1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic 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/ConstantRange.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
190 Instruction *visitFCmpInst(FCmpInst &I);
191 Instruction *visitICmpInst(ICmpInst &I);
192 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
193 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
196 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
197 ConstantInt *DivRHS);
199 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
200 ICmpInst::Predicate Cond, Instruction &I);
201 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
203 Instruction *commonCastTransforms(CastInst &CI);
204 Instruction *commonIntCastTransforms(CastInst &CI);
205 Instruction *commonPointerCastTransforms(CastInst &CI);
206 Instruction *visitTrunc(TruncInst &CI);
207 Instruction *visitZExt(ZExtInst &CI);
208 Instruction *visitSExt(SExtInst &CI);
209 Instruction *visitFPTrunc(FPTruncInst &CI);
210 Instruction *visitFPExt(CastInst &CI);
211 Instruction *visitFPToUI(FPToUIInst &FI);
212 Instruction *visitFPToSI(FPToSIInst &FI);
213 Instruction *visitUIToFP(CastInst &CI);
214 Instruction *visitSIToFP(CastInst &CI);
215 Instruction *visitPtrToInt(CastInst &CI);
216 Instruction *visitIntToPtr(IntToPtrInst &CI);
217 Instruction *visitBitCast(BitCastInst &CI);
218 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
220 Instruction *visitSelectInst(SelectInst &CI);
221 Instruction *visitCallInst(CallInst &CI);
222 Instruction *visitInvokeInst(InvokeInst &II);
223 Instruction *visitPHINode(PHINode &PN);
224 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
225 Instruction *visitAllocationInst(AllocationInst &AI);
226 Instruction *visitFreeInst(FreeInst &FI);
227 Instruction *visitLoadInst(LoadInst &LI);
228 Instruction *visitStoreInst(StoreInst &SI);
229 Instruction *visitBranchInst(BranchInst &BI);
230 Instruction *visitSwitchInst(SwitchInst &SI);
231 Instruction *visitInsertElementInst(InsertElementInst &IE);
232 Instruction *visitExtractElementInst(ExtractElementInst &EI);
233 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
235 // visitInstruction - Specify what to return for unhandled instructions...
236 Instruction *visitInstruction(Instruction &I) { return 0; }
239 Instruction *visitCallSite(CallSite CS);
240 bool transformConstExprCastCall(CallSite CS);
241 Instruction *transformCallThroughTrampoline(CallSite CS);
242 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
243 bool DoXform = true);
244 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
247 // InsertNewInstBefore - insert an instruction New before instruction Old
248 // in the program. Add the new instruction to the worklist.
250 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
251 assert(New && New->getParent() == 0 &&
252 "New instruction already inserted into a basic block!");
253 BasicBlock *BB = Old.getParent();
254 BB->getInstList().insert(&Old, New); // Insert inst
259 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
260 /// This also adds the cast to the worklist. Finally, this returns the
262 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
264 if (V->getType() == Ty) return V;
266 if (Constant *CV = dyn_cast<Constant>(V))
267 return ConstantExpr::getCast(opc, CV, Ty);
269 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
274 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
275 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
279 // ReplaceInstUsesWith - This method is to be used when an instruction is
280 // found to be dead, replacable with another preexisting expression. Here
281 // we add all uses of I to the worklist, replace all uses of I with the new
282 // value, then return I, so that the inst combiner will know that I was
285 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
286 AddUsersToWorkList(I); // Add all modified instrs to worklist
288 I.replaceAllUsesWith(V);
291 // If we are replacing the instruction with itself, this must be in a
292 // segment of unreachable code, so just clobber the instruction.
293 I.replaceAllUsesWith(UndefValue::get(I.getType()));
298 // UpdateValueUsesWith - This method is to be used when an value is
299 // found to be replacable with another preexisting expression or was
300 // updated. Here we add all uses of I to the worklist, replace all uses of
301 // I with the new value (unless the instruction was just updated), then
302 // return true, so that the inst combiner will know that I was modified.
304 bool UpdateValueUsesWith(Value *Old, Value *New) {
305 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
307 Old->replaceAllUsesWith(New);
308 if (Instruction *I = dyn_cast<Instruction>(Old))
310 if (Instruction *I = dyn_cast<Instruction>(New))
315 // EraseInstFromFunction - When dealing with an instruction that has side
316 // effects or produces a void value, we can't rely on DCE to delete the
317 // instruction. Instead, visit methods should return the value returned by
319 Instruction *EraseInstFromFunction(Instruction &I) {
320 assert(I.use_empty() && "Cannot erase instruction that is used!");
321 AddUsesToWorkList(I);
322 RemoveFromWorkList(&I);
324 return 0; // Don't do anything with FI
328 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
329 /// InsertBefore instruction. This is specialized a bit to avoid inserting
330 /// casts that are known to not do anything...
332 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
333 Value *V, const Type *DestTy,
334 Instruction *InsertBefore);
336 /// SimplifyCommutative - This performs a few simplifications for
337 /// commutative operators.
338 bool SimplifyCommutative(BinaryOperator &I);
340 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
341 /// most-complex to least-complex order.
342 bool SimplifyCompare(CmpInst &I);
344 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
345 /// on the demanded bits.
346 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
347 APInt& KnownZero, APInt& KnownOne,
350 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
351 uint64_t &UndefElts, unsigned Depth = 0);
353 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
354 // PHI node as operand #0, see if we can fold the instruction into the PHI
355 // (which is only possible if all operands to the PHI are constants).
356 Instruction *FoldOpIntoPhi(Instruction &I);
358 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
359 // operator and they all are only used by the PHI, PHI together their
360 // inputs, and do the operation once, to the result of the PHI.
361 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
362 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
365 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
366 ConstantInt *AndRHS, BinaryOperator &TheAnd);
368 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
369 bool isSub, Instruction &I);
370 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
371 bool isSigned, bool Inside, Instruction &IB);
372 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
373 Instruction *MatchBSwap(BinaryOperator &I);
374 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
375 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
376 Instruction *SimplifyMemSet(MemSetInst *MI);
379 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
381 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
382 APInt& KnownOne, unsigned Depth = 0) const;
383 bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0);
384 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const;
385 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
387 int &NumCastsRemoved);
388 unsigned GetOrEnforceKnownAlignment(Value *V,
389 unsigned PrefAlign = 0);
393 char InstCombiner::ID = 0;
394 static RegisterPass<InstCombiner>
395 X("instcombine", "Combine redundant instructions");
397 // getComplexity: Assign a complexity or rank value to LLVM Values...
398 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
399 static unsigned getComplexity(Value *V) {
400 if (isa<Instruction>(V)) {
401 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
405 if (isa<Argument>(V)) return 3;
406 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
409 // isOnlyUse - Return true if this instruction will be deleted if we stop using
411 static bool isOnlyUse(Value *V) {
412 return V->hasOneUse() || isa<Constant>(V);
415 // getPromotedType - Return the specified type promoted as it would be to pass
416 // though a va_arg area...
417 static const Type *getPromotedType(const Type *Ty) {
418 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
419 if (ITy->getBitWidth() < 32)
420 return Type::Int32Ty;
425 /// getBitCastOperand - If the specified operand is a CastInst or a constant
426 /// expression bitcast, return the operand value, otherwise return null.
427 static Value *getBitCastOperand(Value *V) {
428 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
429 return I->getOperand(0);
430 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
431 if (CE->getOpcode() == Instruction::BitCast)
432 return CE->getOperand(0);
436 /// This function is a wrapper around CastInst::isEliminableCastPair. It
437 /// simply extracts arguments and returns what that function returns.
438 static Instruction::CastOps
439 isEliminableCastPair(
440 const CastInst *CI, ///< The first cast instruction
441 unsigned opcode, ///< The opcode of the second cast instruction
442 const Type *DstTy, ///< The target type for the second cast instruction
443 TargetData *TD ///< The target data for pointer size
446 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
447 const Type *MidTy = CI->getType(); // B from above
449 // Get the opcodes of the two Cast instructions
450 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
451 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
453 return Instruction::CastOps(
454 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
455 DstTy, TD->getIntPtrType()));
458 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
459 /// in any code being generated. It does not require codegen if V is simple
460 /// enough or if the cast can be folded into other casts.
461 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
462 const Type *Ty, TargetData *TD) {
463 if (V->getType() == Ty || isa<Constant>(V)) return false;
465 // If this is another cast that can be eliminated, it isn't codegen either.
466 if (const CastInst *CI = dyn_cast<CastInst>(V))
467 if (isEliminableCastPair(CI, opcode, Ty, TD))
472 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
473 /// InsertBefore instruction. This is specialized a bit to avoid inserting
474 /// casts that are known to not do anything...
476 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
477 Value *V, const Type *DestTy,
478 Instruction *InsertBefore) {
479 if (V->getType() == DestTy) return V;
480 if (Constant *C = dyn_cast<Constant>(V))
481 return ConstantExpr::getCast(opcode, C, DestTy);
483 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
486 // SimplifyCommutative - This performs a few simplifications for commutative
489 // 1. Order operands such that they are listed from right (least complex) to
490 // left (most complex). This puts constants before unary operators before
493 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
494 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
496 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
497 bool Changed = false;
498 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
499 Changed = !I.swapOperands();
501 if (!I.isAssociative()) return Changed;
502 Instruction::BinaryOps Opcode = I.getOpcode();
503 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
504 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
505 if (isa<Constant>(I.getOperand(1))) {
506 Constant *Folded = ConstantExpr::get(I.getOpcode(),
507 cast<Constant>(I.getOperand(1)),
508 cast<Constant>(Op->getOperand(1)));
509 I.setOperand(0, Op->getOperand(0));
510 I.setOperand(1, Folded);
512 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
513 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
514 isOnlyUse(Op) && isOnlyUse(Op1)) {
515 Constant *C1 = cast<Constant>(Op->getOperand(1));
516 Constant *C2 = cast<Constant>(Op1->getOperand(1));
518 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
519 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
520 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
524 I.setOperand(0, New);
525 I.setOperand(1, Folded);
532 /// SimplifyCompare - For a CmpInst this function just orders the operands
533 /// so that theyare listed from right (least complex) to left (most complex).
534 /// This puts constants before unary operators before binary operators.
535 bool InstCombiner::SimplifyCompare(CmpInst &I) {
536 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
539 // Compare instructions are not associative so there's nothing else we can do.
543 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
544 // if the LHS is a constant zero (which is the 'negate' form).
546 static inline Value *dyn_castNegVal(Value *V) {
547 if (BinaryOperator::isNeg(V))
548 return BinaryOperator::getNegArgument(V);
550 // Constants can be considered to be negated values if they can be folded.
551 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
552 return ConstantExpr::getNeg(C);
556 static inline Value *dyn_castNotVal(Value *V) {
557 if (BinaryOperator::isNot(V))
558 return BinaryOperator::getNotArgument(V);
560 // Constants can be considered to be not'ed values...
561 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
562 return ConstantInt::get(~C->getValue());
566 // dyn_castFoldableMul - If this value is a multiply that can be folded into
567 // other computations (because it has a constant operand), return the
568 // non-constant operand of the multiply, and set CST to point to the multiplier.
569 // Otherwise, return null.
571 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
572 if (V->hasOneUse() && V->getType()->isInteger())
573 if (Instruction *I = dyn_cast<Instruction>(V)) {
574 if (I->getOpcode() == Instruction::Mul)
575 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
576 return I->getOperand(0);
577 if (I->getOpcode() == Instruction::Shl)
578 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
579 // The multiplier is really 1 << CST.
580 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
581 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
582 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
583 return I->getOperand(0);
589 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
590 /// expression, return it.
591 static User *dyn_castGetElementPtr(Value *V) {
592 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
593 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
594 if (CE->getOpcode() == Instruction::GetElementPtr)
595 return cast<User>(V);
599 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
600 /// opcode value. Otherwise return UserOp1.
601 static unsigned getOpcode(Value *V) {
602 if (Instruction *I = dyn_cast<Instruction>(V))
603 return I->getOpcode();
604 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
605 return CE->getOpcode();
606 // Use UserOp1 to mean there's no opcode.
607 return Instruction::UserOp1;
610 /// AddOne - Add one to a ConstantInt
611 static ConstantInt *AddOne(ConstantInt *C) {
612 APInt Val(C->getValue());
613 return ConstantInt::get(++Val);
615 /// SubOne - Subtract one from a ConstantInt
616 static ConstantInt *SubOne(ConstantInt *C) {
617 APInt Val(C->getValue());
618 return ConstantInt::get(--Val);
620 /// Add - Add two ConstantInts together
621 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
622 return ConstantInt::get(C1->getValue() + C2->getValue());
624 /// And - Bitwise AND two ConstantInts together
625 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
626 return ConstantInt::get(C1->getValue() & C2->getValue());
628 /// Subtract - Subtract one ConstantInt from another
629 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
630 return ConstantInt::get(C1->getValue() - C2->getValue());
632 /// Multiply - Multiply two ConstantInts together
633 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
634 return ConstantInt::get(C1->getValue() * C2->getValue());
636 /// MultiplyOverflows - True if the multiply can not be expressed in an int
638 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
639 uint32_t W = C1->getBitWidth();
640 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
649 APInt MulExt = LHSExt * RHSExt;
652 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
653 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
654 return MulExt.slt(Min) || MulExt.sgt(Max);
656 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
659 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
660 /// known to be either zero or one and return them in the KnownZero/KnownOne
661 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
663 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
664 /// we cannot optimize based on the assumption that it is zero without changing
665 /// it to be an explicit zero. If we don't change it to zero, other code could
666 /// optimized based on the contradictory assumption that it is non-zero.
667 /// Because instcombine aggressively folds operations with undef args anyway,
668 /// this won't lose us code quality.
669 void InstCombiner::ComputeMaskedBits(Value *V, const APInt &Mask,
670 APInt& KnownZero, APInt& KnownOne,
671 unsigned Depth) const {
672 assert(V && "No Value?");
673 assert(Depth <= 6 && "Limit Search Depth");
674 uint32_t BitWidth = Mask.getBitWidth();
675 assert((V->getType()->isInteger() || isa<PointerType>(V->getType())) &&
676 "Not integer or pointer type!");
677 assert((!TD || TD->getTypeSizeInBits(V->getType()) == BitWidth) &&
678 (!isa<IntegerType>(V->getType()) ||
679 V->getType()->getPrimitiveSizeInBits() == BitWidth) &&
680 KnownZero.getBitWidth() == BitWidth &&
681 KnownOne.getBitWidth() == BitWidth &&
682 "V, Mask, KnownOne and KnownZero should have same BitWidth");
684 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
685 // We know all of the bits for a constant!
686 KnownOne = CI->getValue() & Mask;
687 KnownZero = ~KnownOne & Mask;
690 // Null is all-zeros.
691 if (isa<ConstantPointerNull>(V)) {
696 // The address of an aligned GlobalValue has trailing zeros.
697 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
698 unsigned Align = GV->getAlignment();
699 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
700 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
702 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
703 CountTrailingZeros_32(Align));
710 KnownZero.clear(); KnownOne.clear(); // Start out not knowing anything.
712 if (Depth == 6 || Mask == 0)
713 return; // Limit search depth.
715 User *I = dyn_cast<User>(V);
718 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
719 switch (getOpcode(I)) {
721 case Instruction::And: {
722 // If either the LHS or the RHS are Zero, the result is zero.
723 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
724 APInt Mask2(Mask & ~KnownZero);
725 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
726 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
727 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
729 // Output known-1 bits are only known if set in both the LHS & RHS.
730 KnownOne &= KnownOne2;
731 // Output known-0 are known to be clear if zero in either the LHS | RHS.
732 KnownZero |= KnownZero2;
735 case Instruction::Or: {
736 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
737 APInt Mask2(Mask & ~KnownOne);
738 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
739 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
740 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
742 // Output known-0 bits are only known if clear in both the LHS & RHS.
743 KnownZero &= KnownZero2;
744 // Output known-1 are known to be set if set in either the LHS | RHS.
745 KnownOne |= KnownOne2;
748 case Instruction::Xor: {
749 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
750 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
751 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
752 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
754 // Output known-0 bits are known if clear or set in both the LHS & RHS.
755 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
756 // Output known-1 are known to be set if set in only one of the LHS, RHS.
757 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
758 KnownZero = KnownZeroOut;
761 case Instruction::Mul: {
762 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
763 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
764 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
765 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
766 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
768 // If low bits are zero in either operand, output low known-0 bits.
769 // Also compute a conserative estimate for high known-0 bits.
770 // More trickiness is possible, but this is sufficient for the
771 // interesting case of alignment computation.
773 unsigned TrailZ = KnownZero.countTrailingOnes() +
774 KnownZero2.countTrailingOnes();
775 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
776 KnownZero2.countLeadingOnes(),
777 BitWidth) - BitWidth;
779 TrailZ = std::min(TrailZ, BitWidth);
780 LeadZ = std::min(LeadZ, BitWidth);
781 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
782 APInt::getHighBitsSet(BitWidth, LeadZ);
786 case Instruction::UDiv: {
787 // For the purposes of computing leading zeros we can conservatively
788 // treat a udiv as a logical right shift by the power of 2 known to
789 // be less than the denominator.
790 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
791 ComputeMaskedBits(I->getOperand(0),
792 AllOnes, KnownZero2, KnownOne2, Depth+1);
793 unsigned LeadZ = KnownZero2.countLeadingOnes();
797 ComputeMaskedBits(I->getOperand(1),
798 AllOnes, KnownZero2, KnownOne2, Depth+1);
799 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
800 if (RHSUnknownLeadingOnes != BitWidth)
801 LeadZ = std::min(BitWidth,
802 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
804 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
807 case Instruction::Select:
808 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
809 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
810 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
811 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
813 // Only known if known in both the LHS and RHS.
814 KnownOne &= KnownOne2;
815 KnownZero &= KnownZero2;
817 case Instruction::FPTrunc:
818 case Instruction::FPExt:
819 case Instruction::FPToUI:
820 case Instruction::FPToSI:
821 case Instruction::SIToFP:
822 case Instruction::UIToFP:
823 return; // Can't work with floating point.
824 case Instruction::PtrToInt:
825 case Instruction::IntToPtr:
826 // We can't handle these if we don't know the pointer size.
828 // FALL THROUGH and handle them the same as zext/trunc.
829 case Instruction::ZExt:
830 case Instruction::Trunc: {
831 // Note that we handle pointer operands here because of inttoptr/ptrtoint
832 // which fall through here.
833 const Type *SrcTy = I->getOperand(0)->getType();
834 uint32_t SrcBitWidth = TD ?
835 TD->getTypeSizeInBits(SrcTy) :
836 SrcTy->getPrimitiveSizeInBits();
838 MaskIn.zextOrTrunc(SrcBitWidth);
839 KnownZero.zextOrTrunc(SrcBitWidth);
840 KnownOne.zextOrTrunc(SrcBitWidth);
841 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
842 KnownZero.zextOrTrunc(BitWidth);
843 KnownOne.zextOrTrunc(BitWidth);
844 // Any top bits are known to be zero.
845 if (BitWidth > SrcBitWidth)
846 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
849 case Instruction::BitCast: {
850 const Type *SrcTy = I->getOperand(0)->getType();
851 if (SrcTy->isInteger() || isa<PointerType>(SrcTy)) {
852 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
857 case Instruction::SExt: {
858 // Compute the bits in the result that are not present in the input.
859 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
860 uint32_t SrcBitWidth = SrcTy->getBitWidth();
863 MaskIn.trunc(SrcBitWidth);
864 KnownZero.trunc(SrcBitWidth);
865 KnownOne.trunc(SrcBitWidth);
866 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
867 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
868 KnownZero.zext(BitWidth);
869 KnownOne.zext(BitWidth);
871 // If the sign bit of the input is known set or clear, then we know the
872 // top bits of the result.
873 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
874 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
875 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
876 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
879 case Instruction::Shl:
880 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
881 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
882 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
883 APInt Mask2(Mask.lshr(ShiftAmt));
884 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
885 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
886 KnownZero <<= ShiftAmt;
887 KnownOne <<= ShiftAmt;
888 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
892 case Instruction::LShr:
893 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
894 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
895 // Compute the new bits that are at the top now.
896 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
898 // Unsigned shift right.
899 APInt Mask2(Mask.shl(ShiftAmt));
900 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
901 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
902 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
903 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
904 // high bits known zero.
905 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
909 case Instruction::AShr:
910 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
911 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
912 // Compute the new bits that are at the top now.
913 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
915 // Signed shift right.
916 APInt Mask2(Mask.shl(ShiftAmt));
917 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
918 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
919 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
920 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
922 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
923 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
924 KnownZero |= HighBits;
925 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
926 KnownOne |= HighBits;
930 case Instruction::Sub: {
931 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
932 // We know that the top bits of C-X are clear if X contains less bits
933 // than C (i.e. no wrap-around can happen). For example, 20-X is
934 // positive if we can prove that X is >= 0 and < 16.
935 if (!CLHS->getValue().isNegative()) {
936 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
937 // NLZ can't be BitWidth with no sign bit
938 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
939 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
942 // If all of the MaskV bits are known to be zero, then we know the
943 // output top bits are zero, because we now know that the output is
945 if ((KnownZero2 & MaskV) == MaskV) {
946 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
947 // Top bits known zero.
948 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
954 case Instruction::Add: {
955 // Output known-0 bits are known if clear or set in both the low clear bits
956 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
958 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
959 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
960 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
961 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
963 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
964 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
965 KnownZeroOut = std::min(KnownZeroOut,
966 KnownZero2.countTrailingOnes());
968 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
971 case Instruction::SRem:
972 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
973 APInt RA = Rem->getValue();
974 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
975 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
976 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
977 ComputeMaskedBits(I->getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
979 // The sign of a remainder is equal to the sign of the first
980 // operand (zero being positive).
981 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
982 KnownZero2 |= ~LowBits;
983 else if (KnownOne2[BitWidth-1])
984 KnownOne2 |= ~LowBits;
986 KnownZero |= KnownZero2 & Mask;
987 KnownOne |= KnownOne2 & Mask;
989 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
993 case Instruction::URem: {
994 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
995 APInt RA = Rem->getValue();
996 if (RA.isPowerOf2()) {
997 APInt LowBits = (RA - 1);
998 APInt Mask2 = LowBits & Mask;
999 KnownZero |= ~LowBits & Mask;
1000 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1001 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1006 // Since the result is less than or equal to either operand, any leading
1007 // zero bits in either operand must also exist in the result.
1008 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1009 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
1011 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
1014 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1015 KnownZero2.countLeadingOnes());
1017 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1021 case Instruction::Alloca:
1022 case Instruction::Malloc: {
1023 AllocationInst *AI = cast<AllocationInst>(V);
1024 unsigned Align = AI->getAlignment();
1025 if (Align == 0 && TD) {
1026 if (isa<AllocaInst>(AI))
1027 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
1028 else if (isa<MallocInst>(AI)) {
1029 // Malloc returns maximally aligned memory.
1030 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1033 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
1036 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
1041 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
1042 CountTrailingZeros_32(Align));
1045 case Instruction::GetElementPtr: {
1046 // Analyze all of the subscripts of this getelementptr instruction
1047 // to determine if we can prove known low zero bits.
1048 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
1049 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1050 ComputeMaskedBits(I->getOperand(0), LocalMask,
1051 LocalKnownZero, LocalKnownOne, Depth+1);
1052 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1054 gep_type_iterator GTI = gep_type_begin(I);
1055 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1056 Value *Index = I->getOperand(i);
1057 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1058 // Handle struct member offset arithmetic.
1060 const StructLayout *SL = TD->getStructLayout(STy);
1061 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1062 uint64_t Offset = SL->getElementOffset(Idx);
1063 TrailZ = std::min(TrailZ,
1064 CountTrailingZeros_64(Offset));
1066 // Handle array index arithmetic.
1067 const Type *IndexedTy = GTI.getIndexedType();
1068 if (!IndexedTy->isSized()) return;
1069 unsigned GEPOpiBits = Index->getType()->getPrimitiveSizeInBits();
1070 uint64_t TypeSize = TD ? TD->getABITypeSize(IndexedTy) : 1;
1071 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
1072 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1073 ComputeMaskedBits(Index, LocalMask,
1074 LocalKnownZero, LocalKnownOne, Depth+1);
1075 TrailZ = std::min(TrailZ,
1076 CountTrailingZeros_64(TypeSize) +
1077 LocalKnownZero.countTrailingOnes());
1081 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
1084 case Instruction::PHI: {
1085 PHINode *P = cast<PHINode>(I);
1086 // Handle the case of a simple two-predecessor recurrence PHI.
1087 // There's a lot more that could theoretically be done here, but
1088 // this is sufficient to catch some interesting cases.
1089 if (P->getNumIncomingValues() == 2) {
1090 for (unsigned i = 0; i != 2; ++i) {
1091 Value *L = P->getIncomingValue(i);
1092 Value *R = P->getIncomingValue(!i);
1093 User *LU = dyn_cast<User>(L);
1094 unsigned Opcode = LU ? getOpcode(LU) : (unsigned)Instruction::UserOp1;
1095 // Check for operations that have the property that if
1096 // both their operands have low zero bits, the result
1097 // will have low zero bits.
1098 if (Opcode == Instruction::Add ||
1099 Opcode == Instruction::Sub ||
1100 Opcode == Instruction::And ||
1101 Opcode == Instruction::Or ||
1102 Opcode == Instruction::Mul) {
1103 Value *LL = LU->getOperand(0);
1104 Value *LR = LU->getOperand(1);
1105 // Find a recurrence.
1112 // Ok, we have a PHI of the form L op= R. Check for low
1114 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1115 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, Depth+1);
1116 Mask2 = APInt::getLowBitsSet(BitWidth,
1117 KnownZero2.countTrailingOnes());
1120 ComputeMaskedBits(L, Mask2, KnownZero2, KnownOne2, Depth+1);
1122 APInt::getLowBitsSet(BitWidth,
1123 KnownZero2.countTrailingOnes());
1130 case Instruction::Call:
1131 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1132 switch (II->getIntrinsicID()) {
1134 case Intrinsic::ctpop:
1135 case Intrinsic::ctlz:
1136 case Intrinsic::cttz: {
1137 unsigned LowBits = Log2_32(BitWidth)+1;
1138 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1147 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1148 /// this predicate to simplify operations downstream. Mask is known to be zero
1149 /// for bits that V cannot have.
1150 bool InstCombiner::MaskedValueIsZero(Value *V, const APInt& Mask,
1152 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1153 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1154 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1155 return (KnownZero & Mask) == Mask;
1158 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1159 /// specified instruction is a constant integer. If so, check to see if there
1160 /// are any bits set in the constant that are not demanded. If so, shrink the
1161 /// constant and return true.
1162 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1164 assert(I && "No instruction?");
1165 assert(OpNo < I->getNumOperands() && "Operand index too large");
1167 // If the operand is not a constant integer, nothing to do.
1168 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1169 if (!OpC) return false;
1171 // If there are no bits set that aren't demanded, nothing to do.
1172 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1173 if ((~Demanded & OpC->getValue()) == 0)
1176 // This instruction is producing bits that are not demanded. Shrink the RHS.
1177 Demanded &= OpC->getValue();
1178 I->setOperand(OpNo, ConstantInt::get(Demanded));
1182 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1183 // set of known zero and one bits, compute the maximum and minimum values that
1184 // could have the specified known zero and known one bits, returning them in
1186 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1187 const APInt& KnownZero,
1188 const APInt& KnownOne,
1189 APInt& Min, APInt& Max) {
1190 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1191 assert(KnownZero.getBitWidth() == BitWidth &&
1192 KnownOne.getBitWidth() == BitWidth &&
1193 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
1194 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1195 APInt UnknownBits = ~(KnownZero|KnownOne);
1197 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1198 // bit if it is unknown.
1200 Max = KnownOne|UnknownBits;
1202 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
1203 Min.set(BitWidth-1);
1204 Max.clear(BitWidth-1);
1208 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1209 // a set of known zero and one bits, compute the maximum and minimum values that
1210 // could have the specified known zero and known one bits, returning them in
1212 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1213 const APInt &KnownZero,
1214 const APInt &KnownOne,
1215 APInt &Min, APInt &Max) {
1216 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
1217 assert(KnownZero.getBitWidth() == BitWidth &&
1218 KnownOne.getBitWidth() == BitWidth &&
1219 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
1220 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1221 APInt UnknownBits = ~(KnownZero|KnownOne);
1223 // The minimum value is when the unknown bits are all zeros.
1225 // The maximum value is when the unknown bits are all ones.
1226 Max = KnownOne|UnknownBits;
1229 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1230 /// value based on the demanded bits. When this function is called, it is known
1231 /// that only the bits set in DemandedMask of the result of V are ever used
1232 /// downstream. Consequently, depending on the mask and V, it may be possible
1233 /// to replace V with a constant or one of its operands. In such cases, this
1234 /// function does the replacement and returns true. In all other cases, it
1235 /// returns false after analyzing the expression and setting KnownOne and known
1236 /// to be one in the expression. KnownZero contains all the bits that are known
1237 /// to be zero in the expression. These are provided to potentially allow the
1238 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1239 /// the expression. KnownOne and KnownZero always follow the invariant that
1240 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1241 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1242 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1243 /// and KnownOne must all be the same.
1244 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1245 APInt& KnownZero, APInt& KnownOne,
1247 assert(V != 0 && "Null pointer of Value???");
1248 assert(Depth <= 6 && "Limit Search Depth");
1249 uint32_t BitWidth = DemandedMask.getBitWidth();
1250 const IntegerType *VTy = cast<IntegerType>(V->getType());
1251 assert(VTy->getBitWidth() == BitWidth &&
1252 KnownZero.getBitWidth() == BitWidth &&
1253 KnownOne.getBitWidth() == BitWidth &&
1254 "Value *V, DemandedMask, KnownZero and KnownOne \
1255 must have same BitWidth");
1256 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1257 // We know all of the bits for a constant!
1258 KnownOne = CI->getValue() & DemandedMask;
1259 KnownZero = ~KnownOne & DemandedMask;
1265 if (!V->hasOneUse()) { // Other users may use these bits.
1266 if (Depth != 0) { // Not at the root.
1267 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1268 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1271 // If this is the root being simplified, allow it to have multiple uses,
1272 // just set the DemandedMask to all bits.
1273 DemandedMask = APInt::getAllOnesValue(BitWidth);
1274 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1275 if (V != UndefValue::get(VTy))
1276 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1278 } else if (Depth == 6) { // Limit search depth.
1282 Instruction *I = dyn_cast<Instruction>(V);
1283 if (!I) return false; // Only analyze instructions.
1285 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1286 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1287 switch (I->getOpcode()) {
1289 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1291 case Instruction::And:
1292 // If either the LHS or the RHS are Zero, the result is zero.
1293 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1294 RHSKnownZero, RHSKnownOne, Depth+1))
1296 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1297 "Bits known to be one AND zero?");
1299 // If something is known zero on the RHS, the bits aren't demanded on the
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1302 LHSKnownZero, LHSKnownOne, Depth+1))
1304 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1305 "Bits known to be one AND zero?");
1307 // If all of the demanded bits are known 1 on one side, return the other.
1308 // These bits cannot contribute to the result of the 'and'.
1309 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1310 (DemandedMask & ~LHSKnownZero))
1311 return UpdateValueUsesWith(I, I->getOperand(0));
1312 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1313 (DemandedMask & ~RHSKnownZero))
1314 return UpdateValueUsesWith(I, I->getOperand(1));
1316 // If all of the demanded bits in the inputs are known zeros, return zero.
1317 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1318 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1320 // If the RHS is a constant, see if we can simplify it.
1321 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1322 return UpdateValueUsesWith(I, I);
1324 // Output known-1 bits are only known if set in both the LHS & RHS.
1325 RHSKnownOne &= LHSKnownOne;
1326 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1327 RHSKnownZero |= LHSKnownZero;
1329 case Instruction::Or:
1330 // If either the LHS or the RHS are One, the result is One.
1331 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1332 RHSKnownZero, RHSKnownOne, Depth+1))
1334 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1335 "Bits known to be one AND zero?");
1336 // If something is known one on the RHS, the bits aren't demanded on the
1338 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1339 LHSKnownZero, LHSKnownOne, Depth+1))
1341 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1342 "Bits known to be one AND zero?");
1344 // If all of the demanded bits are known zero on one side, return the other.
1345 // These bits cannot contribute to the result of the 'or'.
1346 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1347 (DemandedMask & ~LHSKnownOne))
1348 return UpdateValueUsesWith(I, I->getOperand(0));
1349 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1350 (DemandedMask & ~RHSKnownOne))
1351 return UpdateValueUsesWith(I, I->getOperand(1));
1353 // If all of the potentially set bits on one side are known to be set on
1354 // the other side, just use the 'other' side.
1355 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1356 (DemandedMask & (~RHSKnownZero)))
1357 return UpdateValueUsesWith(I, I->getOperand(0));
1358 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1359 (DemandedMask & (~LHSKnownZero)))
1360 return UpdateValueUsesWith(I, I->getOperand(1));
1362 // If the RHS is a constant, see if we can simplify it.
1363 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1364 return UpdateValueUsesWith(I, I);
1366 // Output known-0 bits are only known if clear in both the LHS & RHS.
1367 RHSKnownZero &= LHSKnownZero;
1368 // Output known-1 are known to be set if set in either the LHS | RHS.
1369 RHSKnownOne |= LHSKnownOne;
1371 case Instruction::Xor: {
1372 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1373 RHSKnownZero, RHSKnownOne, Depth+1))
1375 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1376 "Bits known to be one AND zero?");
1377 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1378 LHSKnownZero, LHSKnownOne, Depth+1))
1380 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1381 "Bits known to be one AND zero?");
1383 // If all of the demanded bits are known zero on one side, return the other.
1384 // These bits cannot contribute to the result of the 'xor'.
1385 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1386 return UpdateValueUsesWith(I, I->getOperand(0));
1387 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1388 return UpdateValueUsesWith(I, I->getOperand(1));
1390 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1391 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1392 (RHSKnownOne & LHSKnownOne);
1393 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1394 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1395 (RHSKnownOne & LHSKnownZero);
1397 // If all of the demanded bits are known to be zero on one side or the
1398 // other, turn this into an *inclusive* or.
1399 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1400 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1402 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1404 InsertNewInstBefore(Or, *I);
1405 return UpdateValueUsesWith(I, Or);
1408 // If all of the demanded bits on one side are known, and all of the set
1409 // bits on that side are also known to be set on the other side, turn this
1410 // into an AND, as we know the bits will be cleared.
1411 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1412 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1414 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1415 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1417 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1418 InsertNewInstBefore(And, *I);
1419 return UpdateValueUsesWith(I, And);
1423 // If the RHS is a constant, see if we can simplify it.
1424 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1425 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1426 return UpdateValueUsesWith(I, I);
1428 RHSKnownZero = KnownZeroOut;
1429 RHSKnownOne = KnownOneOut;
1432 case Instruction::Select:
1433 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1434 RHSKnownZero, RHSKnownOne, Depth+1))
1436 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1437 LHSKnownZero, LHSKnownOne, Depth+1))
1439 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1440 "Bits known to be one AND zero?");
1441 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1442 "Bits known to be one AND zero?");
1444 // If the operands are constants, see if we can simplify them.
1445 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1446 return UpdateValueUsesWith(I, I);
1447 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1448 return UpdateValueUsesWith(I, I);
1450 // Only known if known in both the LHS and RHS.
1451 RHSKnownOne &= LHSKnownOne;
1452 RHSKnownZero &= LHSKnownZero;
1454 case Instruction::Trunc: {
1456 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1457 DemandedMask.zext(truncBf);
1458 RHSKnownZero.zext(truncBf);
1459 RHSKnownOne.zext(truncBf);
1460 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1461 RHSKnownZero, RHSKnownOne, Depth+1))
1463 DemandedMask.trunc(BitWidth);
1464 RHSKnownZero.trunc(BitWidth);
1465 RHSKnownOne.trunc(BitWidth);
1466 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1467 "Bits known to be one AND zero?");
1470 case Instruction::BitCast:
1471 if (!I->getOperand(0)->getType()->isInteger())
1474 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1475 RHSKnownZero, RHSKnownOne, Depth+1))
1477 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1478 "Bits known to be one AND zero?");
1480 case Instruction::ZExt: {
1481 // Compute the bits in the result that are not present in the input.
1482 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1483 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1485 DemandedMask.trunc(SrcBitWidth);
1486 RHSKnownZero.trunc(SrcBitWidth);
1487 RHSKnownOne.trunc(SrcBitWidth);
1488 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1489 RHSKnownZero, RHSKnownOne, Depth+1))
1491 DemandedMask.zext(BitWidth);
1492 RHSKnownZero.zext(BitWidth);
1493 RHSKnownOne.zext(BitWidth);
1494 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1495 "Bits known to be one AND zero?");
1496 // The top bits are known to be zero.
1497 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1500 case Instruction::SExt: {
1501 // Compute the bits in the result that are not present in the input.
1502 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1503 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1505 APInt InputDemandedBits = DemandedMask &
1506 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1508 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1509 // If any of the sign extended bits are demanded, we know that the sign
1511 if ((NewBits & DemandedMask) != 0)
1512 InputDemandedBits.set(SrcBitWidth-1);
1514 InputDemandedBits.trunc(SrcBitWidth);
1515 RHSKnownZero.trunc(SrcBitWidth);
1516 RHSKnownOne.trunc(SrcBitWidth);
1517 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1518 RHSKnownZero, RHSKnownOne, Depth+1))
1520 InputDemandedBits.zext(BitWidth);
1521 RHSKnownZero.zext(BitWidth);
1522 RHSKnownOne.zext(BitWidth);
1523 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1524 "Bits known to be one AND zero?");
1526 // If the sign bit of the input is known set or clear, then we know the
1527 // top bits of the result.
1529 // If the input sign bit is known zero, or if the NewBits are not demanded
1530 // convert this into a zero extension.
1531 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1533 // Convert to ZExt cast
1534 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1535 return UpdateValueUsesWith(I, NewCast);
1536 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1537 RHSKnownOne |= NewBits;
1541 case Instruction::Add: {
1542 // Figure out what the input bits are. If the top bits of the and result
1543 // are not demanded, then the add doesn't demand them from its input
1545 uint32_t NLZ = DemandedMask.countLeadingZeros();
1547 // If there is a constant on the RHS, there are a variety of xformations
1549 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1550 // If null, this should be simplified elsewhere. Some of the xforms here
1551 // won't work if the RHS is zero.
1555 // If the top bit of the output is demanded, demand everything from the
1556 // input. Otherwise, we demand all the input bits except NLZ top bits.
1557 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1559 // Find information about known zero/one bits in the input.
1560 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1561 LHSKnownZero, LHSKnownOne, Depth+1))
1564 // If the RHS of the add has bits set that can't affect the input, reduce
1566 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1567 return UpdateValueUsesWith(I, I);
1569 // Avoid excess work.
1570 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1573 // Turn it into OR if input bits are zero.
1574 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1576 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1578 InsertNewInstBefore(Or, *I);
1579 return UpdateValueUsesWith(I, Or);
1582 // We can say something about the output known-zero and known-one bits,
1583 // depending on potential carries from the input constant and the
1584 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1585 // bits set and the RHS constant is 0x01001, then we know we have a known
1586 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1588 // To compute this, we first compute the potential carry bits. These are
1589 // the bits which may be modified. I'm not aware of a better way to do
1591 const APInt& RHSVal = RHS->getValue();
1592 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1594 // Now that we know which bits have carries, compute the known-1/0 sets.
1596 // Bits are known one if they are known zero in one operand and one in the
1597 // other, and there is no input carry.
1598 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1599 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1601 // Bits are known zero if they are known zero in both operands and there
1602 // is no input carry.
1603 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1605 // If the high-bits of this ADD are not demanded, then it does not demand
1606 // the high bits of its LHS or RHS.
1607 if (DemandedMask[BitWidth-1] == 0) {
1608 // Right fill the mask of bits for this ADD to demand the most
1609 // significant bit and all those below it.
1610 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1611 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1612 LHSKnownZero, LHSKnownOne, Depth+1))
1614 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1615 LHSKnownZero, LHSKnownOne, Depth+1))
1621 case Instruction::Sub:
1622 // If the high-bits of this SUB are not demanded, then it does not demand
1623 // the high bits of its LHS or RHS.
1624 if (DemandedMask[BitWidth-1] == 0) {
1625 // Right fill the mask of bits for this SUB to demand the most
1626 // significant bit and all those below it.
1627 uint32_t NLZ = DemandedMask.countLeadingZeros();
1628 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1629 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1630 LHSKnownZero, LHSKnownOne, Depth+1))
1632 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1633 LHSKnownZero, LHSKnownOne, Depth+1))
1636 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1637 // the known zeros and ones.
1638 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1640 case Instruction::Shl:
1641 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1642 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1643 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1644 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1645 RHSKnownZero, RHSKnownOne, Depth+1))
1647 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1648 "Bits known to be one AND zero?");
1649 RHSKnownZero <<= ShiftAmt;
1650 RHSKnownOne <<= ShiftAmt;
1651 // low bits known zero.
1653 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1656 case Instruction::LShr:
1657 // For a logical shift right
1658 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1659 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1661 // Unsigned shift right.
1662 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1663 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1664 RHSKnownZero, RHSKnownOne, Depth+1))
1666 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1667 "Bits known to be one AND zero?");
1668 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1669 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1671 // Compute the new bits that are at the top now.
1672 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1673 RHSKnownZero |= HighBits; // high bits known zero.
1677 case Instruction::AShr:
1678 // If this is an arithmetic shift right and only the low-bit is set, we can
1679 // always convert this into a logical shr, even if the shift amount is
1680 // variable. The low bit of the shift cannot be an input sign bit unless
1681 // the shift amount is >= the size of the datatype, which is undefined.
1682 if (DemandedMask == 1) {
1683 // Perform the logical shift right.
1684 Value *NewVal = BinaryOperator::CreateLShr(
1685 I->getOperand(0), I->getOperand(1), I->getName());
1686 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1687 return UpdateValueUsesWith(I, NewVal);
1690 // If the sign bit is the only bit demanded by this ashr, then there is no
1691 // need to do it, the shift doesn't change the high bit.
1692 if (DemandedMask.isSignBit())
1693 return UpdateValueUsesWith(I, I->getOperand(0));
1695 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1696 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1698 // Signed shift right.
1699 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1700 // If any of the "high bits" are demanded, we should set the sign bit as
1702 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1703 DemandedMaskIn.set(BitWidth-1);
1704 if (SimplifyDemandedBits(I->getOperand(0),
1706 RHSKnownZero, RHSKnownOne, Depth+1))
1708 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1709 "Bits known to be one AND zero?");
1710 // Compute the new bits that are at the top now.
1711 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1712 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1713 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1715 // Handle the sign bits.
1716 APInt SignBit(APInt::getSignBit(BitWidth));
1717 // Adjust to where it is now in the mask.
1718 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1720 // If the input sign bit is known to be zero, or if none of the top bits
1721 // are demanded, turn this into an unsigned shift right.
1722 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1723 (HighBits & ~DemandedMask) == HighBits) {
1724 // Perform the logical shift right.
1725 Value *NewVal = BinaryOperator::CreateLShr(
1726 I->getOperand(0), SA, I->getName());
1727 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1728 return UpdateValueUsesWith(I, NewVal);
1729 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1730 RHSKnownOne |= HighBits;
1734 case Instruction::SRem:
1735 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1736 APInt RA = Rem->getValue();
1737 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1738 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1739 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1740 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1741 LHSKnownZero, LHSKnownOne, Depth+1))
1744 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1745 LHSKnownZero |= ~LowBits;
1746 else if (LHSKnownOne[BitWidth-1])
1747 LHSKnownOne |= ~LowBits;
1749 KnownZero |= LHSKnownZero & DemandedMask;
1750 KnownOne |= LHSKnownOne & DemandedMask;
1752 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1756 case Instruction::URem: {
1757 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1758 APInt RA = Rem->getValue();
1759 if (RA.isPowerOf2()) {
1760 APInt LowBits = (RA - 1);
1761 APInt Mask2 = LowBits & DemandedMask;
1762 KnownZero |= ~LowBits & DemandedMask;
1763 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1764 KnownZero, KnownOne, Depth+1))
1767 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1772 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1773 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1774 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1775 KnownZero2, KnownOne2, Depth+1))
1778 uint32_t Leaders = KnownZero2.countLeadingOnes();
1779 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1780 KnownZero2, KnownOne2, Depth+1))
1783 Leaders = std::max(Leaders,
1784 KnownZero2.countLeadingOnes());
1785 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1790 // If the client is only demanding bits that we know, return the known
1792 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1793 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1798 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1799 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1800 /// actually used by the caller. This method analyzes which elements of the
1801 /// operand are undef and returns that information in UndefElts.
1803 /// If the information about demanded elements can be used to simplify the
1804 /// operation, the operation is simplified, then the resultant value is
1805 /// returned. This returns null if no change was made.
1806 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1807 uint64_t &UndefElts,
1809 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1810 assert(VWidth <= 64 && "Vector too wide to analyze!");
1811 uint64_t EltMask = ~0ULL >> (64-VWidth);
1812 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1813 "Invalid DemandedElts!");
1815 if (isa<UndefValue>(V)) {
1816 // If the entire vector is undefined, just return this info.
1817 UndefElts = EltMask;
1819 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1820 UndefElts = EltMask;
1821 return UndefValue::get(V->getType());
1825 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1826 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1827 Constant *Undef = UndefValue::get(EltTy);
1829 std::vector<Constant*> Elts;
1830 for (unsigned i = 0; i != VWidth; ++i)
1831 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1832 Elts.push_back(Undef);
1833 UndefElts |= (1ULL << i);
1834 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1835 Elts.push_back(Undef);
1836 UndefElts |= (1ULL << i);
1837 } else { // Otherwise, defined.
1838 Elts.push_back(CP->getOperand(i));
1841 // If we changed the constant, return it.
1842 Constant *NewCP = ConstantVector::get(Elts);
1843 return NewCP != CP ? NewCP : 0;
1844 } else if (isa<ConstantAggregateZero>(V)) {
1845 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1847 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1848 Constant *Zero = Constant::getNullValue(EltTy);
1849 Constant *Undef = UndefValue::get(EltTy);
1850 std::vector<Constant*> Elts;
1851 for (unsigned i = 0; i != VWidth; ++i)
1852 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1853 UndefElts = DemandedElts ^ EltMask;
1854 return ConstantVector::get(Elts);
1857 if (!V->hasOneUse()) { // Other users may use these bits.
1858 if (Depth != 0) { // Not at the root.
1859 // TODO: Just compute the UndefElts information recursively.
1863 } else if (Depth == 10) { // Limit search depth.
1867 Instruction *I = dyn_cast<Instruction>(V);
1868 if (!I) return false; // Only analyze instructions.
1870 bool MadeChange = false;
1871 uint64_t UndefElts2;
1873 switch (I->getOpcode()) {
1876 case Instruction::InsertElement: {
1877 // If this is a variable index, we don't know which element it overwrites.
1878 // demand exactly the same input as we produce.
1879 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1881 // Note that we can't propagate undef elt info, because we don't know
1882 // which elt is getting updated.
1883 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1884 UndefElts2, Depth+1);
1885 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1889 // If this is inserting an element that isn't demanded, remove this
1891 unsigned IdxNo = Idx->getZExtValue();
1892 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1893 return AddSoonDeadInstToWorklist(*I, 0);
1895 // Otherwise, the element inserted overwrites whatever was there, so the
1896 // input demanded set is simpler than the output set.
1897 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1898 DemandedElts & ~(1ULL << IdxNo),
1899 UndefElts, Depth+1);
1900 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1902 // The inserted element is defined.
1903 UndefElts |= 1ULL << IdxNo;
1906 case Instruction::BitCast: {
1907 // Vector->vector casts only.
1908 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1910 unsigned InVWidth = VTy->getNumElements();
1911 uint64_t InputDemandedElts = 0;
1914 if (VWidth == InVWidth) {
1915 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1916 // elements as are demanded of us.
1918 InputDemandedElts = DemandedElts;
1919 } else if (VWidth > InVWidth) {
1923 // If there are more elements in the result than there are in the source,
1924 // then an input element is live if any of the corresponding output
1925 // elements are live.
1926 Ratio = VWidth/InVWidth;
1927 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1928 if (DemandedElts & (1ULL << OutIdx))
1929 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1935 // If there are more elements in the source than there are in the result,
1936 // then an input element is live if the corresponding output element is
1938 Ratio = InVWidth/VWidth;
1939 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1940 if (DemandedElts & (1ULL << InIdx/Ratio))
1941 InputDemandedElts |= 1ULL << InIdx;
1944 // div/rem demand all inputs, because they don't want divide by zero.
1945 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1946 UndefElts2, Depth+1);
1948 I->setOperand(0, TmpV);
1952 UndefElts = UndefElts2;
1953 if (VWidth > InVWidth) {
1954 assert(0 && "Unimp");
1955 // If there are more elements in the result than there are in the source,
1956 // then an output element is undef if the corresponding input element is
1958 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1959 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1960 UndefElts |= 1ULL << OutIdx;
1961 } else if (VWidth < InVWidth) {
1962 assert(0 && "Unimp");
1963 // If there are more elements in the source than there are in the result,
1964 // then a result element is undef if all of the corresponding input
1965 // elements are undef.
1966 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1967 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1968 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1969 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1973 case Instruction::And:
1974 case Instruction::Or:
1975 case Instruction::Xor:
1976 case Instruction::Add:
1977 case Instruction::Sub:
1978 case Instruction::Mul:
1979 // div/rem demand all inputs, because they don't want divide by zero.
1980 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1981 UndefElts, Depth+1);
1982 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1983 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1984 UndefElts2, Depth+1);
1985 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1987 // Output elements are undefined if both are undefined. Consider things
1988 // like undef&0. The result is known zero, not undef.
1989 UndefElts &= UndefElts2;
1992 case Instruction::Call: {
1993 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1995 switch (II->getIntrinsicID()) {
1998 // Binary vector operations that work column-wise. A dest element is a
1999 // function of the corresponding input elements from the two inputs.
2000 case Intrinsic::x86_sse_sub_ss:
2001 case Intrinsic::x86_sse_mul_ss:
2002 case Intrinsic::x86_sse_min_ss:
2003 case Intrinsic::x86_sse_max_ss:
2004 case Intrinsic::x86_sse2_sub_sd:
2005 case Intrinsic::x86_sse2_mul_sd:
2006 case Intrinsic::x86_sse2_min_sd:
2007 case Intrinsic::x86_sse2_max_sd:
2008 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
2009 UndefElts, Depth+1);
2010 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
2011 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
2012 UndefElts2, Depth+1);
2013 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
2015 // If only the low elt is demanded and this is a scalarizable intrinsic,
2016 // scalarize it now.
2017 if (DemandedElts == 1) {
2018 switch (II->getIntrinsicID()) {
2020 case Intrinsic::x86_sse_sub_ss:
2021 case Intrinsic::x86_sse_mul_ss:
2022 case Intrinsic::x86_sse2_sub_sd:
2023 case Intrinsic::x86_sse2_mul_sd:
2024 // TODO: Lower MIN/MAX/ABS/etc
2025 Value *LHS = II->getOperand(1);
2026 Value *RHS = II->getOperand(2);
2027 // Extract the element as scalars.
2028 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
2029 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
2031 switch (II->getIntrinsicID()) {
2032 default: assert(0 && "Case stmts out of sync!");
2033 case Intrinsic::x86_sse_sub_ss:
2034 case Intrinsic::x86_sse2_sub_sd:
2035 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
2036 II->getName()), *II);
2038 case Intrinsic::x86_sse_mul_ss:
2039 case Intrinsic::x86_sse2_mul_sd:
2040 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
2041 II->getName()), *II);
2046 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
2048 InsertNewInstBefore(New, *II);
2049 AddSoonDeadInstToWorklist(*II, 0);
2054 // Output elements are undefined if both are undefined. Consider things
2055 // like undef&0. The result is known zero, not undef.
2056 UndefElts &= UndefElts2;
2062 return MadeChange ? I : 0;
2065 /// ComputeNumSignBits - Return the number of times the sign bit of the
2066 /// register is replicated into the other bits. We know that at least 1 bit
2067 /// is always equal to the sign bit (itself), but other cases can give us
2068 /// information. For example, immediately after an "ashr X, 2", we know that
2069 /// the top 3 bits are all equal to each other, so we return 3.
2071 unsigned InstCombiner::ComputeNumSignBits(Value *V, unsigned Depth) const{
2072 const IntegerType *Ty = cast<IntegerType>(V->getType());
2073 unsigned TyBits = Ty->getBitWidth();
2075 unsigned FirstAnswer = 1;
2078 return 1; // Limit search depth.
2080 User *U = dyn_cast<User>(V);
2081 switch (getOpcode(V)) {
2083 case Instruction::SExt:
2084 Tmp = TyBits-cast<IntegerType>(U->getOperand(0)->getType())->getBitWidth();
2085 return ComputeNumSignBits(U->getOperand(0), Depth+1) + Tmp;
2087 case Instruction::AShr:
2088 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2089 // ashr X, C -> adds C sign bits.
2090 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
2091 Tmp += C->getZExtValue();
2092 if (Tmp > TyBits) Tmp = TyBits;
2095 case Instruction::Shl:
2096 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
2097 // shl destroys sign bits.
2098 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2099 if (C->getZExtValue() >= TyBits || // Bad shift.
2100 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2101 return Tmp - C->getZExtValue();
2104 case Instruction::And:
2105 case Instruction::Or:
2106 case Instruction::Xor: // NOT is handled here.
2107 // Logical binary ops preserve the number of sign bits at the worst.
2108 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2110 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2111 FirstAnswer = std::min(Tmp, Tmp2);
2112 // We computed what we know about the sign bits as our first
2113 // answer. Now proceed to the generic code that uses
2114 // ComputeMaskedBits, and pick whichever answer is better.
2118 case Instruction::Select:
2119 Tmp = ComputeNumSignBits(U->getOperand(1), Depth+1);
2120 if (Tmp == 1) return 1; // Early out.
2121 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth+1);
2122 return std::min(Tmp, Tmp2);
2124 case Instruction::Add:
2125 // Add can have at most one carry bit. Thus we know that the output
2126 // is, at worst, one more bit than the inputs.
2127 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2128 if (Tmp == 1) return 1; // Early out.
2130 // Special case decrementing a value (ADD X, -1):
2131 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(0)))
2132 if (CRHS->isAllOnesValue()) {
2133 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2134 APInt Mask = APInt::getAllOnesValue(TyBits);
2135 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2137 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2139 if ((KnownZero | APInt(TyBits, 1)) == Mask)
2142 // If we are subtracting one from a positive number, there is no carry
2143 // out of the result.
2144 if (KnownZero.isNegative())
2148 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2149 if (Tmp2 == 1) return 1;
2150 return std::min(Tmp, Tmp2)-1;
2153 case Instruction::Sub:
2154 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2155 if (Tmp2 == 1) return 1;
2158 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
2159 if (CLHS->isNullValue()) {
2160 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2161 APInt Mask = APInt::getAllOnesValue(TyBits);
2162 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2163 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2165 if ((KnownZero | APInt(TyBits, 1)) == Mask)
2168 // If the input is known to be positive (the sign bit is known clear),
2169 // the output of the NEG has the same number of sign bits as the input.
2170 if (KnownZero.isNegative())
2173 // Otherwise, we treat this like a SUB.
2176 // Sub can have at most one carry bit. Thus we know that the output
2177 // is, at worst, one more bit than the inputs.
2178 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2179 if (Tmp == 1) return 1; // Early out.
2180 return std::min(Tmp, Tmp2)-1;
2182 case Instruction::Trunc:
2183 // FIXME: it's tricky to do anything useful for this, but it is an important
2184 // case for targets like X86.
2188 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2189 // use this information.
2190 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2191 APInt Mask = APInt::getAllOnesValue(TyBits);
2192 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
2194 if (KnownZero.isNegative()) { // sign bit is 0
2196 } else if (KnownOne.isNegative()) { // sign bit is 1;
2203 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2204 // the number of identical bits in the top of the input value.
2206 Mask <<= Mask.getBitWidth()-TyBits;
2207 // Return # leading zeros. We use 'min' here in case Val was zero before
2208 // shifting. We don't want to return '64' as for an i32 "0".
2209 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2213 /// AssociativeOpt - Perform an optimization on an associative operator. This
2214 /// function is designed to check a chain of associative operators for a
2215 /// potential to apply a certain optimization. Since the optimization may be
2216 /// applicable if the expression was reassociated, this checks the chain, then
2217 /// reassociates the expression as necessary to expose the optimization
2218 /// opportunity. This makes use of a special Functor, which must define
2219 /// 'shouldApply' and 'apply' methods.
2221 template<typename Functor>
2222 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2223 unsigned Opcode = Root.getOpcode();
2224 Value *LHS = Root.getOperand(0);
2226 // Quick check, see if the immediate LHS matches...
2227 if (F.shouldApply(LHS))
2228 return F.apply(Root);
2230 // Otherwise, if the LHS is not of the same opcode as the root, return.
2231 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2232 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2233 // Should we apply this transform to the RHS?
2234 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2236 // If not to the RHS, check to see if we should apply to the LHS...
2237 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2238 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2242 // If the functor wants to apply the optimization to the RHS of LHSI,
2243 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2245 BasicBlock *BB = Root.getParent();
2247 // Now all of the instructions are in the current basic block, go ahead
2248 // and perform the reassociation.
2249 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2251 // First move the selected RHS to the LHS of the root...
2252 Root.setOperand(0, LHSI->getOperand(1));
2254 // Make what used to be the LHS of the root be the user of the root...
2255 Value *ExtraOperand = TmpLHSI->getOperand(1);
2256 if (&Root == TmpLHSI) {
2257 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2260 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2261 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2262 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2263 BasicBlock::iterator ARI = &Root; ++ARI;
2264 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2267 // Now propagate the ExtraOperand down the chain of instructions until we
2269 while (TmpLHSI != LHSI) {
2270 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2271 // Move the instruction to immediately before the chain we are
2272 // constructing to avoid breaking dominance properties.
2273 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2274 BB->getInstList().insert(ARI, NextLHSI);
2277 Value *NextOp = NextLHSI->getOperand(1);
2278 NextLHSI->setOperand(1, ExtraOperand);
2280 ExtraOperand = NextOp;
2283 // Now that the instructions are reassociated, have the functor perform
2284 // the transformation...
2285 return F.apply(Root);
2288 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2295 // AddRHS - Implements: X + X --> X << 1
2298 AddRHS(Value *rhs) : RHS(rhs) {}
2299 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2300 Instruction *apply(BinaryOperator &Add) const {
2301 return BinaryOperator::CreateShl(Add.getOperand(0),
2302 ConstantInt::get(Add.getType(), 1));
2306 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2308 struct AddMaskingAnd {
2310 AddMaskingAnd(Constant *c) : C2(c) {}
2311 bool shouldApply(Value *LHS) const {
2313 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2314 ConstantExpr::getAnd(C1, C2)->isNullValue();
2316 Instruction *apply(BinaryOperator &Add) const {
2317 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
2323 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2325 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2326 if (Constant *SOC = dyn_cast<Constant>(SO))
2327 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2329 return IC->InsertNewInstBefore(CastInst::Create(
2330 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2333 // Figure out if the constant is the left or the right argument.
2334 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2335 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2337 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2339 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2340 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2343 Value *Op0 = SO, *Op1 = ConstOperand;
2345 std::swap(Op0, Op1);
2347 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2348 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2349 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2350 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2351 SO->getName()+".cmp");
2353 assert(0 && "Unknown binary instruction type!");
2356 return IC->InsertNewInstBefore(New, I);
2359 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2360 // constant as the other operand, try to fold the binary operator into the
2361 // select arguments. This also works for Cast instructions, which obviously do
2362 // not have a second operand.
2363 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2365 // Don't modify shared select instructions
2366 if (!SI->hasOneUse()) return 0;
2367 Value *TV = SI->getOperand(1);
2368 Value *FV = SI->getOperand(2);
2370 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2371 // Bool selects with constant operands can be folded to logical ops.
2372 if (SI->getType() == Type::Int1Ty) return 0;
2374 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2375 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2377 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2384 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2385 /// node as operand #0, see if we can fold the instruction into the PHI (which
2386 /// is only possible if all operands to the PHI are constants).
2387 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2388 PHINode *PN = cast<PHINode>(I.getOperand(0));
2389 unsigned NumPHIValues = PN->getNumIncomingValues();
2390 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2392 // Check to see if all of the operands of the PHI are constants. If there is
2393 // one non-constant value, remember the BB it is. If there is more than one
2394 // or if *it* is a PHI, bail out.
2395 BasicBlock *NonConstBB = 0;
2396 for (unsigned i = 0; i != NumPHIValues; ++i)
2397 if (!isa<Constant>(PN->getIncomingValue(i))) {
2398 if (NonConstBB) return 0; // More than one non-const value.
2399 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2400 NonConstBB = PN->getIncomingBlock(i);
2402 // If the incoming non-constant value is in I's block, we have an infinite
2404 if (NonConstBB == I.getParent())
2408 // If there is exactly one non-constant value, we can insert a copy of the
2409 // operation in that block. However, if this is a critical edge, we would be
2410 // inserting the computation one some other paths (e.g. inside a loop). Only
2411 // do this if the pred block is unconditionally branching into the phi block.
2413 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2414 if (!BI || !BI->isUnconditional()) return 0;
2417 // Okay, we can do the transformation: create the new PHI node.
2418 PHINode *NewPN = PHINode::Create(I.getType(), "");
2419 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2420 InsertNewInstBefore(NewPN, *PN);
2421 NewPN->takeName(PN);
2423 // Next, add all of the operands to the PHI.
2424 if (I.getNumOperands() == 2) {
2425 Constant *C = cast<Constant>(I.getOperand(1));
2426 for (unsigned i = 0; i != NumPHIValues; ++i) {
2428 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2429 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2430 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2432 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2434 assert(PN->getIncomingBlock(i) == NonConstBB);
2435 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2436 InV = BinaryOperator::Create(BO->getOpcode(),
2437 PN->getIncomingValue(i), C, "phitmp",
2438 NonConstBB->getTerminator());
2439 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2440 InV = CmpInst::Create(CI->getOpcode(),
2442 PN->getIncomingValue(i), C, "phitmp",
2443 NonConstBB->getTerminator());
2445 assert(0 && "Unknown binop!");
2447 AddToWorkList(cast<Instruction>(InV));
2449 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2452 CastInst *CI = cast<CastInst>(&I);
2453 const Type *RetTy = CI->getType();
2454 for (unsigned i = 0; i != NumPHIValues; ++i) {
2456 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2457 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2459 assert(PN->getIncomingBlock(i) == NonConstBB);
2460 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2461 I.getType(), "phitmp",
2462 NonConstBB->getTerminator());
2463 AddToWorkList(cast<Instruction>(InV));
2465 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2468 return ReplaceInstUsesWith(I, NewPN);
2472 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
2473 /// value is never equal to -0.0.
2475 /// Note that this function will need to be revisited when we support nondefault
2478 static bool CannotBeNegativeZero(const Value *V) {
2479 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2480 return !CFP->getValueAPF().isNegZero();
2482 if (const Instruction *I = dyn_cast<Instruction>(V)) {
2483 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2484 if (I->getOpcode() == Instruction::Add &&
2485 isa<ConstantFP>(I->getOperand(1)) &&
2486 cast<ConstantFP>(I->getOperand(1))->isNullValue())
2489 // sitofp and uitofp turn into +0.0 for zero.
2490 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2493 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2494 if (II->getIntrinsicID() == Intrinsic::sqrt)
2495 return CannotBeNegativeZero(II->getOperand(1));
2497 if (const CallInst *CI = dyn_cast<CallInst>(I))
2498 if (const Function *F = CI->getCalledFunction()) {
2499 if (F->isDeclaration()) {
2500 switch (F->getNameLen()) {
2501 case 3: // abs(x) != -0.0
2502 if (!strcmp(F->getNameStart(), "abs")) return true;
2504 case 4: // abs[lf](x) != -0.0
2505 if (!strcmp(F->getNameStart(), "absf")) return true;
2506 if (!strcmp(F->getNameStart(), "absl")) return true;
2516 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2517 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2518 /// This basically requires proving that the add in the original type would not
2519 /// overflow to change the sign bit or have a carry out.
2520 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2521 // There are different heuristics we can use for this. Here are some simple
2524 // Add has the property that adding any two 2's complement numbers can only
2525 // have one carry bit which can change a sign. As such, if LHS and RHS each
2526 // have at least two sign bits, we know that the addition of the two values will
2527 // sign extend fine.
2528 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2532 // If one of the operands only has one non-zero bit, and if the other operand
2533 // has a known-zero bit in a more significant place than it (not including the
2534 // sign bit) the ripple may go up to and fill the zero, but won't change the
2535 // sign. For example, (X & ~4) + 1.
2543 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2544 bool Changed = SimplifyCommutative(I);
2545 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2547 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2548 // X + undef -> undef
2549 if (isa<UndefValue>(RHS))
2550 return ReplaceInstUsesWith(I, RHS);
2553 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2554 if (RHSC->isNullValue())
2555 return ReplaceInstUsesWith(I, LHS);
2556 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2557 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2558 (I.getType())->getValueAPF()))
2559 return ReplaceInstUsesWith(I, LHS);
2562 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2563 // X + (signbit) --> X ^ signbit
2564 const APInt& Val = CI->getValue();
2565 uint32_t BitWidth = Val.getBitWidth();
2566 if (Val == APInt::getSignBit(BitWidth))
2567 return BinaryOperator::CreateXor(LHS, RHS);
2569 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2570 // (X & 254)+1 -> (X&254)|1
2571 if (!isa<VectorType>(I.getType())) {
2572 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2573 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2574 KnownZero, KnownOne))
2579 if (isa<PHINode>(LHS))
2580 if (Instruction *NV = FoldOpIntoPhi(I))
2583 ConstantInt *XorRHS = 0;
2585 if (isa<ConstantInt>(RHSC) &&
2586 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2587 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2588 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2590 uint32_t Size = TySizeBits / 2;
2591 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2592 APInt CFF80Val(-C0080Val);
2594 if (TySizeBits > Size) {
2595 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2596 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2597 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2598 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2599 // This is a sign extend if the top bits are known zero.
2600 if (!MaskedValueIsZero(XorLHS,
2601 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2602 Size = 0; // Not a sign ext, but can't be any others either.
2607 C0080Val = APIntOps::lshr(C0080Val, Size);
2608 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2609 } while (Size >= 1);
2611 // FIXME: This shouldn't be necessary. When the backends can handle types
2612 // with funny bit widths then this switch statement should be removed. It
2613 // is just here to get the size of the "middle" type back up to something
2614 // that the back ends can handle.
2615 const Type *MiddleType = 0;
2618 case 32: MiddleType = Type::Int32Ty; break;
2619 case 16: MiddleType = Type::Int16Ty; break;
2620 case 8: MiddleType = Type::Int8Ty; break;
2623 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2624 InsertNewInstBefore(NewTrunc, I);
2625 return new SExtInst(NewTrunc, I.getType(), I.getName());
2631 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2632 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2634 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2635 if (RHSI->getOpcode() == Instruction::Sub)
2636 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2637 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2639 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2640 if (LHSI->getOpcode() == Instruction::Sub)
2641 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2642 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2647 // -A + -B --> -(A + B)
2648 if (Value *LHSV = dyn_castNegVal(LHS)) {
2649 if (LHS->getType()->isIntOrIntVector()) {
2650 if (Value *RHSV = dyn_castNegVal(RHS)) {
2651 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2652 InsertNewInstBefore(NewAdd, I);
2653 return BinaryOperator::CreateNeg(NewAdd);
2657 return BinaryOperator::CreateSub(RHS, LHSV);
2661 if (!isa<Constant>(RHS))
2662 if (Value *V = dyn_castNegVal(RHS))
2663 return BinaryOperator::CreateSub(LHS, V);
2667 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2668 if (X == RHS) // X*C + X --> X * (C+1)
2669 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2671 // X*C1 + X*C2 --> X * (C1+C2)
2673 if (X == dyn_castFoldableMul(RHS, C1))
2674 return BinaryOperator::CreateMul(X, Add(C1, C2));
2677 // X + X*C --> X * (C+1)
2678 if (dyn_castFoldableMul(RHS, C2) == LHS)
2679 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2681 // X + ~X --> -1 since ~X = -X-1
2682 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2683 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2686 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2687 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2688 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2691 // A+B --> A|B iff A and B have no bits set in common.
2692 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2693 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2694 APInt LHSKnownOne(IT->getBitWidth(), 0);
2695 APInt LHSKnownZero(IT->getBitWidth(), 0);
2696 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2697 if (LHSKnownZero != 0) {
2698 APInt RHSKnownOne(IT->getBitWidth(), 0);
2699 APInt RHSKnownZero(IT->getBitWidth(), 0);
2700 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2702 // No bits in common -> bitwise or.
2703 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2704 return BinaryOperator::CreateOr(LHS, RHS);
2708 // W*X + Y*Z --> W * (X+Z) iff W == Y
2709 if (I.getType()->isIntOrIntVector()) {
2710 Value *W, *X, *Y, *Z;
2711 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2712 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2716 } else if (Y == X) {
2718 } else if (X == Z) {
2725 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2726 LHS->getName()), I);
2727 return BinaryOperator::CreateMul(W, NewAdd);
2732 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2734 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2735 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2737 // (X & FF00) + xx00 -> (X+xx00) & FF00
2738 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2739 Constant *Anded = And(CRHS, C2);
2740 if (Anded == CRHS) {
2741 // See if all bits from the first bit set in the Add RHS up are included
2742 // in the mask. First, get the rightmost bit.
2743 const APInt& AddRHSV = CRHS->getValue();
2745 // Form a mask of all bits from the lowest bit added through the top.
2746 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2748 // See if the and mask includes all of these bits.
2749 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2751 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2752 // Okay, the xform is safe. Insert the new add pronto.
2753 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2754 LHS->getName()), I);
2755 return BinaryOperator::CreateAnd(NewAdd, C2);
2760 // Try to fold constant add into select arguments.
2761 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2762 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2766 // add (cast *A to intptrtype) B ->
2767 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2769 CastInst *CI = dyn_cast<CastInst>(LHS);
2772 CI = dyn_cast<CastInst>(RHS);
2775 if (CI && CI->getType()->isSized() &&
2776 (CI->getType()->getPrimitiveSizeInBits() ==
2777 TD->getIntPtrType()->getPrimitiveSizeInBits())
2778 && isa<PointerType>(CI->getOperand(0)->getType())) {
2780 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2781 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2782 PointerType::get(Type::Int8Ty, AS), I);
2783 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2784 return new PtrToIntInst(I2, CI->getType());
2788 // add (select X 0 (sub n A)) A --> select X A n
2790 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2793 SI = dyn_cast<SelectInst>(RHS);
2796 if (SI && SI->hasOneUse()) {
2797 Value *TV = SI->getTrueValue();
2798 Value *FV = SI->getFalseValue();
2801 // Can we fold the add into the argument of the select?
2802 // We check both true and false select arguments for a matching subtract.
2803 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2804 A == Other) // Fold the add into the true select value.
2805 return SelectInst::Create(SI->getCondition(), N, A);
2806 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2807 A == Other) // Fold the add into the false select value.
2808 return SelectInst::Create(SI->getCondition(), A, N);
2812 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2813 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2814 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2815 return ReplaceInstUsesWith(I, LHS);
2817 // Check for (add (sext x), y), see if we can merge this into an
2818 // integer add followed by a sext.
2819 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2820 // (add (sext x), cst) --> (sext (add x, cst'))
2821 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2823 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2824 if (LHSConv->hasOneUse() &&
2825 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2826 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2827 // Insert the new, smaller add.
2828 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2830 InsertNewInstBefore(NewAdd, I);
2831 return new SExtInst(NewAdd, I.getType());
2835 // (add (sext x), (sext y)) --> (sext (add int x, y))
2836 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2837 // Only do this if x/y have the same type, if at last one of them has a
2838 // single use (so we don't increase the number of sexts), and if the
2839 // integer add will not overflow.
2840 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2841 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2842 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2843 RHSConv->getOperand(0))) {
2844 // Insert the new integer add.
2845 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2846 RHSConv->getOperand(0),
2848 InsertNewInstBefore(NewAdd, I);
2849 return new SExtInst(NewAdd, I.getType());
2854 // Check for (add double (sitofp x), y), see if we can merge this into an
2855 // integer add followed by a promotion.
2856 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2857 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2858 // ... if the constant fits in the integer value. This is useful for things
2859 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2860 // requires a constant pool load, and generally allows the add to be better
2862 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2864 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2865 if (LHSConv->hasOneUse() &&
2866 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2867 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2868 // Insert the new integer add.
2869 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2871 InsertNewInstBefore(NewAdd, I);
2872 return new SIToFPInst(NewAdd, I.getType());
2876 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2877 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2878 // Only do this if x/y have the same type, if at last one of them has a
2879 // single use (so we don't increase the number of int->fp conversions),
2880 // and if the integer add will not overflow.
2881 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2882 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2883 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2884 RHSConv->getOperand(0))) {
2885 // Insert the new integer add.
2886 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2887 RHSConv->getOperand(0),
2889 InsertNewInstBefore(NewAdd, I);
2890 return new SIToFPInst(NewAdd, I.getType());
2895 return Changed ? &I : 0;
2898 // isSignBit - Return true if the value represented by the constant only has the
2899 // highest order bit set.
2900 static bool isSignBit(ConstantInt *CI) {
2901 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2902 return CI->getValue() == APInt::getSignBit(NumBits);
2905 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2906 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2908 if (Op0 == Op1) // sub X, X -> 0
2909 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2911 // If this is a 'B = x-(-A)', change to B = x+A...
2912 if (Value *V = dyn_castNegVal(Op1))
2913 return BinaryOperator::CreateAdd(Op0, V);
2915 if (isa<UndefValue>(Op0))
2916 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2917 if (isa<UndefValue>(Op1))
2918 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2920 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2921 // Replace (-1 - A) with (~A)...
2922 if (C->isAllOnesValue())
2923 return BinaryOperator::CreateNot(Op1);
2925 // C - ~X == X + (1+C)
2927 if (match(Op1, m_Not(m_Value(X))))
2928 return BinaryOperator::CreateAdd(X, AddOne(C));
2930 // -(X >>u 31) -> (X >>s 31)
2931 // -(X >>s 31) -> (X >>u 31)
2933 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2934 if (SI->getOpcode() == Instruction::LShr) {
2935 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2936 // Check to see if we are shifting out everything but the sign bit.
2937 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2938 SI->getType()->getPrimitiveSizeInBits()-1) {
2939 // Ok, the transformation is safe. Insert AShr.
2940 return BinaryOperator::Create(Instruction::AShr,
2941 SI->getOperand(0), CU, SI->getName());
2945 else if (SI->getOpcode() == Instruction::AShr) {
2946 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2947 // Check to see if we are shifting out everything but the sign bit.
2948 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2949 SI->getType()->getPrimitiveSizeInBits()-1) {
2950 // Ok, the transformation is safe. Insert LShr.
2951 return BinaryOperator::CreateLShr(
2952 SI->getOperand(0), CU, SI->getName());
2959 // Try to fold constant sub into select arguments.
2960 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2961 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2964 if (isa<PHINode>(Op0))
2965 if (Instruction *NV = FoldOpIntoPhi(I))
2969 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2970 if (Op1I->getOpcode() == Instruction::Add &&
2971 !Op0->getType()->isFPOrFPVector()) {
2972 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2973 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2974 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2975 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2976 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2977 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2978 // C1-(X+C2) --> (C1-C2)-X
2979 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2980 Op1I->getOperand(0));
2984 if (Op1I->hasOneUse()) {
2985 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2986 // is not used by anyone else...
2988 if (Op1I->getOpcode() == Instruction::Sub &&
2989 !Op1I->getType()->isFPOrFPVector()) {
2990 // Swap the two operands of the subexpr...
2991 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2992 Op1I->setOperand(0, IIOp1);
2993 Op1I->setOperand(1, IIOp0);
2995 // Create the new top level add instruction...
2996 return BinaryOperator::CreateAdd(Op0, Op1);
2999 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
3001 if (Op1I->getOpcode() == Instruction::And &&
3002 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
3003 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
3006 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
3007 return BinaryOperator::CreateAnd(Op0, NewNot);
3010 // 0 - (X sdiv C) -> (X sdiv -C)
3011 if (Op1I->getOpcode() == Instruction::SDiv)
3012 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
3014 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
3015 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
3016 ConstantExpr::getNeg(DivRHS));
3018 // X - X*C --> X * (1-C)
3019 ConstantInt *C2 = 0;
3020 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
3021 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
3022 return BinaryOperator::CreateMul(Op0, CP1);
3025 // X - ((X / Y) * Y) --> X % Y
3026 if (Op1I->getOpcode() == Instruction::Mul)
3027 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
3028 if (Op0 == I->getOperand(0) &&
3029 Op1I->getOperand(1) == I->getOperand(1)) {
3030 if (I->getOpcode() == Instruction::SDiv)
3031 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
3032 if (I->getOpcode() == Instruction::UDiv)
3033 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
3038 if (!Op0->getType()->isFPOrFPVector())
3039 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3040 if (Op0I->getOpcode() == Instruction::Add) {
3041 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
3042 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3043 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
3044 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3045 } else if (Op0I->getOpcode() == Instruction::Sub) {
3046 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
3047 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
3052 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
3053 if (X == Op1) // X*C - X --> X * (C-1)
3054 return BinaryOperator::CreateMul(Op1, SubOne(C1));
3056 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
3057 if (X == dyn_castFoldableMul(Op1, C2))
3058 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
3063 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
3064 /// comparison only checks the sign bit. If it only checks the sign bit, set
3065 /// TrueIfSigned if the result of the comparison is true when the input value is
3067 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
3068 bool &TrueIfSigned) {
3070 case ICmpInst::ICMP_SLT: // True if LHS s< 0
3071 TrueIfSigned = true;
3072 return RHS->isZero();
3073 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
3074 TrueIfSigned = true;
3075 return RHS->isAllOnesValue();
3076 case ICmpInst::ICMP_SGT: // True if LHS s> -1
3077 TrueIfSigned = false;
3078 return RHS->isAllOnesValue();
3079 case ICmpInst::ICMP_UGT:
3080 // True if LHS u> RHS and RHS == high-bit-mask - 1
3081 TrueIfSigned = true;
3082 return RHS->getValue() ==
3083 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
3084 case ICmpInst::ICMP_UGE:
3085 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
3086 TrueIfSigned = true;
3087 return RHS->getValue() ==
3088 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
3094 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
3095 bool Changed = SimplifyCommutative(I);
3096 Value *Op0 = I.getOperand(0);
3098 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
3099 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3101 // Simplify mul instructions with a constant RHS...
3102 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
3103 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3105 // ((X << C1)*C2) == (X * (C2 << C1))
3106 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
3107 if (SI->getOpcode() == Instruction::Shl)
3108 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
3109 return BinaryOperator::CreateMul(SI->getOperand(0),
3110 ConstantExpr::getShl(CI, ShOp));
3113 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
3114 if (CI->equalsInt(1)) // X * 1 == X
3115 return ReplaceInstUsesWith(I, Op0);
3116 if (CI->isAllOnesValue()) // X * -1 == 0 - X
3117 return BinaryOperator::CreateNeg(Op0, I.getName());
3119 const APInt& Val = cast<ConstantInt>(CI)->getValue();
3120 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
3121 return BinaryOperator::CreateShl(Op0,
3122 ConstantInt::get(Op0->getType(), Val.logBase2()));
3124 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
3125 if (Op1F->isNullValue())
3126 return ReplaceInstUsesWith(I, Op1);
3128 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
3129 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
3130 // We need a better interface for long double here.
3131 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
3132 if (Op1F->isExactlyValue(1.0))
3133 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
3136 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3137 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
3138 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
3139 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
3140 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
3142 InsertNewInstBefore(Add, I);
3143 Value *C1C2 = ConstantExpr::getMul(Op1,
3144 cast<Constant>(Op0I->getOperand(1)));
3145 return BinaryOperator::CreateAdd(Add, C1C2);
3149 // Try to fold constant mul into select arguments.
3150 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3151 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3154 if (isa<PHINode>(Op0))
3155 if (Instruction *NV = FoldOpIntoPhi(I))
3159 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3160 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
3161 return BinaryOperator::CreateMul(Op0v, Op1v);
3163 // If one of the operands of the multiply is a cast from a boolean value, then
3164 // we know the bool is either zero or one, so this is a 'masking' multiply.
3165 // See if we can simplify things based on how the boolean was originally
3167 CastInst *BoolCast = 0;
3168 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
3169 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3172 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
3173 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3176 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
3177 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
3178 const Type *SCOpTy = SCIOp0->getType();
3181 // If the icmp is true iff the sign bit of X is set, then convert this
3182 // multiply into a shift/and combination.
3183 if (isa<ConstantInt>(SCIOp1) &&
3184 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
3186 // Shift the X value right to turn it into "all signbits".
3187 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
3188 SCOpTy->getPrimitiveSizeInBits()-1);
3190 InsertNewInstBefore(
3191 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
3192 BoolCast->getOperand(0)->getName()+
3195 // If the multiply type is not the same as the source type, sign extend
3196 // or truncate to the multiply type.
3197 if (I.getType() != V->getType()) {
3198 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
3199 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
3200 Instruction::CastOps opcode =
3201 (SrcBits == DstBits ? Instruction::BitCast :
3202 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
3203 V = InsertCastBefore(opcode, V, I.getType(), I);
3206 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
3207 return BinaryOperator::CreateAnd(V, OtherOp);
3212 return Changed ? &I : 0;
3215 /// This function implements the transforms on div instructions that work
3216 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3217 /// used by the visitors to those instructions.
3218 /// @brief Transforms common to all three div instructions
3219 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3220 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3222 // undef / X -> 0 for integer.
3223 // undef / X -> undef for FP (the undef could be a snan).
3224 if (isa<UndefValue>(Op0)) {
3225 if (Op0->getType()->isFPOrFPVector())
3226 return ReplaceInstUsesWith(I, Op0);
3227 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3230 // X / undef -> undef
3231 if (isa<UndefValue>(Op1))
3232 return ReplaceInstUsesWith(I, Op1);
3234 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3235 // This does not apply for fdiv.
3236 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3237 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
3238 // the same basic block, then we replace the select with Y, and the
3239 // condition of the select with false (if the cond value is in the same BB).
3240 // If the select has uses other than the div, this allows them to be
3241 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
3242 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
3243 if (ST->isNullValue()) {
3244 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3245 if (CondI && CondI->getParent() == I.getParent())
3246 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3247 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3248 I.setOperand(1, SI->getOperand(2));
3250 UpdateValueUsesWith(SI, SI->getOperand(2));
3254 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
3255 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
3256 if (ST->isNullValue()) {
3257 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3258 if (CondI && CondI->getParent() == I.getParent())
3259 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3260 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3261 I.setOperand(1, SI->getOperand(1));
3263 UpdateValueUsesWith(SI, SI->getOperand(1));
3271 /// This function implements the transforms common to both integer division
3272 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3273 /// division instructions.
3274 /// @brief Common integer divide transforms
3275 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3276 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3278 // (sdiv X, X) --> 1 (udiv X, X) --> 1
3280 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
3282 if (Instruction *Common = commonDivTransforms(I))
3285 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3287 if (RHS->equalsInt(1))
3288 return ReplaceInstUsesWith(I, Op0);
3290 // (X / C1) / C2 -> X / (C1*C2)
3291 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3292 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3293 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3294 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
3295 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3297 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3298 Multiply(RHS, LHSRHS));
3301 if (!RHS->isZero()) { // avoid X udiv 0
3302 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3303 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3305 if (isa<PHINode>(Op0))
3306 if (Instruction *NV = FoldOpIntoPhi(I))
3311 // 0 / X == 0, we don't need to preserve faults!
3312 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3313 if (LHS->equalsInt(0))
3314 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3319 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3320 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3322 // Handle the integer div common cases
3323 if (Instruction *Common = commonIDivTransforms(I))
3326 // X udiv C^2 -> X >> C
3327 // Check to see if this is an unsigned division with an exact power of 2,
3328 // if so, convert to a right shift.
3329 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3330 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3331 return BinaryOperator::CreateLShr(Op0,
3332 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3335 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3336 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3337 if (RHSI->getOpcode() == Instruction::Shl &&
3338 isa<ConstantInt>(RHSI->getOperand(0))) {
3339 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3340 if (C1.isPowerOf2()) {
3341 Value *N = RHSI->getOperand(1);
3342 const Type *NTy = N->getType();
3343 if (uint32_t C2 = C1.logBase2()) {
3344 Constant *C2V = ConstantInt::get(NTy, C2);
3345 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3347 return BinaryOperator::CreateLShr(Op0, N);
3352 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3353 // where C1&C2 are powers of two.
3354 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3355 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3356 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3357 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3358 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3359 // Compute the shift amounts
3360 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3361 // Construct the "on true" case of the select
3362 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3363 Instruction *TSI = BinaryOperator::CreateLShr(
3364 Op0, TC, SI->getName()+".t");
3365 TSI = InsertNewInstBefore(TSI, I);
3367 // Construct the "on false" case of the select
3368 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3369 Instruction *FSI = BinaryOperator::CreateLShr(
3370 Op0, FC, SI->getName()+".f");
3371 FSI = InsertNewInstBefore(FSI, I);
3373 // construct the select instruction and return it.
3374 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3380 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3381 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3383 // Handle the integer div common cases
3384 if (Instruction *Common = commonIDivTransforms(I))
3387 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3389 if (RHS->isAllOnesValue())
3390 return BinaryOperator::CreateNeg(Op0);
3393 if (Value *LHSNeg = dyn_castNegVal(Op0))
3394 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3397 // If the sign bits of both operands are zero (i.e. we can prove they are
3398 // unsigned inputs), turn this into a udiv.
3399 if (I.getType()->isInteger()) {
3400 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3401 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3402 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3403 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3410 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3411 return commonDivTransforms(I);
3414 /// This function implements the transforms on rem instructions that work
3415 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3416 /// is used by the visitors to those instructions.
3417 /// @brief Transforms common to all three rem instructions
3418 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3419 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3421 // 0 % X == 0 for integer, we don't need to preserve faults!
3422 if (Constant *LHS = dyn_cast<Constant>(Op0))
3423 if (LHS->isNullValue())
3424 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3426 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3427 if (I.getType()->isFPOrFPVector())
3428 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3429 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3431 if (isa<UndefValue>(Op1))
3432 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3434 // Handle cases involving: rem X, (select Cond, Y, Z)
3435 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3436 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3437 // the same basic block, then we replace the select with Y, and the
3438 // condition of the select with false (if the cond value is in the same
3439 // BB). If the select has uses other than the div, this allows them to be
3441 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3442 if (ST->isNullValue()) {
3443 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3444 if (CondI && CondI->getParent() == I.getParent())
3445 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3446 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3447 I.setOperand(1, SI->getOperand(2));
3449 UpdateValueUsesWith(SI, SI->getOperand(2));
3452 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3453 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3454 if (ST->isNullValue()) {
3455 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3456 if (CondI && CondI->getParent() == I.getParent())
3457 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3458 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3459 I.setOperand(1, SI->getOperand(1));
3461 UpdateValueUsesWith(SI, SI->getOperand(1));
3469 /// This function implements the transforms common to both integer remainder
3470 /// instructions (urem and srem). It is called by the visitors to those integer
3471 /// remainder instructions.
3472 /// @brief Common integer remainder transforms
3473 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3474 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3476 if (Instruction *common = commonRemTransforms(I))
3479 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3480 // X % 0 == undef, we don't need to preserve faults!
3481 if (RHS->equalsInt(0))
3482 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3484 if (RHS->equalsInt(1)) // X % 1 == 0
3485 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3487 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3488 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3489 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3491 } else if (isa<PHINode>(Op0I)) {
3492 if (Instruction *NV = FoldOpIntoPhi(I))
3496 // See if we can fold away this rem instruction.
3497 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3498 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3499 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3500 KnownZero, KnownOne))
3508 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3509 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3511 if (Instruction *common = commonIRemTransforms(I))
3514 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3515 // X urem C^2 -> X and C
3516 // Check to see if this is an unsigned remainder with an exact power of 2,
3517 // if so, convert to a bitwise and.
3518 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3519 if (C->getValue().isPowerOf2())
3520 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3523 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3524 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3525 if (RHSI->getOpcode() == Instruction::Shl &&
3526 isa<ConstantInt>(RHSI->getOperand(0))) {
3527 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3528 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3529 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3531 return BinaryOperator::CreateAnd(Op0, Add);
3536 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3537 // where C1&C2 are powers of two.
3538 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3539 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3540 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3541 // STO == 0 and SFO == 0 handled above.
3542 if ((STO->getValue().isPowerOf2()) &&
3543 (SFO->getValue().isPowerOf2())) {
3544 Value *TrueAnd = InsertNewInstBefore(
3545 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3546 Value *FalseAnd = InsertNewInstBefore(
3547 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3548 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3556 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3557 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3559 // Handle the integer rem common cases
3560 if (Instruction *common = commonIRemTransforms(I))
3563 if (Value *RHSNeg = dyn_castNegVal(Op1))
3564 if (!isa<ConstantInt>(RHSNeg) ||
3565 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
3567 AddUsesToWorkList(I);
3568 I.setOperand(1, RHSNeg);
3572 // If the sign bits of both operands are zero (i.e. we can prove they are
3573 // unsigned inputs), turn this into a urem.
3574 if (I.getType()->isInteger()) {
3575 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3576 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3577 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3578 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3585 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3586 return commonRemTransforms(I);
3589 // isMaxValueMinusOne - return true if this is Max-1
3590 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3591 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3593 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3594 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3597 // isMinValuePlusOne - return true if this is Min+1
3598 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3600 return C->getValue() == 1; // unsigned
3602 // Calculate 1111111111000000000000
3603 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3604 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3607 // isOneBitSet - Return true if there is exactly one bit set in the specified
3609 static bool isOneBitSet(const ConstantInt *CI) {
3610 return CI->getValue().isPowerOf2();
3613 // isHighOnes - Return true if the constant is of the form 1+0+.
3614 // This is the same as lowones(~X).
3615 static bool isHighOnes(const ConstantInt *CI) {
3616 return (~CI->getValue() + 1).isPowerOf2();
3619 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3620 /// are carefully arranged to allow folding of expressions such as:
3622 /// (A < B) | (A > B) --> (A != B)
3624 /// Note that this is only valid if the first and second predicates have the
3625 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3627 /// Three bits are used to represent the condition, as follows:
3632 /// <=> Value Definition
3633 /// 000 0 Always false
3640 /// 111 7 Always true
3642 static unsigned getICmpCode(const ICmpInst *ICI) {
3643 switch (ICI->getPredicate()) {
3645 case ICmpInst::ICMP_UGT: return 1; // 001
3646 case ICmpInst::ICMP_SGT: return 1; // 001
3647 case ICmpInst::ICMP_EQ: return 2; // 010
3648 case ICmpInst::ICMP_UGE: return 3; // 011
3649 case ICmpInst::ICMP_SGE: return 3; // 011
3650 case ICmpInst::ICMP_ULT: return 4; // 100
3651 case ICmpInst::ICMP_SLT: return 4; // 100
3652 case ICmpInst::ICMP_NE: return 5; // 101
3653 case ICmpInst::ICMP_ULE: return 6; // 110
3654 case ICmpInst::ICMP_SLE: return 6; // 110
3657 assert(0 && "Invalid ICmp predicate!");
3662 /// getICmpValue - This is the complement of getICmpCode, which turns an
3663 /// opcode and two operands into either a constant true or false, or a brand
3664 /// new ICmp instruction. The sign is passed in to determine which kind
3665 /// of predicate to use in new icmp instructions.
3666 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3668 default: assert(0 && "Illegal ICmp code!");
3669 case 0: return ConstantInt::getFalse();
3672 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3674 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3675 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3678 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3680 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3683 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3685 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3686 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3689 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3691 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3692 case 7: return ConstantInt::getTrue();
3696 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3697 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3698 (ICmpInst::isSignedPredicate(p1) &&
3699 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3700 (ICmpInst::isSignedPredicate(p2) &&
3701 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3705 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3706 struct FoldICmpLogical {
3709 ICmpInst::Predicate pred;
3710 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3711 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3712 pred(ICI->getPredicate()) {}
3713 bool shouldApply(Value *V) const {
3714 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3715 if (PredicatesFoldable(pred, ICI->getPredicate()))
3716 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3717 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3720 Instruction *apply(Instruction &Log) const {
3721 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3722 if (ICI->getOperand(0) != LHS) {
3723 assert(ICI->getOperand(1) == LHS);
3724 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3727 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3728 unsigned LHSCode = getICmpCode(ICI);
3729 unsigned RHSCode = getICmpCode(RHSICI);
3731 switch (Log.getOpcode()) {
3732 case Instruction::And: Code = LHSCode & RHSCode; break;
3733 case Instruction::Or: Code = LHSCode | RHSCode; break;
3734 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3735 default: assert(0 && "Illegal logical opcode!"); return 0;
3738 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3739 ICmpInst::isSignedPredicate(ICI->getPredicate());
3741 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3742 if (Instruction *I = dyn_cast<Instruction>(RV))
3744 // Otherwise, it's a constant boolean value...
3745 return IC.ReplaceInstUsesWith(Log, RV);
3748 } // end anonymous namespace
3750 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3751 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3752 // guaranteed to be a binary operator.
3753 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3755 ConstantInt *AndRHS,
3756 BinaryOperator &TheAnd) {
3757 Value *X = Op->getOperand(0);
3758 Constant *Together = 0;
3760 Together = And(AndRHS, OpRHS);
3762 switch (Op->getOpcode()) {
3763 case Instruction::Xor:
3764 if (Op->hasOneUse()) {
3765 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3766 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3767 InsertNewInstBefore(And, TheAnd);
3769 return BinaryOperator::CreateXor(And, Together);
3772 case Instruction::Or:
3773 if (Together == AndRHS) // (X | C) & C --> C
3774 return ReplaceInstUsesWith(TheAnd, AndRHS);
3776 if (Op->hasOneUse() && Together != OpRHS) {
3777 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3778 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3779 InsertNewInstBefore(Or, TheAnd);
3781 return BinaryOperator::CreateAnd(Or, AndRHS);
3784 case Instruction::Add:
3785 if (Op->hasOneUse()) {
3786 // Adding a one to a single bit bit-field should be turned into an XOR
3787 // of the bit. First thing to check is to see if this AND is with a
3788 // single bit constant.
3789 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3791 // If there is only one bit set...
3792 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3793 // Ok, at this point, we know that we are masking the result of the
3794 // ADD down to exactly one bit. If the constant we are adding has
3795 // no bits set below this bit, then we can eliminate the ADD.
3796 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3798 // Check to see if any bits below the one bit set in AndRHSV are set.
3799 if ((AddRHS & (AndRHSV-1)) == 0) {
3800 // If not, the only thing that can effect the output of the AND is
3801 // the bit specified by AndRHSV. If that bit is set, the effect of
3802 // the XOR is to toggle the bit. If it is clear, then the ADD has
3804 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3805 TheAnd.setOperand(0, X);
3808 // Pull the XOR out of the AND.
3809 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3810 InsertNewInstBefore(NewAnd, TheAnd);
3811 NewAnd->takeName(Op);
3812 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3819 case Instruction::Shl: {
3820 // We know that the AND will not produce any of the bits shifted in, so if
3821 // the anded constant includes them, clear them now!
3823 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3824 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3825 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3826 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3828 if (CI->getValue() == ShlMask) {
3829 // Masking out bits that the shift already masks
3830 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3831 } else if (CI != AndRHS) { // Reducing bits set in and.
3832 TheAnd.setOperand(1, CI);
3837 case Instruction::LShr:
3839 // We know that the AND will not produce any of the bits shifted in, so if
3840 // the anded constant includes them, clear them now! This only applies to
3841 // unsigned shifts, because a signed shr may bring in set bits!
3843 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3844 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3845 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3846 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3848 if (CI->getValue() == ShrMask) {
3849 // Masking out bits that the shift already masks.
3850 return ReplaceInstUsesWith(TheAnd, Op);
3851 } else if (CI != AndRHS) {
3852 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3857 case Instruction::AShr:
3859 // See if this is shifting in some sign extension, then masking it out
3861 if (Op->hasOneUse()) {
3862 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3863 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3864 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3865 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3866 if (C == AndRHS) { // Masking out bits shifted in.
3867 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3868 // Make the argument unsigned.
3869 Value *ShVal = Op->getOperand(0);
3870 ShVal = InsertNewInstBefore(
3871 BinaryOperator::CreateLShr(ShVal, OpRHS,
3872 Op->getName()), TheAnd);
3873 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3882 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3883 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3884 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3885 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3886 /// insert new instructions.
3887 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3888 bool isSigned, bool Inside,
3890 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3891 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3892 "Lo is not <= Hi in range emission code!");
3895 if (Lo == Hi) // Trivially false.
3896 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3898 // V >= Min && V < Hi --> V < Hi
3899 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3900 ICmpInst::Predicate pred = (isSigned ?
3901 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3902 return new ICmpInst(pred, V, Hi);
3905 // Emit V-Lo <u Hi-Lo
3906 Constant *NegLo = ConstantExpr::getNeg(Lo);
3907 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3908 InsertNewInstBefore(Add, IB);
3909 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3910 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3913 if (Lo == Hi) // Trivially true.
3914 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3916 // V < Min || V >= Hi -> V > Hi-1
3917 Hi = SubOne(cast<ConstantInt>(Hi));
3918 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3919 ICmpInst::Predicate pred = (isSigned ?
3920 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3921 return new ICmpInst(pred, V, Hi);
3924 // Emit V-Lo >u Hi-1-Lo
3925 // Note that Hi has already had one subtracted from it, above.
3926 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3927 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3928 InsertNewInstBefore(Add, IB);
3929 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3930 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3933 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3934 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3935 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3936 // not, since all 1s are not contiguous.
3937 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3938 const APInt& V = Val->getValue();
3939 uint32_t BitWidth = Val->getType()->getBitWidth();
3940 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3942 // look for the first zero bit after the run of ones
3943 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3944 // look for the first non-zero bit
3945 ME = V.getActiveBits();
3949 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3950 /// where isSub determines whether the operator is a sub. If we can fold one of
3951 /// the following xforms:
3953 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3954 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3955 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3957 /// return (A +/- B).
3959 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3960 ConstantInt *Mask, bool isSub,
3962 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3963 if (!LHSI || LHSI->getNumOperands() != 2 ||
3964 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3966 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3968 switch (LHSI->getOpcode()) {
3970 case Instruction::And:
3971 if (And(N, Mask) == Mask) {
3972 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3973 if ((Mask->getValue().countLeadingZeros() +
3974 Mask->getValue().countPopulation()) ==
3975 Mask->getValue().getBitWidth())
3978 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3979 // part, we don't need any explicit masks to take them out of A. If that
3980 // is all N is, ignore it.
3981 uint32_t MB = 0, ME = 0;
3982 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3983 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3984 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3985 if (MaskedValueIsZero(RHS, Mask))
3990 case Instruction::Or:
3991 case Instruction::Xor:
3992 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3993 if ((Mask->getValue().countLeadingZeros() +
3994 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3995 && And(N, Mask)->isZero())
4002 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
4004 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
4005 return InsertNewInstBefore(New, I);
4008 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4009 bool Changed = SimplifyCommutative(I);
4010 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4012 if (isa<UndefValue>(Op1)) // X & undef -> 0
4013 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4017 return ReplaceInstUsesWith(I, Op1);
4019 // See if we can simplify any instructions used by the instruction whose sole
4020 // purpose is to compute bits we don't care about.
4021 if (!isa<VectorType>(I.getType())) {
4022 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4023 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4024 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4025 KnownZero, KnownOne))
4028 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4029 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4030 return ReplaceInstUsesWith(I, I.getOperand(0));
4031 } else if (isa<ConstantAggregateZero>(Op1)) {
4032 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4036 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4037 const APInt& AndRHSMask = AndRHS->getValue();
4038 APInt NotAndRHS(~AndRHSMask);
4040 // Optimize a variety of ((val OP C1) & C2) combinations...
4041 if (isa<BinaryOperator>(Op0)) {
4042 Instruction *Op0I = cast<Instruction>(Op0);
4043 Value *Op0LHS = Op0I->getOperand(0);
4044 Value *Op0RHS = Op0I->getOperand(1);
4045 switch (Op0I->getOpcode()) {
4046 case Instruction::Xor:
4047 case Instruction::Or:
4048 // If the mask is only needed on one incoming arm, push it up.
4049 if (Op0I->hasOneUse()) {
4050 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4051 // Not masking anything out for the LHS, move to RHS.
4052 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4053 Op0RHS->getName()+".masked");
4054 InsertNewInstBefore(NewRHS, I);
4055 return BinaryOperator::Create(
4056 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4058 if (!isa<Constant>(Op0RHS) &&
4059 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4060 // Not masking anything out for the RHS, move to LHS.
4061 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4062 Op0LHS->getName()+".masked");
4063 InsertNewInstBefore(NewLHS, I);
4064 return BinaryOperator::Create(
4065 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4070 case Instruction::Add:
4071 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4072 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4073 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4074 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4075 return BinaryOperator::CreateAnd(V, AndRHS);
4076 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4077 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4080 case Instruction::Sub:
4081 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4082 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4083 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4084 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4085 return BinaryOperator::CreateAnd(V, AndRHS);
4089 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4090 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4092 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4093 // If this is an integer truncation or change from signed-to-unsigned, and
4094 // if the source is an and/or with immediate, transform it. This
4095 // frequently occurs for bitfield accesses.
4096 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4097 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4098 CastOp->getNumOperands() == 2)
4099 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4100 if (CastOp->getOpcode() == Instruction::And) {
4101 // Change: and (cast (and X, C1) to T), C2
4102 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4103 // This will fold the two constants together, which may allow
4104 // other simplifications.
4105 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4106 CastOp->getOperand(0), I.getType(),
4107 CastOp->getName()+".shrunk");
4108 NewCast = InsertNewInstBefore(NewCast, I);
4109 // trunc_or_bitcast(C1)&C2
4110 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4111 C3 = ConstantExpr::getAnd(C3, AndRHS);
4112 return BinaryOperator::CreateAnd(NewCast, C3);
4113 } else if (CastOp->getOpcode() == Instruction::Or) {
4114 // Change: and (cast (or X, C1) to T), C2
4115 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4116 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4117 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
4118 return ReplaceInstUsesWith(I, AndRHS);
4124 // Try to fold constant and into select arguments.
4125 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4126 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4128 if (isa<PHINode>(Op0))
4129 if (Instruction *NV = FoldOpIntoPhi(I))
4133 Value *Op0NotVal = dyn_castNotVal(Op0);
4134 Value *Op1NotVal = dyn_castNotVal(Op1);
4136 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4137 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4139 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4140 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4141 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4142 I.getName()+".demorgan");
4143 InsertNewInstBefore(Or, I);
4144 return BinaryOperator::CreateNot(Or);
4148 Value *A = 0, *B = 0, *C = 0, *D = 0;
4149 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4150 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4151 return ReplaceInstUsesWith(I, Op1);
4153 // (A|B) & ~(A&B) -> A^B
4154 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4155 if ((A == C && B == D) || (A == D && B == C))
4156 return BinaryOperator::CreateXor(A, B);
4160 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4161 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4162 return ReplaceInstUsesWith(I, Op0);
4164 // ~(A&B) & (A|B) -> A^B
4165 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4166 if ((A == C && B == D) || (A == D && B == C))
4167 return BinaryOperator::CreateXor(A, B);
4171 if (Op0->hasOneUse() &&
4172 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4173 if (A == Op1) { // (A^B)&A -> A&(A^B)
4174 I.swapOperands(); // Simplify below
4175 std::swap(Op0, Op1);
4176 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4177 cast<BinaryOperator>(Op0)->swapOperands();
4178 I.swapOperands(); // Simplify below
4179 std::swap(Op0, Op1);
4182 if (Op1->hasOneUse() &&
4183 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4184 if (B == Op0) { // B&(A^B) -> B&(B^A)
4185 cast<BinaryOperator>(Op1)->swapOperands();
4188 if (A == Op0) { // A&(A^B) -> A & ~B
4189 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4190 InsertNewInstBefore(NotB, I);
4191 return BinaryOperator::CreateAnd(A, NotB);
4196 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4197 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4198 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4201 Value *LHSVal, *RHSVal;
4202 ConstantInt *LHSCst, *RHSCst;
4203 ICmpInst::Predicate LHSCC, RHSCC;
4204 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4205 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4206 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
4207 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
4208 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4209 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4210 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4211 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4213 // Don't try to fold ICMP_SLT + ICMP_ULT.
4214 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
4215 ICmpInst::isSignedPredicate(LHSCC) ==
4216 ICmpInst::isSignedPredicate(RHSCC))) {
4217 // Ensure that the larger constant is on the RHS.
4218 ICmpInst::Predicate GT;
4219 if (ICmpInst::isSignedPredicate(LHSCC) ||
4220 (ICmpInst::isEquality(LHSCC) &&
4221 ICmpInst::isSignedPredicate(RHSCC)))
4222 GT = ICmpInst::ICMP_SGT;
4224 GT = ICmpInst::ICMP_UGT;
4226 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4227 ICmpInst *LHS = cast<ICmpInst>(Op0);
4228 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4229 std::swap(LHS, RHS);
4230 std::swap(LHSCst, RHSCst);
4231 std::swap(LHSCC, RHSCC);
4234 // At this point, we know we have have two icmp instructions
4235 // comparing a value against two constants and and'ing the result
4236 // together. Because of the above check, we know that we only have
4237 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4238 // (from the FoldICmpLogical check above), that the two constants
4239 // are not equal and that the larger constant is on the RHS
4240 assert(LHSCst != RHSCst && "Compares not folded above?");
4243 default: assert(0 && "Unknown integer condition code!");
4244 case ICmpInst::ICMP_EQ:
4246 default: assert(0 && "Unknown integer condition code!");
4247 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4248 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4249 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4250 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4251 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4252 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4253 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4254 return ReplaceInstUsesWith(I, LHS);
4256 case ICmpInst::ICMP_NE:
4258 default: assert(0 && "Unknown integer condition code!");
4259 case ICmpInst::ICMP_ULT:
4260 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4261 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
4262 break; // (X != 13 & X u< 15) -> no change
4263 case ICmpInst::ICMP_SLT:
4264 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4265 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
4266 break; // (X != 13 & X s< 15) -> no change
4267 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4268 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4269 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4270 return ReplaceInstUsesWith(I, RHS);
4271 case ICmpInst::ICMP_NE:
4272 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4273 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4274 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4275 LHSVal->getName()+".off");
4276 InsertNewInstBefore(Add, I);
4277 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4278 ConstantInt::get(Add->getType(), 1));
4280 break; // (X != 13 & X != 15) -> no change
4283 case ICmpInst::ICMP_ULT:
4285 default: assert(0 && "Unknown integer condition code!");
4286 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4287 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4288 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4289 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4291 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4292 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4293 return ReplaceInstUsesWith(I, LHS);
4294 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4298 case ICmpInst::ICMP_SLT:
4300 default: assert(0 && "Unknown integer condition code!");
4301 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4302 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4303 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4304 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4306 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4307 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4308 return ReplaceInstUsesWith(I, LHS);
4309 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4313 case ICmpInst::ICMP_UGT:
4315 default: assert(0 && "Unknown integer condition code!");
4316 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
4317 return ReplaceInstUsesWith(I, LHS);
4318 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4319 return ReplaceInstUsesWith(I, RHS);
4320 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4322 case ICmpInst::ICMP_NE:
4323 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4324 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4325 break; // (X u> 13 & X != 15) -> no change
4326 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
4327 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
4329 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4333 case ICmpInst::ICMP_SGT:
4335 default: assert(0 && "Unknown integer condition code!");
4336 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4337 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4338 return ReplaceInstUsesWith(I, RHS);
4339 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4341 case ICmpInst::ICMP_NE:
4342 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4343 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4344 break; // (X s> 13 & X != 15) -> no change
4345 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4346 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4348 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4356 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4357 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4358 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4359 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4360 const Type *SrcTy = Op0C->getOperand(0)->getType();
4361 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4362 // Only do this if the casts both really cause code to be generated.
4363 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4365 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4367 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4368 Op1C->getOperand(0),
4370 InsertNewInstBefore(NewOp, I);
4371 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4375 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4376 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4377 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4378 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4379 SI0->getOperand(1) == SI1->getOperand(1) &&
4380 (SI0->hasOneUse() || SI1->hasOneUse())) {
4381 Instruction *NewOp =
4382 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4384 SI0->getName()), I);
4385 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4386 SI1->getOperand(1));
4390 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4391 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4392 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4393 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4394 RHS->getPredicate() == FCmpInst::FCMP_ORD)
4395 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4396 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4397 // If either of the constants are nans, then the whole thing returns
4399 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4400 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4401 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4402 RHS->getOperand(0));
4407 return Changed ? &I : 0;
4410 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4411 /// in the result. If it does, and if the specified byte hasn't been filled in
4412 /// yet, fill it in and return false.
4413 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4414 Instruction *I = dyn_cast<Instruction>(V);
4415 if (I == 0) return true;
4417 // If this is an or instruction, it is an inner node of the bswap.
4418 if (I->getOpcode() == Instruction::Or)
4419 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4420 CollectBSwapParts(I->getOperand(1), ByteValues);
4422 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
4423 // If this is a shift by a constant int, and it is "24", then its operand
4424 // defines a byte. We only handle unsigned types here.
4425 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4426 // Not shifting the entire input by N-1 bytes?
4427 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
4428 8*(ByteValues.size()-1))
4432 if (I->getOpcode() == Instruction::Shl) {
4433 // X << 24 defines the top byte with the lowest of the input bytes.
4434 DestNo = ByteValues.size()-1;
4436 // X >>u 24 defines the low byte with the highest of the input bytes.
4440 // If the destination byte value is already defined, the values are or'd
4441 // together, which isn't a bswap (unless it's an or of the same bits).
4442 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4444 ByteValues[DestNo] = I->getOperand(0);
4448 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4450 Value *Shift = 0, *ShiftLHS = 0;
4451 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4452 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4453 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4455 Instruction *SI = cast<Instruction>(Shift);
4457 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4458 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
4459 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
4462 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4464 if (AndAmt->getValue().getActiveBits() > 64)
4466 uint64_t AndAmtVal = AndAmt->getZExtValue();
4467 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4468 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
4470 // Unknown mask for bswap.
4471 if (DestByte == ByteValues.size()) return true;
4473 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4475 if (SI->getOpcode() == Instruction::Shl)
4476 SrcByte = DestByte - ShiftBytes;
4478 SrcByte = DestByte + ShiftBytes;
4480 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4481 if (SrcByte != ByteValues.size()-DestByte-1)
4484 // If the destination byte value is already defined, the values are or'd
4485 // together, which isn't a bswap (unless it's an or of the same bits).
4486 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4488 ByteValues[DestByte] = SI->getOperand(0);
4492 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4493 /// If so, insert the new bswap intrinsic and return it.
4494 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4495 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4496 if (!ITy || ITy->getBitWidth() % 16)
4497 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4499 /// ByteValues - For each byte of the result, we keep track of which value
4500 /// defines each byte.
4501 SmallVector<Value*, 8> ByteValues;
4502 ByteValues.resize(ITy->getBitWidth()/8);
4504 // Try to find all the pieces corresponding to the bswap.
4505 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4506 CollectBSwapParts(I.getOperand(1), ByteValues))
4509 // Check to see if all of the bytes come from the same value.
4510 Value *V = ByteValues[0];
4511 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4513 // Check to make sure that all of the bytes come from the same value.
4514 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4515 if (ByteValues[i] != V)
4517 const Type *Tys[] = { ITy };
4518 Module *M = I.getParent()->getParent()->getParent();
4519 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4520 return CallInst::Create(F, V);
4524 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4525 bool Changed = SimplifyCommutative(I);
4526 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4528 if (isa<UndefValue>(Op1)) // X | undef -> -1
4529 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4533 return ReplaceInstUsesWith(I, Op0);
4535 // See if we can simplify any instructions used by the instruction whose sole
4536 // purpose is to compute bits we don't care about.
4537 if (!isa<VectorType>(I.getType())) {
4538 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4539 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4540 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4541 KnownZero, KnownOne))
4543 } else if (isa<ConstantAggregateZero>(Op1)) {
4544 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4545 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4546 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4547 return ReplaceInstUsesWith(I, I.getOperand(1));
4553 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4554 ConstantInt *C1 = 0; Value *X = 0;
4555 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4556 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4557 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4558 InsertNewInstBefore(Or, I);
4560 return BinaryOperator::CreateAnd(Or,
4561 ConstantInt::get(RHS->getValue() | C1->getValue()));
4564 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4565 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4566 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4567 InsertNewInstBefore(Or, I);
4569 return BinaryOperator::CreateXor(Or,
4570 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4573 // Try to fold constant and into select arguments.
4574 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4575 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4577 if (isa<PHINode>(Op0))
4578 if (Instruction *NV = FoldOpIntoPhi(I))
4582 Value *A = 0, *B = 0;
4583 ConstantInt *C1 = 0, *C2 = 0;
4585 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4586 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4587 return ReplaceInstUsesWith(I, Op1);
4588 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4589 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4590 return ReplaceInstUsesWith(I, Op0);
4592 // (A | B) | C and A | (B | C) -> bswap if possible.
4593 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4594 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4595 match(Op1, m_Or(m_Value(), m_Value())) ||
4596 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4597 match(Op1, m_Shift(m_Value(), m_Value())))) {
4598 if (Instruction *BSwap = MatchBSwap(I))
4602 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4603 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4604 MaskedValueIsZero(Op1, C1->getValue())) {
4605 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4606 InsertNewInstBefore(NOr, I);
4608 return BinaryOperator::CreateXor(NOr, C1);
4611 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4612 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4613 MaskedValueIsZero(Op0, C1->getValue())) {
4614 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4615 InsertNewInstBefore(NOr, I);
4617 return BinaryOperator::CreateXor(NOr, C1);
4621 Value *C = 0, *D = 0;
4622 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4623 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4624 Value *V1 = 0, *V2 = 0, *V3 = 0;
4625 C1 = dyn_cast<ConstantInt>(C);
4626 C2 = dyn_cast<ConstantInt>(D);
4627 if (C1 && C2) { // (A & C1)|(B & C2)
4628 // If we have: ((V + N) & C1) | (V & C2)
4629 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4630 // replace with V+N.
4631 if (C1->getValue() == ~C2->getValue()) {
4632 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4633 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4634 // Add commutes, try both ways.
4635 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4636 return ReplaceInstUsesWith(I, A);
4637 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4638 return ReplaceInstUsesWith(I, A);
4640 // Or commutes, try both ways.
4641 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4642 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4643 // Add commutes, try both ways.
4644 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4645 return ReplaceInstUsesWith(I, B);
4646 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4647 return ReplaceInstUsesWith(I, B);
4650 V1 = 0; V2 = 0; V3 = 0;
4653 // Check to see if we have any common things being and'ed. If so, find the
4654 // terms for V1 & (V2|V3).
4655 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4656 if (A == B) // (A & C)|(A & D) == A & (C|D)
4657 V1 = A, V2 = C, V3 = D;
4658 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4659 V1 = A, V2 = B, V3 = C;
4660 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4661 V1 = C, V2 = A, V3 = D;
4662 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4663 V1 = C, V2 = A, V3 = B;
4667 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4668 return BinaryOperator::CreateAnd(V1, Or);
4673 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4674 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4675 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4676 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4677 SI0->getOperand(1) == SI1->getOperand(1) &&
4678 (SI0->hasOneUse() || SI1->hasOneUse())) {
4679 Instruction *NewOp =
4680 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4682 SI0->getName()), I);
4683 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4684 SI1->getOperand(1));
4688 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4689 if (A == Op1) // ~A | A == -1
4690 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4694 // Note, A is still live here!
4695 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4697 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4699 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4700 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4701 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4702 I.getName()+".demorgan"), I);
4703 return BinaryOperator::CreateNot(And);
4707 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4708 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4709 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4712 Value *LHSVal, *RHSVal;
4713 ConstantInt *LHSCst, *RHSCst;
4714 ICmpInst::Predicate LHSCC, RHSCC;
4715 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4716 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4717 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4718 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4719 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4720 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4721 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4722 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4723 // We can't fold (ugt x, C) | (sgt x, C2).
4724 PredicatesFoldable(LHSCC, RHSCC)) {
4725 // Ensure that the larger constant is on the RHS.
4726 ICmpInst *LHS = cast<ICmpInst>(Op0);
4728 if (ICmpInst::isSignedPredicate(LHSCC))
4729 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4731 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4734 std::swap(LHS, RHS);
4735 std::swap(LHSCst, RHSCst);
4736 std::swap(LHSCC, RHSCC);
4739 // At this point, we know we have have two icmp instructions
4740 // comparing a value against two constants and or'ing the result
4741 // together. Because of the above check, we know that we only have
4742 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4743 // FoldICmpLogical check above), that the two constants are not
4745 assert(LHSCst != RHSCst && "Compares not folded above?");
4748 default: assert(0 && "Unknown integer condition code!");
4749 case ICmpInst::ICMP_EQ:
4751 default: assert(0 && "Unknown integer condition code!");
4752 case ICmpInst::ICMP_EQ:
4753 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4754 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4755 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4756 LHSVal->getName()+".off");
4757 InsertNewInstBefore(Add, I);
4758 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4759 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4761 break; // (X == 13 | X == 15) -> no change
4762 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4763 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4765 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4766 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4767 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4768 return ReplaceInstUsesWith(I, RHS);
4771 case ICmpInst::ICMP_NE:
4773 default: assert(0 && "Unknown integer condition code!");
4774 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4775 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4776 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4777 return ReplaceInstUsesWith(I, LHS);
4778 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4779 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4780 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4781 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4784 case ICmpInst::ICMP_ULT:
4786 default: assert(0 && "Unknown integer condition code!");
4787 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4789 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4790 // If RHSCst is [us]MAXINT, it is always false. Not handling
4791 // this can cause overflow.
4792 if (RHSCst->isMaxValue(false))
4793 return ReplaceInstUsesWith(I, LHS);
4794 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4796 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4798 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4799 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4800 return ReplaceInstUsesWith(I, RHS);
4801 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4805 case ICmpInst::ICMP_SLT:
4807 default: assert(0 && "Unknown integer condition code!");
4808 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4810 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4811 // If RHSCst is [us]MAXINT, it is always false. Not handling
4812 // this can cause overflow.
4813 if (RHSCst->isMaxValue(true))
4814 return ReplaceInstUsesWith(I, LHS);
4815 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4817 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4819 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4820 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4821 return ReplaceInstUsesWith(I, RHS);
4822 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4826 case ICmpInst::ICMP_UGT:
4828 default: assert(0 && "Unknown integer condition code!");
4829 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4830 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4831 return ReplaceInstUsesWith(I, LHS);
4832 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4834 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4835 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4836 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4837 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4841 case ICmpInst::ICMP_SGT:
4843 default: assert(0 && "Unknown integer condition code!");
4844 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4845 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4846 return ReplaceInstUsesWith(I, LHS);
4847 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4849 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4850 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4851 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4852 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4860 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4861 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4862 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4863 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4864 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4865 !isa<ICmpInst>(Op1C->getOperand(0))) {
4866 const Type *SrcTy = Op0C->getOperand(0)->getType();
4867 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4868 // Only do this if the casts both really cause code to be
4870 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4872 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4874 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4875 Op1C->getOperand(0),
4877 InsertNewInstBefore(NewOp, I);
4878 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4885 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4886 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4887 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4888 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4889 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4890 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4891 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4892 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4893 // If either of the constants are nans, then the whole thing returns
4895 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4896 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4898 // Otherwise, no need to compare the two constants, compare the
4900 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4901 RHS->getOperand(0));
4906 return Changed ? &I : 0;
4911 // XorSelf - Implements: X ^ X --> 0
4914 XorSelf(Value *rhs) : RHS(rhs) {}
4915 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4916 Instruction *apply(BinaryOperator &Xor) const {
4923 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4924 bool Changed = SimplifyCommutative(I);
4925 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4927 if (isa<UndefValue>(Op1)) {
4928 if (isa<UndefValue>(Op0))
4929 // Handle undef ^ undef -> 0 special case. This is a common
4931 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4932 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4935 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4936 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4937 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4938 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4941 // See if we can simplify any instructions used by the instruction whose sole
4942 // purpose is to compute bits we don't care about.
4943 if (!isa<VectorType>(I.getType())) {
4944 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4945 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4946 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4947 KnownZero, KnownOne))
4949 } else if (isa<ConstantAggregateZero>(Op1)) {
4950 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4953 // Is this a ~ operation?
4954 if (Value *NotOp = dyn_castNotVal(&I)) {
4955 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4956 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4957 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4958 if (Op0I->getOpcode() == Instruction::And ||
4959 Op0I->getOpcode() == Instruction::Or) {
4960 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4961 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4963 BinaryOperator::CreateNot(Op0I->getOperand(1),
4964 Op0I->getOperand(1)->getName()+".not");
4965 InsertNewInstBefore(NotY, I);
4966 if (Op0I->getOpcode() == Instruction::And)
4967 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4969 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4976 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4977 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4978 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4979 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4980 return new ICmpInst(ICI->getInversePredicate(),
4981 ICI->getOperand(0), ICI->getOperand(1));
4983 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4984 return new FCmpInst(FCI->getInversePredicate(),
4985 FCI->getOperand(0), FCI->getOperand(1));
4988 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4989 // ~(c-X) == X-c-1 == X+(-c-1)
4990 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4991 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4992 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4993 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4994 ConstantInt::get(I.getType(), 1));
4995 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4998 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4999 if (Op0I->getOpcode() == Instruction::Add) {
5000 // ~(X-c) --> (-c-1)-X
5001 if (RHS->isAllOnesValue()) {
5002 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5003 return BinaryOperator::CreateSub(
5004 ConstantExpr::getSub(NegOp0CI,
5005 ConstantInt::get(I.getType(), 1)),
5006 Op0I->getOperand(0));
5007 } else if (RHS->getValue().isSignBit()) {
5008 // (X + C) ^ signbit -> (X + C + signbit)
5009 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
5010 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5013 } else if (Op0I->getOpcode() == Instruction::Or) {
5014 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5015 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5016 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5017 // Anything in both C1 and C2 is known to be zero, remove it from
5019 Constant *CommonBits = And(Op0CI, RHS);
5020 NewRHS = ConstantExpr::getAnd(NewRHS,
5021 ConstantExpr::getNot(CommonBits));
5022 AddToWorkList(Op0I);
5023 I.setOperand(0, Op0I->getOperand(0));
5024 I.setOperand(1, NewRHS);
5031 // Try to fold constant and into select arguments.
5032 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5033 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5035 if (isa<PHINode>(Op0))
5036 if (Instruction *NV = FoldOpIntoPhi(I))
5040 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5042 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5044 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5046 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5049 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5052 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5053 if (A == Op0) { // B^(B|A) == (A|B)^B
5054 Op1I->swapOperands();
5056 std::swap(Op0, Op1);
5057 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5058 I.swapOperands(); // Simplified below.
5059 std::swap(Op0, Op1);
5061 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
5062 if (Op0 == A) // A^(A^B) == B
5063 return ReplaceInstUsesWith(I, B);
5064 else if (Op0 == B) // A^(B^A) == B
5065 return ReplaceInstUsesWith(I, A);
5066 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5067 if (A == Op0) { // A^(A&B) -> A^(B&A)
5068 Op1I->swapOperands();
5071 if (B == Op0) { // A^(B&A) -> (B&A)^A
5072 I.swapOperands(); // Simplified below.
5073 std::swap(Op0, Op1);
5078 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5081 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5082 if (A == Op1) // (B|A)^B == (A|B)^B
5084 if (B == Op1) { // (A|B)^B == A & ~B
5086 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5087 return BinaryOperator::CreateAnd(A, NotB);
5089 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
5090 if (Op1 == A) // (A^B)^A == B
5091 return ReplaceInstUsesWith(I, B);
5092 else if (Op1 == B) // (B^A)^A == B
5093 return ReplaceInstUsesWith(I, A);
5094 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5095 if (A == Op1) // (A&B)^A -> (B&A)^A
5097 if (B == Op1 && // (B&A)^A == ~B & A
5098 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5100 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5101 return BinaryOperator::CreateAnd(N, Op1);
5106 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5107 if (Op0I && Op1I && Op0I->isShift() &&
5108 Op0I->getOpcode() == Op1I->getOpcode() &&
5109 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5110 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5111 Instruction *NewOp =
5112 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5113 Op1I->getOperand(0),
5114 Op0I->getName()), I);
5115 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5116 Op1I->getOperand(1));
5120 Value *A, *B, *C, *D;
5121 // (A & B)^(A | B) -> A ^ B
5122 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5123 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5124 if ((A == C && B == D) || (A == D && B == C))
5125 return BinaryOperator::CreateXor(A, B);
5127 // (A | B)^(A & B) -> A ^ B
5128 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5129 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5130 if ((A == C && B == D) || (A == D && B == C))
5131 return BinaryOperator::CreateXor(A, B);
5135 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5136 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5137 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5138 // (X & Y)^(X & Y) -> (Y^Z) & X
5139 Value *X = 0, *Y = 0, *Z = 0;
5141 X = A, Y = B, Z = D;
5143 X = A, Y = B, Z = C;
5145 X = B, Y = A, Z = D;
5147 X = B, Y = A, Z = C;
5150 Instruction *NewOp =
5151 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5152 return BinaryOperator::CreateAnd(NewOp, X);
5157 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5158 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5159 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5162 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5163 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5164 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5165 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5166 const Type *SrcTy = Op0C->getOperand(0)->getType();
5167 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5168 // Only do this if the casts both really cause code to be generated.
5169 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5171 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5173 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5174 Op1C->getOperand(0),
5176 InsertNewInstBefore(NewOp, I);
5177 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5181 return Changed ? &I : 0;
5184 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5185 /// overflowed for this type.
5186 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5187 ConstantInt *In2, bool IsSigned = false) {
5188 Result = cast<ConstantInt>(Add(In1, In2));
5191 if (In2->getValue().isNegative())
5192 return Result->getValue().sgt(In1->getValue());
5194 return Result->getValue().slt(In1->getValue());
5196 return Result->getValue().ult(In1->getValue());
5199 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5200 /// code necessary to compute the offset from the base pointer (without adding
5201 /// in the base pointer). Return the result as a signed integer of intptr size.
5202 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5203 TargetData &TD = IC.getTargetData();
5204 gep_type_iterator GTI = gep_type_begin(GEP);
5205 const Type *IntPtrTy = TD.getIntPtrType();
5206 Value *Result = Constant::getNullValue(IntPtrTy);
5208 // Build a mask for high order bits.
5209 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5210 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5212 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
5213 Value *Op = GEP->getOperand(i);
5214 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5215 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5216 if (OpC->isZero()) continue;
5218 // Handle a struct index, which adds its field offset to the pointer.
5219 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5220 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5222 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5223 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5225 Result = IC.InsertNewInstBefore(
5226 BinaryOperator::CreateAdd(Result,
5227 ConstantInt::get(IntPtrTy, Size),
5228 GEP->getName()+".offs"), I);
5232 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5233 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5234 Scale = ConstantExpr::getMul(OC, Scale);
5235 if (Constant *RC = dyn_cast<Constant>(Result))
5236 Result = ConstantExpr::getAdd(RC, Scale);
5238 // Emit an add instruction.
5239 Result = IC.InsertNewInstBefore(
5240 BinaryOperator::CreateAdd(Result, Scale,
5241 GEP->getName()+".offs"), I);
5245 // Convert to correct type.
5246 if (Op->getType() != IntPtrTy) {
5247 if (Constant *OpC = dyn_cast<Constant>(Op))
5248 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5250 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5251 Op->getName()+".c"), I);
5254 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5255 if (Constant *OpC = dyn_cast<Constant>(Op))
5256 Op = ConstantExpr::getMul(OpC, Scale);
5257 else // We'll let instcombine(mul) convert this to a shl if possible.
5258 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5259 GEP->getName()+".idx"), I);
5262 // Emit an add instruction.
5263 if (isa<Constant>(Op) && isa<Constant>(Result))
5264 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5265 cast<Constant>(Result));
5267 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5268 GEP->getName()+".offs"), I);
5274 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5275 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5276 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5277 /// complex, and scales are involved. The above expression would also be legal
5278 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5279 /// later form is less amenable to optimization though, and we are allowed to
5280 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5282 /// If we can't emit an optimized form for this expression, this returns null.
5284 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5286 TargetData &TD = IC.getTargetData();
5287 gep_type_iterator GTI = gep_type_begin(GEP);
5289 // Check to see if this gep only has a single variable index. If so, and if
5290 // any constant indices are a multiple of its scale, then we can compute this
5291 // in terms of the scale of the variable index. For example, if the GEP
5292 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5293 // because the expression will cross zero at the same point.
5294 unsigned i, e = GEP->getNumOperands();
5296 for (i = 1; i != e; ++i, ++GTI) {
5297 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5298 // Compute the aggregate offset of constant indices.
5299 if (CI->isZero()) continue;
5301 // Handle a struct index, which adds its field offset to the pointer.
5302 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5303 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5305 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5306 Offset += Size*CI->getSExtValue();
5309 // Found our variable index.
5314 // If there are no variable indices, we must have a constant offset, just
5315 // evaluate it the general way.
5316 if (i == e) return 0;
5318 Value *VariableIdx = GEP->getOperand(i);
5319 // Determine the scale factor of the variable element. For example, this is
5320 // 4 if the variable index is into an array of i32.
5321 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5323 // Verify that there are no other variable indices. If so, emit the hard way.
5324 for (++i, ++GTI; i != e; ++i, ++GTI) {
5325 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5328 // Compute the aggregate offset of constant indices.
5329 if (CI->isZero()) continue;
5331 // Handle a struct index, which adds its field offset to the pointer.
5332 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5333 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5335 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5336 Offset += Size*CI->getSExtValue();
5340 // Okay, we know we have a single variable index, which must be a
5341 // pointer/array/vector index. If there is no offset, life is simple, return
5343 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5345 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5346 // we don't need to bother extending: the extension won't affect where the
5347 // computation crosses zero.
5348 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5349 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5350 VariableIdx->getNameStart(), &I);
5354 // Otherwise, there is an index. The computation we will do will be modulo
5355 // the pointer size, so get it.
5356 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5358 Offset &= PtrSizeMask;
5359 VariableScale &= PtrSizeMask;
5361 // To do this transformation, any constant index must be a multiple of the
5362 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5363 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5364 // multiple of the variable scale.
5365 int64_t NewOffs = Offset / (int64_t)VariableScale;
5366 if (Offset != NewOffs*(int64_t)VariableScale)
5369 // Okay, we can do this evaluation. Start by converting the index to intptr.
5370 const Type *IntPtrTy = TD.getIntPtrType();
5371 if (VariableIdx->getType() != IntPtrTy)
5372 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5374 VariableIdx->getNameStart(), &I);
5375 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5376 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5380 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5381 /// else. At this point we know that the GEP is on the LHS of the comparison.
5382 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5383 ICmpInst::Predicate Cond,
5385 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5387 // Look through bitcasts.
5388 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5389 RHS = BCI->getOperand(0);
5391 Value *PtrBase = GEPLHS->getOperand(0);
5392 if (PtrBase == RHS) {
5393 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5394 // This transformation (ignoring the base and scales) is valid because we
5395 // know pointers can't overflow. See if we can output an optimized form.
5396 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5398 // If not, synthesize the offset the hard way.
5400 Offset = EmitGEPOffset(GEPLHS, I, *this);
5401 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5402 Constant::getNullValue(Offset->getType()));
5403 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5404 // If the base pointers are different, but the indices are the same, just
5405 // compare the base pointer.
5406 if (PtrBase != GEPRHS->getOperand(0)) {
5407 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5408 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5409 GEPRHS->getOperand(0)->getType();
5411 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5412 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5413 IndicesTheSame = false;
5417 // If all indices are the same, just compare the base pointers.
5419 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5420 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5422 // Otherwise, the base pointers are different and the indices are
5423 // different, bail out.
5427 // If one of the GEPs has all zero indices, recurse.
5428 bool AllZeros = true;
5429 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5430 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5431 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5436 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5437 ICmpInst::getSwappedPredicate(Cond), I);
5439 // If the other GEP has all zero indices, recurse.
5441 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5442 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5443 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5448 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5450 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5451 // If the GEPs only differ by one index, compare it.
5452 unsigned NumDifferences = 0; // Keep track of # differences.
5453 unsigned DiffOperand = 0; // The operand that differs.
5454 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5455 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5456 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5457 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5458 // Irreconcilable differences.
5462 if (NumDifferences++) break;
5467 if (NumDifferences == 0) // SAME GEP?
5468 return ReplaceInstUsesWith(I, // No comparison is needed here.
5469 ConstantInt::get(Type::Int1Ty,
5470 ICmpInst::isTrueWhenEqual(Cond)));
5472 else if (NumDifferences == 1) {
5473 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5474 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5475 // Make sure we do a signed comparison here.
5476 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5480 // Only lower this if the icmp is the only user of the GEP or if we expect
5481 // the result to fold to a constant!
5482 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5483 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5484 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5485 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5486 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5487 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5493 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5495 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5498 if (!isa<ConstantFP>(RHSC)) return 0;
5499 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5501 // Get the width of the mantissa. We don't want to hack on conversions that
5502 // might lose information from the integer, e.g. "i64 -> float"
5503 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5504 if (MantissaWidth == -1) return 0; // Unknown.
5506 // Check to see that the input is converted from an integer type that is small
5507 // enough that preserves all bits. TODO: check here for "known" sign bits.
5508 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5509 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5511 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5512 if (isa<UIToFPInst>(LHSI))
5515 // If the conversion would lose info, don't hack on this.
5516 if ((int)InputSize > MantissaWidth)
5519 // Otherwise, we can potentially simplify the comparison. We know that it
5520 // will always come through as an integer value and we know the constant is
5521 // not a NAN (it would have been previously simplified).
5522 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5524 ICmpInst::Predicate Pred;
5525 switch (I.getPredicate()) {
5526 default: assert(0 && "Unexpected predicate!");
5527 case FCmpInst::FCMP_UEQ:
5528 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
5529 case FCmpInst::FCMP_UGT:
5530 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
5531 case FCmpInst::FCMP_UGE:
5532 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
5533 case FCmpInst::FCMP_ULT:
5534 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
5535 case FCmpInst::FCMP_ULE:
5536 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
5537 case FCmpInst::FCMP_UNE:
5538 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
5539 case FCmpInst::FCMP_ORD:
5540 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5541 case FCmpInst::FCMP_UNO:
5542 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5545 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5547 // Now we know that the APFloat is a normal number, zero or inf.
5549 // See if the FP constant is too large for the integer. For example,
5550 // comparing an i8 to 300.0.
5551 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5553 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5554 // and large values.
5555 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5556 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5557 APFloat::rmNearestTiesToEven);
5558 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5559 if (ICmpInst::ICMP_NE || ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
5560 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5561 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5564 // See if the RHS value is < SignedMin.
5565 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5566 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5567 APFloat::rmNearestTiesToEven);
5568 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5569 if (ICmpInst::ICMP_NE || ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
5570 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5571 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5574 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
5575 // it may still be fractional. See if it is fractional by casting the FP
5576 // value to the integer value and back, checking for equality. Don't do this
5577 // for zero, because -0.0 is not fractional.
5578 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5579 if (!RHS.isZero() &&
5580 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5581 // If we had a comparison against a fractional value, we have to adjust
5582 // the compare predicate and sometimes the value. RHSC is rounded towards
5583 // zero at this point.
5585 default: assert(0 && "Unexpected integer comparison!");
5586 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5587 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5588 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5589 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5590 case ICmpInst::ICMP_SLE:
5591 // (float)int <= 4.4 --> int <= 4
5592 // (float)int <= -4.4 --> int < -4
5593 if (RHS.isNegative())
5594 Pred = ICmpInst::ICMP_SLT;
5596 case ICmpInst::ICMP_SLT:
5597 // (float)int < -4.4 --> int < -4
5598 // (float)int < 4.4 --> int <= 4
5599 if (!RHS.isNegative())
5600 Pred = ICmpInst::ICMP_SLE;
5602 case ICmpInst::ICMP_SGT:
5603 // (float)int > 4.4 --> int > 4
5604 // (float)int > -4.4 --> int >= -4
5605 if (RHS.isNegative())
5606 Pred = ICmpInst::ICMP_SGE;
5608 case ICmpInst::ICMP_SGE:
5609 // (float)int >= -4.4 --> int >= -4
5610 // (float)int >= 4.4 --> int > 4
5611 if (!RHS.isNegative())
5612 Pred = ICmpInst::ICMP_SGT;
5617 // Lower this FP comparison into an appropriate integer version of the
5619 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5622 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5623 bool Changed = SimplifyCompare(I);
5624 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5626 // Fold trivial predicates.
5627 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5628 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5629 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5630 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5632 // Simplify 'fcmp pred X, X'
5634 switch (I.getPredicate()) {
5635 default: assert(0 && "Unknown predicate!");
5636 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5637 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5638 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5639 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5640 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5641 case FCmpInst::FCMP_OLT: // True if ordered and less than
5642 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5643 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5645 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5646 case FCmpInst::FCMP_ULT: // True if unordered or less than
5647 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5648 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5649 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5650 I.setPredicate(FCmpInst::FCMP_UNO);
5651 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5654 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5655 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5656 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5657 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5658 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5659 I.setPredicate(FCmpInst::FCMP_ORD);
5660 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5665 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5666 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5668 // Handle fcmp with constant RHS
5669 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5670 // If the constant is a nan, see if we can fold the comparison based on it.
5671 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5672 if (CFP->getValueAPF().isNaN()) {
5673 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5674 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5675 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5676 "Comparison must be either ordered or unordered!");
5677 // True if unordered.
5678 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5682 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5683 switch (LHSI->getOpcode()) {
5684 case Instruction::PHI:
5685 if (Instruction *NV = FoldOpIntoPhi(I))
5688 case Instruction::SIToFP:
5689 case Instruction::UIToFP:
5690 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5693 case Instruction::Select:
5694 // If either operand of the select is a constant, we can fold the
5695 // comparison into the select arms, which will cause one to be
5696 // constant folded and the select turned into a bitwise or.
5697 Value *Op1 = 0, *Op2 = 0;
5698 if (LHSI->hasOneUse()) {
5699 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5700 // Fold the known value into the constant operand.
5701 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5702 // Insert a new FCmp of the other select operand.
5703 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5704 LHSI->getOperand(2), RHSC,
5706 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5707 // Fold the known value into the constant operand.
5708 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5709 // Insert a new FCmp of the other select operand.
5710 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5711 LHSI->getOperand(1), RHSC,
5717 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5722 return Changed ? &I : 0;
5725 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5726 bool Changed = SimplifyCompare(I);
5727 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5728 const Type *Ty = Op0->getType();
5732 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5733 I.isTrueWhenEqual()));
5735 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5736 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5738 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5739 // addresses never equal each other! We already know that Op0 != Op1.
5740 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5741 isa<ConstantPointerNull>(Op0)) &&
5742 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5743 isa<ConstantPointerNull>(Op1)))
5744 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5745 !I.isTrueWhenEqual()));
5747 // icmp's with boolean values can always be turned into bitwise operations
5748 if (Ty == Type::Int1Ty) {
5749 switch (I.getPredicate()) {
5750 default: assert(0 && "Invalid icmp instruction!");
5751 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5752 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5753 InsertNewInstBefore(Xor, I);
5754 return BinaryOperator::CreateNot(Xor);
5756 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5757 return BinaryOperator::CreateXor(Op0, Op1);
5759 case ICmpInst::ICMP_UGT:
5760 case ICmpInst::ICMP_SGT:
5761 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5763 case ICmpInst::ICMP_ULT:
5764 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5765 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5766 InsertNewInstBefore(Not, I);
5767 return BinaryOperator::CreateAnd(Not, Op1);
5769 case ICmpInst::ICMP_UGE:
5770 case ICmpInst::ICMP_SGE:
5771 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5773 case ICmpInst::ICMP_ULE:
5774 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5775 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5776 InsertNewInstBefore(Not, I);
5777 return BinaryOperator::CreateOr(Not, Op1);
5782 // See if we are doing a comparison between a constant and an instruction that
5783 // can be folded into the comparison.
5784 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5787 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5788 if (I.isEquality() && CI->isNullValue() &&
5789 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5790 // (icmp cond A B) if cond is equality
5791 return new ICmpInst(I.getPredicate(), A, B);
5794 switch (I.getPredicate()) {
5796 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5797 if (CI->isMinValue(false))
5798 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5799 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5800 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5801 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5802 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5803 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5804 if (CI->isMinValue(true))
5805 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5806 ConstantInt::getAllOnesValue(Op0->getType()));
5810 case ICmpInst::ICMP_SLT:
5811 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5812 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5813 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5814 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5815 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5816 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5819 case ICmpInst::ICMP_UGT:
5820 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5821 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5822 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5823 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5824 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5825 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5827 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5828 if (CI->isMaxValue(true))
5829 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5830 ConstantInt::getNullValue(Op0->getType()));
5833 case ICmpInst::ICMP_SGT:
5834 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5835 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5836 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5837 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5838 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5839 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5842 case ICmpInst::ICMP_ULE:
5843 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5844 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5845 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5846 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5847 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5848 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5851 case ICmpInst::ICMP_SLE:
5852 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5853 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5854 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5855 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5856 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5857 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5860 case ICmpInst::ICMP_UGE:
5861 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5862 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5863 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5864 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5865 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5866 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5869 case ICmpInst::ICMP_SGE:
5870 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5871 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5872 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5873 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5874 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5875 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5879 // If we still have a icmp le or icmp ge instruction, turn it into the
5880 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5881 // already been handled above, this requires little checking.
5883 switch (I.getPredicate()) {
5885 case ICmpInst::ICMP_ULE:
5886 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5887 case ICmpInst::ICMP_SLE:
5888 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5889 case ICmpInst::ICMP_UGE:
5890 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5891 case ICmpInst::ICMP_SGE:
5892 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5895 // See if we can fold the comparison based on bits known to be zero or one
5896 // in the input. If this comparison is a normal comparison, it demands all
5897 // bits, if it is a sign bit comparison, it only demands the sign bit.
5900 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5902 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5903 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5904 if (SimplifyDemandedBits(Op0,
5905 isSignBit ? APInt::getSignBit(BitWidth)
5906 : APInt::getAllOnesValue(BitWidth),
5907 KnownZero, KnownOne, 0))
5910 // Given the known and unknown bits, compute a range that the LHS could be
5912 if ((KnownOne | KnownZero) != 0) {
5913 // Compute the Min, Max and RHS values based on the known bits. For the
5914 // EQ and NE we use unsigned values.
5915 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5916 const APInt& RHSVal = CI->getValue();
5917 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5918 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5921 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5924 switch (I.getPredicate()) { // LE/GE have been folded already.
5925 default: assert(0 && "Unknown icmp opcode!");
5926 case ICmpInst::ICMP_EQ:
5927 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5928 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5930 case ICmpInst::ICMP_NE:
5931 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5932 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5934 case ICmpInst::ICMP_ULT:
5935 if (Max.ult(RHSVal))
5936 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5937 if (Min.uge(RHSVal))
5938 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5940 case ICmpInst::ICMP_UGT:
5941 if (Min.ugt(RHSVal))
5942 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5943 if (Max.ule(RHSVal))
5944 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5946 case ICmpInst::ICMP_SLT:
5947 if (Max.slt(RHSVal))
5948 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5949 if (Min.sgt(RHSVal))
5950 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5952 case ICmpInst::ICMP_SGT:
5953 if (Min.sgt(RHSVal))
5954 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5955 if (Max.sle(RHSVal))
5956 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5961 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5962 // instruction, see if that instruction also has constants so that the
5963 // instruction can be folded into the icmp
5964 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5965 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5969 // Handle icmp with constant (but not simple integer constant) RHS
5970 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5971 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5972 switch (LHSI->getOpcode()) {
5973 case Instruction::GetElementPtr:
5974 if (RHSC->isNullValue()) {
5975 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5976 bool isAllZeros = true;
5977 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5978 if (!isa<Constant>(LHSI->getOperand(i)) ||
5979 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5984 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5985 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5989 case Instruction::PHI:
5990 if (Instruction *NV = FoldOpIntoPhi(I))
5993 case Instruction::Select: {
5994 // If either operand of the select is a constant, we can fold the
5995 // comparison into the select arms, which will cause one to be
5996 // constant folded and the select turned into a bitwise or.
5997 Value *Op1 = 0, *Op2 = 0;
5998 if (LHSI->hasOneUse()) {
5999 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6000 // Fold the known value into the constant operand.
6001 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6002 // Insert a new ICmp of the other select operand.
6003 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6004 LHSI->getOperand(2), RHSC,
6006 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6007 // Fold the known value into the constant operand.
6008 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6009 // Insert a new ICmp of the other select operand.
6010 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6011 LHSI->getOperand(1), RHSC,
6017 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6020 case Instruction::Malloc:
6021 // If we have (malloc != null), and if the malloc has a single use, we
6022 // can assume it is successful and remove the malloc.
6023 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6024 AddToWorkList(LHSI);
6025 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6026 !I.isTrueWhenEqual()));
6032 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6033 if (User *GEP = dyn_castGetElementPtr(Op0))
6034 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6036 if (User *GEP = dyn_castGetElementPtr(Op1))
6037 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6038 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6041 // Test to see if the operands of the icmp are casted versions of other
6042 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6044 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6045 if (isa<PointerType>(Op0->getType()) &&
6046 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6047 // We keep moving the cast from the left operand over to the right
6048 // operand, where it can often be eliminated completely.
6049 Op0 = CI->getOperand(0);
6051 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6052 // so eliminate it as well.
6053 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6054 Op1 = CI2->getOperand(0);
6056 // If Op1 is a constant, we can fold the cast into the constant.
6057 if (Op0->getType() != Op1->getType()) {
6058 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6059 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6061 // Otherwise, cast the RHS right before the icmp
6062 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6065 return new ICmpInst(I.getPredicate(), Op0, Op1);
6069 if (isa<CastInst>(Op0)) {
6070 // Handle the special case of: icmp (cast bool to X), <cst>
6071 // This comes up when you have code like
6074 // For generality, we handle any zero-extension of any operand comparison
6075 // with a constant or another cast from the same type.
6076 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6077 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6081 // ~x < ~y --> y < x
6083 if (match(Op0, m_Not(m_Value(A))) &&
6084 match(Op1, m_Not(m_Value(B))))
6085 return new ICmpInst(I.getPredicate(), B, A);
6088 if (I.isEquality()) {
6089 Value *A, *B, *C, *D;
6091 // -x == -y --> x == y
6092 if (match(Op0, m_Neg(m_Value(A))) &&
6093 match(Op1, m_Neg(m_Value(B))))
6094 return new ICmpInst(I.getPredicate(), A, B);
6096 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6097 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6098 Value *OtherVal = A == Op1 ? B : A;
6099 return new ICmpInst(I.getPredicate(), OtherVal,
6100 Constant::getNullValue(A->getType()));
6103 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6104 // A^c1 == C^c2 --> A == C^(c1^c2)
6105 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
6106 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
6107 if (Op1->hasOneUse()) {
6108 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6109 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6110 return new ICmpInst(I.getPredicate(), A,
6111 InsertNewInstBefore(Xor, I));
6114 // A^B == A^D -> B == D
6115 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6116 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6117 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6118 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6122 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6123 (A == Op0 || B == Op0)) {
6124 // A == (A^B) -> B == 0
6125 Value *OtherVal = A == Op0 ? B : A;
6126 return new ICmpInst(I.getPredicate(), OtherVal,
6127 Constant::getNullValue(A->getType()));
6129 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6130 // (A-B) == A -> B == 0
6131 return new ICmpInst(I.getPredicate(), B,
6132 Constant::getNullValue(B->getType()));
6134 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6135 // A == (A-B) -> B == 0
6136 return new ICmpInst(I.getPredicate(), B,
6137 Constant::getNullValue(B->getType()));
6140 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6141 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6142 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6143 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6144 Value *X = 0, *Y = 0, *Z = 0;
6147 X = B; Y = D; Z = A;
6148 } else if (A == D) {
6149 X = B; Y = C; Z = A;
6150 } else if (B == C) {
6151 X = A; Y = D; Z = B;
6152 } else if (B == D) {
6153 X = A; Y = C; Z = B;
6156 if (X) { // Build (X^Y) & Z
6157 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6158 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6159 I.setOperand(0, Op1);
6160 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6165 return Changed ? &I : 0;
6169 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6170 /// and CmpRHS are both known to be integer constants.
6171 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6172 ConstantInt *DivRHS) {
6173 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6174 const APInt &CmpRHSV = CmpRHS->getValue();
6176 // FIXME: If the operand types don't match the type of the divide
6177 // then don't attempt this transform. The code below doesn't have the
6178 // logic to deal with a signed divide and an unsigned compare (and
6179 // vice versa). This is because (x /s C1) <s C2 produces different
6180 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6181 // (x /u C1) <u C2. Simply casting the operands and result won't
6182 // work. :( The if statement below tests that condition and bails
6184 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6185 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6187 if (DivRHS->isZero())
6188 return 0; // The ProdOV computation fails on divide by zero.
6190 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6191 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6192 // C2 (CI). By solving for X we can turn this into a range check
6193 // instead of computing a divide.
6194 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6196 // Determine if the product overflows by seeing if the product is
6197 // not equal to the divide. Make sure we do the same kind of divide
6198 // as in the LHS instruction that we're folding.
6199 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6200 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6202 // Get the ICmp opcode
6203 ICmpInst::Predicate Pred = ICI.getPredicate();
6205 // Figure out the interval that is being checked. For example, a comparison
6206 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6207 // Compute this interval based on the constants involved and the signedness of
6208 // the compare/divide. This computes a half-open interval, keeping track of
6209 // whether either value in the interval overflows. After analysis each
6210 // overflow variable is set to 0 if it's corresponding bound variable is valid
6211 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6212 int LoOverflow = 0, HiOverflow = 0;
6213 ConstantInt *LoBound = 0, *HiBound = 0;
6216 if (!DivIsSigned) { // udiv
6217 // e.g. X/5 op 3 --> [15, 20)
6219 HiOverflow = LoOverflow = ProdOV;
6221 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6222 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6223 if (CmpRHSV == 0) { // (X / pos) op 0
6224 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6225 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6227 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6228 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6229 HiOverflow = LoOverflow = ProdOV;
6231 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6232 } else { // (X / pos) op neg
6233 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6234 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
6235 LoOverflow = AddWithOverflow(LoBound, Prod,
6236 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
6237 HiBound = AddOne(Prod);
6238 HiOverflow = ProdOV ? -1 : 0;
6240 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6241 if (CmpRHSV == 0) { // (X / neg) op 0
6242 // e.g. X/-5 op 0 --> [-4, 5)
6243 LoBound = AddOne(DivRHS);
6244 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6245 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6246 HiOverflow = 1; // [INTMIN+1, overflow)
6247 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6249 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6250 // e.g. X/-5 op 3 --> [-19, -14)
6251 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6253 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
6254 HiBound = AddOne(Prod);
6255 } else { // (X / neg) op neg
6256 // e.g. X/-5 op -3 --> [15, 20)
6258 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
6259 HiBound = Subtract(Prod, DivRHS);
6262 // Dividing by a negative swaps the condition. LT <-> GT
6263 Pred = ICmpInst::getSwappedPredicate(Pred);
6266 Value *X = DivI->getOperand(0);
6268 default: assert(0 && "Unhandled icmp opcode!");
6269 case ICmpInst::ICMP_EQ:
6270 if (LoOverflow && HiOverflow)
6271 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6272 else if (HiOverflow)
6273 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6274 ICmpInst::ICMP_UGE, X, LoBound);
6275 else if (LoOverflow)
6276 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6277 ICmpInst::ICMP_ULT, X, HiBound);
6279 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6280 case ICmpInst::ICMP_NE:
6281 if (LoOverflow && HiOverflow)
6282 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6283 else if (HiOverflow)
6284 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6285 ICmpInst::ICMP_ULT, X, LoBound);
6286 else if (LoOverflow)
6287 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6288 ICmpInst::ICMP_UGE, X, HiBound);
6290 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6291 case ICmpInst::ICMP_ULT:
6292 case ICmpInst::ICMP_SLT:
6293 if (LoOverflow == +1) // Low bound is greater than input range.
6294 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6295 if (LoOverflow == -1) // Low bound is less than input range.
6296 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6297 return new ICmpInst(Pred, X, LoBound);
6298 case ICmpInst::ICMP_UGT:
6299 case ICmpInst::ICMP_SGT:
6300 if (HiOverflow == +1) // High bound greater than input range.
6301 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6302 else if (HiOverflow == -1) // High bound less than input range.
6303 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6304 if (Pred == ICmpInst::ICMP_UGT)
6305 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6307 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6312 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6314 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6317 const APInt &RHSV = RHS->getValue();
6319 switch (LHSI->getOpcode()) {
6320 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6321 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6322 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6324 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6325 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6326 Value *CompareVal = LHSI->getOperand(0);
6328 // If the sign bit of the XorCST is not set, there is no change to
6329 // the operation, just stop using the Xor.
6330 if (!XorCST->getValue().isNegative()) {
6331 ICI.setOperand(0, CompareVal);
6332 AddToWorkList(LHSI);
6336 // Was the old condition true if the operand is positive?
6337 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6339 // If so, the new one isn't.
6340 isTrueIfPositive ^= true;
6342 if (isTrueIfPositive)
6343 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6345 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6349 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6350 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6351 LHSI->getOperand(0)->hasOneUse()) {
6352 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6354 // If the LHS is an AND of a truncating cast, we can widen the
6355 // and/compare to be the input width without changing the value
6356 // produced, eliminating a cast.
6357 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6358 // We can do this transformation if either the AND constant does not
6359 // have its sign bit set or if it is an equality comparison.
6360 // Extending a relational comparison when we're checking the sign
6361 // bit would not work.
6362 if (Cast->hasOneUse() &&
6363 (ICI.isEquality() ||
6364 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6366 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6367 APInt NewCST = AndCST->getValue();
6368 NewCST.zext(BitWidth);
6370 NewCI.zext(BitWidth);
6371 Instruction *NewAnd =
6372 BinaryOperator::CreateAnd(Cast->getOperand(0),
6373 ConstantInt::get(NewCST),LHSI->getName());
6374 InsertNewInstBefore(NewAnd, ICI);
6375 return new ICmpInst(ICI.getPredicate(), NewAnd,
6376 ConstantInt::get(NewCI));
6380 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6381 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6382 // happens a LOT in code produced by the C front-end, for bitfield
6384 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6385 if (Shift && !Shift->isShift())
6389 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6390 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6391 const Type *AndTy = AndCST->getType(); // Type of the and.
6393 // We can fold this as long as we can't shift unknown bits
6394 // into the mask. This can only happen with signed shift
6395 // rights, as they sign-extend.
6397 bool CanFold = Shift->isLogicalShift();
6399 // To test for the bad case of the signed shr, see if any
6400 // of the bits shifted in could be tested after the mask.
6401 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6402 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6404 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6405 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6406 AndCST->getValue()) == 0)
6412 if (Shift->getOpcode() == Instruction::Shl)
6413 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6415 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6417 // Check to see if we are shifting out any of the bits being
6419 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6420 // If we shifted bits out, the fold is not going to work out.
6421 // As a special case, check to see if this means that the
6422 // result is always true or false now.
6423 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6424 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6425 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6426 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6428 ICI.setOperand(1, NewCst);
6429 Constant *NewAndCST;
6430 if (Shift->getOpcode() == Instruction::Shl)
6431 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6433 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6434 LHSI->setOperand(1, NewAndCST);
6435 LHSI->setOperand(0, Shift->getOperand(0));
6436 AddToWorkList(Shift); // Shift is dead.
6437 AddUsesToWorkList(ICI);
6443 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6444 // preferable because it allows the C<<Y expression to be hoisted out
6445 // of a loop if Y is invariant and X is not.
6446 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6447 ICI.isEquality() && !Shift->isArithmeticShift() &&
6448 isa<Instruction>(Shift->getOperand(0))) {
6451 if (Shift->getOpcode() == Instruction::LShr) {
6452 NS = BinaryOperator::CreateShl(AndCST,
6453 Shift->getOperand(1), "tmp");
6455 // Insert a logical shift.
6456 NS = BinaryOperator::CreateLShr(AndCST,
6457 Shift->getOperand(1), "tmp");
6459 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6461 // Compute X & (C << Y).
6462 Instruction *NewAnd =
6463 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6464 InsertNewInstBefore(NewAnd, ICI);
6466 ICI.setOperand(0, NewAnd);
6472 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6473 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6476 uint32_t TypeBits = RHSV.getBitWidth();
6478 // Check that the shift amount is in range. If not, don't perform
6479 // undefined shifts. When the shift is visited it will be
6481 if (ShAmt->uge(TypeBits))
6484 if (ICI.isEquality()) {
6485 // If we are comparing against bits always shifted out, the
6486 // comparison cannot succeed.
6488 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6489 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6490 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6491 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6492 return ReplaceInstUsesWith(ICI, Cst);
6495 if (LHSI->hasOneUse()) {
6496 // Otherwise strength reduce the shift into an and.
6497 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6499 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6502 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6503 Mask, LHSI->getName()+".mask");
6504 Value *And = InsertNewInstBefore(AndI, ICI);
6505 return new ICmpInst(ICI.getPredicate(), And,
6506 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6510 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6511 bool TrueIfSigned = false;
6512 if (LHSI->hasOneUse() &&
6513 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6514 // (X << 31) <s 0 --> (X&1) != 0
6515 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6516 (TypeBits-ShAmt->getZExtValue()-1));
6518 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6519 Mask, LHSI->getName()+".mask");
6520 Value *And = InsertNewInstBefore(AndI, ICI);
6522 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6523 And, Constant::getNullValue(And->getType()));
6528 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6529 case Instruction::AShr: {
6530 // Only handle equality comparisons of shift-by-constant.
6531 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6532 if (!ShAmt || !ICI.isEquality()) break;
6534 // Check that the shift amount is in range. If not, don't perform
6535 // undefined shifts. When the shift is visited it will be
6537 uint32_t TypeBits = RHSV.getBitWidth();
6538 if (ShAmt->uge(TypeBits))
6541 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6543 // If we are comparing against bits always shifted out, the
6544 // comparison cannot succeed.
6545 APInt Comp = RHSV << ShAmtVal;
6546 if (LHSI->getOpcode() == Instruction::LShr)
6547 Comp = Comp.lshr(ShAmtVal);
6549 Comp = Comp.ashr(ShAmtVal);
6551 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6552 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6553 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6554 return ReplaceInstUsesWith(ICI, Cst);
6557 // Otherwise, check to see if the bits shifted out are known to be zero.
6558 // If so, we can compare against the unshifted value:
6559 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6560 if (LHSI->hasOneUse() &&
6561 MaskedValueIsZero(LHSI->getOperand(0),
6562 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6563 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6564 ConstantExpr::getShl(RHS, ShAmt));
6567 if (LHSI->hasOneUse()) {
6568 // Otherwise strength reduce the shift into an and.
6569 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6570 Constant *Mask = ConstantInt::get(Val);
6573 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6574 Mask, LHSI->getName()+".mask");
6575 Value *And = InsertNewInstBefore(AndI, ICI);
6576 return new ICmpInst(ICI.getPredicate(), And,
6577 ConstantExpr::getShl(RHS, ShAmt));
6582 case Instruction::SDiv:
6583 case Instruction::UDiv:
6584 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6585 // Fold this div into the comparison, producing a range check.
6586 // Determine, based on the divide type, what the range is being
6587 // checked. If there is an overflow on the low or high side, remember
6588 // it, otherwise compute the range [low, hi) bounding the new value.
6589 // See: InsertRangeTest above for the kinds of replacements possible.
6590 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6591 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6596 case Instruction::Add:
6597 // Fold: icmp pred (add, X, C1), C2
6599 if (!ICI.isEquality()) {
6600 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6602 const APInt &LHSV = LHSC->getValue();
6604 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6607 if (ICI.isSignedPredicate()) {
6608 if (CR.getLower().isSignBit()) {
6609 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6610 ConstantInt::get(CR.getUpper()));
6611 } else if (CR.getUpper().isSignBit()) {
6612 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6613 ConstantInt::get(CR.getLower()));
6616 if (CR.getLower().isMinValue()) {
6617 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6618 ConstantInt::get(CR.getUpper()));
6619 } else if (CR.getUpper().isMinValue()) {
6620 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6621 ConstantInt::get(CR.getLower()));
6628 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6629 if (ICI.isEquality()) {
6630 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6632 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6633 // the second operand is a constant, simplify a bit.
6634 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6635 switch (BO->getOpcode()) {
6636 case Instruction::SRem:
6637 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6638 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6639 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6640 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6641 Instruction *NewRem =
6642 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6644 InsertNewInstBefore(NewRem, ICI);
6645 return new ICmpInst(ICI.getPredicate(), NewRem,
6646 Constant::getNullValue(BO->getType()));
6650 case Instruction::Add:
6651 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6652 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6653 if (BO->hasOneUse())
6654 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6655 Subtract(RHS, BOp1C));
6656 } else if (RHSV == 0) {
6657 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6658 // efficiently invertible, or if the add has just this one use.
6659 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6661 if (Value *NegVal = dyn_castNegVal(BOp1))
6662 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6663 else if (Value *NegVal = dyn_castNegVal(BOp0))
6664 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6665 else if (BO->hasOneUse()) {
6666 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6667 InsertNewInstBefore(Neg, ICI);
6669 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6673 case Instruction::Xor:
6674 // For the xor case, we can xor two constants together, eliminating
6675 // the explicit xor.
6676 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6677 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6678 ConstantExpr::getXor(RHS, BOC));
6681 case Instruction::Sub:
6682 // Replace (([sub|xor] A, B) != 0) with (A != B)
6684 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6688 case Instruction::Or:
6689 // If bits are being or'd in that are not present in the constant we
6690 // are comparing against, then the comparison could never succeed!
6691 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6692 Constant *NotCI = ConstantExpr::getNot(RHS);
6693 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6694 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6699 case Instruction::And:
6700 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6701 // If bits are being compared against that are and'd out, then the
6702 // comparison can never succeed!
6703 if ((RHSV & ~BOC->getValue()) != 0)
6704 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6707 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6708 if (RHS == BOC && RHSV.isPowerOf2())
6709 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6710 ICmpInst::ICMP_NE, LHSI,
6711 Constant::getNullValue(RHS->getType()));
6713 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6714 if (isSignBit(BOC)) {
6715 Value *X = BO->getOperand(0);
6716 Constant *Zero = Constant::getNullValue(X->getType());
6717 ICmpInst::Predicate pred = isICMP_NE ?
6718 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6719 return new ICmpInst(pred, X, Zero);
6722 // ((X & ~7) == 0) --> X < 8
6723 if (RHSV == 0 && isHighOnes(BOC)) {
6724 Value *X = BO->getOperand(0);
6725 Constant *NegX = ConstantExpr::getNeg(BOC);
6726 ICmpInst::Predicate pred = isICMP_NE ?
6727 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6728 return new ICmpInst(pred, X, NegX);
6733 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6734 // Handle icmp {eq|ne} <intrinsic>, intcst.
6735 if (II->getIntrinsicID() == Intrinsic::bswap) {
6737 ICI.setOperand(0, II->getOperand(1));
6738 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6742 } else { // Not a ICMP_EQ/ICMP_NE
6743 // If the LHS is a cast from an integral value of the same size,
6744 // then since we know the RHS is a constant, try to simlify.
6745 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6746 Value *CastOp = Cast->getOperand(0);
6747 const Type *SrcTy = CastOp->getType();
6748 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6749 if (SrcTy->isInteger() &&
6750 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6751 // If this is an unsigned comparison, try to make the comparison use
6752 // smaller constant values.
6753 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6754 // X u< 128 => X s> -1
6755 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6756 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6757 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6758 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6759 // X u> 127 => X s< 0
6760 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6761 Constant::getNullValue(SrcTy));
6769 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6770 /// We only handle extending casts so far.
6772 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6773 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6774 Value *LHSCIOp = LHSCI->getOperand(0);
6775 const Type *SrcTy = LHSCIOp->getType();
6776 const Type *DestTy = LHSCI->getType();
6779 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6780 // integer type is the same size as the pointer type.
6781 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6782 getTargetData().getPointerSizeInBits() ==
6783 cast<IntegerType>(DestTy)->getBitWidth()) {
6785 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6786 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6787 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6788 RHSOp = RHSC->getOperand(0);
6789 // If the pointer types don't match, insert a bitcast.
6790 if (LHSCIOp->getType() != RHSOp->getType())
6791 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6795 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6798 // The code below only handles extension cast instructions, so far.
6800 if (LHSCI->getOpcode() != Instruction::ZExt &&
6801 LHSCI->getOpcode() != Instruction::SExt)
6804 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6805 bool isSignedCmp = ICI.isSignedPredicate();
6807 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6808 // Not an extension from the same type?
6809 RHSCIOp = CI->getOperand(0);
6810 if (RHSCIOp->getType() != LHSCIOp->getType())
6813 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6814 // and the other is a zext), then we can't handle this.
6815 if (CI->getOpcode() != LHSCI->getOpcode())
6818 // Deal with equality cases early.
6819 if (ICI.isEquality())
6820 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6822 // A signed comparison of sign extended values simplifies into a
6823 // signed comparison.
6824 if (isSignedCmp && isSignedExt)
6825 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6827 // The other three cases all fold into an unsigned comparison.
6828 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6831 // If we aren't dealing with a constant on the RHS, exit early
6832 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6836 // Compute the constant that would happen if we truncated to SrcTy then
6837 // reextended to DestTy.
6838 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6839 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6841 // If the re-extended constant didn't change...
6843 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6844 // For example, we might have:
6845 // %A = sext short %X to uint
6846 // %B = icmp ugt uint %A, 1330
6847 // It is incorrect to transform this into
6848 // %B = icmp ugt short %X, 1330
6849 // because %A may have negative value.
6851 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6852 // OR operation is EQ/NE.
6853 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6854 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6859 // The re-extended constant changed so the constant cannot be represented
6860 // in the shorter type. Consequently, we cannot emit a simple comparison.
6862 // First, handle some easy cases. We know the result cannot be equal at this
6863 // point so handle the ICI.isEquality() cases
6864 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6865 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6866 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6867 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6869 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6870 // should have been folded away previously and not enter in here.
6873 // We're performing a signed comparison.
6874 if (cast<ConstantInt>(CI)->getValue().isNegative())
6875 Result = ConstantInt::getFalse(); // X < (small) --> false
6877 Result = ConstantInt::getTrue(); // X < (large) --> true
6879 // We're performing an unsigned comparison.
6881 // We're performing an unsigned comp with a sign extended value.
6882 // This is true if the input is >= 0. [aka >s -1]
6883 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6884 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6885 NegOne, ICI.getName()), ICI);
6887 // Unsigned extend & unsigned compare -> always true.
6888 Result = ConstantInt::getTrue();
6892 // Finally, return the value computed.
6893 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6894 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6895 return ReplaceInstUsesWith(ICI, Result);
6897 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6898 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6899 "ICmp should be folded!");
6900 if (Constant *CI = dyn_cast<Constant>(Result))
6901 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6903 return BinaryOperator::CreateNot(Result);
6907 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6908 return commonShiftTransforms(I);
6911 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6912 return commonShiftTransforms(I);
6915 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6916 if (Instruction *R = commonShiftTransforms(I))
6919 Value *Op0 = I.getOperand(0);
6921 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6922 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6923 if (CSI->isAllOnesValue())
6924 return ReplaceInstUsesWith(I, CSI);
6926 // See if we can turn a signed shr into an unsigned shr.
6927 if (MaskedValueIsZero(Op0,
6928 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6929 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6934 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6935 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6936 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6938 // shl X, 0 == X and shr X, 0 == X
6939 // shl 0, X == 0 and shr 0, X == 0
6940 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6941 Op0 == Constant::getNullValue(Op0->getType()))
6942 return ReplaceInstUsesWith(I, Op0);
6944 if (isa<UndefValue>(Op0)) {
6945 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6946 return ReplaceInstUsesWith(I, Op0);
6947 else // undef << X -> 0, undef >>u X -> 0
6948 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6950 if (isa<UndefValue>(Op1)) {
6951 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6952 return ReplaceInstUsesWith(I, Op0);
6953 else // X << undef, X >>u undef -> 0
6954 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6957 // Try to fold constant and into select arguments.
6958 if (isa<Constant>(Op0))
6959 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6960 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6963 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6964 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6969 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6970 BinaryOperator &I) {
6971 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6973 // See if we can simplify any instructions used by the instruction whose sole
6974 // purpose is to compute bits we don't care about.
6975 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6976 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6977 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6978 KnownZero, KnownOne))
6981 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6982 // of a signed value.
6984 if (Op1->uge(TypeBits)) {
6985 if (I.getOpcode() != Instruction::AShr)
6986 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6988 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6993 // ((X*C1) << C2) == (X * (C1 << C2))
6994 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6995 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6996 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6997 return BinaryOperator::CreateMul(BO->getOperand(0),
6998 ConstantExpr::getShl(BOOp, Op1));
7000 // Try to fold constant and into select arguments.
7001 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7002 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7004 if (isa<PHINode>(Op0))
7005 if (Instruction *NV = FoldOpIntoPhi(I))
7008 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7009 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7010 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7011 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7012 // place. Don't try to do this transformation in this case. Also, we
7013 // require that the input operand is a shift-by-constant so that we have
7014 // confidence that the shifts will get folded together. We could do this
7015 // xform in more cases, but it is unlikely to be profitable.
7016 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7017 isa<ConstantInt>(TrOp->getOperand(1))) {
7018 // Okay, we'll do this xform. Make the shift of shift.
7019 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7020 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7022 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7024 // For logical shifts, the truncation has the effect of making the high
7025 // part of the register be zeros. Emulate this by inserting an AND to
7026 // clear the top bits as needed. This 'and' will usually be zapped by
7027 // other xforms later if dead.
7028 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7029 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7030 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7032 // The mask we constructed says what the trunc would do if occurring
7033 // between the shifts. We want to know the effect *after* the second
7034 // shift. We know that it is a logical shift by a constant, so adjust the
7035 // mask as appropriate.
7036 if (I.getOpcode() == Instruction::Shl)
7037 MaskV <<= Op1->getZExtValue();
7039 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7040 MaskV = MaskV.lshr(Op1->getZExtValue());
7043 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7045 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7047 // Return the value truncated to the interesting size.
7048 return new TruncInst(And, I.getType());
7052 if (Op0->hasOneUse()) {
7053 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7054 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7057 switch (Op0BO->getOpcode()) {
7059 case Instruction::Add:
7060 case Instruction::And:
7061 case Instruction::Or:
7062 case Instruction::Xor: {
7063 // These operators commute.
7064 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7065 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7066 match(Op0BO->getOperand(1),
7067 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
7068 Instruction *YS = BinaryOperator::CreateShl(
7069 Op0BO->getOperand(0), Op1,
7071 InsertNewInstBefore(YS, I); // (Y << C)
7073 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7074 Op0BO->getOperand(1)->getName());
7075 InsertNewInstBefore(X, I); // (X + (Y << C))
7076 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7077 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7078 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7081 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7082 Value *Op0BOOp1 = Op0BO->getOperand(1);
7083 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7085 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
7086 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
7088 Instruction *YS = BinaryOperator::CreateShl(
7089 Op0BO->getOperand(0), Op1,
7091 InsertNewInstBefore(YS, I); // (Y << C)
7093 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7094 V1->getName()+".mask");
7095 InsertNewInstBefore(XM, I); // X & (CC << C)
7097 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7102 case Instruction::Sub: {
7103 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7104 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7105 match(Op0BO->getOperand(0),
7106 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
7107 Instruction *YS = BinaryOperator::CreateShl(
7108 Op0BO->getOperand(1), Op1,
7110 InsertNewInstBefore(YS, I); // (Y << C)
7112 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7113 Op0BO->getOperand(0)->getName());
7114 InsertNewInstBefore(X, I); // (X + (Y << C))
7115 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7116 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7117 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7120 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7121 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7122 match(Op0BO->getOperand(0),
7123 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7124 m_ConstantInt(CC))) && V2 == Op1 &&
7125 cast<BinaryOperator>(Op0BO->getOperand(0))
7126 ->getOperand(0)->hasOneUse()) {
7127 Instruction *YS = BinaryOperator::CreateShl(
7128 Op0BO->getOperand(1), Op1,
7130 InsertNewInstBefore(YS, I); // (Y << C)
7132 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7133 V1->getName()+".mask");
7134 InsertNewInstBefore(XM, I); // X & (CC << C)
7136 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7144 // If the operand is an bitwise operator with a constant RHS, and the
7145 // shift is the only use, we can pull it out of the shift.
7146 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7147 bool isValid = true; // Valid only for And, Or, Xor
7148 bool highBitSet = false; // Transform if high bit of constant set?
7150 switch (Op0BO->getOpcode()) {
7151 default: isValid = false; break; // Do not perform transform!
7152 case Instruction::Add:
7153 isValid = isLeftShift;
7155 case Instruction::Or:
7156 case Instruction::Xor:
7159 case Instruction::And:
7164 // If this is a signed shift right, and the high bit is modified
7165 // by the logical operation, do not perform the transformation.
7166 // The highBitSet boolean indicates the value of the high bit of
7167 // the constant which would cause it to be modified for this
7170 if (isValid && I.getOpcode() == Instruction::AShr)
7171 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7174 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7176 Instruction *NewShift =
7177 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7178 InsertNewInstBefore(NewShift, I);
7179 NewShift->takeName(Op0BO);
7181 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7188 // Find out if this is a shift of a shift by a constant.
7189 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7190 if (ShiftOp && !ShiftOp->isShift())
7193 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7194 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7195 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7196 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7197 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7198 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7199 Value *X = ShiftOp->getOperand(0);
7201 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7202 if (AmtSum > TypeBits)
7205 const IntegerType *Ty = cast<IntegerType>(I.getType());
7207 // Check for (X << c1) << c2 and (X >> c1) >> c2
7208 if (I.getOpcode() == ShiftOp->getOpcode()) {
7209 return BinaryOperator::Create(I.getOpcode(), X,
7210 ConstantInt::get(Ty, AmtSum));
7211 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7212 I.getOpcode() == Instruction::AShr) {
7213 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7214 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7215 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7216 I.getOpcode() == Instruction::LShr) {
7217 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7218 Instruction *Shift =
7219 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7220 InsertNewInstBefore(Shift, I);
7222 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7223 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7226 // Okay, if we get here, one shift must be left, and the other shift must be
7227 // right. See if the amounts are equal.
7228 if (ShiftAmt1 == ShiftAmt2) {
7229 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7230 if (I.getOpcode() == Instruction::Shl) {
7231 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7232 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7234 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7235 if (I.getOpcode() == Instruction::LShr) {
7236 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7237 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7239 // We can simplify ((X << C) >>s C) into a trunc + sext.
7240 // NOTE: we could do this for any C, but that would make 'unusual' integer
7241 // types. For now, just stick to ones well-supported by the code
7243 const Type *SExtType = 0;
7244 switch (Ty->getBitWidth() - ShiftAmt1) {
7251 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7256 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7257 InsertNewInstBefore(NewTrunc, I);
7258 return new SExtInst(NewTrunc, Ty);
7260 // Otherwise, we can't handle it yet.
7261 } else if (ShiftAmt1 < ShiftAmt2) {
7262 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7264 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7265 if (I.getOpcode() == Instruction::Shl) {
7266 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7267 ShiftOp->getOpcode() == Instruction::AShr);
7268 Instruction *Shift =
7269 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7270 InsertNewInstBefore(Shift, I);
7272 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7273 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7276 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7277 if (I.getOpcode() == Instruction::LShr) {
7278 assert(ShiftOp->getOpcode() == Instruction::Shl);
7279 Instruction *Shift =
7280 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7281 InsertNewInstBefore(Shift, I);
7283 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7284 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7287 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7289 assert(ShiftAmt2 < ShiftAmt1);
7290 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7292 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7293 if (I.getOpcode() == Instruction::Shl) {
7294 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7295 ShiftOp->getOpcode() == Instruction::AShr);
7296 Instruction *Shift =
7297 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7298 ConstantInt::get(Ty, ShiftDiff));
7299 InsertNewInstBefore(Shift, I);
7301 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7302 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7305 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7306 if (I.getOpcode() == Instruction::LShr) {
7307 assert(ShiftOp->getOpcode() == Instruction::Shl);
7308 Instruction *Shift =
7309 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7310 InsertNewInstBefore(Shift, I);
7312 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7313 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7316 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7323 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7324 /// expression. If so, decompose it, returning some value X, such that Val is
7327 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7329 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7330 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7331 Offset = CI->getZExtValue();
7333 return ConstantInt::get(Type::Int32Ty, 0);
7334 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7335 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7336 if (I->getOpcode() == Instruction::Shl) {
7337 // This is a value scaled by '1 << the shift amt'.
7338 Scale = 1U << RHS->getZExtValue();
7340 return I->getOperand(0);
7341 } else if (I->getOpcode() == Instruction::Mul) {
7342 // This value is scaled by 'RHS'.
7343 Scale = RHS->getZExtValue();
7345 return I->getOperand(0);
7346 } else if (I->getOpcode() == Instruction::Add) {
7347 // We have X+C. Check to see if we really have (X*C2)+C1,
7348 // where C1 is divisible by C2.
7351 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7352 Offset += RHS->getZExtValue();
7359 // Otherwise, we can't look past this.
7366 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7367 /// try to eliminate the cast by moving the type information into the alloc.
7368 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7369 AllocationInst &AI) {
7370 const PointerType *PTy = cast<PointerType>(CI.getType());
7372 // Remove any uses of AI that are dead.
7373 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7375 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7376 Instruction *User = cast<Instruction>(*UI++);
7377 if (isInstructionTriviallyDead(User)) {
7378 while (UI != E && *UI == User)
7379 ++UI; // If this instruction uses AI more than once, don't break UI.
7382 DOUT << "IC: DCE: " << *User;
7383 EraseInstFromFunction(*User);
7387 // Get the type really allocated and the type casted to.
7388 const Type *AllocElTy = AI.getAllocatedType();
7389 const Type *CastElTy = PTy->getElementType();
7390 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7392 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7393 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7394 if (CastElTyAlign < AllocElTyAlign) return 0;
7396 // If the allocation has multiple uses, only promote it if we are strictly
7397 // increasing the alignment of the resultant allocation. If we keep it the
7398 // same, we open the door to infinite loops of various kinds.
7399 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7401 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7402 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7403 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7405 // See if we can satisfy the modulus by pulling a scale out of the array
7407 unsigned ArraySizeScale;
7409 Value *NumElements = // See if the array size is a decomposable linear expr.
7410 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7412 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7414 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7415 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7417 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7422 // If the allocation size is constant, form a constant mul expression
7423 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7424 if (isa<ConstantInt>(NumElements))
7425 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7426 // otherwise multiply the amount and the number of elements
7427 else if (Scale != 1) {
7428 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7429 Amt = InsertNewInstBefore(Tmp, AI);
7433 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7434 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7435 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7436 Amt = InsertNewInstBefore(Tmp, AI);
7439 AllocationInst *New;
7440 if (isa<MallocInst>(AI))
7441 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7443 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7444 InsertNewInstBefore(New, AI);
7447 // If the allocation has multiple uses, insert a cast and change all things
7448 // that used it to use the new cast. This will also hack on CI, but it will
7450 if (!AI.hasOneUse()) {
7451 AddUsesToWorkList(AI);
7452 // New is the allocation instruction, pointer typed. AI is the original
7453 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7454 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7455 InsertNewInstBefore(NewCast, AI);
7456 AI.replaceAllUsesWith(NewCast);
7458 return ReplaceInstUsesWith(CI, New);
7461 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7462 /// and return it as type Ty without inserting any new casts and without
7463 /// changing the computed value. This is used by code that tries to decide
7464 /// whether promoting or shrinking integer operations to wider or smaller types
7465 /// will allow us to eliminate a truncate or extend.
7467 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7468 /// extension operation if Ty is larger.
7469 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7471 int &NumCastsRemoved) {
7472 // We can always evaluate constants in another type.
7473 if (isa<ConstantInt>(V))
7476 Instruction *I = dyn_cast<Instruction>(V);
7477 if (!I) return false;
7479 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7481 // If this is an extension or truncate, we can often eliminate it.
7482 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7483 // If this is a cast from the destination type, we can trivially eliminate
7484 // it, and this will remove a cast overall.
7485 if (I->getOperand(0)->getType() == Ty) {
7486 // If the first operand is itself a cast, and is eliminable, do not count
7487 // this as an eliminable cast. We would prefer to eliminate those two
7489 if (!isa<CastInst>(I->getOperand(0)))
7495 // We can't extend or shrink something that has multiple uses: doing so would
7496 // require duplicating the instruction in general, which isn't profitable.
7497 if (!I->hasOneUse()) return false;
7499 switch (I->getOpcode()) {
7500 case Instruction::Add:
7501 case Instruction::Sub:
7502 case Instruction::And:
7503 case Instruction::Or:
7504 case Instruction::Xor:
7505 // These operators can all arbitrarily be extended or truncated.
7506 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7508 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7511 case Instruction::Mul:
7512 // A multiply can be truncated by truncating its operands.
7513 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
7514 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7516 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7519 case Instruction::Shl:
7520 // If we are truncating the result of this SHL, and if it's a shift of a
7521 // constant amount, we can always perform a SHL in a smaller type.
7522 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7523 uint32_t BitWidth = Ty->getBitWidth();
7524 if (BitWidth < OrigTy->getBitWidth() &&
7525 CI->getLimitedValue(BitWidth) < BitWidth)
7526 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7530 case Instruction::LShr:
7531 // If this is a truncate of a logical shr, we can truncate it to a smaller
7532 // lshr iff we know that the bits we would otherwise be shifting in are
7534 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7535 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7536 uint32_t BitWidth = Ty->getBitWidth();
7537 if (BitWidth < OrigBitWidth &&
7538 MaskedValueIsZero(I->getOperand(0),
7539 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7540 CI->getLimitedValue(BitWidth) < BitWidth) {
7541 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7546 case Instruction::ZExt:
7547 case Instruction::SExt:
7548 case Instruction::Trunc:
7549 // If this is the same kind of case as our original (e.g. zext+zext), we
7550 // can safely replace it. Note that replacing it does not reduce the number
7551 // of casts in the input.
7552 if (I->getOpcode() == CastOpc)
7557 // TODO: Can handle more cases here.
7564 /// EvaluateInDifferentType - Given an expression that
7565 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7566 /// evaluate the expression.
7567 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7569 if (Constant *C = dyn_cast<Constant>(V))
7570 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7572 // Otherwise, it must be an instruction.
7573 Instruction *I = cast<Instruction>(V);
7574 Instruction *Res = 0;
7575 switch (I->getOpcode()) {
7576 case Instruction::Add:
7577 case Instruction::Sub:
7578 case Instruction::Mul:
7579 case Instruction::And:
7580 case Instruction::Or:
7581 case Instruction::Xor:
7582 case Instruction::AShr:
7583 case Instruction::LShr:
7584 case Instruction::Shl: {
7585 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7586 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7587 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7588 LHS, RHS, I->getName());
7591 case Instruction::Trunc:
7592 case Instruction::ZExt:
7593 case Instruction::SExt:
7594 // If the source type of the cast is the type we're trying for then we can
7595 // just return the source. There's no need to insert it because it is not
7597 if (I->getOperand(0)->getType() == Ty)
7598 return I->getOperand(0);
7600 // Otherwise, must be the same type of case, so just reinsert a new one.
7601 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7605 // TODO: Can handle more cases here.
7606 assert(0 && "Unreachable!");
7610 return InsertNewInstBefore(Res, *I);
7613 /// @brief Implement the transforms common to all CastInst visitors.
7614 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7615 Value *Src = CI.getOperand(0);
7617 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7618 // eliminate it now.
7619 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7620 if (Instruction::CastOps opc =
7621 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7622 // The first cast (CSrc) is eliminable so we need to fix up or replace
7623 // the second cast (CI). CSrc will then have a good chance of being dead.
7624 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7628 // If we are casting a select then fold the cast into the select
7629 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7630 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7633 // If we are casting a PHI then fold the cast into the PHI
7634 if (isa<PHINode>(Src))
7635 if (Instruction *NV = FoldOpIntoPhi(CI))
7641 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7642 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7643 Value *Src = CI.getOperand(0);
7645 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7646 // If casting the result of a getelementptr instruction with no offset, turn
7647 // this into a cast of the original pointer!
7648 if (GEP->hasAllZeroIndices()) {
7649 // Changing the cast operand is usually not a good idea but it is safe
7650 // here because the pointer operand is being replaced with another
7651 // pointer operand so the opcode doesn't need to change.
7653 CI.setOperand(0, GEP->getOperand(0));
7657 // If the GEP has a single use, and the base pointer is a bitcast, and the
7658 // GEP computes a constant offset, see if we can convert these three
7659 // instructions into fewer. This typically happens with unions and other
7660 // non-type-safe code.
7661 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7662 if (GEP->hasAllConstantIndices()) {
7663 // We are guaranteed to get a constant from EmitGEPOffset.
7664 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7665 int64_t Offset = OffsetV->getSExtValue();
7667 // Get the base pointer input of the bitcast, and the type it points to.
7668 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7669 const Type *GEPIdxTy =
7670 cast<PointerType>(OrigBase->getType())->getElementType();
7671 if (GEPIdxTy->isSized()) {
7672 SmallVector<Value*, 8> NewIndices;
7674 // Start with the index over the outer type. Note that the type size
7675 // might be zero (even if the offset isn't zero) if the indexed type
7676 // is something like [0 x {int, int}]
7677 const Type *IntPtrTy = TD->getIntPtrType();
7678 int64_t FirstIdx = 0;
7679 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7680 FirstIdx = Offset/TySize;
7683 // Handle silly modulus not returning values values [0..TySize).
7687 assert(Offset >= 0);
7689 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7692 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7694 // Index into the types. If we fail, set OrigBase to null.
7696 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7697 const StructLayout *SL = TD->getStructLayout(STy);
7698 if (Offset < (int64_t)SL->getSizeInBytes()) {
7699 unsigned Elt = SL->getElementContainingOffset(Offset);
7700 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7702 Offset -= SL->getElementOffset(Elt);
7703 GEPIdxTy = STy->getElementType(Elt);
7705 // Otherwise, we can't index into this, bail out.
7709 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7710 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7711 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7712 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7715 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7717 GEPIdxTy = STy->getElementType();
7719 // Otherwise, we can't index into this, bail out.
7725 // If we were able to index down into an element, create the GEP
7726 // and bitcast the result. This eliminates one bitcast, potentially
7728 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7730 NewIndices.end(), "");
7731 InsertNewInstBefore(NGEP, CI);
7732 NGEP->takeName(GEP);
7734 if (isa<BitCastInst>(CI))
7735 return new BitCastInst(NGEP, CI.getType());
7736 assert(isa<PtrToIntInst>(CI));
7737 return new PtrToIntInst(NGEP, CI.getType());
7744 return commonCastTransforms(CI);
7749 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7750 /// integer types. This function implements the common transforms for all those
7752 /// @brief Implement the transforms common to CastInst with integer operands
7753 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7754 if (Instruction *Result = commonCastTransforms(CI))
7757 Value *Src = CI.getOperand(0);
7758 const Type *SrcTy = Src->getType();
7759 const Type *DestTy = CI.getType();
7760 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7761 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7763 // See if we can simplify any instructions used by the LHS whose sole
7764 // purpose is to compute bits we don't care about.
7765 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7766 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7767 KnownZero, KnownOne))
7770 // If the source isn't an instruction or has more than one use then we
7771 // can't do anything more.
7772 Instruction *SrcI = dyn_cast<Instruction>(Src);
7773 if (!SrcI || !Src->hasOneUse())
7776 // Attempt to propagate the cast into the instruction for int->int casts.
7777 int NumCastsRemoved = 0;
7778 if (!isa<BitCastInst>(CI) &&
7779 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7780 CI.getOpcode(), NumCastsRemoved)) {
7781 // If this cast is a truncate, evaluting in a different type always
7782 // eliminates the cast, so it is always a win. If this is a zero-extension,
7783 // we need to do an AND to maintain the clear top-part of the computation,
7784 // so we require that the input have eliminated at least one cast. If this
7785 // is a sign extension, we insert two new casts (to do the extension) so we
7786 // require that two casts have been eliminated.
7788 switch (CI.getOpcode()) {
7790 // All the others use floating point so we shouldn't actually
7791 // get here because of the check above.
7792 assert(0 && "Unknown cast type");
7793 case Instruction::Trunc:
7796 case Instruction::ZExt:
7797 DoXForm = NumCastsRemoved >= 1;
7799 case Instruction::SExt:
7800 DoXForm = NumCastsRemoved >= 2;
7805 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7806 CI.getOpcode() == Instruction::SExt);
7807 assert(Res->getType() == DestTy);
7808 switch (CI.getOpcode()) {
7809 default: assert(0 && "Unknown cast type!");
7810 case Instruction::Trunc:
7811 case Instruction::BitCast:
7812 // Just replace this cast with the result.
7813 return ReplaceInstUsesWith(CI, Res);
7814 case Instruction::ZExt: {
7815 // We need to emit an AND to clear the high bits.
7816 assert(SrcBitSize < DestBitSize && "Not a zext?");
7817 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7819 return BinaryOperator::CreateAnd(Res, C);
7821 case Instruction::SExt:
7822 // We need to emit a cast to truncate, then a cast to sext.
7823 return CastInst::Create(Instruction::SExt,
7824 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7830 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7831 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7833 switch (SrcI->getOpcode()) {
7834 case Instruction::Add:
7835 case Instruction::Mul:
7836 case Instruction::And:
7837 case Instruction::Or:
7838 case Instruction::Xor:
7839 // If we are discarding information, rewrite.
7840 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7841 // Don't insert two casts if they cannot be eliminated. We allow
7842 // two casts to be inserted if the sizes are the same. This could
7843 // only be converting signedness, which is a noop.
7844 if (DestBitSize == SrcBitSize ||
7845 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7846 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7847 Instruction::CastOps opcode = CI.getOpcode();
7848 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7849 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7850 return BinaryOperator::Create(
7851 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7855 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7856 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7857 SrcI->getOpcode() == Instruction::Xor &&
7858 Op1 == ConstantInt::getTrue() &&
7859 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7860 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7861 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7864 case Instruction::SDiv:
7865 case Instruction::UDiv:
7866 case Instruction::SRem:
7867 case Instruction::URem:
7868 // If we are just changing the sign, rewrite.
7869 if (DestBitSize == SrcBitSize) {
7870 // Don't insert two casts if they cannot be eliminated. We allow
7871 // two casts to be inserted if the sizes are the same. This could
7872 // only be converting signedness, which is a noop.
7873 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7874 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7875 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7877 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7879 return BinaryOperator::Create(
7880 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7885 case Instruction::Shl:
7886 // Allow changing the sign of the source operand. Do not allow
7887 // changing the size of the shift, UNLESS the shift amount is a
7888 // constant. We must not change variable sized shifts to a smaller
7889 // size, because it is undefined to shift more bits out than exist
7891 if (DestBitSize == SrcBitSize ||
7892 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7893 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7894 Instruction::BitCast : Instruction::Trunc);
7895 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7896 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7897 return BinaryOperator::CreateShl(Op0c, Op1c);
7900 case Instruction::AShr:
7901 // If this is a signed shr, and if all bits shifted in are about to be
7902 // truncated off, turn it into an unsigned shr to allow greater
7904 if (DestBitSize < SrcBitSize &&
7905 isa<ConstantInt>(Op1)) {
7906 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7907 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7908 // Insert the new logical shift right.
7909 return BinaryOperator::CreateLShr(Op0, Op1);
7917 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7918 if (Instruction *Result = commonIntCastTransforms(CI))
7921 Value *Src = CI.getOperand(0);
7922 const Type *Ty = CI.getType();
7923 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7924 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7926 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7927 switch (SrcI->getOpcode()) {
7929 case Instruction::LShr:
7930 // We can shrink lshr to something smaller if we know the bits shifted in
7931 // are already zeros.
7932 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7933 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7935 // Get a mask for the bits shifting in.
7936 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7937 Value* SrcIOp0 = SrcI->getOperand(0);
7938 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7939 if (ShAmt >= DestBitWidth) // All zeros.
7940 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7942 // Okay, we can shrink this. Truncate the input, then return a new
7944 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7945 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7947 return BinaryOperator::CreateLShr(V1, V2);
7949 } else { // This is a variable shr.
7951 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7952 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7953 // loop-invariant and CSE'd.
7954 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7955 Value *One = ConstantInt::get(SrcI->getType(), 1);
7957 Value *V = InsertNewInstBefore(
7958 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7960 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7961 SrcI->getOperand(0),
7963 Value *Zero = Constant::getNullValue(V->getType());
7964 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7974 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7975 /// in order to eliminate the icmp.
7976 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7978 // If we are just checking for a icmp eq of a single bit and zext'ing it
7979 // to an integer, then shift the bit to the appropriate place and then
7980 // cast to integer to avoid the comparison.
7981 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7982 const APInt &Op1CV = Op1C->getValue();
7984 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7985 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7986 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7987 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7988 if (!DoXform) return ICI;
7990 Value *In = ICI->getOperand(0);
7991 Value *Sh = ConstantInt::get(In->getType(),
7992 In->getType()->getPrimitiveSizeInBits()-1);
7993 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7994 In->getName()+".lobit"),
7996 if (In->getType() != CI.getType())
7997 In = CastInst::CreateIntegerCast(In, CI.getType(),
7998 false/*ZExt*/, "tmp", &CI);
8000 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8001 Constant *One = ConstantInt::get(In->getType(), 1);
8002 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8003 In->getName()+".not"),
8007 return ReplaceInstUsesWith(CI, In);
8012 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8013 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8014 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8015 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8016 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8017 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8018 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8019 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8020 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8021 // This only works for EQ and NE
8022 ICI->isEquality()) {
8023 // If Op1C some other power of two, convert:
8024 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8025 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8026 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8027 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8029 APInt KnownZeroMask(~KnownZero);
8030 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8031 if (!DoXform) return ICI;
8033 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8034 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8035 // (X&4) == 2 --> false
8036 // (X&4) != 2 --> true
8037 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8038 Res = ConstantExpr::getZExt(Res, CI.getType());
8039 return ReplaceInstUsesWith(CI, Res);
8042 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8043 Value *In = ICI->getOperand(0);
8045 // Perform a logical shr by shiftamt.
8046 // Insert the shift to put the result in the low bit.
8047 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8048 ConstantInt::get(In->getType(), ShiftAmt),
8049 In->getName()+".lobit"), CI);
8052 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8053 Constant *One = ConstantInt::get(In->getType(), 1);
8054 In = BinaryOperator::CreateXor(In, One, "tmp");
8055 InsertNewInstBefore(cast<Instruction>(In), CI);
8058 if (CI.getType() == In->getType())
8059 return ReplaceInstUsesWith(CI, In);
8061 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8069 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8070 // If one of the common conversion will work ..
8071 if (Instruction *Result = commonIntCastTransforms(CI))
8074 Value *Src = CI.getOperand(0);
8076 // If this is a cast of a cast
8077 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8078 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8079 // types and if the sizes are just right we can convert this into a logical
8080 // 'and' which will be much cheaper than the pair of casts.
8081 if (isa<TruncInst>(CSrc)) {
8082 // Get the sizes of the types involved
8083 Value *A = CSrc->getOperand(0);
8084 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8085 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8086 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8087 // If we're actually extending zero bits and the trunc is a no-op
8088 if (MidSize < DstSize && SrcSize == DstSize) {
8089 // Replace both of the casts with an And of the type mask.
8090 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8091 Constant *AndConst = ConstantInt::get(AndValue);
8093 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8094 // Unfortunately, if the type changed, we need to cast it back.
8095 if (And->getType() != CI.getType()) {
8096 And->setName(CSrc->getName()+".mask");
8097 InsertNewInstBefore(And, CI);
8098 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8105 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8106 return transformZExtICmp(ICI, CI);
8108 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8109 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8110 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8111 // of the (zext icmp) will be transformed.
8112 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8113 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8114 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8115 (transformZExtICmp(LHS, CI, false) ||
8116 transformZExtICmp(RHS, CI, false))) {
8117 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8118 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8119 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8126 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8127 if (Instruction *I = commonIntCastTransforms(CI))
8130 Value *Src = CI.getOperand(0);
8132 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
8133 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
8134 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
8135 // If we are just checking for a icmp eq of a single bit and zext'ing it
8136 // to an integer, then shift the bit to the appropriate place and then
8137 // cast to integer to avoid the comparison.
8138 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8139 const APInt &Op1CV = Op1C->getValue();
8141 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8142 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8143 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8144 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
8145 Value *In = ICI->getOperand(0);
8146 Value *Sh = ConstantInt::get(In->getType(),
8147 In->getType()->getPrimitiveSizeInBits()-1);
8148 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8149 In->getName()+".lobit"),
8151 if (In->getType() != CI.getType())
8152 In = CastInst::CreateIntegerCast(In, CI.getType(),
8153 true/*SExt*/, "tmp", &CI);
8155 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
8156 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8157 In->getName()+".not"), CI);
8159 return ReplaceInstUsesWith(CI, In);
8164 // See if the value being truncated is already sign extended. If so, just
8165 // eliminate the trunc/sext pair.
8166 if (getOpcode(Src) == Instruction::Trunc) {
8167 Value *Op = cast<User>(Src)->getOperand(0);
8168 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8169 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8170 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8171 unsigned NumSignBits = ComputeNumSignBits(Op);
8173 if (OpBits == DestBits) {
8174 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8175 // bits, it is already ready.
8176 if (NumSignBits > DestBits-MidBits)
8177 return ReplaceInstUsesWith(CI, Op);
8178 } else if (OpBits < DestBits) {
8179 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8180 // bits, just sext from i32.
8181 if (NumSignBits > OpBits-MidBits)
8182 return new SExtInst(Op, CI.getType(), "tmp");
8184 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8185 // bits, just truncate to i32.
8186 if (NumSignBits > OpBits-MidBits)
8187 return new TruncInst(Op, CI.getType(), "tmp");
8194 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8195 /// in the specified FP type without changing its value.
8196 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8197 APFloat F = CFP->getValueAPF();
8198 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
8199 return ConstantFP::get(F);
8203 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8204 /// through it until we get the source value.
8205 static Value *LookThroughFPExtensions(Value *V) {
8206 if (Instruction *I = dyn_cast<Instruction>(V))
8207 if (I->getOpcode() == Instruction::FPExt)
8208 return LookThroughFPExtensions(I->getOperand(0));
8210 // If this value is a constant, return the constant in the smallest FP type
8211 // that can accurately represent it. This allows us to turn
8212 // (float)((double)X+2.0) into x+2.0f.
8213 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8214 if (CFP->getType() == Type::PPC_FP128Ty)
8215 return V; // No constant folding of this.
8216 // See if the value can be truncated to float and then reextended.
8217 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8219 if (CFP->getType() == Type::DoubleTy)
8220 return V; // Won't shrink.
8221 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8223 // Don't try to shrink to various long double types.
8229 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8230 if (Instruction *I = commonCastTransforms(CI))
8233 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8234 // smaller than the destination type, we can eliminate the truncate by doing
8235 // the add as the smaller type. This applies to add/sub/mul/div as well as
8236 // many builtins (sqrt, etc).
8237 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8238 if (OpI && OpI->hasOneUse()) {
8239 switch (OpI->getOpcode()) {
8241 case Instruction::Add:
8242 case Instruction::Sub:
8243 case Instruction::Mul:
8244 case Instruction::FDiv:
8245 case Instruction::FRem:
8246 const Type *SrcTy = OpI->getType();
8247 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8248 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8249 if (LHSTrunc->getType() != SrcTy &&
8250 RHSTrunc->getType() != SrcTy) {
8251 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8252 // If the source types were both smaller than the destination type of
8253 // the cast, do this xform.
8254 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8255 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8256 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8258 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8260 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8269 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8270 return commonCastTransforms(CI);
8273 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8274 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
8275 // mantissa to accurately represent all values of X. For example, do not
8276 // do this with i64->float->i64.
8277 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
8278 if (SrcI->getOperand(0)->getType() == FI.getType() &&
8279 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8280 SrcI->getType()->getFPMantissaWidth())
8281 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
8283 return commonCastTransforms(FI);
8286 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8287 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
8288 // mantissa to accurately represent all values of X. For example, do not
8289 // do this with i64->float->i64.
8290 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
8291 if (SrcI->getOperand(0)->getType() == FI.getType() &&
8292 (int)FI.getType()->getPrimitiveSizeInBits() <=
8293 SrcI->getType()->getFPMantissaWidth())
8294 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
8296 return commonCastTransforms(FI);
8299 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8300 return commonCastTransforms(CI);
8303 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8304 return commonCastTransforms(CI);
8307 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8308 return commonPointerCastTransforms(CI);
8311 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8312 if (Instruction *I = commonCastTransforms(CI))
8315 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8316 if (!DestPointee->isSized()) return 0;
8318 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8321 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8322 m_ConstantInt(Cst)))) {
8323 // If the source and destination operands have the same type, see if this
8324 // is a single-index GEP.
8325 if (X->getType() == CI.getType()) {
8326 // Get the size of the pointee type.
8327 uint64_t Size = TD->getABITypeSize(DestPointee);
8329 // Convert the constant to intptr type.
8330 APInt Offset = Cst->getValue();
8331 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8333 // If Offset is evenly divisible by Size, we can do this xform.
8334 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8335 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8336 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8339 // TODO: Could handle other cases, e.g. where add is indexing into field of
8341 } else if (CI.getOperand(0)->hasOneUse() &&
8342 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8343 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8344 // "inttoptr+GEP" instead of "add+intptr".
8346 // Get the size of the pointee type.
8347 uint64_t Size = TD->getABITypeSize(DestPointee);
8349 // Convert the constant to intptr type.
8350 APInt Offset = Cst->getValue();
8351 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8353 // If Offset is evenly divisible by Size, we can do this xform.
8354 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8355 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8357 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8359 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8365 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8366 // If the operands are integer typed then apply the integer transforms,
8367 // otherwise just apply the common ones.
8368 Value *Src = CI.getOperand(0);
8369 const Type *SrcTy = Src->getType();
8370 const Type *DestTy = CI.getType();
8372 if (SrcTy->isInteger() && DestTy->isInteger()) {
8373 if (Instruction *Result = commonIntCastTransforms(CI))
8375 } else if (isa<PointerType>(SrcTy)) {
8376 if (Instruction *I = commonPointerCastTransforms(CI))
8379 if (Instruction *Result = commonCastTransforms(CI))
8384 // Get rid of casts from one type to the same type. These are useless and can
8385 // be replaced by the operand.
8386 if (DestTy == Src->getType())
8387 return ReplaceInstUsesWith(CI, Src);
8389 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8390 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8391 const Type *DstElTy = DstPTy->getElementType();
8392 const Type *SrcElTy = SrcPTy->getElementType();
8394 // If the address spaces don't match, don't eliminate the bitcast, which is
8395 // required for changing types.
8396 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8399 // If we are casting a malloc or alloca to a pointer to a type of the same
8400 // size, rewrite the allocation instruction to allocate the "right" type.
8401 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8402 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8405 // If the source and destination are pointers, and this cast is equivalent
8406 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8407 // This can enhance SROA and other transforms that want type-safe pointers.
8408 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8409 unsigned NumZeros = 0;
8410 while (SrcElTy != DstElTy &&
8411 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8412 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8413 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8417 // If we found a path from the src to dest, create the getelementptr now.
8418 if (SrcElTy == DstElTy) {
8419 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8420 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8421 ((Instruction*) NULL));
8425 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8426 if (SVI->hasOneUse()) {
8427 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8428 // a bitconvert to a vector with the same # elts.
8429 if (isa<VectorType>(DestTy) &&
8430 cast<VectorType>(DestTy)->getNumElements() ==
8431 SVI->getType()->getNumElements()) {
8433 // If either of the operands is a cast from CI.getType(), then
8434 // evaluating the shuffle in the casted destination's type will allow
8435 // us to eliminate at least one cast.
8436 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8437 Tmp->getOperand(0)->getType() == DestTy) ||
8438 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8439 Tmp->getOperand(0)->getType() == DestTy)) {
8440 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8441 SVI->getOperand(0), DestTy, &CI);
8442 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8443 SVI->getOperand(1), DestTy, &CI);
8444 // Return a new shuffle vector. Use the same element ID's, as we
8445 // know the vector types match #elts.
8446 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8454 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8456 /// %D = select %cond, %C, %A
8458 /// %C = select %cond, %B, 0
8461 /// Assuming that the specified instruction is an operand to the select, return
8462 /// a bitmask indicating which operands of this instruction are foldable if they
8463 /// equal the other incoming value of the select.
8465 static unsigned GetSelectFoldableOperands(Instruction *I) {
8466 switch (I->getOpcode()) {
8467 case Instruction::Add:
8468 case Instruction::Mul:
8469 case Instruction::And:
8470 case Instruction::Or:
8471 case Instruction::Xor:
8472 return 3; // Can fold through either operand.
8473 case Instruction::Sub: // Can only fold on the amount subtracted.
8474 case Instruction::Shl: // Can only fold on the shift amount.
8475 case Instruction::LShr:
8476 case Instruction::AShr:
8479 return 0; // Cannot fold
8483 /// GetSelectFoldableConstant - For the same transformation as the previous
8484 /// function, return the identity constant that goes into the select.
8485 static Constant *GetSelectFoldableConstant(Instruction *I) {
8486 switch (I->getOpcode()) {
8487 default: assert(0 && "This cannot happen!"); abort();
8488 case Instruction::Add:
8489 case Instruction::Sub:
8490 case Instruction::Or:
8491 case Instruction::Xor:
8492 case Instruction::Shl:
8493 case Instruction::LShr:
8494 case Instruction::AShr:
8495 return Constant::getNullValue(I->getType());
8496 case Instruction::And:
8497 return Constant::getAllOnesValue(I->getType());
8498 case Instruction::Mul:
8499 return ConstantInt::get(I->getType(), 1);
8503 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8504 /// have the same opcode and only one use each. Try to simplify this.
8505 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8507 if (TI->getNumOperands() == 1) {
8508 // If this is a non-volatile load or a cast from the same type,
8511 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8514 return 0; // unknown unary op.
8517 // Fold this by inserting a select from the input values.
8518 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8519 FI->getOperand(0), SI.getName()+".v");
8520 InsertNewInstBefore(NewSI, SI);
8521 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8525 // Only handle binary operators here.
8526 if (!isa<BinaryOperator>(TI))
8529 // Figure out if the operations have any operands in common.
8530 Value *MatchOp, *OtherOpT, *OtherOpF;
8532 if (TI->getOperand(0) == FI->getOperand(0)) {
8533 MatchOp = TI->getOperand(0);
8534 OtherOpT = TI->getOperand(1);
8535 OtherOpF = FI->getOperand(1);
8536 MatchIsOpZero = true;
8537 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8538 MatchOp = TI->getOperand(1);
8539 OtherOpT = TI->getOperand(0);
8540 OtherOpF = FI->getOperand(0);
8541 MatchIsOpZero = false;
8542 } else if (!TI->isCommutative()) {
8544 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8545 MatchOp = TI->getOperand(0);
8546 OtherOpT = TI->getOperand(1);
8547 OtherOpF = FI->getOperand(0);
8548 MatchIsOpZero = true;
8549 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8550 MatchOp = TI->getOperand(1);
8551 OtherOpT = TI->getOperand(0);
8552 OtherOpF = FI->getOperand(1);
8553 MatchIsOpZero = true;
8558 // If we reach here, they do have operations in common.
8559 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8560 OtherOpF, SI.getName()+".v");
8561 InsertNewInstBefore(NewSI, SI);
8563 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8565 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8567 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8569 assert(0 && "Shouldn't get here");
8573 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8574 Value *CondVal = SI.getCondition();
8575 Value *TrueVal = SI.getTrueValue();
8576 Value *FalseVal = SI.getFalseValue();
8578 // select true, X, Y -> X
8579 // select false, X, Y -> Y
8580 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8581 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8583 // select C, X, X -> X
8584 if (TrueVal == FalseVal)
8585 return ReplaceInstUsesWith(SI, TrueVal);
8587 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8588 return ReplaceInstUsesWith(SI, FalseVal);
8589 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8590 return ReplaceInstUsesWith(SI, TrueVal);
8591 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8592 if (isa<Constant>(TrueVal))
8593 return ReplaceInstUsesWith(SI, TrueVal);
8595 return ReplaceInstUsesWith(SI, FalseVal);
8598 if (SI.getType() == Type::Int1Ty) {
8599 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8600 if (C->getZExtValue()) {
8601 // Change: A = select B, true, C --> A = or B, C
8602 return BinaryOperator::CreateOr(CondVal, FalseVal);
8604 // Change: A = select B, false, C --> A = and !B, C
8606 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8607 "not."+CondVal->getName()), SI);
8608 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8610 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8611 if (C->getZExtValue() == false) {
8612 // Change: A = select B, C, false --> A = and B, C
8613 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8615 // Change: A = select B, C, true --> A = or !B, C
8617 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8618 "not."+CondVal->getName()), SI);
8619 return BinaryOperator::CreateOr(NotCond, TrueVal);
8623 // select a, b, a -> a&b
8624 // select a, a, b -> a|b
8625 if (CondVal == TrueVal)
8626 return BinaryOperator::CreateOr(CondVal, FalseVal);
8627 else if (CondVal == FalseVal)
8628 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8631 // Selecting between two integer constants?
8632 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8633 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8634 // select C, 1, 0 -> zext C to int
8635 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8636 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8637 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8638 // select C, 0, 1 -> zext !C to int
8640 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8641 "not."+CondVal->getName()), SI);
8642 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8645 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8647 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8649 // (x <s 0) ? -1 : 0 -> ashr x, 31
8650 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8651 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8652 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8653 // The comparison constant and the result are not neccessarily the
8654 // same width. Make an all-ones value by inserting a AShr.
8655 Value *X = IC->getOperand(0);
8656 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8657 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8658 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8660 InsertNewInstBefore(SRA, SI);
8662 // Finally, convert to the type of the select RHS. We figure out
8663 // if this requires a SExt, Trunc or BitCast based on the sizes.
8664 Instruction::CastOps opc = Instruction::BitCast;
8665 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8666 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8667 if (SRASize < SISize)
8668 opc = Instruction::SExt;
8669 else if (SRASize > SISize)
8670 opc = Instruction::Trunc;
8671 return CastInst::Create(opc, SRA, SI.getType());
8676 // If one of the constants is zero (we know they can't both be) and we
8677 // have an icmp instruction with zero, and we have an 'and' with the
8678 // non-constant value, eliminate this whole mess. This corresponds to
8679 // cases like this: ((X & 27) ? 27 : 0)
8680 if (TrueValC->isZero() || FalseValC->isZero())
8681 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8682 cast<Constant>(IC->getOperand(1))->isNullValue())
8683 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8684 if (ICA->getOpcode() == Instruction::And &&
8685 isa<ConstantInt>(ICA->getOperand(1)) &&
8686 (ICA->getOperand(1) == TrueValC ||
8687 ICA->getOperand(1) == FalseValC) &&
8688 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8689 // Okay, now we know that everything is set up, we just don't
8690 // know whether we have a icmp_ne or icmp_eq and whether the
8691 // true or false val is the zero.
8692 bool ShouldNotVal = !TrueValC->isZero();
8693 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8696 V = InsertNewInstBefore(BinaryOperator::Create(
8697 Instruction::Xor, V, ICA->getOperand(1)), SI);
8698 return ReplaceInstUsesWith(SI, V);
8703 // See if we are selecting two values based on a comparison of the two values.
8704 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8705 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8706 // Transform (X == Y) ? X : Y -> Y
8707 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8708 // This is not safe in general for floating point:
8709 // consider X== -0, Y== +0.
8710 // It becomes safe if either operand is a nonzero constant.
8711 ConstantFP *CFPt, *CFPf;
8712 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8713 !CFPt->getValueAPF().isZero()) ||
8714 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8715 !CFPf->getValueAPF().isZero()))
8716 return ReplaceInstUsesWith(SI, FalseVal);
8718 // Transform (X != Y) ? X : Y -> X
8719 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8720 return ReplaceInstUsesWith(SI, TrueVal);
8721 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8723 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8724 // Transform (X == Y) ? Y : X -> X
8725 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8726 // This is not safe in general for floating point:
8727 // consider X== -0, Y== +0.
8728 // It becomes safe if either operand is a nonzero constant.
8729 ConstantFP *CFPt, *CFPf;
8730 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8731 !CFPt->getValueAPF().isZero()) ||
8732 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8733 !CFPf->getValueAPF().isZero()))
8734 return ReplaceInstUsesWith(SI, FalseVal);
8736 // Transform (X != Y) ? Y : X -> Y
8737 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8738 return ReplaceInstUsesWith(SI, TrueVal);
8739 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8743 // See if we are selecting two values based on a comparison of the two values.
8744 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8745 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8746 // Transform (X == Y) ? X : Y -> Y
8747 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8748 return ReplaceInstUsesWith(SI, FalseVal);
8749 // Transform (X != Y) ? X : Y -> X
8750 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8751 return ReplaceInstUsesWith(SI, TrueVal);
8752 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8754 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8755 // Transform (X == Y) ? Y : X -> X
8756 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8757 return ReplaceInstUsesWith(SI, FalseVal);
8758 // Transform (X != Y) ? Y : X -> Y
8759 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8760 return ReplaceInstUsesWith(SI, TrueVal);
8761 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8765 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8766 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8767 if (TI->hasOneUse() && FI->hasOneUse()) {
8768 Instruction *AddOp = 0, *SubOp = 0;
8770 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8771 if (TI->getOpcode() == FI->getOpcode())
8772 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8775 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8776 // even legal for FP.
8777 if (TI->getOpcode() == Instruction::Sub &&
8778 FI->getOpcode() == Instruction::Add) {
8779 AddOp = FI; SubOp = TI;
8780 } else if (FI->getOpcode() == Instruction::Sub &&
8781 TI->getOpcode() == Instruction::Add) {
8782 AddOp = TI; SubOp = FI;
8786 Value *OtherAddOp = 0;
8787 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8788 OtherAddOp = AddOp->getOperand(1);
8789 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8790 OtherAddOp = AddOp->getOperand(0);
8794 // So at this point we know we have (Y -> OtherAddOp):
8795 // select C, (add X, Y), (sub X, Z)
8796 Value *NegVal; // Compute -Z
8797 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8798 NegVal = ConstantExpr::getNeg(C);
8800 NegVal = InsertNewInstBefore(
8801 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8804 Value *NewTrueOp = OtherAddOp;
8805 Value *NewFalseOp = NegVal;
8807 std::swap(NewTrueOp, NewFalseOp);
8808 Instruction *NewSel =
8809 SelectInst::Create(CondVal, NewTrueOp,
8810 NewFalseOp, SI.getName() + ".p");
8812 NewSel = InsertNewInstBefore(NewSel, SI);
8813 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8818 // See if we can fold the select into one of our operands.
8819 if (SI.getType()->isInteger()) {
8820 // See the comment above GetSelectFoldableOperands for a description of the
8821 // transformation we are doing here.
8822 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8823 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8824 !isa<Constant>(FalseVal))
8825 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8826 unsigned OpToFold = 0;
8827 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8829 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8834 Constant *C = GetSelectFoldableConstant(TVI);
8835 Instruction *NewSel =
8836 SelectInst::Create(SI.getCondition(),
8837 TVI->getOperand(2-OpToFold), C);
8838 InsertNewInstBefore(NewSel, SI);
8839 NewSel->takeName(TVI);
8840 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8841 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8843 assert(0 && "Unknown instruction!!");
8848 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8849 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8850 !isa<Constant>(TrueVal))
8851 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8852 unsigned OpToFold = 0;
8853 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8855 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8860 Constant *C = GetSelectFoldableConstant(FVI);
8861 Instruction *NewSel =
8862 SelectInst::Create(SI.getCondition(), C,
8863 FVI->getOperand(2-OpToFold));
8864 InsertNewInstBefore(NewSel, SI);
8865 NewSel->takeName(FVI);
8866 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8867 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8869 assert(0 && "Unknown instruction!!");
8874 if (BinaryOperator::isNot(CondVal)) {
8875 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8876 SI.setOperand(1, FalseVal);
8877 SI.setOperand(2, TrueVal);
8884 /// EnforceKnownAlignment - If the specified pointer points to an object that
8885 /// we control, modify the object's alignment to PrefAlign. This isn't
8886 /// often possible though. If alignment is important, a more reliable approach
8887 /// is to simply align all global variables and allocation instructions to
8888 /// their preferred alignment from the beginning.
8890 static unsigned EnforceKnownAlignment(Value *V,
8891 unsigned Align, unsigned PrefAlign) {
8893 User *U = dyn_cast<User>(V);
8894 if (!U) return Align;
8896 switch (getOpcode(U)) {
8898 case Instruction::BitCast:
8899 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8900 case Instruction::GetElementPtr: {
8901 // If all indexes are zero, it is just the alignment of the base pointer.
8902 bool AllZeroOperands = true;
8903 for (unsigned i = 1, e = U->getNumOperands(); i != e; ++i)
8904 if (!isa<Constant>(U->getOperand(i)) ||
8905 !cast<Constant>(U->getOperand(i))->isNullValue()) {
8906 AllZeroOperands = false;
8910 if (AllZeroOperands) {
8911 // Treat this like a bitcast.
8912 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8918 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8919 // If there is a large requested alignment and we can, bump up the alignment
8921 if (!GV->isDeclaration()) {
8922 GV->setAlignment(PrefAlign);
8925 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8926 // If there is a requested alignment and if this is an alloca, round up. We
8927 // don't do this for malloc, because some systems can't respect the request.
8928 if (isa<AllocaInst>(AI)) {
8929 AI->setAlignment(PrefAlign);
8937 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8938 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8939 /// and it is more than the alignment of the ultimate object, see if we can
8940 /// increase the alignment of the ultimate object, making this check succeed.
8941 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8942 unsigned PrefAlign) {
8943 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8944 sizeof(PrefAlign) * CHAR_BIT;
8945 APInt Mask = APInt::getAllOnesValue(BitWidth);
8946 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8947 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8948 unsigned TrailZ = KnownZero.countTrailingOnes();
8949 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8951 if (PrefAlign > Align)
8952 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8954 // We don't need to make any adjustment.
8958 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8959 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8960 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8961 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8962 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8964 if (CopyAlign < MinAlign) {
8965 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8969 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8971 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8972 if (MemOpLength == 0) return 0;
8974 // Source and destination pointer types are always "i8*" for intrinsic. See
8975 // if the size is something we can handle with a single primitive load/store.
8976 // A single load+store correctly handles overlapping memory in the memmove
8978 unsigned Size = MemOpLength->getZExtValue();
8979 if (Size == 0) return MI; // Delete this mem transfer.
8981 if (Size > 8 || (Size&(Size-1)))
8982 return 0; // If not 1/2/4/8 bytes, exit.
8984 // Use an integer load+store unless we can find something better.
8985 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8987 // Memcpy forces the use of i8* for the source and destination. That means
8988 // that if you're using memcpy to move one double around, you'll get a cast
8989 // from double* to i8*. We'd much rather use a double load+store rather than
8990 // an i64 load+store, here because this improves the odds that the source or
8991 // dest address will be promotable. See if we can find a better type than the
8992 // integer datatype.
8993 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8994 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8995 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8996 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8997 // down through these levels if so.
8998 while (!SrcETy->isSingleValueType()) {
8999 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9000 if (STy->getNumElements() == 1)
9001 SrcETy = STy->getElementType(0);
9004 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9005 if (ATy->getNumElements() == 1)
9006 SrcETy = ATy->getElementType();
9013 if (SrcETy->isSingleValueType())
9014 NewPtrTy = PointerType::getUnqual(SrcETy);
9019 // If the memcpy/memmove provides better alignment info than we can
9021 SrcAlign = std::max(SrcAlign, CopyAlign);
9022 DstAlign = std::max(DstAlign, CopyAlign);
9024 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9025 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9026 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9027 InsertNewInstBefore(L, *MI);
9028 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9030 // Set the size of the copy to 0, it will be deleted on the next iteration.
9031 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9035 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9036 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9037 if (MI->getAlignment()->getZExtValue() < Alignment) {
9038 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9042 // Extract the length and alignment and fill if they are constant.
9043 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9044 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9045 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9047 uint64_t Len = LenC->getZExtValue();
9048 Alignment = MI->getAlignment()->getZExtValue();
9050 // If the length is zero, this is a no-op
9051 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9053 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9054 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9055 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9057 Value *Dest = MI->getDest();
9058 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9060 // Alignment 0 is identity for alignment 1 for memset, but not store.
9061 if (Alignment == 0) Alignment = 1;
9063 // Extract the fill value and store.
9064 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9065 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9068 // Set the size of the copy to 0, it will be deleted on the next iteration.
9069 MI->setLength(Constant::getNullValue(LenC->getType()));
9077 /// visitCallInst - CallInst simplification. This mostly only handles folding
9078 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9079 /// the heavy lifting.
9081 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9082 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9083 if (!II) return visitCallSite(&CI);
9085 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9087 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9088 bool Changed = false;
9090 // memmove/cpy/set of zero bytes is a noop.
9091 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9092 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9094 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9095 if (CI->getZExtValue() == 1) {
9096 // Replace the instruction with just byte operations. We would
9097 // transform other cases to loads/stores, but we don't know if
9098 // alignment is sufficient.
9102 // If we have a memmove and the source operation is a constant global,
9103 // then the source and dest pointers can't alias, so we can change this
9104 // into a call to memcpy.
9105 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9106 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9107 if (GVSrc->isConstant()) {
9108 Module *M = CI.getParent()->getParent()->getParent();
9109 Intrinsic::ID MemCpyID;
9110 if (CI.getOperand(3)->getType() == Type::Int32Ty)
9111 MemCpyID = Intrinsic::memcpy_i32;
9113 MemCpyID = Intrinsic::memcpy_i64;
9114 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
9119 // If we can determine a pointer alignment that is bigger than currently
9120 // set, update the alignment.
9121 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9122 if (Instruction *I = SimplifyMemTransfer(MI))
9124 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9125 if (Instruction *I = SimplifyMemSet(MSI))
9129 if (Changed) return II;
9131 switch (II->getIntrinsicID()) {
9133 case Intrinsic::ppc_altivec_lvx:
9134 case Intrinsic::ppc_altivec_lvxl:
9135 case Intrinsic::x86_sse_loadu_ps:
9136 case Intrinsic::x86_sse2_loadu_pd:
9137 case Intrinsic::x86_sse2_loadu_dq:
9138 // Turn PPC lvx -> load if the pointer is known aligned.
9139 // Turn X86 loadups -> load if the pointer is known aligned.
9140 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9141 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9142 PointerType::getUnqual(II->getType()),
9144 return new LoadInst(Ptr);
9147 case Intrinsic::ppc_altivec_stvx:
9148 case Intrinsic::ppc_altivec_stvxl:
9149 // Turn stvx -> store if the pointer is known aligned.
9150 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9151 const Type *OpPtrTy =
9152 PointerType::getUnqual(II->getOperand(1)->getType());
9153 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9154 return new StoreInst(II->getOperand(1), Ptr);
9157 case Intrinsic::x86_sse_storeu_ps:
9158 case Intrinsic::x86_sse2_storeu_pd:
9159 case Intrinsic::x86_sse2_storeu_dq:
9160 case Intrinsic::x86_sse2_storel_dq:
9161 // Turn X86 storeu -> store if the pointer is known aligned.
9162 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9163 const Type *OpPtrTy =
9164 PointerType::getUnqual(II->getOperand(2)->getType());
9165 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9166 return new StoreInst(II->getOperand(2), Ptr);
9170 case Intrinsic::x86_sse_cvttss2si: {
9171 // These intrinsics only demands the 0th element of its input vector. If
9172 // we can simplify the input based on that, do so now.
9174 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9176 II->setOperand(1, V);
9182 case Intrinsic::ppc_altivec_vperm:
9183 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9184 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9185 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9187 // Check that all of the elements are integer constants or undefs.
9188 bool AllEltsOk = true;
9189 for (unsigned i = 0; i != 16; ++i) {
9190 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9191 !isa<UndefValue>(Mask->getOperand(i))) {
9198 // Cast the input vectors to byte vectors.
9199 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9200 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9201 Value *Result = UndefValue::get(Op0->getType());
9203 // Only extract each element once.
9204 Value *ExtractedElts[32];
9205 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9207 for (unsigned i = 0; i != 16; ++i) {
9208 if (isa<UndefValue>(Mask->getOperand(i)))
9210 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9211 Idx &= 31; // Match the hardware behavior.
9213 if (ExtractedElts[Idx] == 0) {
9215 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9216 InsertNewInstBefore(Elt, CI);
9217 ExtractedElts[Idx] = Elt;
9220 // Insert this value into the result vector.
9221 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9223 InsertNewInstBefore(cast<Instruction>(Result), CI);
9225 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9230 case Intrinsic::stackrestore: {
9231 // If the save is right next to the restore, remove the restore. This can
9232 // happen when variable allocas are DCE'd.
9233 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9234 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9235 BasicBlock::iterator BI = SS;
9237 return EraseInstFromFunction(CI);
9241 // Scan down this block to see if there is another stack restore in the
9242 // same block without an intervening call/alloca.
9243 BasicBlock::iterator BI = II;
9244 TerminatorInst *TI = II->getParent()->getTerminator();
9245 bool CannotRemove = false;
9246 for (++BI; &*BI != TI; ++BI) {
9247 if (isa<AllocaInst>(BI)) {
9248 CannotRemove = true;
9251 if (isa<CallInst>(BI)) {
9252 if (!isa<IntrinsicInst>(BI)) {
9253 CannotRemove = true;
9256 // If there is a stackrestore below this one, remove this one.
9257 return EraseInstFromFunction(CI);
9261 // If the stack restore is in a return/unwind block and if there are no
9262 // allocas or calls between the restore and the return, nuke the restore.
9263 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9264 return EraseInstFromFunction(CI);
9270 return visitCallSite(II);
9273 // InvokeInst simplification
9275 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9276 return visitCallSite(&II);
9279 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9280 /// passed through the varargs area, we can eliminate the use of the cast.
9281 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9282 const CastInst * const CI,
9283 const TargetData * const TD,
9285 if (!CI->isLosslessCast())
9288 // The size of ByVal arguments is derived from the type, so we
9289 // can't change to a type with a different size. If the size were
9290 // passed explicitly we could avoid this check.
9291 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
9295 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9296 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9297 if (!SrcTy->isSized() || !DstTy->isSized())
9299 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9304 // visitCallSite - Improvements for call and invoke instructions.
9306 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9307 bool Changed = false;
9309 // If the callee is a constexpr cast of a function, attempt to move the cast
9310 // to the arguments of the call/invoke.
9311 if (transformConstExprCastCall(CS)) return 0;
9313 Value *Callee = CS.getCalledValue();
9315 if (Function *CalleeF = dyn_cast<Function>(Callee))
9316 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9317 Instruction *OldCall = CS.getInstruction();
9318 // If the call and callee calling conventions don't match, this call must
9319 // be unreachable, as the call is undefined.
9320 new StoreInst(ConstantInt::getTrue(),
9321 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9323 if (!OldCall->use_empty())
9324 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9325 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9326 return EraseInstFromFunction(*OldCall);
9330 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9331 // This instruction is not reachable, just remove it. We insert a store to
9332 // undef so that we know that this code is not reachable, despite the fact
9333 // that we can't modify the CFG here.
9334 new StoreInst(ConstantInt::getTrue(),
9335 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9336 CS.getInstruction());
9338 if (!CS.getInstruction()->use_empty())
9339 CS.getInstruction()->
9340 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9342 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9343 // Don't break the CFG, insert a dummy cond branch.
9344 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9345 ConstantInt::getTrue(), II);
9347 return EraseInstFromFunction(*CS.getInstruction());
9350 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9351 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9352 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9353 return transformCallThroughTrampoline(CS);
9355 const PointerType *PTy = cast<PointerType>(Callee->getType());
9356 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9357 if (FTy->isVarArg()) {
9358 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9359 // See if we can optimize any arguments passed through the varargs area of
9361 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9362 E = CS.arg_end(); I != E; ++I, ++ix) {
9363 CastInst *CI = dyn_cast<CastInst>(*I);
9364 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9365 *I = CI->getOperand(0);
9371 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9372 // Inline asm calls cannot throw - mark them 'nounwind'.
9373 CS.setDoesNotThrow();
9377 return Changed ? CS.getInstruction() : 0;
9380 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9381 // attempt to move the cast to the arguments of the call/invoke.
9383 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9384 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9385 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9386 if (CE->getOpcode() != Instruction::BitCast ||
9387 !isa<Function>(CE->getOperand(0)))
9389 Function *Callee = cast<Function>(CE->getOperand(0));
9390 Instruction *Caller = CS.getInstruction();
9391 const PAListPtr &CallerPAL = CS.getParamAttrs();
9393 // Okay, this is a cast from a function to a different type. Unless doing so
9394 // would cause a type conversion of one of our arguments, change this call to
9395 // be a direct call with arguments casted to the appropriate types.
9397 const FunctionType *FT = Callee->getFunctionType();
9398 const Type *OldRetTy = Caller->getType();
9400 if (isa<StructType>(FT->getReturnType()))
9401 return false; // TODO: Handle multiple return values.
9403 // Check to see if we are changing the return type...
9404 if (OldRetTy != FT->getReturnType()) {
9405 if (Callee->isDeclaration() &&
9406 // Conversion is ok if changing from pointer to int of same size.
9407 !(isa<PointerType>(FT->getReturnType()) &&
9408 TD->getIntPtrType() == OldRetTy))
9409 return false; // Cannot transform this return value.
9411 if (!Caller->use_empty() &&
9412 // void -> non-void is handled specially
9413 FT->getReturnType() != Type::VoidTy &&
9414 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
9415 return false; // Cannot transform this return value.
9417 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9418 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9419 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
9420 return false; // Attribute not compatible with transformed value.
9423 // If the callsite is an invoke instruction, and the return value is used by
9424 // a PHI node in a successor, we cannot change the return type of the call
9425 // because there is no place to put the cast instruction (without breaking
9426 // the critical edge). Bail out in this case.
9427 if (!Caller->use_empty())
9428 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9429 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9431 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9432 if (PN->getParent() == II->getNormalDest() ||
9433 PN->getParent() == II->getUnwindDest())
9437 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9438 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9440 CallSite::arg_iterator AI = CS.arg_begin();
9441 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9442 const Type *ParamTy = FT->getParamType(i);
9443 const Type *ActTy = (*AI)->getType();
9445 if (!CastInst::isCastable(ActTy, ParamTy))
9446 return false; // Cannot transform this parameter value.
9448 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
9449 return false; // Attribute not compatible with transformed value.
9451 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
9452 // Some conversions are safe even if we do not have a body.
9453 // Either we can cast directly, or we can upconvert the argument
9454 bool isConvertible = ActTy == ParamTy ||
9455 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
9456 (ParamTy->isInteger() && ActTy->isInteger() &&
9457 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
9458 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
9459 && c->getValue().isStrictlyPositive());
9460 if (Callee->isDeclaration() && !isConvertible) return false;
9463 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9464 Callee->isDeclaration())
9465 return false; // Do not delete arguments unless we have a function body.
9467 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9468 !CallerPAL.isEmpty())
9469 // In this case we have more arguments than the new function type, but we
9470 // won't be dropping them. Check that these extra arguments have attributes
9471 // that are compatible with being a vararg call argument.
9472 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9473 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9475 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9476 if (PAttrs & ParamAttr::VarArgsIncompatible)
9480 // Okay, we decided that this is a safe thing to do: go ahead and start
9481 // inserting cast instructions as necessary...
9482 std::vector<Value*> Args;
9483 Args.reserve(NumActualArgs);
9484 SmallVector<ParamAttrsWithIndex, 8> attrVec;
9485 attrVec.reserve(NumCommonArgs);
9487 // Get any return attributes.
9488 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9490 // If the return value is not being used, the type may not be compatible
9491 // with the existing attributes. Wipe out any problematic attributes.
9492 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
9494 // Add the new return attributes.
9496 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
9498 AI = CS.arg_begin();
9499 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9500 const Type *ParamTy = FT->getParamType(i);
9501 if ((*AI)->getType() == ParamTy) {
9502 Args.push_back(*AI);
9504 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9505 false, ParamTy, false);
9506 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9507 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9510 // Add any parameter attributes.
9511 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9512 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9515 // If the function takes more arguments than the call was taking, add them
9517 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9518 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9520 // If we are removing arguments to the function, emit an obnoxious warning...
9521 if (FT->getNumParams() < NumActualArgs) {
9522 if (!FT->isVarArg()) {
9523 cerr << "WARNING: While resolving call to function '"
9524 << Callee->getName() << "' arguments were dropped!\n";
9526 // Add all of the arguments in their promoted form to the arg list...
9527 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9528 const Type *PTy = getPromotedType((*AI)->getType());
9529 if (PTy != (*AI)->getType()) {
9530 // Must promote to pass through va_arg area!
9531 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9533 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9534 InsertNewInstBefore(Cast, *Caller);
9535 Args.push_back(Cast);
9537 Args.push_back(*AI);
9540 // Add any parameter attributes.
9541 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9542 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9547 if (FT->getReturnType() == Type::VoidTy)
9548 Caller->setName(""); // Void type should not have a name.
9550 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9553 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9554 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9555 Args.begin(), Args.end(),
9556 Caller->getName(), Caller);
9557 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9558 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9560 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9561 Caller->getName(), Caller);
9562 CallInst *CI = cast<CallInst>(Caller);
9563 if (CI->isTailCall())
9564 cast<CallInst>(NC)->setTailCall();
9565 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9566 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9569 // Insert a cast of the return type as necessary.
9571 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9572 if (NV->getType() != Type::VoidTy) {
9573 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9575 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9577 // If this is an invoke instruction, we should insert it after the first
9578 // non-phi, instruction in the normal successor block.
9579 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9580 BasicBlock::iterator I = II->getNormalDest()->begin();
9581 while (isa<PHINode>(I)) ++I;
9582 InsertNewInstBefore(NC, *I);
9584 // Otherwise, it's a call, just insert cast right after the call instr
9585 InsertNewInstBefore(NC, *Caller);
9587 AddUsersToWorkList(*Caller);
9589 NV = UndefValue::get(Caller->getType());
9593 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9594 Caller->replaceAllUsesWith(NV);
9595 Caller->eraseFromParent();
9596 RemoveFromWorkList(Caller);
9600 // transformCallThroughTrampoline - Turn a call to a function created by the
9601 // init_trampoline intrinsic into a direct call to the underlying function.
9603 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9604 Value *Callee = CS.getCalledValue();
9605 const PointerType *PTy = cast<PointerType>(Callee->getType());
9606 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9607 const PAListPtr &Attrs = CS.getParamAttrs();
9609 // If the call already has the 'nest' attribute somewhere then give up -
9610 // otherwise 'nest' would occur twice after splicing in the chain.
9611 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9614 IntrinsicInst *Tramp =
9615 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9617 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9618 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9619 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9621 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9622 if (!NestAttrs.isEmpty()) {
9623 unsigned NestIdx = 1;
9624 const Type *NestTy = 0;
9625 ParameterAttributes NestAttr = ParamAttr::None;
9627 // Look for a parameter marked with the 'nest' attribute.
9628 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9629 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9630 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9631 // Record the parameter type and any other attributes.
9633 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9638 Instruction *Caller = CS.getInstruction();
9639 std::vector<Value*> NewArgs;
9640 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9642 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9643 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9645 // Insert the nest argument into the call argument list, which may
9646 // mean appending it. Likewise for attributes.
9648 // Add any function result attributes.
9649 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9650 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9654 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9656 if (Idx == NestIdx) {
9657 // Add the chain argument and attributes.
9658 Value *NestVal = Tramp->getOperand(3);
9659 if (NestVal->getType() != NestTy)
9660 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9661 NewArgs.push_back(NestVal);
9662 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9668 // Add the original argument and attributes.
9669 NewArgs.push_back(*I);
9670 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9672 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9678 // The trampoline may have been bitcast to a bogus type (FTy).
9679 // Handle this by synthesizing a new function type, equal to FTy
9680 // with the chain parameter inserted.
9682 std::vector<const Type*> NewTypes;
9683 NewTypes.reserve(FTy->getNumParams()+1);
9685 // Insert the chain's type into the list of parameter types, which may
9686 // mean appending it.
9689 FunctionType::param_iterator I = FTy->param_begin(),
9690 E = FTy->param_end();
9694 // Add the chain's type.
9695 NewTypes.push_back(NestTy);
9700 // Add the original type.
9701 NewTypes.push_back(*I);
9707 // Replace the trampoline call with a direct call. Let the generic
9708 // code sort out any function type mismatches.
9709 FunctionType *NewFTy =
9710 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9711 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9712 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9713 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9715 Instruction *NewCaller;
9716 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9717 NewCaller = InvokeInst::Create(NewCallee,
9718 II->getNormalDest(), II->getUnwindDest(),
9719 NewArgs.begin(), NewArgs.end(),
9720 Caller->getName(), Caller);
9721 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9722 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9724 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9725 Caller->getName(), Caller);
9726 if (cast<CallInst>(Caller)->isTailCall())
9727 cast<CallInst>(NewCaller)->setTailCall();
9728 cast<CallInst>(NewCaller)->
9729 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9730 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9732 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9733 Caller->replaceAllUsesWith(NewCaller);
9734 Caller->eraseFromParent();
9735 RemoveFromWorkList(Caller);
9740 // Replace the trampoline call with a direct call. Since there is no 'nest'
9741 // parameter, there is no need to adjust the argument list. Let the generic
9742 // code sort out any function type mismatches.
9743 Constant *NewCallee =
9744 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9745 CS.setCalledFunction(NewCallee);
9746 return CS.getInstruction();
9749 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9750 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9751 /// and a single binop.
9752 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9753 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9754 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9755 isa<CmpInst>(FirstInst));
9756 unsigned Opc = FirstInst->getOpcode();
9757 Value *LHSVal = FirstInst->getOperand(0);
9758 Value *RHSVal = FirstInst->getOperand(1);
9760 const Type *LHSType = LHSVal->getType();
9761 const Type *RHSType = RHSVal->getType();
9763 // Scan to see if all operands are the same opcode, all have one use, and all
9764 // kill their operands (i.e. the operands have one use).
9765 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9766 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9767 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9768 // Verify type of the LHS matches so we don't fold cmp's of different
9769 // types or GEP's with different index types.
9770 I->getOperand(0)->getType() != LHSType ||
9771 I->getOperand(1)->getType() != RHSType)
9774 // If they are CmpInst instructions, check their predicates
9775 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9776 if (cast<CmpInst>(I)->getPredicate() !=
9777 cast<CmpInst>(FirstInst)->getPredicate())
9780 // Keep track of which operand needs a phi node.
9781 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9782 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9785 // Otherwise, this is safe to transform, determine if it is profitable.
9787 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9788 // Indexes are often folded into load/store instructions, so we don't want to
9789 // hide them behind a phi.
9790 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9793 Value *InLHS = FirstInst->getOperand(0);
9794 Value *InRHS = FirstInst->getOperand(1);
9795 PHINode *NewLHS = 0, *NewRHS = 0;
9797 NewLHS = PHINode::Create(LHSType,
9798 FirstInst->getOperand(0)->getName() + ".pn");
9799 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9800 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9801 InsertNewInstBefore(NewLHS, PN);
9806 NewRHS = PHINode::Create(RHSType,
9807 FirstInst->getOperand(1)->getName() + ".pn");
9808 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9809 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9810 InsertNewInstBefore(NewRHS, PN);
9814 // Add all operands to the new PHIs.
9815 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9817 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9818 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9821 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9822 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9826 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9827 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9828 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9829 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9832 assert(isa<GetElementPtrInst>(FirstInst));
9833 return GetElementPtrInst::Create(LHSVal, RHSVal);
9837 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9838 /// of the block that defines it. This means that it must be obvious the value
9839 /// of the load is not changed from the point of the load to the end of the
9842 /// Finally, it is safe, but not profitable, to sink a load targetting a
9843 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9845 static bool isSafeToSinkLoad(LoadInst *L) {
9846 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9848 for (++BBI; BBI != E; ++BBI)
9849 if (BBI->mayWriteToMemory())
9852 // Check for non-address taken alloca. If not address-taken already, it isn't
9853 // profitable to do this xform.
9854 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9855 bool isAddressTaken = false;
9856 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9858 if (isa<LoadInst>(UI)) continue;
9859 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9860 // If storing TO the alloca, then the address isn't taken.
9861 if (SI->getOperand(1) == AI) continue;
9863 isAddressTaken = true;
9867 if (!isAddressTaken)
9875 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9876 // operator and they all are only used by the PHI, PHI together their
9877 // inputs, and do the operation once, to the result of the PHI.
9878 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9879 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9881 // Scan the instruction, looking for input operations that can be folded away.
9882 // If all input operands to the phi are the same instruction (e.g. a cast from
9883 // the same type or "+42") we can pull the operation through the PHI, reducing
9884 // code size and simplifying code.
9885 Constant *ConstantOp = 0;
9886 const Type *CastSrcTy = 0;
9887 bool isVolatile = false;
9888 if (isa<CastInst>(FirstInst)) {
9889 CastSrcTy = FirstInst->getOperand(0)->getType();
9890 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9891 // Can fold binop, compare or shift here if the RHS is a constant,
9892 // otherwise call FoldPHIArgBinOpIntoPHI.
9893 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9894 if (ConstantOp == 0)
9895 return FoldPHIArgBinOpIntoPHI(PN);
9896 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9897 isVolatile = LI->isVolatile();
9898 // We can't sink the load if the loaded value could be modified between the
9899 // load and the PHI.
9900 if (LI->getParent() != PN.getIncomingBlock(0) ||
9901 !isSafeToSinkLoad(LI))
9903 } else if (isa<GetElementPtrInst>(FirstInst)) {
9904 if (FirstInst->getNumOperands() == 2)
9905 return FoldPHIArgBinOpIntoPHI(PN);
9906 // Can't handle general GEPs yet.
9909 return 0; // Cannot fold this operation.
9912 // Check to see if all arguments are the same operation.
9913 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9914 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9915 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9916 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9919 if (I->getOperand(0)->getType() != CastSrcTy)
9920 return 0; // Cast operation must match.
9921 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9922 // We can't sink the load if the loaded value could be modified between
9923 // the load and the PHI.
9924 if (LI->isVolatile() != isVolatile ||
9925 LI->getParent() != PN.getIncomingBlock(i) ||
9926 !isSafeToSinkLoad(LI))
9929 // If the PHI is volatile and its block has multiple successors, sinking
9930 // it would remove a load of the volatile value from the path through the
9933 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9937 } else if (I->getOperand(1) != ConstantOp) {
9942 // Okay, they are all the same operation. Create a new PHI node of the
9943 // correct type, and PHI together all of the LHS's of the instructions.
9944 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9945 PN.getName()+".in");
9946 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9948 Value *InVal = FirstInst->getOperand(0);
9949 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9951 // Add all operands to the new PHI.
9952 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9953 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9954 if (NewInVal != InVal)
9956 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9961 // The new PHI unions all of the same values together. This is really
9962 // common, so we handle it intelligently here for compile-time speed.
9966 InsertNewInstBefore(NewPN, PN);
9970 // Insert and return the new operation.
9971 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9972 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9973 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9974 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9975 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9976 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9977 PhiVal, ConstantOp);
9978 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9980 // If this was a volatile load that we are merging, make sure to loop through
9981 // and mark all the input loads as non-volatile. If we don't do this, we will
9982 // insert a new volatile load and the old ones will not be deletable.
9984 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9985 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9987 return new LoadInst(PhiVal, "", isVolatile);
9990 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9992 static bool DeadPHICycle(PHINode *PN,
9993 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9994 if (PN->use_empty()) return true;
9995 if (!PN->hasOneUse()) return false;
9997 // Remember this node, and if we find the cycle, return.
9998 if (!PotentiallyDeadPHIs.insert(PN))
10001 // Don't scan crazily complex things.
10002 if (PotentiallyDeadPHIs.size() == 16)
10005 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10006 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10011 /// PHIsEqualValue - Return true if this phi node is always equal to
10012 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10013 /// z = some value; x = phi (y, z); y = phi (x, z)
10014 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10015 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10016 // See if we already saw this PHI node.
10017 if (!ValueEqualPHIs.insert(PN))
10020 // Don't scan crazily complex things.
10021 if (ValueEqualPHIs.size() == 16)
10024 // Scan the operands to see if they are either phi nodes or are equal to
10026 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10027 Value *Op = PN->getIncomingValue(i);
10028 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10029 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10031 } else if (Op != NonPhiInVal)
10039 // PHINode simplification
10041 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10042 // If LCSSA is around, don't mess with Phi nodes
10043 if (MustPreserveLCSSA) return 0;
10045 if (Value *V = PN.hasConstantValue())
10046 return ReplaceInstUsesWith(PN, V);
10048 // If all PHI operands are the same operation, pull them through the PHI,
10049 // reducing code size.
10050 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10051 PN.getIncomingValue(0)->hasOneUse())
10052 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10055 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10056 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10057 // PHI)... break the cycle.
10058 if (PN.hasOneUse()) {
10059 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10060 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10061 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10062 PotentiallyDeadPHIs.insert(&PN);
10063 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10064 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10067 // If this phi has a single use, and if that use just computes a value for
10068 // the next iteration of a loop, delete the phi. This occurs with unused
10069 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10070 // common case here is good because the only other things that catch this
10071 // are induction variable analysis (sometimes) and ADCE, which is only run
10073 if (PHIUser->hasOneUse() &&
10074 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10075 PHIUser->use_back() == &PN) {
10076 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10080 // We sometimes end up with phi cycles that non-obviously end up being the
10081 // same value, for example:
10082 // z = some value; x = phi (y, z); y = phi (x, z)
10083 // where the phi nodes don't necessarily need to be in the same block. Do a
10084 // quick check to see if the PHI node only contains a single non-phi value, if
10085 // so, scan to see if the phi cycle is actually equal to that value.
10087 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10088 // Scan for the first non-phi operand.
10089 while (InValNo != NumOperandVals &&
10090 isa<PHINode>(PN.getIncomingValue(InValNo)))
10093 if (InValNo != NumOperandVals) {
10094 Value *NonPhiInVal = PN.getOperand(InValNo);
10096 // Scan the rest of the operands to see if there are any conflicts, if so
10097 // there is no need to recursively scan other phis.
10098 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10099 Value *OpVal = PN.getIncomingValue(InValNo);
10100 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10104 // If we scanned over all operands, then we have one unique value plus
10105 // phi values. Scan PHI nodes to see if they all merge in each other or
10107 if (InValNo == NumOperandVals) {
10108 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10109 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10110 return ReplaceInstUsesWith(PN, NonPhiInVal);
10117 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10118 Instruction *InsertPoint,
10119 InstCombiner *IC) {
10120 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10121 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10122 // We must cast correctly to the pointer type. Ensure that we
10123 // sign extend the integer value if it is smaller as this is
10124 // used for address computation.
10125 Instruction::CastOps opcode =
10126 (VTySize < PtrSize ? Instruction::SExt :
10127 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10128 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10132 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10133 Value *PtrOp = GEP.getOperand(0);
10134 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10135 // If so, eliminate the noop.
10136 if (GEP.getNumOperands() == 1)
10137 return ReplaceInstUsesWith(GEP, PtrOp);
10139 if (isa<UndefValue>(GEP.getOperand(0)))
10140 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10142 bool HasZeroPointerIndex = false;
10143 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10144 HasZeroPointerIndex = C->isNullValue();
10146 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10147 return ReplaceInstUsesWith(GEP, PtrOp);
10149 // Eliminate unneeded casts for indices.
10150 bool MadeChange = false;
10152 gep_type_iterator GTI = gep_type_begin(GEP);
10153 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
10154 if (isa<SequentialType>(*GTI)) {
10155 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
10156 if (CI->getOpcode() == Instruction::ZExt ||
10157 CI->getOpcode() == Instruction::SExt) {
10158 const Type *SrcTy = CI->getOperand(0)->getType();
10159 // We can eliminate a cast from i32 to i64 iff the target
10160 // is a 32-bit pointer target.
10161 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10163 GEP.setOperand(i, CI->getOperand(0));
10167 // If we are using a wider index than needed for this platform, shrink it
10168 // to what we need. If the incoming value needs a cast instruction,
10169 // insert it. This explicit cast can make subsequent optimizations more
10171 Value *Op = GEP.getOperand(i);
10172 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10173 if (Constant *C = dyn_cast<Constant>(Op)) {
10174 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
10177 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10179 GEP.setOperand(i, Op);
10185 if (MadeChange) return &GEP;
10187 // If this GEP instruction doesn't move the pointer, and if the input operand
10188 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10189 // real input to the dest type.
10190 if (GEP.hasAllZeroIndices()) {
10191 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10192 // If the bitcast is of an allocation, and the allocation will be
10193 // converted to match the type of the cast, don't touch this.
10194 if (isa<AllocationInst>(BCI->getOperand(0))) {
10195 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10196 if (Instruction *I = visitBitCast(*BCI)) {
10199 BCI->getParent()->getInstList().insert(BCI, I);
10200 ReplaceInstUsesWith(*BCI, I);
10205 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10209 // Combine Indices - If the source pointer to this getelementptr instruction
10210 // is a getelementptr instruction, combine the indices of the two
10211 // getelementptr instructions into a single instruction.
10213 SmallVector<Value*, 8> SrcGEPOperands;
10214 if (User *Src = dyn_castGetElementPtr(PtrOp))
10215 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10217 if (!SrcGEPOperands.empty()) {
10218 // Note that if our source is a gep chain itself that we wait for that
10219 // chain to be resolved before we perform this transformation. This
10220 // avoids us creating a TON of code in some cases.
10222 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10223 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10224 return 0; // Wait until our source is folded to completion.
10226 SmallVector<Value*, 8> Indices;
10228 // Find out whether the last index in the source GEP is a sequential idx.
10229 bool EndsWithSequential = false;
10230 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10231 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10232 EndsWithSequential = !isa<StructType>(*I);
10234 // Can we combine the two pointer arithmetics offsets?
10235 if (EndsWithSequential) {
10236 // Replace: gep (gep %P, long B), long A, ...
10237 // With: T = long A+B; gep %P, T, ...
10239 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10240 if (SO1 == Constant::getNullValue(SO1->getType())) {
10242 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10245 // If they aren't the same type, convert both to an integer of the
10246 // target's pointer size.
10247 if (SO1->getType() != GO1->getType()) {
10248 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10249 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10250 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10251 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10253 unsigned PS = TD->getPointerSizeInBits();
10254 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10255 // Convert GO1 to SO1's type.
10256 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10258 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10259 // Convert SO1 to GO1's type.
10260 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10262 const Type *PT = TD->getIntPtrType();
10263 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10264 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10268 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10269 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10271 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10272 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10276 // Recycle the GEP we already have if possible.
10277 if (SrcGEPOperands.size() == 2) {
10278 GEP.setOperand(0, SrcGEPOperands[0]);
10279 GEP.setOperand(1, Sum);
10282 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10283 SrcGEPOperands.end()-1);
10284 Indices.push_back(Sum);
10285 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10287 } else if (isa<Constant>(*GEP.idx_begin()) &&
10288 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10289 SrcGEPOperands.size() != 1) {
10290 // Otherwise we can do the fold if the first index of the GEP is a zero
10291 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10292 SrcGEPOperands.end());
10293 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10296 if (!Indices.empty())
10297 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10298 Indices.end(), GEP.getName());
10300 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10301 // GEP of global variable. If all of the indices for this GEP are
10302 // constants, we can promote this to a constexpr instead of an instruction.
10304 // Scan for nonconstants...
10305 SmallVector<Constant*, 8> Indices;
10306 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10307 for (; I != E && isa<Constant>(*I); ++I)
10308 Indices.push_back(cast<Constant>(*I));
10310 if (I == E) { // If they are all constants...
10311 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10312 &Indices[0],Indices.size());
10314 // Replace all uses of the GEP with the new constexpr...
10315 return ReplaceInstUsesWith(GEP, CE);
10317 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10318 if (!isa<PointerType>(X->getType())) {
10319 // Not interesting. Source pointer must be a cast from pointer.
10320 } else if (HasZeroPointerIndex) {
10321 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10322 // into : GEP [10 x i8]* X, i32 0, ...
10324 // This occurs when the program declares an array extern like "int X[];"
10326 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10327 const PointerType *XTy = cast<PointerType>(X->getType());
10328 if (const ArrayType *XATy =
10329 dyn_cast<ArrayType>(XTy->getElementType()))
10330 if (const ArrayType *CATy =
10331 dyn_cast<ArrayType>(CPTy->getElementType()))
10332 if (CATy->getElementType() == XATy->getElementType()) {
10333 // At this point, we know that the cast source type is a pointer
10334 // to an array of the same type as the destination pointer
10335 // array. Because the array type is never stepped over (there
10336 // is a leading zero) we can fold the cast into this GEP.
10337 GEP.setOperand(0, X);
10340 } else if (GEP.getNumOperands() == 2) {
10341 // Transform things like:
10342 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10343 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10344 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10345 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10346 if (isa<ArrayType>(SrcElTy) &&
10347 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10348 TD->getABITypeSize(ResElTy)) {
10350 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10351 Idx[1] = GEP.getOperand(1);
10352 Value *V = InsertNewInstBefore(
10353 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10354 // V and GEP are both pointer types --> BitCast
10355 return new BitCastInst(V, GEP.getType());
10358 // Transform things like:
10359 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10360 // (where tmp = 8*tmp2) into:
10361 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10363 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10364 uint64_t ArrayEltSize =
10365 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10367 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10368 // allow either a mul, shift, or constant here.
10370 ConstantInt *Scale = 0;
10371 if (ArrayEltSize == 1) {
10372 NewIdx = GEP.getOperand(1);
10373 Scale = ConstantInt::get(NewIdx->getType(), 1);
10374 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10375 NewIdx = ConstantInt::get(CI->getType(), 1);
10377 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10378 if (Inst->getOpcode() == Instruction::Shl &&
10379 isa<ConstantInt>(Inst->getOperand(1))) {
10380 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10381 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10382 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10383 NewIdx = Inst->getOperand(0);
10384 } else if (Inst->getOpcode() == Instruction::Mul &&
10385 isa<ConstantInt>(Inst->getOperand(1))) {
10386 Scale = cast<ConstantInt>(Inst->getOperand(1));
10387 NewIdx = Inst->getOperand(0);
10391 // If the index will be to exactly the right offset with the scale taken
10392 // out, perform the transformation. Note, we don't know whether Scale is
10393 // signed or not. We'll use unsigned version of division/modulo
10394 // operation after making sure Scale doesn't have the sign bit set.
10395 if (Scale && Scale->getSExtValue() >= 0LL &&
10396 Scale->getZExtValue() % ArrayEltSize == 0) {
10397 Scale = ConstantInt::get(Scale->getType(),
10398 Scale->getZExtValue() / ArrayEltSize);
10399 if (Scale->getZExtValue() != 1) {
10400 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10402 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10403 NewIdx = InsertNewInstBefore(Sc, GEP);
10406 // Insert the new GEP instruction.
10408 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10410 Instruction *NewGEP =
10411 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10412 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10413 // The NewGEP must be pointer typed, so must the old one -> BitCast
10414 return new BitCastInst(NewGEP, GEP.getType());
10423 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10424 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10425 if (AI.isArrayAllocation()) { // Check C != 1
10426 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10427 const Type *NewTy =
10428 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10429 AllocationInst *New = 0;
10431 // Create and insert the replacement instruction...
10432 if (isa<MallocInst>(AI))
10433 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10435 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10436 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10439 InsertNewInstBefore(New, AI);
10441 // Scan to the end of the allocation instructions, to skip over a block of
10442 // allocas if possible...
10444 BasicBlock::iterator It = New;
10445 while (isa<AllocationInst>(*It)) ++It;
10447 // Now that I is pointing to the first non-allocation-inst in the block,
10448 // insert our getelementptr instruction...
10450 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10454 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10455 New->getName()+".sub", It);
10457 // Now make everything use the getelementptr instead of the original
10459 return ReplaceInstUsesWith(AI, V);
10460 } else if (isa<UndefValue>(AI.getArraySize())) {
10461 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10465 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10466 // Note that we only do this for alloca's, because malloc should allocate and
10467 // return a unique pointer, even for a zero byte allocation.
10468 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10469 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10470 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10475 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10476 Value *Op = FI.getOperand(0);
10478 // free undef -> unreachable.
10479 if (isa<UndefValue>(Op)) {
10480 // Insert a new store to null because we cannot modify the CFG here.
10481 new StoreInst(ConstantInt::getTrue(),
10482 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10483 return EraseInstFromFunction(FI);
10486 // If we have 'free null' delete the instruction. This can happen in stl code
10487 // when lots of inlining happens.
10488 if (isa<ConstantPointerNull>(Op))
10489 return EraseInstFromFunction(FI);
10491 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10492 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10493 FI.setOperand(0, CI->getOperand(0));
10497 // Change free (gep X, 0,0,0,0) into free(X)
10498 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10499 if (GEPI->hasAllZeroIndices()) {
10500 AddToWorkList(GEPI);
10501 FI.setOperand(0, GEPI->getOperand(0));
10506 // Change free(malloc) into nothing, if the malloc has a single use.
10507 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10508 if (MI->hasOneUse()) {
10509 EraseInstFromFunction(FI);
10510 return EraseInstFromFunction(*MI);
10517 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10518 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10519 const TargetData *TD) {
10520 User *CI = cast<User>(LI.getOperand(0));
10521 Value *CastOp = CI->getOperand(0);
10523 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10524 // Instead of loading constant c string, use corresponding integer value
10525 // directly if string length is small enough.
10526 const std::string &Str = CE->getOperand(0)->getStringValue();
10527 if (!Str.empty()) {
10528 unsigned len = Str.length();
10529 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10530 unsigned numBits = Ty->getPrimitiveSizeInBits();
10531 // Replace LI with immediate integer store.
10532 if ((numBits >> 3) == len + 1) {
10533 APInt StrVal(numBits, 0);
10534 APInt SingleChar(numBits, 0);
10535 if (TD->isLittleEndian()) {
10536 for (signed i = len-1; i >= 0; i--) {
10537 SingleChar = (uint64_t) Str[i];
10538 StrVal = (StrVal << 8) | SingleChar;
10541 for (unsigned i = 0; i < len; i++) {
10542 SingleChar = (uint64_t) Str[i];
10543 StrVal = (StrVal << 8) | SingleChar;
10545 // Append NULL at the end.
10547 StrVal = (StrVal << 8) | SingleChar;
10549 Value *NL = ConstantInt::get(StrVal);
10550 return IC.ReplaceInstUsesWith(LI, NL);
10555 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10556 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10557 const Type *SrcPTy = SrcTy->getElementType();
10559 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10560 isa<VectorType>(DestPTy)) {
10561 // If the source is an array, the code below will not succeed. Check to
10562 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10564 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10565 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10566 if (ASrcTy->getNumElements() != 0) {
10568 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10569 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10570 SrcTy = cast<PointerType>(CastOp->getType());
10571 SrcPTy = SrcTy->getElementType();
10574 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10575 isa<VectorType>(SrcPTy)) &&
10576 // Do not allow turning this into a load of an integer, which is then
10577 // casted to a pointer, this pessimizes pointer analysis a lot.
10578 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10579 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10580 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10582 // Okay, we are casting from one integer or pointer type to another of
10583 // the same size. Instead of casting the pointer before the load, cast
10584 // the result of the loaded value.
10585 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10587 LI.isVolatile()),LI);
10588 // Now cast the result of the load.
10589 return new BitCastInst(NewLoad, LI.getType());
10596 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10597 /// from this value cannot trap. If it is not obviously safe to load from the
10598 /// specified pointer, we do a quick local scan of the basic block containing
10599 /// ScanFrom, to determine if the address is already accessed.
10600 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10601 // If it is an alloca it is always safe to load from.
10602 if (isa<AllocaInst>(V)) return true;
10604 // If it is a global variable it is mostly safe to load from.
10605 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10606 // Don't try to evaluate aliases. External weak GV can be null.
10607 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10609 // Otherwise, be a little bit agressive by scanning the local block where we
10610 // want to check to see if the pointer is already being loaded or stored
10611 // from/to. If so, the previous load or store would have already trapped,
10612 // so there is no harm doing an extra load (also, CSE will later eliminate
10613 // the load entirely).
10614 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10619 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10620 if (LI->getOperand(0) == V) return true;
10621 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10622 if (SI->getOperand(1) == V) return true;
10628 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10629 /// until we find the underlying object a pointer is referring to or something
10630 /// we don't understand. Note that the returned pointer may be offset from the
10631 /// input, because we ignore GEP indices.
10632 static Value *GetUnderlyingObject(Value *Ptr) {
10634 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10635 if (CE->getOpcode() == Instruction::BitCast ||
10636 CE->getOpcode() == Instruction::GetElementPtr)
10637 Ptr = CE->getOperand(0);
10640 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10641 Ptr = BCI->getOperand(0);
10642 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10643 Ptr = GEP->getOperand(0);
10650 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10651 Value *Op = LI.getOperand(0);
10653 // Attempt to improve the alignment.
10654 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10656 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10657 LI.getAlignment()))
10658 LI.setAlignment(KnownAlign);
10660 // load (cast X) --> cast (load X) iff safe
10661 if (isa<CastInst>(Op))
10662 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10665 // None of the following transforms are legal for volatile loads.
10666 if (LI.isVolatile()) return 0;
10668 if (&LI.getParent()->front() != &LI) {
10669 BasicBlock::iterator BBI = &LI; --BBI;
10670 // If the instruction immediately before this is a store to the same
10671 // address, do a simple form of store->load forwarding.
10672 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10673 if (SI->getOperand(1) == LI.getOperand(0))
10674 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10675 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10676 if (LIB->getOperand(0) == LI.getOperand(0))
10677 return ReplaceInstUsesWith(LI, LIB);
10680 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10681 const Value *GEPI0 = GEPI->getOperand(0);
10682 // TODO: Consider a target hook for valid address spaces for this xform.
10683 if (isa<ConstantPointerNull>(GEPI0) &&
10684 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10685 // Insert a new store to null instruction before the load to indicate
10686 // that this code is not reachable. We do this instead of inserting
10687 // an unreachable instruction directly because we cannot modify the
10689 new StoreInst(UndefValue::get(LI.getType()),
10690 Constant::getNullValue(Op->getType()), &LI);
10691 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10695 if (Constant *C = dyn_cast<Constant>(Op)) {
10696 // load null/undef -> undef
10697 // TODO: Consider a target hook for valid address spaces for this xform.
10698 if (isa<UndefValue>(C) || (C->isNullValue() &&
10699 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10700 // Insert a new store to null instruction before the load to indicate that
10701 // this code is not reachable. We do this instead of inserting an
10702 // unreachable instruction directly because we cannot modify the CFG.
10703 new StoreInst(UndefValue::get(LI.getType()),
10704 Constant::getNullValue(Op->getType()), &LI);
10705 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10708 // Instcombine load (constant global) into the value loaded.
10709 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10710 if (GV->isConstant() && !GV->isDeclaration())
10711 return ReplaceInstUsesWith(LI, GV->getInitializer());
10713 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10714 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10715 if (CE->getOpcode() == Instruction::GetElementPtr) {
10716 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10717 if (GV->isConstant() && !GV->isDeclaration())
10719 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10720 return ReplaceInstUsesWith(LI, V);
10721 if (CE->getOperand(0)->isNullValue()) {
10722 // Insert a new store to null instruction before the load to indicate
10723 // that this code is not reachable. We do this instead of inserting
10724 // an unreachable instruction directly because we cannot modify the
10726 new StoreInst(UndefValue::get(LI.getType()),
10727 Constant::getNullValue(Op->getType()), &LI);
10728 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10731 } else if (CE->isCast()) {
10732 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10738 // If this load comes from anywhere in a constant global, and if the global
10739 // is all undef or zero, we know what it loads.
10740 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10741 if (GV->isConstant() && GV->hasInitializer()) {
10742 if (GV->getInitializer()->isNullValue())
10743 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10744 else if (isa<UndefValue>(GV->getInitializer()))
10745 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10749 if (Op->hasOneUse()) {
10750 // Change select and PHI nodes to select values instead of addresses: this
10751 // helps alias analysis out a lot, allows many others simplifications, and
10752 // exposes redundancy in the code.
10754 // Note that we cannot do the transformation unless we know that the
10755 // introduced loads cannot trap! Something like this is valid as long as
10756 // the condition is always false: load (select bool %C, int* null, int* %G),
10757 // but it would not be valid if we transformed it to load from null
10758 // unconditionally.
10760 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10761 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10762 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10763 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10764 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10765 SI->getOperand(1)->getName()+".val"), LI);
10766 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10767 SI->getOperand(2)->getName()+".val"), LI);
10768 return SelectInst::Create(SI->getCondition(), V1, V2);
10771 // load (select (cond, null, P)) -> load P
10772 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10773 if (C->isNullValue()) {
10774 LI.setOperand(0, SI->getOperand(2));
10778 // load (select (cond, P, null)) -> load P
10779 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10780 if (C->isNullValue()) {
10781 LI.setOperand(0, SI->getOperand(1));
10789 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10791 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10792 User *CI = cast<User>(SI.getOperand(1));
10793 Value *CastOp = CI->getOperand(0);
10795 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10796 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10797 const Type *SrcPTy = SrcTy->getElementType();
10799 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10800 // If the source is an array, the code below will not succeed. Check to
10801 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10803 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10804 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10805 if (ASrcTy->getNumElements() != 0) {
10807 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10808 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10809 SrcTy = cast<PointerType>(CastOp->getType());
10810 SrcPTy = SrcTy->getElementType();
10813 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10814 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10815 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10817 // Okay, we are casting from one integer or pointer type to another of
10818 // the same size. Instead of casting the pointer before
10819 // the store, cast the value to be stored.
10821 Value *SIOp0 = SI.getOperand(0);
10822 Instruction::CastOps opcode = Instruction::BitCast;
10823 const Type* CastSrcTy = SIOp0->getType();
10824 const Type* CastDstTy = SrcPTy;
10825 if (isa<PointerType>(CastDstTy)) {
10826 if (CastSrcTy->isInteger())
10827 opcode = Instruction::IntToPtr;
10828 } else if (isa<IntegerType>(CastDstTy)) {
10829 if (isa<PointerType>(SIOp0->getType()))
10830 opcode = Instruction::PtrToInt;
10832 if (Constant *C = dyn_cast<Constant>(SIOp0))
10833 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10835 NewCast = IC.InsertNewInstBefore(
10836 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10838 return new StoreInst(NewCast, CastOp);
10845 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10846 Value *Val = SI.getOperand(0);
10847 Value *Ptr = SI.getOperand(1);
10849 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10850 EraseInstFromFunction(SI);
10855 // If the RHS is an alloca with a single use, zapify the store, making the
10857 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10858 if (isa<AllocaInst>(Ptr)) {
10859 EraseInstFromFunction(SI);
10864 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10865 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10866 GEP->getOperand(0)->hasOneUse()) {
10867 EraseInstFromFunction(SI);
10873 // Attempt to improve the alignment.
10874 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10876 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10877 SI.getAlignment()))
10878 SI.setAlignment(KnownAlign);
10880 // Do really simple DSE, to catch cases where there are several consequtive
10881 // stores to the same location, separated by a few arithmetic operations. This
10882 // situation often occurs with bitfield accesses.
10883 BasicBlock::iterator BBI = &SI;
10884 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10888 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10889 // Prev store isn't volatile, and stores to the same location?
10890 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10893 EraseInstFromFunction(*PrevSI);
10899 // If this is a load, we have to stop. However, if the loaded value is from
10900 // the pointer we're loading and is producing the pointer we're storing,
10901 // then *this* store is dead (X = load P; store X -> P).
10902 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10903 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10904 EraseInstFromFunction(SI);
10908 // Otherwise, this is a load from some other location. Stores before it
10909 // may not be dead.
10913 // Don't skip over loads or things that can modify memory.
10914 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10919 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10921 // store X, null -> turns into 'unreachable' in SimplifyCFG
10922 if (isa<ConstantPointerNull>(Ptr)) {
10923 if (!isa<UndefValue>(Val)) {
10924 SI.setOperand(0, UndefValue::get(Val->getType()));
10925 if (Instruction *U = dyn_cast<Instruction>(Val))
10926 AddToWorkList(U); // Dropped a use.
10929 return 0; // Do not modify these!
10932 // store undef, Ptr -> noop
10933 if (isa<UndefValue>(Val)) {
10934 EraseInstFromFunction(SI);
10939 // If the pointer destination is a cast, see if we can fold the cast into the
10941 if (isa<CastInst>(Ptr))
10942 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10944 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10946 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10950 // If this store is the last instruction in the basic block, and if the block
10951 // ends with an unconditional branch, try to move it to the successor block.
10953 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10954 if (BI->isUnconditional())
10955 if (SimplifyStoreAtEndOfBlock(SI))
10956 return 0; // xform done!
10961 /// SimplifyStoreAtEndOfBlock - Turn things like:
10962 /// if () { *P = v1; } else { *P = v2 }
10963 /// into a phi node with a store in the successor.
10965 /// Simplify things like:
10966 /// *P = v1; if () { *P = v2; }
10967 /// into a phi node with a store in the successor.
10969 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10970 BasicBlock *StoreBB = SI.getParent();
10972 // Check to see if the successor block has exactly two incoming edges. If
10973 // so, see if the other predecessor contains a store to the same location.
10974 // if so, insert a PHI node (if needed) and move the stores down.
10975 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10977 // Determine whether Dest has exactly two predecessors and, if so, compute
10978 // the other predecessor.
10979 pred_iterator PI = pred_begin(DestBB);
10980 BasicBlock *OtherBB = 0;
10981 if (*PI != StoreBB)
10984 if (PI == pred_end(DestBB))
10987 if (*PI != StoreBB) {
10992 if (++PI != pred_end(DestBB))
10996 // Verify that the other block ends in a branch and is not otherwise empty.
10997 BasicBlock::iterator BBI = OtherBB->getTerminator();
10998 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10999 if (!OtherBr || BBI == OtherBB->begin())
11002 // If the other block ends in an unconditional branch, check for the 'if then
11003 // else' case. there is an instruction before the branch.
11004 StoreInst *OtherStore = 0;
11005 if (OtherBr->isUnconditional()) {
11006 // If this isn't a store, or isn't a store to the same location, bail out.
11008 OtherStore = dyn_cast<StoreInst>(BBI);
11009 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11012 // Otherwise, the other block ended with a conditional branch. If one of the
11013 // destinations is StoreBB, then we have the if/then case.
11014 if (OtherBr->getSuccessor(0) != StoreBB &&
11015 OtherBr->getSuccessor(1) != StoreBB)
11018 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11019 // if/then triangle. See if there is a store to the same ptr as SI that
11020 // lives in OtherBB.
11022 // Check to see if we find the matching store.
11023 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11024 if (OtherStore->getOperand(1) != SI.getOperand(1))
11028 // If we find something that may be using the stored value, or if we run
11029 // out of instructions, we can't do the xform.
11030 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
11031 BBI == OtherBB->begin())
11035 // In order to eliminate the store in OtherBr, we have to
11036 // make sure nothing reads the stored value in StoreBB.
11037 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11038 // FIXME: This should really be AA driven.
11039 if (isa<LoadInst>(I) || I->mayWriteToMemory())
11044 // Insert a PHI node now if we need it.
11045 Value *MergedVal = OtherStore->getOperand(0);
11046 if (MergedVal != SI.getOperand(0)) {
11047 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11048 PN->reserveOperandSpace(2);
11049 PN->addIncoming(SI.getOperand(0), SI.getParent());
11050 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11051 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11054 // Advance to a place where it is safe to insert the new store and
11056 BBI = DestBB->begin();
11057 while (isa<PHINode>(BBI)) ++BBI;
11058 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11059 OtherStore->isVolatile()), *BBI);
11061 // Nuke the old stores.
11062 EraseInstFromFunction(SI);
11063 EraseInstFromFunction(*OtherStore);
11069 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11070 // Change br (not X), label True, label False to: br X, label False, True
11072 BasicBlock *TrueDest;
11073 BasicBlock *FalseDest;
11074 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11075 !isa<Constant>(X)) {
11076 // Swap Destinations and condition...
11077 BI.setCondition(X);
11078 BI.setSuccessor(0, FalseDest);
11079 BI.setSuccessor(1, TrueDest);
11083 // Cannonicalize fcmp_one -> fcmp_oeq
11084 FCmpInst::Predicate FPred; Value *Y;
11085 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11086 TrueDest, FalseDest)))
11087 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11088 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11089 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11090 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11091 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11092 NewSCC->takeName(I);
11093 // Swap Destinations and condition...
11094 BI.setCondition(NewSCC);
11095 BI.setSuccessor(0, FalseDest);
11096 BI.setSuccessor(1, TrueDest);
11097 RemoveFromWorkList(I);
11098 I->eraseFromParent();
11099 AddToWorkList(NewSCC);
11103 // Cannonicalize icmp_ne -> icmp_eq
11104 ICmpInst::Predicate IPred;
11105 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11106 TrueDest, FalseDest)))
11107 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11108 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11109 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11110 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11111 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11112 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11113 NewSCC->takeName(I);
11114 // Swap Destinations and condition...
11115 BI.setCondition(NewSCC);
11116 BI.setSuccessor(0, FalseDest);
11117 BI.setSuccessor(1, TrueDest);
11118 RemoveFromWorkList(I);
11119 I->eraseFromParent();;
11120 AddToWorkList(NewSCC);
11127 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11128 Value *Cond = SI.getCondition();
11129 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11130 if (I->getOpcode() == Instruction::Add)
11131 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11132 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11133 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11134 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11136 SI.setOperand(0, I->getOperand(0));
11144 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11145 /// is to leave as a vector operation.
11146 static bool CheapToScalarize(Value *V, bool isConstant) {
11147 if (isa<ConstantAggregateZero>(V))
11149 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11150 if (isConstant) return true;
11151 // If all elts are the same, we can extract.
11152 Constant *Op0 = C->getOperand(0);
11153 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11154 if (C->getOperand(i) != Op0)
11158 Instruction *I = dyn_cast<Instruction>(V);
11159 if (!I) return false;
11161 // Insert element gets simplified to the inserted element or is deleted if
11162 // this is constant idx extract element and its a constant idx insertelt.
11163 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11164 isa<ConstantInt>(I->getOperand(2)))
11166 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11168 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11169 if (BO->hasOneUse() &&
11170 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11171 CheapToScalarize(BO->getOperand(1), isConstant)))
11173 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11174 if (CI->hasOneUse() &&
11175 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11176 CheapToScalarize(CI->getOperand(1), isConstant)))
11182 /// Read and decode a shufflevector mask.
11184 /// It turns undef elements into values that are larger than the number of
11185 /// elements in the input.
11186 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11187 unsigned NElts = SVI->getType()->getNumElements();
11188 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11189 return std::vector<unsigned>(NElts, 0);
11190 if (isa<UndefValue>(SVI->getOperand(2)))
11191 return std::vector<unsigned>(NElts, 2*NElts);
11193 std::vector<unsigned> Result;
11194 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11195 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
11196 if (isa<UndefValue>(CP->getOperand(i)))
11197 Result.push_back(NElts*2); // undef -> 8
11199 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
11203 /// FindScalarElement - Given a vector and an element number, see if the scalar
11204 /// value is already around as a register, for example if it were inserted then
11205 /// extracted from the vector.
11206 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11207 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11208 const VectorType *PTy = cast<VectorType>(V->getType());
11209 unsigned Width = PTy->getNumElements();
11210 if (EltNo >= Width) // Out of range access.
11211 return UndefValue::get(PTy->getElementType());
11213 if (isa<UndefValue>(V))
11214 return UndefValue::get(PTy->getElementType());
11215 else if (isa<ConstantAggregateZero>(V))
11216 return Constant::getNullValue(PTy->getElementType());
11217 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11218 return CP->getOperand(EltNo);
11219 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11220 // If this is an insert to a variable element, we don't know what it is.
11221 if (!isa<ConstantInt>(III->getOperand(2)))
11223 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11225 // If this is an insert to the element we are looking for, return the
11227 if (EltNo == IIElt)
11228 return III->getOperand(1);
11230 // Otherwise, the insertelement doesn't modify the value, recurse on its
11232 return FindScalarElement(III->getOperand(0), EltNo);
11233 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11234 unsigned InEl = getShuffleMask(SVI)[EltNo];
11236 return FindScalarElement(SVI->getOperand(0), InEl);
11237 else if (InEl < Width*2)
11238 return FindScalarElement(SVI->getOperand(1), InEl - Width);
11240 return UndefValue::get(PTy->getElementType());
11243 // Otherwise, we don't know.
11247 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11249 // If vector val is undef, replace extract with scalar undef.
11250 if (isa<UndefValue>(EI.getOperand(0)))
11251 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11253 // If vector val is constant 0, replace extract with scalar 0.
11254 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11255 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11257 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11258 // If vector val is constant with uniform operands, replace EI
11259 // with that operand
11260 Constant *op0 = C->getOperand(0);
11261 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11262 if (C->getOperand(i) != op0) {
11267 return ReplaceInstUsesWith(EI, op0);
11270 // If extracting a specified index from the vector, see if we can recursively
11271 // find a previously computed scalar that was inserted into the vector.
11272 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11273 unsigned IndexVal = IdxC->getZExtValue();
11274 unsigned VectorWidth =
11275 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11277 // If this is extracting an invalid index, turn this into undef, to avoid
11278 // crashing the code below.
11279 if (IndexVal >= VectorWidth)
11280 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11282 // This instruction only demands the single element from the input vector.
11283 // If the input vector has a single use, simplify it based on this use
11285 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11286 uint64_t UndefElts;
11287 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11290 EI.setOperand(0, V);
11295 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11296 return ReplaceInstUsesWith(EI, Elt);
11298 // If the this extractelement is directly using a bitcast from a vector of
11299 // the same number of elements, see if we can find the source element from
11300 // it. In this case, we will end up needing to bitcast the scalars.
11301 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11302 if (const VectorType *VT =
11303 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11304 if (VT->getNumElements() == VectorWidth)
11305 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11306 return new BitCastInst(Elt, EI.getType());
11310 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11311 if (I->hasOneUse()) {
11312 // Push extractelement into predecessor operation if legal and
11313 // profitable to do so
11314 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11315 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11316 if (CheapToScalarize(BO, isConstantElt)) {
11317 ExtractElementInst *newEI0 =
11318 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11319 EI.getName()+".lhs");
11320 ExtractElementInst *newEI1 =
11321 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11322 EI.getName()+".rhs");
11323 InsertNewInstBefore(newEI0, EI);
11324 InsertNewInstBefore(newEI1, EI);
11325 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11327 } else if (isa<LoadInst>(I)) {
11329 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11330 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11331 PointerType::get(EI.getType(), AS),EI);
11332 GetElementPtrInst *GEP =
11333 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11334 InsertNewInstBefore(GEP, EI);
11335 return new LoadInst(GEP);
11338 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11339 // Extracting the inserted element?
11340 if (IE->getOperand(2) == EI.getOperand(1))
11341 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11342 // If the inserted and extracted elements are constants, they must not
11343 // be the same value, extract from the pre-inserted value instead.
11344 if (isa<Constant>(IE->getOperand(2)) &&
11345 isa<Constant>(EI.getOperand(1))) {
11346 AddUsesToWorkList(EI);
11347 EI.setOperand(0, IE->getOperand(0));
11350 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11351 // If this is extracting an element from a shufflevector, figure out where
11352 // it came from and extract from the appropriate input element instead.
11353 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11354 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11356 if (SrcIdx < SVI->getType()->getNumElements())
11357 Src = SVI->getOperand(0);
11358 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
11359 SrcIdx -= SVI->getType()->getNumElements();
11360 Src = SVI->getOperand(1);
11362 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11364 return new ExtractElementInst(Src, SrcIdx);
11371 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11372 /// elements from either LHS or RHS, return the shuffle mask and true.
11373 /// Otherwise, return false.
11374 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11375 std::vector<Constant*> &Mask) {
11376 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11377 "Invalid CollectSingleShuffleElements");
11378 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11380 if (isa<UndefValue>(V)) {
11381 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11383 } else if (V == LHS) {
11384 for (unsigned i = 0; i != NumElts; ++i)
11385 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11387 } else if (V == RHS) {
11388 for (unsigned i = 0; i != NumElts; ++i)
11389 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11391 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11392 // If this is an insert of an extract from some other vector, include it.
11393 Value *VecOp = IEI->getOperand(0);
11394 Value *ScalarOp = IEI->getOperand(1);
11395 Value *IdxOp = IEI->getOperand(2);
11397 if (!isa<ConstantInt>(IdxOp))
11399 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11401 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11402 // Okay, we can handle this if the vector we are insertinting into is
11403 // transitively ok.
11404 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11405 // If so, update the mask to reflect the inserted undef.
11406 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11409 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11410 if (isa<ConstantInt>(EI->getOperand(1)) &&
11411 EI->getOperand(0)->getType() == V->getType()) {
11412 unsigned ExtractedIdx =
11413 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11415 // This must be extracting from either LHS or RHS.
11416 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11417 // Okay, we can handle this if the vector we are insertinting into is
11418 // transitively ok.
11419 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11420 // If so, update the mask to reflect the inserted value.
11421 if (EI->getOperand(0) == LHS) {
11422 Mask[InsertedIdx & (NumElts-1)] =
11423 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11425 assert(EI->getOperand(0) == RHS);
11426 Mask[InsertedIdx & (NumElts-1)] =
11427 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11436 // TODO: Handle shufflevector here!
11441 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11442 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11443 /// that computes V and the LHS value of the shuffle.
11444 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11446 assert(isa<VectorType>(V->getType()) &&
11447 (RHS == 0 || V->getType() == RHS->getType()) &&
11448 "Invalid shuffle!");
11449 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11451 if (isa<UndefValue>(V)) {
11452 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11454 } else if (isa<ConstantAggregateZero>(V)) {
11455 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11457 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11458 // If this is an insert of an extract from some other vector, include it.
11459 Value *VecOp = IEI->getOperand(0);
11460 Value *ScalarOp = IEI->getOperand(1);
11461 Value *IdxOp = IEI->getOperand(2);
11463 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11464 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11465 EI->getOperand(0)->getType() == V->getType()) {
11466 unsigned ExtractedIdx =
11467 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11468 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11470 // Either the extracted from or inserted into vector must be RHSVec,
11471 // otherwise we'd end up with a shuffle of three inputs.
11472 if (EI->getOperand(0) == RHS || RHS == 0) {
11473 RHS = EI->getOperand(0);
11474 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11475 Mask[InsertedIdx & (NumElts-1)] =
11476 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11480 if (VecOp == RHS) {
11481 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11482 // Everything but the extracted element is replaced with the RHS.
11483 for (unsigned i = 0; i != NumElts; ++i) {
11484 if (i != InsertedIdx)
11485 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11490 // If this insertelement is a chain that comes from exactly these two
11491 // vectors, return the vector and the effective shuffle.
11492 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11493 return EI->getOperand(0);
11498 // TODO: Handle shufflevector here!
11500 // Otherwise, can't do anything fancy. Return an identity vector.
11501 for (unsigned i = 0; i != NumElts; ++i)
11502 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11506 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11507 Value *VecOp = IE.getOperand(0);
11508 Value *ScalarOp = IE.getOperand(1);
11509 Value *IdxOp = IE.getOperand(2);
11511 // Inserting an undef or into an undefined place, remove this.
11512 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11513 ReplaceInstUsesWith(IE, VecOp);
11515 // If the inserted element was extracted from some other vector, and if the
11516 // indexes are constant, try to turn this into a shufflevector operation.
11517 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11518 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11519 EI->getOperand(0)->getType() == IE.getType()) {
11520 unsigned NumVectorElts = IE.getType()->getNumElements();
11521 unsigned ExtractedIdx =
11522 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11523 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11525 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11526 return ReplaceInstUsesWith(IE, VecOp);
11528 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11529 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11531 // If we are extracting a value from a vector, then inserting it right
11532 // back into the same place, just use the input vector.
11533 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11534 return ReplaceInstUsesWith(IE, VecOp);
11536 // We could theoretically do this for ANY input. However, doing so could
11537 // turn chains of insertelement instructions into a chain of shufflevector
11538 // instructions, and right now we do not merge shufflevectors. As such,
11539 // only do this in a situation where it is clear that there is benefit.
11540 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11541 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11542 // the values of VecOp, except then one read from EIOp0.
11543 // Build a new shuffle mask.
11544 std::vector<Constant*> Mask;
11545 if (isa<UndefValue>(VecOp))
11546 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11548 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11549 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11552 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11553 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11554 ConstantVector::get(Mask));
11557 // If this insertelement isn't used by some other insertelement, turn it
11558 // (and any insertelements it points to), into one big shuffle.
11559 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11560 std::vector<Constant*> Mask;
11562 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11563 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11564 // We now have a shuffle of LHS, RHS, Mask.
11565 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11574 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11575 Value *LHS = SVI.getOperand(0);
11576 Value *RHS = SVI.getOperand(1);
11577 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11579 bool MadeChange = false;
11581 // Undefined shuffle mask -> undefined value.
11582 if (isa<UndefValue>(SVI.getOperand(2)))
11583 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11585 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11586 // the undef, change them to undefs.
11587 if (isa<UndefValue>(SVI.getOperand(1))) {
11588 // Scan to see if there are any references to the RHS. If so, replace them
11589 // with undef element refs and set MadeChange to true.
11590 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11591 if (Mask[i] >= e && Mask[i] != 2*e) {
11598 // Remap any references to RHS to use LHS.
11599 std::vector<Constant*> Elts;
11600 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11601 if (Mask[i] == 2*e)
11602 Elts.push_back(UndefValue::get(Type::Int32Ty));
11604 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11606 SVI.setOperand(2, ConstantVector::get(Elts));
11610 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11611 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11612 if (LHS == RHS || isa<UndefValue>(LHS)) {
11613 if (isa<UndefValue>(LHS) && LHS == RHS) {
11614 // shuffle(undef,undef,mask) -> undef.
11615 return ReplaceInstUsesWith(SVI, LHS);
11618 // Remap any references to RHS to use LHS.
11619 std::vector<Constant*> Elts;
11620 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11621 if (Mask[i] >= 2*e)
11622 Elts.push_back(UndefValue::get(Type::Int32Ty));
11624 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11625 (Mask[i] < e && isa<UndefValue>(LHS)))
11626 Mask[i] = 2*e; // Turn into undef.
11628 Mask[i] &= (e-1); // Force to LHS.
11629 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11632 SVI.setOperand(0, SVI.getOperand(1));
11633 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11634 SVI.setOperand(2, ConstantVector::get(Elts));
11635 LHS = SVI.getOperand(0);
11636 RHS = SVI.getOperand(1);
11640 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11641 bool isLHSID = true, isRHSID = true;
11643 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11644 if (Mask[i] >= e*2) continue; // Ignore undef values.
11645 // Is this an identity shuffle of the LHS value?
11646 isLHSID &= (Mask[i] == i);
11648 // Is this an identity shuffle of the RHS value?
11649 isRHSID &= (Mask[i]-e == i);
11652 // Eliminate identity shuffles.
11653 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11654 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11656 // If the LHS is a shufflevector itself, see if we can combine it with this
11657 // one without producing an unusual shuffle. Here we are really conservative:
11658 // we are absolutely afraid of producing a shuffle mask not in the input
11659 // program, because the code gen may not be smart enough to turn a merged
11660 // shuffle into two specific shuffles: it may produce worse code. As such,
11661 // we only merge two shuffles if the result is one of the two input shuffle
11662 // masks. In this case, merging the shuffles just removes one instruction,
11663 // which we know is safe. This is good for things like turning:
11664 // (splat(splat)) -> splat.
11665 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11666 if (isa<UndefValue>(RHS)) {
11667 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11669 std::vector<unsigned> NewMask;
11670 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11671 if (Mask[i] >= 2*e)
11672 NewMask.push_back(2*e);
11674 NewMask.push_back(LHSMask[Mask[i]]);
11676 // If the result mask is equal to the src shuffle or this shuffle mask, do
11677 // the replacement.
11678 if (NewMask == LHSMask || NewMask == Mask) {
11679 std::vector<Constant*> Elts;
11680 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11681 if (NewMask[i] >= e*2) {
11682 Elts.push_back(UndefValue::get(Type::Int32Ty));
11684 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11687 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11688 LHSSVI->getOperand(1),
11689 ConstantVector::get(Elts));
11694 return MadeChange ? &SVI : 0;
11700 /// TryToSinkInstruction - Try to move the specified instruction from its
11701 /// current block into the beginning of DestBlock, which can only happen if it's
11702 /// safe to move the instruction past all of the instructions between it and the
11703 /// end of its block.
11704 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11705 assert(I->hasOneUse() && "Invariants didn't hold!");
11707 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11708 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11711 // Do not sink alloca instructions out of the entry block.
11712 if (isa<AllocaInst>(I) && I->getParent() ==
11713 &DestBlock->getParent()->getEntryBlock())
11716 // We can only sink load instructions if there is nothing between the load and
11717 // the end of block that could change the value.
11718 if (I->mayReadFromMemory()) {
11719 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11721 if (Scan->mayWriteToMemory())
11725 BasicBlock::iterator InsertPos = DestBlock->begin();
11726 while (isa<PHINode>(InsertPos)) ++InsertPos;
11728 I->moveBefore(InsertPos);
11734 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11735 /// all reachable code to the worklist.
11737 /// This has a couple of tricks to make the code faster and more powerful. In
11738 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11739 /// them to the worklist (this significantly speeds up instcombine on code where
11740 /// many instructions are dead or constant). Additionally, if we find a branch
11741 /// whose condition is a known constant, we only visit the reachable successors.
11743 static void AddReachableCodeToWorklist(BasicBlock *BB,
11744 SmallPtrSet<BasicBlock*, 64> &Visited,
11746 const TargetData *TD) {
11747 std::vector<BasicBlock*> Worklist;
11748 Worklist.push_back(BB);
11750 while (!Worklist.empty()) {
11751 BB = Worklist.back();
11752 Worklist.pop_back();
11754 // We have now visited this block! If we've already been here, ignore it.
11755 if (!Visited.insert(BB)) continue;
11757 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11758 Instruction *Inst = BBI++;
11760 // DCE instruction if trivially dead.
11761 if (isInstructionTriviallyDead(Inst)) {
11763 DOUT << "IC: DCE: " << *Inst;
11764 Inst->eraseFromParent();
11768 // ConstantProp instruction if trivially constant.
11769 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11770 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11771 Inst->replaceAllUsesWith(C);
11773 Inst->eraseFromParent();
11777 IC.AddToWorkList(Inst);
11780 // Recursively visit successors. If this is a branch or switch on a
11781 // constant, only visit the reachable successor.
11782 TerminatorInst *TI = BB->getTerminator();
11783 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11784 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11785 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11786 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11787 Worklist.push_back(ReachableBB);
11790 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11791 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11792 // See if this is an explicit destination.
11793 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11794 if (SI->getCaseValue(i) == Cond) {
11795 BasicBlock *ReachableBB = SI->getSuccessor(i);
11796 Worklist.push_back(ReachableBB);
11800 // Otherwise it is the default destination.
11801 Worklist.push_back(SI->getSuccessor(0));
11806 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11807 Worklist.push_back(TI->getSuccessor(i));
11811 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11812 bool Changed = false;
11813 TD = &getAnalysis<TargetData>();
11815 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11816 << F.getNameStr() << "\n");
11819 // Do a depth-first traversal of the function, populate the worklist with
11820 // the reachable instructions. Ignore blocks that are not reachable. Keep
11821 // track of which blocks we visit.
11822 SmallPtrSet<BasicBlock*, 64> Visited;
11823 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11825 // Do a quick scan over the function. If we find any blocks that are
11826 // unreachable, remove any instructions inside of them. This prevents
11827 // the instcombine code from having to deal with some bad special cases.
11828 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11829 if (!Visited.count(BB)) {
11830 Instruction *Term = BB->getTerminator();
11831 while (Term != BB->begin()) { // Remove instrs bottom-up
11832 BasicBlock::iterator I = Term; --I;
11834 DOUT << "IC: DCE: " << *I;
11837 if (!I->use_empty())
11838 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11839 I->eraseFromParent();
11844 while (!Worklist.empty()) {
11845 Instruction *I = RemoveOneFromWorkList();
11846 if (I == 0) continue; // skip null values.
11848 // Check to see if we can DCE the instruction.
11849 if (isInstructionTriviallyDead(I)) {
11850 // Add operands to the worklist.
11851 if (I->getNumOperands() < 4)
11852 AddUsesToWorkList(*I);
11855 DOUT << "IC: DCE: " << *I;
11857 I->eraseFromParent();
11858 RemoveFromWorkList(I);
11862 // Instruction isn't dead, see if we can constant propagate it.
11863 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11864 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11866 // Add operands to the worklist.
11867 AddUsesToWorkList(*I);
11868 ReplaceInstUsesWith(*I, C);
11871 I->eraseFromParent();
11872 RemoveFromWorkList(I);
11876 // See if we can trivially sink this instruction to a successor basic block.
11877 // FIXME: Remove GetResultInst test when first class support for aggregates
11879 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11880 BasicBlock *BB = I->getParent();
11881 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11882 if (UserParent != BB) {
11883 bool UserIsSuccessor = false;
11884 // See if the user is one of our successors.
11885 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11886 if (*SI == UserParent) {
11887 UserIsSuccessor = true;
11891 // If the user is one of our immediate successors, and if that successor
11892 // only has us as a predecessors (we'd have to split the critical edge
11893 // otherwise), we can keep going.
11894 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11895 next(pred_begin(UserParent)) == pred_end(UserParent))
11896 // Okay, the CFG is simple enough, try to sink this instruction.
11897 Changed |= TryToSinkInstruction(I, UserParent);
11901 // Now that we have an instruction, try combining it to simplify it...
11905 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11906 if (Instruction *Result = visit(*I)) {
11908 // Should we replace the old instruction with a new one?
11910 DOUT << "IC: Old = " << *I
11911 << " New = " << *Result;
11913 // Everything uses the new instruction now.
11914 I->replaceAllUsesWith(Result);
11916 // Push the new instruction and any users onto the worklist.
11917 AddToWorkList(Result);
11918 AddUsersToWorkList(*Result);
11920 // Move the name to the new instruction first.
11921 Result->takeName(I);
11923 // Insert the new instruction into the basic block...
11924 BasicBlock *InstParent = I->getParent();
11925 BasicBlock::iterator InsertPos = I;
11927 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11928 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11931 InstParent->getInstList().insert(InsertPos, Result);
11933 // Make sure that we reprocess all operands now that we reduced their
11935 AddUsesToWorkList(*I);
11937 // Instructions can end up on the worklist more than once. Make sure
11938 // we do not process an instruction that has been deleted.
11939 RemoveFromWorkList(I);
11941 // Erase the old instruction.
11942 InstParent->getInstList().erase(I);
11945 DOUT << "IC: Mod = " << OrigI
11946 << " New = " << *I;
11949 // If the instruction was modified, it's possible that it is now dead.
11950 // if so, remove it.
11951 if (isInstructionTriviallyDead(I)) {
11952 // Make sure we process all operands now that we are reducing their
11954 AddUsesToWorkList(*I);
11956 // Instructions may end up in the worklist more than once. Erase all
11957 // occurrences of this instruction.
11958 RemoveFromWorkList(I);
11959 I->eraseFromParent();
11962 AddUsersToWorkList(*I);
11969 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11971 // Do an explicit clear, this shrinks the map if needed.
11972 WorklistMap.clear();
11977 bool InstCombiner::runOnFunction(Function &F) {
11978 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11980 bool EverMadeChange = false;
11982 // Iterate while there is work to do.
11983 unsigned Iteration = 0;
11984 while (DoOneIteration(F, Iteration++))
11985 EverMadeChange = true;
11986 return EverMadeChange;
11989 FunctionPass *llvm::createInstructionCombiningPass() {
11990 return new InstCombiner();