//===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This pass implements the Bottom Up SLP vectorizer. It detects consecutive // stores that can be put together into vector-stores. Next, it attempts to // construct vectorizable tree using the use-def chains. If a profitable tree // was found, the SLP vectorizer performs vectorization on the tree. // // The pass is inspired by the work described in the paper: // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. // //===----------------------------------------------------------------------===// #include "llvm/Transforms/Vectorize/SLPVectorizer.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DenseSet.h" #include "llvm/ADT/PostOrderIterator.h" #include "llvm/ADT/PriorityQueue.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SetOperations.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallBitVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallString.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/iterator.h" #include "llvm/ADT/iterator_range.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/DemandedBits.h" #include "llvm/Analysis/GlobalsModRef.h" #include "llvm/Analysis/IVDescriptors.h" #include "llvm/Analysis/LoopAccessAnalysis.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/MemoryLocation.h" #include "llvm/Analysis/OptimizationRemarkEmitter.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/TargetLibraryInfo.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/Analysis/VectorUtils.h" #include "llvm/IR/Attributes.h" #include "llvm/IR/BasicBlock.h" #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DerivedTypes.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/Function.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/InstrTypes.h" #include "llvm/IR/Instruction.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Intrinsics.h" #include "llvm/IR/Module.h" #include "llvm/IR/Operator.h" #include "llvm/IR/PatternMatch.h" #include "llvm/IR/Type.h" #include "llvm/IR/Use.h" #include "llvm/IR/User.h" #include "llvm/IR/Value.h" #include "llvm/IR/ValueHandle.h" #ifdef EXPENSIVE_CHECKS #include "llvm/IR/Verifier.h" #endif #include "llvm/Pass.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/DOTGraphTraits.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/GraphWriter.h" #include "llvm/Support/InstructionCost.h" #include "llvm/Support/KnownBits.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/InjectTLIMappings.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/LoopUtils.h" #include "llvm/Transforms/Vectorize.h" #include #include #include #include #include #include #include #include #include #include #include using namespace llvm; using namespace llvm::PatternMatch; using namespace slpvectorizer; #define SV_NAME "slp-vectorizer" #define DEBUG_TYPE "SLP" STATISTIC(NumVectorInstructions, "Number of vector instructions generated"); cl::opt RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, cl::desc("Run the SLP vectorization passes")); 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(true), cl::Hidden, cl::desc("Attempt to vectorize horizontal reductions")); static cl::opt ShouldStartVectorizeHorAtStore( "slp-vectorize-hor-store", cl::init(false), cl::Hidden, 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")); static cl::opt MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden, cl::desc("Maximum SLP vectorization factor (0=unlimited)")); static cl::opt MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, cl::desc("Maximum depth of the lookup for consecutive stores.")); /// 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")); static cl::opt MinVectorRegSizeOption( "slp-min-reg-size", cl::init(128), cl::Hidden, cl::desc("Attempt to vectorize for this register size in bits")); static cl::opt RecursionMaxDepth( "slp-recursion-max-depth", cl::init(12), cl::Hidden, cl::desc("Limit the recursion depth when building a vectorizable tree")); static cl::opt MinTreeSize( "slp-min-tree-size", cl::init(3), cl::Hidden, cl::desc("Only vectorize small trees if they are fully vectorizable")); // The maximum depth that the look-ahead score heuristic will explore. // The higher this value, the higher the compilation time overhead. static cl::opt LookAheadMaxDepth( "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, cl::desc("The maximum look-ahead depth for operand reordering scores")); // The maximum depth that the look-ahead score heuristic will explore // when it probing among candidates for vectorization tree roots. // The higher this value, the higher the compilation time overhead but unlike // similar limit for operands ordering this is less frequently used, hence // impact of higher value is less noticeable. static cl::opt RootLookAheadMaxDepth( "slp-max-root-look-ahead-depth", cl::init(2), cl::Hidden, cl::desc("The maximum look-ahead depth for searching best rooting option")); static cl::opt ViewSLPTree("view-slp-tree", cl::Hidden, cl::desc("Display the SLP trees with Graphviz")); // 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; // 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; /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling /// regions to be handled. static const int MinScheduleRegionSize = 16; /// 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 True if the value is a constant (but not globals/constant /// expressions). static bool isConstant(Value *V) { return isa(V) && !isa(V); } /// Checks if \p V is one of vector-like instructions, i.e. undef, /// insertelement/extractelement with constant indices for fixed vector type or /// extractvalue instruction. static bool isVectorLikeInstWithConstOps(Value *V) { if (!isa(V) && !isa(V)) return false; auto *I = dyn_cast(V); if (!I || isa(I)) return true; if (!isa(I->getOperand(0)->getType())) return false; if (isa(I)) return isConstant(I->getOperand(1)); assert(isa(V) && "Expected only insertelement."); return isConstant(I->getOperand(2)); } /// \returns true if all of the instructions in \p VL are in the same block or /// false otherwise. static bool allSameBlock(ArrayRef VL) { Instruction *I0 = dyn_cast(VL[0]); if (!I0) return false; if (all_of(VL, isVectorLikeInstWithConstOps)) return true; BasicBlock *BB = I0->getParent(); for (int I = 1, E = VL.size(); I < E; I++) { auto *II = dyn_cast(VL[I]); if (!II) return false; if (BB != II->getParent()) return false; } return true; } /// \returns True if all of the values in \p VL are constants (but not /// globals/constant expressions). static bool allConstant(ArrayRef VL) { // Constant expressions and globals can't be vectorized like normal integer/FP // constants. return all_of(VL, isConstant); } /// \returns True if all of the values in \p VL are identical or some of them /// are UndefValue. static bool isSplat(ArrayRef VL) { Value *FirstNonUndef = nullptr; for (Value *V : VL) { if (isa(V)) continue; if (!FirstNonUndef) { FirstNonUndef = V; continue; } if (V != FirstNonUndef) return false; } return FirstNonUndef != nullptr; } /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. static bool isCommutative(Instruction *I) { if (auto *Cmp = dyn_cast(I)) return Cmp->isCommutative(); if (auto *BO = dyn_cast(I)) return BO->isCommutative(); // TODO: This should check for generic Instruction::isCommutative(), but // we need to confirm that the caller code correctly handles Intrinsics // for example (does not have 2 operands). return false; } /// \returns inserting index of InsertElement or InsertValue instruction, /// using Offset as base offset for index. static std::optional getInsertIndex(const Value *InsertInst, unsigned Offset = 0) { int Index = Offset; if (const auto *IE = dyn_cast(InsertInst)) { const auto *VT = dyn_cast(IE->getType()); if (!VT) return std::nullopt; const auto *CI = dyn_cast(IE->getOperand(2)); if (!CI) return std::nullopt; if (CI->getValue().uge(VT->getNumElements())) return std::nullopt; Index *= VT->getNumElements(); Index += CI->getZExtValue(); return Index; } const auto *IV = cast(InsertInst); Type *CurrentType = IV->getType(); for (unsigned I : IV->indices()) { if (const auto *ST = dyn_cast(CurrentType)) { Index *= ST->getNumElements(); CurrentType = ST->getElementType(I); } else if (const auto *AT = dyn_cast(CurrentType)) { Index *= AT->getNumElements(); CurrentType = AT->getElementType(); } else { return std::nullopt; } Index += I; } return Index; } namespace { /// Specifies the way the mask should be analyzed for undefs/poisonous elements /// in the shuffle mask. enum class UseMask { FirstArg, ///< The mask is expected to be for permutation of 1-2 vectors, ///< check for the mask elements for the first argument (mask ///< indices are in range [0:VF)). SecondArg, ///< The mask is expected to be for permutation of 2 vectors, check ///< for the mask elements for the second argument (mask indices ///< are in range [VF:2*VF)) UndefsAsMask ///< Consider undef mask elements (-1) as placeholders for ///< future shuffle elements and mark them as ones as being used ///< in future. Non-undef elements are considered as unused since ///< they're already marked as used in the mask. }; } // namespace /// Prepares a use bitset for the given mask either for the first argument or /// for the second. static SmallBitVector buildUseMask(int VF, ArrayRef Mask, UseMask MaskArg) { SmallBitVector UseMask(VF, true); for (auto P : enumerate(Mask)) { if (P.value() == UndefMaskElem) { if (MaskArg == UseMask::UndefsAsMask) UseMask.reset(P.index()); continue; } if (MaskArg == UseMask::FirstArg && P.value() < VF) UseMask.reset(P.value()); else if (MaskArg == UseMask::SecondArg && P.value() >= VF) UseMask.reset(P.value() - VF); } return UseMask; } /// Checks if the given value is actually an undefined constant vector. /// Also, if the \p UseMask is not empty, tries to check if the non-masked /// elements actually mask the insertelement buildvector, if any. template static SmallBitVector isUndefVector(const Value *V, const SmallBitVector &UseMask = {}) { SmallBitVector Res(UseMask.empty() ? 1 : UseMask.size(), true); using T = std::conditional_t; if (isa(V)) return Res; auto *VecTy = dyn_cast(V->getType()); if (!VecTy) return Res.reset(); auto *C = dyn_cast(V); if (!C) { if (!UseMask.empty()) { const Value *Base = V; while (auto *II = dyn_cast(Base)) { if (isa(II->getOperand(1))) continue; Base = II->getOperand(0); std::optional Idx = getInsertIndex(II); if (!Idx) continue; if (*Idx < UseMask.size() && !UseMask.test(*Idx)) Res.reset(*Idx); } // TODO: Add analysis for shuffles here too. if (V == Base) { Res.reset(); } else { SmallBitVector SubMask(UseMask.size(), false); Res &= isUndefVector(Base, SubMask); } } else { Res.reset(); } return Res; } for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) { if (Constant *Elem = C->getAggregateElement(I)) if (!isa(Elem) && (UseMask.empty() || (I < UseMask.size() && !UseMask.test(I)))) Res.reset(I); } return Res; } /// Checks if the vector of instructions can be represented as a shuffle, like: /// %x0 = extractelement <4 x i8> %x, i32 0 /// %x3 = extractelement <4 x i8> %x, i32 3 /// %y1 = extractelement <4 x i8> %y, i32 1 /// %y2 = extractelement <4 x i8> %y, i32 2 /// %x0x0 = mul i8 %x0, %x0 /// %x3x3 = mul i8 %x3, %x3 /// %y1y1 = mul i8 %y1, %y1 /// %y2y2 = mul i8 %y2, %y2 /// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0 /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 /// ret <4 x i8> %ins4 /// can be transformed into: /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> /// %2 = mul <4 x i8> %1, %1 /// ret <4 x i8> %2 /// We convert this initially to something like: /// %x0 = extractelement <4 x i8> %x, i32 0 /// %x3 = extractelement <4 x i8> %x, i32 3 /// %y1 = extractelement <4 x i8> %y, i32 1 /// %y2 = extractelement <4 x i8> %y, i32 2 /// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0 /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 /// %5 = mul <4 x i8> %4, %4 /// %6 = extractelement <4 x i8> %5, i32 0 /// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0 /// %7 = extractelement <4 x i8> %5, i32 1 /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 /// %8 = extractelement <4 x i8> %5, i32 2 /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 /// %9 = extractelement <4 x i8> %5, i32 3 /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 /// ret <4 x i8> %ins4 /// InstCombiner transforms this into a shuffle and vector mul /// Mask will return the Shuffle Mask equivalent to the extracted elements. /// TODO: Can we split off and reuse the shuffle mask detection from /// ShuffleVectorInst/getShuffleCost? static std::optional isFixedVectorShuffle(ArrayRef VL, SmallVectorImpl &Mask) { const auto *It = find_if(VL, [](Value *V) { return isa(V); }); if (It == VL.end()) return std::nullopt; auto *EI0 = cast(*It); if (isa(EI0->getVectorOperandType())) return std::nullopt; unsigned Size = cast(EI0->getVectorOperandType())->getNumElements(); Value *Vec1 = nullptr; Value *Vec2 = nullptr; enum ShuffleMode { Unknown, Select, Permute }; ShuffleMode CommonShuffleMode = Unknown; Mask.assign(VL.size(), UndefMaskElem); for (unsigned I = 0, E = VL.size(); I < E; ++I) { // Undef can be represented as an undef element in a vector. if (isa(VL[I])) continue; auto *EI = cast(VL[I]); if (isa(EI->getVectorOperandType())) return std::nullopt; auto *Vec = EI->getVectorOperand(); // We can extractelement from undef or poison vector. if (isUndefVector(Vec).all()) continue; // All vector operands must have the same number of vector elements. if (cast(Vec->getType())->getNumElements() != Size) return std::nullopt; if (isa(EI->getIndexOperand())) continue; auto *Idx = dyn_cast(EI->getIndexOperand()); if (!Idx) return std::nullopt; // Undefined behavior if Idx is negative or >= Size. if (Idx->getValue().uge(Size)) continue; unsigned IntIdx = Idx->getValue().getZExtValue(); Mask[I] = IntIdx; // For correct shuffling we have to have at most 2 different vector operands // in all extractelement instructions. if (!Vec1 || Vec1 == Vec) { Vec1 = Vec; } else if (!Vec2 || Vec2 == Vec) { Vec2 = Vec; Mask[I] += Size; } else { return std::nullopt; } if (CommonShuffleMode == Permute) continue; // If the extract index is not the same as the operation number, it is a // permutation. if (IntIdx != I) { CommonShuffleMode = Permute; continue; } CommonShuffleMode = Select; } // If we're not crossing lanes in different vectors, consider it as blending. if (CommonShuffleMode == Select && Vec2) return TargetTransformInfo::SK_Select; // If Vec2 was never used, we have a permutation of a single vector, otherwise // we have permutation of 2 vectors. return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc : TargetTransformInfo::SK_PermuteSingleSrc; } /// \returns True if Extract{Value,Element} instruction extracts element Idx. static std::optional getExtractIndex(Instruction *E) { unsigned Opcode = E->getOpcode(); assert((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction."); if (Opcode == Instruction::ExtractElement) { auto *CI = dyn_cast(E->getOperand(1)); if (!CI) return std::nullopt; return CI->getZExtValue(); } auto *EI = cast(E); if (EI->getNumIndices() != 1) return std::nullopt; return *EI->idx_begin(); } namespace { /// Main data required for vectorization of instructions. struct InstructionsState { /// The very first instruction in the list with the main opcode. Value *OpValue = nullptr; /// The main/alternate instruction. Instruction *MainOp = nullptr; Instruction *AltOp = nullptr; /// The main/alternate opcodes for the list of instructions. unsigned getOpcode() const { return MainOp ? MainOp->getOpcode() : 0; } unsigned getAltOpcode() const { return AltOp ? AltOp->getOpcode() : 0; } /// Some of the instructions in the list have alternate opcodes. bool isAltShuffle() const { return AltOp != MainOp; } bool isOpcodeOrAlt(Instruction *I) const { unsigned CheckedOpcode = I->getOpcode(); return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; } InstructionsState() = delete; InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} }; } // end anonymous namespace /// Chooses the correct key for scheduling data. If \p Op has the same (or /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p /// OpValue. static Value *isOneOf(const InstructionsState &S, Value *Op) { auto *I = dyn_cast(Op); if (I && S.isOpcodeOrAlt(I)) return Op; return S.OpValue; } /// \returns true if \p Opcode is allowed as part of of the main/alternate /// instruction for SLP vectorization. /// /// Example of unsupported opcode is SDIV that can potentially cause UB if the /// "shuffled out" lane would result in division by zero. static bool isValidForAlternation(unsigned Opcode) { if (Instruction::isIntDivRem(Opcode)) return false; return true; } static InstructionsState getSameOpcode(ArrayRef VL, const TargetLibraryInfo &TLI, unsigned BaseIndex = 0); /// Checks if the provided operands of 2 cmp instructions are compatible, i.e. /// compatible instructions or constants, or just some other regular values. static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0, Value *Op1, const TargetLibraryInfo &TLI) { return (isConstant(BaseOp0) && isConstant(Op0)) || (isConstant(BaseOp1) && isConstant(Op1)) || (!isa(BaseOp0) && !isa(Op0) && !isa(BaseOp1) && !isa(Op1)) || BaseOp0 == Op0 || BaseOp1 == Op1 || getSameOpcode({BaseOp0, Op0}, TLI).getOpcode() || getSameOpcode({BaseOp1, Op1}, TLI).getOpcode(); } /// \returns true if a compare instruction \p CI has similar "look" and /// same predicate as \p BaseCI, "as is" or with its operands and predicate /// swapped, false otherwise. static bool isCmpSameOrSwapped(const CmpInst *BaseCI, const CmpInst *CI, const TargetLibraryInfo &TLI) { assert(BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() && "Assessing comparisons of different types?"); CmpInst::Predicate BasePred = BaseCI->getPredicate(); CmpInst::Predicate Pred = CI->getPredicate(); CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(Pred); Value *BaseOp0 = BaseCI->getOperand(0); Value *BaseOp1 = BaseCI->getOperand(1); Value *Op0 = CI->getOperand(0); Value *Op1 = CI->getOperand(1); return (BasePred == Pred && areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1, TLI)) || (BasePred == SwappedPred && areCompatibleCmpOps(BaseOp0, BaseOp1, Op1, Op0, TLI)); } /// \returns analysis of the Instructions in \p VL described in /// InstructionsState, the Opcode that we suppose the whole list /// could be vectorized even if its structure is diverse. static InstructionsState getSameOpcode(ArrayRef VL, const TargetLibraryInfo &TLI, unsigned BaseIndex) { // Make sure these are all Instructions. if (llvm::any_of(VL, [](Value *V) { return !isa(V); })) return InstructionsState(VL[BaseIndex], nullptr, nullptr); bool IsCastOp = isa(VL[BaseIndex]); bool IsBinOp = isa(VL[BaseIndex]); bool IsCmpOp = isa(VL[BaseIndex]); CmpInst::Predicate BasePred = IsCmpOp ? cast(VL[BaseIndex])->getPredicate() : CmpInst::BAD_ICMP_PREDICATE; unsigned Opcode = cast(VL[BaseIndex])->getOpcode(); unsigned AltOpcode = Opcode; unsigned AltIndex = BaseIndex; // Check for one alternate opcode from another BinaryOperator. // TODO - generalize to support all operators (types, calls etc.). auto *IBase = cast(VL[BaseIndex]); Intrinsic::ID BaseID = 0; SmallVector BaseMappings; if (auto *CallBase = dyn_cast(IBase)) { BaseID = getVectorIntrinsicIDForCall(CallBase, &TLI); BaseMappings = VFDatabase(*CallBase).getMappings(*CallBase); if (!isTriviallyVectorizable(BaseID) && BaseMappings.empty()) return InstructionsState(VL[BaseIndex], nullptr, nullptr); } for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { auto *I = cast(VL[Cnt]); unsigned InstOpcode = I->getOpcode(); if (IsBinOp && isa(I)) { if (InstOpcode == Opcode || InstOpcode == AltOpcode) continue; if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && isValidForAlternation(Opcode)) { AltOpcode = InstOpcode; AltIndex = Cnt; continue; } } else if (IsCastOp && isa(I)) { Value *Op0 = IBase->getOperand(0); Type *Ty0 = Op0->getType(); Value *Op1 = I->getOperand(0); Type *Ty1 = Op1->getType(); if (Ty0 == Ty1) { if (InstOpcode == Opcode || InstOpcode == AltOpcode) continue; if (Opcode == AltOpcode) { assert(isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!"); AltOpcode = InstOpcode; AltIndex = Cnt; continue; } } } else if (auto *Inst = dyn_cast(VL[Cnt]); Inst && IsCmpOp) { auto *BaseInst = cast(VL[BaseIndex]); Type *Ty0 = BaseInst->getOperand(0)->getType(); Type *Ty1 = Inst->getOperand(0)->getType(); if (Ty0 == Ty1) { assert(InstOpcode == Opcode && "Expected same CmpInst opcode."); // Check for compatible operands. If the corresponding operands are not // compatible - need to perform alternate vectorization. CmpInst::Predicate CurrentPred = Inst->getPredicate(); CmpInst::Predicate SwappedCurrentPred = CmpInst::getSwappedPredicate(CurrentPred); if (E == 2 && (BasePred == CurrentPred || BasePred == SwappedCurrentPred)) continue; if (isCmpSameOrSwapped(BaseInst, Inst, TLI)) continue; auto *AltInst = cast(VL[AltIndex]); if (AltIndex != BaseIndex) { if (isCmpSameOrSwapped(AltInst, Inst, TLI)) continue; } else if (BasePred != CurrentPred) { assert( isValidForAlternation(InstOpcode) && "CmpInst isn't safe for alternation, logic needs to be updated!"); AltIndex = Cnt; continue; } CmpInst::Predicate AltPred = AltInst->getPredicate(); if (BasePred == CurrentPred || BasePred == SwappedCurrentPred || AltPred == CurrentPred || AltPred == SwappedCurrentPred) continue; } } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) { if (auto *Gep = dyn_cast(I)) { if (Gep->getNumOperands() != 2 || Gep->getOperand(0)->getType() != IBase->getOperand(0)->getType()) return InstructionsState(VL[BaseIndex], nullptr, nullptr); } else if (auto *EI = dyn_cast(I)) { if (!isVectorLikeInstWithConstOps(EI)) return InstructionsState(VL[BaseIndex], nullptr, nullptr); } else if (auto *LI = dyn_cast(I)) { auto *BaseLI = cast(IBase); if (!LI->isSimple() || !BaseLI->isSimple()) return InstructionsState(VL[BaseIndex], nullptr, nullptr); } else if (auto *Call = dyn_cast(I)) { auto *CallBase = cast(IBase); if (Call->getCalledFunction() != CallBase->getCalledFunction()) return InstructionsState(VL[BaseIndex], nullptr, nullptr); if (Call->hasOperandBundles() && !std::equal(Call->op_begin() + Call->getBundleOperandsStartIndex(), Call->op_begin() + Call->getBundleOperandsEndIndex(), CallBase->op_begin() + CallBase->getBundleOperandsStartIndex())) return InstructionsState(VL[BaseIndex], nullptr, nullptr); Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, &TLI); if (ID != BaseID) return InstructionsState(VL[BaseIndex], nullptr, nullptr); if (!ID) { SmallVector Mappings = VFDatabase(*Call).getMappings(*Call); if (Mappings.size() != BaseMappings.size() || Mappings.front().ISA != BaseMappings.front().ISA || Mappings.front().ScalarName != BaseMappings.front().ScalarName || Mappings.front().VectorName != BaseMappings.front().VectorName || Mappings.front().Shape.VF != BaseMappings.front().Shape.VF || Mappings.front().Shape.Parameters != BaseMappings.front().Shape.Parameters) return InstructionsState(VL[BaseIndex], nullptr, nullptr); } } continue; } return InstructionsState(VL[BaseIndex], nullptr, nullptr); } return InstructionsState(VL[BaseIndex], cast(VL[BaseIndex]), cast(VL[AltIndex])); } /// \returns true if all of the values in \p VL have the same type or false /// otherwise. static bool allSameType(ArrayRef VL) { Type *Ty = VL[0]->getType(); for (int i = 1, e = VL.size(); i < e; i++) if (VL[i]->getType() != Ty) return false; return true; } /// \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) { 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 = getVectorIntrinsicIDForCall(CI, TLI); for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) return (CI->getArgOperand(i) == Scalar); } [[fallthrough]]; } default: return false; } } /// \returns the AA location that is being access by the instruction. static MemoryLocation getLocation(Instruction *I) { if (StoreInst *SI = dyn_cast(I)) return MemoryLocation::get(SI); if (LoadInst *LI = dyn_cast(I)) return MemoryLocation::get(LI); return MemoryLocation(); } /// \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; } /// Shuffles \p Mask in accordance with the given \p SubMask. static void addMask(SmallVectorImpl &Mask, ArrayRef SubMask) { if (SubMask.empty()) return; if (Mask.empty()) { Mask.append(SubMask.begin(), SubMask.end()); return; } SmallVector NewMask(SubMask.size(), UndefMaskElem); int TermValue = std::min(Mask.size(), SubMask.size()); for (int I = 0, E = SubMask.size(); I < E; ++I) { if (SubMask[I] >= TermValue || SubMask[I] == UndefMaskElem || Mask[SubMask[I]] >= TermValue) continue; NewMask[I] = Mask[SubMask[I]]; } Mask.swap(NewMask); } /// Order may have elements assigned special value (size) which is out of /// bounds. Such indices only appear on places which correspond to undef values /// (see canReuseExtract for details) and used in order to avoid undef values /// have effect on operands ordering. /// The first loop below simply finds all unused indices and then the next loop /// nest assigns these indices for undef values positions. /// As an example below Order has two undef positions and they have assigned /// values 3 and 7 respectively: /// before: 6 9 5 4 9 2 1 0 /// after: 6 3 5 4 7 2 1 0 static void fixupOrderingIndices(SmallVectorImpl &Order) { const unsigned Sz = Order.size(); SmallBitVector UnusedIndices(Sz, /*t=*/true); SmallBitVector MaskedIndices(Sz); for (unsigned I = 0; I < Sz; ++I) { if (Order[I] < Sz) UnusedIndices.reset(Order[I]); else MaskedIndices.set(I); } if (MaskedIndices.none()) return; assert(UnusedIndices.count() == MaskedIndices.count() && "Non-synced masked/available indices."); int Idx = UnusedIndices.find_first(); int MIdx = MaskedIndices.find_first(); while (MIdx >= 0) { assert(Idx >= 0 && "Indices must be synced."); Order[MIdx] = Idx; Idx = UnusedIndices.find_next(Idx); MIdx = MaskedIndices.find_next(MIdx); } } namespace llvm { static void inversePermutation(ArrayRef Indices, SmallVectorImpl &Mask) { Mask.clear(); const unsigned E = Indices.size(); Mask.resize(E, UndefMaskElem); for (unsigned I = 0; I < E; ++I) Mask[Indices[I]] = I; } /// Reorders the list of scalars in accordance with the given \p Mask. static void reorderScalars(SmallVectorImpl &Scalars, ArrayRef Mask) { assert(!Mask.empty() && "Expected non-empty mask."); SmallVector Prev(Scalars.size(), UndefValue::get(Scalars.front()->getType())); Prev.swap(Scalars); for (unsigned I = 0, E = Prev.size(); I < E; ++I) if (Mask[I] != UndefMaskElem) Scalars[Mask[I]] = Prev[I]; } /// Checks if the provided value does not require scheduling. It does not /// require scheduling if this is not an instruction or it is an instruction /// that does not read/write memory and all operands are either not instructions /// or phi nodes or instructions from different blocks. static bool areAllOperandsNonInsts(Value *V) { auto *I = dyn_cast(V); if (!I) return true; return !mayHaveNonDefUseDependency(*I) && all_of(I->operands(), [I](Value *V) { auto *IO = dyn_cast(V); if (!IO) return true; return isa(IO) || IO->getParent() != I->getParent(); }); } /// Checks if the provided value does not require scheduling. It does not /// require scheduling if this is not an instruction or it is an instruction /// that does not read/write memory and all users are phi nodes or instructions /// from the different blocks. static bool isUsedOutsideBlock(Value *V) { auto *I = dyn_cast(V); if (!I) return true; // Limits the number of uses to save compile time. constexpr int UsesLimit = 8; return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(UsesLimit) && all_of(I->users(), [I](User *U) { auto *IU = dyn_cast(U); if (!IU) return true; return IU->getParent() != I->getParent() || isa(IU); }); } /// Checks if the specified value does not require scheduling. It does not /// require scheduling if all operands and all users do not need to be scheduled /// in the current basic block. static bool doesNotNeedToBeScheduled(Value *V) { return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V); } /// Checks if the specified array of instructions does not require scheduling. /// It is so if all either instructions have operands that do not require /// scheduling or their users do not require scheduling since they are phis or /// in other basic blocks. static bool doesNotNeedToSchedule(ArrayRef VL) { return !VL.empty() && (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts)); } namespace slpvectorizer { /// Bottom Up SLP Vectorizer. class BoUpSLP { struct TreeEntry; struct ScheduleData; class ShuffleInstructionBuilder; public: using ValueList = SmallVector; using InstrList = SmallVector; using ValueSet = SmallPtrSet; using StoreList = SmallVector; using ExtraValueToDebugLocsMap = MapVector>; using OrdersType = SmallVector; BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, const DataLayout *DL, OptimizationRemarkEmitter *ORE) : BatchAA(*Aa), F(Func), SE(Se), TTI(Tti), TLI(TLi), LI(Li), DT(Dt), AC(AC), DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { CodeMetrics::collectEphemeralValues(F, AC, EphValues); // 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(TargetTransformInfo::RGK_FixedWidthVector) .getFixedValue(); if (MinVectorRegSizeOption.getNumOccurrences()) MinVecRegSize = MinVectorRegSizeOption; else MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); } /// Vectorize the tree that starts with the elements in \p VL. /// Returns the vectorized root. Value *vectorizeTree(); /// Vectorize the tree but with the list of externally used values \p /// ExternallyUsedValues. Values in this MapVector can be replaced but the /// generated extractvalue instructions. Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues, Instruction *ReductionRoot = nullptr); /// \returns the cost incurred by unwanted spills and fills, caused by /// holding live values over call sites. InstructionCost getSpillCost() const; /// \returns the vectorization cost of the subtree that starts at \p VL. /// A negative number means that this is profitable. InstructionCost getTreeCost(ArrayRef VectorizedVals = std::nullopt); /// 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, const SmallDenseSet &UserIgnoreLst); /// Construct a vectorizable tree that starts at \p Roots. void buildTree(ArrayRef Roots); /// Checks if the very first tree node is going to be vectorized. bool isVectorizedFirstNode() const { return !VectorizableTree.empty() && VectorizableTree.front()->State == TreeEntry::Vectorize; } /// Returns the main instruction for the very first node. Instruction *getFirstNodeMainOp() const { assert(!VectorizableTree.empty() && "No tree to get the first node from"); return VectorizableTree.front()->getMainOp(); } /// Returns whether the root node has in-tree uses. bool doesRootHaveInTreeUses() const { return !VectorizableTree.empty() && !VectorizableTree.front()->UserTreeIndices.empty(); } /// Builds external uses of the vectorized scalars, i.e. the list of /// vectorized scalars to be extracted, their lanes and their scalar users. \p /// ExternallyUsedValues contains additional list of external uses to handle /// vectorization of reductions. void buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {}); /// Clear the internal data structures that are created by 'buildTree'. void deleteTree() { VectorizableTree.clear(); ScalarToTreeEntry.clear(); MustGather.clear(); EntryToLastInstruction.clear(); ExternalUses.clear(); for (auto &Iter : BlocksSchedules) { BlockScheduling *BS = Iter.second.get(); BS->clear(); } MinBWs.clear(); InstrElementSize.clear(); UserIgnoreList = nullptr; } unsigned getTreeSize() const { return VectorizableTree.size(); } /// Perform LICM and CSE on the newly generated gather sequences. void optimizeGatherSequence(); /// Checks if the specified gather tree entry \p TE can be represented as a /// shuffled vector entry + (possibly) permutation with other gathers. It /// implements the checks only for possibly ordered scalars (Loads, /// ExtractElement, ExtractValue), which can be part of the graph. std::optional findReusedOrderedScalars(const TreeEntry &TE); /// Sort loads into increasing pointers offsets to allow greater clustering. std::optional findPartiallyOrderedLoads(const TreeEntry &TE); /// Gets reordering data for the given tree entry. If the entry is vectorized /// - just return ReorderIndices, otherwise check if the scalars can be /// reordered and return the most optimal order. /// \param TopToBottom If true, include the order of vectorized stores and /// insertelement nodes, otherwise skip them. std::optional getReorderingData(const TreeEntry &TE, bool TopToBottom); /// Reorders the current graph to the most profitable order starting from the /// root node to the leaf nodes. The best order is chosen only from the nodes /// of the same size (vectorization factor). Smaller nodes are considered /// parts of subgraph with smaller VF and they are reordered independently. We /// can make it because we still need to extend smaller nodes to the wider VF /// and we can merge reordering shuffles with the widening shuffles. void reorderTopToBottom(); /// Reorders the current graph to the most profitable order starting from /// leaves to the root. It allows to rotate small subgraphs and reduce the /// number of reshuffles if the leaf nodes use the same order. In this case we /// can merge the orders and just shuffle user node instead of shuffling its /// operands. Plus, even the leaf nodes have different orders, it allows to /// sink reordering in the graph closer to the root node and merge it later /// during analysis. void reorderBottomToTop(bool IgnoreReorder = false); /// \return The vector element size in bits to use when vectorizing the /// expression tree ending at \p V. If V is a store, the size is the width of /// the stored value. Otherwise, the size is the width of the largest loaded /// value reaching V. This method is used by the vectorizer to calculate /// vectorization factors. unsigned getVectorElementSize(Value *V); /// Compute the minimum type sizes required to represent the entries in a /// vectorizable tree. void computeMinimumValueSizes(); // \returns maximum vector register size as set by TTI or overridden by cl::opt. unsigned getMaxVecRegSize() const { return MaxVecRegSize; } // \returns minimum vector register size as set by cl::opt. unsigned getMinVecRegSize() const { return MinVecRegSize; } unsigned getMinVF(unsigned Sz) const { return std::max(2U, getMinVecRegSize() / Sz); } unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { unsigned MaxVF = MaxVFOption.getNumOccurrences() ? MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode); return MaxVF ? MaxVF : UINT_MAX; } /// Check if homogeneous aggregate is isomorphic to some VectorType. /// Accepts homogeneous multidimensional aggregate of scalars/vectors like /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. /// /// \returns number of elements in vector if isomorphism exists, 0 otherwise. unsigned canMapToVector(Type *T, const DataLayout &DL) const; /// \returns True if the VectorizableTree is both tiny and not fully /// vectorizable. We do not vectorize such trees. bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const; /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values /// can be load combined in the backend. Load combining may not be allowed in /// the IR optimizer, so we do not want to alter the pattern. For example, /// partially transforming a scalar bswap() pattern into vector code is /// effectively impossible for the backend to undo. /// TODO: If load combining is allowed in the IR optimizer, this analysis /// may not be necessary. bool isLoadCombineReductionCandidate(RecurKind RdxKind) const; /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values /// can be load combined in the backend. Load combining may not be allowed in /// the IR optimizer, so we do not want to alter the pattern. For example, /// partially transforming a scalar bswap() pattern into vector code is /// effectively impossible for the backend to undo. /// TODO: If load combining is allowed in the IR optimizer, this analysis /// may not be necessary. bool isLoadCombineCandidate() const; OptimizationRemarkEmitter *getORE() { return ORE; } /// This structure holds any data we need about the edges being traversed /// during buildTree_rec(). We keep track of: /// (i) the user TreeEntry index, and /// (ii) the index of the edge. struct EdgeInfo { EdgeInfo() = default; EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) : UserTE(UserTE), EdgeIdx(EdgeIdx) {} /// The user TreeEntry. TreeEntry *UserTE = nullptr; /// The operand index of the use. unsigned EdgeIdx = UINT_MAX; #ifndef NDEBUG friend inline raw_ostream &operator<<(raw_ostream &OS, const BoUpSLP::EdgeInfo &EI) { EI.dump(OS); return OS; } /// Debug print. void dump(raw_ostream &OS) const { OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") << " EdgeIdx:" << EdgeIdx << "}"; } LLVM_DUMP_METHOD void dump() const { dump(dbgs()); } #endif }; /// A helper class used for scoring candidates for two consecutive lanes. class LookAheadHeuristics { const TargetLibraryInfo &TLI; const DataLayout &DL; ScalarEvolution &SE; const BoUpSLP &R; int NumLanes; // Total number of lanes (aka vectorization factor). int MaxLevel; // The maximum recursion depth for accumulating score. public: LookAheadHeuristics(const TargetLibraryInfo &TLI, const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R, int NumLanes, int MaxLevel) : TLI(TLI), DL(DL), SE(SE), R(R), NumLanes(NumLanes), MaxLevel(MaxLevel) {} // The hard-coded scores listed here are not very important, though it shall // be higher for better matches to improve the resulting cost. When // computing the scores of matching one sub-tree with another, we are // basically counting the number of values that are matching. So even if all // scores are set to 1, we would still get a decent matching result. // However, sometimes we have to break ties. For example we may have to // choose between matching loads vs matching opcodes. This is what these // scores are helping us with: they provide the order of preference. Also, // this is important if the scalar is externally used or used in another // tree entry node in the different lane. /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). static const int ScoreConsecutiveLoads = 4; /// The same load multiple times. This should have a better score than /// `ScoreSplat` because it in x86 for a 2-lane vector we can represent it /// with `movddup (%reg), xmm0` which has a throughput of 0.5 versus 0.5 for /// a vector load and 1.0 for a broadcast. static const int ScoreSplatLoads = 3; /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]). static const int ScoreReversedLoads = 3; /// A load candidate for masked gather. static const int ScoreMaskedGatherCandidate = 1; /// ExtractElementInst from same vector and consecutive indexes. static const int ScoreConsecutiveExtracts = 4; /// ExtractElementInst from same vector and reversed indices. static const int ScoreReversedExtracts = 3; /// Constants. static const int ScoreConstants = 2; /// Instructions with the same opcode. static const int ScoreSameOpcode = 2; /// Instructions with alt opcodes (e.g, add + sub). static const int ScoreAltOpcodes = 1; /// Identical instructions (a.k.a. splat or broadcast). static const int ScoreSplat = 1; /// Matching with an undef is preferable to failing. static const int ScoreUndef = 1; /// Score for failing to find a decent match. static const int ScoreFail = 0; /// Score if all users are vectorized. static const int ScoreAllUserVectorized = 1; /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. /// \p U1 and \p U2 are the users of \p V1 and \p V2. /// Also, checks if \p V1 and \p V2 are compatible with instructions in \p /// MainAltOps. int getShallowScore(Value *V1, Value *V2, Instruction *U1, Instruction *U2, ArrayRef MainAltOps) const { if (!isValidElementType(V1->getType()) || !isValidElementType(V2->getType())) return LookAheadHeuristics::ScoreFail; if (V1 == V2) { if (isa(V1)) { // Retruns true if the users of V1 and V2 won't need to be extracted. auto AllUsersAreInternal = [U1, U2, this](Value *V1, Value *V2) { // Bail out if we have too many uses to save compilation time. static constexpr unsigned Limit = 8; if (V1->hasNUsesOrMore(Limit) || V2->hasNUsesOrMore(Limit)) return false; auto AllUsersVectorized = [U1, U2, this](Value *V) { return llvm::all_of(V->users(), [U1, U2, this](Value *U) { return U == U1 || U == U2 || R.getTreeEntry(U) != nullptr; }); }; return AllUsersVectorized(V1) && AllUsersVectorized(V2); }; // A broadcast of a load can be cheaper on some targets. if (R.TTI->isLegalBroadcastLoad(V1->getType(), ElementCount::getFixed(NumLanes)) && ((int)V1->getNumUses() == NumLanes || AllUsersAreInternal(V1, V2))) return LookAheadHeuristics::ScoreSplatLoads; } return LookAheadHeuristics::ScoreSplat; } auto *LI1 = dyn_cast(V1); auto *LI2 = dyn_cast(V2); if (LI1 && LI2) { if (LI1->getParent() != LI2->getParent() || !LI1->isSimple() || !LI2->isSimple()) return LookAheadHeuristics::ScoreFail; std::optional Dist = getPointersDiff( LI1->getType(), LI1->getPointerOperand(), LI2->getType(), LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true); if (!Dist || *Dist == 0) { if (getUnderlyingObject(LI1->getPointerOperand()) == getUnderlyingObject(LI2->getPointerOperand()) && R.TTI->isLegalMaskedGather( FixedVectorType::get(LI1->getType(), NumLanes), LI1->getAlign())) return LookAheadHeuristics::ScoreMaskedGatherCandidate; return LookAheadHeuristics::ScoreFail; } // The distance is too large - still may be profitable to use masked // loads/gathers. if (std::abs(*Dist) > NumLanes / 2) return LookAheadHeuristics::ScoreMaskedGatherCandidate; // This still will detect consecutive loads, but we might have "holes" // in some cases. It is ok for non-power-2 vectorization and may produce // better results. It should not affect current vectorization. return (*Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveLoads : LookAheadHeuristics::ScoreReversedLoads; } auto *C1 = dyn_cast(V1); auto *C2 = dyn_cast(V2); if (C1 && C2) return LookAheadHeuristics::ScoreConstants; // Extracts from consecutive indexes of the same vector better score as // the extracts could be optimized away. Value *EV1; ConstantInt *Ex1Idx; if (match(V1, m_ExtractElt(m_Value(EV1), m_ConstantInt(Ex1Idx)))) { // Undefs are always profitable for extractelements. if (isa(V2)) return LookAheadHeuristics::ScoreConsecutiveExtracts; Value *EV2 = nullptr; ConstantInt *Ex2Idx = nullptr; if (match(V2, m_ExtractElt(m_Value(EV2), m_CombineOr(m_ConstantInt(Ex2Idx), m_Undef())))) { // Undefs are always profitable for extractelements. if (!Ex2Idx) return LookAheadHeuristics::ScoreConsecutiveExtracts; if (isUndefVector(EV2).all() && EV2->getType() == EV1->getType()) return LookAheadHeuristics::ScoreConsecutiveExtracts; if (EV2 == EV1) { int Idx1 = Ex1Idx->getZExtValue(); int Idx2 = Ex2Idx->getZExtValue(); int Dist = Idx2 - Idx1; // The distance is too large - still may be profitable to use // shuffles. if (std::abs(Dist) == 0) return LookAheadHeuristics::ScoreSplat; if (std::abs(Dist) > NumLanes / 2) return LookAheadHeuristics::ScoreSameOpcode; return (Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveExtracts : LookAheadHeuristics::ScoreReversedExtracts; } return LookAheadHeuristics::ScoreAltOpcodes; } return LookAheadHeuristics::ScoreFail; } auto *I1 = dyn_cast(V1); auto *I2 = dyn_cast(V2); if (I1 && I2) { if (I1->getParent() != I2->getParent()) return LookAheadHeuristics::ScoreFail; SmallVector Ops(MainAltOps.begin(), MainAltOps.end()); Ops.push_back(I1); Ops.push_back(I2); InstructionsState S = getSameOpcode(Ops, TLI); // Note: Only consider instructions with <= 2 operands to avoid // complexity explosion. if (S.getOpcode() && (S.MainOp->getNumOperands() <= 2 || !MainAltOps.empty() || !S.isAltShuffle()) && all_of(Ops, [&S](Value *V) { return cast(V)->getNumOperands() == S.MainOp->getNumOperands(); })) return S.isAltShuffle() ? LookAheadHeuristics::ScoreAltOpcodes : LookAheadHeuristics::ScoreSameOpcode; } if (isa(V2)) return LookAheadHeuristics::ScoreUndef; return LookAheadHeuristics::ScoreFail; } /// Go through the operands of \p LHS and \p RHS recursively until /// MaxLevel, and return the cummulative score. \p U1 and \p U2 are /// the users of \p LHS and \p RHS (that is \p LHS and \p RHS are operands /// of \p U1 and \p U2), except at the beginning of the recursion where /// these are set to nullptr. /// /// For example: /// \verbatim /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] /// \ / \ / \ / \ / /// + + + + /// G1 G2 G3 G4 /// \endverbatim /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at /// each level recursively, accumulating the score. It starts from matching /// the additions at level 0, then moves on to the loads (level 1). The /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and /// {B[0],B[1]} match with LookAheadHeuristics::ScoreConsecutiveLoads, while /// {A[0],C[0]} has a score of LookAheadHeuristics::ScoreFail. /// Please note that the order of the operands does not matter, as we /// evaluate the score of all profitable combinations of operands. In /// other words the score of G1 and G4 is the same as G1 and G2. This /// heuristic is based on ideas described in: /// Look-ahead SLP: Auto-vectorization in the presence of commutative /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, /// Luís F. W. Góes int getScoreAtLevelRec(Value *LHS, Value *RHS, Instruction *U1, Instruction *U2, int CurrLevel, ArrayRef MainAltOps) const { // Get the shallow score of V1 and V2. int ShallowScoreAtThisLevel = getShallowScore(LHS, RHS, U1, U2, MainAltOps); // If reached MaxLevel, // or if V1 and V2 are not instructions, // or if they are SPLAT, // or if they are not consecutive, // or if profitable to vectorize loads or extractelements, early return // the current cost. auto *I1 = dyn_cast(LHS); auto *I2 = dyn_cast(RHS); if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || ShallowScoreAtThisLevel == LookAheadHeuristics::ScoreFail || (((isa(I1) && isa(I2)) || (I1->getNumOperands() > 2 && I2->getNumOperands() > 2) || (isa(I1) && isa(I2))) && ShallowScoreAtThisLevel)) return ShallowScoreAtThisLevel; assert(I1 && I2 && "Should have early exited."); // Contains the I2 operand indexes that got matched with I1 operands. SmallSet Op2Used; // Recursion towards the operands of I1 and I2. We are trying all possible // operand pairs, and keeping track of the best score. for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); OpIdx1 != NumOperands1; ++OpIdx1) { // Try to pair op1I with the best operand of I2. int MaxTmpScore = 0; unsigned MaxOpIdx2 = 0; bool FoundBest = false; // If I2 is commutative try all combinations. unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; unsigned ToIdx = isCommutative(I2) ? I2->getNumOperands() : std::min(I2->getNumOperands(), OpIdx1 + 1); assert(FromIdx <= ToIdx && "Bad index"); for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { // Skip operands already paired with OpIdx1. if (Op2Used.count(OpIdx2)) continue; // Recursively calculate the cost at each level int TmpScore = getScoreAtLevelRec(I1->getOperand(OpIdx1), I2->getOperand(OpIdx2), I1, I2, CurrLevel + 1, std::nullopt); // Look for the best score. if (TmpScore > LookAheadHeuristics::ScoreFail && TmpScore > MaxTmpScore) { MaxTmpScore = TmpScore; MaxOpIdx2 = OpIdx2; FoundBest = true; } } if (FoundBest) { // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. Op2Used.insert(MaxOpIdx2); ShallowScoreAtThisLevel += MaxTmpScore; } } return ShallowScoreAtThisLevel; } }; /// A helper data structure to hold the operands of a vector of instructions. /// This supports a fixed vector length for all operand vectors. class VLOperands { /// For each operand we need (i) the value, and (ii) the opcode that it /// would be attached to if the expression was in a left-linearized form. /// This is required to avoid illegal operand reordering. /// For example: /// \verbatim /// 0 Op1 /// |/ /// Op1 Op2 Linearized + Op2 /// \ / ----------> |/ /// - - /// /// Op1 - Op2 (0 + Op1) - Op2 /// \endverbatim /// /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. /// /// Another way to think of this is to track all the operations across the /// path from the operand all the way to the root of the tree and to /// calculate the operation that corresponds to this path. For example, the /// path from Op2 to the root crosses the RHS of the '-', therefore the /// corresponding operation is a '-' (which matches the one in the /// linearized tree, as shown above). /// /// For lack of a better term, we refer to this operation as Accumulated /// Path Operation (APO). struct OperandData { OperandData() = default; OperandData(Value *V, bool APO, bool IsUsed) : V(V), APO(APO), IsUsed(IsUsed) {} /// The operand value. Value *V = nullptr; /// TreeEntries only allow a single opcode, or an alternate sequence of /// them (e.g, +, -). Therefore, we can safely use a boolean value for the /// APO. It is set to 'true' if 'V' is attached to an inverse operation /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise /// (e.g., Add/Mul) bool APO = false; /// Helper data for the reordering function. bool IsUsed = false; }; /// During operand reordering, we are trying to select the operand at lane /// that matches best with the operand at the neighboring lane. Our /// selection is based on the type of value we are looking for. For example, /// if the neighboring lane has a load, we need to look for a load that is /// accessing a consecutive address. These strategies are summarized in the /// 'ReorderingMode' enumerator. enum class ReorderingMode { Load, ///< Matching loads to consecutive memory addresses Opcode, ///< Matching instructions based on opcode (same or alternate) Constant, ///< Matching constants Splat, ///< Matching the same instruction multiple times (broadcast) Failed, ///< We failed to create a vectorizable group }; using OperandDataVec = SmallVector; /// A vector of operand vectors. SmallVector OpsVec; const TargetLibraryInfo &TLI; const DataLayout &DL; ScalarEvolution &SE; const BoUpSLP &R; /// \returns the operand data at \p OpIdx and \p Lane. OperandData &getData(unsigned OpIdx, unsigned Lane) { return OpsVec[OpIdx][Lane]; } /// \returns the operand data at \p OpIdx and \p Lane. Const version. const OperandData &getData(unsigned OpIdx, unsigned Lane) const { return OpsVec[OpIdx][Lane]; } /// Clears the used flag for all entries. void clearUsed() { for (unsigned OpIdx = 0, NumOperands = getNumOperands(); OpIdx != NumOperands; ++OpIdx) for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; ++Lane) OpsVec[OpIdx][Lane].IsUsed = false; } /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); } /// \param Lane lane of the operands under analysis. /// \param OpIdx operand index in \p Lane lane we're looking the best /// candidate for. /// \param Idx operand index of the current candidate value. /// \returns The additional score due to possible broadcasting of the /// elements in the lane. It is more profitable to have power-of-2 unique /// elements in the lane, it will be vectorized with higher probability /// after removing duplicates. Currently the SLP vectorizer supports only /// vectorization of the power-of-2 number of unique scalars. int getSplatScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { Value *IdxLaneV = getData(Idx, Lane).V; if (!isa(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V) return 0; SmallPtrSet Uniques; for (unsigned Ln = 0, E = getNumLanes(); Ln < E; ++Ln) { if (Ln == Lane) continue; Value *OpIdxLnV = getData(OpIdx, Ln).V; if (!isa(OpIdxLnV)) return 0; Uniques.insert(OpIdxLnV); } int UniquesCount = Uniques.size(); int UniquesCntWithIdxLaneV = Uniques.contains(IdxLaneV) ? UniquesCount : UniquesCount + 1; Value *OpIdxLaneV = getData(OpIdx, Lane).V; int UniquesCntWithOpIdxLaneV = Uniques.contains(OpIdxLaneV) ? UniquesCount : UniquesCount + 1; if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV) return 0; return (PowerOf2Ceil(UniquesCntWithOpIdxLaneV) - UniquesCntWithOpIdxLaneV) - (PowerOf2Ceil(UniquesCntWithIdxLaneV) - UniquesCntWithIdxLaneV); } /// \param Lane lane of the operands under analysis. /// \param OpIdx operand index in \p Lane lane we're looking the best /// candidate for. /// \param Idx operand index of the current candidate value. /// \returns The additional score for the scalar which users are all /// vectorized. int getExternalUseScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { Value *IdxLaneV = getData(Idx, Lane).V; Value *OpIdxLaneV = getData(OpIdx, Lane).V; // Do not care about number of uses for vector-like instructions // (extractelement/extractvalue with constant indices), they are extracts // themselves and already externally used. Vectorization of such // instructions does not add extra extractelement instruction, just may // remove it. if (isVectorLikeInstWithConstOps(IdxLaneV) && isVectorLikeInstWithConstOps(OpIdxLaneV)) return LookAheadHeuristics::ScoreAllUserVectorized; auto *IdxLaneI = dyn_cast(IdxLaneV); if (!IdxLaneI || !isa(OpIdxLaneV)) return 0; return R.areAllUsersVectorized(IdxLaneI, std::nullopt) ? LookAheadHeuristics::ScoreAllUserVectorized : 0; } /// Score scaling factor for fully compatible instructions but with /// different number of external uses. Allows better selection of the /// instructions with less external uses. static const int ScoreScaleFactor = 10; /// \Returns the look-ahead score, which tells us how much the sub-trees /// rooted at \p LHS and \p RHS match, the more they match the higher the /// score. This helps break ties in an informed way when we cannot decide on /// the order of the operands by just considering the immediate /// predecessors. int getLookAheadScore(Value *LHS, Value *RHS, ArrayRef MainAltOps, int Lane, unsigned OpIdx, unsigned Idx, bool &IsUsed) { LookAheadHeuristics LookAhead(TLI, DL, SE, R, getNumLanes(), LookAheadMaxDepth); // Keep track of the instruction stack as we recurse into the operands // during the look-ahead score exploration. int Score = LookAhead.getScoreAtLevelRec(LHS, RHS, /*U1=*/nullptr, /*U2=*/nullptr, /*CurrLevel=*/1, MainAltOps); if (Score) { int SplatScore = getSplatScore(Lane, OpIdx, Idx); if (Score <= -SplatScore) { // Set the minimum score for splat-like sequence to avoid setting // failed state. Score = 1; } else { Score += SplatScore; // Scale score to see the difference between different operands // and similar operands but all vectorized/not all vectorized // uses. It does not affect actual selection of the best // compatible operand in general, just allows to select the // operand with all vectorized uses. Score *= ScoreScaleFactor; Score += getExternalUseScore(Lane, OpIdx, Idx); IsUsed = true; } } return Score; } /// Best defined scores per lanes between the passes. Used to choose the /// best operand (with the highest score) between the passes. /// The key - {Operand Index, Lane}. /// The value - the best score between the passes for the lane and the /// operand. SmallDenseMap, unsigned, 8> BestScoresPerLanes; // Search all operands in Ops[*][Lane] for the one that matches best // Ops[OpIdx][LastLane] and return its opreand index. // If no good match can be found, return std::nullopt. std::optional getBestOperand(unsigned OpIdx, int Lane, int LastLane, ArrayRef ReorderingModes, ArrayRef MainAltOps) { unsigned NumOperands = getNumOperands(); // The operand of the previous lane at OpIdx. Value *OpLastLane = getData(OpIdx, LastLane).V; // Our strategy mode for OpIdx. ReorderingMode RMode = ReorderingModes[OpIdx]; if (RMode == ReorderingMode::Failed) return std::nullopt; // The linearized opcode of the operand at OpIdx, Lane. bool OpIdxAPO = getData(OpIdx, Lane).APO; // The best operand index and its score. // Sometimes we have more than one option (e.g., Opcode and Undefs), so we // are using the score to differentiate between the two. struct BestOpData { std::optional Idx; unsigned Score = 0; } BestOp; BestOp.Score = BestScoresPerLanes.try_emplace(std::make_pair(OpIdx, Lane), 0) .first->second; // Track if the operand must be marked as used. If the operand is set to // Score 1 explicitly (because of non power-of-2 unique scalars, we may // want to reestimate the operands again on the following iterations). bool IsUsed = RMode == ReorderingMode::Splat || RMode == ReorderingMode::Constant; // Iterate through all unused operands and look for the best. for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { // Get the operand at Idx and Lane. OperandData &OpData = getData(Idx, Lane); Value *Op = OpData.V; bool OpAPO = OpData.APO; // Skip already selected operands. if (OpData.IsUsed) continue; // Skip if we are trying to move the operand to a position with a // different opcode in the linearized tree form. This would break the // semantics. if (OpAPO != OpIdxAPO) continue; // Look for an operand that matches the current mode. switch (RMode) { case ReorderingMode::Load: case ReorderingMode::Constant: case ReorderingMode::Opcode: { bool LeftToRight = Lane > LastLane; Value *OpLeft = (LeftToRight) ? OpLastLane : Op; Value *OpRight = (LeftToRight) ? Op : OpLastLane; int Score = getLookAheadScore(OpLeft, OpRight, MainAltOps, Lane, OpIdx, Idx, IsUsed); if (Score > static_cast(BestOp.Score)) { BestOp.Idx = Idx; BestOp.Score = Score; BestScoresPerLanes[std::make_pair(OpIdx, Lane)] = Score; } break; } case ReorderingMode::Splat: if (Op == OpLastLane) BestOp.Idx = Idx; break; case ReorderingMode::Failed: llvm_unreachable("Not expected Failed reordering mode."); } } if (BestOp.Idx) { getData(*BestOp.Idx, Lane).IsUsed = IsUsed; return BestOp.Idx; } // If we could not find a good match return std::nullopt. return std::nullopt; } /// Helper for reorderOperandVecs. /// \returns the lane that we should start reordering from. This is the one /// which has the least number of operands that can freely move about or /// less profitable because it already has the most optimal set of operands. unsigned getBestLaneToStartReordering() const { unsigned Min = UINT_MAX; unsigned SameOpNumber = 0; // std::pair is used to implement a simple voting // algorithm and choose the lane with the least number of operands that // can freely move about or less profitable because it already has the // most optimal set of operands. The first unsigned is a counter for // voting, the second unsigned is the counter of lanes with instructions // with same/alternate opcodes and same parent basic block. MapVector> HashMap; // Try to be closer to the original results, if we have multiple lanes // with same cost. If 2 lanes have the same cost, use the one with the // lowest index. for (int I = getNumLanes(); I > 0; --I) { unsigned Lane = I - 1; OperandsOrderData NumFreeOpsHash = getMaxNumOperandsThatCanBeReordered(Lane); // Compare the number of operands that can move and choose the one with // the least number. if (NumFreeOpsHash.NumOfAPOs < Min) { Min = NumFreeOpsHash.NumOfAPOs; SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; HashMap.clear(); HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); } else if (NumFreeOpsHash.NumOfAPOs == Min && NumFreeOpsHash.NumOpsWithSameOpcodeParent < SameOpNumber) { // Select the most optimal lane in terms of number of operands that // should be moved around. SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); } else if (NumFreeOpsHash.NumOfAPOs == Min && NumFreeOpsHash.NumOpsWithSameOpcodeParent == SameOpNumber) { auto It = HashMap.find(NumFreeOpsHash.Hash); if (It == HashMap.end()) HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); else ++It->second.first; } } // Select the lane with the minimum counter. unsigned BestLane = 0; unsigned CntMin = UINT_MAX; for (const auto &Data : reverse(HashMap)) { if (Data.second.first < CntMin) { CntMin = Data.second.first; BestLane = Data.second.second; } } return BestLane; } /// Data structure that helps to reorder operands. struct OperandsOrderData { /// The best number of operands with the same APOs, which can be /// reordered. unsigned NumOfAPOs = UINT_MAX; /// Number of operands with the same/alternate instruction opcode and /// parent. unsigned NumOpsWithSameOpcodeParent = 0; /// Hash for the actual operands ordering. /// Used to count operands, actually their position id and opcode /// value. It is used in the voting mechanism to find the lane with the /// least number of operands that can freely move about or less profitable /// because it already has the most optimal set of operands. Can be /// replaced with SmallVector instead but hash code is faster /// and requires less memory. unsigned Hash = 0; }; /// \returns the maximum number of operands that are allowed to be reordered /// for \p Lane and the number of compatible instructions(with the same /// parent/opcode). This is used as a heuristic for selecting the first lane /// to start operand reordering. OperandsOrderData getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { unsigned CntTrue = 0; unsigned NumOperands = getNumOperands(); // Operands with the same APO can be reordered. We therefore need to count // how many of them we have for each APO, like this: Cnt[APO] = x. // Since we only have two APOs, namely true and false, we can avoid using // a map. Instead we can simply count the number of operands that // correspond to one of them (in this case the 'true' APO), and calculate // the other by subtracting it from the total number of operands. // Operands with the same instruction opcode and parent are more // profitable since we don't need to move them in many cases, with a high // probability such lane already can be vectorized effectively. bool AllUndefs = true; unsigned NumOpsWithSameOpcodeParent = 0; Instruction *OpcodeI = nullptr; BasicBlock *Parent = nullptr; unsigned Hash = 0; for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { const OperandData &OpData = getData(OpIdx, Lane); if (OpData.APO) ++CntTrue; // Use Boyer-Moore majority voting for finding the majority opcode and // the number of times it occurs. if (auto *I = dyn_cast(OpData.V)) { if (!OpcodeI || !getSameOpcode({OpcodeI, I}, TLI).getOpcode() || I->getParent() != Parent) { if (NumOpsWithSameOpcodeParent == 0) { NumOpsWithSameOpcodeParent = 1; OpcodeI = I; Parent = I->getParent(); } else { --NumOpsWithSameOpcodeParent; } } else { ++NumOpsWithSameOpcodeParent; } } Hash = hash_combine( Hash, hash_value((OpIdx + 1) * (OpData.V->getValueID() + 1))); AllUndefs = AllUndefs && isa(OpData.V); } if (AllUndefs) return {}; OperandsOrderData Data; Data.NumOfAPOs = std::max(CntTrue, NumOperands - CntTrue); Data.NumOpsWithSameOpcodeParent = NumOpsWithSameOpcodeParent; Data.Hash = Hash; return Data; } /// Go through the instructions in VL and append their operands. void appendOperandsOfVL(ArrayRef VL) { assert(!VL.empty() && "Bad VL"); assert((empty() || VL.size() == getNumLanes()) && "Expected same number of lanes"); assert(isa(VL[0]) && "Expected instruction"); unsigned NumOperands = cast(VL[0])->getNumOperands(); OpsVec.resize(NumOperands); unsigned NumLanes = VL.size(); for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { OpsVec[OpIdx].resize(NumLanes); for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { assert(isa(VL[Lane]) && "Expected instruction"); // Our tree has just 3 nodes: the root and two operands. // It is therefore trivial to get the APO. We only need to check the // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or // RHS operand. The LHS operand of both add and sub is never attached // to an inversese operation in the linearized form, therefore its APO // is false. The RHS is true only if VL[Lane] is an inverse operation. // Since operand reordering is performed on groups of commutative // operations or alternating sequences (e.g., +, -), we can safely // tell the inverse operations by checking commutativity. bool IsInverseOperation = !isCommutative(cast(VL[Lane])); bool APO = (OpIdx == 0) ? false : IsInverseOperation; OpsVec[OpIdx][Lane] = {cast(VL[Lane])->getOperand(OpIdx), APO, false}; } } } /// \returns the number of operands. unsigned getNumOperands() const { return OpsVec.size(); } /// \returns the number of lanes. unsigned getNumLanes() const { return OpsVec[0].size(); } /// \returns the operand value at \p OpIdx and \p Lane. Value *getValue(unsigned OpIdx, unsigned Lane) const { return getData(OpIdx, Lane).V; } /// \returns true if the data structure is empty. bool empty() const { return OpsVec.empty(); } /// Clears the data. void clear() { OpsVec.clear(); } /// \Returns true if there are enough operands identical to \p Op to fill /// the whole vector. /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { bool OpAPO = getData(OpIdx, Lane).APO; for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { if (Ln == Lane) continue; // This is set to true if we found a candidate for broadcast at Lane. bool FoundCandidate = false; for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { OperandData &Data = getData(OpI, Ln); if (Data.APO != OpAPO || Data.IsUsed) continue; if (Data.V == Op) { FoundCandidate = true; Data.IsUsed = true; break; } } if (!FoundCandidate) return false; } return true; } public: /// Initialize with all the operands of the instruction vector \p RootVL. VLOperands(ArrayRef RootVL, const TargetLibraryInfo &TLI, const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R) : TLI(TLI), DL(DL), SE(SE), R(R) { // Append all the operands of RootVL. appendOperandsOfVL(RootVL); } /// \Returns a value vector with the operands across all lanes for the /// opearnd at \p OpIdx. ValueList getVL(unsigned OpIdx) const { ValueList OpVL(OpsVec[OpIdx].size()); assert(OpsVec[OpIdx].size() == getNumLanes() && "Expected same num of lanes across all operands"); for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) OpVL[Lane] = OpsVec[OpIdx][Lane].V; return OpVL; } // Performs operand reordering for 2 or more operands. // The original operands are in OrigOps[OpIdx][Lane]. // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. void reorder() { unsigned NumOperands = getNumOperands(); unsigned NumLanes = getNumLanes(); // Each operand has its own mode. We are using this mode to help us select // the instructions for each lane, so that they match best with the ones // we have selected so far. SmallVector ReorderingModes(NumOperands); // This is a greedy single-pass algorithm. We are going over each lane // once and deciding on the best order right away with no back-tracking. // However, in order to increase its effectiveness, we start with the lane // that has operands that can move the least. For example, given the // following lanes: // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd // Lane 1 : A[1] = C[1] - B[1] // Visited 1st // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd // Lane 3 : A[3] = C[3] - B[3] // Visited 4th // we will start at Lane 1, since the operands of the subtraction cannot // be reordered. Then we will visit the rest of the lanes in a circular // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. // Find the first lane that we will start our search from. unsigned FirstLane = getBestLaneToStartReordering(); // Initialize the modes. for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { Value *OpLane0 = getValue(OpIdx, FirstLane); // Keep track if we have instructions with all the same opcode on one // side. if (isa(OpLane0)) ReorderingModes[OpIdx] = ReorderingMode::Load; else if (isa(OpLane0)) { // Check if OpLane0 should be broadcast. if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) ReorderingModes[OpIdx] = ReorderingMode::Splat; else ReorderingModes[OpIdx] = ReorderingMode::Opcode; } else if (isa(OpLane0)) ReorderingModes[OpIdx] = ReorderingMode::Constant; else if (isa(OpLane0)) // Our best hope is a Splat. It may save some cost in some cases. ReorderingModes[OpIdx] = ReorderingMode::Splat; else // NOTE: This should be unreachable. ReorderingModes[OpIdx] = ReorderingMode::Failed; } // Check that we don't have same operands. No need to reorder if operands // are just perfect diamond or shuffled diamond match. Do not do it only // for possible broadcasts or non-power of 2 number of scalars (just for // now). auto &&SkipReordering = [this]() { SmallPtrSet UniqueValues; ArrayRef Op0 = OpsVec.front(); for (const OperandData &Data : Op0) UniqueValues.insert(Data.V); for (ArrayRef Op : drop_begin(OpsVec, 1)) { if (any_of(Op, [&UniqueValues](const OperandData &Data) { return !UniqueValues.contains(Data.V); })) return false; } // TODO: Check if we can remove a check for non-power-2 number of // scalars after full support of non-power-2 vectorization. return UniqueValues.size() != 2 && isPowerOf2_32(UniqueValues.size()); }; // If the initial strategy fails for any of the operand indexes, then we // perform reordering again in a second pass. This helps avoid assigning // high priority to the failed strategy, and should improve reordering for // the non-failed operand indexes. for (int Pass = 0; Pass != 2; ++Pass) { // Check if no need to reorder operands since they're are perfect or // shuffled diamond match. // Need to to do it to avoid extra external use cost counting for // shuffled matches, which may cause regressions. if (SkipReordering()) break; // Skip the second pass if the first pass did not fail. bool StrategyFailed = false; // Mark all operand data as free to use. clearUsed(); // We keep the original operand order for the FirstLane, so reorder the // rest of the lanes. We are visiting the nodes in a circular fashion, // using FirstLane as the center point and increasing the radius // distance. SmallVector> MainAltOps(NumOperands); for (unsigned I = 0; I < NumOperands; ++I) MainAltOps[I].push_back(getData(I, FirstLane).V); for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { // Visit the lane on the right and then the lane on the left. for (int Direction : {+1, -1}) { int Lane = FirstLane + Direction * Distance; if (Lane < 0 || Lane >= (int)NumLanes) continue; int LastLane = Lane - Direction; assert(LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds"); // Look for a good match for each operand. for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { // Search for the operand that matches SortedOps[OpIdx][Lane-1]. std::optional BestIdx = getBestOperand( OpIdx, Lane, LastLane, ReorderingModes, MainAltOps[OpIdx]); // By not selecting a value, we allow the operands that follow to // select a better matching value. We will get a non-null value in // the next run of getBestOperand(). if (BestIdx) { // Swap the current operand with the one returned by // getBestOperand(). swap(OpIdx, *BestIdx, Lane); } else { // We failed to find a best operand, set mode to 'Failed'. ReorderingModes[OpIdx] = ReorderingMode::Failed; // Enable the second pass. StrategyFailed = true; } // Try to get the alternate opcode and follow it during analysis. if (MainAltOps[OpIdx].size() != 2) { OperandData &AltOp = getData(OpIdx, Lane); InstructionsState OpS = getSameOpcode({MainAltOps[OpIdx].front(), AltOp.V}, TLI); if (OpS.getOpcode() && OpS.isAltShuffle()) MainAltOps[OpIdx].push_back(AltOp.V); } } } } // Skip second pass if the strategy did not fail. if (!StrategyFailed) break; } } #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) LLVM_DUMP_METHOD static StringRef getModeStr(ReorderingMode RMode) { switch (RMode) { case ReorderingMode::Load: return "Load"; case ReorderingMode::Opcode: return "Opcode"; case ReorderingMode::Constant: return "Constant"; case ReorderingMode::Splat: return "Splat"; case ReorderingMode::Failed: return "Failed"; } llvm_unreachable("Unimplemented Reordering Type"); } LLVM_DUMP_METHOD static raw_ostream &printMode(ReorderingMode RMode, raw_ostream &OS) { return OS << getModeStr(RMode); } /// Debug print. LLVM_DUMP_METHOD static void dumpMode(ReorderingMode RMode) { printMode(RMode, dbgs()); } friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { return printMode(RMode, OS); } LLVM_DUMP_METHOD raw_ostream &print(raw_ostream &OS) const { const unsigned Indent = 2; unsigned Cnt = 0; for (const OperandDataVec &OpDataVec : OpsVec) { OS << "Operand " << Cnt++ << "\n"; for (const OperandData &OpData : OpDataVec) { OS.indent(Indent) << "{"; if (Value *V = OpData.V) OS << *V; else OS << "null"; OS << ", APO:" << OpData.APO << "}\n"; } OS << "\n"; } return OS; } /// Debug print. LLVM_DUMP_METHOD void dump() const { print(dbgs()); } #endif }; /// Evaluate each pair in \p Candidates and return index into \p Candidates /// for a pair which have highest score deemed to have best chance to form /// root of profitable tree to vectorize. Return std::nullopt if no candidate /// scored above the LookAheadHeuristics::ScoreFail. \param Limit Lower limit /// of the cost, considered to be good enough score. std::optional findBestRootPair(ArrayRef> Candidates, int Limit = LookAheadHeuristics::ScoreFail) { LookAheadHeuristics LookAhead(*TLI, *DL, *SE, *this, /*NumLanes=*/2, RootLookAheadMaxDepth); int BestScore = Limit; std::optional Index; for (int I : seq(0, Candidates.size())) { int Score = LookAhead.getScoreAtLevelRec(Candidates[I].first, Candidates[I].second, /*U1=*/nullptr, /*U2=*/nullptr, /*Level=*/1, std::nullopt); if (Score > BestScore) { BestScore = Score; Index = I; } } return Index; } /// Checks if the instruction is marked for deletion. bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } /// 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. void eraseInstruction(Instruction *I) { DeletedInstructions.insert(I); } /// Checks if the instruction was already analyzed for being possible /// reduction root. bool isAnalyzedReductionRoot(Instruction *I) const { return AnalyzedReductionsRoots.count(I); } /// Register given instruction as already analyzed for being possible /// reduction root. void analyzedReductionRoot(Instruction *I) { AnalyzedReductionsRoots.insert(I); } /// Checks if the provided list of reduced values was checked already for /// vectorization. bool areAnalyzedReductionVals(ArrayRef VL) const { return AnalyzedReductionVals.contains(hash_value(VL)); } /// Adds the list of reduced values to list of already checked values for the /// vectorization. void analyzedReductionVals(ArrayRef VL) { AnalyzedReductionVals.insert(hash_value(VL)); } /// Clear the list of the analyzed reduction root instructions. void clearReductionData() { AnalyzedReductionsRoots.clear(); AnalyzedReductionVals.clear(); } /// Checks if the given value is gathered in one of the nodes. bool isAnyGathered(const SmallDenseSet &Vals) const { return any_of(MustGather, [&](Value *V) { return Vals.contains(V); }); } /// Check if the value is vectorized in the tree. bool isVectorized(Value *V) const { return getTreeEntry(V); } ~BoUpSLP(); private: /// Check if the operands on the edges \p Edges of the \p UserTE allows /// reordering (i.e. the operands can be reordered because they have only one /// user and reordarable). /// \param ReorderableGathers List of all gather nodes that require reordering /// (e.g., gather of extractlements or partially vectorizable loads). /// \param GatherOps List of gather operand nodes for \p UserTE that require /// reordering, subset of \p NonVectorized. bool canReorderOperands(TreeEntry *UserTE, SmallVectorImpl> &Edges, ArrayRef ReorderableGathers, SmallVectorImpl &GatherOps); /// Checks if the given \p TE is a gather node with clustered reused scalars /// and reorders it per given \p Mask. void reorderNodeWithReuses(TreeEntry &TE, ArrayRef Mask) const; /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, /// if any. If it is not vectorized (gather node), returns nullptr. TreeEntry *getVectorizedOperand(TreeEntry *UserTE, unsigned OpIdx) { ArrayRef VL = UserTE->getOperand(OpIdx); TreeEntry *TE = nullptr; const auto *It = find_if(VL, [this, &TE](Value *V) { TE = getTreeEntry(V); return TE; }); if (It != VL.end() && TE->isSame(VL)) return TE; return nullptr; } /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, /// if any. If it is not vectorized (gather node), returns nullptr. const TreeEntry *getVectorizedOperand(const TreeEntry *UserTE, unsigned OpIdx) const { return const_cast(this)->getVectorizedOperand( const_cast(UserTE), OpIdx); } /// Checks if all users of \p I are the part of the vectorization tree. bool areAllUsersVectorized(Instruction *I, ArrayRef VectorizedVals) const; /// Return information about the vector formed for the specified index /// of a vector of (the same) instruction. TargetTransformInfo::OperandValueInfo getOperandInfo(ArrayRef VL, unsigned OpIdx); /// \returns the cost of the vectorizable entry. InstructionCost getEntryCost(const TreeEntry *E, ArrayRef VectorizedVals); /// This is the recursive part of buildTree. void buildTree_rec(ArrayRef Roots, unsigned Depth, const EdgeInfo &EI); /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can /// be vectorized to use the original vector (or aggregate "bitcast" to a /// vector) and sets \p CurrentOrder to the identity permutation; otherwise /// returns false, setting \p CurrentOrder to either an empty vector or a /// non-identity permutation that allows to reuse extract instructions. bool canReuseExtract(ArrayRef VL, Value *OpValue, SmallVectorImpl &CurrentOrder) const; /// Vectorize a single entry in the tree. Value *vectorizeTree(TreeEntry *E); /// Vectorize a single entry in the tree, the \p Idx-th operand of the entry /// \p E. Value *vectorizeOperand(TreeEntry *E, unsigned NodeIdx); /// Create a new vector from a list of scalar values. Produces a sequence /// which exploits values reused across lanes, and arranges the inserts /// for ease of later optimization. Value *createBuildVector(const TreeEntry *E); /// \returns the scalarization cost for this type. Scalarization in this /// context means the creation of vectors from a group of scalars. If \p /// NeedToShuffle is true, need to add a cost of reshuffling some of the /// vector elements. InstructionCost getGatherCost(FixedVectorType *Ty, const APInt &ShuffledIndices, bool NeedToShuffle) const; /// Returns the instruction in the bundle, which can be used as a base point /// for scheduling. Usually it is the last instruction in the bundle, except /// for the case when all operands are external (in this case, it is the first /// instruction in the list). Instruction &getLastInstructionInBundle(const TreeEntry *E); /// Checks if the gathered \p VL can be represented as shuffle(s) of previous /// tree entries. /// \param TE Tree entry checked for permutation. /// \param VL List of scalars (a subset of the TE scalar), checked for /// permutations. /// \returns ShuffleKind, if gathered values can be represented as shuffles of /// previous tree entries. \p Mask is filled with the shuffle mask. std::optional isGatherShuffledEntry(const TreeEntry *TE, ArrayRef VL, SmallVectorImpl &Mask, SmallVectorImpl &Entries); /// \returns the scalarization cost for this list of values. Assuming that /// this subtree gets vectorized, we may need to extract the values from the /// roots. This method calculates the cost of extracting the values. InstructionCost getGatherCost(ArrayRef VL) const; /// Set the Builder insert point to one after the last instruction in /// the bundle void setInsertPointAfterBundle(const TreeEntry *E); /// \returns a vector from a collection of scalars in \p VL. Value *gather(ArrayRef VL); /// \returns whether the VectorizableTree is fully vectorizable and will /// be beneficial even the tree height is tiny. bool isFullyVectorizableTinyTree(bool ForReduction) const; /// Reorder commutative or alt operands to get better probability of /// generating vectorized code. static void reorderInputsAccordingToOpcode( ArrayRef VL, SmallVectorImpl &Left, SmallVectorImpl &Right, const TargetLibraryInfo &TLI, const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R); /// Helper for `findExternalStoreUsersReorderIndices()`. It iterates over the /// users of \p TE and collects the stores. It returns the map from the store /// pointers to the collected stores. DenseMap> collectUserStores(const BoUpSLP::TreeEntry *TE) const; /// Helper for `findExternalStoreUsersReorderIndices()`. It checks if the /// stores in \p StoresVec can form a vector instruction. If so it returns true /// and populates \p ReorderIndices with the shuffle indices of the the stores /// when compared to the sorted vector. bool canFormVector(const SmallVector &StoresVec, OrdersType &ReorderIndices) const; /// Iterates through the users of \p TE, looking for scalar stores that can be /// potentially vectorized in a future SLP-tree. If found, it keeps track of /// their order and builds an order index vector for each store bundle. It /// returns all these order vectors found. /// We run this after the tree has formed, otherwise we may come across user /// instructions that are not yet in the tree. SmallVector findExternalStoreUsersReorderIndices(TreeEntry *TE) const; struct TreeEntry { using VecTreeTy = SmallVector, 8>; TreeEntry(VecTreeTy &Container) : Container(Container) {} /// \returns true if the scalars in VL are equal to this entry. bool isSame(ArrayRef VL) const { auto &&IsSame = [VL](ArrayRef Scalars, ArrayRef Mask) { if (Mask.size() != VL.size() && VL.size() == Scalars.size()) return std::equal(VL.begin(), VL.end(), Scalars.begin()); return VL.size() == Mask.size() && std::equal(VL.begin(), VL.end(), Mask.begin(), [Scalars](Value *V, int Idx) { return (isa(V) && Idx == UndefMaskElem) || (Idx != UndefMaskElem && V == Scalars[Idx]); }); }; if (!ReorderIndices.empty()) { // TODO: implement matching if the nodes are just reordered, still can // treat the vector as the same if the list of scalars matches VL // directly, without reordering. SmallVector Mask; inversePermutation(ReorderIndices, Mask); if (VL.size() == Scalars.size()) return IsSame(Scalars, Mask); if (VL.size() == ReuseShuffleIndices.size()) { ::addMask(Mask, ReuseShuffleIndices); return IsSame(Scalars, Mask); } return false; } return IsSame(Scalars, ReuseShuffleIndices); } bool isOperandGatherNode(const EdgeInfo &UserEI) const { return State == TreeEntry::NeedToGather && UserTreeIndices.front().EdgeIdx == UserEI.EdgeIdx && UserTreeIndices.front().UserTE == UserEI.UserTE; } /// \returns true if current entry has same operands as \p TE. bool hasEqualOperands(const TreeEntry &TE) const { if (TE.getNumOperands() != getNumOperands()) return false; SmallBitVector Used(getNumOperands()); for (unsigned I = 0, E = getNumOperands(); I < E; ++I) { unsigned PrevCount = Used.count(); for (unsigned K = 0; K < E; ++K) { if (Used.test(K)) continue; if (getOperand(K) == TE.getOperand(I)) { Used.set(K); break; } } // Check if we actually found the matching operand. if (PrevCount == Used.count()) return false; } return true; } /// \return Final vectorization factor for the node. Defined by the total /// number of vectorized scalars, including those, used several times in the /// entry and counted in the \a ReuseShuffleIndices, if any. unsigned getVectorFactor() const { if (!ReuseShuffleIndices.empty()) return ReuseShuffleIndices.size(); return Scalars.size(); }; /// A vector of scalars. ValueList Scalars; /// The Scalars are vectorized into this value. It is initialized to Null. Value *VectorizedValue = nullptr; /// Do we need to gather this sequence or vectorize it /// (either with vector instruction or with scatter/gather /// intrinsics for store/load)? enum EntryState { Vectorize, ScatterVectorize, NeedToGather }; EntryState State; /// Does this sequence require some shuffling? SmallVector ReuseShuffleIndices; /// Does this entry require reordering? SmallVector ReorderIndices; /// Points back to the VectorizableTree. /// /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has /// to be a pointer and needs to be able to initialize the child iterator. /// Thus we need a reference back to the container to translate the indices /// to entries. VecTreeTy &Container; /// The TreeEntry index containing the user of this entry. We can actually /// have multiple users so the data structure is not truly a tree. SmallVector UserTreeIndices; /// The index of this treeEntry in VectorizableTree. int Idx = -1; private: /// The operands of each instruction in each lane Operands[op_index][lane]. /// Note: This helps avoid the replication of the code that performs the /// reordering of operands during buildTree_rec() and vectorizeTree(). SmallVector Operands; /// The main/alternate instruction. Instruction *MainOp = nullptr; Instruction *AltOp = nullptr; public: /// Set this bundle's \p OpIdx'th operand to \p OpVL. void setOperand(unsigned OpIdx, ArrayRef OpVL) { if (Operands.size() < OpIdx + 1) Operands.resize(OpIdx + 1); assert(Operands[OpIdx].empty() && "Already resized?"); assert(OpVL.size() <= Scalars.size() && "Number of operands is greater than the number of scalars."); Operands[OpIdx].resize(OpVL.size()); copy(OpVL, Operands[OpIdx].begin()); } /// Set the operands of this bundle in their original order. void setOperandsInOrder() { assert(Operands.empty() && "Already initialized?"); auto *I0 = cast(Scalars[0]); Operands.resize(I0->getNumOperands()); unsigned NumLanes = Scalars.size(); for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); OpIdx != NumOperands; ++OpIdx) { Operands[OpIdx].resize(NumLanes); for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { auto *I = cast(Scalars[Lane]); assert(I->getNumOperands() == NumOperands && "Expected same number of operands"); Operands[OpIdx][Lane] = I->getOperand(OpIdx); } } } /// Reorders operands of the node to the given mask \p Mask. void reorderOperands(ArrayRef Mask) { for (ValueList &Operand : Operands) reorderScalars(Operand, Mask); } /// \returns the \p OpIdx operand of this TreeEntry. ValueList &getOperand(unsigned OpIdx) { assert(OpIdx < Operands.size() && "Off bounds"); return Operands[OpIdx]; } /// \returns the \p OpIdx operand of this TreeEntry. ArrayRef getOperand(unsigned OpIdx) const { assert(OpIdx < Operands.size() && "Off bounds"); return Operands[OpIdx]; } /// \returns the number of operands. unsigned getNumOperands() const { return Operands.size(); } /// \return the single \p OpIdx operand. Value *getSingleOperand(unsigned OpIdx) const { assert(OpIdx < Operands.size() && "Off bounds"); assert(!Operands[OpIdx].empty() && "No operand available"); return Operands[OpIdx][0]; } /// Some of the instructions in the list have alternate opcodes. bool isAltShuffle() const { return MainOp != AltOp; } bool isOpcodeOrAlt(Instruction *I) const { unsigned CheckedOpcode = I->getOpcode(); return (getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode); } /// Chooses the correct key for scheduling data. If \p Op has the same (or /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is /// \p OpValue. Value *isOneOf(Value *Op) const { auto *I = dyn_cast(Op); if (I && isOpcodeOrAlt(I)) return Op; return MainOp; } void setOperations(const InstructionsState &S) { MainOp = S.MainOp; AltOp = S.AltOp; } Instruction *getMainOp() const { return MainOp; } Instruction *getAltOp() const { return AltOp; } /// The main/alternate opcodes for the list of instructions. unsigned getOpcode() const { return MainOp ? MainOp->getOpcode() : 0; } unsigned getAltOpcode() const { return AltOp ? AltOp->getOpcode() : 0; } /// When ReuseReorderShuffleIndices is empty it just returns position of \p /// V within vector of Scalars. Otherwise, try to remap on its reuse index. int findLaneForValue(Value *V) const { unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V)); assert(FoundLane < Scalars.size() && "Couldn't find extract lane"); if (!ReorderIndices.empty()) FoundLane = ReorderIndices[FoundLane]; assert(FoundLane < Scalars.size() && "Couldn't find extract lane"); if (!ReuseShuffleIndices.empty()) { FoundLane = std::distance(ReuseShuffleIndices.begin(), find(ReuseShuffleIndices, FoundLane)); } return FoundLane; } #ifndef NDEBUG /// Debug printer. LLVM_DUMP_METHOD void dump() const { dbgs() << Idx << ".\n"; for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { dbgs() << "Operand " << OpI << ":\n"; for (const Value *V : Operands[OpI]) dbgs().indent(2) << *V << "\n"; } dbgs() << "Scalars: \n"; for (Value *V : Scalars) dbgs().indent(2) << *V << "\n"; dbgs() << "State: "; switch (State) { case Vectorize: dbgs() << "Vectorize\n"; break; case ScatterVectorize: dbgs() << "ScatterVectorize\n"; break; case NeedToGather: dbgs() << "NeedToGather\n"; break; } dbgs() << "MainOp: "; if (MainOp) dbgs() << *MainOp << "\n"; else dbgs() << "NULL\n"; dbgs() << "AltOp: "; if (AltOp) dbgs() << *AltOp << "\n"; else dbgs() << "NULL\n"; dbgs() << "VectorizedValue: "; if (VectorizedValue) dbgs() << *VectorizedValue << "\n"; else dbgs() << "NULL\n"; dbgs() << "ReuseShuffleIndices: "; if (ReuseShuffleIndices.empty()) dbgs() << "Empty"; else for (int ReuseIdx : ReuseShuffleIndices) dbgs() << ReuseIdx << ", "; dbgs() << "\n"; dbgs() << "ReorderIndices: "; for (unsigned ReorderIdx : ReorderIndices) dbgs() << ReorderIdx << ", "; dbgs() << "\n"; dbgs() << "UserTreeIndices: "; for (const auto &EInfo : UserTreeIndices) dbgs() << EInfo << ", "; dbgs() << "\n"; } #endif }; #ifndef NDEBUG void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost, InstructionCost VecCost, InstructionCost ScalarCost) const { dbgs() << "SLP: Calculated costs for Tree:\n"; E->dump(); dbgs() << "SLP: Costs:\n"; dbgs() << "SLP: ReuseShuffleCost = " << ReuseShuffleCost << "\n"; dbgs() << "SLP: VectorCost = " << VecCost << "\n"; dbgs() << "SLP: ScalarCost = " << ScalarCost << "\n"; dbgs() << "SLP: ReuseShuffleCost + VecCost - ScalarCost = " << ReuseShuffleCost + VecCost - ScalarCost << "\n"; } #endif /// Create a new VectorizableTree entry. TreeEntry *newTreeEntry(ArrayRef VL, std::optional Bundle, const InstructionsState &S, const EdgeInfo &UserTreeIdx, ArrayRef ReuseShuffleIndices = std::nullopt, ArrayRef ReorderIndices = std::nullopt) { TreeEntry::EntryState EntryState = Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather; return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx, ReuseShuffleIndices, ReorderIndices); } TreeEntry *newTreeEntry(ArrayRef VL, TreeEntry::EntryState EntryState, std::optional Bundle, const InstructionsState &S, const EdgeInfo &UserTreeIdx, ArrayRef ReuseShuffleIndices = std::nullopt, ArrayRef ReorderIndices = std::nullopt) { assert(((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?"); VectorizableTree.push_back(std::make_unique(VectorizableTree)); TreeEntry *Last = VectorizableTree.back().get(); Last->Idx = VectorizableTree.size() - 1; Last->State = EntryState; Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), ReuseShuffleIndices.end()); if (ReorderIndices.empty()) { Last->Scalars.assign(VL.begin(), VL.end()); Last->setOperations(S); } else { // Reorder scalars and build final mask. Last->Scalars.assign(VL.size(), nullptr); transform(ReorderIndices, Last->Scalars.begin(), [VL](unsigned Idx) -> Value * { if (Idx >= VL.size()) return UndefValue::get(VL.front()->getType()); return VL[Idx]; }); InstructionsState S = getSameOpcode(Last->Scalars, *TLI); Last->setOperations(S); Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); } if (Last->State != TreeEntry::NeedToGather) { for (Value *V : VL) { assert(!getTreeEntry(V) && "Scalar already in tree!"); ScalarToTreeEntry[V] = Last; } // Update the scheduler bundle to point to this TreeEntry. ScheduleData *BundleMember = *Bundle; assert((BundleMember || isa(S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule(VL)) && "Bundle and VL out of sync"); if (BundleMember) { for (Value *V : VL) { if (doesNotNeedToBeScheduled(V)) continue; assert(BundleMember && "Unexpected end of bundle."); BundleMember->TE = Last; BundleMember = BundleMember->NextInBundle; } } assert(!BundleMember && "Bundle and VL out of sync"); } else { MustGather.insert(VL.begin(), VL.end()); } if (UserTreeIdx.UserTE) Last->UserTreeIndices.push_back(UserTreeIdx); return Last; } /// -- Vectorization State -- /// Holds all of the tree entries. TreeEntry::VecTreeTy VectorizableTree; #ifndef NDEBUG /// Debug printer. LLVM_DUMP_METHOD void dumpVectorizableTree() const { for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { VectorizableTree[Id]->dump(); dbgs() << "\n"; } } #endif TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); } const TreeEntry *getTreeEntry(Value *V) const { return ScalarToTreeEntry.lookup(V); } /// Maps a specific scalar to its tree entry. SmallDenseMap ScalarToTreeEntry; /// Maps a value to the proposed vectorizable size. SmallDenseMap InstrElementSize; /// A list of scalars that we found that we need to keep as scalars. ValueSet MustGather; /// A map between the vectorized entries and the last instructions in the /// bundles. The bundles are built in use order, not in the def order of the /// instructions. So, we cannot rely directly on the last instruction in the /// bundle being the last instruction in the program order during /// vectorization process since the basic blocks are affected, need to /// pre-gather them before. DenseMap EntryToLastInstruction; /// 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) {} // Which scalar in our function. Value *Scalar; // Which user that uses the scalar. llvm::User *User; // Which lane does the scalar belong to. int Lane; }; using UserList = SmallVector; /// 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); std::optional &result = AliasCache[key]; if (result) { return *result; } bool aliased = true; if (Loc1.Ptr && isSimple(Inst1)) aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1)); // Store the result in the cache. result = aliased; return aliased; } using AliasCacheKey = std::pair; /// Cache for alias results. /// TODO: consider moving this to the AliasAnalysis itself. DenseMap> AliasCache; // Cache for pointerMayBeCaptured calls inside AA. This is preserved // globally through SLP because we don't perform any action which // invalidates capture results. BatchAAResults BatchAA; /// Temporary store for deleted instructions. Instructions will be deleted /// eventually when the BoUpSLP is destructed. The deferral 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. DenseSet DeletedInstructions; /// Set of the instruction, being analyzed already for reductions. SmallPtrSet AnalyzedReductionsRoots; /// Set of hashes for the list of reduction values already being analyzed. DenseSet AnalyzedReductionVals; /// A list of values that need to extracted out of the tree. /// This list holds pairs of (Internal Scalar : External User). External User /// can be nullptr, it means that this Internal Scalar will be used later, /// after vectorization. UserList ExternalUses; /// Values used only by @llvm.assume calls. SmallPtrSet EphValues; /// Holds all of the instructions that we gathered, shuffle instructions and /// extractelements. SetVector GatherShuffleExtractSeq; /// A list of blocks that we are going to CSE. SetVector CSEBlocks; /// 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() = default; void init(int BlockSchedulingRegionID, Value *OpVal) { FirstInBundle = this; NextInBundle = nullptr; NextLoadStore = nullptr; IsScheduled = false; SchedulingRegionID = BlockSchedulingRegionID; clearDependencies(); OpValue = OpVal; TE = nullptr; } /// Verify basic self consistency properties void verify() { if (hasValidDependencies()) { assert(UnscheduledDeps <= Dependencies && "invariant"); } else { assert(UnscheduledDeps == Dependencies && "invariant"); } if (IsScheduled) { assert(isSchedulingEntity() && "unexpected scheduled state"); for (const ScheduleData *BundleMember = this; BundleMember; BundleMember = BundleMember->NextInBundle) { assert(BundleMember->hasValidDependencies() && BundleMember->UnscheduledDeps == 0 && "unexpected scheduled state"); assert((BundleMember == this || !BundleMember->IsScheduled) && "only bundle is marked scheduled"); } } assert(Inst->getParent() == FirstInBundle->Inst->getParent() && "all bundle members must be in same basic block"); } /// Returns true if the dependency information has been calculated. /// Note that depenendency validity can vary between instructions within /// a single bundle. 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 || TE; } /// 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 for this instruction, /// and returns the number of remaining dependencies for the containing /// bundle. int incrementUnscheduledDeps(int Incr) { assert(hasValidDependencies() && "increment of unscheduled deps would be meaningless"); UnscheduledDeps += Incr; return FirstInBundle->unscheduledDepsInBundle(); } /// Sets the number of unscheduled dependencies to the number of /// dependencies. void resetUnscheduledDeps() { UnscheduledDeps = Dependencies; } /// Clears all dependency information. void clearDependencies() { Dependencies = InvalidDeps; resetUnscheduledDeps(); MemoryDependencies.clear(); ControlDependencies.clear(); } int unscheduledDepsInBundle() const { assert(isSchedulingEntity() && "only meaningful on the bundle"); int Sum = 0; for (const ScheduleData *BundleMember = this; BundleMember; BundleMember = BundleMember->NextInBundle) { if (BundleMember->UnscheduledDeps == InvalidDeps) return InvalidDeps; Sum += BundleMember->UnscheduledDeps; } return Sum; } 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 = nullptr; /// Opcode of the current instruction in the schedule data. Value *OpValue = nullptr; /// The TreeEntry that this instruction corresponds to. TreeEntry *TE = nullptr; /// Points to the head in an instruction bundle (and always to this for /// single instructions). ScheduleData *FirstInBundle = nullptr; /// Single linked list of all instructions in a bundle. Null if it is a /// single instruction. ScheduleData *NextInBundle = nullptr; /// Single linked list of all memory instructions (e.g. load, store, call) /// in the block - until the end of the scheduling region. ScheduleData *NextLoadStore = nullptr; /// The dependent memory instructions. /// This list is derived on demand in calculateDependencies(). SmallVector MemoryDependencies; /// List of instructions which this instruction could be control dependent /// on. Allowing such nodes to be scheduled below this one could introduce /// a runtime fault which didn't exist in the original program. /// ex: this is a load or udiv following a readonly call which inf loops SmallVector ControlDependencies; /// This ScheduleData is in the current scheduling region if this matches /// the current SchedulingRegionID of BlockScheduling. int SchedulingRegionID = 0; /// Used for getting a "good" final ordering of instructions. int SchedulingPriority = 0; /// 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 = InvalidDeps; /// 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 = InvalidDeps; /// True if this instruction is scheduled (or considered as scheduled in the /// dry-run). bool IsScheduled = false; }; #ifndef NDEBUG friend inline raw_ostream &operator<<(raw_ostream &os, const BoUpSLP::ScheduleData &SD) { SD.dump(os); return os; } #endif friend struct GraphTraits; friend struct DOTGraphTraits; /// Contains all scheduling data for a basic block. /// It does not schedules instructions, which are not memory read/write /// instructions and their operands are either constants, or arguments, or /// phis, or instructions from others blocks, or their users are phis or from /// the other blocks. The resulting vector instructions can be placed at the /// beginning of the basic block without scheduling (if operands does not need /// to be scheduled) or at the end of the block (if users are outside of the /// block). It allows to save some compile time and memory used by the /// compiler. /// ScheduleData is assigned for each instruction in between the boundaries of /// the tree entry, even for those, which are not part of the graph. It is /// required to correctly follow the dependencies between the instructions and /// their correct scheduling. The ScheduleData is not allocated for the /// instructions, which do not require scheduling, like phis, nodes with /// extractelements/insertelements only or nodes with instructions, with /// uses/operands outside of the block. struct BlockScheduling { BlockScheduling(BasicBlock *BB) : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} void clear() { ReadyInsts.clear(); ScheduleStart = nullptr; ScheduleEnd = nullptr; FirstLoadStoreInRegion = nullptr; LastLoadStoreInRegion = nullptr; RegionHasStackSave = false; // 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(Instruction *I) { if (BB != I->getParent()) // Avoid lookup if can't possibly be in map. return nullptr; ScheduleData *SD = ScheduleDataMap.lookup(I); if (SD && isInSchedulingRegion(SD)) return SD; return nullptr; } ScheduleData *getScheduleData(Value *V) { if (auto *I = dyn_cast(V)) return getScheduleData(I); return nullptr; } ScheduleData *getScheduleData(Value *V, Value *Key) { if (V == Key) return getScheduleData(V); auto I = ExtraScheduleDataMap.find(V); if (I != ExtraScheduleDataMap.end()) { ScheduleData *SD = I->second.lookup(Key); if (SD && isInSchedulingRegion(SD)) return SD; } return nullptr; } bool isInSchedulingRegion(ScheduleData *SD) const { 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; LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n"); for (ScheduleData *BundleMember = SD; BundleMember; BundleMember = BundleMember->NextInBundle) { if (BundleMember->Inst != BundleMember->OpValue) continue; // Handle the def-use chain dependencies. // Decrement the unscheduled counter and insert to ready list if ready. auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { 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); LLVM_DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"); } }); }; // If BundleMember is a vector bundle, its operands may have been // reordered during buildTree(). We therefore need to get its operands // through the TreeEntry. if (TreeEntry *TE = BundleMember->TE) { // Need to search for the lane since the tree entry can be reordered. int Lane = std::distance(TE->Scalars.begin(), find(TE->Scalars, BundleMember->Inst)); assert(Lane >= 0 && "Lane not set"); // Since vectorization tree is being built recursively this assertion // ensures that the tree entry has all operands set before reaching // this code. Couple of exceptions known at the moment are extracts // where their second (immediate) operand is not added. Since // immediates do not affect scheduler behavior this is considered // okay. auto *In = BundleMember->Inst; assert(In && (isa(In) || In->getNumOperands() == TE->getNumOperands()) && "Missed TreeEntry operands?"); (void)In; // fake use to avoid build failure when assertions disabled for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); OpIdx != NumOperands; ++OpIdx) if (auto *I = dyn_cast(TE->getOperand(OpIdx)[Lane])) DecrUnsched(I); } else { // If BundleMember is a stand-alone instruction, no operand reordering // has taken place, so we directly access its operands. for (Use &U : BundleMember->Inst->operands()) if (auto *I = dyn_cast(U.get())) DecrUnsched(I); } // Handle the memory dependencies. for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { if (MemoryDepSD->hasValidDependencies() && 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); LLVM_DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"); } } // Handle the control dependencies. for (ScheduleData *DepSD : BundleMember->ControlDependencies) { if (DepSD->incrementUnscheduledDeps(-1) == 0) { // There are no more unscheduled dependencies after decrementing, // so we can put the dependent instruction into the ready list. ScheduleData *DepBundle = DepSD->FirstInBundle; assert(!DepBundle->IsScheduled && "already scheduled bundle gets ready"); ReadyList.insert(DepBundle); LLVM_DEBUG(dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"); } } } } /// Verify basic self consistency properties of the data structure. void verify() { if (!ScheduleStart) return; assert(ScheduleStart->getParent() == ScheduleEnd->getParent() && ScheduleStart->comesBefore(ScheduleEnd) && "Not a valid scheduling region?"); for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { auto *SD = getScheduleData(I); if (!SD) continue; assert(isInSchedulingRegion(SD) && "primary schedule data not in window?"); assert(isInSchedulingRegion(SD->FirstInBundle) && "entire bundle in window!"); (void)SD; doForAllOpcodes(I, [](ScheduleData *SD) { SD->verify(); }); } for (auto *SD : ReadyInsts) { assert(SD->isSchedulingEntity() && SD->isReady() && "item in ready list not ready?"); (void)SD; } } void doForAllOpcodes(Value *V, function_ref Action) { if (ScheduleData *SD = getScheduleData(V)) Action(SD); auto I = ExtraScheduleDataMap.find(V); if (I != ExtraScheduleDataMap.end()) for (auto &P : I->second) if (isInSchedulingRegion(P.second)) Action(P.second); } /// 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()) { doForAllOpcodes(I, [&](ScheduleData *SD) { if (SD->isSchedulingEntity() && SD->hasValidDependencies() && SD->isReady()) { ReadyList.insert(SD); LLVM_DEBUG(dbgs() << "SLP: initially in ready list: " << *SD << "\n"); } }); } } /// Build a bundle from the ScheduleData nodes corresponding to the /// scalar instruction for each lane. ScheduleData *buildBundle(ArrayRef VL); /// 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. /// \returns the scheduling bundle. The returned Optional value is not /// std::nullopt if \p VL is allowed to be scheduled. std::optional tryScheduleBundle(ArrayRef VL, BoUpSLP *SLP, const InstructionsState &S); /// Un-bundles a group of instructions. void cancelScheduling(ArrayRef VL, Value *OpValue); /// Allocates schedule data chunk. ScheduleData *allocateScheduleDataChunks(); /// 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, const InstructionsState &S); /// 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; /// Attaches ScheduleData to Instruction with the leading key. DenseMap> ExtraScheduleDataMap; /// The ready-list for scheduling (only used for the dry-run). SetVector ReadyInsts; /// The first instruction of the scheduling region. Instruction *ScheduleStart = nullptr; /// The first instruction _after_ the scheduling region. Instruction *ScheduleEnd = nullptr; /// The first memory accessing instruction in the scheduling region /// (can be null). ScheduleData *FirstLoadStoreInRegion = nullptr; /// The last memory accessing instruction in the scheduling region /// (can be null). ScheduleData *LastLoadStoreInRegion = nullptr; /// Is there an llvm.stacksave or llvm.stackrestore in the scheduling /// region? Used to optimize the dependence calculation for the /// common case where there isn't. bool RegionHasStackSave = false; /// The current size of the scheduling region. int ScheduleRegionSize = 0; /// The maximum size allowed for the scheduling region. int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; /// The ID of the scheduling region. For a new vectorization iteration this /// is incremented which "removes" all ScheduleData from the region. /// Make sure that the initial SchedulingRegionID is greater than the /// initial SchedulingRegionID in ScheduleData (which is 0). int SchedulingRegionID = 1; }; /// 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. const SmallDenseSet *UserIgnoreList = nullptr; /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of /// sorted SmallVectors of unsigned. struct OrdersTypeDenseMapInfo { static OrdersType getEmptyKey() { OrdersType V; V.push_back(~1U); return V; } static OrdersType getTombstoneKey() { OrdersType V; V.push_back(~2U); return V; } static unsigned getHashValue(const OrdersType &V) { return static_cast(hash_combine_range(V.begin(), V.end())); } static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { return LHS == RHS; } }; // Analysis and block reference. Function *F; ScalarEvolution *SE; TargetTransformInfo *TTI; TargetLibraryInfo *TLI; LoopInfo *LI; DominatorTree *DT; AssumptionCache *AC; DemandedBits *DB; const DataLayout *DL; OptimizationRemarkEmitter *ORE; unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. unsigned MinVecRegSize; // Set by cl::opt (default: 128). /// Instruction builder to construct the vectorized tree. IRBuilder<> Builder; /// A map of scalar integer values to the smallest bit width with which they /// can legally be represented. The values map to (width, signed) pairs, /// where "width" indicates the minimum bit width and "signed" is True if the /// value must be signed-extended, rather than zero-extended, back to its /// original width. MapVector> MinBWs; }; } // end namespace slpvectorizer template <> struct GraphTraits { using TreeEntry = BoUpSLP::TreeEntry; /// NodeRef has to be a pointer per the GraphWriter. using NodeRef = TreeEntry *; using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; /// Add the VectorizableTree to the index iterator to be able to return /// TreeEntry pointers. struct ChildIteratorType : public iterator_adaptor_base< ChildIteratorType, SmallVector::iterator> { ContainerTy &VectorizableTree; ChildIteratorType(SmallVector::iterator W, ContainerTy &VT) : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} NodeRef operator*() { return I->UserTE; } }; static NodeRef getEntryNode(BoUpSLP &R) { return R.VectorizableTree[0].get(); } static ChildIteratorType child_begin(NodeRef N) { return {N->UserTreeIndices.begin(), N->Container}; } static ChildIteratorType child_end(NodeRef N) { return {N->UserTreeIndices.end(), N->Container}; } /// For the node iterator we just need to turn the TreeEntry iterator into a /// TreeEntry* iterator so that it dereferences to NodeRef. class nodes_iterator { using ItTy = ContainerTy::iterator; ItTy It; public: nodes_iterator(const ItTy &It2) : It(It2) {} NodeRef operator*() { return It->get(); } nodes_iterator operator++() { ++It; return *this; } bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } }; static nodes_iterator nodes_begin(BoUpSLP *R) { return nodes_iterator(R->VectorizableTree.begin()); } static nodes_iterator nodes_end(BoUpSLP *R) { return nodes_iterator(R->VectorizableTree.end()); } static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } }; template <> struct DOTGraphTraits : public DefaultDOTGraphTraits { using TreeEntry = BoUpSLP::TreeEntry; DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { std::string Str; raw_string_ostream OS(Str); OS << Entry->Idx << ".\n"; if (isSplat(Entry->Scalars)) OS << " "; for (auto *V : Entry->Scalars) { OS << *V; if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; })) OS << " "; OS << "\n"; } return Str; } static std::string getNodeAttributes(const TreeEntry *Entry, const BoUpSLP *) { if (Entry->State == TreeEntry::NeedToGather) return "color=red"; if (Entry->State == TreeEntry::ScatterVectorize) return "color=blue"; return ""; } }; } // end namespace llvm BoUpSLP::~BoUpSLP() { SmallVector DeadInsts; for (auto *I : DeletedInstructions) { for (Use &U : I->operands()) { auto *Op = dyn_cast(U.get()); if (Op && !DeletedInstructions.count(Op) && Op->hasOneUser() && wouldInstructionBeTriviallyDead(Op, TLI)) DeadInsts.emplace_back(Op); } I->dropAllReferences(); } for (auto *I : DeletedInstructions) { assert(I->use_empty() && "trying to erase instruction with users."); I->eraseFromParent(); } // Cleanup any dead scalar code feeding the vectorized instructions RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI); #ifdef EXPENSIVE_CHECKS // If we could guarantee that this call is not extremely slow, we could // remove the ifdef limitation (see PR47712). assert(!verifyFunction(*F, &dbgs())); #endif } /// Reorders the given \p Reuses mask according to the given \p Mask. \p Reuses /// contains original mask for the scalars reused in the node. Procedure /// transform this mask in accordance with the given \p Mask. static void reorderReuses(SmallVectorImpl &Reuses, ArrayRef Mask) { assert(!Mask.empty() && Reuses.size() == Mask.size() && "Expected non-empty mask."); SmallVector Prev(Reuses.begin(), Reuses.end()); Prev.swap(Reuses); for (unsigned I = 0, E = Prev.size(); I < E; ++I) if (Mask[I] != UndefMaskElem) Reuses[Mask[I]] = Prev[I]; } /// Reorders the given \p Order according to the given \p Mask. \p Order - is /// the original order of the scalars. Procedure transforms the provided order /// in accordance with the given \p Mask. If the resulting \p Order is just an /// identity order, \p Order is cleared. static void reorderOrder(SmallVectorImpl &Order, ArrayRef Mask) { assert(!Mask.empty() && "Expected non-empty mask."); SmallVector MaskOrder; if (Order.empty()) { MaskOrder.resize(Mask.size()); std::iota(MaskOrder.begin(), MaskOrder.end(), 0); } else { inversePermutation(Order, MaskOrder); } reorderReuses(MaskOrder, Mask); if (ShuffleVectorInst::isIdentityMask(MaskOrder)) { Order.clear(); return; } Order.assign(Mask.size(), Mask.size()); for (unsigned I = 0, E = Mask.size(); I < E; ++I) if (MaskOrder[I] != UndefMaskElem) Order[MaskOrder[I]] = I; fixupOrderingIndices(Order); } std::optional BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE) { assert(TE.State == TreeEntry::NeedToGather && "Expected gather node only."); unsigned NumScalars = TE.Scalars.size(); OrdersType CurrentOrder(NumScalars, NumScalars); SmallVector Positions; SmallBitVector UsedPositions(NumScalars); const TreeEntry *STE = nullptr; // Try to find all gathered scalars that are gets vectorized in other // vectorize node. Here we can have only one single tree vector node to // correctly identify order of the gathered scalars. for (unsigned I = 0; I < NumScalars; ++I) { Value *V = TE.Scalars[I]; if (!isa(V)) continue; if (const auto *LocalSTE = getTreeEntry(V)) { if (!STE) STE = LocalSTE; else if (STE != LocalSTE) // Take the order only from the single vector node. return std::nullopt; unsigned Lane = std::distance(STE->Scalars.begin(), find(STE->Scalars, V)); if (Lane >= NumScalars) return std::nullopt; if (CurrentOrder[Lane] != NumScalars) { if (Lane != I) continue; UsedPositions.reset(CurrentOrder[Lane]); } // The partial identity (where only some elements of the gather node are // in the identity order) is good. CurrentOrder[Lane] = I; UsedPositions.set(I); } } // Need to keep the order if we have a vector entry and at least 2 scalars or // the vectorized entry has just 2 scalars. if (STE && (UsedPositions.count() > 1 || STE->Scalars.size() == 2)) { auto &&IsIdentityOrder = [NumScalars](ArrayRef CurrentOrder) { for (unsigned I = 0; I < NumScalars; ++I) if (CurrentOrder[I] != I && CurrentOrder[I] != NumScalars) return false; return true; }; if (IsIdentityOrder(CurrentOrder)) { CurrentOrder.clear(); return CurrentOrder; } auto *It = CurrentOrder.begin(); for (unsigned I = 0; I < NumScalars;) { if (UsedPositions.test(I)) { ++I; continue; } if (*It == NumScalars) { *It = I; ++I; } ++It; } return CurrentOrder; } return std::nullopt; } namespace { /// Tracks the state we can represent the loads in the given sequence. enum class LoadsState { Gather, Vectorize, ScatterVectorize }; } // anonymous namespace static bool arePointersCompatible(Value *Ptr1, Value *Ptr2, const TargetLibraryInfo &TLI, bool CompareOpcodes = true) { if (getUnderlyingObject(Ptr1) != getUnderlyingObject(Ptr2)) return false; auto *GEP1 = dyn_cast(Ptr1); if (!GEP1) return false; auto *GEP2 = dyn_cast(Ptr2); if (!GEP2) return false; return GEP1->getNumOperands() == 2 && GEP2->getNumOperands() == 2 && ((isConstant(GEP1->getOperand(1)) && isConstant(GEP2->getOperand(1))) || !CompareOpcodes || getSameOpcode({GEP1->getOperand(1), GEP2->getOperand(1)}, TLI) .getOpcode()); } /// Checks if the given array of loads can be represented as a vectorized, /// scatter or just simple gather. static LoadsState canVectorizeLoads(ArrayRef VL, const Value *VL0, const TargetTransformInfo &TTI, const DataLayout &DL, ScalarEvolution &SE, LoopInfo &LI, const TargetLibraryInfo &TLI, SmallVectorImpl &Order, SmallVectorImpl &PointerOps) { // Check that a vectorized load would load the same memory as a scalar // load. For example, we don't want to 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. Type *ScalarTy = VL0->getType(); if (DL.getTypeSizeInBits(ScalarTy) != DL.getTypeAllocSizeInBits(ScalarTy)) return LoadsState::Gather; // Make sure all loads in the bundle are simple - we can't vectorize // atomic or volatile loads. PointerOps.clear(); PointerOps.resize(VL.size()); auto *POIter = PointerOps.begin(); for (Value *V : VL) { auto *L = cast(V); if (!L->isSimple()) return LoadsState::Gather; *POIter = L->getPointerOperand(); ++POIter; } Order.clear(); // Check the order of pointer operands or that all pointers are the same. bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order); if (IsSorted || all_of(PointerOps, [&](Value *P) { return arePointersCompatible(P, PointerOps.front(), TLI); })) { if (IsSorted) { Value *Ptr0; Value *PtrN; if (Order.empty()) { Ptr0 = PointerOps.front(); PtrN = PointerOps.back(); } else { Ptr0 = PointerOps[Order.front()]; PtrN = PointerOps[Order.back()]; } std::optional Diff = getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE); // Check that the sorted loads are consecutive. if (static_cast(*Diff) == VL.size() - 1) return LoadsState::Vectorize; } // TODO: need to improve analysis of the pointers, if not all of them are // GEPs or have > 2 operands, we end up with a gather node, which just // increases the cost. Loop *L = LI.getLoopFor(cast(VL0)->getParent()); bool ProfitableGatherPointers = static_cast(count_if(PointerOps, [L](Value *V) { return L && L->isLoopInvariant(V); })) <= VL.size() / 2 && VL.size() > 2; if (ProfitableGatherPointers || all_of(PointerOps, [IsSorted](Value *P) { auto *GEP = dyn_cast(P); return (IsSorted && !GEP && doesNotNeedToBeScheduled(P)) || (GEP && GEP->getNumOperands() == 2); })) { Align CommonAlignment = cast(VL0)->getAlign(); for (Value *V : VL) CommonAlignment = std::min(CommonAlignment, cast(V)->getAlign()); auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); if (TTI.isLegalMaskedGather(VecTy, CommonAlignment) && !TTI.forceScalarizeMaskedGather(VecTy, CommonAlignment)) return LoadsState::ScatterVectorize; } } return LoadsState::Gather; } bool clusterSortPtrAccesses(ArrayRef VL, Type *ElemTy, const DataLayout &DL, ScalarEvolution &SE, SmallVectorImpl &SortedIndices) { assert(llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands."); // Map from bases to a vector of (Ptr, Offset, OrigIdx), which we insert each // Ptr into, sort and return the sorted indices with values next to one // another. MapVector>> Bases; Bases[VL[0]].push_back(std::make_tuple(VL[0], 0U, 0U)); unsigned Cnt = 1; for (Value *Ptr : VL.drop_front()) { bool Found = any_of(Bases, [&](auto &Base) { std::optional Diff = getPointersDiff(ElemTy, Base.first, ElemTy, Ptr, DL, SE, /*StrictCheck=*/true); if (!Diff) return false; Base.second.emplace_back(Ptr, *Diff, Cnt++); return true; }); if (!Found) { // If we haven't found enough to usefully cluster, return early. if (Bases.size() > VL.size() / 2 - 1) return false; // Not found already - add a new Base Bases[Ptr].emplace_back(Ptr, 0, Cnt++); } } // For each of the bases sort the pointers by Offset and check if any of the // base become consecutively allocated. bool AnyConsecutive = false; for (auto &Base : Bases) { auto &Vec = Base.second; if (Vec.size() > 1) { llvm::stable_sort(Vec, [](const std::tuple &X, const std::tuple &Y) { return std::get<1>(X) < std::get<1>(Y); }); int InitialOffset = std::get<1>(Vec[0]); AnyConsecutive |= all_of(enumerate(Vec), [InitialOffset](auto &P) { return std::get<1>(P.value()) == int(P.index()) + InitialOffset; }); } } // Fill SortedIndices array only if it looks worth-while to sort the ptrs. SortedIndices.clear(); if (!AnyConsecutive) return false; for (auto &Base : Bases) { for (auto &T : Base.second) SortedIndices.push_back(std::get<2>(T)); } assert(SortedIndices.size() == VL.size() && "Expected SortedIndices to be the size of VL"); return true; } std::optional BoUpSLP::findPartiallyOrderedLoads(const BoUpSLP::TreeEntry &TE) { assert(TE.State == TreeEntry::NeedToGather && "Expected gather node only."); Type *ScalarTy = TE.Scalars[0]->getType(); SmallVector Ptrs; Ptrs.reserve(TE.Scalars.size()); for (Value *V : TE.Scalars) { auto *L = dyn_cast(V); if (!L || !L->isSimple()) return std::nullopt; Ptrs.push_back(L->getPointerOperand()); } BoUpSLP::OrdersType Order; if (clusterSortPtrAccesses(Ptrs, ScalarTy, *DL, *SE, Order)) return Order; return std::nullopt; } /// Check if two insertelement instructions are from the same buildvector. static bool areTwoInsertFromSameBuildVector( InsertElementInst *VU, InsertElementInst *V, function_ref GetBaseOperand) { // Instructions must be from the same basic blocks. if (VU->getParent() != V->getParent()) return false; // Checks if 2 insertelements are from the same buildvector. if (VU->getType() != V->getType()) return false; // Multiple used inserts are separate nodes. if (!VU->hasOneUse() && !V->hasOneUse()) return false; auto *IE1 = VU; auto *IE2 = V; std::optional Idx1 = getInsertIndex(IE1); std::optional Idx2 = getInsertIndex(IE2); if (Idx1 == std::nullopt || Idx2 == std::nullopt) return false; // Go through the vector operand of insertelement instructions trying to find // either VU as the original vector for IE2 or V as the original vector for // IE1. do { if (IE2 == VU) return VU->hasOneUse(); if (IE1 == V) return V->hasOneUse(); if (IE1) { if ((IE1 != VU && !IE1->hasOneUse()) || getInsertIndex(IE1).value_or(*Idx2) == *Idx2) IE1 = nullptr; else IE1 = dyn_cast_or_null(GetBaseOperand(IE1)); } if (IE2) { if ((IE2 != V && !IE2->hasOneUse()) || getInsertIndex(IE2).value_or(*Idx1) == *Idx1) IE2 = nullptr; else IE2 = dyn_cast_or_null(GetBaseOperand(IE2)); } } while (IE1 || IE2); return false; } std::optional BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom) { // No need to reorder if need to shuffle reuses, still need to shuffle the // node. if (!TE.ReuseShuffleIndices.empty()) { // Check if reuse shuffle indices can be improved by reordering. // For this, check that reuse mask is "clustered", i.e. each scalar values // is used once in each submask of size . // Example: 4 scalar values. // ReuseShuffleIndices mask: 0, 1, 2, 3, 3, 2, 0, 1 - clustered. // 0, 1, 2, 3, 3, 3, 1, 0 - not clustered, because // element 3 is used twice in the second submask. unsigned Sz = TE.Scalars.size(); if (!ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, Sz)) return std::nullopt; unsigned VF = TE.getVectorFactor(); // Try build correct order for extractelement instructions. SmallVector ReusedMask(TE.ReuseShuffleIndices.begin(), TE.ReuseShuffleIndices.end()); if (TE.getOpcode() == Instruction::ExtractElement && !TE.isAltShuffle() && all_of(TE.Scalars, [Sz](Value *V) { std::optional Idx = getExtractIndex(cast(V)); return Idx && *Idx < Sz; })) { SmallVector ReorderMask(Sz, UndefMaskElem); if (TE.ReorderIndices.empty()) std::iota(ReorderMask.begin(), ReorderMask.end(), 0); else inversePermutation(TE.ReorderIndices, ReorderMask); for (unsigned I = 0; I < VF; ++I) { int &Idx = ReusedMask[I]; if (Idx == UndefMaskElem) continue; Value *V = TE.Scalars[ReorderMask[Idx]]; std::optional EI = getExtractIndex(cast(V)); Idx = std::distance(ReorderMask.begin(), find(ReorderMask, *EI)); } } // Build the order of the VF size, need to reorder reuses shuffles, they are // always of VF size. OrdersType ResOrder(VF); std::iota(ResOrder.begin(), ResOrder.end(), 0); auto *It = ResOrder.begin(); for (unsigned K = 0; K < VF; K += Sz) { OrdersType CurrentOrder(TE.ReorderIndices); SmallVector SubMask{ArrayRef(ReusedMask).slice(K, Sz)}; if (SubMask.front() == UndefMaskElem) std::iota(SubMask.begin(), SubMask.end(), 0); reorderOrder(CurrentOrder, SubMask); transform(CurrentOrder, It, [K](unsigned Pos) { return Pos + K; }); std::advance(It, Sz); } if (all_of(enumerate(ResOrder), [](const auto &Data) { return Data.index() == Data.value(); })) return {}; // Use identity order. return ResOrder; } if (TE.State == TreeEntry::Vectorize && (isa(TE.getMainOp()) || (TopToBottom && isa(TE.getMainOp()))) && !TE.isAltShuffle()) return TE.ReorderIndices; if (TE.State == TreeEntry::Vectorize && TE.getOpcode() == Instruction::PHI) { auto PHICompare = [](llvm::Value *V1, llvm::Value *V2) { if (!V1->hasOneUse() || !V2->hasOneUse()) return false; auto *FirstUserOfPhi1 = cast(*V1->user_begin()); auto *FirstUserOfPhi2 = cast(*V2->user_begin()); if (auto *IE1 = dyn_cast(FirstUserOfPhi1)) if (auto *IE2 = dyn_cast(FirstUserOfPhi2)) { if (!areTwoInsertFromSameBuildVector( IE1, IE2, [](InsertElementInst *II) { return II->getOperand(0); })) return false; std::optional Idx1 = getInsertIndex(IE1); std::optional Idx2 = getInsertIndex(IE2); if (Idx1 == std::nullopt || Idx2 == std::nullopt) return false; return *Idx1 < *Idx2; } if (auto *EE1 = dyn_cast(FirstUserOfPhi1)) if (auto *EE2 = dyn_cast(FirstUserOfPhi2)) { if (EE1->getOperand(0) != EE2->getOperand(0)) return false; std::optional Idx1 = getExtractIndex(EE1); std::optional Idx2 = getExtractIndex(EE2); if (Idx1 == std::nullopt || Idx2 == std::nullopt) return false; return *Idx1 < *Idx2; } return false; }; auto IsIdentityOrder = [](const OrdersType &Order) { for (unsigned Idx : seq(0, Order.size())) if (Idx != Order[Idx]) return false; return true; }; if (!TE.ReorderIndices.empty()) return TE.ReorderIndices; DenseMap PhiToId; SmallVector Phis; OrdersType ResOrder(TE.Scalars.size()); for (unsigned Id = 0, Sz = TE.Scalars.size(); Id < Sz; ++Id) { PhiToId[TE.Scalars[Id]] = Id; Phis.push_back(TE.Scalars[Id]); } llvm::stable_sort(Phis, PHICompare); for (unsigned Id = 0, Sz = Phis.size(); Id < Sz; ++Id) ResOrder[Id] = PhiToId[Phis[Id]]; if (IsIdentityOrder(ResOrder)) return {}; return ResOrder; } if (TE.State == TreeEntry::NeedToGather) { // TODO: add analysis of other gather nodes with extractelement // instructions and other values/instructions, not only undefs. if (((TE.getOpcode() == Instruction::ExtractElement && !TE.isAltShuffle()) || (all_of(TE.Scalars, [](Value *V) { return isa(V); }) && any_of(TE.Scalars, [](Value *V) { return isa(V); }))) && all_of(TE.Scalars, [](Value *V) { auto *EE = dyn_cast(V); return !EE || isa(EE->getVectorOperandType()); }) && allSameType(TE.Scalars)) { // Check that gather of extractelements can be represented as // just a shuffle of a single vector. OrdersType CurrentOrder; bool Reuse = canReuseExtract(TE.Scalars, TE.getMainOp(), CurrentOrder); if (Reuse || !CurrentOrder.empty()) { if (!CurrentOrder.empty()) fixupOrderingIndices(CurrentOrder); return CurrentOrder; } } if (std::optional CurrentOrder = findReusedOrderedScalars(TE)) return CurrentOrder; if (TE.Scalars.size() >= 4) if (std::optional Order = findPartiallyOrderedLoads(TE)) return Order; } return std::nullopt; } /// Checks if the given mask is a "clustered" mask with the same clusters of /// size \p Sz, which are not identity submasks. static bool isRepeatedNonIdentityClusteredMask(ArrayRef Mask, unsigned Sz) { ArrayRef FirstCluster = Mask.slice(0, Sz); if (ShuffleVectorInst::isIdentityMask(FirstCluster)) return false; for (unsigned I = Sz, E = Mask.size(); I < E; I += Sz) { ArrayRef Cluster = Mask.slice(I, Sz); if (Cluster != FirstCluster) return false; } return true; } void BoUpSLP::reorderNodeWithReuses(TreeEntry &TE, ArrayRef Mask) const { // Reorder reuses mask. reorderReuses(TE.ReuseShuffleIndices, Mask); const unsigned Sz = TE.Scalars.size(); // For vectorized and non-clustered reused no need to do anything else. if (TE.State != TreeEntry::NeedToGather || !ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, Sz) || !isRepeatedNonIdentityClusteredMask(TE.ReuseShuffleIndices, Sz)) return; SmallVector NewMask; inversePermutation(TE.ReorderIndices, NewMask); addMask(NewMask, TE.ReuseShuffleIndices); // Clear reorder since it is going to be applied to the new mask. TE.ReorderIndices.clear(); // Try to improve gathered nodes with clustered reuses, if possible. ArrayRef Slice = ArrayRef(NewMask).slice(0, Sz); SmallVector NewOrder(Slice.begin(), Slice.end()); inversePermutation(NewOrder, NewMask); reorderScalars(TE.Scalars, NewMask); // Fill the reuses mask with the identity submasks. for (auto *It = TE.ReuseShuffleIndices.begin(), *End = TE.ReuseShuffleIndices.end(); It != End; std::advance(It, Sz)) std::iota(It, std::next(It, Sz), 0); } void BoUpSLP::reorderTopToBottom() { // Maps VF to the graph nodes. DenseMap> VFToOrderedEntries; // ExtractElement gather nodes which can be vectorized and need to handle // their ordering. DenseMap GathersToOrders; // Phi nodes can have preferred ordering based on their result users DenseMap PhisToOrders; // AltShuffles can also have a preferred ordering that leads to fewer // instructions, e.g., the addsub instruction in x86. DenseMap AltShufflesToOrders; // Maps a TreeEntry to the reorder indices of external users. DenseMap> ExternalUserReorderMap; // FIXME: Workaround for syntax error reported by MSVC buildbots. TargetTransformInfo &TTIRef = *TTI; // Find all reorderable nodes with the given VF. // Currently the are vectorized stores,loads,extracts + some gathering of // extracts. for_each(VectorizableTree, [this, &TTIRef, &VFToOrderedEntries, &GathersToOrders, &ExternalUserReorderMap, &AltShufflesToOrders, &PhisToOrders]( const std::unique_ptr &TE) { // Look for external users that will probably be vectorized. SmallVector ExternalUserReorderIndices = findExternalStoreUsersReorderIndices(TE.get()); if (!ExternalUserReorderIndices.empty()) { VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); ExternalUserReorderMap.try_emplace(TE.get(), std::move(ExternalUserReorderIndices)); } // Patterns like [fadd,fsub] can be combined into a single instruction in // x86. Reordering them into [fsub,fadd] blocks this pattern. So we need // to take into account their order when looking for the most used order. if (TE->isAltShuffle()) { VectorType *VecTy = FixedVectorType::get(TE->Scalars[0]->getType(), TE->Scalars.size()); unsigned Opcode0 = TE->getOpcode(); unsigned Opcode1 = TE->getAltOpcode(); // The opcode mask selects between the two opcodes. SmallBitVector OpcodeMask(TE->Scalars.size(), false); for (unsigned Lane : seq(0, TE->Scalars.size())) if (cast(TE->Scalars[Lane])->getOpcode() == Opcode1) OpcodeMask.set(Lane); // If this pattern is supported by the target then we consider the order. if (TTIRef.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask)) { VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); AltShufflesToOrders.try_emplace(TE.get(), OrdersType()); } // TODO: Check the reverse order too. } if (std::optional CurrentOrder = getReorderingData(*TE, /*TopToBottom=*/true)) { // Do not include ordering for nodes used in the alt opcode vectorization, // better to reorder them during bottom-to-top stage. If follow the order // here, it causes reordering of the whole graph though actually it is // profitable just to reorder the subgraph that starts from the alternate // opcode vectorization node. Such nodes already end-up with the shuffle // instruction and it is just enough to change this shuffle rather than // rotate the scalars for the whole graph. unsigned Cnt = 0; const TreeEntry *UserTE = TE.get(); while (UserTE && Cnt < RecursionMaxDepth) { if (UserTE->UserTreeIndices.size() != 1) break; if (all_of(UserTE->UserTreeIndices, [](const EdgeInfo &EI) { return EI.UserTE->State == TreeEntry::Vectorize && EI.UserTE->isAltShuffle() && EI.UserTE->Idx != 0; })) return; UserTE = UserTE->UserTreeIndices.back().UserTE; ++Cnt; } VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) GathersToOrders.try_emplace(TE.get(), *CurrentOrder); if (TE->State == TreeEntry::Vectorize && TE->getOpcode() == Instruction::PHI) PhisToOrders.try_emplace(TE.get(), *CurrentOrder); } }); // Reorder the graph nodes according to their vectorization factor. for (unsigned VF = VectorizableTree.front()->getVectorFactor(); VF > 1; VF /= 2) { auto It = VFToOrderedEntries.find(VF); if (It == VFToOrderedEntries.end()) continue; // Try to find the most profitable order. We just are looking for the most // used order and reorder scalar elements in the nodes according to this // mostly used order. ArrayRef OrderedEntries = It->second.getArrayRef(); // All operands are reordered and used only in this node - propagate the // most used order to the user node. MapVector> OrdersUses; SmallPtrSet VisitedOps; for (const TreeEntry *OpTE : OrderedEntries) { // No need to reorder this nodes, still need to extend and to use shuffle, // just need to merge reordering shuffle and the reuse shuffle. if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) continue; // Count number of orders uses. const auto &Order = [OpTE, &GathersToOrders, &AltShufflesToOrders, &PhisToOrders]() -> const OrdersType & { if (OpTE->State == TreeEntry::NeedToGather || !OpTE->ReuseShuffleIndices.empty()) { auto It = GathersToOrders.find(OpTE); if (It != GathersToOrders.end()) return It->second; } if (OpTE->isAltShuffle()) { auto It = AltShufflesToOrders.find(OpTE); if (It != AltShufflesToOrders.end()) return It->second; } if (OpTE->State == TreeEntry::Vectorize && OpTE->getOpcode() == Instruction::PHI) { auto It = PhisToOrders.find(OpTE); if (It != PhisToOrders.end()) return It->second; } return OpTE->ReorderIndices; }(); // First consider the order of the external scalar users. auto It = ExternalUserReorderMap.find(OpTE); if (It != ExternalUserReorderMap.end()) { const auto &ExternalUserReorderIndices = It->second; // If the OpTE vector factor != number of scalars - use natural order, // it is an attempt to reorder node with reused scalars but with // external uses. if (OpTE->getVectorFactor() != OpTE->Scalars.size()) { OrdersUses.insert(std::make_pair(OrdersType(), 0)).first->second += ExternalUserReorderIndices.size(); } else { for (const OrdersType &ExtOrder : ExternalUserReorderIndices) ++OrdersUses.insert(std::make_pair(ExtOrder, 0)).first->second; } // No other useful reorder data in this entry. if (Order.empty()) continue; } // Stores actually store the mask, not the order, need to invert. if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && OpTE->getOpcode() == Instruction::Store && !Order.empty()) { SmallVector Mask; inversePermutation(Order, Mask); unsigned E = Order.size(); OrdersType CurrentOrder(E, E); transform(Mask, CurrentOrder.begin(), [E](int Idx) { return Idx == UndefMaskElem ? E : static_cast(Idx); }); fixupOrderingIndices(CurrentOrder); ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second; } else { ++OrdersUses.insert(std::make_pair(Order, 0)).first->second; } } // Set order of the user node. if (OrdersUses.empty()) continue; // Choose the most used order. ArrayRef BestOrder = OrdersUses.front().first; unsigned Cnt = OrdersUses.front().second; for (const auto &Pair : drop_begin(OrdersUses)) { if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { BestOrder = Pair.first; Cnt = Pair.second; } } // Set order of the user node. if (BestOrder.empty()) continue; SmallVector Mask; inversePermutation(BestOrder, Mask); SmallVector MaskOrder(BestOrder.size(), UndefMaskElem); unsigned E = BestOrder.size(); transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { return I < E ? static_cast(I) : UndefMaskElem; }); // Do an actual reordering, if profitable. for (std::unique_ptr &TE : VectorizableTree) { // Just do the reordering for the nodes with the given VF. if (TE->Scalars.size() != VF) { if (TE->ReuseShuffleIndices.size() == VF) { // Need to reorder the reuses masks of the operands with smaller VF to // be able to find the match between the graph nodes and scalar // operands of the given node during vectorization/cost estimation. assert(all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && "All users must be of VF size."); // Update ordering of the operands with the smaller VF than the given // one. reorderNodeWithReuses(*TE, Mask); } continue; } if (TE->State == TreeEntry::Vectorize && isa(TE->getMainOp()) && !TE->isAltShuffle()) { // Build correct orders for extract{element,value}, loads and // stores. reorderOrder(TE->ReorderIndices, Mask); if (isa(TE->getMainOp())) TE->reorderOperands(Mask); } else { // Reorder the node and its operands. TE->reorderOperands(Mask); assert(TE->ReorderIndices.empty() && "Expected empty reorder sequence."); reorderScalars(TE->Scalars, Mask); } if (!TE->ReuseShuffleIndices.empty()) { // Apply reversed order to keep the original ordering of the reused // elements to avoid extra reorder indices shuffling. OrdersType CurrentOrder; reorderOrder(CurrentOrder, MaskOrder); SmallVector NewReuses; inversePermutation(CurrentOrder, NewReuses); addMask(NewReuses, TE->ReuseShuffleIndices); TE->ReuseShuffleIndices.swap(NewReuses); } } } } bool BoUpSLP::canReorderOperands( TreeEntry *UserTE, SmallVectorImpl> &Edges, ArrayRef ReorderableGathers, SmallVectorImpl &GatherOps) { for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) { if (any_of(Edges, [I](const std::pair &OpData) { return OpData.first == I && OpData.second->State == TreeEntry::Vectorize; })) continue; if (TreeEntry *TE = getVectorizedOperand(UserTE, I)) { // Do not reorder if operand node is used by many user nodes. if (any_of(TE->UserTreeIndices, [UserTE](const EdgeInfo &EI) { return EI.UserTE != UserTE; })) return false; // Add the node to the list of the ordered nodes with the identity // order. Edges.emplace_back(I, TE); // Add ScatterVectorize nodes to the list of operands, where just // reordering of the scalars is required. Similar to the gathers, so // simply add to the list of gathered ops. // If there are reused scalars, process this node as a regular vectorize // node, just reorder reuses mask. if (TE->State != TreeEntry::Vectorize && TE->ReuseShuffleIndices.empty()) GatherOps.push_back(TE); continue; } TreeEntry *Gather = nullptr; if (count_if(ReorderableGathers, [&Gather, UserTE, I](TreeEntry *TE) { assert(TE->State != TreeEntry::Vectorize && "Only non-vectorized nodes are expected."); if (any_of(TE->UserTreeIndices, [UserTE, I](const EdgeInfo &EI) { return EI.UserTE == UserTE && EI.EdgeIdx == I; })) { assert(TE->isSame(UserTE->getOperand(I)) && "Operand entry does not match operands."); Gather = TE; return true; } return false; }) > 1 && !all_of(UserTE->getOperand(I), isConstant)) return false; if (Gather) GatherOps.push_back(Gather); } return true; } void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) { SetVector OrderedEntries; DenseMap GathersToOrders; // Find all reorderable leaf nodes with the given VF. // Currently the are vectorized loads,extracts without alternate operands + // some gathering of extracts. SmallVector NonVectorized; for_each(VectorizableTree, [this, &OrderedEntries, &GathersToOrders, &NonVectorized]( const std::unique_ptr &TE) { if (TE->State != TreeEntry::Vectorize) NonVectorized.push_back(TE.get()); if (std::optional CurrentOrder = getReorderingData(*TE, /*TopToBottom=*/false)) { OrderedEntries.insert(TE.get()); if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) GathersToOrders.try_emplace(TE.get(), *CurrentOrder); } }); // 1. Propagate order to the graph nodes, which use only reordered nodes. // I.e., if the node has operands, that are reordered, try to make at least // one operand order in the natural order and reorder others + reorder the // user node itself. SmallPtrSet Visited; while (!OrderedEntries.empty()) { // 1. Filter out only reordered nodes. // 2. If the entry has multiple uses - skip it and jump to the next node. DenseMap>> Users; SmallVector Filtered; for (TreeEntry *TE : OrderedEntries) { if (!(TE->State == TreeEntry::Vectorize || (TE->State == TreeEntry::NeedToGather && GathersToOrders.count(TE))) || TE->UserTreeIndices.empty() || !TE->ReuseShuffleIndices.empty() || !all_of(drop_begin(TE->UserTreeIndices), [TE](const EdgeInfo &EI) { return EI.UserTE == TE->UserTreeIndices.front().UserTE; }) || !Visited.insert(TE).second) { Filtered.push_back(TE); continue; } // Build a map between user nodes and their operands order to speedup // search. The graph currently does not provide this dependency directly. for (EdgeInfo &EI : TE->UserTreeIndices) { TreeEntry *UserTE = EI.UserTE; auto It = Users.find(UserTE); if (It == Users.end()) It = Users.insert({UserTE, {}}).first; It->second.emplace_back(EI.EdgeIdx, TE); } } // Erase filtered entries. for_each(Filtered, [&OrderedEntries](TreeEntry *TE) { OrderedEntries.remove(TE); }); SmallVector< std::pair>>> UsersVec(Users.begin(), Users.end()); sort(UsersVec, [](const auto &Data1, const auto &Data2) { return Data1.first->Idx > Data2.first->Idx; }); for (auto &Data : UsersVec) { // Check that operands are used only in the User node. SmallVector GatherOps; if (!canReorderOperands(Data.first, Data.second, NonVectorized, GatherOps)) { for_each(Data.second, [&OrderedEntries](const std::pair &Op) { OrderedEntries.remove(Op.second); }); continue; } // All operands are reordered and used only in this node - propagate the // most used order to the user node. MapVector> OrdersUses; // Do the analysis for each tree entry only once, otherwise the order of // the same node my be considered several times, though might be not // profitable. SmallPtrSet VisitedOps; SmallPtrSet VisitedUsers; for (const auto &Op : Data.second) { TreeEntry *OpTE = Op.second; if (!VisitedOps.insert(OpTE).second) continue; if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) continue; const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & { if (OpTE->State == TreeEntry::NeedToGather || !OpTE->ReuseShuffleIndices.empty()) return GathersToOrders.find(OpTE)->second; return OpTE->ReorderIndices; }(); unsigned NumOps = count_if( Data.second, [OpTE](const std::pair &P) { return P.second == OpTE; }); // Stores actually store the mask, not the order, need to invert. if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && OpTE->getOpcode() == Instruction::Store && !Order.empty()) { SmallVector Mask; inversePermutation(Order, Mask); unsigned E = Order.size(); OrdersType CurrentOrder(E, E); transform(Mask, CurrentOrder.begin(), [E](int Idx) { return Idx == UndefMaskElem ? E : static_cast(Idx); }); fixupOrderingIndices(CurrentOrder); OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second += NumOps; } else { OrdersUses.insert(std::make_pair(Order, 0)).first->second += NumOps; } auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0)); const auto &&AllowsReordering = [IgnoreReorder, &GathersToOrders]( const TreeEntry *TE) { if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() || (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) || (IgnoreReorder && TE->Idx == 0)) return true; if (TE->State == TreeEntry::NeedToGather) { auto It = GathersToOrders.find(TE); if (It != GathersToOrders.end()) return !It->second.empty(); return true; } return false; }; for (const EdgeInfo &EI : OpTE->UserTreeIndices) { TreeEntry *UserTE = EI.UserTE; if (!VisitedUsers.insert(UserTE).second) continue; // May reorder user node if it requires reordering, has reused // scalars, is an alternate op vectorize node or its op nodes require // reordering. if (AllowsReordering(UserTE)) continue; // Check if users allow reordering. // Currently look up just 1 level of operands to avoid increase of // the compile time. // Profitable to reorder if definitely more operands allow // reordering rather than those with natural order. ArrayRef> Ops = Users[UserTE]; if (static_cast(count_if( Ops, [UserTE, &AllowsReordering]( const std::pair &Op) { return AllowsReordering(Op.second) && all_of(Op.second->UserTreeIndices, [UserTE](const EdgeInfo &EI) { return EI.UserTE == UserTE; }); })) <= Ops.size() / 2) ++Res.first->second; } } // If no orders - skip current nodes and jump to the next one, if any. if (OrdersUses.empty()) { for_each(Data.second, [&OrderedEntries](const std::pair &Op) { OrderedEntries.remove(Op.second); }); continue; } // Choose the best order. ArrayRef BestOrder = OrdersUses.front().first; unsigned Cnt = OrdersUses.front().second; for (const auto &Pair : drop_begin(OrdersUses)) { if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { BestOrder = Pair.first; Cnt = Pair.second; } } // Set order of the user node (reordering of operands and user nodes). if (BestOrder.empty()) { for_each(Data.second, [&OrderedEntries](const std::pair &Op) { OrderedEntries.remove(Op.second); }); continue; } // Erase operands from OrderedEntries list and adjust their orders. VisitedOps.clear(); SmallVector Mask; inversePermutation(BestOrder, Mask); SmallVector MaskOrder(BestOrder.size(), UndefMaskElem); unsigned E = BestOrder.size(); transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { return I < E ? static_cast(I) : UndefMaskElem; }); for (const std::pair &Op : Data.second) { TreeEntry *TE = Op.second; OrderedEntries.remove(TE); if (!VisitedOps.insert(TE).second) continue; if (TE->ReuseShuffleIndices.size() == BestOrder.size()) { reorderNodeWithReuses(*TE, Mask); continue; } // Gathers are processed separately. if (TE->State != TreeEntry::Vectorize) continue; assert((BestOrder.size() == TE->ReorderIndices.size() || TE->ReorderIndices.empty()) && "Non-matching sizes of user/operand entries."); reorderOrder(TE->ReorderIndices, Mask); if (IgnoreReorder && TE == VectorizableTree.front().get()) IgnoreReorder = false; } // For gathers just need to reorder its scalars. for (TreeEntry *Gather : GatherOps) { assert(Gather->ReorderIndices.empty() && "Unexpected reordering of gathers."); if (!Gather->ReuseShuffleIndices.empty()) { // Just reorder reuses indices. reorderReuses(Gather->ReuseShuffleIndices, Mask); continue; } reorderScalars(Gather->Scalars, Mask); OrderedEntries.remove(Gather); } // Reorder operands of the user node and set the ordering for the user // node itself. if (Data.first->State != TreeEntry::Vectorize || !isa( Data.first->getMainOp()) || Data.first->isAltShuffle()) Data.first->reorderOperands(Mask); if (!isa(Data.first->getMainOp()) || Data.first->isAltShuffle()) { reorderScalars(Data.first->Scalars, Mask); reorderOrder(Data.first->ReorderIndices, MaskOrder); if (Data.first->ReuseShuffleIndices.empty() && !Data.first->ReorderIndices.empty() && !Data.first->isAltShuffle()) { // Insert user node to the list to try to sink reordering deeper in // the graph. OrderedEntries.insert(Data.first); } } else { reorderOrder(Data.first->ReorderIndices, Mask); } } } // If the reordering is unnecessary, just remove the reorder. if (IgnoreReorder && !VectorizableTree.front()->ReorderIndices.empty() && VectorizableTree.front()->ReuseShuffleIndices.empty()) VectorizableTree.front()->ReorderIndices.clear(); } void BoUpSLP::buildExternalUses( const ExtraValueToDebugLocsMap &ExternallyUsedValues) { // Collect the values that we need to extract from the tree. for (auto &TEPtr : VectorizableTree) { TreeEntry *Entry = TEPtr.get(); // No need to handle users of gathered values. if (Entry->State == TreeEntry::NeedToGather) continue; // For each lane: for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { Value *Scalar = Entry->Scalars[Lane]; int FoundLane = Entry->findLaneForValue(Scalar); // Check if the scalar is externally used as an extra arg. auto ExtI = ExternallyUsedValues.find(Scalar); if (ExtI != ExternallyUsedValues.end()) { LLVM_DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n"); ExternalUses.emplace_back(Scalar, nullptr, FoundLane); } for (User *U : Scalar->users()) { LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n"); Instruction *UserInst = dyn_cast(U); if (!UserInst) continue; if (isDeleted(UserInst)) continue; // Skip in-tree scalars that become vectors if (TreeEntry *UseEntry = getTreeEntry(U)) { 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 || UseEntry->State == TreeEntry::ScatterVectorize || !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { LLVM_DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"); assert(UseEntry->State != TreeEntry::NeedToGather && "Bad state"); continue; } } // Ignore users in the user ignore list. if (UserIgnoreList && UserIgnoreList->contains(UserInst)) continue; LLVM_DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " << Lane << " from " << *Scalar << ".\n"); ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); } } } } DenseMap> BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const { DenseMap> PtrToStoresMap; for (unsigned Lane : seq(0, TE->Scalars.size())) { Value *V = TE->Scalars[Lane]; // To save compilation time we don't visit if we have too many users. static constexpr unsigned UsersLimit = 4; if (V->hasNUsesOrMore(UsersLimit)) break; // Collect stores per pointer object. for (User *U : V->users()) { auto *SI = dyn_cast(U); if (SI == nullptr || !SI->isSimple() || !isValidElementType(SI->getValueOperand()->getType())) continue; // Skip entry if already if (getTreeEntry(U)) continue; Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); auto &StoresVec = PtrToStoresMap[Ptr]; // For now just keep one store per pointer object per lane. // TODO: Extend this to support multiple stores per pointer per lane if (StoresVec.size() > Lane) continue; // Skip if in different BBs. if (!StoresVec.empty() && SI->getParent() != StoresVec.back()->getParent()) continue; // Make sure that the stores are of the same type. if (!StoresVec.empty() && SI->getValueOperand()->getType() != StoresVec.back()->getValueOperand()->getType()) continue; StoresVec.push_back(SI); } } return PtrToStoresMap; } bool BoUpSLP::canFormVector(const SmallVector &StoresVec, OrdersType &ReorderIndices) const { // We check whether the stores in StoreVec can form a vector by sorting them // and checking whether they are consecutive. // To avoid calling getPointersDiff() while sorting we create a vector of // pairs {store, offset from first} and sort this instead. SmallVector, 4> StoreOffsetVec(StoresVec.size()); StoreInst *S0 = StoresVec[0]; StoreOffsetVec[0] = {S0, 0}; Type *S0Ty = S0->getValueOperand()->getType(); Value *S0Ptr = S0->getPointerOperand(); for (unsigned Idx : seq(1, StoresVec.size())) { StoreInst *SI = StoresVec[Idx]; std::optional Diff = getPointersDiff(S0Ty, S0Ptr, SI->getValueOperand()->getType(), SI->getPointerOperand(), *DL, *SE, /*StrictCheck=*/true); // We failed to compare the pointers so just abandon this StoresVec. if (!Diff) return false; StoreOffsetVec[Idx] = {StoresVec[Idx], *Diff}; } // Sort the vector based on the pointers. We create a copy because we may // need the original later for calculating the reorder (shuffle) indices. stable_sort(StoreOffsetVec, [](const std::pair &Pair1, const std::pair &Pair2) { int Offset1 = Pair1.second; int Offset2 = Pair2.second; return Offset1 < Offset2; }); // Check if the stores are consecutive by checking if their difference is 1. for (unsigned Idx : seq(1, StoreOffsetVec.size())) if (StoreOffsetVec[Idx].second != StoreOffsetVec[Idx-1].second + 1) return false; // Calculate the shuffle indices according to their offset against the sorted // StoreOffsetVec. ReorderIndices.reserve(StoresVec.size()); for (StoreInst *SI : StoresVec) { unsigned Idx = find_if(StoreOffsetVec, [SI](const std::pair &Pair) { return Pair.first == SI; }) - StoreOffsetVec.begin(); ReorderIndices.push_back(Idx); } // Identity order (e.g., {0,1,2,3}) is modeled as an empty OrdersType in // reorderTopToBottom() and reorderBottomToTop(), so we are following the // same convention here. auto IsIdentityOrder = [](const OrdersType &Order) { for (unsigned Idx : seq(0, Order.size())) if (Idx != Order[Idx]) return false; return true; }; if (IsIdentityOrder(ReorderIndices)) ReorderIndices.clear(); return true; } #ifndef NDEBUG LLVM_DUMP_METHOD static void dumpOrder(const BoUpSLP::OrdersType &Order) { for (unsigned Idx : Order) dbgs() << Idx << ", "; dbgs() << "\n"; } #endif SmallVector BoUpSLP::findExternalStoreUsersReorderIndices(TreeEntry *TE) const { unsigned NumLanes = TE->Scalars.size(); DenseMap> PtrToStoresMap = collectUserStores(TE); // Holds the reorder indices for each candidate store vector that is a user of // the current TreeEntry. SmallVector ExternalReorderIndices; // Now inspect the stores collected per pointer and look for vectorization // candidates. For each candidate calculate the reorder index vector and push // it into `ExternalReorderIndices` for (const auto &Pair : PtrToStoresMap) { auto &StoresVec = Pair.second; // If we have fewer than NumLanes stores, then we can't form a vector. if (StoresVec.size() != NumLanes) continue; // If the stores are not consecutive then abandon this StoresVec. OrdersType ReorderIndices; if (!canFormVector(StoresVec, ReorderIndices)) continue; // We now know that the scalars in StoresVec can form a vector instruction, // so set the reorder indices. ExternalReorderIndices.push_back(ReorderIndices); } return ExternalReorderIndices; } void BoUpSLP::buildTree(ArrayRef Roots, const SmallDenseSet &UserIgnoreLst) { deleteTree(); UserIgnoreList = &UserIgnoreLst; if (!allSameType(Roots)) return; buildTree_rec(Roots, 0, EdgeInfo()); } void BoUpSLP::buildTree(ArrayRef Roots) { deleteTree(); if (!allSameType(Roots)) return; buildTree_rec(Roots, 0, EdgeInfo()); } /// \return true if the specified list of values has only one instruction that /// requires scheduling, false otherwise. #ifndef NDEBUG static bool needToScheduleSingleInstruction(ArrayRef VL) { Value *NeedsScheduling = nullptr; for (Value *V : VL) { if (doesNotNeedToBeScheduled(V)) continue; if (!NeedsScheduling) { NeedsScheduling = V; continue; } return false; } return NeedsScheduling; } #endif /// Generates key/subkey pair for the given value to provide effective sorting /// of the values and better detection of the vectorizable values sequences. The /// keys/subkeys can be used for better sorting of the values themselves (keys) /// and in values subgroups (subkeys). static std::pair generateKeySubkey( Value *V, const TargetLibraryInfo *TLI, function_ref LoadsSubkeyGenerator, bool AllowAlternate) { hash_code Key = hash_value(V->getValueID() + 2); hash_code SubKey = hash_value(0); // Sort the loads by the distance between the pointers. if (auto *LI = dyn_cast(V)) { Key = hash_combine(LI->getType(), hash_value(Instruction::Load), Key); if (LI->isSimple()) SubKey = hash_value(LoadsSubkeyGenerator(Key, LI)); else Key = SubKey = hash_value(LI); } else if (isVectorLikeInstWithConstOps(V)) { // Sort extracts by the vector operands. if (isa(V)) Key = hash_value(Value::UndefValueVal + 1); if (auto *EI = dyn_cast(V)) { if (!isUndefVector(EI->getVectorOperand()).all() && !isa(EI->getIndexOperand())) SubKey = hash_value(EI->getVectorOperand()); } } else if (auto *I = dyn_cast(V)) { // Sort other instructions just by the opcodes except for CMPInst. // For CMP also sort by the predicate kind. if ((isa(I)) && isValidForAlternation(I->getOpcode())) { if (AllowAlternate) Key = hash_value(isa(I) ? 1 : 0); else Key = hash_combine(hash_value(I->getOpcode()), Key); SubKey = hash_combine( hash_value(I->getOpcode()), hash_value(I->getType()), hash_value(isa(I) ? I->getType() : cast(I)->getOperand(0)->getType())); // For casts, look through the only operand to improve compile time. if (isa(I)) { std::pair OpVals = generateKeySubkey(I->getOperand(0), TLI, LoadsSubkeyGenerator, /*AllowAlternate=*/true); Key = hash_combine(OpVals.first, Key); SubKey = hash_combine(OpVals.first, SubKey); } } else if (auto *CI = dyn_cast(I)) { CmpInst::Predicate Pred = CI->getPredicate(); if (CI->isCommutative()) Pred = std::min(Pred, CmpInst::getInversePredicate(Pred)); CmpInst::Predicate SwapPred = CmpInst::getSwappedPredicate(Pred); SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Pred), hash_value(SwapPred), hash_value(CI->getOperand(0)->getType())); } else if (auto *Call = dyn_cast(I)) { Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, TLI); if (isTriviallyVectorizable(ID)) { SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(ID)); } else if (!VFDatabase(*Call).getMappings(*Call).empty()) { SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Call->getCalledFunction())); } else { Key = hash_combine(hash_value(Call), Key); SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Call)); } for (const CallBase::BundleOpInfo &Op : Call->bundle_op_infos()) SubKey = hash_combine(hash_value(Op.Begin), hash_value(Op.End), hash_value(Op.Tag), SubKey); } else if (auto *Gep = dyn_cast(I)) { if (Gep->getNumOperands() == 2 && isa(Gep->getOperand(1))) SubKey = hash_value(Gep->getPointerOperand()); else SubKey = hash_value(Gep); } else if (BinaryOperator::isIntDivRem(I->getOpcode()) && !isa(I->getOperand(1))) { // Do not try to vectorize instructions with potentially high cost. SubKey = hash_value(I); } else { SubKey = hash_value(I->getOpcode()); } Key = hash_combine(hash_value(I->getParent()), Key); } return std::make_pair(Key, SubKey); } /// Checks if the specified instruction \p I is an alternate operation for /// the given \p MainOp and \p AltOp instructions. static bool isAlternateInstruction(const Instruction *I, const Instruction *MainOp, const Instruction *AltOp, const TargetLibraryInfo &TLI); void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth, const EdgeInfo &UserTreeIdx) { assert((allConstant(VL) || allSameType(VL)) && "Invalid types!"); SmallVector ReuseShuffleIndicies; SmallVector UniqueValues; auto &&TryToFindDuplicates = [&VL, &ReuseShuffleIndicies, &UniqueValues, &UserTreeIdx, this](const InstructionsState &S) { // Check that every instruction appears once in this bundle. DenseMap UniquePositions(VL.size()); for (Value *V : VL) { if (isConstant(V)) { ReuseShuffleIndicies.emplace_back( isa(V) ? UndefMaskElem : UniqueValues.size()); UniqueValues.emplace_back(V); continue; } auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); ReuseShuffleIndicies.emplace_back(Res.first->second); if (Res.second) UniqueValues.emplace_back(V); } size_t NumUniqueScalarValues = UniqueValues.size(); if (NumUniqueScalarValues == VL.size()) { ReuseShuffleIndicies.clear(); } else { LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n"); if (NumUniqueScalarValues <= 1 || (UniquePositions.size() == 1 && all_of(UniqueValues, [](Value *V) { return isa(V) || !isConstant(V); })) || !llvm::isPowerOf2_32(NumUniqueScalarValues)) { LLVM_DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return false; } VL = UniqueValues; } return true; }; InstructionsState S = getSameOpcode(VL, *TLI); // Gather if we hit the RecursionMaxDepth, unless this is a load (or z/sext of // a load), in which case peek through to include it in the tree, without // ballooning over-budget. if (Depth >= RecursionMaxDepth && !(S.MainOp && isa(S.MainOp) && S.MainOp == S.AltOp && VL.size() >= 4 && (match(S.MainOp, m_Load(m_Value())) || all_of(VL, [&S](const Value *I) { return match(I, m_OneUse(m_ZExtOrSExt(m_OneUse(m_Load(m_Value()))))) && cast(I)->getOpcode() == cast(S.MainOp)->getOpcode(); })))) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } // Don't handle scalable vectors if (S.getOpcode() == Instruction::ExtractElement && isa( cast(S.OpValue)->getVectorOperandType())) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to scalable vector type.\n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } // Don't handle vectors. if (S.OpValue->getType()->isVectorTy() && !isa(S.OpValue)) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return; } if (StoreInst *SI = dyn_cast(S.OpValue)) if (SI->getValueOperand()->getType()->isVectorTy()) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return; } // If all of the operands are identical or constant we have a simple solution. // If we deal with insert/extract instructions, they all must have constant // indices, otherwise we should gather them, not try to vectorize. // If alternate op node with 2 elements with gathered operands - do not // vectorize. auto &&NotProfitableForVectorization = [&S, this, Depth](ArrayRef VL) { if (!S.getOpcode() || !S.isAltShuffle() || VL.size() > 2) return false; if (VectorizableTree.size() < MinTreeSize) return false; if (Depth >= RecursionMaxDepth - 1) return true; // Check if all operands are extracts, part of vector node or can build a // regular vectorize node. SmallVector InstsCount(VL.size(), 0); for (Value *V : VL) { auto *I = cast(V); InstsCount.push_back(count_if(I->operand_values(), [](Value *Op) { return isa(Op) || isVectorLikeInstWithConstOps(Op); })); } bool IsCommutative = isCommutative(S.MainOp) || isCommutative(S.AltOp); if ((IsCommutative && std::accumulate(InstsCount.begin(), InstsCount.end(), 0) < 2) || (!IsCommutative && all_of(InstsCount, [](unsigned ICnt) { return ICnt < 2; }))) return true; assert(VL.size() == 2 && "Expected only 2 alternate op instructions."); SmallVector>> Candidates; auto *I1 = cast(VL.front()); auto *I2 = cast(VL.back()); for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) Candidates.emplace_back().emplace_back(I1->getOperand(Op), I2->getOperand(Op)); if (static_cast(count_if( Candidates, [this](ArrayRef> Cand) { return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); })) >= S.MainOp->getNumOperands() / 2) return false; if (S.MainOp->getNumOperands() > 2) return true; if (IsCommutative) { // Check permuted operands. Candidates.clear(); for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) Candidates.emplace_back().emplace_back(I1->getOperand(Op), I2->getOperand((Op + 1) % E)); if (any_of( Candidates, [this](ArrayRef> Cand) { return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); })) return false; } return true; }; SmallVector SortedIndices; BasicBlock *BB = nullptr; bool IsScatterVectorizeUserTE = UserTreeIdx.UserTE && UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize; bool AreAllSameInsts = (S.getOpcode() && allSameBlock(VL)) || (S.OpValue->getType()->isPointerTy() && IsScatterVectorizeUserTE && VL.size() > 2 && all_of(VL, [&BB](Value *V) { auto *I = dyn_cast(V); if (!I) return doesNotNeedToBeScheduled(V); if (!BB) BB = I->getParent(); return BB == I->getParent() && I->getNumOperands() == 2; }) && BB && sortPtrAccesses(VL, UserTreeIdx.UserTE->getMainOp()->getType(), *DL, *SE, SortedIndices)); if (!AreAllSameInsts || allConstant(VL) || isSplat(VL) || (isa( S.OpValue) && !all_of(VL, isVectorLikeInstWithConstOps)) || NotProfitableForVectorization(VL)) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O, small shuffle. \n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } // We now know that this is a vector of instructions of the same type from // the same block. // Don't vectorize ephemeral values. if (!EphValues.empty()) { for (Value *V : VL) { if (EphValues.count(V)) { LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V << ") is ephemeral.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return; } } } // Check if this is a duplicate of another entry. if (TreeEntry *E = getTreeEntry(S.OpValue)) { LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n"); if (!E->isSame(VL)) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } // Record the reuse of the tree node. FIXME, currently this is only used to // properly draw the graph rather than for the actual vectorization. E->UserTreeIndices.push_back(UserTreeIdx); LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"); return; } // Check that none of the instructions in the bundle are already in the tree. for (Value *V : VL) { if (!IsScatterVectorizeUserTE && !isa(V)) continue; if (getTreeEntry(V)) { LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V << ") is already in tree.\n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } // The reduction nodes (stored in UserIgnoreList) also should stay scalar. if (UserIgnoreList && !UserIgnoreList->empty()) { for (Value *V : VL) { if (UserIgnoreList && UserIgnoreList->contains(V)) { LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n"); if (TryToFindDuplicates(S)) newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } } // Special processing for sorted pointers for ScatterVectorize node with // constant indeces only. if (AreAllSameInsts && UserTreeIdx.UserTE && UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize && !(S.getOpcode() && allSameBlock(VL))) { assert(S.OpValue->getType()->isPointerTy() && count_if(VL, [](Value *V) { return isa(V); }) >= 2 && "Expected pointers only."); // Reset S to make it GetElementPtr kind of node. const auto *It = find_if(VL, [](Value *V) { return isa(V); }); assert(It != VL.end() && "Expected at least one GEP."); S = getSameOpcode(*It, *TLI); } // Check that all of the users of the scalars that we want to vectorize are // schedulable. auto *VL0 = cast(S.OpValue); BB = VL0->getParent(); if (!DT->isReachableFromEntry(BB)) { // Don't go into unreachable blocks. They may contain instructions with // dependency cycles which confuse the final scheduling. LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return; } // Don't go into catchswitch blocks, which can happen with PHIs. // Such blocks can only have PHIs and the catchswitch. There is no // place to insert a shuffle if we need to, so just avoid that issue. if (isa(BB->getTerminator())) { LLVM_DEBUG(dbgs() << "SLP: bundle in catchswitch block.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); return; } // Check that every instruction appears once in this bundle. if (!TryToFindDuplicates(S)) return; auto &BSRef = BlocksSchedules[BB]; if (!BSRef) BSRef = std::make_unique(BB); BlockScheduling &BS = *BSRef; std::optional Bundle = BS.tryScheduleBundle(VL, this, S); #ifdef EXPENSIVE_CHECKS // Make sure we didn't break any internal invariants BS.verify(); #endif if (!Bundle) { LLVM_DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n"); assert((!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } LLVM_DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); unsigned ShuffleOrOp = S.isAltShuffle() ? (unsigned) Instruction::ShuffleVector : S.getOpcode(); switch (ShuffleOrOp) { case Instruction::PHI: { auto *PH = cast(VL0); // Check for terminator values (e.g. invoke). for (Value *V : VL) for (Value *Incoming : cast(V)->incoming_values()) { Instruction *Term = dyn_cast(Incoming); if (Term && Term->isTerminator()) { LLVM_DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n"); BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } TreeEntry *TE = newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n"); // Keeps the reordered operands to avoid code duplication. SmallVector OperandsVec; for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { if (!DT->isReachableFromEntry(PH->getIncomingBlock(I))) { ValueList Operands(VL.size(), PoisonValue::get(PH->getType())); TE->setOperand(I, Operands); OperandsVec.push_back(Operands); continue; } ValueList Operands; // Prepare the operand vector. for (Value *V : VL) Operands.push_back(cast(V)->getIncomingValueForBlock( PH->getIncomingBlock(I))); TE->setOperand(I, Operands); OperandsVec.push_back(Operands); } for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); return; } case Instruction::ExtractValue: case Instruction::ExtractElement: { OrdersType CurrentOrder; bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); if (Reuse) { LLVM_DEBUG(dbgs() << "SLP: Reusing or shuffling extract sequence.\n"); newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); // This is a special case, as it does not gather, but at the same time // we are not extending buildTree_rec() towards the operands. ValueList Op0; Op0.assign(VL.size(), VL0->getOperand(0)); VectorizableTree.back()->setOperand(0, Op0); return; } if (!CurrentOrder.empty()) { LLVM_DEBUG({ dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }); fixupOrderingIndices(CurrentOrder); // Insert new order with initial value 0, if it does not exist, // otherwise return the iterator to the existing one. newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies, CurrentOrder); // This is a special case, as it does not gather, but at the same time // we are not extending buildTree_rec() towards the operands. ValueList Op0; Op0.assign(VL.size(), VL0->getOperand(0)); VectorizableTree.back()->setOperand(0, Op0); return; } LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); BS.cancelScheduling(VL, VL0); return; } case Instruction::InsertElement: { assert(ReuseShuffleIndicies.empty() && "All inserts should be unique"); // Check that we have a buildvector and not a shuffle of 2 or more // different vectors. ValueSet SourceVectors; for (Value *V : VL) { SourceVectors.insert(cast(V)->getOperand(0)); assert(getInsertIndex(V) != std::nullopt && "Non-constant or undef index?"); } if (count_if(VL, [&SourceVectors](Value *V) { return !SourceVectors.contains(V); }) >= 2) { // Found 2nd source vector - cancel. LLVM_DEBUG(dbgs() << "SLP: Gather of insertelement vectors with " "different source vectors.\n"); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); BS.cancelScheduling(VL, VL0); return; } auto OrdCompare = [](const std::pair &P1, const std::pair &P2) { return P1.first > P2.first; }; PriorityQueue, SmallVector>, decltype(OrdCompare)> Indices(OrdCompare); for (int I = 0, E = VL.size(); I < E; ++I) { unsigned Idx = *getInsertIndex(VL[I]); Indices.emplace(Idx, I); } OrdersType CurrentOrder(VL.size(), VL.size()); bool IsIdentity = true; for (int I = 0, E = VL.size(); I < E; ++I) { CurrentOrder[Indices.top().second] = I; IsIdentity &= Indices.top().second == I; Indices.pop(); } if (IsIdentity) CurrentOrder.clear(); TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, std::nullopt, CurrentOrder); LLVM_DEBUG(dbgs() << "SLP: added inserts bundle.\n"); constexpr int NumOps = 2; ValueList VectorOperands[NumOps]; for (int I = 0; I < NumOps; ++I) { for (Value *V : VL) VectorOperands[I].push_back(cast(V)->getOperand(I)); TE->setOperand(I, VectorOperands[I]); } buildTree_rec(VectorOperands[NumOps - 1], Depth + 1, {TE, NumOps - 1}); return; } case Instruction::Load: { // Check that a vectorized load would load the same memory as a scalar // load. For example, we don't want to 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. SmallVector PointerOps; OrdersType CurrentOrder; TreeEntry *TE = nullptr; switch (canVectorizeLoads(VL, VL0, *TTI, *DL, *SE, *LI, *TLI, CurrentOrder, PointerOps)) { case LoadsState::Vectorize: if (CurrentOrder.empty()) { // Original loads are consecutive and does not require reordering. TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of loads.\n"); } else { fixupOrderingIndices(CurrentOrder); // Need to reorder. TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies, CurrentOrder); LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled loads.\n"); } TE->setOperandsInOrder(); break; case LoadsState::ScatterVectorize: // Vectorizing non-consecutive loads with `llvm.masked.gather`. TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); TE->setOperandsInOrder(); buildTree_rec(PointerOps, Depth + 1, {TE, 0}); LLVM_DEBUG(dbgs() << "SLP: added a vector of non-consecutive loads.\n"); break; case LoadsState::Gather: BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); #ifndef NDEBUG Type *ScalarTy = VL0->getType(); if (DL->getTypeSizeInBits(ScalarTy) != DL->getTypeAllocSizeInBits(ScalarTy)) LLVM_DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n"); else if (any_of(VL, [](Value *V) { return !cast(V)->isSimple(); })) LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n"); else LLVM_DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n"); #endif // NDEBUG break; } return; } case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: case Instruction::FPToSI: case Instruction::FPExt: case Instruction::PtrToInt: case Instruction::IntToPtr: case Instruction::SIToFP: case Instruction::UIToFP: case Instruction::Trunc: case Instruction::FPTrunc: case Instruction::BitCast: { Type *SrcTy = VL0->getOperand(0)->getType(); for (Value *V : VL) { Type *Ty = cast(V)->getOperand(0)->getType(); if (Ty != SrcTy || !isValidElementType(Ty)) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n"); return; } } TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of casts.\n"); TE->setOperandsInOrder(); for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { ValueList Operands; // Prepare the operand vector. for (Value *V : VL) Operands.push_back(cast(V)->getOperand(i)); buildTree_rec(Operands, Depth + 1, {TE, i}); } return; } case Instruction::ICmp: case Instruction::FCmp: { // Check that all of the compares have the same predicate. CmpInst::Predicate P0 = cast(VL0)->getPredicate(); CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); Type *ComparedTy = VL0->getOperand(0)->getType(); for (Value *V : VL) { CmpInst *Cmp = cast(V); if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || Cmp->getOperand(0)->getType() != ComparedTy) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n"); return; } } TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of compares.\n"); ValueList Left, Right; if (cast(VL0)->isCommutative()) { // Commutative predicate - collect + sort operands of the instructions // so that each side is more likely to have the same opcode. assert(P0 == SwapP0 && "Commutative Predicate mismatch"); reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, *this); } else { // Collect operands - commute if it uses the swapped predicate. for (Value *V : VL) { auto *Cmp = cast(V); Value *LHS = Cmp->getOperand(0); Value *RHS = Cmp->getOperand(1); if (Cmp->getPredicate() != P0) std::swap(LHS, RHS); Left.push_back(LHS); Right.push_back(RHS); } } TE->setOperand(0, Left); TE->setOperand(1, Right); buildTree_rec(Left, Depth + 1, {TE, 0}); buildTree_rec(Right, Depth + 1, {TE, 1}); return; } case Instruction::Select: case Instruction::FNeg: case Instruction::Add: case Instruction::FAdd: case Instruction::Sub: case Instruction::FSub: case Instruction::Mul: case Instruction::FMul: case Instruction::UDiv: case Instruction::SDiv: case Instruction::FDiv: case Instruction::URem: case Instruction::SRem: case Instruction::FRem: case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::And: case Instruction::Or: case Instruction::Xor: { TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of un/bin op.\n"); // Sort operands of the instructions so that each side is more likely to // have the same opcode. if (isa(VL0) && VL0->isCommutative()) { ValueList Left, Right; reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, *this); TE->setOperand(0, Left); TE->setOperand(1, Right); buildTree_rec(Left, Depth + 1, {TE, 0}); buildTree_rec(Right, Depth + 1, {TE, 1}); return; } TE->setOperandsInOrder(); for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { ValueList Operands; // Prepare the operand vector. for (Value *V : VL) Operands.push_back(cast(V)->getOperand(i)); buildTree_rec(Operands, Depth + 1, {TE, i}); } return; } case Instruction::GetElementPtr: { // We don't combine GEPs with complicated (nested) indexing. for (Value *V : VL) { auto *I = dyn_cast(V); if (!I) continue; if (I->getNumOperands() != 2) { LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n"); BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } // We can't combine several GEPs into one vector if they operate on // different types. Type *Ty0 = cast(VL0)->getSourceElementType(); for (Value *V : VL) { auto *GEP = dyn_cast(V); if (!GEP) continue; Type *CurTy = GEP->getSourceElementType(); if (Ty0 != CurTy) { LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n"); BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } // We don't combine GEPs with non-constant indexes. Type *Ty1 = VL0->getOperand(1)->getType(); for (Value *V : VL) { auto *I = dyn_cast(V); if (!I) continue; auto *Op = I->getOperand(1); if ((!IsScatterVectorizeUserTE && !isa(Op)) || (Op->getType() != Ty1 && ((IsScatterVectorizeUserTE && !isa(Op)) || Op->getType()->getScalarSizeInBits() > DL->getIndexSizeInBits( V->getType()->getPointerAddressSpace())))) { LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"); BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); return; } } TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a vector of GEPs.\n"); SmallVector Operands(2); // Prepare the operand vector for pointer operands. for (Value *V : VL) { auto *GEP = dyn_cast(V); if (!GEP) { Operands.front().push_back(V); continue; } Operands.front().push_back(GEP->getPointerOperand()); } TE->setOperand(0, Operands.front()); // Need to cast all indices to the same type before vectorization to // avoid crash. // Required to be able to find correct matches between different gather // nodes and reuse the vectorized values rather than trying to gather them // again. int IndexIdx = 1; Type *VL0Ty = VL0->getOperand(IndexIdx)->getType(); Type *Ty = all_of(VL, [VL0Ty, IndexIdx](Value *V) { auto *GEP = dyn_cast(V); if (!GEP) return true; return VL0Ty == GEP->getOperand(IndexIdx)->getType(); }) ? VL0Ty : DL->getIndexType(cast(VL0) ->getPointerOperandType() ->getScalarType()); // Prepare the operand vector. for (Value *V : VL) { auto *I = dyn_cast(V); if (!I) { Operands.back().push_back( ConstantInt::get(Ty, 0, /*isSigned=*/false)); continue; } auto *Op = I->getOperand(IndexIdx); auto *CI = dyn_cast(Op); if (!CI) Operands.back().push_back(Op); else Operands.back().push_back(ConstantExpr::getIntegerCast( CI, Ty, CI->getValue().isSignBitSet())); } TE->setOperand(IndexIdx, Operands.back()); for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I) buildTree_rec(Operands[I], Depth + 1, {TE, I}); return; } case Instruction::Store: { // Check if the stores are consecutive or if we need to swizzle them. llvm::Type *ScalarTy = cast(VL0)->getValueOperand()->getType(); // Avoid types that are padded when being allocated as scalars, while // being packed together in a vector (such as i1). if (DL->getTypeSizeInBits(ScalarTy) != DL->getTypeAllocSizeInBits(ScalarTy)) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Gathering stores of non-packed type.\n"); return; } // Make sure all stores in the bundle are simple - we can't vectorize // atomic or volatile stores. SmallVector PointerOps(VL.size()); ValueList Operands(VL.size()); auto POIter = PointerOps.begin(); auto OIter = Operands.begin(); for (Value *V : VL) { auto *SI = cast(V); if (!SI->isSimple()) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple stores.\n"); return; } *POIter = SI->getPointerOperand(); *OIter = SI->getValueOperand(); ++POIter; ++OIter; } OrdersType CurrentOrder; // Check the order of pointer operands. if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) { Value *Ptr0; Value *PtrN; if (CurrentOrder.empty()) { Ptr0 = PointerOps.front(); PtrN = PointerOps.back(); } else { Ptr0 = PointerOps[CurrentOrder.front()]; PtrN = PointerOps[CurrentOrder.back()]; } std::optional Dist = getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE); // Check that the sorted pointer operands are consecutive. if (static_cast(*Dist) == VL.size() - 1) { if (CurrentOrder.empty()) { // Original stores are consecutive and does not require reordering. TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); TE->setOperandsInOrder(); buildTree_rec(Operands, Depth + 1, {TE, 0}); LLVM_DEBUG(dbgs() << "SLP: added a vector of stores.\n"); } else { fixupOrderingIndices(CurrentOrder); TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies, CurrentOrder); TE->setOperandsInOrder(); buildTree_rec(Operands, Depth + 1, {TE, 0}); LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled stores.\n"); } return; } } BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n"); return; } case Instruction::Call: { // Check if the calls are all to the same vectorizable intrinsic or // library function. CallInst *CI = cast(VL0); Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); VFShape Shape = VFShape::get( *CI, ElementCount::getFixed(static_cast(VL.size())), false /*HasGlobalPred*/); Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); if (!VecFunc && !isTriviallyVectorizable(ID)) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n"); return; } Function *F = CI->getCalledFunction(); unsigned NumArgs = CI->arg_size(); SmallVector ScalarArgs(NumArgs, nullptr); for (unsigned j = 0; j != NumArgs; ++j) if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) ScalarArgs[j] = CI->getArgOperand(j); for (Value *V : VL) { CallInst *CI2 = dyn_cast(V); if (!CI2 || CI2->getCalledFunction() != F || getVectorIntrinsicIDForCall(CI2, TLI) != ID || (VecFunc && VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || !CI->hasIdenticalOperandBundleSchema(*CI2)) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *V << "\n"); return; } // Some intrinsics have scalar arguments and should be same in order for // them to be vectorized. for (unsigned j = 0; j != NumArgs; ++j) { if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) { Value *A1J = CI2->getArgOperand(j); if (ScalarArgs[j] != A1J) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"); return; } } } // Verify that the bundle operands are identical between the two calls. if (CI->hasOperandBundles() && !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), CI->op_begin() + CI->getBundleOperandsEndIndex(), CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'); return; } } TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); TE->setOperandsInOrder(); for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { // For scalar operands no need to to create an entry since no need to // vectorize it. if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) continue; ValueList Operands; // Prepare the operand vector. for (Value *V : VL) { auto *CI2 = cast(V); Operands.push_back(CI2->getArgOperand(i)); } buildTree_rec(Operands, Depth + 1, {TE, i}); } return; } case Instruction::ShuffleVector: { // If this is not an alternate sequence of opcode like add-sub // then do not vectorize this instruction. if (!S.isAltShuffle()) { BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n"); return; } TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n"); // Reorder operands if reordering would enable vectorization. auto *CI = dyn_cast(VL0); if (isa(VL0) || CI) { ValueList Left, Right; if (!CI || all_of(VL, [](Value *V) { return cast(V)->isCommutative(); })) { reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, *this); } else { auto *MainCI = cast(S.MainOp); auto *AltCI = cast(S.AltOp); CmpInst::Predicate MainP = MainCI->getPredicate(); CmpInst::Predicate AltP = AltCI->getPredicate(); assert(MainP != AltP && "Expected different main/alternate predicates."); // Collect operands - commute if it uses the swapped predicate or // alternate operation. for (Value *V : VL) { auto *Cmp = cast(V); Value *LHS = Cmp->getOperand(0); Value *RHS = Cmp->getOperand(1); if (isAlternateInstruction(Cmp, MainCI, AltCI, *TLI)) { if (AltP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) std::swap(LHS, RHS); } else { if (MainP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) std::swap(LHS, RHS); } Left.push_back(LHS); Right.push_back(RHS); } } TE->setOperand(0, Left); TE->setOperand(1, Right); buildTree_rec(Left, Depth + 1, {TE, 0}); buildTree_rec(Right, Depth + 1, {TE, 1}); return; } TE->setOperandsInOrder(); for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { ValueList Operands; // Prepare the operand vector. for (Value *V : VL) Operands.push_back(cast(V)->getOperand(i)); buildTree_rec(Operands, Depth + 1, {TE, i}); } return; } default: BS.cancelScheduling(VL, VL0); newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, ReuseShuffleIndicies); LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n"); return; } } unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { unsigned N = 1; Type *EltTy = T; while (isa(EltTy)) { if (auto *ST = dyn_cast(EltTy)) { // Check that struct is homogeneous. for (const auto *Ty : ST->elements()) if (Ty != *ST->element_begin()) return 0; N *= ST->getNumElements(); EltTy = *ST->element_begin(); } else if (auto *AT = dyn_cast(EltTy)) { N *= AT->getNumElements(); EltTy = AT->getElementType(); } else { auto *VT = cast(EltTy); N *= VT->getNumElements(); EltTy = VT->getElementType(); } } if (!isValidElementType(EltTy)) return 0; uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) return 0; return N; } bool BoUpSLP::canReuseExtract(ArrayRef VL, Value *OpValue, SmallVectorImpl &CurrentOrder) const { const auto *It = find_if(VL, [](Value *V) { return isa(V); }); assert(It != VL.end() && "Expected at least one extract instruction."); auto *E0 = cast(*It); assert(all_of(VL, [](Value *V) { return isa( V); }) && "Invalid opcode"); // Check if all of the extracts come from the same vector and from the // correct offset. Value *Vec = E0->getOperand(0); CurrentOrder.clear(); // We have to extract from a vector/aggregate with the same number of elements. unsigned NElts; if (E0->getOpcode() == Instruction::ExtractValue) { const DataLayout &DL = E0->getModule()->getDataLayout(); NElts = canMapToVector(Vec->getType(), DL); if (!NElts) return false; // Check if load can be rewritten as load of vector. LoadInst *LI = dyn_cast(Vec); if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) return false; } else { NElts = cast(Vec->getType())->getNumElements(); } if (NElts != VL.size()) return false; // Check that all of the indices extract from the correct offset. bool ShouldKeepOrder = true; unsigned E = VL.size(); // Assign to all items the initial value E + 1 so we can check if the extract // instruction index was used already. // Also, later we can check that all the indices are used and we have a // consecutive access in the extract instructions, by checking that no // element of CurrentOrder still has value E + 1. CurrentOrder.assign(E, E); unsigned I = 0; for (; I < E; ++I) { auto *Inst = dyn_cast(VL[I]); if (!Inst) continue; if (Inst->getOperand(0) != Vec) break; if (auto *EE = dyn_cast(Inst)) if (isa(EE->getIndexOperand())) continue; std::optional Idx = getExtractIndex(Inst); if (!Idx) break; const unsigned ExtIdx = *Idx; if (ExtIdx != I) { if (ExtIdx >= E || CurrentOrder[ExtIdx] != E) break; ShouldKeepOrder = false; CurrentOrder[ExtIdx] = I; } else { if (CurrentOrder[I] != E) break; CurrentOrder[I] = I; } } if (I < E) { CurrentOrder.clear(); return false; } if (ShouldKeepOrder) CurrentOrder.clear(); return ShouldKeepOrder; } bool BoUpSLP::areAllUsersVectorized(Instruction *I, ArrayRef VectorizedVals) const { return (I->hasOneUse() && is_contained(VectorizedVals, I)) || all_of(I->users(), [this](User *U) { return ScalarToTreeEntry.count(U) > 0 || isVectorLikeInstWithConstOps(U) || (isa(U) && MustGather.contains(U)); }); } static std::pair getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); // Calculate the cost of the scalar and vector calls. SmallVector VecTys; for (Use &Arg : CI->args()) VecTys.push_back( FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); FastMathFlags FMF; if (auto *FPCI = dyn_cast(CI)) FMF = FPCI->getFastMathFlags(); SmallVector Arguments(CI->args()); IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF, dyn_cast(CI)); auto IntrinsicCost = TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast( VecTy->getNumElements())), false /*HasGlobalPred*/); Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); auto LibCost = IntrinsicCost; if (!CI->isNoBuiltin() && VecFunc) { // Calculate the cost of the vector library call. // If the corresponding vector call is cheaper, return its cost. LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, TTI::TCK_RecipThroughput); } return {IntrinsicCost, LibCost}; } /// Compute the cost of creating a vector of type \p VecTy containing the /// extracted values from \p VL. static InstructionCost computeExtractCost(ArrayRef VL, FixedVectorType *VecTy, TargetTransformInfo::ShuffleKind ShuffleKind, ArrayRef Mask, TargetTransformInfo &TTI) { unsigned NumOfParts = TTI.getNumberOfParts(VecTy); if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || !NumOfParts || VecTy->getNumElements() < NumOfParts) return TTI.getShuffleCost(ShuffleKind, VecTy, Mask); bool AllConsecutive = true; unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts; unsigned Idx = -1; InstructionCost Cost = 0; // Process extracts in blocks of EltsPerVector to check if the source vector // operand can be re-used directly. If not, add the cost of creating a shuffle // to extract the values into a vector register. SmallVector RegMask(EltsPerVector, UndefMaskElem); for (auto *V : VL) { ++Idx; // Reached the start of a new vector registers. if (Idx % EltsPerVector == 0) { RegMask.assign(EltsPerVector, UndefMaskElem); AllConsecutive = true; continue; } // Need to exclude undefs from analysis. if (isa(V) || Mask[Idx] == UndefMaskElem) continue; // Check all extracts for a vector register on the target directly // extract values in order. unsigned CurrentIdx = *getExtractIndex(cast(V)); if (!isa(VL[Idx - 1]) && Mask[Idx - 1] != UndefMaskElem) { unsigned PrevIdx = *getExtractIndex(cast(VL[Idx - 1])); AllConsecutive &= PrevIdx + 1 == CurrentIdx && CurrentIdx % EltsPerVector == Idx % EltsPerVector; RegMask[Idx % EltsPerVector] = CurrentIdx % EltsPerVector; } if (AllConsecutive) continue; // Skip all indices, except for the last index per vector block. if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size()) continue; // If we have a series of extracts which are not consecutive and hence // cannot re-use the source vector register directly, compute the shuffle // cost to extract the vector with EltsPerVector elements. Cost += TTI.getShuffleCost( TargetTransformInfo::SK_PermuteSingleSrc, FixedVectorType::get(VecTy->getElementType(), EltsPerVector), RegMask); } return Cost; } /// Build shuffle mask for shuffle graph entries and lists of main and alternate /// operations operands. static void buildShuffleEntryMask(ArrayRef VL, ArrayRef ReorderIndices, ArrayRef ReusesIndices, const function_ref IsAltOp, SmallVectorImpl &Mask, SmallVectorImpl *OpScalars = nullptr, SmallVectorImpl *AltScalars = nullptr) { unsigned Sz = VL.size(); Mask.assign(Sz, UndefMaskElem); SmallVector OrderMask; if (!ReorderIndices.empty()) inversePermutation(ReorderIndices, OrderMask); for (unsigned I = 0; I < Sz; ++I) { unsigned Idx = I; if (!ReorderIndices.empty()) Idx = OrderMask[I]; auto *OpInst = cast(VL[Idx]); if (IsAltOp(OpInst)) { Mask[I] = Sz + Idx; if (AltScalars) AltScalars->push_back(OpInst); } else { Mask[I] = Idx; if (OpScalars) OpScalars->push_back(OpInst); } } if (!ReusesIndices.empty()) { SmallVector NewMask(ReusesIndices.size(), UndefMaskElem); transform(ReusesIndices, NewMask.begin(), [&Mask](int Idx) { return Idx != UndefMaskElem ? Mask[Idx] : UndefMaskElem; }); Mask.swap(NewMask); } } static bool isAlternateInstruction(const Instruction *I, const Instruction *MainOp, const Instruction *AltOp, const TargetLibraryInfo &TLI) { if (auto *MainCI = dyn_cast(MainOp)) { auto *AltCI = cast(AltOp); CmpInst::Predicate MainP = MainCI->getPredicate(); CmpInst::Predicate AltP = AltCI->getPredicate(); assert(MainP != AltP && "Expected different main/alternate predicates."); auto *CI = cast(I); if (isCmpSameOrSwapped(MainCI, CI, TLI)) return false; if (isCmpSameOrSwapped(AltCI, CI, TLI)) return true; CmpInst::Predicate P = CI->getPredicate(); CmpInst::Predicate SwappedP = CmpInst::getSwappedPredicate(P); assert((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && "CmpInst expected to match either main or alternate predicate or " "their swap."); (void)AltP; return MainP != P && MainP != SwappedP; } return I->getOpcode() == AltOp->getOpcode(); } TTI::OperandValueInfo BoUpSLP::getOperandInfo(ArrayRef VL, unsigned OpIdx) { assert(!VL.empty()); const auto *I0 = cast(*find_if(VL, Instruction::classof)); const auto *Op0 = I0->getOperand(OpIdx); const bool IsConstant = all_of(VL, [&](Value *V) { // TODO: We should allow undef elements here const auto *I = dyn_cast(V); if (!I) return true; auto *Op = I->getOperand(OpIdx); return isConstant(Op) && !isa(Op); }); const bool IsUniform = all_of(VL, [&](Value *V) { // TODO: We should allow undef elements here const auto *I = dyn_cast(V); if (!I) return false; return I->getOperand(OpIdx) == Op0; }); const bool IsPowerOfTwo = all_of(VL, [&](Value *V) { // TODO: We should allow undef elements here const auto *I = dyn_cast(V); if (!I) { assert((isa(V) || I0->getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP."); return true; } auto *Op = I->getOperand(OpIdx); if (auto *CI = dyn_cast(Op)) return CI->getValue().isPowerOf2(); return false; }); const bool IsNegatedPowerOfTwo = all_of(VL, [&](Value *V) { // TODO: We should allow undef elements here const auto *I = dyn_cast(V); if (!I) { assert((isa(V) || I0->getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP."); return true; } const auto *Op = I->getOperand(OpIdx); if (auto *CI = dyn_cast(Op)) return CI->getValue().isNegatedPowerOf2(); return false; }); TTI::OperandValueKind VK = TTI::OK_AnyValue; if (IsConstant && IsUniform) VK = TTI::OK_UniformConstantValue; else if (IsConstant) VK = TTI::OK_NonUniformConstantValue; else if (IsUniform) VK = TTI::OK_UniformValue; TTI::OperandValueProperties VP = TTI::OP_None; VP = IsPowerOfTwo ? TTI::OP_PowerOf2 : VP; VP = IsNegatedPowerOfTwo ? TTI::OP_NegatedPowerOf2 : VP; return {VK, VP}; } namespace { /// The base class for shuffle instruction emission and shuffle cost estimation. class BaseShuffleAnalysis { protected: /// Checks if the mask is an identity mask. /// \param IsStrict if is true the function returns false if mask size does /// not match vector size. static bool isIdentityMask(ArrayRef Mask, const FixedVectorType *VecTy, bool IsStrict) { int Limit = Mask.size(); int VF = VecTy->getNumElements(); return (VF == Limit || !IsStrict) && all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) && ShuffleVectorInst::isIdentityMask(Mask); } /// Tries to combine 2 different masks into single one. /// \param LocalVF Vector length of the permuted input vector. \p Mask may /// change the size of the vector, \p LocalVF is the original size of the /// shuffled vector. static void combineMasks(unsigned LocalVF, SmallVectorImpl &Mask, ArrayRef ExtMask) { unsigned VF = Mask.size(); SmallVector NewMask(ExtMask.size(), UndefMaskElem); for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) { if (ExtMask[I] == UndefMaskElem) continue; int MaskedIdx = Mask[ExtMask[I] % VF]; NewMask[I] = MaskedIdx == UndefMaskElem ? UndefMaskElem : MaskedIdx % LocalVF; } Mask.swap(NewMask); } /// Looks through shuffles trying to reduce final number of shuffles in the /// code. The function looks through the previously emitted shuffle /// instructions and properly mark indices in mask as undef. /// For example, given the code /// \code /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0> /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0> /// \endcode /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will /// look through %s1 and %s2 and select vectors %0 and %1 with mask /// <0, 1, 2, 3> for the shuffle. /// If 2 operands are of different size, the smallest one will be resized and /// the mask recalculated properly. /// For example, given the code /// \code /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0> /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0> /// \endcode /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will /// look through %s1 and %s2 and select vectors %0 and %1 with mask /// <0, 1, 2, 3> for the shuffle. /// So, it tries to transform permutations to simple vector merge, if /// possible. /// \param V The input vector which must be shuffled using the given \p Mask. /// If the better candidate is found, \p V is set to this best candidate /// vector. /// \param Mask The input mask for the shuffle. If the best candidate is found /// during looking-through-shuffles attempt, it is updated accordingly. /// \param SinglePermute true if the shuffle operation is originally a /// single-value-permutation. In this case the look-through-shuffles procedure /// may look for resizing shuffles as the best candidates. /// \return true if the shuffle results in the non-resizing identity shuffle /// (and thus can be ignored), false - otherwise. static bool peekThroughShuffles(Value *&V, SmallVectorImpl &Mask, bool SinglePermute) { Value *Op = V; ShuffleVectorInst *IdentityOp = nullptr; SmallVector IdentityMask; while (auto *SV = dyn_cast(Op)) { // Exit if not a fixed vector type or changing size shuffle. auto *SVTy = dyn_cast(SV->getType()); if (!SVTy) break; // Remember the identity or broadcast mask, if it is not a resizing // shuffle. If no better candidates are found, this Op and Mask will be // used in the final shuffle. if (isIdentityMask(Mask, SVTy, /*IsStrict=*/false)) { if (!IdentityOp || !SinglePermute || (isIdentityMask(Mask, SVTy, /*IsStrict=*/true) && !ShuffleVectorInst::isZeroEltSplatMask(IdentityMask))) { IdentityOp = SV; // Store current mask in the IdentityMask so later we did not lost // this info if IdentityOp is selected as the best candidate for the // permutation. IdentityMask.assign(Mask); } } // Remember the broadcast mask. If no better candidates are found, this Op // and Mask will be used in the final shuffle. // Zero splat can be used as identity too, since it might be used with // mask <0, 1, 2, ...>, i.e. identity mask without extra reshuffling. // E.g. if need to shuffle the vector with the mask <3, 1, 2, 0>, which is // expensive, the analysis founds out, that the source vector is just a // broadcast, this original mask can be transformed to identity mask <0, // 1, 2, 3>. // \code // %0 = shuffle %v, poison, zeroinitalizer // %res = shuffle %0, poison, <3, 1, 2, 0> // \endcode // may be transformed to // \code // %0 = shuffle %v, poison, zeroinitalizer // %res = shuffle %0, poison, <0, 1, 2, 3> // \endcode if (SV->isZeroEltSplat()) { IdentityOp = SV; IdentityMask.assign(Mask); } int LocalVF = Mask.size(); if (auto *SVOpTy = dyn_cast(SV->getOperand(0)->getType())) LocalVF = SVOpTy->getNumElements(); SmallVector ExtMask(Mask.size(), UndefMaskElem); for (auto [Idx, I] : enumerate(Mask)) { if (I == UndefMaskElem) continue; ExtMask[Idx] = SV->getMaskValue(I); } bool IsOp1Undef = isUndefVector(SV->getOperand(0), buildUseMask(LocalVF, ExtMask, UseMask::FirstArg)) .all(); bool IsOp2Undef = isUndefVector(SV->getOperand(1), buildUseMask(LocalVF, ExtMask, UseMask::SecondArg)) .all(); if (!IsOp1Undef && !IsOp2Undef) { // Update mask and mark undef elems. for (int &I : Mask) { if (I == UndefMaskElem) continue; if (SV->getMaskValue(I % SV->getShuffleMask().size()) == UndefMaskElem) I = UndefMaskElem; } break; } SmallVector ShuffleMask(SV->getShuffleMask().begin(), SV->getShuffleMask().end()); combineMasks(LocalVF, ShuffleMask, Mask); Mask.swap(ShuffleMask); if (IsOp2Undef) Op = SV->getOperand(0); else Op = SV->getOperand(1); } if (auto *OpTy = dyn_cast(Op->getType()); !OpTy || !isIdentityMask(Mask, OpTy, SinglePermute)) { if (IdentityOp) { V = IdentityOp; assert(Mask.size() == IdentityMask.size() && "Expected masks of same sizes."); // Clear known poison elements. for (auto [I, Idx] : enumerate(Mask)) if (Idx == UndefMaskElem) IdentityMask[I] = UndefMaskElem; Mask.swap(IdentityMask); auto *Shuffle = dyn_cast(V); return SinglePermute && (isIdentityMask(Mask, cast(V->getType()), /*IsStrict=*/true) || (Shuffle && Mask.size() == Shuffle->getShuffleMask().size() && Shuffle->isZeroEltSplat() && ShuffleVectorInst::isZeroEltSplatMask(Mask))); } V = Op; return false; } V = Op; return true; } /// Smart shuffle instruction emission, walks through shuffles trees and /// tries to find the best matching vector for the actual shuffle /// instruction. template static Value *createShuffle(Value *V1, Value *V2, ArrayRef Mask, ShuffleBuilderTy &Builder) { assert(V1 && "Expected at least one vector value."); int VF = Mask.size(); if (auto *FTy = dyn_cast(V1->getType())) VF = FTy->getNumElements(); if (V2 && !isUndefVector(V2, buildUseMask(VF, Mask, UseMask::SecondArg)).all()) { // Peek through shuffles. Value *Op1 = V1; Value *Op2 = V2; int VF = cast(V1->getType())->getElementCount().getKnownMinValue(); SmallVector CombinedMask1(Mask.size(), UndefMaskElem); SmallVector CombinedMask2(Mask.size(), UndefMaskElem); for (int I = 0, E = Mask.size(); I < E; ++I) { if (Mask[I] < VF) CombinedMask1[I] = Mask[I]; else CombinedMask2[I] = Mask[I] - VF; } Value *PrevOp1; Value *PrevOp2; do { PrevOp1 = Op1; PrevOp2 = Op2; (void)peekThroughShuffles(Op1, CombinedMask1, /*SinglePermute=*/false); (void)peekThroughShuffles(Op2, CombinedMask2, /*SinglePermute=*/false); // Check if we have 2 resizing shuffles - need to peek through operands // again. if (auto *SV1 = dyn_cast(Op1)) if (auto *SV2 = dyn_cast(Op2)) { SmallVector ExtMask1(Mask.size(), UndefMaskElem); for (auto [Idx, I] : enumerate(CombinedMask1)) { if (I == UndefMaskElem) continue; ExtMask1[Idx] = SV1->getMaskValue(I); } SmallBitVector UseMask1 = buildUseMask( cast(SV1->getOperand(1)->getType()) ->getNumElements(), ExtMask1, UseMask::SecondArg); SmallVector ExtMask2(CombinedMask2.size(), UndefMaskElem); for (auto [Idx, I] : enumerate(CombinedMask2)) { if (I == UndefMaskElem) continue; ExtMask2[Idx] = SV2->getMaskValue(I); } SmallBitVector UseMask2 = buildUseMask( cast(SV2->getOperand(1)->getType()) ->getNumElements(), ExtMask2, UseMask::SecondArg); if (SV1->getOperand(0)->getType() == SV2->getOperand(0)->getType() && SV1->getOperand(0)->getType() != SV1->getType() && isUndefVector(SV1->getOperand(1), UseMask1).all() && isUndefVector(SV2->getOperand(1), UseMask2).all()) { Op1 = SV1->getOperand(0); Op2 = SV2->getOperand(0); SmallVector ShuffleMask1(SV1->getShuffleMask().begin(), SV1->getShuffleMask().end()); int LocalVF = ShuffleMask1.size(); if (auto *FTy = dyn_cast(Op1->getType())) LocalVF = FTy->getNumElements(); combineMasks(LocalVF, ShuffleMask1, CombinedMask1); CombinedMask1.swap(ShuffleMask1); SmallVector ShuffleMask2(SV2->getShuffleMask().begin(), SV2->getShuffleMask().end()); LocalVF = ShuffleMask2.size(); if (auto *FTy = dyn_cast(Op2->getType())) LocalVF = FTy->getNumElements(); combineMasks(LocalVF, ShuffleMask2, CombinedMask2); CombinedMask2.swap(ShuffleMask2); } } } while (PrevOp1 != Op1 || PrevOp2 != Op2); Builder.resizeToMatch(Op1, Op2); VF = std::max(cast(Op1->getType()) ->getElementCount() .getKnownMinValue(), cast(Op2->getType()) ->getElementCount() .getKnownMinValue()); for (int I = 0, E = Mask.size(); I < E; ++I) { if (CombinedMask2[I] != UndefMaskElem) { assert(CombinedMask1[I] == UndefMaskElem && "Expected undefined mask element"); CombinedMask1[I] = CombinedMask2[I] + (Op1 == Op2 ? 0 : VF); } } return Builder.createShuffleVector( Op1, Op1 == Op2 ? PoisonValue::get(Op1->getType()) : Op2, CombinedMask1); } if (isa(V1)) return PoisonValue::get(FixedVectorType::get( cast(V1->getType())->getElementType(), Mask.size())); SmallVector NewMask(Mask.begin(), Mask.end()); bool IsIdentity = peekThroughShuffles(V1, NewMask, /*SinglePermute=*/true); assert(V1 && "Expected non-null value after looking through shuffles."); if (!IsIdentity) return Builder.createShuffleVector(V1, NewMask); return V1; } }; } // namespace InstructionCost BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef VectorizedVals) { ArrayRef VL = E->Scalars; Type *ScalarTy = VL[0]->getType(); if (StoreInst *SI = dyn_cast(VL[0])) ScalarTy = SI->getValueOperand()->getType(); else if (CmpInst *CI = dyn_cast(VL[0])) ScalarTy = CI->getOperand(0)->getType(); else if (auto *IE = dyn_cast(VL[0])) ScalarTy = IE->getOperand(1)->getType(); auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; // If we have computed a smaller type for the expression, update VecTy so // that the costs will be accurate. if (MinBWs.count(VL[0])) VecTy = FixedVectorType::get( IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); unsigned EntryVF = E->getVectorFactor(); auto *FinalVecTy = FixedVectorType::get(VecTy->getElementType(), EntryVF); bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); // FIXME: it tries to fix a problem with MSVC buildbots. TargetTransformInfo *TTI = this->TTI; auto AdjustExtractsCost = [=](InstructionCost &Cost) { // If the resulting type is scalarized, do not adjust the cost. unsigned VecNumParts = TTI->getNumberOfParts(VecTy); if (VecNumParts == VecTy->getNumElements()) return; DenseMap ExtractVectorsTys; SmallPtrSet CheckedExtracts; for (auto *V : VL) { if (isa(V)) continue; // If all users of instruction are going to be vectorized and this // instruction itself is not going to be vectorized, consider this // instruction as dead and remove its cost from the final cost of the // vectorized tree. // Also, avoid adjusting the cost for extractelements with multiple uses // in different graph entries. const TreeEntry *VE = getTreeEntry(V); if (!CheckedExtracts.insert(V).second || !areAllUsersVectorized(cast(V), VectorizedVals) || (VE && VE != E)) continue; auto *EE = cast(V); std::optional EEIdx = getExtractIndex(EE); if (!EEIdx) continue; unsigned Idx = *EEIdx; if (VecNumParts != TTI->getNumberOfParts(EE->getVectorOperandType())) { auto It = ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first; It->getSecond() = std::min(It->second, Idx); } // Take credit for instruction that will become dead. if (EE->hasOneUse()) { Instruction *Ext = EE->user_back(); if (isa(Ext) && all_of(Ext->users(), [](User *U) { return isa(U); })) { // Use getExtractWithExtendCost() to calculate the cost of // extractelement/ext pair. Cost -= TTI->getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(), EE->getVectorOperandType(), Idx); // Add back the cost of s|zext which is subtracted separately. Cost += TTI->getCastInstrCost( Ext->getOpcode(), Ext->getType(), EE->getType(), TTI::getCastContextHint(Ext), CostKind, Ext); continue; } } Cost -= TTI->getVectorInstrCost(*EE, EE->getVectorOperandType(), CostKind, Idx); } // Add a cost for subvector extracts/inserts if required. for (const auto &Data : ExtractVectorsTys) { auto *EEVTy = cast(Data.first->getType()); unsigned NumElts = VecTy->getNumElements(); if (Data.second % NumElts == 0) continue; if (TTI->getNumberOfParts(EEVTy) > VecNumParts) { unsigned Idx = (Data.second / NumElts) * NumElts; unsigned EENumElts = EEVTy->getNumElements(); if (Idx + NumElts <= EENumElts) { Cost += TTI->getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, EEVTy, std::nullopt, CostKind, Idx, VecTy); } else { // Need to round up the subvector type vectorization factor to avoid a // crash in cost model functions. Make SubVT so that Idx + VF of SubVT // <= EENumElts. auto *SubVT = FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx); Cost += TTI->getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, EEVTy, std::nullopt, CostKind, Idx, SubVT); } } else { Cost += TTI->getShuffleCost(TargetTransformInfo::SK_InsertSubvector, VecTy, std::nullopt, CostKind, 0, EEVTy); } } }; if (E->State == TreeEntry::NeedToGather) { if (allConstant(VL)) return 0; if (isa(VL[0])) return InstructionCost::getInvalid(); SmallVector GatheredScalars(E->Scalars.begin(), E->Scalars.end()); // Build a mask out of the reorder indices and reorder scalars per this // mask. SmallVector ReorderMask; inversePermutation(E->ReorderIndices, ReorderMask); if (!ReorderMask.empty()) reorderScalars(GatheredScalars, ReorderMask); SmallVector Mask; std::optional GatherShuffle; SmallVector Entries; // Do not try to look for reshuffled loads for gathered loads (they will be // handled later), for vectorized scalars, and cases, which are definitely // not profitable (splats and small gather nodes.) if (E->getOpcode() != Instruction::Load || E->isAltShuffle() || all_of(E->Scalars, [this](Value *V) { return getTreeEntry(V); }) || isSplat(E->Scalars) || (E->Scalars != GatheredScalars && GatheredScalars.size() <= 2)) GatherShuffle = isGatherShuffledEntry(E, GatheredScalars, Mask, Entries); if (GatherShuffle) { // Remove shuffled elements from list of gathers. for (int I = 0, Sz = Mask.size(); I < Sz; ++I) { if (Mask[I] != UndefMaskElem) GatheredScalars[I] = PoisonValue::get(ScalarTy); } assert((Entries.size() == 1 || Entries.size() == 2) && "Expected shuffle of 1 or 2 entries."); InstructionCost GatherCost = 0; int Limit = Mask.size() * 2; if (all_of(Mask, [=](int Idx) { return Idx < Limit; }) && ShuffleVectorInst::isIdentityMask(Mask)) { // Perfect match in the graph, will reuse the previously vectorized // node. Cost is 0. LLVM_DEBUG( dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"); if (NeedToShuffleReuses) GatherCost = TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices); } else { LLVM_DEBUG(dbgs() << "SLP: shuffled " << Entries.size() << " entries for bundle that starts with " << *VL.front() << ".\n"); // Detected that instead of gather we can emit a shuffle of single/two // previously vectorized nodes. Add the cost of the permutation rather // than gather. ::addMask(Mask, E->ReuseShuffleIndices); GatherCost = TTI->getShuffleCost(*GatherShuffle, FinalVecTy, Mask); } if (!all_of(GatheredScalars, UndefValue::classof)) GatherCost += getGatherCost(GatheredScalars); return GatherCost; } if ((E->getOpcode() == Instruction::ExtractElement || all_of(E->Scalars, [](Value *V) { return isa(V); })) && allSameType(VL)) { // Check that gather of extractelements can be represented as just a // shuffle of a single/two vectors the scalars are extracted from. SmallVector Mask; std::optional ShuffleKind = isFixedVectorShuffle(VL, Mask); if (ShuffleKind) { // Found the bunch of extractelement instructions that must be gathered // into a vector and can be represented as a permutation elements in a // single input vector or of 2 input vectors. InstructionCost Cost = computeExtractCost(VL, VecTy, *ShuffleKind, Mask, *TTI); AdjustExtractsCost(Cost); if (NeedToShuffleReuses) Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices); return Cost; } } if (isSplat(VL)) { // Found the broadcasting of the single scalar, calculate the cost as the // broadcast. assert(VecTy == FinalVecTy && "No reused scalars expected for broadcast."); const auto *It = find_if(VL, [](Value *V) { return !isa(V); }); // If all values are undefs - consider cost free. if (It == VL.end()) return TTI::TCC_Free; // Add broadcast for non-identity shuffle only. bool NeedShuffle = VL.front() != *It || !all_of(VL.drop_front(), UndefValue::classof); InstructionCost InsertCost = TTI->getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind, /*Index=*/0, PoisonValue::get(VecTy), *It); return InsertCost + (NeedShuffle ? TTI->getShuffleCost( TargetTransformInfo::SK_Broadcast, VecTy, /*Mask=*/std::nullopt, CostKind, /*Index=*/0, /*SubTp=*/nullptr, /*Args=*/VL[0]) : TTI::TCC_Free); } InstructionCost ReuseShuffleCost = 0; if (NeedToShuffleReuses) ReuseShuffleCost = TTI->getShuffleCost( TTI::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices); // Improve gather cost for gather of loads, if we can group some of the // loads into vector loads. if (VL.size() > 2 && E->getOpcode() == Instruction::Load && !E->isAltShuffle()) { BoUpSLP::ValueSet VectorizedLoads; unsigned StartIdx = 0; unsigned VF = VL.size() / 2; unsigned VectorizedCnt = 0; unsigned ScatterVectorizeCnt = 0; const unsigned Sz = DL->getTypeSizeInBits(E->getMainOp()->getType()); for (unsigned MinVF = getMinVF(2 * Sz); VF >= MinVF; VF /= 2) { for (unsigned Cnt = StartIdx, End = VL.size(); Cnt + VF <= End; Cnt += VF) { ArrayRef Slice = VL.slice(Cnt, VF); if (!VectorizedLoads.count(Slice.front()) && !VectorizedLoads.count(Slice.back()) && allSameBlock(Slice)) { SmallVector PointerOps; OrdersType CurrentOrder; LoadsState LS = canVectorizeLoads(Slice, Slice.front(), *TTI, *DL, *SE, *LI, *TLI, CurrentOrder, PointerOps); switch (LS) { case LoadsState::Vectorize: case LoadsState::ScatterVectorize: // Mark the vectorized loads so that we don't vectorize them // again. if (LS == LoadsState::Vectorize) ++VectorizedCnt; else ++ScatterVectorizeCnt; VectorizedLoads.insert(Slice.begin(), Slice.end()); // If we vectorized initial block, no need to try to vectorize it // again. if (Cnt == StartIdx) StartIdx += VF; break; case LoadsState::Gather: break; } } } // Check if the whole array was vectorized already - exit. if (StartIdx >= VL.size()) break; // Found vectorizable parts - exit. if (!VectorizedLoads.empty()) break; } if (!VectorizedLoads.empty()) { InstructionCost GatherCost = 0; unsigned NumParts = TTI->getNumberOfParts(VecTy); bool NeedInsertSubvectorAnalysis = !NumParts || (VL.size() / VF) > NumParts; // Get the cost for gathered loads. for (unsigned I = 0, End = VL.size(); I < End; I += VF) { if (VectorizedLoads.contains(VL[I])) continue; GatherCost += getGatherCost(VL.slice(I, VF)); } // The cost for vectorized loads. InstructionCost ScalarsCost = 0; for (Value *V : VectorizedLoads) { auto *LI = cast(V); ScalarsCost += TTI->getMemoryOpCost(Instruction::Load, LI->getType(), LI->getAlign(), LI->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo(), LI); } auto *LI = cast(E->getMainOp()); auto *LoadTy = FixedVectorType::get(LI->getType(), VF); Align Alignment = LI->getAlign(); GatherCost += VectorizedCnt * TTI->getMemoryOpCost(Instruction::Load, LoadTy, Alignment, LI->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo(), LI); GatherCost += ScatterVectorizeCnt * TTI->getGatherScatterOpCost( Instruction::Load, LoadTy, LI->getPointerOperand(), /*VariableMask=*/false, Alignment, CostKind, LI); if (NeedInsertSubvectorAnalysis) { // Add the cost for the subvectors insert. for (int I = VF, E = VL.size(); I < E; I += VF) GatherCost += TTI->getShuffleCost(TTI::SK_InsertSubvector, VecTy, std::nullopt, CostKind, I, LoadTy); } return ReuseShuffleCost + GatherCost - ScalarsCost; } } return ReuseShuffleCost + getGatherCost(VL); } InstructionCost CommonCost = 0; SmallVector Mask; if (!E->ReorderIndices.empty()) { SmallVector NewMask; if (E->getOpcode() == Instruction::Store) { // For stores the order is actually a mask. NewMask.resize(E->ReorderIndices.size()); copy(E->ReorderIndices, NewMask.begin()); } else { inversePermutation(E->ReorderIndices, NewMask); } ::addMask(Mask, NewMask); } if (NeedToShuffleReuses) ::addMask(Mask, E->ReuseShuffleIndices); if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask)) CommonCost = TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FinalVecTy, Mask); assert((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state"); assert(E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && "Invalid VL"); Instruction *VL0 = E->getMainOp(); unsigned ShuffleOrOp = E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); const unsigned Sz = VL.size(); auto GetCostDiff = [=](function_ref ScalarEltCost, function_ref VectorCost) { // Calculate the cost of this instruction. InstructionCost ScalarCost = 0; if (isa(VL0)) { // For some of the instructions no need to calculate cost for each // particular instruction, we can use the cost of the single // instruction x total number of scalar instructions. ScalarCost = Sz * ScalarEltCost(0); } else { for (unsigned I = 0; I < Sz; ++I) ScalarCost += ScalarEltCost(I); } InstructionCost VecCost = VectorCost(CommonCost); LLVM_DEBUG( dumpTreeCosts(E, CommonCost, VecCost - CommonCost, ScalarCost)); // Disable warnings for `this` and `E` are unused. Required for // `dumpTreeCosts`. (void)this; (void)E; return VecCost - ScalarCost; }; // Calculate cost difference from vectorizing set of GEPs. // Negative value means vectorizing is profitable. auto GetGEPCostDiff = [=](ArrayRef Ptrs, Value *BasePtr) { InstructionCost CostSavings = 0; for (Value *V : Ptrs) { if (V == BasePtr) continue; auto *Ptr = dyn_cast(V); // GEPs may contain just addresses without instructions, considered free. // GEPs with all constant indices also considered to have zero cost. if (!Ptr || Ptr->hasAllConstantIndices()) continue; // Here we differentiate two cases: when GEPs represent a regular // vectorization tree node (and hence vectorized) and when the set is // arguments of a set of loads or stores being vectorized. In the former // case all the scalar GEPs will be removed as a result of vectorization. // For any external uses of some lanes extract element instructions will // be generated (which cost is estimated separately). For the latter case // since the set of GEPs itself is not vectorized those used more than // once will remain staying in vectorized code as well. So we should not // count them as savings. if (!Ptr->hasOneUse() && isa(VL0)) continue; // TODO: it is target dependent, so need to implement and then use a TTI // interface. CostSavings += TTI->getArithmeticInstrCost(Instruction::Add, Ptr->getType(), CostKind); } LLVM_DEBUG(dbgs() << "SLP: Calculated GEPs cost savings or Tree:\n"; E->dump()); LLVM_DEBUG(dbgs() << "SLP: GEP cost saving = " << CostSavings << "\n"); return InstructionCost() - CostSavings; }; switch (ShuffleOrOp) { case Instruction::PHI: { // Count reused scalars. InstructionCost ScalarCost = 0; SmallPtrSet CountedOps; for (Value *V : VL) { auto *PHI = dyn_cast(V); if (!PHI) continue; ValueList Operands(PHI->getNumIncomingValues(), nullptr); for (unsigned I = 0, N = PHI->getNumIncomingValues(); I < N; ++I) { Value *Op = PHI->getIncomingValue(I); Operands[I] = Op; } if (const TreeEntry *OpTE = getTreeEntry(Operands.front())) if (OpTE->isSame(Operands) && CountedOps.insert(OpTE).second) if (!OpTE->ReuseShuffleIndices.empty()) ScalarCost += TTI::TCC_Basic * (OpTE->ReuseShuffleIndices.size() - OpTE->Scalars.size()); } return CommonCost - ScalarCost; } case Instruction::ExtractValue: case Instruction::ExtractElement: { auto GetScalarCost = [=](unsigned Idx) { auto *I = cast(VL[Idx]); VectorType *SrcVecTy; if (ShuffleOrOp == Instruction::ExtractElement) { auto *EE = cast(I); SrcVecTy = EE->getVectorOperandType(); } else { auto *EV = cast(I); Type *AggregateTy = EV->getAggregateOperand()->getType(); unsigned NumElts; if (auto *ATy = dyn_cast(AggregateTy)) NumElts = ATy->getNumElements(); else NumElts = AggregateTy->getStructNumElements(); SrcVecTy = FixedVectorType::get(ScalarTy, NumElts); } if (I->hasOneUse()) { Instruction *Ext = I->user_back(); if ((isa(Ext) || isa(Ext)) && all_of(Ext->users(), [](User *U) { return isa(U); })) { // Use getExtractWithExtendCost() to calculate the cost of // extractelement/ext pair. InstructionCost Cost = TTI->getExtractWithExtendCost( Ext->getOpcode(), Ext->getType(), SrcVecTy, *getExtractIndex(I)); // Subtract the cost of s|zext which is subtracted separately. Cost -= TTI->getCastInstrCost( Ext->getOpcode(), Ext->getType(), I->getType(), TTI::getCastContextHint(Ext), CostKind, Ext); return Cost; } } return TTI->getVectorInstrCost(Instruction::ExtractElement, SrcVecTy, CostKind, *getExtractIndex(I)); }; auto GetVectorCost = [](InstructionCost CommonCost) { return CommonCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } case Instruction::InsertElement: { assert(E->ReuseShuffleIndices.empty() && "Unique insertelements only are expected."); auto *SrcVecTy = cast(VL0->getType()); unsigned const NumElts = SrcVecTy->getNumElements(); unsigned const NumScalars = VL.size(); unsigned NumOfParts = TTI->getNumberOfParts(SrcVecTy); SmallVector InsertMask(NumElts, UndefMaskElem); unsigned OffsetBeg = *getInsertIndex(VL.front()); unsigned OffsetEnd = OffsetBeg; InsertMask[OffsetBeg] = 0; for (auto [I, V] : enumerate(VL.drop_front())) { unsigned Idx = *getInsertIndex(V); if (OffsetBeg > Idx) OffsetBeg = Idx; else if (OffsetEnd < Idx) OffsetEnd = Idx; InsertMask[Idx] = I + 1; } unsigned VecScalarsSz = PowerOf2Ceil(NumElts); if (NumOfParts > 0) VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts); unsigned VecSz = (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) * VecScalarsSz; unsigned Offset = VecScalarsSz * (OffsetBeg / VecScalarsSz); unsigned InsertVecSz = std::min( PowerOf2Ceil(OffsetEnd - OffsetBeg + 1), ((OffsetEnd - OffsetBeg + VecScalarsSz) / VecScalarsSz) * VecScalarsSz); bool IsWholeSubvector = OffsetBeg == Offset && ((OffsetEnd + 1) % VecScalarsSz == 0); // Check if we can safely insert a subvector. If it is not possible, just // generate a whole-sized vector and shuffle the source vector and the new // subvector. if (OffsetBeg + InsertVecSz > VecSz) { // Align OffsetBeg to generate correct mask. OffsetBeg = alignDown(OffsetBeg, VecSz, Offset); InsertVecSz = VecSz; } APInt DemandedElts = APInt::getZero(NumElts); // TODO: Add support for Instruction::InsertValue. SmallVector Mask; if (!E->ReorderIndices.empty()) { inversePermutation(E->ReorderIndices, Mask); Mask.append(InsertVecSz - Mask.size(), UndefMaskElem); } else { Mask.assign(VecSz, UndefMaskElem); std::iota(Mask.begin(), std::next(Mask.begin(), InsertVecSz), 0); } bool IsIdentity = true; SmallVector PrevMask(InsertVecSz, UndefMaskElem); Mask.swap(PrevMask); for (unsigned I = 0; I < NumScalars; ++I) { unsigned InsertIdx = *getInsertIndex(VL[PrevMask[I]]); DemandedElts.setBit(InsertIdx); IsIdentity &= InsertIdx - OffsetBeg == I; Mask[InsertIdx - OffsetBeg] = I; } assert(Offset < NumElts && "Failed to find vector index offset"); InstructionCost Cost = 0; Cost -= TTI->getScalarizationOverhead(SrcVecTy, DemandedElts, /*Insert*/ true, /*Extract*/ false, CostKind); // First cost - resize to actual vector size if not identity shuffle or // need to shift the vector. // Do not calculate the cost if the actual size is the register size and // we can merge this shuffle with the following SK_Select. auto *InsertVecTy = FixedVectorType::get(SrcVecTy->getElementType(), InsertVecSz); if (!IsIdentity) Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, InsertVecTy, Mask); auto *FirstInsert = cast(*find_if(E->Scalars, [E](Value *V) { return !is_contained(E->Scalars, cast(V)->getOperand(0)); })); // Second cost - permutation with subvector, if some elements are from the // initial vector or inserting a subvector. // TODO: Implement the analysis of the FirstInsert->getOperand(0) // subvector of ActualVecTy. SmallBitVector InMask = isUndefVector(FirstInsert->getOperand(0), buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask)); if (!InMask.all() && NumScalars != NumElts && !IsWholeSubvector) { if (InsertVecSz != VecSz) { auto *ActualVecTy = FixedVectorType::get(SrcVecTy->getElementType(), VecSz); Cost += TTI->getShuffleCost(TTI::SK_InsertSubvector, ActualVecTy, std::nullopt, CostKind, OffsetBeg - Offset, InsertVecTy); } else { for (unsigned I = 0, End = OffsetBeg - Offset; I < End; ++I) Mask[I] = InMask.test(I) ? UndefMaskElem : I; for (unsigned I = OffsetBeg - Offset, End = OffsetEnd - Offset; I <= End; ++I) if (Mask[I] != UndefMaskElem) Mask[I] = I + VecSz; for (unsigned I = OffsetEnd + 1 - Offset; I < VecSz; ++I) Mask[I] = ((I >= InMask.size()) || InMask.test(I)) ? UndefMaskElem : I; Cost += TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, InsertVecTy, Mask); } } return Cost; } case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: case Instruction::FPToSI: case Instruction::FPExt: case Instruction::PtrToInt: case Instruction::IntToPtr: case Instruction::SIToFP: case Instruction::UIToFP: case Instruction::Trunc: case Instruction::FPTrunc: case Instruction::BitCast: { auto GetScalarCost = [=](unsigned Idx) { auto *VI = cast(VL[Idx]); return TTI->getCastInstrCost(E->getOpcode(), ScalarTy, VI->getOperand(0)->getType(), TTI::getCastContextHint(VI), CostKind, VI); }; auto GetVectorCost = [=](InstructionCost CommonCost) { Type *SrcTy = VL0->getOperand(0)->getType(); auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); InstructionCost VecCost = CommonCost; // Check if the values are candidates to demote. if (!MinBWs.count(VL0) || VecTy != SrcVecTy) VecCost += TTI->getCastInstrCost(E->getOpcode(), VecTy, SrcVecTy, TTI::getCastContextHint(VL0), CostKind, VL0); return VecCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } case Instruction::FCmp: case Instruction::ICmp: case Instruction::Select: { CmpInst::Predicate VecPred, SwappedVecPred; auto MatchCmp = m_Cmp(VecPred, m_Value(), m_Value()); if (match(VL0, m_Select(MatchCmp, m_Value(), m_Value())) || match(VL0, MatchCmp)) SwappedVecPred = CmpInst::getSwappedPredicate(VecPred); else SwappedVecPred = VecPred = ScalarTy->isFloatingPointTy() ? CmpInst::BAD_FCMP_PREDICATE : CmpInst::BAD_ICMP_PREDICATE; auto GetScalarCost = [&](unsigned Idx) { auto *VI = cast(VL[Idx]); CmpInst::Predicate CurrentPred = ScalarTy->isFloatingPointTy() ? CmpInst::BAD_FCMP_PREDICATE : CmpInst::BAD_ICMP_PREDICATE; auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value()); if ((!match(VI, m_Select(MatchCmp, m_Value(), m_Value())) && !match(VI, MatchCmp)) || (CurrentPred != VecPred && CurrentPred != SwappedVecPred)) VecPred = SwappedVecPred = ScalarTy->isFloatingPointTy() ? CmpInst::BAD_FCMP_PREDICATE : CmpInst::BAD_ICMP_PREDICATE; return TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(), CurrentPred, CostKind, VI); }; auto GetVectorCost = [&](InstructionCost CommonCost) { auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); InstructionCost VecCost = TTI->getCmpSelInstrCost( E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0); // Check if it is possible and profitable to use min/max for selects // in VL. // auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL); if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) { IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy, {VecTy, VecTy}); InstructionCost IntrinsicCost = TTI->getIntrinsicInstrCost(CostAttrs, CostKind); // If the selects are the only uses of the compares, they will be // dead and we can adjust the cost by removing their cost. if (IntrinsicAndUse.second) IntrinsicCost -= TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, MaskTy, VecPred, CostKind); VecCost = std::min(VecCost, IntrinsicCost); } return VecCost + CommonCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } case Instruction::FNeg: case Instruction::Add: case Instruction::FAdd: case Instruction::Sub: case Instruction::FSub: case Instruction::Mul: case Instruction::FMul: case Instruction::UDiv: case Instruction::SDiv: case Instruction::FDiv: case Instruction::URem: case Instruction::SRem: case Instruction::FRem: case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::And: case Instruction::Or: case Instruction::Xor: { auto GetScalarCost = [=](unsigned Idx) { auto *VI = cast(VL[Idx]); unsigned OpIdx = isa(VI) ? 0 : 1; TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(VI->getOperand(0)); TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(VI->getOperand(OpIdx)); SmallVector Operands(VI->operand_values()); return TTI->getArithmeticInstrCost(ShuffleOrOp, ScalarTy, CostKind, Op1Info, Op2Info, Operands, VI); }; auto GetVectorCost = [=](InstructionCost CommonCost) { unsigned OpIdx = isa(VL0) ? 0 : 1; TTI::OperandValueInfo Op1Info = getOperandInfo(VL, 0); TTI::OperandValueInfo Op2Info = getOperandInfo(VL, OpIdx); return TTI->getArithmeticInstrCost(ShuffleOrOp, VecTy, CostKind, Op1Info, Op2Info) + CommonCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } case Instruction::GetElementPtr: { return CommonCost + GetGEPCostDiff(VL, VL0); } case Instruction::Load: { auto GetScalarCost = [=](unsigned Idx) { auto *VI = cast(VL[Idx]); return TTI->getMemoryOpCost(Instruction::Load, ScalarTy, VI->getAlign(), VI->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo(), VI); }; auto *LI0 = cast(VL0); auto GetVectorCost = [=](InstructionCost CommonCost) { InstructionCost VecLdCost; if (E->State == TreeEntry::Vectorize) { VecLdCost = TTI->getMemoryOpCost( Instruction::Load, VecTy, LI0->getAlign(), LI0->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo()); } else { assert(E->State == TreeEntry::ScatterVectorize && "Unknown EntryState"); Align CommonAlignment = LI0->getAlign(); for (Value *V : VL) CommonAlignment = std::min(CommonAlignment, cast(V)->getAlign()); VecLdCost = TTI->getGatherScatterOpCost( Instruction::Load, VecTy, LI0->getPointerOperand(), /*VariableMask=*/false, CommonAlignment, CostKind); } return VecLdCost + CommonCost; }; InstructionCost Cost = GetCostDiff(GetScalarCost, GetVectorCost); // If this node generates masked gather load then it is not a terminal node. // Hence address operand cost is estimated separately. if (E->State == TreeEntry::ScatterVectorize) return Cost; // Estimate cost of GEPs since this tree node is a terminator. SmallVector PointerOps(VL.size()); for (auto [I, V] : enumerate(VL)) PointerOps[I] = cast(V)->getPointerOperand(); return Cost + GetGEPCostDiff(PointerOps, LI0->getPointerOperand()); } case Instruction::Store: { bool IsReorder = !E->ReorderIndices.empty(); auto GetScalarCost = [=](unsigned Idx) { auto *VI = cast(VL[Idx]); TTI::OperandValueInfo OpInfo = getOperandInfo(VI, 0); return TTI->getMemoryOpCost(Instruction::Store, ScalarTy, VI->getAlign(), VI->getPointerAddressSpace(), CostKind, OpInfo, VI); }; auto *BaseSI = cast(IsReorder ? VL[E->ReorderIndices.front()] : VL0); auto GetVectorCost = [=](InstructionCost CommonCost) { // We know that we can merge the stores. Calculate the cost. TTI::OperandValueInfo OpInfo = getOperandInfo(VL, 0); return TTI->getMemoryOpCost(Instruction::Store, VecTy, BaseSI->getAlign(), BaseSI->getPointerAddressSpace(), CostKind, OpInfo) + CommonCost; }; SmallVector PointerOps(VL.size()); for (auto [I, V] : enumerate(VL)) { unsigned Idx = IsReorder ? E->ReorderIndices[I] : I; PointerOps[Idx] = cast(V)->getPointerOperand(); } return GetCostDiff(GetScalarCost, GetVectorCost) + GetGEPCostDiff(PointerOps, BaseSI->getPointerOperand()); } case Instruction::Call: { auto GetScalarCost = [=](unsigned Idx) { auto *CI = cast(VL[Idx]); Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); if (ID != Intrinsic::not_intrinsic) { IntrinsicCostAttributes CostAttrs(ID, *CI, 1); return TTI->getIntrinsicInstrCost(CostAttrs, CostKind); } return TTI->getCallInstrCost(CI->getCalledFunction(), CI->getFunctionType()->getReturnType(), CI->getFunctionType()->params(), CostKind); }; auto GetVectorCost = [=](InstructionCost CommonCost) { auto *CI = cast(VL0); auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); return std::min(VecCallCosts.first, VecCallCosts.second) + CommonCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } case Instruction::ShuffleVector: { assert(E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa(VL0) && isa(E->getAltOp()))) && "Invalid Shuffle Vector Operand"); // Try to find the previous shuffle node with the same operands and same // main/alternate ops. auto TryFindNodeWithEqualOperands = [=]() { for (const std::unique_ptr &TE : VectorizableTree) { if (TE.get() == E) break; if (TE->isAltShuffle() && ((TE->getOpcode() == E->getOpcode() && TE->getAltOpcode() == E->getAltOpcode()) || (TE->getOpcode() == E->getAltOpcode() && TE->getAltOpcode() == E->getOpcode())) && TE->hasEqualOperands(*E)) return true; } return false; }; auto GetScalarCost = [=](unsigned Idx) { auto *VI = cast(VL[Idx]); assert(E->isOpcodeOrAlt(VI) && "Unexpected main/alternate opcode"); (void)E; return TTI->getInstructionCost(VI, CostKind); }; // Need to clear CommonCost since the final shuffle cost is included into // vector cost. auto GetVectorCost = [&](InstructionCost) { // VecCost is equal to sum of the cost of creating 2 vectors // and the cost of creating shuffle. InstructionCost VecCost = 0; if (TryFindNodeWithEqualOperands()) { LLVM_DEBUG({ dbgs() << "SLP: diamond match for alternate node found.\n"; E->dump(); }); // No need to add new vector costs here since we're going to reuse // same main/alternate vector ops, just do different shuffling. } else if (Instruction::isBinaryOp(E->getOpcode())) { VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); VecCost += TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, CostKind); } else if (auto *CI0 = dyn_cast(VL0)) { VecCost = TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(), CI0->getPredicate(), CostKind, VL0); VecCost += TTI->getCmpSelInstrCost( E->getOpcode(), ScalarTy, Builder.getInt1Ty(), cast(E->getAltOp())->getPredicate(), CostKind, E->getAltOp()); } else { Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, TTI::CastContextHint::None, CostKind); VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, TTI::CastContextHint::None, CostKind); } if (E->ReuseShuffleIndices.empty()) { VecCost += TTI->getShuffleCost(TargetTransformInfo::SK_Select, FinalVecTy); } else { SmallVector Mask; buildShuffleEntryMask( E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, [E](Instruction *I) { assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode"); return I->getOpcode() == E->getAltOpcode(); }, Mask); VecCost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, FinalVecTy, Mask); } return VecCost; }; return GetCostDiff(GetScalarCost, GetVectorCost); } default: llvm_unreachable("Unknown instruction"); } } bool BoUpSLP::isFullyVectorizableTinyTree(bool ForReduction) const { LLVM_DEBUG(dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n"); auto &&AreVectorizableGathers = [this](const TreeEntry *TE, unsigned Limit) { SmallVector Mask; return TE->State == TreeEntry::NeedToGather && !any_of(TE->Scalars, [this](Value *V) { return EphValues.contains(V); }) && (allConstant(TE->Scalars) || isSplat(TE->Scalars) || TE->Scalars.size() < Limit || ((TE->getOpcode() == Instruction::ExtractElement || all_of(TE->Scalars, [](Value *V) { return isa(V); })) && isFixedVectorShuffle(TE->Scalars, Mask)) || (TE->State == TreeEntry::NeedToGather && TE->getOpcode() == Instruction::Load && !TE->isAltShuffle())); }; // We only handle trees of heights 1 and 2. if (VectorizableTree.size() == 1 && (VectorizableTree[0]->State == TreeEntry::Vectorize || (ForReduction && AreVectorizableGathers(VectorizableTree[0].get(), VectorizableTree[0]->Scalars.size()) && VectorizableTree[0]->getVectorFactor() > 2))) return true; if (VectorizableTree.size() != 2) return false; // Handle splat and all-constants stores. Also try to vectorize tiny trees // with the second gather nodes if they have less scalar operands rather than // the initial tree element (may be profitable to shuffle the second gather) // or they are extractelements, which form shuffle. SmallVector Mask; if (VectorizableTree[0]->State == TreeEntry::Vectorize && AreVectorizableGathers(VectorizableTree[1].get(), VectorizableTree[0]->Scalars.size())) return true; // Gathering cost would be too much for tiny trees. if (VectorizableTree[0]->State == TreeEntry::NeedToGather || (VectorizableTree[1]->State == TreeEntry::NeedToGather && VectorizableTree[0]->State != TreeEntry::ScatterVectorize)) return false; return true; } static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, TargetTransformInfo *TTI, bool MustMatchOrInst) { // Look past the root to find a source value. Arbitrarily follow the // path through operand 0 of any 'or'. Also, peek through optional // shift-left-by-multiple-of-8-bits. Value *ZextLoad = Root; const APInt *ShAmtC; bool FoundOr = false; while (!isa(ZextLoad) && (match(ZextLoad, m_Or(m_Value(), m_Value())) || (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && ShAmtC->urem(8) == 0))) { auto *BinOp = cast(ZextLoad); ZextLoad = BinOp->getOperand(0); if (BinOp->getOpcode() == Instruction::Or) FoundOr = true; } // Check if the input is an extended load of the required or/shift expression. Value *Load; if ((MustMatchOrInst && !FoundOr) || ZextLoad == Root || !match(ZextLoad, m_ZExt(m_Value(Load))) || !isa(Load)) return false; // Require that the total load bit width is a legal integer type. // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. Type *SrcTy = Load->getType(); unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) return false; // Everything matched - assume that we can fold the whole sequence using // load combining. LLVM_DEBUG(dbgs() << "SLP: Assume load combining for tree starting at " << *(cast(Root)) << "\n"); return true; } bool BoUpSLP::isLoadCombineReductionCandidate(RecurKind RdxKind) const { if (RdxKind != RecurKind::Or) return false; unsigned NumElts = VectorizableTree[0]->Scalars.size(); Value *FirstReduced = VectorizableTree[0]->Scalars[0]; return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI, /* MatchOr */ false); } bool BoUpSLP::isLoadCombineCandidate() const { // Peek through a final sequence of stores and check if all operations are // likely to be load-combined. unsigned NumElts = VectorizableTree[0]->Scalars.size(); for (Value *Scalar : VectorizableTree[0]->Scalars) { Value *X; if (!match(Scalar, m_Store(m_Value(X), m_Value())) || !isLoadCombineCandidateImpl(X, NumElts, TTI, /* MatchOr */ true)) return false; } return true; } bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const { // No need to vectorize inserts of gathered values. if (VectorizableTree.size() == 2 && isa(VectorizableTree[0]->Scalars[0]) && VectorizableTree[1]->State == TreeEntry::NeedToGather && (VectorizableTree[1]->getVectorFactor() <= 2 || !(isSplat(VectorizableTree[1]->Scalars) || allConstant(VectorizableTree[1]->Scalars)))) return true; // We can vectorize the tree if its size is greater than or equal to the // minimum size specified by the MinTreeSize command line option. if (VectorizableTree.size() >= MinTreeSize) return false; // If we have a tiny tree (a tree whose size is less than MinTreeSize), we // can vectorize it if we can prove it fully vectorizable. if (isFullyVectorizableTinyTree(ForReduction)) return false; assert(VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users"); // Otherwise, we can't vectorize the tree. It is both tiny and not fully // vectorizable. return true; } InstructionCost BoUpSLP::getSpillCost() const { // 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(); InstructionCost Cost = 0; SmallPtrSet LiveValues; Instruction *PrevInst = nullptr; // The entries in VectorizableTree are not necessarily ordered by their // position in basic blocks. Collect them and order them by dominance so later // instructions are guaranteed to be visited first. For instructions in // different basic blocks, we only scan to the beginning of the block, so // their order does not matter, as long as all instructions in a basic block // are grouped together. Using dominance ensures a deterministic order. SmallVector OrderedScalars; for (const auto &TEPtr : VectorizableTree) { Instruction *Inst = dyn_cast(TEPtr->Scalars[0]); if (!Inst) continue; OrderedScalars.push_back(Inst); } llvm::sort(OrderedScalars, [&](Instruction *A, Instruction *B) { auto *NodeA = DT->getNode(A->getParent()); auto *NodeB = DT->getNode(B->getParent()); assert(NodeA && "Should only process reachable instructions"); assert(NodeB && "Should only process reachable instructions"); assert((NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) && "Different nodes should have different DFS numbers"); if (NodeA != NodeB) return NodeA->getDFSNumIn() < NodeB->getDFSNumIn(); return B->comesBefore(A); }); for (Instruction *Inst : OrderedScalars) { if (!PrevInst) { PrevInst = Inst; continue; } // Update LiveValues. LiveValues.erase(PrevInst); for (auto &J : PrevInst->operands()) { if (isa(&*J) && getTreeEntry(&*J)) LiveValues.insert(cast(&*J)); } LLVM_DEBUG({ dbgs() << "SLP: #LV: " << LiveValues.size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump(); }); // Now find the sequence of instructions between PrevInst and Inst. unsigned NumCalls = 0; BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), PrevInstIt = PrevInst->getIterator().getReverse(); while (InstIt != PrevInstIt) { if (PrevInstIt == PrevInst->getParent()->rend()) { PrevInstIt = Inst->getParent()->rbegin(); continue; } auto NoCallIntrinsic = [this](Instruction *I) { if (auto *II = dyn_cast(I)) { if (II->isAssumeLikeIntrinsic()) return true; FastMathFlags FMF; SmallVector Tys; for (auto &ArgOp : II->args()) Tys.push_back(ArgOp->getType()); if (auto *FPMO = dyn_cast(II)) FMF = FPMO->getFastMathFlags(); IntrinsicCostAttributes ICA(II->getIntrinsicID(), II->getType(), Tys, FMF); InstructionCost IntrCost = TTI->getIntrinsicInstrCost(ICA, TTI::TCK_RecipThroughput); InstructionCost CallCost = TTI->getCallInstrCost( nullptr, II->getType(), Tys, TTI::TCK_RecipThroughput); if (IntrCost < CallCost) return true; } return false; }; // Debug information does not impact spill cost. if (isa(&*PrevInstIt) && !NoCallIntrinsic(&*PrevInstIt) && &*PrevInstIt != PrevInst) NumCalls++; ++PrevInstIt; } if (NumCalls) { SmallVector V; for (auto *II : LiveValues) { auto *ScalarTy = II->getType(); if (auto *VectorTy = dyn_cast(ScalarTy)) ScalarTy = VectorTy->getElementType(); V.push_back(FixedVectorType::get(ScalarTy, BundleWidth)); } Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); } PrevInst = Inst; } return Cost; } /// Checks if the \p IE1 instructions is followed by \p IE2 instruction in the /// buildvector sequence. static bool isFirstInsertElement(const InsertElementInst *IE1, const InsertElementInst *IE2) { if (IE1 == IE2) return false; const auto *I1 = IE1; const auto *I2 = IE2; const InsertElementInst *PrevI1; const InsertElementInst *PrevI2; unsigned Idx1 = *getInsertIndex(IE1); unsigned Idx2 = *getInsertIndex(IE2); do { if (I2 == IE1) return true; if (I1 == IE2) return false; PrevI1 = I1; PrevI2 = I2; if (I1 && (I1 == IE1 || I1->hasOneUse()) && getInsertIndex(I1).value_or(Idx2) != Idx2) I1 = dyn_cast(I1->getOperand(0)); if (I2 && ((I2 == IE2 || I2->hasOneUse())) && getInsertIndex(I2).value_or(Idx1) != Idx1) I2 = dyn_cast(I2->getOperand(0)); } while ((I1 && PrevI1 != I1) || (I2 && PrevI2 != I2)); llvm_unreachable("Two different buildvectors not expected."); } namespace { /// Returns incoming Value *, if the requested type is Value * too, or a default /// value, otherwise. struct ValueSelect { template static std::enable_if_t, Value *> get(Value *V) { return V; } template static std::enable_if_t, U> get(Value *) { return U(); } }; } // namespace /// Does the analysis of the provided shuffle masks and performs the requested /// actions on the vectors with the given shuffle masks. It tries to do it in /// several steps. /// 1. If the Base vector is not undef vector, resizing the very first mask to /// have common VF and perform action for 2 input vectors (including non-undef /// Base). Other shuffle masks are combined with the resulting after the 1 stage /// and processed as a shuffle of 2 elements. /// 2. If the Base is undef vector and have only 1 shuffle mask, perform the /// action only for 1 vector with the given mask, if it is not the identity /// mask. /// 3. If > 2 masks are used, perform the remaining shuffle actions for 2 /// vectors, combing the masks properly between the steps. template static T *performExtractsShuffleAction( MutableArrayRef>> ShuffleMask, Value *Base, function_ref GetVF, function_ref(T *, ArrayRef, bool)> ResizeAction, function_ref, ArrayRef)> Action) { assert(!ShuffleMask.empty() && "Empty list of shuffles for inserts."); SmallVector Mask(ShuffleMask.begin()->second); auto VMIt = std::next(ShuffleMask.begin()); T *Prev = nullptr; SmallBitVector UseMask = buildUseMask(Mask.size(), Mask, UseMask::UndefsAsMask); SmallBitVector IsBaseUndef = isUndefVector(Base, UseMask); if (!IsBaseUndef.all()) { // Base is not undef, need to combine it with the next subvectors. std::pair Res = ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/false); SmallBitVector IsBasePoison = isUndefVector(Base, UseMask); for (unsigned Idx = 0, VF = Mask.size(); Idx < VF; ++Idx) { if (Mask[Idx] == UndefMaskElem) Mask[Idx] = IsBasePoison.test(Idx) ? UndefMaskElem : Idx; else Mask[Idx] = (Res.second ? Idx : Mask[Idx]) + VF; } auto *V = ValueSelect::get(Base); (void)V; assert((!V || GetVF(V) == Mask.size()) && "Expected base vector of VF number of elements."); Prev = Action(Mask, {nullptr, Res.first}); } else if (ShuffleMask.size() == 1) { // Base is undef and only 1 vector is shuffled - perform the action only for // single vector, if the mask is not the identity mask. std::pair Res = ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/true); if (Res.second) // Identity mask is found. Prev = Res.first; else Prev = Action(Mask, {ShuffleMask.begin()->first}); } else { // Base is undef and at least 2 input vectors shuffled - perform 2 vectors // shuffles step by step, combining shuffle between the steps. unsigned Vec1VF = GetVF(ShuffleMask.begin()->first); unsigned Vec2VF = GetVF(VMIt->first); if (Vec1VF == Vec2VF) { // No need to resize the input vectors since they are of the same size, we // can shuffle them directly. ArrayRef SecMask = VMIt->second; for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { if (SecMask[I] != UndefMaskElem) { assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars."); Mask[I] = SecMask[I] + Vec1VF; } } Prev = Action(Mask, {ShuffleMask.begin()->first, VMIt->first}); } else { // Vectors of different sizes - resize and reshuffle. std::pair Res1 = ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/false); std::pair Res2 = ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); ArrayRef SecMask = VMIt->second; for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { if (Mask[I] != UndefMaskElem) { assert(SecMask[I] == UndefMaskElem && "Multiple uses of scalars."); if (Res1.second) Mask[I] = I; } else if (SecMask[I] != UndefMaskElem) { assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars."); Mask[I] = (Res2.second ? I : SecMask[I]) + VF; } } Prev = Action(Mask, {Res1.first, Res2.first}); } VMIt = std::next(VMIt); } bool IsBaseNotUndef = !IsBaseUndef.all(); (void)IsBaseNotUndef; // Perform requested actions for the remaining masks/vectors. for (auto E = ShuffleMask.end(); VMIt != E; ++VMIt) { // Shuffle other input vectors, if any. std::pair Res = ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); ArrayRef SecMask = VMIt->second; for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { if (SecMask[I] != UndefMaskElem) { assert((Mask[I] == UndefMaskElem || IsBaseNotUndef) && "Multiple uses of scalars."); Mask[I] = (Res.second ? I : SecMask[I]) + VF; } else if (Mask[I] != UndefMaskElem) { Mask[I] = I; } } Prev = Action(Mask, {Prev, Res.first}); } return Prev; } InstructionCost BoUpSLP::getTreeCost(ArrayRef VectorizedVals) { InstructionCost Cost = 0; LLVM_DEBUG(dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"); unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { TreeEntry &TE = *VectorizableTree[I]; if (TE.State == TreeEntry::NeedToGather) { if (const TreeEntry *E = getTreeEntry(TE.getMainOp()); E && E->getVectorFactor() == TE.getVectorFactor() && E->isSame(TE.Scalars)) { // Some gather nodes might be absolutely the same as some vectorizable // nodes after reordering, need to handle it. LLVM_DEBUG(dbgs() << "SLP: Adding cost 0 for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"); continue; } } InstructionCost C = getEntryCost(&TE, VectorizedVals); Cost += C; LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"); } SmallPtrSet ExtractCostCalculated; InstructionCost ExtractCost = 0; SmallVector>> ShuffleMasks; SmallVector> FirstUsers; SmallVector DemandedElts; for (ExternalUser &EU : ExternalUses) { // We only add extract cost once for the same scalar. if (!isa_and_nonnull(EU.User) && !ExtractCostCalculated.insert(EU.Scalar).second) continue; // 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; // No extract cost for vector "scalar" if (isa(EU.Scalar->getType())) continue; // If found user is an insertelement, do not calculate extract cost but try // to detect it as a final shuffled/identity match. if (auto *VU = dyn_cast_or_null(EU.User)) { if (auto *FTy = dyn_cast(VU->getType())) { std::optional InsertIdx = getInsertIndex(VU); if (InsertIdx) { const TreeEntry *ScalarTE = getTreeEntry(EU.Scalar); auto *It = find_if( FirstUsers, [this, VU](const std::pair &Pair) { return areTwoInsertFromSameBuildVector( VU, cast(Pair.first), [this](InsertElementInst *II) -> Value * { Value *Op0 = II->getOperand(0); if (getTreeEntry(II) && !getTreeEntry(Op0)) return nullptr; return Op0; }); }); int VecId = -1; if (It == FirstUsers.end()) { (void)ShuffleMasks.emplace_back(); SmallVectorImpl &Mask = ShuffleMasks.back()[ScalarTE]; if (Mask.empty()) Mask.assign(FTy->getNumElements(), UndefMaskElem); // Find the insertvector, vectorized in tree, if any. Value *Base = VU; while (auto *IEBase = dyn_cast(Base)) { if (IEBase != EU.User && (!IEBase->hasOneUse() || getInsertIndex(IEBase).value_or(*InsertIdx) == *InsertIdx)) break; // Build the mask for the vectorized insertelement instructions. if (const TreeEntry *E = getTreeEntry(IEBase)) { VU = IEBase; do { IEBase = cast(Base); int Idx = *getInsertIndex(IEBase); assert(Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already."); Mask[Idx] = Idx; Base = IEBase->getOperand(0); } while (E == getTreeEntry(Base)); break; } Base = cast(Base)->getOperand(0); } FirstUsers.emplace_back(VU, ScalarTE); DemandedElts.push_back(APInt::getZero(FTy->getNumElements())); VecId = FirstUsers.size() - 1; } else { if (isFirstInsertElement(VU, cast(It->first))) It->first = VU; VecId = std::distance(FirstUsers.begin(), It); } int InIdx = *InsertIdx; SmallVectorImpl &Mask = ShuffleMasks[VecId][ScalarTE]; if (Mask.empty()) Mask.assign(FTy->getNumElements(), UndefMaskElem); Mask[InIdx] = EU.Lane; DemandedElts[VecId].setBit(InIdx); continue; } } } // If we plan to rewrite the tree in a smaller type, we will need to sign // extend the extracted value back to the original type. Here, we account // for the extract and the added cost of the sign extend if needed. auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; if (MinBWs.count(ScalarRoot)) { auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); auto Extend = MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; VecTy = FixedVectorType::get(MinTy, BundleWidth); ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), VecTy, EU.Lane); } else { ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind, EU.Lane); } } InstructionCost SpillCost = getSpillCost(); Cost += SpillCost + ExtractCost; auto &&ResizeToVF = [this, &Cost](const TreeEntry *TE, ArrayRef Mask, bool) { InstructionCost C = 0; unsigned VF = Mask.size(); unsigned VecVF = TE->getVectorFactor(); if (VF != VecVF && (any_of(Mask, [VF](int Idx) { return Idx >= static_cast(VF); }) || (all_of(Mask, [VF](int Idx) { return Idx < 2 * static_cast(VF); }) && !ShuffleVectorInst::isIdentityMask(Mask)))) { SmallVector OrigMask(VecVF, UndefMaskElem); std::copy(Mask.begin(), std::next(Mask.begin(), std::min(VF, VecVF)), OrigMask.begin()); C = TTI->getShuffleCost( TTI::SK_PermuteSingleSrc, FixedVectorType::get(TE->getMainOp()->getType(), VecVF), OrigMask); LLVM_DEBUG( dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users.\n"; TE->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"); Cost += C; return std::make_pair(TE, true); } return std::make_pair(TE, false); }; // Calculate the cost of the reshuffled vectors, if any. for (int I = 0, E = FirstUsers.size(); I < E; ++I) { Value *Base = cast(FirstUsers[I].first)->getOperand(0); unsigned VF = ShuffleMasks[I].begin()->second.size(); auto *FTy = FixedVectorType::get( cast(FirstUsers[I].first->getType())->getElementType(), VF); auto Vector = ShuffleMasks[I].takeVector(); auto &&EstimateShufflesCost = [this, FTy, &Cost](ArrayRef Mask, ArrayRef TEs) { assert((TEs.size() == 1 || TEs.size() == 2) && "Expected exactly 1 or 2 tree entries."); if (TEs.size() == 1) { int Limit = 2 * Mask.size(); if (!all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) || !ShuffleVectorInst::isIdentityMask(Mask)) { InstructionCost C = TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FTy, Mask); LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement " "external users.\n"; TEs.front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"); Cost += C; } } else { InstructionCost C = TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, FTy, Mask); LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users.\n"; if (TEs.front()) { TEs.front()->dump(); } TEs.back()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"); Cost += C; } return TEs.back(); }; (void)performExtractsShuffleAction( MutableArrayRef(Vector.data(), Vector.size()), Base, [](const TreeEntry *E) { return E->getVectorFactor(); }, ResizeToVF, EstimateShufflesCost); InstructionCost InsertCost = TTI->getScalarizationOverhead( cast(FirstUsers[I].first->getType()), DemandedElts[I], /*Insert*/ true, /*Extract*/ false, TTI::TCK_RecipThroughput); Cost -= InsertCost; } #ifndef NDEBUG SmallString<256> Str; { raw_svector_ostream OS(Str); OS << "SLP: Spill Cost = " << SpillCost << ".\n" << "SLP: Extract Cost = " << ExtractCost << ".\n" << "SLP: Total Cost = " << Cost << ".\n"; } LLVM_DEBUG(dbgs() << Str); if (ViewSLPTree) ViewGraph(this, "SLP" + F->getName(), false, Str); #endif return Cost; } std::optional BoUpSLP::isGatherShuffledEntry(const TreeEntry *TE, ArrayRef VL, SmallVectorImpl &Mask, SmallVectorImpl &Entries) { Entries.clear(); // No need to check for the topmost gather node. if (TE == VectorizableTree.front().get()) return std::nullopt; Mask.assign(VL.size(), UndefMaskElem); assert(TE->UserTreeIndices.size() == 1 && "Expected only single user of the gather node."); // TODO: currently checking only for Scalars in the tree entry, need to count // reused elements too for better cost estimation. Instruction &UserInst = getLastInstructionInBundle(TE->UserTreeIndices.front().UserTE); auto *PHI = dyn_cast(&UserInst); auto *NodeUI = DT->getNode( PHI ? PHI->getIncomingBlock(TE->UserTreeIndices.front().EdgeIdx) : UserInst.getParent()); assert(NodeUI && "Should only process reachable instructions"); SmallPtrSet GatheredScalars(VL.begin(), VL.end()); auto CheckOrdering = [&](Instruction *LastEI) { // Check if the user node of the TE comes after user node of EntryPtr, // otherwise EntryPtr depends on TE. // Gather nodes usually are not scheduled and inserted before their first // user node. So, instead of checking dependency between the gather nodes // themselves, we check the dependency between their user nodes. // If one user node comes before the second one, we cannot use the second // gather node as the source vector for the first gather node, because in // the list of instructions it will be emitted later. auto *EntryParent = LastEI->getParent(); auto *NodeEUI = DT->getNode(EntryParent); if (!NodeEUI) return false; assert((NodeUI == NodeEUI) == (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) && "Different nodes should have different DFS numbers"); // Check the order of the gather nodes users. if (UserInst.getParent() != EntryParent && (DT->dominates(NodeUI, NodeEUI) || !DT->dominates(NodeEUI, NodeUI))) return false; if (UserInst.getParent() == EntryParent && UserInst.comesBefore(LastEI)) return false; return true; }; // Build a lists of values to tree entries. DenseMap> ValueToTEs; for (const std::unique_ptr &EntryPtr : VectorizableTree) { if (EntryPtr.get() == TE) continue; if (EntryPtr->State != TreeEntry::NeedToGather) continue; if (!any_of(EntryPtr->Scalars, [&GatheredScalars](Value *V) { return GatheredScalars.contains(V); })) continue; assert(EntryPtr->UserTreeIndices.size() == 1 && "Expected only single user of the gather node."); Instruction &EntryUserInst = getLastInstructionInBundle(EntryPtr->UserTreeIndices.front().UserTE); if (&UserInst == &EntryUserInst) { // If 2 gathers are operands of the same entry, compare operands indices, // use the earlier one as the base. if (TE->UserTreeIndices.front().UserTE == EntryPtr->UserTreeIndices.front().UserTE && TE->UserTreeIndices.front().EdgeIdx < EntryPtr->UserTreeIndices.front().EdgeIdx) continue; } // Check if the user node of the TE comes after user node of EntryPtr, // otherwise EntryPtr depends on TE. auto *EntryPHI = dyn_cast(&EntryUserInst); auto *EntryI = EntryPHI ? EntryPHI ->getIncomingBlock(EntryPtr->UserTreeIndices.front().EdgeIdx) ->getTerminator() : &EntryUserInst; if (!CheckOrdering(EntryI)) continue; for (Value *V : EntryPtr->Scalars) if (!isConstant(V)) ValueToTEs.try_emplace(V).first->getSecond().insert(EntryPtr.get()); } // Find all tree entries used by the gathered values. If no common entries // found - not a shuffle. // Here we build a set of tree nodes for each gathered value and trying to // find the intersection between these sets. If we have at least one common // tree node for each gathered value - we have just a permutation of the // single vector. If we have 2 different sets, we're in situation where we // have a permutation of 2 input vectors. SmallVector> UsedTEs; DenseMap UsedValuesEntry; for (Value *V : TE->Scalars) { if (isConstant(V)) continue; // Build a list of tree entries where V is used. SmallPtrSet VToTEs; auto It = ValueToTEs.find(V); if (It != ValueToTEs.end()) VToTEs = It->second; if (const TreeEntry *VTE = getTreeEntry(V)) VToTEs.insert(VTE); if (VToTEs.empty()) continue; if (UsedTEs.empty()) { // The first iteration, just insert the list of nodes to vector. UsedTEs.push_back(VToTEs); UsedValuesEntry.try_emplace(V, 0); } else { // Need to check if there are any previously used tree nodes which use V. // If there are no such nodes, consider that we have another one input // vector. SmallPtrSet SavedVToTEs(VToTEs); unsigned Idx = 0; for (SmallPtrSet &Set : UsedTEs) { // Do we have a non-empty intersection of previously listed tree entries // and tree entries using current V? set_intersect(VToTEs, Set); if (!VToTEs.empty()) { // Yes, write the new subset and continue analysis for the next // scalar. Set.swap(VToTEs); break; } VToTEs = SavedVToTEs; ++Idx; } // No non-empty intersection found - need to add a second set of possible // source vectors. if (Idx == UsedTEs.size()) { // If the number of input vectors is greater than 2 - not a permutation, // fallback to the regular gather. // TODO: support multiple reshuffled nodes. if (UsedTEs.size() == 2) continue; UsedTEs.push_back(SavedVToTEs); Idx = UsedTEs.size() - 1; } UsedValuesEntry.try_emplace(V, Idx); } } if (UsedTEs.empty()) return std::nullopt; unsigned VF = 0; if (UsedTEs.size() == 1) { // Keep the order to avoid non-determinism. SmallVector FirstEntries(UsedTEs.front().begin(), UsedTEs.front().end()); sort(FirstEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) { return TE1->Idx < TE2->Idx; }); // Try to find the perfect match in another gather node at first. auto *It = find_if(FirstEntries, [=](const TreeEntry *EntryPtr) { return EntryPtr->isSame(VL) || EntryPtr->isSame(TE->Scalars); }); if (It != FirstEntries.end()) { Entries.push_back(*It); std::iota(Mask.begin(), Mask.end(), 0); // Clear undef scalars. for (int I = 0, Sz = VL.size(); I < Sz; ++I) if (isa(TE->Scalars[I])) Mask[I] = UndefMaskElem; return TargetTransformInfo::SK_PermuteSingleSrc; } // No perfect match, just shuffle, so choose the first tree node from the // tree. Entries.push_back(FirstEntries.front()); } else { // Try to find nodes with the same vector factor. assert(UsedTEs.size() == 2 && "Expected at max 2 permuted entries."); // Keep the order of tree nodes to avoid non-determinism. DenseMap VFToTE; for (const TreeEntry *TE : UsedTEs.front()) { unsigned VF = TE->getVectorFactor(); auto It = VFToTE.find(VF); if (It != VFToTE.end()) { if (It->second->Idx > TE->Idx) It->getSecond() = TE; continue; } VFToTE.try_emplace(VF, TE); } // Same, keep the order to avoid non-determinism. SmallVector SecondEntries(UsedTEs.back().begin(), UsedTEs.back().end()); sort(SecondEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) { return TE1->Idx < TE2->Idx; }); for (const TreeEntry *TE : SecondEntries) { auto It = VFToTE.find(TE->getVectorFactor()); if (It != VFToTE.end()) { VF = It->first; Entries.push_back(It->second); Entries.push_back(TE); break; } } // No 2 source vectors with the same vector factor - give up and do regular // gather. if (Entries.empty()) return std::nullopt; } bool IsSplatOrUndefs = isSplat(VL) || all_of(VL, UndefValue::classof); // Checks if the 2 PHIs are compatible in terms of high possibility to be // vectorized. auto AreCompatiblePHIs = [&](Value *V, Value *V1) { auto *PHI = cast(V); auto *PHI1 = cast(V1); // Check that all incoming values are compatible/from same parent (if they // are instructions). // The incoming values are compatible if they all are constants, or // instruction with the same/alternate opcodes from the same basic block. for (int I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) { Value *In = PHI->getIncomingValue(I); Value *In1 = PHI1->getIncomingValue(I); if (isConstant(In) && isConstant(In1)) continue; if (!getSameOpcode({In, In1}, *TLI).getOpcode()) return false; if (cast(In)->getParent() != cast(In1)->getParent()) return false; } return true; }; // Check if the value can be ignored during analysis for shuffled gathers. // We suppose it is better to ignore instruction, which do not form splats, // are not vectorized/not extractelements (these instructions will be handled // by extractelements processing) or may form vector node in future. auto MightBeIgnored = [=](Value *V) { auto *I = dyn_cast(V); SmallVector IgnoredVals; if (UserIgnoreList) IgnoredVals.assign(UserIgnoreList->begin(), UserIgnoreList->end()); return I && !IsSplatOrUndefs && !ScalarToTreeEntry.count(I) && !isVectorLikeInstWithConstOps(I) && !areAllUsersVectorized(I, IgnoredVals) && isSimple(I); }; // Check that the neighbor instruction may form a full vector node with the // current instruction V. It is possible, if they have same/alternate opcode // and same parent basic block. auto NeighborMightBeIgnored = [&](Value *V, int Idx) { Value *V1 = VL[Idx]; bool UsedInSameVTE = false; auto It = UsedValuesEntry.find(V1); if (It != UsedValuesEntry.end()) UsedInSameVTE = It->second == UsedValuesEntry.find(V)->second; return V != V1 && MightBeIgnored(V1) && !UsedInSameVTE && getSameOpcode({V, V1}, *TLI).getOpcode() && cast(V)->getParent() == cast(V1)->getParent() && (!isa(V1) || AreCompatiblePHIs(V, V1)); }; // Build a shuffle mask for better cost estimation and vector emission. SmallBitVector UsedIdxs(Entries.size()); SmallVector> EntryLanes; for (int I = 0, E = VL.size(); I < E; ++I) { Value *V = VL[I]; auto It = UsedValuesEntry.find(V); if (It == UsedValuesEntry.end()) continue; // Do not try to shuffle scalars, if they are constants, or instructions // that can be vectorized as a result of the following vector build // vectorization. if (isConstant(V) || (MightBeIgnored(V) && ((I > 0 && NeighborMightBeIgnored(V, I - 1)) || (I != E - 1 && NeighborMightBeIgnored(V, I + 1))))) continue; unsigned Idx = It->second; EntryLanes.emplace_back(Idx, I); UsedIdxs.set(Idx); } // Iterate through all shuffled scalars and select entries, which can be used // for final shuffle. SmallVector TempEntries; for (unsigned I = 0, Sz = Entries.size(); I < Sz; ++I) { if (!UsedIdxs.test(I)) continue; // Fix the entry number for the given scalar. If it is the first entry, set // Pair.first to 0, otherwise to 1 (currently select at max 2 nodes). // These indices are used when calculating final shuffle mask as the vector // offset. for (std::pair &Pair : EntryLanes) if (Pair.first == I) Pair.first = TempEntries.size(); TempEntries.push_back(Entries[I]); } Entries.swap(TempEntries); if (EntryLanes.size() == Entries.size() && !VL.equals(TE->Scalars)) { // We may have here 1 or 2 entries only. If the number of scalars is equal // to the number of entries, no need to do the analysis, it is not very // profitable. Since VL is not the same as TE->Scalars, it means we already // have some shuffles before. Cut off not profitable case. Entries.clear(); return std::nullopt; } // Build the final mask, check for the identity shuffle, if possible. bool IsIdentity = Entries.size() == 1; // Pair.first is the offset to the vector, while Pair.second is the index of // scalar in the list. for (const std::pair &Pair : EntryLanes) { Mask[Pair.second] = Pair.first * VF + Entries[Pair.first]->findLaneForValue(VL[Pair.second]); IsIdentity &= Mask[Pair.second] == Pair.second; } switch (Entries.size()) { case 1: if (IsIdentity || EntryLanes.size() > 1 || VL.size() <= 2) return TargetTransformInfo::SK_PermuteSingleSrc; break; case 2: if (EntryLanes.size() > 2 || VL.size() <= 2) return TargetTransformInfo::SK_PermuteTwoSrc; break; default: break; } Entries.clear(); return std::nullopt; } InstructionCost BoUpSLP::getGatherCost(FixedVectorType *Ty, const APInt &ShuffledIndices, bool NeedToShuffle) const { TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; InstructionCost Cost = TTI->getScalarizationOverhead(Ty, ~ShuffledIndices, /*Insert*/ true, /*Extract*/ false, CostKind); if (NeedToShuffle) Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); return Cost; } InstructionCost BoUpSLP::getGatherCost(ArrayRef VL) const { // Find the type of the operands in VL. Type *ScalarTy = VL[0]->getType(); if (StoreInst *SI = dyn_cast(VL[0])) ScalarTy = SI->getValueOperand()->getType(); auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); bool DuplicateNonConst = false; // Find the cost of inserting/extracting values from the vector. // Check if the same elements are inserted several times and count them as // shuffle candidates. APInt ShuffledElements = APInt::getZero(VL.size()); DenseSet UniqueElements; // Iterate in reverse order to consider insert elements with the high cost. for (unsigned I = VL.size(); I > 0; --I) { unsigned Idx = I - 1; // No need to shuffle duplicates for constants. if (isConstant(VL[Idx])) { ShuffledElements.setBit(Idx); continue; } if (!UniqueElements.insert(VL[Idx]).second) { DuplicateNonConst = true; ShuffledElements.setBit(Idx); } } return getGatherCost(VecTy, ShuffledElements, DuplicateNonConst); } // Perform operand reordering on the instructions in VL and return the reordered // operands in Left and Right. void BoUpSLP::reorderInputsAccordingToOpcode( ArrayRef VL, SmallVectorImpl &Left, SmallVectorImpl &Right, const TargetLibraryInfo &TLI, const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R) { if (VL.empty()) return; VLOperands Ops(VL, TLI, DL, SE, R); // Reorder the operands in place. Ops.reorder(); Left = Ops.getVL(0); Right = Ops.getVL(1); } Instruction &BoUpSLP::getLastInstructionInBundle(const TreeEntry *E) { // Get the basic block this bundle is in. All instructions in the bundle // should be in this block (except for extractelement-like instructions with // constant indeces). auto *Front = E->getMainOp(); auto *BB = Front->getParent(); assert(llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa(V)) return true; auto *I = cast(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })); auto &&FindLastInst = [E, Front, this, &BB]() { Instruction *LastInst = Front; for (Value *V : E->Scalars) { auto *I = dyn_cast(V); if (!I) continue; if (LastInst->getParent() == I->getParent()) { if (LastInst->comesBefore(I)) LastInst = I; continue; } assert(isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I) && "Expected vector-like insts only."); if (!DT->isReachableFromEntry(LastInst->getParent())) { LastInst = I; continue; } if (!DT->isReachableFromEntry(I->getParent())) continue; auto *NodeA = DT->getNode(LastInst->getParent()); auto *NodeB = DT->getNode(I->getParent()); assert(NodeA && "Should only process reachable instructions"); assert(NodeB && "Should only process reachable instructions"); assert((NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) && "Different nodes should have different DFS numbers"); if (NodeA->getDFSNumIn() < NodeB->getDFSNumIn()) LastInst = I; } BB = LastInst->getParent(); return LastInst; }; auto &&FindFirstInst = [E, Front, this]() { Instruction *FirstInst = Front; for (Value *V : E->Scalars) { auto *I = dyn_cast(V); if (!I) continue; if (FirstInst->getParent() == I->getParent()) { if (I->comesBefore(FirstInst)) FirstInst = I; continue; } assert(isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I) && "Expected vector-like insts only."); if (!DT->isReachableFromEntry(FirstInst->getParent())) { FirstInst = I; continue; } if (!DT->isReachableFromEntry(I->getParent())) continue; auto *NodeA = DT->getNode(FirstInst->getParent()); auto *NodeB = DT->getNode(I->getParent()); assert(NodeA && "Should only process reachable instructions"); assert(NodeB && "Should only process reachable instructions"); assert((NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) && "Different nodes should have different DFS numbers"); if (NodeA->getDFSNumIn() > NodeB->getDFSNumIn()) FirstInst = I; } return FirstInst; }; // Set the insert point to the beginning of the basic block if the entry // should not be scheduled. if (E->State != TreeEntry::NeedToGather && (doesNotNeedToSchedule(E->Scalars) || all_of(E->Scalars, isVectorLikeInstWithConstOps))) { Instruction *InsertInst; if (all_of(E->Scalars, [](Value *V) { return !isVectorLikeInstWithConstOps(V) && isUsedOutsideBlock(V); })) InsertInst = FindLastInst(); else InsertInst = FindFirstInst(); return *InsertInst; } // The last instruction in the bundle in program order. Instruction *LastInst = nullptr; // Find the last instruction. The common case should be that BB has been // scheduled, and the last instruction is VL.back(). So we start with // VL.back() and iterate over schedule data until we reach the end of the // bundle. The end of the bundle is marked by null ScheduleData. if (BlocksSchedules.count(BB)) { Value *V = E->isOneOf(E->Scalars.back()); if (doesNotNeedToBeScheduled(V)) V = *find_if_not(E->Scalars, doesNotNeedToBeScheduled); auto *Bundle = BlocksSchedules[BB]->getScheduleData(V); if (Bundle && Bundle->isPartOfBundle()) for (; Bundle; Bundle = Bundle->NextInBundle) if (Bundle->OpValue == Bundle->Inst) LastInst = Bundle->Inst; } // LastInst can still be null at this point if there's either not an entry // for BB in BlocksSchedules or there's no ScheduleData available for // VL.back(). This can be the case if buildTree_rec aborts for various // reasons (e.g., the maximum recursion depth is reached, the maximum region // size is reached, etc.). ScheduleData is initialized in the scheduling // "dry-run". // // If this happens, we can still find the last instruction by brute force. We // iterate forwards from Front (inclusive) until we either see all // instructions in the bundle or reach the end of the block. If Front is the // last instruction in program order, LastInst will be set to Front, and we // will visit all the remaining instructions in the block. // // One of the reasons we exit early from buildTree_rec is to place an upper // bound on compile-time. Thus, taking an additional compile-time hit here is // not ideal. However, this should be exceedingly rare since it requires that // we both exit early from buildTree_rec and that the bundle be out-of-order // (causing us to iterate all the way to the end of the block). if (!LastInst) LastInst = FindLastInst(); assert(LastInst && "Failed to find last instruction in bundle"); return *LastInst; } void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) { auto *Front = E->getMainOp(); Instruction *LastInst = EntryToLastInstruction.lookup(E); assert(LastInst && "Failed to find last instruction in bundle"); // If the instruction is PHI, set the insert point after all the PHIs. bool IsPHI = isa(LastInst); if (IsPHI) LastInst = LastInst->getParent()->getFirstNonPHI(); if (IsPHI || (E->State != TreeEntry::NeedToGather && doesNotNeedToSchedule(E->Scalars))) { Builder.SetInsertPoint(LastInst); } else { // Set the insertion point after the last instruction in the bundle. Set the // debug location to Front. Builder.SetInsertPoint(LastInst->getParent(), std::next(LastInst->getIterator())); } Builder.SetCurrentDebugLocation(Front->getDebugLoc()); } Value *BoUpSLP::gather(ArrayRef VL) { // List of instructions/lanes from current block and/or the blocks which are // part of the current loop. These instructions will be inserted at the end to // make it possible to optimize loops and hoist invariant instructions out of // the loops body with better chances for success. SmallVector, 4> PostponedInsts; SmallSet PostponedIndices; Loop *L = LI->getLoopFor(Builder.GetInsertBlock()); auto &&CheckPredecessor = [](BasicBlock *InstBB, BasicBlock *InsertBB) { SmallPtrSet Visited; while (InsertBB && InsertBB != InstBB && Visited.insert(InsertBB).second) InsertBB = InsertBB->getSinglePredecessor(); return InsertBB && InsertBB == InstBB; }; for (int I = 0, E = VL.size(); I < E; ++I) { if (auto *Inst = dyn_cast(VL[I])) if ((CheckPredecessor(Inst->getParent(), Builder.GetInsertBlock()) || getTreeEntry(Inst) || (L && (L->contains(Inst)))) && PostponedIndices.insert(I).second) PostponedInsts.emplace_back(Inst, I); } auto &&CreateInsertElement = [this](Value *Vec, Value *V, unsigned Pos) { Vec = Builder.CreateInsertElement(Vec, V, Builder.getInt32(Pos)); auto *InsElt = dyn_cast(Vec); if (!InsElt) return Vec; GatherShuffleExtractSeq.insert(InsElt); CSEBlocks.insert(InsElt->getParent()); // Add to our 'need-to-extract' list. if (TreeEntry *Entry = getTreeEntry(V)) { // Find which lane we need to extract. unsigned FoundLane = Entry->findLaneForValue(V); ExternalUses.emplace_back(V, InsElt, FoundLane); } return Vec; }; Value *Val0 = isa(VL[0]) ? cast(VL[0])->getValueOperand() : VL[0]; FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); Value *Vec = PoisonValue::get(VecTy); SmallVector NonConsts; // Insert constant values at first. for (int I = 0, E = VL.size(); I < E; ++I) { if (PostponedIndices.contains(I)) continue; if (!isConstant(VL[I])) { NonConsts.push_back(I); continue; } Vec = CreateInsertElement(Vec, VL[I], I); } // Insert non-constant values. for (int I : NonConsts) Vec = CreateInsertElement(Vec, VL[I], I); // Append instructions, which are/may be part of the loop, in the end to make // it possible to hoist non-loop-based instructions. for (const std::pair &Pair : PostponedInsts) Vec = CreateInsertElement(Vec, Pair.first, Pair.second); return Vec; } /// Merges shuffle masks and emits final shuffle instruction, if required. It /// supports shuffling of 2 input vectors. It implements lazy shuffles emission, /// when the actual shuffle instruction is generated only if this is actually /// required. Otherwise, the shuffle instruction emission is delayed till the /// end of the process, to reduce the number of emitted instructions and further /// analysis/transformations. /// The class also will look through the previously emitted shuffle instructions /// and properly mark indices in mask as undef. /// For example, given the code /// \code /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0> /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0> /// \endcode /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will /// look through %s1 and %s2 and emit /// \code /// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3> /// \endcode /// instead. /// If 2 operands are of different size, the smallest one will be resized and /// the mask recalculated properly. /// For example, given the code /// \code /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0> /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0> /// \endcode /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will /// look through %s1 and %s2 and emit /// \code /// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3> /// \endcode /// instead. class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis { bool IsFinalized = false; /// Combined mask for all applied operands and masks. It is built during /// analysis and actual emission of shuffle vector instructions. SmallVector CommonMask; /// List of operands for the shuffle vector instruction. It hold at max 2 /// operands, if the 3rd is going to be added, the first 2 are combined into /// shuffle with \p CommonMask mask, the first operand sets to be the /// resulting shuffle and the second operand sets to be the newly added /// operand. The \p CommonMask is transformed in the proper way after that. SmallVector InVectors; IRBuilderBase &Builder; BoUpSLP &R; class ShuffleIRBuilder { IRBuilderBase &Builder; /// Holds all of the instructions that we gathered. SetVector &GatherShuffleExtractSeq; /// A list of blocks that we are going to CSE. SetVector &CSEBlocks; public: ShuffleIRBuilder(IRBuilderBase &Builder, SetVector &GatherShuffleExtractSeq, SetVector &CSEBlocks) : Builder(Builder), GatherShuffleExtractSeq(GatherShuffleExtractSeq), CSEBlocks(CSEBlocks) {} ~ShuffleIRBuilder() = default; /// Creates shufflevector for the 2 operands with the given mask. Value *createShuffleVector(Value *V1, Value *V2, ArrayRef Mask) { Value *Vec = Builder.CreateShuffleVector(V1, V2, Mask); if (auto *I = dyn_cast(Vec)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } return Vec; } /// Creates permutation of the single vector operand with the given mask, if /// it is not identity mask. Value *createShuffleVector(Value *V1, ArrayRef Mask) { if (Mask.empty()) return V1; unsigned VF = Mask.size(); unsigned LocalVF = cast(V1->getType())->getNumElements(); if (VF == LocalVF && ShuffleVectorInst::isIdentityMask(Mask)) return V1; Value *Vec = Builder.CreateShuffleVector(V1, Mask); if (auto *I = dyn_cast(Vec)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } return Vec; } /// Resizes 2 input vector to match the sizes, if the they are not equal /// yet. The smallest vector is resized to the size of the larger vector. void resizeToMatch(Value *&V1, Value *&V2) { if (V1->getType() == V2->getType()) return; int V1VF = cast(V1->getType())->getNumElements(); int V2VF = cast(V2->getType())->getNumElements(); int VF = std::max(V1VF, V2VF); int MinVF = std::min(V1VF, V2VF); SmallVector IdentityMask(VF, UndefMaskElem); std::iota(IdentityMask.begin(), std::next(IdentityMask.begin(), MinVF), 0); Value *&Op = MinVF == V1VF ? V1 : V2; Op = Builder.CreateShuffleVector(Op, IdentityMask); if (auto *I = dyn_cast(Op)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } if (MinVF == V1VF) V1 = Op; else V2 = Op; } }; /// Smart shuffle instruction emission, walks through shuffles trees and /// tries to find the best matching vector for the actual shuffle /// instruction. Value *createShuffle(Value *V1, Value *V2, ArrayRef Mask) { assert(V1 && "Expected at least one vector value."); ShuffleIRBuilder ShuffleBuilder(Builder, R.GatherShuffleExtractSeq, R.CSEBlocks); return BaseShuffleAnalysis::createShuffle(V1, V2, Mask, ShuffleBuilder); } /// Transforms mask \p CommonMask per given \p Mask to make proper set after /// shuffle emission. static void transformMaskAfterShuffle(MutableArrayRef CommonMask, ArrayRef Mask) { for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) if (Mask[Idx] != UndefMaskElem) CommonMask[Idx] = Idx; } public: ShuffleInstructionBuilder(IRBuilderBase &Builder, BoUpSLP &R) : Builder(Builder), R(R) {} /// Adds 2 input vectors and the mask for their shuffling. void add(Value *V1, Value *V2, ArrayRef Mask) { assert(V1 && V2 && !Mask.empty() && "Expected non-empty input vectors."); if (InVectors.empty()) { InVectors.push_back(V1); InVectors.push_back(V2); CommonMask.assign(Mask.begin(), Mask.end()); return; } Value *Vec = InVectors.front(); if (InVectors.size() == 2) { Vec = createShuffle(Vec, InVectors.back(), CommonMask); transformMaskAfterShuffle(CommonMask, Mask); } else if (cast(Vec->getType())->getNumElements() != Mask.size()) { Vec = createShuffle(Vec, nullptr, CommonMask); transformMaskAfterShuffle(CommonMask, Mask); } V1 = createShuffle(V1, V2, Mask); for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) if (Mask[Idx] != UndefMaskElem) CommonMask[Idx] = Idx + Sz; InVectors.front() = Vec; if (InVectors.size() == 2) InVectors.back() = V1; else InVectors.push_back(V1); } /// Adds another one input vector and the mask for the shuffling. void add(Value *V1, ArrayRef Mask) { if (InVectors.empty()) { if (!isa(V1->getType())) { V1 = createShuffle(V1, nullptr, CommonMask); CommonMask.assign(Mask.size(), UndefMaskElem); transformMaskAfterShuffle(CommonMask, Mask); } InVectors.push_back(V1); CommonMask.assign(Mask.begin(), Mask.end()); return; } const auto *It = find(InVectors, V1); if (It == InVectors.end()) { if (InVectors.size() == 2 || InVectors.front()->getType() != V1->getType() || !isa(V1->getType())) { Value *V = InVectors.front(); if (InVectors.size() == 2) { V = createShuffle(InVectors.front(), InVectors.back(), CommonMask); transformMaskAfterShuffle(CommonMask, CommonMask); } else if (cast(V->getType())->getNumElements() != CommonMask.size()) { V = createShuffle(InVectors.front(), nullptr, CommonMask); transformMaskAfterShuffle(CommonMask, CommonMask); } for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) if (CommonMask[Idx] == UndefMaskElem && Mask[Idx] != UndefMaskElem) CommonMask[Idx] = V->getType() != V1->getType() ? Idx + Sz : Mask[Idx] + cast(V1->getType()) ->getNumElements(); if (V->getType() != V1->getType()) V1 = createShuffle(V1, nullptr, Mask); InVectors.front() = V; if (InVectors.size() == 2) InVectors.back() = V1; else InVectors.push_back(V1); return; } // Check if second vector is required if the used elements are already // used from the first one. for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) if (Mask[Idx] != UndefMaskElem && CommonMask[Idx] == UndefMaskElem) { InVectors.push_back(V1); break; } } int VF = CommonMask.size(); if (auto *FTy = dyn_cast(V1->getType())) VF = FTy->getNumElements(); for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) if (Mask[Idx] != UndefMaskElem && CommonMask[Idx] == UndefMaskElem) CommonMask[Idx] = Mask[Idx] + (It == InVectors.begin() ? 0 : VF); } /// Adds another one input vector and the mask for the shuffling. void addOrdered(Value *V1, ArrayRef Order) { SmallVector NewMask; inversePermutation(Order, NewMask); add(V1, NewMask); } /// Finalize emission of the shuffles. Value * finalize(ArrayRef ExtMask = std::nullopt) { IsFinalized = true; if (!ExtMask.empty()) { if (CommonMask.empty()) { CommonMask.assign(ExtMask.begin(), ExtMask.end()); } else { SmallVector NewMask(ExtMask.size(), UndefMaskElem); for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) { if (ExtMask[I] == UndefMaskElem) continue; NewMask[I] = CommonMask[ExtMask[I]]; } CommonMask.swap(NewMask); } } if (CommonMask.empty()) { assert(InVectors.size() == 1 && "Expected only one vector with no mask"); return InVectors.front(); } if (InVectors.size() == 2) return createShuffle(InVectors.front(), InVectors.back(), CommonMask); return createShuffle(InVectors.front(), nullptr, CommonMask); } ~ShuffleInstructionBuilder() { assert((IsFinalized || CommonMask.empty()) && "Shuffle construction must be finalized."); } }; Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx) { ArrayRef VL = E->getOperand(NodeIdx); const unsigned VF = VL.size(); InstructionsState S = getSameOpcode(VL, *TLI); // Special processing for GEPs bundle, which may include non-gep values. if (!S.getOpcode() && VL.front()->getType()->isPointerTy()) { const auto *It = find_if(VL, [](Value *V) { return isa(V); }); if (It != VL.end()) S = getSameOpcode(*It, *TLI); } if (S.getOpcode()) { if (TreeEntry *VE = getTreeEntry(S.OpValue); VE && VE->isSame(VL) && (any_of(VE->UserTreeIndices, [E, NodeIdx](const EdgeInfo &EI) { return EI.UserTE == E && EI.EdgeIdx == NodeIdx; }) || any_of(VectorizableTree, [E, NodeIdx, VE](const std::unique_ptr &TE) { return TE->isOperandGatherNode({E, NodeIdx}) && VE->isSame(TE->Scalars); }))) { auto FinalShuffle = [&](Value *V, ArrayRef Mask) { ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); ShuffleBuilder.add(V, Mask); return ShuffleBuilder.finalize(std::nullopt); }; Value *V = vectorizeTree(VE); if (VF != cast(V->getType())->getNumElements()) { if (!VE->ReuseShuffleIndices.empty()) { // Reshuffle to get only unique values. // If some of the scalars are duplicated in the vectorization // tree entry, we do not vectorize them but instead generate a // mask for the reuses. But if there are several users of the // same entry, they may have different vectorization factors. // This is especially important for PHI nodes. In this case, we // need to adapt the resulting instruction for the user // vectorization factor and have to reshuffle it again to take // only unique elements of the vector. Without this code the // function incorrectly returns reduced vector instruction with // the same elements, not with the unique ones. // block: // %phi = phi <2 x > { .., %entry} {%shuffle, %block} // %2 = shuffle <2 x > %phi, poison, <4 x > <1, 1, 0, 0> // ... (use %2) // %shuffle = shuffle <2 x> %2, poison, <2 x> {2, 0} // br %block SmallVector UniqueIdxs(VF, UndefMaskElem); SmallSet UsedIdxs; int Pos = 0; for (int Idx : VE->ReuseShuffleIndices) { if (Idx != static_cast(VF) && Idx != UndefMaskElem && UsedIdxs.insert(Idx).second) UniqueIdxs[Idx] = Pos; ++Pos; } assert(VF >= UsedIdxs.size() && "Expected vectorization factor " "less than original vector size."); UniqueIdxs.append(VF - UsedIdxs.size(), UndefMaskElem); V = FinalShuffle(V, UniqueIdxs); } else { assert(VF < cast(V->getType())->getNumElements() && "Expected vectorization factor less " "than original vector size."); SmallVector UniformMask(VF, 0); std::iota(UniformMask.begin(), UniformMask.end(), 0); V = FinalShuffle(V, UniformMask); } } return V; } } // Find the corresponding gather entry and vectorize it. // Allows to be more accurate with tree/graph transformations, checks for the // correctness of the transformations in many cases. auto *I = find_if(VectorizableTree, [E, NodeIdx](const std::unique_ptr &TE) { return TE->isOperandGatherNode({E, NodeIdx}); }); assert(I != VectorizableTree.end() && "Gather node is not in the graph."); assert(I->get()->UserTreeIndices.size() == 1 && "Expected only single user for the gather node."); assert(I->get()->isSame(VL) && "Expected same list of scalars."); IRBuilder<>::InsertPointGuard Guard(Builder); if (E->getOpcode() != Instruction::InsertElement && E->getOpcode() != Instruction::PHI) { Instruction *LastInst = EntryToLastInstruction.lookup(E); assert(LastInst && "Failed to find last instruction in bundle"); Builder.SetInsertPoint(LastInst); } return vectorizeTree(I->get()); } Value *BoUpSLP::createBuildVector(const TreeEntry *E) { assert(E->State == TreeEntry::NeedToGather && "Expected gather node."); unsigned VF = E->getVectorFactor(); ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); SmallVector Gathered( VF, PoisonValue::get(E->Scalars.front()->getType())); bool NeedFreeze = false; SmallVector VL(E->Scalars.begin(), E->Scalars.end()); // Build a mask out of the redorder indices and reorder scalars per this mask. SmallVector ReorderMask; inversePermutation(E->ReorderIndices, ReorderMask); if (!ReorderMask.empty()) reorderScalars(VL, ReorderMask); SmallVector ReuseMask(VF, UndefMaskElem); if (!allConstant(VL)) { // For splats with can emit broadcasts instead of gathers, so try to find // such sequences. bool IsSplat = isSplat(VL) && (VL.size() > 2 || VL.front() == VL.back()); SmallVector UndefPos; DenseMap UniquePositions; // Gather unique non-const values and all constant values. // For repeated values, just shuffle them. for (auto [I, V] : enumerate(VL)) { if (isa(V)) { if (!isa(V)) { Gathered[I] = V; ReuseMask[I] = I; UndefPos.push_back(I); } continue; } if (isConstant(V)) { Gathered[I] = V; ReuseMask[I] = I; continue; } if (IsSplat) { Gathered.front() = V; ReuseMask[I] = 0; } else { const auto Res = UniquePositions.try_emplace(V, I); Gathered[Res.first->second] = V; ReuseMask[I] = Res.first->second; } } if (!UndefPos.empty() && IsSplat) { // For undef values, try to replace them with the simple broadcast. // We can do it if the broadcasted value is guaranteed to be // non-poisonous, or by freezing the incoming scalar value first. auto *It = find_if(Gathered, [this, E](Value *V) { return !isa(V) && (getTreeEntry(V) || isGuaranteedNotToBePoison(V) || any_of(V->uses(), [E](const Use &U) { // Check if the value already used in the same operation in // one of the nodes already. return E->UserTreeIndices.size() == 1 && is_contained( E->UserTreeIndices.front().UserTE->Scalars, U.getUser()) && E->UserTreeIndices.front().EdgeIdx != U.getOperandNo(); })); }); if (It != Gathered.end()) { // Replace undefs by the non-poisoned scalars and emit broadcast. int Pos = std::distance(Gathered.begin(), It); for_each(UndefPos, [&](int I) { // Set the undef position to the non-poisoned scalar. ReuseMask[I] = Pos; // Replace the undef by the poison, in the mask it is replaced by non-poisoned scalar already. if (I != Pos) Gathered[I] = PoisonValue::get(Gathered[I]->getType()); }); } else { // Replace undefs by the poisons, emit broadcast and then emit // freeze. for_each(UndefPos, [&](int I) { ReuseMask[I] = UndefMaskElem; if (isa(Gathered[I])) Gathered[I] = PoisonValue::get(Gathered[I]->getType()); }); NeedFreeze = true; } } } else { ReuseMask.clear(); copy(VL, Gathered.begin()); } // Gather unique scalars and all constants. Value *Vec = gather(Gathered); ShuffleBuilder.add(Vec, ReuseMask); Vec = ShuffleBuilder.finalize(E->ReuseShuffleIndices); if (NeedFreeze) Vec = Builder.CreateFreeze(Vec); return Vec; } Value *BoUpSLP::vectorizeTree(TreeEntry *E) { IRBuilder<>::InsertPointGuard Guard(Builder); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n"); return E->VectorizedValue; } auto FinalShuffle = [&](Value *V, const TreeEntry *E) { ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); if (E->State != TreeEntry::NeedToGather && E->getOpcode() == Instruction::Store) { ArrayRef Mask = ArrayRef(reinterpret_cast(E->ReorderIndices.begin()), E->ReorderIndices.size()); ShuffleBuilder.add(V, Mask); } else { ShuffleBuilder.addOrdered(V, E->ReorderIndices); } return ShuffleBuilder.finalize(E->ReuseShuffleIndices); }; if (E->State == TreeEntry::NeedToGather) { if (E->Idx > 0) { // We are in the middle of a vectorizable chain. We need to gather the // scalars from the users. Value *Vec = createBuildVector(E); E->VectorizedValue = Vec; return Vec; } if (E->getMainOp()) setInsertPointAfterBundle(E); SmallVector GatheredScalars(E->Scalars.begin(), E->Scalars.end()); // Build a mask out of the reorder indices and reorder scalars per this // mask. SmallVector ReorderMask; inversePermutation(E->ReorderIndices, ReorderMask); if (!ReorderMask.empty()) reorderScalars(GatheredScalars, ReorderMask); Value *Vec; SmallVector Mask; SmallVector Entries; std::optional Shuffle = isGatherShuffledEntry(E, GatheredScalars, Mask, Entries); if (Shuffle) { assert((Entries.size() == 1 || Entries.size() == 2) && "Expected shuffle of 1 or 2 entries."); Vec = Builder.CreateShuffleVector(Entries.front()->VectorizedValue, Entries.back()->VectorizedValue, Mask); if (auto *I = dyn_cast(Vec)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } } else { Vec = gather(E->Scalars); } Vec = FinalShuffle(Vec, E); E->VectorizedValue = Vec; return Vec; } assert((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state"); unsigned ShuffleOrOp = E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); Instruction *VL0 = E->getMainOp(); Type *ScalarTy = VL0->getType(); if (auto *Store = dyn_cast(VL0)) ScalarTy = Store->getValueOperand()->getType(); else if (auto *IE = dyn_cast(VL0)) ScalarTy = IE->getOperand(1)->getType(); auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); switch (ShuffleOrOp) { case Instruction::PHI: { assert((E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices.empty()) && "PHI reordering is free."); auto *PH = cast(VL0); Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); Builder.SetCurrentDebugLocation(PH->getDebugLoc()); PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); Value *V = NewPhi; // Adjust insertion point once all PHI's have been generated. Builder.SetInsertPoint(&*PH->getParent()->getFirstInsertionPt()); Builder.SetCurrentDebugLocation(PH->getDebugLoc()); V = FinalShuffle(V, E); E->VectorizedValue = V; // PHINodes may have multiple entries from the same block. We want to // visit every block once. SmallPtrSet VisitedBBs; for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { ValueList Operands; BasicBlock *IBB = PH->getIncomingBlock(i); // Stop emission if all incoming values are generated. if (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues()) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return V; } if (!VisitedBBs.insert(IBB).second) { NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); continue; } Builder.SetInsertPoint(IBB->getTerminator()); Builder.SetCurrentDebugLocation(PH->getDebugLoc()); Value *Vec = vectorizeOperand(E, i); NewPhi->addIncoming(Vec, IBB); } assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values"); return V; } case Instruction::ExtractElement: { Value *V = E->getSingleOperand(0); setInsertPointAfterBundle(E); V = FinalShuffle(V, E); E->VectorizedValue = V; return V; } case Instruction::ExtractValue: { auto *LI = cast(E->getSingleOperand(0)); Builder.SetInsertPoint(LI); auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); Value *NewV = propagateMetadata(V, E->Scalars); NewV = FinalShuffle(NewV, E); E->VectorizedValue = NewV; return NewV; } case Instruction::InsertElement: { assert(E->ReuseShuffleIndices.empty() && "All inserts should be unique"); Builder.SetInsertPoint(cast(E->Scalars.back())); Value *V = vectorizeOperand(E, 1); // Create InsertVector shuffle if necessary auto *FirstInsert = cast(*find_if(E->Scalars, [E](Value *V) { return !is_contained(E->Scalars, cast(V)->getOperand(0)); })); const unsigned NumElts = cast(FirstInsert->getType())->getNumElements(); const unsigned NumScalars = E->Scalars.size(); unsigned Offset = *getInsertIndex(VL0); assert(Offset < NumElts && "Failed to find vector index offset"); // Create shuffle to resize vector SmallVector Mask; if (!E->ReorderIndices.empty()) { inversePermutation(E->ReorderIndices, Mask); Mask.append(NumElts - NumScalars, UndefMaskElem); } else { Mask.assign(NumElts, UndefMaskElem); std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0); } // Create InsertVector shuffle if necessary bool IsIdentity = true; SmallVector PrevMask(NumElts, UndefMaskElem); Mask.swap(PrevMask); for (unsigned I = 0; I < NumScalars; ++I) { Value *Scalar = E->Scalars[PrevMask[I]]; unsigned InsertIdx = *getInsertIndex(Scalar); IsIdentity &= InsertIdx - Offset == I; Mask[InsertIdx - Offset] = I; } if (!IsIdentity || NumElts != NumScalars) { V = Builder.CreateShuffleVector(V, Mask); if (auto *I = dyn_cast(V)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } } SmallVector InsertMask(NumElts, UndefMaskElem); for (unsigned I = 0; I < NumElts; I++) { if (Mask[I] != UndefMaskElem) InsertMask[Offset + I] = I; } SmallBitVector UseMask = buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask); SmallBitVector IsFirstUndef = isUndefVector(FirstInsert->getOperand(0), UseMask); if ((!IsIdentity || Offset != 0 || !IsFirstUndef.all()) && NumElts != NumScalars) { if (IsFirstUndef.all()) { if (!ShuffleVectorInst::isIdentityMask(InsertMask)) { SmallBitVector IsFirstPoison = isUndefVector(FirstInsert->getOperand(0), UseMask); if (!IsFirstPoison.all()) { for (unsigned I = 0; I < NumElts; I++) { if (InsertMask[I] == UndefMaskElem && !IsFirstPoison.test(I)) InsertMask[I] = I + NumElts; } } V = Builder.CreateShuffleVector( V, IsFirstPoison.all() ? PoisonValue::get(V->getType()) : FirstInsert->getOperand(0), InsertMask, cast(E->Scalars.back())->getName()); if (auto *I = dyn_cast(V)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } } } else { SmallBitVector IsFirstPoison = isUndefVector(FirstInsert->getOperand(0), UseMask); for (unsigned I = 0; I < NumElts; I++) { if (InsertMask[I] == UndefMaskElem) InsertMask[I] = IsFirstPoison.test(I) ? UndefMaskElem : I; else InsertMask[I] += NumElts; } V = Builder.CreateShuffleVector( FirstInsert->getOperand(0), V, InsertMask, cast(E->Scalars.back())->getName()); if (auto *I = dyn_cast(V)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } } } ++NumVectorInstructions; E->VectorizedValue = V; return V; } case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: case Instruction::FPToSI: case Instruction::FPExt: case Instruction::PtrToInt: case Instruction::IntToPtr: case Instruction::SIToFP: case Instruction::UIToFP: case Instruction::Trunc: case Instruction::FPTrunc: case Instruction::BitCast: { setInsertPointAfterBundle(E); Value *InVec = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } auto *CI = cast(VL0); Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::FCmp: case Instruction::ICmp: { setInsertPointAfterBundle(E); Value *L = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *R = vectorizeOperand(E, 1); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } CmpInst::Predicate P0 = cast(VL0)->getPredicate(); Value *V = Builder.CreateCmp(P0, L, R); propagateIRFlags(V, E->Scalars, VL0); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::Select: { setInsertPointAfterBundle(E); Value *Cond = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *True = vectorizeOperand(E, 1); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *False = vectorizeOperand(E, 2); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *V = Builder.CreateSelect(Cond, True, False); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::FNeg: { setInsertPointAfterBundle(E); Value *Op = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *V = Builder.CreateUnOp( static_cast(E->getOpcode()), Op); propagateIRFlags(V, E->Scalars, VL0); if (auto *I = dyn_cast(V)) V = propagateMetadata(I, E->Scalars); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::Add: case Instruction::FAdd: case Instruction::Sub: case Instruction::FSub: case Instruction::Mul: case Instruction::FMul: case Instruction::UDiv: case Instruction::SDiv: case Instruction::FDiv: case Instruction::URem: case Instruction::SRem: case Instruction::FRem: case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::And: case Instruction::Or: case Instruction::Xor: { setInsertPointAfterBundle(E); Value *LHS = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *RHS = vectorizeOperand(E, 1); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *V = Builder.CreateBinOp( static_cast(E->getOpcode()), LHS, RHS); propagateIRFlags(V, E->Scalars, VL0); if (auto *I = dyn_cast(V)) V = propagateMetadata(I, E->Scalars); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::Load: { // Loads are inserted at the head of the tree because we don't want to // sink them all the way down past store instructions. setInsertPointAfterBundle(E); LoadInst *LI = cast(VL0); Instruction *NewLI; unsigned AS = LI->getPointerAddressSpace(); Value *PO = LI->getPointerOperand(); if (E->State == TreeEntry::Vectorize) { Value *VecPtr = Builder.CreateBitCast(PO, VecTy->getPointerTo(AS)); NewLI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); // The pointer operand uses an in-tree scalar so we add the new BitCast // or LoadInst to ExternalUses list to make sure that an extract will // be generated in the future. if (TreeEntry *Entry = getTreeEntry(PO)) { // Find which lane we need to extract. unsigned FoundLane = Entry->findLaneForValue(PO); ExternalUses.emplace_back( PO, PO != VecPtr ? cast(VecPtr) : NewLI, FoundLane); } } else { assert(E->State == TreeEntry::ScatterVectorize && "Unhandled state"); Value *VecPtr = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } // Use the minimum alignment of the gathered loads. Align CommonAlignment = LI->getAlign(); for (Value *V : E->Scalars) CommonAlignment = std::min(CommonAlignment, cast(V)->getAlign()); NewLI = Builder.CreateMaskedGather(VecTy, VecPtr, CommonAlignment); } Value *V = propagateMetadata(NewLI, E->Scalars); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::Store: { auto *SI = cast(VL0); unsigned AS = SI->getPointerAddressSpace(); setInsertPointAfterBundle(E); Value *VecValue = vectorizeOperand(E, 0); VecValue = FinalShuffle(VecValue, E); Value *ScalarPtr = SI->getPointerOperand(); Value *VecPtr = Builder.CreateBitCast( ScalarPtr, VecValue->getType()->getPointerTo(AS)); StoreInst *ST = Builder.CreateAlignedStore(VecValue, VecPtr, SI->getAlign()); // The pointer operand uses an in-tree scalar, so add the new BitCast or // StoreInst to ExternalUses to make sure that an extract will be // generated in the future. if (TreeEntry *Entry = getTreeEntry(ScalarPtr)) { // Find which lane we need to extract. unsigned FoundLane = Entry->findLaneForValue(ScalarPtr); ExternalUses.push_back(ExternalUser( ScalarPtr, ScalarPtr != VecPtr ? cast(VecPtr) : ST, FoundLane)); } Value *V = propagateMetadata(ST, E->Scalars); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::GetElementPtr: { auto *GEP0 = cast(VL0); setInsertPointAfterBundle(E); Value *Op0 = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } SmallVector OpVecs; for (int J = 1, N = GEP0->getNumOperands(); J < N; ++J) { Value *OpVec = vectorizeOperand(E, J); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } OpVecs.push_back(OpVec); } Value *V = Builder.CreateGEP(GEP0->getSourceElementType(), Op0, OpVecs); if (Instruction *I = dyn_cast(V)) { SmallVector GEPs; for (Value *V : E->Scalars) { if (isa(V)) GEPs.push_back(V); } V = propagateMetadata(I, GEPs); } V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::Call: { CallInst *CI = cast(VL0); setInsertPointAfterBundle(E); Intrinsic::ID IID = Intrinsic::not_intrinsic; if (Function *FI = CI->getCalledFunction()) IID = FI->getIntrinsicID(); Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); bool UseIntrinsic = ID != Intrinsic::not_intrinsic && VecCallCosts.first <= VecCallCosts.second; Value *ScalarArg = nullptr; std::vector OpVecs; SmallVector TysForDecl = {FixedVectorType::get(CI->getType(), E->Scalars.size())}; for (int j = 0, e = CI->arg_size(); j < e; ++j) { ValueList OpVL; // Some intrinsics have scalar arguments. This argument should not be // vectorized. if (UseIntrinsic && isVectorIntrinsicWithScalarOpAtArg(IID, j)) { CallInst *CEI = cast(VL0); ScalarArg = CEI->getArgOperand(j); OpVecs.push_back(CEI->getArgOperand(j)); if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) TysForDecl.push_back(ScalarArg->getType()); continue; } Value *OpVec = vectorizeOperand(E, j); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"); OpVecs.push_back(OpVec); if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) TysForDecl.push_back(OpVec->getType()); } Function *CF; if (!UseIntrinsic) { VFShape Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast( VecTy->getNumElements())), false /*HasGlobalPred*/); CF = VFDatabase(*CI).getVectorizedFunction(Shape); } else { CF = Intrinsic::getDeclaration(F->getParent(), ID, TysForDecl); } SmallVector OpBundles; CI->getOperandBundlesAsDefs(OpBundles); Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); // 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) { if (TreeEntry *Entry = getTreeEntry(ScalarArg)) { // Find which lane we need to extract. unsigned FoundLane = Entry->findLaneForValue(ScalarArg); ExternalUses.push_back( ExternalUser(ScalarArg, cast(V), FoundLane)); } } propagateIRFlags(V, E->Scalars, VL0); V = FinalShuffle(V, E); E->VectorizedValue = V; ++NumVectorInstructions; return V; } case Instruction::ShuffleVector: { assert(E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa(VL0) && isa(E->getAltOp()))) && "Invalid Shuffle Vector Operand"); Value *LHS = nullptr, *RHS = nullptr; if (Instruction::isBinaryOp(E->getOpcode()) || isa(VL0)) { setInsertPointAfterBundle(E); LHS = vectorizeOperand(E, 0); if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } RHS = vectorizeOperand(E, 1); } else { setInsertPointAfterBundle(E); LHS = vectorizeOperand(E, 0); } if (E->VectorizedValue) { LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"); return E->VectorizedValue; } Value *V0, *V1; if (Instruction::isBinaryOp(E->getOpcode())) { V0 = Builder.CreateBinOp( static_cast(E->getOpcode()), LHS, RHS); V1 = Builder.CreateBinOp( static_cast(E->getAltOpcode()), LHS, RHS); } else if (auto *CI0 = dyn_cast(VL0)) { V0 = Builder.CreateCmp(CI0->getPredicate(), LHS, RHS); auto *AltCI = cast(E->getAltOp()); CmpInst::Predicate AltPred = AltCI->getPredicate(); V1 = Builder.CreateCmp(AltPred, LHS, RHS); } else { V0 = Builder.CreateCast( static_cast(E->getOpcode()), LHS, VecTy); V1 = Builder.CreateCast( static_cast(E->getAltOpcode()), LHS, VecTy); } // Add V0 and V1 to later analysis to try to find and remove matching // instruction, if any. for (Value *V : {V0, V1}) { if (auto *I = dyn_cast(V)) { GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } } // Create shuffle to take alternate operations from the vector. // Also, gather up main and alt scalar ops to propagate IR flags to // each vector operation. ValueList OpScalars, AltScalars; SmallVector Mask; buildShuffleEntryMask( E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, [E, this](Instruction *I) { assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode"); return isAlternateInstruction(I, E->getMainOp(), E->getAltOp(), *TLI); }, Mask, &OpScalars, &AltScalars); propagateIRFlags(V0, OpScalars); propagateIRFlags(V1, AltScalars); Value *V = Builder.CreateShuffleVector(V0, V1, Mask); if (auto *I = dyn_cast(V)) { V = propagateMetadata(I, E->Scalars); GatherShuffleExtractSeq.insert(I); CSEBlocks.insert(I->getParent()); } E->VectorizedValue = V; ++NumVectorInstructions; return V; } default: llvm_unreachable("unknown inst"); } return nullptr; } Value *BoUpSLP::vectorizeTree() { ExtraValueToDebugLocsMap ExternallyUsedValues; return vectorizeTree(ExternallyUsedValues); } namespace { /// Data type for handling buildvector sequences with the reused scalars from /// other tree entries. struct ShuffledInsertData { /// List of insertelements to be replaced by shuffles. SmallVector InsertElements; /// The parent vectors and shuffle mask for the given list of inserts. MapVector> ValueMasks; }; } // namespace Value *BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues, Instruction *ReductionRoot) { // All blocks must be scheduled before any instructions are inserted. for (auto &BSIter : BlocksSchedules) { scheduleBlock(BSIter.second.get()); } // Pre-gather last instructions. for (const std::unique_ptr &E : VectorizableTree) { if ((E->State == TreeEntry::NeedToGather && (!E->getMainOp() || E->Idx > 0)) || (E->State != TreeEntry::NeedToGather && E->getOpcode() == Instruction::ExtractValue) || E->getOpcode() == Instruction::InsertElement) continue; Instruction *LastInst = &getLastInstructionInBundle(E.get()); EntryToLastInstruction.try_emplace(E.get(), LastInst); } Builder.SetInsertPoint(ReductionRoot ? ReductionRoot : &F->getEntryBlock().front()); auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); // If the vectorized tree can be rewritten in a smaller type, we truncate the // vectorized root. InstCombine will then rewrite the entire expression. We // sign extend the extracted values below. auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; if (MinBWs.count(ScalarRoot)) { if (auto *I = dyn_cast(VectorRoot)) { // If current instr is a phi and not the last phi, insert it after the // last phi node. if (isa(I)) Builder.SetInsertPoint(&*I->getParent()->getFirstInsertionPt()); else Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); } auto BundleWidth = VectorizableTree[0]->Scalars.size(); auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); VectorizableTree[0]->VectorizedValue = Trunc; } LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n"); SmallVector ShuffledInserts; // Maps vector instruction to original insertelement instruction DenseMap VectorToInsertElement; // Maps extract Scalar to the corresponding extractelement instruction in the // basic block. Only one extractelement per block should be emitted. DenseMap> ScalarToEEs; // Extract all of the elements with the external uses. for (const auto &ExternalUse : ExternalUses) { Value *Scalar = ExternalUse.Scalar; llvm::User *User = ExternalUse.User; // Skip users that we already RAUW. This happens when one instruction // has multiple uses of the same value. if (User && !is_contained(Scalar->users(), User)) continue; TreeEntry *E = getTreeEntry(Scalar); assert(E && "Invalid scalar"); assert(E->State != TreeEntry::NeedToGather && "Extracting from a gather list"); // Non-instruction pointers are not deleted, just skip them. if (E->getOpcode() == Instruction::GetElementPtr && !isa(Scalar)) continue; Value *Vec = E->VectorizedValue; assert(Vec && "Can't find vectorizable value"); Value *Lane = Builder.getInt32(ExternalUse.Lane); auto ExtractAndExtendIfNeeded = [&](Value *Vec) { if (Scalar->getType() != Vec->getType()) { Value *Ex = nullptr; auto It = ScalarToEEs.find(Scalar); if (It != ScalarToEEs.end()) { // No need to emit many extracts, just move the only one in the // current block. auto EEIt = It->second.find(Builder.GetInsertBlock()); if (EEIt != It->second.end()) { Instruction *I = EEIt->second; if (Builder.GetInsertPoint() != Builder.GetInsertBlock()->end() && Builder.GetInsertPoint()->comesBefore(I)) I->moveBefore(&*Builder.GetInsertPoint()); Ex = I; } } if (!Ex) { // "Reuse" the existing extract to improve final codegen. if (auto *ES = dyn_cast(Scalar)) { Ex = Builder.CreateExtractElement(ES->getOperand(0), ES->getOperand(1)); } else { Ex = Builder.CreateExtractElement(Vec, Lane); } if (auto *I = dyn_cast(Ex)) ScalarToEEs[Scalar].try_emplace(Builder.GetInsertBlock(), I); } // The then branch of the previous if may produce constants, since 0 // operand might be a constant. if (auto *ExI = dyn_cast(Ex)) { GatherShuffleExtractSeq.insert(ExI); CSEBlocks.insert(ExI->getParent()); } // If necessary, sign-extend or zero-extend ScalarRoot // to the larger type. if (!MinBWs.count(ScalarRoot)) return Ex; if (MinBWs[ScalarRoot].second) return Builder.CreateSExt(Ex, Scalar->getType()); return Builder.CreateZExt(Ex, Scalar->getType()); } assert(isa(Scalar->getType()) && isa(Scalar) && "In-tree scalar of vector type is not insertelement?"); auto *IE = cast(Scalar); VectorToInsertElement.try_emplace(Vec, IE); return Vec; }; // If User == nullptr, the Scalar is used as extra arg. Generate // ExtractElement instruction and update the record for this scalar in // ExternallyUsedValues. if (!User) { assert(ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map"); if (auto *VecI = dyn_cast(Vec)) { if (auto *PHI = dyn_cast(VecI)) Builder.SetInsertPoint(PHI->getParent()->getFirstNonPHI()); else Builder.SetInsertPoint(VecI->getParent(), std::next(VecI->getIterator())); } else { Builder.SetInsertPoint(&F->getEntryBlock().front()); } Value *NewInst = ExtractAndExtendIfNeeded(Vec); auto &NewInstLocs = ExternallyUsedValues[NewInst]; auto It = ExternallyUsedValues.find(Scalar); assert(It != ExternallyUsedValues.end() && "Externally used scalar is not found in ExternallyUsedValues"); NewInstLocs.append(It->second); ExternallyUsedValues.erase(Scalar); // Required to update internally referenced instructions. Scalar->replaceAllUsesWith(NewInst); continue; } if (auto *VU = dyn_cast(User)) { // Skip if the scalar is another vector op or Vec is not an instruction. if (!Scalar->getType()->isVectorTy() && isa(Vec)) { if (auto *FTy = dyn_cast(User->getType())) { std::optional InsertIdx = getInsertIndex(VU); if (InsertIdx) { // Need to use original vector, if the root is truncated. if (MinBWs.count(Scalar) && VectorizableTree[0]->VectorizedValue == Vec) Vec = VectorRoot; auto *It = find_if(ShuffledInserts, [VU](const ShuffledInsertData &Data) { // Checks if 2 insertelements are from the same buildvector. InsertElementInst *VecInsert = Data.InsertElements.front(); return areTwoInsertFromSameBuildVector( VU, VecInsert, [](InsertElementInst *II) { return II->getOperand(0); }); }); unsigned Idx = *InsertIdx; if (It == ShuffledInserts.end()) { (void)ShuffledInserts.emplace_back(); It = std::next(ShuffledInserts.begin(), ShuffledInserts.size() - 1); SmallVectorImpl &Mask = It->ValueMasks[Vec]; if (Mask.empty()) Mask.assign(FTy->getNumElements(), UndefMaskElem); // Find the insertvector, vectorized in tree, if any. Value *Base = VU; while (auto *IEBase = dyn_cast(Base)) { if (IEBase != User && (!IEBase->hasOneUse() || getInsertIndex(IEBase).value_or(Idx) == Idx)) break; // Build the mask for the vectorized insertelement instructions. if (const TreeEntry *E = getTreeEntry(IEBase)) { do { IEBase = cast(Base); int IEIdx = *getInsertIndex(IEBase); assert(Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already."); Mask[IEIdx] = IEIdx; Base = IEBase->getOperand(0); } while (E == getTreeEntry(Base)); break; } Base = cast(Base)->getOperand(0); // After the vectorization the def-use chain has changed, need // to look through original insertelement instructions, if they // get replaced by vector instructions. auto It = VectorToInsertElement.find(Base); if (It != VectorToInsertElement.end()) Base = It->second; } } SmallVectorImpl &Mask = It->ValueMasks[Vec]; if (Mask.empty()) Mask.assign(FTy->getNumElements(), UndefMaskElem); Mask[Idx] = ExternalUse.Lane; It->InsertElements.push_back(cast(User)); continue; } } } } // Generate extracts for out-of-tree users. // Find the insertion point for the extractelement lane. if (auto *VecI = dyn_cast(Vec)) { if (PHINode *PH = dyn_cast(User)) { for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { if (PH->getIncomingValue(i) == Scalar) { Instruction *IncomingTerminator = PH->getIncomingBlock(i)->getTerminator(); if (isa(IncomingTerminator)) { Builder.SetInsertPoint(VecI->getParent(), std::next(VecI->getIterator())); } else { Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); } Value *NewInst = ExtractAndExtendIfNeeded(Vec); PH->setOperand(i, NewInst); } } } else { Builder.SetInsertPoint(cast(User)); Value *NewInst = ExtractAndExtendIfNeeded(Vec); User->replaceUsesOfWith(Scalar, NewInst); } } else { Builder.SetInsertPoint(&F->getEntryBlock().front()); Value *NewInst = ExtractAndExtendIfNeeded(Vec); User->replaceUsesOfWith(Scalar, NewInst); } LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n"); } auto CreateShuffle = [&](Value *V1, Value *V2, ArrayRef Mask) { SmallVector CombinedMask1(Mask.size(), UndefMaskElem); SmallVector CombinedMask2(Mask.size(), UndefMaskElem); int VF = cast(V1->getType())->getNumElements(); for (int I = 0, E = Mask.size(); I < E; ++I) { if (Mask[I] < VF) CombinedMask1[I] = Mask[I]; else CombinedMask2[I] = Mask[I] - VF; } ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); ShuffleBuilder.add(V1, CombinedMask1); if (V2) ShuffleBuilder.add(V2, CombinedMask2); return ShuffleBuilder.finalize(std::nullopt); }; auto &&ResizeToVF = [&CreateShuffle](Value *Vec, ArrayRef Mask, bool ForSingleMask) { unsigned VF = Mask.size(); unsigned VecVF = cast(Vec->getType())->getNumElements(); if (VF != VecVF) { if (any_of(Mask, [VF](int Idx) { return Idx >= static_cast(VF); })) { Vec = CreateShuffle(Vec, nullptr, Mask); return std::make_pair(Vec, true); } if (!ForSingleMask) { SmallVector ResizeMask(VF, UndefMaskElem); for (unsigned I = 0; I < VF; ++I) { if (Mask[I] != UndefMaskElem) ResizeMask[Mask[I]] = Mask[I]; } Vec = CreateShuffle(Vec, nullptr, ResizeMask); } } return std::make_pair(Vec, false); }; // Perform shuffling of the vectorize tree entries for better handling of // external extracts. for (int I = 0, E = ShuffledInserts.size(); I < E; ++I) { // Find the first and the last instruction in the list of insertelements. sort(ShuffledInserts[I].InsertElements, isFirstInsertElement); InsertElementInst *FirstInsert = ShuffledInserts[I].InsertElements.front(); InsertElementInst *LastInsert = ShuffledInserts[I].InsertElements.back(); Builder.SetInsertPoint(LastInsert); auto Vector = ShuffledInserts[I].ValueMasks.takeVector(); Value *NewInst = performExtractsShuffleAction( MutableArrayRef(Vector.data(), Vector.size()), FirstInsert->getOperand(0), [](Value *Vec) { return cast(Vec->getType()) ->getElementCount() .getKnownMinValue(); }, ResizeToVF, [FirstInsert, &CreateShuffle](ArrayRef Mask, ArrayRef Vals) { assert((Vals.size() == 1 || Vals.size() == 2) && "Expected exactly 1 or 2 input values."); if (Vals.size() == 1) { // Do not create shuffle if the mask is a simple identity // non-resizing mask. if (Mask.size() != cast(Vals.front()->getType()) ->getNumElements() || !ShuffleVectorInst::isIdentityMask(Mask)) return CreateShuffle(Vals.front(), nullptr, Mask); return Vals.front(); } return CreateShuffle(Vals.front() ? Vals.front() : FirstInsert->getOperand(0), Vals.back(), Mask); }); auto It = ShuffledInserts[I].InsertElements.rbegin(); // Rebuild buildvector chain. InsertElementInst *II = nullptr; if (It != ShuffledInserts[I].InsertElements.rend()) II = *It; SmallVector Inserts; while (It != ShuffledInserts[I].InsertElements.rend()) { assert(II && "Must be an insertelement instruction."); if (*It == II) ++It; else Inserts.push_back(cast(II)); II = dyn_cast(II->getOperand(0)); } for (Instruction *II : reverse(Inserts)) { II->replaceUsesOfWith(II->getOperand(0), NewInst); if (auto *NewI = dyn_cast(NewInst)) if (II->getParent() == NewI->getParent() && II->comesBefore(NewI)) II->moveAfter(NewI); NewInst = II; } LastInsert->replaceAllUsesWith(NewInst); for (InsertElementInst *IE : reverse(ShuffledInserts[I].InsertElements)) { IE->replaceUsesOfWith(IE->getOperand(0), PoisonValue::get(IE->getOperand(0)->getType())); IE->replaceUsesOfWith(IE->getOperand(1), PoisonValue::get(IE->getOperand(1)->getType())); eraseInstruction(IE); } CSEBlocks.insert(LastInsert->getParent()); } SmallVector RemovedInsts; // For each vectorized value: for (auto &TEPtr : VectorizableTree) { TreeEntry *Entry = TEPtr.get(); // No need to handle users of gathered values. if (Entry->State == TreeEntry::NeedToGather) continue; assert(Entry->VectorizedValue && "Can't find vectorizable value"); // For each lane: for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { Value *Scalar = Entry->Scalars[Lane]; if (Entry->getOpcode() == Instruction::GetElementPtr && !isa(Scalar)) continue; #ifndef NDEBUG Type *Ty = Scalar->getType(); if (!Ty->isVoidTy()) { for (User *U : Scalar->users()) { LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n"); // It is legal to delete users in the ignorelist. assert((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull(U) && isDeleted(cast(U)))) && "Deleting out-of-tree value"); } } #endif LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n"); eraseInstruction(cast(Scalar)); // Retain to-be-deleted instructions for some debug-info // bookkeeping. NOTE: eraseInstruction only marks the instruction for // deletion - instructions are not deleted until later. RemovedInsts.push_back(cast(Scalar)); } } // Merge the DIAssignIDs from the about-to-be-deleted instructions into the // new vector instruction. if (auto *V = dyn_cast(VectorizableTree[0]->VectorizedValue)) V->mergeDIAssignID(RemovedInsts); Builder.ClearInsertionPoint(); InstrElementSize.clear(); return VectorizableTree[0]->VectorizedValue; } void BoUpSLP::optimizeGatherSequence() { LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherShuffleExtractSeq.size() << " gather sequences instructions.\n"); // LICM InsertElementInst sequences. for (Instruction *I : GatherShuffleExtractSeq) { if (isDeleted(I)) continue; // Check if this block is inside a loop. Loop *L = LI->getLoopFor(I->getParent()); if (!L) continue; // Check if it has a preheader. BasicBlock *PreHeader = L->getLoopPreheader(); if (!PreHeader) continue; // If the vector or the element that we insert into it are // instructions that are defined in this basic block then we can't // hoist this instruction. if (any_of(I->operands(), [L](Value *V) { auto *OpI = dyn_cast(V); return OpI && L->contains(OpI); })) continue; // We can hoist this instruction. Move it to the pre-header. I->moveBefore(PreHeader->getTerminator()); CSEBlocks.insert(PreHeader); } // 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. llvm::sort(CSEWorkList, [](const DomTreeNode *A, const DomTreeNode *B) { assert((A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) && "Different nodes should have different DFS numbers"); return A->getDFSNumIn() < B->getDFSNumIn(); }); // Less defined shuffles can be replaced by the more defined copies. // Between two shuffles one is less defined if it has the same vector operands // and its mask indeces are the same as in the first one or undefs. E.g. // shuffle %0, poison, <0, 0, 0, undef> is less defined than shuffle %0, // poison, <0, 0, 0, 0>. auto &&IsIdenticalOrLessDefined = [this](Instruction *I1, Instruction *I2, SmallVectorImpl &NewMask) { if (I1->getType() != I2->getType()) return false; auto *SI1 = dyn_cast(I1); auto *SI2 = dyn_cast(I2); if (!SI1 || !SI2) return I1->isIdenticalTo(I2); if (SI1->isIdenticalTo(SI2)) return true; for (int I = 0, E = SI1->getNumOperands(); I < E; ++I) if (SI1->getOperand(I) != SI2->getOperand(I)) return false; // Check if the second instruction is more defined than the first one. NewMask.assign(SI2->getShuffleMask().begin(), SI2->getShuffleMask().end()); ArrayRef SM1 = SI1->getShuffleMask(); // Count trailing undefs in the mask to check the final number of used // registers. unsigned LastUndefsCnt = 0; for (int I = 0, E = NewMask.size(); I < E; ++I) { if (SM1[I] == UndefMaskElem) ++LastUndefsCnt; else LastUndefsCnt = 0; if (NewMask[I] != UndefMaskElem && SM1[I] != UndefMaskElem && NewMask[I] != SM1[I]) return false; if (NewMask[I] == UndefMaskElem) NewMask[I] = SM1[I]; } // Check if the last undefs actually change the final number of used vector // registers. return SM1.size() - LastUndefsCnt > 1 && TTI->getNumberOfParts(SI1->getType()) == TTI->getNumberOfParts( FixedVectorType::get(SI1->getType()->getElementType(), SM1.size() - LastUndefsCnt)); }; // Perform O(N^2) search over the gather/shuffle sequences and merge identical // instructions. TODO: We can further optimize this scan if we split the // instructions into different buckets based on the insert lane. SmallVector Visited; for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { assert(*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!"); BasicBlock *BB = (*I)->getBlock(); // For all instructions in blocks containing gather sequences: for (Instruction &In : llvm::make_early_inc_range(*BB)) { if (isDeleted(&In)) continue; if (!isa(&In) && !GatherShuffleExtractSeq.contains(&In)) continue; // Check if we can replace this instruction with any of the // visited instructions. bool Replaced = false; for (Instruction *&V : Visited) { SmallVector NewMask; if (IsIdenticalOrLessDefined(&In, V, NewMask) && DT->dominates(V->getParent(), In.getParent())) { In.replaceAllUsesWith(V); eraseInstruction(&In); if (auto *SI = dyn_cast(V)) if (!NewMask.empty()) SI->setShuffleMask(NewMask); Replaced = true; break; } if (isa(In) && isa(V) && GatherShuffleExtractSeq.contains(V) && IsIdenticalOrLessDefined(V, &In, NewMask) && DT->dominates(In.getParent(), V->getParent())) { In.moveAfter(V); V->replaceAllUsesWith(&In); eraseInstruction(V); if (auto *SI = dyn_cast(&In)) if (!NewMask.empty()) SI->setShuffleMask(NewMask); V = &In; Replaced = true; break; } } if (!Replaced) { assert(!is_contained(Visited, &In)); Visited.push_back(&In); } } } CSEBlocks.clear(); GatherShuffleExtractSeq.clear(); } BoUpSLP::ScheduleData * BoUpSLP::BlockScheduling::buildBundle(ArrayRef VL) { ScheduleData *Bundle = nullptr; ScheduleData *PrevInBundle = nullptr; for (Value *V : VL) { if (doesNotNeedToBeScheduled(V)) continue; ScheduleData *BundleMember = getScheduleData(V); assert(BundleMember && "no ScheduleData for bundle member " "(maybe not in same basic block)"); assert(BundleMember->isSchedulingEntity() && "bundle member already part of other bundle"); if (PrevInBundle) { PrevInBundle->NextInBundle = BundleMember; } else { Bundle = BundleMember; } // Group the instructions to a bundle. BundleMember->FirstInBundle = Bundle; PrevInBundle = BundleMember; } assert(Bundle && "Failed to find schedule bundle"); return Bundle; } // Groups the instructions to a bundle (which is then a single scheduling entity) // and schedules instructions until the bundle gets ready. std::optional BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef VL, BoUpSLP *SLP, const InstructionsState &S) { // No need to schedule PHIs, insertelement, extractelement and extractvalue // instructions. if (isa(S.OpValue) || isVectorLikeInstWithConstOps(S.OpValue) || doesNotNeedToSchedule(VL)) return nullptr; // Initialize the instruction bundle. Instruction *OldScheduleEnd = ScheduleEnd; LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n"); auto TryScheduleBundleImpl = [this, OldScheduleEnd, SLP](bool ReSchedule, ScheduleData *Bundle) { // 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. if (ScheduleEnd != OldScheduleEnd) { for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) doForAllOpcodes(I, [](ScheduleData *SD) { SD->clearDependencies(); }); ReSchedule = true; } if (Bundle) { LLVM_DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"); calculateDependencies(Bundle, /*InsertInReadyList=*/true, SLP); } if (ReSchedule) { resetSchedule(); initialFillReadyList(ReadyInsts); } // Now try to schedule the new bundle or (if no bundle) just calculate // dependencies. 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 && ReSchedule) || (Bundle && !Bundle->isReady())) && !ReadyInsts.empty()) { ScheduleData *Picked = ReadyInsts.pop_back_val(); assert(Picked->isSchedulingEntity() && Picked->isReady() && "must be ready to schedule"); schedule(Picked, ReadyInsts); } }; // Make sure that the scheduling region contains all // instructions of the bundle. for (Value *V : VL) { if (doesNotNeedToBeScheduled(V)) continue; if (!extendSchedulingRegion(V, S)) { // If 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. // Otherwise the compiler may crash trying to incorrectly calculate // dependencies and emit instruction in the wrong order at the actual // scheduling. TryScheduleBundleImpl(/*ReSchedule=*/false, nullptr); return std::nullopt; } } bool ReSchedule = false; for (Value *V : VL) { if (doesNotNeedToBeScheduled(V)) continue; ScheduleData *BundleMember = getScheduleData(V); assert(BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)"); // Make sure we don't leave the pieces of the bundle in the ready list when // whole bundle might not be ready. ReadyInsts.remove(BundleMember); if (!BundleMember->IsScheduled) continue; // 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. LLVM_DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"); ReSchedule = true; } auto *Bundle = buildBundle(VL); TryScheduleBundleImpl(ReSchedule, Bundle); if (!Bundle->isReady()) { cancelScheduling(VL, S.OpValue); return std::nullopt; } return Bundle; } void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef VL, Value *OpValue) { if (isa(OpValue) || isVectorLikeInstWithConstOps(OpValue) || doesNotNeedToSchedule(VL)) return; if (doesNotNeedToBeScheduled(OpValue)) OpValue = *find_if_not(VL, doesNotNeedToBeScheduled); ScheduleData *Bundle = getScheduleData(OpValue); LLVM_DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"); assert(!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled"); assert(Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction(VL)) && "tried to unbundle something which is not a bundle"); // Remove the bundle from the ready list. if (Bundle->isReady()) ReadyInsts.remove(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->TE = nullptr; if (BundleMember->unscheduledDepsInBundle() == 0) { ReadyInsts.insert(BundleMember); } BundleMember = Next; } } BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { // Allocate a new ScheduleData for the instruction. if (ChunkPos >= ChunkSize) { ScheduleDataChunks.push_back(std::make_unique(ChunkSize)); ChunkPos = 0; } return &(ScheduleDataChunks.back()[ChunkPos++]); } bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, const InstructionsState &S) { if (getScheduleData(V, isOneOf(S, V))) return true; Instruction *I = dyn_cast(V); assert(I && "bundle member must be an instruction"); assert(!isa(I) && !isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled(I) && "phi nodes/insertelements/extractelements/extractvalues don't need to " "be scheduled"); auto &&CheckScheduleForI = [this, &S](Instruction *I) -> bool { ScheduleData *ISD = getScheduleData(I); if (!ISD) return false; assert(isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region"); ScheduleData *SD = allocateScheduleDataChunks(); SD->Inst = I; SD->init(SchedulingRegionID, S.OpValue); ExtraScheduleDataMap[I][S.OpValue] = SD; return true; }; if (CheckScheduleForI(I)) return true; if (!ScheduleStart) { // It's the first instruction in the new region. initScheduleData(I, I->getNextNode(), nullptr, nullptr); ScheduleStart = I; ScheduleEnd = I->getNextNode(); if (isOneOf(S, I) != I) CheckScheduleForI(I); assert(ScheduleEnd && "tried to vectorize a terminator?"); LLVM_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().getReverse(); BasicBlock::reverse_iterator UpperEnd = BB->rend(); BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); BasicBlock::iterator LowerEnd = BB->end(); while (UpIter != UpperEnd && DownIter != LowerEnd && &*UpIter != I && &*DownIter != I) { if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { LLVM_DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n"); return false; } ++UpIter; ++DownIter; } if (DownIter == LowerEnd || (UpIter != UpperEnd && &*UpIter == I)) { assert(I->getParent() == ScheduleStart->getParent() && "Instruction is in wrong basic block."); initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); ScheduleStart = I; if (isOneOf(S, I) != I) CheckScheduleForI(I); LLVM_DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n"); return true; } assert((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end."); assert(I->getParent() == ScheduleEnd->getParent() && "Instruction is in wrong basic block."); initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, nullptr); ScheduleEnd = I->getNextNode(); if (isOneOf(S, I) != I) CheckScheduleForI(I); assert(ScheduleEnd && "tried to vectorize a terminator?"); LLVM_DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n"); 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()) { // No need to allocate data for non-schedulable instructions. if (doesNotNeedToBeScheduled(I)) continue; ScheduleData *SD = ScheduleDataMap.lookup(I); if (!SD) { SD = allocateScheduleDataChunks(); ScheduleDataMap[I] = SD; SD->Inst = I; } assert(!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region"); SD->init(SchedulingRegionID, I); if (I->mayReadOrWriteMemory() && (!isa(I) || (cast(I)->getIntrinsicID() != Intrinsic::sideeffect && cast(I)->getIntrinsicID() != Intrinsic::pseudoprobe))) { // Update the linked list of memory accessing instructions. if (CurrentLoadStore) { CurrentLoadStore->NextLoadStore = SD; } else { FirstLoadStoreInRegion = SD; } CurrentLoadStore = SD; } if (match(I, m_Intrinsic()) || match(I, m_Intrinsic())) RegionHasStackSave = true; } 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.pop_back_val(); for (ScheduleData *BundleMember = SD; BundleMember; BundleMember = BundleMember->NextInBundle) { assert(isInSchedulingRegion(BundleMember)); if (BundleMember->hasValidDependencies()) continue; LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n"); BundleMember->Dependencies = 0; BundleMember->resetUnscheduledDeps(); // Handle def-use chain dependencies. if (BundleMember->OpValue != BundleMember->Inst) { if (ScheduleData *UseSD = getScheduleData(BundleMember->Inst)) { BundleMember->Dependencies++; ScheduleData *DestBundle = UseSD->FirstInBundle; if (!DestBundle->IsScheduled) BundleMember->incrementUnscheduledDeps(1); if (!DestBundle->hasValidDependencies()) WorkList.push_back(DestBundle); } } else { for (User *U : BundleMember->Inst->users()) { if (ScheduleData *UseSD = getScheduleData(cast(U))) { BundleMember->Dependencies++; ScheduleData *DestBundle = UseSD->FirstInBundle; if (!DestBundle->IsScheduled) BundleMember->incrementUnscheduledDeps(1); if (!DestBundle->hasValidDependencies()) WorkList.push_back(DestBundle); } } } auto makeControlDependent = [&](Instruction *I) { auto *DepDest = getScheduleData(I); assert(DepDest && "must be in schedule window"); DepDest->ControlDependencies.push_back(BundleMember); BundleMember->Dependencies++; ScheduleData *DestBundle = DepDest->FirstInBundle; if (!DestBundle->IsScheduled) BundleMember->incrementUnscheduledDeps(1); if (!DestBundle->hasValidDependencies()) WorkList.push_back(DestBundle); }; // Any instruction which isn't safe to speculate at the beginning of the // block is control dependend on any early exit or non-willreturn call // which proceeds it. if (!isGuaranteedToTransferExecutionToSuccessor(BundleMember->Inst)) { for (Instruction *I = BundleMember->Inst->getNextNode(); I != ScheduleEnd; I = I->getNextNode()) { if (isSafeToSpeculativelyExecute(I, &*BB->begin(), SLP->AC)) continue; // Add the dependency makeControlDependent(I); if (!isGuaranteedToTransferExecutionToSuccessor(I)) // Everything past here must be control dependent on I. break; } } if (RegionHasStackSave) { // If we have an inalloc alloca instruction, it needs to be scheduled // after any preceeding stacksave. We also need to prevent any alloca // from reordering above a preceeding stackrestore. if (match(BundleMember->Inst, m_Intrinsic()) || match(BundleMember->Inst, m_Intrinsic())) { for (Instruction *I = BundleMember->Inst->getNextNode(); I != ScheduleEnd; I = I->getNextNode()) { if (match(I, m_Intrinsic()) || match(I, m_Intrinsic())) // Any allocas past here must be control dependent on I, and I // must be memory dependend on BundleMember->Inst. break; if (!isa(I)) continue; // Add the dependency makeControlDependent(I); } } // In addition to the cases handle just above, we need to prevent // allocas and loads/stores from moving below a stacksave or a // stackrestore. Avoiding moving allocas below stackrestore is currently // thought to be conservatism. Moving loads/stores below a stackrestore // can lead to incorrect code. if (isa(BundleMember->Inst) || BundleMember->Inst->mayReadOrWriteMemory()) { for (Instruction *I = BundleMember->Inst->getNextNode(); I != ScheduleEnd; I = I->getNextNode()) { if (!match(I, m_Intrinsic()) && !match(I, m_Intrinsic())) continue; // Add the dependency makeControlDependent(I); break; } } } // Handle the memory dependencies (if any). ScheduleData *DepDest = BundleMember->NextLoadStore; if (!DepDest) continue; Instruction *SrcInst = BundleMember->Inst; assert(SrcInst->mayReadOrWriteMemory() && "NextLoadStore list for non memory effecting bundle?"); MemoryLocation SrcLoc = getLocation(SrcInst); bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); unsigned numAliased = 0; unsigned DistToSrc = 1; for ( ; DepDest; DepDest = DepDest->NextLoadStore) { 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); } } // 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++; } } if (InsertInReadyList && SD->isReady()) { ReadyInsts.insert(SD); LLVM_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()) { doForAllOpcodes(I, [&](ScheduleData *SD) { assert(isInSchedulingRegion(SD) && "ScheduleData not in scheduling region"); SD->IsScheduled = false; SD->resetUnscheduledDeps(); }); } ReadyInsts.clear(); } void BoUpSLP::scheduleBlock(BlockScheduling *BS) { if (!BS->ScheduleStart) return; LLVM_DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n"); // A key point - if we got here, pre-scheduling was able to find a valid // scheduling of the sub-graph of the scheduling window which consists // of all vector bundles and their transitive users. As such, we do not // need to reschedule anything *outside of* that subgraph. 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. // WARNING: If changing this order causes a correctness issue, that means // there is some missing dependence edge in the schedule data graph. struct ScheduleDataCompare { bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { return SD2->SchedulingPriority < SD1->SchedulingPriority; } }; std::set ReadyInsts; // Ensure that all dependency data is updated (for nodes in the sub-graph) // and fill the ready-list with initial instructions. int Idx = 0; for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; I = I->getNextNode()) { BS->doForAllOpcodes(I, [this, &Idx, BS](ScheduleData *SD) { TreeEntry *SDTE = getTreeEntry(SD->Inst); (void)SDTE; assert((isVectorLikeInstWithConstOps(SD->Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) && "scheduler and vectorizer bundle mismatch"); SD->FirstInBundle->SchedulingPriority = Idx++; if (SD->isSchedulingEntity() && SD->isPartOfBundle()) BS->calculateDependencies(SD, false, this); }); } 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. for (ScheduleData *BundleMember = picked; BundleMember; BundleMember = BundleMember->NextInBundle) { Instruction *pickedInst = BundleMember->Inst; if (pickedInst->getNextNode() != LastScheduledInst) pickedInst->moveBefore(LastScheduledInst); LastScheduledInst = pickedInst; } BS->schedule(picked, ReadyInsts); } // Check that we didn't break any of our invariants. #ifdef EXPENSIVE_CHECKS BS->verify(); #endif #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) // Check that all schedulable entities got scheduled for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; I = I->getNextNode()) { BS->doForAllOpcodes(I, [&](ScheduleData *SD) { if (SD->isSchedulingEntity() && SD->hasValidDependencies()) { assert(SD->IsScheduled && "must be scheduled at this point"); } }); } #endif // Avoid duplicate scheduling of the block. BS->ScheduleStart = nullptr; } unsigned BoUpSLP::getVectorElementSize(Value *V) { // If V is a store, just return the width of the stored value (or value // truncated just before storing) without traversing the expression tree. // This is the common case. if (auto *Store = dyn_cast(V)) return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); if (auto *IEI = dyn_cast(V)) return getVectorElementSize(IEI->getOperand(1)); auto E = InstrElementSize.find(V); if (E != InstrElementSize.end()) return E->second; // If V is not a store, we can traverse the expression tree to find loads // that feed it. The type of the loaded value may indicate a more suitable // width than V's type. We want to base the vector element size on the width // of memory operations where possible. SmallVector, 16> Worklist; SmallPtrSet Visited; if (auto *I = dyn_cast(V)) { Worklist.emplace_back(I, I->getParent()); Visited.insert(I); } // Traverse the expression tree in bottom-up order looking for loads. If we // encounter an instruction we don't yet handle, we give up. auto Width = 0u; while (!Worklist.empty()) { Instruction *I; BasicBlock *Parent; std::tie(I, Parent) = Worklist.pop_back_val(); // We should only be looking at scalar instructions here. If the current // instruction has a vector type, skip. auto *Ty = I->getType(); if (isa(Ty)) continue; // If the current instruction is a load, update MaxWidth to reflect the // width of the loaded value. if (isa(I)) Width = std::max(Width, DL->getTypeSizeInBits(Ty)); // Otherwise, we need to visit the operands of the instruction. We only // handle the interesting cases from buildTree here. If an operand is an // instruction we haven't yet visited and from the same basic block as the // user or the use is a PHI node, we add it to the worklist. else if (isa(I)) { for (Use &U : I->operands()) if (auto *J = dyn_cast(U.get())) if (Visited.insert(J).second && (isa(I) || J->getParent() == Parent)) Worklist.emplace_back(J, J->getParent()); } else { break; } } // If we didn't encounter a memory access in the expression tree, or if we // gave up for some reason, just return the width of V. Otherwise, return the // maximum width we found. if (!Width) { if (auto *CI = dyn_cast(V)) V = CI->getOperand(0); Width = DL->getTypeSizeInBits(V->getType()); } for (Instruction *I : Visited) InstrElementSize[I] = Width; return Width; } // Determine if a value V in a vectorizable expression Expr can be demoted to a // smaller type with a truncation. We collect the values that will be demoted // in ToDemote and additional roots that require investigating in Roots. static bool collectValuesToDemote(Value *V, SmallPtrSetImpl &Expr, SmallVectorImpl &ToDemote, SmallVectorImpl &Roots) { // We can always demote constants. if (isa(V)) { ToDemote.push_back(V); return true; } // If the value is not an instruction in the expression with only one use, it // cannot be demoted. auto *I = dyn_cast(V); if (!I || !I->hasOneUse() || !Expr.count(I)) return false; switch (I->getOpcode()) { // We can always demote truncations and extensions. Since truncations can // seed additional demotion, we save the truncated value. case Instruction::Trunc: Roots.push_back(I->getOperand(0)); break; case Instruction::ZExt: case Instruction::SExt: if (isa(I->getOperand(0))) return false; break; // We can demote certain binary operations if we can demote both of their // operands. case Instruction::Add: case Instruction::Sub: case Instruction::Mul: case Instruction::And: case Instruction::Or: case Instruction::Xor: if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) return false; break; // We can demote selects if we can demote their true and false values. case Instruction::Select: { SelectInst *SI = cast(I); if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) return false; break; } // We can demote phis if we can demote all their incoming operands. Note that // we don't need to worry about cycles since we ensure single use above. case Instruction::PHI: { PHINode *PN = cast(I); for (Value *IncValue : PN->incoming_values()) if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) return false; break; } // Otherwise, conservatively give up. default: return false; } // Record the value that we can demote. ToDemote.push_back(V); return true; } void BoUpSLP::computeMinimumValueSizes() { // If there are no external uses, the expression tree must be rooted by a // store. We can't demote in-memory values, so there is nothing to do here. if (ExternalUses.empty()) return; // We only attempt to truncate integer expressions. auto &TreeRoot = VectorizableTree[0]->Scalars; auto *TreeRootIT = dyn_cast(TreeRoot[0]->getType()); if (!TreeRootIT) return; // If the expression is not rooted by a store, these roots should have // external uses. We will rely on InstCombine to rewrite the expression in // the narrower type. However, InstCombine only rewrites single-use values. // This means that if a tree entry other than a root is used externally, it // must have multiple uses and InstCombine will not rewrite it. The code // below ensures that only the roots are used externally. SmallPtrSet Expr(TreeRoot.begin(), TreeRoot.end()); for (auto &EU : ExternalUses) if (!Expr.erase(EU.Scalar)) return; if (!Expr.empty()) return; // Collect the scalar values of the vectorizable expression. We will use this // context to determine which values can be demoted. If we see a truncation, // we mark it as seeding another demotion. for (auto &EntryPtr : VectorizableTree) Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); // Ensure the roots of the vectorizable tree don't form a cycle. They must // have a single external user that is not in the vectorizable tree. for (auto *Root : TreeRoot) if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) return; // Conservatively determine if we can actually truncate the roots of the // expression. Collect the values that can be demoted in ToDemote and // additional roots that require investigating in Roots. SmallVector ToDemote; SmallVector Roots; for (auto *Root : TreeRoot) if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) return; // The maximum bit width required to represent all the values that can be // demoted without loss of precision. It would be safe to truncate the roots // of the expression to this width. auto MaxBitWidth = 8u; // We first check if all the bits of the roots are demanded. If they're not, // we can truncate the roots to this narrower type. for (auto *Root : TreeRoot) { auto Mask = DB->getDemandedBits(cast(Root)); MaxBitWidth = std::max( Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); } // True if the roots can be zero-extended back to their original type, rather // than sign-extended. We know that if the leading bits are not demanded, we // can safely zero-extend. So we initialize IsKnownPositive to True. bool IsKnownPositive = true; // If all the bits of the roots are demanded, we can try a little harder to // compute a narrower type. This can happen, for example, if the roots are // getelementptr indices. InstCombine promotes these indices to the pointer // width. Thus, all their bits are technically demanded even though the // address computation might be vectorized in a smaller type. // // We start by looking at each entry that can be demoted. We compute the // maximum bit width required to store the scalar by using ValueTracking to // compute the number of high-order bits we can truncate. if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && llvm::all_of(TreeRoot, [](Value *R) { assert(R->hasOneUse() && "Root should have only one use!"); return isa(R->user_back()); })) { MaxBitWidth = 8u; // Determine if the sign bit of all the roots is known to be zero. If not, // IsKnownPositive is set to False. IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { KnownBits Known = computeKnownBits(R, *DL); return Known.isNonNegative(); }); // Determine the maximum number of bits required to store the scalar // values. for (auto *Scalar : ToDemote) { auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); MaxBitWidth = std::max(NumTypeBits - NumSignBits, MaxBitWidth); } // If we can't prove that the sign bit is zero, we must add one to the // maximum bit width to account for the unknown sign bit. This preserves // the existing sign bit so we can safely sign-extend the root back to the // original type. Otherwise, if we know the sign bit is zero, we will // zero-extend the root instead. // // FIXME: This is somewhat suboptimal, as there will be cases where adding // one to the maximum bit width will yield a larger-than-necessary // type. In general, we need to add an extra bit only if we can't // prove that the upper bit of the original type is equal to the // upper bit of the proposed smaller type. If these two bits are the // same (either zero or one) we know that sign-extending from the // smaller type will result in the same value. Here, since we can't // yet prove this, we are just making the proposed smaller type // larger to ensure correctness. if (!IsKnownPositive) ++MaxBitWidth; } // Round MaxBitWidth up to the next power-of-two. if (!isPowerOf2_64(MaxBitWidth)) MaxBitWidth = NextPowerOf2(MaxBitWidth); // If the maximum bit width we compute is less than the with of the roots' // type, we can proceed with the narrowing. Otherwise, do nothing. if (MaxBitWidth >= TreeRootIT->getBitWidth()) return; // If we can truncate the root, we must collect additional values that might // be demoted as a result. That is, those seeded by truncations we will // modify. while (!Roots.empty()) collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); // Finally, map the values we can demote to the maximum bit with we computed. for (auto *Scalar : ToDemote) MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); } namespace { /// The SLPVectorizer Pass. struct SLPVectorizer : public FunctionPass { SLPVectorizerPass Impl; /// Pass identification, replacement for typeid static char ID; explicit SLPVectorizer() : FunctionPass(ID) { initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); } bool doInitialization(Module &M) override { return false; } bool runOnFunction(Function &F) override { if (skipFunction(F)) return false; auto *SE = &getAnalysis().getSE(); auto *TTI = &getAnalysis().getTTI(F); auto *TLIP = getAnalysisIfAvailable(); auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; auto *AA = &getAnalysis().getAAResults(); auto *LI = &getAnalysis().getLoopInfo(); auto *DT = &getAnalysis().getDomTree(); auto *AC = &getAnalysis().getAssumptionCache(F); auto *DB = &getAnalysis().getDemandedBits(); auto *ORE = &getAnalysis().getORE(); return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); } 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.addPreserved(); AU.addPreserved(); AU.addPreserved(); AU.addPreserved(); AU.setPreservesCFG(); } }; } // end anonymous namespace PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { auto *SE = &AM.getResult(F); auto *TTI = &AM.getResult(F); auto *TLI = AM.getCachedResult(F); auto *AA = &AM.getResult(F); auto *LI = &AM.getResult(F); auto *DT = &AM.getResult(F); auto *AC = &AM.getResult(F); auto *DB = &AM.getResult(F); auto *ORE = &AM.getResult(F); bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); if (!Changed) return PreservedAnalyses::all(); PreservedAnalyses PA; PA.preserveSet(); return PA; } bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, TargetTransformInfo *TTI_, TargetLibraryInfo *TLI_, AAResults *AA_, LoopInfo *LI_, DominatorTree *DT_, AssumptionCache *AC_, DemandedBits *DB_, OptimizationRemarkEmitter *ORE_) { if (!RunSLPVectorization) return false; SE = SE_; TTI = TTI_; TLI = TLI_; AA = AA_; LI = LI_; DT = DT_; AC = AC_; DB = DB_; DL = &F.getParent()->getDataLayout(); Stores.clear(); GEPs.clear(); bool Changed = false; // If the target claims to have no vector registers don't attempt // vectorization. if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) { LLVM_DEBUG( dbgs() << "SLP: Didn't find any vector registers for target, abort.\n"); return false; } // Don't vectorize when the attribute NoImplicitFloat is used. if (F.hasFnAttribute(Attribute::NoImplicitFloat)) return false; LLVM_DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"); // Use the bottom up slp vectorizer to construct chains that start with // store instructions. BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to // delete instructions. // Update DFS numbers now so that we can use them for ordering. DT->updateDFSNumbers(); // Scan the blocks in the function in post order. for (auto *BB : post_order(&F.getEntryBlock())) { // Start new block - clear the list of reduction roots. R.clearReductionData(); collectSeedInstructions(BB); // Vectorize trees that end at stores. if (!Stores.empty()) { LLVM_DEBUG(dbgs() << "SLP: Found stores for " << Stores.size() << " underlying objects.\n"); Changed |= vectorizeStoreChains(R); } // Vectorize trees that end at reductions. Changed |= vectorizeChainsInBlock(BB, R); // Vectorize the index computations of getelementptr instructions. This // is primarily intended to catch gather-like idioms ending at // non-consecutive loads. if (!GEPs.empty()) { LLVM_DEBUG(dbgs() << "SLP: Found GEPs for " << GEPs.size() << " underlying objects.\n"); Changed |= vectorizeGEPIndices(BB, R); } } if (Changed) { R.optimizeGatherSequence(); LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n"); } return Changed; } bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef Chain, BoUpSLP &R, unsigned Idx, unsigned MinVF) { LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"); const unsigned Sz = R.getVectorElementSize(Chain[0]); unsigned VF = Chain.size(); if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) return false; LLVM_DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"); R.buildTree(Chain); if (R.isTreeTinyAndNotFullyVectorizable()) return false; if (R.isLoadCombineCandidate()) return false; R.reorderTopToBottom(); R.reorderBottomToTop(); R.buildExternalUses(); R.computeMinimumValueSizes(); InstructionCost Cost = R.getTreeCost(); LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for VF=" << VF << "\n"); if (Cost < -SLPCostThreshold) { LLVM_DEBUG(dbgs() << "SLP: Decided to vectorize cost = " << Cost << "\n"); using namespace ore; R.getORE()->emit(OptimizationRemark(SV_NAME, "StoresVectorized", cast(Chain[0])) << "Stores SLP vectorized with cost " << NV("Cost", Cost) << " and with tree size " << NV("TreeSize", R.getTreeSize())); R.vectorizeTree(); return true; } return false; } bool SLPVectorizerPass::vectorizeStores(ArrayRef Stores, BoUpSLP &R) { // 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. BoUpSLP::ValueSet VectorizedStores; bool Changed = false; int E = Stores.size(); SmallBitVector Tails(E, false); int MaxIter = MaxStoreLookup.getValue(); SmallVector, 16> ConsecutiveChain( E, std::make_pair(E, INT_MAX)); SmallVector CheckedPairs(E, SmallBitVector(E, false)); int IterCnt; auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, &CheckedPairs, &ConsecutiveChain](int K, int Idx) { if (IterCnt >= MaxIter) return true; if (CheckedPairs[Idx].test(K)) return ConsecutiveChain[K].second == 1 && ConsecutiveChain[K].first == Idx; ++IterCnt; CheckedPairs[Idx].set(K); CheckedPairs[K].set(Idx); std::optional Diff = getPointersDiff( Stores[K]->getValueOperand()->getType(), Stores[K]->getPointerOperand(), Stores[Idx]->getValueOperand()->getType(), Stores[Idx]->getPointerOperand(), *DL, *SE, /*StrictCheck=*/true); if (!Diff || *Diff == 0) return false; int Val = *Diff; if (Val < 0) { if (ConsecutiveChain[Idx].second > -Val) { Tails.set(K); ConsecutiveChain[Idx] = std::make_pair(K, -Val); } return false; } if (ConsecutiveChain[K].second <= Val) return false; Tails.set(Idx); ConsecutiveChain[K] = std::make_pair(Idx, Val); return Val == 1; }; // Do a quadratic search on all of the given stores in reverse order and find // all of the pairs of stores that follow each other. for (int Idx = E - 1; Idx >= 0; --Idx) { // If a store has multiple consecutive store candidates, search according // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... // This is because usually pairing with immediate succeeding or preceding // candidate create the best chance to find slp vectorization opportunity. const int MaxLookDepth = std::max(E - Idx, Idx + 1); IterCnt = 0; for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) break; } // Tracks if we tried to vectorize stores starting from the given tail // already. SmallBitVector TriedTails(E, false); // For stores that start but don't end a link in the chain: for (int Cnt = E; Cnt > 0; --Cnt) { int I = Cnt - 1; if (ConsecutiveChain[I].first == E || Tails.test(I)) continue; // We found a store instr that starts a chain. Now follow the chain and try // to vectorize it. BoUpSLP::ValueList Operands; // Collect the chain into a list. while (I != E && !VectorizedStores.count(Stores[I])) { Operands.push_back(Stores[I]); Tails.set(I); if (ConsecutiveChain[I].second != 1) { // Mark the new end in the chain and go back, if required. It might be // required if the original stores come in reversed order, for example. if (ConsecutiveChain[I].first != E && Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) && !VectorizedStores.count(Stores[ConsecutiveChain[I].first])) { TriedTails.set(I); Tails.reset(ConsecutiveChain[I].first); if (Cnt < ConsecutiveChain[I].first + 2) Cnt = ConsecutiveChain[I].first + 2; } break; } // Move to the next value in the chain. I = ConsecutiveChain[I].first; } assert(!Operands.empty() && "Expected non-empty list of stores."); unsigned MaxVecRegSize = R.getMaxVecRegSize(); unsigned EltSize = R.getVectorElementSize(Operands[0]); unsigned MaxElts = llvm::PowerOf2Floor(MaxVecRegSize / EltSize); unsigned MaxVF = std::min(R.getMaximumVF(EltSize, Instruction::Store), MaxElts); auto *Store = cast(Operands[0]); Type *StoreTy = Store->getValueOperand()->getType(); Type *ValueTy = StoreTy; if (auto *Trunc = dyn_cast(Store->getValueOperand())) ValueTy = Trunc->getSrcTy(); unsigned MinVF = TTI->getStoreMinimumVF( R.getMinVF(DL->getTypeSizeInBits(ValueTy)), StoreTy, ValueTy); if (MaxVF <= MinVF) { LLVM_DEBUG(dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF << ") <= " << "MinVF (" << MinVF << ")\n"); } // FIXME: Is division-by-2 the correct step? Should we assert that the // register size is a power-of-2? unsigned StartIdx = 0; for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) { for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { ArrayRef Slice = ArrayRef(Operands).slice(Cnt, Size); if (!VectorizedStores.count(Slice.front()) && !VectorizedStores.count(Slice.back()) && vectorizeStoreChain(Slice, R, Cnt, MinVF)) { // Mark the vectorized stores so that we don't vectorize them again. VectorizedStores.insert(Slice.begin(), Slice.end()); Changed = true; // If we vectorized initial block, no need to try to vectorize it // again. if (Cnt == StartIdx) StartIdx += Size; Cnt += Size; continue; } ++Cnt; } // Check if the whole array was vectorized already - exit. if (StartIdx >= Operands.size()) break; } } return Changed; } void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { // Initialize the collections. We will make a single pass over the block. Stores.clear(); GEPs.clear(); // Visit the store and getelementptr instructions in BB and organize them in // Stores and GEPs according to the underlying objects of their pointer // operands. for (Instruction &I : *BB) { // Ignore store instructions that are volatile or have a pointer operand // that doesn't point to a scalar type. if (auto *SI = dyn_cast(&I)) { if (!SI->isSimple()) continue; if (!isValidElementType(SI->getValueOperand()->getType())) continue; Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); } // Ignore getelementptr instructions that have more than one index, a // constant index, or a pointer operand that doesn't point to a scalar // type. else if (auto *GEP = dyn_cast(&I)) { auto Idx = GEP->idx_begin()->get(); if (GEP->getNumIndices() > 1 || isa(Idx)) continue; if (!isValidElementType(Idx->getType())) continue; if (GEP->getType()->isVectorTy()) continue; GEPs[GEP->getPointerOperand()].push_back(GEP); } } } bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { if (!A || !B) return false; if (isa(A) || isa(B)) return false; Value *VL[] = {A, B}; return tryToVectorizeList(VL, R); } bool SLPVectorizerPass::tryToVectorizeList(ArrayRef VL, BoUpSLP &R, bool LimitForRegisterSize) { if (VL.size() < 2) return false; LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"); // Check that all of the parts are instructions of the same type, // we permit an alternate opcode via InstructionsState. InstructionsState S = getSameOpcode(VL, *TLI); if (!S.getOpcode()) return false; Instruction *I0 = cast(S.OpValue); // Make sure invalid types (including vector type) are rejected before // determining vectorization factor for scalar instructions. for (Value *V : VL) { Type *Ty = V->getType(); if (!isa(V) && !isValidElementType(Ty)) { // NOTE: the following will give user internal llvm type name, which may // not be useful. R.getORE()->emit([&]() { std::string type_str; llvm::raw_string_ostream rso(type_str); Ty->print(rso); return OptimizationRemarkMissed(SV_NAME, "UnsupportedType", I0) << "Cannot SLP vectorize list: type " << rso.str() + " is unsupported by vectorizer"; }); return false; } } unsigned Sz = R.getVectorElementSize(I0); unsigned MinVF = R.getMinVF(Sz); unsigned MaxVF = std::max(PowerOf2Floor(VL.size()), MinVF); MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF); if (MaxVF < 2) { R.getORE()->emit([&]() { return OptimizationRemarkMissed(SV_NAME, "SmallVF", I0) << "Cannot SLP vectorize list: vectorization factor " << "less than 2 is not supported"; }); return false; } bool Changed = false; bool CandidateFound = false; InstructionCost MinCost = SLPCostThreshold.getValue(); Type *ScalarTy = VL[0]->getType(); if (auto *IE = dyn_cast(VL[0])) ScalarTy = IE->getOperand(1)->getType(); unsigned NextInst = 0, MaxInst = VL.size(); for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { // No actual vectorization should happen, if number of parts is the same as // provided vectorization factor (i.e. the scalar type is used for vector // code during codegen). auto *VecTy = FixedVectorType::get(ScalarTy, VF); if (TTI->getNumberOfParts(VecTy) == VF) continue; for (unsigned I = NextInst; I < MaxInst; ++I) { unsigned OpsWidth = 0; if (I + VF > MaxInst) OpsWidth = MaxInst - I; else OpsWidth = VF; if (!isPowerOf2_32(OpsWidth)) continue; if ((LimitForRegisterSize && OpsWidth < MaxVF) || (VF > MinVF && OpsWidth <= VF / 2) || (VF == MinVF && OpsWidth < 2)) break; ArrayRef Ops = VL.slice(I, OpsWidth); // Check that a previous iteration of this loop did not delete the Value. if (llvm::any_of(Ops, [&R](Value *V) { auto *I = dyn_cast(V); return I && R.isDeleted(I); })) continue; LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"); R.buildTree(Ops); if (R.isTreeTinyAndNotFullyVectorizable()) continue; R.reorderTopToBottom(); R.reorderBottomToTop( /*IgnoreReorder=*/!isa(Ops.front()) && !R.doesRootHaveInTreeUses()); R.buildExternalUses(); R.computeMinimumValueSizes(); InstructionCost Cost = R.getTreeCost(); CandidateFound = true; MinCost = std::min(MinCost, Cost); LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for VF=" << OpsWidth << "\n"); if (Cost < -SLPCostThreshold) { LLVM_DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n"); R.getORE()->emit(OptimizationRemark(SV_NAME, "VectorizedList", cast(Ops[0])) << "SLP vectorized with cost " << ore::NV("Cost", Cost) << " and with tree size " << ore::NV("TreeSize", R.getTreeSize())); R.vectorizeTree(); // Move to the next bundle. I += VF - 1; NextInst = I + 1; Changed = true; } } } if (!Changed && CandidateFound) { R.getORE()->emit([&]() { return OptimizationRemarkMissed(SV_NAME, "NotBeneficial", I0) << "List vectorization was possible but not beneficial with cost " << ore::NV("Cost", MinCost) << " >= " << ore::NV("Treshold", -SLPCostThreshold); }); } else if (!Changed) { R.getORE()->emit([&]() { return OptimizationRemarkMissed(SV_NAME, "NotPossible", I0) << "Cannot SLP vectorize list: vectorization was impossible" << " with available vectorization factors"; }); } return Changed; } bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { if (!I) return false; if (!isa(I) || isa(I->getType())) return false; Value *P = I->getParent(); // Vectorize in current basic block only. auto *Op0 = dyn_cast(I->getOperand(0)); auto *Op1 = dyn_cast(I->getOperand(1)); if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) return false; // First collect all possible candidates SmallVector, 4> Candidates; Candidates.emplace_back(Op0, Op1); auto *A = dyn_cast(Op0); auto *B = dyn_cast(Op1); // Try to skip B. if (A && B && B->hasOneUse()) { auto *B0 = dyn_cast(B->getOperand(0)); auto *B1 = dyn_cast(B->getOperand(1)); if (B0 && B0->getParent() == P) Candidates.emplace_back(A, B0); if (B1 && B1->getParent() == P) Candidates.emplace_back(A, B1); } // Try to skip A. if (B && A && A->hasOneUse()) { auto *A0 = dyn_cast(A->getOperand(0)); auto *A1 = dyn_cast(A->getOperand(1)); if (A0 && A0->getParent() == P) Candidates.emplace_back(A0, B); if (A1 && A1->getParent() == P) Candidates.emplace_back(A1, B); } if (Candidates.size() == 1) return tryToVectorizePair(Op0, Op1, R); // We have multiple options. Try to pick the single best. std::optional BestCandidate = R.findBestRootPair(Candidates); if (!BestCandidate) return false; return tryToVectorizePair(Candidates[*BestCandidate].first, Candidates[*BestCandidate].second, R); } namespace { /// Model horizontal reductions. /// /// A horizontal reduction is a tree of reduction instructions that has values /// that can be put into a vector as its leaves. For example: /// /// mul mul mul mul /// \ / \ / /// + + /// \ / /// + /// This tree has "mul" as its leaf values and "+" as its reduction /// instructions. A reduction can feed into a store or a binary operation /// feeding a phi. /// ... /// \ / /// + /// | /// phi += /// /// Or: /// ... /// \ / /// + /// | /// *p = /// class HorizontalReduction { using ReductionOpsType = SmallVector; using ReductionOpsListType = SmallVector; ReductionOpsListType ReductionOps; /// List of possibly reduced values. SmallVector> ReducedVals; /// Maps reduced value to the corresponding reduction operation. DenseMap> ReducedValsToOps; // Use map vector to make stable output. MapVector ExtraArgs; WeakTrackingVH ReductionRoot; /// The type of reduction operation. RecurKind RdxKind; static bool isCmpSelMinMax(Instruction *I) { return match(I, m_Select(m_Cmp(), m_Value(), m_Value())) && RecurrenceDescriptor::isMinMaxRecurrenceKind(getRdxKind(I)); } // And/or are potentially poison-safe logical patterns like: // select x, y, false // select x, true, y static bool isBoolLogicOp(Instruction *I) { return isa(I) && (match(I, m_LogicalAnd()) || match(I, m_LogicalOr())); } /// Checks if instruction is associative and can be vectorized. static bool isVectorizable(RecurKind Kind, Instruction *I) { if (Kind == RecurKind::None) return false; // Integer ops that map to select instructions or intrinsics are fine. if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind) || isBoolLogicOp(I)) return true; if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) { // FP min/max are associative except for NaN and -0.0. We do not // have to rule out -0.0 here because the intrinsic semantics do not // specify a fixed result for it. return I->getFastMathFlags().noNaNs(); } return I->isAssociative(); } static Value *getRdxOperand(Instruction *I, unsigned Index) { // Poison-safe 'or' takes the form: select X, true, Y // To make that work with the normal operand processing, we skip the // true value operand. // TODO: Change the code and data structures to handle this without a hack. if (getRdxKind(I) == RecurKind::Or && isa(I) && Index == 1) return I->getOperand(2); return I->getOperand(Index); } /// Creates reduction operation with the current opcode. static Value *createOp(IRBuilder<> &Builder, RecurKind Kind, Value *LHS, Value *RHS, const Twine &Name, bool UseSelect) { unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(Kind); switch (Kind) { case RecurKind::Or: if (UseSelect && LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) return Builder.CreateSelect(LHS, Builder.getTrue(), RHS, Name); return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, Name); case RecurKind::And: if (UseSelect && LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) return Builder.CreateSelect(LHS, RHS, Builder.getFalse(), Name); return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, Name); case RecurKind::Add: case RecurKind::Mul: case RecurKind::Xor: case RecurKind::FAdd: case RecurKind::FMul: return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, Name); case RecurKind::FMax: return Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS); case RecurKind::FMin: return Builder.CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS); case RecurKind::SMax: if (UseSelect) { Value *Cmp = Builder.CreateICmpSGT(LHS, RHS, Name); return Builder.CreateSelect(Cmp, LHS, RHS, Name); } return Builder.CreateBinaryIntrinsic(Intrinsic::smax, LHS, RHS); case RecurKind::SMin: if (UseSelect) { Value *Cmp = Builder.CreateICmpSLT(LHS, RHS, Name); return Builder.CreateSelect(Cmp, LHS, RHS, Name); } return Builder.CreateBinaryIntrinsic(Intrinsic::smin, LHS, RHS); case RecurKind::UMax: if (UseSelect) { Value *Cmp = Builder.CreateICmpUGT(LHS, RHS, Name); return Builder.CreateSelect(Cmp, LHS, RHS, Name); } return Builder.CreateBinaryIntrinsic(Intrinsic::umax, LHS, RHS); case RecurKind::UMin: if (UseSelect) { Value *Cmp = Builder.CreateICmpULT(LHS, RHS, Name); return Builder.CreateSelect(Cmp, LHS, RHS, Name); } return Builder.CreateBinaryIntrinsic(Intrinsic::umin, LHS, RHS); default: llvm_unreachable("Unknown reduction operation."); } } /// Creates reduction operation with the current opcode with the IR flags /// from \p ReductionOps, dropping nuw/nsw flags. static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, Value *RHS, const Twine &Name, const ReductionOpsListType &ReductionOps) { bool UseSelect = ReductionOps.size() == 2 || // Logical or/and. (ReductionOps.size() == 1 && isa(ReductionOps.front().front())); assert((!UseSelect || ReductionOps.size() != 2 || isa(ReductionOps[1][0])) && "Expected cmp + select pairs for reduction"); Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, UseSelect); if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { if (auto *Sel = dyn_cast(Op)) { propagateIRFlags(Sel->getCondition(), ReductionOps[0], nullptr, /*IncludeWrapFlags=*/false); propagateIRFlags(Op, ReductionOps[1], nullptr, /*IncludeWrapFlags=*/false); return Op; } } propagateIRFlags(Op, ReductionOps[0], nullptr, /*IncludeWrapFlags=*/false); return Op; } static RecurKind getRdxKind(Value *V) { auto *I = dyn_cast(V); if (!I) return RecurKind::None; if (match(I, m_Add(m_Value(), m_Value()))) return RecurKind::Add; if (match(I, m_Mul(m_Value(), m_Value()))) return RecurKind::Mul; if (match(I, m_And(m_Value(), m_Value())) || match(I, m_LogicalAnd(m_Value(), m_Value()))) return RecurKind::And; if (match(I, m_Or(m_Value(), m_Value())) || match(I, m_LogicalOr(m_Value(), m_Value()))) return RecurKind::Or; if (match(I, m_Xor(m_Value(), m_Value()))) return RecurKind::Xor; if (match(I, m_FAdd(m_Value(), m_Value()))) return RecurKind::FAdd; if (match(I, m_FMul(m_Value(), m_Value()))) return RecurKind::FMul; if (match(I, m_Intrinsic(m_Value(), m_Value()))) return RecurKind::FMax; if (match(I, m_Intrinsic(m_Value(), m_Value()))) return RecurKind::FMin; // This matches either cmp+select or intrinsics. SLP is expected to handle // either form. // TODO: If we are canonicalizing to intrinsics, we can remove several // special-case paths that deal with selects. if (match(I, m_SMax(m_Value(), m_Value()))) return RecurKind::SMax; if (match(I, m_SMin(m_Value(), m_Value()))) return RecurKind::SMin; if (match(I, m_UMax(m_Value(), m_Value()))) return RecurKind::UMax; if (match(I, m_UMin(m_Value(), m_Value()))) return RecurKind::UMin; if (auto *Select = dyn_cast(I)) { // Try harder: look for min/max pattern based on instructions producing // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). // During the intermediate stages of SLP, it's very common to have // pattern like this (since optimizeGatherSequence is run only once // at the end): // %1 = extractelement <2 x i32> %a, i32 0 // %2 = extractelement <2 x i32> %a, i32 1 // %cond = icmp sgt i32 %1, %2 // %3 = extractelement <2 x i32> %a, i32 0 // %4 = extractelement <2 x i32> %a, i32 1 // %select = select i1 %cond, i32 %3, i32 %4 CmpInst::Predicate Pred; Instruction *L1; Instruction *L2; Value *LHS = Select->getTrueValue(); Value *RHS = Select->getFalseValue(); Value *Cond = Select->getCondition(); // TODO: Support inverse predicates. if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { if (!isa(RHS) || !L2->isIdenticalTo(cast(RHS))) return RecurKind::None; } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { if (!isa(LHS) || !L1->isIdenticalTo(cast(LHS))) return RecurKind::None; } else { if (!isa(LHS) || !isa(RHS)) return RecurKind::None; if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || !L1->isIdenticalTo(cast(LHS)) || !L2->isIdenticalTo(cast(RHS))) return RecurKind::None; } switch (Pred) { default: return RecurKind::None; case CmpInst::ICMP_SGT: case CmpInst::ICMP_SGE: return RecurKind::SMax; case CmpInst::ICMP_SLT: case CmpInst::ICMP_SLE: return RecurKind::SMin; case CmpInst::ICMP_UGT: case CmpInst::ICMP_UGE: return RecurKind::UMax; case CmpInst::ICMP_ULT: case CmpInst::ICMP_ULE: return RecurKind::UMin; } } return RecurKind::None; } /// Get the index of the first operand. static unsigned getFirstOperandIndex(Instruction *I) { return isCmpSelMinMax(I) ? 1 : 0; } /// Total number of operands in the reduction operation. static unsigned getNumberOfOperands(Instruction *I) { return isCmpSelMinMax(I) ? 3 : 2; } /// Checks if the instruction is in basic block \p BB. /// For a cmp+sel min/max reduction check that both ops are in \p BB. static bool hasSameParent(Instruction *I, BasicBlock *BB) { if (isCmpSelMinMax(I) || isBoolLogicOp(I)) { auto *Sel = cast(I); auto *Cmp = dyn_cast(Sel->getCondition()); return Sel->getParent() == BB && Cmp && Cmp->getParent() == BB; } return I->getParent() == BB; } /// Expected number of uses for reduction operations/reduced values. static bool hasRequiredNumberOfUses(bool IsCmpSelMinMax, Instruction *I) { if (IsCmpSelMinMax) { // SelectInst must be used twice while the condition op must have single // use only. if (auto *Sel = dyn_cast(I)) return Sel->hasNUses(2) && Sel->getCondition()->hasOneUse(); return I->hasNUses(2); } // Arithmetic reduction operation must be used once only. return I->hasOneUse(); } /// Initializes the list of reduction operations. void initReductionOps(Instruction *I) { if (isCmpSelMinMax(I)) ReductionOps.assign(2, ReductionOpsType()); else ReductionOps.assign(1, ReductionOpsType()); } /// Add all reduction operations for the reduction instruction \p I. void addReductionOps(Instruction *I) { if (isCmpSelMinMax(I)) { ReductionOps[0].emplace_back(cast(I)->getCondition()); ReductionOps[1].emplace_back(I); } else { ReductionOps[0].emplace_back(I); } } static Value *getLHS(RecurKind Kind, Instruction *I) { if (Kind == RecurKind::None) return nullptr; return I->getOperand(getFirstOperandIndex(I)); } static Value *getRHS(RecurKind Kind, Instruction *I) { if (Kind == RecurKind::None) return nullptr; return I->getOperand(getFirstOperandIndex(I) + 1); } static bool isGoodForReduction(ArrayRef Data) { int Sz = Data.size(); auto *I = dyn_cast(Data.front()); return Sz > 1 || isConstant(Data.front()) || (I && !isa(I) && isValidForAlternation(I->getOpcode())); } public: HorizontalReduction() = default; /// Try to find a reduction tree. bool matchAssociativeReduction(PHINode *Phi, Instruction *Inst, ScalarEvolution &SE, const DataLayout &DL, const TargetLibraryInfo &TLI) { assert((!Phi || is_contained(Phi->operands(), Inst)) && "Phi needs to use the binary operator"); assert((isa(Inst) || isa(Inst) || isa(Inst)) && "Expected binop, select, or intrinsic for reduction matching"); RdxKind = getRdxKind(Inst); // We could have a initial reductions that is not an add. // r *= v1 + v2 + v3 + v4 // In such a case start looking for a tree rooted in the first '+'. if (Phi) { if (getLHS(RdxKind, Inst) == Phi) { Phi = nullptr; Inst = dyn_cast(getRHS(RdxKind, Inst)); if (!Inst) return false; RdxKind = getRdxKind(Inst); } else if (getRHS(RdxKind, Inst) == Phi) { Phi = nullptr; Inst = dyn_cast(getLHS(RdxKind, Inst)); if (!Inst) return false; RdxKind = getRdxKind(Inst); } } if (!isVectorizable(RdxKind, Inst)) return false; // Analyze "regular" integer/FP types for reductions - no target-specific // types or pointers. Type *Ty = Inst->getType(); if (!isValidElementType(Ty) || Ty->isPointerTy()) return false; // Though the ultimate reduction may have multiple uses, its condition must // have only single use. if (auto *Sel = dyn_cast(Inst)) if (!Sel->getCondition()->hasOneUse()) return false; ReductionRoot = Inst; // Iterate through all the operands of the possible reduction tree and // gather all the reduced values, sorting them by their value id. BasicBlock *BB = Inst->getParent(); bool IsCmpSelMinMax = isCmpSelMinMax(Inst); SmallVector Worklist(1, Inst); // Checks if the operands of the \p TreeN instruction are also reduction // operations or should be treated as reduced values or an extra argument, // which is not part of the reduction. auto &&CheckOperands = [this, IsCmpSelMinMax, BB](Instruction *TreeN, SmallVectorImpl &ExtraArgs, SmallVectorImpl &PossibleReducedVals, SmallVectorImpl &ReductionOps) { for (int I = getFirstOperandIndex(TreeN), End = getNumberOfOperands(TreeN); I < End; ++I) { Value *EdgeVal = getRdxOperand(TreeN, I); ReducedValsToOps[EdgeVal].push_back(TreeN); auto *EdgeInst = dyn_cast(EdgeVal); // Edge has wrong parent - mark as an extra argument. if (EdgeInst && !isVectorLikeInstWithConstOps(EdgeInst) && !hasSameParent(EdgeInst, BB)) { ExtraArgs.push_back(EdgeVal); continue; } // If the edge is not an instruction, or it is different from the main // reduction opcode or has too many uses - possible reduced value. if (!EdgeInst || getRdxKind(EdgeInst) != RdxKind || IsCmpSelMinMax != isCmpSelMinMax(EdgeInst) || !hasRequiredNumberOfUses(IsCmpSelMinMax, EdgeInst) || !isVectorizable(getRdxKind(EdgeInst), EdgeInst)) { PossibleReducedVals.push_back(EdgeVal); continue; } ReductionOps.push_back(EdgeInst); } }; // Try to regroup reduced values so that it gets more profitable to try to // reduce them. Values are grouped by their value ids, instructions - by // instruction op id and/or alternate op id, plus do extra analysis for // loads (grouping them by the distabce between pointers) and cmp // instructions (grouping them by the predicate). MapVector>> PossibleReducedVals; initReductionOps(Inst); DenseMap> LoadsMap; SmallSet LoadKeyUsed; SmallPtrSet DoNotReverseVals; while (!Worklist.empty()) { Instruction *TreeN = Worklist.pop_back_val(); SmallVector Args; SmallVector PossibleRedVals; SmallVector PossibleReductionOps; CheckOperands(TreeN, Args, PossibleRedVals, PossibleReductionOps); // If too many extra args - mark the instruction itself as a reduction // value, not a reduction operation. if (Args.size() < 2) { addReductionOps(TreeN); // Add extra args. if (!Args.empty()) { assert(Args.size() == 1 && "Expected only single argument."); ExtraArgs[TreeN] = Args.front(); } // Add reduction values. The values are sorted for better vectorization // results. for (Value *V : PossibleRedVals) { size_t Key, Idx; std::tie(Key, Idx) = generateKeySubkey( V, &TLI, [&](size_t Key, LoadInst *LI) { Value *Ptr = getUnderlyingObject(LI->getPointerOperand()); if (LoadKeyUsed.contains(Key)) { auto LIt = LoadsMap.find(Ptr); if (LIt != LoadsMap.end()) { for (LoadInst *RLI: LIt->second) { if (getPointersDiff( RLI->getType(), RLI->getPointerOperand(), LI->getType(), LI->getPointerOperand(), DL, SE, /*StrictCheck=*/true)) return hash_value(RLI->getPointerOperand()); } for (LoadInst *RLI : LIt->second) { if (arePointersCompatible(RLI->getPointerOperand(), LI->getPointerOperand(), TLI)) { hash_code SubKey = hash_value(RLI->getPointerOperand()); DoNotReverseVals.insert(RLI); return SubKey; } } if (LIt->second.size() > 2) { hash_code SubKey = hash_value(LIt->second.back()->getPointerOperand()); DoNotReverseVals.insert(LIt->second.back()); return SubKey; } } } LoadKeyUsed.insert(Key); LoadsMap.try_emplace(Ptr).first->second.push_back(LI); return hash_value(LI->getPointerOperand()); }, /*AllowAlternate=*/false); ++PossibleReducedVals[Key][Idx] .insert(std::make_pair(V, 0)) .first->second; } Worklist.append(PossibleReductionOps.rbegin(), PossibleReductionOps.rend()); } else { size_t Key, Idx; std::tie(Key, Idx) = generateKeySubkey( TreeN, &TLI, [&](size_t Key, LoadInst *LI) { Value *Ptr = getUnderlyingObject(LI->getPointerOperand()); if (LoadKeyUsed.contains(Key)) { auto LIt = LoadsMap.find(Ptr); if (LIt != LoadsMap.end()) { for (LoadInst *RLI: LIt->second) { if (getPointersDiff(RLI->getType(), RLI->getPointerOperand(), LI->getType(), LI->getPointerOperand(), DL, SE, /*StrictCheck=*/true)) return hash_value(RLI->getPointerOperand()); } for (LoadInst *RLI : LIt->second) { if (arePointersCompatible(RLI->getPointerOperand(), LI->getPointerOperand(), TLI)) { hash_code SubKey = hash_value(RLI->getPointerOperand()); DoNotReverseVals.insert(RLI); return SubKey; } } if (LIt->second.size() > 2) { hash_code SubKey = hash_value(LIt->second.back()->getPointerOperand()); DoNotReverseVals.insert(LIt->second.back()); return SubKey; } } } LoadKeyUsed.insert(Key); LoadsMap.try_emplace(Ptr).first->second.push_back(LI); return hash_value(LI->getPointerOperand()); }, /*AllowAlternate=*/false); ++PossibleReducedVals[Key][Idx] .insert(std::make_pair(TreeN, 0)) .first->second; } } auto PossibleReducedValsVect = PossibleReducedVals.takeVector(); // Sort values by the total number of values kinds to start the reduction // from the longest possible reduced values sequences. for (auto &PossibleReducedVals : PossibleReducedValsVect) { auto PossibleRedVals = PossibleReducedVals.second.takeVector(); SmallVector> PossibleRedValsVect; for (auto It = PossibleRedVals.begin(), E = PossibleRedVals.end(); It != E; ++It) { PossibleRedValsVect.emplace_back(); auto RedValsVect = It->second.takeVector(); stable_sort(RedValsVect, llvm::less_second()); for (const std::pair &Data : RedValsVect) PossibleRedValsVect.back().append(Data.second, Data.first); } stable_sort(PossibleRedValsVect, [](const auto &P1, const auto &P2) { return P1.size() > P2.size(); }); int NewIdx = -1; for (ArrayRef Data : PossibleRedValsVect) { if (isGoodForReduction(Data) || (isa(Data.front()) && NewIdx >= 0 && isa(ReducedVals[NewIdx].front()) && getUnderlyingObject( cast(Data.front())->getPointerOperand()) == getUnderlyingObject(cast(ReducedVals[NewIdx].front()) ->getPointerOperand()))) { if (NewIdx < 0) { NewIdx = ReducedVals.size(); ReducedVals.emplace_back(); } if (DoNotReverseVals.contains(Data.front())) ReducedVals[NewIdx].append(Data.begin(), Data.end()); else ReducedVals[NewIdx].append(Data.rbegin(), Data.rend()); } else { ReducedVals.emplace_back().append(Data.rbegin(), Data.rend()); } } } // Sort the reduced values by number of same/alternate opcode and/or pointer // operand. stable_sort(ReducedVals, [](ArrayRef P1, ArrayRef P2) { return P1.size() > P2.size(); }); return true; } /// Attempt to vectorize the tree found by matchAssociativeReduction. Value *tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI, const TargetLibraryInfo &TLI) { constexpr int ReductionLimit = 4; constexpr unsigned RegMaxNumber = 4; constexpr unsigned RedValsMaxNumber = 128; // If there are a sufficient number of reduction values, reduce // to a nearby power-of-2. We can safely generate oversized // vectors and rely on the backend to split them to legal sizes. size_t NumReducedVals = std::accumulate(ReducedVals.begin(), ReducedVals.end(), 0, [](size_t Num, ArrayRef Vals) { if (!isGoodForReduction(Vals)) return Num; return Num + Vals.size(); }); if (NumReducedVals < ReductionLimit) { for (ReductionOpsType &RdxOps : ReductionOps) for (Value *RdxOp : RdxOps) V.analyzedReductionRoot(cast(RdxOp)); return nullptr; } IRBuilder<> Builder(cast(ReductionRoot)); // Track the reduced values in case if they are replaced by extractelement // because of the vectorization. DenseMap TrackedVals( ReducedVals.size() * ReducedVals.front().size() + ExtraArgs.size()); BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; ExternallyUsedValues.reserve(ExtraArgs.size() + 1); // The same extra argument may be used several times, so log each attempt // to use it. for (const std::pair &Pair : ExtraArgs) { assert(Pair.first && "DebugLoc must be set."); ExternallyUsedValues[Pair.second].push_back(Pair.first); TrackedVals.try_emplace(Pair.second, Pair.second); } // The compare instruction of a min/max is the insertion point for new // instructions and may be replaced with a new compare instruction. auto &&GetCmpForMinMaxReduction = [](Instruction *RdxRootInst) { assert(isa(RdxRootInst) && "Expected min/max reduction to have select root instruction"); Value *ScalarCond = cast(RdxRootInst)->getCondition(); assert(isa(ScalarCond) && "Expected min/max reduction to have compare condition"); return cast(ScalarCond); }; // The reduction root is used as the insertion point for new instructions, // so set it as externally used to prevent it from being deleted. ExternallyUsedValues[ReductionRoot]; SmallDenseSet IgnoreList(ReductionOps.size() * ReductionOps.front().size()); for (ReductionOpsType &RdxOps : ReductionOps) for (Value *RdxOp : RdxOps) { if (!RdxOp) continue; IgnoreList.insert(RdxOp); } bool IsCmpSelMinMax = isCmpSelMinMax(cast(ReductionRoot)); // Need to track reduced vals, they may be changed during vectorization of // subvectors. for (ArrayRef Candidates : ReducedVals) for (Value *V : Candidates) TrackedVals.try_emplace(V, V); DenseMap VectorizedVals(ReducedVals.size()); // List of the values that were reduced in other trees as part of gather // nodes and thus requiring extract if fully vectorized in other trees. SmallPtrSet RequiredExtract; Value *VectorizedTree = nullptr; bool CheckForReusedReductionOps = false; // Try to vectorize elements based on their type. for (unsigned I = 0, E = ReducedVals.size(); I < E; ++I) { ArrayRef OrigReducedVals = ReducedVals[I]; InstructionsState S = getSameOpcode(OrigReducedVals, TLI); SmallVector Candidates; Candidates.reserve(2 * OrigReducedVals.size()); DenseMap TrackedToOrig(2 * OrigReducedVals.size()); for (unsigned Cnt = 0, Sz = OrigReducedVals.size(); Cnt < Sz; ++Cnt) { Value *RdxVal = TrackedVals.find(OrigReducedVals[Cnt])->second; // Check if the reduction value was not overriden by the extractelement // instruction because of the vectorization and exclude it, if it is not // compatible with other values. if (auto *Inst = dyn_cast(RdxVal)) if (isVectorLikeInstWithConstOps(Inst) && (!S.getOpcode() || !S.isOpcodeOrAlt(Inst))) continue; Candidates.push_back(RdxVal); TrackedToOrig.try_emplace(RdxVal, OrigReducedVals[Cnt]); } bool ShuffledExtracts = false; // Try to handle shuffled extractelements. if (S.getOpcode() == Instruction::ExtractElement && !S.isAltShuffle() && I + 1 < E) { InstructionsState NextS = getSameOpcode(ReducedVals[I + 1], TLI); if (NextS.getOpcode() == Instruction::ExtractElement && !NextS.isAltShuffle()) { SmallVector CommonCandidates(Candidates); for (Value *RV : ReducedVals[I + 1]) { Value *RdxVal = TrackedVals.find(RV)->second; // Check if the reduction value was not overriden by the // extractelement instruction because of the vectorization and // exclude it, if it is not compatible with other values. if (auto *Inst = dyn_cast(RdxVal)) if (!NextS.getOpcode() || !NextS.isOpcodeOrAlt(Inst)) continue; CommonCandidates.push_back(RdxVal); TrackedToOrig.try_emplace(RdxVal, RV); } SmallVector Mask; if (isFixedVectorShuffle(CommonCandidates, Mask)) { ++I; Candidates.swap(CommonCandidates); ShuffledExtracts = true; } } } unsigned NumReducedVals = Candidates.size(); if (NumReducedVals < ReductionLimit) continue; unsigned MaxVecRegSize = V.getMaxVecRegSize(); unsigned EltSize = V.getVectorElementSize(Candidates[0]); unsigned MaxElts = RegMaxNumber * PowerOf2Floor(MaxVecRegSize / EltSize); unsigned ReduxWidth = std::min( PowerOf2Floor(NumReducedVals), std::max(RedValsMaxNumber, MaxElts)); unsigned Start = 0; unsigned Pos = Start; // Restarts vectorization attempt with lower vector factor. unsigned PrevReduxWidth = ReduxWidth; bool CheckForReusedReductionOpsLocal = false; auto &&AdjustReducedVals = [&Pos, &Start, &ReduxWidth, NumReducedVals, &CheckForReusedReductionOpsLocal, &PrevReduxWidth, &V, &IgnoreList](bool IgnoreVL = false) { bool IsAnyRedOpGathered = !IgnoreVL && V.isAnyGathered(IgnoreList); if (!CheckForReusedReductionOpsLocal && PrevReduxWidth == ReduxWidth) { // Check if any of the reduction ops are gathered. If so, worth // trying again with less number of reduction ops. CheckForReusedReductionOpsLocal |= IsAnyRedOpGathered; } ++Pos; if (Pos < NumReducedVals - ReduxWidth + 1) return IsAnyRedOpGathered; Pos = Start; ReduxWidth /= 2; return IsAnyRedOpGathered; }; while (Pos < NumReducedVals - ReduxWidth + 1 && ReduxWidth >= ReductionLimit) { // Dependency in tree of the reduction ops - drop this attempt, try // later. if (CheckForReusedReductionOpsLocal && PrevReduxWidth != ReduxWidth && Start == 0) { CheckForReusedReductionOps = true; break; } PrevReduxWidth = ReduxWidth; ArrayRef VL(std::next(Candidates.begin(), Pos), ReduxWidth); // Beeing analyzed already - skip. if (V.areAnalyzedReductionVals(VL)) { (void)AdjustReducedVals(/*IgnoreVL=*/true); continue; } // Early exit if any of the reduction values were deleted during // previous vectorization attempts. if (any_of(VL, [&V](Value *RedVal) { auto *RedValI = dyn_cast(RedVal); if (!RedValI) return false; return V.isDeleted(RedValI); })) break; V.buildTree(VL, IgnoreList); if (V.isTreeTinyAndNotFullyVectorizable(/*ForReduction=*/true)) { if (!AdjustReducedVals()) V.analyzedReductionVals(VL); continue; } if (V.isLoadCombineReductionCandidate(RdxKind)) { if (!AdjustReducedVals()) V.analyzedReductionVals(VL); continue; } V.reorderTopToBottom(); // No need to reorder the root node at all. V.reorderBottomToTop(/*IgnoreReorder=*/true); // Keep extracted other reduction values, if they are used in the // vectorization trees. BoUpSLP::ExtraValueToDebugLocsMap LocalExternallyUsedValues( ExternallyUsedValues); for (unsigned Cnt = 0, Sz = ReducedVals.size(); Cnt < Sz; ++Cnt) { if (Cnt == I || (ShuffledExtracts && Cnt == I - 1)) continue; for_each(ReducedVals[Cnt], [&LocalExternallyUsedValues, &TrackedVals](Value *V) { if (isa(V)) LocalExternallyUsedValues[TrackedVals[V]]; }); } // Number of uses of the candidates in the vector of values. SmallDenseMap NumUses(Candidates.size()); for (unsigned Cnt = 0; Cnt < Pos; ++Cnt) { Value *V = Candidates[Cnt]; ++NumUses.try_emplace(V, 0).first->getSecond(); } for (unsigned Cnt = Pos + ReduxWidth; Cnt < NumReducedVals; ++Cnt) { Value *V = Candidates[Cnt]; ++NumUses.try_emplace(V, 0).first->getSecond(); } SmallPtrSet VLScalars(VL.begin(), VL.end()); // Gather externally used values. SmallPtrSet Visited; for (unsigned Cnt = 0; Cnt < Pos; ++Cnt) { Value *RdxVal = Candidates[Cnt]; if (!Visited.insert(RdxVal).second) continue; // Check if the scalar was vectorized as part of the vectorization // tree but not the top node. if (!VLScalars.contains(RdxVal) && V.isVectorized(RdxVal)) { LocalExternallyUsedValues[RdxVal]; continue; } unsigned NumOps = VectorizedVals.lookup(RdxVal) + NumUses[RdxVal]; if (NumOps != ReducedValsToOps.find(RdxVal)->second.size()) LocalExternallyUsedValues[RdxVal]; } for (unsigned Cnt = Pos + ReduxWidth; Cnt < NumReducedVals; ++Cnt) { Value *RdxVal = Candidates[Cnt]; if (!Visited.insert(RdxVal).second) continue; // Check if the scalar was vectorized as part of the vectorization // tree but not the top node. if (!VLScalars.contains(RdxVal) && V.isVectorized(RdxVal)) { LocalExternallyUsedValues[RdxVal]; continue; } unsigned NumOps = VectorizedVals.lookup(RdxVal) + NumUses[RdxVal]; if (NumOps != ReducedValsToOps.find(RdxVal)->second.size()) LocalExternallyUsedValues[RdxVal]; } for (Value *RdxVal : VL) if (RequiredExtract.contains(RdxVal)) LocalExternallyUsedValues[RdxVal]; V.buildExternalUses(LocalExternallyUsedValues); V.computeMinimumValueSizes(); // Intersect the fast-math-flags from all reduction operations. FastMathFlags RdxFMF; RdxFMF.set(); for (Value *U : IgnoreList) if (auto *FPMO = dyn_cast(U)) RdxFMF &= FPMO->getFastMathFlags(); // Estimate cost. InstructionCost TreeCost = V.getTreeCost(VL); InstructionCost ReductionCost = getReductionCost(TTI, VL, ReduxWidth, RdxFMF); if (V.isVectorizedFirstNode() && isa(VL.front())) { Instruction *MainOp = V.getFirstNodeMainOp(); for (Value *V : VL) { auto *VI = dyn_cast(V); // Add the costs of scalar GEP pointers, to be removed from the // code. if (!VI || VI == MainOp) continue; auto *Ptr = dyn_cast(VI->getPointerOperand()); if (!Ptr || !Ptr->hasOneUse() || Ptr->hasAllConstantIndices()) continue; TreeCost -= TTI->getArithmeticInstrCost( Instruction::Add, Ptr->getType(), TTI::TCK_RecipThroughput); } } InstructionCost Cost = TreeCost + ReductionCost; LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for reduction\n"); if (!Cost.isValid()) return nullptr; if (Cost >= -SLPCostThreshold) { V.getORE()->emit([&]() { return OptimizationRemarkMissed( SV_NAME, "HorSLPNotBeneficial", ReducedValsToOps.find(VL[0])->second.front()) << "Vectorizing horizontal reduction is possible " << "but not beneficial with cost " << ore::NV("Cost", Cost) << " and threshold " << ore::NV("Threshold", -SLPCostThreshold); }); if (!AdjustReducedVals()) V.analyzedReductionVals(VL); continue; } LLVM_DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"); V.getORE()->emit([&]() { return OptimizationRemark( SV_NAME, "VectorizedHorizontalReduction", ReducedValsToOps.find(VL[0])->second.front()) << "Vectorized horizontal reduction with cost " << ore::NV("Cost", Cost) << " and with tree size " << ore::NV("TreeSize", V.getTreeSize()); }); Builder.setFastMathFlags(RdxFMF); // Emit a reduction. If the root is a select (min/max idiom), the insert // point is the compare condition of that select. Instruction *RdxRootInst = cast(ReductionRoot); Instruction *InsertPt = RdxRootInst; if (IsCmpSelMinMax) InsertPt = GetCmpForMinMaxReduction(RdxRootInst); // Vectorize a tree. Value *VectorizedRoot = V.vectorizeTree(LocalExternallyUsedValues, InsertPt); Builder.SetInsertPoint(InsertPt); // To prevent poison from leaking across what used to be sequential, // safe, scalar boolean logic operations, the reduction operand must be // frozen. if (isBoolLogicOp(RdxRootInst)) VectorizedRoot = Builder.CreateFreeze(VectorizedRoot); Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); if (!VectorizedTree) { // Initialize the final value in the reduction. VectorizedTree = ReducedSubTree; } else { // Update the final value in the reduction. Builder.SetCurrentDebugLocation( cast(ReductionOps.front().front())->getDebugLoc()); VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, ReducedSubTree, "op.rdx", ReductionOps); } // Count vectorized reduced values to exclude them from final reduction. for (Value *RdxVal : VL) { ++VectorizedVals.try_emplace(TrackedToOrig.find(RdxVal)->second, 0) .first->getSecond(); if (!V.isVectorized(RdxVal)) RequiredExtract.insert(RdxVal); } Pos += ReduxWidth; Start = Pos; ReduxWidth = PowerOf2Floor(NumReducedVals - Pos); } } if (VectorizedTree) { // Reorder operands of bool logical op in the natural order to avoid // possible problem with poison propagation. If not possible to reorder // (both operands are originally RHS), emit an extra freeze instruction // for the LHS operand. //I.e., if we have original code like this: // RedOp1 = select i1 ?, i1 LHS, i1 false // RedOp2 = select i1 RHS, i1 ?, i1 false // Then, we swap LHS/RHS to create a new op that matches the poison // semantics of the original code. // If we have original code like this and both values could be poison: // RedOp1 = select i1 ?, i1 LHS, i1 false // RedOp2 = select i1 ?, i1 RHS, i1 false // Then, we must freeze LHS in the new op. auto &&FixBoolLogicalOps = [&Builder, VectorizedTree](Value *&LHS, Value *&RHS, Instruction *RedOp1, Instruction *RedOp2) { if (!isBoolLogicOp(RedOp1)) return; if (LHS == VectorizedTree || getRdxOperand(RedOp1, 0) == LHS || isGuaranteedNotToBePoison(LHS)) return; if (!isBoolLogicOp(RedOp2)) return; if (RHS == VectorizedTree || getRdxOperand(RedOp2, 0) == RHS || isGuaranteedNotToBePoison(RHS)) { std::swap(LHS, RHS); return; } LHS = Builder.CreateFreeze(LHS); }; // Finish the reduction. // Need to add extra arguments and not vectorized possible reduction // values. // Try to avoid dependencies between the scalar remainders after // reductions. auto &&FinalGen = [this, &Builder, &TrackedVals, &FixBoolLogicalOps]( ArrayRef> InstVals) { unsigned Sz = InstVals.size(); SmallVector> ExtraReds(Sz / 2 + Sz % 2); for (unsigned I = 0, E = (Sz / 2) * 2; I < E; I += 2) { Instruction *RedOp = InstVals[I + 1].first; Builder.SetCurrentDebugLocation(RedOp->getDebugLoc()); Value *RdxVal1 = InstVals[I].second; Value *StableRdxVal1 = RdxVal1; auto It1 = TrackedVals.find(RdxVal1); if (It1 != TrackedVals.end()) StableRdxVal1 = It1->second; Value *RdxVal2 = InstVals[I + 1].second; Value *StableRdxVal2 = RdxVal2; auto It2 = TrackedVals.find(RdxVal2); if (It2 != TrackedVals.end()) StableRdxVal2 = It2->second; // To prevent poison from leaking across what used to be // sequential, safe, scalar boolean logic operations, the // reduction operand must be frozen. FixBoolLogicalOps(StableRdxVal1, StableRdxVal2, InstVals[I].first, RedOp); Value *ExtraRed = createOp(Builder, RdxKind, StableRdxVal1, StableRdxVal2, "op.rdx", ReductionOps); ExtraReds[I / 2] = std::make_pair(InstVals[I].first, ExtraRed); } if (Sz % 2 == 1) ExtraReds[Sz / 2] = InstVals.back(); return ExtraReds; }; SmallVector> ExtraReductions; ExtraReductions.emplace_back(cast(ReductionRoot), VectorizedTree); SmallPtrSet Visited; for (ArrayRef Candidates : ReducedVals) { for (Value *RdxVal : Candidates) { if (!Visited.insert(RdxVal).second) continue; unsigned NumOps = VectorizedVals.lookup(RdxVal); for (Instruction *RedOp : ArrayRef(ReducedValsToOps.find(RdxVal)->second) .drop_back(NumOps)) ExtraReductions.emplace_back(RedOp, RdxVal); } } for (auto &Pair : ExternallyUsedValues) { // Add each externally used value to the final reduction. for (auto *I : Pair.second) ExtraReductions.emplace_back(I, Pair.first); } // Iterate through all not-vectorized reduction values/extra arguments. while (ExtraReductions.size() > 1) { VectorizedTree = ExtraReductions.front().second; SmallVector> NewReds = FinalGen(ExtraReductions); ExtraReductions.swap(NewReds); } VectorizedTree = ExtraReductions.front().second; ReductionRoot->replaceAllUsesWith(VectorizedTree); // The original scalar reduction is expected to have no remaining // uses outside the reduction tree itself. Assert that we got this // correct, replace internal uses with undef, and mark for eventual // deletion. #ifndef NDEBUG SmallSet IgnoreSet; for (ArrayRef RdxOps : ReductionOps) IgnoreSet.insert(RdxOps.begin(), RdxOps.end()); #endif for (ArrayRef RdxOps : ReductionOps) { for (Value *Ignore : RdxOps) { if (!Ignore) continue; #ifndef NDEBUG for (auto *U : Ignore->users()) { assert(IgnoreSet.count(U) && "All users must be either in the reduction ops list."); } #endif if (!Ignore->use_empty()) { Value *Undef = UndefValue::get(Ignore->getType()); Ignore->replaceAllUsesWith(Undef); } V.eraseInstruction(cast(Ignore)); } } } else if (!CheckForReusedReductionOps) { for (ReductionOpsType &RdxOps : ReductionOps) for (Value *RdxOp : RdxOps) V.analyzedReductionRoot(cast(RdxOp)); } return VectorizedTree; } private: /// Calculate the cost of a reduction. InstructionCost getReductionCost(TargetTransformInfo *TTI, ArrayRef ReducedVals, unsigned ReduxWidth, FastMathFlags FMF) { TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; Value *FirstReducedVal = ReducedVals.front(); Type *ScalarTy = FirstReducedVal->getType(); FixedVectorType *VectorTy = FixedVectorType::get(ScalarTy, ReduxWidth); InstructionCost VectorCost = 0, ScalarCost; // If all of the reduced values are constant, the vector cost is 0, since // the reduction value can be calculated at the compile time. bool AllConsts = all_of(ReducedVals, isConstant); switch (RdxKind) { case RecurKind::Add: case RecurKind::Mul: case RecurKind::Or: case RecurKind::And: case RecurKind::Xor: case RecurKind::FAdd: case RecurKind::FMul: { unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind); if (!AllConsts) VectorCost = TTI->getArithmeticReductionCost(RdxOpcode, VectorTy, FMF, CostKind); ScalarCost = TTI->getArithmeticInstrCost(RdxOpcode, ScalarTy, CostKind); break; } case RecurKind::FMax: case RecurKind::FMin: { auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); if (!AllConsts) { auto *VecCondTy = cast(CmpInst::makeCmpResultType(VectorTy)); VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, /*IsUnsigned=*/false, CostKind); } CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); ScalarCost = TTI->getCmpSelInstrCost(Instruction::FCmp, ScalarTy, SclCondTy, RdxPred, CostKind) + TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, SclCondTy, RdxPred, CostKind); break; } case RecurKind::SMax: case RecurKind::SMin: case RecurKind::UMax: case RecurKind::UMin: { auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); if (!AllConsts) { auto *VecCondTy = cast(CmpInst::makeCmpResultType(VectorTy)); bool IsUnsigned = RdxKind == RecurKind::UMax || RdxKind == RecurKind::UMin; VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, IsUnsigned, CostKind); } CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); ScalarCost = TTI->getCmpSelInstrCost(Instruction::ICmp, ScalarTy, SclCondTy, RdxPred, CostKind) + TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, SclCondTy, RdxPred, CostKind); break; } default: llvm_unreachable("Expected arithmetic or min/max reduction operation"); } // Scalar cost is repeated for N-1 elements. ScalarCost *= (ReduxWidth - 1); LLVM_DEBUG(dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n"); return VectorCost - ScalarCost; } /// Emit a horizontal reduction of the vectorized value. Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, unsigned ReduxWidth, const TargetTransformInfo *TTI) { assert(VectorizedValue && "Need to have a vectorized tree node"); assert(isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now"); assert(RdxKind != RecurKind::FMulAdd && "A call to the llvm.fmuladd intrinsic is not handled yet"); ++NumVectorInstructions; return createSimpleTargetReduction(Builder, TTI, VectorizedValue, RdxKind); } }; } // end anonymous namespace static std::optional getAggregateSize(Instruction *InsertInst) { if (auto *IE = dyn_cast(InsertInst)) return cast(IE->getType())->getNumElements(); unsigned AggregateSize = 1; auto *IV = cast(InsertInst); Type *CurrentType = IV->getType(); do { if (auto *ST = dyn_cast(CurrentType)) { for (auto *Elt : ST->elements()) if (Elt != ST->getElementType(0)) // check homogeneity return std::nullopt; AggregateSize *= ST->getNumElements(); CurrentType = ST->getElementType(0); } else if (auto *AT = dyn_cast(CurrentType)) { AggregateSize *= AT->getNumElements(); CurrentType = AT->getElementType(); } else if (auto *VT = dyn_cast(CurrentType)) { AggregateSize *= VT->getNumElements(); return AggregateSize; } else if (CurrentType->isSingleValueType()) { return AggregateSize; } else { return std::nullopt; } } while (true); } static void findBuildAggregate_rec(Instruction *LastInsertInst, TargetTransformInfo *TTI, SmallVectorImpl &BuildVectorOpds, SmallVectorImpl &InsertElts, unsigned OperandOffset) { do { Value *InsertedOperand = LastInsertInst->getOperand(1); std::optional OperandIndex = getInsertIndex(LastInsertInst, OperandOffset); if (!OperandIndex) return; if (isa(InsertedOperand)) { findBuildAggregate_rec(cast(InsertedOperand), TTI, BuildVectorOpds, InsertElts, *OperandIndex); } else { BuildVectorOpds[*OperandIndex] = InsertedOperand; InsertElts[*OperandIndex] = LastInsertInst; } LastInsertInst = dyn_cast(LastInsertInst->getOperand(0)); } while (LastInsertInst != nullptr && isa(LastInsertInst) && LastInsertInst->hasOneUse()); } /// Recognize construction of vectors like /// %ra = insertelement <4 x float> poison, float %s0, i32 0 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 /// starting from the last insertelement or insertvalue instruction. /// /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. /// /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. /// /// \return true if it matches. static bool findBuildAggregate(Instruction *LastInsertInst, TargetTransformInfo *TTI, SmallVectorImpl &BuildVectorOpds, SmallVectorImpl &InsertElts) { assert((isa(LastInsertInst) || isa(LastInsertInst)) && "Expected insertelement or insertvalue instruction!"); assert((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!"); std::optional AggregateSize = getAggregateSize(LastInsertInst); if (!AggregateSize) return false; BuildVectorOpds.resize(*AggregateSize); InsertElts.resize(*AggregateSize); findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, 0); llvm::erase_value(BuildVectorOpds, nullptr); llvm::erase_value(InsertElts, nullptr); if (BuildVectorOpds.size() >= 2) return true; return false; } /// 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 isa(R) && DT->dominates(P->getParent(), 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 occasionally 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; } static bool matchRdxBop(Instruction *I, Value *&V0, Value *&V1) { if (match(I, m_BinOp(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; if (match(I, m_Intrinsic(m_Value(V0), m_Value(V1)))) return true; return false; } bool SLPVectorizerPass::vectorizeHorReduction( PHINode *P, Value *V, BasicBlock *BB, BoUpSLP &R, TargetTransformInfo *TTI, SmallVectorImpl &PostponedInsts) { if (!ShouldVectorizeHor) return false; auto *Root = dyn_cast_or_null(V); if (!Root) return false; if (!isa(Root)) P = nullptr; if (Root->getParent() != BB || isa(Root)) return false; // Start analysis starting from Root instruction. If horizontal reduction is // found, try to vectorize it. If it is not a horizontal reduction or // vectorization is not possible or not effective, and currently analyzed // instruction is a binary operation, try to vectorize the operands, using // pre-order DFS traversal order. If the operands were not vectorized, repeat // the same procedure considering each operand as a possible root of the // horizontal reduction. // Interrupt the process if the Root instruction itself was vectorized or all // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. // If a horizintal reduction was not matched or vectorized we collect // instructions for possible later attempts for vectorization. std::queue> Stack; Stack.emplace(Root, 0); SmallPtrSet VisitedInstrs; bool Res = false; auto &&TryToReduce = [this, TTI, &P, &R](Instruction *Inst, Value *&B0, Value *&B1) -> Value * { if (R.isAnalyzedReductionRoot(Inst)) return nullptr; bool IsBinop = matchRdxBop(Inst, B0, B1); bool IsSelect = match(Inst, m_Select(m_Value(), m_Value(), m_Value())); if (IsBinop || IsSelect) { HorizontalReduction HorRdx; if (HorRdx.matchAssociativeReduction(P, Inst, *SE, *DL, *TLI)) return HorRdx.tryToReduce(R, TTI, *TLI); } return nullptr; }; while (!Stack.empty()) { Instruction *Inst; unsigned Level; std::tie(Inst, Level) = Stack.front(); Stack.pop(); // Do not try to analyze instruction that has already been vectorized. // This may happen when we vectorize instruction operands on a previous // iteration while stack was populated before that happened. if (R.isDeleted(Inst)) continue; Value *B0 = nullptr, *B1 = nullptr; if (Value *V = TryToReduce(Inst, B0, B1)) { Res = true; // Set P to nullptr to avoid re-analysis of phi node in // matchAssociativeReduction function unless this is the root node. P = nullptr; if (auto *I = dyn_cast(V)) { // Try to find another reduction. Stack.emplace(I, Level); continue; } } else { bool IsBinop = B0 && B1; if (P && IsBinop) { Inst = dyn_cast(B0); if (Inst == P) Inst = dyn_cast(B1); if (!Inst) { // Set P to nullptr to avoid re-analysis of phi node in // matchAssociativeReduction function unless this is the root node. P = nullptr; continue; } } // Set P to nullptr to avoid re-analysis of phi node in // matchAssociativeReduction function unless this is the root node. P = nullptr; // Do not collect CmpInst or InsertElementInst/InsertValueInst as their // analysis is done separately. if (!isa(Inst)) PostponedInsts.push_back(Inst); } // Try to vectorize operands. // Continue analysis for the instruction from the same basic block only to // save compile time. if (++Level < RecursionMaxDepth) for (auto *Op : Inst->operand_values()) if (VisitedInstrs.insert(Op).second) if (auto *I = dyn_cast(Op)) // Do not try to vectorize CmpInst operands, this is done // separately. if (!isa(I) && !R.isDeleted(I) && I->getParent() == BB) Stack.emplace(I, Level); } return Res; } bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, BasicBlock *BB, BoUpSLP &R, TargetTransformInfo *TTI) { SmallVector PostponedInsts; bool Res = vectorizeHorReduction(P, V, BB, R, TTI, PostponedInsts); Res |= tryToVectorize(PostponedInsts, R); return Res; } bool SLPVectorizerPass::tryToVectorize(ArrayRef Insts, BoUpSLP &R) { bool Res = false; for (Value *V : Insts) if (auto *Inst = dyn_cast(V); Inst && !R.isDeleted(Inst)) Res |= tryToVectorize(Inst, R); return Res; } bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, BasicBlock *BB, BoUpSLP &R) { const DataLayout &DL = BB->getModule()->getDataLayout(); if (!R.canMapToVector(IVI->getType(), DL)) return false; SmallVector BuildVectorOpds; SmallVector BuildVectorInsts; if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) return false; LLVM_DEBUG(dbgs() << "SLP: array mappable to vector: " << *IVI << "\n"); // Aggregate value is unlikely to be processed in vector register. return tryToVectorizeList(BuildVectorOpds, R); } bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, BasicBlock *BB, BoUpSLP &R) { SmallVector BuildVectorInsts; SmallVector BuildVectorOpds; SmallVector Mask; if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || (llvm::all_of( BuildVectorOpds, [](Value *V) { return isa(V); }) && isFixedVectorShuffle(BuildVectorOpds, Mask))) return false; LLVM_DEBUG(dbgs() << "SLP: array mappable to vector: " << *IEI << "\n"); return tryToVectorizeList(BuildVectorInsts, R); } template static bool tryToVectorizeSequence(SmallVectorImpl &Incoming, function_ref Limit, function_ref Comparator, function_ref AreCompatible, function_ref, bool)> TryToVectorizeHelper, bool LimitForRegisterSize) { bool Changed = false; // Sort by type, parent, operands. stable_sort(Incoming, Comparator); // Try to vectorize elements base on their type. SmallVector Candidates; for (auto *IncIt = Incoming.begin(), *E = Incoming.end(); IncIt != E;) { // Look for the next elements with the same type, parent and operand // kinds. auto *SameTypeIt = IncIt; while (SameTypeIt != E && AreCompatible(*SameTypeIt, *IncIt)) ++SameTypeIt; // Try to vectorize them. unsigned NumElts = (SameTypeIt - IncIt); LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize starting at nodes (" << NumElts << ")\n"); // The vectorization is a 3-state attempt: // 1. Try to vectorize instructions with the same/alternate opcodes with the // size of maximal register at first. // 2. Try to vectorize remaining instructions with the same type, if // possible. This may result in the better vectorization results rather than // if we try just to vectorize instructions with the same/alternate opcodes. // 3. Final attempt to try to vectorize all instructions with the // same/alternate ops only, this may result in some extra final // vectorization. if (NumElts > 1 && TryToVectorizeHelper(ArrayRef(IncIt, NumElts), LimitForRegisterSize)) { // Success start over because instructions might have been changed. Changed = true; } else if (NumElts < Limit(*IncIt) && (Candidates.empty() || Candidates.front()->getType() == (*IncIt)->getType())) { Candidates.append(IncIt, std::next(IncIt, NumElts)); } // Final attempt to vectorize instructions with the same types. if (Candidates.size() > 1 && (SameTypeIt == E || (*SameTypeIt)->getType() != (*IncIt)->getType())) { if (TryToVectorizeHelper(Candidates, /*LimitForRegisterSize=*/false)) { // Success start over because instructions might have been changed. Changed = true; } else if (LimitForRegisterSize) { // Try to vectorize using small vectors. for (auto *It = Candidates.begin(), *End = Candidates.end(); It != End;) { auto *SameTypeIt = It; while (SameTypeIt != End && AreCompatible(*SameTypeIt, *It)) ++SameTypeIt; unsigned NumElts = (SameTypeIt - It); if (NumElts > 1 && TryToVectorizeHelper(ArrayRef(It, NumElts), /*LimitForRegisterSize=*/false)) Changed = true; It = SameTypeIt; } } Candidates.clear(); } // Start over at the next instruction of a different type (or the end). IncIt = SameTypeIt; } return Changed; } /// Compare two cmp instructions. If IsCompatibility is true, function returns /// true if 2 cmps have same/swapped predicates and mos compatible corresponding /// operands. If IsCompatibility is false, function implements strict weak /// ordering relation between two cmp instructions, returning true if the first /// instruction is "less" than the second, i.e. its predicate is less than the /// predicate of the second or the operands IDs are less than the operands IDs /// of the second cmp instruction. template static bool compareCmp(Value *V, Value *V2, TargetLibraryInfo &TLI, function_ref IsDeleted) { auto *CI1 = cast(V); auto *CI2 = cast(V2); if (IsDeleted(CI2) || !isValidElementType(CI2->getType())) return false; if (CI1->getOperand(0)->getType()->getTypeID() < CI2->getOperand(0)->getType()->getTypeID()) return !IsCompatibility; if (CI1->getOperand(0)->getType()->getTypeID() > CI2->getOperand(0)->getType()->getTypeID()) return false; CmpInst::Predicate Pred1 = CI1->getPredicate(); CmpInst::Predicate Pred2 = CI2->getPredicate(); CmpInst::Predicate SwapPred1 = CmpInst::getSwappedPredicate(Pred1); CmpInst::Predicate SwapPred2 = CmpInst::getSwappedPredicate(Pred2); CmpInst::Predicate BasePred1 = std::min(Pred1, SwapPred1); CmpInst::Predicate BasePred2 = std::min(Pred2, SwapPred2); if (BasePred1 < BasePred2) return !IsCompatibility; if (BasePred1 > BasePred2) return false; // Compare operands. bool LEPreds = Pred1 <= Pred2; bool GEPreds = Pred1 >= Pred2; for (int I = 0, E = CI1->getNumOperands(); I < E; ++I) { auto *Op1 = CI1->getOperand(LEPreds ? I : E - I - 1); auto *Op2 = CI2->getOperand(GEPreds ? I : E - I - 1); if (Op1->getValueID() < Op2->getValueID()) return !IsCompatibility; if (Op1->getValueID() > Op2->getValueID()) return false; if (auto *I1 = dyn_cast(Op1)) if (auto *I2 = dyn_cast(Op2)) { if (I1->getParent() != I2->getParent()) return false; InstructionsState S = getSameOpcode({I1, I2}, TLI); if (S.getOpcode()) continue; return false; } } return IsCompatibility; } bool SLPVectorizerPass::vectorizeSimpleInstructions(InstSetVector &Instructions, BasicBlock *BB, BoUpSLP &R, bool AtTerminator) { bool OpsChanged = false; SmallVector PostponedCmps; SmallVector PostponedInsts; // pass1 - try to vectorize reductions only for (auto *I : reverse(Instructions)) { if (R.isDeleted(I)) continue; if (isa(I)) { PostponedCmps.push_back(I); continue; } OpsChanged |= vectorizeHorReduction(nullptr, I, BB, R, TTI, PostponedInsts); } // pass2 - try to match and vectorize a buildvector sequence. for (auto *I : reverse(Instructions)) { if (R.isDeleted(I) || isa(I)) continue; if (auto *LastInsertValue = dyn_cast(I)) { OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); } else if (auto *LastInsertElem = dyn_cast(I)) { OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); } } // Now try to vectorize postponed instructions. OpsChanged |= tryToVectorize(PostponedInsts, R); if (AtTerminator) { // Try to find reductions first. for (Instruction *I : PostponedCmps) { if (R.isDeleted(I)) continue; for (Value *Op : I->operands()) OpsChanged |= vectorizeRootInstruction(nullptr, Op, BB, R, TTI); } // Try to vectorize operands as vector bundles. for (Instruction *I : PostponedCmps) { if (R.isDeleted(I)) continue; OpsChanged |= tryToVectorize(I, R); } // Try to vectorize list of compares. // Sort by type, compare predicate, etc. auto CompareSorter = [&](Value *V, Value *V2) { return compareCmp(V, V2, *TLI, [&R](Instruction *I) { return R.isDeleted(I); }); }; auto AreCompatibleCompares = [&](Value *V1, Value *V2) { if (V1 == V2) return true; return compareCmp(V1, V2, *TLI, [&R](Instruction *I) { return R.isDeleted(I); }); }; auto Limit = [&R](Value *V) { unsigned EltSize = R.getVectorElementSize(V); return std::max(2U, R.getMaxVecRegSize() / EltSize); }; SmallVector Vals(PostponedCmps.begin(), PostponedCmps.end()); OpsChanged |= tryToVectorizeSequence( Vals, Limit, CompareSorter, AreCompatibleCompares, [this, &R](ArrayRef Candidates, bool LimitForRegisterSize) { // Exclude possible reductions from other blocks. bool ArePossiblyReducedInOtherBlock = any_of(Candidates, [](Value *V) { return any_of(V->users(), [V](User *U) { return isa(U) && cast(U)->getParent() != cast(V)->getParent(); }); }); if (ArePossiblyReducedInOtherBlock) return false; return tryToVectorizeList(Candidates, R, LimitForRegisterSize); }, /*LimitForRegisterSize=*/true); Instructions.clear(); } else { Instructions.clear(); // Insert in reverse order since the PostponedCmps vector was filled in // reverse order. Instructions.insert(PostponedCmps.rbegin(), PostponedCmps.rend()); } return OpsChanged; } bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { bool Changed = false; SmallVector Incoming; SmallPtrSet VisitedInstrs; // Maps phi nodes to the non-phi nodes found in the use tree for each phi // node. Allows better to identify the chains that can be vectorized in the // better way. DenseMap> PHIToOpcodes; auto PHICompare = [this, &PHIToOpcodes](Value *V1, Value *V2) { assert(isValidElementType(V1->getType()) && isValidElementType(V2->getType()) && "Expected vectorizable types only."); // It is fine to compare type IDs here, since we expect only vectorizable // types, like ints, floats and pointers, we don't care about other type. if (V1->getType()->getTypeID() < V2->getType()->getTypeID()) return true; if (V1->getType()->getTypeID() > V2->getType()->getTypeID()) return false; ArrayRef Opcodes1 = PHIToOpcodes[V1]; ArrayRef Opcodes2 = PHIToOpcodes[V2]; if (Opcodes1.size() < Opcodes2.size()) return true; if (Opcodes1.size() > Opcodes2.size()) return false; std::optional ConstOrder; for (int I = 0, E = Opcodes1.size(); I < E; ++I) { // Undefs are compatible with any other value. if (isa(Opcodes1[I]) || isa(Opcodes2[I])) { if (!ConstOrder) ConstOrder = !isa(Opcodes1[I]) && isa(Opcodes2[I]); continue; } if (auto *I1 = dyn_cast(Opcodes1[I])) if (auto *I2 = dyn_cast(Opcodes2[I])) { DomTreeNodeBase *NodeI1 = DT->getNode(I1->getParent()); DomTreeNodeBase *NodeI2 = DT->getNode(I2->getParent()); if (!NodeI1) return NodeI2 != nullptr; if (!NodeI2) return false; assert((NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers"); if (NodeI1 != NodeI2) return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); InstructionsState S = getSameOpcode({I1, I2}, *TLI); if (S.getOpcode()) continue; return I1->getOpcode() < I2->getOpcode(); } if (isa(Opcodes1[I]) && isa(Opcodes2[I])) { if (!ConstOrder) ConstOrder = Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID(); continue; } if (Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID()) return true; if (Opcodes1[I]->getValueID() > Opcodes2[I]->getValueID()) return false; } return ConstOrder && *ConstOrder; }; auto AreCompatiblePHIs = [&PHIToOpcodes, this](Value *V1, Value *V2) { if (V1 == V2) return true; if (V1->getType() != V2->getType()) return false; ArrayRef Opcodes1 = PHIToOpcodes[V1]; ArrayRef Opcodes2 = PHIToOpcodes[V2]; if (Opcodes1.size() != Opcodes2.size()) return false; for (int I = 0, E = Opcodes1.size(); I < E; ++I) { // Undefs are compatible with any other value. if (isa(Opcodes1[I]) || isa(Opcodes2[I])) continue; if (auto *I1 = dyn_cast(Opcodes1[I])) if (auto *I2 = dyn_cast(Opcodes2[I])) { if (I1->getParent() != I2->getParent()) return false; InstructionsState S = getSameOpcode({I1, I2}, *TLI); if (S.getOpcode()) continue; return false; } if (isa(Opcodes1[I]) && isa(Opcodes2[I])) continue; if (Opcodes1[I]->getValueID() != Opcodes2[I]->getValueID()) return false; } return true; }; auto Limit = [&R](Value *V) { unsigned EltSize = R.getVectorElementSize(V); return std::max(2U, R.getMaxVecRegSize() / EltSize); }; bool HaveVectorizedPhiNodes = false; do { // Collect the incoming values from the PHIs. Incoming.clear(); for (Instruction &I : *BB) { PHINode *P = dyn_cast(&I); if (!P) break; // No need to analyze deleted, vectorized and non-vectorizable // instructions. if (!VisitedInstrs.count(P) && !R.isDeleted(P) && isValidElementType(P->getType())) Incoming.push_back(P); } // Find the corresponding non-phi nodes for better matching when trying to // build the tree. for (Value *V : Incoming) { SmallVectorImpl &Opcodes = PHIToOpcodes.try_emplace(V).first->getSecond(); if (!Opcodes.empty()) continue; SmallVector Nodes(1, V); SmallPtrSet Visited; while (!Nodes.empty()) { auto *PHI = cast(Nodes.pop_back_val()); if (!Visited.insert(PHI).second) continue; for (Value *V : PHI->incoming_values()) { if (auto *PHI1 = dyn_cast((V))) { Nodes.push_back(PHI1); continue; } Opcodes.emplace_back(V); } } } HaveVectorizedPhiNodes = tryToVectorizeSequence( Incoming, Limit, PHICompare, AreCompatiblePHIs, [this, &R](ArrayRef Candidates, bool LimitForRegisterSize) { return tryToVectorizeList(Candidates, R, LimitForRegisterSize); }, /*LimitForRegisterSize=*/true); Changed |= HaveVectorizedPhiNodes; VisitedInstrs.insert(Incoming.begin(), Incoming.end()); } while (HaveVectorizedPhiNodes); VisitedInstrs.clear(); InstSetVector PostProcessInstructions; SmallDenseSet KeyNodes; for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { // Skip instructions with scalable type. The num of elements is unknown at // compile-time for scalable type. if (isa(it->getType())) continue; // Skip instructions marked for the deletion. if (R.isDeleted(&*it)) continue; // We may go through BB multiple times so skip the one we have checked. if (!VisitedInstrs.insert(&*it).second) { if (it->use_empty() && KeyNodes.contains(&*it) && vectorizeSimpleInstructions(PostProcessInstructions, BB, R, it->isTerminator())) { // We would like to start over since some instructions are deleted // and the iterator may become invalid value. Changed = true; it = BB->begin(); e = BB->end(); } continue; } if (isa(it)) continue; // Try to vectorize reductions that use PHINodes. if (PHINode *P = dyn_cast(it)) { // Check that the PHI is a reduction PHI. if (P->getNumIncomingValues() == 2) { // Try to match and vectorize a horizontal reduction. if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, TTI)) { Changed = true; it = BB->begin(); e = BB->end(); continue; } } // Try to vectorize the incoming values of the PHI, to catch reductions // that feed into PHIs. for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) { // Skip if the incoming block is the current BB for now. Also, bypass // unreachable IR for efficiency and to avoid crashing. // TODO: Collect the skipped incoming values and try to vectorize them // after processing BB. if (BB == P->getIncomingBlock(I) || !DT->isReachableFromEntry(P->getIncomingBlock(I))) continue; // Postponed instructions should not be vectorized here, delay their // vectorization. if (auto *PI = dyn_cast(P->getIncomingValue(I)); PI && !PostProcessInstructions.contains(PI)) Changed |= vectorizeRootInstruction(nullptr, P->getIncomingValue(I), P->getIncomingBlock(I), R, TTI); } continue; } // Ran into an instruction without users, like terminator, or function call // with ignored return value, store. Ignore unused instructions (basing on // instruction type, except for CallInst and InvokeInst). if (it->use_empty() && (it->getType()->isVoidTy() || isa(it))) { KeyNodes.insert(&*it); bool OpsChanged = false; auto *SI = dyn_cast(it); bool TryToVectorizeRoot = ShouldStartVectorizeHorAtStore || !SI; if (SI) { auto I = Stores.find(getUnderlyingObject(SI->getPointerOperand())); // Try to vectorize chain in store, if this is the only store to the // address in the block. // TODO: This is just a temporarily solution to save compile time. Need // to investigate if we can safely turn on slp-vectorize-hor-store // instead to allow lookup for reduction chains in all non-vectorized // stores (need to check side effects and compile time). TryToVectorizeRoot = (I == Stores.end() || I->second.size() == 1) && SI->getValueOperand()->hasOneUse(); } if (TryToVectorizeRoot) { for (auto *V : it->operand_values()) { // Postponed instructions should not be vectorized here, delay their // vectorization. if (auto *VI = dyn_cast(V); VI && !PostProcessInstructions.contains(VI)) // Try to match and vectorize a horizontal reduction. OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); } } // Start vectorization of post-process list of instructions from the // top-tree instructions to try to vectorize as many instructions as // possible. OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R, it->isTerminator()); if (OpsChanged) { // We would like to start over since some instructions are deleted // and the iterator may become invalid value. Changed = true; it = BB->begin(); e = BB->end(); continue; } } if (isa(it)) PostProcessInstructions.insert(&*it); } return Changed; } bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { auto Changed = false; for (auto &Entry : GEPs) { // If the getelementptr list has fewer than two elements, there's nothing // to do. if (Entry.second.size() < 2) continue; LLVM_DEBUG(dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"); // Process the GEP list in chunks suitable for the target's supported // vector size. If a vector register can't hold 1 element, we are done. We // are trying to vectorize the index computations, so the maximum number of // elements is based on the size of the index expression, rather than the // size of the GEP itself (the target's pointer size). unsigned MaxVecRegSize = R.getMaxVecRegSize(); unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); if (MaxVecRegSize < EltSize) continue; unsigned MaxElts = MaxVecRegSize / EltSize; for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { auto Len = std::min(BE - BI, MaxElts); ArrayRef GEPList(&Entry.second[BI], Len); // Initialize a set a candidate getelementptrs. Note that we use a // SetVector here to preserve program order. If the index computations // are vectorizable and begin with loads, we want to minimize the chance // of having to reorder them later. SetVector Candidates(GEPList.begin(), GEPList.end()); // Some of the candidates may have already been vectorized after we // initially collected them. If so, they are marked as deleted, so remove // them from the set of candidates. Candidates.remove_if( [&R](Value *I) { return R.isDeleted(cast(I)); }); // Remove from the set of candidates all pairs of getelementptrs with // constant differences. Such getelementptrs are likely not good // candidates for vectorization in a bottom-up phase since one can be // computed from the other. We also ensure all candidate getelementptr // indices are unique. for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { auto *GEPI = GEPList[I]; if (!Candidates.count(GEPI)) continue; auto *SCEVI = SE->getSCEV(GEPList[I]); for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { auto *GEPJ = GEPList[J]; auto *SCEVJ = SE->getSCEV(GEPList[J]); if (isa(SE->getMinusSCEV(SCEVI, SCEVJ))) { Candidates.remove(GEPI); Candidates.remove(GEPJ); } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { Candidates.remove(GEPJ); } } } // We break out of the above computation as soon as we know there are // fewer than two candidates remaining. if (Candidates.size() < 2) continue; // Add the single, non-constant index of each candidate to the bundle. We // ensured the indices met these constraints when we originally collected // the getelementptrs. SmallVector Bundle(Candidates.size()); auto BundleIndex = 0u; for (auto *V : Candidates) { auto *GEP = cast(V); auto *GEPIdx = GEP->idx_begin()->get(); assert(GEP->getNumIndices() == 1 || !isa(GEPIdx)); Bundle[BundleIndex++] = GEPIdx; } // Try and vectorize the indices. We are currently only interested in // gather-like cases of the form: // // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... // // where the loads of "a", the loads of "b", and the subtractions can be // performed in parallel. It's likely that detecting this pattern in a // bottom-up phase will be simpler and less costly than building a // full-blown top-down phase beginning at the consecutive loads. Changed |= tryToVectorizeList(Bundle, R); } } return Changed; } bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { bool Changed = false; // Sort by type, base pointers and values operand. Value operands must be // compatible (have the same opcode, same parent), otherwise it is // definitely not profitable to try to vectorize them. auto &&StoreSorter = [this](StoreInst *V, StoreInst *V2) { if (V->getPointerOperandType()->getTypeID() < V2->getPointerOperandType()->getTypeID()) return true; if (V->getPointerOperandType()->getTypeID() > V2->getPointerOperandType()->getTypeID()) return false; // UndefValues are compatible with all other values. if (isa(V->getValueOperand()) || isa(V2->getValueOperand())) return false; if (auto *I1 = dyn_cast(V->getValueOperand())) if (auto *I2 = dyn_cast(V2->getValueOperand())) { DomTreeNodeBase *NodeI1 = DT->getNode(I1->getParent()); DomTreeNodeBase *NodeI2 = DT->getNode(I2->getParent()); assert(NodeI1 && "Should only process reachable instructions"); assert(NodeI2 && "Should only process reachable instructions"); assert((NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers"); if (NodeI1 != NodeI2) return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); InstructionsState S = getSameOpcode({I1, I2}, *TLI); if (S.getOpcode()) return false; return I1->getOpcode() < I2->getOpcode(); } if (isa(V->getValueOperand()) && isa(V2->getValueOperand())) return false; return V->getValueOperand()->getValueID() < V2->getValueOperand()->getValueID(); }; auto &&AreCompatibleStores = [this](StoreInst *V1, StoreInst *V2) { if (V1 == V2) return true; if (V1->getPointerOperandType() != V2->getPointerOperandType()) return false; // Undefs are compatible with any other value. if (isa(V1->getValueOperand()) || isa(V2->getValueOperand())) return true; if (auto *I1 = dyn_cast(V1->getValueOperand())) if (auto *I2 = dyn_cast(V2->getValueOperand())) { if (I1->getParent() != I2->getParent()) return false; InstructionsState S = getSameOpcode({I1, I2}, *TLI); return S.getOpcode() > 0; } if (isa(V1->getValueOperand()) && isa(V2->getValueOperand())) return true; return V1->getValueOperand()->getValueID() == V2->getValueOperand()->getValueID(); }; auto Limit = [&R, this](StoreInst *SI) { unsigned EltSize = DL->getTypeSizeInBits(SI->getValueOperand()->getType()); return R.getMinVF(EltSize); }; // Attempt to sort and vectorize each of the store-groups. for (auto &Pair : Stores) { if (Pair.second.size() < 2) continue; LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << Pair.second.size() << ".\n"); if (!isValidElementType(Pair.second.front()->getValueOperand()->getType())) continue; Changed |= tryToVectorizeSequence( Pair.second, Limit, StoreSorter, AreCompatibleStores, [this, &R](ArrayRef Candidates, bool) { return vectorizeStores(Candidates, R); }, /*LimitForRegisterSize=*/false); } return Changed; } char SLPVectorizer::ID = 0; static const char lv_name[] = "SLP Vectorizer"; INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false) INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass) INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy) INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false) Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); }