diff options
Diffstat (limited to 'lib/Transforms')
-rw-r--r-- | lib/Transforms/IPO/GlobalOpt.cpp | 6 | ||||
-rw-r--r-- | lib/Transforms/InstCombine/InstCombineMulDivRem.cpp | 2 | ||||
-rw-r--r-- | lib/Transforms/ObjCARC/ObjCARCOpts.cpp | 131 | ||||
-rw-r--r-- | lib/Transforms/Vectorize/LoopVectorize.cpp | 162 | ||||
-rw-r--r-- | lib/Transforms/Vectorize/VecUtils.cpp | 44 |
5 files changed, 291 insertions, 54 deletions
diff --git a/lib/Transforms/IPO/GlobalOpt.cpp b/lib/Transforms/IPO/GlobalOpt.cpp index 0ef900e2b9..4a9cb27b03 100644 --- a/lib/Transforms/IPO/GlobalOpt.cpp +++ b/lib/Transforms/IPO/GlobalOpt.cpp @@ -3323,8 +3323,6 @@ bool GlobalOpt::runOnModule(Module &M) { // Try to find the llvm.globalctors list. GlobalVariable *GlobalCtors = FindGlobalCtors(M); - Function *CXAAtExitFn = FindCXAAtExit(M, TLI); - bool LocalChange = true; while (LocalChange) { LocalChange = false; @@ -3342,7 +3340,9 @@ bool GlobalOpt::runOnModule(Module &M) { // Resolve aliases, when possible. LocalChange |= OptimizeGlobalAliases(M); - // Try to remove trivial global destructors. + // Try to remove trivial global destructors if they are not removed + // already. + Function *CXAAtExitFn = FindCXAAtExit(M, TLI); if (CXAAtExitFn) LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn); diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index 87d56214a3..51ca29bc07 100644 --- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -846,7 +846,7 @@ Instruction *InstCombiner::visitSDiv(BinaryOperator &I) { /// FP value and: /// 1) 1/C is exact, or /// 2) reciprocal is allowed. -/// If the convertion was successful, the simplified expression "X * 1/C" is +/// If the conversion was successful, the simplified expression "X * 1/C" is /// returned; otherwise, NULL is returned. /// static Instruction *CvtFDivConstToReciprocal(Value *Dividend, diff --git a/lib/Transforms/ObjCARC/ObjCARCOpts.cpp b/lib/Transforms/ObjCARC/ObjCARCOpts.cpp index 43e2e20035..4bf25facc6 100644 --- a/lib/Transforms/ObjCARC/ObjCARCOpts.cpp +++ b/lib/Transforms/ObjCARC/ObjCARCOpts.cpp @@ -107,6 +107,12 @@ namespace { return std::make_pair(Vector.begin() + Pair.first->second, false); } + iterator find(const KeyT &Key) { + typename MapTy::iterator It = Map.find(Key); + if (It == Map.end()) return Vector.end(); + return Vector.begin() + It->second; + } + const_iterator find(const KeyT &Key) const { typename MapTy::const_iterator It = Map.find(Key); if (It == Map.end()) return Vector.end(); @@ -253,6 +259,40 @@ static bool DoesRetainableObjPtrEscape(const User *Ptr) { return false; } +/// This is a wrapper around getUnderlyingObjCPtr along the lines of +/// GetUnderlyingObjects except that it returns early when it sees the first +/// alloca. +static inline bool AreAnyUnderlyingObjectsAnAlloca(const Value *V) { + SmallPtrSet<const Value *, 4> Visited; + SmallVector<const Value *, 4> Worklist; + Worklist.push_back(V); + do { + const Value *P = Worklist.pop_back_val(); + P = GetUnderlyingObjCPtr(P); + + if (isa<AllocaInst>(P)) + return true; + + if (!Visited.insert(P)) + continue; + + if (const SelectInst *SI = dyn_cast<const SelectInst>(P)) { + Worklist.push_back(SI->getTrueValue()); + Worklist.push_back(SI->getFalseValue()); + continue; + } + + if (const PHINode *PN = dyn_cast<const PHINode>(P)) { + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) + Worklist.push_back(PN->getIncomingValue(i)); + continue; + } + } while (!Worklist.empty()); + + return false; +} + + /// @} /// /// \defgroup ARCOpt ARC Optimization. @@ -300,18 +340,18 @@ STATISTIC(NumNoops, "Number of no-op objc calls eliminated"); STATISTIC(NumPartialNoops, "Number of partially no-op objc calls eliminated"); STATISTIC(NumAutoreleases,"Number of autoreleases converted to releases"); STATISTIC(NumRets, "Number of return value forwarding " - "retain+autoreleaes eliminated"); + "retain+autoreleases eliminated"); STATISTIC(NumRRs, "Number of retain+release paths eliminated"); STATISTIC(NumPeeps, "Number of calls peephole-optimized"); +#ifndef NDEBUG STATISTIC(NumRetainsBeforeOpt, - "Number of retains before optimization."); + "Number of retains before optimization"); STATISTIC(NumReleasesBeforeOpt, - "Number of releases before optimization."); -#ifndef NDEBUG + "Number of releases before optimization"); STATISTIC(NumRetainsAfterOpt, - "Number of retains after optimization."); + "Number of retains after optimization"); STATISTIC(NumReleasesAfterOpt, - "Number of releases after optimization."); + "Number of releases after optimization"); #endif namespace { @@ -414,8 +454,18 @@ namespace { /// sequence. SmallPtrSet<Instruction *, 2> ReverseInsertPts; + /// Does this pointer have multiple owners? + /// + /// In the presence of multiple owners with the same provenance caused by + /// allocas, we can not assume that the frontend will emit balanced code + /// since it could put the release on the pointer loaded from the + /// alloca. This confuses the optimizer so we must be more conservative in + /// that case. + bool MultipleOwners; + RRInfo() : - KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0) {} + KnownSafe(false), IsTailCallRelease(false), ReleaseMetadata(0), + MultipleOwners(false) {} void clear(); @@ -428,6 +478,7 @@ namespace { void RRInfo::clear() { KnownSafe = false; IsTailCallRelease = false; + MultipleOwners = false; ReleaseMetadata = 0; Calls.clear(); ReverseInsertPts.clear(); @@ -457,10 +508,12 @@ namespace { Seq(S_None) {} void SetKnownPositiveRefCount() { + DEBUG(dbgs() << "Setting Known Positive.\n"); KnownPositiveRefCount = true; } void ClearKnownPositiveRefCount() { + DEBUG(dbgs() << "Clearing Known Positive.\n"); KnownPositiveRefCount = false; } @@ -516,6 +569,7 @@ PtrState::Merge(const PtrState &Other, bool TopDown) { RRI.IsTailCallRelease = RRI.IsTailCallRelease && Other.RRI.IsTailCallRelease; RRI.Calls.insert(Other.RRI.Calls.begin(), Other.RRI.Calls.end()); + RRI.MultipleOwners |= Other.RRI.MultipleOwners; // Merge the insert point sets. If there are any differences, // that makes this a partial merge. @@ -587,14 +641,26 @@ namespace { /// definition. void SetAsExit() { BottomUpPathCount = 1; } + /// Attempt to find the PtrState object describing the top down state for + /// pointer Arg. Return a new initialized PtrState describing the top down + /// state for Arg if we do not find one. PtrState &getPtrTopDownState(const Value *Arg) { return PerPtrTopDown[Arg]; } + /// Attempt to find the PtrState object describing the bottom up state for + /// pointer Arg. Return a new initialized PtrState describing the bottom up + /// state for Arg if we do not find one. PtrState &getPtrBottomUpState(const Value *Arg) { return PerPtrBottomUp[Arg]; } + /// Attempt to find the PtrState object describing the bottom up state for + /// pointer Arg. + ptr_iterator findPtrBottomUpState(const Value *Arg) { + return PerPtrBottomUp.find(Arg); + } + void clearBottomUpPointers() { PerPtrBottomUp.clear(); } @@ -1440,11 +1506,7 @@ void ObjCARCOpt::OptimizeIndividualCalls(Function &F) { case IC_RetainBlock: // If we strength reduce an objc_retainBlock to an objc_retain, continue // onto the objc_retain peephole optimizations. Otherwise break. - if (!OptimizeRetainBlockCall(F, Inst, Class)) - break; - // FALLTHROUGH - case IC_Retain: - ++NumRetainsBeforeOpt; + OptimizeRetainBlockCall(F, Inst, Class); break; case IC_RetainRV: if (OptimizeRetainRVCall(F, Inst)) @@ -1453,9 +1515,6 @@ void ObjCARCOpt::OptimizeIndividualCalls(Function &F) { case IC_AutoreleaseRV: OptimizeAutoreleaseRVCall(F, Inst, Class); break; - case IC_Release: - ++NumReleasesBeforeOpt; - break; } // objc_autorelease(x) -> objc_release(x) if x is otherwise unused. @@ -1866,6 +1925,28 @@ ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst, case IC_None: // These are irrelevant. return NestingDetected; + case IC_User: + // If we have a store into an alloca of a pointer we are tracking, the + // pointer has multiple owners implying that we must be more conservative. + // + // This comes up in the context of a pointer being ``KnownSafe''. In the + // presense of a block being initialized, the frontend will emit the + // objc_retain on the original pointer and the release on the pointer loaded + // from the alloca. The optimizer will through the provenance analysis + // realize that the two are related, but since we only require KnownSafe in + // one direction, will match the inner retain on the original pointer with + // the guard release on the original pointer. This is fixed by ensuring that + // in the presense of allocas we only unconditionally remove pointers if + // both our retain and our release are KnownSafe. + if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { + if (AreAnyUnderlyingObjectsAnAlloca(SI->getPointerOperand())) { + BBState::ptr_iterator I = MyStates.findPtrBottomUpState( + StripPointerCastsAndObjCCalls(SI->getValueOperand())); + if (I != MyStates.bottom_up_ptr_end()) + I->second.RRI.MultipleOwners = true; + } + } + break; default: break; } @@ -2412,8 +2493,10 @@ ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> bool KnownSafe, bool &AnyPairsCompletelyEliminated) { // If a pair happens in a region where it is known that the reference count - // is already incremented, we can similarly ignore possible decrements. + // is already incremented, we can similarly ignore possible decrements unless + // we are dealing with a retainable object with multiple provenance sources. bool KnownSafeTD = true, KnownSafeBU = true; + bool MultipleOwners = false; // Connect the dots between the top-down-collected RetainsToMove and // bottom-up-collected ReleasesToMove to form sets of related calls. @@ -2432,6 +2515,7 @@ ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> assert(It != Retains.end()); const RRInfo &NewRetainRRI = It->second; KnownSafeTD &= NewRetainRRI.KnownSafe; + MultipleOwners |= NewRetainRRI.MultipleOwners; for (SmallPtrSet<Instruction *, 2>::const_iterator LI = NewRetainRRI.Calls.begin(), LE = NewRetainRRI.Calls.end(); LI != LE; ++LI) { @@ -2525,9 +2609,12 @@ ObjCARCOpt::ConnectTDBUTraversals(DenseMap<const BasicBlock *, BBState> if (NewRetains.empty()) break; } - // If the pointer is known incremented or nested, we can safely delete the - // pair regardless of what's between them. - if (KnownSafeTD || KnownSafeBU) { + // If the pointer is known incremented in 1 direction and we do not have + // MultipleOwners, we can safely remove the retain/releases. Otherwise we need + // to be known safe in both directions. + bool UnconditionallySafe = (KnownSafeTD && KnownSafeBU) || + ((KnownSafeTD || KnownSafeBU) && !MultipleOwners); + if (UnconditionallySafe) { RetainsToMove.ReverseInsertPts.clear(); ReleasesToMove.ReverseInsertPts.clear(); NewCount = 0; @@ -3050,6 +3137,12 @@ bool ObjCARCOpt::runOnFunction(Function &F) { PA.setAA(&getAnalysis<AliasAnalysis>()); +#ifndef NDEBUG + if (AreStatisticsEnabled()) { + GatherStatistics(F, false); + } +#endif + // This pass performs several distinct transformations. As a compile-time aid // when compiling code that isn't ObjC, skip these if the relevant ObjC // library functions aren't declared. diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp index 0dd6abb1ae..58a1a74655 100644 --- a/lib/Transforms/Vectorize/LoopVectorize.cpp +++ b/lib/Transforms/Vectorize/LoopVectorize.cpp @@ -318,6 +318,93 @@ private: ValueMap WidenMap; }; +/// \brief Check if conditionally executed loads are hoistable. +/// +/// This class has two functions: isHoistableLoad and canHoistAllLoads. +/// isHoistableLoad should be called on all load instructions that are executed +/// conditionally. After all conditional loads are processed, the client should +/// call canHoistAllLoads to determine if all of the conditional executed loads +/// have an unconditional memory access to the same memory address in the loop. +class LoadHoisting { + typedef SmallPtrSet<Value *, 8> MemorySet; + + Loop *TheLoop; + DominatorTree *DT; + MemorySet CondLoadAddrSet; + +public: + LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {} + + /// \brief Check if the instruction is a load with a identifiable address. + bool isHoistableLoad(Instruction *L); + + /// \brief Check if all of the conditional loads are hoistable because there + /// exists an unconditional memory access to the same address in the loop. + bool canHoistAllLoads(); +}; + +bool LoadHoisting::isHoistableLoad(Instruction *L) { + LoadInst *LI = dyn_cast<LoadInst>(L); + if (!LI) + return false; + + CondLoadAddrSet.insert(LI->getPointerOperand()); + return true; +} + +static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) { + for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) { + Instruction *I = &*BI; + Value *Addr = 0; + + // Try a load. + LoadInst *LI = dyn_cast<LoadInst>(I); + if (LI) { + Addr = LI->getPointerOperand(); + Set.insert(Addr); + continue; + } + + // Try a store. + StoreInst *SI = dyn_cast<StoreInst>(I); + if (!SI) + continue; + + Addr = SI->getPointerOperand(); + Set.insert(Addr); + } +} + +bool LoadHoisting::canHoistAllLoads() { + // No conditional loads. + if (CondLoadAddrSet.empty()) + return true; + + MemorySet UncondMemAccesses; + std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); + BasicBlock *LoopLatch = TheLoop->getLoopLatch(); + + // Iterate over the unconditional blocks and collect memory access addresses. + for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { + BasicBlock *BB = LoopBlocks[i]; + + // Ignore conditional blocks. + if (BB != LoopLatch && !DT->dominates(BB, LoopLatch)) + continue; + + addMemAccesses(BB, UncondMemAccesses); + } + + // And make sure there is a matching unconditional access for every + // conditional load. + for (MemorySet::iterator MI = CondLoadAddrSet.begin(), + ME = CondLoadAddrSet.end(); MI != ME; ++MI) + if (!UncondMemAccesses.count(*MI)) + return false; + + return true; +} + /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and /// to what vectorization factor. /// This class does not look at the profitability of vectorization, only the @@ -337,7 +424,8 @@ public: DominatorTree *DT, TargetTransformInfo* TTI, AliasAnalysis *AA, TargetLibraryInfo *TLI) : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI), - Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false) {} + Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false), + LoadSpeculation(L, DT) {} /// This enum represents the kinds of reductions that we support. enum ReductionKind { @@ -598,6 +686,9 @@ private: RuntimePointerCheck PtrRtCheck; /// Can we assume the absence of NaNs. bool HasFunNoNaNAttr; + + /// Utility to determine whether loads can be speculated. + LoadHoisting LoadSpeculation; }; /// LoopVectorizationCostModel - estimates the expected speedups due to @@ -1389,9 +1480,10 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { case LoopVectorizationLegality::IK_NoInduction: llvm_unreachable("Unknown induction"); case LoopVectorizationLegality::IK_IntInduction: { - // Handle the integer induction counter: + // Handle the integer induction counter. assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); - assert(OrigPhi == OldInduction && "Unknown integer PHI"); + + // We have the canonical induction variable. if (OrigPhi == OldInduction) { // Create a truncated version of the resume value for the scalar loop, // we might have promoted the type to a larger width. @@ -1402,11 +1494,20 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); TruncResumeVal->addIncoming(EndValue, VecBody); + + // We know what the end value is. + EndValue = IdxEndRoundDown; + // We also know which PHI node holds it. + ResumeIndex = ResumeVal; + break; } - // We know what the end value is. - EndValue = IdxEndRoundDown; - // We also know which PHI node holds it. - ResumeIndex = ResumeVal; + + // Not the canonical induction variable - add the vector loop count to the + // start value. + Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, + II.StartValue->getType(), + "cast.crd"); + EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end"); break; } case LoopVectorizationLegality::IK_ReverseIntInduction: { @@ -2056,12 +2157,25 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, case LoopVectorizationLegality::IK_NoInduction: llvm_unreachable("Unknown induction"); case LoopVectorizationLegality::IK_IntInduction: { - assert(P == OldInduction && "Unexpected PHI"); - // We might have had to extend the type. - Value *Trunc = Builder.CreateTrunc(Induction, P->getType()); - Value *Broadcasted = getBroadcastInstrs(Trunc); - // After broadcasting the induction variable we need to make the - // vector consecutive by adding 0, 1, 2 ... + assert(P->getType() == II.StartValue->getType() && "Types must match"); + Type *PhiTy = P->getType(); + Value *Broadcasted; + if (P == OldInduction) { + // Handle the canonical induction variable. We might have had to + // extend the type. + Broadcasted = Builder.CreateTrunc(Induction, PhiTy); + } else { + // Handle other induction variables that are now based on the + // canonical one. + Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, + "normalized.idx"); + NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); + Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx, + "offset.idx"); + } + Broadcasted = getBroadcastInstrs(Broadcasted); + // After broadcasting the induction variable we need to make the vector + // consecutive by adding 0, 1, 2, etc. for (unsigned part = 0; part < UF; ++part) Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); continue; @@ -2466,11 +2580,11 @@ bool LoopVectorizationLegality::canVectorizeInstrs() { // Int inductions are special because we only allow one IV. if (IK == IK_IntInduction) { - if (Induction) { - DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); - return false; - } - Induction = Phi; + // Use the phi node with the widest type as induction. Use the last + // one if there are multiple (no good reason for doing this other + // than it is expedient). + if (!Induction || PhiTy == WidestIndTy) + Induction = Phi; } DEBUG(dbgs() << "LV: Found an induction variable.\n"); @@ -3236,8 +3350,12 @@ bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { - // We don't predicate loads/stores at the moment. - if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow()) + // We might be able to hoist the load. + if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it)) + return false; + + // We don't predicate stores at the moment. + if (it->mayWriteToMemory() || it->mayThrow()) return false; // The instructions below can trap. @@ -3251,6 +3369,10 @@ bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { } } + // Check that we can actually speculate the hoistable loads. + if (!LoadSpeculation.canHoistAllLoads()) + return false; + return true; } diff --git a/lib/Transforms/Vectorize/VecUtils.cpp b/lib/Transforms/Vectorize/VecUtils.cpp index 55adf8a816..50d2af0f65 100644 --- a/lib/Transforms/Vectorize/VecUtils.cpp +++ b/lib/Transforms/Vectorize/VecUtils.cpp @@ -282,6 +282,7 @@ int BoUpSLP::getTreeCost(ArrayRef<Value *> VL) { DEBUG(dbgs()<<"SLP: Adding to MustExtract " "because of a safe out of tree usage.\n"); MustExtract.insert(*it); + continue; } if (Lane == -1) Lane = LaneMap[*I]; if (Lane != LaneMap[*I]) { @@ -610,6 +611,9 @@ Value *BoUpSLP::Scalarize(ArrayRef<Value *> VL, VectorType *Ty) { GatherInstructions.push_back(Vec); } + for (unsigned i = 0; i < Ty->getNumElements(); ++i) + VectorizedValues[VL[i]] = Vec; + return Vec; } @@ -617,6 +621,7 @@ Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL, int VF) { Value *V = vectorizeTree_rec(VL, VF); Instruction *LastInstr = GetLastInstr(VL, VL.size()); + int LastInstrIdx = InstrIdx[LastInstr]; IRBuilder<> Builder(LastInstr); for (ValueSet::iterator it = MustExtract.begin(), e = MustExtract.end(); it != e; ++it) { @@ -625,7 +630,16 @@ Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL, int VF) { assert(LaneMap.count(I) && "Unable to find the lane for the external use"); Value *Idx = Builder.getInt32(LaneMap[I]); Value *Extract = Builder.CreateExtractElement(Vec, Idx); - I->replaceAllUsesWith(Extract); + bool Replaced = false; + for (Value::use_iterator U = I->use_begin(), UE = U->use_end(); U != UE; + ++U) { + Instruction *UI = cast<Instruction>(*U); + if (UI->getParent() != I->getParent() || InstrIdx[UI] > LastInstrIdx) + UI->replaceUsesOfWith(I ,Extract); + Replaced = true; + } + assert(Replaced && "Must replace at least one outside user"); + (void)Replaced; } // We moved some instructions around. We have to number them again @@ -633,6 +647,7 @@ Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL, int VF) { numberInstructions(); MustScalarize.clear(); MustExtract.clear(); + VectorizedValues.clear(); return V; } @@ -690,7 +705,10 @@ Value *BoUpSLP::vectorizeTree_rec(ArrayRef<Value *> VL, int VF) { IRBuilder<> Builder(GetLastInstr(VL, VF)); CastInst *CI = dyn_cast<CastInst>(VL0); Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); - VectorizedValues[VL0] = V; + + for (int i = 0; i < VF; ++i) + VectorizedValues[VL[i]] = V; + return V; } case Instruction::Add: @@ -713,16 +731,19 @@ Value *BoUpSLP::vectorizeTree_rec(ArrayRef<Value *> VL, int VF) { case Instruction::Xor: { ValueList LHSVL, RHSVL; for (int i = 0; i < VF; ++i) { - RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0)); - LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1)); + LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0)); + RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1)); } - Value *RHS = vectorizeTree_rec(RHSVL, VF); Value *LHS = vectorizeTree_rec(LHSVL, VF); + Value *RHS = vectorizeTree_rec(RHSVL, VF); IRBuilder<> Builder(GetLastInstr(VL, VF)); BinaryOperator *BinOp = cast<BinaryOperator>(VL0); - Value *V = Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS); - VectorizedValues[VL0] = V; + Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS,RHS); + + for (int i = 0; i < VF; ++i) + VectorizedValues[VL[i]] = V; + return V; } case Instruction::Load: { @@ -739,7 +760,10 @@ Value *BoUpSLP::vectorizeTree_rec(ArrayRef<Value *> VL, int VF) { VecTy->getPointerTo()); LI = Builder.CreateLoad(VecPtr); LI->setAlignment(Alignment); - VectorizedValues[VL0] = LI; + + for (int i = 0; i < VF; ++i) + VectorizedValues[VL[i]] = LI; + return LI; } case Instruction::Store: { @@ -762,9 +786,7 @@ Value *BoUpSLP::vectorizeTree_rec(ArrayRef<Value *> VL, int VF) { return 0; } default: - Value *S = Scalarize(VL, VecTy); - VectorizedValues[VL0] = S; - return S; + return Scalarize(VL, VecTy); } } |