB). If the select has uses other than the div, this allows them to be
2615 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2616 if (ST->isNullValue()) {
2617 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2618 if (CondI && CondI->getParent() == I.getParent())
2619 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2620 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2621 I.setOperand(1, SI->getOperand(2));
2623 UpdateValueUsesWith(SI, SI->getOperand(2));
2626 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2627 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2628 if (ST->isNullValue()) {
2629 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2630 if (CondI && CondI->getParent() == I.getParent())
2631 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2632 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2633 I.setOperand(1, SI->getOperand(1));
2635 UpdateValueUsesWith(SI, SI->getOperand(1));
2643 /// This function implements the transforms common to both integer remainder
2644 /// instructions (urem and srem). It is called by the visitors to those integer
2645 /// remainder instructions.
2646 /// @brief Common integer remainder transforms
2647 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2648 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2650 if (Instruction *common = commonRemTransforms(I))
2653 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2654 // X % 0 == undef, we don't need to preserve faults!
2655 if (RHS->equalsInt(0))
2656 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2658 if (RHS->equalsInt(1)) // X % 1 == 0
2659 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2661 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2662 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2663 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2665 } else if (isa<PHINode>(Op0I)) {
2666 if (Instruction *NV = FoldOpIntoPhi(I))
2669 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2670 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2671 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2678 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2679 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2681 if (Instruction *common = commonIRemTransforms(I))
2684 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2685 // X urem C^2 -> X and C
2686 // Check to see if this is an unsigned remainder with an exact power of 2,
2687 // if so, convert to a bitwise and.
2688 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2689 if (C->getValue().isPowerOf2())
2690 return BinaryOperator::createAnd(Op0, SubOne(C));
2693 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2694 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2695 if (RHSI->getOpcode() == Instruction::Shl &&
2696 isa<ConstantInt>(RHSI->getOperand(0))) {
2697 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2698 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2699 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2701 return BinaryOperator::createAnd(Op0, Add);
2706 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2707 // where C1&C2 are powers of two.
2708 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2709 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2710 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2711 // STO == 0 and SFO == 0 handled above.
2712 if ((STO->getValue().isPowerOf2()) &&
2713 (SFO->getValue().isPowerOf2())) {
2714 Value *TrueAnd = InsertNewInstBefore(
2715 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2716 Value *FalseAnd = InsertNewInstBefore(
2717 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2718 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2726 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2727 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2729 if (Instruction *common = commonIRemTransforms(I))
2732 if (Value *RHSNeg = dyn_castNegVal(Op1))
2733 if (!isa<ConstantInt>(RHSNeg) ||
2734 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2736 AddUsesToWorkList(I);
2737 I.setOperand(1, RHSNeg);
2741 // If the top bits of both operands are zero (i.e. we can prove they are
2742 // unsigned inputs), turn this into a urem.
2743 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2744 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2745 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2746 return BinaryOperator::createURem(Op0, Op1, I.getName());
2752 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2753 return commonRemTransforms(I);
2756 // isMaxValueMinusOne - return true if this is Max-1
2757 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2758 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2760 // Calculate 0111111111..11111
2761 APInt Val(APInt::getSignedMaxValue(TypeBits));
2762 return C->getValue() == Val-1;
2764 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2767 // isMinValuePlusOne - return true if this is Min+1
2768 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2770 // Calculate 1111111111000000000000
2771 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2772 APInt Val(APInt::getSignedMinValue(TypeBits));
2773 return C->getValue() == Val+1;
2775 return C->getValue() == 1; // unsigned
2778 // isOneBitSet - Return true if there is exactly one bit set in the specified
2780 static bool isOneBitSet(const ConstantInt *CI) {
2781 return CI->getValue().isPowerOf2();
2784 // isHighOnes - Return true if the constant is of the form 1+0+.
2785 // This is the same as lowones(~X).
2786 static bool isHighOnes(const ConstantInt *CI) {
2787 return (~CI->getValue() + 1).isPowerOf2();
2790 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2791 /// are carefully arranged to allow folding of expressions such as:
2793 /// (A < B) | (A > B) --> (A != B)
2795 /// Note that this is only valid if the first and second predicates have the
2796 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2798 /// Three bits are used to represent the condition, as follows:
2803 /// <=> Value Definition
2804 /// 000 0 Always false
2811 /// 111 7 Always true
2813 static unsigned getICmpCode(const ICmpInst *ICI) {
2814 switch (ICI->getPredicate()) {
2816 case ICmpInst::ICMP_UGT: return 1; // 001
2817 case ICmpInst::ICMP_SGT: return 1; // 001
2818 case ICmpInst::ICMP_EQ: return 2; // 010
2819 case ICmpInst::ICMP_UGE: return 3; // 011
2820 case ICmpInst::ICMP_SGE: return 3; // 011
2821 case ICmpInst::ICMP_ULT: return 4; // 100
2822 case ICmpInst::ICMP_SLT: return 4; // 100
2823 case ICmpInst::ICMP_NE: return 5; // 101
2824 case ICmpInst::ICMP_ULE: return 6; // 110
2825 case ICmpInst::ICMP_SLE: return 6; // 110
2828 assert(0 && "Invalid ICmp predicate!");
2833 /// getICmpValue - This is the complement of getICmpCode, which turns an
2834 /// opcode and two operands into either a constant true or false, or a brand
2835 /// new /// ICmp instruction. The sign is passed in to determine which kind
2836 /// of predicate to use in new icmp instructions.
2837 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2839 default: assert(0 && "Illegal ICmp code!");
2840 case 0: return ConstantInt::getFalse();
2843 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2845 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2846 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2849 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2851 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2854 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2856 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2857 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2860 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2862 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2863 case 7: return ConstantInt::getTrue();
2867 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2868 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2869 (ICmpInst::isSignedPredicate(p1) &&
2870 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2871 (ICmpInst::isSignedPredicate(p2) &&
2872 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2876 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2877 struct FoldICmpLogical {
2880 ICmpInst::Predicate pred;
2881 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2882 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2883 pred(ICI->getPredicate()) {}
2884 bool shouldApply(Value *V) const {
2885 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2886 if (PredicatesFoldable(pred, ICI->getPredicate()))
2887 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2888 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2891 Instruction *apply(Instruction &Log) const {
2892 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2893 if (ICI->getOperand(0) != LHS) {
2894 assert(ICI->getOperand(1) == LHS);
2895 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2898 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2899 unsigned LHSCode = getICmpCode(ICI);
2900 unsigned RHSCode = getICmpCode(RHSICI);
2902 switch (Log.getOpcode()) {
2903 case Instruction::And: Code = LHSCode & RHSCode; break;
2904 case Instruction::Or: Code = LHSCode | RHSCode; break;
2905 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2906 default: assert(0 && "Illegal logical opcode!"); return 0;
2909 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2910 ICmpInst::isSignedPredicate(ICI->getPredicate());
2912 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2913 if (Instruction *I = dyn_cast<Instruction>(RV))
2915 // Otherwise, it's a constant boolean value...
2916 return IC.ReplaceInstUsesWith(Log, RV);
2919 } // end anonymous namespace
2921 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2922 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2923 // guaranteed to be a binary operator.
2924 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2926 ConstantInt *AndRHS,
2927 BinaryOperator &TheAnd) {
2928 Value *X = Op->getOperand(0);
2929 Constant *Together = 0;
2931 Together = And(AndRHS, OpRHS);
2933 switch (Op->getOpcode()) {
2934 case Instruction::Xor:
2935 if (Op->hasOneUse()) {
2936 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2937 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2938 InsertNewInstBefore(And, TheAnd);
2940 return BinaryOperator::createXor(And, Together);
2943 case Instruction::Or:
2944 if (Together == AndRHS) // (X | C) & C --> C
2945 return ReplaceInstUsesWith(TheAnd, AndRHS);
2947 if (Op->hasOneUse() && Together != OpRHS) {
2948 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2949 Instruction *Or = BinaryOperator::createOr(X, Together);
2950 InsertNewInstBefore(Or, TheAnd);
2952 return BinaryOperator::createAnd(Or, AndRHS);
2955 case Instruction::Add:
2956 if (Op->hasOneUse()) {
2957 // Adding a one to a single bit bit-field should be turned into an XOR
2958 // of the bit. First thing to check is to see if this AND is with a
2959 // single bit constant.
2960 APInt AndRHSV(cast<ConstantInt>(AndRHS)->getValue());
2962 // If there is only one bit set...
2963 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2964 // Ok, at this point, we know that we are masking the result of the
2965 // ADD down to exactly one bit. If the constant we are adding has
2966 // no bits set below this bit, then we can eliminate the ADD.
2967 APInt AddRHS(cast<ConstantInt>(OpRHS)->getValue());
2969 // Check to see if any bits below the one bit set in AndRHSV are set.
2970 if ((AddRHS & (AndRHSV-1)) == 0) {
2971 // If not, the only thing that can effect the output of the AND is
2972 // the bit specified by AndRHSV. If that bit is set, the effect of
2973 // the XOR is to toggle the bit. If it is clear, then the ADD has
2975 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2976 TheAnd.setOperand(0, X);
2979 // Pull the XOR out of the AND.
2980 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2981 InsertNewInstBefore(NewAnd, TheAnd);
2982 NewAnd->takeName(Op);
2983 return BinaryOperator::createXor(NewAnd, AndRHS);
2990 case Instruction::Shl: {
2991 // We know that the AND will not produce any of the bits shifted in, so if
2992 // the anded constant includes them, clear them now!
2994 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2995 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2996 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2998 if (CI == ShlMask) { // Masking out bits that the shift already masks
2999 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3000 } else if (CI != AndRHS) { // Reducing bits set in and.
3001 TheAnd.setOperand(1, CI);
3006 case Instruction::LShr:
3008 // We know that the AND will not produce any of the bits shifted in, so if
3009 // the anded constant includes them, clear them now! This only applies to
3010 // unsigned shifts, because a signed shr may bring in set bits!
3012 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3013 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3014 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
3016 if (CI == ShrMask) { // Masking out bits that the shift already masks.
3017 return ReplaceInstUsesWith(TheAnd, Op);
3018 } else if (CI != AndRHS) {
3019 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3024 case Instruction::AShr:
3026 // See if this is shifting in some sign extension, then masking it out
3028 if (Op->hasOneUse()) {
3029 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3030 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3031 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
3032 if (C == AndRHS) { // Masking out bits shifted in.
3033 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3034 // Make the argument unsigned.
3035 Value *ShVal = Op->getOperand(0);
3036 ShVal = InsertNewInstBefore(
3037 BinaryOperator::createLShr(ShVal, OpRHS,
3038 Op->getName()), TheAnd);
3039 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3048 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3049 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3050 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3051 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3052 /// insert new instructions.
3053 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3054 bool isSigned, bool Inside,
3056 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3057 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3058 "Lo is not <= Hi in range emission code!");
3061 if (Lo == Hi) // Trivially false.
3062 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3064 // V >= Min && V < Hi --> V < Hi
3065 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3066 ICmpInst::Predicate pred = (isSigned ?
3067 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3068 return new ICmpInst(pred, V, Hi);
3071 // Emit V-Lo <u Hi-Lo
3072 Constant *NegLo = ConstantExpr::getNeg(Lo);
3073 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3074 InsertNewInstBefore(Add, IB);
3075 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3076 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3079 if (Lo == Hi) // Trivially true.
3080 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3082 // V < Min || V >= Hi -> V > Hi-1
3083 Hi = SubOne(cast<ConstantInt>(Hi));
3084 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3085 ICmpInst::Predicate pred = (isSigned ?
3086 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3087 return new ICmpInst(pred, V, Hi);
3090 // Emit V-Lo >u Hi-1-Lo
3091 // Note that Hi has already had one subtracted from it, above.
3092 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3093 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3094 InsertNewInstBefore(Add, IB);
3095 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3096 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3099 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3100 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3101 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3102 // not, since all 1s are not contiguous.
3103 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3104 APInt V = Val->getValue();
3105 uint32_t BitWidth = Val->getType()->getBitWidth();
3106 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3108 // look for the first zero bit after the run of ones
3109 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3110 // look for the first non-zero bit
3111 ME = V.getActiveBits();
3115 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3116 /// where isSub determines whether the operator is a sub. If we can fold one of
3117 /// the following xforms:
3119 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3120 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3121 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3123 /// return (A +/- B).
3125 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3126 ConstantInt *Mask, bool isSub,
3128 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3129 if (!LHSI || LHSI->getNumOperands() != 2 ||
3130 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3132 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3134 switch (LHSI->getOpcode()) {
3136 case Instruction::And:
3137 if (And(N, Mask) == Mask) {
3138 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3139 if ((Mask->getValue().countLeadingZeros() +
3140 Mask->getValue().countPopulation()) ==
3141 Mask->getValue().getBitWidth())
3144 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3145 // part, we don't need any explicit masks to take them out of A. If that
3146 // is all N is, ignore it.
3147 unsigned MB = 0, ME = 0;
3148 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3149 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3150 APInt Mask(APInt::getAllOnesValue(BitWidth));
3151 Mask = Mask.lshr(BitWidth-MB+1);
3152 if (MaskedValueIsZero(RHS, Mask))
3157 case Instruction::Or:
3158 case Instruction::Xor:
3159 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3160 if ((Mask->getValue().countLeadingZeros() +
3161 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3162 && And(N, Mask)->isZero())
3169 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3171 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3172 return InsertNewInstBefore(New, I);
3175 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3176 bool Changed = SimplifyCommutative(I);
3177 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3179 if (isa<UndefValue>(Op1)) // X & undef -> 0
3180 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3184 return ReplaceInstUsesWith(I, Op1);
3186 // See if we can simplify any instructions used by the instruction whose sole
3187 // purpose is to compute bits we don't care about.
3188 if (!isa<VectorType>(I.getType())) {
3189 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3190 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3191 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3192 KnownZero, KnownOne))
3195 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3196 if (CP->isAllOnesValue())
3197 return ReplaceInstUsesWith(I, I.getOperand(0));
3201 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3202 APInt AndRHSMask(AndRHS->getValue());
3203 APInt TypeMask(cast<IntegerType>(Op0->getType())->getMask());
3204 APInt NotAndRHS = AndRHSMask^TypeMask;
3206 // Optimize a variety of ((val OP C1) & C2) combinations...
3207 if (isa<BinaryOperator>(Op0)) {
3208 Instruction *Op0I = cast<Instruction>(Op0);
3209 Value *Op0LHS = Op0I->getOperand(0);
3210 Value *Op0RHS = Op0I->getOperand(1);
3211 switch (Op0I->getOpcode()) {
3212 case Instruction::Xor:
3213 case Instruction::Or:
3214 // If the mask is only needed on one incoming arm, push it up.
3215 if (Op0I->hasOneUse()) {
3216 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3217 // Not masking anything out for the LHS, move to RHS.
3218 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3219 Op0RHS->getName()+".masked");
3220 InsertNewInstBefore(NewRHS, I);
3221 return BinaryOperator::create(
3222 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3224 if (!isa<Constant>(Op0RHS) &&
3225 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3226 // Not masking anything out for the RHS, move to LHS.
3227 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3228 Op0LHS->getName()+".masked");
3229 InsertNewInstBefore(NewLHS, I);
3230 return BinaryOperator::create(
3231 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3236 case Instruction::Add:
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, false, I))
3241 return BinaryOperator::createAnd(V, AndRHS);
3242 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3243 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3246 case Instruction::Sub:
3247 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3248 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3249 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3250 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3251 return BinaryOperator::createAnd(V, AndRHS);
3255 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3256 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3258 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3259 // If this is an integer truncation or change from signed-to-unsigned, and
3260 // if the source is an and/or with immediate, transform it. This
3261 // frequently occurs for bitfield accesses.
3262 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3263 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3264 CastOp->getNumOperands() == 2)
3265 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3266 if (CastOp->getOpcode() == Instruction::And) {
3267 // Change: and (cast (and X, C1) to T), C2
3268 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3269 // This will fold the two constants together, which may allow
3270 // other simplifications.
3271 Instruction *NewCast = CastInst::createTruncOrBitCast(
3272 CastOp->getOperand(0), I.getType(),
3273 CastOp->getName()+".shrunk");
3274 NewCast = InsertNewInstBefore(NewCast, I);
3275 // trunc_or_bitcast(C1)&C2
3276 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3277 C3 = ConstantExpr::getAnd(C3, AndRHS);
3278 return BinaryOperator::createAnd(NewCast, C3);
3279 } else if (CastOp->getOpcode() == Instruction::Or) {
3280 // Change: and (cast (or X, C1) to T), C2
3281 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3282 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3283 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3284 return ReplaceInstUsesWith(I, AndRHS);
3289 // Try to fold constant and into select arguments.
3290 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3291 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3293 if (isa<PHINode>(Op0))
3294 if (Instruction *NV = FoldOpIntoPhi(I))
3298 Value *Op0NotVal = dyn_castNotVal(Op0);
3299 Value *Op1NotVal = dyn_castNotVal(Op1);
3301 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3302 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3304 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3305 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3306 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3307 I.getName()+".demorgan");
3308 InsertNewInstBefore(Or, I);
3309 return BinaryOperator::createNot(Or);
3313 Value *A = 0, *B = 0;
3314 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3315 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3316 return ReplaceInstUsesWith(I, Op1);
3317 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3318 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3319 return ReplaceInstUsesWith(I, Op0);
3321 if (Op0->hasOneUse() &&
3322 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3323 if (A == Op1) { // (A^B)&A -> A&(A^B)
3324 I.swapOperands(); // Simplify below
3325 std::swap(Op0, Op1);
3326 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3327 cast<BinaryOperator>(Op0)->swapOperands();
3328 I.swapOperands(); // Simplify below
3329 std::swap(Op0, Op1);
3332 if (Op1->hasOneUse() &&
3333 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3334 if (B == Op0) { // B&(A^B) -> B&(B^A)
3335 cast<BinaryOperator>(Op1)->swapOperands();
3338 if (A == Op0) { // A&(A^B) -> A & ~B
3339 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3340 InsertNewInstBefore(NotB, I);
3341 return BinaryOperator::createAnd(A, NotB);
3346 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3347 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3348 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3351 Value *LHSVal, *RHSVal;
3352 ConstantInt *LHSCst, *RHSCst;
3353 ICmpInst::Predicate LHSCC, RHSCC;
3354 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3355 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3356 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3357 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3358 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3359 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3360 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3361 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3362 // Ensure that the larger constant is on the RHS.
3363 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3364 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3365 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3366 ICmpInst *LHS = cast<ICmpInst>(Op0);
3367 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3368 std::swap(LHS, RHS);
3369 std::swap(LHSCst, RHSCst);
3370 std::swap(LHSCC, RHSCC);
3373 // At this point, we know we have have two icmp instructions
3374 // comparing a value against two constants and and'ing the result
3375 // together. Because of the above check, we know that we only have
3376 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3377 // (from the FoldICmpLogical check above), that the two constants
3378 // are not equal and that the larger constant is on the RHS
3379 assert(LHSCst != RHSCst && "Compares not folded above?");
3382 default: assert(0 && "Unknown integer condition code!");
3383 case ICmpInst::ICMP_EQ:
3385 default: assert(0 && "Unknown integer condition code!");
3386 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3387 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3388 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3389 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3390 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3391 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3392 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3393 return ReplaceInstUsesWith(I, LHS);
3395 case ICmpInst::ICMP_NE:
3397 default: assert(0 && "Unknown integer condition code!");
3398 case ICmpInst::ICMP_ULT:
3399 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3400 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3401 break; // (X != 13 & X u< 15) -> no change
3402 case ICmpInst::ICMP_SLT:
3403 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3404 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3405 break; // (X != 13 & X s< 15) -> no change
3406 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3407 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3408 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3409 return ReplaceInstUsesWith(I, RHS);
3410 case ICmpInst::ICMP_NE:
3411 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3412 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3413 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3414 LHSVal->getName()+".off");
3415 InsertNewInstBefore(Add, I);
3416 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3417 ConstantInt::get(Add->getType(), 1));
3419 break; // (X != 13 & X != 15) -> no change
3422 case ICmpInst::ICMP_ULT:
3424 default: assert(0 && "Unknown integer condition code!");
3425 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3426 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3427 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3428 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3430 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3431 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3432 return ReplaceInstUsesWith(I, LHS);
3433 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3437 case ICmpInst::ICMP_SLT:
3439 default: assert(0 && "Unknown integer condition code!");
3440 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3441 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3442 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3443 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3445 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3446 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3447 return ReplaceInstUsesWith(I, LHS);
3448 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3452 case ICmpInst::ICMP_UGT:
3454 default: assert(0 && "Unknown integer condition code!");
3455 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3456 return ReplaceInstUsesWith(I, LHS);
3457 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3458 return ReplaceInstUsesWith(I, RHS);
3459 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3461 case ICmpInst::ICMP_NE:
3462 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3463 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3464 break; // (X u> 13 & X != 15) -> no change
3465 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3466 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3468 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3472 case ICmpInst::ICMP_SGT:
3474 default: assert(0 && "Unknown integer condition code!");
3475 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3476 return ReplaceInstUsesWith(I, LHS);
3477 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3478 return ReplaceInstUsesWith(I, RHS);
3479 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3481 case ICmpInst::ICMP_NE:
3482 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3483 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3484 break; // (X s> 13 & X != 15) -> no change
3485 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3486 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3488 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3496 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3497 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3498 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3499 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3500 const Type *SrcTy = Op0C->getOperand(0)->getType();
3501 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3502 // Only do this if the casts both really cause code to be generated.
3503 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3505 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3507 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3508 Op1C->getOperand(0),
3510 InsertNewInstBefore(NewOp, I);
3511 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3515 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3516 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3517 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3518 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3519 SI0->getOperand(1) == SI1->getOperand(1) &&
3520 (SI0->hasOneUse() || SI1->hasOneUse())) {
3521 Instruction *NewOp =
3522 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3524 SI0->getName()), I);
3525 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3526 SI1->getOperand(1));
3530 return Changed ? &I : 0;
3533 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3534 /// in the result. If it does, and if the specified byte hasn't been filled in
3535 /// yet, fill it in and return false.
3536 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3537 Instruction *I = dyn_cast<Instruction>(V);
3538 if (I == 0) return true;
3540 // If this is an or instruction, it is an inner node of the bswap.
3541 if (I->getOpcode() == Instruction::Or)
3542 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3543 CollectBSwapParts(I->getOperand(1), ByteValues);
3545 // If this is a shift by a constant int, and it is "24", then its operand
3546 // defines a byte. We only handle unsigned types here.
3547 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3548 // Not shifting the entire input by N-1 bytes?
3549 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3550 8*(ByteValues.size()-1))
3554 if (I->getOpcode() == Instruction::Shl) {
3555 // X << 24 defines the top byte with the lowest of the input bytes.
3556 DestNo = ByteValues.size()-1;
3558 // X >>u 24 defines the low byte with the highest of the input bytes.
3562 // If the destination byte value is already defined, the values are or'd
3563 // together, which isn't a bswap (unless it's an or of the same bits).
3564 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3566 ByteValues[DestNo] = I->getOperand(0);
3570 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3572 Value *Shift = 0, *ShiftLHS = 0;
3573 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3574 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3575 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3577 Instruction *SI = cast<Instruction>(Shift);
3579 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3580 if (ShiftAmt->getZExtValue() & 7 ||
3581 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3584 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3586 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3587 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3589 // Unknown mask for bswap.
3590 if (DestByte == ByteValues.size()) return true;
3592 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3594 if (SI->getOpcode() == Instruction::Shl)
3595 SrcByte = DestByte - ShiftBytes;
3597 SrcByte = DestByte + ShiftBytes;
3599 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3600 if (SrcByte != ByteValues.size()-DestByte-1)
3603 // If the destination byte value is already defined, the values are or'd
3604 // together, which isn't a bswap (unless it's an or of the same bits).
3605 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3607 ByteValues[DestByte] = SI->getOperand(0);
3611 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3612 /// If so, insert the new bswap intrinsic and return it.
3613 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3614 // We cannot bswap one byte.
3615 if (I.getType() == Type::Int8Ty)
3618 /// ByteValues - For each byte of the result, we keep track of which value
3619 /// defines each byte.
3620 SmallVector<Value*, 8> ByteValues;
3621 ByteValues.resize(TD->getTypeSize(I.getType()));
3623 // Try to find all the pieces corresponding to the bswap.
3624 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3625 CollectBSwapParts(I.getOperand(1), ByteValues))
3628 // Check to see if all of the bytes come from the same value.
3629 Value *V = ByteValues[0];
3630 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3632 // Check to make sure that all of the bytes come from the same value.
3633 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3634 if (ByteValues[i] != V)
3637 // If they do then *success* we can turn this into a bswap. Figure out what
3638 // bswap to make it into.
3639 Module *M = I.getParent()->getParent()->getParent();
3640 const char *FnName = 0;
3641 if (I.getType() == Type::Int16Ty)
3642 FnName = "llvm.bswap.i16";
3643 else if (I.getType() == Type::Int32Ty)
3644 FnName = "llvm.bswap.i32";
3645 else if (I.getType() == Type::Int64Ty)
3646 FnName = "llvm.bswap.i64";
3648 assert(0 && "Unknown integer type!");
3649 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3650 return new CallInst(F, V);
3654 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3655 bool Changed = SimplifyCommutative(I);
3656 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3658 if (isa<UndefValue>(Op1)) // X | undef -> -1
3659 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
3663 return ReplaceInstUsesWith(I, Op0);
3665 // See if we can simplify any instructions used by the instruction whose sole
3666 // purpose is to compute bits we don't care about.
3667 if (!isa<VectorType>(I.getType())) {
3668 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3669 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3670 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3671 KnownZero, KnownOne))
3676 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3677 ConstantInt *C1 = 0; Value *X = 0;
3678 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3679 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3680 Instruction *Or = BinaryOperator::createOr(X, RHS);
3681 InsertNewInstBefore(Or, I);
3683 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3686 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3687 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3688 Instruction *Or = BinaryOperator::createOr(X, RHS);
3689 InsertNewInstBefore(Or, I);
3691 return BinaryOperator::createXor(Or,
3692 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3695 // Try to fold constant and into select arguments.
3696 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3697 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3699 if (isa<PHINode>(Op0))
3700 if (Instruction *NV = FoldOpIntoPhi(I))
3704 Value *A = 0, *B = 0;
3705 ConstantInt *C1 = 0, *C2 = 0;
3707 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3708 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3709 return ReplaceInstUsesWith(I, Op1);
3710 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3711 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3712 return ReplaceInstUsesWith(I, Op0);
3714 // (A | B) | C and A | (B | C) -> bswap if possible.
3715 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3716 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3717 match(Op1, m_Or(m_Value(), m_Value())) ||
3718 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3719 match(Op1, m_Shift(m_Value(), m_Value())))) {
3720 if (Instruction *BSwap = MatchBSwap(I))
3724 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3725 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3726 MaskedValueIsZero(Op1, C1->getValue())) {
3727 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3728 InsertNewInstBefore(NOr, I);
3730 return BinaryOperator::createXor(NOr, C1);
3733 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3734 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3735 MaskedValueIsZero(Op0, C1->getValue())) {
3736 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3737 InsertNewInstBefore(NOr, I);
3739 return BinaryOperator::createXor(NOr, C1);
3742 // (A & C1)|(B & C2)
3743 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3744 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3746 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3747 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3750 // If we have: ((V + N) & C1) | (V & C2)
3751 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3752 // replace with V+N.
3753 if (C1 == ConstantExpr::getNot(C2)) {
3754 Value *V1 = 0, *V2 = 0;
3755 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3756 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3757 // Add commutes, try both ways.
3758 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3759 return ReplaceInstUsesWith(I, A);
3760 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3761 return ReplaceInstUsesWith(I, A);
3763 // Or commutes, try both ways.
3764 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3765 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3766 // Add commutes, try both ways.
3767 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3768 return ReplaceInstUsesWith(I, B);
3769 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3770 return ReplaceInstUsesWith(I, B);
3775 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3776 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3777 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3778 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3779 SI0->getOperand(1) == SI1->getOperand(1) &&
3780 (SI0->hasOneUse() || SI1->hasOneUse())) {
3781 Instruction *NewOp =
3782 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3784 SI0->getName()), I);
3785 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3786 SI1->getOperand(1));
3790 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3791 if (A == Op1) // ~A | A == -1
3792 return ReplaceInstUsesWith(I,
3793 ConstantInt::getAllOnesValue(I.getType()));
3797 // Note, A is still live here!
3798 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3800 return ReplaceInstUsesWith(I,
3801 ConstantInt::getAllOnesValue(I.getType()));
3803 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3804 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3805 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3806 I.getName()+".demorgan"), I);
3807 return BinaryOperator::createNot(And);
3811 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3812 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3813 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3816 Value *LHSVal, *RHSVal;
3817 ConstantInt *LHSCst, *RHSCst;
3818 ICmpInst::Predicate LHSCC, RHSCC;
3819 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3820 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3821 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3822 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3823 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3824 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3825 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3826 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3827 // Ensure that the larger constant is on the RHS.
3828 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3829 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3830 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3831 ICmpInst *LHS = cast<ICmpInst>(Op0);
3832 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3833 std::swap(LHS, RHS);
3834 std::swap(LHSCst, RHSCst);
3835 std::swap(LHSCC, RHSCC);
3838 // At this point, we know we have have two icmp instructions
3839 // comparing a value against two constants and or'ing the result
3840 // together. Because of the above check, we know that we only have
3841 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3842 // FoldICmpLogical check above), that the two constants are not
3844 assert(LHSCst != RHSCst && "Compares not folded above?");
3847 default: assert(0 && "Unknown integer condition code!");
3848 case ICmpInst::ICMP_EQ:
3850 default: assert(0 && "Unknown integer condition code!");
3851 case ICmpInst::ICMP_EQ:
3852 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3853 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3854 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3855 LHSVal->getName()+".off");
3856 InsertNewInstBefore(Add, I);
3857 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3858 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3860 break; // (X == 13 | X == 15) -> no change
3861 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3862 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3864 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3865 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3866 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3867 return ReplaceInstUsesWith(I, RHS);
3870 case ICmpInst::ICMP_NE:
3872 default: assert(0 && "Unknown integer condition code!");
3873 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3874 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3875 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3876 return ReplaceInstUsesWith(I, LHS);
3877 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3878 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3879 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3880 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3883 case ICmpInst::ICMP_ULT:
3885 default: assert(0 && "Unknown integer condition code!");
3886 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3888 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3889 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3891 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3893 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3894 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3895 return ReplaceInstUsesWith(I, RHS);
3896 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3900 case ICmpInst::ICMP_SLT:
3902 default: assert(0 && "Unknown integer condition code!");
3903 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3905 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3906 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3908 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3910 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3911 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3912 return ReplaceInstUsesWith(I, RHS);
3913 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3917 case ICmpInst::ICMP_UGT:
3919 default: assert(0 && "Unknown integer condition code!");
3920 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3921 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3922 return ReplaceInstUsesWith(I, LHS);
3923 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3925 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3926 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3927 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3928 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3932 case ICmpInst::ICMP_SGT:
3934 default: assert(0 && "Unknown integer condition code!");
3935 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3936 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3937 return ReplaceInstUsesWith(I, LHS);
3938 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3940 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3941 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3942 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3943 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3951 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3952 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3953 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3954 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3955 const Type *SrcTy = Op0C->getOperand(0)->getType();
3956 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3957 // Only do this if the casts both really cause code to be generated.
3958 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3960 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3962 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3963 Op1C->getOperand(0),
3965 InsertNewInstBefore(NewOp, I);
3966 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3971 return Changed ? &I : 0;
3974 // XorSelf - Implements: X ^ X --> 0
3977 XorSelf(Value *rhs) : RHS(rhs) {}
3978 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3979 Instruction *apply(BinaryOperator &Xor) const {
3985 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3986 bool Changed = SimplifyCommutative(I);
3987 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3989 if (isa<UndefValue>(Op1))
3990 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3992 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3993 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3994 assert(Result == &I && "AssociativeOpt didn't work?");
3995 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3998 // See if we can simplify any instructions used by the instruction whose sole
3999 // purpose is to compute bits we don't care about.
4000 if (!isa<VectorType>(I.getType())) {
4001 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4002 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4003 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4004 KnownZero, KnownOne))
4008 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4009 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4010 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4011 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4012 return new ICmpInst(ICI->getInversePredicate(),
4013 ICI->getOperand(0), ICI->getOperand(1));
4015 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4016 // ~(c-X) == X-c-1 == X+(-c-1)
4017 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4018 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4019 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4020 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4021 ConstantInt::get(I.getType(), 1));
4022 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4025 // ~(~X & Y) --> (X | ~Y)
4026 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
4027 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4028 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4030 BinaryOperator::createNot(Op0I->getOperand(1),
4031 Op0I->getOperand(1)->getName()+".not");
4032 InsertNewInstBefore(NotY, I);
4033 return BinaryOperator::createOr(Op0NotVal, NotY);
4037 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4038 if (Op0I->getOpcode() == Instruction::Add) {
4039 // ~(X-c) --> (-c-1)-X
4040 if (RHS->isAllOnesValue()) {
4041 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4042 return BinaryOperator::createSub(
4043 ConstantExpr::getSub(NegOp0CI,
4044 ConstantInt::get(I.getType(), 1)),
4045 Op0I->getOperand(0));
4047 } else if (Op0I->getOpcode() == Instruction::Or) {
4048 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4049 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4050 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4051 // Anything in both C1 and C2 is known to be zero, remove it from
4053 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
4054 NewRHS = ConstantExpr::getAnd(NewRHS,
4055 ConstantExpr::getNot(CommonBits));
4056 AddToWorkList(Op0I);
4057 I.setOperand(0, Op0I->getOperand(0));
4058 I.setOperand(1, NewRHS);
4064 // Try to fold constant and into select arguments.
4065 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4066 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4068 if (isa<PHINode>(Op0))
4069 if (Instruction *NV = FoldOpIntoPhi(I))
4073 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4075 return ReplaceInstUsesWith(I,
4076 ConstantInt::getAllOnesValue(I.getType()));
4078 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4080 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4083 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4086 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4087 if (A == Op0) { // B^(B|A) == (A|B)^B
4088 Op1I->swapOperands();
4090 std::swap(Op0, Op1);
4091 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4092 I.swapOperands(); // Simplified below.
4093 std::swap(Op0, Op1);
4095 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4096 if (Op0 == A) // A^(A^B) == B
4097 return ReplaceInstUsesWith(I, B);
4098 else if (Op0 == B) // A^(B^A) == B
4099 return ReplaceInstUsesWith(I, A);
4100 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4101 if (A == Op0) // A^(A&B) -> A^(B&A)
4102 Op1I->swapOperands();
4103 if (B == Op0) { // A^(B&A) -> (B&A)^A
4104 I.swapOperands(); // Simplified below.
4105 std::swap(Op0, Op1);
4110 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4113 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4114 if (A == Op1) // (B|A)^B == (A|B)^B
4116 if (B == Op1) { // (A|B)^B == A & ~B
4118 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4119 return BinaryOperator::createAnd(A, NotB);
4121 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4122 if (Op1 == A) // (A^B)^A == B
4123 return ReplaceInstUsesWith(I, B);
4124 else if (Op1 == B) // (B^A)^A == B
4125 return ReplaceInstUsesWith(I, A);
4126 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4127 if (A == Op1) // (A&B)^A -> (B&A)^A
4129 if (B == Op1 && // (B&A)^A == ~B & A
4130 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4132 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4133 return BinaryOperator::createAnd(N, Op1);
4138 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4139 if (Op0I && Op1I && Op0I->isShift() &&
4140 Op0I->getOpcode() == Op1I->getOpcode() &&
4141 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4142 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4143 Instruction *NewOp =
4144 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4145 Op1I->getOperand(0),
4146 Op0I->getName()), I);
4147 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4148 Op1I->getOperand(1));
4152 Value *A, *B, *C, *D;
4153 // (A & B)^(A | B) -> A ^ B
4154 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4155 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4156 if ((A == C && B == D) || (A == D && B == C))
4157 return BinaryOperator::createXor(A, B);
4159 // (A | B)^(A & B) -> A ^ B
4160 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4161 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4162 if ((A == C && B == D) || (A == D && B == C))
4163 return BinaryOperator::createXor(A, B);
4167 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4168 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4169 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4170 // (X & Y)^(X & Y) -> (Y^Z) & X
4171 Value *X = 0, *Y = 0, *Z = 0;
4173 X = A, Y = B, Z = D;
4175 X = A, Y = B, Z = C;
4177 X = B, Y = A, Z = D;
4179 X = B, Y = A, Z = C;
4182 Instruction *NewOp =
4183 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4184 return BinaryOperator::createAnd(NewOp, X);
4189 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4190 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4191 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4194 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4195 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4196 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4197 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4198 const Type *SrcTy = Op0C->getOperand(0)->getType();
4199 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4200 // Only do this if the casts both really cause code to be generated.
4201 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4203 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4205 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4206 Op1C->getOperand(0),
4208 InsertNewInstBefore(NewOp, I);
4209 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4213 return Changed ? &I : 0;
4216 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4217 /// overflowed for this type.
4218 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4219 ConstantInt *In2, bool IsSigned = false) {
4220 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4223 if (In2->getValue().isNegative())
4224 return Result->getValue().sgt(In1->getValue());
4226 return Result->getValue().slt(In1->getValue());
4228 return Result->getValue().ult(In1->getValue());
4231 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4232 /// code necessary to compute the offset from the base pointer (without adding
4233 /// in the base pointer). Return the result as a signed integer of intptr size.
4234 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4235 TargetData &TD = IC.getTargetData();
4236 gep_type_iterator GTI = gep_type_begin(GEP);
4237 const Type *IntPtrTy = TD.getIntPtrType();
4238 Value *Result = Constant::getNullValue(IntPtrTy);
4240 // Build a mask for high order bits.
4241 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4243 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4244 Value *Op = GEP->getOperand(i);
4245 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4246 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4247 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4248 if (!OpC->isNullValue()) {
4249 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4250 Scale = ConstantExpr::getMul(OpC, Scale);
4251 if (Constant *RC = dyn_cast<Constant>(Result))
4252 Result = ConstantExpr::getAdd(RC, Scale);
4254 // Emit an add instruction.
4255 Result = IC.InsertNewInstBefore(
4256 BinaryOperator::createAdd(Result, Scale,
4257 GEP->getName()+".offs"), I);
4261 // Convert to correct type.
4262 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4263 Op->getName()+".c"), I);
4265 // We'll let instcombine(mul) convert this to a shl if possible.
4266 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4267 GEP->getName()+".idx"), I);
4269 // Emit an add instruction.
4270 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4271 GEP->getName()+".offs"), I);
4277 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4278 /// else. At this point we know that the GEP is on the LHS of the comparison.
4279 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4280 ICmpInst::Predicate Cond,
4282 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4284 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4285 if (isa<PointerType>(CI->getOperand(0)->getType()))
4286 RHS = CI->getOperand(0);
4288 Value *PtrBase = GEPLHS->getOperand(0);
4289 if (PtrBase == RHS) {
4290 // As an optimization, we don't actually have to compute the actual value of
4291 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4292 // each index is zero or not.
4293 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4294 Instruction *InVal = 0;
4295 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4296 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4298 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4299 if (isa<UndefValue>(C)) // undef index -> undef.
4300 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4301 if (C->isNullValue())
4303 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4304 EmitIt = false; // This is indexing into a zero sized array?
4305 } else if (isa<ConstantInt>(C))
4306 return ReplaceInstUsesWith(I, // No comparison is needed here.
4307 ConstantInt::get(Type::Int1Ty,
4308 Cond == ICmpInst::ICMP_NE));
4313 new ICmpInst(Cond, GEPLHS->getOperand(i),
4314 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4318 InVal = InsertNewInstBefore(InVal, I);
4319 InsertNewInstBefore(Comp, I);
4320 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4321 InVal = BinaryOperator::createOr(InVal, Comp);
4322 else // True if all are equal
4323 InVal = BinaryOperator::createAnd(InVal, Comp);
4331 // No comparison is needed here, all indexes = 0
4332 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4333 Cond == ICmpInst::ICMP_EQ));
4336 // Only lower this if the icmp is the only user of the GEP or if we expect
4337 // the result to fold to a constant!
4338 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4339 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4340 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4341 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4342 Constant::getNullValue(Offset->getType()));
4344 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4345 // If the base pointers are different, but the indices are the same, just
4346 // compare the base pointer.
4347 if (PtrBase != GEPRHS->getOperand(0)) {
4348 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4349 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4350 GEPRHS->getOperand(0)->getType();
4352 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4353 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4354 IndicesTheSame = false;
4358 // If all indices are the same, just compare the base pointers.
4360 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4361 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4363 // Otherwise, the base pointers are different and the indices are
4364 // different, bail out.
4368 // If one of the GEPs has all zero indices, recurse.
4369 bool AllZeros = true;
4370 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4371 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4372 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4377 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4378 ICmpInst::getSwappedPredicate(Cond), I);
4380 // If the other GEP has all zero indices, recurse.
4382 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4383 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4384 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4389 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4391 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4392 // If the GEPs only differ by one index, compare it.
4393 unsigned NumDifferences = 0; // Keep track of # differences.
4394 unsigned DiffOperand = 0; // The operand that differs.
4395 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4396 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4397 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4398 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4399 // Irreconcilable differences.
4403 if (NumDifferences++) break;
4408 if (NumDifferences == 0) // SAME GEP?
4409 return ReplaceInstUsesWith(I, // No comparison is needed here.
4410 ConstantInt::get(Type::Int1Ty,
4411 Cond == ICmpInst::ICMP_EQ));
4412 else if (NumDifferences == 1) {
4413 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4414 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4415 // Make sure we do a signed comparison here.
4416 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4420 // Only lower this if the icmp is the only user of the GEP or if we expect
4421 // the result to fold to a constant!
4422 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4423 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4424 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4425 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4426 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4427 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4433 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4434 bool Changed = SimplifyCompare(I);
4435 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4437 // Fold trivial predicates.
4438 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4439 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4440 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4441 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4443 // Simplify 'fcmp pred X, X'
4445 switch (I.getPredicate()) {
4446 default: assert(0 && "Unknown predicate!");
4447 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4448 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4449 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4450 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4451 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4452 case FCmpInst::FCMP_OLT: // True if ordered and less than
4453 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4454 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4456 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4457 case FCmpInst::FCMP_ULT: // True if unordered or less than
4458 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4459 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4460 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4461 I.setPredicate(FCmpInst::FCMP_UNO);
4462 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4465 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4466 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4467 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4468 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4469 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4470 I.setPredicate(FCmpInst::FCMP_ORD);
4471 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4476 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4477 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4479 // Handle fcmp with constant RHS
4480 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4481 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4482 switch (LHSI->getOpcode()) {
4483 case Instruction::PHI:
4484 if (Instruction *NV = FoldOpIntoPhi(I))
4487 case Instruction::Select:
4488 // If either operand of the select is a constant, we can fold the
4489 // comparison into the select arms, which will cause one to be
4490 // constant folded and the select turned into a bitwise or.
4491 Value *Op1 = 0, *Op2 = 0;
4492 if (LHSI->hasOneUse()) {
4493 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4494 // Fold the known value into the constant operand.
4495 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4496 // Insert a new FCmp of the other select operand.
4497 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4498 LHSI->getOperand(2), RHSC,
4500 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4501 // Fold the known value into the constant operand.
4502 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4503 // Insert a new FCmp of the other select operand.
4504 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4505 LHSI->getOperand(1), RHSC,
4511 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4516 return Changed ? &I : 0;
4519 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4520 bool Changed = SimplifyCompare(I);
4521 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4522 const Type *Ty = Op0->getType();
4526 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4527 isTrueWhenEqual(I)));
4529 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4530 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4532 // icmp of GlobalValues can never equal each other as long as they aren't
4533 // external weak linkage type.
4534 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4535 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4536 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4537 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4538 !isTrueWhenEqual(I)));
4540 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4541 // addresses never equal each other! We already know that Op0 != Op1.
4542 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4543 isa<ConstantPointerNull>(Op0)) &&
4544 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4545 isa<ConstantPointerNull>(Op1)))
4546 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4547 !isTrueWhenEqual(I)));
4549 // icmp's with boolean values can always be turned into bitwise operations
4550 if (Ty == Type::Int1Ty) {
4551 switch (I.getPredicate()) {
4552 default: assert(0 && "Invalid icmp instruction!");
4553 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4554 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4555 InsertNewInstBefore(Xor, I);
4556 return BinaryOperator::createNot(Xor);
4558 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4559 return BinaryOperator::createXor(Op0, Op1);
4561 case ICmpInst::ICMP_UGT:
4562 case ICmpInst::ICMP_SGT:
4563 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4565 case ICmpInst::ICMP_ULT:
4566 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4567 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4568 InsertNewInstBefore(Not, I);
4569 return BinaryOperator::createAnd(Not, Op1);
4571 case ICmpInst::ICMP_UGE:
4572 case ICmpInst::ICMP_SGE:
4573 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4575 case ICmpInst::ICMP_ULE:
4576 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4577 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4578 InsertNewInstBefore(Not, I);
4579 return BinaryOperator::createOr(Not, Op1);
4584 // See if we are doing a comparison between a constant and an instruction that
4585 // can be folded into the comparison.
4586 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4587 switch (I.getPredicate()) {
4589 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4590 if (CI->isMinValue(false))
4591 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4592 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4593 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4594 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4595 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4598 case ICmpInst::ICMP_SLT:
4599 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4600 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4601 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4602 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4603 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4604 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4607 case ICmpInst::ICMP_UGT:
4608 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4609 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4610 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4611 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4612 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4613 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4616 case ICmpInst::ICMP_SGT:
4617 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4618 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4619 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4620 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4621 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4622 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4625 case ICmpInst::ICMP_ULE:
4626 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4627 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4628 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4629 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4630 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4631 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4634 case ICmpInst::ICMP_SLE:
4635 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4636 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4637 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4638 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4639 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4640 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4643 case ICmpInst::ICMP_UGE:
4644 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4645 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4646 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4647 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4648 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4649 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4652 case ICmpInst::ICMP_SGE:
4653 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4654 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4655 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4656 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4657 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4658 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4662 // If we still have a icmp le or icmp ge instruction, turn it into the
4663 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4664 // already been handled above, this requires little checking.
4666 switch (I.getPredicate()) {
4668 case ICmpInst::ICMP_ULE:
4669 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4670 case ICmpInst::ICMP_SLE:
4671 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4672 case ICmpInst::ICMP_UGE:
4673 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4674 case ICmpInst::ICMP_SGE:
4675 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4678 // See if we can fold the comparison based on bits known to be zero or one
4680 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4681 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4682 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
4683 KnownZero, KnownOne, 0))
4686 // Given the known and unknown bits, compute a range that the LHS could be
4688 if ((KnownOne | KnownZero) != 0) {
4689 // Compute the Min, Max and RHS values based on the known bits. For the
4690 // EQ and NE we use unsigned values.
4691 APInt Min(BitWidth, 0), Max(BitWidth, 0), RHSVal(CI->getValue());
4692 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4693 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4696 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4699 switch (I.getPredicate()) { // LE/GE have been folded already.
4700 default: assert(0 && "Unknown icmp opcode!");
4701 case ICmpInst::ICMP_EQ:
4702 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4703 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4705 case ICmpInst::ICMP_NE:
4706 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4707 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4709 case ICmpInst::ICMP_ULT:
4710 if (Max.ult(RHSVal))
4711 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4712 if (Min.ugt(RHSVal))
4713 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4715 case ICmpInst::ICMP_UGT:
4716 if (Min.ugt(RHSVal))
4717 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4718 if (Max.ult(RHSVal))
4719 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4721 case ICmpInst::ICMP_SLT:
4722 if (Max.slt(RHSVal))
4723 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4724 if (Min.sgt(RHSVal))
4725 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4727 case ICmpInst::ICMP_SGT:
4728 if (Min.sgt(RHSVal))
4729 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4730 if (Max.slt(RHSVal))
4731 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4736 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4737 // instruction, see if that instruction also has constants so that the
4738 // instruction can be folded into the icmp
4739 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4740 switch (LHSI->getOpcode()) {
4741 case Instruction::And:
4742 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4743 LHSI->getOperand(0)->hasOneUse()) {
4744 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4746 // If the LHS is an AND of a truncating cast, we can widen the
4747 // and/compare to be the input width without changing the value
4748 // produced, eliminating a cast.
4749 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4750 // We can do this transformation if either the AND constant does not
4751 // have its sign bit set or if it is an equality comparison.
4752 // Extending a relational comparison when we're checking the sign
4753 // bit would not work.
4754 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4755 (I.isEquality() || AndCST->getValue().isPositive() &&
4756 CI->getValue().isPositive())) {
4757 ConstantInt *NewCST;
4759 APInt NewCSTVal(AndCST->getValue()), NewCIVal(CI->getValue());
4760 uint32_t BitWidth = cast<IntegerType>(
4761 Cast->getOperand(0)->getType())->getBitWidth();
4762 NewCST = ConstantInt::get(NewCSTVal.zext(BitWidth));
4763 NewCI = ConstantInt::get(NewCIVal.zext(BitWidth));
4764 Instruction *NewAnd =
4765 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4767 InsertNewInstBefore(NewAnd, I);
4768 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4772 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4773 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4774 // happens a LOT in code produced by the C front-end, for bitfield
4776 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4777 if (Shift && !Shift->isShift())
4781 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4782 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4783 const Type *AndTy = AndCST->getType(); // Type of the and.
4785 // We can fold this as long as we can't shift unknown bits
4786 // into the mask. This can only happen with signed shift
4787 // rights, as they sign-extend.
4789 bool CanFold = Shift->isLogicalShift();
4791 // To test for the bad case of the signed shr, see if any
4792 // of the bits shifted in could be tested after the mask.
4793 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4794 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4796 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4798 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4800 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4806 if (Shift->getOpcode() == Instruction::Shl)
4807 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4809 NewCst = ConstantExpr::getShl(CI, ShAmt);
4811 // Check to see if we are shifting out any of the bits being
4813 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4814 // If we shifted bits out, the fold is not going to work out.
4815 // As a special case, check to see if this means that the
4816 // result is always true or false now.
4817 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4818 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4819 if (I.getPredicate() == ICmpInst::ICMP_NE)
4820 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4822 I.setOperand(1, NewCst);
4823 Constant *NewAndCST;
4824 if (Shift->getOpcode() == Instruction::Shl)
4825 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4827 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4828 LHSI->setOperand(1, NewAndCST);
4829 LHSI->setOperand(0, Shift->getOperand(0));
4830 AddToWorkList(Shift); // Shift is dead.
4831 AddUsesToWorkList(I);
4837 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4838 // preferable because it allows the C<<Y expression to be hoisted out
4839 // of a loop if Y is invariant and X is not.
4840 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4841 I.isEquality() && !Shift->isArithmeticShift() &&
4842 isa<Instruction>(Shift->getOperand(0))) {
4845 if (Shift->getOpcode() == Instruction::LShr) {
4846 NS = BinaryOperator::createShl(AndCST,
4847 Shift->getOperand(1), "tmp");
4849 // Insert a logical shift.
4850 NS = BinaryOperator::createLShr(AndCST,
4851 Shift->getOperand(1), "tmp");
4853 InsertNewInstBefore(cast<Instruction>(NS), I);
4855 // Compute X & (C << Y).
4856 Instruction *NewAnd = BinaryOperator::createAnd(
4857 Shift->getOperand(0), NS, LHSI->getName());
4858 InsertNewInstBefore(NewAnd, I);
4860 I.setOperand(0, NewAnd);
4866 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4867 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4868 if (I.isEquality()) {
4869 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4871 // Check that the shift amount is in range. If not, don't perform
4872 // undefined shifts. When the shift is visited it will be
4874 if (ShAmt->getZExtValue() >= TypeBits)
4877 // If we are comparing against bits always shifted out, the
4878 // comparison cannot succeed.
4880 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4881 if (Comp != CI) {// Comparing against a bit that we know is zero.
4882 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4883 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4884 return ReplaceInstUsesWith(I, Cst);
4887 if (LHSI->hasOneUse()) {
4888 // Otherwise strength reduce the shift into an and.
4889 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4890 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4891 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4894 BinaryOperator::createAnd(LHSI->getOperand(0),
4895 Mask, LHSI->getName()+".mask");
4896 Value *And = InsertNewInstBefore(AndI, I);
4897 return new ICmpInst(I.getPredicate(), And,
4898 ConstantExpr::getLShr(CI, ShAmt));
4904 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4905 case Instruction::AShr:
4906 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4907 if (I.isEquality()) {
4908 // Check that the shift amount is in range. If not, don't perform
4909 // undefined shifts. When the shift is visited it will be
4911 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4912 if (ShAmt->getZExtValue() >= TypeBits)
4915 // If we are comparing against bits always shifted out, the
4916 // comparison cannot succeed.
4918 if (LHSI->getOpcode() == Instruction::LShr)
4919 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4922 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4925 if (Comp != CI) {// Comparing against a bit that we know is zero.
4926 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4927 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4928 return ReplaceInstUsesWith(I, Cst);
4931 if (LHSI->hasOneUse() || CI->isNullValue()) {
4932 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4934 // Otherwise strength reduce the shift into an and.
4935 APInt Val(APInt::getAllOnesValue(TypeBits).shl(ShAmtVal));
4936 Constant *Mask = ConstantInt::get(Val);
4939 BinaryOperator::createAnd(LHSI->getOperand(0),
4940 Mask, LHSI->getName()+".mask");
4941 Value *And = InsertNewInstBefore(AndI, I);
4942 return new ICmpInst(I.getPredicate(), And,
4943 ConstantExpr::getShl(CI, ShAmt));
4949 case Instruction::SDiv:
4950 case Instruction::UDiv:
4951 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4952 // Fold this div into the comparison, producing a range check.
4953 // Determine, based on the divide type, what the range is being
4954 // checked. If there is an overflow on the low or high side, remember
4955 // it, otherwise compute the range [low, hi) bounding the new value.
4956 // See: InsertRangeTest above for the kinds of replacements possible.
4957 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4958 // FIXME: If the operand types don't match the type of the divide
4959 // then don't attempt this transform. The code below doesn't have the
4960 // logic to deal with a signed divide and an unsigned compare (and
4961 // vice versa). This is because (x /s C1) <s C2 produces different
4962 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4963 // (x /u C1) <u C2. Simply casting the operands and result won't
4964 // work. :( The if statement below tests that condition and bails
4966 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4967 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4969 if (DivRHS->isZero())
4970 break; // Don't hack on div by zero
4972 // Initialize the variables that will indicate the nature of the
4974 bool LoOverflow = false, HiOverflow = false;
4975 ConstantInt *LoBound = 0, *HiBound = 0;
4977 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4978 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4979 // C2 (CI). By solving for X we can turn this into a range check
4980 // instead of computing a divide.
4982 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4984 // Determine if the product overflows by seeing if the product is
4985 // not equal to the divide. Make sure we do the same kind of divide
4986 // as in the LHS instruction that we're folding.
4987 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4988 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4990 // Get the ICmp opcode
4991 ICmpInst::Predicate predicate = I.getPredicate();
4993 if (!DivIsSigned) { // udiv
4995 LoOverflow = ProdOV;
4996 HiOverflow = ProdOV ||
4997 AddWithOverflow(HiBound, LoBound, DivRHS, false);
4998 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
4999 if (CI->isNullValue()) { // (X / pos) op 0
5001 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5003 } else if (CI->getValue().isPositive()) { // (X / pos) op pos
5005 LoOverflow = ProdOV;
5006 HiOverflow = ProdOV ||
5007 AddWithOverflow(HiBound, Prod, DivRHS, true);
5008 } else { // (X / pos) op neg
5009 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5010 LoOverflow = AddWithOverflow(LoBound, Prod,
5011 cast<ConstantInt>(DivRHSH), true);
5012 HiBound = AddOne(Prod);
5013 HiOverflow = ProdOV;
5015 } else { // Divisor is < 0.
5016 if (CI->isNullValue()) { // (X / neg) op 0
5017 LoBound = AddOne(DivRHS);
5018 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5019 if (HiBound == DivRHS)
5020 LoBound = 0; // - INTMIN = INTMIN
5021 } else if (CI->getValue().isPositive()) { // (X / neg) op pos
5022 HiOverflow = LoOverflow = ProdOV;
5024 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5026 HiBound = AddOne(Prod);
5027 } else { // (X / neg) op neg
5029 LoOverflow = HiOverflow = ProdOV;
5030 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
5033 // Dividing by a negate swaps the condition.
5034 predicate = ICmpInst::getSwappedPredicate(predicate);
5038 Value *X = LHSI->getOperand(0);
5039 switch (predicate) {
5040 default: assert(0 && "Unhandled icmp opcode!");
5041 case ICmpInst::ICMP_EQ:
5042 if (LoOverflow && HiOverflow)
5043 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5044 else if (HiOverflow)
5045 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5046 ICmpInst::ICMP_UGE, X, LoBound);
5047 else if (LoOverflow)
5048 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5049 ICmpInst::ICMP_ULT, X, HiBound);
5051 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5053 case ICmpInst::ICMP_NE:
5054 if (LoOverflow && HiOverflow)
5055 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5056 else if (HiOverflow)
5057 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5058 ICmpInst::ICMP_ULT, X, LoBound);
5059 else if (LoOverflow)
5060 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5061 ICmpInst::ICMP_UGE, X, HiBound);
5063 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5065 case ICmpInst::ICMP_ULT:
5066 case ICmpInst::ICMP_SLT:
5068 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5069 return new ICmpInst(predicate, X, LoBound);
5070 case ICmpInst::ICMP_UGT:
5071 case ICmpInst::ICMP_SGT:
5073 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5074 if (predicate == ICmpInst::ICMP_UGT)
5075 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5077 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5084 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5085 if (I.isEquality()) {
5086 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5088 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5089 // the second operand is a constant, simplify a bit.
5090 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
5091 switch (BO->getOpcode()) {
5092 case Instruction::SRem:
5093 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5094 if (CI->isZero() && isa<ConstantInt>(BO->getOperand(1)) &&
5096 APInt V(cast<ConstantInt>(BO->getOperand(1))->getValue());
5097 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5098 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
5099 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
5100 return new ICmpInst(I.getPredicate(), NewRem,
5101 Constant::getNullValue(BO->getType()));
5105 case Instruction::Add:
5106 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5107 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5108 if (BO->hasOneUse())
5109 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5110 ConstantExpr::getSub(CI, BOp1C));
5111 } else if (CI->isNullValue()) {
5112 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5113 // efficiently invertible, or if the add has just this one use.
5114 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5116 if (Value *NegVal = dyn_castNegVal(BOp1))
5117 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
5118 else if (Value *NegVal = dyn_castNegVal(BOp0))
5119 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
5120 else if (BO->hasOneUse()) {
5121 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5122 InsertNewInstBefore(Neg, I);
5124 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5128 case Instruction::Xor:
5129 // For the xor case, we can xor two constants together, eliminating
5130 // the explicit xor.
5131 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5132 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5133 ConstantExpr::getXor(CI, BOC));
5136 case Instruction::Sub:
5137 // Replace (([sub|xor] A, B) != 0) with (A != B)
5139 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5143 case Instruction::Or:
5144 // If bits are being or'd in that are not present in the constant we
5145 // are comparing against, then the comparison could never succeed!
5146 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5147 Constant *NotCI = ConstantExpr::getNot(CI);
5148 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5149 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5154 case Instruction::And:
5155 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5156 // If bits are being compared against that are and'd out, then the
5157 // comparison can never succeed!
5158 if (!ConstantExpr::getAnd(CI,
5159 ConstantExpr::getNot(BOC))->isNullValue())
5160 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5163 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5164 if (CI == BOC && isOneBitSet(CI))
5165 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5166 ICmpInst::ICMP_NE, Op0,
5167 Constant::getNullValue(CI->getType()));
5169 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5170 if (isSignBit(BOC)) {
5171 Value *X = BO->getOperand(0);
5172 Constant *Zero = Constant::getNullValue(X->getType());
5173 ICmpInst::Predicate pred = isICMP_NE ?
5174 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5175 return new ICmpInst(pred, X, Zero);
5178 // ((X & ~7) == 0) --> X < 8
5179 if (CI->isNullValue() && isHighOnes(BOC)) {
5180 Value *X = BO->getOperand(0);
5181 Constant *NegX = ConstantExpr::getNeg(BOC);
5182 ICmpInst::Predicate pred = isICMP_NE ?
5183 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5184 return new ICmpInst(pred, X, NegX);
5190 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5191 // Handle set{eq|ne} <intrinsic>, intcst.
5192 switch (II->getIntrinsicID()) {
5194 case Intrinsic::bswap_i16:
5195 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5196 AddToWorkList(II); // Dead?
5197 I.setOperand(0, II->getOperand(1));
5198 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5199 ByteSwap_16(CI->getZExtValue())));
5201 case Intrinsic::bswap_i32:
5202 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5203 AddToWorkList(II); // Dead?
5204 I.setOperand(0, II->getOperand(1));
5205 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5206 ByteSwap_32(CI->getZExtValue())));
5208 case Intrinsic::bswap_i64:
5209 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5210 AddToWorkList(II); // Dead?
5211 I.setOperand(0, II->getOperand(1));
5212 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5213 ByteSwap_64(CI->getZExtValue())));
5217 } else { // Not a ICMP_EQ/ICMP_NE
5218 // If the LHS is a cast from an integral value of the same size, then
5219 // since we know the RHS is a constant, try to simlify.
5220 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5221 Value *CastOp = Cast->getOperand(0);
5222 const Type *SrcTy = CastOp->getType();
5223 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5224 if (SrcTy->isInteger() &&
5225 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5226 // If this is an unsigned comparison, try to make the comparison use
5227 // smaller constant values.
5228 switch (I.getPredicate()) {
5230 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5231 ConstantInt *CUI = cast<ConstantInt>(CI);
5232 if (CUI->getValue() == APInt::getSignBit(SrcTySize))
5233 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5234 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5237 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5238 ConstantInt *CUI = cast<ConstantInt>(CI);
5239 if (CUI->getValue() == APInt::getSignedMaxValue(SrcTySize))
5240 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5241 Constant::getNullValue(SrcTy));
5251 // Handle icmp with constant RHS
5252 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5253 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5254 switch (LHSI->getOpcode()) {
5255 case Instruction::GetElementPtr:
5256 if (RHSC->isNullValue()) {
5257 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5258 bool isAllZeros = true;
5259 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5260 if (!isa<Constant>(LHSI->getOperand(i)) ||
5261 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5266 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5267 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5271 case Instruction::PHI:
5272 if (Instruction *NV = FoldOpIntoPhi(I))
5275 case Instruction::Select:
5276 // If either operand of the select is a constant, we can fold the
5277 // comparison into the select arms, which will cause one to be
5278 // constant folded and the select turned into a bitwise or.
5279 Value *Op1 = 0, *Op2 = 0;
5280 if (LHSI->hasOneUse()) {
5281 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5282 // Fold the known value into the constant operand.
5283 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5284 // Insert a new ICmp of the other select operand.
5285 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5286 LHSI->getOperand(2), RHSC,
5288 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5289 // Fold the known value into the constant operand.
5290 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5291 // Insert a new ICmp of the other select operand.
5292 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5293 LHSI->getOperand(1), RHSC,
5299 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5304 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5305 if (User *GEP = dyn_castGetElementPtr(Op0))
5306 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5308 if (User *GEP = dyn_castGetElementPtr(Op1))
5309 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5310 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5313 // Test to see if the operands of the icmp are casted versions of other
5314 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5316 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5317 if (isa<PointerType>(Op0->getType()) &&
5318 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5319 // We keep moving the cast from the left operand over to the right
5320 // operand, where it can often be eliminated completely.
5321 Op0 = CI->getOperand(0);
5323 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5324 // so eliminate it as well.
5325 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5326 Op1 = CI2->getOperand(0);
5328 // If Op1 is a constant, we can fold the cast into the constant.
5329 if (Op0->getType() != Op1->getType())
5330 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5331 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5333 // Otherwise, cast the RHS right before the icmp
5334 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5336 return new ICmpInst(I.getPredicate(), Op0, Op1);
5340 if (isa<CastInst>(Op0)) {
5341 // Handle the special case of: icmp (cast bool to X), <cst>
5342 // This comes up when you have code like
5345 // For generality, we handle any zero-extension of any operand comparison
5346 // with a constant or another cast from the same type.
5347 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5348 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5352 if (I.isEquality()) {
5353 Value *A, *B, *C, *D;
5354 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5355 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5356 Value *OtherVal = A == Op1 ? B : A;
5357 return new ICmpInst(I.getPredicate(), OtherVal,
5358 Constant::getNullValue(A->getType()));
5361 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5362 // A^c1 == C^c2 --> A == C^(c1^c2)
5363 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5364 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5365 if (Op1->hasOneUse()) {
5366 Constant *NC = ConstantExpr::getXor(C1, C2);
5367 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5368 return new ICmpInst(I.getPredicate(), A,
5369 InsertNewInstBefore(Xor, I));
5372 // A^B == A^D -> B == D
5373 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5374 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5375 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5376 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5380 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5381 (A == Op0 || B == Op0)) {
5382 // A == (A^B) -> B == 0
5383 Value *OtherVal = A == Op0 ? B : A;
5384 return new ICmpInst(I.getPredicate(), OtherVal,
5385 Constant::getNullValue(A->getType()));
5387 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5388 // (A-B) == A -> B == 0
5389 return new ICmpInst(I.getPredicate(), B,
5390 Constant::getNullValue(B->getType()));
5392 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5393 // A == (A-B) -> B == 0
5394 return new ICmpInst(I.getPredicate(), B,
5395 Constant::getNullValue(B->getType()));
5398 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5399 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5400 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5401 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5402 Value *X = 0, *Y = 0, *Z = 0;
5405 X = B; Y = D; Z = A;
5406 } else if (A == D) {
5407 X = B; Y = C; Z = A;
5408 } else if (B == C) {
5409 X = A; Y = D; Z = B;
5410 } else if (B == D) {
5411 X = A; Y = C; Z = B;
5414 if (X) { // Build (X^Y) & Z
5415 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5416 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5417 I.setOperand(0, Op1);
5418 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5423 return Changed ? &I : 0;
5426 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5427 // We only handle extending casts so far.
5429 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5430 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5431 Value *LHSCIOp = LHSCI->getOperand(0);
5432 const Type *SrcTy = LHSCIOp->getType();
5433 const Type *DestTy = LHSCI->getType();
5436 // We only handle extension cast instructions, so far. Enforce this.
5437 if (LHSCI->getOpcode() != Instruction::ZExt &&
5438 LHSCI->getOpcode() != Instruction::SExt)
5441 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5442 bool isSignedCmp = ICI.isSignedPredicate();
5444 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5445 // Not an extension from the same type?
5446 RHSCIOp = CI->getOperand(0);
5447 if (RHSCIOp->getType() != LHSCIOp->getType())
5450 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5451 // and the other is a zext), then we can't handle this.
5452 if (CI->getOpcode() != LHSCI->getOpcode())
5455 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5456 // then we can't handle this.
5457 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5460 // Okay, just insert a compare of the reduced operands now!
5461 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5464 // If we aren't dealing with a constant on the RHS, exit early
5465 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5469 // Compute the constant that would happen if we truncated to SrcTy then
5470 // reextended to DestTy.
5471 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5472 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5474 // If the re-extended constant didn't change...
5476 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5477 // For example, we might have:
5478 // %A = sext short %X to uint
5479 // %B = icmp ugt uint %A, 1330
5480 // It is incorrect to transform this into
5481 // %B = icmp ugt short %X, 1330
5482 // because %A may have negative value.
5484 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5485 // OR operation is EQ/NE.
5486 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5487 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5492 // The re-extended constant changed so the constant cannot be represented
5493 // in the shorter type. Consequently, we cannot emit a simple comparison.
5495 // First, handle some easy cases. We know the result cannot be equal at this
5496 // point so handle the ICI.isEquality() cases
5497 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5498 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5499 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5500 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5502 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5503 // should have been folded away previously and not enter in here.
5506 // We're performing a signed comparison.
5507 if (cast<ConstantInt>(CI)->getValue().isNegative())
5508 Result = ConstantInt::getFalse(); // X < (small) --> false
5510 Result = ConstantInt::getTrue(); // X < (large) --> true
5512 // We're performing an unsigned comparison.
5514 // We're performing an unsigned comp with a sign extended value.
5515 // This is true if the input is >= 0. [aka >s -1]
5516 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5517 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5518 NegOne, ICI.getName()), ICI);
5520 // Unsigned extend & unsigned compare -> always true.
5521 Result = ConstantInt::getTrue();
5525 // Finally, return the value computed.
5526 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5527 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5528 return ReplaceInstUsesWith(ICI, Result);
5530 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5531 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5532 "ICmp should be folded!");
5533 if (Constant *CI = dyn_cast<Constant>(Result))
5534 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5536 return BinaryOperator::createNot(Result);
5540 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5541 return commonShiftTransforms(I);
5544 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5545 return commonShiftTransforms(I);
5548 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5549 return commonShiftTransforms(I);
5552 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5553 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5554 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5556 // shl X, 0 == X and shr X, 0 == X
5557 // shl 0, X == 0 and shr 0, X == 0
5558 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5559 Op0 == Constant::getNullValue(Op0->getType()))
5560 return ReplaceInstUsesWith(I, Op0);
5562 if (isa<UndefValue>(Op0)) {
5563 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5564 return ReplaceInstUsesWith(I, Op0);
5565 else // undef << X -> 0, undef >>u X -> 0
5566 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5568 if (isa<UndefValue>(Op1)) {
5569 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5570 return ReplaceInstUsesWith(I, Op0);
5571 else // X << undef, X >>u undef -> 0
5572 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5575 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5576 if (I.getOpcode() == Instruction::AShr)
5577 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5578 if (CSI->isAllOnesValue())
5579 return ReplaceInstUsesWith(I, CSI);
5581 // Try to fold constant and into select arguments.
5582 if (isa<Constant>(Op0))
5583 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5584 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5587 // See if we can turn a signed shr into an unsigned shr.
5588 if (I.isArithmeticShift()) {
5589 if (MaskedValueIsZero(Op0,
5590 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5591 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5595 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5596 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5601 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5602 BinaryOperator &I) {
5603 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5605 // See if we can simplify any instructions used by the instruction whose sole
5606 // purpose is to compute bits we don't care about.
5607 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5608 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5609 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5610 KnownZero, KnownOne))
5613 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5614 // of a signed value.
5616 if (Op1->getZExtValue() >= TypeBits) { // shift amount always <= 32 bits
5617 if (I.getOpcode() != Instruction::AShr)
5618 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5620 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5625 // ((X*C1) << C2) == (X * (C1 << C2))
5626 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5627 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5628 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5629 return BinaryOperator::createMul(BO->getOperand(0),
5630 ConstantExpr::getShl(BOOp, Op1));
5632 // Try to fold constant and into select arguments.
5633 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5634 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5636 if (isa<PHINode>(Op0))
5637 if (Instruction *NV = FoldOpIntoPhi(I))
5640 if (Op0->hasOneUse()) {
5641 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5642 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5645 switch (Op0BO->getOpcode()) {
5647 case Instruction::Add:
5648 case Instruction::And:
5649 case Instruction::Or:
5650 case Instruction::Xor: {
5651 // These operators commute.
5652 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5653 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5654 match(Op0BO->getOperand(1),
5655 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5656 Instruction *YS = BinaryOperator::createShl(
5657 Op0BO->getOperand(0), Op1,
5659 InsertNewInstBefore(YS, I); // (Y << C)
5661 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5662 Op0BO->getOperand(1)->getName());
5663 InsertNewInstBefore(X, I); // (X + (Y << C))
5664 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5665 C2 = ConstantExpr::getShl(C2, Op1);
5666 return BinaryOperator::createAnd(X, C2);
5669 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5670 Value *Op0BOOp1 = Op0BO->getOperand(1);
5671 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5673 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5674 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5676 Instruction *YS = BinaryOperator::createShl(
5677 Op0BO->getOperand(0), Op1,
5679 InsertNewInstBefore(YS, I); // (Y << C)
5681 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5682 V1->getName()+".mask");
5683 InsertNewInstBefore(XM, I); // X & (CC << C)
5685 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5690 case Instruction::Sub: {
5691 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5692 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5693 match(Op0BO->getOperand(0),
5694 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5695 Instruction *YS = BinaryOperator::createShl(
5696 Op0BO->getOperand(1), Op1,
5698 InsertNewInstBefore(YS, I); // (Y << C)
5700 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5701 Op0BO->getOperand(0)->getName());
5702 InsertNewInstBefore(X, I); // (X + (Y << C))
5703 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5704 C2 = ConstantExpr::getShl(C2, Op1);
5705 return BinaryOperator::createAnd(X, C2);
5708 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5709 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5710 match(Op0BO->getOperand(0),
5711 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5712 m_ConstantInt(CC))) && V2 == Op1 &&
5713 cast<BinaryOperator>(Op0BO->getOperand(0))
5714 ->getOperand(0)->hasOneUse()) {
5715 Instruction *YS = BinaryOperator::createShl(
5716 Op0BO->getOperand(1), Op1,
5718 InsertNewInstBefore(YS, I); // (Y << C)
5720 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5721 V1->getName()+".mask");
5722 InsertNewInstBefore(XM, I); // X & (CC << C)
5724 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5732 // If the operand is an bitwise operator with a constant RHS, and the
5733 // shift is the only use, we can pull it out of the shift.
5734 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5735 bool isValid = true; // Valid only for And, Or, Xor
5736 bool highBitSet = false; // Transform if high bit of constant set?
5738 switch (Op0BO->getOpcode()) {
5739 default: isValid = false; break; // Do not perform transform!
5740 case Instruction::Add:
5741 isValid = isLeftShift;
5743 case Instruction::Or:
5744 case Instruction::Xor:
5747 case Instruction::And:
5752 // If this is a signed shift right, and the high bit is modified
5753 // by the logical operation, do not perform the transformation.
5754 // The highBitSet boolean indicates the value of the high bit of
5755 // the constant which would cause it to be modified for this
5758 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5759 isValid = ((Op0C->getValue() & APInt::getSignBit(TypeBits)) != 0) ==
5764 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5766 Instruction *NewShift =
5767 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5768 InsertNewInstBefore(NewShift, I);
5769 NewShift->takeName(Op0BO);
5771 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5778 // Find out if this is a shift of a shift by a constant.
5779 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5780 if (ShiftOp && !ShiftOp->isShift())
5783 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5784 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5785 // These shift amounts are always <= 32 bits.
5786 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5787 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5788 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5789 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5790 Value *X = ShiftOp->getOperand(0);
5792 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5793 if (AmtSum > TypeBits)
5796 const IntegerType *Ty = cast<IntegerType>(I.getType());
5798 // Check for (X << c1) << c2 and (X >> c1) >> c2
5799 if (I.getOpcode() == ShiftOp->getOpcode()) {
5800 return BinaryOperator::create(I.getOpcode(), X,
5801 ConstantInt::get(Ty, AmtSum));
5802 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5803 I.getOpcode() == Instruction::AShr) {
5804 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5805 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5806 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5807 I.getOpcode() == Instruction::LShr) {
5808 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5809 Instruction *Shift =
5810 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5811 InsertNewInstBefore(Shift, I);
5813 APInt Mask(Ty->getMask().lshr(ShiftAmt2));
5814 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5817 // Okay, if we get here, one shift must be left, and the other shift must be
5818 // right. See if the amounts are equal.
5819 if (ShiftAmt1 == ShiftAmt2) {
5820 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5821 if (I.getOpcode() == Instruction::Shl) {
5822 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
5823 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5825 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5826 if (I.getOpcode() == Instruction::LShr) {
5827 APInt Mask(Ty->getMask().lshr(ShiftAmt1));
5828 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5830 // We can simplify ((X << C) >>s C) into a trunc + sext.
5831 // NOTE: we could do this for any C, but that would make 'unusual' integer
5832 // types. For now, just stick to ones well-supported by the code
5834 const Type *SExtType = 0;
5835 switch (Ty->getBitWidth() - ShiftAmt1) {
5836 case 1 : SExtType = Type::Int1Ty; break;
5837 case 8 : SExtType = Type::Int8Ty; break;
5838 case 16 : SExtType = Type::Int16Ty; break;
5839 case 32 : SExtType = Type::Int32Ty; break;
5840 case 64 : SExtType = Type::Int64Ty; break;
5844 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5845 InsertNewInstBefore(NewTrunc, I);
5846 return new SExtInst(NewTrunc, Ty);
5848 // Otherwise, we can't handle it yet.
5849 } else if (ShiftAmt1 < ShiftAmt2) {
5850 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5852 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5853 if (I.getOpcode() == Instruction::Shl) {
5854 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5855 ShiftOp->getOpcode() == Instruction::AShr);
5856 Instruction *Shift =
5857 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5858 InsertNewInstBefore(Shift, I);
5860 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5861 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5864 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5865 if (I.getOpcode() == Instruction::LShr) {
5866 assert(ShiftOp->getOpcode() == Instruction::Shl);
5867 Instruction *Shift =
5868 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5869 InsertNewInstBefore(Shift, I);
5871 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5872 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5875 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5877 assert(ShiftAmt2 < ShiftAmt1);
5878 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5880 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5881 if (I.getOpcode() == Instruction::Shl) {
5882 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5883 ShiftOp->getOpcode() == Instruction::AShr);
5884 Instruction *Shift =
5885 BinaryOperator::create(ShiftOp->getOpcode(), X,
5886 ConstantInt::get(Ty, ShiftDiff));
5887 InsertNewInstBefore(Shift, I);
5889 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5890 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5893 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5894 if (I.getOpcode() == Instruction::LShr) {
5895 assert(ShiftOp->getOpcode() == Instruction::Shl);
5896 Instruction *Shift =
5897 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5898 InsertNewInstBefore(Shift, I);
5900 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5901 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5904 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5911 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5912 /// expression. If so, decompose it, returning some value X, such that Val is
5915 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5917 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5918 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5919 Offset = CI->getZExtValue();
5921 return ConstantInt::get(Type::Int32Ty, 0);
5922 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5923 if (I->getNumOperands() == 2) {
5924 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5925 if (I->getOpcode() == Instruction::Shl) {
5926 // This is a value scaled by '1 << the shift amt'.
5927 Scale = 1U << CUI->getZExtValue();
5929 return I->getOperand(0);
5930 } else if (I->getOpcode() == Instruction::Mul) {
5931 // This value is scaled by 'CUI'.
5932 Scale = CUI->getZExtValue();
5934 return I->getOperand(0);
5935 } else if (I->getOpcode() == Instruction::Add) {
5936 // We have X+C. Check to see if we really have (X*C2)+C1,
5937 // where C1 is divisible by C2.
5940 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5941 Offset += CUI->getZExtValue();
5942 if (SubScale > 1 && (Offset % SubScale == 0)) {
5951 // Otherwise, we can't look past this.
5958 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5959 /// try to eliminate the cast by moving the type information into the alloc.
5960 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5961 AllocationInst &AI) {
5962 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5963 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5965 // Remove any uses of AI that are dead.
5966 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5968 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5969 Instruction *User = cast<Instruction>(*UI++);
5970 if (isInstructionTriviallyDead(User)) {
5971 while (UI != E && *UI == User)
5972 ++UI; // If this instruction uses AI more than once, don't break UI.
5975 DOUT << "IC: DCE: " << *User;
5976 EraseInstFromFunction(*User);
5980 // Get the type really allocated and the type casted to.
5981 const Type *AllocElTy = AI.getAllocatedType();
5982 const Type *CastElTy = PTy->getElementType();
5983 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5985 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5986 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5987 if (CastElTyAlign < AllocElTyAlign) return 0;
5989 // If the allocation has multiple uses, only promote it if we are strictly
5990 // increasing the alignment of the resultant allocation. If we keep it the
5991 // same, we open the door to infinite loops of various kinds.
5992 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5994 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5995 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5996 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5998 // See if we can satisfy the modulus by pulling a scale out of the array
6000 unsigned ArraySizeScale, ArrayOffset;
6001 Value *NumElements = // See if the array size is a decomposable linear expr.
6002 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6004 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6006 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6007 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6009 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6014 // If the allocation size is constant, form a constant mul expression
6015 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6016 if (isa<ConstantInt>(NumElements))
6017 Amt = ConstantExpr::getMul(
6018 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6019 // otherwise multiply the amount and the number of elements
6020 else if (Scale != 1) {
6021 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6022 Amt = InsertNewInstBefore(Tmp, AI);
6026 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6027 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
6028 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6029 Amt = InsertNewInstBefore(Tmp, AI);
6032 AllocationInst *New;
6033 if (isa<MallocInst>(AI))
6034 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6036 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6037 InsertNewInstBefore(New, AI);
6040 // If the allocation has multiple uses, insert a cast and change all things
6041 // that used it to use the new cast. This will also hack on CI, but it will
6043 if (!AI.hasOneUse()) {
6044 AddUsesToWorkList(AI);
6045 // New is the allocation instruction, pointer typed. AI is the original
6046 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6047 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6048 InsertNewInstBefore(NewCast, AI);
6049 AI.replaceAllUsesWith(NewCast);
6051 return ReplaceInstUsesWith(CI, New);
6054 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6055 /// and return it as type Ty without inserting any new casts and without
6056 /// changing the computed value. This is used by code that tries to decide
6057 /// whether promoting or shrinking integer operations to wider or smaller types
6058 /// will allow us to eliminate a truncate or extend.
6060 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6061 /// extension operation if Ty is larger.
6062 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6063 int &NumCastsRemoved) {
6064 // We can always evaluate constants in another type.
6065 if (isa<ConstantInt>(V))
6068 Instruction *I = dyn_cast<Instruction>(V);
6069 if (!I) return false;
6071 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6073 switch (I->getOpcode()) {
6074 case Instruction::Add:
6075 case Instruction::Sub:
6076 case Instruction::And:
6077 case Instruction::Or:
6078 case Instruction::Xor:
6079 if (!I->hasOneUse()) return false;
6080 // These operators can all arbitrarily be extended or truncated.
6081 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6082 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6084 case Instruction::Shl:
6085 if (!I->hasOneUse()) return false;
6086 // If we are truncating the result of this SHL, and if it's a shift of a
6087 // constant amount, we can always perform a SHL in a smaller type.
6088 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6089 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6090 CI->getZExtValue() < Ty->getBitWidth())
6091 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6094 case Instruction::LShr:
6095 if (!I->hasOneUse()) return false;
6096 // If this is a truncate of a logical shr, we can truncate it to a smaller
6097 // lshr iff we know that the bits we would otherwise be shifting in are
6099 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6100 uint32_t BitWidth = OrigTy->getBitWidth();
6101 if (Ty->getBitWidth() < BitWidth &&
6102 MaskedValueIsZero(I->getOperand(0),
6103 APInt::getAllOnesValue(BitWidth) &
6104 APInt::getAllOnesValue(Ty->getBitWidth()).zextOrTrunc(BitWidth).flip())
6105 && CI->getZExtValue() < Ty->getBitWidth()) {
6106 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
6110 case Instruction::Trunc:
6111 case Instruction::ZExt:
6112 case Instruction::SExt:
6113 // If this is a cast from the destination type, we can trivially eliminate
6114 // it, and this will remove a cast overall.
6115 if (I->getOperand(0)->getType() == Ty) {
6116 // If the first operand is itself a cast, and is eliminable, do not count
6117 // this as an eliminable cast. We would prefer to eliminate those two
6119 if (isa<CastInst>(I->getOperand(0)))
6127 // TODO: Can handle more cases here.
6134 /// EvaluateInDifferentType - Given an expression that
6135 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6136 /// evaluate the expression.
6137 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6139 if (Constant *C = dyn_cast<Constant>(V))
6140 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6142 // Otherwise, it must be an instruction.
6143 Instruction *I = cast<Instruction>(V);
6144 Instruction *Res = 0;
6145 switch (I->getOpcode()) {
6146 case Instruction::Add:
6147 case Instruction::Sub:
6148 case Instruction::And:
6149 case Instruction::Or:
6150 case Instruction::Xor:
6151 case Instruction::AShr:
6152 case Instruction::LShr:
6153 case Instruction::Shl: {
6154 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6155 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6156 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6157 LHS, RHS, I->getName());
6160 case Instruction::Trunc:
6161 case Instruction::ZExt:
6162 case Instruction::SExt:
6163 case Instruction::BitCast:
6164 // If the source type of the cast is the type we're trying for then we can
6165 // just return the source. There's no need to insert it because its not new.
6166 if (I->getOperand(0)->getType() == Ty)
6167 return I->getOperand(0);
6169 // Some other kind of cast, which shouldn't happen, so just ..
6172 // TODO: Can handle more cases here.
6173 assert(0 && "Unreachable!");
6177 return InsertNewInstBefore(Res, *I);
6180 /// @brief Implement the transforms common to all CastInst visitors.
6181 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6182 Value *Src = CI.getOperand(0);
6184 // Casting undef to anything results in undef so might as just replace it and
6185 // get rid of the cast.
6186 if (isa<UndefValue>(Src)) // cast undef -> undef
6187 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6189 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6190 // eliminate it now.
6191 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6192 if (Instruction::CastOps opc =
6193 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6194 // The first cast (CSrc) is eliminable so we need to fix up or replace
6195 // the second cast (CI). CSrc will then have a good chance of being dead.
6196 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6200 // If casting the result of a getelementptr instruction with no offset, turn
6201 // this into a cast of the original pointer!
6203 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6204 bool AllZeroOperands = true;
6205 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6206 if (!isa<Constant>(GEP->getOperand(i)) ||
6207 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6208 AllZeroOperands = false;
6211 if (AllZeroOperands) {
6212 // Changing the cast operand is usually not a good idea but it is safe
6213 // here because the pointer operand is being replaced with another
6214 // pointer operand so the opcode doesn't need to change.
6215 CI.setOperand(0, GEP->getOperand(0));
6220 // If we are casting a malloc or alloca to a pointer to a type of the same
6221 // size, rewrite the allocation instruction to allocate the "right" type.
6222 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6223 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6226 // If we are casting a select then fold the cast into the select
6227 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6228 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6231 // If we are casting a PHI then fold the cast into the PHI
6232 if (isa<PHINode>(Src))
6233 if (Instruction *NV = FoldOpIntoPhi(CI))
6239 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6240 /// integer types. This function implements the common transforms for all those
6242 /// @brief Implement the transforms common to CastInst with integer operands
6243 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6244 if (Instruction *Result = commonCastTransforms(CI))
6247 Value *Src = CI.getOperand(0);
6248 const Type *SrcTy = Src->getType();
6249 const Type *DestTy = CI.getType();
6250 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6251 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6253 // See if we can simplify any instructions used by the LHS whose sole
6254 // purpose is to compute bits we don't care about.
6255 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6256 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6257 KnownZero, KnownOne))
6260 // If the source isn't an instruction or has more than one use then we
6261 // can't do anything more.
6262 Instruction *SrcI = dyn_cast<Instruction>(Src);
6263 if (!SrcI || !Src->hasOneUse())
6266 // Attempt to propagate the cast into the instruction for int->int casts.
6267 int NumCastsRemoved = 0;
6268 if (!isa<BitCastInst>(CI) &&
6269 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6271 // If this cast is a truncate, evaluting in a different type always
6272 // eliminates the cast, so it is always a win. If this is a noop-cast
6273 // this just removes a noop cast which isn't pointful, but simplifies
6274 // the code. If this is a zero-extension, we need to do an AND to
6275 // maintain the clear top-part of the computation, so we require that
6276 // the input have eliminated at least one cast. If this is a sign
6277 // extension, we insert two new casts (to do the extension) so we
6278 // require that two casts have been eliminated.
6280 switch (CI.getOpcode()) {
6282 // All the others use floating point so we shouldn't actually
6283 // get here because of the check above.
6284 assert(0 && "Unknown cast type");
6285 case Instruction::Trunc:
6288 case Instruction::ZExt:
6289 DoXForm = NumCastsRemoved >= 1;
6291 case Instruction::SExt:
6292 DoXForm = NumCastsRemoved >= 2;
6294 case Instruction::BitCast:
6300 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6301 CI.getOpcode() == Instruction::SExt);
6302 assert(Res->getType() == DestTy);
6303 switch (CI.getOpcode()) {
6304 default: assert(0 && "Unknown cast type!");
6305 case Instruction::Trunc:
6306 case Instruction::BitCast:
6307 // Just replace this cast with the result.
6308 return ReplaceInstUsesWith(CI, Res);
6309 case Instruction::ZExt: {
6310 // We need to emit an AND to clear the high bits.
6311 assert(SrcBitSize < DestBitSize && "Not a zext?");
6312 Constant *C = ConstantInt::get(APInt::getAllOnesValue(SrcBitSize));
6313 C = ConstantExpr::getZExt(C, DestTy);
6314 return BinaryOperator::createAnd(Res, C);
6316 case Instruction::SExt:
6317 // We need to emit a cast to truncate, then a cast to sext.
6318 return CastInst::create(Instruction::SExt,
6319 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6325 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6326 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6328 switch (SrcI->getOpcode()) {
6329 case Instruction::Add:
6330 case Instruction::Mul:
6331 case Instruction::And:
6332 case Instruction::Or:
6333 case Instruction::Xor:
6334 // If we are discarding information, or just changing the sign,
6336 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6337 // Don't insert two casts if they cannot be eliminated. We allow
6338 // two casts to be inserted if the sizes are the same. This could
6339 // only be converting signedness, which is a noop.
6340 if (DestBitSize == SrcBitSize ||
6341 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6342 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6343 Instruction::CastOps opcode = CI.getOpcode();
6344 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6345 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6346 return BinaryOperator::create(
6347 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6351 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6352 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6353 SrcI->getOpcode() == Instruction::Xor &&
6354 Op1 == ConstantInt::getTrue() &&
6355 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6356 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6357 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6360 case Instruction::SDiv:
6361 case Instruction::UDiv:
6362 case Instruction::SRem:
6363 case Instruction::URem:
6364 // If we are just changing the sign, rewrite.
6365 if (DestBitSize == SrcBitSize) {
6366 // Don't insert two casts if they cannot be eliminated. We allow
6367 // two casts to be inserted if the sizes are the same. This could
6368 // only be converting signedness, which is a noop.
6369 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6370 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6371 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6373 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6375 return BinaryOperator::create(
6376 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6381 case Instruction::Shl:
6382 // Allow changing the sign of the source operand. Do not allow
6383 // changing the size of the shift, UNLESS the shift amount is a
6384 // constant. We must not change variable sized shifts to a smaller
6385 // size, because it is undefined to shift more bits out than exist
6387 if (DestBitSize == SrcBitSize ||
6388 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6389 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6390 Instruction::BitCast : Instruction::Trunc);
6391 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6392 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6393 return BinaryOperator::createShl(Op0c, Op1c);
6396 case Instruction::AShr:
6397 // If this is a signed shr, and if all bits shifted in are about to be
6398 // truncated off, turn it into an unsigned shr to allow greater
6400 if (DestBitSize < SrcBitSize &&
6401 isa<ConstantInt>(Op1)) {
6402 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6403 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6404 // Insert the new logical shift right.
6405 return BinaryOperator::createLShr(Op0, Op1);
6410 case Instruction::ICmp:
6411 // If we are just checking for a icmp eq of a single bit and casting it
6412 // to an integer, then shift the bit to the appropriate place and then
6413 // cast to integer to avoid the comparison.
6414 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6415 APInt Op1CV(Op1C->getValue());
6416 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6417 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6418 // cast (X == 1) to int --> X iff X has only the low bit set.
6419 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6420 // cast (X != 0) to int --> X iff X has only the low bit set.
6421 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6422 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6423 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6424 if (Op1CV == 0 || Op1CV.isPowerOf2()) {
6425 // If Op1C some other power of two, convert:
6426 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6427 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6428 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6429 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6431 // This only works for EQ and NE
6432 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6433 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6436 APInt KnownZeroMask(KnownZero ^ TypeMask);
6437 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6438 bool isNE = pred == ICmpInst::ICMP_NE;
6439 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6440 // (X&4) == 2 --> false
6441 // (X&4) != 2 --> true
6442 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6443 Res = ConstantExpr::getZExt(Res, CI.getType());
6444 return ReplaceInstUsesWith(CI, Res);
6447 unsigned ShiftAmt = KnownZeroMask.logBase2();
6450 // Perform a logical shr by shiftamt.
6451 // Insert the shift to put the result in the low bit.
6452 In = InsertNewInstBefore(
6453 BinaryOperator::createLShr(In,
6454 ConstantInt::get(In->getType(), ShiftAmt),
6455 In->getName()+".lobit"), CI);
6458 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6459 Constant *One = ConstantInt::get(In->getType(), 1);
6460 In = BinaryOperator::createXor(In, One, "tmp");
6461 InsertNewInstBefore(cast<Instruction>(In), CI);
6464 if (CI.getType() == In->getType())
6465 return ReplaceInstUsesWith(CI, In);
6467 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6476 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6477 if (Instruction *Result = commonIntCastTransforms(CI))
6480 Value *Src = CI.getOperand(0);
6481 const Type *Ty = CI.getType();
6482 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6483 unsigned SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6485 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6486 switch (SrcI->getOpcode()) {
6488 case Instruction::LShr:
6489 // We can shrink lshr to something smaller if we know the bits shifted in
6490 // are already zeros.
6491 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6492 unsigned ShAmt = ShAmtV->getZExtValue();
6494 // Get a mask for the bits shifting in.
6495 APInt Mask(APInt::getAllOnesValue(SrcBitWidth).lshr(
6496 SrcBitWidth-ShAmt).shl(DestBitWidth));
6497 Value* SrcIOp0 = SrcI->getOperand(0);
6498 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6499 if (ShAmt >= DestBitWidth) // All zeros.
6500 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6502 // Okay, we can shrink this. Truncate the input, then return a new
6504 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6505 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6507 return BinaryOperator::createLShr(V1, V2);
6509 } else { // This is a variable shr.
6511 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6512 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6513 // loop-invariant and CSE'd.
6514 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6515 Value *One = ConstantInt::get(SrcI->getType(), 1);
6517 Value *V = InsertNewInstBefore(
6518 BinaryOperator::createShl(One, SrcI->getOperand(1),
6520 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6521 SrcI->getOperand(0),
6523 Value *Zero = Constant::getNullValue(V->getType());
6524 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6534 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6535 // If one of the common conversion will work ..
6536 if (Instruction *Result = commonIntCastTransforms(CI))
6539 Value *Src = CI.getOperand(0);
6541 // If this is a cast of a cast
6542 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6543 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6544 // types and if the sizes are just right we can convert this into a logical
6545 // 'and' which will be much cheaper than the pair of casts.
6546 if (isa<TruncInst>(CSrc)) {
6547 // Get the sizes of the types involved
6548 Value *A = CSrc->getOperand(0);
6549 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6550 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6551 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6552 // If we're actually extending zero bits and the trunc is a no-op
6553 if (MidSize < DstSize && SrcSize == DstSize) {
6554 // Replace both of the casts with an And of the type mask.
6555 APInt AndValue(APInt::getAllOnesValue(MidSize).zext(SrcSize));
6556 Constant *AndConst = ConstantInt::get(AndValue);
6558 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6559 // Unfortunately, if the type changed, we need to cast it back.
6560 if (And->getType() != CI.getType()) {
6561 And->setName(CSrc->getName()+".mask");
6562 InsertNewInstBefore(And, CI);
6563 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6573 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6574 return commonIntCastTransforms(CI);
6577 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6578 return commonCastTransforms(CI);
6581 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6582 return commonCastTransforms(CI);
6585 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6586 return commonCastTransforms(CI);
6589 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6590 return commonCastTransforms(CI);
6593 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6594 return commonCastTransforms(CI);
6597 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6598 return commonCastTransforms(CI);
6601 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6602 return commonCastTransforms(CI);
6605 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6606 return commonCastTransforms(CI);
6609 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6611 // If the operands are integer typed then apply the integer transforms,
6612 // otherwise just apply the common ones.
6613 Value *Src = CI.getOperand(0);
6614 const Type *SrcTy = Src->getType();
6615 const Type *DestTy = CI.getType();
6617 if (SrcTy->isInteger() && DestTy->isInteger()) {
6618 if (Instruction *Result = commonIntCastTransforms(CI))
6621 if (Instruction *Result = commonCastTransforms(CI))
6626 // Get rid of casts from one type to the same type. These are useless and can
6627 // be replaced by the operand.
6628 if (DestTy == Src->getType())
6629 return ReplaceInstUsesWith(CI, Src);
6631 // If the source and destination are pointers, and this cast is equivalent to
6632 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6633 // This can enhance SROA and other transforms that want type-safe pointers.
6634 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6635 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6636 const Type *DstElTy = DstPTy->getElementType();
6637 const Type *SrcElTy = SrcPTy->getElementType();
6639 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6640 unsigned NumZeros = 0;
6641 while (SrcElTy != DstElTy &&
6642 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6643 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6644 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6648 // If we found a path from the src to dest, create the getelementptr now.
6649 if (SrcElTy == DstElTy) {
6650 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6651 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6656 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6657 if (SVI->hasOneUse()) {
6658 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6659 // a bitconvert to a vector with the same # elts.
6660 if (isa<VectorType>(DestTy) &&
6661 cast<VectorType>(DestTy)->getNumElements() ==
6662 SVI->getType()->getNumElements()) {
6664 // If either of the operands is a cast from CI.getType(), then
6665 // evaluating the shuffle in the casted destination's type will allow
6666 // us to eliminate at least one cast.
6667 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6668 Tmp->getOperand(0)->getType() == DestTy) ||
6669 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6670 Tmp->getOperand(0)->getType() == DestTy)) {
6671 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6672 SVI->getOperand(0), DestTy, &CI);
6673 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6674 SVI->getOperand(1), DestTy, &CI);
6675 // Return a new shuffle vector. Use the same element ID's, as we
6676 // know the vector types match #elts.
6677 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6685 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6687 /// %D = select %cond, %C, %A
6689 /// %C = select %cond, %B, 0
6692 /// Assuming that the specified instruction is an operand to the select, return
6693 /// a bitmask indicating which operands of this instruction are foldable if they
6694 /// equal the other incoming value of the select.
6696 static unsigned GetSelectFoldableOperands(Instruction *I) {
6697 switch (I->getOpcode()) {
6698 case Instruction::Add:
6699 case Instruction::Mul:
6700 case Instruction::And:
6701 case Instruction::Or:
6702 case Instruction::Xor:
6703 return 3; // Can fold through either operand.
6704 case Instruction::Sub: // Can only fold on the amount subtracted.
6705 case Instruction::Shl: // Can only fold on the shift amount.
6706 case Instruction::LShr:
6707 case Instruction::AShr:
6710 return 0; // Cannot fold
6714 /// GetSelectFoldableConstant - For the same transformation as the previous
6715 /// function, return the identity constant that goes into the select.
6716 static Constant *GetSelectFoldableConstant(Instruction *I) {
6717 switch (I->getOpcode()) {
6718 default: assert(0 && "This cannot happen!"); abort();
6719 case Instruction::Add:
6720 case Instruction::Sub:
6721 case Instruction::Or:
6722 case Instruction::Xor:
6723 case Instruction::Shl:
6724 case Instruction::LShr:
6725 case Instruction::AShr:
6726 return Constant::getNullValue(I->getType());
6727 case Instruction::And:
6728 return ConstantInt::getAllOnesValue(I->getType());
6729 case Instruction::Mul:
6730 return ConstantInt::get(I->getType(), 1);
6734 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6735 /// have the same opcode and only one use each. Try to simplify this.
6736 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6738 if (TI->getNumOperands() == 1) {
6739 // If this is a non-volatile load or a cast from the same type,
6742 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6745 return 0; // unknown unary op.
6748 // Fold this by inserting a select from the input values.
6749 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6750 FI->getOperand(0), SI.getName()+".v");
6751 InsertNewInstBefore(NewSI, SI);
6752 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6756 // Only handle binary operators here.
6757 if (!isa<BinaryOperator>(TI))
6760 // Figure out if the operations have any operands in common.
6761 Value *MatchOp, *OtherOpT, *OtherOpF;
6763 if (TI->getOperand(0) == FI->getOperand(0)) {
6764 MatchOp = TI->getOperand(0);
6765 OtherOpT = TI->getOperand(1);
6766 OtherOpF = FI->getOperand(1);
6767 MatchIsOpZero = true;
6768 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6769 MatchOp = TI->getOperand(1);
6770 OtherOpT = TI->getOperand(0);
6771 OtherOpF = FI->getOperand(0);
6772 MatchIsOpZero = false;
6773 } else if (!TI->isCommutative()) {
6775 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6776 MatchOp = TI->getOperand(0);
6777 OtherOpT = TI->getOperand(1);
6778 OtherOpF = FI->getOperand(0);
6779 MatchIsOpZero = true;
6780 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6781 MatchOp = TI->getOperand(1);
6782 OtherOpT = TI->getOperand(0);
6783 OtherOpF = FI->getOperand(1);
6784 MatchIsOpZero = true;
6789 // If we reach here, they do have operations in common.
6790 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6791 OtherOpF, SI.getName()+".v");
6792 InsertNewInstBefore(NewSI, SI);
6794 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6796 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6798 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6800 assert(0 && "Shouldn't get here");
6804 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6805 Value *CondVal = SI.getCondition();
6806 Value *TrueVal = SI.getTrueValue();
6807 Value *FalseVal = SI.getFalseValue();
6809 // select true, X, Y -> X
6810 // select false, X, Y -> Y
6811 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6812 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6814 // select C, X, X -> X
6815 if (TrueVal == FalseVal)
6816 return ReplaceInstUsesWith(SI, TrueVal);
6818 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6819 return ReplaceInstUsesWith(SI, FalseVal);
6820 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6821 return ReplaceInstUsesWith(SI, TrueVal);
6822 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6823 if (isa<Constant>(TrueVal))
6824 return ReplaceInstUsesWith(SI, TrueVal);
6826 return ReplaceInstUsesWith(SI, FalseVal);
6829 if (SI.getType() == Type::Int1Ty) {
6830 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6831 if (C->getZExtValue()) {
6832 // Change: A = select B, true, C --> A = or B, C
6833 return BinaryOperator::createOr(CondVal, FalseVal);
6835 // Change: A = select B, false, C --> A = and !B, C
6837 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6838 "not."+CondVal->getName()), SI);
6839 return BinaryOperator::createAnd(NotCond, FalseVal);
6841 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6842 if (C->getZExtValue() == false) {
6843 // Change: A = select B, C, false --> A = and B, C
6844 return BinaryOperator::createAnd(CondVal, TrueVal);
6846 // Change: A = select B, C, true --> A = or !B, C
6848 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6849 "not."+CondVal->getName()), SI);
6850 return BinaryOperator::createOr(NotCond, TrueVal);
6855 // Selecting between two integer constants?
6856 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6857 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6858 // select C, 1, 0 -> cast C to int
6859 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
6860 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6861 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
6862 // select C, 0, 1 -> cast !C to int
6864 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6865 "not."+CondVal->getName()), SI);
6866 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6869 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6871 // (x <s 0) ? -1 : 0 -> ashr x, 31
6872 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6873 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
6874 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6875 bool CanXForm = false;
6876 if (IC->isSignedPredicate())
6877 CanXForm = CmpCst->isZero() &&
6878 IC->getPredicate() == ICmpInst::ICMP_SLT;
6880 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6881 CanXForm = CmpCst->getValue() == APInt::getSignedMaxValue(Bits) &&
6882 IC->getPredicate() == ICmpInst::ICMP_UGT;
6886 // The comparison constant and the result are not neccessarily the
6887 // same width. Make an all-ones value by inserting a AShr.
6888 Value *X = IC->getOperand(0);
6889 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6890 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6891 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6893 InsertNewInstBefore(SRA, SI);
6895 // Finally, convert to the type of the select RHS. We figure out
6896 // if this requires a SExt, Trunc or BitCast based on the sizes.
6897 Instruction::CastOps opc = Instruction::BitCast;
6898 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6899 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6900 if (SRASize < SISize)
6901 opc = Instruction::SExt;
6902 else if (SRASize > SISize)
6903 opc = Instruction::Trunc;
6904 return CastInst::create(opc, SRA, SI.getType());
6909 // If one of the constants is zero (we know they can't both be) and we
6910 // have a fcmp instruction with zero, and we have an 'and' with the
6911 // non-constant value, eliminate this whole mess. This corresponds to
6912 // cases like this: ((X & 27) ? 27 : 0)
6913 if (TrueValC->isZero() || FalseValC->isZero())
6914 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6915 cast<Constant>(IC->getOperand(1))->isNullValue())
6916 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6917 if (ICA->getOpcode() == Instruction::And &&
6918 isa<ConstantInt>(ICA->getOperand(1)) &&
6919 (ICA->getOperand(1) == TrueValC ||
6920 ICA->getOperand(1) == FalseValC) &&
6921 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6922 // Okay, now we know that everything is set up, we just don't
6923 // know whether we have a icmp_ne or icmp_eq and whether the
6924 // true or false val is the zero.
6925 bool ShouldNotVal = !TrueValC->isZero();
6926 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6929 V = InsertNewInstBefore(BinaryOperator::create(
6930 Instruction::Xor, V, ICA->getOperand(1)), SI);
6931 return ReplaceInstUsesWith(SI, V);
6936 // See if we are selecting two values based on a comparison of the two values.
6937 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6938 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6939 // Transform (X == Y) ? X : Y -> Y
6940 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6941 return ReplaceInstUsesWith(SI, FalseVal);
6942 // Transform (X != Y) ? X : Y -> X
6943 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6944 return ReplaceInstUsesWith(SI, TrueVal);
6945 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6947 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6948 // Transform (X == Y) ? Y : X -> X
6949 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6950 return ReplaceInstUsesWith(SI, FalseVal);
6951 // Transform (X != Y) ? Y : X -> Y
6952 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6953 return ReplaceInstUsesWith(SI, TrueVal);
6954 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6958 // See if we are selecting two values based on a comparison of the two values.
6959 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6960 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6961 // Transform (X == Y) ? X : Y -> Y
6962 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6963 return ReplaceInstUsesWith(SI, FalseVal);
6964 // Transform (X != Y) ? X : Y -> X
6965 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6966 return ReplaceInstUsesWith(SI, TrueVal);
6967 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6969 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6970 // Transform (X == Y) ? Y : X -> X
6971 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6972 return ReplaceInstUsesWith(SI, FalseVal);
6973 // Transform (X != Y) ? Y : X -> Y
6974 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6975 return ReplaceInstUsesWith(SI, TrueVal);
6976 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6980 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6981 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6982 if (TI->hasOneUse() && FI->hasOneUse()) {
6983 Instruction *AddOp = 0, *SubOp = 0;
6985 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6986 if (TI->getOpcode() == FI->getOpcode())
6987 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6990 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6991 // even legal for FP.
6992 if (TI->getOpcode() == Instruction::Sub &&
6993 FI->getOpcode() == Instruction::Add) {
6994 AddOp = FI; SubOp = TI;
6995 } else if (FI->getOpcode() == Instruction::Sub &&
6996 TI->getOpcode() == Instruction::Add) {
6997 AddOp = TI; SubOp = FI;
7001 Value *OtherAddOp = 0;
7002 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7003 OtherAddOp = AddOp->getOperand(1);
7004 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7005 OtherAddOp = AddOp->getOperand(0);
7009 // So at this point we know we have (Y -> OtherAddOp):
7010 // select C, (add X, Y), (sub X, Z)
7011 Value *NegVal; // Compute -Z
7012 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7013 NegVal = ConstantExpr::getNeg(C);
7015 NegVal = InsertNewInstBefore(
7016 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7019 Value *NewTrueOp = OtherAddOp;
7020 Value *NewFalseOp = NegVal;
7022 std::swap(NewTrueOp, NewFalseOp);
7023 Instruction *NewSel =
7024 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7026 NewSel = InsertNewInstBefore(NewSel, SI);
7027 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7032 // See if we can fold the select into one of our operands.
7033 if (SI.getType()->isInteger()) {
7034 // See the comment above GetSelectFoldableOperands for a description of the
7035 // transformation we are doing here.
7036 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7037 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7038 !isa<Constant>(FalseVal))
7039 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7040 unsigned OpToFold = 0;
7041 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7043 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7048 Constant *C = GetSelectFoldableConstant(TVI);
7049 Instruction *NewSel =
7050 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7051 InsertNewInstBefore(NewSel, SI);
7052 NewSel->takeName(TVI);
7053 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7054 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7056 assert(0 && "Unknown instruction!!");
7061 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7062 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7063 !isa<Constant>(TrueVal))
7064 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7065 unsigned OpToFold = 0;
7066 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7068 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7073 Constant *C = GetSelectFoldableConstant(FVI);
7074 Instruction *NewSel =
7075 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7076 InsertNewInstBefore(NewSel, SI);
7077 NewSel->takeName(FVI);
7078 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7079 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7081 assert(0 && "Unknown instruction!!");
7086 if (BinaryOperator::isNot(CondVal)) {
7087 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7088 SI.setOperand(1, FalseVal);
7089 SI.setOperand(2, TrueVal);
7096 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7097 /// determine, return it, otherwise return 0.
7098 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7099 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7100 unsigned Align = GV->getAlignment();
7101 if (Align == 0 && TD)
7102 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7104 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7105 unsigned Align = AI->getAlignment();
7106 if (Align == 0 && TD) {
7107 if (isa<AllocaInst>(AI))
7108 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7109 else if (isa<MallocInst>(AI)) {
7110 // Malloc returns maximally aligned memory.
7111 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7114 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7117 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7121 } else if (isa<BitCastInst>(V) ||
7122 (isa<ConstantExpr>(V) &&
7123 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7124 User *CI = cast<User>(V);
7125 if (isa<PointerType>(CI->getOperand(0)->getType()))
7126 return GetKnownAlignment(CI->getOperand(0), TD);
7128 } else if (isa<GetElementPtrInst>(V) ||
7129 (isa<ConstantExpr>(V) &&
7130 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7131 User *GEPI = cast<User>(V);
7132 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7133 if (BaseAlignment == 0) return 0;
7135 // If all indexes are zero, it is just the alignment of the base pointer.
7136 bool AllZeroOperands = true;
7137 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7138 if (!isa<Constant>(GEPI->getOperand(i)) ||
7139 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7140 AllZeroOperands = false;
7143 if (AllZeroOperands)
7144 return BaseAlignment;
7146 // Otherwise, if the base alignment is >= the alignment we expect for the
7147 // base pointer type, then we know that the resultant pointer is aligned at
7148 // least as much as its type requires.
7151 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7152 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7153 if (TD->getABITypeAlignment(PtrTy->getElementType())
7155 const Type *GEPTy = GEPI->getType();
7156 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7157 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7165 /// visitCallInst - CallInst simplification. This mostly only handles folding
7166 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7167 /// the heavy lifting.
7169 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7170 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7171 if (!II) return visitCallSite(&CI);
7173 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7175 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7176 bool Changed = false;
7178 // memmove/cpy/set of zero bytes is a noop.
7179 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7180 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7182 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7183 if (CI->getZExtValue() == 1) {
7184 // Replace the instruction with just byte operations. We would
7185 // transform other cases to loads/stores, but we don't know if
7186 // alignment is sufficient.
7190 // If we have a memmove and the source operation is a constant global,
7191 // then the source and dest pointers can't alias, so we can change this
7192 // into a call to memcpy.
7193 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7194 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7195 if (GVSrc->isConstant()) {
7196 Module *M = CI.getParent()->getParent()->getParent();
7198 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7200 Name = "llvm.memcpy.i32";
7202 Name = "llvm.memcpy.i64";
7203 Constant *MemCpy = M->getOrInsertFunction(Name,
7204 CI.getCalledFunction()->getFunctionType());
7205 CI.setOperand(0, MemCpy);
7210 // If we can determine a pointer alignment that is bigger than currently
7211 // set, update the alignment.
7212 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7213 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7214 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7215 unsigned Align = std::min(Alignment1, Alignment2);
7216 if (MI->getAlignment()->getZExtValue() < Align) {
7217 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7220 } else if (isa<MemSetInst>(MI)) {
7221 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7222 if (MI->getAlignment()->getZExtValue() < Alignment) {
7223 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7228 if (Changed) return II;
7230 switch (II->getIntrinsicID()) {
7232 case Intrinsic::ppc_altivec_lvx:
7233 case Intrinsic::ppc_altivec_lvxl:
7234 case Intrinsic::x86_sse_loadu_ps:
7235 case Intrinsic::x86_sse2_loadu_pd:
7236 case Intrinsic::x86_sse2_loadu_dq:
7237 // Turn PPC lvx -> load if the pointer is known aligned.
7238 // Turn X86 loadups -> load if the pointer is known aligned.
7239 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7240 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7241 PointerType::get(II->getType()), CI);
7242 return new LoadInst(Ptr);
7245 case Intrinsic::ppc_altivec_stvx:
7246 case Intrinsic::ppc_altivec_stvxl:
7247 // Turn stvx -> store if the pointer is known aligned.
7248 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7249 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7250 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7252 return new StoreInst(II->getOperand(1), Ptr);
7255 case Intrinsic::x86_sse_storeu_ps:
7256 case Intrinsic::x86_sse2_storeu_pd:
7257 case Intrinsic::x86_sse2_storeu_dq:
7258 case Intrinsic::x86_sse2_storel_dq:
7259 // Turn X86 storeu -> store if the pointer is known aligned.
7260 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7261 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7262 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7264 return new StoreInst(II->getOperand(2), Ptr);
7268 case Intrinsic::x86_sse_cvttss2si: {
7269 // These intrinsics only demands the 0th element of its input vector. If
7270 // we can simplify the input based on that, do so now.
7272 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7274 II->setOperand(1, V);
7280 case Intrinsic::ppc_altivec_vperm:
7281 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7282 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7283 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7285 // Check that all of the elements are integer constants or undefs.
7286 bool AllEltsOk = true;
7287 for (unsigned i = 0; i != 16; ++i) {
7288 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7289 !isa<UndefValue>(Mask->getOperand(i))) {
7296 // Cast the input vectors to byte vectors.
7297 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7298 II->getOperand(1), Mask->getType(), CI);
7299 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7300 II->getOperand(2), Mask->getType(), CI);
7301 Value *Result = UndefValue::get(Op0->getType());
7303 // Only extract each element once.
7304 Value *ExtractedElts[32];
7305 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7307 for (unsigned i = 0; i != 16; ++i) {
7308 if (isa<UndefValue>(Mask->getOperand(i)))
7310 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7311 Idx &= 31; // Match the hardware behavior.
7313 if (ExtractedElts[Idx] == 0) {
7315 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7316 InsertNewInstBefore(Elt, CI);
7317 ExtractedElts[Idx] = Elt;
7320 // Insert this value into the result vector.
7321 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7322 InsertNewInstBefore(cast<Instruction>(Result), CI);
7324 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7329 case Intrinsic::stackrestore: {
7330 // If the save is right next to the restore, remove the restore. This can
7331 // happen when variable allocas are DCE'd.
7332 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7333 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7334 BasicBlock::iterator BI = SS;
7336 return EraseInstFromFunction(CI);
7340 // If the stack restore is in a return/unwind block and if there are no
7341 // allocas or calls between the restore and the return, nuke the restore.
7342 TerminatorInst *TI = II->getParent()->getTerminator();
7343 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7344 BasicBlock::iterator BI = II;
7345 bool CannotRemove = false;
7346 for (++BI; &*BI != TI; ++BI) {
7347 if (isa<AllocaInst>(BI) ||
7348 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7349 CannotRemove = true;
7354 return EraseInstFromFunction(CI);
7361 return visitCallSite(II);
7364 // InvokeInst simplification
7366 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7367 return visitCallSite(&II);
7370 // visitCallSite - Improvements for call and invoke instructions.
7372 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7373 bool Changed = false;
7375 // If the callee is a constexpr cast of a function, attempt to move the cast
7376 // to the arguments of the call/invoke.
7377 if (transformConstExprCastCall(CS)) return 0;
7379 Value *Callee = CS.getCalledValue();
7381 if (Function *CalleeF = dyn_cast<Function>(Callee))
7382 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7383 Instruction *OldCall = CS.getInstruction();
7384 // If the call and callee calling conventions don't match, this call must
7385 // be unreachable, as the call is undefined.
7386 new StoreInst(ConstantInt::getTrue(),
7387 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7388 if (!OldCall->use_empty())
7389 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7390 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7391 return EraseInstFromFunction(*OldCall);
7395 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7396 // This instruction is not reachable, just remove it. We insert a store to
7397 // undef so that we know that this code is not reachable, despite the fact
7398 // that we can't modify the CFG here.
7399 new StoreInst(ConstantInt::getTrue(),
7400 UndefValue::get(PointerType::get(Type::Int1Ty)),
7401 CS.getInstruction());
7403 if (!CS.getInstruction()->use_empty())
7404 CS.getInstruction()->
7405 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7407 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7408 // Don't break the CFG, insert a dummy cond branch.
7409 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7410 ConstantInt::getTrue(), II);
7412 return EraseInstFromFunction(*CS.getInstruction());
7415 const PointerType *PTy = cast<PointerType>(Callee->getType());
7416 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7417 if (FTy->isVarArg()) {
7418 // See if we can optimize any arguments passed through the varargs area of
7420 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7421 E = CS.arg_end(); I != E; ++I)
7422 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7423 // If this cast does not effect the value passed through the varargs
7424 // area, we can eliminate the use of the cast.
7425 Value *Op = CI->getOperand(0);
7426 if (CI->isLosslessCast()) {
7433 return Changed ? CS.getInstruction() : 0;
7436 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7437 // attempt to move the cast to the arguments of the call/invoke.
7439 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7440 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7441 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7442 if (CE->getOpcode() != Instruction::BitCast ||
7443 !isa<Function>(CE->getOperand(0)))
7445 Function *Callee = cast<Function>(CE->getOperand(0));
7446 Instruction *Caller = CS.getInstruction();
7448 // Okay, this is a cast from a function to a different type. Unless doing so
7449 // would cause a type conversion of one of our arguments, change this call to
7450 // be a direct call with arguments casted to the appropriate types.
7452 const FunctionType *FT = Callee->getFunctionType();
7453 const Type *OldRetTy = Caller->getType();
7455 // Check to see if we are changing the return type...
7456 if (OldRetTy != FT->getReturnType()) {
7457 if (Callee->isDeclaration() && !Caller->use_empty() &&
7458 // Conversion is ok if changing from pointer to int of same size.
7459 !(isa<PointerType>(FT->getReturnType()) &&
7460 TD->getIntPtrType() == OldRetTy))
7461 return false; // Cannot transform this return value.
7463 // If the callsite is an invoke instruction, and the return value is used by
7464 // a PHI node in a successor, we cannot change the return type of the call
7465 // because there is no place to put the cast instruction (without breaking
7466 // the critical edge). Bail out in this case.
7467 if (!Caller->use_empty())
7468 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7469 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7471 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7472 if (PN->getParent() == II->getNormalDest() ||
7473 PN->getParent() == II->getUnwindDest())
7477 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7478 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7480 CallSite::arg_iterator AI = CS.arg_begin();
7481 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7482 const Type *ParamTy = FT->getParamType(i);
7483 const Type *ActTy = (*AI)->getType();
7484 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7485 //Either we can cast directly, or we can upconvert the argument
7486 bool isConvertible = ActTy == ParamTy ||
7487 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7488 (ParamTy->isInteger() && ActTy->isInteger() &&
7489 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7490 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7491 && c->getValue().isStrictlyPositive());
7492 if (Callee->isDeclaration() && !isConvertible) return false;
7495 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7496 Callee->isDeclaration())
7497 return false; // Do not delete arguments unless we have a function body...
7499 // Okay, we decided that this is a safe thing to do: go ahead and start
7500 // inserting cast instructions as necessary...
7501 std::vector<Value*> Args;
7502 Args.reserve(NumActualArgs);
7504 AI = CS.arg_begin();
7505 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7506 const Type *ParamTy = FT->getParamType(i);
7507 if ((*AI)->getType() == ParamTy) {
7508 Args.push_back(*AI);
7510 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7511 false, ParamTy, false);
7512 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7513 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7517 // If the function takes more arguments than the call was taking, add them
7519 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7520 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7522 // If we are removing arguments to the function, emit an obnoxious warning...
7523 if (FT->getNumParams() < NumActualArgs)
7524 if (!FT->isVarArg()) {
7525 cerr << "WARNING: While resolving call to function '"
7526 << Callee->getName() << "' arguments were dropped!\n";
7528 // Add all of the arguments in their promoted form to the arg list...
7529 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7530 const Type *PTy = getPromotedType((*AI)->getType());
7531 if (PTy != (*AI)->getType()) {
7532 // Must promote to pass through va_arg area!
7533 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7535 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7536 InsertNewInstBefore(Cast, *Caller);
7537 Args.push_back(Cast);
7539 Args.push_back(*AI);
7544 if (FT->getReturnType() == Type::VoidTy)
7545 Caller->setName(""); // Void type should not have a name.
7548 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7549 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7550 &Args[0], Args.size(), Caller->getName(), Caller);
7551 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7553 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7554 if (cast<CallInst>(Caller)->isTailCall())
7555 cast<CallInst>(NC)->setTailCall();
7556 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7559 // Insert a cast of the return type as necessary.
7561 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7562 if (NV->getType() != Type::VoidTy) {
7563 const Type *CallerTy = Caller->getType();
7564 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7566 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7568 // If this is an invoke instruction, we should insert it after the first
7569 // non-phi, instruction in the normal successor block.
7570 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7571 BasicBlock::iterator I = II->getNormalDest()->begin();
7572 while (isa<PHINode>(I)) ++I;
7573 InsertNewInstBefore(NC, *I);
7575 // Otherwise, it's a call, just insert cast right after the call instr
7576 InsertNewInstBefore(NC, *Caller);
7578 AddUsersToWorkList(*Caller);
7580 NV = UndefValue::get(Caller->getType());
7584 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7585 Caller->replaceAllUsesWith(NV);
7586 Caller->eraseFromParent();
7587 RemoveFromWorkList(Caller);
7591 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7592 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7593 /// and a single binop.
7594 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7595 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7596 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7597 isa<CmpInst>(FirstInst));
7598 unsigned Opc = FirstInst->getOpcode();
7599 Value *LHSVal = FirstInst->getOperand(0);
7600 Value *RHSVal = FirstInst->getOperand(1);
7602 const Type *LHSType = LHSVal->getType();
7603 const Type *RHSType = RHSVal->getType();
7605 // Scan to see if all operands are the same opcode, all have one use, and all
7606 // kill their operands (i.e. the operands have one use).
7607 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7608 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7609 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7610 // Verify type of the LHS matches so we don't fold cmp's of different
7611 // types or GEP's with different index types.
7612 I->getOperand(0)->getType() != LHSType ||
7613 I->getOperand(1)->getType() != RHSType)
7616 // If they are CmpInst instructions, check their predicates
7617 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7618 if (cast<CmpInst>(I)->getPredicate() !=
7619 cast<CmpInst>(FirstInst)->getPredicate())
7622 // Keep track of which operand needs a phi node.
7623 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7624 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7627 // Otherwise, this is safe to transform, determine if it is profitable.
7629 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7630 // Indexes are often folded into load/store instructions, so we don't want to
7631 // hide them behind a phi.
7632 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7635 Value *InLHS = FirstInst->getOperand(0);
7636 Value *InRHS = FirstInst->getOperand(1);
7637 PHINode *NewLHS = 0, *NewRHS = 0;
7639 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7640 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7641 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7642 InsertNewInstBefore(NewLHS, PN);
7647 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7648 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7649 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7650 InsertNewInstBefore(NewRHS, PN);
7654 // Add all operands to the new PHIs.
7655 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7657 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7658 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7661 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7662 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7666 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7667 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7668 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7669 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7672 assert(isa<GetElementPtrInst>(FirstInst));
7673 return new GetElementPtrInst(LHSVal, RHSVal);
7677 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7678 /// of the block that defines it. This means that it must be obvious the value
7679 /// of the load is not changed from the point of the load to the end of the
7682 /// Finally, it is safe, but not profitable, to sink a load targetting a
7683 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7685 static bool isSafeToSinkLoad(LoadInst *L) {
7686 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7688 for (++BBI; BBI != E; ++BBI)
7689 if (BBI->mayWriteToMemory())
7692 // Check for non-address taken alloca. If not address-taken already, it isn't
7693 // profitable to do this xform.
7694 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7695 bool isAddressTaken = false;
7696 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7698 if (isa<LoadInst>(UI)) continue;
7699 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7700 // If storing TO the alloca, then the address isn't taken.
7701 if (SI->getOperand(1) == AI) continue;
7703 isAddressTaken = true;
7707 if (!isAddressTaken)
7715 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7716 // operator and they all are only used by the PHI, PHI together their
7717 // inputs, and do the operation once, to the result of the PHI.
7718 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7719 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7721 // Scan the instruction, looking for input operations that can be folded away.
7722 // If all input operands to the phi are the same instruction (e.g. a cast from
7723 // the same type or "+42") we can pull the operation through the PHI, reducing
7724 // code size and simplifying code.
7725 Constant *ConstantOp = 0;
7726 const Type *CastSrcTy = 0;
7727 bool isVolatile = false;
7728 if (isa<CastInst>(FirstInst)) {
7729 CastSrcTy = FirstInst->getOperand(0)->getType();
7730 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7731 // Can fold binop, compare or shift here if the RHS is a constant,
7732 // otherwise call FoldPHIArgBinOpIntoPHI.
7733 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7734 if (ConstantOp == 0)
7735 return FoldPHIArgBinOpIntoPHI(PN);
7736 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7737 isVolatile = LI->isVolatile();
7738 // We can't sink the load if the loaded value could be modified between the
7739 // load and the PHI.
7740 if (LI->getParent() != PN.getIncomingBlock(0) ||
7741 !isSafeToSinkLoad(LI))
7743 } else if (isa<GetElementPtrInst>(FirstInst)) {
7744 if (FirstInst->getNumOperands() == 2)
7745 return FoldPHIArgBinOpIntoPHI(PN);
7746 // Can't handle general GEPs yet.
7749 return 0; // Cannot fold this operation.
7752 // Check to see if all arguments are the same operation.
7753 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7754 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7755 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7756 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7759 if (I->getOperand(0)->getType() != CastSrcTy)
7760 return 0; // Cast operation must match.
7761 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7762 // We can't sink the load if the loaded value could be modified between
7763 // the load and the PHI.
7764 if (LI->isVolatile() != isVolatile ||
7765 LI->getParent() != PN.getIncomingBlock(i) ||
7766 !isSafeToSinkLoad(LI))
7768 } else if (I->getOperand(1) != ConstantOp) {
7773 // Okay, they are all the same operation. Create a new PHI node of the
7774 // correct type, and PHI together all of the LHS's of the instructions.
7775 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7776 PN.getName()+".in");
7777 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7779 Value *InVal = FirstInst->getOperand(0);
7780 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7782 // Add all operands to the new PHI.
7783 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7784 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7785 if (NewInVal != InVal)
7787 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7792 // The new PHI unions all of the same values together. This is really
7793 // common, so we handle it intelligently here for compile-time speed.
7797 InsertNewInstBefore(NewPN, PN);
7801 // Insert and return the new operation.
7802 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7803 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7804 else if (isa<LoadInst>(FirstInst))
7805 return new LoadInst(PhiVal, "", isVolatile);
7806 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7807 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7808 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7809 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7810 PhiVal, ConstantOp);
7812 assert(0 && "Unknown operation");
7816 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7818 static bool DeadPHICycle(PHINode *PN,
7819 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
7820 if (PN->use_empty()) return true;
7821 if (!PN->hasOneUse()) return false;
7823 // Remember this node, and if we find the cycle, return.
7824 if (!PotentiallyDeadPHIs.insert(PN))
7827 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7828 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7833 // PHINode simplification
7835 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7836 // If LCSSA is around, don't mess with Phi nodes
7837 if (MustPreserveLCSSA) return 0;
7839 if (Value *V = PN.hasConstantValue())
7840 return ReplaceInstUsesWith(PN, V);
7842 // If all PHI operands are the same operation, pull them through the PHI,
7843 // reducing code size.
7844 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7845 PN.getIncomingValue(0)->hasOneUse())
7846 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7849 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7850 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7851 // PHI)... break the cycle.
7852 if (PN.hasOneUse()) {
7853 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7854 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7855 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
7856 PotentiallyDeadPHIs.insert(&PN);
7857 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7858 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7861 // If this phi has a single use, and if that use just computes a value for
7862 // the next iteration of a loop, delete the phi. This occurs with unused
7863 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7864 // common case here is good because the only other things that catch this
7865 // are induction variable analysis (sometimes) and ADCE, which is only run
7867 if (PHIUser->hasOneUse() &&
7868 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7869 PHIUser->use_back() == &PN) {
7870 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7877 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7878 Instruction *InsertPoint,
7880 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7881 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7882 // We must cast correctly to the pointer type. Ensure that we
7883 // sign extend the integer value if it is smaller as this is
7884 // used for address computation.
7885 Instruction::CastOps opcode =
7886 (VTySize < PtrSize ? Instruction::SExt :
7887 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7888 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7892 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7893 Value *PtrOp = GEP.getOperand(0);
7894 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7895 // If so, eliminate the noop.
7896 if (GEP.getNumOperands() == 1)
7897 return ReplaceInstUsesWith(GEP, PtrOp);
7899 if (isa<UndefValue>(GEP.getOperand(0)))
7900 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7902 bool HasZeroPointerIndex = false;
7903 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7904 HasZeroPointerIndex = C->isNullValue();
7906 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7907 return ReplaceInstUsesWith(GEP, PtrOp);
7909 // Keep track of whether all indices are zero constants integers.
7910 bool AllZeroIndices = true;
7912 // Eliminate unneeded casts for indices.
7913 bool MadeChange = false;
7915 gep_type_iterator GTI = gep_type_begin(GEP);
7916 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
7917 // Track whether this GEP has all zero indices, if so, it doesn't move the
7918 // input pointer, it just changes its type.
7919 if (AllZeroIndices) {
7920 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(i)))
7921 AllZeroIndices = CI->isNullValue();
7923 AllZeroIndices = false;
7925 if (isa<SequentialType>(*GTI)) {
7926 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7927 if (CI->getOpcode() == Instruction::ZExt ||
7928 CI->getOpcode() == Instruction::SExt) {
7929 const Type *SrcTy = CI->getOperand(0)->getType();
7930 // We can eliminate a cast from i32 to i64 iff the target
7931 // is a 32-bit pointer target.
7932 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7934 GEP.setOperand(i, CI->getOperand(0));
7938 // If we are using a wider index than needed for this platform, shrink it
7939 // to what we need. If the incoming value needs a cast instruction,
7940 // insert it. This explicit cast can make subsequent optimizations more
7942 Value *Op = GEP.getOperand(i);
7943 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7944 if (Constant *C = dyn_cast<Constant>(Op)) {
7945 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7948 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7950 GEP.setOperand(i, Op);
7955 if (MadeChange) return &GEP;
7957 // If this GEP instruction doesn't move the pointer, and if the input operand
7958 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
7959 // real input to the dest type.
7960 if (AllZeroIndices && isa<BitCastInst>(GEP.getOperand(0)))
7961 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
7964 // Combine Indices - If the source pointer to this getelementptr instruction
7965 // is a getelementptr instruction, combine the indices of the two
7966 // getelementptr instructions into a single instruction.
7968 SmallVector<Value*, 8> SrcGEPOperands;
7969 if (User *Src = dyn_castGetElementPtr(PtrOp))
7970 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
7972 if (!SrcGEPOperands.empty()) {
7973 // Note that if our source is a gep chain itself that we wait for that
7974 // chain to be resolved before we perform this transformation. This
7975 // avoids us creating a TON of code in some cases.
7977 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7978 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7979 return 0; // Wait until our source is folded to completion.
7981 SmallVector<Value*, 8> Indices;
7983 // Find out whether the last index in the source GEP is a sequential idx.
7984 bool EndsWithSequential = false;
7985 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7986 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7987 EndsWithSequential = !isa<StructType>(*I);
7989 // Can we combine the two pointer arithmetics offsets?
7990 if (EndsWithSequential) {
7991 // Replace: gep (gep %P, long B), long A, ...
7992 // With: T = long A+B; gep %P, T, ...
7994 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7995 if (SO1 == Constant::getNullValue(SO1->getType())) {
7997 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8000 // If they aren't the same type, convert both to an integer of the
8001 // target's pointer size.
8002 if (SO1->getType() != GO1->getType()) {
8003 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8004 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8005 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8006 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8008 unsigned PS = TD->getPointerSize();
8009 if (TD->getTypeSize(SO1->getType()) == PS) {
8010 // Convert GO1 to SO1's type.
8011 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8013 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8014 // Convert SO1 to GO1's type.
8015 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8017 const Type *PT = TD->getIntPtrType();
8018 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8019 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8023 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8024 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8026 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8027 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8031 // Recycle the GEP we already have if possible.
8032 if (SrcGEPOperands.size() == 2) {
8033 GEP.setOperand(0, SrcGEPOperands[0]);
8034 GEP.setOperand(1, Sum);
8037 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8038 SrcGEPOperands.end()-1);
8039 Indices.push_back(Sum);
8040 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8042 } else if (isa<Constant>(*GEP.idx_begin()) &&
8043 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8044 SrcGEPOperands.size() != 1) {
8045 // Otherwise we can do the fold if the first index of the GEP is a zero
8046 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8047 SrcGEPOperands.end());
8048 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8051 if (!Indices.empty())
8052 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8053 Indices.size(), GEP.getName());
8055 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8056 // GEP of global variable. If all of the indices for this GEP are
8057 // constants, we can promote this to a constexpr instead of an instruction.
8059 // Scan for nonconstants...
8060 SmallVector<Constant*, 8> Indices;
8061 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8062 for (; I != E && isa<Constant>(*I); ++I)
8063 Indices.push_back(cast<Constant>(*I));
8065 if (I == E) { // If they are all constants...
8066 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8067 &Indices[0],Indices.size());
8069 // Replace all uses of the GEP with the new constexpr...
8070 return ReplaceInstUsesWith(GEP, CE);
8072 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8073 if (!isa<PointerType>(X->getType())) {
8074 // Not interesting. Source pointer must be a cast from pointer.
8075 } else if (HasZeroPointerIndex) {
8076 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8077 // into : GEP [10 x ubyte]* X, long 0, ...
8079 // This occurs when the program declares an array extern like "int X[];"
8081 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8082 const PointerType *XTy = cast<PointerType>(X->getType());
8083 if (const ArrayType *XATy =
8084 dyn_cast<ArrayType>(XTy->getElementType()))
8085 if (const ArrayType *CATy =
8086 dyn_cast<ArrayType>(CPTy->getElementType()))
8087 if (CATy->getElementType() == XATy->getElementType()) {
8088 // At this point, we know that the cast source type is a pointer
8089 // to an array of the same type as the destination pointer
8090 // array. Because the array type is never stepped over (there
8091 // is a leading zero) we can fold the cast into this GEP.
8092 GEP.setOperand(0, X);
8095 } else if (GEP.getNumOperands() == 2) {
8096 // Transform things like:
8097 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8098 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8099 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8100 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8101 if (isa<ArrayType>(SrcElTy) &&
8102 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8103 TD->getTypeSize(ResElTy)) {
8104 Value *V = InsertNewInstBefore(
8105 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8106 GEP.getOperand(1), GEP.getName()), GEP);
8107 // V and GEP are both pointer types --> BitCast
8108 return new BitCastInst(V, GEP.getType());
8111 // Transform things like:
8112 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8113 // (where tmp = 8*tmp2) into:
8114 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8116 if (isa<ArrayType>(SrcElTy) &&
8117 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8118 uint64_t ArrayEltSize =
8119 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8121 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8122 // allow either a mul, shift, or constant here.
8124 ConstantInt *Scale = 0;
8125 if (ArrayEltSize == 1) {
8126 NewIdx = GEP.getOperand(1);
8127 Scale = ConstantInt::get(NewIdx->getType(), 1);
8128 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8129 NewIdx = ConstantInt::get(CI->getType(), 1);
8131 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8132 if (Inst->getOpcode() == Instruction::Shl &&
8133 isa<ConstantInt>(Inst->getOperand(1))) {
8135 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
8136 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
8137 NewIdx = Inst->getOperand(0);
8138 } else if (Inst->getOpcode() == Instruction::Mul &&
8139 isa<ConstantInt>(Inst->getOperand(1))) {
8140 Scale = cast<ConstantInt>(Inst->getOperand(1));
8141 NewIdx = Inst->getOperand(0);
8145 // If the index will be to exactly the right offset with the scale taken
8146 // out, perform the transformation.
8147 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8148 if (isa<ConstantInt>(Scale))
8149 Scale = ConstantInt::get(Scale->getType(),
8150 Scale->getZExtValue() / ArrayEltSize);
8151 if (Scale->getZExtValue() != 1) {
8152 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8154 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8155 NewIdx = InsertNewInstBefore(Sc, GEP);
8158 // Insert the new GEP instruction.
8159 Instruction *NewGEP =
8160 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8161 NewIdx, GEP.getName());
8162 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8163 // The NewGEP must be pointer typed, so must the old one -> BitCast
8164 return new BitCastInst(NewGEP, GEP.getType());
8173 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8174 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8175 if (AI.isArrayAllocation()) // Check C != 1
8176 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8178 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8179 AllocationInst *New = 0;
8181 // Create and insert the replacement instruction...
8182 if (isa<MallocInst>(AI))
8183 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8185 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8186 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8189 InsertNewInstBefore(New, AI);
8191 // Scan to the end of the allocation instructions, to skip over a block of
8192 // allocas if possible...
8194 BasicBlock::iterator It = New;
8195 while (isa<AllocationInst>(*It)) ++It;
8197 // Now that I is pointing to the first non-allocation-inst in the block,
8198 // insert our getelementptr instruction...
8200 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8201 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8202 New->getName()+".sub", It);
8204 // Now make everything use the getelementptr instead of the original
8206 return ReplaceInstUsesWith(AI, V);
8207 } else if (isa<UndefValue>(AI.getArraySize())) {
8208 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8211 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8212 // Note that we only do this for alloca's, because malloc should allocate and
8213 // return a unique pointer, even for a zero byte allocation.
8214 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8215 TD->getTypeSize(AI.getAllocatedType()) == 0)
8216 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8221 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8222 Value *Op = FI.getOperand(0);
8224 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8225 if (CastInst *CI = dyn_cast<CastInst>(Op))
8226 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8227 FI.setOperand(0, CI->getOperand(0));
8231 // free undef -> unreachable.
8232 if (isa<UndefValue>(Op)) {
8233 // Insert a new store to null because we cannot modify the CFG here.
8234 new StoreInst(ConstantInt::getTrue(),
8235 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8236 return EraseInstFromFunction(FI);
8239 // If we have 'free null' delete the instruction. This can happen in stl code
8240 // when lots of inlining happens.
8241 if (isa<ConstantPointerNull>(Op))
8242 return EraseInstFromFunction(FI);
8248 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8249 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8250 User *CI = cast<User>(LI.getOperand(0));
8251 Value *CastOp = CI->getOperand(0);
8253 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8254 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8255 const Type *SrcPTy = SrcTy->getElementType();
8257 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8258 isa<VectorType>(DestPTy)) {
8259 // If the source is an array, the code below will not succeed. Check to
8260 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8262 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8263 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8264 if (ASrcTy->getNumElements() != 0) {
8266 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8267 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8268 SrcTy = cast<PointerType>(CastOp->getType());
8269 SrcPTy = SrcTy->getElementType();
8272 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8273 isa<VectorType>(SrcPTy)) &&
8274 // Do not allow turning this into a load of an integer, which is then
8275 // casted to a pointer, this pessimizes pointer analysis a lot.
8276 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8277 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8278 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8280 // Okay, we are casting from one integer or pointer type to another of
8281 // the same size. Instead of casting the pointer before the load, cast
8282 // the result of the loaded value.
8283 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8285 LI.isVolatile()),LI);
8286 // Now cast the result of the load.
8287 return new BitCastInst(NewLoad, LI.getType());
8294 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8295 /// from this value cannot trap. If it is not obviously safe to load from the
8296 /// specified pointer, we do a quick local scan of the basic block containing
8297 /// ScanFrom, to determine if the address is already accessed.
8298 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8299 // If it is an alloca or global variable, it is always safe to load from.
8300 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8302 // Otherwise, be a little bit agressive by scanning the local block where we
8303 // want to check to see if the pointer is already being loaded or stored
8304 // from/to. If so, the previous load or store would have already trapped,
8305 // so there is no harm doing an extra load (also, CSE will later eliminate
8306 // the load entirely).
8307 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8312 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8313 if (LI->getOperand(0) == V) return true;
8314 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8315 if (SI->getOperand(1) == V) return true;
8321 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8322 Value *Op = LI.getOperand(0);
8324 // load (cast X) --> cast (load X) iff safe
8325 if (isa<CastInst>(Op))
8326 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8329 // None of the following transforms are legal for volatile loads.
8330 if (LI.isVolatile()) return 0;
8332 if (&LI.getParent()->front() != &LI) {
8333 BasicBlock::iterator BBI = &LI; --BBI;
8334 // If the instruction immediately before this is a store to the same
8335 // address, do a simple form of store->load forwarding.
8336 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8337 if (SI->getOperand(1) == LI.getOperand(0))
8338 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8339 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8340 if (LIB->getOperand(0) == LI.getOperand(0))
8341 return ReplaceInstUsesWith(LI, LIB);
8344 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8345 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8346 isa<UndefValue>(GEPI->getOperand(0))) {
8347 // Insert a new store to null instruction before the load to indicate
8348 // that this code is not reachable. We do this instead of inserting
8349 // an unreachable instruction directly because we cannot modify the
8351 new StoreInst(UndefValue::get(LI.getType()),
8352 Constant::getNullValue(Op->getType()), &LI);
8353 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8356 if (Constant *C = dyn_cast<Constant>(Op)) {
8357 // load null/undef -> undef
8358 if ((C->isNullValue() || isa<UndefValue>(C))) {
8359 // Insert a new store to null instruction before the load to indicate that
8360 // this code is not reachable. We do this instead of inserting an
8361 // unreachable instruction directly because we cannot modify the CFG.
8362 new StoreInst(UndefValue::get(LI.getType()),
8363 Constant::getNullValue(Op->getType()), &LI);
8364 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8367 // Instcombine load (constant global) into the value loaded.
8368 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8369 if (GV->isConstant() && !GV->isDeclaration())
8370 return ReplaceInstUsesWith(LI, GV->getInitializer());
8372 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8373 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8374 if (CE->getOpcode() == Instruction::GetElementPtr) {
8375 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8376 if (GV->isConstant() && !GV->isDeclaration())
8378 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8379 return ReplaceInstUsesWith(LI, V);
8380 if (CE->getOperand(0)->isNullValue()) {
8381 // Insert a new store to null instruction before the load to indicate
8382 // that this code is not reachable. We do this instead of inserting
8383 // an unreachable instruction directly because we cannot modify the
8385 new StoreInst(UndefValue::get(LI.getType()),
8386 Constant::getNullValue(Op->getType()), &LI);
8387 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8390 } else if (CE->isCast()) {
8391 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8396 if (Op->hasOneUse()) {
8397 // Change select and PHI nodes to select values instead of addresses: this
8398 // helps alias analysis out a lot, allows many others simplifications, and
8399 // exposes redundancy in the code.
8401 // Note that we cannot do the transformation unless we know that the
8402 // introduced loads cannot trap! Something like this is valid as long as
8403 // the condition is always false: load (select bool %C, int* null, int* %G),
8404 // but it would not be valid if we transformed it to load from null
8407 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8408 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8409 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8410 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8411 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8412 SI->getOperand(1)->getName()+".val"), LI);
8413 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8414 SI->getOperand(2)->getName()+".val"), LI);
8415 return new SelectInst(SI->getCondition(), V1, V2);
8418 // load (select (cond, null, P)) -> load P
8419 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8420 if (C->isNullValue()) {
8421 LI.setOperand(0, SI->getOperand(2));
8425 // load (select (cond, P, null)) -> load P
8426 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8427 if (C->isNullValue()) {
8428 LI.setOperand(0, SI->getOperand(1));
8436 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8438 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8439 User *CI = cast<User>(SI.getOperand(1));
8440 Value *CastOp = CI->getOperand(0);
8442 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8443 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8444 const Type *SrcPTy = SrcTy->getElementType();
8446 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8447 // If the source is an array, the code below will not succeed. Check to
8448 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8450 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8451 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8452 if (ASrcTy->getNumElements() != 0) {
8454 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8455 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8456 SrcTy = cast<PointerType>(CastOp->getType());
8457 SrcPTy = SrcTy->getElementType();
8460 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8461 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8462 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8464 // Okay, we are casting from one integer or pointer type to another of
8465 // the same size. Instead of casting the pointer before
8466 // the store, cast the value to be stored.
8468 Value *SIOp0 = SI.getOperand(0);
8469 Instruction::CastOps opcode = Instruction::BitCast;
8470 const Type* CastSrcTy = SIOp0->getType();
8471 const Type* CastDstTy = SrcPTy;
8472 if (isa<PointerType>(CastDstTy)) {
8473 if (CastSrcTy->isInteger())
8474 opcode = Instruction::IntToPtr;
8475 } else if (isa<IntegerType>(CastDstTy)) {
8476 if (isa<PointerType>(SIOp0->getType()))
8477 opcode = Instruction::PtrToInt;
8479 if (Constant *C = dyn_cast<Constant>(SIOp0))
8480 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8482 NewCast = IC.InsertNewInstBefore(
8483 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8485 return new StoreInst(NewCast, CastOp);
8492 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8493 Value *Val = SI.getOperand(0);
8494 Value *Ptr = SI.getOperand(1);
8496 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8497 EraseInstFromFunction(SI);
8502 // If the RHS is an alloca with a single use, zapify the store, making the
8504 if (Ptr->hasOneUse()) {
8505 if (isa<AllocaInst>(Ptr)) {
8506 EraseInstFromFunction(SI);
8511 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8512 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8513 GEP->getOperand(0)->hasOneUse()) {
8514 EraseInstFromFunction(SI);
8520 // Do really simple DSE, to catch cases where there are several consequtive
8521 // stores to the same location, separated by a few arithmetic operations. This
8522 // situation often occurs with bitfield accesses.
8523 BasicBlock::iterator BBI = &SI;
8524 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8528 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8529 // Prev store isn't volatile, and stores to the same location?
8530 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8533 EraseInstFromFunction(*PrevSI);
8539 // If this is a load, we have to stop. However, if the loaded value is from
8540 // the pointer we're loading and is producing the pointer we're storing,
8541 // then *this* store is dead (X = load P; store X -> P).
8542 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8543 if (LI == Val && LI->getOperand(0) == Ptr) {
8544 EraseInstFromFunction(SI);
8548 // Otherwise, this is a load from some other location. Stores before it
8553 // Don't skip over loads or things that can modify memory.
8554 if (BBI->mayWriteToMemory())
8559 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8561 // store X, null -> turns into 'unreachable' in SimplifyCFG
8562 if (isa<ConstantPointerNull>(Ptr)) {
8563 if (!isa<UndefValue>(Val)) {
8564 SI.setOperand(0, UndefValue::get(Val->getType()));
8565 if (Instruction *U = dyn_cast<Instruction>(Val))
8566 AddToWorkList(U); // Dropped a use.
8569 return 0; // Do not modify these!
8572 // store undef, Ptr -> noop
8573 if (isa<UndefValue>(Val)) {
8574 EraseInstFromFunction(SI);
8579 // If the pointer destination is a cast, see if we can fold the cast into the
8581 if (isa<CastInst>(Ptr))
8582 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8584 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8586 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8590 // If this store is the last instruction in the basic block, and if the block
8591 // ends with an unconditional branch, try to move it to the successor block.
8593 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8594 if (BI->isUnconditional()) {
8595 // Check to see if the successor block has exactly two incoming edges. If
8596 // so, see if the other predecessor contains a store to the same location.
8597 // if so, insert a PHI node (if needed) and move the stores down.
8598 BasicBlock *Dest = BI->getSuccessor(0);
8600 pred_iterator PI = pred_begin(Dest);
8601 BasicBlock *Other = 0;
8602 if (*PI != BI->getParent())
8605 if (PI != pred_end(Dest)) {
8606 if (*PI != BI->getParent())
8611 if (++PI != pred_end(Dest))
8614 if (Other) { // If only one other pred...
8615 BBI = Other->getTerminator();
8616 // Make sure this other block ends in an unconditional branch and that
8617 // there is an instruction before the branch.
8618 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8619 BBI != Other->begin()) {
8621 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8623 // If this instruction is a store to the same location.
8624 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8625 // Okay, we know we can perform this transformation. Insert a PHI
8626 // node now if we need it.
8627 Value *MergedVal = OtherStore->getOperand(0);
8628 if (MergedVal != SI.getOperand(0)) {
8629 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8630 PN->reserveOperandSpace(2);
8631 PN->addIncoming(SI.getOperand(0), SI.getParent());
8632 PN->addIncoming(OtherStore->getOperand(0), Other);
8633 MergedVal = InsertNewInstBefore(PN, Dest->front());
8636 // Advance to a place where it is safe to insert the new store and
8638 BBI = Dest->begin();
8639 while (isa<PHINode>(BBI)) ++BBI;
8640 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8641 OtherStore->isVolatile()), *BBI);
8643 // Nuke the old stores.
8644 EraseInstFromFunction(SI);
8645 EraseInstFromFunction(*OtherStore);
8657 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8658 // Change br (not X), label True, label False to: br X, label False, True
8660 BasicBlock *TrueDest;
8661 BasicBlock *FalseDest;
8662 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8663 !isa<Constant>(X)) {
8664 // Swap Destinations and condition...
8666 BI.setSuccessor(0, FalseDest);
8667 BI.setSuccessor(1, TrueDest);
8671 // Cannonicalize fcmp_one -> fcmp_oeq
8672 FCmpInst::Predicate FPred; Value *Y;
8673 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8674 TrueDest, FalseDest)))
8675 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8676 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8677 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8678 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8679 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8680 NewSCC->takeName(I);
8681 // Swap Destinations and condition...
8682 BI.setCondition(NewSCC);
8683 BI.setSuccessor(0, FalseDest);
8684 BI.setSuccessor(1, TrueDest);
8685 RemoveFromWorkList(I);
8686 I->eraseFromParent();
8687 AddToWorkList(NewSCC);
8691 // Cannonicalize icmp_ne -> icmp_eq
8692 ICmpInst::Predicate IPred;
8693 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8694 TrueDest, FalseDest)))
8695 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8696 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8697 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8698 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8699 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8700 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8701 NewSCC->takeName(I);
8702 // Swap Destinations and condition...
8703 BI.setCondition(NewSCC);
8704 BI.setSuccessor(0, FalseDest);
8705 BI.setSuccessor(1, TrueDest);
8706 RemoveFromWorkList(I);
8707 I->eraseFromParent();;
8708 AddToWorkList(NewSCC);
8715 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8716 Value *Cond = SI.getCondition();
8717 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8718 if (I->getOpcode() == Instruction::Add)
8719 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8720 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8721 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8722 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8724 SI.setOperand(0, I->getOperand(0));
8732 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8733 /// is to leave as a vector operation.
8734 static bool CheapToScalarize(Value *V, bool isConstant) {
8735 if (isa<ConstantAggregateZero>(V))
8737 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8738 if (isConstant) return true;
8739 // If all elts are the same, we can extract.
8740 Constant *Op0 = C->getOperand(0);
8741 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8742 if (C->getOperand(i) != Op0)
8746 Instruction *I = dyn_cast<Instruction>(V);
8747 if (!I) return false;
8749 // Insert element gets simplified to the inserted element or is deleted if
8750 // this is constant idx extract element and its a constant idx insertelt.
8751 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8752 isa<ConstantInt>(I->getOperand(2)))
8754 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8756 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8757 if (BO->hasOneUse() &&
8758 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8759 CheapToScalarize(BO->getOperand(1), isConstant)))
8761 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8762 if (CI->hasOneUse() &&
8763 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8764 CheapToScalarize(CI->getOperand(1), isConstant)))
8770 /// Read and decode a shufflevector mask.
8772 /// It turns undef elements into values that are larger than the number of
8773 /// elements in the input.
8774 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8775 unsigned NElts = SVI->getType()->getNumElements();
8776 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8777 return std::vector<unsigned>(NElts, 0);
8778 if (isa<UndefValue>(SVI->getOperand(2)))
8779 return std::vector<unsigned>(NElts, 2*NElts);
8781 std::vector<unsigned> Result;
8782 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8783 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8784 if (isa<UndefValue>(CP->getOperand(i)))
8785 Result.push_back(NElts*2); // undef -> 8
8787 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8791 /// FindScalarElement - Given a vector and an element number, see if the scalar
8792 /// value is already around as a register, for example if it were inserted then
8793 /// extracted from the vector.
8794 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8795 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8796 const VectorType *PTy = cast<VectorType>(V->getType());
8797 unsigned Width = PTy->getNumElements();
8798 if (EltNo >= Width) // Out of range access.
8799 return UndefValue::get(PTy->getElementType());
8801 if (isa<UndefValue>(V))
8802 return UndefValue::get(PTy->getElementType());
8803 else if (isa<ConstantAggregateZero>(V))
8804 return Constant::getNullValue(PTy->getElementType());
8805 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8806 return CP->getOperand(EltNo);
8807 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8808 // If this is an insert to a variable element, we don't know what it is.
8809 if (!isa<ConstantInt>(III->getOperand(2)))
8811 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8813 // If this is an insert to the element we are looking for, return the
8816 return III->getOperand(1);
8818 // Otherwise, the insertelement doesn't modify the value, recurse on its
8820 return FindScalarElement(III->getOperand(0), EltNo);
8821 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8822 unsigned InEl = getShuffleMask(SVI)[EltNo];
8824 return FindScalarElement(SVI->getOperand(0), InEl);
8825 else if (InEl < Width*2)
8826 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8828 return UndefValue::get(PTy->getElementType());
8831 // Otherwise, we don't know.
8835 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8837 // If packed val is undef, replace extract with scalar undef.
8838 if (isa<UndefValue>(EI.getOperand(0)))
8839 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8841 // If packed val is constant 0, replace extract with scalar 0.
8842 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8843 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8845 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8846 // If packed val is constant with uniform operands, replace EI
8847 // with that operand
8848 Constant *op0 = C->getOperand(0);
8849 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8850 if (C->getOperand(i) != op0) {
8855 return ReplaceInstUsesWith(EI, op0);
8858 // If extracting a specified index from the vector, see if we can recursively
8859 // find a previously computed scalar that was inserted into the vector.
8860 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8861 // This instruction only demands the single element from the input vector.
8862 // If the input vector has a single use, simplify it based on this use
8864 uint64_t IndexVal = IdxC->getZExtValue();
8865 if (EI.getOperand(0)->hasOneUse()) {
8867 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8870 EI.setOperand(0, V);
8875 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8876 return ReplaceInstUsesWith(EI, Elt);
8879 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8880 if (I->hasOneUse()) {
8881 // Push extractelement into predecessor operation if legal and
8882 // profitable to do so
8883 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8884 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8885 if (CheapToScalarize(BO, isConstantElt)) {
8886 ExtractElementInst *newEI0 =
8887 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8888 EI.getName()+".lhs");
8889 ExtractElementInst *newEI1 =
8890 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8891 EI.getName()+".rhs");
8892 InsertNewInstBefore(newEI0, EI);
8893 InsertNewInstBefore(newEI1, EI);
8894 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8896 } else if (isa<LoadInst>(I)) {
8897 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8898 PointerType::get(EI.getType()), EI);
8899 GetElementPtrInst *GEP =
8900 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8901 InsertNewInstBefore(GEP, EI);
8902 return new LoadInst(GEP);
8905 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8906 // Extracting the inserted element?
8907 if (IE->getOperand(2) == EI.getOperand(1))
8908 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8909 // If the inserted and extracted elements are constants, they must not
8910 // be the same value, extract from the pre-inserted value instead.
8911 if (isa<Constant>(IE->getOperand(2)) &&
8912 isa<Constant>(EI.getOperand(1))) {
8913 AddUsesToWorkList(EI);
8914 EI.setOperand(0, IE->getOperand(0));
8917 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8918 // If this is extracting an element from a shufflevector, figure out where
8919 // it came from and extract from the appropriate input element instead.
8920 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8921 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8923 if (SrcIdx < SVI->getType()->getNumElements())
8924 Src = SVI->getOperand(0);
8925 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8926 SrcIdx -= SVI->getType()->getNumElements();
8927 Src = SVI->getOperand(1);
8929 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8931 return new ExtractElementInst(Src, SrcIdx);
8938 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8939 /// elements from either LHS or RHS, return the shuffle mask and true.
8940 /// Otherwise, return false.
8941 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8942 std::vector<Constant*> &Mask) {
8943 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8944 "Invalid CollectSingleShuffleElements");
8945 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8947 if (isa<UndefValue>(V)) {
8948 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8950 } else if (V == LHS) {
8951 for (unsigned i = 0; i != NumElts; ++i)
8952 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8954 } else if (V == RHS) {
8955 for (unsigned i = 0; i != NumElts; ++i)
8956 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8958 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8959 // If this is an insert of an extract from some other vector, include it.
8960 Value *VecOp = IEI->getOperand(0);
8961 Value *ScalarOp = IEI->getOperand(1);
8962 Value *IdxOp = IEI->getOperand(2);
8964 if (!isa<ConstantInt>(IdxOp))
8966 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8968 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8969 // Okay, we can handle this if the vector we are insertinting into is
8971 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8972 // If so, update the mask to reflect the inserted undef.
8973 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8976 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8977 if (isa<ConstantInt>(EI->getOperand(1)) &&
8978 EI->getOperand(0)->getType() == V->getType()) {
8979 unsigned ExtractedIdx =
8980 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8982 // This must be extracting from either LHS or RHS.
8983 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8984 // Okay, we can handle this if the vector we are insertinting into is
8986 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8987 // If so, update the mask to reflect the inserted value.
8988 if (EI->getOperand(0) == LHS) {
8989 Mask[InsertedIdx & (NumElts-1)] =
8990 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8992 assert(EI->getOperand(0) == RHS);
8993 Mask[InsertedIdx & (NumElts-1)] =
8994 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9003 // TODO: Handle shufflevector here!
9008 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9009 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9010 /// that computes V and the LHS value of the shuffle.
9011 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9013 assert(isa<VectorType>(V->getType()) &&
9014 (RHS == 0 || V->getType() == RHS->getType()) &&
9015 "Invalid shuffle!");
9016 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9018 if (isa<UndefValue>(V)) {
9019 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9021 } else if (isa<ConstantAggregateZero>(V)) {
9022 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9024 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9025 // If this is an insert of an extract from some other vector, include it.
9026 Value *VecOp = IEI->getOperand(0);
9027 Value *ScalarOp = IEI->getOperand(1);
9028 Value *IdxOp = IEI->getOperand(2);
9030 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9031 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9032 EI->getOperand(0)->getType() == V->getType()) {
9033 unsigned ExtractedIdx =
9034 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9035 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9037 // Either the extracted from or inserted into vector must be RHSVec,
9038 // otherwise we'd end up with a shuffle of three inputs.
9039 if (EI->getOperand(0) == RHS || RHS == 0) {
9040 RHS = EI->getOperand(0);
9041 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9042 Mask[InsertedIdx & (NumElts-1)] =
9043 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9048 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9049 // Everything but the extracted element is replaced with the RHS.
9050 for (unsigned i = 0; i != NumElts; ++i) {
9051 if (i != InsertedIdx)
9052 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9057 // If this insertelement is a chain that comes from exactly these two
9058 // vectors, return the vector and the effective shuffle.
9059 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9060 return EI->getOperand(0);
9065 // TODO: Handle shufflevector here!
9067 // Otherwise, can't do anything fancy. Return an identity vector.
9068 for (unsigned i = 0; i != NumElts; ++i)
9069 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9073 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9074 Value *VecOp = IE.getOperand(0);
9075 Value *ScalarOp = IE.getOperand(1);
9076 Value *IdxOp = IE.getOperand(2);
9078 // If the inserted element was extracted from some other vector, and if the
9079 // indexes are constant, try to turn this into a shufflevector operation.
9080 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9081 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9082 EI->getOperand(0)->getType() == IE.getType()) {
9083 unsigned NumVectorElts = IE.getType()->getNumElements();
9084 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9085 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9087 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9088 return ReplaceInstUsesWith(IE, VecOp);
9090 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9091 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9093 // If we are extracting a value from a vector, then inserting it right
9094 // back into the same place, just use the input vector.
9095 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9096 return ReplaceInstUsesWith(IE, VecOp);
9098 // We could theoretically do this for ANY input. However, doing so could
9099 // turn chains of insertelement instructions into a chain of shufflevector
9100 // instructions, and right now we do not merge shufflevectors. As such,
9101 // only do this in a situation where it is clear that there is benefit.
9102 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9103 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9104 // the values of VecOp, except then one read from EIOp0.
9105 // Build a new shuffle mask.
9106 std::vector<Constant*> Mask;
9107 if (isa<UndefValue>(VecOp))
9108 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9110 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9111 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9114 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9115 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9116 ConstantVector::get(Mask));
9119 // If this insertelement isn't used by some other insertelement, turn it
9120 // (and any insertelements it points to), into one big shuffle.
9121 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9122 std::vector<Constant*> Mask;
9124 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9125 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9126 // We now have a shuffle of LHS, RHS, Mask.
9127 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9136 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9137 Value *LHS = SVI.getOperand(0);
9138 Value *RHS = SVI.getOperand(1);
9139 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9141 bool MadeChange = false;
9143 // Undefined shuffle mask -> undefined value.
9144 if (isa<UndefValue>(SVI.getOperand(2)))
9145 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9147 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9148 // the undef, change them to undefs.
9149 if (isa<UndefValue>(SVI.getOperand(1))) {
9150 // Scan to see if there are any references to the RHS. If so, replace them
9151 // with undef element refs and set MadeChange to true.
9152 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9153 if (Mask[i] >= e && Mask[i] != 2*e) {
9160 // Remap any references to RHS to use LHS.
9161 std::vector<Constant*> Elts;
9162 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9164 Elts.push_back(UndefValue::get(Type::Int32Ty));
9166 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9168 SVI.setOperand(2, ConstantVector::get(Elts));
9172 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9173 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9174 if (LHS == RHS || isa<UndefValue>(LHS)) {
9175 if (isa<UndefValue>(LHS) && LHS == RHS) {
9176 // shuffle(undef,undef,mask) -> undef.
9177 return ReplaceInstUsesWith(SVI, LHS);
9180 // Remap any references to RHS to use LHS.
9181 std::vector<Constant*> Elts;
9182 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9184 Elts.push_back(UndefValue::get(Type::Int32Ty));
9186 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9187 (Mask[i] < e && isa<UndefValue>(LHS)))
9188 Mask[i] = 2*e; // Turn into undef.
9190 Mask[i] &= (e-1); // Force to LHS.
9191 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9194 SVI.setOperand(0, SVI.getOperand(1));
9195 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9196 SVI.setOperand(2, ConstantVector::get(Elts));
9197 LHS = SVI.getOperand(0);
9198 RHS = SVI.getOperand(1);
9202 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9203 bool isLHSID = true, isRHSID = true;
9205 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9206 if (Mask[i] >= e*2) continue; // Ignore undef values.
9207 // Is this an identity shuffle of the LHS value?
9208 isLHSID &= (Mask[i] == i);
9210 // Is this an identity shuffle of the RHS value?
9211 isRHSID &= (Mask[i]-e == i);
9214 // Eliminate identity shuffles.
9215 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9216 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9218 // If the LHS is a shufflevector itself, see if we can combine it with this
9219 // one without producing an unusual shuffle. Here we are really conservative:
9220 // we are absolutely afraid of producing a shuffle mask not in the input
9221 // program, because the code gen may not be smart enough to turn a merged
9222 // shuffle into two specific shuffles: it may produce worse code. As such,
9223 // we only merge two shuffles if the result is one of the two input shuffle
9224 // masks. In this case, merging the shuffles just removes one instruction,
9225 // which we know is safe. This is good for things like turning:
9226 // (splat(splat)) -> splat.
9227 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9228 if (isa<UndefValue>(RHS)) {
9229 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9231 std::vector<unsigned> NewMask;
9232 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9234 NewMask.push_back(2*e);
9236 NewMask.push_back(LHSMask[Mask[i]]);
9238 // If the result mask is equal to the src shuffle or this shuffle mask, do
9240 if (NewMask == LHSMask || NewMask == Mask) {
9241 std::vector<Constant*> Elts;
9242 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9243 if (NewMask[i] >= e*2) {
9244 Elts.push_back(UndefValue::get(Type::Int32Ty));
9246 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9249 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9250 LHSSVI->getOperand(1),
9251 ConstantVector::get(Elts));
9256 return MadeChange ? &SVI : 0;
9262 /// TryToSinkInstruction - Try to move the specified instruction from its
9263 /// current block into the beginning of DestBlock, which can only happen if it's
9264 /// safe to move the instruction past all of the instructions between it and the
9265 /// end of its block.
9266 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9267 assert(I->hasOneUse() && "Invariants didn't hold!");
9269 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9270 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9272 // Do not sink alloca instructions out of the entry block.
9273 if (isa<AllocaInst>(I) && I->getParent() ==
9274 &DestBlock->getParent()->getEntryBlock())
9277 // We can only sink load instructions if there is nothing between the load and
9278 // the end of block that could change the value.
9279 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9280 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9282 if (Scan->mayWriteToMemory())
9286 BasicBlock::iterator InsertPos = DestBlock->begin();
9287 while (isa<PHINode>(InsertPos)) ++InsertPos;
9289 I->moveBefore(InsertPos);
9295 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9296 /// all reachable code to the worklist.
9298 /// This has a couple of tricks to make the code faster and more powerful. In
9299 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9300 /// them to the worklist (this significantly speeds up instcombine on code where
9301 /// many instructions are dead or constant). Additionally, if we find a branch
9302 /// whose condition is a known constant, we only visit the reachable successors.
9304 static void AddReachableCodeToWorklist(BasicBlock *BB,
9305 SmallPtrSet<BasicBlock*, 64> &Visited,
9307 const TargetData *TD) {
9308 std::vector<BasicBlock*> Worklist;
9309 Worklist.push_back(BB);
9311 while (!Worklist.empty()) {
9312 BB = Worklist.back();
9313 Worklist.pop_back();
9315 // We have now visited this block! If we've already been here, ignore it.
9316 if (!Visited.insert(BB)) continue;
9318 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9319 Instruction *Inst = BBI++;
9321 // DCE instruction if trivially dead.
9322 if (isInstructionTriviallyDead(Inst)) {
9324 DOUT << "IC: DCE: " << *Inst;
9325 Inst->eraseFromParent();
9329 // ConstantProp instruction if trivially constant.
9330 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9331 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9332 Inst->replaceAllUsesWith(C);
9334 Inst->eraseFromParent();
9338 IC.AddToWorkList(Inst);
9341 // Recursively visit successors. If this is a branch or switch on a
9342 // constant, only visit the reachable successor.
9343 TerminatorInst *TI = BB->getTerminator();
9344 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9345 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9346 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9347 Worklist.push_back(BI->getSuccessor(!CondVal));
9350 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9351 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9352 // See if this is an explicit destination.
9353 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9354 if (SI->getCaseValue(i) == Cond) {
9355 Worklist.push_back(SI->getSuccessor(i));
9359 // Otherwise it is the default destination.
9360 Worklist.push_back(SI->getSuccessor(0));
9365 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9366 Worklist.push_back(TI->getSuccessor(i));
9370 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9371 bool Changed = false;
9372 TD = &getAnalysis<TargetData>();
9374 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9375 << F.getNameStr() << "\n");
9378 // Do a depth-first traversal of the function, populate the worklist with
9379 // the reachable instructions. Ignore blocks that are not reachable. Keep
9380 // track of which blocks we visit.
9381 SmallPtrSet<BasicBlock*, 64> Visited;
9382 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9384 // Do a quick scan over the function. If we find any blocks that are
9385 // unreachable, remove any instructions inside of them. This prevents
9386 // the instcombine code from having to deal with some bad special cases.
9387 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9388 if (!Visited.count(BB)) {
9389 Instruction *Term = BB->getTerminator();
9390 while (Term != BB->begin()) { // Remove instrs bottom-up
9391 BasicBlock::iterator I = Term; --I;
9393 DOUT << "IC: DCE: " << *I;
9396 if (!I->use_empty())
9397 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9398 I->eraseFromParent();
9403 while (!Worklist.empty()) {
9404 Instruction *I = RemoveOneFromWorkList();
9405 if (I == 0) continue; // skip null values.
9407 // Check to see if we can DCE the instruction.
9408 if (isInstructionTriviallyDead(I)) {
9409 // Add operands to the worklist.
9410 if (I->getNumOperands() < 4)
9411 AddUsesToWorkList(*I);
9414 DOUT << "IC: DCE: " << *I;
9416 I->eraseFromParent();
9417 RemoveFromWorkList(I);
9421 // Instruction isn't dead, see if we can constant propagate it.
9422 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9423 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9425 // Add operands to the worklist.
9426 AddUsesToWorkList(*I);
9427 ReplaceInstUsesWith(*I, C);
9430 I->eraseFromParent();
9431 RemoveFromWorkList(I);
9435 // See if we can trivially sink this instruction to a successor basic block.
9436 if (I->hasOneUse()) {
9437 BasicBlock *BB = I->getParent();
9438 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9439 if (UserParent != BB) {
9440 bool UserIsSuccessor = false;
9441 // See if the user is one of our successors.
9442 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9443 if (*SI == UserParent) {
9444 UserIsSuccessor = true;
9448 // If the user is one of our immediate successors, and if that successor
9449 // only has us as a predecessors (we'd have to split the critical edge
9450 // otherwise), we can keep going.
9451 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9452 next(pred_begin(UserParent)) == pred_end(UserParent))
9453 // Okay, the CFG is simple enough, try to sink this instruction.
9454 Changed |= TryToSinkInstruction(I, UserParent);
9458 // Now that we have an instruction, try combining it to simplify it...
9462 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
9463 if (Instruction *Result = visit(*I)) {
9465 // Should we replace the old instruction with a new one?
9467 DOUT << "IC: Old = " << *I
9468 << " New = " << *Result;
9470 // Everything uses the new instruction now.
9471 I->replaceAllUsesWith(Result);
9473 // Push the new instruction and any users onto the worklist.
9474 AddToWorkList(Result);
9475 AddUsersToWorkList(*Result);
9477 // Move the name to the new instruction first.
9478 Result->takeName(I);
9480 // Insert the new instruction into the basic block...
9481 BasicBlock *InstParent = I->getParent();
9482 BasicBlock::iterator InsertPos = I;
9484 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9485 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9488 InstParent->getInstList().insert(InsertPos, Result);
9490 // Make sure that we reprocess all operands now that we reduced their
9492 AddUsesToWorkList(*I);
9494 // Instructions can end up on the worklist more than once. Make sure
9495 // we do not process an instruction that has been deleted.
9496 RemoveFromWorkList(I);
9498 // Erase the old instruction.
9499 InstParent->getInstList().erase(I);
9502 DOUT << "IC: Mod = " << OrigI
9506 // If the instruction was modified, it's possible that it is now dead.
9507 // if so, remove it.
9508 if (isInstructionTriviallyDead(I)) {
9509 // Make sure we process all operands now that we are reducing their
9511 AddUsesToWorkList(*I);
9513 // Instructions may end up in the worklist more than once. Erase all
9514 // occurrences of this instruction.
9515 RemoveFromWorkList(I);
9516 I->eraseFromParent();
9519 AddUsersToWorkList(*I);
9526 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9531 bool InstCombiner::runOnFunction(Function &F) {
9532 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
9534 bool EverMadeChange = false;
9536 // Iterate while there is work to do.
9537 unsigned Iteration = 0;
9538 while (DoOneIteration(F, Iteration++))
9539 EverMadeChange = true;
9540 return EverMadeChange;
9543 FunctionPass *llvm::createInstructionCombiningPass() {
9544 return new InstCombiner();