X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FVectorize%2FSLPVectorizer.cpp;h=f69a4e52c7e1e86441bfd91351cd338374202892;hp=b49b1b0ff5ca2936fa2fcac0777e9390addbe84a;hb=8bb7acb4c17874dd85a9acd009295c5d5cb0fbf9;hpb=06ace3c8697a80f87773f51eb355c9f85fb01e61 diff --git a/lib/Transforms/Vectorize/SLPVectorizer.cpp b/lib/Transforms/Vectorize/SLPVectorizer.cpp index b49b1b0ff5c..f69a4e52c7e 100644 --- a/lib/Transforms/Vectorize/SLPVectorizer.cpp +++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp @@ -17,9 +17,14 @@ //===----------------------------------------------------------------------===// #include "llvm/Transforms/Vectorize.h" #include "llvm/ADT/MapVector.h" +#include "llvm/ADT/Optional.h" #include "llvm/ADT/PostOrderIterator.h" #include "llvm/ADT/SetVector.h" +#include "llvm/ADT/Statistic.h" #include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/GlobalsModRef.h" +#include "llvm/Analysis/AssumptionCache.h" +#include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" @@ -31,6 +36,7 @@ #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Module.h" +#include "llvm/IR/NoFolder.h" #include "llvm/IR/Type.h" #include "llvm/IR/Value.h" #include "llvm/IR/Verifier.h" @@ -38,22 +44,25 @@ #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Transforms/Utils/VectorUtils.h" +#include "llvm/Analysis/VectorUtils.h" #include #include +#include using namespace llvm; #define SV_NAME "slp-vectorizer" #define DEBUG_TYPE "SLP" +STATISTIC(NumVectorInstructions, "Number of vector instructions generated"); + static cl::opt SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, cl::desc("Only vectorize if you gain more than this " "number ")); static cl::opt -ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden, +ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, cl::desc("Attempt to vectorize horizontal reductions")); static cl::opt ShouldStartVectorizeHorAtStore( @@ -61,60 +70,49 @@ static cl::opt ShouldStartVectorizeHorAtStore( cl::desc( "Attempt to vectorize horizontal reductions feeding into a store")); +static cl::opt +MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, + cl::desc("Attempt to vectorize for this register size in bits")); + +/// Limits the size of scheduling regions in a block. +/// It avoid long compile times for _very_ large blocks where vector +/// instructions are spread over a wide range. +/// This limit is way higher than needed by real-world functions. +static cl::opt +ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, + cl::desc("Limit the size of the SLP scheduling region per block")); + namespace { +// FIXME: Set this via cl::opt to allow overriding. static const unsigned MinVecRegSize = 128; static const unsigned RecursionMaxDepth = 12; -/// A helper class for numbering instructions in multiple blocks. -/// Numbers start at zero for each basic block. -struct BlockNumbering { - - BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {} - - BlockNumbering() : BB(nullptr), Valid(false) {} - - void numberInstructions() { - unsigned Loc = 0; - InstrIdx.clear(); - InstrVec.clear(); - // Number the instructions in the block. - for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { - InstrIdx[it] = Loc++; - InstrVec.push_back(it); - assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation"); - } - Valid = true; - } - - int getIndex(Instruction *I) { - assert(I->getParent() == BB && "Invalid instruction"); - if (!Valid) - numberInstructions(); - assert(InstrIdx.count(I) && "Unknown instruction"); - return InstrIdx[I]; - } +// Limit the number of alias checks. The limit is chosen so that +// it has no negative effect on the llvm benchmarks. +static const unsigned AliasedCheckLimit = 10; - Instruction *getInstruction(unsigned loc) { - if (!Valid) - numberInstructions(); - assert(InstrVec.size() > loc && "Invalid Index"); - return InstrVec[loc]; - } +// Another limit for the alias checks: The maximum distance between load/store +// instructions where alias checks are done. +// This limit is useful for very large basic blocks. +static const unsigned MaxMemDepDistance = 160; - void forget() { Valid = false; } +/// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling +/// regions to be handled. +static const int MinScheduleRegionSize = 16; -private: - /// The block we are numbering. - BasicBlock *BB; - /// Is the block numbered. - bool Valid; - /// Maps instructions to numbers and back. - SmallDenseMap InstrIdx; - /// Maps integers to Instructions. - SmallVector InstrVec; -}; +/// \brief Predicate for the element types that the SLP vectorizer supports. +/// +/// The most important thing to filter here are types which are invalid in LLVM +/// vectors. We also filter target specific types which have absolutely no +/// meaningful vectorization path such as x86_fp80 and ppc_f128. This just +/// avoids spending time checking the cost model and realizing that they will +/// be inevitably scalarized. +static bool isValidElementType(Type *Ty) { + return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && + !Ty->isPPC_FP128Ty(); +} /// \returns the parent basic block if all of the instructions in \p VL /// are in the same block or null otherwise. @@ -150,6 +148,46 @@ static bool isSplat(ArrayRef VL) { return true; } +///\returns Opcode that can be clubbed with \p Op to create an alternate +/// sequence which can later be merged as a ShuffleVector instruction. +static unsigned getAltOpcode(unsigned Op) { + switch (Op) { + case Instruction::FAdd: + return Instruction::FSub; + case Instruction::FSub: + return Instruction::FAdd; + case Instruction::Add: + return Instruction::Sub; + case Instruction::Sub: + return Instruction::Add; + default: + return 0; + } +} + +///\returns bool representing if Opcode \p Op can be part +/// of an alternate sequence which can later be merged as +/// a ShuffleVector instruction. +static bool canCombineAsAltInst(unsigned Op) { + return Op == Instruction::FAdd || Op == Instruction::FSub || + Op == Instruction::Sub || Op == Instruction::Add; +} + +/// \returns ShuffleVector instruction if instructions in \p VL have +/// alternate fadd,fsub / fsub,fadd/add,sub/sub,add sequence. +/// (i.e. e.g. opcodes of fadd,fsub,fadd,fsub...) +static unsigned isAltInst(ArrayRef VL) { + Instruction *I0 = dyn_cast(VL[0]); + unsigned Opcode = I0->getOpcode(); + unsigned AltOpcode = getAltOpcode(Opcode); + for (int i = 1, e = VL.size(); i < e; i++) { + Instruction *I = dyn_cast(VL[i]); + if (!I || I->getOpcode() != ((i & 1) ? AltOpcode : Opcode)) + return 0; + } + return Instruction::ShuffleVector; +} + /// \returns The opcode if all of the Instructions in \p VL have the same /// opcode, or zero. static unsigned getSameOpcode(ArrayRef VL) { @@ -159,12 +197,32 @@ static unsigned getSameOpcode(ArrayRef VL) { unsigned Opcode = I0->getOpcode(); for (int i = 1, e = VL.size(); i < e; i++) { Instruction *I = dyn_cast(VL[i]); - if (!I || Opcode != I->getOpcode()) + if (!I || Opcode != I->getOpcode()) { + if (canCombineAsAltInst(Opcode) && i == 1) + return isAltInst(VL); return 0; + } } return Opcode; } +/// Get the intersection (logical and) of all of the potential IR flags +/// of each scalar operation (VL) that will be converted into a vector (I). +/// Flag set: NSW, NUW, exact, and all of fast-math. +static void propagateIRFlags(Value *I, ArrayRef VL) { + if (auto *VecOp = dyn_cast(I)) { + if (auto *Intersection = dyn_cast(VL[0])) { + // Intersection is initialized to the 0th scalar, + // so start counting from index '1'. + for (int i = 1, e = VL.size(); i < e; ++i) { + if (auto *Scalar = dyn_cast(VL[i])) + Intersection->andIRFlags(Scalar); + } + VecOp->copyIRFlags(Intersection); + } + } +} + /// \returns \p I after propagating metadata from \p VL. static Instruction *propagateMetadata(Instruction *I, ArrayRef VL) { Instruction *I0 = cast(VL[0]); @@ -186,9 +244,18 @@ static Instruction *propagateMetadata(Instruction *I, ArrayRef VL) { case LLVMContext::MD_tbaa: MD = MDNode::getMostGenericTBAA(MD, IMD); break; + case LLVMContext::MD_alias_scope: + MD = MDNode::getMostGenericAliasScope(MD, IMD); + break; + case LLVMContext::MD_noalias: + MD = MDNode::intersect(MD, IMD); + break; case LLVMContext::MD_fpmath: MD = MDNode::getMostGenericFPMath(MD, IMD); break; + case LLVMContext::MD_nontemporal: + MD = MDNode::intersect(MD, IMD); + break; } } I->setMetadata(Kind, MD); @@ -239,102 +306,51 @@ static bool CanReuseExtract(ArrayRef VL) { return true; } -static void reorderInputsAccordingToOpcode(ArrayRef VL, - SmallVectorImpl &Left, - SmallVectorImpl &Right) { - - SmallVector OrigLeft, OrigRight; - - bool AllSameOpcodeLeft = true; - bool AllSameOpcodeRight = true; - for (unsigned i = 0, e = VL.size(); i != e; ++i) { - Instruction *I = cast(VL[i]); - Value *V0 = I->getOperand(0); - Value *V1 = I->getOperand(1); +/// \returns True if in-tree use also needs extract. This refers to +/// possible scalar operand in vectorized instruction. +static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, + TargetLibraryInfo *TLI) { - OrigLeft.push_back(V0); - OrigRight.push_back(V1); - - Instruction *I0 = dyn_cast(V0); - Instruction *I1 = dyn_cast(V1); - - // Check whether all operands on one side have the same opcode. In this case - // we want to preserve the original order and not make things worse by - // reordering. - AllSameOpcodeLeft = I0; - AllSameOpcodeRight = I1; - - if (i && AllSameOpcodeLeft) { - if(Instruction *P0 = dyn_cast(OrigLeft[i-1])) { - if(P0->getOpcode() != I0->getOpcode()) - AllSameOpcodeLeft = false; - } else - AllSameOpcodeLeft = false; - } - if (i && AllSameOpcodeRight) { - if(Instruction *P1 = dyn_cast(OrigRight[i-1])) { - if(P1->getOpcode() != I1->getOpcode()) - AllSameOpcodeRight = false; - } else - AllSameOpcodeRight = false; - } - - // Sort two opcodes. In the code below we try to preserve the ability to use - // broadcast of values instead of individual inserts. - // vl1 = load - // vl2 = phi - // vr1 = load - // vr2 = vr2 - // = vl1 x vr1 - // = vl2 x vr2 - // If we just sorted according to opcode we would leave the first line in - // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load). - // = vl1 x vr1 - // = vr2 x vl2 - // Because vr2 and vr1 are from the same load we loose the opportunity of a - // broadcast for the packed right side in the backend: we have [vr1, vl2] - // instead of [vr1, vr2=vr1]. - if (I0 && I1) { - if(!i && I0->getOpcode() > I1->getOpcode()) { - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) { - // Try not to destroy a broad cast for no apparent benefit. - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) { - // Try preserve broadcasts. - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) { - // Try preserve broadcasts. - Left.push_back(I1); - Right.push_back(I0); - } else { - Left.push_back(I0); - Right.push_back(I1); - } - continue; - } - // One opcode, put the instruction on the right. - if (I0) { - Left.push_back(V1); - Right.push_back(I0); - continue; + unsigned Opcode = UserInst->getOpcode(); + switch (Opcode) { + case Instruction::Load: { + LoadInst *LI = cast(UserInst); + return (LI->getPointerOperand() == Scalar); + } + case Instruction::Store: { + StoreInst *SI = cast(UserInst); + return (SI->getPointerOperand() == Scalar); + } + case Instruction::Call: { + CallInst *CI = cast(UserInst); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); + if (hasVectorInstrinsicScalarOpd(ID, 1)) { + return (CI->getArgOperand(1) == Scalar); } - Left.push_back(V0); - Right.push_back(V1); } + default: + return false; + } +} - bool LeftBroadcast = isSplat(Left); - bool RightBroadcast = isSplat(Right); +/// \returns the AA location that is being access by the instruction. +static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) { + if (StoreInst *SI = dyn_cast(I)) + return MemoryLocation::get(SI); + if (LoadInst *LI = dyn_cast(I)) + return MemoryLocation::get(LI); + return MemoryLocation(); +} - // Don't reorder if the operands where good to begin with. - if (!(LeftBroadcast || RightBroadcast) && - (AllSameOpcodeRight || AllSameOpcodeLeft)) { - Left = OrigLeft; - Right = OrigRight; - } +/// \returns True if the instruction is not a volatile or atomic load/store. +static bool isSimple(Instruction *I) { + if (LoadInst *LI = dyn_cast(I)) + return LI->isSimple(); + if (StoreInst *SI = dyn_cast(I)) + return SI->isSimple(); + if (MemIntrinsic *MI = dyn_cast(I)) + return !MI->isVolatile(); + return true; } /// Bottom Up SLP Vectorizer. @@ -345,46 +361,57 @@ public: typedef SmallPtrSet ValueSet; typedef SmallVector StoreList; - BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl, - TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li, - DominatorTree *Dt) : - F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt), - Builder(Se->getContext()) { - // Setup the block numbering utility for all of the blocks in the - // function. - for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) { - BasicBlock *BB = it; - BlocksNumbers[BB] = BlockNumbering(BB); - } - } + BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, + TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li, + DominatorTree *Dt, AssumptionCache *AC) + : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), F(Func), + SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), + Builder(Se->getContext()) { + CodeMetrics::collectEphemeralValues(F, AC, EphValues); + } /// \brief Vectorize the tree that starts with the elements in \p VL. /// Returns the vectorized root. Value *vectorizeTree(); + /// \returns the cost incurred by unwanted spills and fills, caused by + /// holding live values over call sites. + int getSpillCost(); + /// \returns the vectorization cost of the subtree that starts at \p VL. /// A negative number means that this is profitable. int getTreeCost(); - /// Construct a vectorizable tree that starts at \p Roots and is possibly - /// used by a reduction of \p RdxOps. - void buildTree(ArrayRef Roots, ValueSet *RdxOps = 0); + /// Construct a vectorizable tree that starts at \p Roots, ignoring users for + /// the purpose of scheduling and extraction in the \p UserIgnoreLst. + void buildTree(ArrayRef Roots, + ArrayRef UserIgnoreLst = None); /// Clear the internal data structures that are created by 'buildTree'. void deleteTree() { - RdxOps = 0; VectorizableTree.clear(); ScalarToTreeEntry.clear(); MustGather.clear(); ExternalUses.clear(); - MemBarrierIgnoreList.clear(); + NumLoadsWantToKeepOrder = 0; + NumLoadsWantToChangeOrder = 0; + for (auto &Iter : BlocksSchedules) { + BlockScheduling *BS = Iter.second.get(); + BS->clear(); + } } /// \returns true if the memory operations A and B are consecutive. - bool isConsecutiveAccess(Value *A, Value *B); + bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL); /// \brief Perform LICM and CSE on the newly generated gather sequences. void optimizeGatherSequence(); + + /// \returns true if it is beneficial to reverse the vector order. + bool shouldReorder() const { + return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder; + } + private: struct TreeEntry; @@ -421,20 +448,6 @@ private: /// roots. This method calculates the cost of extracting the values. int getGatherCost(ArrayRef VL); - /// \returns the AA location that is being access by the instruction. - AliasAnalysis::Location getLocation(Instruction *I); - - /// \brief Checks if it is possible to sink an instruction from - /// \p Src to \p Dst. - /// \returns the pointer to the barrier instruction if we can't sink. - Value *getSinkBarrier(Instruction *Src, Instruction *Dst); - - /// \returns the index of the last instruction in the BB from \p VL. - int getLastIndex(ArrayRef VL); - - /// \returns the Instruction in the bundle \p VL. - Instruction *getLastInstruction(ArrayRef VL); - /// \brief Set the Builder insert point to one after the last instruction in /// the bundle void setInsertPointAfterBundle(ArrayRef VL); @@ -442,12 +455,22 @@ private: /// \returns a vector from a collection of scalars in \p VL. Value *Gather(ArrayRef VL, VectorType *Ty); - /// \returns whether the VectorizableTree is fully vectoriable and will + /// \returns whether the VectorizableTree is fully vectorizable and will /// be beneficial even the tree height is tiny. bool isFullyVectorizableTinyTree(); + /// \reorder commutative operands in alt shuffle if they result in + /// vectorized code. + void reorderAltShuffleOperands(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right); + /// \reorder commutative operands to get better probability of + /// generating vectorized code. + void reorderInputsAccordingToOpcode(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right); struct TreeEntry { - TreeEntry() : Scalars(), VectorizedValue(nullptr), LastScalarIndex(0), + TreeEntry() : Scalars(), VectorizedValue(nullptr), NeedToGather(0) {} /// \returns true if the scalars in VL are equal to this entry. @@ -462,28 +485,23 @@ private: /// The Scalars are vectorized into this value. It is initialized to Null. Value *VectorizedValue; - /// The index in the basic block of the last scalar. - int LastScalarIndex; - /// Do we need to gather this sequence ? bool NeedToGather; }; /// Create a new VectorizableTree entry. TreeEntry *newTreeEntry(ArrayRef VL, bool Vectorized) { - VectorizableTree.push_back(TreeEntry()); + VectorizableTree.emplace_back(); int idx = VectorizableTree.size() - 1; TreeEntry *Last = &VectorizableTree[idx]; Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); Last->NeedToGather = !Vectorized; if (Vectorized) { - Last->LastScalarIndex = getLastIndex(VL); for (int i = 0, e = VL.size(); i != e; ++i) { assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!"); ScalarToTreeEntry[VL[i]] = idx; } } else { - Last->LastScalarIndex = 0; MustGather.insert(VL.begin(), VL.end()); } return Last; @@ -502,7 +520,7 @@ private: /// This POD struct describes one external user in the vectorized tree. struct ExternalUser { ExternalUser (Value *S, llvm::User *U, int L) : - Scalar(S), User(U), Lane(L){}; + Scalar(S), User(U), Lane(L){} // Which scalar in our function. Value *Scalar; // Which user that uses the scalar. @@ -512,30 +530,390 @@ private: }; typedef SmallVector UserList; + /// Checks if two instructions may access the same memory. + /// + /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it + /// is invariant in the calling loop. + bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, + Instruction *Inst2) { + + // First check if the result is already in the cache. + AliasCacheKey key = std::make_pair(Inst1, Inst2); + Optional &result = AliasCache[key]; + if (result.hasValue()) { + return result.getValue(); + } + MemoryLocation Loc2 = getLocation(Inst2, AA); + bool aliased = true; + if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { + // Do the alias check. + aliased = AA->alias(Loc1, Loc2); + } + // Store the result in the cache. + result = aliased; + return aliased; + } + + typedef std::pair AliasCacheKey; + + /// Cache for alias results. + /// TODO: consider moving this to the AliasAnalysis itself. + DenseMap> AliasCache; + + /// Removes an instruction from its block and eventually deletes it. + /// It's like Instruction::eraseFromParent() except that the actual deletion + /// is delayed until BoUpSLP is destructed. + /// This is required to ensure that there are no incorrect collisions in the + /// AliasCache, which can happen if a new instruction is allocated at the + /// same address as a previously deleted instruction. + void eraseInstruction(Instruction *I) { + I->removeFromParent(); + I->dropAllReferences(); + DeletedInstructions.push_back(std::unique_ptr(I)); + } + + /// Temporary store for deleted instructions. Instructions will be deleted + /// eventually when the BoUpSLP is destructed. + SmallVector, 8> DeletedInstructions; + /// A list of values that need to extracted out of the tree. /// This list holds pairs of (Internal Scalar : External User). UserList ExternalUses; - /// A list of instructions to ignore while sinking - /// memory instructions. This map must be reset between runs of getCost. - ValueSet MemBarrierIgnoreList; + /// Values used only by @llvm.assume calls. + SmallPtrSet EphValues; /// Holds all of the instructions that we gathered. SetVector GatherSeq; /// A list of blocks that we are going to CSE. SetVector CSEBlocks; - /// Numbers instructions in different blocks. - DenseMap BlocksNumbers; + /// Contains all scheduling relevant data for an instruction. + /// A ScheduleData either represents a single instruction or a member of an + /// instruction bundle (= a group of instructions which is combined into a + /// vector instruction). + struct ScheduleData { + + // The initial value for the dependency counters. It means that the + // dependencies are not calculated yet. + enum { InvalidDeps = -1 }; + + ScheduleData() + : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr), + NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0), + Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps), + UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {} + + void init(int BlockSchedulingRegionID) { + FirstInBundle = this; + NextInBundle = nullptr; + NextLoadStore = nullptr; + IsScheduled = false; + SchedulingRegionID = BlockSchedulingRegionID; + UnscheduledDepsInBundle = UnscheduledDeps; + clearDependencies(); + } - /// Reduction operators. - ValueSet *RdxOps; + /// Returns true if the dependency information has been calculated. + bool hasValidDependencies() const { return Dependencies != InvalidDeps; } + + /// Returns true for single instructions and for bundle representatives + /// (= the head of a bundle). + bool isSchedulingEntity() const { return FirstInBundle == this; } + + /// Returns true if it represents an instruction bundle and not only a + /// single instruction. + bool isPartOfBundle() const { + return NextInBundle != nullptr || FirstInBundle != this; + } + + /// Returns true if it is ready for scheduling, i.e. it has no more + /// unscheduled depending instructions/bundles. + bool isReady() const { + assert(isSchedulingEntity() && + "can't consider non-scheduling entity for ready list"); + return UnscheduledDepsInBundle == 0 && !IsScheduled; + } + + /// Modifies the number of unscheduled dependencies, also updating it for + /// the whole bundle. + int incrementUnscheduledDeps(int Incr) { + UnscheduledDeps += Incr; + return FirstInBundle->UnscheduledDepsInBundle += Incr; + } + + /// Sets the number of unscheduled dependencies to the number of + /// dependencies. + void resetUnscheduledDeps() { + incrementUnscheduledDeps(Dependencies - UnscheduledDeps); + } + + /// Clears all dependency information. + void clearDependencies() { + Dependencies = InvalidDeps; + resetUnscheduledDeps(); + MemoryDependencies.clear(); + } + + void dump(raw_ostream &os) const { + if (!isSchedulingEntity()) { + os << "/ " << *Inst; + } else if (NextInBundle) { + os << '[' << *Inst; + ScheduleData *SD = NextInBundle; + while (SD) { + os << ';' << *SD->Inst; + SD = SD->NextInBundle; + } + os << ']'; + } else { + os << *Inst; + } + } + + Instruction *Inst; + + /// Points to the head in an instruction bundle (and always to this for + /// single instructions). + ScheduleData *FirstInBundle; + + /// Single linked list of all instructions in a bundle. Null if it is a + /// single instruction. + ScheduleData *NextInBundle; + + /// Single linked list of all memory instructions (e.g. load, store, call) + /// in the block - until the end of the scheduling region. + ScheduleData *NextLoadStore; + + /// The dependent memory instructions. + /// This list is derived on demand in calculateDependencies(). + SmallVector MemoryDependencies; + + /// This ScheduleData is in the current scheduling region if this matches + /// the current SchedulingRegionID of BlockScheduling. + int SchedulingRegionID; + + /// Used for getting a "good" final ordering of instructions. + int SchedulingPriority; + + /// The number of dependencies. Constitutes of the number of users of the + /// instruction plus the number of dependent memory instructions (if any). + /// This value is calculated on demand. + /// If InvalidDeps, the number of dependencies is not calculated yet. + /// + int Dependencies; + + /// The number of dependencies minus the number of dependencies of scheduled + /// instructions. As soon as this is zero, the instruction/bundle gets ready + /// for scheduling. + /// Note that this is negative as long as Dependencies is not calculated. + int UnscheduledDeps; + + /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for + /// single instructions. + int UnscheduledDepsInBundle; + + /// True if this instruction is scheduled (or considered as scheduled in the + /// dry-run). + bool IsScheduled; + }; + +#ifndef NDEBUG + friend raw_ostream &operator<<(raw_ostream &os, + const BoUpSLP::ScheduleData &SD); +#endif + + /// Contains all scheduling data for a basic block. + /// + struct BlockScheduling { + + BlockScheduling(BasicBlock *BB) + : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize), + ScheduleStart(nullptr), ScheduleEnd(nullptr), + FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr), + ScheduleRegionSize(0), + ScheduleRegionSizeLimit(ScheduleRegionSizeBudget), + // Make sure that the initial SchedulingRegionID is greater than the + // initial SchedulingRegionID in ScheduleData (which is 0). + SchedulingRegionID(1) {} + + void clear() { + ReadyInsts.clear(); + ScheduleStart = nullptr; + ScheduleEnd = nullptr; + FirstLoadStoreInRegion = nullptr; + LastLoadStoreInRegion = nullptr; + + // Reduce the maximum schedule region size by the size of the + // previous scheduling run. + ScheduleRegionSizeLimit -= ScheduleRegionSize; + if (ScheduleRegionSizeLimit < MinScheduleRegionSize) + ScheduleRegionSizeLimit = MinScheduleRegionSize; + ScheduleRegionSize = 0; + + // Make a new scheduling region, i.e. all existing ScheduleData is not + // in the new region yet. + ++SchedulingRegionID; + } + + ScheduleData *getScheduleData(Value *V) { + ScheduleData *SD = ScheduleDataMap[V]; + if (SD && SD->SchedulingRegionID == SchedulingRegionID) + return SD; + return nullptr; + } + + bool isInSchedulingRegion(ScheduleData *SD) { + return SD->SchedulingRegionID == SchedulingRegionID; + } + + /// Marks an instruction as scheduled and puts all dependent ready + /// instructions into the ready-list. + template + void schedule(ScheduleData *SD, ReadyListType &ReadyList) { + SD->IsScheduled = true; + DEBUG(dbgs() << "SLP: schedule " << *SD << "\n"); + + ScheduleData *BundleMember = SD; + while (BundleMember) { + // Handle the def-use chain dependencies. + for (Use &U : BundleMember->Inst->operands()) { + ScheduleData *OpDef = getScheduleData(U.get()); + if (OpDef && OpDef->hasValidDependencies() && + OpDef->incrementUnscheduledDeps(-1) == 0) { + // There are no more unscheduled dependencies after decrementing, + // so we can put the dependent instruction into the ready list. + ScheduleData *DepBundle = OpDef->FirstInBundle; + assert(!DepBundle->IsScheduled && + "already scheduled bundle gets ready"); + ReadyList.insert(DepBundle); + DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"); + } + } + // Handle the memory dependencies. + for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { + if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { + // There are no more unscheduled dependencies after decrementing, + // so we can put the dependent instruction into the ready list. + ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; + assert(!DepBundle->IsScheduled && + "already scheduled bundle gets ready"); + ReadyList.insert(DepBundle); + DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"); + } + } + BundleMember = BundleMember->NextInBundle; + } + } + + /// Put all instructions into the ReadyList which are ready for scheduling. + template + void initialFillReadyList(ReadyListType &ReadyList) { + for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + if (SD->isSchedulingEntity() && SD->isReady()) { + ReadyList.insert(SD); + DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n"); + } + } + } + + /// Checks if a bundle of instructions can be scheduled, i.e. has no + /// cyclic dependencies. This is only a dry-run, no instructions are + /// actually moved at this stage. + bool tryScheduleBundle(ArrayRef VL, BoUpSLP *SLP); + + /// Un-bundles a group of instructions. + void cancelScheduling(ArrayRef VL); + + /// Extends the scheduling region so that V is inside the region. + /// \returns true if the region size is within the limit. + bool extendSchedulingRegion(Value *V); + + /// Initialize the ScheduleData structures for new instructions in the + /// scheduling region. + void initScheduleData(Instruction *FromI, Instruction *ToI, + ScheduleData *PrevLoadStore, + ScheduleData *NextLoadStore); + + /// Updates the dependency information of a bundle and of all instructions/ + /// bundles which depend on the original bundle. + void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, + BoUpSLP *SLP); + + /// Sets all instruction in the scheduling region to un-scheduled. + void resetSchedule(); + + BasicBlock *BB; + + /// Simple memory allocation for ScheduleData. + std::vector> ScheduleDataChunks; + + /// The size of a ScheduleData array in ScheduleDataChunks. + int ChunkSize; + + /// The allocator position in the current chunk, which is the last entry + /// of ScheduleDataChunks. + int ChunkPos; + + /// Attaches ScheduleData to Instruction. + /// Note that the mapping survives during all vectorization iterations, i.e. + /// ScheduleData structures are recycled. + DenseMap ScheduleDataMap; + + struct ReadyList : SmallVector { + void insert(ScheduleData *SD) { push_back(SD); } + }; + + /// The ready-list for scheduling (only used for the dry-run). + ReadyList ReadyInsts; + + /// The first instruction of the scheduling region. + Instruction *ScheduleStart; + + /// The first instruction _after_ the scheduling region. + Instruction *ScheduleEnd; + + /// The first memory accessing instruction in the scheduling region + /// (can be null). + ScheduleData *FirstLoadStoreInRegion; + + /// The last memory accessing instruction in the scheduling region + /// (can be null). + ScheduleData *LastLoadStoreInRegion; + + /// The current size of the scheduling region. + int ScheduleRegionSize; + + /// The maximum size allowed for the scheduling region. + int ScheduleRegionSizeLimit; + + /// The ID of the scheduling region. For a new vectorization iteration this + /// is incremented which "removes" all ScheduleData from the region. + int SchedulingRegionID; + }; + + /// Attaches the BlockScheduling structures to basic blocks. + MapVector> BlocksSchedules; + + /// Performs the "real" scheduling. Done before vectorization is actually + /// performed in a basic block. + void scheduleBlock(BlockScheduling *BS); + + /// List of users to ignore during scheduling and that don't need extracting. + ArrayRef UserIgnoreList; + + // Number of load-bundles, which contain consecutive loads. + int NumLoadsWantToKeepOrder; + + // Number of load-bundles of size 2, which are consecutive loads if reversed. + int NumLoadsWantToChangeOrder; // Analysis and block reference. Function *F; ScalarEvolution *SE; - const DataLayout *DL; TargetTransformInfo *TTI; + TargetLibraryInfo *TLI; AliasAnalysis *AA; LoopInfo *LI; DominatorTree *DT; @@ -543,9 +921,17 @@ private: IRBuilder<> Builder; }; -void BoUpSLP::buildTree(ArrayRef Roots, ValueSet *Rdx) { +#ifndef NDEBUG +raw_ostream &operator<<(raw_ostream &os, const BoUpSLP::ScheduleData &SD) { + SD.dump(os); + return os; +} +#endif + +void BoUpSLP::buildTree(ArrayRef Roots, + ArrayRef UserIgnoreLst) { deleteTree(); - RdxOps = Rdx; + UserIgnoreList = UserIgnoreLst; if (!getSameType(Roots)) return; buildTree_rec(Roots, 0); @@ -565,20 +951,30 @@ void BoUpSLP::buildTree(ArrayRef Roots, ValueSet *Rdx) { for (User *U : Scalar->users()) { DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n"); - // Skip in-tree scalars that become vectors. - if (ScalarToTreeEntry.count(U)) { - DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << - *U << ".\n"); - int Idx = ScalarToTreeEntry[U]; (void) Idx; - assert(!VectorizableTree[Idx].NeedToGather && "Bad state"); - continue; - } Instruction *UserInst = dyn_cast(U); if (!UserInst) continue; - // Ignore uses that are part of the reduction. - if (Rdx && std::find(Rdx->begin(), Rdx->end(), UserInst) != Rdx->end()) + // Skip in-tree scalars that become vectors + if (ScalarToTreeEntry.count(U)) { + int Idx = ScalarToTreeEntry[U]; + TreeEntry *UseEntry = &VectorizableTree[Idx]; + Value *UseScalar = UseEntry->Scalars[0]; + // Some in-tree scalars will remain as scalar in vectorized + // instructions. If that is the case, the one in Lane 0 will + // be used. + if (UseScalar != U || + !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { + DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *U + << ".\n"); + assert(!VectorizableTree[Idx].NeedToGather && "Bad state"); + continue; + } + } + + // Ignore users in the user ignore list. + if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UserInst) != + UserIgnoreList.end()) continue; DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " << @@ -592,6 +988,7 @@ void BoUpSLP::buildTree(ArrayRef Roots, ValueSet *Rdx) { void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { bool SameTy = getSameType(VL); (void)SameTy; + bool isAltShuffle = false; assert(SameTy && "Invalid types!"); if (Depth == RecursionMaxDepth) { @@ -613,10 +1010,19 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { newTreeEntry(VL, false); return; } + unsigned Opcode = getSameOpcode(VL); + + // Check that this shuffle vector refers to the alternate + // sequence of opcodes. + if (Opcode == Instruction::ShuffleVector) { + Instruction *I0 = dyn_cast(VL[0]); + unsigned Op = I0->getOpcode(); + if (Op != Instruction::ShuffleVector) + isAltShuffle = true; + } // If all of the operands are identical or constant we have a simple solution. - if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) || - !getSameOpcode(VL)) { + if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) || !Opcode) { DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n"); newTreeEntry(VL, false); return; @@ -625,6 +1031,16 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // We now know that this is a vector of instructions of the same type from // the same block. + // Don't vectorize ephemeral values. + for (unsigned i = 0, e = VL.size(); i != e; ++i) { + if (EphValues.count(VL[i])) { + DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] << + ") is ephemeral.\n"); + newTreeEntry(VL, false); + return; + } + } + // Check if this is a duplicate of another entry. if (ScalarToTreeEntry.count(VL[0])) { int Idx = ScalarToTreeEntry[VL[0]]; @@ -651,11 +1067,11 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { } } - // If any of the scalars appears in the table OR it is marked as a value that - // needs to stat scalar then we need to gather the scalars. + // If any of the scalars is marked as a value that needs to stay scalar then + // we need to gather the scalars. for (unsigned i = 0, e = VL.size(); i != e; ++i) { - if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) { - DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n"); + if (MustGather.count(VL[i])) { + DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n"); newTreeEntry(VL, false); return; } @@ -664,66 +1080,14 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // Check that all of the users of the scalars that we want to vectorize are // schedulable. Instruction *VL0 = cast(VL[0]); - int MyLastIndex = getLastIndex(VL); BasicBlock *BB = cast(VL0)->getParent(); - for (unsigned i = 0, e = VL.size(); i != e; ++i) { - Instruction *Scalar = cast(VL[i]); - DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n"); - for (User *U : Scalar->users()) { - DEBUG(dbgs() << "SLP: \tUser " << *U << ". \n"); - Instruction *UI = dyn_cast(U); - if (!UI) { - DEBUG(dbgs() << "SLP: Gathering due unknown user. \n"); - newTreeEntry(VL, false); - return; - } - - // We don't care if the user is in a different basic block. - BasicBlock *UserBlock = UI->getParent(); - if (UserBlock != BB) { - DEBUG(dbgs() << "SLP: User from a different basic block " - << *UI << ". \n"); - continue; - } - - // If this is a PHINode within this basic block then we can place the - // extract wherever we want. - if (isa(*UI)) { - DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *UI << ". \n"); - continue; - } - - // Check if this is a safe in-tree user. - if (ScalarToTreeEntry.count(UI)) { - int Idx = ScalarToTreeEntry[UI]; - int VecLocation = VectorizableTree[Idx].LastScalarIndex; - if (VecLocation <= MyLastIndex) { - DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n"); - newTreeEntry(VL, false); - return; - } - DEBUG(dbgs() << "SLP: In-tree user (" << *UI << ") at #" << - VecLocation << " vector value (" << *Scalar << ") at #" - << MyLastIndex << ".\n"); - continue; - } - - // This user is part of the reduction. - if (RdxOps && RdxOps->count(UI)) - continue; - - // Make sure that we can schedule this unknown user. - BlockNumbering &BN = BlocksNumbers[BB]; - int UserIndex = BN.getIndex(UI); - if (UserIndex < MyLastIndex) { - - DEBUG(dbgs() << "SLP: Can't schedule extractelement for " - << *UI << ". \n"); - newTreeEntry(VL, false); - return; - } - } + if (!DT->isReachableFromEntry(BB)) { + // Don't go into unreachable blocks. They may contain instructions with + // dependency cycles which confuse the final scheduling. + DEBUG(dbgs() << "SLP: bundle in unreachable block.\n"); + newTreeEntry(VL, false); + return; } // Check that every instructions appears once in this bundle. @@ -735,40 +1099,21 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { return; } - // Check that instructions in this bundle don't reference other instructions. - // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4. - for (unsigned i = 0, e = VL.size(); i < e; ++i) { - for (User *U : VL[i]->users()) { - for (unsigned j = 0; j < e; ++j) { - if (i != j && U == VL[j]) { - DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n"); - newTreeEntry(VL, false); - return; - } - } - } + auto &BSRef = BlocksSchedules[BB]; + if (!BSRef) { + BSRef = llvm::make_unique(BB); } + BlockScheduling &BS = *BSRef.get(); - DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); - - unsigned Opcode = getSameOpcode(VL); - - // Check if it is safe to sink the loads or the stores. - if (Opcode == Instruction::Load || Opcode == Instruction::Store) { - Instruction *Last = getLastInstruction(VL); - - for (unsigned i = 0, e = VL.size(); i < e; ++i) { - if (VL[i] == Last) - continue; - Value *Barrier = getSinkBarrier(cast(VL[i]), Last); - if (Barrier) { - DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last - << "\n because of " << *Barrier << ". Gathering.\n"); - newTreeEntry(VL, false); - return; - } - } + if (!BS.tryScheduleBundle(VL, this)) { + DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n"); + assert((!BS.getScheduleData(VL[0]) || + !BS.getScheduleData(VL[0])->isPartOfBundle()) && + "tryScheduleBundle should cancelScheduling on failure"); + newTreeEntry(VL, false); + return; } + DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); switch (Opcode) { case Instruction::PHI: { @@ -781,6 +1126,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { cast(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i))); if (Term) { DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n"); + BS.cancelScheduling(VL); newTreeEntry(VL, false); return; } @@ -804,20 +1150,51 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { bool Reuse = CanReuseExtract(VL); if (Reuse) { DEBUG(dbgs() << "SLP: Reusing extract sequence.\n"); + } else { + BS.cancelScheduling(VL); } newTreeEntry(VL, Reuse); return; } case Instruction::Load: { + // Check that a vectorized load would load the same memory as a scalar + // load. + // For example we don't want vectorize loads that are smaller than 8 bit. + // Even though we have a packed struct {} LLVM treats + // loading/storing it as an i8 struct. If we vectorize loads/stores from + // such a struct we read/write packed bits disagreeing with the + // unvectorized version. + const DataLayout &DL = F->getParent()->getDataLayout(); + Type *ScalarTy = VL[0]->getType(); + + if (DL.getTypeSizeInBits(ScalarTy) != + DL.getTypeAllocSizeInBits(ScalarTy)) { + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n"); + return; + } // Check if the loads are consecutive or of we need to swizzle them. for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { LoadInst *L = cast(VL[i]); - if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) { + if (!L->isSimple()) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); - DEBUG(dbgs() << "SLP: Need to swizzle loads.\n"); + DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n"); + return; + } + + if (!isConsecutiveAccess(VL[i], VL[i + 1], DL)) { + if (VL.size() == 2 && isConsecutiveAccess(VL[1], VL[0], DL)) { + ++NumLoadsWantToChangeOrder; + } + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n"); return; } } + ++NumLoadsWantToKeepOrder; newTreeEntry(VL, true); DEBUG(dbgs() << "SLP: added a vector of loads.\n"); return; @@ -837,7 +1214,8 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { Type *SrcTy = VL0->getOperand(0)->getType(); for (unsigned i = 0; i < VL.size(); ++i) { Type *Ty = cast(VL[i])->getOperand(0)->getType(); - if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) { + if (Ty != SrcTy || !isValidElementType(Ty)) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n"); return; @@ -859,12 +1237,13 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { case Instruction::ICmp: case Instruction::FCmp: { // Check that all of the compares have the same predicate. - CmpInst::Predicate P0 = dyn_cast(VL0)->getPredicate(); + CmpInst::Predicate P0 = cast(VL0)->getPredicate(); Type *ComparedTy = cast(VL[0])->getOperand(0)->getType(); for (unsigned i = 1, e = VL.size(); i < e; ++i) { CmpInst *Cmp = cast(VL[i]); if (Cmp->getPredicate() != P0 || Cmp->getOperand(0)->getType() != ComparedTy) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n"); return; @@ -926,10 +1305,60 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { } return; } + case Instruction::GetElementPtr: { + // We don't combine GEPs with complicated (nested) indexing. + for (unsigned j = 0; j < VL.size(); ++j) { + if (cast(VL[j])->getNumOperands() != 2) { + DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n"); + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + return; + } + } + + // We can't combine several GEPs into one vector if they operate on + // different types. + Type *Ty0 = cast(VL0)->getOperand(0)->getType(); + for (unsigned j = 0; j < VL.size(); ++j) { + Type *CurTy = cast(VL[j])->getOperand(0)->getType(); + if (Ty0 != CurTy) { + DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n"); + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + return; + } + } + + // We don't combine GEPs with non-constant indexes. + for (unsigned j = 0; j < VL.size(); ++j) { + auto Op = cast(VL[j])->getOperand(1); + if (!isa(Op)) { + DEBUG( + dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"); + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + return; + } + } + + newTreeEntry(VL, true); + DEBUG(dbgs() << "SLP: added a vector of GEPs.\n"); + for (unsigned i = 0, e = 2; i < e; ++i) { + ValueList Operands; + // Prepare the operand vector. + for (unsigned j = 0; j < VL.size(); ++j) + Operands.push_back(cast(VL[j])->getOperand(i)); + + buildTree_rec(Operands, Depth + 1); + } + return; + } case Instruction::Store: { + const DataLayout &DL = F->getParent()->getDataLayout(); // Check if the stores are consecutive or of we need to swizzle them. for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) - if (!isConsecutiveAccess(VL[i], VL[i + 1])) { + if (!isConsecutiveAccess(VL[i], VL[i + 1], DL)) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Non-consecutive store.\n"); return; @@ -942,47 +1371,95 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { for (unsigned j = 0; j < VL.size(); ++j) Operands.push_back(cast(VL[j])->getOperand(0)); - // We can ignore these values because we are sinking them down. - MemBarrierIgnoreList.insert(VL.begin(), VL.end()); buildTree_rec(Operands, Depth + 1); return; } case Instruction::Call: { // Check if the calls are all to the same vectorizable intrinsic. - IntrinsicInst *II = dyn_cast(VL[0]); - Intrinsic::ID ID = II ? II->getIntrinsicID() : Intrinsic::not_intrinsic; - + CallInst *CI = cast(VL[0]); + // Check if this is an Intrinsic call or something that can be + // represented by an intrinsic call + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); if (!isTriviallyVectorizable(ID)) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Non-vectorizable call.\n"); return; } - - Function *Int = II->getCalledFunction(); - + Function *Int = CI->getCalledFunction(); + Value *A1I = nullptr; + if (hasVectorInstrinsicScalarOpd(ID, 1)) + A1I = CI->getArgOperand(1); for (unsigned i = 1, e = VL.size(); i != e; ++i) { - IntrinsicInst *II2 = dyn_cast(VL[i]); - if (!II2 || II2->getCalledFunction() != Int) { + CallInst *CI2 = dyn_cast(VL[i]); + if (!CI2 || CI2->getCalledFunction() != Int || + getIntrinsicIDForCall(CI2, TLI) != ID) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); - DEBUG(dbgs() << "SLP: mismatched calls:" << *II << "!=" << *VL[i] + DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i] << "\n"); return; } + // ctlz,cttz and powi are special intrinsics whose second argument + // should be same in order for them to be vectorized. + if (hasVectorInstrinsicScalarOpd(ID, 1)) { + Value *A1J = CI2->getArgOperand(1); + if (A1I != A1J) { + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI + << " argument "<< A1I<<"!=" << A1J + << "\n"); + return; + } + } } newTreeEntry(VL, true); - for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i) { + for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { ValueList Operands; // Prepare the operand vector. for (unsigned j = 0; j < VL.size(); ++j) { - IntrinsicInst *II2 = dyn_cast(VL[j]); - Operands.push_back(II2->getArgOperand(i)); + CallInst *CI2 = dyn_cast(VL[j]); + Operands.push_back(CI2->getArgOperand(i)); } buildTree_rec(Operands, Depth + 1); } return; } + case Instruction::ShuffleVector: { + // If this is not an alternate sequence of opcode like add-sub + // then do not vectorize this instruction. + if (!isAltShuffle) { + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n"); + return; + } + newTreeEntry(VL, true); + DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n"); + + // Reorder operands if reordering would enable vectorization. + if (isa(VL0)) { + ValueList Left, Right; + reorderAltShuffleOperands(VL, Left, Right); + buildTree_rec(Left, Depth + 1); + buildTree_rec(Right, Depth + 1); + return; + } + + for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { + ValueList Operands; + // Prepare the operand vector. + for (unsigned j = 0; j < VL.size(); ++j) + Operands.push_back(cast(VL[j])->getOperand(i)); + + buildTree_rec(Operands, Depth + 1); + } + return; + } default: + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n"); return; @@ -1005,11 +1482,9 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { } return getGatherCost(E->Scalars); } - - assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) && - "Invalid VL"); + unsigned Opcode = getSameOpcode(VL); + assert(Opcode && getSameType(VL) && getSameBlock(VL) && "Invalid VL"); Instruction *VL0 = cast(VL[0]); - unsigned Opcode = VL0->getOpcode(); switch (Opcode) { case Instruction::PHI: { return 0; @@ -1087,6 +1562,10 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { TargetTransformInfo::OK_AnyValue; TargetTransformInfo::OperandValueKind Op2VK = TargetTransformInfo::OK_UniformConstantValue; + TargetTransformInfo::OperandValueProperties Op1VP = + TargetTransformInfo::OP_None; + TargetTransformInfo::OperandValueProperties Op2VP = + TargetTransformInfo::OP_None; // If all operands are exactly the same ConstantInt then set the // operand kind to OK_UniformConstantValue. @@ -1108,14 +1587,34 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { CInt != cast(I->getOperand(1))) Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; } + // FIXME: Currently cost of model modification for division by + // power of 2 is handled only for X86. Add support for other targets. + if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt && + CInt->getValue().isPowerOf2()) + Op2VP = TargetTransformInfo::OP_PowerOf2; - ScalarCost = - VecTy->getNumElements() * - TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK); - VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK); + ScalarCost = VecTy->getNumElements() * + TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK, + Op1VP, Op2VP); + VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK, + Op1VP, Op2VP); } return VecCost - ScalarCost; } + case Instruction::GetElementPtr: { + TargetTransformInfo::OperandValueKind Op1VK = + TargetTransformInfo::OK_AnyValue; + TargetTransformInfo::OperandValueKind Op2VK = + TargetTransformInfo::OK_UniformConstantValue; + + int ScalarCost = + VecTy->getNumElements() * + TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK); + int VecCost = + TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK); + + return VecCost - ScalarCost; + } case Instruction::Load: { // Cost of wide load - cost of scalar loads. int ScalarLdCost = VecTy->getNumElements() * @@ -1132,12 +1631,11 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { } case Instruction::Call: { CallInst *CI = cast(VL0); - IntrinsicInst *II = cast(CI); - Intrinsic::ID ID = II->getIntrinsicID(); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); // Calculate the cost of the scalar and vector calls. SmallVector ScalarTys, VecTys; - for (unsigned op = 0, opc = II->getNumArgOperands(); op!= opc; ++op) { + for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) { ScalarTys.push_back(CI->getArgOperand(op)->getType()); VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(), VecTy->getNumElements())); @@ -1150,10 +1648,36 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" - << " for " << *II << "\n"); + << " for " << *CI << "\n"); return VecCallCost - ScalarCallCost; } + case Instruction::ShuffleVector: { + TargetTransformInfo::OperandValueKind Op1VK = + TargetTransformInfo::OK_AnyValue; + TargetTransformInfo::OperandValueKind Op2VK = + TargetTransformInfo::OK_AnyValue; + int ScalarCost = 0; + int VecCost = 0; + for (unsigned i = 0; i < VL.size(); ++i) { + Instruction *I = cast(VL[i]); + if (!I) + break; + ScalarCost += + TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK); + } + // VecCost is equal to sum of the cost of creating 2 vectors + // and the cost of creating shuffle. + Instruction *I0 = cast(VL[0]); + VecCost = + TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK); + Instruction *I1 = cast(VL[1]); + VecCost += + TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK); + VecCost += + TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0); + return VecCost - ScalarCost; + } default: llvm_unreachable("Unknown instruction"); } @@ -1167,8 +1691,10 @@ bool BoUpSLP::isFullyVectorizableTinyTree() { if (VectorizableTree.size() != 2) return false; - // Handle splat stores. - if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars)) + // Handle splat and all-constants stores. + if (!VectorizableTree[0].NeedToGather && + (allConstant(VectorizableTree[1].Scalars) || + isSplat(VectorizableTree[1].Scalars))) return true; // Gathering cost would be too much for tiny trees. @@ -1178,6 +1704,69 @@ bool BoUpSLP::isFullyVectorizableTinyTree() { return true; } +int BoUpSLP::getSpillCost() { + // Walk from the bottom of the tree to the top, tracking which values are + // live. When we see a call instruction that is not part of our tree, + // query TTI to see if there is a cost to keeping values live over it + // (for example, if spills and fills are required). + unsigned BundleWidth = VectorizableTree.front().Scalars.size(); + int Cost = 0; + + SmallPtrSet LiveValues; + Instruction *PrevInst = nullptr; + + for (unsigned N = 0; N < VectorizableTree.size(); ++N) { + Instruction *Inst = dyn_cast(VectorizableTree[N].Scalars[0]); + if (!Inst) + continue; + + if (!PrevInst) { + PrevInst = Inst; + continue; + } + + DEBUG( + dbgs() << "SLP: #LV: " << LiveValues.size(); + for (auto *X : LiveValues) + dbgs() << " " << X->getName(); + dbgs() << ", Looking at "; + Inst->dump(); + ); + + // Update LiveValues. + LiveValues.erase(PrevInst); + for (auto &J : PrevInst->operands()) { + if (isa(&*J) && ScalarToTreeEntry.count(&*J)) + LiveValues.insert(cast(&*J)); + } + + // Now find the sequence of instructions between PrevInst and Inst. + BasicBlock::reverse_iterator InstIt(Inst->getIterator()), + PrevInstIt(PrevInst->getIterator()); + --PrevInstIt; + while (InstIt != PrevInstIt) { + if (PrevInstIt == PrevInst->getParent()->rend()) { + PrevInstIt = Inst->getParent()->rbegin(); + continue; + } + + if (isa(&*PrevInstIt) && &*PrevInstIt != PrevInst) { + SmallVector V; + for (auto *II : LiveValues) + V.push_back(VectorType::get(II->getType(), BundleWidth)); + Cost += TTI->getCostOfKeepingLiveOverCall(V); + } + + ++PrevInstIt; + } + + PrevInst = Inst; + } + + DEBUG(dbgs() << "SLP: SpillCost=" << Cost << "\n"); + return Cost; +} + int BoUpSLP::getTreeCost() { int Cost = 0; DEBUG(dbgs() << "SLP: Calculating cost for tree of size " << @@ -1185,7 +1774,7 @@ int BoUpSLP::getTreeCost() { // We only vectorize tiny trees if it is fully vectorizable. if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) { - if (!VectorizableTree.size()) { + if (VectorizableTree.empty()) { assert(!ExternalUses.size() && "We should not have any external users"); } return INT_MAX; @@ -1193,26 +1782,33 @@ int BoUpSLP::getTreeCost() { unsigned BundleWidth = VectorizableTree[0].Scalars.size(); - for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) { - int C = getEntryCost(&VectorizableTree[i]); + for (TreeEntry &TE : VectorizableTree) { + int C = getEntryCost(&TE); DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " - << *VectorizableTree[i].Scalars[0] << " .\n"); + << TE.Scalars[0] << " .\n"); Cost += C; } SmallSet ExtractCostCalculated; int ExtractCost = 0; - for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end(); - I != E; ++I) { + for (ExternalUser &EU : ExternalUses) { // We only add extract cost once for the same scalar. - if (!ExtractCostCalculated.insert(I->Scalar)) + if (!ExtractCostCalculated.insert(EU.Scalar).second) continue; - VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth); + // Uses by ephemeral values are free (because the ephemeral value will be + // removed prior to code generation, and so the extraction will be + // removed as well). + if (EphValues.count(EU.User)) + continue; + + VectorType *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth); ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, - I->Lane); + EU.Lane); } + Cost += getSpillCost(); + DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n"); return Cost + ExtractCost; } @@ -1234,14 +1830,6 @@ int BoUpSLP::getGatherCost(ArrayRef VL) { return getGatherCost(VecTy); } -AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) { - if (StoreInst *SI = dyn_cast(I)) - return AA->getLocation(SI); - if (LoadInst *LI = dyn_cast(I)) - return AA->getLocation(LI); - return AliasAnalysis::Location(); -} - Value *BoUpSLP::getPointerOperand(Value *I) { if (LoadInst *LI = dyn_cast(I)) return LI->getPointerOperand(); @@ -1258,7 +1846,7 @@ unsigned BoUpSLP::getAddressSpaceOperand(Value *I) { return -1; } -bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) { +bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL) { Value *PtrA = getPointerOperand(A); Value *PtrB = getPointerOperand(B); unsigned ASA = getAddressSpaceOperand(A); @@ -1272,13 +1860,13 @@ bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) { if (PtrA == PtrB || PtrA->getType() != PtrB->getType()) return false; - unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA); + unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA); Type *Ty = cast(PtrA->getType())->getElementType(); - APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty)); + APInt Size(PtrBitWidth, DL.getTypeStoreSize(Ty)); APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0); - PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA); - PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB); + PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); + PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); APInt OffsetDelta = OffsetB - OffsetA; @@ -1299,59 +1887,221 @@ bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) { return X == PtrSCEVB; } -Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) { - assert(Src->getParent() == Dst->getParent() && "Not the same BB"); - BasicBlock::iterator I = Src, E = Dst; - /// Scan all of the instruction from SRC to DST and check if - /// the source may alias. - for (++I; I != E; ++I) { - // Ignore store instructions that are marked as 'ignore'. - if (MemBarrierIgnoreList.count(I)) - continue; - if (Src->mayWriteToMemory()) /* Write */ { - if (!I->mayReadOrWriteMemory()) - continue; - } else /* Read */ { - if (!I->mayWriteToMemory()) - continue; +// Reorder commutative operations in alternate shuffle if the resulting vectors +// are consecutive loads. This would allow us to vectorize the tree. +// If we have something like- +// load a[0] - load b[0] +// load b[1] + load a[1] +// load a[2] - load b[2] +// load a[3] + load b[3] +// Reordering the second load b[1] load a[1] would allow us to vectorize this +// code. +void BoUpSLP::reorderAltShuffleOperands(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right) { + const DataLayout &DL = F->getParent()->getDataLayout(); + + // Push left and right operands of binary operation into Left and Right + for (unsigned i = 0, e = VL.size(); i < e; ++i) { + Left.push_back(cast(VL[i])->getOperand(0)); + Right.push_back(cast(VL[i])->getOperand(1)); + } + + // Reorder if we have a commutative operation and consecutive access + // are on either side of the alternate instructions. + for (unsigned j = 0; j < VL.size() - 1; ++j) { + if (LoadInst *L = dyn_cast(Left[j])) { + if (LoadInst *L1 = dyn_cast(Right[j + 1])) { + Instruction *VL1 = cast(VL[j]); + Instruction *VL2 = cast(VL[j + 1]); + if (isConsecutiveAccess(L, L1, DL) && VL1->isCommutative()) { + std::swap(Left[j], Right[j]); + continue; + } else if (isConsecutiveAccess(L, L1, DL) && VL2->isCommutative()) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + // else unchanged + } + } + if (LoadInst *L = dyn_cast(Right[j])) { + if (LoadInst *L1 = dyn_cast(Left[j + 1])) { + Instruction *VL1 = cast(VL[j]); + Instruction *VL2 = cast(VL[j + 1]); + if (isConsecutiveAccess(L, L1, DL) && VL1->isCommutative()) { + std::swap(Left[j], Right[j]); + continue; + } else if (isConsecutiveAccess(L, L1, DL) && VL2->isCommutative()) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + // else unchanged + } + } + } +} + +// Return true if I should be commuted before adding it's left and right +// operands to the arrays Left and Right. +// +// The vectorizer is trying to either have all elements one side being +// instruction with the same opcode to enable further vectorization, or having +// a splat to lower the vectorizing cost. +static bool shouldReorderOperands(int i, Instruction &I, + SmallVectorImpl &Left, + SmallVectorImpl &Right, + bool AllSameOpcodeLeft, + bool AllSameOpcodeRight, bool SplatLeft, + bool SplatRight) { + Value *VLeft = I.getOperand(0); + Value *VRight = I.getOperand(1); + // If we have "SplatRight", try to see if commuting is needed to preserve it. + if (SplatRight) { + if (VRight == Right[i - 1]) + // Preserve SplatRight + return false; + if (VLeft == Right[i - 1]) { + // Commuting would preserve SplatRight, but we don't want to break + // SplatLeft either, i.e. preserve the original order if possible. + // (FIXME: why do we care?) + if (SplatLeft && VLeft == Left[i - 1]) + return false; + return true; } - AliasAnalysis::Location A = getLocation(&*I); - AliasAnalysis::Location B = getLocation(Src); + } + // Symmetrically handle Right side. + if (SplatLeft) { + if (VLeft == Left[i - 1]) + // Preserve SplatLeft + return false; + if (VRight == Left[i - 1]) + return true; + } + + Instruction *ILeft = dyn_cast(VLeft); + Instruction *IRight = dyn_cast(VRight); - if (!A.Ptr || !B.Ptr || AA->alias(A, B)) - return I; + // If we have "AllSameOpcodeRight", try to see if the left operands preserves + // it and not the right, in this case we want to commute. + if (AllSameOpcodeRight) { + unsigned RightPrevOpcode = cast(Right[i - 1])->getOpcode(); + if (IRight && RightPrevOpcode == IRight->getOpcode()) + // Do not commute, a match on the right preserves AllSameOpcodeRight + return false; + if (ILeft && RightPrevOpcode == ILeft->getOpcode()) { + // We have a match and may want to commute, but first check if there is + // not also a match on the existing operands on the Left to preserve + // AllSameOpcodeLeft, i.e. preserve the original order if possible. + // (FIXME: why do we care?) + if (AllSameOpcodeLeft && ILeft && + cast(Left[i - 1])->getOpcode() == ILeft->getOpcode()) + return false; + return true; + } } - return nullptr; + // Symmetrically handle Left side. + if (AllSameOpcodeLeft) { + unsigned LeftPrevOpcode = cast(Left[i - 1])->getOpcode(); + if (ILeft && LeftPrevOpcode == ILeft->getOpcode()) + return false; + if (IRight && LeftPrevOpcode == IRight->getOpcode()) + return true; + } + return false; } -int BoUpSLP::getLastIndex(ArrayRef VL) { - BasicBlock *BB = cast(VL[0])->getParent(); - assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block"); - BlockNumbering &BN = BlocksNumbers[BB]; +void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right) { + + if (VL.size()) { + // Peel the first iteration out of the loop since there's nothing + // interesting to do anyway and it simplifies the checks in the loop. + auto VLeft = cast(VL[0])->getOperand(0); + auto VRight = cast(VL[0])->getOperand(1); + if (!isa(VRight) && isa(VLeft)) + // Favor having instruction to the right. FIXME: why? + std::swap(VLeft, VRight); + Left.push_back(VLeft); + Right.push_back(VRight); + } - int MaxIdx = BN.getIndex(BB->getFirstNonPHI()); - for (unsigned i = 0, e = VL.size(); i < e; ++i) - MaxIdx = std::max(MaxIdx, BN.getIndex(cast(VL[i]))); - return MaxIdx; -} + // Keep track if we have instructions with all the same opcode on one side. + bool AllSameOpcodeLeft = isa(Left[0]); + bool AllSameOpcodeRight = isa(Right[0]); + // Keep track if we have one side with all the same value (broadcast). + bool SplatLeft = true; + bool SplatRight = true; -Instruction *BoUpSLP::getLastInstruction(ArrayRef VL) { - BasicBlock *BB = cast(VL[0])->getParent(); - assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block"); - BlockNumbering &BN = BlocksNumbers[BB]; + for (unsigned i = 1, e = VL.size(); i != e; ++i) { + Instruction *I = cast(VL[i]); + assert(I->isCommutative() && "Can only process commutative instruction"); + // Commute to favor either a splat or maximizing having the same opcodes on + // one side. + if (shouldReorderOperands(i, *I, Left, Right, AllSameOpcodeLeft, + AllSameOpcodeRight, SplatLeft, SplatRight)) { + Left.push_back(I->getOperand(1)); + Right.push_back(I->getOperand(0)); + } else { + Left.push_back(I->getOperand(0)); + Right.push_back(I->getOperand(1)); + } + // Update Splat* and AllSameOpcode* after the insertion. + SplatRight = SplatRight && (Right[i - 1] == Right[i]); + SplatLeft = SplatLeft && (Left[i - 1] == Left[i]); + AllSameOpcodeLeft = AllSameOpcodeLeft && isa(Left[i]) && + (cast(Left[i - 1])->getOpcode() == + cast(Left[i])->getOpcode()); + AllSameOpcodeRight = AllSameOpcodeRight && isa(Right[i]) && + (cast(Right[i - 1])->getOpcode() == + cast(Right[i])->getOpcode()); + } - int MaxIdx = BN.getIndex(cast(VL[0])); - for (unsigned i = 1, e = VL.size(); i < e; ++i) - MaxIdx = std::max(MaxIdx, BN.getIndex(cast(VL[i]))); - Instruction *I = BN.getInstruction(MaxIdx); - assert(I && "bad location"); - return I; + // If one operand end up being broadcast, return this operand order. + if (SplatRight || SplatLeft) + return; + + const DataLayout &DL = F->getParent()->getDataLayout(); + + // Finally check if we can get longer vectorizable chain by reordering + // without breaking the good operand order detected above. + // E.g. If we have something like- + // load a[0] load b[0] + // load b[1] load a[1] + // load a[2] load b[2] + // load a[3] load b[3] + // Reordering the second load b[1] load a[1] would allow us to vectorize + // this code and we still retain AllSameOpcode property. + // FIXME: This load reordering might break AllSameOpcode in some rare cases + // such as- + // add a[0],c[0] load b[0] + // add a[1],c[2] load b[1] + // b[2] load b[2] + // add a[3],c[3] load b[3] + for (unsigned j = 0; j < VL.size() - 1; ++j) { + if (LoadInst *L = dyn_cast(Left[j])) { + if (LoadInst *L1 = dyn_cast(Right[j + 1])) { + if (isConsecutiveAccess(L, L1, DL)) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + } + } + if (LoadInst *L = dyn_cast(Right[j])) { + if (LoadInst *L1 = dyn_cast(Left[j + 1])) { + if (isConsecutiveAccess(L, L1, DL)) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + } + } + // else unchanged + } } void BoUpSLP::setInsertPointAfterBundle(ArrayRef VL) { Instruction *VL0 = cast(VL[0]); - Instruction *LastInst = getLastInstruction(VL); - BasicBlock::iterator NextInst = LastInst; + BasicBlock::iterator NextInst(VL0); ++NextInst; Builder.SetInsertPoint(VL0->getParent(), NextInst); Builder.SetCurrentDebugLocation(VL0->getDebugLoc()); @@ -1435,8 +2185,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { return Gather(E->Scalars, VecTy); } - unsigned Opcode = VL0->getOpcode(); - assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode"); + const DataLayout &DL = F->getParent()->getDataLayout(); + unsigned Opcode = getSameOpcode(E->Scalars); switch (Opcode) { case Instruction::PHI: { @@ -1454,15 +2204,14 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { ValueList Operands; BasicBlock *IBB = PH->getIncomingBlock(i); - if (!VisitedBBs.insert(IBB)) { + if (!VisitedBBs.insert(IBB).second) { NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); continue; } // Prepare the operand vector. - for (unsigned j = 0; j < E->Scalars.size(); ++j) - Operands.push_back(cast(E->Scalars[j])-> - getIncomingValueForBlock(IBB)); + for (Value *V : E->Scalars) + Operands.push_back(cast(V)->getIncomingValueForBlock(IBB)); Builder.SetInsertPoint(IBB->getTerminator()); Builder.SetCurrentDebugLocation(PH->getDebugLoc()); @@ -1496,8 +2245,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { case Instruction::FPTrunc: case Instruction::BitCast: { ValueList INVL; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) - INVL.push_back(cast(E->Scalars[i])->getOperand(0)); + for (Value *V : E->Scalars) + INVL.push_back(cast(V)->getOperand(0)); setInsertPointAfterBundle(E->Scalars); @@ -1509,14 +2258,15 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { CastInst *CI = dyn_cast(VL0); Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::FCmp: case Instruction::ICmp: { ValueList LHSV, RHSV; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) { - LHSV.push_back(cast(E->Scalars[i])->getOperand(0)); - RHSV.push_back(cast(E->Scalars[i])->getOperand(1)); + for (Value *V : E->Scalars) { + LHSV.push_back(cast(V)->getOperand(0)); + RHSV.push_back(cast(V)->getOperand(1)); } setInsertPointAfterBundle(E->Scalars); @@ -1527,7 +2277,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { if (Value *V = alreadyVectorized(E->Scalars)) return V; - CmpInst::Predicate P0 = dyn_cast(VL0)->getPredicate(); + CmpInst::Predicate P0 = cast(VL0)->getPredicate(); Value *V; if (Opcode == Instruction::FCmp) V = Builder.CreateFCmp(P0, L, R); @@ -1535,14 +2285,15 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { V = Builder.CreateICmp(P0, L, R); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::Select: { ValueList TrueVec, FalseVec, CondVec; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) { - CondVec.push_back(cast(E->Scalars[i])->getOperand(0)); - TrueVec.push_back(cast(E->Scalars[i])->getOperand(1)); - FalseVec.push_back(cast(E->Scalars[i])->getOperand(2)); + for (Value *V : E->Scalars) { + CondVec.push_back(cast(V)->getOperand(0)); + TrueVec.push_back(cast(V)->getOperand(1)); + FalseVec.push_back(cast(V)->getOperand(2)); } setInsertPointAfterBundle(E->Scalars); @@ -1556,6 +2307,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Value *V = Builder.CreateSelect(Cond, True, False); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::Add: @@ -1580,9 +2332,9 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { if (isa(VL0) && VL0->isCommutative()) reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL); else - for (int i = 0, e = E->Scalars.size(); i < e; ++i) { - LHSVL.push_back(cast(E->Scalars[i])->getOperand(0)); - RHSVL.push_back(cast(E->Scalars[i])->getOperand(1)); + for (Value *V : E->Scalars) { + LHSVL.push_back(cast(V)->getOperand(0)); + RHSVL.push_back(cast(V)->getOperand(1)); } setInsertPointAfterBundle(E->Scalars); @@ -1600,6 +2352,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { BinaryOperator *BinOp = cast(VL0); Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS); E->VectorizedValue = V; + propagateIRFlags(E->VectorizedValue, E->Scalars); + ++NumVectorInstructions; if (Instruction *I = dyn_cast(V)) return propagateMetadata(I, E->Scalars); @@ -1612,14 +2366,27 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { setInsertPointAfterBundle(E->Scalars); LoadInst *LI = cast(VL0); + Type *ScalarLoadTy = LI->getType(); unsigned AS = LI->getPointerAddressSpace(); Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), VecTy->getPointerTo(AS)); + + // The pointer operand uses an in-tree scalar so we add the new BitCast to + // ExternalUses list to make sure that an extract will be generated in the + // future. + if (ScalarToTreeEntry.count(LI->getPointerOperand())) + ExternalUses.push_back( + ExternalUser(LI->getPointerOperand(), cast(VecPtr), 0)); + unsigned Alignment = LI->getAlignment(); LI = Builder.CreateLoad(VecPtr); + if (!Alignment) { + Alignment = DL.getABITypeAlignment(ScalarLoadTy); + } LI->setAlignment(Alignment); E->VectorizedValue = LI; + ++NumVectorInstructions; return propagateMetadata(LI, E->Scalars); } case Instruction::Store: { @@ -1628,8 +2395,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { unsigned AS = SI->getPointerAddressSpace(); ValueList ValueOp; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) - ValueOp.push_back(cast(E->Scalars[i])->getValueOperand()); + for (Value *V : E->Scalars) + ValueOp.push_back(cast(V)->getValueOperand()); setInsertPointAfterBundle(E->Scalars); @@ -1637,19 +2404,74 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(), VecTy->getPointerTo(AS)); StoreInst *S = Builder.CreateStore(VecValue, VecPtr); + + // The pointer operand uses an in-tree scalar so we add the new BitCast to + // ExternalUses list to make sure that an extract will be generated in the + // future. + if (ScalarToTreeEntry.count(SI->getPointerOperand())) + ExternalUses.push_back( + ExternalUser(SI->getPointerOperand(), cast(VecPtr), 0)); + + if (!Alignment) { + Alignment = DL.getABITypeAlignment(SI->getValueOperand()->getType()); + } S->setAlignment(Alignment); E->VectorizedValue = S; + ++NumVectorInstructions; return propagateMetadata(S, E->Scalars); } + case Instruction::GetElementPtr: { + setInsertPointAfterBundle(E->Scalars); + + ValueList Op0VL; + for (Value *V : E->Scalars) + Op0VL.push_back(cast(V)->getOperand(0)); + + Value *Op0 = vectorizeTree(Op0VL); + + std::vector OpVecs; + for (int j = 1, e = cast(VL0)->getNumOperands(); j < e; + ++j) { + ValueList OpVL; + for (Value *V : E->Scalars) + OpVL.push_back(cast(V)->getOperand(j)); + + Value *OpVec = vectorizeTree(OpVL); + OpVecs.push_back(OpVec); + } + + Value *V = Builder.CreateGEP( + cast(VL0)->getSourceElementType(), Op0, OpVecs); + E->VectorizedValue = V; + ++NumVectorInstructions; + + if (Instruction *I = dyn_cast(V)) + return propagateMetadata(I, E->Scalars); + + return V; + } case Instruction::Call: { CallInst *CI = cast(VL0); - setInsertPointAfterBundle(E->Scalars); + Function *FI; + Intrinsic::ID IID = Intrinsic::not_intrinsic; + Value *ScalarArg = nullptr; + if (CI && (FI = CI->getCalledFunction())) { + IID = FI->getIntrinsicID(); + } std::vector OpVecs; for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { ValueList OpVL; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) { - CallInst *CEI = cast(E->Scalars[i]); + // ctlz,cttz and powi are special intrinsics whose second argument is + // a scalar. This argument should not be vectorized. + if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) { + CallInst *CEI = cast(E->Scalars[0]); + ScalarArg = CEI->getArgOperand(j); + OpVecs.push_back(CEI->getArgOperand(j)); + continue; + } + for (Value *V : E->Scalars) { + CallInst *CEI = cast(V); OpVL.push_back(CEI->getArgOperand(j)); } @@ -1659,12 +2481,68 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { } Module *M = F->getParent(); - IntrinsicInst *II = cast(CI); - Intrinsic::ID ID = II->getIntrinsicID(); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) }; Function *CF = Intrinsic::getDeclaration(M, ID, Tys); Value *V = Builder.CreateCall(CF, OpVecs); + + // The scalar argument uses an in-tree scalar so we add the new vectorized + // call to ExternalUses list to make sure that an extract will be + // generated in the future. + if (ScalarArg && ScalarToTreeEntry.count(ScalarArg)) + ExternalUses.push_back(ExternalUser(ScalarArg, cast(V), 0)); + + E->VectorizedValue = V; + ++NumVectorInstructions; + return V; + } + case Instruction::ShuffleVector: { + ValueList LHSVL, RHSVL; + assert(isa(VL0) && "Invalid Shuffle Vector Operand"); + reorderAltShuffleOperands(E->Scalars, LHSVL, RHSVL); + setInsertPointAfterBundle(E->Scalars); + + Value *LHS = vectorizeTree(LHSVL); + Value *RHS = vectorizeTree(RHSVL); + + if (Value *V = alreadyVectorized(E->Scalars)) + return V; + + // Create a vector of LHS op1 RHS + BinaryOperator *BinOp0 = cast(VL0); + Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS); + + // Create a vector of LHS op2 RHS + Instruction *VL1 = cast(E->Scalars[1]); + BinaryOperator *BinOp1 = cast(VL1); + Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS); + + // Create shuffle to take alternate operations from the vector. + // Also, gather up odd and even scalar ops to propagate IR flags to + // each vector operation. + ValueList OddScalars, EvenScalars; + unsigned e = E->Scalars.size(); + SmallVector Mask(e); + for (unsigned i = 0; i < e; ++i) { + if (i & 1) { + Mask[i] = Builder.getInt32(e + i); + OddScalars.push_back(E->Scalars[i]); + } else { + Mask[i] = Builder.getInt32(i); + EvenScalars.push_back(E->Scalars[i]); + } + } + + Value *ShuffleMask = ConstantVector::get(Mask); + propagateIRFlags(V0, EvenScalars); + propagateIRFlags(V1, OddScalars); + + Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask); E->VectorizedValue = V; + ++NumVectorInstructions; + if (Instruction *I = dyn_cast(V)) + return propagateMetadata(I, E->Scalars); + return V; } default: @@ -1674,7 +2552,13 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { } Value *BoUpSLP::vectorizeTree() { - Builder.SetInsertPoint(F->getEntryBlock().begin()); + + // All blocks must be scheduled before any instructions are inserted. + for (auto &BSIter : BlocksSchedules) { + scheduleBlock(BSIter.second.get()); + } + + Builder.SetInsertPoint(&F->getEntryBlock().front()); vectorizeTree(&VectorizableTree[0]); DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n"); @@ -1719,7 +2603,7 @@ Value *BoUpSLP::vectorizeTree() { User->replaceUsesOfWith(Scalar, Ex); } } else { - Builder.SetInsertPoint(F->getEntryBlock().begin()); + Builder.SetInsertPoint(&F->getEntryBlock().front()); Value *Ex = Builder.CreateExtractElement(Vec, Lane); CSEBlocks.insert(&F->getEntryBlock()); User->replaceUsesOfWith(Scalar, Ex); @@ -1735,7 +2619,6 @@ Value *BoUpSLP::vectorizeTree() { // For each lane: for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { Value *Scalar = Entry->Scalars[Lane]; - // No need to handle users of gathered values. if (Entry->NeedToGather) continue; @@ -1749,8 +2632,9 @@ Value *BoUpSLP::vectorizeTree() { DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n"); assert((ScalarToTreeEntry.count(U) || - // It is legal to replace the reduction users by undef. - (RdxOps && RdxOps->count(U))) && + // It is legal to replace users in the ignorelist by undef. + (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) != + UserIgnoreList.end())) && "Replacing out-of-tree value with undef"); } #endif @@ -1758,13 +2642,10 @@ Value *BoUpSLP::vectorizeTree() { Scalar->replaceAllUsesWith(Undef); } DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n"); - cast(Scalar)->eraseFromParent(); + eraseInstruction(cast(Scalar)); } } - for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) { - BlocksNumbers[it].forget(); - } Builder.ClearInsertionPoint(); return VectorizableTree[0].VectorizedValue; @@ -1805,11 +2686,19 @@ void BoUpSLP::optimizeGatherSequence() { Insert->moveBefore(PreHeader->getTerminator()); } + // Make a list of all reachable blocks in our CSE queue. + SmallVector CSEWorkList; + CSEWorkList.reserve(CSEBlocks.size()); + for (BasicBlock *BB : CSEBlocks) + if (DomTreeNode *N = DT->getNode(BB)) { + assert(DT->isReachableFromEntry(N)); + CSEWorkList.push_back(N); + } + // Sort blocks by domination. This ensures we visit a block after all blocks // dominating it are visited. - SmallVector CSEWorkList(CSEBlocks.begin(), CSEBlocks.end()); std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), - [this](const BasicBlock *A, const BasicBlock *B) { + [this](const DomTreeNode *A, const DomTreeNode *B) { return DT->properlyDominates(A, B); }); @@ -1817,15 +2706,13 @@ void BoUpSLP::optimizeGatherSequence() { // instructions. TODO: We can further optimize this scan if we split the // instructions into different buckets based on the insert lane. SmallVector Visited; - for (SmallVectorImpl::iterator I = CSEWorkList.begin(), - E = CSEWorkList.end(); - I != E; ++I) { + for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!"); - BasicBlock *BB = *I; + BasicBlock *BB = (*I)->getBlock(); // For all instructions in blocks containing gather sequences: for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { - Instruction *In = it++; + Instruction *In = &*it++; if (!isa(In) && !isa(In)) continue; @@ -1837,7 +2724,7 @@ void BoUpSLP::optimizeGatherSequence() { if (In->isIdenticalTo(*v) && DT->dominates((*v)->getParent(), In->getParent())) { In->replaceAllUsesWith(*v); - In->eraseFromParent(); + eraseInstruction(In); In = nullptr; break; } @@ -1852,6 +2739,406 @@ void BoUpSLP::optimizeGatherSequence() { GatherSeq.clear(); } +// Groups the instructions to a bundle (which is then a single scheduling entity) +// and schedules instructions until the bundle gets ready. +bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef VL, + BoUpSLP *SLP) { + if (isa(VL[0])) + return true; + + // Initialize the instruction bundle. + Instruction *OldScheduleEnd = ScheduleEnd; + ScheduleData *PrevInBundle = nullptr; + ScheduleData *Bundle = nullptr; + bool ReSchedule = false; + DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n"); + + // Make sure that the scheduling region contains all + // instructions of the bundle. + for (Value *V : VL) { + if (!extendSchedulingRegion(V)) + return false; + } + + for (Value *V : VL) { + ScheduleData *BundleMember = getScheduleData(V); + assert(BundleMember && + "no ScheduleData for bundle member (maybe not in same basic block)"); + if (BundleMember->IsScheduled) { + // A bundle member was scheduled as single instruction before and now + // needs to be scheduled as part of the bundle. We just get rid of the + // existing schedule. + DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMember + << " was already scheduled\n"); + ReSchedule = true; + } + assert(BundleMember->isSchedulingEntity() && + "bundle member already part of other bundle"); + if (PrevInBundle) { + PrevInBundle->NextInBundle = BundleMember; + } else { + Bundle = BundleMember; + } + BundleMember->UnscheduledDepsInBundle = 0; + Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; + + // Group the instructions to a bundle. + BundleMember->FirstInBundle = Bundle; + PrevInBundle = BundleMember; + } + if (ScheduleEnd != OldScheduleEnd) { + // The scheduling region got new instructions at the lower end (or it is a + // new region for the first bundle). This makes it necessary to + // recalculate all dependencies. + // It is seldom that this needs to be done a second time after adding the + // initial bundle to the region. + for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + SD->clearDependencies(); + } + ReSchedule = true; + } + if (ReSchedule) { + resetSchedule(); + initialFillReadyList(ReadyInsts); + } + + DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " + << BB->getName() << "\n"); + + calculateDependencies(Bundle, true, SLP); + + // Now try to schedule the new bundle. As soon as the bundle is "ready" it + // means that there are no cyclic dependencies and we can schedule it. + // Note that's important that we don't "schedule" the bundle yet (see + // cancelScheduling). + while (!Bundle->isReady() && !ReadyInsts.empty()) { + + ScheduleData *pickedSD = ReadyInsts.back(); + ReadyInsts.pop_back(); + + if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { + schedule(pickedSD, ReadyInsts); + } + } + if (!Bundle->isReady()) { + cancelScheduling(VL); + return false; + } + return true; +} + +void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef VL) { + if (isa(VL[0])) + return; + + ScheduleData *Bundle = getScheduleData(VL[0]); + DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"); + assert(!Bundle->IsScheduled && + "Can't cancel bundle which is already scheduled"); + assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && + "tried to unbundle something which is not a bundle"); + + // Un-bundle: make single instructions out of the bundle. + ScheduleData *BundleMember = Bundle; + while (BundleMember) { + assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links"); + BundleMember->FirstInBundle = BundleMember; + ScheduleData *Next = BundleMember->NextInBundle; + BundleMember->NextInBundle = nullptr; + BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; + if (BundleMember->UnscheduledDepsInBundle == 0) { + ReadyInsts.insert(BundleMember); + } + BundleMember = Next; + } +} + +bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) { + if (getScheduleData(V)) + return true; + Instruction *I = dyn_cast(V); + assert(I && "bundle member must be an instruction"); + assert(!isa(I) && "phi nodes don't need to be scheduled"); + if (!ScheduleStart) { + // It's the first instruction in the new region. + initScheduleData(I, I->getNextNode(), nullptr, nullptr); + ScheduleStart = I; + ScheduleEnd = I->getNextNode(); + assert(ScheduleEnd && "tried to vectorize a TerminatorInst?"); + DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n"); + return true; + } + // Search up and down at the same time, because we don't know if the new + // instruction is above or below the existing scheduling region. + BasicBlock::reverse_iterator UpIter(ScheduleStart->getIterator()); + BasicBlock::reverse_iterator UpperEnd = BB->rend(); + BasicBlock::iterator DownIter(ScheduleEnd); + BasicBlock::iterator LowerEnd = BB->end(); + for (;;) { + if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { + DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n"); + return false; + } + + if (UpIter != UpperEnd) { + if (&*UpIter == I) { + initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); + ScheduleStart = I; + DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n"); + return true; + } + UpIter++; + } + if (DownIter != LowerEnd) { + if (&*DownIter == I) { + initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, + nullptr); + ScheduleEnd = I->getNextNode(); + assert(ScheduleEnd && "tried to vectorize a TerminatorInst?"); + DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n"); + return true; + } + DownIter++; + } + assert((UpIter != UpperEnd || DownIter != LowerEnd) && + "instruction not found in block"); + } + return true; +} + +void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, + Instruction *ToI, + ScheduleData *PrevLoadStore, + ScheduleData *NextLoadStore) { + ScheduleData *CurrentLoadStore = PrevLoadStore; + for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { + ScheduleData *SD = ScheduleDataMap[I]; + if (!SD) { + // Allocate a new ScheduleData for the instruction. + if (ChunkPos >= ChunkSize) { + ScheduleDataChunks.push_back( + llvm::make_unique(ChunkSize)); + ChunkPos = 0; + } + SD = &(ScheduleDataChunks.back()[ChunkPos++]); + ScheduleDataMap[I] = SD; + SD->Inst = I; + } + assert(!isInSchedulingRegion(SD) && + "new ScheduleData already in scheduling region"); + SD->init(SchedulingRegionID); + + if (I->mayReadOrWriteMemory()) { + // Update the linked list of memory accessing instructions. + if (CurrentLoadStore) { + CurrentLoadStore->NextLoadStore = SD; + } else { + FirstLoadStoreInRegion = SD; + } + CurrentLoadStore = SD; + } + } + if (NextLoadStore) { + if (CurrentLoadStore) + CurrentLoadStore->NextLoadStore = NextLoadStore; + } else { + LastLoadStoreInRegion = CurrentLoadStore; + } +} + +void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, + bool InsertInReadyList, + BoUpSLP *SLP) { + assert(SD->isSchedulingEntity()); + + SmallVector WorkList; + WorkList.push_back(SD); + + while (!WorkList.empty()) { + ScheduleData *SD = WorkList.back(); + WorkList.pop_back(); + + ScheduleData *BundleMember = SD; + while (BundleMember) { + assert(isInSchedulingRegion(BundleMember)); + if (!BundleMember->hasValidDependencies()) { + + DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n"); + BundleMember->Dependencies = 0; + BundleMember->resetUnscheduledDeps(); + + // Handle def-use chain dependencies. + for (User *U : BundleMember->Inst->users()) { + if (isa(U)) { + ScheduleData *UseSD = getScheduleData(U); + if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { + BundleMember->Dependencies++; + ScheduleData *DestBundle = UseSD->FirstInBundle; + if (!DestBundle->IsScheduled) { + BundleMember->incrementUnscheduledDeps(1); + } + if (!DestBundle->hasValidDependencies()) { + WorkList.push_back(DestBundle); + } + } + } else { + // I'm not sure if this can ever happen. But we need to be safe. + // This lets the instruction/bundle never be scheduled and + // eventually disable vectorization. + BundleMember->Dependencies++; + BundleMember->incrementUnscheduledDeps(1); + } + } + + // Handle the memory dependencies. + ScheduleData *DepDest = BundleMember->NextLoadStore; + if (DepDest) { + Instruction *SrcInst = BundleMember->Inst; + MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); + bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); + unsigned numAliased = 0; + unsigned DistToSrc = 1; + + while (DepDest) { + assert(isInSchedulingRegion(DepDest)); + + // We have two limits to reduce the complexity: + // 1) AliasedCheckLimit: It's a small limit to reduce calls to + // SLP->isAliased (which is the expensive part in this loop). + // 2) MaxMemDepDistance: It's for very large blocks and it aborts + // the whole loop (even if the loop is fast, it's quadratic). + // It's important for the loop break condition (see below) to + // check this limit even between two read-only instructions. + if (DistToSrc >= MaxMemDepDistance || + ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && + (numAliased >= AliasedCheckLimit || + SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { + + // We increment the counter only if the locations are aliased + // (instead of counting all alias checks). This gives a better + // balance between reduced runtime and accurate dependencies. + numAliased++; + + DepDest->MemoryDependencies.push_back(BundleMember); + BundleMember->Dependencies++; + ScheduleData *DestBundle = DepDest->FirstInBundle; + if (!DestBundle->IsScheduled) { + BundleMember->incrementUnscheduledDeps(1); + } + if (!DestBundle->hasValidDependencies()) { + WorkList.push_back(DestBundle); + } + } + DepDest = DepDest->NextLoadStore; + + // Example, explaining the loop break condition: Let's assume our + // starting instruction is i0 and MaxMemDepDistance = 3. + // + // +--------v--v--v + // i0,i1,i2,i3,i4,i5,i6,i7,i8 + // +--------^--^--^ + // + // MaxMemDepDistance let us stop alias-checking at i3 and we add + // dependencies from i0 to i3,i4,.. (even if they are not aliased). + // Previously we already added dependencies from i3 to i6,i7,i8 + // (because of MaxMemDepDistance). As we added a dependency from + // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 + // and we can abort this loop at i6. + if (DistToSrc >= 2 * MaxMemDepDistance) + break; + DistToSrc++; + } + } + } + BundleMember = BundleMember->NextInBundle; + } + if (InsertInReadyList && SD->isReady()) { + ReadyInsts.push_back(SD); + DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"); + } + } +} + +void BoUpSLP::BlockScheduling::resetSchedule() { + assert(ScheduleStart && + "tried to reset schedule on block which has not been scheduled"); + for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + assert(isInSchedulingRegion(SD)); + SD->IsScheduled = false; + SD->resetUnscheduledDeps(); + } + ReadyInsts.clear(); +} + +void BoUpSLP::scheduleBlock(BlockScheduling *BS) { + + if (!BS->ScheduleStart) + return; + + DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n"); + + BS->resetSchedule(); + + // For the real scheduling we use a more sophisticated ready-list: it is + // sorted by the original instruction location. This lets the final schedule + // be as close as possible to the original instruction order. + struct ScheduleDataCompare { + bool operator()(ScheduleData *SD1, ScheduleData *SD2) { + return SD2->SchedulingPriority < SD1->SchedulingPriority; + } + }; + std::set ReadyInsts; + + // Ensure that all dependency data is updated and fill the ready-list with + // initial instructions. + int Idx = 0; + int NumToSchedule = 0; + for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; + I = I->getNextNode()) { + ScheduleData *SD = BS->getScheduleData(I); + assert( + SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && + "scheduler and vectorizer have different opinion on what is a bundle"); + SD->FirstInBundle->SchedulingPriority = Idx++; + if (SD->isSchedulingEntity()) { + BS->calculateDependencies(SD, false, this); + NumToSchedule++; + } + } + BS->initialFillReadyList(ReadyInsts); + + Instruction *LastScheduledInst = BS->ScheduleEnd; + + // Do the "real" scheduling. + while (!ReadyInsts.empty()) { + ScheduleData *picked = *ReadyInsts.begin(); + ReadyInsts.erase(ReadyInsts.begin()); + + // Move the scheduled instruction(s) to their dedicated places, if not + // there yet. + ScheduleData *BundleMember = picked; + while (BundleMember) { + Instruction *pickedInst = BundleMember->Inst; + if (LastScheduledInst->getNextNode() != pickedInst) { + BS->BB->getInstList().remove(pickedInst); + BS->BB->getInstList().insert(LastScheduledInst->getIterator(), + pickedInst); + } + LastScheduledInst = pickedInst; + BundleMember = BundleMember->NextInBundle; + } + + BS->schedule(picked, ReadyInsts); + NumToSchedule--; + } + assert(NumToSchedule == 0 && "could not schedule all instructions"); + + // Avoid duplicate scheduling of the block. + BS->ScheduleStart = nullptr; +} + /// The SLPVectorizer Pass. struct SLPVectorizer : public FunctionPass { typedef SmallVector StoreList; @@ -1865,23 +3152,25 @@ struct SLPVectorizer : public FunctionPass { } ScalarEvolution *SE; - const DataLayout *DL; TargetTransformInfo *TTI; + TargetLibraryInfo *TLI; AliasAnalysis *AA; LoopInfo *LI; DominatorTree *DT; + AssumptionCache *AC; bool runOnFunction(Function &F) override { if (skipOptnoneFunction(F)) return false; - SE = &getAnalysis(); - DataLayoutPass *DLP = getAnalysisIfAvailable(); - DL = DLP ? &DLP->getDataLayout() : nullptr; - TTI = &getAnalysis(); - AA = &getAnalysis(); - LI = &getAnalysis(); + SE = &getAnalysis().getSE(); + TTI = &getAnalysis().getTTI(F); + auto *TLIP = getAnalysisIfAvailable(); + TLI = TLIP ? &TLIP->getTLI() : nullptr; + AA = &getAnalysis().getAAResults(); + LI = &getAnalysis().getLoopInfo(); DT = &getAnalysis().getDomTree(); + AC = &getAnalysis().getAssumptionCache(F); StoreRefs.clear(); bool Changed = false; @@ -1891,10 +3180,16 @@ struct SLPVectorizer : public FunctionPass { if (!TTI->getNumberOfRegisters(true)) return false; - // Must have DataLayout. We can't require it because some tests run w/o - // triple. - if (!DL) - return false; + // Use the vector register size specified by the target unless overridden + // by a command-line option. + // TODO: It would be better to limit the vectorization factor based on + // data type rather than just register size. For example, x86 AVX has + // 256-bit registers, but it does not support integer operations + // at that width (that requires AVX2). + if (MaxVectorRegSizeOption.getNumOccurrences()) + MaxVecRegSize = MaxVectorRegSizeOption; + else + MaxVecRegSize = TTI->getRegisterBitWidth(true); // Don't vectorize when the attribute NoImplicitFloat is used. if (F.hasFnAttribute(Attribute::NoImplicitFloat)) @@ -1903,14 +3198,14 @@ struct SLPVectorizer : public FunctionPass { DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"); // Use the bottom up slp vectorizer to construct chains that start with - // he store instructions. - BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT); + // store instructions. + BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC); - // Scan the blocks in the function in post order. - for (po_iterator it = po_begin(&F.getEntryBlock()), - e = po_end(&F.getEntryBlock()); it != e; ++it) { - BasicBlock *BB = *it; + // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to + // delete instructions. + // Scan the blocks in the function in post order. + for (auto BB : post_order(&F.getEntryBlock())) { // Vectorize trees that end at stores. if (unsigned count = collectStores(BB, R)) { (void)count; @@ -1932,13 +3227,16 @@ struct SLPVectorizer : public FunctionPass { void getAnalysisUsage(AnalysisUsage &AU) const override { FunctionPass::getAnalysisUsage(AU); - AU.addRequired(); - AU.addRequired(); - AU.addRequired(); - AU.addRequired(); + AU.addRequired(); + AU.addRequired(); + AU.addRequired(); + AU.addRequired(); + AU.addRequired(); AU.addRequired(); - AU.addPreserved(); + AU.addPreserved(); AU.addPreserved(); + AU.addPreserved(); + AU.addPreserved(); AU.setPreservesCFG(); } @@ -1954,8 +3252,12 @@ private: bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R); /// \brief Try to vectorize a list of operands. + /// \@param BuildVector A list of users to ignore for the purpose of + /// scheduling and that don't need extracting. /// \returns true if a value was vectorized. - bool tryToVectorizeList(ArrayRef VL, BoUpSLP &R); + bool tryToVectorizeList(ArrayRef VL, BoUpSLP &R, + ArrayRef BuildVector = None, + bool allowReorder = false); /// \brief Try to vectorize a chain that may start at the operands of \V; bool tryToVectorize(BinaryOperator *V, BoUpSLP &R); @@ -1968,37 +3270,36 @@ private: bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R); bool vectorizeStoreChain(ArrayRef Chain, int CostThreshold, - BoUpSLP &R); + BoUpSLP &R, unsigned VecRegSize); bool vectorizeStores(ArrayRef Stores, int costThreshold, BoUpSLP &R); private: StoreListMap StoreRefs; + unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. }; /// \brief Check that the Values in the slice in VL array are still existent in /// the WeakVH array. /// Vectorization of part of the VL array may cause later values in the VL array /// to become invalid. We track when this has happened in the WeakVH array. -static bool hasValueBeenRAUWed(ArrayRef &VL, - SmallVectorImpl &VH, - unsigned SliceBegin, - unsigned SliceSize) { - for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i) - if (VH[i] != VL[i]) - return true; - - return false; +static bool hasValueBeenRAUWed(ArrayRef VL, ArrayRef VH, + unsigned SliceBegin, unsigned SliceSize) { + VL = VL.slice(SliceBegin, SliceSize); + VH = VH.slice(SliceBegin, SliceSize); + return !std::equal(VL.begin(), VL.end(), VH.begin()); } bool SLPVectorizer::vectorizeStoreChain(ArrayRef Chain, - int CostThreshold, BoUpSLP &R) { + int CostThreshold, BoUpSLP &R, + unsigned VecRegSize) { unsigned ChainLen = Chain.size(); DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"); Type *StoreTy = cast(Chain[0])->getValueOperand()->getType(); - unsigned Sz = DL->getTypeSizeInBits(StoreTy); - unsigned VF = MinVecRegSize / Sz; + auto &DL = cast(Chain[0])->getModule()->getDataLayout(); + unsigned Sz = DL.getTypeSizeInBits(StoreTy); + unsigned VF = VecRegSize / Sz; if (!isPowerOf2_32(Sz) || VF < 2) return false; @@ -2040,8 +3341,8 @@ bool SLPVectorizer::vectorizeStoreChain(ArrayRef Chain, bool SLPVectorizer::vectorizeStores(ArrayRef Stores, int costThreshold, BoUpSLP &R) { - SetVector Heads, Tails; - SmallDenseMap ConsecutiveChain; + SetVector Heads, Tails; + SmallDenseMap ConsecutiveChain; // We may run into multiple chains that merge into a single chain. We mark the // stores that we vectorized so that we don't visit the same store twice. @@ -2050,21 +3351,32 @@ bool SLPVectorizer::vectorizeStores(ArrayRef Stores, // Do a quadratic search on all of the given stores and find // all of the pairs of stores that follow each other. + SmallVector IndexQueue; for (unsigned i = 0, e = Stores.size(); i < e; ++i) { - for (unsigned j = 0; j < e; ++j) { - if (i == j) - continue; - - if (R.isConsecutiveAccess(Stores[i], Stores[j])) { - Tails.insert(Stores[j]); + const DataLayout &DL = Stores[i]->getModule()->getDataLayout(); + IndexQueue.clear(); + // If a store has multiple consecutive store candidates, search Stores + // array according to the sequence: from i+1 to e, then from i-1 to 0. + // This is because usually pairing with immediate succeeding or preceding + // candidate create the best chance to find slp vectorization opportunity. + unsigned j = 0; + for (j = i + 1; j < e; ++j) + IndexQueue.push_back(j); + for (j = i; j > 0; --j) + IndexQueue.push_back(j - 1); + + for (auto &k : IndexQueue) { + if (R.isConsecutiveAccess(Stores[i], Stores[k], DL)) { + Tails.insert(Stores[k]); Heads.insert(Stores[i]); - ConsecutiveChain[Stores[i]] = Stores[j]; + ConsecutiveChain[Stores[i]] = Stores[k]; + break; } } } // For stores that start but don't end a link in the chain: - for (SetVector::iterator it = Heads.begin(), e = Heads.end(); + for (SetVector::iterator it = Heads.begin(), e = Heads.end(); it != e; ++it) { if (Tails.count(*it)) continue; @@ -2072,7 +3384,7 @@ bool SLPVectorizer::vectorizeStores(ArrayRef Stores, // We found a store instr that starts a chain. Now follow the chain and try // to vectorize it. BoUpSLP::ValueList Operands; - Value *I = *it; + StoreInst *I = *it; // Collect the chain into a list. while (Tails.count(I) || Heads.count(I)) { if (VectorizedStores.count(I)) @@ -2082,12 +3394,16 @@ bool SLPVectorizer::vectorizeStores(ArrayRef Stores, I = ConsecutiveChain[I]; } - bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R); - - // Mark the vectorized stores so that we don't vectorize them again. - if (Vectorized) - VectorizedStores.insert(Operands.begin(), Operands.end()); - Changed |= Vectorized; + // FIXME: Is division-by-2 the correct step? Should we assert that the + // register size is a power-of-2? + for (unsigned Size = MaxVecRegSize; Size >= MinVecRegSize; Size /= 2) { + if (vectorizeStoreChain(Operands, costThreshold, R, Size)) { + // Mark the vectorized stores so that we don't vectorize them again. + VectorizedStores.insert(Operands.begin(), Operands.end()); + Changed = true; + break; + } + } } return Changed; @@ -2097,8 +3413,9 @@ bool SLPVectorizer::vectorizeStores(ArrayRef Stores, unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) { unsigned count = 0; StoreRefs.clear(); - for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { - StoreInst *SI = dyn_cast(it); + const DataLayout &DL = BB->getModule()->getDataLayout(); + for (Instruction &I : *BB) { + StoreInst *SI = dyn_cast(&I); if (!SI) continue; @@ -2108,7 +3425,7 @@ unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) { // Check that the pointer points to scalars. Type *Ty = SI->getValueOperand()->getType(); - if (Ty->isAggregateType() || Ty->isVectorTy()) + if (!isValidElementType(Ty)) continue; // Find the base pointer. @@ -2125,10 +3442,12 @@ bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { if (!A || !B) return false; Value *VL[] = { A, B }; - return tryToVectorizeList(VL, R); + return tryToVectorizeList(VL, R, None, true); } -bool SLPVectorizer::tryToVectorizeList(ArrayRef VL, BoUpSLP &R) { +bool SLPVectorizer::tryToVectorizeList(ArrayRef VL, BoUpSLP &R, + ArrayRef BuildVector, + bool allowReorder) { if (VL.size() < 2) return false; @@ -2140,16 +3459,19 @@ bool SLPVectorizer::tryToVectorizeList(ArrayRef VL, BoUpSLP &R) { return false; unsigned Opcode0 = I0->getOpcode(); + const DataLayout &DL = I0->getModule()->getDataLayout(); Type *Ty0 = I0->getType(); - unsigned Sz = DL->getTypeSizeInBits(Ty0); + unsigned Sz = DL.getTypeSizeInBits(Ty0); + // FIXME: Register size should be a parameter to this function, so we can + // try different vectorization factors. unsigned VF = MinVecRegSize / Sz; - for (int i = 0, e = VL.size(); i < e; ++i) { - Type *Ty = VL[i]->getType(); - if (Ty->isAggregateType() || Ty->isVectorTy()) + for (Value *V : VL) { + Type *Ty = V->getType(); + if (!isValidElementType(Ty)) return false; - Instruction *Inst = dyn_cast(VL[i]); + Instruction *Inst = dyn_cast(V); if (!Inst || Inst->getOpcode() != Opcode0) return false; } @@ -2178,13 +3500,46 @@ bool SLPVectorizer::tryToVectorizeList(ArrayRef VL, BoUpSLP &R) { << "\n"); ArrayRef Ops = VL.slice(i, OpsWidth); - R.buildTree(Ops); + ArrayRef BuildVectorSlice; + if (!BuildVector.empty()) + BuildVectorSlice = BuildVector.slice(i, OpsWidth); + + R.buildTree(Ops, BuildVectorSlice); + // TODO: check if we can allow reordering also for other cases than + // tryToVectorizePair() + if (allowReorder && R.shouldReorder()) { + assert(Ops.size() == 2); + assert(BuildVectorSlice.empty()); + Value *ReorderedOps[] = { Ops[1], Ops[0] }; + R.buildTree(ReorderedOps, None); + } int Cost = R.getTreeCost(); if (Cost < -SLPCostThreshold) { DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n"); - R.vectorizeTree(); - + Value *VectorizedRoot = R.vectorizeTree(); + + // Reconstruct the build vector by extracting the vectorized root. This + // way we handle the case where some elements of the vector are undefined. + // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2)) + if (!BuildVectorSlice.empty()) { + // The insert point is the last build vector instruction. The vectorized + // root will precede it. This guarantees that we get an instruction. The + // vectorized tree could have been constant folded. + Instruction *InsertAfter = cast(BuildVectorSlice.back()); + unsigned VecIdx = 0; + for (auto &V : BuildVectorSlice) { + IRBuilder Builder( + InsertAfter->getParent(), ++BasicBlock::iterator(InsertAfter)); + InsertElementInst *IE = cast(V); + Instruction *Extract = cast(Builder.CreateExtractElement( + VectorizedRoot, Builder.getInt32(VecIdx++))); + IE->setOperand(1, Extract); + IE->removeFromParent(); + IE->insertAfter(Extract); + InsertAfter = IE; + } + } // Move to the next bundle. i += VF - 1; Changed = true; @@ -2209,11 +3564,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { BinaryOperator *B0 = dyn_cast(B->getOperand(0)); BinaryOperator *B1 = dyn_cast(B->getOperand(1)); if (tryToVectorizePair(A, B0, R)) { - B->moveBefore(V); return true; } if (tryToVectorizePair(A, B1, R)) { - B->moveBefore(V); return true; } } @@ -2223,11 +3576,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { BinaryOperator *A0 = dyn_cast(A->getOperand(0)); BinaryOperator *A1 = dyn_cast(A->getOperand(1)); if (tryToVectorizePair(A0, B, R)) { - A->moveBefore(V); return true; } if (tryToVectorizePair(A1, B, R)) { - A->moveBefore(V); return true; } } @@ -2240,7 +3591,7 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { /// \param NumEltsToRdx The number of elements that should be reduced in the /// vector. /// \param IsPairwise Whether the reduction is a pairwise or splitting -/// reduction. A pairwise reduction will generate a mask of +/// reduction. A pairwise reduction will generate a mask of /// <0,2,...> or <1,3,..> while a splitting reduction will generate /// <2,3, undef,undef> for a vector of 4 and NumElts = 2. /// \param IsLeft True will generate a mask of even elements, odd otherwise. @@ -2293,7 +3644,7 @@ static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx, /// *p = /// class HorizontalReduction { - SmallPtrSet ReductionOps; + SmallVector ReductionOps; SmallVector ReducedVals; BinaryOperator *ReductionRoot; @@ -2303,20 +3654,20 @@ class HorizontalReduction { unsigned ReductionOpcode; /// The opcode of the values we perform a reduction on. unsigned ReducedValueOpcode; - /// The width of one full horizontal reduction operation. - unsigned ReduxWidth; /// Should we model this reduction as a pairwise reduction tree or a tree that /// splits the vector in halves and adds those halves. bool IsPairwiseReduction; public: + /// The width of one full horizontal reduction operation. + unsigned ReduxWidth; + HorizontalReduction() : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0), - ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {} + ReducedValueOpcode(0), IsPairwiseReduction(false), ReduxWidth(0) {} /// \brief Try to find a reduction tree. - bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B, - const DataLayout *DL) { + bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B) { assert((!Phi || std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) && "Thi phi needs to use the binary operator"); @@ -2338,12 +3689,15 @@ public: return false; Type *Ty = B->getType(); - if (Ty->isVectorTy()) + if (!isValidElementType(Ty)) return false; + const DataLayout &DL = B->getModule()->getDataLayout(); ReductionOpcode = B->getOpcode(); ReducedValueOpcode = 0; - ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty); + // FIXME: Register size should be a parameter to this function, so we can + // try different vectorization factors. + ReduxWidth = MinVecRegSize / DL.getTypeSizeInBits(Ty); ReductionRoot = B; ReductionPHI = Phi; @@ -2356,11 +3710,11 @@ public: return false; // Post order traverse the reduction tree starting at B. We only handle true - // trees containing only binary operators. - SmallVector, 32> Stack; + // trees containing only binary operators or selects. + SmallVector, 32> Stack; Stack.push_back(std::make_pair(B, 0)); while (!Stack.empty()) { - BinaryOperator *TreeN = Stack.back().first; + Instruction *TreeN = Stack.back().first; unsigned EdgeToVist = Stack.back().second++; bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode; @@ -2387,7 +3741,7 @@ public: // We need to be able to reassociate the adds. if (!TreeN->isAssociative()) return false; - ReductionOps.insert(TreeN); + ReductionOps.push_back(TreeN); } // Retract. Stack.pop_back(); @@ -2396,9 +3750,10 @@ public: // Visit left or right. Value *NextV = TreeN->getOperand(EdgeToVist); - BinaryOperator *Next = dyn_cast(NextV); - if (Next) - Stack.push_back(std::make_pair(Next, 0)); + // We currently only allow BinaryOperator's and SelectInst's as reduction + // values in our tree. + if (isa(NextV) || isa(NextV)) + Stack.push_back(std::make_pair(cast(NextV), 0)); else if (NextV != Phi) return false; } @@ -2419,12 +3774,11 @@ public: IRBuilder<> Builder(ReductionRoot); FastMathFlags Unsafe; Unsafe.setUnsafeAlgebra(); - Builder.SetFastMathFlags(Unsafe); + Builder.setFastMathFlags(Unsafe); unsigned i = 0; for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) { - ArrayRef ValsToReduce(&ReducedVals[i], ReduxWidth); - V.buildTree(ValsToReduce, &ReductionOps); + V.buildTree(makeArrayRef(&ReducedVals[i], ReduxWidth), ReductionOps); // Estimate cost. int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]); @@ -2467,9 +3821,12 @@ public: return VectorizedTree != nullptr; } -private: + unsigned numReductionValues() const { + return ReducedVals.size(); + } - /// \brief Calcuate the cost of a reduction. +private: + /// \brief Calculate the cost of a reduction. int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) { Type *ScalarTy = FirstReducedVal->getType(); Type *VecTy = VectorType::get(ScalarTy, ReduxWidth); @@ -2502,11 +3859,10 @@ private: /// \brief Emit a horizontal reduction of the vectorized value. Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) { assert(VectorizedValue && "Need to have a vectorized tree node"); - Instruction *ValToReduce = dyn_cast(VectorizedValue); assert(isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now"); - Value *TmpVec = ValToReduce; + Value *TmpVec = VectorizedValue; for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { if (IsPairwiseReduction) { Value *LeftMask = @@ -2543,13 +3899,16 @@ private: /// /// Returns true if it matches /// -static bool findBuildVector(InsertElementInst *IE, - SmallVectorImpl &Ops) { - if (!isa(IE->getOperand(0))) +static bool findBuildVector(InsertElementInst *FirstInsertElem, + SmallVectorImpl &BuildVector, + SmallVectorImpl &BuildVectorOpds) { + if (!isa(FirstInsertElem->getOperand(0))) return false; + InsertElementInst *IE = FirstInsertElem; while (true) { - Ops.push_back(IE->getOperand(1)); + BuildVector.push_back(IE); + BuildVectorOpds.push_back(IE->getOperand(1)); if (IE->use_empty()) return false; @@ -2573,6 +3932,82 @@ static bool PhiTypeSorterFunc(Value *V, Value *V2) { return V->getType() < V2->getType(); } +/// \brief Try and get a reduction value from a phi node. +/// +/// Given a phi node \p P in a block \p ParentBB, consider possible reductions +/// if they come from either \p ParentBB or a containing loop latch. +/// +/// \returns A candidate reduction value if possible, or \code nullptr \endcode +/// if not possible. +static Value *getReductionValue(const DominatorTree *DT, PHINode *P, + BasicBlock *ParentBB, LoopInfo *LI) { + // There are situations where the reduction value is not dominated by the + // reduction phi. Vectorizing such cases has been reported to cause + // miscompiles. See PR25787. + auto DominatedReduxValue = [&](Value *R) { + return ( + dyn_cast(R) && + DT->dominates(P->getParent(), dyn_cast(R)->getParent())); + }; + + Value *Rdx = nullptr; + + // Return the incoming value if it comes from the same BB as the phi node. + if (P->getIncomingBlock(0) == ParentBB) { + Rdx = P->getIncomingValue(0); + } else if (P->getIncomingBlock(1) == ParentBB) { + Rdx = P->getIncomingValue(1); + } + + if (Rdx && DominatedReduxValue(Rdx)) + return Rdx; + + // Otherwise, check whether we have a loop latch to look at. + Loop *BBL = LI->getLoopFor(ParentBB); + if (!BBL) + return nullptr; + BasicBlock *BBLatch = BBL->getLoopLatch(); + if (!BBLatch) + return nullptr; + + // There is a loop latch, return the incoming value if it comes from + // that. This reduction pattern occassionaly turns up. + if (P->getIncomingBlock(0) == BBLatch) { + Rdx = P->getIncomingValue(0); + } else if (P->getIncomingBlock(1) == BBLatch) { + Rdx = P->getIncomingValue(1); + } + + if (Rdx && DominatedReduxValue(Rdx)) + return Rdx; + + return nullptr; +} + +/// \brief Attempt to reduce a horizontal reduction. +/// If it is legal to match a horizontal reduction feeding +/// the phi node P with reduction operators BI, then check if it +/// can be done. +/// \returns true if a horizontal reduction was matched and reduced. +/// \returns false if a horizontal reduction was not matched. +static bool canMatchHorizontalReduction(PHINode *P, BinaryOperator *BI, + BoUpSLP &R, TargetTransformInfo *TTI) { + if (!ShouldVectorizeHor) + return false; + + HorizontalReduction HorRdx; + if (!HorRdx.matchAssociativeReduction(P, BI)) + return false; + + // If there is a sufficient number of reduction values, reduce + // to a nearby power-of-2. Can safely generate oversized + // vectors and rely on the backend to split them to legal sizes. + HorRdx.ReduxWidth = + std::max((uint64_t)4, PowerOf2Floor(HorRdx.numReductionValues())); + + return HorRdx.tryToReduce(R, TTI); +} + bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { bool Changed = false; SmallVector Incoming; @@ -2584,9 +4019,8 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { // Collect the incoming values from the PHIs. Incoming.clear(); - for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie; - ++instr) { - PHINode *P = dyn_cast(instr); + for (Instruction &I : *BB) { + PHINode *P = dyn_cast(&I); if (!P) break; @@ -2613,8 +4047,7 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { // Try to vectorize them. unsigned NumElts = (SameTypeIt - IncIt); DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"); - if (NumElts > 1 && - tryToVectorizeList(ArrayRef(IncIt, NumElts), R)) { + if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R)) { // Success start over because instructions might have been changed. HaveVectorizedPhiNodes = true; Changed = true; @@ -2630,7 +4063,7 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { // We may go through BB multiple times so skip the one we have checked. - if (!VisitedInstrs.insert(it)) + if (!VisitedInstrs.insert(&*it).second) continue; if (isa(it)) @@ -2641,21 +4074,16 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { // Check that the PHI is a reduction PHI. if (P->getNumIncomingValues() != 2) return Changed; - Value *Rdx = - (P->getIncomingBlock(0) == BB - ? (P->getIncomingValue(0)) - : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) - : nullptr)); + + Value *Rdx = getReductionValue(DT, P, BB, LI); + // Check if this is a Binary Operator. BinaryOperator *BI = dyn_cast_or_null(Rdx); if (!BI) continue; // Try to match and vectorize a horizontal reduction. - HorizontalReduction HorRdx; - if (ShouldVectorizeHor && - HorRdx.matchAssociativeReduction(P, BI, DL) && - HorRdx.tryToReduce(R, TTI)) { + if (canMatchHorizontalReduction(P, BI, R, TTI)) { Changed = true; it = BB->begin(); e = BB->end(); @@ -2678,15 +4106,27 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { continue; } - // Try to vectorize horizontal reductions feeding into a store. if (ShouldStartVectorizeHorAtStore) if (StoreInst *SI = dyn_cast(it)) if (BinaryOperator *BinOp = dyn_cast(SI->getValueOperand())) { - HorizontalReduction HorRdx; - if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) && - HorRdx.tryToReduce(R, TTI)) || - tryToVectorize(BinOp, R))) { + if (canMatchHorizontalReduction(nullptr, BinOp, R, TTI) || + tryToVectorize(BinOp, R)) { + Changed = true; + it = BB->begin(); + e = BB->end(); + continue; + } + } + + // Try to vectorize horizontal reductions feeding into a return. + if (ReturnInst *RI = dyn_cast(it)) + if (RI->getNumOperands() != 0) + if (BinaryOperator *BinOp = + dyn_cast(RI->getOperand(0))) { + DEBUG(dbgs() << "SLP: Found a return to vectorize.\n"); + if (tryToVectorizePair(BinOp->getOperand(0), + BinOp->getOperand(1), R)) { Changed = true; it = BB->begin(); e = BB->end(); @@ -2706,26 +4146,31 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { } for (int i = 0; i < 2; ++i) { - if (BinaryOperator *BI = dyn_cast(CI->getOperand(i))) { - if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) { - Changed = true; - // We would like to start over since some instructions are deleted - // and the iterator may become invalid value. - it = BB->begin(); - e = BB->end(); - } - } + if (BinaryOperator *BI = dyn_cast(CI->getOperand(i))) { + if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) { + Changed = true; + // We would like to start over since some instructions are deleted + // and the iterator may become invalid value. + it = BB->begin(); + e = BB->end(); + break; + } + } } continue; } // Try to vectorize trees that start at insertelement instructions. - if (InsertElementInst *IE = dyn_cast(it)) { - SmallVector Ops; - if (!findBuildVector(IE, Ops)) + if (InsertElementInst *FirstInsertElem = dyn_cast(it)) { + SmallVector BuildVector; + SmallVector BuildVectorOpds; + if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds)) continue; - if (tryToVectorizeList(Ops, R)) { + // Vectorize starting with the build vector operands ignoring the + // BuildVector instructions for the purpose of scheduling and user + // extraction. + if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) { Changed = true; it = BB->begin(); e = BB->end(); @@ -2750,10 +4195,13 @@ bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) { << it->second.size() << ".\n"); // Process the stores in chunks of 16. + // TODO: The limit of 16 inhibits greater vectorization factors. + // For example, AVX2 supports v32i8. Increasing this limit, however, + // may cause a significant compile-time increase. for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { unsigned Len = std::min(CE - CI, 16); - ArrayRef Chunk(&it->second[CI], Len); - Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R); + Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), + -SLPCostThreshold, R); } } return Changed; @@ -2764,9 +4212,10 @@ bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) { char SLPVectorizer::ID = 0; static const char lv_name[] = "SLP Vectorizer"; INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false) -INITIALIZE_AG_DEPENDENCY(AliasAnalysis) -INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) -INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) +INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) +INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) +INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)