diff options
Diffstat (limited to 'contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp')
| -rw-r--r-- | contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 7418 |
1 files changed, 3509 insertions, 3909 deletions
diff --git a/contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp index 46ff0994e04e..dd596c567cd4 100644 --- a/contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp +++ b/contrib/llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp @@ -27,7 +27,7 @@ // // There is a development effort going on to migrate loop vectorizer to the // VPlan infrastructure and to introduce outer loop vectorization support (see -// docs/Proposal/VectorizationPlan.rst and +// docs/VectorizationPlan.rst and // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this // purpose, we temporarily introduced the VPlan-native vectorization path: an // alternative vectorization path that is natively implemented on top of the @@ -57,8 +57,8 @@ #include "LoopVectorizationPlanner.h" #include "VPRecipeBuilder.h" #include "VPlan.h" +#include "VPlanAnalysis.h" #include "VPlanHCFGBuilder.h" -#include "VPlanPredicator.h" #include "VPlanTransforms.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/ArrayRef.h" @@ -66,8 +66,6 @@ #include "llvm/ADT/DenseMapInfo.h" #include "llvm/ADT/Hashing.h" #include "llvm/ADT/MapVector.h" -#include "llvm/ADT/None.h" -#include "llvm/ADT/Optional.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallSet.h" @@ -93,6 +91,7 @@ #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" @@ -100,6 +99,7 @@ #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DataLayout.h" +#include "llvm/IR/DebugInfo.h" #include "llvm/IR/DebugInfoMetadata.h" #include "llvm/IR/DebugLoc.h" #include "llvm/IR/DerivedTypes.h" @@ -112,19 +112,18 @@ #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Intrinsics.h" -#include "llvm/IR/LLVMContext.h" +#include "llvm/IR/MDBuilder.h" #include "llvm/IR/Metadata.h" #include "llvm/IR/Module.h" #include "llvm/IR/Operator.h" #include "llvm/IR/PatternMatch.h" +#include "llvm/IR/ProfDataUtils.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" #include "llvm/IR/Verifier.h" -#include "llvm/InitializePasses.h" -#include "llvm/Pass.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" @@ -143,11 +142,12 @@ #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" #include <algorithm> #include <cassert> +#include <cmath> #include <cstdint> -#include <cstdlib> #include <functional> #include <iterator> #include <limits> +#include <map> #include <memory> #include <string> #include <tuple> @@ -198,10 +198,9 @@ static cl::opt<unsigned> TinyTripCountVectorThreshold( "value are vectorized only if no scalar iteration overheads " "are incurred.")); -static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold( - "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden, - cl::desc("The maximum allowed number of runtime memory checks with a " - "vectorize(enable) pragma.")); +static cl::opt<unsigned> VectorizeMemoryCheckThreshold( + "vectorize-memory-check-threshold", cl::init(128), cl::Hidden, + cl::desc("The maximum allowed number of runtime memory checks")); // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, // that predication is preferred, and this lists all options. I.e., the @@ -234,6 +233,25 @@ static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( "prefers tail-folding, don't attempt vectorization if " "tail-folding fails."))); +static cl::opt<TailFoldingStyle> ForceTailFoldingStyle( + "force-tail-folding-style", cl::desc("Force the tail folding style"), + cl::init(TailFoldingStyle::None), + cl::values( + clEnumValN(TailFoldingStyle::None, "none", "Disable tail folding"), + clEnumValN( + TailFoldingStyle::Data, "data", + "Create lane mask for data only, using active.lane.mask intrinsic"), + clEnumValN(TailFoldingStyle::DataWithoutLaneMask, + "data-without-lane-mask", + "Create lane mask with compare/stepvector"), + clEnumValN(TailFoldingStyle::DataAndControlFlow, "data-and-control", + "Create lane mask using active.lane.mask intrinsic, and use " + "it for both data and control flow"), + clEnumValN( + TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck, + "data-and-control-without-rt-check", + "Similar to data-and-control, but remove the runtime check"))); + static cl::opt<bool> MaximizeBandwidth( "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, cl::desc("Maximize bandwidth when selecting vectorization factor which " @@ -341,17 +359,12 @@ static cl::opt<bool> PreferPredicatedReductionSelect( cl::desc( "Prefer predicating a reduction operation over an after loop select.")); +namespace llvm { cl::opt<bool> EnableVPlanNativePath( - "enable-vplan-native-path", cl::init(false), cl::Hidden, + "enable-vplan-native-path", cl::Hidden, cl::desc("Enable VPlan-native vectorization path with " "support for outer loop vectorization.")); - -// FIXME: Remove this switch once we have divergence analysis. Currently we -// assume divergent non-backedge branches when this switch is true. -cl::opt<bool> EnableVPlanPredication( - "enable-vplan-predication", cl::init(false), cl::Hidden, - cl::desc("Enable VPlan-native vectorization path predicator with " - "support for outer loop vectorization.")); +} // This flag enables the stress testing of the VPlan H-CFG construction in the // VPlan-native vectorization path. It must be used in conjuction with @@ -371,10 +384,30 @@ cl::opt<bool> llvm::EnableLoopVectorization( "vectorize-loops", cl::init(true), cl::Hidden, cl::desc("Run the Loop vectorization passes")); -cl::opt<bool> PrintVPlansInDotFormat( - "vplan-print-in-dot-format", cl::init(false), cl::Hidden, +static cl::opt<bool> PrintVPlansInDotFormat( + "vplan-print-in-dot-format", cl::Hidden, cl::desc("Use dot format instead of plain text when dumping VPlans")); +static cl::opt<cl::boolOrDefault> ForceSafeDivisor( + "force-widen-divrem-via-safe-divisor", cl::Hidden, + cl::desc( + "Override cost based safe divisor widening for div/rem instructions")); + +static cl::opt<bool> UseWiderVFIfCallVariantsPresent( + "vectorizer-maximize-bandwidth-for-vector-calls", cl::init(true), + cl::Hidden, + cl::desc("Try wider VFs if they enable the use of vector variants")); + +// Likelyhood of bypassing the vectorized loop because assumptions about SCEV +// variables not overflowing do not hold. See `emitSCEVChecks`. +static constexpr uint32_t SCEVCheckBypassWeights[] = {1, 127}; +// Likelyhood of bypassing the vectorized loop because pointers overlap. See +// `emitMemRuntimeChecks`. +static constexpr uint32_t MemCheckBypassWeights[] = {1, 127}; +// Likelyhood of bypassing the vectorized loop because there are zero trips left +// after prolog. See `emitIterationCountCheck`. +static constexpr uint32_t MinItersBypassWeights[] = {1, 127}; + /// A helper function that returns true if the given type is irregular. The /// type is irregular if its allocated size doesn't equal the store size of an /// element of the corresponding vector type. @@ -393,20 +426,14 @@ static bool hasIrregularType(Type *Ty, const DataLayout &DL) { /// we always assume predicated blocks have a 50% chance of executing. static unsigned getReciprocalPredBlockProb() { return 2; } -/// A helper function that returns an integer or floating-point constant with -/// value C. -static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { - return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) - : ConstantFP::get(Ty, C); -} - /// Returns "best known" trip count for the specified loop \p L as defined by /// the following procedure: /// 1) Returns exact trip count if it is known. /// 2) Returns expected trip count according to profile data if any. /// 3) Returns upper bound estimate if it is known. -/// 4) Returns None if all of the above failed. -static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { +/// 4) Returns std::nullopt if all of the above failed. +static std::optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, + Loop *L) { // Check if exact trip count is known. if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) return ExpectedTC; @@ -414,18 +441,53 @@ static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { // Check if there is an expected trip count available from profile data. if (LoopVectorizeWithBlockFrequency) if (auto EstimatedTC = getLoopEstimatedTripCount(L)) - return EstimatedTC; + return *EstimatedTC; // Check if upper bound estimate is known. if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) return ExpectedTC; - return None; + return std::nullopt; } +/// Return a vector containing interleaved elements from multiple +/// smaller input vectors. +static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals, + const Twine &Name) { + unsigned Factor = Vals.size(); + assert(Factor > 1 && "Tried to interleave invalid number of vectors"); + + VectorType *VecTy = cast<VectorType>(Vals[0]->getType()); +#ifndef NDEBUG + for (Value *Val : Vals) + assert(Val->getType() == VecTy && "Tried to interleave mismatched types"); +#endif + + // Scalable vectors cannot use arbitrary shufflevectors (only splats), so + // must use intrinsics to interleave. + if (VecTy->isScalableTy()) { + VectorType *WideVecTy = VectorType::getDoubleElementsVectorType(VecTy); + return Builder.CreateIntrinsic( + WideVecTy, Intrinsic::experimental_vector_interleave2, Vals, + /*FMFSource=*/nullptr, Name); + } + + // Fixed length. Start by concatenating all vectors into a wide vector. + Value *WideVec = concatenateVectors(Builder, Vals); + + // Interleave the elements into the wide vector. + const unsigned NumElts = VecTy->getElementCount().getFixedValue(); + return Builder.CreateShuffleVector( + WideVec, createInterleaveMask(NumElts, Factor), Name); +} + +namespace { // Forward declare GeneratedRTChecks. class GeneratedRTChecks; +using SCEV2ValueTy = DenseMap<const SCEV *, Value *>; +} // namespace + namespace llvm { AnalysisKey ShouldRunExtraVectorPasses::Key; @@ -451,6 +513,7 @@ public: const TargetLibraryInfo *TLI, const TargetTransformInfo *TTI, AssumptionCache *AC, OptimizationRemarkEmitter *ORE, ElementCount VecWidth, + ElementCount MinProfitableTripCount, unsigned UnrollFactor, LoopVectorizationLegality *LVL, LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) @@ -462,6 +525,11 @@ public: // of the original loop header may change as the transformation happens. OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); + + if (MinProfitableTripCount.isZero()) + this->MinProfitableTripCount = VecWidth; + else + this->MinProfitableTripCount = MinProfitableTripCount; } virtual ~InnerLoopVectorizer() = default; @@ -473,15 +541,13 @@ public: /// loop and the start value for the canonical induction, if it is != 0. The /// latter is the case when vectorizing the epilogue loop. In the case of /// epilogue vectorization, this function is overriden to handle the more - /// complex control flow around the loops. - virtual std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton(); - - /// Widen a single call instruction within the innermost loop. - void widenCallInstruction(CallInst &I, VPValue *Def, VPUser &ArgOperands, - VPTransformState &State); + /// complex control flow around the loops. \p ExpandedSCEVs is used to + /// look up SCEV expansions for expressions needed during skeleton creation. + virtual std::pair<BasicBlock *, Value *> + createVectorizedLoopSkeleton(const SCEV2ValueTy &ExpandedSCEVs); /// Fix the vectorized code, taking care of header phi's, live-outs, and more. - void fixVectorizedLoop(VPTransformState &State); + void fixVectorizedLoop(VPTransformState &State, VPlan &Plan); // Return true if any runtime check is added. bool areSafetyChecksAdded() { return AddedSafetyChecks; } @@ -491,32 +557,16 @@ public: /// new unrolled loop, where UF is the unroll factor. using VectorParts = SmallVector<Value *, 2>; - /// Vectorize a single first-order recurrence or pointer induction PHINode in - /// a block. This method handles the induction variable canonicalization. It - /// supports both VF = 1 for unrolled loops and arbitrary length vectors. - void widenPHIInstruction(Instruction *PN, VPWidenPHIRecipe *PhiR, - VPTransformState &State); - /// A helper function to scalarize a single Instruction in the innermost loop. /// Generates a sequence of scalar instances for each lane between \p MinLane /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, /// inclusive. Uses the VPValue operands from \p RepRecipe instead of \p /// Instr's operands. - void scalarizeInstruction(Instruction *Instr, VPReplicateRecipe *RepRecipe, - const VPIteration &Instance, bool IfPredicateInstr, + void scalarizeInstruction(const Instruction *Instr, + VPReplicateRecipe *RepRecipe, + const VPIteration &Instance, VPTransformState &State); - /// Widen an integer or floating-point induction variable \p IV. If \p Trunc - /// is provided, the integer induction variable will first be truncated to - /// the corresponding type. \p CanonicalIV is the scalar value generated for - /// the canonical induction variable. - void widenIntOrFpInduction(PHINode *IV, VPWidenIntOrFpInductionRecipe *Def, - VPTransformState &State, Value *CanonicalIV); - - /// Construct the vector value of a scalarized value \p V one lane at a time. - void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, - VPTransformState &State); - /// Try to vectorize interleaved access group \p Group with the base address /// given in \p Addr, optionally masking the vector operations if \p /// BlockInMask is non-null. Use \p State to translate given VPValues to IR @@ -525,41 +575,34 @@ public: ArrayRef<VPValue *> VPDefs, VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, - VPValue *BlockInMask = nullptr); - - /// Set the debug location in the builder \p Ptr using the debug location in - /// \p V. If \p Ptr is None then it uses the class member's Builder. - void setDebugLocFromInst(const Value *V, - Optional<IRBuilder<> *> CustomBuilder = None); + VPValue *BlockInMask, bool NeedsMaskForGaps); - /// Fix the non-induction PHIs in the OrigPHIsToFix vector. - void fixNonInductionPHIs(VPTransformState &State); + /// Fix the non-induction PHIs in \p Plan. + void fixNonInductionPHIs(VPlan &Plan, VPTransformState &State); /// Returns true if the reordering of FP operations is not allowed, but we are /// able to vectorize with strict in-order reductions for the given RdxDesc. bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc); - /// Create a broadcast instruction. This method generates a broadcast - /// instruction (shuffle) for loop invariant values and for the induction - /// value. If this is the induction variable then we extend it to N, N+1, ... - /// this is needed because each iteration in the loop corresponds to a SIMD - /// element. - virtual Value *getBroadcastInstrs(Value *V); - - /// Add metadata from one instruction to another. - /// - /// This includes both the original MDs from \p From and additional ones (\see - /// addNewMetadata). Use this for *newly created* instructions in the vector - /// loop. - void addMetadata(Instruction *To, Instruction *From); + /// Create a new phi node for the induction variable \p OrigPhi to resume + /// iteration count in the scalar epilogue, from where the vectorized loop + /// left off. \p Step is the SCEV-expanded induction step to use. In cases + /// where the loop skeleton is more complicated (i.e., epilogue vectorization) + /// and the resume values can come from an additional bypass block, the \p + /// AdditionalBypass pair provides information about the bypass block and the + /// end value on the edge from bypass to this loop. + PHINode *createInductionResumeValue( + PHINode *OrigPhi, const InductionDescriptor &ID, Value *Step, + ArrayRef<BasicBlock *> BypassBlocks, + std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); - /// Similar to the previous function but it adds the metadata to a - /// vector of instructions. - void addMetadata(ArrayRef<Value *> To, Instruction *From); + /// Returns the original loop trip count. + Value *getTripCount() const { return TripCount; } - // Returns the resume value (bc.merge.rdx) for a reduction as - // generated by fixReduction. - PHINode *getReductionResumeValue(const RecurrenceDescriptor &RdxDesc); + /// Used to set the trip count after ILV's construction and after the + /// preheader block has been executed. Note that this always holds the trip + /// count of the original loop for both main loop and epilogue vectorization. + void setTripCount(Value *TC) { TripCount = TC; } protected: friend class LoopVectorizationPlanner; @@ -575,68 +618,24 @@ protected: /// Set up the values of the IVs correctly when exiting the vector loop. void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, - Value *CountRoundDown, Value *EndValue, - BasicBlock *MiddleBlock); - - /// Introduce a conditional branch (on true, condition to be set later) at the - /// end of the header=latch connecting it to itself (across the backedge) and - /// to the exit block of \p L. - void createHeaderBranch(Loop *L); - - /// Handle all cross-iteration phis in the header. - void fixCrossIterationPHIs(VPTransformState &State); + Value *VectorTripCount, Value *EndValue, + BasicBlock *MiddleBlock, BasicBlock *VectorHeader, + VPlan &Plan, VPTransformState &State); /// Create the exit value of first order recurrences in the middle block and /// update their users. - void fixFirstOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, + void fixFixedOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State); /// Create code for the loop exit value of the reduction. void fixReduction(VPReductionPHIRecipe *Phi, VPTransformState &State); - /// Clear NSW/NUW flags from reduction instructions if necessary. - void clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, - VPTransformState &State); - - /// Fixup the LCSSA phi nodes in the unique exit block. This simply - /// means we need to add the appropriate incoming value from the middle - /// block as exiting edges from the scalar epilogue loop (if present) are - /// already in place, and we exit the vector loop exclusively to the middle - /// block. - void fixLCSSAPHIs(VPTransformState &State); - /// Iteratively sink the scalarized operands of a predicated instruction into /// the block that was created for it. void sinkScalarOperands(Instruction *PredInst); - /// Shrinks vector element sizes to the smallest bitwidth they can be legally - /// represented as. - void truncateToMinimalBitwidths(VPTransformState &State); - - /// Compute scalar induction steps. \p ScalarIV is the scalar induction - /// variable on which to base the steps, \p Step is the size of the step, and - /// \p EntryVal is the value from the original loop that maps to the steps. - /// Note that \p EntryVal doesn't have to be an induction variable - it - /// can also be a truncate instruction. - void buildScalarSteps(Value *ScalarIV, Value *Step, Instruction *EntryVal, - const InductionDescriptor &ID, VPValue *Def, - VPTransformState &State); - - /// Create a vector induction phi node based on an existing scalar one. \p - /// EntryVal is the value from the original loop that maps to the vector phi - /// node, and \p Step is the loop-invariant step. If \p EntryVal is a - /// truncate instruction, instead of widening the original IV, we widen a - /// version of the IV truncated to \p EntryVal's type. - void createVectorIntOrFpInductionPHI(const InductionDescriptor &II, - Value *Step, Value *Start, - Instruction *EntryVal, VPValue *Def, - VPTransformState &State); - - /// Returns (and creates if needed) the original loop trip count. - Value *getOrCreateTripCount(Loop *NewLoop); - /// Returns (and creates if needed) the trip count of the widened loop. - Value *getOrCreateVectorTripCount(Loop *NewLoop); + Value *getOrCreateVectorTripCount(BasicBlock *InsertBlock); /// Returns a bitcasted value to the requested vector type. /// Also handles bitcasts of vector<float> <-> vector<pointer> types. @@ -645,33 +644,21 @@ protected: /// Emit a bypass check to see if the vector trip count is zero, including if /// it overflows. - void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); + void emitIterationCountCheck(BasicBlock *Bypass); /// Emit a bypass check to see if all of the SCEV assumptions we've /// had to make are correct. Returns the block containing the checks or /// nullptr if no checks have been added. - BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass); + BasicBlock *emitSCEVChecks(BasicBlock *Bypass); /// Emit bypass checks to check any memory assumptions we may have made. /// Returns the block containing the checks or nullptr if no checks have been /// added. - BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); - - /// Compute the transformed value of Index at offset StartValue using step - /// StepValue. - /// For integer induction, returns StartValue + Index * StepValue. - /// For pointer induction, returns StartValue[Index * StepValue]. - /// FIXME: The newly created binary instructions should contain nsw/nuw - /// flags, which can be found from the original scalar operations. - Value *emitTransformedIndex(IRBuilder<> &B, Value *Index, ScalarEvolution *SE, - const DataLayout &DL, - const InductionDescriptor &ID, - BasicBlock *VectorHeader) const; + BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass); /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, - /// vector loop preheader, middle block and scalar preheader. Also - /// allocate a loop object for the new vector loop and return it. - Loop *createVectorLoopSkeleton(StringRef Prefix); + /// vector loop preheader, middle block and scalar preheader. + void createVectorLoopSkeleton(StringRef Prefix); /// Create new phi nodes for the induction variables to resume iteration count /// in the scalar epilogue, from where the vectorized loop left off. @@ -680,21 +667,13 @@ protected: /// block, the \p AdditionalBypass pair provides information about the bypass /// block and the end value on the edge from bypass to this loop. void createInductionResumeValues( - Loop *L, + const SCEV2ValueTy &ExpandedSCEVs, std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); /// Complete the loop skeleton by adding debug MDs, creating appropriate /// conditional branches in the middle block, preparing the builder and - /// running the verifier. Take in the vector loop \p L as argument, and return - /// the preheader of the completed vector loop. - BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID); - - /// Add additional metadata to \p To that was not present on \p Orig. - /// - /// Currently this is used to add the noalias annotations based on the - /// inserted memchecks. Use this for instructions that are *cloned* into the - /// vector loop. - void addNewMetadata(Instruction *To, const Instruction *Orig); + /// running the verifier. Return the preheader of the completed vector loop. + BasicBlock *completeLoopSkeleton(); /// Collect poison-generating recipes that may generate a poison value that is /// used after vectorization, even when their operands are not poison. Those @@ -726,9 +705,6 @@ protected: /// Dominator Tree. DominatorTree *DT; - /// Alias Analysis. - AAResults *AA; - /// Target Library Info. const TargetLibraryInfo *TLI; @@ -741,17 +717,12 @@ protected: /// Interface to emit optimization remarks. OptimizationRemarkEmitter *ORE; - /// LoopVersioning. It's only set up (non-null) if memchecks were - /// used. - /// - /// This is currently only used to add no-alias metadata based on the - /// memchecks. The actually versioning is performed manually. - std::unique_ptr<LoopVersioning> LVer; - /// The vectorization SIMD factor to use. Each vector will have this many /// vector elements. ElementCount VF; + ElementCount MinProfitableTripCount; + /// The vectorization unroll factor to use. Each scalar is vectorized to this /// many different vector instructions. unsigned UF; @@ -774,9 +745,6 @@ protected: /// there can be multiple exiting edges reaching this block. BasicBlock *LoopExitBlock; - /// The vector loop body. - BasicBlock *LoopVectorBody; - /// The scalar loop body. BasicBlock *LoopScalarBody; @@ -805,10 +773,6 @@ protected: // so we can later fix-up the external users of the induction variables. DenseMap<PHINode *, Value *> IVEndValues; - // Vector of original scalar PHIs whose corresponding widened PHIs need to be - // fixed up at the end of vector code generation. - SmallVector<PHINode *, 8> OrigPHIsToFix; - /// BFI and PSI are used to check for profile guided size optimizations. BlockFrequencyInfo *BFI; ProfileSummaryInfo *PSI; @@ -838,11 +802,9 @@ public: LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, + ElementCount::getFixed(1), ElementCount::getFixed(1), UnrollFactor, LVL, CM, BFI, PSI, Check) {} - -private: - Value *getBroadcastInstrs(Value *V) override; }; /// Encapsulate information regarding vectorization of a loop and its epilogue. @@ -886,22 +848,22 @@ public: BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks) : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, - EPI.MainLoopVF, EPI.MainLoopUF, LVL, CM, BFI, PSI, - Checks), + EPI.MainLoopVF, EPI.MainLoopVF, EPI.MainLoopUF, LVL, + CM, BFI, PSI, Checks), EPI(EPI) {} // Override this function to handle the more complex control flow around the // three loops. - std::pair<BasicBlock *, Value *> - createVectorizedLoopSkeleton() final override { - return createEpilogueVectorizedLoopSkeleton(); + std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton( + const SCEV2ValueTy &ExpandedSCEVs) final { + return createEpilogueVectorizedLoopSkeleton(ExpandedSCEVs); } /// The interface for creating a vectorized skeleton using one of two /// different strategies, each corresponding to one execution of the vplan /// as described above. virtual std::pair<BasicBlock *, Value *> - createEpilogueVectorizedLoopSkeleton() = 0; + createEpilogueVectorizedLoopSkeleton(const SCEV2ValueTy &ExpandedSCEVs) = 0; /// Holds and updates state information required to vectorize the main loop /// and its epilogue in two separate passes. This setup helps us avoid @@ -930,14 +892,13 @@ public: /// Implements the interface for creating a vectorized skeleton using the /// *main loop* strategy (ie the first pass of vplan execution). std::pair<BasicBlock *, Value *> - createEpilogueVectorizedLoopSkeleton() final override; + createEpilogueVectorizedLoopSkeleton(const SCEV2ValueTy &ExpandedSCEVs) final; protected: /// Emits an iteration count bypass check once for the main loop (when \p /// ForEpilogue is false) and once for the epilogue loop (when \p /// ForEpilogue is true). - BasicBlock *emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass, - bool ForEpilogue); + BasicBlock *emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue); void printDebugTracesAtStart() override; void printDebugTracesAtEnd() override; }; @@ -956,17 +917,19 @@ public: BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks) : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, - EPI, LVL, CM, BFI, PSI, Checks) {} + EPI, LVL, CM, BFI, PSI, Checks) { + TripCount = EPI.TripCount; + } /// Implements the interface for creating a vectorized skeleton using the /// *epilogue loop* strategy (ie the second pass of vplan execution). std::pair<BasicBlock *, Value *> - createEpilogueVectorizedLoopSkeleton() final override; + createEpilogueVectorizedLoopSkeleton(const SCEV2ValueTy &ExpandedSCEVs) final; protected: /// Emits an iteration count bypass check after the main vector loop has /// finished to see if there are any iterations left to execute by either /// the vector epilogue or the scalar epilogue. - BasicBlock *emitMinimumVectorEpilogueIterCountCheck(Loop *L, + BasicBlock *emitMinimumVectorEpilogueIterCountCheck( BasicBlock *Bypass, BasicBlock *Insert); void printDebugTracesAtStart() override; @@ -976,46 +939,21 @@ protected: /// Look for a meaningful debug location on the instruction or it's /// operands. -static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { +static DebugLoc getDebugLocFromInstOrOperands(Instruction *I) { if (!I) - return I; + return DebugLoc(); DebugLoc Empty; if (I->getDebugLoc() != Empty) - return I; + return I->getDebugLoc(); for (Use &Op : I->operands()) { if (Instruction *OpInst = dyn_cast<Instruction>(Op)) if (OpInst->getDebugLoc() != Empty) - return OpInst; + return OpInst->getDebugLoc(); } - return I; -} - -void InnerLoopVectorizer::setDebugLocFromInst( - const Value *V, Optional<IRBuilder<> *> CustomBuilder) { - IRBuilder<> *B = (CustomBuilder == None) ? &Builder : *CustomBuilder; - if (const Instruction *Inst = dyn_cast_or_null<Instruction>(V)) { - const DILocation *DIL = Inst->getDebugLoc(); - - // When a FSDiscriminator is enabled, we don't need to add the multiply - // factors to the discriminators. - if (DIL && Inst->getFunction()->isDebugInfoForProfiling() && - !isa<DbgInfoIntrinsic>(Inst) && !EnableFSDiscriminator) { - // FIXME: For scalable vectors, assume vscale=1. - auto NewDIL = - DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue()); - if (NewDIL) - B->SetCurrentDebugLocation(NewDIL.getValue()); - else - LLVM_DEBUG(dbgs() - << "Failed to create new discriminator: " - << DIL->getFilename() << " Line: " << DIL->getLine()); - } else - B->SetCurrentDebugLocation(DIL); - } else - B->SetCurrentDebugLocation(DebugLoc()); + return I->getDebugLoc(); } /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I @@ -1059,24 +997,24 @@ static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, namespace llvm { /// Return a value for Step multiplied by VF. -Value *createStepForVF(IRBuilder<> &B, Type *Ty, ElementCount VF, +Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, int64_t Step) { assert(Ty->isIntegerTy() && "Expected an integer step"); - Constant *StepVal = ConstantInt::get(Ty, Step * VF.getKnownMinValue()); - return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; + return B.CreateElementCount(Ty, VF.multiplyCoefficientBy(Step)); } /// Return the runtime value for VF. -Value *getRuntimeVF(IRBuilder<> &B, Type *Ty, ElementCount VF) { - Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); - return VF.isScalable() ? B.CreateVScale(EC) : EC; +Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { + return B.CreateElementCount(Ty, VF); } -static Value *getRuntimeVFAsFloat(IRBuilder<> &B, Type *FTy, ElementCount VF) { - assert(FTy->isFloatingPointTy() && "Expected floating point type!"); - Type *IntTy = IntegerType::get(FTy->getContext(), FTy->getScalarSizeInBits()); - Value *RuntimeVF = getRuntimeVF(B, IntTy, VF); - return B.CreateUIToFP(RuntimeVF, FTy); +const SCEV *createTripCountSCEV(Type *IdxTy, PredicatedScalarEvolution &PSE, + Loop *OrigLoop) { + const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); + assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && "Invalid loop count"); + + ScalarEvolution &SE = *PSE.getSE(); + return SE.getTripCountFromExitCount(BackedgeTakenCount, IdxTy, OrigLoop); } void reportVectorizationFailure(const StringRef DebugMsg, @@ -1100,6 +1038,23 @@ void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, << Msg); } +/// Report successful vectorization of the loop. In case an outer loop is +/// vectorized, prepend "outer" to the vectorization remark. +static void reportVectorization(OptimizationRemarkEmitter *ORE, Loop *TheLoop, + VectorizationFactor VF, unsigned IC) { + LLVM_DEBUG(debugVectorizationMessage( + "Vectorizing: ", TheLoop->isInnermost() ? "innermost loop" : "outer loop", + nullptr)); + StringRef LoopType = TheLoop->isInnermost() ? "" : "outer "; + ORE->emit([&]() { + return OptimizationRemark(LV_NAME, "Vectorized", TheLoop->getStartLoc(), + TheLoop->getHeader()) + << "vectorized " << LoopType << "loop (vectorization width: " + << ore::NV("VectorizationFactor", VF.Width) + << ", interleaved count: " << ore::NV("InterleaveCount", IC) << ")"; + }); +} + } // end namespace llvm #ifndef NDEBUG @@ -1119,14 +1074,6 @@ static std::string getDebugLocString(const Loop *L) { } #endif -void InnerLoopVectorizer::addNewMetadata(Instruction *To, - const Instruction *Orig) { - // If the loop was versioned with memchecks, add the corresponding no-alias - // metadata. - if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig))) - LVer->annotateInstWithNoAlias(To, Orig); -} - void InnerLoopVectorizer::collectPoisonGeneratingRecipes( VPTransformState &State) { @@ -1151,39 +1098,46 @@ void InnerLoopVectorizer::collectPoisonGeneratingRecipes( // handled. if (isa<VPWidenMemoryInstructionRecipe>(CurRec) || isa<VPInterleaveRecipe>(CurRec) || - isa<VPCanonicalIVPHIRecipe>(CurRec)) + isa<VPScalarIVStepsRecipe>(CurRec) || + isa<VPCanonicalIVPHIRecipe>(CurRec) || + isa<VPActiveLaneMaskPHIRecipe>(CurRec)) continue; // This recipe contributes to the address computation of a widen - // load/store. Collect recipe if its underlying instruction has - // poison-generating flags. - Instruction *Instr = CurRec->getUnderlyingInstr(); - if (Instr && Instr->hasPoisonGeneratingFlags()) - State.MayGeneratePoisonRecipes.insert(CurRec); + // load/store. If the underlying instruction has poison-generating flags, + // drop them directly. + if (auto *RecWithFlags = dyn_cast<VPRecipeWithIRFlags>(CurRec)) { + RecWithFlags->dropPoisonGeneratingFlags(); + } else { + Instruction *Instr = dyn_cast_or_null<Instruction>( + CurRec->getVPSingleValue()->getUnderlyingValue()); + (void)Instr; + assert((!Instr || !Instr->hasPoisonGeneratingFlags()) && + "found instruction with poison generating flags not covered by " + "VPRecipeWithIRFlags"); + } // Add new definitions to the worklist. for (VPValue *operand : CurRec->operands()) - if (VPDef *OpDef = operand->getDef()) - Worklist.push_back(cast<VPRecipeBase>(OpDef)); + if (VPRecipeBase *OpDef = operand->getDefiningRecipe()) + Worklist.push_back(OpDef); } }); // Traverse all the recipes in the VPlan and collect the poison-generating // recipes in the backward slice starting at the address of a VPWidenRecipe or // VPInterleaveRecipe. - auto Iter = depth_first( - VPBlockRecursiveTraversalWrapper<VPBlockBase *>(State.Plan->getEntry())); + auto Iter = vp_depth_first_deep(State.Plan->getEntry()); for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { for (VPRecipeBase &Recipe : *VPBB) { if (auto *WidenRec = dyn_cast<VPWidenMemoryInstructionRecipe>(&Recipe)) { - Instruction *UnderlyingInstr = WidenRec->getUnderlyingInstr(); - VPDef *AddrDef = WidenRec->getAddr()->getDef(); - if (AddrDef && WidenRec->isConsecutive() && UnderlyingInstr && - Legal->blockNeedsPredication(UnderlyingInstr->getParent())) - collectPoisonGeneratingInstrsInBackwardSlice( - cast<VPRecipeBase>(AddrDef)); + Instruction &UnderlyingInstr = WidenRec->getIngredient(); + VPRecipeBase *AddrDef = WidenRec->getAddr()->getDefiningRecipe(); + if (AddrDef && WidenRec->isConsecutive() && + Legal->blockNeedsPredication(UnderlyingInstr.getParent())) + collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) { - VPDef *AddrDef = InterleaveRec->getAddr()->getDef(); + VPRecipeBase *AddrDef = InterleaveRec->getAddr()->getDefiningRecipe(); if (AddrDef) { // Check if any member of the interleave group needs predication. const InterleaveGroup<Instruction> *InterGroup = @@ -1198,36 +1152,13 @@ void InnerLoopVectorizer::collectPoisonGeneratingRecipes( } if (NeedPredication) - collectPoisonGeneratingInstrsInBackwardSlice( - cast<VPRecipeBase>(AddrDef)); + collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); } } } } } -void InnerLoopVectorizer::addMetadata(Instruction *To, - Instruction *From) { - propagateMetadata(To, From); - addNewMetadata(To, From); -} - -void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To, - Instruction *From) { - for (Value *V : To) { - if (Instruction *I = dyn_cast<Instruction>(V)) - addMetadata(I, From); - } -} - -PHINode *InnerLoopVectorizer::getReductionResumeValue( - const RecurrenceDescriptor &RdxDesc) { - auto It = ReductionResumeValues.find(&RdxDesc); - assert(It != ReductionResumeValues.end() && - "Expected to find a resume value for the reduction."); - return It->second; -} - namespace llvm { // Loop vectorization cost-model hints how the scalar epilogue loop should be @@ -1253,15 +1184,7 @@ enum ScalarEpilogueLowering { CM_ScalarEpilogueNotAllowedUsePredicate }; -/// ElementCountComparator creates a total ordering for ElementCount -/// for the purposes of using it in a set structure. -struct ElementCountComparator { - bool operator()(const ElementCount &LHS, const ElementCount &RHS) const { - return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) < - std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue()); - } -}; -using ElementCountSet = SmallSet<ElementCount, 16, ElementCountComparator>; +using InstructionVFPair = std::pair<Instruction *, ElementCount>; /// LoopVectorizationCostModel - estimates the expected speedups due to /// vectorization. @@ -1294,17 +1217,6 @@ public: /// otherwise. bool runtimeChecksRequired(); - /// \return The most profitable vectorization factor and the cost of that VF. - /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO - /// then this vectorization factor will be selected if vectorization is - /// possible. - VectorizationFactor - selectVectorizationFactor(const ElementCountSet &CandidateVFs); - - VectorizationFactor - selectEpilogueVectorizationFactor(const ElementCount MaxVF, - const LoopVectorizationPlanner &LVP); - /// Setup cost-based decisions for user vectorization factor. /// \return true if the UserVF is a feasible VF to be chosen. bool selectUserVectorizationFactor(ElementCount UserVF) { @@ -1322,7 +1234,7 @@ public: /// If interleave count has been specified by metadata it will be returned. /// Otherwise, the interleave count is computed and returned. VF and LoopCost /// are the selected vectorization factor and the cost of the selected VF. - unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost); + unsigned selectInterleaveCount(ElementCount VF, InstructionCost LoopCost); /// Memory access instruction may be vectorized in more than one way. /// Form of instruction after vectorization depends on cost. @@ -1333,6 +1245,13 @@ public: /// avoid redundant calculations. void setCostBasedWideningDecision(ElementCount VF); + /// A call may be vectorized in different ways depending on whether we have + /// vectorized variants available and whether the target supports masking. + /// This function analyzes all calls in the function at the supplied VF, + /// makes a decision based on the costs of available options, and stores that + /// decision in a map for use in planning and plan execution. + void setVectorizedCallDecision(ElementCount VF); + /// A struct that represents some properties of the register usage /// of a loop. struct RegisterUsage { @@ -1356,14 +1275,14 @@ public: void collectElementTypesForWidening(); /// Split reductions into those that happen in the loop, and those that happen - /// outside. In loop reductions are collected into InLoopReductionChains. + /// outside. In loop reductions are collected into InLoopReductions. void collectInLoopReductions(); /// Returns true if we should use strict in-order reductions for the given /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, /// the IsOrdered flag of RdxDesc is set and we do not allow reordering /// of FP operations. - bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) { + bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const { return !Hints->allowReordering() && RdxDesc.isOrdered(); } @@ -1388,11 +1307,17 @@ public: auto Scalars = InstsToScalarize.find(VF); assert(Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability"); - return Scalars->second.find(I) != Scalars->second.end(); + return Scalars->second.contains(I); } /// Returns true if \p I is known to be uniform after vectorization. bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { + // Pseudo probe needs to be duplicated for each unrolled iteration and + // vector lane so that profiled loop trip count can be accurately + // accumulated instead of being under counted. + if (isa<PseudoProbeInst>(I)) + return false; + if (VF.isScalar()) return true; @@ -1426,7 +1351,7 @@ public: /// \returns True if instruction \p I can be truncated to a smaller bitwidth /// for vectorization factor \p VF. bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { - return VF.isVector() && MinBWs.find(I) != MinBWs.end() && + return VF.isVector() && MinBWs.contains(I) && !isProfitableToScalarize(I, VF) && !isScalarAfterVectorization(I, VF); } @@ -1438,7 +1363,9 @@ public: CM_Widen_Reverse, // For consecutive accesses with stride -1. CM_Interleave, CM_GatherScatter, - CM_Scalarize + CM_Scalarize, + CM_VectorCall, + CM_IntrinsicCall }; /// Save vectorization decision \p W and \p Cost taken by the cost model for @@ -1489,11 +1416,34 @@ public: InstructionCost getWideningCost(Instruction *I, ElementCount VF) { assert(VF.isVector() && "Expected VF >=2"); std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); - assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() && + assert(WideningDecisions.contains(InstOnVF) && "The cost is not calculated"); return WideningDecisions[InstOnVF].second; } + struct CallWideningDecision { + InstWidening Kind; + Function *Variant; + Intrinsic::ID IID; + std::optional<unsigned> MaskPos; + InstructionCost Cost; + }; + + void setCallWideningDecision(CallInst *CI, ElementCount VF, InstWidening Kind, + Function *Variant, Intrinsic::ID IID, + std::optional<unsigned> MaskPos, + InstructionCost Cost) { + assert(!VF.isScalar() && "Expected vector VF"); + CallWideningDecisions[std::make_pair(CI, VF)] = {Kind, Variant, IID, + MaskPos, Cost}; + } + + CallWideningDecision getCallWideningDecision(CallInst *CI, + ElementCount VF) const { + assert(!VF.isScalar() && "Expected vector VF"); + return CallWideningDecisions.at(std::make_pair(CI, VF)); + } + /// Return True if instruction \p I is an optimizable truncate whose operand /// is an induction variable. Such a truncate will be removed by adding a new /// induction variable with the destination type. @@ -1527,11 +1477,15 @@ public: /// Collect Uniform and Scalar values for the given \p VF. /// The sets depend on CM decision for Load/Store instructions /// that may be vectorized as interleave, gather-scatter or scalarized. + /// Also make a decision on what to do about call instructions in the loop + /// at that VF -- scalarize, call a known vector routine, or call a + /// vector intrinsic. void collectUniformsAndScalars(ElementCount VF) { // Do the analysis once. - if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) + if (VF.isScalar() || Uniforms.contains(VF)) return; setCostBasedWideningDecision(VF); + setVectorizedCallDecision(VF); collectLoopUniforms(VF); collectLoopScalars(VF); } @@ -1552,8 +1506,7 @@ public: /// Returns true if the target machine can represent \p V as a masked gather /// or scatter operation. - bool isLegalGatherOrScatter(Value *V, - ElementCount VF = ElementCount::getFixed(1)) { + bool isLegalGatherOrScatter(Value *V, ElementCount VF) { bool LI = isa<LoadInst>(V); bool SI = isa<StoreInst>(V); if (!LI && !SI) @@ -1575,48 +1528,49 @@ public: })); } - /// Returns true if \p I is an instruction that will be scalarized with - /// predication when vectorizing \p I with vectorization factor \p VF. Such - /// instructions include conditional stores and instructions that may divide - /// by zero. - bool isScalarWithPredication(Instruction *I, ElementCount VF) const; - - // Returns true if \p I is an instruction that will be predicated either - // through scalar predication or masked load/store or masked gather/scatter. - // \p VF is the vectorization factor that will be used to vectorize \p I. - // Superset of instructions that return true for isScalarWithPredication. - bool isPredicatedInst(Instruction *I, ElementCount VF, - bool IsKnownUniform = false) { - // When we know the load is uniform and the original scalar loop was not - // predicated we don't need to mark it as a predicated instruction. Any - // vectorised blocks created when tail-folding are something artificial we - // have introduced and we know there is always at least one active lane. - // That's why we call Legal->blockNeedsPredication here because it doesn't - // query tail-folding. - if (IsKnownUniform && isa<LoadInst>(I) && - !Legal->blockNeedsPredication(I->getParent())) - return false; - if (!blockNeedsPredicationForAnyReason(I->getParent())) + /// Given costs for both strategies, return true if the scalar predication + /// lowering should be used for div/rem. This incorporates an override + /// option so it is not simply a cost comparison. + bool isDivRemScalarWithPredication(InstructionCost ScalarCost, + InstructionCost SafeDivisorCost) const { + switch (ForceSafeDivisor) { + case cl::BOU_UNSET: + return ScalarCost < SafeDivisorCost; + case cl::BOU_TRUE: return false; - // Loads and stores that need some form of masked operation are predicated - // instructions. - if (isa<LoadInst>(I) || isa<StoreInst>(I)) - return Legal->isMaskRequired(I); - return isScalarWithPredication(I, VF); + case cl::BOU_FALSE: + return true; + }; + llvm_unreachable("impossible case value"); } + /// Returns true if \p I is an instruction which requires predication and + /// for which our chosen predication strategy is scalarization (i.e. we + /// don't have an alternate strategy such as masking available). + /// \p VF is the vectorization factor that will be used to vectorize \p I. + bool isScalarWithPredication(Instruction *I, ElementCount VF) const; + + /// Returns true if \p I is an instruction that needs to be predicated + /// at runtime. The result is independent of the predication mechanism. + /// Superset of instructions that return true for isScalarWithPredication. + bool isPredicatedInst(Instruction *I) const; + + /// Return the costs for our two available strategies for lowering a + /// div/rem operation which requires speculating at least one lane. + /// First result is for scalarization (will be invalid for scalable + /// vectors); second is for the safe-divisor strategy. + std::pair<InstructionCost, InstructionCost> + getDivRemSpeculationCost(Instruction *I, + ElementCount VF) const; + /// Returns true if \p I is a memory instruction with consecutive memory /// access that can be widened. - bool - memoryInstructionCanBeWidened(Instruction *I, - ElementCount VF = ElementCount::getFixed(1)); + bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF); /// Returns true if \p I is a memory instruction in an interleaved-group /// of memory accesses that can be vectorized with wide vector loads/stores /// and shuffles. - bool - interleavedAccessCanBeWidened(Instruction *I, - ElementCount VF = ElementCount::getFixed(1)); + bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF); /// Check if \p Instr belongs to any interleaved access group. bool isAccessInterleaved(Instruction *Instr) { @@ -1631,14 +1585,29 @@ public: /// Returns true if we're required to use a scalar epilogue for at least /// the final iteration of the original loop. - bool requiresScalarEpilogue(ElementCount VF) const { + bool requiresScalarEpilogue(bool IsVectorizing) const { if (!isScalarEpilogueAllowed()) return false; // If we might exit from anywhere but the latch, must run the exiting // iteration in scalar form. if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) return true; - return VF.isVector() && InterleaveInfo.requiresScalarEpilogue(); + return IsVectorizing && InterleaveInfo.requiresScalarEpilogue(); + } + + /// Returns true if we're required to use a scalar epilogue for at least + /// the final iteration of the original loop for all VFs in \p Range. + /// A scalar epilogue must either be required for all VFs in \p Range or for + /// none. + bool requiresScalarEpilogue(VFRange Range) const { + auto RequiresScalarEpilogue = [this](ElementCount VF) { + return requiresScalarEpilogue(VF.isVector()); + }; + bool IsRequired = all_of(Range, RequiresScalarEpilogue); + assert( + (IsRequired || none_of(Range, RequiresScalarEpilogue)) && + "all VFs in range must agree on whether a scalar epilogue is required"); + return IsRequired; } /// Returns true if a scalar epilogue is not allowed due to optsize or a @@ -1647,8 +1616,22 @@ public: return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; } + /// Returns the TailFoldingStyle that is best for the current loop. + TailFoldingStyle + getTailFoldingStyle(bool IVUpdateMayOverflow = true) const { + if (!CanFoldTailByMasking) + return TailFoldingStyle::None; + + if (ForceTailFoldingStyle.getNumOccurrences()) + return ForceTailFoldingStyle; + + return TTI.getPreferredTailFoldingStyle(IVUpdateMayOverflow); + } + /// Returns true if all loop blocks should be masked to fold tail loop. - bool foldTailByMasking() const { return FoldTailByMasking; } + bool foldTailByMasking() const { + return getTailFoldingStyle() != TailFoldingStyle::None; + } /// Returns true if the instructions in this block requires predication /// for any reason, e.g. because tail folding now requires a predicate @@ -1657,20 +1640,9 @@ public: return foldTailByMasking() || Legal->blockNeedsPredication(BB); } - /// A SmallMapVector to store the InLoop reduction op chains, mapping phi - /// nodes to the chain of instructions representing the reductions. Uses a - /// MapVector to ensure deterministic iteration order. - using ReductionChainMap = - SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; - - /// Return the chain of instructions representing an inloop reduction. - const ReductionChainMap &getInLoopReductionChains() const { - return InLoopReductionChains; - } - /// Returns true if the Phi is part of an inloop reduction. bool isInLoopReduction(PHINode *Phi) const { - return InLoopReductionChains.count(Phi); + return InLoopReductions.contains(Phi); } /// Estimate cost of an intrinsic call instruction CI if it were vectorized @@ -1680,77 +1652,66 @@ public: /// Estimate cost of a call instruction CI if it were vectorized with factor /// VF. Return the cost of the instruction, including scalarization overhead - /// if it's needed. The flag NeedToScalarize shows if the call needs to be - /// scalarized - - /// i.e. either vector version isn't available, or is too expensive. - InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, - bool &NeedToScalarize) const; - - /// Returns true if the per-lane cost of VectorizationFactor A is lower than - /// that of B. - bool isMoreProfitable(const VectorizationFactor &A, - const VectorizationFactor &B) const; + /// if it's needed. + InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF) const; /// Invalidates decisions already taken by the cost model. void invalidateCostModelingDecisions() { WideningDecisions.clear(); + CallWideningDecisions.clear(); Uniforms.clear(); Scalars.clear(); } + /// The vectorization cost is a combination of the cost itself and a boolean + /// indicating whether any of the contributing operations will actually + /// operate on vector values after type legalization in the backend. If this + /// latter value is false, then all operations will be scalarized (i.e. no + /// vectorization has actually taken place). + using VectorizationCostTy = std::pair<InstructionCost, bool>; + + /// Returns the expected execution cost. The unit of the cost does + /// not matter because we use the 'cost' units to compare different + /// vector widths. The cost that is returned is *not* normalized by + /// the factor width. If \p Invalid is not nullptr, this function + /// will add a pair(Instruction*, ElementCount) to \p Invalid for + /// each instruction that has an Invalid cost for the given VF. + VectorizationCostTy + expectedCost(ElementCount VF, + SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); + + bool hasPredStores() const { return NumPredStores > 0; } + + /// Returns true if epilogue vectorization is considered profitable, and + /// false otherwise. + /// \p VF is the vectorization factor chosen for the original loop. + bool isEpilogueVectorizationProfitable(const ElementCount VF) const; + private: unsigned NumPredStores = 0; - /// Convenience function that returns the value of vscale_range iff - /// vscale_range.min == vscale_range.max or otherwise returns the value - /// returned by the corresponding TLI method. - Optional<unsigned> getVScaleForTuning() const; - /// \return An upper bound for the vectorization factors for both /// fixed and scalable vectorization, where the minimum-known number of /// elements is a power-of-2 larger than zero. If scalable vectorization is /// disabled or unsupported, then the scalable part will be equal to /// ElementCount::getScalable(0). - FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount, + FixedScalableVFPair computeFeasibleMaxVF(unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking); /// \return the maximized element count based on the targets vector /// registers and the loop trip-count, but limited to a maximum safe VF. /// This is a helper function of computeFeasibleMaxVF. - /// FIXME: MaxSafeVF is currently passed by reference to avoid some obscure - /// issue that occurred on one of the buildbots which cannot be reproduced - /// without having access to the properietary compiler (see comments on - /// D98509). The issue is currently under investigation and this workaround - /// will be removed as soon as possible. - ElementCount getMaximizedVFForTarget(unsigned ConstTripCount, + ElementCount getMaximizedVFForTarget(unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType, - const ElementCount &MaxSafeVF, + ElementCount MaxSafeVF, bool FoldTailByMasking); /// \return the maximum legal scalable VF, based on the safe max number /// of elements. ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); - /// The vectorization cost is a combination of the cost itself and a boolean - /// indicating whether any of the contributing operations will actually - /// operate on vector values after type legalization in the backend. If this - /// latter value is false, then all operations will be scalarized (i.e. no - /// vectorization has actually taken place). - using VectorizationCostTy = std::pair<InstructionCost, bool>; - - /// Returns the expected execution cost. The unit of the cost does - /// not matter because we use the 'cost' units to compare different - /// vector widths. The cost that is returned is *not* normalized by - /// the factor width. If \p Invalid is not nullptr, this function - /// will add a pair(Instruction*, ElementCount) to \p Invalid for - /// each instruction that has an Invalid cost for the given VF. - using InstructionVFPair = std::pair<Instruction *, ElementCount>; - VectorizationCostTy - expectedCost(ElementCount VF, - SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); - /// Returns the execution time cost of an instruction for a given vector /// width. Vector width of one means scalar. VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); @@ -1762,9 +1723,9 @@ private: /// Return the cost of instructions in an inloop reduction pattern, if I is /// part of that pattern. - Optional<InstructionCost> + std::optional<InstructionCost> getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy, - TTI::TargetCostKind CostKind); + TTI::TargetCostKind CostKind) const; /// Calculate vectorization cost of memory instruction \p I. InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); @@ -1790,12 +1751,8 @@ private: /// Estimate the overhead of scalarizing an instruction. This is a /// convenience wrapper for the type-based getScalarizationOverhead API. - InstructionCost getScalarizationOverhead(Instruction *I, - ElementCount VF) const; - - /// Returns whether the instruction is a load or store and will be a emitted - /// as a vector operation. - bool isConsecutiveLoadOrStore(Instruction *I); + InstructionCost getScalarizationOverhead(Instruction *I, ElementCount VF, + TTI::TargetCostKind CostKind) const; /// Returns true if an artificially high cost for emulated masked memrefs /// should be used. @@ -1813,7 +1770,8 @@ private: /// A set containing all BasicBlocks that are known to present after /// vectorization as a predicated block. - SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization; + DenseMap<ElementCount, SmallPtrSet<BasicBlock *, 4>> + PredicatedBBsAfterVectorization; /// Records whether it is allowed to have the original scalar loop execute at /// least once. This may be needed as a fallback loop in case runtime @@ -1825,7 +1783,7 @@ private: ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; /// All blocks of loop are to be masked to fold tail of scalar iterations. - bool FoldTailByMasking = false; + bool CanFoldTailByMasking = false; /// A map holding scalar costs for different vectorization factors. The /// presence of a cost for an instruction in the mapping indicates that the @@ -1845,15 +1803,12 @@ private: /// scalarized. DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; - /// PHINodes of the reductions that should be expanded in-loop along with - /// their associated chains of reduction operations, in program order from top - /// (PHI) to bottom - ReductionChainMap InLoopReductionChains; + /// PHINodes of the reductions that should be expanded in-loop. + SmallPtrSet<PHINode *, 4> InLoopReductions; /// A Map of inloop reduction operations and their immediate chain operand. /// FIXME: This can be removed once reductions can be costed correctly in - /// vplan. This was added to allow quick lookup to the inloop operations, - /// without having to loop through InLoopReductionChains. + /// VPlan. This was added to allow quick lookup of the inloop operations. DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; /// Returns the expected difference in cost from scalarizing the expression @@ -1861,8 +1816,9 @@ private: /// scalarize and their scalar costs are collected in \p ScalarCosts. A /// non-negative return value implies the expression will be scalarized. /// Currently, only single-use chains are considered for scalarization. - int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, - ElementCount VF); + InstructionCost computePredInstDiscount(Instruction *PredInst, + ScalarCostsTy &ScalarCosts, + ElementCount VF); /// Collect the instructions that are uniform after vectorization. An /// instruction is uniform if we represent it with a single scalar value in @@ -1891,6 +1847,11 @@ private: DecisionList WideningDecisions; + using CallDecisionList = + DenseMap<std::pair<CallInst *, ElementCount>, CallWideningDecision>; + + CallDecisionList CallWideningDecisions; + /// Returns true if \p V is expected to be vectorized and it needs to be /// extracted. bool needsExtract(Value *V, ElementCount VF) const { @@ -1905,8 +1866,7 @@ private: // the scalars are collected. That should be a safe assumption in most // cases, because we check if the operands have vectorizable types // beforehand in LoopVectorizationLegality. - return Scalars.find(VF) == Scalars.end() || - !isScalarAfterVectorization(I, VF); + return !Scalars.contains(VF) || !isScalarAfterVectorization(I, VF); }; /// Returns a range containing only operands needing to be extracted. @@ -1916,16 +1876,6 @@ private: Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); } - /// Determines if we have the infrastructure to vectorize loop \p L and its - /// epilogue, assuming the main loop is vectorized by \p VF. - bool isCandidateForEpilogueVectorization(const Loop &L, - const ElementCount VF) const; - - /// Returns true if epilogue vectorization is considered profitable, and - /// false otherwise. - /// \p VF is the vectorization factor chosen for the original loop. - bool isEpilogueVectorizationProfitable(const ElementCount VF) const; - public: /// The loop that we evaluate. Loop *TheLoop; @@ -1971,12 +1921,10 @@ public: /// All element types found in the loop. SmallPtrSet<Type *, 16> ElementTypesInLoop; - - /// Profitable vector factors. - SmallVector<VectorizationFactor, 8> ProfitableVFs; }; } // end namespace llvm +namespace { /// Helper struct to manage generating runtime checks for vectorization. /// /// The runtime checks are created up-front in temporary blocks to allow better @@ -2001,15 +1949,22 @@ class GeneratedRTChecks { DominatorTree *DT; LoopInfo *LI; + TargetTransformInfo *TTI; SCEVExpander SCEVExp; SCEVExpander MemCheckExp; + bool CostTooHigh = false; + const bool AddBranchWeights; + + Loop *OuterLoop = nullptr; + public: GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, - const DataLayout &DL) - : DT(DT), LI(LI), SCEVExp(SE, DL, "scev.check"), - MemCheckExp(SE, DL, "scev.check") {} + TargetTransformInfo *TTI, const DataLayout &DL, + bool AddBranchWeights) + : DT(DT), LI(LI), TTI(TTI), SCEVExp(SE, DL, "scev.check"), + MemCheckExp(SE, DL, "scev.check"), AddBranchWeights(AddBranchWeights) {} /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can /// accurately estimate the cost of the runtime checks. The blocks are @@ -2017,7 +1972,16 @@ public: /// there is no vector code generation, the check blocks are removed /// completely. void Create(Loop *L, const LoopAccessInfo &LAI, - const SCEVUnionPredicate &UnionPred) { + const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) { + + // Hard cutoff to limit compile-time increase in case a very large number of + // runtime checks needs to be generated. + // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to + // profile info. + CostTooHigh = + LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold; + if (CostTooHigh) + return; BasicBlock *LoopHeader = L->getHeader(); BasicBlock *Preheader = L->getLoopPreheader(); @@ -2040,9 +2004,22 @@ public: MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, "vector.memcheck"); - MemRuntimeCheckCond = - addRuntimeChecks(MemCheckBlock->getTerminator(), L, - RtPtrChecking.getChecks(), MemCheckExp); + auto DiffChecks = RtPtrChecking.getDiffChecks(); + if (DiffChecks) { + Value *RuntimeVF = nullptr; + MemRuntimeCheckCond = addDiffRuntimeChecks( + MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp, + [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) { + if (!RuntimeVF) + RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF); + return RuntimeVF; + }, + IC); + } else { + MemRuntimeCheckCond = addRuntimeChecks( + MemCheckBlock->getTerminator(), L, RtPtrChecking.getChecks(), + MemCheckExp, VectorizerParams::HoistRuntimeChecks); + } assert(MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required"); @@ -2078,6 +2055,92 @@ public: DT->eraseNode(SCEVCheckBlock); LI->removeBlock(SCEVCheckBlock); } + + // Outer loop is used as part of the later cost calculations. + OuterLoop = L->getParentLoop(); + } + + InstructionCost getCost() { + if (SCEVCheckBlock || MemCheckBlock) + LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n"); + + if (CostTooHigh) { + InstructionCost Cost; + Cost.setInvalid(); + LLVM_DEBUG(dbgs() << " number of checks exceeded threshold\n"); + return Cost; + } + + InstructionCost RTCheckCost = 0; + if (SCEVCheckBlock) + for (Instruction &I : *SCEVCheckBlock) { + if (SCEVCheckBlock->getTerminator() == &I) + continue; + InstructionCost C = + TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); + LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n"); + RTCheckCost += C; + } + if (MemCheckBlock) { + InstructionCost MemCheckCost = 0; + for (Instruction &I : *MemCheckBlock) { + if (MemCheckBlock->getTerminator() == &I) + continue; + InstructionCost C = + TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); + LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n"); + MemCheckCost += C; + } + + // If the runtime memory checks are being created inside an outer loop + // we should find out if these checks are outer loop invariant. If so, + // the checks will likely be hoisted out and so the effective cost will + // reduce according to the outer loop trip count. + if (OuterLoop) { + ScalarEvolution *SE = MemCheckExp.getSE(); + // TODO: If profitable, we could refine this further by analysing every + // individual memory check, since there could be a mixture of loop + // variant and invariant checks that mean the final condition is + // variant. + const SCEV *Cond = SE->getSCEV(MemRuntimeCheckCond); + if (SE->isLoopInvariant(Cond, OuterLoop)) { + // It seems reasonable to assume that we can reduce the effective + // cost of the checks even when we know nothing about the trip + // count. Assume that the outer loop executes at least twice. + unsigned BestTripCount = 2; + + // If exact trip count is known use that. + if (unsigned SmallTC = SE->getSmallConstantTripCount(OuterLoop)) + BestTripCount = SmallTC; + else if (LoopVectorizeWithBlockFrequency) { + // Else use profile data if available. + if (auto EstimatedTC = getLoopEstimatedTripCount(OuterLoop)) + BestTripCount = *EstimatedTC; + } + + InstructionCost NewMemCheckCost = MemCheckCost / BestTripCount; + + // Let's ensure the cost is always at least 1. + NewMemCheckCost = std::max(*NewMemCheckCost.getValue(), + (InstructionCost::CostType)1); + + LLVM_DEBUG(dbgs() + << "We expect runtime memory checks to be hoisted " + << "out of the outer loop. Cost reduced from " + << MemCheckCost << " to " << NewMemCheckCost << '\n'); + + MemCheckCost = NewMemCheckCost; + } + } + + RTCheckCost += MemCheckCost; + } + + if (SCEVCheckBlock || MemCheckBlock) + LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCost + << "\n"); + + return RTCheckCost; } /// Remove the created SCEV & memory runtime check blocks & instructions, if @@ -2114,12 +2177,16 @@ public: /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and /// adjusts the branches to branch to the vector preheader or \p Bypass, /// depending on the generated condition. - BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass, + BasicBlock *emitSCEVChecks(BasicBlock *Bypass, BasicBlock *LoopVectorPreHeader, BasicBlock *LoopExitBlock) { if (!SCEVCheckCond) return nullptr; - if (auto *C = dyn_cast<ConstantInt>(SCEVCheckCond)) + + Value *Cond = SCEVCheckCond; + // Mark the check as used, to prevent it from being removed during cleanup. + SCEVCheckCond = nullptr; + if (auto *C = dyn_cast<ConstantInt>(Cond)) if (C->isZero()) return nullptr; @@ -2127,8 +2194,8 @@ public: BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); // Create new preheader for vector loop. - if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) - PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); + if (OuterLoop) + OuterLoop->addBasicBlockToLoop(SCEVCheckBlock, *LI); SCEVCheckBlock->getTerminator()->eraseFromParent(); SCEVCheckBlock->moveBefore(LoopVectorPreHeader); @@ -2138,18 +2205,17 @@ public: DT->addNewBlock(SCEVCheckBlock, Pred); DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); - ReplaceInstWithInst( - SCEVCheckBlock->getTerminator(), - BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheckCond)); - // Mark the check as used, to prevent it from being removed during cleanup. - SCEVCheckCond = nullptr; + BranchInst &BI = *BranchInst::Create(Bypass, LoopVectorPreHeader, Cond); + if (AddBranchWeights) + setBranchWeights(BI, SCEVCheckBypassWeights); + ReplaceInstWithInst(SCEVCheckBlock->getTerminator(), &BI); return SCEVCheckBlock; } /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts /// the branches to branch to the vector preheader or \p Bypass, depending on /// the generated condition. - BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass, + BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass, BasicBlock *LoopVectorPreHeader) { // Check if we generated code that checks in runtime if arrays overlap. if (!MemRuntimeCheckCond) @@ -2163,12 +2229,15 @@ public: DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); MemCheckBlock->moveBefore(LoopVectorPreHeader); - if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) - PL->addBasicBlockToLoop(MemCheckBlock, *LI); + if (OuterLoop) + OuterLoop->addBasicBlockToLoop(MemCheckBlock, *LI); - ReplaceInstWithInst( - MemCheckBlock->getTerminator(), - BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); + BranchInst &BI = + *BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond); + if (AddBranchWeights) { + setBranchWeights(BI, MemCheckBypassWeights); + } + ReplaceInstWithInst(MemCheckBlock->getTerminator(), &BI); MemCheckBlock->getTerminator()->setDebugLoc( Pred->getTerminator()->getDebugLoc()); @@ -2177,6 +2246,18 @@ public: return MemCheckBlock; } }; +} // namespace + +static bool useActiveLaneMask(TailFoldingStyle Style) { + return Style == TailFoldingStyle::Data || + Style == TailFoldingStyle::DataAndControlFlow || + Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; +} + +static bool useActiveLaneMaskForControlFlow(TailFoldingStyle Style) { + return Style == TailFoldingStyle::DataAndControlFlow || + Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; +} // Return true if \p OuterLp is an outer loop annotated with hints for explicit // vectorization. The loop needs to be annotated with #pragma omp simd @@ -2245,426 +2326,139 @@ static void collectSupportedLoops(Loop &L, LoopInfo *LI, collectSupportedLoops(*InnerL, LI, ORE, V); } -namespace { - -/// The LoopVectorize Pass. -struct LoopVectorize : public FunctionPass { - /// Pass identification, replacement for typeid - static char ID; - - LoopVectorizePass Impl; - - explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, - bool VectorizeOnlyWhenForced = false) - : FunctionPass(ID), - Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { - initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); - } - - bool runOnFunction(Function &F) override { - if (skipFunction(F)) - return false; - - auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); - auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); - auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); - auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); - auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); - auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); - auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; - auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); - auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); - auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); - auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); - auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); - auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); - - std::function<const LoopAccessInfo &(Loop &)> GetLAA = - [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; - - return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC, - GetLAA, *ORE, PSI).MadeAnyChange; - } - - void getAnalysisUsage(AnalysisUsage &AU) const override { - AU.addRequired<AssumptionCacheTracker>(); - AU.addRequired<BlockFrequencyInfoWrapperPass>(); - AU.addRequired<DominatorTreeWrapperPass>(); - AU.addRequired<LoopInfoWrapperPass>(); - AU.addRequired<ScalarEvolutionWrapperPass>(); - AU.addRequired<TargetTransformInfoWrapperPass>(); - AU.addRequired<AAResultsWrapperPass>(); - AU.addRequired<LoopAccessLegacyAnalysis>(); - AU.addRequired<DemandedBitsWrapperPass>(); - AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); - AU.addRequired<InjectTLIMappingsLegacy>(); - - // We currently do not preserve loopinfo/dominator analyses with outer loop - // vectorization. Until this is addressed, mark these analyses as preserved - // only for non-VPlan-native path. - // TODO: Preserve Loop and Dominator analyses for VPlan-native path. - if (!EnableVPlanNativePath) { - AU.addPreserved<LoopInfoWrapperPass>(); - AU.addPreserved<DominatorTreeWrapperPass>(); - } - - AU.addPreserved<BasicAAWrapperPass>(); - AU.addPreserved<GlobalsAAWrapperPass>(); - AU.addRequired<ProfileSummaryInfoWrapperPass>(); - } -}; - -} // end anonymous namespace - //===----------------------------------------------------------------------===// // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and // LoopVectorizationCostModel and LoopVectorizationPlanner. //===----------------------------------------------------------------------===// -Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { - // We need to place the broadcast of invariant variables outside the loop, - // but only if it's proven safe to do so. Else, broadcast will be inside - // vector loop body. - Instruction *Instr = dyn_cast<Instruction>(V); - bool SafeToHoist = OrigLoop->isLoopInvariant(V) && - (!Instr || - DT->dominates(Instr->getParent(), LoopVectorPreHeader)); - // Place the code for broadcasting invariant variables in the new preheader. - IRBuilder<>::InsertPointGuard Guard(Builder); - if (SafeToHoist) - Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); - - // Broadcast the scalar into all locations in the vector. - Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); - - return Shuf; -} - -/// This function adds -/// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...) -/// to each vector element of Val. The sequence starts at StartIndex. -/// \p Opcode is relevant for FP induction variable. -static Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, - Instruction::BinaryOps BinOp, ElementCount VF, - IRBuilder<> &Builder) { - assert(VF.isVector() && "only vector VFs are supported"); - - // Create and check the types. - auto *ValVTy = cast<VectorType>(Val->getType()); - ElementCount VLen = ValVTy->getElementCount(); - - Type *STy = Val->getType()->getScalarType(); - assert((STy->isIntegerTy() || STy->isFloatingPointTy()) && - "Induction Step must be an integer or FP"); - assert(Step->getType() == STy && "Step has wrong type"); - - SmallVector<Constant *, 8> Indices; - - // Create a vector of consecutive numbers from zero to VF. - VectorType *InitVecValVTy = ValVTy; - Type *InitVecValSTy = STy; - if (STy->isFloatingPointTy()) { - InitVecValSTy = - IntegerType::get(STy->getContext(), STy->getScalarSizeInBits()); - InitVecValVTy = VectorType::get(InitVecValSTy, VLen); - } - Value *InitVec = Builder.CreateStepVector(InitVecValVTy); - - // Splat the StartIdx - Value *StartIdxSplat = Builder.CreateVectorSplat(VLen, StartIdx); - - if (STy->isIntegerTy()) { - InitVec = Builder.CreateAdd(InitVec, StartIdxSplat); - Step = Builder.CreateVectorSplat(VLen, Step); - assert(Step->getType() == Val->getType() && "Invalid step vec"); - // FIXME: The newly created binary instructions should contain nsw/nuw - // flags, which can be found from the original scalar operations. - Step = Builder.CreateMul(InitVec, Step); - return Builder.CreateAdd(Val, Step, "induction"); - } - - // Floating point induction. - assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && - "Binary Opcode should be specified for FP induction"); - InitVec = Builder.CreateUIToFP(InitVec, ValVTy); - InitVec = Builder.CreateFAdd(InitVec, StartIdxSplat); - - Step = Builder.CreateVectorSplat(VLen, Step); - Value *MulOp = Builder.CreateFMul(InitVec, Step); - return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); -} - -void InnerLoopVectorizer::createVectorIntOrFpInductionPHI( - const InductionDescriptor &II, Value *Step, Value *Start, - Instruction *EntryVal, VPValue *Def, VPTransformState &State) { - IRBuilder<> &Builder = State.Builder; - assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && - "Expected either an induction phi-node or a truncate of it!"); - - // Construct the initial value of the vector IV in the vector loop preheader - auto CurrIP = Builder.saveIP(); - Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); - if (isa<TruncInst>(EntryVal)) { - assert(Start->getType()->isIntegerTy() && - "Truncation requires an integer type"); - auto *TruncType = cast<IntegerType>(EntryVal->getType()); - Step = Builder.CreateTrunc(Step, TruncType); - Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); - } - - Value *Zero = getSignedIntOrFpConstant(Start->getType(), 0); - Value *SplatStart = Builder.CreateVectorSplat(State.VF, Start); - Value *SteppedStart = getStepVector( - SplatStart, Zero, Step, II.getInductionOpcode(), State.VF, State.Builder); - - // We create vector phi nodes for both integer and floating-point induction - // variables. Here, we determine the kind of arithmetic we will perform. - Instruction::BinaryOps AddOp; - Instruction::BinaryOps MulOp; - if (Step->getType()->isIntegerTy()) { - AddOp = Instruction::Add; - MulOp = Instruction::Mul; - } else { - AddOp = II.getInductionOpcode(); - MulOp = Instruction::FMul; +/// Compute the transformed value of Index at offset StartValue using step +/// StepValue. +/// For integer induction, returns StartValue + Index * StepValue. +/// For pointer induction, returns StartValue[Index * StepValue]. +/// FIXME: The newly created binary instructions should contain nsw/nuw +/// flags, which can be found from the original scalar operations. +static Value * +emitTransformedIndex(IRBuilderBase &B, Value *Index, Value *StartValue, + Value *Step, + InductionDescriptor::InductionKind InductionKind, + const BinaryOperator *InductionBinOp) { + Type *StepTy = Step->getType(); + Value *CastedIndex = StepTy->isIntegerTy() + ? B.CreateSExtOrTrunc(Index, StepTy) + : B.CreateCast(Instruction::SIToFP, Index, StepTy); + if (CastedIndex != Index) { + CastedIndex->setName(CastedIndex->getName() + ".cast"); + Index = CastedIndex; } - // Multiply the vectorization factor by the step using integer or - // floating-point arithmetic as appropriate. - Type *StepType = Step->getType(); - Value *RuntimeVF; - if (Step->getType()->isFloatingPointTy()) - RuntimeVF = getRuntimeVFAsFloat(Builder, StepType, State.VF); - else - RuntimeVF = getRuntimeVF(Builder, StepType, State.VF); - Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF); - - // Create a vector splat to use in the induction update. - // - // FIXME: If the step is non-constant, we create the vector splat with - // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't - // handle a constant vector splat. - Value *SplatVF = isa<Constant>(Mul) - ? ConstantVector::getSplat(State.VF, cast<Constant>(Mul)) - : Builder.CreateVectorSplat(State.VF, Mul); - Builder.restoreIP(CurrIP); - - // We may need to add the step a number of times, depending on the unroll - // factor. The last of those goes into the PHI. - PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", - &*LoopVectorBody->getFirstInsertionPt()); - VecInd->setDebugLoc(EntryVal->getDebugLoc()); - Instruction *LastInduction = VecInd; - for (unsigned Part = 0; Part < UF; ++Part) { - State.set(Def, LastInduction, Part); - - if (isa<TruncInst>(EntryVal)) - addMetadata(LastInduction, EntryVal); - - LastInduction = cast<Instruction>( - Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); - LastInduction->setDebugLoc(EntryVal->getDebugLoc()); - } - - // Move the last step to the end of the latch block. This ensures consistent - // placement of all induction updates. - auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); - auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator()); - LastInduction->moveBefore(Br); - LastInduction->setName("vec.ind.next"); - - VecInd->addIncoming(SteppedStart, LoopVectorPreHeader); - VecInd->addIncoming(LastInduction, LoopVectorLatch); -} - -void InnerLoopVectorizer::widenIntOrFpInduction( - PHINode *IV, VPWidenIntOrFpInductionRecipe *Def, VPTransformState &State, - Value *CanonicalIV) { - Value *Start = Def->getStartValue()->getLiveInIRValue(); - const InductionDescriptor &ID = Def->getInductionDescriptor(); - TruncInst *Trunc = Def->getTruncInst(); - IRBuilder<> &Builder = State.Builder; - assert(IV->getType() == ID.getStartValue()->getType() && "Types must match"); - assert(!State.VF.isZero() && "VF must be non-zero"); - - // The value from the original loop to which we are mapping the new induction - // variable. - Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; - - auto &DL = EntryVal->getModule()->getDataLayout(); - - // Generate code for the induction step. Note that induction steps are - // required to be loop-invariant - auto CreateStepValue = [&](const SCEV *Step) -> Value * { - assert(PSE.getSE()->isLoopInvariant(Step, OrigLoop) && - "Induction step should be loop invariant"); - if (PSE.getSE()->isSCEVable(IV->getType())) { - SCEVExpander Exp(*PSE.getSE(), DL, "induction"); - return Exp.expandCodeFor(Step, Step->getType(), - State.CFG.VectorPreHeader->getTerminator()); - } - return cast<SCEVUnknown>(Step)->getValue(); + // Note: the IR at this point is broken. We cannot use SE to create any new + // SCEV and then expand it, hoping that SCEV's simplification will give us + // a more optimal code. Unfortunately, attempt of doing so on invalid IR may + // lead to various SCEV crashes. So all we can do is to use builder and rely + // on InstCombine for future simplifications. Here we handle some trivial + // cases only. + auto CreateAdd = [&B](Value *X, Value *Y) { + assert(X->getType() == Y->getType() && "Types don't match!"); + if (auto *CX = dyn_cast<ConstantInt>(X)) + if (CX->isZero()) + return Y; + if (auto *CY = dyn_cast<ConstantInt>(Y)) + if (CY->isZero()) + return X; + return B.CreateAdd(X, Y); }; - // The scalar value to broadcast. This is derived from the canonical - // induction variable. If a truncation type is given, truncate the canonical - // induction variable and step. Otherwise, derive these values from the - // induction descriptor. - auto CreateScalarIV = [&](Value *&Step) -> Value * { - Value *ScalarIV = CanonicalIV; - Type *NeededType = IV->getType(); - if (!Def->isCanonical() || ScalarIV->getType() != NeededType) { - ScalarIV = - NeededType->isIntegerTy() - ? Builder.CreateSExtOrTrunc(ScalarIV, NeededType) - : Builder.CreateCast(Instruction::SIToFP, ScalarIV, NeededType); - ScalarIV = emitTransformedIndex(Builder, ScalarIV, PSE.getSE(), DL, ID, - State.CFG.PrevBB); - ScalarIV->setName("offset.idx"); - } - if (Trunc) { - auto *TruncType = cast<IntegerType>(Trunc->getType()); - assert(Step->getType()->isIntegerTy() && - "Truncation requires an integer step"); - ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType); - Step = Builder.CreateTrunc(Step, TruncType); - } - return ScalarIV; + // We allow X to be a vector type, in which case Y will potentially be + // splatted into a vector with the same element count. + auto CreateMul = [&B](Value *X, Value *Y) { + assert(X->getType()->getScalarType() == Y->getType() && + "Types don't match!"); + if (auto *CX = dyn_cast<ConstantInt>(X)) + if (CX->isOne()) + return Y; + if (auto *CY = dyn_cast<ConstantInt>(Y)) + if (CY->isOne()) + return X; + VectorType *XVTy = dyn_cast<VectorType>(X->getType()); + if (XVTy && !isa<VectorType>(Y->getType())) + Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); + return B.CreateMul(X, Y); }; - // Fast-math-flags propagate from the original induction instruction. - IRBuilder<>::FastMathFlagGuard FMFG(Builder); - if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) - Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); - - // Now do the actual transformations, and start with creating the step value. - Value *Step = CreateStepValue(ID.getStep()); - if (State.VF.isScalar()) { - Value *ScalarIV = CreateScalarIV(Step); - Type *ScalarTy = IntegerType::get(ScalarIV->getContext(), - Step->getType()->getScalarSizeInBits()); - - for (unsigned Part = 0; Part < UF; ++Part) { - Value *StartIdx = ConstantInt::get(ScalarTy, Part); - Value *EntryPart; - if (Step->getType()->isFloatingPointTy()) { - StartIdx = Builder.CreateUIToFP(StartIdx, Step->getType()); - Value *MulOp = Builder.CreateFMul(StartIdx, Step); - EntryPart = Builder.CreateBinOp(ID.getInductionOpcode(), ScalarIV, - MulOp, "induction"); - } else { - EntryPart = Builder.CreateAdd( - ScalarIV, Builder.CreateMul(StartIdx, Step), "induction"); - } - State.set(Def, EntryPart, Part); - if (Trunc) { - assert(!Step->getType()->isFloatingPointTy() && - "fp inductions shouldn't be truncated"); - addMetadata(EntryPart, Trunc); - } - } - return; + switch (InductionKind) { + case InductionDescriptor::IK_IntInduction: { + assert(!isa<VectorType>(Index->getType()) && + "Vector indices not supported for integer inductions yet"); + assert(Index->getType() == StartValue->getType() && + "Index type does not match StartValue type"); + if (isa<ConstantInt>(Step) && cast<ConstantInt>(Step)->isMinusOne()) + return B.CreateSub(StartValue, Index); + auto *Offset = CreateMul(Index, Step); + return CreateAdd(StartValue, Offset); } + case InductionDescriptor::IK_PtrInduction: + return B.CreatePtrAdd(StartValue, CreateMul(Index, Step)); + case InductionDescriptor::IK_FpInduction: { + assert(!isa<VectorType>(Index->getType()) && + "Vector indices not supported for FP inductions yet"); + assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value"); + assert(InductionBinOp && + (InductionBinOp->getOpcode() == Instruction::FAdd || + InductionBinOp->getOpcode() == Instruction::FSub) && + "Original bin op should be defined for FP induction"); - // Create a new independent vector induction variable, if one is needed. - if (Def->needsVectorIV()) - createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, State); - - if (Def->needsScalarIV()) { - // Create scalar steps that can be used by instructions we will later - // scalarize. Note that the addition of the scalar steps will not increase - // the number of instructions in the loop in the common case prior to - // InstCombine. We will be trading one vector extract for each scalar step. - Value *ScalarIV = CreateScalarIV(Step); - buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, State); - } -} - -void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step, - Instruction *EntryVal, - const InductionDescriptor &ID, - VPValue *Def, - VPTransformState &State) { - IRBuilder<> &Builder = State.Builder; - // We shouldn't have to build scalar steps if we aren't vectorizing. - assert(State.VF.isVector() && "VF should be greater than one"); - // Get the value type and ensure it and the step have the same integer type. - Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); - assert(ScalarIVTy == Step->getType() && - "Val and Step should have the same type"); - - // We build scalar steps for both integer and floating-point induction - // variables. Here, we determine the kind of arithmetic we will perform. - Instruction::BinaryOps AddOp; - Instruction::BinaryOps MulOp; - if (ScalarIVTy->isIntegerTy()) { - AddOp = Instruction::Add; - MulOp = Instruction::Mul; - } else { - AddOp = ID.getInductionOpcode(); - MulOp = Instruction::FMul; - } - - // Determine the number of scalars we need to generate for each unroll - // iteration. - bool FirstLaneOnly = vputils::onlyFirstLaneUsed(Def); - unsigned Lanes = FirstLaneOnly ? 1 : State.VF.getKnownMinValue(); - // Compute the scalar steps and save the results in State. - Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), - ScalarIVTy->getScalarSizeInBits()); - Type *VecIVTy = nullptr; - Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr; - if (!FirstLaneOnly && State.VF.isScalable()) { - VecIVTy = VectorType::get(ScalarIVTy, State.VF); - UnitStepVec = - Builder.CreateStepVector(VectorType::get(IntStepTy, State.VF)); - SplatStep = Builder.CreateVectorSplat(State.VF, Step); - SplatIV = Builder.CreateVectorSplat(State.VF, ScalarIV); + Value *MulExp = B.CreateFMul(Step, Index); + return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, + "induction"); } + case InductionDescriptor::IK_NoInduction: + return nullptr; + } + llvm_unreachable("invalid enum"); +} - for (unsigned Part = 0; Part < State.UF; ++Part) { - Value *StartIdx0 = createStepForVF(Builder, IntStepTy, State.VF, Part); - - if (!FirstLaneOnly && State.VF.isScalable()) { - auto *SplatStartIdx = Builder.CreateVectorSplat(State.VF, StartIdx0); - auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec); - if (ScalarIVTy->isFloatingPointTy()) - InitVec = Builder.CreateSIToFP(InitVec, VecIVTy); - auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep); - auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul); - State.set(Def, Add, Part); - // It's useful to record the lane values too for the known minimum number - // of elements so we do those below. This improves the code quality when - // trying to extract the first element, for example. +std::optional<unsigned> getMaxVScale(const Function &F, + const TargetTransformInfo &TTI) { + if (std::optional<unsigned> MaxVScale = TTI.getMaxVScale()) + return MaxVScale; + + if (F.hasFnAttribute(Attribute::VScaleRange)) + return F.getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); + + return std::nullopt; +} + +/// For the given VF and UF and maximum trip count computed for the loop, return +/// whether the induction variable might overflow in the vectorized loop. If not, +/// then we know a runtime overflow check always evaluates to false and can be +/// removed. +static bool isIndvarOverflowCheckKnownFalse( + const LoopVectorizationCostModel *Cost, + ElementCount VF, std::optional<unsigned> UF = std::nullopt) { + // Always be conservative if we don't know the exact unroll factor. + unsigned MaxUF = UF ? *UF : Cost->TTI.getMaxInterleaveFactor(VF); + + Type *IdxTy = Cost->Legal->getWidestInductionType(); + APInt MaxUIntTripCount = cast<IntegerType>(IdxTy)->getMask(); + + // We know the runtime overflow check is known false iff the (max) trip-count + // is known and (max) trip-count + (VF * UF) does not overflow in the type of + // the vector loop induction variable. + if (unsigned TC = + Cost->PSE.getSE()->getSmallConstantMaxTripCount(Cost->TheLoop)) { + uint64_t MaxVF = VF.getKnownMinValue(); + if (VF.isScalable()) { + std::optional<unsigned> MaxVScale = + getMaxVScale(*Cost->TheFunction, Cost->TTI); + if (!MaxVScale) + return false; + MaxVF *= *MaxVScale; } - if (ScalarIVTy->isFloatingPointTy()) - StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy); - - for (unsigned Lane = 0; Lane < Lanes; ++Lane) { - Value *StartIdx = Builder.CreateBinOp( - AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane)); - // The step returned by `createStepForVF` is a runtime-evaluated value - // when VF is scalable. Otherwise, it should be folded into a Constant. - assert((State.VF.isScalable() || isa<Constant>(StartIdx)) && - "Expected StartIdx to be folded to a constant when VF is not " - "scalable"); - auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); - auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); - State.set(Def, Add, VPIteration(Part, Lane)); - } + return (MaxUIntTripCount - TC).ugt(MaxVF * MaxUF); } -} -void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, - const VPIteration &Instance, - VPTransformState &State) { - Value *ScalarInst = State.get(Def, Instance); - Value *VectorValue = State.get(Def, Instance.Part); - VectorValue = Builder.CreateInsertElement( - VectorValue, ScalarInst, - Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); - State.set(Def, VectorValue, Instance.Part); + return false; } // Return whether we allow using masked interleave-groups (for dealing with @@ -2709,14 +2503,13 @@ static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { void InnerLoopVectorizer::vectorizeInterleaveGroup( const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, - VPValue *BlockInMask) { + VPValue *BlockInMask, bool NeedsMaskForGaps) { Instruction *Instr = Group->getInsertPos(); const DataLayout &DL = Instr->getModule()->getDataLayout(); // Prepare for the vector type of the interleaved load/store. Type *ScalarTy = getLoadStoreType(Instr); unsigned InterleaveFactor = Group->getFactor(); - assert(!VF.isScalable() && "scalable vectors not yet supported."); auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); // Prepare for the new pointers. @@ -2727,18 +2520,26 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( assert((!BlockInMask || !Group->isReverse()) && "Reversed masked interleave-group not supported."); + Value *Idx; // If the group is reverse, adjust the index to refer to the last vector lane // instead of the first. We adjust the index from the first vector lane, // rather than directly getting the pointer for lane VF - 1, because the // pointer operand of the interleaved access is supposed to be uniform. For // uniform instructions, we're only required to generate a value for the // first vector lane in each unroll iteration. - if (Group->isReverse()) - Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); + if (Group->isReverse()) { + Value *RuntimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), VF); + Idx = Builder.CreateSub(RuntimeVF, Builder.getInt32(1)); + Idx = Builder.CreateMul(Idx, Builder.getInt32(Group->getFactor())); + Idx = Builder.CreateAdd(Idx, Builder.getInt32(Index)); + Idx = Builder.CreateNeg(Idx); + } else + Idx = Builder.getInt32(-Index); for (unsigned Part = 0; Part < UF; Part++) { Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); - setDebugLocFromInst(AddrPart); + if (auto *I = dyn_cast<Instruction>(AddrPart)) + State.setDebugLocFrom(I->getDebugLoc()); // Notice current instruction could be any index. Need to adjust the address // to the member of index 0. @@ -2755,26 +2556,50 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( bool InBounds = false; if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) InBounds = gep->isInBounds(); - AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); - cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); - - // Cast to the vector pointer type. - unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); - Type *PtrTy = VecTy->getPointerTo(AddressSpace); - AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); + AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Idx, "", InBounds); + AddrParts.push_back(AddrPart); } - setDebugLocFromInst(Instr); + State.setDebugLocFrom(Instr->getDebugLoc()); Value *PoisonVec = PoisonValue::get(VecTy); - Value *MaskForGaps = nullptr; - if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { - MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); - assert(MaskForGaps && "Mask for Gaps is required but it is null"); - } + auto CreateGroupMask = [this, &BlockInMask, &State, &InterleaveFactor]( + unsigned Part, Value *MaskForGaps) -> Value * { + if (VF.isScalable()) { + assert(!MaskForGaps && "Interleaved groups with gaps are not supported."); + assert(InterleaveFactor == 2 && + "Unsupported deinterleave factor for scalable vectors"); + auto *BlockInMaskPart = State.get(BlockInMask, Part); + SmallVector<Value *, 2> Ops = {BlockInMaskPart, BlockInMaskPart}; + auto *MaskTy = + VectorType::get(Builder.getInt1Ty(), VF.getKnownMinValue() * 2, true); + return Builder.CreateIntrinsic( + MaskTy, Intrinsic::experimental_vector_interleave2, Ops, + /*FMFSource=*/nullptr, "interleaved.mask"); + } + + if (!BlockInMask) + return MaskForGaps; + + Value *BlockInMaskPart = State.get(BlockInMask, Part); + Value *ShuffledMask = Builder.CreateShuffleVector( + BlockInMaskPart, + createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), + "interleaved.mask"); + return MaskForGaps ? Builder.CreateBinOp(Instruction::And, ShuffledMask, + MaskForGaps) + : ShuffledMask; + }; // Vectorize the interleaved load group. if (isa<LoadInst>(Instr)) { + Value *MaskForGaps = nullptr; + if (NeedsMaskForGaps) { + MaskForGaps = + createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); + assert(MaskForGaps && "Mask for Gaps is required but it is null"); + } + // For each unroll part, create a wide load for the group. SmallVector<Value *, 2> NewLoads; for (unsigned Part = 0; Part < UF; Part++) { @@ -2782,18 +2607,7 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( if (BlockInMask || MaskForGaps) { assert(useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed."); - Value *GroupMask = MaskForGaps; - if (BlockInMask) { - Value *BlockInMaskPart = State.get(BlockInMask, Part); - Value *ShuffledMask = Builder.CreateShuffleVector( - BlockInMaskPart, - createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), - "interleaved.mask"); - GroupMask = MaskForGaps - ? Builder.CreateBinOp(Instruction::And, ShuffledMask, - MaskForGaps) - : ShuffledMask; - } + Value *GroupMask = CreateGroupMask(Part, MaskForGaps); NewLoad = Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(), GroupMask, PoisonVec, "wide.masked.vec"); @@ -2805,6 +2619,41 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( NewLoads.push_back(NewLoad); } + if (VecTy->isScalableTy()) { + assert(InterleaveFactor == 2 && + "Unsupported deinterleave factor for scalable vectors"); + + for (unsigned Part = 0; Part < UF; ++Part) { + // Scalable vectors cannot use arbitrary shufflevectors (only splats), + // so must use intrinsics to deinterleave. + Value *DI = Builder.CreateIntrinsic( + Intrinsic::experimental_vector_deinterleave2, VecTy, NewLoads[Part], + /*FMFSource=*/nullptr, "strided.vec"); + unsigned J = 0; + for (unsigned I = 0; I < InterleaveFactor; ++I) { + Instruction *Member = Group->getMember(I); + + if (!Member) + continue; + + Value *StridedVec = Builder.CreateExtractValue(DI, I); + // If this member has different type, cast the result type. + if (Member->getType() != ScalarTy) { + VectorType *OtherVTy = VectorType::get(Member->getType(), VF); + StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); + } + + if (Group->isReverse()) + StridedVec = Builder.CreateVectorReverse(StridedVec, "reverse"); + + State.set(VPDefs[J], StridedVec, Part); + ++J; + } + } + + return; + } + // For each member in the group, shuffle out the appropriate data from the // wide loads. unsigned J = 0; @@ -2842,7 +2691,8 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( auto *SubVT = VectorType::get(ScalarTy, VF); // Vectorize the interleaved store group. - MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); + Value *MaskForGaps = + createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); assert((!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) && "masked interleaved groups are not allowed."); assert((!MaskForGaps || !VF.isScalable()) && @@ -2850,6 +2700,7 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( for (unsigned Part = 0; Part < UF; Part++) { // Collect the stored vector from each member. SmallVector<Value *, 4> StoredVecs; + unsigned StoredIdx = 0; for (unsigned i = 0; i < InterleaveFactor; i++) { assert((Group->getMember(i) || MaskForGaps) && "Fail to get a member from an interleaved store group"); @@ -2862,7 +2713,8 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( continue; } - Value *StoredVec = State.get(StoredValues[i], Part); + Value *StoredVec = State.get(StoredValues[StoredIdx], Part); + ++StoredIdx; if (Group->isReverse()) StoredVec = Builder.CreateVectorReverse(StoredVec, "reverse"); @@ -2875,27 +2727,11 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( StoredVecs.push_back(StoredVec); } - // Concatenate all vectors into a wide vector. - Value *WideVec = concatenateVectors(Builder, StoredVecs); - - // Interleave the elements in the wide vector. - Value *IVec = Builder.CreateShuffleVector( - WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), - "interleaved.vec"); - + // Interleave all the smaller vectors into one wider vector. + Value *IVec = interleaveVectors(Builder, StoredVecs, "interleaved.vec"); Instruction *NewStoreInstr; if (BlockInMask || MaskForGaps) { - Value *GroupMask = MaskForGaps; - if (BlockInMask) { - Value *BlockInMaskPart = State.get(BlockInMask, Part); - Value *ShuffledMask = Builder.CreateShuffleVector( - BlockInMaskPart, - createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), - "interleaved.mask"); - GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And, - ShuffledMask, MaskForGaps) - : ShuffledMask; - } + Value *GroupMask = CreateGroupMask(Part, MaskForGaps); NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part], Group->getAlign(), GroupMask); } else @@ -2906,10 +2742,9 @@ void InnerLoopVectorizer::vectorizeInterleaveGroup( } } -void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, +void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr, VPReplicateRecipe *RepRecipe, const VPIteration &Instance, - bool IfPredicateInstr, VPTransformState &State) { assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); @@ -2919,39 +2754,38 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, if (!Instance.isFirstIteration()) return; - setDebugLocFromInst(Instr); - // Does this instruction return a value ? bool IsVoidRetTy = Instr->getType()->isVoidTy(); Instruction *Cloned = Instr->clone(); - if (!IsVoidRetTy) + if (!IsVoidRetTy) { Cloned->setName(Instr->getName() + ".cloned"); +#if !defined(NDEBUG) + // Verify that VPlan type inference results agree with the type of the + // generated values. + assert(State.TypeAnalysis.inferScalarType(RepRecipe) == Cloned->getType() && + "inferred type and type from generated instructions do not match"); +#endif + } + + RepRecipe->setFlags(Cloned); + + if (auto DL = Instr->getDebugLoc()) + State.setDebugLocFrom(DL); - // If the scalarized instruction contributes to the address computation of a - // widen masked load/store which was in a basic block that needed predication - // and is not predicated after vectorization, we can't propagate - // poison-generating flags (nuw/nsw, exact, inbounds, etc.). The scalarized - // instruction could feed a poison value to the base address of the widen - // load/store. - if (State.MayGeneratePoisonRecipes.contains(RepRecipe)) - Cloned->dropPoisonGeneratingFlags(); - - State.Builder.SetInsertPoint(Builder.GetInsertBlock(), - Builder.GetInsertPoint()); // Replace the operands of the cloned instructions with their scalar // equivalents in the new loop. - for (auto &I : enumerate(RepRecipe->operands())) { + for (const auto &I : enumerate(RepRecipe->operands())) { auto InputInstance = Instance; VPValue *Operand = I.value(); - if (State.Plan->isUniformAfterVectorization(Operand)) + if (vputils::isUniformAfterVectorization(Operand)) InputInstance.Lane = VPLane::getFirstLane(); Cloned->setOperand(I.index(), State.get(Operand, InputInstance)); } - addNewMetadata(Cloned, Instr); + State.addNewMetadata(Cloned, Instr); // Place the cloned scalar in the new loop. - Builder.Insert(Cloned); + State.Builder.Insert(Cloned); State.set(RepRecipe, Cloned, Instance); @@ -2960,80 +2794,18 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, AC->registerAssumption(II); // End if-block. + bool IfPredicateInstr = RepRecipe->getParent()->getParent()->isReplicator(); if (IfPredicateInstr) PredicatedInstructions.push_back(Cloned); } -void InnerLoopVectorizer::createHeaderBranch(Loop *L) { - BasicBlock *Header = L->getHeader(); - assert(!L->getLoopLatch() && "loop should not have a latch at this point"); - - IRBuilder<> B(Header->getTerminator()); - Instruction *OldInst = - getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); - setDebugLocFromInst(OldInst, &B); - - // Connect the header to the exit and header blocks and replace the old - // terminator. - B.CreateCondBr(B.getTrue(), L->getUniqueExitBlock(), Header); - - // Now we have two terminators. Remove the old one from the block. - Header->getTerminator()->eraseFromParent(); -} - -Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { - if (TripCount) - return TripCount; - - assert(L && "Create Trip Count for null loop."); - IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); - // Find the loop boundaries. - ScalarEvolution *SE = PSE.getSE(); - const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); - assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && - "Invalid loop count"); - - Type *IdxTy = Legal->getWidestInductionType(); - assert(IdxTy && "No type for induction"); - - // The exit count might have the type of i64 while the phi is i32. This can - // happen if we have an induction variable that is sign extended before the - // compare. The only way that we get a backedge taken count is that the - // induction variable was signed and as such will not overflow. In such a case - // truncation is legal. - if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) > - IdxTy->getPrimitiveSizeInBits()) - BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); - BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); - - // Get the total trip count from the count by adding 1. - const SCEV *ExitCount = SE->getAddExpr( - BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); - - const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); - - // Expand the trip count and place the new instructions in the preheader. - // Notice that the pre-header does not change, only the loop body. - SCEVExpander Exp(*SE, DL, "induction"); - - // Count holds the overall loop count (N). - TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), - L->getLoopPreheader()->getTerminator()); - - if (TripCount->getType()->isPointerTy()) - TripCount = - CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", - L->getLoopPreheader()->getTerminator()); - - return TripCount; -} - -Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { +Value * +InnerLoopVectorizer::getOrCreateVectorTripCount(BasicBlock *InsertBlock) { if (VectorTripCount) return VectorTripCount; - Value *TC = getOrCreateTripCount(L); - IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); + Value *TC = getTripCount(); + IRBuilder<> Builder(InsertBlock->getTerminator()); Type *Ty = TC->getType(); // This is where we can make the step a runtime constant. @@ -3045,6 +2817,8 @@ Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { // overflows: the vector induction variable will eventually wrap to zero given // that it starts at zero and its Step is a power of two; the loop will then // exit, with the last early-exit vector comparison also producing all-true. + // For scalable vectors the VF is not guaranteed to be a power of 2, but this + // is accounted for in emitIterationCountCheck that adds an overflow check. if (Cost->foldTailByMasking()) { assert(isPowerOf2_32(VF.getKnownMinValue() * UF) && "VF*UF must be a power of 2 when folding tail by masking"); @@ -3066,7 +2840,7 @@ Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { // the step does not evenly divide the trip count, no adjustment is necessary // since there will already be scalar iterations. Note that the minimum // iterations check ensures that N >= Step. - if (Cost->requiresScalarEpilogue(VF)) { + if (Cost->requiresScalarEpilogue(VF.isVector())) { auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); R = Builder.CreateSelect(IsZero, Step, R); } @@ -3079,10 +2853,10 @@ Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, const DataLayout &DL) { // Verify that V is a vector type with same number of elements as DstVTy. - auto *DstFVTy = cast<FixedVectorType>(DstVTy); - unsigned VF = DstFVTy->getNumElements(); - auto *SrcVecTy = cast<FixedVectorType>(V->getType()); - assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match"); + auto *DstFVTy = cast<VectorType>(DstVTy); + auto VF = DstFVTy->getElementCount(); + auto *SrcVecTy = cast<VectorType>(V->getType()); + assert(VF == SrcVecTy->getElementCount() && "Vector dimensions do not match"); Type *SrcElemTy = SrcVecTy->getElementType(); Type *DstElemTy = DstFVTy->getElementType(); assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && @@ -3102,14 +2876,13 @@ Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, "Only one type should be a floating point type"); Type *IntTy = IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); - auto *VecIntTy = FixedVectorType::get(IntTy, VF); + auto *VecIntTy = VectorType::get(IntTy, VF); Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); } -void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, - BasicBlock *Bypass) { - Value *Count = getOrCreateTripCount(L); +void InnerLoopVectorizer::emitIterationCountCheck(BasicBlock *Bypass) { + Value *Count = getTripCount(); // Reuse existing vector loop preheader for TC checks. // Note that new preheader block is generated for vector loop. BasicBlock *const TCCheckBlock = LoopVectorPreHeader; @@ -3120,15 +2893,45 @@ void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, // vector trip count is zero. This check also covers the case where adding one // to the backedge-taken count overflowed leading to an incorrect trip count // of zero. In this case we will also jump to the scalar loop. - auto P = Cost->requiresScalarEpilogue(VF) ? ICmpInst::ICMP_ULE - : ICmpInst::ICMP_ULT; + auto P = Cost->requiresScalarEpilogue(VF.isVector()) ? ICmpInst::ICMP_ULE + : ICmpInst::ICMP_ULT; // If tail is to be folded, vector loop takes care of all iterations. + Type *CountTy = Count->getType(); Value *CheckMinIters = Builder.getFalse(); - if (!Cost->foldTailByMasking()) { - Value *Step = createStepForVF(Builder, Count->getType(), VF, UF); - CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check"); + auto CreateStep = [&]() -> Value * { + // Create step with max(MinProTripCount, UF * VF). + if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue()) + return createStepForVF(Builder, CountTy, VF, UF); + + Value *MinProfTC = + createStepForVF(Builder, CountTy, MinProfitableTripCount, 1); + if (!VF.isScalable()) + return MinProfTC; + return Builder.CreateBinaryIntrinsic( + Intrinsic::umax, MinProfTC, createStepForVF(Builder, CountTy, VF, UF)); + }; + + TailFoldingStyle Style = Cost->getTailFoldingStyle(); + if (Style == TailFoldingStyle::None) + CheckMinIters = + Builder.CreateICmp(P, Count, CreateStep(), "min.iters.check"); + else if (VF.isScalable() && + !isIndvarOverflowCheckKnownFalse(Cost, VF, UF) && + Style != TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck) { + // vscale is not necessarily a power-of-2, which means we cannot guarantee + // an overflow to zero when updating induction variables and so an + // additional overflow check is required before entering the vector loop. + + // Get the maximum unsigned value for the type. + Value *MaxUIntTripCount = + ConstantInt::get(CountTy, cast<IntegerType>(CountTy)->getMask()); + Value *LHS = Builder.CreateSub(MaxUIntTripCount, Count); + + // Don't execute the vector loop if (UMax - n) < (VF * UF). + CheckMinIters = Builder.CreateICmp(ICmpInst::ICMP_ULT, LHS, CreateStep()); } + // Create new preheader for vector loop. LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, @@ -3140,22 +2943,23 @@ void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, // Update dominator for Bypass & LoopExit (if needed). DT->changeImmediateDominator(Bypass, TCCheckBlock); - if (!Cost->requiresScalarEpilogue(VF)) + if (!Cost->requiresScalarEpilogue(VF.isVector())) // If there is an epilogue which must run, there's no edge from the // middle block to exit blocks and thus no need to update the immediate // dominator of the exit blocks. DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); - ReplaceInstWithInst( - TCCheckBlock->getTerminator(), - BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); + BranchInst &BI = + *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); + if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) + setBranchWeights(BI, MinItersBypassWeights); + ReplaceInstWithInst(TCCheckBlock->getTerminator(), &BI); LoopBypassBlocks.push_back(TCCheckBlock); } -BasicBlock *InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { - +BasicBlock *InnerLoopVectorizer::emitSCEVChecks(BasicBlock *Bypass) { BasicBlock *const SCEVCheckBlock = - RTChecks.emitSCEVChecks(L, Bypass, LoopVectorPreHeader, LoopExitBlock); + RTChecks.emitSCEVChecks(Bypass, LoopVectorPreHeader, LoopExitBlock); if (!SCEVCheckBlock) return nullptr; @@ -3168,7 +2972,7 @@ BasicBlock *InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { // Update dominator only if this is first RT check. if (LoopBypassBlocks.empty()) { DT->changeImmediateDominator(Bypass, SCEVCheckBlock); - if (!Cost->requiresScalarEpilogue(VF)) + if (!Cost->requiresScalarEpilogue(VF.isVector())) // If there is an epilogue which must run, there's no edge from the // middle block to exit blocks and thus no need to update the immediate // dominator of the exit blocks. @@ -3180,14 +2984,13 @@ BasicBlock *InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { return SCEVCheckBlock; } -BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, - BasicBlock *Bypass) { +BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(BasicBlock *Bypass) { // VPlan-native path does not do any analysis for runtime checks currently. if (EnableVPlanNativePath) return nullptr; BasicBlock *const MemCheckBlock = - RTChecks.emitMemRuntimeChecks(L, Bypass, LoopVectorPreHeader); + RTChecks.emitMemRuntimeChecks(Bypass, LoopVectorPreHeader); // Check if we generated code that checks in runtime if arrays overlap. We put // the checks into a separate block to make the more common case of few @@ -3201,7 +3004,8 @@ BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, "to vectorize."); ORE->emit([&]() { return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize", - L->getStartLoc(), L->getHeader()) + OrigLoop->getStartLoc(), + OrigLoop->getHeader()) << "Code-size may be reduced by not forcing " "vectorization, or by source-code modifications " "eliminating the need for runtime checks " @@ -3213,121 +3017,15 @@ BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, AddedSafetyChecks = true; - // We currently don't use LoopVersioning for the actual loop cloning but we - // still use it to add the noalias metadata. - LVer = std::make_unique<LoopVersioning>( - *Legal->getLAI(), - Legal->getLAI()->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, - DT, PSE.getSE()); - LVer->prepareNoAliasMetadata(); return MemCheckBlock; } -Value *InnerLoopVectorizer::emitTransformedIndex( - IRBuilder<> &B, Value *Index, ScalarEvolution *SE, const DataLayout &DL, - const InductionDescriptor &ID, BasicBlock *VectorHeader) const { - - SCEVExpander Exp(*SE, DL, "induction"); - auto Step = ID.getStep(); - auto StartValue = ID.getStartValue(); - assert(Index->getType()->getScalarType() == Step->getType() && - "Index scalar type does not match StepValue type"); - - // Note: the IR at this point is broken. We cannot use SE to create any new - // SCEV and then expand it, hoping that SCEV's simplification will give us - // a more optimal code. Unfortunately, attempt of doing so on invalid IR may - // lead to various SCEV crashes. So all we can do is to use builder and rely - // on InstCombine for future simplifications. Here we handle some trivial - // cases only. - auto CreateAdd = [&B](Value *X, Value *Y) { - assert(X->getType() == Y->getType() && "Types don't match!"); - if (auto *CX = dyn_cast<ConstantInt>(X)) - if (CX->isZero()) - return Y; - if (auto *CY = dyn_cast<ConstantInt>(Y)) - if (CY->isZero()) - return X; - return B.CreateAdd(X, Y); - }; - - // We allow X to be a vector type, in which case Y will potentially be - // splatted into a vector with the same element count. - auto CreateMul = [&B](Value *X, Value *Y) { - assert(X->getType()->getScalarType() == Y->getType() && - "Types don't match!"); - if (auto *CX = dyn_cast<ConstantInt>(X)) - if (CX->isOne()) - return Y; - if (auto *CY = dyn_cast<ConstantInt>(Y)) - if (CY->isOne()) - return X; - VectorType *XVTy = dyn_cast<VectorType>(X->getType()); - if (XVTy && !isa<VectorType>(Y->getType())) - Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); - return B.CreateMul(X, Y); - }; - - // Get a suitable insert point for SCEV expansion. For blocks in the vector - // loop, choose the end of the vector loop header (=VectorHeader), because - // the DomTree is not kept up-to-date for additional blocks generated in the - // vector loop. By using the header as insertion point, we guarantee that the - // expanded instructions dominate all their uses. - auto GetInsertPoint = [this, &B, VectorHeader]() { - BasicBlock *InsertBB = B.GetInsertPoint()->getParent(); - if (InsertBB != LoopVectorBody && - LI->getLoopFor(VectorHeader) == LI->getLoopFor(InsertBB)) - return VectorHeader->getTerminator(); - return &*B.GetInsertPoint(); - }; - - switch (ID.getKind()) { - case InductionDescriptor::IK_IntInduction: { - assert(!isa<VectorType>(Index->getType()) && - "Vector indices not supported for integer inductions yet"); - assert(Index->getType() == StartValue->getType() && - "Index type does not match StartValue type"); - if (ID.getConstIntStepValue() && ID.getConstIntStepValue()->isMinusOne()) - return B.CreateSub(StartValue, Index); - auto *Offset = CreateMul( - Index, Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint())); - return CreateAdd(StartValue, Offset); - } - case InductionDescriptor::IK_PtrInduction: { - assert(isa<SCEVConstant>(Step) && - "Expected constant step for pointer induction"); - return B.CreateGEP( - ID.getElementType(), StartValue, - CreateMul(Index, - Exp.expandCodeFor(Step, Index->getType()->getScalarType(), - GetInsertPoint()))); - } - case InductionDescriptor::IK_FpInduction: { - assert(!isa<VectorType>(Index->getType()) && - "Vector indices not supported for FP inductions yet"); - assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value"); - auto InductionBinOp = ID.getInductionBinOp(); - assert(InductionBinOp && - (InductionBinOp->getOpcode() == Instruction::FAdd || - InductionBinOp->getOpcode() == Instruction::FSub) && - "Original bin op should be defined for FP induction"); - - Value *StepValue = cast<SCEVUnknown>(Step)->getValue(); - Value *MulExp = B.CreateFMul(StepValue, Index); - return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, - "induction"); - } - case InductionDescriptor::IK_NoInduction: - return nullptr; - } - llvm_unreachable("invalid enum"); -} - -Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { +void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { LoopScalarBody = OrigLoop->getHeader(); LoopVectorPreHeader = OrigLoop->getLoopPreheader(); assert(LoopVectorPreHeader && "Invalid loop structure"); LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr - assert((LoopExitBlock || Cost->requiresScalarEpilogue(VF)) && + assert((LoopExitBlock || Cost->requiresScalarEpilogue(VF.isVector())) && "multiple exit loop without required epilogue?"); LoopMiddleBlock = @@ -3343,54 +3041,105 @@ Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { // 1) If we know that we must execute the scalar epilogue, emit an // unconditional branch. // 2) Otherwise, we must have a single unique exit block (due to how we - // implement the multiple exit case). In this case, set up a conditonal + // implement the multiple exit case). In this case, set up a conditional // branch from the middle block to the loop scalar preheader, and the // exit block. completeLoopSkeleton will update the condition to use an // iteration check, if required to decide whether to execute the remainder. - BranchInst *BrInst = Cost->requiresScalarEpilogue(VF) ? - BranchInst::Create(LoopScalarPreHeader) : - BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, - Builder.getTrue()); + BranchInst *BrInst = + Cost->requiresScalarEpilogue(VF.isVector()) + ? BranchInst::Create(LoopScalarPreHeader) + : BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, + Builder.getTrue()); BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); - // We intentionally don't let SplitBlock to update LoopInfo since - // LoopVectorBody should belong to another loop than LoopVectorPreHeader. - // LoopVectorBody is explicitly added to the correct place few lines later. - LoopVectorBody = - SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, - nullptr, nullptr, Twine(Prefix) + "vector.body"); - - // Update dominator for loop exit. - if (!Cost->requiresScalarEpilogue(VF)) + // Update dominator for loop exit. During skeleton creation, only the vector + // pre-header and the middle block are created. The vector loop is entirely + // created during VPlan exection. + if (!Cost->requiresScalarEpilogue(VF.isVector())) // If there is an epilogue which must run, there's no edge from the // middle block to exit blocks and thus no need to update the immediate // dominator of the exit blocks. DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); +} - // Create and register the new vector loop. - Loop *Lp = LI->AllocateLoop(); - Loop *ParentLoop = OrigLoop->getParentLoop(); +PHINode *InnerLoopVectorizer::createInductionResumeValue( + PHINode *OrigPhi, const InductionDescriptor &II, Value *Step, + ArrayRef<BasicBlock *> BypassBlocks, + std::pair<BasicBlock *, Value *> AdditionalBypass) { + Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); + assert(VectorTripCount && "Expected valid arguments"); - // Insert the new loop into the loop nest and register the new basic blocks - // before calling any utilities such as SCEV that require valid LoopInfo. - if (ParentLoop) { - ParentLoop->addChildLoop(Lp); + Instruction *OldInduction = Legal->getPrimaryInduction(); + Value *&EndValue = IVEndValues[OrigPhi]; + Value *EndValueFromAdditionalBypass = AdditionalBypass.second; + if (OrigPhi == OldInduction) { + // We know what the end value is. + EndValue = VectorTripCount; } else { - LI->addTopLevelLoop(Lp); + IRBuilder<> B(LoopVectorPreHeader->getTerminator()); + + // Fast-math-flags propagate from the original induction instruction. + if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) + B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); + + EndValue = emitTransformedIndex(B, VectorTripCount, II.getStartValue(), + Step, II.getKind(), II.getInductionBinOp()); + EndValue->setName("ind.end"); + + // Compute the end value for the additional bypass (if applicable). + if (AdditionalBypass.first) { + B.SetInsertPoint(AdditionalBypass.first, + AdditionalBypass.first->getFirstInsertionPt()); + EndValueFromAdditionalBypass = + emitTransformedIndex(B, AdditionalBypass.second, II.getStartValue(), + Step, II.getKind(), II.getInductionBinOp()); + EndValueFromAdditionalBypass->setName("ind.end"); + } } - Lp->addBasicBlockToLoop(LoopVectorBody, *LI); - return Lp; + + // Create phi nodes to merge from the backedge-taken check block. + PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", + LoopScalarPreHeader->getTerminator()); + // Copy original phi DL over to the new one. + BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); + + // The new PHI merges the original incoming value, in case of a bypass, + // or the value at the end of the vectorized loop. + BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); + + // Fix the scalar body counter (PHI node). + // The old induction's phi node in the scalar body needs the truncated + // value. + for (BasicBlock *BB : BypassBlocks) + BCResumeVal->addIncoming(II.getStartValue(), BB); + + if (AdditionalBypass.first) + BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, + EndValueFromAdditionalBypass); + return BCResumeVal; +} + +/// Return the expanded step for \p ID using \p ExpandedSCEVs to look up SCEV +/// expansion results. +static Value *getExpandedStep(const InductionDescriptor &ID, + const SCEV2ValueTy &ExpandedSCEVs) { + const SCEV *Step = ID.getStep(); + if (auto *C = dyn_cast<SCEVConstant>(Step)) + return C->getValue(); + if (auto *U = dyn_cast<SCEVUnknown>(Step)) + return U->getValue(); + auto I = ExpandedSCEVs.find(Step); + assert(I != ExpandedSCEVs.end() && "SCEV must be expanded at this point"); + return I->second; } void InnerLoopVectorizer::createInductionResumeValues( - Loop *L, std::pair<BasicBlock *, Value *> AdditionalBypass) { + const SCEV2ValueTy &ExpandedSCEVs, + std::pair<BasicBlock *, Value *> AdditionalBypass) { assert(((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass."); - - Value *VectorTripCount = getOrCreateVectorTripCount(L); - assert(VectorTripCount && L && "Expected valid arguments"); // We are going to resume the execution of the scalar loop. // Go over all of the induction variables that we found and fix the // PHIs that are left in the scalar version of the loop. @@ -3398,75 +3147,20 @@ void InnerLoopVectorizer::createInductionResumeValues( // iteration in the vectorized loop. // If we come from a bypass edge then we need to start from the original // start value. - Instruction *OldInduction = Legal->getPrimaryInduction(); - for (auto &InductionEntry : Legal->getInductionVars()) { + for (const auto &InductionEntry : Legal->getInductionVars()) { PHINode *OrigPhi = InductionEntry.first; - InductionDescriptor II = InductionEntry.second; - - // Create phi nodes to merge from the backedge-taken check block. - PHINode *BCResumeVal = - PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", - LoopScalarPreHeader->getTerminator()); - // Copy original phi DL over to the new one. - BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); - Value *&EndValue = IVEndValues[OrigPhi]; - Value *EndValueFromAdditionalBypass = AdditionalBypass.second; - if (OrigPhi == OldInduction) { - // We know what the end value is. - EndValue = VectorTripCount; - } else { - IRBuilder<> B(L->getLoopPreheader()->getTerminator()); - - // Fast-math-flags propagate from the original induction instruction. - if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) - B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); - - Type *StepType = II.getStep()->getType(); - Instruction::CastOps CastOp = - CastInst::getCastOpcode(VectorTripCount, true, StepType, true); - Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd"); - const DataLayout &DL = LoopScalarBody->getModule()->getDataLayout(); - EndValue = - emitTransformedIndex(B, CRD, PSE.getSE(), DL, II, LoopVectorBody); - EndValue->setName("ind.end"); - - // Compute the end value for the additional bypass (if applicable). - if (AdditionalBypass.first) { - B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); - CastOp = CastInst::getCastOpcode(AdditionalBypass.second, true, - StepType, true); - CRD = - B.CreateCast(CastOp, AdditionalBypass.second, StepType, "cast.crd"); - EndValueFromAdditionalBypass = - emitTransformedIndex(B, CRD, PSE.getSE(), DL, II, LoopVectorBody); - EndValueFromAdditionalBypass->setName("ind.end"); - } - } - // The new PHI merges the original incoming value, in case of a bypass, - // or the value at the end of the vectorized loop. - BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); - - // Fix the scalar body counter (PHI node). - // The old induction's phi node in the scalar body needs the truncated - // value. - for (BasicBlock *BB : LoopBypassBlocks) - BCResumeVal->addIncoming(II.getStartValue(), BB); - - if (AdditionalBypass.first) - BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, - EndValueFromAdditionalBypass); - + const InductionDescriptor &II = InductionEntry.second; + PHINode *BCResumeVal = createInductionResumeValue( + OrigPhi, II, getExpandedStep(II, ExpandedSCEVs), LoopBypassBlocks, + AdditionalBypass); OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); } } -BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, - MDNode *OrigLoopID) { - assert(L && "Expected valid loop."); - +BasicBlock *InnerLoopVectorizer::completeLoopSkeleton() { // The trip counts should be cached by now. - Value *Count = getOrCreateTripCount(L); - Value *VectorTripCount = getOrCreateVectorTripCount(L); + Value *Count = getTripCount(); + Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); @@ -3478,41 +3172,47 @@ BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, // Thus if tail is to be folded, we know we don't need to run the // remainder and we can use the previous value for the condition (true). // 3) Otherwise, construct a runtime check. - if (!Cost->requiresScalarEpilogue(VF) && !Cost->foldTailByMasking()) { - Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, - Count, VectorTripCount, "cmp.n", - LoopMiddleBlock->getTerminator()); - + if (!Cost->requiresScalarEpilogue(VF.isVector()) && + !Cost->foldTailByMasking()) { // Here we use the same DebugLoc as the scalar loop latch terminator instead // of the corresponding compare because they may have ended up with // different line numbers and we want to avoid awkward line stepping while // debugging. Eg. if the compare has got a line number inside the loop. - CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); - cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); + // TODO: At the moment, CreateICmpEQ will simplify conditions with constant + // operands. Perform simplification directly on VPlan once the branch is + // modeled there. + IRBuilder<> B(LoopMiddleBlock->getTerminator()); + B.SetCurrentDebugLocation(ScalarLatchTerm->getDebugLoc()); + Value *CmpN = B.CreateICmpEQ(Count, VectorTripCount, "cmp.n"); + BranchInst &BI = *cast<BranchInst>(LoopMiddleBlock->getTerminator()); + BI.setCondition(CmpN); + if (hasBranchWeightMD(*ScalarLatchTerm)) { + // Assume that `Count % VectorTripCount` is equally distributed. + unsigned TripCount = UF * VF.getKnownMinValue(); + assert(TripCount > 0 && "trip count should not be zero"); + const uint32_t Weights[] = {1, TripCount - 1}; + setBranchWeights(BI, Weights); + } } - // Get ready to start creating new instructions into the vectorized body. - assert(LoopVectorPreHeader == L->getLoopPreheader() && - "Inconsistent vector loop preheader"); - Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt()); - #ifdef EXPENSIVE_CHECKS assert(DT->verify(DominatorTree::VerificationLevel::Fast)); - LI->verify(*DT); #endif return LoopVectorPreHeader; } std::pair<BasicBlock *, Value *> -InnerLoopVectorizer::createVectorizedLoopSkeleton() { +InnerLoopVectorizer::createVectorizedLoopSkeleton( + const SCEV2ValueTy &ExpandedSCEVs) { /* In this function we generate a new loop. The new loop will contain the vectorized instructions while the old loop will continue to run the scalar remainder. - [ ] <-- loop iteration number check. - / | + [ ] <-- old preheader - loop iteration number check and SCEVs in Plan's + / | preheader are expanded here. Eventually all required SCEV + / | expansion should happen here. / v | [ ] <-- vector loop bypass (may consist of multiple blocks). | / | @@ -3521,7 +3221,7 @@ InnerLoopVectorizer::createVectorizedLoopSkeleton() { |/ | | v | [ ] \ - | [ ]_| <-- vector loop. + | [ ]_| <-- vector loop (created during VPlan execution). | | | v \ -[ ] <--- middle-block. @@ -3538,44 +3238,30 @@ InnerLoopVectorizer::createVectorizedLoopSkeleton() { ... */ - // Get the metadata of the original loop before it gets modified. - MDNode *OrigLoopID = OrigLoop->getLoopID(); - - // Workaround! Compute the trip count of the original loop and cache it - // before we start modifying the CFG. This code has a systemic problem - // wherein it tries to run analysis over partially constructed IR; this is - // wrong, and not simply for SCEV. The trip count of the original loop - // simply happens to be prone to hitting this in practice. In theory, we - // can hit the same issue for any SCEV, or ValueTracking query done during - // mutation. See PR49900. - getOrCreateTripCount(OrigLoop); - // Create an empty vector loop, and prepare basic blocks for the runtime // checks. - Loop *Lp = createVectorLoopSkeleton(""); + createVectorLoopSkeleton(""); // Now, compare the new count to zero. If it is zero skip the vector loop and // jump to the scalar loop. This check also covers the case where the // backedge-taken count is uint##_max: adding one to it will overflow leading // to an incorrect trip count of zero. In this (rare) case we will also jump // to the scalar loop. - emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader); + emitIterationCountCheck(LoopScalarPreHeader); // Generate the code to check any assumptions that we've made for SCEV // expressions. - emitSCEVChecks(Lp, LoopScalarPreHeader); + emitSCEVChecks(LoopScalarPreHeader); // Generate the code that checks in runtime if arrays overlap. We put the // checks into a separate block to make the more common case of few elements // faster. - emitMemRuntimeChecks(Lp, LoopScalarPreHeader); - - createHeaderBranch(Lp); + emitMemRuntimeChecks(LoopScalarPreHeader); // Emit phis for the new starting index of the scalar loop. - createInductionResumeValues(Lp); + createInductionResumeValues(ExpandedSCEVs); - return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; + return {completeLoopSkeleton(), nullptr}; } // Fix up external users of the induction variable. At this point, we are @@ -3584,8 +3270,10 @@ InnerLoopVectorizer::createVectorizedLoopSkeleton() { // value for the IV when arriving directly from the middle block. void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, - Value *CountRoundDown, Value *EndValue, - BasicBlock *MiddleBlock) { + Value *VectorTripCount, Value *EndValue, + BasicBlock *MiddleBlock, + BasicBlock *VectorHeader, VPlan &Plan, + VPTransformState &State) { // There are two kinds of external IV usages - those that use the value // computed in the last iteration (the PHI) and those that use the penultimate // value (the value that feeds into the phi from the loop latch). @@ -3612,10 +3300,7 @@ void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, for (User *U : OrigPhi->users()) { auto *UI = cast<Instruction>(U); if (!OrigLoop->contains(UI)) { - const DataLayout &DL = - OrigLoop->getHeader()->getModule()->getDataLayout(); assert(isa<PHINode>(UI) && "Expected LCSSA form"); - IRBuilder<> B(MiddleBlock->getTerminator()); // Fast-math-flags propagate from the original induction instruction. @@ -3623,15 +3308,16 @@ void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); Value *CountMinusOne = B.CreateSub( - CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1)); - Value *CMO = - !II.getStep()->getType()->isIntegerTy() - ? B.CreateCast(Instruction::SIToFP, CountMinusOne, - II.getStep()->getType()) - : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType()); - CMO->setName("cast.cmo"); + VectorTripCount, ConstantInt::get(VectorTripCount->getType(), 1)); + CountMinusOne->setName("cmo"); + + VPValue *StepVPV = Plan.getSCEVExpansion(II.getStep()); + assert(StepVPV && "step must have been expanded during VPlan execution"); + Value *Step = StepVPV->isLiveIn() ? StepVPV->getLiveInIRValue() + : State.get(StepVPV, {0, 0}); Value *Escape = - emitTransformedIndex(B, CMO, PSE.getSE(), DL, II, LoopVectorBody); + emitTransformedIndex(B, CountMinusOne, II.getStartValue(), Step, + II.getKind(), II.getInductionBinOp()); Escape->setName("ind.escape"); MissingVals[UI] = Escape; } @@ -3644,8 +3330,10 @@ void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, // In this case, if IV1 has an external use, we need to avoid adding both // "last value of IV1" and "penultimate value of IV2". So, verify that we // don't already have an incoming value for the middle block. - if (PHI->getBasicBlockIndex(MiddleBlock) == -1) + if (PHI->getBasicBlockIndex(MiddleBlock) == -1) { PHI->addIncoming(I.second, MiddleBlock); + Plan.removeLiveOut(PHI); + } } } @@ -3702,52 +3390,32 @@ static void cse(BasicBlock *BB) { } InstructionCost -LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, - bool &NeedToScalarize) const { - Function *F = CI->getCalledFunction(); - Type *ScalarRetTy = CI->getType(); - SmallVector<Type *, 4> Tys, ScalarTys; - for (auto &ArgOp : CI->args()) - ScalarTys.push_back(ArgOp->getType()); +LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, + ElementCount VF) const { + // We only need to calculate a cost if the VF is scalar; for actual vectors + // we should already have a pre-calculated cost at each VF. + if (!VF.isScalar()) + return CallWideningDecisions.at(std::make_pair(CI, VF)).Cost; - // Estimate cost of scalarized vector call. The source operands are assumed - // to be vectors, so we need to extract individual elements from there, - // execute VF scalar calls, and then gather the result into the vector return - // value. - InstructionCost ScalarCallCost = - TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput); - if (VF.isScalar()) - return ScalarCallCost; - - // Compute corresponding vector type for return value and arguments. - Type *RetTy = ToVectorTy(ScalarRetTy, VF); - for (Type *ScalarTy : ScalarTys) - Tys.push_back(ToVectorTy(ScalarTy, VF)); - - // Compute costs of unpacking argument values for the scalar calls and - // packing the return values to a vector. - InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); - - InstructionCost Cost = - ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; + TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; + Type *RetTy = CI->getType(); + if (RecurrenceDescriptor::isFMulAddIntrinsic(CI)) + if (auto RedCost = getReductionPatternCost(CI, VF, RetTy, CostKind)) + return *RedCost; - // If we can't emit a vector call for this function, then the currently found - // cost is the cost we need to return. - NeedToScalarize = true; - VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); - Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); + SmallVector<Type *, 4> Tys; + for (auto &ArgOp : CI->args()) + Tys.push_back(ArgOp->getType()); - if (!TLI || CI->isNoBuiltin() || !VecFunc) - return Cost; + InstructionCost ScalarCallCost = + TTI.getCallInstrCost(CI->getCalledFunction(), RetTy, Tys, CostKind); - // If the corresponding vector cost is cheaper, return its cost. - InstructionCost VectorCallCost = - TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput); - if (VectorCallCost < Cost) { - NeedToScalarize = false; - Cost = VectorCallCost; + // If this is an intrinsic we may have a lower cost for it. + if (getVectorIntrinsicIDForCall(CI, TLI)) { + InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); + return std::min(ScalarCallCost, IntrinsicCost); } - return Cost; + return ScalarCallCost; } static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { @@ -3791,179 +3459,72 @@ static Type *largestIntegerVectorType(Type *T1, Type *T2) { return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; } -void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { - // For every instruction `I` in MinBWs, truncate the operands, create a - // truncated version of `I` and reextend its result. InstCombine runs - // later and will remove any ext/trunc pairs. - SmallPtrSet<Value *, 4> Erased; - for (const auto &KV : Cost->getMinimalBitwidths()) { - // If the value wasn't vectorized, we must maintain the original scalar - // type. The absence of the value from State indicates that it - // wasn't vectorized. - // FIXME: Should not rely on getVPValue at this point. - VPValue *Def = State.Plan->getVPValue(KV.first, true); - if (!State.hasAnyVectorValue(Def)) - continue; - for (unsigned Part = 0; Part < UF; ++Part) { - Value *I = State.get(Def, Part); - if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) - continue; - Type *OriginalTy = I->getType(); - Type *ScalarTruncatedTy = - IntegerType::get(OriginalTy->getContext(), KV.second); - auto *TruncatedTy = VectorType::get( - ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount()); - if (TruncatedTy == OriginalTy) - continue; - - IRBuilder<> B(cast<Instruction>(I)); - auto ShrinkOperand = [&](Value *V) -> Value * { - if (auto *ZI = dyn_cast<ZExtInst>(V)) - if (ZI->getSrcTy() == TruncatedTy) - return ZI->getOperand(0); - return B.CreateZExtOrTrunc(V, TruncatedTy); - }; - - // The actual instruction modification depends on the instruction type, - // unfortunately. - Value *NewI = nullptr; - if (auto *BO = dyn_cast<BinaryOperator>(I)) { - NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), - ShrinkOperand(BO->getOperand(1))); - - // Any wrapping introduced by shrinking this operation shouldn't be - // considered undefined behavior. So, we can't unconditionally copy - // arithmetic wrapping flags to NewI. - cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); - } else if (auto *CI = dyn_cast<ICmpInst>(I)) { - NewI = - B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), - ShrinkOperand(CI->getOperand(1))); - } else if (auto *SI = dyn_cast<SelectInst>(I)) { - NewI = B.CreateSelect(SI->getCondition(), - ShrinkOperand(SI->getTrueValue()), - ShrinkOperand(SI->getFalseValue())); - } else if (auto *CI = dyn_cast<CastInst>(I)) { - switch (CI->getOpcode()) { - default: - llvm_unreachable("Unhandled cast!"); - case Instruction::Trunc: - NewI = ShrinkOperand(CI->getOperand(0)); - break; - case Instruction::SExt: - NewI = B.CreateSExtOrTrunc( - CI->getOperand(0), - smallestIntegerVectorType(OriginalTy, TruncatedTy)); - break; - case Instruction::ZExt: - NewI = B.CreateZExtOrTrunc( - CI->getOperand(0), - smallestIntegerVectorType(OriginalTy, TruncatedTy)); - break; - } - } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { - auto Elements0 = - cast<VectorType>(SI->getOperand(0)->getType())->getElementCount(); - auto *O0 = B.CreateZExtOrTrunc( - SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0)); - auto Elements1 = - cast<VectorType>(SI->getOperand(1)->getType())->getElementCount(); - auto *O1 = B.CreateZExtOrTrunc( - SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1)); - - NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); - } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { - // Don't do anything with the operands, just extend the result. - continue; - } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { - auto Elements = - cast<VectorType>(IE->getOperand(0)->getType())->getElementCount(); - auto *O0 = B.CreateZExtOrTrunc( - IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); - auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); - NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); - } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { - auto Elements = - cast<VectorType>(EE->getOperand(0)->getType())->getElementCount(); - auto *O0 = B.CreateZExtOrTrunc( - EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); - NewI = B.CreateExtractElement(O0, EE->getOperand(2)); - } else { - // If we don't know what to do, be conservative and don't do anything. - continue; - } - - // Lastly, extend the result. - NewI->takeName(cast<Instruction>(I)); - Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); - I->replaceAllUsesWith(Res); - cast<Instruction>(I)->eraseFromParent(); - Erased.insert(I); - State.reset(Def, Res, Part); - } - } - - // We'll have created a bunch of ZExts that are now parentless. Clean up. - for (const auto &KV : Cost->getMinimalBitwidths()) { - // If the value wasn't vectorized, we must maintain the original scalar - // type. The absence of the value from State indicates that it - // wasn't vectorized. - // FIXME: Should not rely on getVPValue at this point. - VPValue *Def = State.Plan->getVPValue(KV.first, true); - if (!State.hasAnyVectorValue(Def)) - continue; - for (unsigned Part = 0; Part < UF; ++Part) { - Value *I = State.get(Def, Part); - ZExtInst *Inst = dyn_cast<ZExtInst>(I); - if (Inst && Inst->use_empty()) { - Value *NewI = Inst->getOperand(0); - Inst->eraseFromParent(); - State.reset(Def, NewI, Part); - } - } - } -} - -void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { - // Insert truncates and extends for any truncated instructions as hints to - // InstCombine. - if (VF.isVector()) - truncateToMinimalBitwidths(State); - +void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State, + VPlan &Plan) { // Fix widened non-induction PHIs by setting up the PHI operands. - if (OrigPHIsToFix.size()) { - assert(EnableVPlanNativePath && - "Unexpected non-induction PHIs for fixup in non VPlan-native path"); - fixNonInductionPHIs(State); - } + if (EnableVPlanNativePath) + fixNonInductionPHIs(Plan, State); // At this point every instruction in the original loop is widened to a // vector form. Now we need to fix the recurrences in the loop. These PHI // nodes are currently empty because we did not want to introduce cycles. - // This is the second stage of vectorizing recurrences. - fixCrossIterationPHIs(State); + // This is the second stage of vectorizing recurrences. Note that fixing + // reduction phis are already modeled in VPlan. + // TODO: Also model fixing fixed-order recurrence phis in VPlan. + VPRegionBlock *VectorRegion = State.Plan->getVectorLoopRegion(); + VPBasicBlock *HeaderVPBB = VectorRegion->getEntryBasicBlock(); + for (VPRecipeBase &R : HeaderVPBB->phis()) { + if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) + fixFixedOrderRecurrence(FOR, State); + } // Forget the original basic block. PSE.getSE()->forgetLoop(OrigLoop); + PSE.getSE()->forgetBlockAndLoopDispositions(); + + // After vectorization, the exit blocks of the original loop will have + // additional predecessors. Invalidate SCEVs for the exit phis in case SE + // looked through single-entry phis. + SmallVector<BasicBlock *> ExitBlocks; + OrigLoop->getExitBlocks(ExitBlocks); + for (BasicBlock *Exit : ExitBlocks) + for (PHINode &PN : Exit->phis()) + PSE.getSE()->forgetLcssaPhiWithNewPredecessor(OrigLoop, &PN); + + VPBasicBlock *LatchVPBB = VectorRegion->getExitingBasicBlock(); + Loop *VectorLoop = LI->getLoopFor(State.CFG.VPBB2IRBB[LatchVPBB]); + if (Cost->requiresScalarEpilogue(VF.isVector())) { + // No edge from the middle block to the unique exit block has been inserted + // and there is nothing to fix from vector loop; phis should have incoming + // from scalar loop only. + } else { + // TODO: Check VPLiveOuts to see if IV users need fixing instead of checking + // the cost model. + + // If we inserted an edge from the middle block to the unique exit block, + // update uses outside the loop (phis) to account for the newly inserted + // edge. - // If we inserted an edge from the middle block to the unique exit block, - // update uses outside the loop (phis) to account for the newly inserted - // edge. - if (!Cost->requiresScalarEpilogue(VF)) { // Fix-up external users of the induction variables. - for (auto &Entry : Legal->getInductionVars()) + for (const auto &Entry : Legal->getInductionVars()) fixupIVUsers(Entry.first, Entry.second, - getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)), - IVEndValues[Entry.first], LoopMiddleBlock); - - fixLCSSAPHIs(State); + getOrCreateVectorTripCount(VectorLoop->getLoopPreheader()), + IVEndValues[Entry.first], LoopMiddleBlock, + VectorLoop->getHeader(), Plan, State); } + // Fix LCSSA phis not already fixed earlier. Extracts may need to be generated + // in the exit block, so update the builder. + State.Builder.SetInsertPoint(State.CFG.ExitBB, + State.CFG.ExitBB->getFirstNonPHIIt()); + for (const auto &KV : Plan.getLiveOuts()) + KV.second->fixPhi(Plan, State); + for (Instruction *PI : PredicatedInstructions) sinkScalarOperands(&*PI); // Remove redundant induction instructions. - cse(LoopVectorBody); + cse(VectorLoop->getHeader()); // Set/update profile weights for the vector and remainder loops as original // loop iterations are now distributed among them. Note that original loop @@ -3978,28 +3539,12 @@ void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { // For scalable vectorization we can't know at compile time how many iterations // of the loop are handled in one vector iteration, so instead assume a pessimistic // vscale of '1'. - setProfileInfoAfterUnrolling( - LI->getLoopFor(LoopScalarBody), LI->getLoopFor(LoopVectorBody), - LI->getLoopFor(LoopScalarBody), VF.getKnownMinValue() * UF); -} - -void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { - // In order to support recurrences we need to be able to vectorize Phi nodes. - // Phi nodes have cycles, so we need to vectorize them in two stages. This is - // stage #2: We now need to fix the recurrences by adding incoming edges to - // the currently empty PHI nodes. At this point every instruction in the - // original loop is widened to a vector form so we can use them to construct - // the incoming edges. - VPBasicBlock *Header = State.Plan->getEntry()->getEntryBasicBlock(); - for (VPRecipeBase &R : Header->phis()) { - if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) - fixReduction(ReductionPhi, State); - else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) - fixFirstOrderRecurrence(FOR, State); - } + setProfileInfoAfterUnrolling(LI->getLoopFor(LoopScalarBody), VectorLoop, + LI->getLoopFor(LoopScalarBody), + VF.getKnownMinValue() * UF); } -void InnerLoopVectorizer::fixFirstOrderRecurrence( +void InnerLoopVectorizer::fixFixedOrderRecurrence( VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) { // This is the second phase of vectorizing first-order recurrences. An // overview of the transformation is described below. Suppose we have the @@ -4056,34 +3601,56 @@ void InnerLoopVectorizer::fixFirstOrderRecurrence( Value *Incoming = State.get(PreviousDef, UF - 1); auto *ExtractForScalar = Incoming; auto *IdxTy = Builder.getInt32Ty(); + Value *RuntimeVF = nullptr; if (VF.isVector()) { auto *One = ConstantInt::get(IdxTy, 1); Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); - auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); + RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); auto *LastIdx = Builder.CreateSub(RuntimeVF, One); - ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx, - "vector.recur.extract"); - } - // Extract the second last element in the middle block if the - // Phi is used outside the loop. We need to extract the phi itself - // and not the last element (the phi update in the current iteration). This - // will be the value when jumping to the exit block from the LoopMiddleBlock, - // when the scalar loop is not run at all. - Value *ExtractForPhiUsedOutsideLoop = nullptr; - if (VF.isVector()) { - auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); - auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); - ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( - Incoming, Idx, "vector.recur.extract.for.phi"); - } else if (UF > 1) - // When loop is unrolled without vectorizing, initialize - // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value - // of `Incoming`. This is analogous to the vectorized case above: extracting - // the second last element when VF > 1. - ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); + ExtractForScalar = + Builder.CreateExtractElement(Incoming, LastIdx, "vector.recur.extract"); + } + + auto RecurSplice = cast<VPInstruction>(*PhiR->user_begin()); + assert(PhiR->getNumUsers() == 1 && + RecurSplice->getOpcode() == + VPInstruction::FirstOrderRecurrenceSplice && + "recurrence phi must have a single user: FirstOrderRecurrenceSplice"); + SmallVector<VPLiveOut *> LiveOuts; + for (VPUser *U : RecurSplice->users()) + if (auto *LiveOut = dyn_cast<VPLiveOut>(U)) + LiveOuts.push_back(LiveOut); + + if (!LiveOuts.empty()) { + // Extract the second last element in the middle block if the + // Phi is used outside the loop. We need to extract the phi itself + // and not the last element (the phi update in the current iteration). This + // will be the value when jumping to the exit block from the + // LoopMiddleBlock, when the scalar loop is not run at all. + Value *ExtractForPhiUsedOutsideLoop = nullptr; + if (VF.isVector()) { + auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); + ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( + Incoming, Idx, "vector.recur.extract.for.phi"); + } else { + assert(UF > 1 && "VF and UF cannot both be 1"); + // When loop is unrolled without vectorizing, initialize + // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled + // value of `Incoming`. This is analogous to the vectorized case above: + // extracting the second last element when VF > 1. + ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); + } + + for (VPLiveOut *LiveOut : LiveOuts) { + assert(!Cost->requiresScalarEpilogue(VF.isVector())); + PHINode *LCSSAPhi = LiveOut->getPhi(); + LCSSAPhi->addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); + State.Plan->removeLiveOut(LCSSAPhi); + } + } // Fix the initial value of the original recurrence in the scalar loop. - Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); + Builder.SetInsertPoint(LoopScalarPreHeader, LoopScalarPreHeader->begin()); PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue()); auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue(); @@ -4094,261 +3661,6 @@ void InnerLoopVectorizer::fixFirstOrderRecurrence( Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); Phi->setName("scalar.recur"); - - // Finally, fix users of the recurrence outside the loop. The users will need - // either the last value of the scalar recurrence or the last value of the - // vector recurrence we extracted in the middle block. Since the loop is in - // LCSSA form, we just need to find all the phi nodes for the original scalar - // recurrence in the exit block, and then add an edge for the middle block. - // Note that LCSSA does not imply single entry when the original scalar loop - // had multiple exiting edges (as we always run the last iteration in the - // scalar epilogue); in that case, there is no edge from middle to exit and - // and thus no phis which needed updated. - if (!Cost->requiresScalarEpilogue(VF)) - for (PHINode &LCSSAPhi : LoopExitBlock->phis()) - if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi)) - LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); -} - -void InnerLoopVectorizer::fixReduction(VPReductionPHIRecipe *PhiR, - VPTransformState &State) { - PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); - // Get it's reduction variable descriptor. - assert(Legal->isReductionVariable(OrigPhi) && - "Unable to find the reduction variable"); - const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); - - RecurKind RK = RdxDesc.getRecurrenceKind(); - TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); - Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); - setDebugLocFromInst(ReductionStartValue); - - VPValue *LoopExitInstDef = PhiR->getBackedgeValue(); - // This is the vector-clone of the value that leaves the loop. - Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); - - // Wrap flags are in general invalid after vectorization, clear them. - clearReductionWrapFlags(RdxDesc, State); - - // Before each round, move the insertion point right between - // the PHIs and the values we are going to write. - // This allows us to write both PHINodes and the extractelement - // instructions. - Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); - - setDebugLocFromInst(LoopExitInst); - - Type *PhiTy = OrigPhi->getType(); - // If tail is folded by masking, the vector value to leave the loop should be - // a Select choosing between the vectorized LoopExitInst and vectorized Phi, - // instead of the former. For an inloop reduction the reduction will already - // be predicated, and does not need to be handled here. - if (Cost->foldTailByMasking() && !PhiR->isInLoop()) { - for (unsigned Part = 0; Part < UF; ++Part) { - Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); - Value *Sel = nullptr; - for (User *U : VecLoopExitInst->users()) { - if (isa<SelectInst>(U)) { - assert(!Sel && "Reduction exit feeding two selects"); - Sel = U; - } else - assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select"); - } - assert(Sel && "Reduction exit feeds no select"); - State.reset(LoopExitInstDef, Sel, Part); - - // If the target can create a predicated operator for the reduction at no - // extra cost in the loop (for example a predicated vadd), it can be - // cheaper for the select to remain in the loop than be sunk out of it, - // and so use the select value for the phi instead of the old - // LoopExitValue. - if (PreferPredicatedReductionSelect || - TTI->preferPredicatedReductionSelect( - RdxDesc.getOpcode(), PhiTy, - TargetTransformInfo::ReductionFlags())) { - auto *VecRdxPhi = - cast<PHINode>(State.get(PhiR, Part)); - VecRdxPhi->setIncomingValueForBlock( - LI->getLoopFor(LoopVectorBody)->getLoopLatch(), Sel); - } - } - } - - // If the vector reduction can be performed in a smaller type, we truncate - // then extend the loop exit value to enable InstCombine to evaluate the - // entire expression in the smaller type. - if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { - assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!"); - Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); - Builder.SetInsertPoint( - LI->getLoopFor(LoopVectorBody)->getLoopLatch()->getTerminator()); - VectorParts RdxParts(UF); - for (unsigned Part = 0; Part < UF; ++Part) { - RdxParts[Part] = State.get(LoopExitInstDef, Part); - Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); - Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) - : Builder.CreateZExt(Trunc, VecTy); - for (User *U : llvm::make_early_inc_range(RdxParts[Part]->users())) - if (U != Trunc) { - U->replaceUsesOfWith(RdxParts[Part], Extnd); - RdxParts[Part] = Extnd; - } - } - Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); - for (unsigned Part = 0; Part < UF; ++Part) { - RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); - State.reset(LoopExitInstDef, RdxParts[Part], Part); - } - } - - // Reduce all of the unrolled parts into a single vector. - Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); - unsigned Op = RecurrenceDescriptor::getOpcode(RK); - - // The middle block terminator has already been assigned a DebugLoc here (the - // OrigLoop's single latch terminator). We want the whole middle block to - // appear to execute on this line because: (a) it is all compiler generated, - // (b) these instructions are always executed after evaluating the latch - // conditional branch, and (c) other passes may add new predecessors which - // terminate on this line. This is the easiest way to ensure we don't - // accidentally cause an extra step back into the loop while debugging. - setDebugLocFromInst(LoopMiddleBlock->getTerminator()); - if (PhiR->isOrdered()) - ReducedPartRdx = State.get(LoopExitInstDef, UF - 1); - else { - // Floating-point operations should have some FMF to enable the reduction. - IRBuilderBase::FastMathFlagGuard FMFG(Builder); - Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); - for (unsigned Part = 1; Part < UF; ++Part) { - Value *RdxPart = State.get(LoopExitInstDef, Part); - if (Op != Instruction::ICmp && Op != Instruction::FCmp) { - ReducedPartRdx = Builder.CreateBinOp( - (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); - } else if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) - ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK, - ReducedPartRdx, RdxPart); - else - ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); - } - } - - // Create the reduction after the loop. Note that inloop reductions create the - // target reduction in the loop using a Reduction recipe. - if (VF.isVector() && !PhiR->isInLoop()) { - ReducedPartRdx = - createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi); - // If the reduction can be performed in a smaller type, we need to extend - // the reduction to the wider type before we branch to the original loop. - if (PhiTy != RdxDesc.getRecurrenceType()) - ReducedPartRdx = RdxDesc.isSigned() - ? Builder.CreateSExt(ReducedPartRdx, PhiTy) - : Builder.CreateZExt(ReducedPartRdx, PhiTy); - } - - PHINode *ResumePhi = - dyn_cast<PHINode>(PhiR->getStartValue()->getUnderlyingValue()); - - // Create a phi node that merges control-flow from the backedge-taken check - // block and the middle block. - PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx", - LoopScalarPreHeader->getTerminator()); - - // If we are fixing reductions in the epilogue loop then we should already - // have created a bc.merge.rdx Phi after the main vector body. Ensure that - // we carry over the incoming values correctly. - for (auto *Incoming : predecessors(LoopScalarPreHeader)) { - if (Incoming == LoopMiddleBlock) - BCBlockPhi->addIncoming(ReducedPartRdx, Incoming); - else if (ResumePhi && llvm::is_contained(ResumePhi->blocks(), Incoming)) - BCBlockPhi->addIncoming(ResumePhi->getIncomingValueForBlock(Incoming), - Incoming); - else - BCBlockPhi->addIncoming(ReductionStartValue, Incoming); - } - - // Set the resume value for this reduction - ReductionResumeValues.insert({&RdxDesc, BCBlockPhi}); - - // Now, we need to fix the users of the reduction variable - // inside and outside of the scalar remainder loop. - - // We know that the loop is in LCSSA form. We need to update the PHI nodes - // in the exit blocks. See comment on analogous loop in - // fixFirstOrderRecurrence for a more complete explaination of the logic. - if (!Cost->requiresScalarEpilogue(VF)) - for (PHINode &LCSSAPhi : LoopExitBlock->phis()) - if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst)) - LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); - - // Fix the scalar loop reduction variable with the incoming reduction sum - // from the vector body and from the backedge value. - int IncomingEdgeBlockIdx = - OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); - assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); - // Pick the other block. - int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); - OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); - OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); -} - -void InnerLoopVectorizer::clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, - VPTransformState &State) { - RecurKind RK = RdxDesc.getRecurrenceKind(); - if (RK != RecurKind::Add && RK != RecurKind::Mul) - return; - - Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr(); - assert(LoopExitInstr && "null loop exit instruction"); - SmallVector<Instruction *, 8> Worklist; - SmallPtrSet<Instruction *, 8> Visited; - Worklist.push_back(LoopExitInstr); - Visited.insert(LoopExitInstr); - - while (!Worklist.empty()) { - Instruction *Cur = Worklist.pop_back_val(); - if (isa<OverflowingBinaryOperator>(Cur)) - for (unsigned Part = 0; Part < UF; ++Part) { - // FIXME: Should not rely on getVPValue at this point. - Value *V = State.get(State.Plan->getVPValue(Cur, true), Part); - cast<Instruction>(V)->dropPoisonGeneratingFlags(); - } - - for (User *U : Cur->users()) { - Instruction *UI = cast<Instruction>(U); - if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) && - Visited.insert(UI).second) - Worklist.push_back(UI); - } - } -} - -void InnerLoopVectorizer::fixLCSSAPHIs(VPTransformState &State) { - for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { - if (LCSSAPhi.getBasicBlockIndex(LoopMiddleBlock) != -1) - // Some phis were already hand updated by the reduction and recurrence - // code above, leave them alone. - continue; - - auto *IncomingValue = LCSSAPhi.getIncomingValue(0); - // Non-instruction incoming values will have only one value. - - VPLane Lane = VPLane::getFirstLane(); - if (isa<Instruction>(IncomingValue) && - !Cost->isUniformAfterVectorization(cast<Instruction>(IncomingValue), - VF)) - Lane = VPLane::getLastLaneForVF(VF); - - // Can be a loop invariant incoming value or the last scalar value to be - // extracted from the vectorized loop. - // FIXME: Should not rely on getVPValue at this point. - Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); - Value *lastIncomingValue = - OrigLoop->isLoopInvariant(IncomingValue) - ? IncomingValue - : State.get(State.Plan->getVPValue(IncomingValue, true), - VPIteration(UF - 1, Lane)); - LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock); - } } void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { @@ -4390,10 +3702,11 @@ void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); // We can't sink an instruction if it is a phi node, is not in the loop, - // or may have side effects. + // may have side effects or may read from memory. + // TODO Could dor more granular checking to allow sinking a load past non-store instructions. if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) || - I->mayHaveSideEffects()) - continue; + I->mayHaveSideEffects() || I->mayReadFromMemory()) + continue; // If the instruction is already in PredBB, check if we can sink its // operands. In that case, VPlan's sinkScalarOperands() succeeded in @@ -4425,17 +3738,22 @@ void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { } while (Changed); } -void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { - for (PHINode *OrigPhi : OrigPHIsToFix) { - VPWidenPHIRecipe *VPPhi = - cast<VPWidenPHIRecipe>(State.Plan->getVPValue(OrigPhi)); - PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); - // Make sure the builder has a valid insert point. - Builder.SetInsertPoint(NewPhi); - for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { - VPValue *Inc = VPPhi->getIncomingValue(i); - VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); - NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); +void InnerLoopVectorizer::fixNonInductionPHIs(VPlan &Plan, + VPTransformState &State) { + auto Iter = vp_depth_first_deep(Plan.getEntry()); + for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { + for (VPRecipeBase &P : VPBB->phis()) { + VPWidenPHIRecipe *VPPhi = dyn_cast<VPWidenPHIRecipe>(&P); + if (!VPPhi) + continue; + PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); + // Make sure the builder has a valid insert point. + Builder.SetInsertPoint(NewPhi); + for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { + VPValue *Inc = VPPhi->getIncomingValue(i); + VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); + NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); + } } } } @@ -4445,237 +3763,21 @@ bool InnerLoopVectorizer::useOrderedReductions( return Cost->useOrderedReductions(RdxDesc); } -void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, - VPWidenPHIRecipe *PhiR, - VPTransformState &State) { - PHINode *P = cast<PHINode>(PN); - if (EnableVPlanNativePath) { - // Currently we enter here in the VPlan-native path for non-induction - // PHIs where all control flow is uniform. We simply widen these PHIs. - // Create a vector phi with no operands - the vector phi operands will be - // set at the end of vector code generation. - Type *VecTy = (State.VF.isScalar()) - ? PN->getType() - : VectorType::get(PN->getType(), State.VF); - Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi"); - State.set(PhiR, VecPhi, 0); - OrigPHIsToFix.push_back(P); - - return; - } - - assert(PN->getParent() == OrigLoop->getHeader() && - "Non-header phis should have been handled elsewhere"); - - // In order to support recurrences we need to be able to vectorize Phi nodes. - // Phi nodes have cycles, so we need to vectorize them in two stages. This is - // stage #1: We create a new vector PHI node with no incoming edges. We'll use - // this value when we vectorize all of the instructions that use the PHI. - - assert(!Legal->isReductionVariable(P) && - "reductions should be handled elsewhere"); - - setDebugLocFromInst(P); - - // This PHINode must be an induction variable. - // Make sure that we know about it. - assert(Legal->getInductionVars().count(P) && "Not an induction variable"); - - InductionDescriptor II = Legal->getInductionVars().lookup(P); - const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); - - auto *IVR = PhiR->getParent()->getPlan()->getCanonicalIV(); - PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); - - // FIXME: The newly created binary instructions should contain nsw/nuw flags, - // which can be found from the original scalar operations. - switch (II.getKind()) { - case InductionDescriptor::IK_NoInduction: - llvm_unreachable("Unknown induction"); - case InductionDescriptor::IK_IntInduction: - case InductionDescriptor::IK_FpInduction: - llvm_unreachable("Integer/fp induction is handled elsewhere."); - case InductionDescriptor::IK_PtrInduction: { - // Handle the pointer induction variable case. - assert(P->getType()->isPointerTy() && "Unexpected type."); - - if (Cost->isScalarAfterVectorization(P, State.VF)) { - // This is the normalized GEP that starts counting at zero. - Value *PtrInd = - Builder.CreateSExtOrTrunc(CanonicalIV, II.getStep()->getType()); - // Determine the number of scalars we need to generate for each unroll - // iteration. If the instruction is uniform, we only need to generate the - // first lane. Otherwise, we generate all VF values. - bool IsUniform = vputils::onlyFirstLaneUsed(PhiR); - assert((IsUniform || !State.VF.isScalable()) && - "Cannot scalarize a scalable VF"); - unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); - - for (unsigned Part = 0; Part < UF; ++Part) { - Value *PartStart = - createStepForVF(Builder, PtrInd->getType(), VF, Part); - - for (unsigned Lane = 0; Lane < Lanes; ++Lane) { - Value *Idx = Builder.CreateAdd( - PartStart, ConstantInt::get(PtrInd->getType(), Lane)); - Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); - Value *SclrGep = emitTransformedIndex(Builder, GlobalIdx, PSE.getSE(), - DL, II, State.CFG.PrevBB); - SclrGep->setName("next.gep"); - State.set(PhiR, SclrGep, VPIteration(Part, Lane)); - } - } - return; - } - assert(isa<SCEVConstant>(II.getStep()) && - "Induction step not a SCEV constant!"); - Type *PhiType = II.getStep()->getType(); - - // Build a pointer phi - Value *ScalarStartValue = PhiR->getStartValue()->getLiveInIRValue(); - Type *ScStValueType = ScalarStartValue->getType(); - PHINode *NewPointerPhi = - PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); - NewPointerPhi->addIncoming(ScalarStartValue, LoopVectorPreHeader); - - // A pointer induction, performed by using a gep - BasicBlock *LoopLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); - Instruction *InductionLoc = LoopLatch->getTerminator(); - const SCEV *ScalarStep = II.getStep(); - SCEVExpander Exp(*PSE.getSE(), DL, "induction"); - Value *ScalarStepValue = - Exp.expandCodeFor(ScalarStep, PhiType, InductionLoc); - Value *RuntimeVF = getRuntimeVF(Builder, PhiType, VF); - Value *NumUnrolledElems = - Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); - Value *InductionGEP = GetElementPtrInst::Create( - II.getElementType(), NewPointerPhi, - Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", - InductionLoc); - NewPointerPhi->addIncoming(InductionGEP, LoopLatch); - - // Create UF many actual address geps that use the pointer - // phi as base and a vectorized version of the step value - // (<step*0, ..., step*N>) as offset. - for (unsigned Part = 0; Part < State.UF; ++Part) { - Type *VecPhiType = VectorType::get(PhiType, State.VF); - Value *StartOffsetScalar = - Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); - Value *StartOffset = - Builder.CreateVectorSplat(State.VF, StartOffsetScalar); - // Create a vector of consecutive numbers from zero to VF. - StartOffset = - Builder.CreateAdd(StartOffset, Builder.CreateStepVector(VecPhiType)); - - Value *GEP = Builder.CreateGEP( - II.getElementType(), NewPointerPhi, - Builder.CreateMul( - StartOffset, Builder.CreateVectorSplat(State.VF, ScalarStepValue), - "vector.gep")); - State.set(PhiR, GEP, Part); - } - } - } -} - -/// A helper function for checking whether an integer division-related -/// instruction may divide by zero (in which case it must be predicated if -/// executed conditionally in the scalar code). -/// TODO: It may be worthwhile to generalize and check isKnownNonZero(). -/// Non-zero divisors that are non compile-time constants will not be -/// converted into multiplication, so we will still end up scalarizing -/// the division, but can do so w/o predication. -static bool mayDivideByZero(Instruction &I) { - assert((I.getOpcode() == Instruction::UDiv || - I.getOpcode() == Instruction::SDiv || - I.getOpcode() == Instruction::URem || - I.getOpcode() == Instruction::SRem) && - "Unexpected instruction"); - Value *Divisor = I.getOperand(1); - auto *CInt = dyn_cast<ConstantInt>(Divisor); - return !CInt || CInt->isZero(); -} - -void InnerLoopVectorizer::widenCallInstruction(CallInst &I, VPValue *Def, - VPUser &ArgOperands, - VPTransformState &State) { - assert(!isa<DbgInfoIntrinsic>(I) && - "DbgInfoIntrinsic should have been dropped during VPlan construction"); - setDebugLocFromInst(&I); - - Module *M = I.getParent()->getParent()->getParent(); - auto *CI = cast<CallInst>(&I); - - SmallVector<Type *, 4> Tys; - for (Value *ArgOperand : CI->args()) - Tys.push_back(ToVectorTy(ArgOperand->getType(), VF.getKnownMinValue())); - - Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); - - // The flag shows whether we use Intrinsic or a usual Call for vectorized - // version of the instruction. - // Is it beneficial to perform intrinsic call compared to lib call? - bool NeedToScalarize = false; - InstructionCost CallCost = Cost->getVectorCallCost(CI, VF, NeedToScalarize); - InstructionCost IntrinsicCost = ID ? Cost->getVectorIntrinsicCost(CI, VF) : 0; - bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; - assert((UseVectorIntrinsic || !NeedToScalarize) && - "Instruction should be scalarized elsewhere."); - assert((IntrinsicCost.isValid() || CallCost.isValid()) && - "Either the intrinsic cost or vector call cost must be valid"); - - for (unsigned Part = 0; Part < UF; ++Part) { - SmallVector<Type *, 2> TysForDecl = {CI->getType()}; - SmallVector<Value *, 4> Args; - for (auto &I : enumerate(ArgOperands.operands())) { - // Some intrinsics have a scalar argument - don't replace it with a - // vector. - Value *Arg; - if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, I.index())) - Arg = State.get(I.value(), Part); - else { - Arg = State.get(I.value(), VPIteration(0, 0)); - if (hasVectorInstrinsicOverloadedScalarOpd(ID, I.index())) - TysForDecl.push_back(Arg->getType()); - } - Args.push_back(Arg); - } - - Function *VectorF; - if (UseVectorIntrinsic) { - // Use vector version of the intrinsic. - if (VF.isVector()) - TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); - VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); - assert(VectorF && "Can't retrieve vector intrinsic."); - } else { - // Use vector version of the function call. - const VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); -#ifndef NDEBUG - assert(VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && - "Can't create vector function."); -#endif - VectorF = VFDatabase(*CI).getVectorizedFunction(Shape); - } - SmallVector<OperandBundleDef, 1> OpBundles; - CI->getOperandBundlesAsDefs(OpBundles); - CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles); - - if (isa<FPMathOperator>(V)) - V->copyFastMathFlags(CI); - - State.set(Def, V, Part); - addMetadata(V, &I); - } -} - void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { // We should not collect Scalars more than once per VF. Right now, this // function is called from collectUniformsAndScalars(), which already does // this check. Collecting Scalars for VF=1 does not make any sense. - assert(VF.isVector() && Scalars.find(VF) == Scalars.end() && + assert(VF.isVector() && !Scalars.contains(VF) && "This function should not be visited twice for the same VF"); + // This avoids any chances of creating a REPLICATE recipe during planning + // since that would result in generation of scalarized code during execution, + // which is not supported for scalable vectors. + if (VF.isScalable()) { + Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end()); + return; + } + SmallSetVector<Instruction *, 8> Worklist; // These sets are used to seed the analysis with pointers used by memory @@ -4765,12 +3867,14 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { } // Insert the forced scalars. - // FIXME: Currently widenPHIInstruction() often creates a dead vector + // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector // induction variable when the PHI user is scalarized. auto ForcedScalar = ForcedScalars.find(VF); if (ForcedScalar != ForcedScalars.end()) - for (auto *I : ForcedScalar->second) + for (auto *I : ForcedScalar->second) { + LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n"); Worklist.insert(I); + } // Expand the worklist by looking through any bitcasts and getelementptr // instructions we've already identified as scalar. This is similar to the @@ -4795,7 +3899,7 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { // An induction variable will remain scalar if all users of the induction // variable and induction variable update remain scalar. - for (auto &Induction : Legal->getInductionVars()) { + for (const auto &Induction : Legal->getInductionVars()) { auto *Ind = Induction.first; auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); @@ -4848,15 +3952,21 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { bool LoopVectorizationCostModel::isScalarWithPredication( Instruction *I, ElementCount VF) const { - if (!blockNeedsPredicationForAnyReason(I->getParent())) + if (!isPredicatedInst(I)) return false; + + // Do we have a non-scalar lowering for this predicated + // instruction? No - it is scalar with predication. switch(I->getOpcode()) { default: - break; + return true; + case Instruction::Call: + if (VF.isScalar()) + return true; + return CallWideningDecisions.at(std::make_pair(cast<CallInst>(I), VF)) + .Kind == CM_Scalarize; case Instruction::Load: case Instruction::Store: { - if (!Legal->isMaskRequired(I)) - return false; auto *Ptr = getLoadStorePointerOperand(I); auto *Ty = getLoadStoreType(I); Type *VTy = Ty; @@ -4871,10 +3981,122 @@ bool LoopVectorizationCostModel::isScalarWithPredication( case Instruction::UDiv: case Instruction::SDiv: case Instruction::SRem: - case Instruction::URem: - return mayDivideByZero(*I); + case Instruction::URem: { + // We have the option to use the safe-divisor idiom to avoid predication. + // The cost based decision here will always select safe-divisor for + // scalable vectors as scalarization isn't legal. + const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); + return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost); } - return false; + } +} + +bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const { + if (!blockNeedsPredicationForAnyReason(I->getParent())) + return false; + + // Can we prove this instruction is safe to unconditionally execute? + // If not, we must use some form of predication. + switch(I->getOpcode()) { + default: + return false; + case Instruction::Load: + case Instruction::Store: { + if (!Legal->isMaskRequired(I)) + return false; + // When we know the load's address is loop invariant and the instruction + // in the original scalar loop was unconditionally executed then we + // don't need to mark it as a predicated instruction. Tail folding may + // introduce additional predication, but we're guaranteed to always have + // at least one active lane. We call Legal->blockNeedsPredication here + // because it doesn't query tail-folding. For stores, we need to prove + // both speculation safety (which follows from the same argument as loads), + // but also must prove the value being stored is correct. The easiest + // form of the later is to require that all values stored are the same. + if (Legal->isInvariant(getLoadStorePointerOperand(I)) && + (isa<LoadInst>(I) || + (isa<StoreInst>(I) && + TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()))) && + !Legal->blockNeedsPredication(I->getParent())) + return false; + return true; + } + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::SRem: + case Instruction::URem: + // TODO: We can use the loop-preheader as context point here and get + // context sensitive reasoning + return !isSafeToSpeculativelyExecute(I); + case Instruction::Call: + return Legal->isMaskRequired(I); + } +} + +std::pair<InstructionCost, InstructionCost> +LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I, + ElementCount VF) const { + assert(I->getOpcode() == Instruction::UDiv || + I->getOpcode() == Instruction::SDiv || + I->getOpcode() == Instruction::SRem || + I->getOpcode() == Instruction::URem); + assert(!isSafeToSpeculativelyExecute(I)); + + const TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; + + // Scalarization isn't legal for scalable vector types + InstructionCost ScalarizationCost = InstructionCost::getInvalid(); + if (!VF.isScalable()) { + // Get the scalarization cost and scale this amount by the probability of + // executing the predicated block. If the instruction is not predicated, + // we fall through to the next case. + ScalarizationCost = 0; + + // These instructions have a non-void type, so account for the phi nodes + // that we will create. This cost is likely to be zero. The phi node + // cost, if any, should be scaled by the block probability because it + // models a copy at the end of each predicated block. + ScalarizationCost += VF.getKnownMinValue() * + TTI.getCFInstrCost(Instruction::PHI, CostKind); + + // The cost of the non-predicated instruction. + ScalarizationCost += VF.getKnownMinValue() * + TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind); + + // The cost of insertelement and extractelement instructions needed for + // scalarization. + ScalarizationCost += getScalarizationOverhead(I, VF, CostKind); + + // Scale the cost by the probability of executing the predicated blocks. + // This assumes the predicated block for each vector lane is equally + // likely. + ScalarizationCost = ScalarizationCost / getReciprocalPredBlockProb(); + } + InstructionCost SafeDivisorCost = 0; + + auto *VecTy = ToVectorTy(I->getType(), VF); + + // The cost of the select guard to ensure all lanes are well defined + // after we speculate above any internal control flow. + SafeDivisorCost += TTI.getCmpSelInstrCost( + Instruction::Select, VecTy, + ToVectorTy(Type::getInt1Ty(I->getContext()), VF), + CmpInst::BAD_ICMP_PREDICATE, CostKind); + + // Certain instructions can be cheaper to vectorize if they have a constant + // second vector operand. One example of this are shifts on x86. + Value *Op2 = I->getOperand(1); + auto Op2Info = TTI.getOperandInfo(Op2); + if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && + Legal->isInvariant(Op2)) + Op2Info.Kind = TargetTransformInfo::OK_UniformValue; + + SmallVector<const Value *, 4> Operands(I->operand_values()); + SafeDivisorCost += TTI.getArithmeticInstrCost( + I->getOpcode(), VecTy, CostKind, + {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, + Op2Info, Operands, I); + return {ScalarizationCost, SafeDivisorCost}; } bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( @@ -4892,6 +4114,27 @@ bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( if (hasIrregularType(ScalarTy, DL)) return false; + // If the group involves a non-integral pointer, we may not be able to + // losslessly cast all values to a common type. + unsigned InterleaveFactor = Group->getFactor(); + bool ScalarNI = DL.isNonIntegralPointerType(ScalarTy); + for (unsigned i = 0; i < InterleaveFactor; i++) { + Instruction *Member = Group->getMember(i); + if (!Member) + continue; + auto *MemberTy = getLoadStoreType(Member); + bool MemberNI = DL.isNonIntegralPointerType(MemberTy); + // Don't coerce non-integral pointers to integers or vice versa. + if (MemberNI != ScalarNI) { + // TODO: Consider adding special nullptr value case here + return false; + } else if (MemberNI && ScalarNI && + ScalarTy->getPointerAddressSpace() != + MemberTy->getPointerAddressSpace()) { + return false; + } + } + // Check if masking is required. // A Group may need masking for one of two reasons: it resides in a block that // needs predication, or it was decided to use masking to deal with gaps @@ -4957,7 +4200,7 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { // already does this check. Collecting Uniforms for VF=1 does not make any // sense. - assert(VF.isVector() && Uniforms.find(VF) == Uniforms.end() && + assert(VF.isVector() && !Uniforms.contains(VF) && "This function should not be visited twice for the same VF"); // Visit the list of Uniforms. If we'll not find any uniform value, we'll @@ -5006,28 +4249,48 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) addToWorklistIfAllowed(Cmp); + auto PrevVF = VF.divideCoefficientBy(2); + // Return true if all lanes perform the same memory operation, and we can + // thus chose to execute only one. + auto isUniformMemOpUse = [&](Instruction *I) { + // If the value was already known to not be uniform for the previous + // (smaller VF), it cannot be uniform for the larger VF. + if (PrevVF.isVector()) { + auto Iter = Uniforms.find(PrevVF); + if (Iter != Uniforms.end() && !Iter->second.contains(I)) + return false; + } + if (!Legal->isUniformMemOp(*I, VF)) + return false; + if (isa<LoadInst>(I)) + // Loading the same address always produces the same result - at least + // assuming aliasing and ordering which have already been checked. + return true; + // Storing the same value on every iteration. + return TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()); + }; + auto isUniformDecision = [&](Instruction *I, ElementCount VF) { InstWidening WideningDecision = getWideningDecision(I, VF); assert(WideningDecision != CM_Unknown && "Widening decision should be ready at this moment"); - // A uniform memory op is itself uniform. We exclude uniform stores - // here as they demand the last lane, not the first one. - if (isa<LoadInst>(I) && Legal->isUniformMemOp(*I)) { - assert(WideningDecision == CM_Scalarize); + if (isUniformMemOpUse(I)) return true; - } return (WideningDecision == CM_Widen || WideningDecision == CM_Widen_Reverse || WideningDecision == CM_Interleave); }; - // Returns true if Ptr is the pointer operand of a memory access instruction - // I, and I is known to not require scalarization. + // I, I is known to not require scalarization, and the pointer is not also + // stored. auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { - return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); + if (isa<StoreInst>(I) && I->getOperand(0) == Ptr) + return false; + return getLoadStorePointerOperand(I) == Ptr && + (isUniformDecision(I, VF) || Legal->isInvariant(Ptr)); }; // Holds a list of values which are known to have at least one uniform use. @@ -5070,15 +4333,11 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { if (!Ptr) continue; - // A uniform memory op is itself uniform. We exclude uniform stores - // here as they demand the last lane, not the first one. - if (isa<LoadInst>(I) && Legal->isUniformMemOp(I)) + if (isUniformMemOpUse(&I)) addToWorklistIfAllowed(&I); - if (isUniformDecision(&I, VF)) { - assert(isVectorizedMemAccessUse(&I, Ptr) && "consistency check"); + if (isVectorizedMemAccessUse(&I, Ptr)) HasUniformUse.insert(Ptr); - } } // Add to the worklist any operands which have *only* uniform (e.g. lane 0 @@ -5103,14 +4362,14 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { while (idx != Worklist.size()) { Instruction *I = Worklist[idx++]; - for (auto OV : I->operand_values()) { + for (auto *OV : I->operand_values()) { // isOutOfScope operands cannot be uniform instructions. if (isOutOfScope(OV)) continue; // First order recurrence Phi's should typically be considered // non-uniform. auto *OP = dyn_cast<PHINode>(OV); - if (OP && Legal->isFirstOrderRecurrence(OP)) + if (OP && Legal->isFixedOrderRecurrence(OP)) continue; // If all the users of the operand are uniform, then add the // operand into the uniform worklist. @@ -5129,7 +4388,7 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { // nodes separately. An induction variable will remain uniform if all users // of the induction variable and induction variable update remain uniform. // The code below handles both pointer and non-pointer induction variables. - for (auto &Induction : Legal->getInductionVars()) { + for (const auto &Induction : Legal->getInductionVars()) { auto *Ind = Induction.first; auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); @@ -5174,7 +4433,7 @@ bool LoopVectorizationCostModel::runtimeChecksRequired() { return true; } - if (!PSE.getUnionPredicate().getPredicates().empty()) { + if (!PSE.getPredicate().isAlwaysTrue()) { reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", "runtime SCEV checks needed. Enable vectorization of this " "loop with '#pragma clang loop vectorize(enable)' when " @@ -5242,12 +4501,11 @@ LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { return MaxScalableVF; // Limit MaxScalableVF by the maximum safe dependence distance. - Optional<unsigned> MaxVScale = TTI.getMaxVScale(); - if (!MaxVScale && TheFunction->hasFnAttribute(Attribute::VScaleRange)) - MaxVScale = - TheFunction->getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); - MaxScalableVF = ElementCount::getScalable( - MaxVScale ? (MaxSafeElements / MaxVScale.getValue()) : 0); + if (std::optional<unsigned> MaxVScale = getMaxVScale(*TheFunction, TTI)) + MaxScalableVF = ElementCount::getScalable(MaxSafeElements / *MaxVScale); + else + MaxScalableVF = ElementCount::getScalable(0); + if (!MaxScalableVF) reportVectorizationInfo( "Max legal vector width too small, scalable vectorization " @@ -5258,7 +4516,7 @@ LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { } FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( - unsigned ConstTripCount, ElementCount UserVF, bool FoldTailByMasking) { + unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking) { MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); unsigned SmallestType, WidestType; std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); @@ -5268,7 +4526,7 @@ FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( // the memory accesses that is most restrictive (involved in the smallest // dependence distance). unsigned MaxSafeElements = - PowerOf2Floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); + llvm::bit_floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements); auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements); @@ -5346,12 +4604,12 @@ FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( FixedScalableVFPair Result(ElementCount::getFixed(1), ElementCount::getScalable(0)); if (auto MaxVF = - getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, + getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, MaxSafeFixedVF, FoldTailByMasking)) Result.FixedVF = MaxVF; if (auto MaxVF = - getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, + getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, MaxSafeScalableVF, FoldTailByMasking)) if (MaxVF.isScalable()) { Result.ScalableVF = MaxVF; @@ -5375,6 +4633,7 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { } unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); + unsigned MaxTC = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); if (TC == 1) { reportVectorizationFailure("Single iteration (non) loop", @@ -5385,9 +4644,9 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { switch (ScalarEpilogueStatus) { case CM_ScalarEpilogueAllowed: - return computeFeasibleMaxVF(TC, UserVF, false); + return computeFeasibleMaxVF(MaxTC, UserVF, false); case CM_ScalarEpilogueNotAllowedUsePredicate: - LLVM_FALLTHROUGH; + [[fallthrough]]; case CM_ScalarEpilogueNotNeededUsePredicate: LLVM_DEBUG( dbgs() << "LV: vector predicate hint/switch found.\n" @@ -5423,7 +4682,7 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"); ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; - return computeFeasibleMaxVF(TC, UserVF, false); + return computeFeasibleMaxVF(MaxTC, UserVF, false); } return FixedScalableVFPair::getNone(); } @@ -5440,17 +4699,27 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); } - FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(TC, UserVF, true); + FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(MaxTC, UserVF, true); + // Avoid tail folding if the trip count is known to be a multiple of any VF - // we chose. - // FIXME: The condition below pessimises the case for fixed-width vectors, - // when scalable VFs are also candidates for vectorization. - if (MaxFactors.FixedVF.isVector() && !MaxFactors.ScalableVF) { - ElementCount MaxFixedVF = MaxFactors.FixedVF; - assert((UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) && + // we choose. + std::optional<unsigned> MaxPowerOf2RuntimeVF = + MaxFactors.FixedVF.getFixedValue(); + if (MaxFactors.ScalableVF) { + std::optional<unsigned> MaxVScale = getMaxVScale(*TheFunction, TTI); + if (MaxVScale && TTI.isVScaleKnownToBeAPowerOfTwo()) { + MaxPowerOf2RuntimeVF = std::max<unsigned>( + *MaxPowerOf2RuntimeVF, + *MaxVScale * MaxFactors.ScalableVF.getKnownMinValue()); + } else + MaxPowerOf2RuntimeVF = std::nullopt; // Stick with tail-folding for now. + } + + if (MaxPowerOf2RuntimeVF && *MaxPowerOf2RuntimeVF > 0) { + assert((UserVF.isNonZero() || isPowerOf2_32(*MaxPowerOf2RuntimeVF)) && "MaxFixedVF must be a power of 2"); - unsigned MaxVFtimesIC = UserIC ? MaxFixedVF.getFixedValue() * UserIC - : MaxFixedVF.getFixedValue(); + unsigned MaxVFtimesIC = + UserIC ? *MaxPowerOf2RuntimeVF * UserIC : *MaxPowerOf2RuntimeVF; ScalarEvolution *SE = PSE.getSE(); const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); const SCEV *ExitCount = SE->getAddExpr( @@ -5465,20 +4734,12 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { } } - // For scalable vectors don't use tail folding for low trip counts or - // optimizing for code size. We only permit this if the user has explicitly - // requested it. - if (ScalarEpilogueStatus != CM_ScalarEpilogueNotNeededUsePredicate && - ScalarEpilogueStatus != CM_ScalarEpilogueNotAllowedUsePredicate && - MaxFactors.ScalableVF.isVector()) - MaxFactors.ScalableVF = ElementCount::getScalable(0); - // If we don't know the precise trip count, or if the trip count that we // found modulo the vectorization factor is not zero, try to fold the tail // by masking. // FIXME: look for a smaller MaxVF that does divide TC rather than masking. if (Legal->prepareToFoldTailByMasking()) { - FoldTailByMasking = true; + CanFoldTailByMasking = true; return MaxFactors; } @@ -5514,10 +4775,10 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { } ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( - unsigned ConstTripCount, unsigned SmallestType, unsigned WidestType, - const ElementCount &MaxSafeVF, bool FoldTailByMasking) { + unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType, + ElementCount MaxSafeVF, bool FoldTailByMasking) { bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); - TypeSize WidestRegister = TTI.getRegisterBitWidth( + const TypeSize WidestRegister = TTI.getRegisterBitWidth( ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector : TargetTransformInfo::RGK_FixedWidthVector); @@ -5531,7 +4792,7 @@ ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( // Ensure MaxVF is a power of 2; the dependence distance bound may not be. // Note that both WidestRegister and WidestType may not be a powers of 2. auto MaxVectorElementCount = ElementCount::get( - PowerOf2Floor(WidestRegister.getKnownMinSize() / WidestType), + llvm::bit_floor(WidestRegister.getKnownMinValue() / WidestType), ComputeScalableMaxVF); MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: " @@ -5544,27 +4805,46 @@ ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( return ElementCount::getFixed(1); } - const auto TripCountEC = ElementCount::getFixed(ConstTripCount); - if (ConstTripCount && - ElementCount::isKnownLE(TripCountEC, MaxVectorElementCount) && - (!FoldTailByMasking || isPowerOf2_32(ConstTripCount))) { - // If loop trip count (TC) is known at compile time there is no point in - // choosing VF greater than TC (as done in the loop below). Select maximum - // power of two which doesn't exceed TC. - // If MaxVectorElementCount is scalable, we only fall back on a fixed VF - // when the TC is less than or equal to the known number of lanes. - auto ClampedConstTripCount = PowerOf2Floor(ConstTripCount); + unsigned WidestRegisterMinEC = MaxVectorElementCount.getKnownMinValue(); + if (MaxVectorElementCount.isScalable() && + TheFunction->hasFnAttribute(Attribute::VScaleRange)) { + auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); + auto Min = Attr.getVScaleRangeMin(); + WidestRegisterMinEC *= Min; + } + + // When a scalar epilogue is required, at least one iteration of the scalar + // loop has to execute. Adjust MaxTripCount accordingly to avoid picking a + // max VF that results in a dead vector loop. + if (MaxTripCount > 0 && requiresScalarEpilogue(true)) + MaxTripCount -= 1; + + if (MaxTripCount && MaxTripCount <= WidestRegisterMinEC && + (!FoldTailByMasking || isPowerOf2_32(MaxTripCount))) { + // If upper bound loop trip count (TC) is known at compile time there is no + // point in choosing VF greater than TC (as done in the loop below). Select + // maximum power of two which doesn't exceed TC. If MaxVectorElementCount is + // scalable, we only fall back on a fixed VF when the TC is less than or + // equal to the known number of lanes. + auto ClampedUpperTripCount = llvm::bit_floor(MaxTripCount); LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " - << ClampedConstTripCount << "\n"); - return ElementCount::getFixed(ClampedConstTripCount); + << ClampedUpperTripCount << "\n"); + return ElementCount::get( + ClampedUpperTripCount, + FoldTailByMasking ? MaxVectorElementCount.isScalable() : false); } + TargetTransformInfo::RegisterKind RegKind = + ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector + : TargetTransformInfo::RGK_FixedWidthVector; ElementCount MaxVF = MaxVectorElementCount; - if (TTI.shouldMaximizeVectorBandwidth() || - (MaximizeBandwidth && isScalarEpilogueAllowed())) { + if (MaximizeBandwidth || + (MaximizeBandwidth.getNumOccurrences() == 0 && + (TTI.shouldMaximizeVectorBandwidth(RegKind) || + (UseWiderVFIfCallVariantsPresent && Legal->hasVectorCallVariants())))) { auto MaxVectorElementCountMaxBW = ElementCount::get( - PowerOf2Floor(WidestRegister.getKnownMinSize() / SmallestType), + llvm::bit_floor(WidestRegister.getKnownMinValue() / SmallestType), ComputeScalableMaxVF); MaxVectorElementCountMaxBW = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); @@ -5600,13 +4880,23 @@ ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( MaxVF = MinVF; } } + + // Invalidate any widening decisions we might have made, in case the loop + // requires prediction (decided later), but we have already made some + // load/store widening decisions. + invalidateCostModelingDecisions(); } return MaxVF; } -Optional<unsigned> LoopVectorizationCostModel::getVScaleForTuning() const { - if (TheFunction->hasFnAttribute(Attribute::VScaleRange)) { - auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); +/// Convenience function that returns the value of vscale_range iff +/// vscale_range.min == vscale_range.max or otherwise returns the value +/// returned by the corresponding TTI method. +static std::optional<unsigned> +getVScaleForTuning(const Loop *L, const TargetTransformInfo &TTI) { + const Function *Fn = L->getHeader()->getParent(); + if (Fn->hasFnAttribute(Attribute::VScaleRange)) { + auto Attr = Fn->getFnAttribute(Attribute::VScaleRange); auto Min = Attr.getVScaleRangeMin(); auto Max = Attr.getVScaleRangeMax(); if (Max && Min == Max) @@ -5616,35 +4906,43 @@ Optional<unsigned> LoopVectorizationCostModel::getVScaleForTuning() const { return TTI.getVScaleForTuning(); } -bool LoopVectorizationCostModel::isMoreProfitable( +bool LoopVectorizationPlanner::isMoreProfitable( const VectorizationFactor &A, const VectorizationFactor &B) const { InstructionCost CostA = A.Cost; InstructionCost CostB = B.Cost; - unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); - - if (!A.Width.isScalable() && !B.Width.isScalable() && FoldTailByMasking && - MaxTripCount) { - // If we are folding the tail and the trip count is a known (possibly small) - // constant, the trip count will be rounded up to an integer number of - // iterations. The total cost will be PerIterationCost*ceil(TripCount/VF), - // which we compare directly. When not folding the tail, the total cost will - // be PerIterationCost*floor(TC/VF) + Scalar remainder cost, and so is - // approximated with the per-lane cost below instead of using the tripcount - // as here. - auto RTCostA = CostA * divideCeil(MaxTripCount, A.Width.getFixedValue()); - auto RTCostB = CostB * divideCeil(MaxTripCount, B.Width.getFixedValue()); + unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(OrigLoop); + + if (!A.Width.isScalable() && !B.Width.isScalable() && MaxTripCount) { + // If the trip count is a known (possibly small) constant, the trip count + // will be rounded up to an integer number of iterations under + // FoldTailByMasking. The total cost in that case will be + // VecCost*ceil(TripCount/VF). When not folding the tail, the total + // cost will be VecCost*floor(TC/VF) + ScalarCost*(TC%VF). There will be + // some extra overheads, but for the purpose of comparing the costs of + // different VFs we can use this to compare the total loop-body cost + // expected after vectorization. + auto GetCostForTC = [MaxTripCount, this](unsigned VF, + InstructionCost VectorCost, + InstructionCost ScalarCost) { + return CM.foldTailByMasking() ? VectorCost * divideCeil(MaxTripCount, VF) + : VectorCost * (MaxTripCount / VF) + + ScalarCost * (MaxTripCount % VF); + }; + auto RTCostA = GetCostForTC(A.Width.getFixedValue(), CostA, A.ScalarCost); + auto RTCostB = GetCostForTC(B.Width.getFixedValue(), CostB, B.ScalarCost); + return RTCostA < RTCostB; } // Improve estimate for the vector width if it is scalable. unsigned EstimatedWidthA = A.Width.getKnownMinValue(); unsigned EstimatedWidthB = B.Width.getKnownMinValue(); - if (Optional<unsigned> VScale = getVScaleForTuning()) { + if (std::optional<unsigned> VScale = getVScaleForTuning(OrigLoop, TTI)) { if (A.Width.isScalable()) - EstimatedWidthA *= VScale.getValue(); + EstimatedWidthA *= *VScale; if (B.Width.isScalable()) - EstimatedWidthB *= VScale.getValue(); + EstimatedWidthB *= *VScale; } // Assume vscale may be larger than 1 (or the value being tuned for), @@ -5659,18 +4957,84 @@ bool LoopVectorizationCostModel::isMoreProfitable( return (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA); } -VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( +static void emitInvalidCostRemarks(SmallVector<InstructionVFPair> InvalidCosts, + OptimizationRemarkEmitter *ORE, + Loop *TheLoop) { + if (InvalidCosts.empty()) + return; + + // Emit a report of VFs with invalid costs in the loop. + + // Group the remarks per instruction, keeping the instruction order from + // InvalidCosts. + std::map<Instruction *, unsigned> Numbering; + unsigned I = 0; + for (auto &Pair : InvalidCosts) + if (!Numbering.count(Pair.first)) + Numbering[Pair.first] = I++; + + // Sort the list, first on instruction(number) then on VF. + sort(InvalidCosts, [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { + if (Numbering[A.first] != Numbering[B.first]) + return Numbering[A.first] < Numbering[B.first]; + ElementCountComparator ECC; + return ECC(A.second, B.second); + }); + + // For a list of ordered instruction-vf pairs: + // [(load, vf1), (load, vf2), (store, vf1)] + // Group the instructions together to emit separate remarks for: + // load (vf1, vf2) + // store (vf1) + auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); + auto Subset = ArrayRef<InstructionVFPair>(); + do { + if (Subset.empty()) + Subset = Tail.take_front(1); + + Instruction *I = Subset.front().first; + + // If the next instruction is different, or if there are no other pairs, + // emit a remark for the collated subset. e.g. + // [(load, vf1), (load, vf2))] + // to emit: + // remark: invalid costs for 'load' at VF=(vf, vf2) + if (Subset == Tail || Tail[Subset.size()].first != I) { + std::string OutString; + raw_string_ostream OS(OutString); + assert(!Subset.empty() && "Unexpected empty range"); + OS << "Instruction with invalid costs prevented vectorization at VF=("; + for (const auto &Pair : Subset) + OS << (Pair.second == Subset.front().second ? "" : ", ") << Pair.second; + OS << "):"; + if (auto *CI = dyn_cast<CallInst>(I)) + OS << " call to " << CI->getCalledFunction()->getName(); + else + OS << " " << I->getOpcodeName(); + OS.flush(); + reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); + Tail = Tail.drop_front(Subset.size()); + Subset = {}; + } else + // Grow the subset by one element + Subset = Tail.take_front(Subset.size() + 1); + } while (!Tail.empty()); +} + +VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor( const ElementCountSet &VFCandidates) { - InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first; + InstructionCost ExpectedCost = + CM.expectedCost(ElementCount::getFixed(1)).first; LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n"); assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop"); assert(VFCandidates.count(ElementCount::getFixed(1)) && "Expected Scalar VF to be a candidate"); - const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost); + const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost, + ExpectedCost); VectorizationFactor ChosenFactor = ScalarCost; - bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; + bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; if (ForceVectorization && VFCandidates.size() > 1) { // Ignore scalar width, because the user explicitly wants vectorization. // Initialize cost to max so that VF = 2 is, at least, chosen during cost @@ -5684,13 +5048,13 @@ VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( if (i.isScalar()) continue; - VectorizationCostTy C = expectedCost(i, &InvalidCosts); - VectorizationFactor Candidate(i, C.first); + LoopVectorizationCostModel::VectorizationCostTy C = + CM.expectedCost(i, &InvalidCosts); + VectorizationFactor Candidate(i, C.first, ScalarCost.ScalarCost); #ifndef NDEBUG - unsigned AssumedMinimumVscale = 1; - if (Optional<unsigned> VScale = getVScaleForTuning()) - AssumedMinimumVscale = VScale.getValue(); + unsigned AssumedMinimumVscale = + getVScaleForTuning(OrigLoop, TTI).value_or(1); unsigned Width = Candidate.Width.isScalable() ? Candidate.Width.getKnownMinValue() * AssumedMinimumVscale @@ -5718,115 +5082,51 @@ VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( ChosenFactor = Candidate; } - // Emit a report of VFs with invalid costs in the loop. - if (!InvalidCosts.empty()) { - // Group the remarks per instruction, keeping the instruction order from - // InvalidCosts. - std::map<Instruction *, unsigned> Numbering; - unsigned I = 0; - for (auto &Pair : InvalidCosts) - if (!Numbering.count(Pair.first)) - Numbering[Pair.first] = I++; - - // Sort the list, first on instruction(number) then on VF. - llvm::sort(InvalidCosts, - [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { - if (Numbering[A.first] != Numbering[B.first]) - return Numbering[A.first] < Numbering[B.first]; - ElementCountComparator ECC; - return ECC(A.second, B.second); - }); - - // For a list of ordered instruction-vf pairs: - // [(load, vf1), (load, vf2), (store, vf1)] - // Group the instructions together to emit separate remarks for: - // load (vf1, vf2) - // store (vf1) - auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); - auto Subset = ArrayRef<InstructionVFPair>(); - do { - if (Subset.empty()) - Subset = Tail.take_front(1); - - Instruction *I = Subset.front().first; - - // If the next instruction is different, or if there are no other pairs, - // emit a remark for the collated subset. e.g. - // [(load, vf1), (load, vf2))] - // to emit: - // remark: invalid costs for 'load' at VF=(vf, vf2) - if (Subset == Tail || Tail[Subset.size()].first != I) { - std::string OutString; - raw_string_ostream OS(OutString); - assert(!Subset.empty() && "Unexpected empty range"); - OS << "Instruction with invalid costs prevented vectorization at VF=("; - for (auto &Pair : Subset) - OS << (Pair.second == Subset.front().second ? "" : ", ") - << Pair.second; - OS << "):"; - if (auto *CI = dyn_cast<CallInst>(I)) - OS << " call to " << CI->getCalledFunction()->getName(); - else - OS << " " << I->getOpcodeName(); - OS.flush(); - reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); - Tail = Tail.drop_front(Subset.size()); - Subset = {}; - } else - // Grow the subset by one element - Subset = Tail.take_front(Subset.size() + 1); - } while (!Tail.empty()); - } + emitInvalidCostRemarks(InvalidCosts, ORE, OrigLoop); - if (!EnableCondStoresVectorization && NumPredStores) { - reportVectorizationFailure("There are conditional stores.", + if (!EnableCondStoresVectorization && CM.hasPredStores()) { + reportVectorizationFailure( + "There are conditional stores.", "store that is conditionally executed prevents vectorization", - "ConditionalStore", ORE, TheLoop); + "ConditionalStore", ORE, OrigLoop); ChosenFactor = ScalarCost; } LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() && - ChosenFactor.Cost >= ScalarCost.Cost) dbgs() + !isMoreProfitable(ChosenFactor, ScalarCost)) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"); LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << ChosenFactor.Width << ".\n"); return ChosenFactor; } -bool LoopVectorizationCostModel::isCandidateForEpilogueVectorization( - const Loop &L, ElementCount VF) const { +bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization( + ElementCount VF) const { // Cross iteration phis such as reductions need special handling and are // currently unsupported. - if (any_of(L.getHeader()->phis(), - [&](PHINode &Phi) { return Legal->isFirstOrderRecurrence(&Phi); })) + if (any_of(OrigLoop->getHeader()->phis(), + [&](PHINode &Phi) { return Legal->isFixedOrderRecurrence(&Phi); })) return false; // Phis with uses outside of the loop require special handling and are // currently unsupported. - for (auto &Entry : Legal->getInductionVars()) { + for (const auto &Entry : Legal->getInductionVars()) { // Look for uses of the value of the induction at the last iteration. - Value *PostInc = Entry.first->getIncomingValueForBlock(L.getLoopLatch()); + Value *PostInc = + Entry.first->getIncomingValueForBlock(OrigLoop->getLoopLatch()); for (User *U : PostInc->users()) - if (!L.contains(cast<Instruction>(U))) + if (!OrigLoop->contains(cast<Instruction>(U))) return false; // Look for uses of penultimate value of the induction. for (User *U : Entry.first->users()) - if (!L.contains(cast<Instruction>(U))) + if (!OrigLoop->contains(cast<Instruction>(U))) return false; } - // Induction variables that are widened require special handling that is - // currently not supported. - if (any_of(Legal->getInductionVars(), [&](auto &Entry) { - return !(this->isScalarAfterVectorization(Entry.first, VF) || - this->isProfitableToScalarize(Entry.first, VF)); - })) - return false; - // Epilogue vectorization code has not been auditted to ensure it handles // non-latch exits properly. It may be fine, but it needs auditted and // tested. - if (L.getExitingBlock() != L.getLoopLatch()) + if (OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch()) return false; return true; @@ -5838,64 +5138,66 @@ bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( // as register pressure, code size increase and cost of extra branches into // account. For now we apply a very crude heuristic and only consider loops // with vectorization factors larger than a certain value. + + // Allow the target to opt out entirely. + if (!TTI.preferEpilogueVectorization()) + return false; + // We also consider epilogue vectorization unprofitable for targets that don't // consider interleaving beneficial (eg. MVE). - if (TTI.getMaxInterleaveFactor(VF.getKnownMinValue()) <= 1) + if (TTI.getMaxInterleaveFactor(VF) <= 1) return false; - // FIXME: We should consider changing the threshold for scalable - // vectors to take VScaleForTuning into account. - if (VF.getKnownMinValue() >= EpilogueVectorizationMinVF) + + unsigned Multiplier = 1; + if (VF.isScalable()) + Multiplier = getVScaleForTuning(TheLoop, TTI).value_or(1); + if ((Multiplier * VF.getKnownMinValue()) >= EpilogueVectorizationMinVF) return true; return false; } -VectorizationFactor -LoopVectorizationCostModel::selectEpilogueVectorizationFactor( - const ElementCount MainLoopVF, const LoopVectorizationPlanner &LVP) { +VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor( + const ElementCount MainLoopVF, unsigned IC) { VectorizationFactor Result = VectorizationFactor::Disabled(); if (!EnableEpilogueVectorization) { - LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n";); + LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n"); return Result; } - if (!isScalarEpilogueAllowed()) { - LLVM_DEBUG( - dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " - "allowed.\n";); + if (!CM.isScalarEpilogueAllowed()) { + LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because no " + "epilogue is allowed.\n"); return Result; } // Not really a cost consideration, but check for unsupported cases here to // simplify the logic. - if (!isCandidateForEpilogueVectorization(*TheLoop, MainLoopVF)) { - LLVM_DEBUG( - dbgs() << "LEV: Unable to vectorize epilogue because the loop is " - "not a supported candidate.\n";); + if (!isCandidateForEpilogueVectorization(MainLoopVF)) { + LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because the loop " + "is not a supported candidate.\n"); return Result; } if (EpilogueVectorizationForceVF > 1) { - LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n";); + LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n"); ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); - if (LVP.hasPlanWithVF(ForcedEC)) - return {ForcedEC, 0}; + if (hasPlanWithVF(ForcedEC)) + return {ForcedEC, 0, 0}; else { - LLVM_DEBUG( - dbgs() - << "LEV: Epilogue vectorization forced factor is not viable.\n";); + LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization forced factor is not " + "viable.\n"); return Result; } } - if (TheLoop->getHeader()->getParent()->hasOptSize() || - TheLoop->getHeader()->getParent()->hasMinSize()) { + if (OrigLoop->getHeader()->getParent()->hasOptSize() || + OrigLoop->getHeader()->getParent()->hasMinSize()) { LLVM_DEBUG( - dbgs() - << "LEV: Epilogue vectorization skipped due to opt for size.\n";); + dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n"); return Result; } - if (!isEpilogueVectorizationProfitable(MainLoopVF)) { + if (!CM.isEpilogueVectorizationProfitable(MainLoopVF)) { LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"); return Result; @@ -5907,21 +5209,48 @@ LoopVectorizationCostModel::selectEpilogueVectorizationFactor( ElementCount EstimatedRuntimeVF = MainLoopVF; if (MainLoopVF.isScalable()) { EstimatedRuntimeVF = ElementCount::getFixed(MainLoopVF.getKnownMinValue()); - if (Optional<unsigned> VScale = getVScaleForTuning()) - EstimatedRuntimeVF *= VScale.getValue(); + if (std::optional<unsigned> VScale = getVScaleForTuning(OrigLoop, TTI)) + EstimatedRuntimeVF *= *VScale; } - for (auto &NextVF : ProfitableVFs) - if (((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && - ElementCount::isKnownLT(NextVF.Width, EstimatedRuntimeVF)) || - ElementCount::isKnownLT(NextVF.Width, MainLoopVF)) && - (Result.Width.isScalar() || isMoreProfitable(NextVF, Result)) && - LVP.hasPlanWithVF(NextVF.Width)) + ScalarEvolution &SE = *PSE.getSE(); + Type *TCType = Legal->getWidestInductionType(); + const SCEV *RemainingIterations = nullptr; + for (auto &NextVF : ProfitableVFs) { + // Skip candidate VFs without a corresponding VPlan. + if (!hasPlanWithVF(NextVF.Width)) + continue; + + // Skip candidate VFs with widths >= the estimate runtime VF (scalable + // vectors) or the VF of the main loop (fixed vectors). + if ((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && + ElementCount::isKnownGE(NextVF.Width, EstimatedRuntimeVF)) || + ElementCount::isKnownGE(NextVF.Width, MainLoopVF)) + continue; + + // If NextVF is greater than the number of remaining iterations, the + // epilogue loop would be dead. Skip such factors. + if (!MainLoopVF.isScalable() && !NextVF.Width.isScalable()) { + // TODO: extend to support scalable VFs. + if (!RemainingIterations) { + const SCEV *TC = createTripCountSCEV(TCType, PSE, OrigLoop); + RemainingIterations = SE.getURemExpr( + TC, SE.getConstant(TCType, MainLoopVF.getKnownMinValue() * IC)); + } + if (SE.isKnownPredicate( + CmpInst::ICMP_UGT, + SE.getConstant(TCType, NextVF.Width.getKnownMinValue()), + RemainingIterations)) + continue; + } + + if (Result.Width.isScalar() || isMoreProfitable(NextVF, Result)) Result = NextVF; + } if (Result != VectorizationFactor::Disabled()) LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = " - << Result.Width << "\n";); + << Result.Width << "\n"); return Result; } @@ -5937,7 +5266,7 @@ LoopVectorizationCostModel::getSmallestAndWidestTypes() { // Reset MaxWidth so that we can find the smallest type used by recurrences // in the loop. MaxWidth = -1U; - for (auto &PhiDescriptorPair : Legal->getReductionVars()) { + for (const auto &PhiDescriptorPair : Legal->getReductionVars()) { const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; // When finding the min width used by the recurrence we need to account // for casts on the input operands of the recurrence. @@ -5949,9 +5278,9 @@ LoopVectorizationCostModel::getSmallestAndWidestTypes() { } else { for (Type *T : ElementTypesInLoop) { MinWidth = std::min<unsigned>( - MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); + MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); MaxWidth = std::max<unsigned>( - MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); + MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); } } return {MinWidth, MaxWidth}; @@ -6000,8 +5329,9 @@ void LoopVectorizationCostModel::collectElementTypesForWidening() { } } -unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, - unsigned LoopCost) { +unsigned +LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, + InstructionCost LoopCost) { // -- The interleave heuristics -- // We interleave the loop in order to expose ILP and reduce the loop overhead. // There are many micro-architectural considerations that we can't predict @@ -6020,7 +5350,7 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, return 1; // We used the distance for the interleave count. - if (Legal->getMaxSafeDepDistBytes() != -1U) + if (!Legal->isSafeForAnyVectorWidth()) return 1; auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); @@ -6037,9 +5367,8 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, // If we did not calculate the cost for VF (because the user selected the VF) // then we calculate the cost of VF here. if (LoopCost == 0) { - InstructionCost C = expectedCost(VF).first; - assert(C.isValid() && "Expected to have chosen a VF with valid cost"); - LoopCost = *C.getValue(); + LoopCost = expectedCost(VF).first; + assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost"); // Loop body is free and there is no need for interleaving. if (LoopCost == 0) @@ -6083,20 +5412,19 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; - unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); + unsigned TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs) / + MaxLocalUsers); // Don't count the induction variable as interleaved. if (EnableIndVarRegisterHeur) { - TmpIC = - PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / - std::max(1U, (MaxLocalUsers - 1))); + TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs - 1) / + std::max(1U, (MaxLocalUsers - 1))); } IC = std::min(IC, TmpIC); } // Clamp the interleave ranges to reasonable counts. - unsigned MaxInterleaveCount = - TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); + unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF); // Check if the user has overridden the max. if (VF.isScalar()) { @@ -6107,21 +5435,45 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; } - // If trip count is known or estimated compile time constant, limit the - // interleave count to be less than the trip count divided by VF, provided it - // is at least 1. - // - // For scalable vectors we can't know if interleaving is beneficial. It may - // not be beneficial for small loops if none of the lanes in the second vector - // iterations is enabled. However, for larger loops, there is likely to be a - // similar benefit as for fixed-width vectors. For now, we choose to leave - // the InterleaveCount as if vscale is '1', although if some information about - // the vector is known (e.g. min vector size), we can make a better decision. - if (BestKnownTC) { - MaxInterleaveCount = - std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); - // Make sure MaxInterleaveCount is greater than 0. - MaxInterleaveCount = std::max(1u, MaxInterleaveCount); + unsigned EstimatedVF = VF.getKnownMinValue(); + if (VF.isScalable()) { + if (std::optional<unsigned> VScale = getVScaleForTuning(TheLoop, TTI)) + EstimatedVF *= *VScale; + } + assert(EstimatedVF >= 1 && "Estimated VF shouldn't be less than 1"); + + unsigned KnownTC = PSE.getSE()->getSmallConstantTripCount(TheLoop); + if (KnownTC) { + // If trip count is known we select between two prospective ICs, where + // 1) the aggressive IC is capped by the trip count divided by VF + // 2) the conservative IC is capped by the trip count divided by (VF * 2) + // The final IC is selected in a way that the epilogue loop trip count is + // minimized while maximizing the IC itself, so that we either run the + // vector loop at least once if it generates a small epilogue loop, or else + // we run the vector loop at least twice. + + unsigned InterleaveCountUB = bit_floor( + std::max(1u, std::min(KnownTC / EstimatedVF, MaxInterleaveCount))); + unsigned InterleaveCountLB = bit_floor(std::max( + 1u, std::min(KnownTC / (EstimatedVF * 2), MaxInterleaveCount))); + MaxInterleaveCount = InterleaveCountLB; + + if (InterleaveCountUB != InterleaveCountLB) { + unsigned TailTripCountUB = (KnownTC % (EstimatedVF * InterleaveCountUB)); + unsigned TailTripCountLB = (KnownTC % (EstimatedVF * InterleaveCountLB)); + // If both produce same scalar tail, maximize the IC to do the same work + // in fewer vector loop iterations + if (TailTripCountUB == TailTripCountLB) + MaxInterleaveCount = InterleaveCountUB; + } + } else if (BestKnownTC) { + // If trip count is an estimated compile time constant, limit the + // IC to be capped by the trip count divided by VF * 2, such that the vector + // loop runs at least twice to make interleaving seem profitable when there + // is an epilogue loop present. Since exact Trip count is not known we + // choose to be conservative in our IC estimate. + MaxInterleaveCount = bit_floor(std::max( + 1u, std::min(*BestKnownTC / (EstimatedVF * 2), MaxInterleaveCount))); } assert(MaxInterleaveCount > 0 && @@ -6144,9 +5496,15 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, return IC; } - // Note that if we've already vectorized the loop we will have done the - // runtime check and so interleaving won't require further checks. - bool InterleavingRequiresRuntimePointerCheck = + // For any scalar loop that either requires runtime checks or predication we + // are better off leaving this to the unroller. Note that if we've already + // vectorized the loop we will have done the runtime check and so interleaving + // won't require further checks. + bool ScalarInterleavingRequiresPredication = + (VF.isScalar() && any_of(TheLoop->blocks(), [this](BasicBlock *BB) { + return Legal->blockNeedsPredication(BB); + })); + bool ScalarInterleavingRequiresRuntimePointerCheck = (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); // We want to interleave small loops in order to reduce the loop overhead and @@ -6156,12 +5514,13 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, << "LV: VF is " << VF << '\n'); const bool AggressivelyInterleaveReductions = TTI.enableAggressiveInterleaving(HasReductions); - if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) { + if (!ScalarInterleavingRequiresRuntimePointerCheck && + !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { // We assume that the cost overhead is 1 and we use the cost model // to estimate the cost of the loop and interleave until the cost of the // loop overhead is about 5% of the cost of the loop. - unsigned SmallIC = - std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); + unsigned SmallIC = std::min(IC, (unsigned)llvm::bit_floor<uint64_t>( + SmallLoopCost / *LoopCost.getValue())); // Interleave until store/load ports (estimated by max interleave count) are // saturated. @@ -6178,7 +5537,7 @@ unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, HasReductions && any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { const RecurrenceDescriptor &RdxDesc = Reduction.second; - return RecurrenceDescriptor::isSelectCmpRecurrenceKind( + return RecurrenceDescriptor::isAnyOfRecurrenceKind( RdxDesc.getRecurrenceKind()); }); if (HasSelectCmpReductions) { @@ -6276,9 +5635,10 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { IntervalMap EndPoint; // Saves the list of instruction indices that are used in the loop. SmallPtrSet<Instruction *, 8> Ends; - // Saves the list of values that are used in the loop but are - // defined outside the loop, such as arguments and constants. - SmallPtrSet<Value *, 8> LoopInvariants; + // Saves the list of values that are used in the loop but are defined outside + // the loop (not including non-instruction values such as arguments and + // constants). + SmallSetVector<Instruction *, 8> LoopInvariants; for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { for (Instruction &I : BB->instructionsWithoutDebug()) { @@ -6289,6 +5649,9 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { auto *Instr = dyn_cast<Instruction>(U); // Ignore non-instruction values such as arguments, constants, etc. + // FIXME: Might need some motivation why these values are ignored. If + // for example an argument is used inside the loop it will increase the + // register pressure (so shouldn't we add it to LoopInvariants). if (!Instr) continue; @@ -6319,16 +5682,11 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { LLVM_DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); - // A lambda that gets the register usage for the given type and VF. const auto &TTICapture = TTI; auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned { if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty)) return 0; - InstructionCost::CostType RegUsage = - *TTICapture.getRegUsageForType(VectorType::get(Ty, VF)).getValue(); - assert(RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && - "Nonsensical values for register usage."); - return RegUsage; + return TTICapture.getRegUsageForType(VectorType::get(Ty, VF)); }; for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { @@ -6347,46 +5705,48 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { if (ValuesToIgnore.count(I)) continue; + collectInLoopReductions(); + // For each VF find the maximum usage of registers. for (unsigned j = 0, e = VFs.size(); j < e; ++j) { - // Count the number of live intervals. + // Count the number of registers used, per register class, given all open + // intervals. + // Note that elements in this SmallMapVector will be default constructed + // as 0. So we can use "RegUsage[ClassID] += n" in the code below even if + // there is no previous entry for ClassID. SmallMapVector<unsigned, unsigned, 4> RegUsage; if (VFs[j].isScalar()) { - for (auto Inst : OpenIntervals) { - unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); - if (RegUsage.find(ClassID) == RegUsage.end()) - RegUsage[ClassID] = 1; - else - RegUsage[ClassID] += 1; + for (auto *Inst : OpenIntervals) { + unsigned ClassID = + TTI.getRegisterClassForType(false, Inst->getType()); + // FIXME: The target might use more than one register for the type + // even in the scalar case. + RegUsage[ClassID] += 1; } } else { collectUniformsAndScalars(VFs[j]); - for (auto Inst : OpenIntervals) { + for (auto *Inst : OpenIntervals) { // Skip ignored values for VF > 1. if (VecValuesToIgnore.count(Inst)) continue; if (isScalarAfterVectorization(Inst, VFs[j])) { - unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); - if (RegUsage.find(ClassID) == RegUsage.end()) - RegUsage[ClassID] = 1; - else - RegUsage[ClassID] += 1; + unsigned ClassID = + TTI.getRegisterClassForType(false, Inst->getType()); + // FIXME: The target might use more than one register for the type + // even in the scalar case. + RegUsage[ClassID] += 1; } else { - unsigned ClassID = TTI.getRegisterClassForType(true, Inst->getType()); - if (RegUsage.find(ClassID) == RegUsage.end()) - RegUsage[ClassID] = GetRegUsage(Inst->getType(), VFs[j]); - else - RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); + unsigned ClassID = + TTI.getRegisterClassForType(true, Inst->getType()); + RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); } } } for (auto& pair : RegUsage) { - if (MaxUsages[j].find(pair.first) != MaxUsages[j].end()) - MaxUsages[j][pair.first] = std::max(MaxUsages[j][pair.first], pair.second); - else - MaxUsages[j][pair.first] = pair.second; + auto &Entry = MaxUsages[j][pair.first]; + Entry = std::max(Entry, pair.second); } } @@ -6398,17 +5758,24 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { } for (unsigned i = 0, e = VFs.size(); i < e; ++i) { + // Note that elements in this SmallMapVector will be default constructed + // as 0. So we can use "Invariant[ClassID] += n" in the code below even if + // there is no previous entry for ClassID. SmallMapVector<unsigned, unsigned, 4> Invariant; - for (auto Inst : LoopInvariants) { - unsigned Usage = - VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); + for (auto *Inst : LoopInvariants) { + // FIXME: The target might use more than one register for the type + // even in the scalar case. + bool IsScalar = all_of(Inst->users(), [&](User *U) { + auto *I = cast<Instruction>(U); + return TheLoop != LI->getLoopFor(I->getParent()) || + isScalarAfterVectorization(I, VFs[i]); + }); + + ElementCount VF = IsScalar ? ElementCount::getFixed(1) : VFs[i]; unsigned ClassID = - TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); - if (Invariant.find(ClassID) == Invariant.end()) - Invariant[ClassID] = Usage; - else - Invariant[ClassID] += Usage; + TTI.getRegisterClassForType(VF.isVector(), Inst->getType()); + Invariant[ClassID] += GetRegUsage(Inst->getType(), VF); } LLVM_DEBUG({ @@ -6447,7 +5814,8 @@ bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, // from moving "masked load/store" check from legality to cost model. // Masked Load/Gather emulation was previously never allowed. // Limited number of Masked Store/Scatter emulation was allowed. - assert(isPredicatedInst(I, VF) && "Expecting a scalar emulated instruction"); + assert((isPredicatedInst(I)) && + "Expecting a scalar emulated instruction"); return isa<LoadInst>(I) || (isa<StoreInst>(I) && NumPredStores > NumberOfStoresToPredicate); @@ -6458,8 +5826,7 @@ void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { // instructions to scalarize, there's nothing to do. Collection may already // have occurred if we have a user-selected VF and are now computing the // expected cost for interleaving. - if (VF.isScalar() || VF.isZero() || - InstsToScalarize.find(VF) != InstsToScalarize.end()) + if (VF.isScalar() || VF.isZero() || InstsToScalarize.contains(VF)) return; // Initialize a mapping for VF in InstsToScalalarize. If we find that it's @@ -6467,6 +5834,8 @@ void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { // map will indicate that we've analyzed it already. ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; + PredicatedBBsAfterVectorization[VF].clear(); + // Find all the instructions that are scalar with predication in the loop and // determine if it would be better to not if-convert the blocks they are in. // If so, we also record the instructions to scalarize. @@ -6484,12 +5853,12 @@ void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { computePredInstDiscount(&I, ScalarCosts, VF) >= 0) ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); // Remember that BB will remain after vectorization. - PredicatedBBsAfterVectorization.insert(BB); + PredicatedBBsAfterVectorization[VF].insert(BB); } } } -int LoopVectorizationCostModel::computePredInstDiscount( +InstructionCost LoopVectorizationCostModel::computePredInstDiscount( Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { assert(!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated"); @@ -6546,7 +5915,7 @@ int LoopVectorizationCostModel::computePredInstDiscount( Instruction *I = Worklist.pop_back_val(); // If we've already analyzed the instruction, there's nothing to do. - if (ScalarCosts.find(I) != ScalarCosts.end()) + if (ScalarCosts.contains(I)) continue; // Compute the cost of the vector instruction. Note that this cost already @@ -6563,13 +5932,14 @@ int LoopVectorizationCostModel::computePredInstDiscount( // Compute the scalarization overhead of needed insertelement instructions // and phi nodes. + TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { ScalarCost += TTI.getScalarizationOverhead( cast<VectorType>(ToVectorTy(I->getType(), VF)), - APInt::getAllOnes(VF.getFixedValue()), true, false); + APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ true, + /*Extract*/ false, CostKind); ScalarCost += - VF.getFixedValue() * - TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput); + VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind); } // Compute the scalarization overhead of needed extractelement @@ -6585,7 +5955,8 @@ int LoopVectorizationCostModel::computePredInstDiscount( else if (needsExtract(J, VF)) { ScalarCost += TTI.getScalarizationOverhead( cast<VectorType>(ToVectorTy(J->getType(), VF)), - APInt::getAllOnes(VF.getFixedValue()), false, true); + APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false, + /*Extract*/ true, CostKind); } } @@ -6598,7 +5969,7 @@ int LoopVectorizationCostModel::computePredInstDiscount( ScalarCosts[I] = ScalarCost; } - return *Discount.getValue(); + return Discount; } LoopVectorizationCostModel::VectorizationCostTy @@ -6682,11 +6053,6 @@ static const SCEV *getAddressAccessSCEV( return PSE.getSCEV(Ptr); } -static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { - return Legal->hasStride(I->getOperand(0)) || - Legal->hasStride(I->getOperand(1)); -} - InstructionCost LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, ElementCount VF) { @@ -6714,19 +6080,20 @@ LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, // Don't pass *I here, since it is scalar but will actually be part of a // vectorized loop where the user of it is a vectorized instruction. + TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; const Align Alignment = getLoadStoreAlignment(I); - Cost += VF.getKnownMinValue() * - TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, - AS, TTI::TCK_RecipThroughput); + Cost += VF.getKnownMinValue() * TTI.getMemoryOpCost(I->getOpcode(), + ValTy->getScalarType(), + Alignment, AS, CostKind); // Get the overhead of the extractelement and insertelement instructions // we might create due to scalarization. - Cost += getScalarizationOverhead(I, VF); + Cost += getScalarizationOverhead(I, VF, CostKind); // If we have a predicated load/store, it will need extra i1 extracts and // conditional branches, but may not be executed for each vector lane. Scale // the cost by the probability of executing the predicated block. - if (isPredicatedInst(I, VF)) { + if (isPredicatedInst(I)) { Cost /= getReciprocalPredBlockProb(); // Add the cost of an i1 extract and a branch @@ -6734,8 +6101,8 @@ LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); Cost += TTI.getScalarizationOverhead( Vec_i1Ty, APInt::getAllOnes(VF.getKnownMinValue()), - /*Insert=*/false, /*Extract=*/true); - Cost += TTI.getCFInstrCost(Instruction::Br, TTI::TCK_RecipThroughput); + /*Insert=*/false, /*Extract=*/true, CostKind); + Cost += TTI.getCFInstrCost(Instruction::Br, CostKind); if (useEmulatedMaskMemRefHack(I, VF)) // Artificially setting to a high enough value to practically disable @@ -6760,24 +6127,26 @@ LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, "Stride should be 1 or -1 for consecutive memory access"); const Align Alignment = getLoadStoreAlignment(I); InstructionCost Cost = 0; - if (Legal->isMaskRequired(I)) + if (Legal->isMaskRequired(I)) { Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, CostKind); - else + } else { + TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, - CostKind, I); + CostKind, OpInfo, I); + } bool Reverse = ConsecutiveStride < 0; if (Reverse) - Cost += - TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); + Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, + std::nullopt, CostKind, 0); return Cost; } InstructionCost LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, ElementCount VF) { - assert(Legal->isUniformMemOp(*I)); + assert(Legal->isUniformMemOp(*I, VF)); Type *ValTy = getLoadStoreType(I); auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); @@ -6792,14 +6161,14 @@ LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, } StoreInst *SI = cast<StoreInst>(I); - bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); + bool isLoopInvariantStoreValue = Legal->isInvariant(SI->getValueOperand()); return TTI.getAddressComputationCost(ValTy) + TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, CostKind) + (isLoopInvariantStoreValue ? 0 : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, - VF.getKnownMinValue() - 1)); + CostKind, VF.getKnownMinValue() - 1)); } InstructionCost @@ -6819,14 +6188,10 @@ LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, InstructionCost LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, ElementCount VF) { - // TODO: Once we have support for interleaving with scalable vectors - // we can calculate the cost properly here. - if (VF.isScalable()) - return InstructionCost::getInvalid(); - Type *ValTy = getLoadStoreType(I); auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); unsigned AS = getLoadStoreAddressSpace(I); + enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; auto Group = getInterleavedAccessGroup(I); assert(Group && "Fail to get an interleaved access group."); @@ -6846,25 +6211,27 @@ LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, (isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor())); InstructionCost Cost = TTI.getInterleavedMemoryOpCost( I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), - AS, TTI::TCK_RecipThroughput, Legal->isMaskRequired(I), UseMaskForGaps); + AS, CostKind, Legal->isMaskRequired(I), UseMaskForGaps); if (Group->isReverse()) { // TODO: Add support for reversed masked interleaved access. assert(!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported."); - Cost += - Group->getNumMembers() * - TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); + Cost += Group->getNumMembers() * + TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, + std::nullopt, CostKind, 0); } return Cost; } -Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( - Instruction *I, ElementCount VF, Type *Ty, TTI::TargetCostKind CostKind) { +std::optional<InstructionCost> +LoopVectorizationCostModel::getReductionPatternCost( + Instruction *I, ElementCount VF, Type *Ty, + TTI::TargetCostKind CostKind) const { using namespace llvm::PatternMatch; // Early exit for no inloop reductions - if (InLoopReductionChains.empty() || VF.isScalar() || !isa<VectorType>(Ty)) - return None; + if (InLoopReductions.empty() || VF.isScalar() || !isa<VectorType>(Ty)) + return std::nullopt; auto *VectorTy = cast<VectorType>(Ty); // We are looking for a pattern of, and finding the minimal acceptable cost: @@ -6882,27 +6249,26 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( Instruction *RetI = I; if (match(RetI, m_ZExtOrSExt(m_Value()))) { if (!RetI->hasOneUser()) - return None; + return std::nullopt; RetI = RetI->user_back(); } - if (match(RetI, m_Mul(m_Value(), m_Value())) && + + if (match(RetI, m_OneUse(m_Mul(m_Value(), m_Value()))) && RetI->user_back()->getOpcode() == Instruction::Add) { - if (!RetI->hasOneUser()) - return None; RetI = RetI->user_back(); } // Test if the found instruction is a reduction, and if not return an invalid // cost specifying the parent to use the original cost modelling. if (!InLoopReductionImmediateChains.count(RetI)) - return None; + return std::nullopt; // Find the reduction this chain is a part of and calculate the basic cost of // the reduction on its own. - Instruction *LastChain = InLoopReductionImmediateChains[RetI]; + Instruction *LastChain = InLoopReductionImmediateChains.at(RetI); Instruction *ReductionPhi = LastChain; while (!isa<PHINode>(ReductionPhi)) - ReductionPhi = InLoopReductionImmediateChains[ReductionPhi]; + ReductionPhi = InLoopReductionImmediateChains.at(ReductionPhi); const RecurrenceDescriptor &RdxDesc = Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second; @@ -6931,7 +6297,7 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); Instruction *Op0, *Op1; - if (RedOp && + if (RedOp && RdxDesc.getOpcode() == Instruction::Add && match(RedOp, m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && match(Op0, m_ZExtOrSExt(m_Value())) && @@ -6940,7 +6306,7 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { - // Matched reduce(ext(mul(ext(A), ext(B))) + // Matched reduce.add(ext(mul(ext(A), ext(B))) // Note that the extend opcodes need to all match, or if A==B they will have // been converted to zext(mul(sext(A), sext(A))) as it is known positive, // which is equally fine. @@ -6957,9 +6323,8 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, TTI::CastContextHint::None, CostKind, RedOp); - InstructionCost RedCost = TTI.getExtendedAddReductionCost( - /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, - CostKind); + InstructionCost RedCost = TTI.getMulAccReductionCost( + IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); if (RedCost.isValid() && RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) @@ -6969,16 +6334,16 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( // Matched reduce(ext(A)) bool IsUnsigned = isa<ZExtInst>(RedOp); auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); - InstructionCost RedCost = TTI.getExtendedAddReductionCost( - /*IsMLA=*/false, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, - CostKind); + InstructionCost RedCost = TTI.getExtendedReductionCost( + RdxDesc.getOpcode(), IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, + RdxDesc.getFastMathFlags(), CostKind); InstructionCost ExtCost = TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, TTI::CastContextHint::None, CostKind, RedOp); if (RedCost.isValid() && RedCost < BaseCost + ExtCost) return I == RetI ? RedCost : 0; - } else if (RedOp && + } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add && match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { if (match(Op0, m_ZExtOrSExt(m_Value())) && Op0->getOpcode() == Op1->getOpcode() && @@ -6991,7 +6356,7 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( : Op0Ty; auto *ExtType = VectorType::get(LargestOpTy, VectorTy); - // Matched reduce(mul(ext(A), ext(B))), where the two ext may be of + // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of // different sizes. We take the largest type as the ext to reduce, and add // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). InstructionCost ExtCost0 = TTI.getCastInstrCost( @@ -7003,9 +6368,8 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( InstructionCost MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); - InstructionCost RedCost = TTI.getExtendedAddReductionCost( - /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, - CostKind); + InstructionCost RedCost = TTI.getMulAccReductionCost( + IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); InstructionCost ExtraExtCost = 0; if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; @@ -7019,20 +6383,19 @@ Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) return I == RetI ? RedCost : 0; } else if (!match(I, m_ZExtOrSExt(m_Value()))) { - // Matched reduce(mul()) + // Matched reduce.add(mul()) InstructionCost MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); - InstructionCost RedCost = TTI.getExtendedAddReductionCost( - /*IsMLA=*/true, true, RdxDesc.getRecurrenceType(), VectorTy, - CostKind); + InstructionCost RedCost = TTI.getMulAccReductionCost( + true, RdxDesc.getRecurrenceType(), VectorTy, CostKind); if (RedCost.isValid() && RedCost < MulCost + BaseCost) return I == RetI ? RedCost : 0; } } - return I == RetI ? Optional<InstructionCost>(BaseCost) : None; + return I == RetI ? std::optional<InstructionCost>(BaseCost) : std::nullopt; } InstructionCost @@ -7045,9 +6408,10 @@ LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, const Align Alignment = getLoadStoreAlignment(I); unsigned AS = getLoadStoreAddressSpace(I); + TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); return TTI.getAddressComputationCost(ValTy) + TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, - TTI::TCK_RecipThroughput, I); + TTI::TCK_RecipThroughput, OpInfo, I); } return getWideningCost(I, VF); } @@ -7079,18 +6443,24 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, bool TypeNotScalarized = false; if (VF.isVector() && VectorTy->isVectorTy()) { - unsigned NumParts = TTI.getNumberOfParts(VectorTy); - if (NumParts) - TypeNotScalarized = NumParts < VF.getKnownMinValue(); - else + if (unsigned NumParts = TTI.getNumberOfParts(VectorTy)) { + if (VF.isScalable()) + // <vscale x 1 x iN> is assumed to be profitable over iN because + // scalable registers are a distinct register class from scalar ones. + // If we ever find a target which wants to lower scalable vectors + // back to scalars, we'll need to update this code to explicitly + // ask TTI about the register class uses for each part. + TypeNotScalarized = NumParts <= VF.getKnownMinValue(); + else + TypeNotScalarized = NumParts < VF.getKnownMinValue(); + } else C = InstructionCost::getInvalid(); } return VectorizationCostTy(C, TypeNotScalarized); } -InstructionCost -LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, - ElementCount VF) const { +InstructionCost LoopVectorizationCostModel::getScalarizationOverhead( + Instruction *I, ElementCount VF, TTI::TargetCostKind CostKind) const { // There is no mechanism yet to create a scalable scalarization loop, // so this is currently Invalid. @@ -7105,8 +6475,9 @@ LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, if (!RetTy->isVoidTy() && (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) Cost += TTI.getScalarizationOverhead( - cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), true, - false); + cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), + /*Insert*/ true, + /*Extract*/ false, CostKind); // Some targets keep addresses scalar. if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) @@ -7126,7 +6497,7 @@ LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, for (auto *V : filterExtractingOperands(Ops, VF)) Tys.push_back(MaybeVectorizeType(V->getType(), VF)); return Cost + TTI.getOperandsScalarizationOverhead( - filterExtractingOperands(Ops, VF), Tys); + filterExtractingOperands(Ops, VF), Tys, CostKind); } void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { @@ -7147,22 +6518,48 @@ void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF)) NumPredStores++; - if (Legal->isUniformMemOp(I)) { - // TODO: Avoid replicating loads and stores instead of - // relying on instcombine to remove them. + if (Legal->isUniformMemOp(I, VF)) { + auto isLegalToScalarize = [&]() { + if (!VF.isScalable()) + // Scalarization of fixed length vectors "just works". + return true; + + // We have dedicated lowering for unpredicated uniform loads and + // stores. Note that even with tail folding we know that at least + // one lane is active (i.e. generalized predication is not possible + // here), and the logic below depends on this fact. + if (!foldTailByMasking()) + return true; + + // For scalable vectors, a uniform memop load is always + // uniform-by-parts and we know how to scalarize that. + if (isa<LoadInst>(I)) + return true; + + // A uniform store isn't neccessarily uniform-by-part + // and we can't assume scalarization. + auto &SI = cast<StoreInst>(I); + return TheLoop->isLoopInvariant(SI.getValueOperand()); + }; + + const InstructionCost GatherScatterCost = + isLegalGatherOrScatter(&I, VF) ? + getGatherScatterCost(&I, VF) : InstructionCost::getInvalid(); + // Load: Scalar load + broadcast // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract - InstructionCost Cost; - if (isa<StoreInst>(&I) && VF.isScalable() && - isLegalGatherOrScatter(&I, VF)) { - Cost = getGatherScatterCost(&I, VF); - setWideningDecision(&I, VF, CM_GatherScatter, Cost); - } else { - assert((isa<LoadInst>(&I) || !VF.isScalable()) && - "Cannot yet scalarize uniform stores"); - Cost = getUniformMemOpCost(&I, VF); - setWideningDecision(&I, VF, CM_Scalarize, Cost); - } + // FIXME: This cost is a significant under-estimate for tail folded + // memory ops. + const InstructionCost ScalarizationCost = isLegalToScalarize() ? + getUniformMemOpCost(&I, VF) : InstructionCost::getInvalid(); + + // Choose better solution for the current VF, Note that Invalid + // costs compare as maximumal large. If both are invalid, we get + // scalable invalid which signals a failure and a vectorization abort. + if (GatherScatterCost < ScalarizationCost) + setWideningDecision(&I, VF, CM_GatherScatter, GatherScatterCost); + else + setWideningDecision(&I, VF, CM_Scalarize, ScalarizationCost); continue; } @@ -7289,6 +6686,168 @@ void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { } } +void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) { + assert(!VF.isScalar() && + "Trying to set a vectorization decision for a scalar VF"); + + for (BasicBlock *BB : TheLoop->blocks()) { + // For each instruction in the old loop. + for (Instruction &I : *BB) { + CallInst *CI = dyn_cast<CallInst>(&I); + + if (!CI) + continue; + + InstructionCost ScalarCost = InstructionCost::getInvalid(); + InstructionCost VectorCost = InstructionCost::getInvalid(); + InstructionCost IntrinsicCost = InstructionCost::getInvalid(); + TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; + + Function *ScalarFunc = CI->getCalledFunction(); + Type *ScalarRetTy = CI->getType(); + SmallVector<Type *, 4> Tys, ScalarTys; + bool MaskRequired = Legal->isMaskRequired(CI); + for (auto &ArgOp : CI->args()) + ScalarTys.push_back(ArgOp->getType()); + + // Compute corresponding vector type for return value and arguments. + Type *RetTy = ToVectorTy(ScalarRetTy, VF); + for (Type *ScalarTy : ScalarTys) + Tys.push_back(ToVectorTy(ScalarTy, VF)); + + // An in-loop reduction using an fmuladd intrinsic is a special case; + // we don't want the normal cost for that intrinsic. + if (RecurrenceDescriptor::isFMulAddIntrinsic(CI)) + if (auto RedCost = getReductionPatternCost(CI, VF, RetTy, CostKind)) { + setCallWideningDecision(CI, VF, CM_IntrinsicCall, nullptr, + getVectorIntrinsicIDForCall(CI, TLI), + std::nullopt, *RedCost); + continue; + } + + // Estimate cost of scalarized vector call. The source operands are + // assumed to be vectors, so we need to extract individual elements from + // there, execute VF scalar calls, and then gather the result into the + // vector return value. + InstructionCost ScalarCallCost = + TTI.getCallInstrCost(ScalarFunc, ScalarRetTy, ScalarTys, CostKind); + + // Compute costs of unpacking argument values for the scalar calls and + // packing the return values to a vector. + InstructionCost ScalarizationCost = + getScalarizationOverhead(CI, VF, CostKind); + + ScalarCost = ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; + + // Find the cost of vectorizing the call, if we can find a suitable + // vector variant of the function. + bool UsesMask = false; + VFInfo FuncInfo; + Function *VecFunc = nullptr; + // Search through any available variants for one we can use at this VF. + for (VFInfo &Info : VFDatabase::getMappings(*CI)) { + // Must match requested VF. + if (Info.Shape.VF != VF) + continue; + + // Must take a mask argument if one is required + if (MaskRequired && !Info.isMasked()) + continue; + + // Check that all parameter kinds are supported + bool ParamsOk = true; + for (VFParameter Param : Info.Shape.Parameters) { + switch (Param.ParamKind) { + case VFParamKind::Vector: + break; + case VFParamKind::OMP_Uniform: { + Value *ScalarParam = CI->getArgOperand(Param.ParamPos); + // Make sure the scalar parameter in the loop is invariant. + if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(ScalarParam), + TheLoop)) + ParamsOk = false; + break; + } + case VFParamKind::OMP_Linear: { + Value *ScalarParam = CI->getArgOperand(Param.ParamPos); + // Find the stride for the scalar parameter in this loop and see if + // it matches the stride for the variant. + // TODO: do we need to figure out the cost of an extract to get the + // first lane? Or do we hope that it will be folded away? + ScalarEvolution *SE = PSE.getSE(); + const auto *SAR = + dyn_cast<SCEVAddRecExpr>(SE->getSCEV(ScalarParam)); + + if (!SAR || SAR->getLoop() != TheLoop) { + ParamsOk = false; + break; + } + + const SCEVConstant *Step = + dyn_cast<SCEVConstant>(SAR->getStepRecurrence(*SE)); + + if (!Step || + Step->getAPInt().getSExtValue() != Param.LinearStepOrPos) + ParamsOk = false; + + break; + } + case VFParamKind::GlobalPredicate: + UsesMask = true; + break; + default: + ParamsOk = false; + break; + } + } + + if (!ParamsOk) + continue; + + // Found a suitable candidate, stop here. + VecFunc = CI->getModule()->getFunction(Info.VectorName); + FuncInfo = Info; + break; + } + + // Add in the cost of synthesizing a mask if one wasn't required. + InstructionCost MaskCost = 0; + if (VecFunc && UsesMask && !MaskRequired) + MaskCost = TTI.getShuffleCost( + TargetTransformInfo::SK_Broadcast, + VectorType::get(IntegerType::getInt1Ty( + VecFunc->getFunctionType()->getContext()), + VF)); + + if (TLI && VecFunc && !CI->isNoBuiltin()) + VectorCost = + TTI.getCallInstrCost(nullptr, RetTy, Tys, CostKind) + MaskCost; + + // Find the cost of an intrinsic; some targets may have instructions that + // perform the operation without needing an actual call. + Intrinsic::ID IID = getVectorIntrinsicIDForCall(CI, TLI); + if (IID != Intrinsic::not_intrinsic) + IntrinsicCost = getVectorIntrinsicCost(CI, VF); + + InstructionCost Cost = ScalarCost; + InstWidening Decision = CM_Scalarize; + + if (VectorCost <= Cost) { + Cost = VectorCost; + Decision = CM_VectorCall; + } + + if (IntrinsicCost <= Cost) { + Cost = IntrinsicCost; + Decision = CM_IntrinsicCall; + } + + setCallWideningDecision(CI, VF, Decision, VecFunc, IID, + FuncInfo.getParamIndexForOptionalMask(), Cost); + } + } +} + InstructionCost LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, Type *&VectorTy) { @@ -7318,7 +6877,7 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, // With the exception of GEPs and PHIs, after scalarization there should // only be one copy of the instruction generated in the loop. This is // because the VF is either 1, or any instructions that need scalarizing - // have already been dealt with by the the time we get here. As a result, + // have already been dealt with by the time we get here. As a result, // it means we don't have to multiply the instruction cost by VF. assert(I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || @@ -7344,8 +6903,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, bool ScalarPredicatedBB = false; BranchInst *BI = cast<BranchInst>(I); if (VF.isVector() && BI->isConditional() && - (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) || - PredicatedBBsAfterVectorization.count(BI->getSuccessor(1)))) + (PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(0)) || + PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(1)))) ScalarPredicatedBB = true; if (ScalarPredicatedBB) { @@ -7357,7 +6916,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); return ( TTI.getScalarizationOverhead( - Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), false, true) + + Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), + /*Insert*/ false, /*Extract*/ true, CostKind) + (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) // The back-edge branch will remain, as will all scalar branches. @@ -7373,11 +6933,13 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, auto *Phi = cast<PHINode>(I); // First-order recurrences are replaced by vector shuffles inside the loop. - // NOTE: Don't use ToVectorTy as SK_ExtractSubvector expects a vector type. - if (VF.isVector() && Legal->isFirstOrderRecurrence(Phi)) - return TTI.getShuffleCost( - TargetTransformInfo::SK_ExtractSubvector, cast<VectorType>(VectorTy), - None, VF.getKnownMinValue() - 1, FixedVectorType::get(RetTy, 1)); + if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) { + SmallVector<int> Mask(VF.getKnownMinValue()); + std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1); + return TTI.getShuffleCost(TargetTransformInfo::SK_Splice, + cast<VectorType>(VectorTy), Mask, CostKind, + VF.getKnownMinValue() - 1); + } // Phi nodes in non-header blocks (not inductions, reductions, etc.) are // converted into select instructions. We require N - 1 selects per phi @@ -7395,34 +6957,13 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, case Instruction::SDiv: case Instruction::URem: case Instruction::SRem: - // If we have a predicated instruction, it may not be executed for each - // vector lane. Get the scalarization cost and scale this amount by the - // probability of executing the predicated block. If the instruction is not - // predicated, we fall through to the next case. - if (VF.isVector() && isScalarWithPredication(I, VF)) { - InstructionCost Cost = 0; - - // These instructions have a non-void type, so account for the phi nodes - // that we will create. This cost is likely to be zero. The phi node - // cost, if any, should be scaled by the block probability because it - // models a copy at the end of each predicated block. - Cost += VF.getKnownMinValue() * - TTI.getCFInstrCost(Instruction::PHI, CostKind); - - // The cost of the non-predicated instruction. - Cost += VF.getKnownMinValue() * - TTI.getArithmeticInstrCost(I->getOpcode(), RetTy, CostKind); - - // The cost of insertelement and extractelement instructions needed for - // scalarization. - Cost += getScalarizationOverhead(I, VF); - - // Scale the cost by the probability of executing the predicated blocks. - // This assumes the predicated block for each vector lane is equally - // likely. - return Cost / getReciprocalPredBlockProb(); + if (VF.isVector() && isPredicatedInst(I)) { + const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); + return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ? + ScalarCost : SafeDivisorCost; } - LLVM_FALLTHROUGH; + // We've proven all lanes safe to speculate, fall through. + [[fallthrough]]; case Instruction::Add: case Instruction::FAdd: case Instruction::Sub: @@ -7437,8 +6978,12 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, case Instruction::And: case Instruction::Or: case Instruction::Xor: { - // Since we will replace the stride by 1 the multiplication should go away. - if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) + // If we're speculating on the stride being 1, the multiplication may + // fold away. We can generalize this for all operations using the notion + // of neutral elements. (TODO) + if (I->getOpcode() == Instruction::Mul && + (PSE.getSCEV(I->getOperand(0))->isOne() || + PSE.getSCEV(I->getOperand(1))->isOne())) return 0; // Detect reduction patterns @@ -7448,22 +6993,38 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, // Certain instructions can be cheaper to vectorize if they have a constant // second vector operand. One example of this are shifts on x86. Value *Op2 = I->getOperand(1); - TargetTransformInfo::OperandValueProperties Op2VP; - TargetTransformInfo::OperandValueKind Op2VK = - TTI.getOperandInfo(Op2, Op2VP); - if (Op2VK == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) - Op2VK = TargetTransformInfo::OK_UniformValue; + auto Op2Info = TTI.getOperandInfo(Op2); + if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && + Legal->isInvariant(Op2)) + Op2Info.Kind = TargetTransformInfo::OK_UniformValue; SmallVector<const Value *, 4> Operands(I->operand_values()); - return TTI.getArithmeticInstrCost( - I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, - Op2VK, TargetTransformInfo::OP_None, Op2VP, Operands, I); + auto InstrCost = TTI.getArithmeticInstrCost( + I->getOpcode(), VectorTy, CostKind, + {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, + Op2Info, Operands, I); + + // Some targets can replace frem with vector library calls. + InstructionCost VecCallCost = InstructionCost::getInvalid(); + if (I->getOpcode() == Instruction::FRem) { + LibFunc Func; + if (TLI->getLibFunc(I->getOpcode(), I->getType(), Func) && + TLI->isFunctionVectorizable(TLI->getName(Func), VF)) { + SmallVector<Type *, 4> OpTypes; + for (auto &Op : I->operands()) + OpTypes.push_back(Op->getType()); + VecCallCost = + TTI.getCallInstrCost(nullptr, VectorTy, OpTypes, CostKind); + } + } + return std::min(InstrCost, VecCallCost); } case Instruction::FNeg: { return TTI.getArithmeticInstrCost( - I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, - TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None, - TargetTransformInfo::OP_None, I->getOperand(0), I); + I->getOpcode(), VectorTy, CostKind, + {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, + {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, + I->getOperand(0), I); } case Instruction::Select: { SelectInst *SI = cast<SelectInst>(I); @@ -7476,17 +7037,15 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { // select x, y, false --> x & y // select x, true, y --> x | y - TTI::OperandValueProperties Op1VP = TTI::OP_None; - TTI::OperandValueProperties Op2VP = TTI::OP_None; - TTI::OperandValueKind Op1VK = TTI::getOperandInfo(Op0, Op1VP); - TTI::OperandValueKind Op2VK = TTI::getOperandInfo(Op1, Op2VP); + const auto [Op1VK, Op1VP] = TTI::getOperandInfo(Op0); + const auto [Op2VK, Op2VP] = TTI::getOperandInfo(Op1); assert(Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1); SmallVector<const Value *, 2> Operands{Op0, Op1}; return TTI.getArithmeticInstrCost( match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, - CostKind, Op1VK, Op2VK, Op1VP, Op2VP, Operands, I); + CostKind, {Op1VK, Op1VP}, {Op2VK, Op2VP}, Operands, I); } Type *CondTy = SI->getCondition()->getType(); @@ -7517,6 +7076,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, InstWidening Decision = getWideningDecision(I, Width); assert(Decision != CM_Unknown && "CM decision should be taken at this point"); + if (getWideningCost(I, VF) == InstructionCost::getInvalid()) + return InstructionCost::getInvalid(); if (Decision == CM_Scalarize) Width = ElementCount::getFixed(1); } @@ -7526,7 +7087,7 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, case Instruction::BitCast: if (I->getType()->isPointerTy()) return 0; - LLVM_FALLTHROUGH; + [[fallthrough]]; case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: @@ -7559,6 +7120,9 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, return TTI::CastContextHint::Reversed; case LoopVectorizationCostModel::CM_Unknown: llvm_unreachable("Instr did not go through cost modelling?"); + case LoopVectorizationCostModel::CM_VectorCall: + case LoopVectorizationCostModel::CM_IntrinsicCall: + llvm_unreachable_internal("Instr has invalid widening decision"); } llvm_unreachable("Unhandled case!"); @@ -7607,7 +7171,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { - SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); + // Leave SrcVecTy unchanged - we only shrink the destination element + // type. VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); } @@ -7615,19 +7180,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); } - case Instruction::Call: { - if (RecurrenceDescriptor::isFMulAddIntrinsic(I)) - if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) - return *RedCost; - bool NeedToScalarize; - CallInst *CI = cast<CallInst>(I); - InstructionCost CallCost = getVectorCallCost(CI, VF, NeedToScalarize); - if (getVectorIntrinsicIDForCall(CI, TLI)) { - InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); - return std::min(CallCost, IntrinsicCost); - } - return CallCost; - } + case Instruction::Call: + return getVectorCallCost(cast<CallInst>(I), VF); case Instruction::ExtractValue: return TTI.getInstructionCost(I, TTI::TCK_RecipThroughput); case Instruction::Alloca: @@ -7635,67 +7189,37 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, // the result would need to be a vector of pointers. if (VF.isScalable()) return InstructionCost::getInvalid(); - LLVM_FALLTHROUGH; + [[fallthrough]]; default: // This opcode is unknown. Assume that it is the same as 'mul'. return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); } // end of switch. } -char LoopVectorize::ID = 0; - -static const char lv_name[] = "Loop Vectorization"; - -INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) -INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) -INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass) -INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) -INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) -INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) -INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) -INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) -INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) -INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) -INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis) -INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass) -INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) -INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) -INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy) -INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) - -namespace llvm { - -Pass *createLoopVectorizePass() { return new LoopVectorize(); } - -Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, - bool VectorizeOnlyWhenForced) { - return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); -} - -} // end namespace llvm - -bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { - // Check if the pointer operand of a load or store instruction is - // consecutive. - if (auto *Ptr = getLoadStorePointerOperand(Inst)) - return Legal->isConsecutivePtr(getLoadStoreType(Inst), Ptr); - return false; -} - void LoopVectorizationCostModel::collectValuesToIgnore() { // Ignore ephemeral values. CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); + // Find all stores to invariant variables. Since they are going to sink + // outside the loop we do not need calculate cost for them. + for (BasicBlock *BB : TheLoop->blocks()) + for (Instruction &I : *BB) { + StoreInst *SI; + if ((SI = dyn_cast<StoreInst>(&I)) && + Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) + ValuesToIgnore.insert(&I); + } + // Ignore type-promoting instructions we identified during reduction // detection. - for (auto &Reduction : Legal->getReductionVars()) { + for (const auto &Reduction : Legal->getReductionVars()) { const RecurrenceDescriptor &RedDes = Reduction.second; const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); VecValuesToIgnore.insert(Casts.begin(), Casts.end()); } // Ignore type-casting instructions we identified during induction // detection. - for (auto &Induction : Legal->getInductionVars()) { + for (const auto &Induction : Legal->getInductionVars()) { const InductionDescriptor &IndDes = Induction.second; const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); VecValuesToIgnore.insert(Casts.begin(), Casts.end()); @@ -7703,7 +7227,7 @@ void LoopVectorizationCostModel::collectValuesToIgnore() { } void LoopVectorizationCostModel::collectInLoopReductions() { - for (auto &Reduction : Legal->getReductionVars()) { + for (const auto &Reduction : Legal->getReductionVars()) { PHINode *Phi = Reduction.first; const RecurrenceDescriptor &RdxDesc = Reduction.second; @@ -7725,8 +7249,9 @@ void LoopVectorizationCostModel::collectInLoopReductions() { SmallVector<Instruction *, 4> ReductionOperations = RdxDesc.getReductionOpChain(Phi, TheLoop); bool InLoop = !ReductionOperations.empty(); + if (InLoop) { - InLoopReductionChains[Phi] = ReductionOperations; + InLoopReductions.insert(Phi); // Add the elements to InLoopReductionImmediateChains for cost modelling. Instruction *LastChain = Phi; for (auto *I : ReductionOperations) { @@ -7739,21 +7264,38 @@ void LoopVectorizationCostModel::collectInLoopReductions() { } } +VPValue *VPBuilder::createICmp(CmpInst::Predicate Pred, VPValue *A, VPValue *B, + DebugLoc DL, const Twine &Name) { + assert(Pred >= CmpInst::FIRST_ICMP_PREDICATE && + Pred <= CmpInst::LAST_ICMP_PREDICATE && "invalid predicate"); + return tryInsertInstruction( + new VPInstruction(Instruction::ICmp, Pred, A, B, DL, Name)); +} + +// This function will select a scalable VF if the target supports scalable +// vectors and a fixed one otherwise. // TODO: we could return a pair of values that specify the max VF and // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment // doesn't have a cost model that can choose which plan to execute if // more than one is generated. -static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, - LoopVectorizationCostModel &CM) { +static ElementCount determineVPlanVF(const TargetTransformInfo &TTI, + LoopVectorizationCostModel &CM) { unsigned WidestType; std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); - return WidestVectorRegBits / WidestType; + + TargetTransformInfo::RegisterKind RegKind = + TTI.enableScalableVectorization() + ? TargetTransformInfo::RGK_ScalableVector + : TargetTransformInfo::RGK_FixedWidthVector; + + TypeSize RegSize = TTI.getRegisterBitWidth(RegKind); + unsigned N = RegSize.getKnownMinValue() / WidestType; + return ElementCount::get(N, RegSize.isScalable()); } VectorizationFactor LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { - assert(!UserVF.isScalable() && "scalable vectors not yet supported"); ElementCount VF = UserVF; // Outer loop handling: They may require CFG and instruction level // transformations before even evaluating whether vectorization is profitable. @@ -7763,10 +7305,7 @@ LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { // If the user doesn't provide a vectorization factor, determine a // reasonable one. if (UserVF.isZero()) { - VF = ElementCount::getFixed(determineVPlanVF( - TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) - .getFixedSize(), - CM)); + VF = determineVPlanVF(TTI, CM); LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n"); // Make sure we have a VF > 1 for stress testing. @@ -7775,6 +7314,17 @@ LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { << "overriding computed VF.\n"); VF = ElementCount::getFixed(4); } + } else if (UserVF.isScalable() && !TTI.supportsScalableVectors() && + !ForceTargetSupportsScalableVectors) { + LLVM_DEBUG(dbgs() << "LV: Not vectorizing. Scalable VF requested, but " + << "not supported by the target.\n"); + reportVectorizationFailure( + "Scalable vectorization requested but not supported by the target", + "the scalable user-specified vectorization width for outer-loop " + "vectorization cannot be used because the target does not support " + "scalable vectors.", + "ScalableVFUnfeasible", ORE, OrigLoop); + return VectorizationFactor::Disabled(); } assert(EnableVPlanNativePath && "VPlan-native path is not enabled."); assert(isPowerOf2_32(VF.getKnownMinValue()) && @@ -7787,7 +7337,7 @@ LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { if (VPlanBuildStressTest) return VectorizationFactor::Disabled(); - return {VF, 0 /*Cost*/}; + return {VF, 0 /*Cost*/, 0 /* ScalarCost */}; } LLVM_DEBUG( @@ -7796,16 +7346,19 @@ LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { return VectorizationFactor::Disabled(); } -Optional<VectorizationFactor> +std::optional<VectorizationFactor> LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { assert(OrigLoop->isInnermost() && "Inner loop expected."); + CM.collectValuesToIgnore(); + CM.collectElementTypesForWidening(); + FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. - return None; + return std::nullopt; // Invalidate interleave groups if all blocks of loop will be predicated. if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && - !useMaskedInterleavedAccesses(*TTI)) { + !useMaskedInterleavedAccesses(TTI)) { LLVM_DEBUG( dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " @@ -7825,12 +7378,18 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { "VF needs to be a power of two"); // Collect the instructions (and their associated costs) that will be more // profitable to scalarize. + CM.collectInLoopReductions(); if (CM.selectUserVectorizationFactor(UserVF)) { LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); - CM.collectInLoopReductions(); buildVPlansWithVPRecipes(UserVF, UserVF); + if (!hasPlanWithVF(UserVF)) { + LLVM_DEBUG(dbgs() << "LV: No VPlan could be built for " << UserVF + << ".\n"); + return std::nullopt; + } + LLVM_DEBUG(printPlans(dbgs())); - return {{UserVF, 0}}; + return {{UserVF, 0, 0}}; } else reportVectorizationInfo("UserVF ignored because of invalid costs.", "InvalidCost", ORE, OrigLoop); @@ -7845,6 +7404,7 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) VFCandidates.insert(VF); + CM.collectInLoopReductions(); for (const auto &VF : VFCandidates) { // Collect Uniform and Scalar instructions after vectorization with VF. CM.collectUniformsAndScalars(VF); @@ -7855,7 +7415,6 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { CM.collectInstsToScalarize(VF); } - CM.collectInLoopReductions(); buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); @@ -7864,30 +7423,14 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { return VectorizationFactor::Disabled(); // Select the optimal vectorization factor. - auto SelectedVF = CM.selectVectorizationFactor(VFCandidates); - - // Check if it is profitable to vectorize with runtime checks. - unsigned NumRuntimePointerChecks = Requirements.getNumRuntimePointerChecks(); - if (SelectedVF.Width.getKnownMinValue() > 1 && NumRuntimePointerChecks) { - bool PragmaThresholdReached = - NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold; - bool ThresholdReached = - NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold; - if ((ThresholdReached && !Hints.allowReordering()) || - PragmaThresholdReached) { - ORE->emit([&]() { - return OptimizationRemarkAnalysisAliasing( - DEBUG_TYPE, "CantReorderMemOps", OrigLoop->getStartLoc(), - OrigLoop->getHeader()) - << "loop not vectorized: cannot prove it is safe to reorder " - "memory operations"; - }); - LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n"); - Hints.emitRemarkWithHints(); - return VectorizationFactor::Disabled(); - } + VectorizationFactor VF = selectVectorizationFactor(VFCandidates); + assert((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero."); + if (!hasPlanWithVF(VF.Width)) { + LLVM_DEBUG(dbgs() << "LV: No VPlan could be built for " << VF.Width + << ".\n"); + return std::nullopt; } - return SelectedVF; + return VF; } VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const { @@ -7916,7 +7459,7 @@ static void AddRuntimeUnrollDisableMetaData(Loop *L) { if (MD) { const auto *S = dyn_cast<MDString>(MD->getOperand(0)); IsUnrollMetadata = - S && S->getString().startswith("llvm.loop.unroll.disable"); + S && S->getString().starts_with("llvm.loop.unroll.disable"); } MDs.push_back(LoopID->getOperand(i)); } @@ -7937,20 +7480,124 @@ static void AddRuntimeUnrollDisableMetaData(Loop *L) { } } -void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, - VPlan &BestVPlan, - InnerLoopVectorizer &ILV, - DominatorTree *DT) { +// Check if \p RedResult is a ComputeReductionResult instruction, and if it is +// create a merge phi node for it and add it to \p ReductionResumeValues. +static void createAndCollectMergePhiForReduction( + VPInstruction *RedResult, + DenseMap<const RecurrenceDescriptor *, Value *> &ReductionResumeValues, + VPTransformState &State, Loop *OrigLoop, BasicBlock *LoopMiddleBlock) { + if (!RedResult || + RedResult->getOpcode() != VPInstruction::ComputeReductionResult) + return; + + auto *PhiR = cast<VPReductionPHIRecipe>(RedResult->getOperand(0)); + const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); + + TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); + Value *FinalValue = + State.get(RedResult, VPIteration(State.UF - 1, VPLane::getFirstLane())); + auto *ResumePhi = + dyn_cast<PHINode>(PhiR->getStartValue()->getUnderlyingValue()); + + // TODO: bc.merge.rdx should not be created here, instead it should be + // modeled in VPlan. + BasicBlock *LoopScalarPreHeader = OrigLoop->getLoopPreheader(); + // Create a phi node that merges control-flow from the backedge-taken check + // block and the middle block. + auto *BCBlockPhi = PHINode::Create(FinalValue->getType(), 2, "bc.merge.rdx", + LoopScalarPreHeader->getTerminator()); + + // If we are fixing reductions in the epilogue loop then we should already + // have created a bc.merge.rdx Phi after the main vector body. Ensure that + // we carry over the incoming values correctly. + for (auto *Incoming : predecessors(LoopScalarPreHeader)) { + if (Incoming == LoopMiddleBlock) + BCBlockPhi->addIncoming(FinalValue, Incoming); + else if (ResumePhi && is_contained(ResumePhi->blocks(), Incoming)) + BCBlockPhi->addIncoming(ResumePhi->getIncomingValueForBlock(Incoming), + Incoming); + else + BCBlockPhi->addIncoming(ReductionStartValue, Incoming); + } + + auto *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); + // TODO: This fixup should instead be modeled in VPlan. + // Fix the scalar loop reduction variable with the incoming reduction sum + // from the vector body and from the backedge value. + int IncomingEdgeBlockIdx = + OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); + assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); + // Pick the other block. + int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); + OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); + Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); + OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); + + ReductionResumeValues[&RdxDesc] = BCBlockPhi; +} + +std::pair<DenseMap<const SCEV *, Value *>, + DenseMap<const RecurrenceDescriptor *, Value *>> +LoopVectorizationPlanner::executePlan( + ElementCount BestVF, unsigned BestUF, VPlan &BestVPlan, + InnerLoopVectorizer &ILV, DominatorTree *DT, bool IsEpilogueVectorization, + const DenseMap<const SCEV *, Value *> *ExpandedSCEVs) { + assert(BestVPlan.hasVF(BestVF) && + "Trying to execute plan with unsupported VF"); + assert(BestVPlan.hasUF(BestUF) && + "Trying to execute plan with unsupported UF"); + assert( + (IsEpilogueVectorization || !ExpandedSCEVs) && + "expanded SCEVs to reuse can only be used during epilogue vectorization"); + LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n'); + if (!IsEpilogueVectorization) + VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE); + // Perform the actual loop transformation. + VPTransformState State(BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan, + OrigLoop->getHeader()->getContext()); + + // 0. Generate SCEV-dependent code into the preheader, including TripCount, + // before making any changes to the CFG. + if (!BestVPlan.getPreheader()->empty()) { + State.CFG.PrevBB = OrigLoop->getLoopPreheader(); + State.Builder.SetInsertPoint(OrigLoop->getLoopPreheader()->getTerminator()); + BestVPlan.getPreheader()->execute(&State); + } + if (!ILV.getTripCount()) + ILV.setTripCount(State.get(BestVPlan.getTripCount(), {0, 0})); + else + assert(IsEpilogueVectorization && "should only re-use the existing trip " + "count during epilogue vectorization"); - // 1. Create a new empty loop. Unlink the old loop and connect the new one. - VPTransformState State{BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan}; + // 1. Set up the skeleton for vectorization, including vector pre-header and + // middle block. The vector loop is created during VPlan execution. Value *CanonicalIVStartValue; std::tie(State.CFG.PrevBB, CanonicalIVStartValue) = - ILV.createVectorizedLoopSkeleton(); + ILV.createVectorizedLoopSkeleton(ExpandedSCEVs ? *ExpandedSCEVs + : State.ExpandedSCEVs); + + // Only use noalias metadata when using memory checks guaranteeing no overlap + // across all iterations. + const LoopAccessInfo *LAI = ILV.Legal->getLAI(); + std::unique_ptr<LoopVersioning> LVer = nullptr; + if (LAI && !LAI->getRuntimePointerChecking()->getChecks().empty() && + !LAI->getRuntimePointerChecking()->getDiffChecks()) { + + // We currently don't use LoopVersioning for the actual loop cloning but we + // still use it to add the noalias metadata. + // TODO: Find a better way to re-use LoopVersioning functionality to add + // metadata. + LVer = std::make_unique<LoopVersioning>( + *LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT, + PSE.getSE()); + State.LVer = &*LVer; + State.LVer->prepareNoAliasMetadata(); + } + ILV.collectPoisonGeneratingRecipes(State); ILV.printDebugTracesAtStart(); @@ -7964,22 +7611,36 @@ void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, //===------------------------------------------------===// // 2. Copy and widen instructions from the old loop into the new loop. - BestVPlan.prepareToExecute(ILV.getOrCreateTripCount(nullptr), + BestVPlan.prepareToExecute(ILV.getTripCount(), ILV.getOrCreateVectorTripCount(nullptr), CanonicalIVStartValue, State); + BestVPlan.execute(&State); + // 2.5 Collect reduction resume values. + DenseMap<const RecurrenceDescriptor *, Value *> ReductionResumeValues; + auto *ExitVPBB = + cast<VPBasicBlock>(BestVPlan.getVectorLoopRegion()->getSingleSuccessor()); + for (VPRecipeBase &R : *ExitVPBB) { + createAndCollectMergePhiForReduction(dyn_cast<VPInstruction>(&R), + ReductionResumeValues, State, OrigLoop, + State.CFG.VPBB2IRBB[ExitVPBB]); + } + + // 2.6. Maintain Loop Hints // Keep all loop hints from the original loop on the vector loop (we'll // replace the vectorizer-specific hints below). MDNode *OrigLoopID = OrigLoop->getLoopID(); - Optional<MDNode *> VectorizedLoopID = + std::optional<MDNode *> VectorizedLoopID = makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, LLVMLoopVectorizeFollowupVectorized}); - Loop *L = LI->getLoopFor(State.CFG.PrevBB); - if (VectorizedLoopID.hasValue()) - L->setLoopID(VectorizedLoopID.getValue()); + VPBasicBlock *HeaderVPBB = + BestVPlan.getVectorLoopRegion()->getEntryBasicBlock(); + Loop *L = LI->getLoopFor(State.CFG.VPBB2IRBB[HeaderVPBB]); + if (VectorizedLoopID) + L->setLoopID(*VectorizedLoopID); else { // Keep all loop hints from the original loop on the vector loop (we'll // replace the vectorizer-specific hints below). @@ -7989,15 +7650,18 @@ void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, LoopVectorizeHints Hints(L, true, *ORE); Hints.setAlreadyVectorized(); } - // Disable runtime unrolling when vectorizing the epilogue loop. - if (CanonicalIVStartValue) + TargetTransformInfo::UnrollingPreferences UP; + TTI.getUnrollingPreferences(L, *PSE.getSE(), UP, ORE); + if (!UP.UnrollVectorizedLoop || CanonicalIVStartValue) AddRuntimeUnrollDisableMetaData(L); // 3. Fix the vectorized code: take care of header phi's, live-outs, // predication, updating analyses. - ILV.fixVectorizedLoop(State); + ILV.fixVectorizedLoop(State, BestVPlan); ILV.printDebugTracesAtEnd(); + + return {State.ExpandedSCEVs, ReductionResumeValues}; } #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) @@ -8010,53 +7674,6 @@ void LoopVectorizationPlanner::printPlans(raw_ostream &O) { } #endif -void LoopVectorizationPlanner::collectTriviallyDeadInstructions( - SmallPtrSetImpl<Instruction *> &DeadInstructions) { - - // We create new control-flow for the vectorized loop, so the original exit - // conditions will be dead after vectorization if it's only used by the - // terminator - SmallVector<BasicBlock*> ExitingBlocks; - OrigLoop->getExitingBlocks(ExitingBlocks); - for (auto *BB : ExitingBlocks) { - auto *Cmp = dyn_cast<Instruction>(BB->getTerminator()->getOperand(0)); - if (!Cmp || !Cmp->hasOneUse()) - continue; - - // TODO: we should introduce a getUniqueExitingBlocks on Loop - if (!DeadInstructions.insert(Cmp).second) - continue; - - // The operands of the icmp is often a dead trunc, used by IndUpdate. - // TODO: can recurse through operands in general - for (Value *Op : Cmp->operands()) { - if (isa<TruncInst>(Op) && Op->hasOneUse()) - DeadInstructions.insert(cast<Instruction>(Op)); - } - } - - // We create new "steps" for induction variable updates to which the original - // induction variables map. An original update instruction will be dead if - // all its users except the induction variable are dead. - auto *Latch = OrigLoop->getLoopLatch(); - for (auto &Induction : Legal->getInductionVars()) { - PHINode *Ind = Induction.first; - auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); - - // If the tail is to be folded by masking, the primary induction variable, - // if exists, isn't dead: it will be used for masking. Don't kill it. - if (CM.foldTailByMasking() && IndUpdate == Legal->getPrimaryInduction()) - continue; - - if (llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { - return U == Ind || DeadInstructions.count(cast<Instruction>(U)); - })) - DeadInstructions.insert(IndUpdate); - } -} - -Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } - //===--------------------------------------------------------------------===// // EpilogueVectorizerMainLoop //===--------------------------------------------------------------------===// @@ -8064,24 +7681,24 @@ Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } /// This function is partially responsible for generating the control flow /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. std::pair<BasicBlock *, Value *> -EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { - MDNode *OrigLoopID = OrigLoop->getLoopID(); - Loop *Lp = createVectorLoopSkeleton(""); +EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton( + const SCEV2ValueTy &ExpandedSCEVs) { + createVectorLoopSkeleton(""); // Generate the code to check the minimum iteration count of the vector // epilogue (see below). EPI.EpilogueIterationCountCheck = - emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, true); + emitIterationCountCheck(LoopScalarPreHeader, true); EPI.EpilogueIterationCountCheck->setName("iter.check"); // Generate the code to check any assumptions that we've made for SCEV // expressions. - EPI.SCEVSafetyCheck = emitSCEVChecks(Lp, LoopScalarPreHeader); + EPI.SCEVSafetyCheck = emitSCEVChecks(LoopScalarPreHeader); // Generate the code that checks at runtime if arrays overlap. We put the // checks into a separate block to make the more common case of few elements // faster. - EPI.MemSafetyCheck = emitMemRuntimeChecks(Lp, LoopScalarPreHeader); + EPI.MemSafetyCheck = emitMemRuntimeChecks(LoopScalarPreHeader); // Generate the iteration count check for the main loop, *after* the check // for the epilogue loop, so that the path-length is shorter for the case @@ -8090,19 +7707,17 @@ EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { // trip count. Note: the branch will get updated later on when we vectorize // the epilogue. EPI.MainLoopIterationCountCheck = - emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, false); + emitIterationCountCheck(LoopScalarPreHeader, false); // Generate the induction variable. - Value *CountRoundDown = getOrCreateVectorTripCount(Lp); - EPI.VectorTripCount = CountRoundDown; - createHeaderBranch(Lp); + EPI.VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); // Skip induction resume value creation here because they will be created in - // the second pass. If we created them here, they wouldn't be used anyway, - // because the vplan in the second pass still contains the inductions from the - // original loop. + // the second pass for the scalar loop. The induction resume values for the + // inductions in the epilogue loop are created before executing the plan for + // the epilogue loop. - return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; + return {completeLoopSkeleton(), nullptr}; } void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { @@ -8122,13 +7737,13 @@ void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { }); } -BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( - Loop *L, BasicBlock *Bypass, bool ForEpilogue) { - assert(L && "Expected valid Loop."); +BasicBlock * +EpilogueVectorizerMainLoop::emitIterationCountCheck(BasicBlock *Bypass, + bool ForEpilogue) { assert(Bypass && "Expected valid bypass basic block."); ElementCount VFactor = ForEpilogue ? EPI.EpilogueVF : VF; unsigned UFactor = ForEpilogue ? EPI.EpilogueUF : UF; - Value *Count = getOrCreateTripCount(L); + Value *Count = getTripCount(); // Reuse existing vector loop preheader for TC checks. // Note that new preheader block is generated for vector loop. BasicBlock *const TCCheckBlock = LoopVectorPreHeader; @@ -8136,8 +7751,10 @@ BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( // Generate code to check if the loop's trip count is less than VF * UF of the // main vector loop. - auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF : VF) ? - ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; + auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF.isVector() + : VF.isVector()) + ? ICmpInst::ICMP_ULE + : ICmpInst::ICMP_ULT; Value *CheckMinIters = Builder.CreateICmp( P, Count, createStepForVF(Builder, Count->getType(), VFactor, UFactor), @@ -8157,7 +7774,7 @@ BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( // Update dominator for Bypass & LoopExit. DT->changeImmediateDominator(Bypass, TCCheckBlock); - if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) + if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF.isVector())) // For loops with multiple exits, there's no edge from the middle block // to exit blocks (as the epilogue must run) and thus no need to update // the immediate dominator of the exit blocks. @@ -8171,9 +7788,11 @@ BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( EPI.TripCount = Count; } - ReplaceInstWithInst( - TCCheckBlock->getTerminator(), - BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); + BranchInst &BI = + *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); + if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) + setBranchWeights(BI, MinItersBypassWeights); + ReplaceInstWithInst(TCCheckBlock->getTerminator(), &BI); return TCCheckBlock; } @@ -8185,9 +7804,9 @@ BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( /// This function is partially responsible for generating the control flow /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. std::pair<BasicBlock *, Value *> -EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { - MDNode *OrigLoopID = OrigLoop->getLoopID(); - Loop *Lp = createVectorLoopSkeleton("vec.epilog."); +EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton( + const SCEV2ValueTy &ExpandedSCEVs) { + createVectorLoopSkeleton("vec.epilog."); // Now, compare the remaining count and if there aren't enough iterations to // execute the vectorized epilogue skip to the scalar part. @@ -8196,7 +7815,7 @@ EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { LoopVectorPreHeader = SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, LI, nullptr, "vec.epilog.ph"); - emitMinimumVectorEpilogueIterCountCheck(Lp, LoopScalarPreHeader, + emitMinimumVectorEpilogueIterCountCheck(LoopScalarPreHeader, VecEpilogueIterationCountCheck); // Adjust the control flow taking the state info from the main loop @@ -8225,52 +7844,58 @@ EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { DT->changeImmediateDominator(LoopScalarPreHeader, EPI.EpilogueIterationCountCheck); - if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) + if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF.isVector())) // If there is an epilogue which must run, there's no edge from the // middle block to exit blocks and thus no need to update the immediate // dominator of the exit blocks. DT->changeImmediateDominator(LoopExitBlock, EPI.EpilogueIterationCountCheck); - // Keep track of bypass blocks, as they feed start values to the induction - // phis in the scalar loop preheader. + // Keep track of bypass blocks, as they feed start values to the induction and + // reduction phis in the scalar loop preheader. if (EPI.SCEVSafetyCheck) LoopBypassBlocks.push_back(EPI.SCEVSafetyCheck); if (EPI.MemSafetyCheck) LoopBypassBlocks.push_back(EPI.MemSafetyCheck); LoopBypassBlocks.push_back(EPI.EpilogueIterationCountCheck); - // The vec.epilog.iter.check block may contain Phi nodes from reductions which - // merge control-flow from the latch block and the middle block. Update the - // incoming values here and move the Phi into the preheader. + // The vec.epilog.iter.check block may contain Phi nodes from inductions or + // reductions which merge control-flow from the latch block and the middle + // block. Update the incoming values here and move the Phi into the preheader. SmallVector<PHINode *, 4> PhisInBlock; for (PHINode &Phi : VecEpilogueIterationCountCheck->phis()) PhisInBlock.push_back(&Phi); for (PHINode *Phi : PhisInBlock) { + Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHI()); Phi->replaceIncomingBlockWith( VecEpilogueIterationCountCheck->getSinglePredecessor(), VecEpilogueIterationCountCheck); + + // If the phi doesn't have an incoming value from the + // EpilogueIterationCountCheck, we are done. Otherwise remove the incoming + // value and also those from other check blocks. This is needed for + // reduction phis only. + if (none_of(Phi->blocks(), [&](BasicBlock *IncB) { + return EPI.EpilogueIterationCountCheck == IncB; + })) + continue; Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck); if (EPI.SCEVSafetyCheck) Phi->removeIncomingValue(EPI.SCEVSafetyCheck); if (EPI.MemSafetyCheck) Phi->removeIncomingValue(EPI.MemSafetyCheck); - Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHI()); } // Generate a resume induction for the vector epilogue and put it in the // vector epilogue preheader Type *IdxTy = Legal->getWidestInductionType(); - PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val", - LoopVectorPreHeader->getFirstNonPHI()); + PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val"); + EPResumeVal->insertBefore(LoopVectorPreHeader->getFirstNonPHIIt()); EPResumeVal->addIncoming(EPI.VectorTripCount, VecEpilogueIterationCountCheck); EPResumeVal->addIncoming(ConstantInt::get(IdxTy, 0), EPI.MainLoopIterationCountCheck); - // Generate the induction variable. - createHeaderBranch(Lp); - // Generate induction resume values. These variables save the new starting // indexes for the scalar loop. They are used to test if there are any tail // iterations left once the vector loop has completed. @@ -8278,15 +7903,16 @@ EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { // check, then the resume value for the induction variable comes from // the trip count of the main vector loop, hence passing the AdditionalBypass // argument. - createInductionResumeValues(Lp, {VecEpilogueIterationCountCheck, - EPI.VectorTripCount} /* AdditionalBypass */); + createInductionResumeValues(ExpandedSCEVs, + {VecEpilogueIterationCountCheck, + EPI.VectorTripCount} /* AdditionalBypass */); - return {completeLoopSkeleton(Lp, OrigLoopID), EPResumeVal}; + return {completeLoopSkeleton(), EPResumeVal}; } BasicBlock * EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( - Loop *L, BasicBlock *Bypass, BasicBlock *Insert) { + BasicBlock *Bypass, BasicBlock *Insert) { assert(EPI.TripCount && "Expected trip count to have been safed in the first pass."); @@ -8300,8 +7926,9 @@ EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( // Generate code to check if the loop's trip count is less than VF * UF of the // vector epilogue loop. - auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF) ? - ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; + auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF.isVector()) + ? ICmpInst::ICMP_ULE + : ICmpInst::ICMP_ULT; Value *CheckMinIters = Builder.CreateICmp(P, Count, @@ -8309,9 +7936,22 @@ EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( EPI.EpilogueVF, EPI.EpilogueUF), "min.epilog.iters.check"); - ReplaceInstWithInst( - Insert->getTerminator(), - BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); + BranchInst &BI = + *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); + if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) { + unsigned MainLoopStep = UF * VF.getKnownMinValue(); + unsigned EpilogueLoopStep = + EPI.EpilogueUF * EPI.EpilogueVF.getKnownMinValue(); + // We assume the remaining `Count` is equally distributed in + // [0, MainLoopStep) + // So the probability for `Count < EpilogueLoopStep` should be + // min(MainLoopStep, EpilogueLoopStep) / MainLoopStep + unsigned EstimatedSkipCount = std::min(MainLoopStep, EpilogueLoopStep); + const uint32_t Weights[] = {EstimatedSkipCount, + MainLoopStep - EstimatedSkipCount}; + setBranchWeights(BI, Weights); + } + ReplaceInstWithInst(Insert->getTerminator(), &BI); LoopBypassBlocks.push_back(Insert); return Insert; @@ -8336,8 +7976,7 @@ bool LoopVectorizationPlanner::getDecisionAndClampRange( assert(!Range.isEmpty() && "Trying to test an empty VF range."); bool PredicateAtRangeStart = Predicate(Range.Start); - for (ElementCount TmpVF = Range.Start * 2; - ElementCount::isKnownLT(TmpVF, Range.End); TmpVF *= 2) + for (ElementCount TmpVF : VFRange(Range.Start * 2, Range.End)) if (Predicate(TmpVF) != PredicateAtRangeStart) { Range.End = TmpVF; break; @@ -8353,16 +7992,16 @@ bool LoopVectorizationPlanner::getDecisionAndClampRange( /// buildVPlan(). void LoopVectorizationPlanner::buildVPlans(ElementCount MinVF, ElementCount MaxVF) { - auto MaxVFPlusOne = MaxVF.getWithIncrement(1); - for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { - VFRange SubRange = {VF, MaxVFPlusOne}; + auto MaxVFTimes2 = MaxVF * 2; + for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) { + VFRange SubRange = {VF, MaxVFTimes2}; VPlans.push_back(buildVPlan(SubRange)); VF = SubRange.End; } } VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, - VPlanPtr &Plan) { + VPlan &Plan) { assert(is_contained(predecessors(Dst), Src) && "Invalid edge"); // Look for cached value. @@ -8371,7 +8010,7 @@ VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, if (ECEntryIt != EdgeMaskCache.end()) return ECEntryIt->second; - VPValue *SrcMask = createBlockInMask(Src, Plan); + VPValue *SrcMask = getBlockInMask(Src); // The terminator has to be a branch inst! BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); @@ -8386,7 +8025,7 @@ VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, if (OrigLoop->isLoopExiting(Src)) return EdgeMaskCache[Edge] = SrcMask; - VPValue *EdgeMask = Plan->getOrAddVPValue(BI->getCondition()); + VPValue *EdgeMask = Plan.getVPValueOrAddLiveIn(BI->getCondition()); assert(EdgeMask && "No Edge Mask found for condition"); if (BI->getSuccessor(0) != Dst) @@ -8397,7 +8036,7 @@ VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, // 'select i1 SrcMask, i1 EdgeMask, i1 false'. // The select version does not introduce new UB if SrcMask is false and // EdgeMask is poison. Using 'and' here introduces undefined behavior. - VPValue *False = Plan->getOrAddVPValue( + VPValue *False = Plan.getVPValueOrAddLiveIn( ConstantInt::getFalse(BI->getCondition()->getType())); EdgeMask = Builder.createSelect(SrcMask, EdgeMask, False, BI->getDebugLoc()); @@ -8406,49 +8045,57 @@ VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, return EdgeMaskCache[Edge] = EdgeMask; } -VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { - assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); +void VPRecipeBuilder::createHeaderMask(VPlan &Plan) { + BasicBlock *Header = OrigLoop->getHeader(); - // Look for cached value. - BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); - if (BCEntryIt != BlockMaskCache.end()) - return BCEntryIt->second; + // When not folding the tail, use nullptr to model all-true mask. + if (!CM.foldTailByMasking()) { + BlockMaskCache[Header] = nullptr; + return; + } - // All-one mask is modelled as no-mask following the convention for masked - // load/store/gather/scatter. Initialize BlockMask to no-mask. + // Introduce the early-exit compare IV <= BTC to form header block mask. + // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by + // constructing the desired canonical IV in the header block as its first + // non-phi instructions. + + VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock(); + auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); + auto *IV = new VPWidenCanonicalIVRecipe(Plan.getCanonicalIV()); + HeaderVPBB->insert(IV, NewInsertionPoint); + + VPBuilder::InsertPointGuard Guard(Builder); + Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); VPValue *BlockMask = nullptr; + VPValue *BTC = Plan.getOrCreateBackedgeTakenCount(); + BlockMask = Builder.createICmp(CmpInst::ICMP_ULE, IV, BTC); + BlockMaskCache[Header] = BlockMask; +} - if (OrigLoop->getHeader() == BB) { - if (!CM.blockNeedsPredicationForAnyReason(BB)) - return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. - - // Introduce the early-exit compare IV <= BTC to form header block mask. - // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by - // constructing the desired canonical IV in the header block as its first - // non-phi instructions. - assert(CM.foldTailByMasking() && "must fold the tail"); - VPBasicBlock *HeaderVPBB = Plan->getEntry()->getEntryBasicBlock(); - auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); - auto *IV = new VPWidenCanonicalIVRecipe(Plan->getCanonicalIV()); - HeaderVPBB->insert(IV, HeaderVPBB->getFirstNonPhi()); - - VPBuilder::InsertPointGuard Guard(Builder); - Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); - if (CM.TTI.emitGetActiveLaneMask()) { - VPValue *TC = Plan->getOrCreateTripCount(); - BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV, TC}); - } else { - VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); - BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); - } - return BlockMaskCache[BB] = BlockMask; - } +VPValue *VPRecipeBuilder::getBlockInMask(BasicBlock *BB) const { + // Return the cached value. + BlockMaskCacheTy::const_iterator BCEntryIt = BlockMaskCache.find(BB); + assert(BCEntryIt != BlockMaskCache.end() && + "Trying to access mask for block without one."); + return BCEntryIt->second; +} + +void VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlan &Plan) { + assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); + assert(BlockMaskCache.count(BB) == 0 && "Mask for block already computed"); + assert(OrigLoop->getHeader() != BB && + "Loop header must have cached block mask"); + // All-one mask is modelled as no-mask following the convention for masked + // load/store/gather/scatter. Initialize BlockMask to no-mask. + VPValue *BlockMask = nullptr; // This is the block mask. We OR all incoming edges. for (auto *Predecessor : predecessors(BB)) { VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); - if (!EdgeMask) // Mask of predecessor is all-one so mask of block is too. - return BlockMaskCache[BB] = EdgeMask; + if (!EdgeMask) { // Mask of predecessor is all-one so mask of block is too. + BlockMaskCache[BB] = EdgeMask; + return; + } if (!BlockMask) { // BlockMask has its initialized nullptr value. BlockMask = EdgeMask; @@ -8458,7 +8105,7 @@ VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { BlockMask = Builder.createOr(BlockMask, EdgeMask, {}); } - return BlockMaskCache[BB] = BlockMask; + BlockMaskCache[BB] = BlockMask; } VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, @@ -8469,8 +8116,6 @@ VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, "Must be called with either a load or store"); auto willWiden = [&](ElementCount VF) -> bool { - if (VF.isScalar()) - return false; LoopVectorizationCostModel::InstWidening Decision = CM.getWideningDecision(I, VF); assert(Decision != LoopVectorizationCostModel::CM_Unknown && @@ -8488,7 +8133,7 @@ VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, VPValue *Mask = nullptr; if (Legal->isMaskRequired(I)) - Mask = createBlockInMask(I->getParent(), Plan); + Mask = getBlockInMask(I->getParent()); // Determine if the pointer operand of the access is either consecutive or // reverse consecutive. @@ -8498,70 +8143,72 @@ VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, bool Consecutive = Reverse || Decision == LoopVectorizationCostModel::CM_Widen; + VPValue *Ptr = isa<LoadInst>(I) ? Operands[0] : Operands[1]; + if (Consecutive) { + auto *GEP = dyn_cast<GetElementPtrInst>( + Ptr->getUnderlyingValue()->stripPointerCasts()); + auto *VectorPtr = new VPVectorPointerRecipe( + Ptr, getLoadStoreType(I), Reverse, GEP ? GEP->isInBounds() : false, + I->getDebugLoc()); + Builder.getInsertBlock()->appendRecipe(VectorPtr); + Ptr = VectorPtr; + } if (LoadInst *Load = dyn_cast<LoadInst>(I)) - return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask, - Consecutive, Reverse); + return new VPWidenMemoryInstructionRecipe(*Load, Ptr, Mask, Consecutive, + Reverse); StoreInst *Store = cast<StoreInst>(I); - return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0], - Mask, Consecutive, Reverse); + return new VPWidenMemoryInstructionRecipe(*Store, Ptr, Operands[0], Mask, + Consecutive, Reverse); } +/// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also +/// insert a recipe to expand the step for the induction recipe. static VPWidenIntOrFpInductionRecipe * -createWidenInductionRecipe(PHINode *Phi, Instruction *PhiOrTrunc, - VPValue *Start, const InductionDescriptor &IndDesc, - LoopVectorizationCostModel &CM, Loop &OrigLoop, - VFRange &Range) { - // Returns true if an instruction \p I should be scalarized instead of - // vectorized for the chosen vectorization factor. - auto ShouldScalarizeInstruction = [&CM](Instruction *I, ElementCount VF) { - return CM.isScalarAfterVectorization(I, VF) || - CM.isProfitableToScalarize(I, VF); - }; - - bool NeedsScalarIV = LoopVectorizationPlanner::getDecisionAndClampRange( - [&](ElementCount VF) { - // Returns true if we should generate a scalar version of \p IV. - if (ShouldScalarizeInstruction(PhiOrTrunc, VF)) - return true; - auto isScalarInst = [&](User *U) -> bool { - auto *I = cast<Instruction>(U); - return OrigLoop.contains(I) && ShouldScalarizeInstruction(I, VF); - }; - return any_of(PhiOrTrunc->users(), isScalarInst); - }, - Range); - bool NeedsScalarIVOnly = LoopVectorizationPlanner::getDecisionAndClampRange( - [&](ElementCount VF) { - return ShouldScalarizeInstruction(PhiOrTrunc, VF); - }, - Range); +createWidenInductionRecipes(PHINode *Phi, Instruction *PhiOrTrunc, + VPValue *Start, const InductionDescriptor &IndDesc, + VPlan &Plan, ScalarEvolution &SE, Loop &OrigLoop, + VFRange &Range) { assert(IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())); + assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && + "step must be loop invariant"); + + VPValue *Step = + vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep(), SE); if (auto *TruncI = dyn_cast<TruncInst>(PhiOrTrunc)) { - return new VPWidenIntOrFpInductionRecipe(Phi, Start, IndDesc, TruncI, - NeedsScalarIV, !NeedsScalarIVOnly); + return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc, TruncI); } assert(isa<PHINode>(PhiOrTrunc) && "must be a phi node here"); - return new VPWidenIntOrFpInductionRecipe(Phi, Start, IndDesc, NeedsScalarIV, - !NeedsScalarIVOnly); + return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc); } -VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionPHI( - PHINode *Phi, ArrayRef<VPValue *> Operands, VFRange &Range) const { +VPRecipeBase *VPRecipeBuilder::tryToOptimizeInductionPHI( + PHINode *Phi, ArrayRef<VPValue *> Operands, VPlan &Plan, VFRange &Range) { // Check if this is an integer or fp induction. If so, build the recipe that // produces its scalar and vector values. if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) - return createWidenInductionRecipe(Phi, Phi, Operands[0], *II, CM, *OrigLoop, - Range); - + return createWidenInductionRecipes(Phi, Phi, Operands[0], *II, Plan, + *PSE.getSE(), *OrigLoop, Range); + + // Check if this is pointer induction. If so, build the recipe for it. + if (auto *II = Legal->getPointerInductionDescriptor(Phi)) { + VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, II->getStep(), + *PSE.getSE()); + return new VPWidenPointerInductionRecipe( + Phi, Operands[0], Step, *II, + LoopVectorizationPlanner::getDecisionAndClampRange( + [&](ElementCount VF) { + return CM.isScalarAfterVectorization(Phi, VF); + }, + Range)); + } return nullptr; } VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( - TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, - VPlan &Plan) const { + TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, VPlan &Plan) { // Optimize the special case where the source is a constant integer // induction variable. Notice that we can only optimize the 'trunc' case // because (a) FP conversions lose precision, (b) sext/zext may wrap, and @@ -8581,8 +8228,9 @@ VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( auto *Phi = cast<PHINode>(I->getOperand(0)); const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); - VPValue *Start = Plan.getOrAddVPValue(II.getStartValue()); - return createWidenInductionRecipe(Phi, I, Start, II, CM, *OrigLoop, Range); + VPValue *Start = Plan.getVPValueOrAddLiveIn(II.getStartValue()); + return createWidenInductionRecipes(Phi, I, Start, II, Plan, *PSE.getSE(), + *OrigLoop, Range); } return nullptr; } @@ -8592,24 +8240,37 @@ VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, VPlanPtr &Plan) { // If all incoming values are equal, the incoming VPValue can be used directly // instead of creating a new VPBlendRecipe. - VPValue *FirstIncoming = Operands[0]; - if (all_of(Operands, [FirstIncoming](const VPValue *Inc) { - return FirstIncoming == Inc; - })) { + if (llvm::all_equal(Operands)) return Operands[0]; + + unsigned NumIncoming = Phi->getNumIncomingValues(); + // For in-loop reductions, we do not need to create an additional select. + VPValue *InLoopVal = nullptr; + for (unsigned In = 0; In < NumIncoming; In++) { + PHINode *PhiOp = + dyn_cast_or_null<PHINode>(Operands[In]->getUnderlyingValue()); + if (PhiOp && CM.isInLoopReduction(PhiOp)) { + assert(!InLoopVal && "Found more than one in-loop reduction!"); + InLoopVal = Operands[In]; + } } + assert((!InLoopVal || NumIncoming == 2) && + "Found an in-loop reduction for PHI with unexpected number of " + "incoming values"); + if (InLoopVal) + return Operands[Operands[0] == InLoopVal ? 1 : 0]; + // We know that all PHIs in non-header blocks are converted into selects, so // we don't have to worry about the insertion order and we can just use the // builder. At this point we generate the predication tree. There may be // duplications since this is a simple recursive scan, but future // optimizations will clean it up. SmallVector<VPValue *, 2> OperandsWithMask; - unsigned NumIncoming = Phi->getNumIncomingValues(); for (unsigned In = 0; In < NumIncoming; In++) { VPValue *EdgeMask = - createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); + createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), *Plan); assert((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask"); OperandsWithMask.push_back(Operands[In]); @@ -8621,8 +8282,8 @@ VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, ArrayRef<VPValue *> Operands, - VFRange &Range) const { - + VFRange &Range, + VPlanPtr &Plan) { bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( [this, CI](ElementCount VF) { return CM.isScalarWithPredication(CI, VF); @@ -8639,24 +8300,76 @@ VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, ID == Intrinsic::experimental_noalias_scope_decl)) return nullptr; - auto willWiden = [&](ElementCount VF) -> bool { - Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); - // The following case may be scalarized depending on the VF. - // The flag shows whether we use Intrinsic or a usual Call for vectorized - // version of the instruction. - // Is it beneficial to perform intrinsic call compared to lib call? - bool NeedToScalarize = false; - InstructionCost CallCost = CM.getVectorCallCost(CI, VF, NeedToScalarize); - InstructionCost IntrinsicCost = ID ? CM.getVectorIntrinsicCost(CI, VF) : 0; - bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; - return UseVectorIntrinsic || !NeedToScalarize; - }; + SmallVector<VPValue *, 4> Ops(Operands.take_front(CI->arg_size())); - if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) - return nullptr; + // Is it beneficial to perform intrinsic call compared to lib call? + bool ShouldUseVectorIntrinsic = + ID && LoopVectorizationPlanner::getDecisionAndClampRange( + [&](ElementCount VF) -> bool { + return CM.getCallWideningDecision(CI, VF).Kind == + LoopVectorizationCostModel::CM_IntrinsicCall; + }, + Range); + if (ShouldUseVectorIntrinsic) + return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), ID, + CI->getDebugLoc()); + + Function *Variant = nullptr; + std::optional<unsigned> MaskPos; + // Is better to call a vectorized version of the function than to to scalarize + // the call? + auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange( + [&](ElementCount VF) -> bool { + // The following case may be scalarized depending on the VF. + // The flag shows whether we can use a usual Call for vectorized + // version of the instruction. + + // If we've found a variant at a previous VF, then stop looking. A + // vectorized variant of a function expects input in a certain shape + // -- basically the number of input registers, the number of lanes + // per register, and whether there's a mask required. + // We store a pointer to the variant in the VPWidenCallRecipe, so + // once we have an appropriate variant it's only valid for that VF. + // This will force a different vplan to be generated for each VF that + // finds a valid variant. + if (Variant) + return false; + LoopVectorizationCostModel::CallWideningDecision Decision = + CM.getCallWideningDecision(CI, VF); + if (Decision.Kind == LoopVectorizationCostModel::CM_VectorCall) { + Variant = Decision.Variant; + MaskPos = Decision.MaskPos; + return true; + } - ArrayRef<VPValue *> Ops = Operands.take_front(CI->arg_size()); - return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end())); + return false; + }, + Range); + if (ShouldUseVectorCall) { + if (MaskPos.has_value()) { + // We have 2 cases that would require a mask: + // 1) The block needs to be predicated, either due to a conditional + // in the scalar loop or use of an active lane mask with + // tail-folding, and we use the appropriate mask for the block. + // 2) No mask is required for the block, but the only available + // vector variant at this VF requires a mask, so we synthesize an + // all-true mask. + VPValue *Mask = nullptr; + if (Legal->isMaskRequired(CI)) + Mask = getBlockInMask(CI->getParent()); + else + Mask = Plan->getVPValueOrAddLiveIn(ConstantInt::getTrue( + IntegerType::getInt1Ty(Variant->getFunctionType()->getContext()))); + + Ops.insert(Ops.begin() + *MaskPos, Mask); + } + + return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), + Intrinsic::not_intrinsic, CI->getDebugLoc(), + Variant); + } + + return nullptr; } bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { @@ -8673,54 +8386,53 @@ bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { Range); } -VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, - ArrayRef<VPValue *> Operands) const { - auto IsVectorizableOpcode = [](unsigned Opcode) { - switch (Opcode) { - case Instruction::Add: - case Instruction::And: - case Instruction::AShr: - case Instruction::BitCast: - case Instruction::FAdd: - case Instruction::FCmp: - case Instruction::FDiv: - case Instruction::FMul: - case Instruction::FNeg: - case Instruction::FPExt: - case Instruction::FPToSI: - case Instruction::FPToUI: - case Instruction::FPTrunc: - case Instruction::FRem: - case Instruction::FSub: - case Instruction::ICmp: - case Instruction::IntToPtr: - case Instruction::LShr: - case Instruction::Mul: - case Instruction::Or: - case Instruction::PtrToInt: - case Instruction::SDiv: - case Instruction::Select: - case Instruction::SExt: - case Instruction::Shl: - case Instruction::SIToFP: - case Instruction::SRem: - case Instruction::Sub: - case Instruction::Trunc: - case Instruction::UDiv: - case Instruction::UIToFP: - case Instruction::URem: - case Instruction::Xor: - case Instruction::ZExt: - return true; +VPRecipeBase *VPRecipeBuilder::tryToWiden(Instruction *I, + ArrayRef<VPValue *> Operands, + VPBasicBlock *VPBB, VPlanPtr &Plan) { + switch (I->getOpcode()) { + default: + return nullptr; + case Instruction::SDiv: + case Instruction::UDiv: + case Instruction::SRem: + case Instruction::URem: { + // If not provably safe, use a select to form a safe divisor before widening the + // div/rem operation itself. Otherwise fall through to general handling below. + if (CM.isPredicatedInst(I)) { + SmallVector<VPValue *> Ops(Operands.begin(), Operands.end()); + VPValue *Mask = getBlockInMask(I->getParent()); + VPValue *One = Plan->getVPValueOrAddLiveIn( + ConstantInt::get(I->getType(), 1u, false)); + auto *SafeRHS = + new VPInstruction(Instruction::Select, {Mask, Ops[1], One}, + I->getDebugLoc()); + VPBB->appendRecipe(SafeRHS); + Ops[1] = SafeRHS; + return new VPWidenRecipe(*I, make_range(Ops.begin(), Ops.end())); } - return false; + [[fallthrough]]; + } + case Instruction::Add: + case Instruction::And: + case Instruction::AShr: + case Instruction::FAdd: + case Instruction::FCmp: + case Instruction::FDiv: + case Instruction::FMul: + case Instruction::FNeg: + case Instruction::FRem: + case Instruction::FSub: + case Instruction::ICmp: + case Instruction::LShr: + case Instruction::Mul: + case Instruction::Or: + case Instruction::Select: + case Instruction::Shl: + case Instruction::Sub: + case Instruction::Xor: + case Instruction::Freeze: + return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); }; - - if (!IsVectorizableOpcode(I->getOpcode())) - return nullptr; - - // Success: widen this instruction. - return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); } void VPRecipeBuilder::fixHeaderPhis() { @@ -8733,16 +8445,14 @@ void VPRecipeBuilder::fixHeaderPhis() { } } -VPBasicBlock *VPRecipeBuilder::handleReplication( - Instruction *I, VFRange &Range, VPBasicBlock *VPBB, - VPlanPtr &Plan) { +VPRecipeOrVPValueTy VPRecipeBuilder::handleReplication(Instruction *I, + VFRange &Range, + VPlan &Plan) { bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, Range); - bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( - [&](ElementCount VF) { return CM.isPredicatedInst(I, VF, IsUniform); }, - Range); + bool IsPredicated = CM.isPredicatedInst(I); // Even if the instruction is not marked as uniform, there are certain // intrinsic calls that can be effectively treated as such, so we check for @@ -8774,130 +8484,73 @@ VPBasicBlock *VPRecipeBuilder::handleReplication( break; } } - - auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), - IsUniform, IsPredicated); - setRecipe(I, Recipe); - Plan->addVPValue(I, Recipe); - - // Find if I uses a predicated instruction. If so, it will use its scalar - // value. Avoid hoisting the insert-element which packs the scalar value into - // a vector value, as that happens iff all users use the vector value. - for (VPValue *Op : Recipe->operands()) { - auto *PredR = dyn_cast_or_null<VPPredInstPHIRecipe>(Op->getDef()); - if (!PredR) - continue; - auto *RepR = - cast_or_null<VPReplicateRecipe>(PredR->getOperand(0)->getDef()); - assert(RepR->isPredicated() && - "expected Replicate recipe to be predicated"); - RepR->setAlsoPack(false); - } - - // Finalize the recipe for Instr, first if it is not predicated. + VPValue *BlockInMask = nullptr; if (!IsPredicated) { + // Finalize the recipe for Instr, first if it is not predicated. LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n"); - VPBB->appendRecipe(Recipe); - return VPBB; - } - LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n"); - - VPBlockBase *SingleSucc = VPBB->getSingleSuccessor(); - assert(SingleSucc && "VPBB must have a single successor when handling " - "predicated replication."); - VPBlockUtils::disconnectBlocks(VPBB, SingleSucc); - // Record predicated instructions for above packing optimizations. - VPBlockBase *Region = createReplicateRegion(I, Recipe, Plan); - VPBlockUtils::insertBlockAfter(Region, VPBB); - auto *RegSucc = new VPBasicBlock(); - VPBlockUtils::insertBlockAfter(RegSucc, Region); - VPBlockUtils::connectBlocks(RegSucc, SingleSucc); - return RegSucc; -} - -VPRegionBlock *VPRecipeBuilder::createReplicateRegion(Instruction *Instr, - VPRecipeBase *PredRecipe, - VPlanPtr &Plan) { - // Instructions marked for predication are replicated and placed under an - // if-then construct to prevent side-effects. - - // Generate recipes to compute the block mask for this region. - VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); - - // Build the triangular if-then region. - std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); - assert(Instr->getParent() && "Predicated instruction not in any basic block"); - auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); - auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); - auto *PHIRecipe = Instr->getType()->isVoidTy() - ? nullptr - : new VPPredInstPHIRecipe(Plan->getOrAddVPValue(Instr)); - if (PHIRecipe) { - Plan->removeVPValueFor(Instr); - Plan->addVPValue(Instr, PHIRecipe); - } - auto *Exit = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); - auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); - VPRegionBlock *Region = new VPRegionBlock(Entry, Exit, RegionName, true); - - // Note: first set Entry as region entry and then connect successors starting - // from it in order, to propagate the "parent" of each VPBasicBlock. - VPBlockUtils::insertTwoBlocksAfter(Pred, Exit, BlockInMask, Entry); - VPBlockUtils::connectBlocks(Pred, Exit); - - return Region; + } else { + LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n"); + // Instructions marked for predication are replicated and a mask operand is + // added initially. Masked replicate recipes will later be placed under an + // if-then construct to prevent side-effects. Generate recipes to compute + // the block mask for this region. + BlockInMask = getBlockInMask(I->getParent()); + } + + auto *Recipe = new VPReplicateRecipe(I, Plan.mapToVPValues(I->operands()), + IsUniform, BlockInMask); + return toVPRecipeResult(Recipe); } VPRecipeOrVPValueTy VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, ArrayRef<VPValue *> Operands, - VFRange &Range, VPlanPtr &Plan) { - // First, check for specific widening recipes that deal with calls, memory - // operations, inductions and Phi nodes. - if (auto *CI = dyn_cast<CallInst>(Instr)) - return toVPRecipeResult(tryToWidenCall(CI, Operands, Range)); - - if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) - return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); - + VFRange &Range, VPBasicBlock *VPBB, + VPlanPtr &Plan) { + // First, check for specific widening recipes that deal with inductions, Phi + // nodes, calls and memory operations. VPRecipeBase *Recipe; if (auto Phi = dyn_cast<PHINode>(Instr)) { if (Phi->getParent() != OrigLoop->getHeader()) return tryToBlend(Phi, Operands, Plan); - if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, Range))) + + // Always record recipes for header phis. Later first-order recurrence phis + // can have earlier phis as incoming values. + recordRecipeOf(Phi); + + if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, *Plan, Range))) return toVPRecipeResult(Recipe); VPHeaderPHIRecipe *PhiRecipe = nullptr; - if (Legal->isReductionVariable(Phi) || Legal->isFirstOrderRecurrence(Phi)) { - VPValue *StartV = Operands[0]; - if (Legal->isReductionVariable(Phi)) { - const RecurrenceDescriptor &RdxDesc = - Legal->getReductionVars().find(Phi)->second; - assert(RdxDesc.getRecurrenceStartValue() == - Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())); - PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, - CM.isInLoopReduction(Phi), - CM.useOrderedReductions(RdxDesc)); - } else { - PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); - } - - // Record the incoming value from the backedge, so we can add the incoming - // value from the backedge after all recipes have been created. - recordRecipeOf(cast<Instruction>( - Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch()))); - PhisToFix.push_back(PhiRecipe); + assert((Legal->isReductionVariable(Phi) || + Legal->isFixedOrderRecurrence(Phi)) && + "can only widen reductions and fixed-order recurrences here"); + VPValue *StartV = Operands[0]; + if (Legal->isReductionVariable(Phi)) { + const RecurrenceDescriptor &RdxDesc = + Legal->getReductionVars().find(Phi)->second; + assert(RdxDesc.getRecurrenceStartValue() == + Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())); + PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, + CM.isInLoopReduction(Phi), + CM.useOrderedReductions(RdxDesc)); } else { - // TODO: record backedge value for remaining pointer induction phis. - assert(Phi->getType()->isPointerTy() && - "only pointer phis should be handled here"); - assert(Legal->getInductionVars().count(Phi) && - "Not an induction variable"); - InductionDescriptor II = Legal->getInductionVars().lookup(Phi); - VPValue *Start = Plan->getOrAddVPValue(II.getStartValue()); - PhiRecipe = new VPWidenPHIRecipe(Phi, Start); + // TODO: Currently fixed-order recurrences are modeled as chains of + // first-order recurrences. If there are no users of the intermediate + // recurrences in the chain, the fixed order recurrence should be modeled + // directly, enabling more efficient codegen. + PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); } + // Record the incoming value from the backedge, so we can add the incoming + // value from the backedge after all recipes have been created. + auto *Inc = cast<Instruction>( + Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); + auto RecipeIter = Ingredient2Recipe.find(Inc); + if (RecipeIter == Ingredient2Recipe.end()) + recordRecipeOf(Inc); + + PhisToFix.push_back(PhiRecipe); return toVPRecipeResult(PhiRecipe); } @@ -8906,112 +8559,108 @@ VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, Range, *Plan))) return toVPRecipeResult(Recipe); + // All widen recipes below deal only with VF > 1. + if (LoopVectorizationPlanner::getDecisionAndClampRange( + [&](ElementCount VF) { return VF.isScalar(); }, Range)) + return nullptr; + + if (auto *CI = dyn_cast<CallInst>(Instr)) + return toVPRecipeResult(tryToWidenCall(CI, Operands, Range, Plan)); + + if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) + return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); + if (!shouldWiden(Instr, Range)) return nullptr; if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) return toVPRecipeResult(new VPWidenGEPRecipe( - GEP, make_range(Operands.begin(), Operands.end()), OrigLoop)); + GEP, make_range(Operands.begin(), Operands.end()))); if (auto *SI = dyn_cast<SelectInst>(Instr)) { - bool InvariantCond = - PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); return toVPRecipeResult(new VPWidenSelectRecipe( - *SI, make_range(Operands.begin(), Operands.end()), InvariantCond)); + *SI, make_range(Operands.begin(), Operands.end()))); + } + + if (auto *CI = dyn_cast<CastInst>(Instr)) { + return toVPRecipeResult(new VPWidenCastRecipe(CI->getOpcode(), Operands[0], + CI->getType(), *CI)); } - return toVPRecipeResult(tryToWiden(Instr, Operands)); + return toVPRecipeResult(tryToWiden(Instr, Operands, VPBB, Plan)); } void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, ElementCount MaxVF) { assert(OrigLoop->isInnermost() && "Inner loop expected."); - // Collect instructions from the original loop that will become trivially dead - // in the vectorized loop. We don't need to vectorize these instructions. For - // example, original induction update instructions can become dead because we - // separately emit induction "steps" when generating code for the new loop. - // Similarly, we create a new latch condition when setting up the structure - // of the new loop, so the old one can become dead. - SmallPtrSet<Instruction *, 4> DeadInstructions; - collectTriviallyDeadInstructions(DeadInstructions); - - // Add assume instructions we need to drop to DeadInstructions, to prevent - // them from being added to the VPlan. - // TODO: We only need to drop assumes in blocks that get flattend. If the - // control flow is preserved, we should keep them. - auto &ConditionalAssumes = Legal->getConditionalAssumes(); - DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); - - MapVector<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); - // Dead instructions do not need sinking. Remove them from SinkAfter. - for (Instruction *I : DeadInstructions) - SinkAfter.erase(I); - - // Cannot sink instructions after dead instructions (there won't be any - // recipes for them). Instead, find the first non-dead previous instruction. - for (auto &P : Legal->getSinkAfter()) { - Instruction *SinkTarget = P.second; - Instruction *FirstInst = &*SinkTarget->getParent()->begin(); - (void)FirstInst; - while (DeadInstructions.contains(SinkTarget)) { - assert( - SinkTarget != FirstInst && - "Must find a live instruction (at least the one feeding the " - "first-order recurrence PHI) before reaching beginning of the block"); - SinkTarget = SinkTarget->getPrevNode(); - assert(SinkTarget != P.first && - "sink source equals target, no sinking required"); + auto MaxVFTimes2 = MaxVF * 2; + for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) { + VFRange SubRange = {VF, MaxVFTimes2}; + if (auto Plan = tryToBuildVPlanWithVPRecipes(SubRange)) { + // Now optimize the initial VPlan. + if (!Plan->hasVF(ElementCount::getFixed(1))) + VPlanTransforms::truncateToMinimalBitwidths( + *Plan, CM.getMinimalBitwidths(), PSE.getSE()->getContext()); + VPlanTransforms::optimize(*Plan, *PSE.getSE()); + assert(VPlanVerifier::verifyPlanIsValid(*Plan) && "VPlan is invalid"); + VPlans.push_back(std::move(Plan)); } - P.second = SinkTarget; - } - - auto MaxVFPlusOne = MaxVF.getWithIncrement(1); - for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { - VFRange SubRange = {VF, MaxVFPlusOne}; - VPlans.push_back( - buildVPlanWithVPRecipes(SubRange, DeadInstructions, SinkAfter)); VF = SubRange.End; } } -// Add a VPCanonicalIVPHIRecipe starting at 0 to the header, a -// CanonicalIVIncrement{NUW} VPInstruction to increment it by VF * UF and a -// BranchOnCount VPInstruction to the latch. -static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, DebugLoc DL, - bool HasNUW, bool IsVPlanNative) { +// Add the necessary canonical IV and branch recipes required to control the +// loop. +static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, bool HasNUW, + DebugLoc DL) { Value *StartIdx = ConstantInt::get(IdxTy, 0); - auto *StartV = Plan.getOrAddVPValue(StartIdx); + auto *StartV = Plan.getVPValueOrAddLiveIn(StartIdx); + // Add a VPCanonicalIVPHIRecipe starting at 0 to the header. auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL); VPRegionBlock *TopRegion = Plan.getVectorLoopRegion(); VPBasicBlock *Header = TopRegion->getEntryBasicBlock(); - if (IsVPlanNative) - Header = cast<VPBasicBlock>(Header->getSingleSuccessor()); Header->insert(CanonicalIVPHI, Header->begin()); + // Add a CanonicalIVIncrement{NUW} VPInstruction to increment the scalar + // IV by VF * UF. auto *CanonicalIVIncrement = - new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementNUW - : VPInstruction::CanonicalIVIncrement, - {CanonicalIVPHI}, DL); + new VPInstruction(Instruction::Add, {CanonicalIVPHI, &Plan.getVFxUF()}, + {HasNUW, false}, DL, "index.next"); CanonicalIVPHI->addOperand(CanonicalIVIncrement); - VPBasicBlock *EB = TopRegion->getExitBasicBlock(); - if (IsVPlanNative) { - EB = cast<VPBasicBlock>(EB->getSinglePredecessor()); - EB->setCondBit(nullptr); - } + VPBasicBlock *EB = TopRegion->getExitingBasicBlock(); EB->appendRecipe(CanonicalIVIncrement); - auto *BranchOnCount = + // Add the BranchOnCount VPInstruction to the latch. + VPInstruction *BranchBack = new VPInstruction(VPInstruction::BranchOnCount, {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL); - EB->appendRecipe(BranchOnCount); + EB->appendRecipe(BranchBack); } -VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( - VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, - const MapVector<Instruction *, Instruction *> &SinkAfter) { +// Add exit values to \p Plan. VPLiveOuts are added for each LCSSA phi in the +// original exit block. +static void addUsersInExitBlock(VPBasicBlock *HeaderVPBB, Loop *OrigLoop, + VPlan &Plan) { + BasicBlock *ExitBB = OrigLoop->getUniqueExitBlock(); + BasicBlock *ExitingBB = OrigLoop->getExitingBlock(); + // Only handle single-exit loops with unique exit blocks for now. + if (!ExitBB || !ExitBB->getSinglePredecessor() || !ExitingBB) + return; + + // Introduce VPUsers modeling the exit values. + for (PHINode &ExitPhi : ExitBB->phis()) { + Value *IncomingValue = + ExitPhi.getIncomingValueForBlock(ExitingBB); + VPValue *V = Plan.getVPValueOrAddLiveIn(IncomingValue); + Plan.addLiveOut(&ExitPhi, V); + } +} + +VPlanPtr +LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range) { SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; @@ -9022,39 +8671,21 @@ VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( // process after constructing the initial VPlan. // --------------------------------------------------------------------------- - // Mark instructions we'll need to sink later and their targets as - // ingredients whose recipe we'll need to record. - for (auto &Entry : SinkAfter) { - RecipeBuilder.recordRecipeOf(Entry.first); - RecipeBuilder.recordRecipeOf(Entry.second); - } - for (auto &Reduction : CM.getInLoopReductionChains()) { - PHINode *Phi = Reduction.first; - RecurKind Kind = - Legal->getReductionVars().find(Phi)->second.getRecurrenceKind(); - const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; - - RecipeBuilder.recordRecipeOf(Phi); - for (auto &R : ReductionOperations) { - RecipeBuilder.recordRecipeOf(R); - // For min/max reducitons, where we have a pair of icmp/select, we also - // need to record the ICmp recipe, so it can be removed later. - assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && - "Only min/max recurrences allowed for inloop reductions"); - if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) - RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); - } - } - // For each interleave group which is relevant for this (possibly trimmed) // Range, add it to the set of groups to be later applied to the VPlan and add // placeholders for its members' Recipes which we'll be replacing with a // single VPInterleaveRecipe. for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { auto applyIG = [IG, this](ElementCount VF) -> bool { - return (VF.isVector() && // Query is illegal for VF == 1 - CM.getWideningDecision(IG->getInsertPos(), VF) == - LoopVectorizationCostModel::CM_Interleave); + bool Result = (VF.isVector() && // Query is illegal for VF == 1 + CM.getWideningDecision(IG->getInsertPos(), VF) == + LoopVectorizationCostModel::CM_Interleave); + // For scalable vectors, the only interleave factor currently supported + // is 2 since we require the (de)interleave2 intrinsics instead of + // shufflevectors. + assert((!Result || !VF.isScalable() || IG->getFactor() == 2) && + "Unsupported interleave factor for scalable vectors"); + return Result; }; if (!getDecisionAndClampRange(applyIG, Range)) continue; @@ -9069,18 +8700,34 @@ VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( // visit each basic block after having visited its predecessor basic blocks. // --------------------------------------------------------------------------- - // Create initial VPlan skeleton, with separate header and latch blocks. - VPBasicBlock *HeaderVPBB = new VPBasicBlock(); + // Create initial VPlan skeleton, having a basic block for the pre-header + // which contains SCEV expansions that need to happen before the CFG is + // modified; a basic block for the vector pre-header, followed by a region for + // the vector loop, followed by the middle basic block. The skeleton vector + // loop region contains a header and latch basic blocks. + VPlanPtr Plan = VPlan::createInitialVPlan( + createTripCountSCEV(Legal->getWidestInductionType(), PSE, OrigLoop), + *PSE.getSE()); + VPBasicBlock *HeaderVPBB = new VPBasicBlock("vector.body"); VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch"); VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB); - auto *TopRegion = new VPRegionBlock(HeaderVPBB, LatchVPBB, "vector loop"); - auto Plan = std::make_unique<VPlan>(TopRegion); - - Instruction *DLInst = - getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); - addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), - DLInst ? DLInst->getDebugLoc() : DebugLoc(), - !CM.foldTailByMasking(), false); + Plan->getVectorLoopRegion()->setEntry(HeaderVPBB); + Plan->getVectorLoopRegion()->setExiting(LatchVPBB); + + // Don't use getDecisionAndClampRange here, because we don't know the UF + // so this function is better to be conservative, rather than to split + // it up into different VPlans. + // TODO: Consider using getDecisionAndClampRange here to split up VPlans. + bool IVUpdateMayOverflow = false; + for (ElementCount VF : Range) + IVUpdateMayOverflow |= !isIndvarOverflowCheckKnownFalse(&CM, VF); + + DebugLoc DL = getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); + TailFoldingStyle Style = CM.getTailFoldingStyle(IVUpdateMayOverflow); + // When not folding the tail, we know that the induction increment will not + // overflow. + bool HasNUW = Style == TailFoldingStyle::None; + addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), HasNUW, DL); // Scan the body of the loop in a topological order to visit each basic block // after having visited its predecessor basic blocks. @@ -9088,91 +8735,98 @@ VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( DFS.perform(LI); VPBasicBlock *VPBB = HeaderVPBB; - SmallVector<VPWidenIntOrFpInductionRecipe *> InductionsToMove; + bool NeedsMasks = CM.foldTailByMasking() || + any_of(OrigLoop->blocks(), [this](BasicBlock *BB) { + return Legal->blockNeedsPredication(BB); + }); for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { // Relevant instructions from basic block BB will be grouped into VPRecipe // ingredients and fill a new VPBasicBlock. - unsigned VPBBsForBB = 0; - VPBB->setName(BB->getName()); + if (VPBB != HeaderVPBB) + VPBB->setName(BB->getName()); Builder.setInsertPoint(VPBB); + if (VPBB == HeaderVPBB) + RecipeBuilder.createHeaderMask(*Plan); + else if (NeedsMasks) + RecipeBuilder.createBlockInMask(BB, *Plan); + // Introduce each ingredient into VPlan. - // TODO: Model and preserve debug instrinsics in VPlan. - for (Instruction &I : BB->instructionsWithoutDebug()) { + // TODO: Model and preserve debug intrinsics in VPlan. + for (Instruction &I : drop_end(BB->instructionsWithoutDebug(false))) { Instruction *Instr = &I; - - // First filter out irrelevant instructions, to ensure no recipes are - // built for them. - if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) - continue; - SmallVector<VPValue *, 4> Operands; auto *Phi = dyn_cast<PHINode>(Instr); if (Phi && Phi->getParent() == OrigLoop->getHeader()) { - Operands.push_back(Plan->getOrAddVPValue( + Operands.push_back(Plan->getVPValueOrAddLiveIn( Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))); } else { auto OpRange = Plan->mapToVPValues(Instr->operands()); Operands = {OpRange.begin(), OpRange.end()}; } - if (auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( - Instr, Operands, Range, Plan)) { - // If Instr can be simplified to an existing VPValue, use it. - if (RecipeOrValue.is<VPValue *>()) { - auto *VPV = RecipeOrValue.get<VPValue *>(); - Plan->addVPValue(Instr, VPV); - // If the re-used value is a recipe, register the recipe for the - // instruction, in case the recipe for Instr needs to be recorded. - if (auto *R = dyn_cast_or_null<VPRecipeBase>(VPV->getDef())) - RecipeBuilder.setRecipe(Instr, R); - continue; - } - // Otherwise, add the new recipe. - VPRecipeBase *Recipe = RecipeOrValue.get<VPRecipeBase *>(); - for (auto *Def : Recipe->definedValues()) { - auto *UV = Def->getUnderlyingValue(); - Plan->addVPValue(UV, Def); - } - if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && - HeaderVPBB->getFirstNonPhi() != VPBB->end()) { - // Keep track of VPWidenIntOrFpInductionRecipes not in the phi section - // of the header block. That can happen for truncates of induction - // variables. Those recipes are moved to the phi section of the header - // block after applying SinkAfter, which relies on the original - // position of the trunc. - assert(isa<TruncInst>(Instr)); - InductionsToMove.push_back( - cast<VPWidenIntOrFpInductionRecipe>(Recipe)); - } - RecipeBuilder.setRecipe(Instr, Recipe); - VPBB->appendRecipe(Recipe); + // Invariant stores inside loop will be deleted and a single store + // with the final reduction value will be added to the exit block + StoreInst *SI; + if ((SI = dyn_cast<StoreInst>(&I)) && + Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) continue; - } - // Otherwise, if all widening options failed, Instruction is to be - // replicated. This may create a successor for VPBB. - VPBasicBlock *NextVPBB = - RecipeBuilder.handleReplication(Instr, Range, VPBB, Plan); - if (NextVPBB != VPBB) { - VPBB = NextVPBB; - VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) - : ""); + auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( + Instr, Operands, Range, VPBB, Plan); + if (!RecipeOrValue) + RecipeOrValue = RecipeBuilder.handleReplication(Instr, Range, *Plan); + // If Instr can be simplified to an existing VPValue, use it. + if (isa<VPValue *>(RecipeOrValue)) { + auto *VPV = cast<VPValue *>(RecipeOrValue); + Plan->addVPValue(Instr, VPV); + // If the re-used value is a recipe, register the recipe for the + // instruction, in case the recipe for Instr needs to be recorded. + if (VPRecipeBase *R = VPV->getDefiningRecipe()) + RecipeBuilder.setRecipe(Instr, R); + continue; + } + // Otherwise, add the new recipe. + VPRecipeBase *Recipe = cast<VPRecipeBase *>(RecipeOrValue); + for (auto *Def : Recipe->definedValues()) { + auto *UV = Def->getUnderlyingValue(); + Plan->addVPValue(UV, Def); } + + RecipeBuilder.setRecipe(Instr, Recipe); + if (isa<VPHeaderPHIRecipe>(Recipe)) { + // VPHeaderPHIRecipes must be kept in the phi section of HeaderVPBB. In + // the following cases, VPHeaderPHIRecipes may be created after non-phi + // recipes and need to be moved to the phi section of HeaderVPBB: + // * tail-folding (non-phi recipes computing the header mask are + // introduced earlier than regular header phi recipes, and should appear + // after them) + // * Optimizing truncates to VPWidenIntOrFpInductionRecipe. + + assert((HeaderVPBB->getFirstNonPhi() == VPBB->end() || + CM.foldTailByMasking() || isa<TruncInst>(Instr)) && + "unexpected recipe needs moving"); + Recipe->insertBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); + } else + VPBB->appendRecipe(Recipe); } VPBlockUtils::insertBlockAfter(new VPBasicBlock(), VPBB); VPBB = cast<VPBasicBlock>(VPBB->getSingleSuccessor()); } - // Fold the last, empty block into its predecessor. - VPBB = VPBlockUtils::tryToMergeBlockIntoPredecessor(VPBB); - assert(VPBB && "expected to fold last (empty) block"); // After here, VPBB should not be used. VPBB = nullptr; - assert(isa<VPRegionBlock>(Plan->getEntry()) && - !Plan->getEntry()->getEntryBasicBlock()->empty() && + if (CM.requiresScalarEpilogue(Range)) { + // No edge from the middle block to the unique exit block has been inserted + // and there is nothing to fix from vector loop; phis should have incoming + // from scalar loop only. + } else + addUsersInExitBlock(HeaderVPBB, OrigLoop, *Plan); + + assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && + !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock"); RecipeBuilder.fixHeaderPhis(); @@ -9182,110 +8836,13 @@ VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( // bring the VPlan to its final state. // --------------------------------------------------------------------------- - // Apply Sink-After legal constraints. - auto GetReplicateRegion = [](VPRecipeBase *R) -> VPRegionBlock * { - auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent()); - if (Region && Region->isReplicator()) { - assert(Region->getNumSuccessors() == 1 && - Region->getNumPredecessors() == 1 && "Expected SESE region!"); - assert(R->getParent()->size() == 1 && - "A recipe in an original replicator region must be the only " - "recipe in its block"); - return Region; - } - return nullptr; - }; - for (auto &Entry : SinkAfter) { - VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); - VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); - - auto *TargetRegion = GetReplicateRegion(Target); - auto *SinkRegion = GetReplicateRegion(Sink); - if (!SinkRegion) { - // If the sink source is not a replicate region, sink the recipe directly. - if (TargetRegion) { - // The target is in a replication region, make sure to move Sink to - // the block after it, not into the replication region itself. - VPBasicBlock *NextBlock = - cast<VPBasicBlock>(TargetRegion->getSuccessors().front()); - Sink->moveBefore(*NextBlock, NextBlock->getFirstNonPhi()); - } else - Sink->moveAfter(Target); - continue; - } - - // The sink source is in a replicate region. Unhook the region from the CFG. - auto *SinkPred = SinkRegion->getSinglePredecessor(); - auto *SinkSucc = SinkRegion->getSingleSuccessor(); - VPBlockUtils::disconnectBlocks(SinkPred, SinkRegion); - VPBlockUtils::disconnectBlocks(SinkRegion, SinkSucc); - VPBlockUtils::connectBlocks(SinkPred, SinkSucc); - - if (TargetRegion) { - // The target recipe is also in a replicate region, move the sink region - // after the target region. - auto *TargetSucc = TargetRegion->getSingleSuccessor(); - VPBlockUtils::disconnectBlocks(TargetRegion, TargetSucc); - VPBlockUtils::connectBlocks(TargetRegion, SinkRegion); - VPBlockUtils::connectBlocks(SinkRegion, TargetSucc); - } else { - // The sink source is in a replicate region, we need to move the whole - // replicate region, which should only contain a single recipe in the - // main block. - auto *SplitBlock = - Target->getParent()->splitAt(std::next(Target->getIterator())); - - auto *SplitPred = SplitBlock->getSinglePredecessor(); - - VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock); - VPBlockUtils::connectBlocks(SplitPred, SinkRegion); - VPBlockUtils::connectBlocks(SinkRegion, SplitBlock); - } - } - - VPlanTransforms::removeRedundantCanonicalIVs(*Plan); - VPlanTransforms::removeRedundantInductionCasts(*Plan); - - // Now that sink-after is done, move induction recipes for optimized truncates - // to the phi section of the header block. - for (VPWidenIntOrFpInductionRecipe *Ind : InductionsToMove) - Ind->moveBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); - // Adjust the recipes for any inloop reductions. - adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExit()), Plan, - RecipeBuilder, Range.Start); - - // Introduce a recipe to combine the incoming and previous values of a - // first-order recurrence. - for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { - auto *RecurPhi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R); - if (!RecurPhi) - continue; - - VPRecipeBase *PrevRecipe = RecurPhi->getBackedgeRecipe(); - VPBasicBlock *InsertBlock = PrevRecipe->getParent(); - auto *Region = GetReplicateRegion(PrevRecipe); - if (Region) - InsertBlock = cast<VPBasicBlock>(Region->getSingleSuccessor()); - if (Region || PrevRecipe->isPhi()) - Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi()); - else - Builder.setInsertPoint(InsertBlock, std::next(PrevRecipe->getIterator())); - - auto *RecurSplice = cast<VPInstruction>( - Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice, - {RecurPhi, RecurPhi->getBackedgeValue()})); - - RecurPhi->replaceAllUsesWith(RecurSplice); - // Set the first operand of RecurSplice to RecurPhi again, after replacing - // all users. - RecurSplice->setOperand(0, RecurPhi); - } + adjustRecipesForReductions(LatchVPBB, Plan, RecipeBuilder, Range.Start); // Interleave memory: for each Interleave Group we marked earlier as relevant // for this VPlan, replace the Recipes widening its memory instructions with a // single VPInterleaveRecipe at its insertion point. - for (auto IG : InterleaveGroups) { + for (const auto *IG : InterleaveGroups) { auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( RecipeBuilder.getRecipe(IG->getInsertPos())); SmallVector<VPValue *, 4> StoredValues; @@ -9296,49 +8853,62 @@ VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( StoredValues.push_back(StoreR->getStoredValue()); } + bool NeedsMaskForGaps = + IG->requiresScalarEpilogue() && !CM.isScalarEpilogueAllowed(); auto *VPIG = new VPInterleaveRecipe(IG, Recipe->getAddr(), StoredValues, - Recipe->getMask()); + Recipe->getMask(), NeedsMaskForGaps); VPIG->insertBefore(Recipe); unsigned J = 0; for (unsigned i = 0; i < IG->getFactor(); ++i) if (Instruction *Member = IG->getMember(i)) { + VPRecipeBase *MemberR = RecipeBuilder.getRecipe(Member); if (!Member->getType()->isVoidTy()) { - VPValue *OriginalV = Plan->getVPValue(Member); - Plan->removeVPValueFor(Member); - Plan->addVPValue(Member, VPIG->getVPValue(J)); + VPValue *OriginalV = MemberR->getVPSingleValue(); OriginalV->replaceAllUsesWith(VPIG->getVPValue(J)); J++; } - RecipeBuilder.getRecipe(Member)->eraseFromParent(); + MemberR->eraseFromParent(); } } + for (ElementCount VF : Range) + Plan->addVF(VF); + Plan->setName("Initial VPlan"); + + // Replace VPValues for known constant strides guaranteed by predicate scalar + // evolution. + for (auto [_, Stride] : Legal->getLAI()->getSymbolicStrides()) { + auto *StrideV = cast<SCEVUnknown>(Stride)->getValue(); + auto *ScevStride = dyn_cast<SCEVConstant>(PSE.getSCEV(StrideV)); + // Only handle constant strides for now. + if (!ScevStride) + continue; + Constant *CI = ConstantInt::get(Stride->getType(), ScevStride->getAPInt()); + + auto *ConstVPV = Plan->getVPValueOrAddLiveIn(CI); + // The versioned value may not be used in the loop directly, so just add a + // new live-in in those cases. + Plan->getVPValueOrAddLiveIn(StrideV)->replaceAllUsesWith(ConstVPV); + } + // From this point onwards, VPlan-to-VPlan transformations may change the plan // in ways that accessing values using original IR values is incorrect. Plan->disableValue2VPValue(); - VPlanTransforms::sinkScalarOperands(*Plan); - VPlanTransforms::mergeReplicateRegions(*Plan); + // Sink users of fixed-order recurrence past the recipe defining the previous + // value and introduce FirstOrderRecurrenceSplice VPInstructions. + if (!VPlanTransforms::adjustFixedOrderRecurrences(*Plan, Builder)) + return nullptr; - std::string PlanName; - raw_string_ostream RSO(PlanName); - ElementCount VF = Range.Start; - Plan->addVF(VF); - RSO << "Initial VPlan for VF={" << VF; - for (VF *= 2; ElementCount::isKnownLT(VF, Range.End); VF *= 2) { - Plan->addVF(VF); - RSO << "," << VF; + if (useActiveLaneMask(Style)) { + // TODO: Move checks to VPlanTransforms::addActiveLaneMask once + // TailFoldingStyle is visible there. + bool ForControlFlow = useActiveLaneMaskForControlFlow(Style); + bool WithoutRuntimeCheck = + Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; + VPlanTransforms::addActiveLaneMask(*Plan, ForControlFlow, + WithoutRuntimeCheck); } - RSO << "},UF>=1"; - RSO.flush(); - Plan->setName(PlanName); - - // Fold Exit block into its predecessor if possible. - // TODO: Fold block earlier once all VPlan transforms properly maintain a - // VPBasicBlock as exit. - VPBlockUtils::tryToMergeBlockIntoPredecessor(TopRegion->getExit()); - - assert(VPlanVerifier::verifyPlanIsValid(*Plan) && "VPlan is invalid"); return Plan; } @@ -9351,33 +8921,33 @@ VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { assert(EnableVPlanNativePath && "VPlan-native path is not enabled."); // Create new empty VPlan - auto Plan = std::make_unique<VPlan>(); + auto Plan = VPlan::createInitialVPlan( + createTripCountSCEV(Legal->getWidestInductionType(), PSE, OrigLoop), + *PSE.getSE()); // Build hierarchical CFG VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); HCFGBuilder.buildHierarchicalCFG(); - for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); - VF *= 2) + for (ElementCount VF : Range) Plan->addVF(VF); - if (EnableVPlanPredication) { - VPlanPredicator VPP(*Plan); - VPP.predicate(); - - // Avoid running transformation to recipes until masked code generation in - // VPlan-native path is in place. - return Plan; - } - - SmallPtrSet<Instruction *, 1> DeadInstructions; VPlanTransforms::VPInstructionsToVPRecipes( - OrigLoop, Plan, + Plan, [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, - DeadInstructions, *PSE.getSE()); - - addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), DebugLoc(), - true, true); + *PSE.getSE(), *TLI); + + // Remove the existing terminator of the exiting block of the top-most region. + // A BranchOnCount will be added instead when adding the canonical IV recipes. + auto *Term = + Plan->getVectorLoopRegion()->getExitingBasicBlock()->getTerminator(); + Term->eraseFromParent(); + + // Tail folding is not supported for outer loops, so the induction increment + // is guaranteed to not wrap. + bool HasNUW = true; + addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), HasNUW, + DebugLoc()); return Plan; } @@ -9386,103 +8956,242 @@ VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { // to reductions, with one operand being vector and the other being the scalar // reduction chain. For other reductions, a select is introduced between the phi // and live-out recipes when folding the tail. +// +// A ComputeReductionResult recipe is added to the middle block, also for +// in-loop reductions which compute their result in-loop, because generating +// the subsequent bc.merge.rdx phi is driven by ComputeReductionResult recipes. void LoopVectorizationPlanner::adjustRecipesForReductions( VPBasicBlock *LatchVPBB, VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, ElementCount MinVF) { - for (auto &Reduction : CM.getInLoopReductionChains()) { - PHINode *Phi = Reduction.first; - const RecurrenceDescriptor &RdxDesc = - Legal->getReductionVars().find(Phi)->second; - const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; + VPRegionBlock *VectorLoopRegion = Plan->getVectorLoopRegion(); + VPBasicBlock *Header = VectorLoopRegion->getEntryBasicBlock(); + // Gather all VPReductionPHIRecipe and sort them so that Intermediate stores + // sank outside of the loop would keep the same order as they had in the + // original loop. + SmallVector<VPReductionPHIRecipe *> ReductionPHIList; + for (VPRecipeBase &R : Header->phis()) { + if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) + ReductionPHIList.emplace_back(ReductionPhi); + } + bool HasIntermediateStore = false; + stable_sort(ReductionPHIList, + [this, &HasIntermediateStore](const VPReductionPHIRecipe *R1, + const VPReductionPHIRecipe *R2) { + auto *IS1 = R1->getRecurrenceDescriptor().IntermediateStore; + auto *IS2 = R2->getRecurrenceDescriptor().IntermediateStore; + HasIntermediateStore |= IS1 || IS2; + + // If neither of the recipes has an intermediate store, keep the + // order the same. + if (!IS1 && !IS2) + return false; + + // If only one of the recipes has an intermediate store, then + // move it towards the beginning of the list. + if (IS1 && !IS2) + return true; + + if (!IS1 && IS2) + return false; + + // If both recipes have an intermediate store, then the recipe + // with the later store should be processed earlier. So it + // should go to the beginning of the list. + return DT->dominates(IS2, IS1); + }); + + if (HasIntermediateStore && ReductionPHIList.size() > 1) + for (VPRecipeBase *R : ReductionPHIList) + R->moveBefore(*Header, Header->getFirstNonPhi()); - if (MinVF.isScalar() && !CM.useOrderedReductions(RdxDesc)) + for (VPRecipeBase &R : Header->phis()) { + auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); + if (!PhiR || !PhiR->isInLoop() || (MinVF.isScalar() && !PhiR->isOrdered())) continue; - // ReductionOperations are orders top-down from the phi's use to the - // LoopExitValue. We keep a track of the previous item (the Chain) to tell - // which of the two operands will remain scalar and which will be reduced. - // For minmax the chain will be the select instructions. - Instruction *Chain = Phi; - for (Instruction *R : ReductionOperations) { - VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); - RecurKind Kind = RdxDesc.getRecurrenceKind(); - - VPValue *ChainOp = Plan->getVPValue(Chain); - unsigned FirstOpId; - assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && - "Only min/max recurrences allowed for inloop reductions"); - // Recognize a call to the llvm.fmuladd intrinsic. - bool IsFMulAdd = (Kind == RecurKind::FMulAdd); - assert((!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) && - "Expected instruction to be a call to the llvm.fmuladd intrinsic"); - if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { - assert(isa<VPWidenSelectRecipe>(WidenRecipe) && - "Expected to replace a VPWidenSelectSC"); - FirstOpId = 1; - } else { - assert((MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || - (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && - "Expected to replace a VPWidenSC"); - FirstOpId = 0; + const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); + RecurKind Kind = RdxDesc.getRecurrenceKind(); + assert(!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) && + "AnyOf reductions are not allowed for in-loop reductions"); + + // Collect the chain of "link" recipes for the reduction starting at PhiR. + SetVector<VPSingleDefRecipe *> Worklist; + Worklist.insert(PhiR); + for (unsigned I = 0; I != Worklist.size(); ++I) { + VPSingleDefRecipe *Cur = Worklist[I]; + for (VPUser *U : Cur->users()) { + auto *UserRecipe = dyn_cast<VPSingleDefRecipe>(U); + if (!UserRecipe) { + assert(isa<VPLiveOut>(U) && + "U must either be a VPSingleDef or VPLiveOut"); + continue; + } + Worklist.insert(UserRecipe); } - unsigned VecOpId = - R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; - VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); - - auto *CondOp = CM.foldTailByMasking() - ? RecipeBuilder.createBlockInMask(R->getParent(), Plan) - : nullptr; + } + // Visit operation "Links" along the reduction chain top-down starting from + // the phi until LoopExitValue. We keep track of the previous item + // (PreviousLink) to tell which of the two operands of a Link will remain + // scalar and which will be reduced. For minmax by select(cmp), Link will be + // the select instructions. + VPSingleDefRecipe *PreviousLink = PhiR; // Aka Worklist[0]. + for (VPSingleDefRecipe *CurrentLink : Worklist.getArrayRef().drop_front()) { + Instruction *CurrentLinkI = CurrentLink->getUnderlyingInstr(); + + // Index of the first operand which holds a non-mask vector operand. + unsigned IndexOfFirstOperand; + // Recognize a call to the llvm.fmuladd intrinsic. + bool IsFMulAdd = (Kind == RecurKind::FMulAdd); + VPValue *VecOp; + VPBasicBlock *LinkVPBB = CurrentLink->getParent(); if (IsFMulAdd) { + assert( + RecurrenceDescriptor::isFMulAddIntrinsic(CurrentLinkI) && + "Expected instruction to be a call to the llvm.fmuladd intrinsic"); + assert(((MinVF.isScalar() && isa<VPReplicateRecipe>(CurrentLink)) || + isa<VPWidenCallRecipe>(CurrentLink)) && + CurrentLink->getOperand(2) == PreviousLink && + "expected a call where the previous link is the added operand"); + // If the instruction is a call to the llvm.fmuladd intrinsic then we - // need to create an fmul recipe to use as the vector operand for the - // fadd reduction. + // need to create an fmul recipe (multiplying the first two operands of + // the fmuladd together) to use as the vector operand for the fadd + // reduction. VPInstruction *FMulRecipe = new VPInstruction( - Instruction::FMul, {VecOp, Plan->getVPValue(R->getOperand(1))}); - FMulRecipe->setFastMathFlags(R->getFastMathFlags()); - WidenRecipe->getParent()->insert(FMulRecipe, - WidenRecipe->getIterator()); + Instruction::FMul, + {CurrentLink->getOperand(0), CurrentLink->getOperand(1)}, + CurrentLinkI->getFastMathFlags()); + LinkVPBB->insert(FMulRecipe, CurrentLink->getIterator()); VecOp = FMulRecipe; + } else { + if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { + if (isa<VPWidenRecipe>(CurrentLink)) { + assert(isa<CmpInst>(CurrentLinkI) && + "need to have the compare of the select"); + continue; + } + assert(isa<VPWidenSelectRecipe>(CurrentLink) && + "must be a select recipe"); + IndexOfFirstOperand = 1; + } else { + assert((MinVF.isScalar() || isa<VPWidenRecipe>(CurrentLink)) && + "Expected to replace a VPWidenSC"); + IndexOfFirstOperand = 0; + } + // Note that for non-commutable operands (cmp-selects), the semantics of + // the cmp-select are captured in the recurrence kind. + unsigned VecOpId = + CurrentLink->getOperand(IndexOfFirstOperand) == PreviousLink + ? IndexOfFirstOperand + 1 + : IndexOfFirstOperand; + VecOp = CurrentLink->getOperand(VecOpId); + assert(VecOp != PreviousLink && + CurrentLink->getOperand(CurrentLink->getNumOperands() - 1 - + (VecOpId - IndexOfFirstOperand)) == + PreviousLink && + "PreviousLink must be the operand other than VecOp"); } - VPReductionRecipe *RedRecipe = - new VPReductionRecipe(&RdxDesc, R, ChainOp, VecOp, CondOp, TTI); - WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); - Plan->removeVPValueFor(R); - Plan->addVPValue(R, RedRecipe); - WidenRecipe->getParent()->insert(RedRecipe, WidenRecipe->getIterator()); - WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); - WidenRecipe->eraseFromParent(); - - if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { - VPRecipeBase *CompareRecipe = - RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); - assert(isa<VPWidenRecipe>(CompareRecipe) && - "Expected to replace a VPWidenSC"); - assert(cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 && - "Expected no remaining users"); - CompareRecipe->eraseFromParent(); + + BasicBlock *BB = CurrentLinkI->getParent(); + VPValue *CondOp = nullptr; + if (CM.blockNeedsPredicationForAnyReason(BB)) { + VPBuilder::InsertPointGuard Guard(Builder); + Builder.setInsertPoint(CurrentLink); + CondOp = RecipeBuilder.getBlockInMask(BB); } - Chain = R; + + VPReductionRecipe *RedRecipe = new VPReductionRecipe( + RdxDesc, CurrentLinkI, PreviousLink, VecOp, CondOp); + // Append the recipe to the end of the VPBasicBlock because we need to + // ensure that it comes after all of it's inputs, including CondOp. + // Note that this transformation may leave over dead recipes (including + // CurrentLink), which will be cleaned by a later VPlan transform. + LinkVPBB->appendRecipe(RedRecipe); + CurrentLink->replaceAllUsesWith(RedRecipe); + PreviousLink = RedRecipe; } } + Builder.setInsertPoint(&*LatchVPBB->begin()); + for (VPRecipeBase &R : + Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { + VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); + if (!PhiR) + continue; - // If tail is folded by masking, introduce selects between the phi - // and the live-out instruction of each reduction, at the beginning of the - // dedicated latch block. - if (CM.foldTailByMasking()) { - Builder.setInsertPoint(LatchVPBB, LatchVPBB->begin()); - for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { - VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); - if (!PhiR || PhiR->isInLoop()) - continue; - VPValue *Cond = - RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); - VPValue *Red = PhiR->getBackedgeValue(); - assert(cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB && + const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); + // If tail is folded by masking, introduce selects between the phi + // and the live-out instruction of each reduction, at the beginning of the + // dedicated latch block. + auto *OrigExitingVPV = PhiR->getBackedgeValue(); + auto *NewExitingVPV = PhiR->getBackedgeValue(); + if (!PhiR->isInLoop() && CM.foldTailByMasking()) { + VPValue *Cond = RecipeBuilder.getBlockInMask(OrigLoop->getHeader()); + assert(OrigExitingVPV->getDefiningRecipe()->getParent() != LatchVPBB && "reduction recipe must be defined before latch"); - Builder.createNaryOp(Instruction::Select, {Cond, Red, PhiR}); + Type *PhiTy = PhiR->getOperand(0)->getLiveInIRValue()->getType(); + std::optional<FastMathFlags> FMFs = + PhiTy->isFloatingPointTy() + ? std::make_optional(RdxDesc.getFastMathFlags()) + : std::nullopt; + NewExitingVPV = + Builder.createSelect(Cond, OrigExitingVPV, PhiR, {}, "", FMFs); + OrigExitingVPV->replaceUsesWithIf(NewExitingVPV, [](VPUser &U, unsigned) { + return isa<VPInstruction>(&U) && + cast<VPInstruction>(&U)->getOpcode() == + VPInstruction::ComputeReductionResult; + }); + if (PreferPredicatedReductionSelect || + TTI.preferPredicatedReductionSelect( + PhiR->getRecurrenceDescriptor().getOpcode(), PhiTy, + TargetTransformInfo::ReductionFlags())) + PhiR->setOperand(1, NewExitingVPV); + } + + // If the vector reduction can be performed in a smaller type, we truncate + // then extend the loop exit value to enable InstCombine to evaluate the + // entire expression in the smaller type. + Type *PhiTy = PhiR->getStartValue()->getLiveInIRValue()->getType(); + if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { + assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!"); + Type *RdxTy = RdxDesc.getRecurrenceType(); + auto *Trunc = + new VPWidenCastRecipe(Instruction::Trunc, NewExitingVPV, RdxTy); + auto *Extnd = + RdxDesc.isSigned() + ? new VPWidenCastRecipe(Instruction::SExt, Trunc, PhiTy) + : new VPWidenCastRecipe(Instruction::ZExt, Trunc, PhiTy); + + Trunc->insertAfter(NewExitingVPV->getDefiningRecipe()); + Extnd->insertAfter(Trunc); + if (PhiR->getOperand(1) == NewExitingVPV) + PhiR->setOperand(1, Extnd->getVPSingleValue()); + NewExitingVPV = Extnd; } + + // We want code in the middle block to appear to execute on the location of + // the scalar loop's latch terminator because: (a) it is all compiler + // generated, (b) these instructions are always executed after evaluating + // the latch conditional branch, and (c) other passes may add new + // predecessors which terminate on this line. This is the easiest way to + // ensure we don't accidentally cause an extra step back into the loop while + // debugging. + DebugLoc ExitDL = OrigLoop->getLoopLatch()->getTerminator()->getDebugLoc(); + + // TODO: At the moment ComputeReductionResult also drives creation of the + // bc.merge.rdx phi nodes, hence it needs to be created unconditionally here + // even for in-loop reductions, until the reduction resume value handling is + // also modeled in VPlan. + auto *FinalReductionResult = new VPInstruction( + VPInstruction::ComputeReductionResult, {PhiR, NewExitingVPV}, ExitDL); + cast<VPBasicBlock>(VectorLoopRegion->getSingleSuccessor()) + ->appendRecipe(FinalReductionResult); + OrigExitingVPV->replaceUsesWithIf( + FinalReductionResult, + [](VPUser &User, unsigned) { return isa<VPLiveOut>(&User); }); } + + VPlanTransforms::clearReductionWrapFlags(*Plan); } #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) @@ -9516,433 +9225,243 @@ void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, } #endif -void VPWidenCallRecipe::execute(VPTransformState &State) { - State.ILV->widenCallInstruction(*cast<CallInst>(getUnderlyingInstr()), this, - *this, State); -} - -void VPWidenSelectRecipe::execute(VPTransformState &State) { - auto &I = *cast<SelectInst>(getUnderlyingInstr()); - State.ILV->setDebugLocFromInst(&I); +void VPWidenPointerInductionRecipe::execute(VPTransformState &State) { + assert(IndDesc.getKind() == InductionDescriptor::IK_PtrInduction && + "Not a pointer induction according to InductionDescriptor!"); + assert(cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() && + "Unexpected type."); - // The condition can be loop invariant but still defined inside the - // loop. This means that we can't just use the original 'cond' value. - // We have to take the 'vectorized' value and pick the first lane. - // Instcombine will make this a no-op. - auto *InvarCond = - InvariantCond ? State.get(getOperand(0), VPIteration(0, 0)) : nullptr; - - for (unsigned Part = 0; Part < State.UF; ++Part) { - Value *Cond = InvarCond ? InvarCond : State.get(getOperand(0), Part); - Value *Op0 = State.get(getOperand(1), Part); - Value *Op1 = State.get(getOperand(2), Part); - Value *Sel = State.Builder.CreateSelect(Cond, Op0, Op1); - State.set(this, Sel, Part); - State.ILV->addMetadata(Sel, &I); - } -} - -void VPWidenRecipe::execute(VPTransformState &State) { - auto &I = *cast<Instruction>(getUnderlyingValue()); - auto &Builder = State.Builder; - switch (I.getOpcode()) { - case Instruction::Call: - case Instruction::Br: - case Instruction::PHI: - case Instruction::GetElementPtr: - case Instruction::Select: - llvm_unreachable("This instruction is handled by a different recipe."); - case Instruction::UDiv: - case Instruction::SDiv: - case Instruction::SRem: - case Instruction::URem: - case Instruction::Add: - case Instruction::FAdd: - case Instruction::Sub: - case Instruction::FSub: - case Instruction::FNeg: - case Instruction::Mul: - case Instruction::FMul: - case Instruction::FDiv: - case Instruction::FRem: - case Instruction::Shl: - case Instruction::LShr: - case Instruction::AShr: - case Instruction::And: - case Instruction::Or: - case Instruction::Xor: { - // Just widen unops and binops. - State.ILV->setDebugLocFromInst(&I); + auto *IVR = getParent()->getPlan()->getCanonicalIV(); + PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); - for (unsigned Part = 0; Part < State.UF; ++Part) { - SmallVector<Value *, 2> Ops; - for (VPValue *VPOp : operands()) - Ops.push_back(State.get(VPOp, Part)); - - Value *V = Builder.CreateNAryOp(I.getOpcode(), Ops); - - if (auto *VecOp = dyn_cast<Instruction>(V)) { - VecOp->copyIRFlags(&I); - - // If the instruction is vectorized and was in a basic block that needed - // predication, we can't propagate poison-generating flags (nuw/nsw, - // exact, etc.). The control flow has been linearized and the - // instruction is no longer guarded by the predicate, which could make - // the flag properties to no longer hold. - if (State.MayGeneratePoisonRecipes.contains(this)) - VecOp->dropPoisonGeneratingFlags(); - } + if (onlyScalarsGenerated(State.VF)) { + // This is the normalized GEP that starts counting at zero. + Value *PtrInd = State.Builder.CreateSExtOrTrunc( + CanonicalIV, IndDesc.getStep()->getType()); + // Determine the number of scalars we need to generate for each unroll + // iteration. If the instruction is uniform, we only need to generate the + // first lane. Otherwise, we generate all VF values. + bool IsUniform = vputils::onlyFirstLaneUsed(this); + assert((IsUniform || !State.VF.isScalable()) && + "Cannot scalarize a scalable VF"); + unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); - // Use this vector value for all users of the original instruction. - State.set(this, V, Part); - State.ILV->addMetadata(V, &I); - } - - break; - } - case Instruction::ICmp: - case Instruction::FCmp: { - // Widen compares. Generate vector compares. - bool FCmp = (I.getOpcode() == Instruction::FCmp); - auto *Cmp = cast<CmpInst>(&I); - State.ILV->setDebugLocFromInst(Cmp); for (unsigned Part = 0; Part < State.UF; ++Part) { - Value *A = State.get(getOperand(0), Part); - Value *B = State.get(getOperand(1), Part); - Value *C = nullptr; - if (FCmp) { - // Propagate fast math flags. - IRBuilder<>::FastMathFlagGuard FMFG(Builder); - Builder.setFastMathFlags(Cmp->getFastMathFlags()); - C = Builder.CreateFCmp(Cmp->getPredicate(), A, B); - } else { - C = Builder.CreateICmp(Cmp->getPredicate(), A, B); + Value *PartStart = + createStepForVF(State.Builder, PtrInd->getType(), State.VF, Part); + + for (unsigned Lane = 0; Lane < Lanes; ++Lane) { + Value *Idx = State.Builder.CreateAdd( + PartStart, ConstantInt::get(PtrInd->getType(), Lane)); + Value *GlobalIdx = State.Builder.CreateAdd(PtrInd, Idx); + + Value *Step = State.get(getOperand(1), VPIteration(Part, Lane)); + Value *SclrGep = emitTransformedIndex( + State.Builder, GlobalIdx, IndDesc.getStartValue(), Step, + IndDesc.getKind(), IndDesc.getInductionBinOp()); + SclrGep->setName("next.gep"); + State.set(this, SclrGep, VPIteration(Part, Lane)); } - State.set(this, C, Part); - State.ILV->addMetadata(C, &I); } - - 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: { - auto *CI = cast<CastInst>(&I); - State.ILV->setDebugLocFromInst(CI); - - /// Vectorize casts. - Type *DestTy = (State.VF.isScalar()) - ? CI->getType() - : VectorType::get(CI->getType(), State.VF); - - for (unsigned Part = 0; Part < State.UF; ++Part) { - Value *A = State.get(getOperand(0), Part); - Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy); - State.set(this, Cast, Part); - State.ILV->addMetadata(Cast, &I); - } - break; - } - default: - // This instruction is not vectorized by simple widening. - LLVM_DEBUG(dbgs() << "LV: Found an unhandled instruction: " << I); - llvm_unreachable("Unhandled instruction!"); - } // end of switch. -} + Type *PhiType = IndDesc.getStep()->getType(); + + // Build a pointer phi + Value *ScalarStartValue = getStartValue()->getLiveInIRValue(); + Type *ScStValueType = ScalarStartValue->getType(); + PHINode *NewPointerPhi = + PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); + + BasicBlock *VectorPH = State.CFG.getPreheaderBBFor(this); + NewPointerPhi->addIncoming(ScalarStartValue, VectorPH); + + // A pointer induction, performed by using a gep + Instruction *InductionLoc = &*State.Builder.GetInsertPoint(); + + Value *ScalarStepValue = State.get(getOperand(1), VPIteration(0, 0)); + Value *RuntimeVF = getRuntimeVF(State.Builder, PhiType, State.VF); + Value *NumUnrolledElems = + State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); + Value *InductionGEP = GetElementPtrInst::Create( + State.Builder.getInt8Ty(), NewPointerPhi, + State.Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", + InductionLoc); + // Add induction update using an incorrect block temporarily. The phi node + // will be fixed after VPlan execution. Note that at this point the latch + // block cannot be used, as it does not exist yet. + // TODO: Model increment value in VPlan, by turning the recipe into a + // multi-def and a subclass of VPHeaderPHIRecipe. + NewPointerPhi->addIncoming(InductionGEP, VectorPH); + + // Create UF many actual address geps that use the pointer + // phi as base and a vectorized version of the step value + // (<step*0, ..., step*N>) as offset. + for (unsigned Part = 0; Part < State.UF; ++Part) { + Type *VecPhiType = VectorType::get(PhiType, State.VF); + Value *StartOffsetScalar = + State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); + Value *StartOffset = + State.Builder.CreateVectorSplat(State.VF, StartOffsetScalar); + // Create a vector of consecutive numbers from zero to VF. + StartOffset = State.Builder.CreateAdd( + StartOffset, State.Builder.CreateStepVector(VecPhiType)); + + assert(ScalarStepValue == State.get(getOperand(1), VPIteration(Part, 0)) && + "scalar step must be the same across all parts"); + Value *GEP = State.Builder.CreateGEP( + State.Builder.getInt8Ty(), NewPointerPhi, + State.Builder.CreateMul( + StartOffset, + State.Builder.CreateVectorSplat(State.VF, ScalarStepValue), + "vector.gep")); + State.set(this, GEP, Part); + } +} + +void VPDerivedIVRecipe::execute(VPTransformState &State) { + assert(!State.Instance && "VPDerivedIVRecipe being replicated."); -void VPWidenGEPRecipe::execute(VPTransformState &State) { - auto *GEP = cast<GetElementPtrInst>(getUnderlyingInstr()); - // Construct a vector GEP by widening the operands of the scalar GEP as - // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP - // results in a vector of pointers when at least one operand of the GEP - // is vector-typed. Thus, to keep the representation compact, we only use - // vector-typed operands for loop-varying values. - - if (State.VF.isVector() && IsPtrLoopInvariant && IsIndexLoopInvariant.all()) { - // If we are vectorizing, but the GEP has only loop-invariant operands, - // the GEP we build (by only using vector-typed operands for - // loop-varying values) would be a scalar pointer. Thus, to ensure we - // produce a vector of pointers, we need to either arbitrarily pick an - // operand to broadcast, or broadcast a clone of the original GEP. - // Here, we broadcast a clone of the original. - // - // TODO: If at some point we decide to scalarize instructions having - // loop-invariant operands, this special case will no longer be - // required. We would add the scalarization decision to - // collectLoopScalars() and teach getVectorValue() to broadcast - // the lane-zero scalar value. - auto *Clone = State.Builder.Insert(GEP->clone()); - for (unsigned Part = 0; Part < State.UF; ++Part) { - Value *EntryPart = State.Builder.CreateVectorSplat(State.VF, Clone); - State.set(this, EntryPart, Part); - State.ILV->addMetadata(EntryPart, GEP); - } - } else { - // If the GEP has at least one loop-varying operand, we are sure to - // produce a vector of pointers. But if we are only unrolling, we want - // to produce a scalar GEP for each unroll part. Thus, the GEP we - // produce with the code below will be scalar (if VF == 1) or vector - // (otherwise). Note that for the unroll-only case, we still maintain - // values in the vector mapping with initVector, as we do for other - // instructions. - for (unsigned Part = 0; Part < State.UF; ++Part) { - // The pointer operand of the new GEP. If it's loop-invariant, we - // won't broadcast it. - auto *Ptr = IsPtrLoopInvariant - ? State.get(getOperand(0), VPIteration(0, 0)) - : State.get(getOperand(0), Part); - - // Collect all the indices for the new GEP. If any index is - // loop-invariant, we won't broadcast it. - SmallVector<Value *, 4> Indices; - for (unsigned I = 1, E = getNumOperands(); I < E; I++) { - VPValue *Operand = getOperand(I); - if (IsIndexLoopInvariant[I - 1]) - Indices.push_back(State.get(Operand, VPIteration(0, 0))); - else - Indices.push_back(State.get(Operand, Part)); - } + // Fast-math-flags propagate from the original induction instruction. + IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); + if (FPBinOp) + State.Builder.setFastMathFlags(FPBinOp->getFastMathFlags()); - // If the GEP instruction is vectorized and was in a basic block that - // needed predication, we can't propagate the poison-generating 'inbounds' - // flag. The control flow has been linearized and the GEP is no longer - // guarded by the predicate, which could make the 'inbounds' properties to - // no longer hold. - bool IsInBounds = - GEP->isInBounds() && State.MayGeneratePoisonRecipes.count(this) == 0; - - // Create the new GEP. Note that this GEP may be a scalar if VF == 1, - // but it should be a vector, otherwise. - auto *NewGEP = IsInBounds - ? State.Builder.CreateInBoundsGEP( - GEP->getSourceElementType(), Ptr, Indices) - : State.Builder.CreateGEP(GEP->getSourceElementType(), - Ptr, Indices); - assert((State.VF.isScalar() || NewGEP->getType()->isVectorTy()) && - "NewGEP is not a pointer vector"); - State.set(this, NewGEP, Part); - State.ILV->addMetadata(NewGEP, GEP); - } + Value *Step = State.get(getStepValue(), VPIteration(0, 0)); + Value *CanonicalIV = State.get(getCanonicalIV(), VPIteration(0, 0)); + Value *DerivedIV = emitTransformedIndex( + State.Builder, CanonicalIV, getStartValue()->getLiveInIRValue(), Step, + Kind, cast_if_present<BinaryOperator>(FPBinOp)); + DerivedIV->setName("offset.idx"); + if (TruncResultTy) { + assert(TruncResultTy != DerivedIV->getType() && + Step->getType()->isIntegerTy() && + "Truncation requires an integer step"); + DerivedIV = State.Builder.CreateTrunc(DerivedIV, TruncResultTy); } -} - -void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { - assert(!State.Instance && "Int or FP induction being replicated."); - auto *CanonicalIV = State.get(getParent()->getPlan()->getCanonicalIV(), 0); - State.ILV->widenIntOrFpInduction(IV, this, State, CanonicalIV); -} + assert(DerivedIV != CanonicalIV && "IV didn't need transforming?"); -void VPWidenPHIRecipe::execute(VPTransformState &State) { - State.ILV->widenPHIInstruction(cast<PHINode>(getUnderlyingValue()), this, - State); -} - -void VPBlendRecipe::execute(VPTransformState &State) { - State.ILV->setDebugLocFromInst(Phi, &State.Builder); - // We know that all PHIs in non-header blocks are converted into - // selects, so we don't have to worry about the insertion order and we - // can just use the builder. - // At this point we generate the predication tree. There may be - // duplications since this is a simple recursive scan, but future - // optimizations will clean it up. - - unsigned NumIncoming = getNumIncomingValues(); - - // Generate a sequence of selects of the form: - // SELECT(Mask3, In3, - // SELECT(Mask2, In2, - // SELECT(Mask1, In1, - // In0))) - // Note that Mask0 is never used: lanes for which no path reaches this phi and - // are essentially undef are taken from In0. - InnerLoopVectorizer::VectorParts Entry(State.UF); - for (unsigned In = 0; In < NumIncoming; ++In) { - for (unsigned Part = 0; Part < State.UF; ++Part) { - // We might have single edge PHIs (blocks) - use an identity - // 'select' for the first PHI operand. - Value *In0 = State.get(getIncomingValue(In), Part); - if (In == 0) - Entry[Part] = In0; // Initialize with the first incoming value. - else { - // Select between the current value and the previous incoming edge - // based on the incoming mask. - Value *Cond = State.get(getMask(In), Part); - Entry[Part] = - State.Builder.CreateSelect(Cond, In0, Entry[Part], "predphi"); - } - } - } - for (unsigned Part = 0; Part < State.UF; ++Part) - State.set(this, Entry[Part], Part); + State.set(this, DerivedIV, VPIteration(0, 0)); } void VPInterleaveRecipe::execute(VPTransformState &State) { assert(!State.Instance && "Interleave group being replicated."); State.ILV->vectorizeInterleaveGroup(IG, definedValues(), State, getAddr(), - getStoredValues(), getMask()); + getStoredValues(), getMask(), + NeedsMaskForGaps); } void VPReductionRecipe::execute(VPTransformState &State) { assert(!State.Instance && "Reduction being replicated."); Value *PrevInChain = State.get(getChainOp(), 0); - RecurKind Kind = RdxDesc->getRecurrenceKind(); - bool IsOrdered = State.ILV->useOrderedReductions(*RdxDesc); + RecurKind Kind = RdxDesc.getRecurrenceKind(); + bool IsOrdered = State.ILV->useOrderedReductions(RdxDesc); // Propagate the fast-math flags carried by the underlying instruction. IRBuilderBase::FastMathFlagGuard FMFGuard(State.Builder); - State.Builder.setFastMathFlags(RdxDesc->getFastMathFlags()); + State.Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); for (unsigned Part = 0; Part < State.UF; ++Part) { Value *NewVecOp = State.get(getVecOp(), Part); if (VPValue *Cond = getCondOp()) { - Value *NewCond = State.get(Cond, Part); - VectorType *VecTy = cast<VectorType>(NewVecOp->getType()); - Value *Iden = RdxDesc->getRecurrenceIdentity( - Kind, VecTy->getElementType(), RdxDesc->getFastMathFlags()); - Value *IdenVec = - State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); - Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, IdenVec); + Value *NewCond = State.VF.isVector() ? State.get(Cond, Part) + : State.get(Cond, {Part, 0}); + VectorType *VecTy = dyn_cast<VectorType>(NewVecOp->getType()); + Type *ElementTy = VecTy ? VecTy->getElementType() : NewVecOp->getType(); + Value *Iden = RdxDesc.getRecurrenceIdentity(Kind, ElementTy, + RdxDesc.getFastMathFlags()); + if (State.VF.isVector()) { + Iden = + State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); + } + + Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, Iden); NewVecOp = Select; } Value *NewRed; Value *NextInChain; if (IsOrdered) { if (State.VF.isVector()) - NewRed = createOrderedReduction(State.Builder, *RdxDesc, NewVecOp, + NewRed = createOrderedReduction(State.Builder, RdxDesc, NewVecOp, PrevInChain); else NewRed = State.Builder.CreateBinOp( - (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), PrevInChain, + (Instruction::BinaryOps)RdxDesc.getOpcode(Kind), PrevInChain, NewVecOp); PrevInChain = NewRed; } else { PrevInChain = State.get(getChainOp(), Part); - NewRed = createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp); + NewRed = createTargetReduction(State.Builder, RdxDesc, NewVecOp); } if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { - NextInChain = - createMinMaxOp(State.Builder, RdxDesc->getRecurrenceKind(), - NewRed, PrevInChain); + NextInChain = createMinMaxOp(State.Builder, RdxDesc.getRecurrenceKind(), + NewRed, PrevInChain); } else if (IsOrdered) NextInChain = NewRed; else NextInChain = State.Builder.CreateBinOp( - (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), NewRed, - PrevInChain); + (Instruction::BinaryOps)RdxDesc.getOpcode(Kind), NewRed, PrevInChain); State.set(this, NextInChain, Part); } } void VPReplicateRecipe::execute(VPTransformState &State) { + Instruction *UI = getUnderlyingInstr(); if (State.Instance) { // Generate a single instance. assert(!State.VF.isScalable() && "Can't scalarize a scalable vector"); - State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, *State.Instance, - IsPredicated, State); + State.ILV->scalarizeInstruction(UI, this, *State.Instance, State); // Insert scalar instance packing it into a vector. - if (AlsoPack && State.VF.isVector()) { + if (State.VF.isVector() && shouldPack()) { // If we're constructing lane 0, initialize to start from poison. if (State.Instance->Lane.isFirstLane()) { assert(!State.VF.isScalable() && "VF is assumed to be non scalable."); Value *Poison = PoisonValue::get( - VectorType::get(getUnderlyingValue()->getType(), State.VF)); + VectorType::get(UI->getType(), State.VF)); State.set(this, Poison, State.Instance->Part); } - State.ILV->packScalarIntoVectorValue(this, *State.Instance, State); + State.packScalarIntoVectorValue(this, *State.Instance); } return; } - // Generate scalar instances for all VF lanes of all UF parts, unless the - // instruction is uniform inwhich case generate only the first lane for each - // of the UF parts. - unsigned EndLane = IsUniform ? 1 : State.VF.getKnownMinValue(); - assert((!State.VF.isScalable() || IsUniform) && - "Can't scalarize a scalable vector"); + if (IsUniform) { + // If the recipe is uniform across all parts (instead of just per VF), only + // generate a single instance. + if ((isa<LoadInst>(UI) || isa<StoreInst>(UI)) && + all_of(operands(), [](VPValue *Op) { + return Op->isDefinedOutsideVectorRegions(); + })) { + State.ILV->scalarizeInstruction(UI, this, VPIteration(0, 0), State); + if (user_begin() != user_end()) { + for (unsigned Part = 1; Part < State.UF; ++Part) + State.set(this, State.get(this, VPIteration(0, 0)), + VPIteration(Part, 0)); + } + return; + } + + // Uniform within VL means we need to generate lane 0 only for each + // unrolled copy. + for (unsigned Part = 0; Part < State.UF; ++Part) + State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, 0), State); + return; + } + + // A store of a loop varying value to a uniform address only needs the last + // copy of the store. + if (isa<StoreInst>(UI) && + vputils::isUniformAfterVectorization(getOperand(1))) { + auto Lane = VPLane::getLastLaneForVF(State.VF); + State.ILV->scalarizeInstruction(UI, this, VPIteration(State.UF - 1, Lane), + State); + return; + } + + // Generate scalar instances for all VF lanes of all UF parts. + assert(!State.VF.isScalable() && "Can't scalarize a scalable vector"); + const unsigned EndLane = State.VF.getKnownMinValue(); for (unsigned Part = 0; Part < State.UF; ++Part) for (unsigned Lane = 0; Lane < EndLane; ++Lane) - State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, - VPIteration(Part, Lane), IsPredicated, - State); -} - -void VPBranchOnMaskRecipe::execute(VPTransformState &State) { - assert(State.Instance && "Branch on Mask works only on single instance."); - - unsigned Part = State.Instance->Part; - unsigned Lane = State.Instance->Lane.getKnownLane(); - - Value *ConditionBit = nullptr; - VPValue *BlockInMask = getMask(); - if (BlockInMask) { - ConditionBit = State.get(BlockInMask, Part); - if (ConditionBit->getType()->isVectorTy()) - ConditionBit = State.Builder.CreateExtractElement( - ConditionBit, State.Builder.getInt32(Lane)); - } else // Block in mask is all-one. - ConditionBit = State.Builder.getTrue(); - - // Replace the temporary unreachable terminator with a new conditional branch, - // whose two destinations will be set later when they are created. - auto *CurrentTerminator = State.CFG.PrevBB->getTerminator(); - assert(isa<UnreachableInst>(CurrentTerminator) && - "Expected to replace unreachable terminator with conditional branch."); - auto *CondBr = BranchInst::Create(State.CFG.PrevBB, nullptr, ConditionBit); - CondBr->setSuccessor(0, nullptr); - ReplaceInstWithInst(CurrentTerminator, CondBr); -} - -void VPPredInstPHIRecipe::execute(VPTransformState &State) { - assert(State.Instance && "Predicated instruction PHI works per instance."); - Instruction *ScalarPredInst = - cast<Instruction>(State.get(getOperand(0), *State.Instance)); - BasicBlock *PredicatedBB = ScalarPredInst->getParent(); - BasicBlock *PredicatingBB = PredicatedBB->getSinglePredecessor(); - assert(PredicatingBB && "Predicated block has no single predecessor."); - assert(isa<VPReplicateRecipe>(getOperand(0)) && - "operand must be VPReplicateRecipe"); - - // By current pack/unpack logic we need to generate only a single phi node: if - // a vector value for the predicated instruction exists at this point it means - // the instruction has vector users only, and a phi for the vector value is - // needed. In this case the recipe of the predicated instruction is marked to - // also do that packing, thereby "hoisting" the insert-element sequence. - // Otherwise, a phi node for the scalar value is needed. - unsigned Part = State.Instance->Part; - if (State.hasVectorValue(getOperand(0), Part)) { - Value *VectorValue = State.get(getOperand(0), Part); - InsertElementInst *IEI = cast<InsertElementInst>(VectorValue); - PHINode *VPhi = State.Builder.CreatePHI(IEI->getType(), 2); - VPhi->addIncoming(IEI->getOperand(0), PredicatingBB); // Unmodified vector. - VPhi->addIncoming(IEI, PredicatedBB); // New vector with inserted element. - if (State.hasVectorValue(this, Part)) - State.reset(this, VPhi, Part); - else - State.set(this, VPhi, Part); - // NOTE: Currently we need to update the value of the operand, so the next - // predicated iteration inserts its generated value in the correct vector. - State.reset(getOperand(0), VPhi, Part); - } else { - Type *PredInstType = getOperand(0)->getUnderlyingValue()->getType(); - PHINode *Phi = State.Builder.CreatePHI(PredInstType, 2); - Phi->addIncoming(PoisonValue::get(ScalarPredInst->getType()), - PredicatingBB); - Phi->addIncoming(ScalarPredInst, PredicatedBB); - if (State.hasScalarValue(this, *State.Instance)) - State.reset(this, Phi, *State.Instance); - else - State.set(this, Phi, *State.Instance); - // NOTE: Currently we need to update the value of the operand, so the next - // predicated iteration inserts its generated value in the correct vector. - State.reset(getOperand(0), Phi, *State.Instance); - } + State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, Lane), State); } void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { @@ -9960,56 +9479,25 @@ void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { auto *DataTy = VectorType::get(ScalarDataTy, State.VF); const Align Alignment = getLoadStoreAlignment(&Ingredient); - bool CreateGatherScatter = !Consecutive; + bool CreateGatherScatter = !isConsecutive(); auto &Builder = State.Builder; InnerLoopVectorizer::VectorParts BlockInMaskParts(State.UF); bool isMaskRequired = getMask(); - if (isMaskRequired) - for (unsigned Part = 0; Part < State.UF; ++Part) - BlockInMaskParts[Part] = State.get(getMask(), Part); - - const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { - // Calculate the pointer for the specific unroll-part. - GetElementPtrInst *PartPtr = nullptr; - - bool InBounds = false; - if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) - InBounds = gep->isInBounds(); - if (Reverse) { - // If the address is consecutive but reversed, then the - // wide store needs to start at the last vector element. - // RunTimeVF = VScale * VF.getKnownMinValue() - // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue() - Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), State.VF); - // NumElt = -Part * RunTimeVF - Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF); - // LastLane = 1 - RunTimeVF - Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF); - PartPtr = - cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt)); - PartPtr->setIsInBounds(InBounds); - PartPtr = cast<GetElementPtrInst>( - Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane)); - PartPtr->setIsInBounds(InBounds); - if (isMaskRequired) // Reverse of a null all-one mask is a null mask. - BlockInMaskParts[Part] = - Builder.CreateVectorReverse(BlockInMaskParts[Part], "reverse"); - } else { - Value *Increment = - createStepForVF(Builder, Builder.getInt32Ty(), State.VF, Part); - PartPtr = cast<GetElementPtrInst>( - Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); - PartPtr->setIsInBounds(InBounds); + if (isMaskRequired) { + // Mask reversal is only needed for non-all-one (null) masks, as reverse of + // a null all-one mask is a null mask. + for (unsigned Part = 0; Part < State.UF; ++Part) { + Value *Mask = State.get(getMask(), Part); + if (isReverse()) + Mask = Builder.CreateVectorReverse(Mask, "reverse"); + BlockInMaskParts[Part] = Mask; } - - unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); - return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); - }; + } // Handle Stores: if (SI) { - State.ILV->setDebugLocFromInst(SI); + State.setDebugLocFrom(SI->getDebugLoc()); for (unsigned Part = 0; Part < State.UF; ++Part) { Instruction *NewSI = nullptr; @@ -10020,29 +9508,28 @@ void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, MaskPart); } else { - if (Reverse) { + if (isReverse()) { // If we store to reverse consecutive memory locations, then we need // to reverse the order of elements in the stored value. StoredVal = Builder.CreateVectorReverse(StoredVal, "reverse"); // We don't want to update the value in the map as it might be used in // another expression. So don't call resetVectorValue(StoredVal). } - auto *VecPtr = - CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); + auto *VecPtr = State.get(getAddr(), Part); if (isMaskRequired) NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, BlockInMaskParts[Part]); else NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); } - State.ILV->addMetadata(NewSI, SI); + State.addMetadata(NewSI, SI); } return; } // Handle loads. assert(LI && "Must have a load instruction"); - State.ILV->setDebugLocFromInst(LI); + State.setDebugLocFrom(LI->getDebugLoc()); for (unsigned Part = 0; Part < State.UF; ++Part) { Value *NewLI; if (CreateGatherScatter) { @@ -10050,10 +9537,9 @@ void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { Value *VectorGep = State.get(getAddr(), Part); NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart, nullptr, "wide.masked.gather"); - State.ILV->addMetadata(NewLI, LI); + State.addMetadata(NewLI, LI); } else { - auto *VecPtr = - CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); + auto *VecPtr = State.get(getAddr(), Part); if (isMaskRequired) NewLI = Builder.CreateMaskedLoad( DataTy, VecPtr, Alignment, BlockInMaskParts[Part], @@ -10063,12 +9549,12 @@ void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); // Add metadata to the load, but setVectorValue to the reverse shuffle. - State.ILV->addMetadata(NewLI, LI); + State.addMetadata(NewLI, LI); if (Reverse) NewLI = Builder.CreateVectorReverse(NewLI, "reverse"); } - State.set(this, NewLI, Part); + State.set(getVPSingleValue(), NewLI, Part); } } @@ -10079,8 +9565,7 @@ void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { static ScalarEpilogueLowering getScalarEpilogueLowering( Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, - AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, - LoopVectorizationLegality &LVL) { + LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) { // 1) OptSize takes precedence over all other options, i.e. if this is set, // don't look at hints or options, and don't request a scalar epilogue. // (For PGSO, as shouldOptimizeForSize isn't currently accessible from @@ -10115,80 +9600,13 @@ static ScalarEpilogueLowering getScalarEpilogueLowering( }; // 4) if the TTI hook indicates this is profitable, request predication. - if (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, - LVL.getLAI())) + TailFoldingInfo TFI(TLI, &LVL, IAI); + if (TTI->preferPredicateOverEpilogue(&TFI)) return CM_ScalarEpilogueNotNeededUsePredicate; return CM_ScalarEpilogueAllowed; } -Value *VPTransformState::get(VPValue *Def, unsigned Part) { - // If Values have been set for this Def return the one relevant for \p Part. - if (hasVectorValue(Def, Part)) - return Data.PerPartOutput[Def][Part]; - - if (!hasScalarValue(Def, {Part, 0})) { - Value *IRV = Def->getLiveInIRValue(); - Value *B = ILV->getBroadcastInstrs(IRV); - set(Def, B, Part); - return B; - } - - Value *ScalarValue = get(Def, {Part, 0}); - // If we aren't vectorizing, we can just copy the scalar map values over - // to the vector map. - if (VF.isScalar()) { - set(Def, ScalarValue, Part); - return ScalarValue; - } - - auto *RepR = dyn_cast<VPReplicateRecipe>(Def); - bool IsUniform = RepR && RepR->isUniform(); - - unsigned LastLane = IsUniform ? 0 : VF.getKnownMinValue() - 1; - // Check if there is a scalar value for the selected lane. - if (!hasScalarValue(Def, {Part, LastLane})) { - // At the moment, VPWidenIntOrFpInductionRecipes can also be uniform. - assert(isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) && - "unexpected recipe found to be invariant"); - IsUniform = true; - LastLane = 0; - } - - auto *LastInst = cast<Instruction>(get(Def, {Part, LastLane})); - // Set the insert point after the last scalarized instruction or after the - // last PHI, if LastInst is a PHI. This ensures the insertelement sequence - // will directly follow the scalar definitions. - auto OldIP = Builder.saveIP(); - auto NewIP = - isa<PHINode>(LastInst) - ? BasicBlock::iterator(LastInst->getParent()->getFirstNonPHI()) - : std::next(BasicBlock::iterator(LastInst)); - Builder.SetInsertPoint(&*NewIP); - - // However, if we are vectorizing, we need to construct the vector values. - // If the value is known to be uniform after vectorization, we can just - // broadcast the scalar value corresponding to lane zero for each unroll - // iteration. Otherwise, we construct the vector values using - // insertelement instructions. Since the resulting vectors are stored in - // State, we will only generate the insertelements once. - Value *VectorValue = nullptr; - if (IsUniform) { - VectorValue = ILV->getBroadcastInstrs(ScalarValue); - set(Def, VectorValue, Part); - } else { - // Initialize packing with insertelements to start from undef. - assert(!VF.isScalable() && "VF is assumed to be non scalable."); - Value *Undef = PoisonValue::get(VectorType::get(LastInst->getType(), VF)); - set(Def, Undef, Part); - for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) - ILV->packScalarIntoVectorValue(Def, {Part, Lane}, *this); - VectorValue = get(Def, Part); - } - Builder.restoreIP(OldIP); - return VectorValue; -} - // Process the loop in the VPlan-native vectorization path. This path builds // VPlan upfront in the vectorization pipeline, which allows to apply // VPlan-to-VPlan transformations from the very beginning without modifying the @@ -10209,16 +9627,16 @@ static bool processLoopInVPlanNativePath( Function *F = L->getHeader()->getParent(); InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); - ScalarEpilogueLowering SEL = getScalarEpilogueLowering( - F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL); + ScalarEpilogueLowering SEL = + getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, *LVL, &IAI); LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, &Hints, IAI); // Use the planner for outer loop vectorization. // TODO: CM is not used at this point inside the planner. Turn CM into an // optional argument if we don't need it in the future. - LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE, Hints, - Requirements, ORE); + LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, LVL, CM, IAI, PSE, Hints, + ORE); // Get user vectorization factor. ElementCount UserVF = Hints.getWidth(); @@ -10231,22 +9649,25 @@ static bool processLoopInVPlanNativePath( // If we are stress testing VPlan builds, do not attempt to generate vector // code. Masked vector code generation support will follow soon. // Also, do not attempt to vectorize if no vector code will be produced. - if (VPlanBuildStressTest || EnableVPlanPredication || - VectorizationFactor::Disabled() == VF) + if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF) return false; VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); { - GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, - F->getParent()->getDataLayout()); - InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL, - &CM, BFI, PSI, Checks); + bool AddBranchWeights = + hasBranchWeightMD(*L->getLoopLatch()->getTerminator()); + GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, + F->getParent()->getDataLayout(), AddBranchWeights); + InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, + VF.Width, 1, LVL, &CM, BFI, PSI, Checks); LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"); - LVP.executePlan(VF.Width, 1, BestPlan, LB, DT); + LVP.executePlan(VF.Width, 1, BestPlan, LB, DT, false); } + reportVectorization(ORE, L, VF, 1); + // Mark the loop as already vectorized to avoid vectorizing again. Hints.setAlreadyVectorized(); assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs())); @@ -10298,6 +9719,108 @@ static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { } } +static bool areRuntimeChecksProfitable(GeneratedRTChecks &Checks, + VectorizationFactor &VF, + std::optional<unsigned> VScale, Loop *L, + ScalarEvolution &SE, + ScalarEpilogueLowering SEL) { + InstructionCost CheckCost = Checks.getCost(); + if (!CheckCost.isValid()) + return false; + + // When interleaving only scalar and vector cost will be equal, which in turn + // would lead to a divide by 0. Fall back to hard threshold. + if (VF.Width.isScalar()) { + if (CheckCost > VectorizeMemoryCheckThreshold) { + LLVM_DEBUG( + dbgs() + << "LV: Interleaving only is not profitable due to runtime checks\n"); + return false; + } + return true; + } + + // The scalar cost should only be 0 when vectorizing with a user specified VF/IC. In those cases, runtime checks should always be generated. + double ScalarC = *VF.ScalarCost.getValue(); + if (ScalarC == 0) + return true; + + // First, compute the minimum iteration count required so that the vector + // loop outperforms the scalar loop. + // The total cost of the scalar loop is + // ScalarC * TC + // where + // * TC is the actual trip count of the loop. + // * ScalarC is the cost of a single scalar iteration. + // + // The total cost of the vector loop is + // RtC + VecC * (TC / VF) + EpiC + // where + // * RtC is the cost of the generated runtime checks + // * VecC is the cost of a single vector iteration. + // * TC is the actual trip count of the loop + // * VF is the vectorization factor + // * EpiCost is the cost of the generated epilogue, including the cost + // of the remaining scalar operations. + // + // Vectorization is profitable once the total vector cost is less than the + // total scalar cost: + // RtC + VecC * (TC / VF) + EpiC < ScalarC * TC + // + // Now we can compute the minimum required trip count TC as + // (RtC + EpiC) / (ScalarC - (VecC / VF)) < TC + // + // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that + // the computations are performed on doubles, not integers and the result + // is rounded up, hence we get an upper estimate of the TC. + unsigned IntVF = VF.Width.getKnownMinValue(); + if (VF.Width.isScalable()) { + unsigned AssumedMinimumVscale = 1; + if (VScale) + AssumedMinimumVscale = *VScale; + IntVF *= AssumedMinimumVscale; + } + double VecCOverVF = double(*VF.Cost.getValue()) / IntVF; + double RtC = *CheckCost.getValue(); + double MinTC1 = RtC / (ScalarC - VecCOverVF); + + // Second, compute a minimum iteration count so that the cost of the + // runtime checks is only a fraction of the total scalar loop cost. This + // adds a loop-dependent bound on the overhead incurred if the runtime + // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC + // * TC. To bound the runtime check to be a fraction 1/X of the scalar + // cost, compute + // RtC < ScalarC * TC * (1 / X) ==> RtC * X / ScalarC < TC + double MinTC2 = RtC * 10 / ScalarC; + + // Now pick the larger minimum. If it is not a multiple of VF and a scalar + // epilogue is allowed, choose the next closest multiple of VF. This should + // partly compensate for ignoring the epilogue cost. + uint64_t MinTC = std::ceil(std::max(MinTC1, MinTC2)); + if (SEL == CM_ScalarEpilogueAllowed) + MinTC = alignTo(MinTC, IntVF); + VF.MinProfitableTripCount = ElementCount::getFixed(MinTC); + + LLVM_DEBUG( + dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" + << VF.MinProfitableTripCount << "\n"); + + // Skip vectorization if the expected trip count is less than the minimum + // required trip count. + if (auto ExpectedTC = getSmallBestKnownTC(SE, L)) { + if (ElementCount::isKnownLT(ElementCount::getFixed(*ExpectedTC), + VF.MinProfitableTripCount)) { + LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected " + "trip count < minimum profitable VF (" + << *ExpectedTC << " < " << VF.MinProfitableTripCount + << ")\n"); + + return false; + } + } + return true; +} + LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || !EnableLoopInterleaving), @@ -10312,8 +9835,8 @@ bool LoopVectorizePass::processLoop(Loop *L) { const std::string DebugLocStr = getDebugLocString(L); #endif /* NDEBUG */ - LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in \"" - << L->getHeader()->getParent()->getName() << "\" from " + LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '" + << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"); LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); @@ -10349,7 +9872,7 @@ bool LoopVectorizePass::processLoop(Loop *L) { // Check if it is legal to vectorize the loop. LoopVectorizationRequirements Requirements; - LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, AA, F, GetLAA, LI, ORE, + LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE, &Requirements, &Hints, DB, AC, BFI, PSI); if (!LVL.canVectorize(EnableVPlanNativePath)) { LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n"); @@ -10357,11 +9880,6 @@ bool LoopVectorizePass::processLoop(Loop *L) { return false; } - // Check the function attributes and profiles to find out if this function - // should be optimized for size. - ScalarEpilogueLowering SEL = getScalarEpilogueLowering( - F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL); - // Entrance to the VPlan-native vectorization path. Outer loops are processed // here. They may require CFG and instruction level transformations before // even evaluating whether vectorization is profitable. Since we cannot modify @@ -10373,6 +9891,22 @@ bool LoopVectorizePass::processLoop(Loop *L) { assert(L->isInnermost() && "Inner loop expected."); + InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); + bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); + + // If an override option has been passed in for interleaved accesses, use it. + if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) + UseInterleaved = EnableInterleavedMemAccesses; + + // Analyze interleaved memory accesses. + if (UseInterleaved) + IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); + + // Check the function attributes and profiles to find out if this function + // should be optimized for size. + ScalarEpilogueLowering SEL = + getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, LVL, &IAI); + // Check the loop for a trip count threshold: vectorize loops with a tiny trip // count by optimizing for size, to minimize overheads. auto ExpectedTC = getSmallBestKnownTC(*SE, L); @@ -10383,15 +9917,31 @@ bool LoopVectorizePass::processLoop(Loop *L) { if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n"); else { - LLVM_DEBUG(dbgs() << "\n"); - SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; + if (*ExpectedTC > TTI->getMinTripCountTailFoldingThreshold()) { + LLVM_DEBUG(dbgs() << "\n"); + // Predicate tail-folded loops are efficient even when the loop + // iteration count is low. However, setting the epilogue policy to + // `CM_ScalarEpilogueNotAllowedLowTripLoop` prevents vectorizing loops + // with runtime checks. It's more effective to let + // `areRuntimeChecksProfitable` determine if vectorization is beneficial + // for the loop. + if (SEL != CM_ScalarEpilogueNotNeededUsePredicate) + SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; + } else { + LLVM_DEBUG(dbgs() << " But the target considers the trip count too " + "small to consider vectorizing.\n"); + reportVectorizationFailure( + "The trip count is below the minial threshold value.", + "loop trip count is too low, avoiding vectorization", + "LowTripCount", ORE, L); + Hints.emitRemarkWithHints(); + return false; + } } } - // Check the function attributes to see if implicit floats are allowed. - // FIXME: This check doesn't seem possibly correct -- what if the loop is - // an integer loop and the vector instructions selected are purely integer - // vector instructions? + // Check the function attributes to see if implicit floats or vectors are + // allowed. if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { reportVectorizationFailure( "Can't vectorize when the NoImplicitFloat attribute is used", @@ -10436,42 +9986,55 @@ bool LoopVectorizePass::processLoop(Loop *L) { return false; } - bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); - InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); - - // If an override option has been passed in for interleaved accesses, use it. - if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) - UseInterleaved = EnableInterleavedMemAccesses; - - // Analyze interleaved memory accesses. - if (UseInterleaved) { - IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); - } - // Use the cost model. LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, F, &Hints, IAI); - CM.collectValuesToIgnore(); - CM.collectElementTypesForWidening(); - // Use the planner for vectorization. - LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE, Hints, - Requirements, ORE); + LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, &LVL, CM, IAI, PSE, Hints, + ORE); // Get user vectorization factor and interleave count. ElementCount UserVF = Hints.getWidth(); unsigned UserIC = Hints.getInterleave(); // Plan how to best vectorize, return the best VF and its cost. - Optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); + std::optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); VectorizationFactor VF = VectorizationFactor::Disabled(); unsigned IC = 1; + bool AddBranchWeights = + hasBranchWeightMD(*L->getLoopLatch()->getTerminator()); + GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, + F->getParent()->getDataLayout(), AddBranchWeights); if (MaybeVF) { VF = *MaybeVF; // Select the interleave count. - IC = CM.selectInterleaveCount(VF.Width, *VF.Cost.getValue()); + IC = CM.selectInterleaveCount(VF.Width, VF.Cost); + + unsigned SelectedIC = std::max(IC, UserIC); + // Optimistically generate runtime checks if they are needed. Drop them if + // they turn out to not be profitable. + if (VF.Width.isVector() || SelectedIC > 1) + Checks.Create(L, *LVL.getLAI(), PSE.getPredicate(), VF.Width, SelectedIC); + + // Check if it is profitable to vectorize with runtime checks. + bool ForceVectorization = + Hints.getForce() == LoopVectorizeHints::FK_Enabled; + if (!ForceVectorization && + !areRuntimeChecksProfitable(Checks, VF, getVScaleForTuning(L, *TTI), L, + *PSE.getSE(), SEL)) { + ORE->emit([&]() { + return OptimizationRemarkAnalysisAliasing( + DEBUG_TYPE, "CantReorderMemOps", L->getStartLoc(), + L->getHeader()) + << "loop not vectorized: cannot prove it is safe to reorder " + "memory operations"; + }); + LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n"); + Hints.emitRemarkWithHints(); + return false; + } } // Identify the diagnostic messages that should be produced. @@ -10559,14 +10122,6 @@ bool LoopVectorizePass::processLoop(Loop *L) { bool DisableRuntimeUnroll = false; MDNode *OrigLoopID = L->getLoopID(); { - // Optimistically generate runtime checks. Drop them if they turn out to not - // be profitable. Limit the scope of Checks, so the cleanup happens - // immediately after vector codegeneration is done. - GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, - F->getParent()->getDataLayout()); - if (!VF.Width.isScalar() || IC > 1) - Checks.Create(L, *LVL.getLAI(), PSE.getUnionPredicate()); - using namespace ore; if (!VectorizeLoop) { assert(IC > 1 && "interleave count should not be 1 or 0"); @@ -10576,7 +10131,7 @@ bool LoopVectorizePass::processLoop(Loop *L) { &CM, BFI, PSI, Checks); VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); - LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT); + LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT, false); ORE->emit([&]() { return OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(), @@ -10589,7 +10144,7 @@ bool LoopVectorizePass::processLoop(Loop *L) { // Consider vectorizing the epilogue too if it's profitable. VectorizationFactor EpilogueVF = - CM.selectEpilogueVectorizationFactor(VF.Width, LVP); + LVP.selectEpilogueVectorizationFactor(VF.Width, IC); if (EpilogueVF.Width.isVector()) { // The first pass vectorizes the main loop and creates a scalar epilogue @@ -10600,13 +10155,10 @@ bool LoopVectorizePass::processLoop(Loop *L) { EPI, &LVL, &CM, BFI, PSI, Checks); VPlan &BestMainPlan = LVP.getBestPlanFor(EPI.MainLoopVF); - LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, - DT); + const auto &[ExpandedSCEVs, ReductionResumeValues] = LVP.executePlan( + EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, DT, true); ++LoopsVectorized; - simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); - formLCSSARecursively(*L, *DT, LI, SE); - // Second pass vectorizes the epilogue and adjusts the control flow // edges from the first pass. EPI.MainLoopVF = EPI.EpilogueVF; @@ -10616,33 +10168,75 @@ bool LoopVectorizePass::processLoop(Loop *L) { Checks); VPlan &BestEpiPlan = LVP.getBestPlanFor(EPI.EpilogueVF); + VPRegionBlock *VectorLoop = BestEpiPlan.getVectorLoopRegion(); + VPBasicBlock *Header = VectorLoop->getEntryBasicBlock(); + Header->setName("vec.epilog.vector.body"); + + // Re-use the trip count and steps expanded for the main loop, as + // skeleton creation needs it as a value that dominates both the scalar + // and vector epilogue loops + // TODO: This is a workaround needed for epilogue vectorization and it + // should be removed once induction resume value creation is done + // directly in VPlan. + EpilogILV.setTripCount(MainILV.getTripCount()); + for (auto &R : make_early_inc_range(*BestEpiPlan.getPreheader())) { + auto *ExpandR = cast<VPExpandSCEVRecipe>(&R); + auto *ExpandedVal = BestEpiPlan.getVPValueOrAddLiveIn( + ExpandedSCEVs.find(ExpandR->getSCEV())->second); + ExpandR->replaceAllUsesWith(ExpandedVal); + ExpandR->eraseFromParent(); + } - // Ensure that the start values for any VPReductionPHIRecipes are - // updated before vectorising the epilogue loop. - VPBasicBlock *Header = BestEpiPlan.getEntry()->getEntryBasicBlock(); + // Ensure that the start values for any VPWidenIntOrFpInductionRecipe, + // VPWidenPointerInductionRecipe and VPReductionPHIRecipes are updated + // before vectorizing the epilogue loop. for (VPRecipeBase &R : Header->phis()) { + if (isa<VPCanonicalIVPHIRecipe>(&R)) + continue; + + Value *ResumeV = nullptr; + // TODO: Move setting of resume values to prepareToExecute. if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) { - if (auto *Resume = MainILV.getReductionResumeValue( - ReductionPhi->getRecurrenceDescriptor())) { - VPValue *StartVal = new VPValue(Resume); - BestEpiPlan.addExternalDef(StartVal); - ReductionPhi->setOperand(0, StartVal); + ResumeV = ReductionResumeValues + .find(&ReductionPhi->getRecurrenceDescriptor()) + ->second; + } else { + // Create induction resume values for both widened pointer and + // integer/fp inductions and update the start value of the induction + // recipes to use the resume value. + PHINode *IndPhi = nullptr; + const InductionDescriptor *ID; + if (auto *Ind = dyn_cast<VPWidenPointerInductionRecipe>(&R)) { + IndPhi = cast<PHINode>(Ind->getUnderlyingValue()); + ID = &Ind->getInductionDescriptor(); + } else { + auto *WidenInd = cast<VPWidenIntOrFpInductionRecipe>(&R); + IndPhi = WidenInd->getPHINode(); + ID = &WidenInd->getInductionDescriptor(); } + + ResumeV = MainILV.createInductionResumeValue( + IndPhi, *ID, getExpandedStep(*ID, ExpandedSCEVs), + {EPI.MainLoopIterationCountCheck}); } + assert(ResumeV && "Must have a resume value"); + VPValue *StartVal = BestEpiPlan.getVPValueOrAddLiveIn(ResumeV); + cast<VPHeaderPHIRecipe>(&R)->setStartValue(StartVal); } LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, - DT); + DT, true, &ExpandedSCEVs); ++LoopsEpilogueVectorized; if (!MainILV.areSafetyChecksAdded()) DisableRuntimeUnroll = true; } else { - InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC, - &LVL, &CM, BFI, PSI, Checks); + InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, + VF.MinProfitableTripCount, IC, &LVL, &CM, BFI, + PSI, Checks); VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); - LVP.executePlan(VF.Width, IC, BestPlan, LB, DT); + LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false); ++LoopsVectorized; // Add metadata to disable runtime unrolling a scalar loop when there @@ -10652,24 +10246,18 @@ bool LoopVectorizePass::processLoop(Loop *L) { DisableRuntimeUnroll = true; } // Report the vectorization decision. - ORE->emit([&]() { - return OptimizationRemark(LV_NAME, "Vectorized", L->getStartLoc(), - L->getHeader()) - << "vectorized loop (vectorization width: " - << NV("VectorizationFactor", VF.Width) - << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; - }); + reportVectorization(ORE, L, VF, IC); } if (ORE->allowExtraAnalysis(LV_NAME)) checkMixedPrecision(L, ORE); } - Optional<MDNode *> RemainderLoopID = + std::optional<MDNode *> RemainderLoopID = makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, LLVMLoopVectorizeFollowupEpilogue}); - if (RemainderLoopID.hasValue()) { - L->setLoopID(RemainderLoopID.getValue()); + if (RemainderLoopID) { + L->setLoopID(*RemainderLoopID); } else { if (DisableRuntimeUnroll) AddRuntimeUnrollDisableMetaData(L); @@ -10684,19 +10272,17 @@ bool LoopVectorizePass::processLoop(Loop *L) { LoopVectorizeResult LoopVectorizePass::runImpl( Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, - DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, - DemandedBits &DB_, AAResults &AA_, AssumptionCache &AC_, - std::function<const LoopAccessInfo &(Loop &)> &GetLAA_, + DominatorTree &DT_, BlockFrequencyInfo *BFI_, TargetLibraryInfo *TLI_, + DemandedBits &DB_, AssumptionCache &AC_, LoopAccessInfoManager &LAIs_, OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { SE = &SE_; LI = &LI_; TTI = &TTI_; DT = &DT_; - BFI = &BFI_; + BFI = BFI_; TLI = TLI_; - AA = &AA_; AC = &AC_; - GetLAA = &GetLAA_; + LAIs = &LAIs_; DB = &DB_; ORE = &ORE_; PSI = PSI_; @@ -10709,7 +10295,7 @@ LoopVectorizeResult LoopVectorizePass::runImpl( // vector registers, loop vectorization may still enable scalar // interleaving. if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && - TTI->getMaxInterleaveFactor(1) < 2) + TTI->getMaxInterleaveFactor(ElementCount::getFixed(1)) < 2) return LoopVectorizeResult(false, false); bool Changed = false, CFGChanged = false; @@ -10719,7 +10305,7 @@ LoopVectorizeResult LoopVectorizePass::runImpl( // legality and profitability checks. This means running the loop vectorizer // will simplify all loops, regardless of whether anything end up being // vectorized. - for (auto &L : *LI) + for (const auto &L : *LI) Changed |= CFGChanged |= simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); @@ -10742,6 +10328,15 @@ LoopVectorizeResult LoopVectorizePass::runImpl( Changed |= formLCSSARecursively(*L, *DT, LI, SE); Changed |= CFGChanged |= processLoop(L); + + if (Changed) { + LAIs->clear(); + +#ifndef NDEBUG + if (VerifySCEV) + SE->verify(); +#endif + } } // Process each loop nest in the function. @@ -10750,33 +10345,37 @@ LoopVectorizeResult LoopVectorizePass::runImpl( PreservedAnalyses LoopVectorizePass::run(Function &F, FunctionAnalysisManager &AM) { - auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); auto &LI = AM.getResult<LoopAnalysis>(F); + // There are no loops in the function. Return before computing other expensive + // analyses. + if (LI.empty()) + return PreservedAnalyses::all(); + auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); auto &TTI = AM.getResult<TargetIRAnalysis>(F); auto &DT = AM.getResult<DominatorTreeAnalysis>(F); - auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); - auto &AA = AM.getResult<AAManager>(F); auto &AC = AM.getResult<AssumptionAnalysis>(F); auto &DB = AM.getResult<DemandedBitsAnalysis>(F); auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); - auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); - std::function<const LoopAccessInfo &(Loop &)> GetLAA = - [&](Loop &L) -> const LoopAccessInfo & { - LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, - TLI, TTI, nullptr, nullptr, nullptr}; - return LAM.getResult<LoopAccessAnalysis>(L, AR); - }; + LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(F); auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); ProfileSummaryInfo *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); + BlockFrequencyInfo *BFI = nullptr; + if (PSI && PSI->hasProfileSummary()) + BFI = &AM.getResult<BlockFrequencyAnalysis>(F); LoopVectorizeResult Result = - runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE, PSI); + runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AC, LAIs, ORE, PSI); if (!Result.MadeAnyChange) return PreservedAnalyses::all(); PreservedAnalyses PA; + if (isAssignmentTrackingEnabled(*F.getParent())) { + for (auto &BB : F) + RemoveRedundantDbgInstrs(&BB); + } + // We currently do not preserve loopinfo/dominator analyses with outer loop // vectorization. Until this is addressed, mark these analyses as preserved // only for non-VPlan-native path. @@ -10784,6 +10383,7 @@ PreservedAnalyses LoopVectorizePass::run(Function &F, if (!EnableVPlanNativePath) { PA.preserve<LoopAnalysis>(); PA.preserve<DominatorTreeAnalysis>(); + PA.preserve<ScalarEvolutionAnalysis>(); } if (Result.MadeCFGChange) { @@ -10804,8 +10404,8 @@ void LoopVectorizePass::printPipeline( static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( OS, MapClassName2PassName); - OS << "<"; + OS << '<'; OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; - OS << ">"; + OS << '>'; } |