X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FVectorize%2FLoopVectorize.cpp;h=a8e1f4579c7f4242acd65c4c0583774576902d66;hb=1386692ef64d3151da8986589eadf0c58aba5c50;hp=d571903984c4a6abec84827edf6dd45dbcbc3ede;hpb=9e5329d77e590f757dbd8384f418e44df9dbf91a;p=oota-llvm.git diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp index d571903984c..a8e1f4579c7 100644 --- a/lib/Transforms/Vectorize/LoopVectorize.cpp +++ b/lib/Transforms/Vectorize/LoopVectorize.cpp @@ -6,7 +6,51 @@ // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// -#include "LoopVectorize.h" +// +// This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops +// and generates target-independent LLVM-IR. +// The vectorizer uses the TargetTransformInfo analysis to estimate the costs +// of instructions in order to estimate the profitability of vectorization. +// +// The loop vectorizer combines consecutive loop iterations into a single +// 'wide' iteration. After this transformation the index is incremented +// by the SIMD vector width, and not by one. +// +// This pass has three parts: +// 1. The main loop pass that drives the different parts. +// 2. LoopVectorizationLegality - A unit that checks for the legality +// of the vectorization. +// 3. InnerLoopVectorizer - A unit that performs the actual +// widening of instructions. +// 4. LoopVectorizationCostModel - A unit that checks for the profitability +// of vectorization. It decides on the optimal vector width, which +// can be one, if vectorization is not profitable. +// +//===----------------------------------------------------------------------===// +// +// The reduction-variable vectorization is based on the paper: +// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. +// +// Variable uniformity checks are inspired by: +// Karrenberg, R. and Hack, S. Whole Function Vectorization. +// +// Other ideas/concepts are from: +// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. +// +// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of +// Vectorizing Compilers. +// +//===----------------------------------------------------------------------===// + +#define LV_NAME "loop-vectorize" +#define DEBUG_TYPE LV_NAME + +#include "llvm/Transforms/Vectorize.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/ADT/MapVector.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringExtras.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/AliasSetTracker.h" @@ -14,40 +58,729 @@ #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/LoopIterator.h" #include "llvm/Analysis/LoopPass.h" +#include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpander.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" +#include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/Analysis/Verifier.h" -#include "llvm/Constants.h" -#include "llvm/DataLayout.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Function.h" -#include "llvm/Instructions.h" -#include "llvm/IntrinsicInst.h" -#include "llvm/LLVMContext.h" -#include "llvm/Module.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Function.h" +#include "llvm/IR/IRBuilder.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/LLVMContext.h" +#include "llvm/IR/Module.h" +#include "llvm/IR/Type.h" +#include "llvm/IR/Value.h" #include "llvm/Pass.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/PatternMatch.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/TargetTransformInfo.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Scalar.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Local.h" -#include "llvm/Transforms/Vectorize.h" -#include "llvm/Type.h" -#include "llvm/Value.h" +#include +#include + +using namespace llvm; +using namespace llvm::PatternMatch; static cl::opt VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, cl::desc("Sets the SIMD width. Zero is autoselect.")); +static cl::opt +VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden, + cl::desc("Sets the vectorization unroll count. " + "Zero is autoselect.")); + static cl::opt EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, cl::desc("Enable if-conversion during vectorization.")); +/// We don't vectorize loops with a known constant trip count below this number. +static cl::opt +TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), + cl::Hidden, + cl::desc("Don't vectorize loops with a constant " + "trip count that is smaller than this " + "value.")); + +/// We don't unroll loops with a known constant trip count below this number. +static const unsigned TinyTripCountUnrollThreshold = 128; + +/// When performing memory disambiguation checks at runtime do not make more +/// than this number of comparisons. +static const unsigned RuntimeMemoryCheckThreshold = 8; + +/// We use a metadata with this name to indicate that a scalar loop was +/// vectorized and that we don't need to re-vectorize it if we run into it +/// again. +static const char* +AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized"; + namespace { +// Forward declarations. +class LoopVectorizationLegality; +class LoopVectorizationCostModel; + +/// InnerLoopVectorizer vectorizes loops which contain only one basic +/// block to a specified vectorization factor (VF). +/// This class performs the widening of scalars into vectors, or multiple +/// scalars. This class also implements the following features: +/// * It inserts an epilogue loop for handling loops that don't have iteration +/// counts that are known to be a multiple of the vectorization factor. +/// * It handles the code generation for reduction variables. +/// * Scalarization (implementation using scalars) of un-vectorizable +/// instructions. +/// InnerLoopVectorizer does not perform any vectorization-legality +/// checks, and relies on the caller to check for the different legality +/// aspects. The InnerLoopVectorizer relies on the +/// LoopVectorizationLegality class to provide information about the induction +/// and reduction variables that were found to a given vectorization factor. +class InnerLoopVectorizer { +public: + InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, + DominatorTree *DT, DataLayout *DL, + const TargetLibraryInfo *TLI, unsigned VecWidth, + unsigned UnrollFactor) + : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI), + VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0), + OldInduction(0), WidenMap(UnrollFactor) {} + + // Perform the actual loop widening (vectorization). + void vectorize(LoopVectorizationLegality *Legal) { + // Create a new empty loop. Unlink the old loop and connect the new one. + createEmptyLoop(Legal); + // Widen each instruction in the old loop to a new one in the new loop. + // Use the Legality module to find the induction and reduction variables. + vectorizeLoop(Legal); + // Register the new loop and update the analysis passes. + updateAnalysis(); + } + +private: + /// A small list of PHINodes. + typedef SmallVector PhiVector; + /// When we unroll loops we have multiple vector values for each scalar. + /// This data structure holds the unrolled and vectorized values that + /// originated from one scalar instruction. + typedef SmallVector VectorParts; + + /// Add code that checks at runtime if the accessed arrays overlap. + /// Returns the comparator value or NULL if no check is needed. + Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal, + Instruction *Loc); + /// Create an empty loop, based on the loop ranges of the old loop. + void createEmptyLoop(LoopVectorizationLegality *Legal); + /// Copy and widen the instructions from the old loop. + void vectorizeLoop(LoopVectorizationLegality *Legal); + + /// A helper function that computes the predicate of the block BB, assuming + /// that the header block of the loop is set to True. It returns the *entry* + /// mask for the block BB. + VectorParts createBlockInMask(BasicBlock *BB); + /// A helper function that computes the predicate of the edge between SRC + /// and DST. + VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); + + /// A helper function to vectorize a single BB within the innermost loop. + void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, + PhiVector *PV); + + /// Insert the new loop to the loop hierarchy and pass manager + /// and update the analysis passes. + void updateAnalysis(); + + /// This instruction is un-vectorizable. Implement it as a sequence + /// of scalars. + void scalarizeInstruction(Instruction *Instr); + + /// Vectorize Load and Store instructions, + void vectorizeMemoryInstruction(Instruction *Instr, + LoopVectorizationLegality *Legal); + + /// 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. + Value *getBroadcastInstrs(Value *V); + + /// This function adds 0, 1, 2 ... to each vector element, starting at zero. + /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). + /// The sequence starts at StartIndex. + Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate); + + /// When we go over instructions in the basic block we rely on previous + /// values within the current basic block or on loop invariant values. + /// When we widen (vectorize) values we place them in the map. If the values + /// are not within the map, they have to be loop invariant, so we simply + /// broadcast them into a vector. + VectorParts &getVectorValue(Value *V); + + /// Generate a shuffle sequence that will reverse the vector Vec. + Value *reverseVector(Value *Vec); + + /// This is a helper class that holds the vectorizer state. It maps scalar + /// instructions to vector instructions. When the code is 'unrolled' then + /// then a single scalar value is mapped to multiple vector parts. The parts + /// are stored in the VectorPart type. + struct ValueMap { + /// C'tor. UnrollFactor controls the number of vectors ('parts') that + /// are mapped. + ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} + + /// \return True if 'Key' is saved in the Value Map. + bool has(Value *Key) const { return MapStorage.count(Key); } + + /// Initializes a new entry in the map. Sets all of the vector parts to the + /// save value in 'Val'. + /// \return A reference to a vector with splat values. + VectorParts &splat(Value *Key, Value *Val) { + VectorParts &Entry = MapStorage[Key]; + Entry.assign(UF, Val); + return Entry; + } + + ///\return A reference to the value that is stored at 'Key'. + VectorParts &get(Value *Key) { + VectorParts &Entry = MapStorage[Key]; + if (Entry.empty()) + Entry.resize(UF); + assert(Entry.size() == UF); + return Entry; + } + + private: + /// The unroll factor. Each entry in the map stores this number of vector + /// elements. + unsigned UF; + + /// Map storage. We use std::map and not DenseMap because insertions to a + /// dense map invalidates its iterators. + std::map MapStorage; + }; + + /// The original loop. + Loop *OrigLoop; + /// Scev analysis to use. + ScalarEvolution *SE; + /// Loop Info. + LoopInfo *LI; + /// Dominator Tree. + DominatorTree *DT; + /// Data Layout. + DataLayout *DL; + /// Target Library Info. + const TargetLibraryInfo *TLI; + + /// The vectorization SIMD factor to use. Each vector will have this many + /// vector elements. + unsigned VF; + /// The vectorization unroll factor to use. Each scalar is vectorized to this + /// many different vector instructions. + unsigned UF; + + /// The builder that we use + IRBuilder<> Builder; + + // --- Vectorization state --- + + /// The vector-loop preheader. + BasicBlock *LoopVectorPreHeader; + /// The scalar-loop preheader. + BasicBlock *LoopScalarPreHeader; + /// Middle Block between the vector and the scalar. + BasicBlock *LoopMiddleBlock; + ///The ExitBlock of the scalar loop. + BasicBlock *LoopExitBlock; + ///The vector loop body. + BasicBlock *LoopVectorBody; + ///The scalar loop body. + BasicBlock *LoopScalarBody; + /// A list of all bypass blocks. The first block is the entry of the loop. + SmallVector LoopBypassBlocks; + + /// The new Induction variable which was added to the new block. + PHINode *Induction; + /// The induction variable of the old basic block. + PHINode *OldInduction; + /// Holds the extended (to the widest induction type) start index. + Value *ExtendedIdx; + /// Maps scalars to widened vectors. + ValueMap WidenMap; +}; + +/// \brief Check if conditionally executed loads are hoistable. +/// +/// This class has two functions. isHoistableLoad and canHoistAllLoads. +/// isHoistableLoad should be called on all load instructions that are executed +/// conditionally. After all conditional loads are processed, the client should +/// call canHoistAllLoads to determine if all of the conditional execute loads +/// have an unconditional memory access in the loop. +class LoadHoisting { + typedef SmallPtrSet MemorySet; + + Loop *TheLoop; + DominatorTree *DT; + MemorySet CondLoadAddrSet; + +public: + LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {} + + /// \brief Check if the instruction is a load with a identifiable address. + bool isHoistableLoad(Instruction *L); + + /// \brief Check if all of the conditional loads are hoistable because there + /// exists an unconditional memory access to the same address in the loop. + bool canHoistAllLoads(); +}; + +bool LoadHoisting::isHoistableLoad(Instruction *L) { + LoadInst *LI = dyn_cast(L); + if (!LI) + return false; + + CondLoadAddrSet.insert(LI->getPointerOperand()); + return true; +} + +static void addMemAccesses(BasicBlock *BB, SmallPtrSet &Set) { + for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) { + Instruction *I = &*BI; + Value *Addr = 0; + + // Try a load. + LoadInst *LI = dyn_cast(I); + if (LI) { + Addr = LI->getPointerOperand(); + Set.insert(Addr); + continue; + } + + // Try a store. + StoreInst *SI = dyn_cast(I); + if (!SI) + continue; + + Addr = SI->getPointerOperand(); + Set.insert(Addr); + } +} + +bool LoadHoisting::canHoistAllLoads() { + // No conditional loads. + if (CondLoadAddrSet.empty()) + return true; + + MemorySet UncondMemAccesses; + std::vector &LoopBlocks = TheLoop->getBlocksVector(); + BasicBlock *LoopLatch = TheLoop->getLoopLatch(); + + // Iterate over the unconditional blocks and collect memory access addresses. + for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { + BasicBlock *BB = LoopBlocks[i]; + + // Ignore conditional blocks. + if (BB != LoopLatch && !DT->dominates(BB, LoopLatch)) + continue; + + addMemAccesses(BB, UncondMemAccesses); + } + + // And make sure there is a matching unconditional access for every + // conditional load. + for (MemorySet::iterator MI = CondLoadAddrSet.begin(), + ME = CondLoadAddrSet.end(); MI != ME; ++MI) + if (!UncondMemAccesses.count(*MI)) + return false; + + return true; +} + +/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and +/// to what vectorization factor. +/// This class does not look at the profitability of vectorization, only the +/// legality. This class has two main kinds of checks: +/// * Memory checks - The code in canVectorizeMemory checks if vectorization +/// will change the order of memory accesses in a way that will change the +/// correctness of the program. +/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory +/// checks for a number of different conditions, such as the availability of a +/// single induction variable, that all types are supported and vectorize-able, +/// etc. This code reflects the capabilities of InnerLoopVectorizer. +/// This class is also used by InnerLoopVectorizer for identifying +/// induction variable and the different reduction variables. +class LoopVectorizationLegality { +public: + LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL, + DominatorTree *DT, TargetTransformInfo* TTI, + AliasAnalysis *AA, TargetLibraryInfo *TLI) + : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI), + Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false), + LoadSpeculation(L, DT) {} + + /// This enum represents the kinds of reductions that we support. + enum ReductionKind { + RK_NoReduction, ///< Not a reduction. + RK_IntegerAdd, ///< Sum of integers. + RK_IntegerMult, ///< Product of integers. + RK_IntegerOr, ///< Bitwise or logical OR of numbers. + RK_IntegerAnd, ///< Bitwise or logical AND of numbers. + RK_IntegerXor, ///< Bitwise or logical XOR of numbers. + RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()). + RK_FloatAdd, ///< Sum of floats. + RK_FloatMult, ///< Product of floats. + RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()). + }; + + /// This enum represents the kinds of inductions that we support. + enum InductionKind { + IK_NoInduction, ///< Not an induction variable. + IK_IntInduction, ///< Integer induction variable. Step = 1. + IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1. + IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem). + IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem). + }; + + // This enum represents the kind of minmax reduction. + enum MinMaxReductionKind { + MRK_Invalid, + MRK_UIntMin, + MRK_UIntMax, + MRK_SIntMin, + MRK_SIntMax, + MRK_FloatMin, + MRK_FloatMax + }; + + /// This POD struct holds information about reduction variables. + struct ReductionDescriptor { + ReductionDescriptor() : StartValue(0), LoopExitInstr(0), + Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} + + ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, + MinMaxReductionKind MK) + : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} + + // The starting value of the reduction. + // It does not have to be zero! + Value *StartValue; + // The instruction who's value is used outside the loop. + Instruction *LoopExitInstr; + // The kind of the reduction. + ReductionKind Kind; + // If this a min/max reduction the kind of reduction. + MinMaxReductionKind MinMaxKind; + }; + + /// This POD struct holds information about a potential reduction operation. + struct ReductionInstDesc { + ReductionInstDesc(bool IsRedux, Instruction *I) : + IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} + + ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : + IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} + + // Is this instruction a reduction candidate. + bool IsReduction; + // The last instruction in a min/max pattern (select of the select(icmp()) + // pattern), or the current reduction instruction otherwise. + Instruction *PatternLastInst; + // If this is a min/max pattern the comparison predicate. + MinMaxReductionKind MinMaxKind; + }; + + // This POD struct holds information about the memory runtime legality + // check that a group of pointers do not overlap. + struct RuntimePointerCheck { + RuntimePointerCheck() : Need(false) {} + + /// Reset the state of the pointer runtime information. + void reset() { + Need = false; + Pointers.clear(); + Starts.clear(); + Ends.clear(); + } + + /// Insert a pointer and calculate the start and end SCEVs. + void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr); + + /// This flag indicates if we need to add the runtime check. + bool Need; + /// Holds the pointers that we need to check. + SmallVector Pointers; + /// Holds the pointer value at the beginning of the loop. + SmallVector Starts; + /// Holds the pointer value at the end of the loop. + SmallVector Ends; + /// Holds the information if this pointer is used for writing to memory. + SmallVector IsWritePtr; + }; + + /// A POD for saving information about induction variables. + struct InductionInfo { + InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} + InductionInfo() : StartValue(0), IK(IK_NoInduction) {} + /// Start value. + Value *StartValue; + /// Induction kind. + InductionKind IK; + }; + + /// ReductionList contains the reduction descriptors for all + /// of the reductions that were found in the loop. + typedef DenseMap ReductionList; + + /// InductionList saves induction variables and maps them to the + /// induction descriptor. + typedef MapVector InductionList; + + /// Alias(Multi)Map stores the values (GEPs or underlying objects and their + /// respective Store/Load instruction(s) to calculate aliasing. + typedef MapVector AliasMap; + typedef DenseMap > AliasMultiMap; + + /// Returns true if it is legal to vectorize this loop. + /// This does not mean that it is profitable to vectorize this + /// loop, only that it is legal to do so. + bool canVectorize(); + + /// Returns the Induction variable. + PHINode *getInduction() { return Induction; } + + /// Returns the reduction variables found in the loop. + ReductionList *getReductionVars() { return &Reductions; } + + /// Returns the induction variables found in the loop. + InductionList *getInductionVars() { return &Inductions; } + + /// Returns the widest induction type. + Type *getWidestInductionType() { return WidestIndTy; } + + /// Returns True if V is an induction variable in this loop. + bool isInductionVariable(const Value *V); + + /// Return true if the block BB needs to be predicated in order for the loop + /// to be vectorized. + bool blockNeedsPredication(BasicBlock *BB); + + /// Check if this pointer is consecutive when vectorizing. This happens + /// when the last index of the GEP is the induction variable, or that the + /// pointer itself is an induction variable. + /// This check allows us to vectorize A[idx] into a wide load/store. + /// Returns: + /// 0 - Stride is unknown or non consecutive. + /// 1 - Address is consecutive. + /// -1 - Address is consecutive, and decreasing. + int isConsecutivePtr(Value *Ptr); + + /// Returns true if the value V is uniform within the loop. + bool isUniform(Value *V); + + /// Returns true if this instruction will remain scalar after vectorization. + bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } + + /// Returns the information that we collected about runtime memory check. + RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } + + /// This function returns the identity element (or neutral element) for + /// the operation K. + static Constant *getReductionIdentity(ReductionKind K, Type *Tp); +private: + /// Check if a single basic block loop is vectorizable. + /// At this point we know that this is a loop with a constant trip count + /// and we only need to check individual instructions. + bool canVectorizeInstrs(); + + /// When we vectorize loops we may change the order in which + /// we read and write from memory. This method checks if it is + /// legal to vectorize the code, considering only memory constrains. + /// Returns true if the loop is vectorizable + bool canVectorizeMemory(); + + /// Return true if we can vectorize this loop using the IF-conversion + /// transformation. + bool canVectorizeWithIfConvert(); + + /// Collect the variables that need to stay uniform after vectorization. + void collectLoopUniforms(); + + /// Return true if all of the instructions in the block can be speculatively + /// executed. + bool blockCanBePredicated(BasicBlock *BB); + + /// Returns True, if 'Phi' is the kind of reduction variable for type + /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. + bool AddReductionVar(PHINode *Phi, ReductionKind Kind); + /// Returns a struct describing if the instruction 'I' can be a reduction + /// variable of type 'Kind'. If the reduction is a min/max pattern of + /// select(icmp()) this function advances the instruction pointer 'I' from the + /// compare instruction to the select instruction and stores this pointer in + /// 'PatternLastInst' member of the returned struct. + ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, + ReductionInstDesc &Desc); + /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction + /// pattern corresponding to a min(X, Y) or max(X, Y). + static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, + ReductionInstDesc &Prev); + /// Returns the induction kind of Phi. This function may return NoInduction + /// if the PHI is not an induction variable. + InductionKind isInductionVariable(PHINode *Phi); + /// Return true if can compute the address bounds of Ptr within the loop. + bool hasComputableBounds(Value *Ptr); + /// Return true if there is the chance of write reorder. + bool hasPossibleGlobalWriteReorder(Value *Object, + Instruction *Inst, + AliasMultiMap &WriteObjects, + unsigned MaxByteWidth); + /// Return the AA location for a load or a store. + AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst); + + + /// The loop that we evaluate. + Loop *TheLoop; + /// Scev analysis. + ScalarEvolution *SE; + /// DataLayout analysis. + DataLayout *DL; + /// Dominators. + DominatorTree *DT; + /// Target Info. + TargetTransformInfo *TTI; + /// Alias Analysis. + AliasAnalysis *AA; + /// Target Library Info. + TargetLibraryInfo *TLI; + + // --- vectorization state --- // + + /// Holds the integer induction variable. This is the counter of the + /// loop. + PHINode *Induction; + /// Holds the reduction variables. + ReductionList Reductions; + /// Holds all of the induction variables that we found in the loop. + /// Notice that inductions don't need to start at zero and that induction + /// variables can be pointers. + InductionList Inductions; + /// Holds the widest induction type encountered. + Type *WidestIndTy; + + /// Allowed outside users. This holds the reduction + /// vars which can be accessed from outside the loop. + SmallPtrSet AllowedExit; + /// This set holds the variables which are known to be uniform after + /// vectorization. + SmallPtrSet Uniforms; + /// We need to check that all of the pointers in this list are disjoint + /// at runtime. + RuntimePointerCheck PtrRtCheck; + /// Can we assume the absence of NaNs. + bool HasFunNoNaNAttr; + + /// Utility to determine whether loads can be speculated. + LoadHoisting LoadSpeculation; +}; + +/// LoopVectorizationCostModel - estimates the expected speedups due to +/// vectorization. +/// In many cases vectorization is not profitable. This can happen because of +/// a number of reasons. In this class we mainly attempt to predict the +/// expected speedup/slowdowns due to the supported instruction set. We use the +/// TargetTransformInfo to query the different backends for the cost of +/// different operations. +class LoopVectorizationCostModel { +public: + LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, + LoopVectorizationLegality *Legal, + const TargetTransformInfo &TTI, + DataLayout *DL, const TargetLibraryInfo *TLI) + : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {} + + /// Information about vectorization costs + struct VectorizationFactor { + unsigned Width; // Vector width with best cost + unsigned Cost; // Cost of the loop with that width + }; + /// \return The most profitable vectorization factor and the cost of that VF. + /// This method checks every power of two up to VF. If UserVF is not ZERO + /// then this vectorization factor will be selected if vectorization is + /// possible. + VectorizationFactor selectVectorizationFactor(bool OptForSize, + unsigned UserVF); + + /// \return The size (in bits) of the widest type in the code that + /// needs to be vectorized. We ignore values that remain scalar such as + /// 64 bit loop indices. + unsigned getWidestType(); + + /// \return The most profitable unroll factor. + /// If UserUF is non-zero then this method finds the best unroll-factor + /// based on register pressure and other parameters. + /// VF and LoopCost are the selected vectorization factor and the cost of the + /// selected VF. + unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF, + unsigned LoopCost); + + /// \brief A struct that represents some properties of the register usage + /// of a loop. + struct RegisterUsage { + /// Holds the number of loop invariant values that are used in the loop. + unsigned LoopInvariantRegs; + /// Holds the maximum number of concurrent live intervals in the loop. + unsigned MaxLocalUsers; + /// Holds the number of instructions in the loop. + unsigned NumInstructions; + }; + + /// \return information about the register usage of the loop. + RegisterUsage calculateRegisterUsage(); + +private: + /// 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. + unsigned expectedCost(unsigned VF); + + /// Returns the execution time cost of an instruction for a given vector + /// width. Vector width of one means scalar. + unsigned getInstructionCost(Instruction *I, unsigned VF); + + /// A helper function for converting Scalar types to vector types. + /// If the incoming type is void, we return void. If the VF is 1, we return + /// the scalar type. + static Type* ToVectorTy(Type *Scalar, unsigned VF); + + /// Returns whether the instruction is a load or store and will be a emitted + /// as a vector operation. + bool isConsecutiveLoadOrStore(Instruction *I); + + /// The loop that we evaluate. + Loop *TheLoop; + /// Scev analysis. + ScalarEvolution *SE; + /// Loop Info analysis. + LoopInfo *LI; + /// Vectorization legality. + LoopVectorizationLegality *Legal; + /// Vector target information. + const TargetTransformInfo &TTI; + /// Target data layout information. + DataLayout *DL; + /// Target Library Info. + const TargetLibraryInfo *TLI; +}; + /// The LoopVectorize Pass. struct LoopVectorize : public LoopPass { /// Pass identification, replacement for typeid @@ -62,6 +795,8 @@ struct LoopVectorize : public LoopPass { LoopInfo *LI; TargetTransformInfo *TTI; DominatorTree *DT; + AliasAnalysis *AA; + TargetLibraryInfo *TLI; virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { // We only vectorize innermost loops. @@ -71,44 +806,62 @@ struct LoopVectorize : public LoopPass { SE = &getAnalysis(); DL = getAnalysisIfAvailable(); LI = &getAnalysis(); - TTI = getAnalysisIfAvailable(); + TTI = &getAnalysis(); DT = &getAnalysis(); + AA = getAnalysisIfAvailable(); + TLI = getAnalysisIfAvailable(); + + if (DL == NULL) { + DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout"); + return false; + } DEBUG(dbgs() << "LV: Checking a loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"); // Check if it is legal to vectorize the loop. - LoopVectorizationLegality LVL(L, SE, DL, DT); + LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI); if (!LVL.canVectorize()) { DEBUG(dbgs() << "LV: Not vectorizing.\n"); return false; } - // Select the preffered vectorization factor. - const VectorTargetTransformInfo *VTTI = 0; - if (TTI) - VTTI = TTI->getVectorTargetTransformInfo(); // Use the cost model. - LoopVectorizationCostModel CM(L, SE, &LVL, VTTI); + LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI); - // Check the function attribues to find out if this function should be + // Check the function attributes to find out if this function should be // optimized for size. Function *F = L->getHeader()->getParent(); - Attribute::AttrKind SzAttr= Attribute::OptimizeForSize; - bool OptForSize = F->getFnAttributes().hasAttribute(SzAttr); + Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; + Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; + unsigned FnIndex = AttributeSet::FunctionIndex; + bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); + bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); + + if (NoFloat) { + DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" + "attribute is used.\n"); + return false; + } - unsigned VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); + // Select the optimal vectorization factor. + LoopVectorizationCostModel::VectorizationFactor VF; + VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); + // Select the unroll factor. + unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll, + VF.Width, VF.Cost); - if (VF == 1) { + if (VF.Width == 1) { DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); return false; } - DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<< + DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<< F->getParent()->getModuleIdentifier()<<"\n"); + DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n"); - // If we decided that it is *legal* to vectorizer the loop then do it. - InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF); + // If we decided that it is *legal* to vectorize the loop then do it. + InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF); LB.vectorize(&LVL); DEBUG(verifyFunction(*L->getHeader()->getParent())); @@ -119,16 +872,17 @@ struct LoopVectorize : public LoopPass { LoopPass::getAnalysisUsage(AU); AU.addRequiredID(LoopSimplifyID); AU.addRequiredID(LCSSAID); + AU.addRequired(); AU.addRequired(); AU.addRequired(); - AU.addRequired(); + AU.addRequired(); AU.addPreserved(); AU.addPreserved(); } }; -}// namespace +} // end anonymous namespace //===----------------------------------------------------------------------===// // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and @@ -137,7 +891,8 @@ struct LoopVectorize : public LoopPass { void LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, - Loop *Lp, Value *Ptr) { + Loop *Lp, Value *Ptr, + bool WritePtr) { const SCEV *Sc = SE->getSCEV(Ptr); const SCEVAddRecExpr *AR = dyn_cast(Sc); assert(AR && "Invalid addrec expression"); @@ -146,14 +901,10 @@ LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, Pointers.push_back(Ptr); Starts.push_back(AR->getStart()); Ends.push_back(ScEnd); + IsWritePtr.push_back(WritePtr); } Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { - // Create the types. - LLVMContext &C = V->getContext(); - Type *VTy = VectorType::get(V->getType(), VF); - Type *I32 = IntegerType::getInt32Ty(C); - // Save the current insertion location. Instruction *Loc = Builder.GetInsertPoint(); @@ -166,14 +917,8 @@ Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { if (Invariant) Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); - Constant *Zero = ConstantInt::get(I32, 0); - Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF)); - Value *UndefVal = UndefValue::get(VTy); - // Insert the value into a new vector. - Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero); // Broadcast the scalar into all locations in the vector. - Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros, - "broadcast"); + Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); // Restore the builder insertion point. if (Invariant) @@ -182,7 +927,8 @@ Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { return Shuf; } -Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) { +Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx, + bool Negate) { assert(Val->getType()->isVectorTy() && "Must be a vector"); assert(Val->getType()->getScalarType()->isIntegerTy() && "Elem must be an integer"); @@ -193,8 +939,10 @@ Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) { SmallVector Indices; // Create a vector of consecutive numbers from zero to VF. - for (int i = 0; i < VLen; ++i) - Indices.push_back(ConstantInt::get(ITy, Negate ? (-i): i )); + for (int i = 0; i < VLen; ++i) { + int64_t Idx = Negate ? (-i) : i; + Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate)); + } // Add the consecutive indices to the vector value. Constant *Cv = ConstantVector::get(Indices); @@ -202,28 +950,56 @@ Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) { return Builder.CreateAdd(Val, Cv, "induction"); } -bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { +int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); + // Make sure that the pointer does not point to structs. + if (cast(Ptr->getType())->getElementType()->isAggregateType()) + return 0; // If this value is a pointer induction variable we know it is consecutive. PHINode *Phi = dyn_cast_or_null(Ptr); if (Phi && Inductions.count(Phi)) { InductionInfo II = Inductions[Phi]; - if (PtrInduction == II.IK) - return true; + if (IK_PtrInduction == II.IK) + return 1; + else if (IK_ReversePtrInduction == II.IK) + return -1; } GetElementPtrInst *Gep = dyn_cast_or_null(Ptr); if (!Gep) - return false; + return 0; unsigned NumOperands = Gep->getNumOperands(); Value *LastIndex = Gep->getOperand(NumOperands - 1); + Value *GpPtr = Gep->getPointerOperand(); + // If this GEP value is a consecutive pointer induction variable and all of + // the indices are constant then we know it is consecutive. We can + Phi = dyn_cast(GpPtr); + if (Phi && Inductions.count(Phi)) { + + // Make sure that the pointer does not point to structs. + PointerType *GepPtrType = cast(GpPtr->getType()); + if (GepPtrType->getElementType()->isAggregateType()) + return 0; + + // Make sure that all of the index operands are loop invariant. + for (unsigned i = 1; i < NumOperands; ++i) + if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) + return 0; + + InductionInfo II = Inductions[Phi]; + if (IK_PtrInduction == II.IK) + return 1; + else if (IK_ReversePtrInduction == II.IK) + return -1; + } + // Check that all of the gep indices are uniform except for the last. for (unsigned i = 0; i < NumOperands - 1; ++i) if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) - return false; + return 0; // We can emit wide load/stores only if the last index is the induction // variable. @@ -234,39 +1010,159 @@ bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { // The memory is consecutive because the last index is consecutive // and all other indices are loop invariant. if (Step->isOne()) - return true; + return 1; + if (Step->isAllOnesValue()) + return -1; } - return false; + return 0; } bool LoopVectorizationLegality::isUniform(Value *V) { return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); } -Value *InnerLoopVectorizer::getVectorValue(Value *V) { +InnerLoopVectorizer::VectorParts& +InnerLoopVectorizer::getVectorValue(Value *V) { assert(V != Induction && "The new induction variable should not be used."); assert(!V->getType()->isVectorTy() && "Can't widen a vector"); - // If we saved a vectorized copy of V, use it. - Value *&MapEntry = WidenMap[V]; - if (MapEntry) - return MapEntry; - // Broadcast V and save the value for future uses. + // If we have this scalar in the map, return it. + if (WidenMap.has(V)) + return WidenMap.get(V); + + // If this scalar is unknown, assume that it is a constant or that it is + // loop invariant. Broadcast V and save the value for future uses. Value *B = getBroadcastInstrs(V); - MapEntry = B; - return B; + return WidenMap.splat(V, B); } -Constant* -InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) { - return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true)); +Value *InnerLoopVectorizer::reverseVector(Value *Vec) { + assert(Vec->getType()->isVectorTy() && "Invalid type"); + SmallVector ShuffleMask; + for (unsigned i = 0; i < VF; ++i) + ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); + + return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), + ConstantVector::get(ShuffleMask), + "reverse"); +} + + +void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, + LoopVectorizationLegality *Legal) { + // Attempt to issue a wide load. + LoadInst *LI = dyn_cast(Instr); + StoreInst *SI = dyn_cast(Instr); + + assert((LI || SI) && "Invalid Load/Store instruction"); + + Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); + Type *DataTy = VectorType::get(ScalarDataTy, VF); + Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); + unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); + + unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy); + unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF; + + if (ScalarAllocatedSize != VectorElementSize) + return scalarizeInstruction(Instr); + + // If the pointer is loop invariant or if it is non consecutive, + // scalarize the load. + int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); + bool Reverse = ConsecutiveStride < 0; + bool UniformLoad = LI && Legal->isUniform(Ptr); + if (!ConsecutiveStride || UniformLoad) + return scalarizeInstruction(Instr); + + Constant *Zero = Builder.getInt32(0); + VectorParts &Entry = WidenMap.get(Instr); + + // Handle consecutive loads/stores. + GetElementPtrInst *Gep = dyn_cast(Ptr); + if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { + Value *PtrOperand = Gep->getPointerOperand(); + Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; + FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); + + // Create the new GEP with the new induction variable. + GetElementPtrInst *Gep2 = cast(Gep->clone()); + Gep2->setOperand(0, FirstBasePtr); + Gep2->setName("gep.indvar.base"); + Ptr = Builder.Insert(Gep2); + } else if (Gep) { + assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), + OrigLoop) && "Base ptr must be invariant"); + + // The last index does not have to be the induction. It can be + // consecutive and be a function of the index. For example A[I+1]; + unsigned NumOperands = Gep->getNumOperands(); + + Value *LastGepOperand = Gep->getOperand(NumOperands - 1); + VectorParts &GEPParts = getVectorValue(LastGepOperand); + Value *LastIndex = GEPParts[0]; + LastIndex = Builder.CreateExtractElement(LastIndex, Zero); + + // Create the new GEP with the new induction variable. + GetElementPtrInst *Gep2 = cast(Gep->clone()); + Gep2->setOperand(NumOperands - 1, LastIndex); + Gep2->setName("gep.indvar.idx"); + Ptr = Builder.Insert(Gep2); + } else { + // Use the induction element ptr. + assert(isa(Ptr) && "Invalid induction ptr"); + VectorParts &PtrVal = getVectorValue(Ptr); + Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); + } + + // Handle Stores: + if (SI) { + assert(!Legal->isUniform(SI->getPointerOperand()) && + "We do not allow storing to uniform addresses"); + + VectorParts &StoredVal = getVectorValue(SI->getValueOperand()); + for (unsigned Part = 0; Part < UF; ++Part) { + // Calculate the pointer for the specific unroll-part. + Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); + + if (Reverse) { + // If we store to reverse consecutive memory locations then we need + // to reverse the order of elements in the stored value. + StoredVal[Part] = reverseVector(StoredVal[Part]); + // If the address is consecutive but reversed, then the + // wide store needs to start at the last vector element. + PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); + PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); + } + + Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); + Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); + } + } + + for (unsigned Part = 0; Part < UF; ++Part) { + // Calculate the pointer for the specific unroll-part. + Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); + + if (Reverse) { + // If the address is consecutive but reversed, then the + // wide store needs to start at the last vector element. + PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); + PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); + } + + Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); + Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); + cast(LI)->setAlignment(Alignment); + Entry[Part] = Reverse ? reverseVector(LI) : LI; + } } void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); // Holds vector parameters or scalars, in case of uniform vals. - SmallVector Params; + SmallVector Params; // Find all of the vectorized parameters. for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { @@ -284,12 +1180,14 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { // If the src is an instruction that appeared earlier in the basic block // then it should already be vectorized. if (SrcInst && OrigLoop->contains(SrcInst)) { - assert(WidenMap.count(SrcInst) && "Source operand is unavailable"); + assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); // The parameter is a vector value from earlier. - Params.push_back(WidenMap[SrcInst]); + Params.push_back(WidenMap.get(SrcInst)); } else { // The parameter is a scalar from outside the loop. Maybe even a constant. - Params.push_back(SrcOp); + VectorParts Scalars; + Scalars.append(UF, SrcOp); + Params.push_back(Scalars); } } @@ -298,42 +1196,41 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { // Does this instruction return a value ? bool IsVoidRetTy = Instr->getType()->isVoidTy(); - Value *VecResults = 0; - - // If we have a return value, create an empty vector. We place the scalarized - // instructions in this vector. - if (!IsVoidRetTy) - VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF)); - - // For each scalar that we create: - for (unsigned i = 0; i < VF; ++i) { - Instruction *Cloned = Instr->clone(); - if (!IsVoidRetTy) - Cloned->setName(Instr->getName() + ".cloned"); - // Replace the operands of the cloned instrucions with extracted scalars. - for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { - Value *Op = Params[op]; - // Param is a vector. Need to extract the right lane. - if (Op->getType()->isVectorTy()) - Op = Builder.CreateExtractElement(Op, Builder.getInt32(i)); - Cloned->setOperand(op, Op); - } - // Place the cloned scalar in the new loop. - Builder.Insert(Cloned); + Value *UndefVec = IsVoidRetTy ? 0 : + UndefValue::get(VectorType::get(Instr->getType(), VF)); + // Create a new entry in the WidenMap and initialize it to Undef or Null. + VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); + + // For each vector unroll 'part': + for (unsigned Part = 0; Part < UF; ++Part) { + // For each scalar that we create: + for (unsigned Width = 0; Width < VF; ++Width) { + Instruction *Cloned = Instr->clone(); + if (!IsVoidRetTy) + Cloned->setName(Instr->getName() + ".cloned"); + // Replace the operands of the cloned instrucions with extracted scalars. + for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { + Value *Op = Params[op][Part]; + // Param is a vector. Need to extract the right lane. + if (Op->getType()->isVectorTy()) + Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); + Cloned->setOperand(op, Op); + } - // If the original scalar returns a value we need to place it in a vector - // so that future users will be able to use it. - if (!IsVoidRetTy) - VecResults = Builder.CreateInsertElement(VecResults, Cloned, - Builder.getInt32(i)); - } + // Place the cloned scalar in the new loop. + Builder.Insert(Cloned); - if (!IsVoidRetTy) - WidenMap[Instr] = VecResults; + // If the original scalar returns a value we need to place it in a vector + // so that future users will be able to use it. + if (!IsVoidRetTy) + VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, + Builder.getInt32(Width)); + } + } } -Value* +Instruction * InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, Instruction *Loc) { LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = @@ -342,7 +1239,7 @@ InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, if (!PtrRtCheck->Need) return NULL; - Value *MemoryRuntimeCheck = 0; + Instruction *MemoryRuntimeCheck = 0; unsigned NumPointers = PtrRtCheck->Pointers.size(); SmallVector Starts; SmallVector Ends; @@ -371,28 +1268,27 @@ InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, } } + IRBuilder<> ChkBuilder(Loc); + for (unsigned i = 0; i < NumPointers; ++i) { for (unsigned j = i+1; j < NumPointers; ++j) { - Instruction::CastOps Op = Instruction::BitCast; - Value *Start0 = CastInst::Create(Op, Starts[i], PtrArithTy, "bc", Loc); - Value *Start1 = CastInst::Create(Op, Starts[j], PtrArithTy, "bc", Loc); - Value *End0 = CastInst::Create(Op, Ends[i], PtrArithTy, "bc", Loc); - Value *End1 = CastInst::Create(Op, Ends[j], PtrArithTy, "bc", Loc); - - Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, - Start0, End1, "bound0", Loc); - Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, - Start1, End0, "bound1", Loc); - Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1, - "found.conflict", Loc); + // No need to check if two readonly pointers intersect. + if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j]) + continue; + + Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc"); + Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc"); + Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc"); + Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc"); + + Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0"); + Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1"); + Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); if (MemoryRuntimeCheck) - MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or, - MemoryRuntimeCheck, - IsConflict, - "conflict.rdx", Loc); - else - MemoryRuntimeCheck = IsConflict; + IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, + "conflict.rdx"); + MemoryRuntimeCheck = cast(IsConflict); } } @@ -406,7 +1302,7 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { the vectorized instructions while the old loop will continue to run the scalar remainder. - [ ] <-- vector loop bypass. + [ ] <-- vector loop bypass (may consist of multiple blocks). / | / v | [ ] <-- vector pre header. @@ -435,13 +1331,17 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { BasicBlock *ExitBlock = OrigLoop->getExitBlock(); assert(ExitBlock && "Must have an exit block"); + // Mark the old scalar loop with metadata that tells us not to vectorize this + // loop again if we run into it. + MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None); + OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD); + // Some loops have a single integer induction variable, while other loops // don't. One example is c++ iterators that often have multiple pointer // induction variables. In the code below we also support a case where we // don't have a single induction variable. OldInduction = Legal->getInduction(); - Type *IdxTy = OldInduction ? OldInduction->getType() : - DL->getIntPtrType(SE->getContext()); + Type *IdxTy = Legal->getWidestInductionType(); // Find the loop boundaries. const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch()); @@ -462,15 +1362,14 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { // The loop index does not have to start at Zero. Find the original start // value from the induction PHI node. If we don't have an induction variable // then we know that it starts at zero. - Value *StartIdx = OldInduction ? - OldInduction->getIncomingValueForBlock(BypassBlock): - ConstantInt::get(IdxTy, 0); + Builder.SetInsertPoint(BypassBlock->getTerminator()); + Value *StartIdx = ExtendedIdx = OldInduction ? + Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock), + IdxTy): + ConstantInt::get(IdxTy, 0); assert(BypassBlock && "Invalid loop structure"); - - // Generate the code that checks in runtime if arrays overlap. - Value *MemoryRuntimeCheck = addRuntimeCheck(Legal, - BypassBlock->getTerminator()); + LoopBypassBlocks.push_back(BypassBlock); // Split the single block loop into the two loop structure described above. BasicBlock *VectorPH = @@ -482,17 +1381,19 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { BasicBlock *ScalarPH = MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); - // This is the location in which we add all of the logic for bypassing - // the new vector loop. - Instruction *Loc = BypassBlock->getTerminator(); - // Use this IR builder to create the loop instructions (Phi, Br, Cmp) // inside the loop. Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); // Generate the induction variable. Induction = Builder.CreatePHI(IdxTy, 2, "index"); - Constant *Step = ConstantInt::get(IdxTy, VF); + // The loop step is equal to the vectorization factor (num of SIMD elements) + // times the unroll factor (num of SIMD instructions). + Constant *Step = ConstantInt::get(IdxTy, VF * UF); + + // This is the IR builder that we use to add all of the logic for bypassing + // the new vector loop. + IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); // We may need to extend the index in case there is a type mismatch. // We know that the count starts at zero and does not overflow. @@ -500,37 +1401,52 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { // The exit count can be of pointer type. Convert it to the correct // integer type. if (ExitCount->getType()->isPointerTy()) - Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc); + Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); else - Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc); + Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); } // Add the start index to the loop count to get the new end index. - Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc); + Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); // Now we need to generate the expression for N - (N % VF), which is // the part that the vectorized body will execute. - Constant *CIVF = ConstantInt::get(IdxTy, VF); - Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc); - Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc); - Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx, - "end.idx.rnd.down", Loc); + Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); + Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); + Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, + "end.idx.rnd.down"); // Now, compare the new count to zero. If it is zero skip the vector loop and // jump to the scalar loop. - Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, - IdxEndRoundDown, - StartIdx, - "cmp.zero", Loc); - - // If we are using memory runtime checks, include them in. - if (MemoryRuntimeCheck) - Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck, - "CntOrMem", Loc); + Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, + "cmp.zero"); + + BasicBlock *LastBypassBlock = BypassBlock; + + // 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. + Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, + BypassBlock->getTerminator()); + if (MemRuntimeCheck) { + // Create a new block containing the memory check. + BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, + "vector.memcheck"); + LoopBypassBlocks.push_back(CheckBlock); + + // Replace the branch into the memory check block with a conditional branch + // for the "few elements case". + Instruction *OldTerm = BypassBlock->getTerminator(); + BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); + OldTerm->eraseFromParent(); + + Cmp = MemRuntimeCheck; + LastBypassBlock = CheckBlock; + } - BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc); - // Remove the old terminator. - Loc->eraseFromParent(); + LastBypassBlock->getTerminator()->eraseFromParent(); + BranchInst::Create(MiddleBlock, VectorPH, Cmp, + LastBypassBlock); // We are going to resume the execution of the scalar loop. // Go over all of the induction variables that we found and fix the @@ -544,61 +1460,101 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { PHINode *ResumeIndex = 0; LoopVectorizationLegality::InductionList::iterator I, E; LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); + // Set builder to point to last bypass block. + BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator()); for (I = List->begin(), E = List->end(); I != E; ++I) { PHINode *OrigPhi = I->first; LoopVectorizationLegality::InductionInfo II = I->second; - PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val", + + Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType(); + PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val", MiddleBlock->getTerminator()); + // We might have extended the type of the induction variable but we need a + // truncated version for the scalar loop. + PHINode *TruncResumeVal = (OrigPhi == OldInduction) ? + PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val", + MiddleBlock->getTerminator()) : 0; + Value *EndValue = 0; switch (II.IK) { - case LoopVectorizationLegality::NoInduction: + case LoopVectorizationLegality::IK_NoInduction: llvm_unreachable("Unknown induction"); - case LoopVectorizationLegality::IntInduction: { - // Handle the integer induction counter: + case LoopVectorizationLegality::IK_IntInduction: { + // Handle the integer induction counter. assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); - assert(OrigPhi == OldInduction && "Unknown integer PHI"); - // We know what the end value is. - EndValue = IdxEndRoundDown; - // We also know which PHI node holds it. - ResumeIndex = ResumeVal; + + // We have the canonical induction variable. + if (OrigPhi == OldInduction) { + // Create a truncated version of the resume value for the scalar loop, + // we might have promoted the type to a larger width. + EndValue = + BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType()); + // The new PHI merges the original incoming value, in case of a bypass, + // or the value at the end of the vectorized loop. + for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) + TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); + TruncResumeVal->addIncoming(EndValue, VecBody); + + // We know what the end value is. + EndValue = IdxEndRoundDown; + // We also know which PHI node holds it. + ResumeIndex = ResumeVal; + break; + } + + // Not the canonical induction variable - add the vector loop count to the + // start value. + Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, + II.StartValue->getType(), + "cast.crd"); + EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end"); break; } - case LoopVectorizationLegality::ReverseIntInduction: { + case LoopVectorizationLegality::IK_ReverseIntInduction: { // Convert the CountRoundDown variable to the PHI size. - unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits(); - unsigned IISize = II.StartValue->getType()->getScalarSizeInBits(); - Value *CRD = CountRoundDown; - if (CRDSize > IISize) - CRD = CastInst::Create(Instruction::Trunc, CountRoundDown, - II.StartValue->getType(), - "tr.crd", BypassBlock->getTerminator()); - else if (CRDSize < IISize) - CRD = CastInst::Create(Instruction::SExt, CountRoundDown, - II.StartValue->getType(), - "sext.crd", BypassBlock->getTerminator()); - // Handle reverse integer induction counter: - EndValue = BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end", - BypassBlock->getTerminator()); + Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, + II.StartValue->getType(), + "cast.crd"); + // Handle reverse integer induction counter. + EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end"); break; } - case LoopVectorizationLegality::PtrInduction: { + case LoopVectorizationLegality::IK_PtrInduction: { // For pointer induction variables, calculate the offset using // the end index. - EndValue = GetElementPtrInst::Create(II.StartValue, CountRoundDown, - "ptr.ind.end", - BypassBlock->getTerminator()); + EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown, + "ptr.ind.end"); + break; + } + case LoopVectorizationLegality::IK_ReversePtrInduction: { + // The value at the end of the loop for the reverse pointer is calculated + // by creating a GEP with a negative index starting from the start value. + Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); + Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown, + "rev.ind.end"); + EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx, + "rev.ptr.ind.end"); break; } }// end of case // The new PHI merges the original incoming value, in case of a bypass, // or the value at the end of the vectorized loop. - ResumeVal->addIncoming(II.StartValue, BypassBlock); + for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) { + if (OrigPhi == OldInduction) + ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]); + else + ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); + } ResumeVal->addIncoming(EndValue, VecBody); // Fix the scalar body counter (PHI node). unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); - OrigPhi->setIncomingValue(BlockIdx, ResumeVal); + // The old inductions phi node in the scalar body needs the truncated value. + if (OrigPhi == OldInduction) + OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal); + else + OrigPhi->setIncomingValue(BlockIdx, ResumeVal); } // If we are generating a new induction variable then we also need to @@ -609,7 +1565,8 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { assert(!ResumeIndex && "Unexpected resume value found"); ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", MiddleBlock->getTerminator()); - ResumeIndex->addIncoming(StartIdx, BypassBlock); + for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) + ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); } @@ -649,6 +1606,8 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { // Insert the new loop into the loop nest and register the new basic blocks. if (ParentLoop) { ParentLoop->addChildLoop(Lp); + for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) + ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase()); ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); @@ -665,57 +1624,202 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { LoopExitBlock = ExitBlock; LoopVectorBody = VecBody; LoopScalarBody = OldBasicBlock; - LoopBypassBlock = BypassBlock; } /// This function returns the identity element (or neutral element) for /// the operation K. -static unsigned -getReductionIdentity(LoopVectorizationLegality::ReductionKind K) { +Constant* +LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { switch (K) { - case LoopVectorizationLegality::IntegerXor: - case LoopVectorizationLegality::IntegerAdd: - case LoopVectorizationLegality::IntegerOr: + case RK_IntegerXor: + case RK_IntegerAdd: + case RK_IntegerOr: // Adding, Xoring, Oring zero to a number does not change it. - return 0; - case LoopVectorizationLegality::IntegerMult: + return ConstantInt::get(Tp, 0); + case RK_IntegerMult: // Multiplying a number by 1 does not change it. - return 1; - case LoopVectorizationLegality::IntegerAnd: + return ConstantInt::get(Tp, 1); + case RK_IntegerAnd: // AND-ing a number with an all-1 value does not change it. - return -1; + return ConstantInt::get(Tp, -1, true); + case RK_FloatMult: + // Multiplying a number by 1 does not change it. + return ConstantFP::get(Tp, 1.0L); + case RK_FloatAdd: + // Adding zero to a number does not change it. + return ConstantFP::get(Tp, 0.0L); default: llvm_unreachable("Unknown reduction kind"); } } -static bool -isTriviallyVectorizableIntrinsic(Instruction *Inst) { - IntrinsicInst *II = dyn_cast(Inst); - if (!II) - return false; - switch (II->getIntrinsicID()) { - case Intrinsic::sqrt: - case Intrinsic::sin: - case Intrinsic::cos: - case Intrinsic::exp: - case Intrinsic::exp2: - case Intrinsic::log: - case Intrinsic::log10: - case Intrinsic::log2: - case Intrinsic::fabs: - case Intrinsic::floor: - case Intrinsic::ceil: - case Intrinsic::trunc: - case Intrinsic::rint: - case Intrinsic::nearbyint: - case Intrinsic::pow: - case Intrinsic::fma: - return true; +static Intrinsic::ID +getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { + // If we have an intrinsic call, check if it is trivially vectorizable. + if (IntrinsicInst *II = dyn_cast(CI)) { + switch (II->getIntrinsicID()) { + case Intrinsic::sqrt: + case Intrinsic::sin: + case Intrinsic::cos: + case Intrinsic::exp: + case Intrinsic::exp2: + case Intrinsic::log: + case Intrinsic::log10: + case Intrinsic::log2: + case Intrinsic::fabs: + case Intrinsic::floor: + case Intrinsic::ceil: + case Intrinsic::trunc: + case Intrinsic::rint: + case Intrinsic::nearbyint: + case Intrinsic::pow: + case Intrinsic::fma: + case Intrinsic::fmuladd: + return II->getIntrinsicID(); + default: + return Intrinsic::not_intrinsic; + } + } + + if (!TLI) + return Intrinsic::not_intrinsic; + + LibFunc::Func Func; + Function *F = CI->getCalledFunction(); + // We're going to make assumptions on the semantics of the functions, check + // that the target knows that it's available in this environment. + if (!F || !TLI->getLibFunc(F->getName(), Func)) + return Intrinsic::not_intrinsic; + + // Otherwise check if we have a call to a function that can be turned into a + // vector intrinsic. + switch (Func) { default: - return false; + break; + case LibFunc::sin: + case LibFunc::sinf: + case LibFunc::sinl: + return Intrinsic::sin; + case LibFunc::cos: + case LibFunc::cosf: + case LibFunc::cosl: + return Intrinsic::cos; + case LibFunc::exp: + case LibFunc::expf: + case LibFunc::expl: + return Intrinsic::exp; + case LibFunc::exp2: + case LibFunc::exp2f: + case LibFunc::exp2l: + return Intrinsic::exp2; + case LibFunc::log: + case LibFunc::logf: + case LibFunc::logl: + return Intrinsic::log; + case LibFunc::log10: + case LibFunc::log10f: + case LibFunc::log10l: + return Intrinsic::log10; + case LibFunc::log2: + case LibFunc::log2f: + case LibFunc::log2l: + return Intrinsic::log2; + case LibFunc::fabs: + case LibFunc::fabsf: + case LibFunc::fabsl: + return Intrinsic::fabs; + case LibFunc::floor: + case LibFunc::floorf: + case LibFunc::floorl: + return Intrinsic::floor; + case LibFunc::ceil: + case LibFunc::ceilf: + case LibFunc::ceill: + return Intrinsic::ceil; + case LibFunc::trunc: + case LibFunc::truncf: + case LibFunc::truncl: + return Intrinsic::trunc; + case LibFunc::rint: + case LibFunc::rintf: + case LibFunc::rintl: + return Intrinsic::rint; + case LibFunc::nearbyint: + case LibFunc::nearbyintf: + case LibFunc::nearbyintl: + return Intrinsic::nearbyint; + case LibFunc::pow: + case LibFunc::powf: + case LibFunc::powl: + return Intrinsic::pow; } - return false; + + return Intrinsic::not_intrinsic; +} + +/// This function translates the reduction kind to an LLVM binary operator. +static unsigned +getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { + switch (Kind) { + case LoopVectorizationLegality::RK_IntegerAdd: + return Instruction::Add; + case LoopVectorizationLegality::RK_IntegerMult: + return Instruction::Mul; + case LoopVectorizationLegality::RK_IntegerOr: + return Instruction::Or; + case LoopVectorizationLegality::RK_IntegerAnd: + return Instruction::And; + case LoopVectorizationLegality::RK_IntegerXor: + return Instruction::Xor; + case LoopVectorizationLegality::RK_FloatMult: + return Instruction::FMul; + case LoopVectorizationLegality::RK_FloatAdd: + return Instruction::FAdd; + case LoopVectorizationLegality::RK_IntegerMinMax: + return Instruction::ICmp; + case LoopVectorizationLegality::RK_FloatMinMax: + return Instruction::FCmp; + default: + llvm_unreachable("Unknown reduction operation"); + } +} + +Value *createMinMaxOp(IRBuilder<> &Builder, + LoopVectorizationLegality::MinMaxReductionKind RK, + Value *Left, + Value *Right) { + CmpInst::Predicate P = CmpInst::ICMP_NE; + switch (RK) { + default: + llvm_unreachable("Unknown min/max reduction kind"); + case LoopVectorizationLegality::MRK_UIntMin: + P = CmpInst::ICMP_ULT; + break; + case LoopVectorizationLegality::MRK_UIntMax: + P = CmpInst::ICMP_UGT; + break; + case LoopVectorizationLegality::MRK_SIntMin: + P = CmpInst::ICMP_SLT; + break; + case LoopVectorizationLegality::MRK_SIntMax: + P = CmpInst::ICMP_SGT; + break; + case LoopVectorizationLegality::MRK_FloatMin: + P = CmpInst::FCMP_OLT; + break; + case LoopVectorizationLegality::MRK_FloatMax: + P = CmpInst::FCMP_OGT; + break; + } + + Value *Cmp; + if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax) + Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); + else + Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); + + Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); + return Select; } void @@ -727,9 +1831,7 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { // the cost-model. // //===------------------------------------------------===// - BasicBlock &BB = *OrigLoop->getHeader(); - Constant *Zero = - ConstantInt::get(IntegerType::getInt32Ty(BB.getContext()), 0); + Constant *Zero = Builder.getInt32(0); // In order to support reduction variables we need to be able to vectorize // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two @@ -763,7 +1865,6 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); it != e; ++it) { PHINode *RdxPhi = *it; - PHINode *VecRdxPhi = dyn_cast(WidenMap[RdxPhi]); assert(RdxPhi && "Unable to recover vectorized PHI"); // Find the reduction variable descriptor. @@ -776,21 +1877,32 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { // To do so, we need to generate the 'identity' vector and overide // one of the elements with the incoming scalar reduction. We need // to do it in the vector-loop preheader. - Builder.SetInsertPoint(LoopBypassBlock->getTerminator()); + Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator()); // This is the vector-clone of the value that leaves the loop. - Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr); - Type *VecTy = VectorExit->getType(); + VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); + Type *VecTy = VectorExit[0]->getType(); // Find the reduction identity variable. Zero for addition, or, xor, // one for multiplication, -1 for And. - Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind), - VecTy->getScalarType()); - - // This vector is the Identity vector where the first element is the - // incoming scalar reduction. - Value *VectorStart = Builder.CreateInsertElement(Identity, - RdxDesc.StartValue, Zero); + Value *Identity; + Value *VectorStart; + if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || + RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { + // MinMax reduction have the start value as their identify. + VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue, + "minmax.ident"); + } else { + Constant *Iden = + LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, + VecTy->getScalarType()); + Identity = ConstantVector::getSplat(VF, Iden); + + // This vector is the Identity vector where the first element is the + // incoming scalar reduction. + VectorStart = Builder.CreateInsertElement(Identity, + RdxDesc.StartValue, Zero); + } // Fix the vector-loop phi. // We created the induction variable so we know that the @@ -799,10 +1911,17 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { // Reductions do not have to start at zero. They can start with // any loop invariant values. - VecRdxPhi->addIncoming(VectorStart, VecPreheader); - Value *Val = - getVectorValue(RdxPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); - VecRdxPhi->addIncoming(Val, LoopVectorBody); + VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); + BasicBlock *Latch = OrigLoop->getLoopLatch(); + Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); + VectorParts &Val = getVectorValue(LoopVal); + for (unsigned part = 0; part < UF; ++part) { + // Make sure to add the reduction stat value only to the + // first unroll part. + Value *StartVal = (part == 0) ? VectorStart : Identity; + cast(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); + cast(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); + } // Before each round, move the insertion point right between // the PHIs and the values we are going to write. @@ -810,18 +1929,38 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { // instructions. Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); - // This PHINode contains the vectorized reduction variable, or - // the initial value vector, if we bypass the vector loop. - PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); - NewPhi->addIncoming(VectorStart, LoopBypassBlock); - NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody); + VectorParts RdxParts; + for (unsigned part = 0; part < UF; ++part) { + // This PHINode contains the vectorized reduction variable, or + // the initial value vector, if we bypass the vector loop. + VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); + PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); + Value *StartVal = (part == 0) ? VectorStart : Identity; + for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) + NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); + NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); + RdxParts.push_back(NewPhi); + } + + // Reduce all of the unrolled parts into a single vector. + Value *ReducedPartRdx = RdxParts[0]; + unsigned Op = getReductionBinOp(RdxDesc.Kind); + for (unsigned part = 1; part < UF; ++part) { + if (Op != Instruction::ICmp && Op != Instruction::FCmp) + ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, + RdxParts[part], ReducedPartRdx, + "bin.rdx"); + else + ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, + ReducedPartRdx, RdxParts[part]); + } // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles // and vector ops, reducing the set of values being computed by half each // round. assert(isPowerOf2_32(VF) && "Reduction emission only supported for pow2 vectors!"); - Value *TmpVec = NewPhi; + Value *TmpVec = ReducedPartRdx; SmallVector ShuffleMask(VF, 0); for (unsigned i = VF; i != 1; i >>= 1) { // Move the upper half of the vector to the lower half. @@ -838,26 +1977,11 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { ConstantVector::get(ShuffleMask), "rdx.shuf"); - // Emit the operation on the shuffled value. - switch (RdxDesc.Kind) { - case LoopVectorizationLegality::IntegerAdd: - TmpVec = Builder.CreateAdd(TmpVec, Shuf, "add.rdx"); - break; - case LoopVectorizationLegality::IntegerMult: - TmpVec = Builder.CreateMul(TmpVec, Shuf, "mul.rdx"); - break; - case LoopVectorizationLegality::IntegerOr: - TmpVec = Builder.CreateOr(TmpVec, Shuf, "or.rdx"); - break; - case LoopVectorizationLegality::IntegerAnd: - TmpVec = Builder.CreateAnd(TmpVec, Shuf, "and.rdx"); - break; - case LoopVectorizationLegality::IntegerXor: - TmpVec = Builder.CreateXor(TmpVec, Shuf, "xor.rdx"); - break; - default: - llvm_unreachable("Unknown reduction operation"); - } + if (Op != Instruction::ICmp && Op != Instruction::FCmp) + TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, + "bin.rdx"); + else + TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); } // The result is in the first element of the vector. @@ -895,29 +2019,49 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); }// end of for each redux variable. + + // The Loop exit block may have single value PHI nodes where the incoming + // value is 'undef'. While vectorizing we only handled real values that + // were defined inside the loop. Here we handle the 'undef case'. + // See PR14725. + for (BasicBlock::iterator LEI = LoopExitBlock->begin(), + LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { + PHINode *LCSSAPhi = dyn_cast(LEI); + if (!LCSSAPhi) continue; + if (LCSSAPhi->getNumIncomingValues() == 1) + LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), + LoopMiddleBlock); + } } -Value *InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { +InnerLoopVectorizer::VectorParts +InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && "Invalid edge"); - Value *SrcMask = createBlockInMask(Src); + VectorParts SrcMask = createBlockInMask(Src); // The terminator has to be a branch inst! BranchInst *BI = dyn_cast(Src->getTerminator()); assert(BI && "Unexpected terminator found"); - Value *EdgeMask = SrcMask; if (BI->isConditional()) { - EdgeMask = getVectorValue(BI->getCondition()); + VectorParts EdgeMask = getVectorValue(BI->getCondition()); + if (BI->getSuccessor(0) != Dst) - EdgeMask = Builder.CreateNot(EdgeMask); + for (unsigned part = 0; part < UF; ++part) + EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); + + for (unsigned part = 0; part < UF; ++part) + EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); + return EdgeMask; } - return Builder.CreateAnd(EdgeMask, SrcMask); + return SrcMask; } -Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { +InnerLoopVectorizer::VectorParts +InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); // Loop incoming mask is all-one. @@ -928,11 +2072,14 @@ Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { // This is the block mask. We OR all incoming edges, and with zero. Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); - Value *BlockMask = getVectorValue(Zero); + VectorParts BlockMask = getVectorValue(Zero); // For each pred: - for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) - BlockMask = Builder.CreateOr(BlockMask, createEdgeMask(*it, BB)); + for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { + VectorParts EM = createEdgeMask(*it, BB); + for (unsigned part = 0; part < UF; ++part) + BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); + } return BlockMask; } @@ -940,11 +2087,9 @@ Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { void InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, PhiVector *PV) { - Constant *Zero = - ConstantInt::get(IntegerType::getInt32Ty(BB->getContext()), 0); - // For each instruction in the old loop. for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { + VectorParts &Entry = WidenMap.get(it); switch (it->getOpcode()) { case Instruction::Br: // Nothing to do for PHIs and BR, since we already took care of the @@ -954,11 +2099,12 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, PHINode* P = cast(it); // Handle reduction variables: if (Legal->getReductionVars()->count(P)) { - // This is phase one of vectorizing PHIs. - Type *VecTy = VectorType::get(it->getType(), VF); - WidenMap[it] = - PHINode::Create(VecTy, 2, "vec.phi", - LoopVectorBody->getFirstInsertionPt()); + for (unsigned part = 0; part < UF; ++part) { + // This is phase one of vectorizing PHIs. + Type *VecTy = VectorType::get(it->getType(), VF); + Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", + LoopVectorBody-> getFirstInsertionPt()); + } PV->push_back(P); continue; } @@ -968,16 +2114,34 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, // 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. - Value *Cond = createEdgeMask(P->getIncomingBlock(0), P->getParent()); - WidenMap[P] = - Builder.CreateSelect(Cond, - getVectorValue(P->getIncomingValue(0)), - getVectorValue(P->getIncomingValue(1)), - "predphi"); + + unsigned NumIncoming = P->getNumIncomingValues(); + assert(NumIncoming > 1 && "Invalid PHI"); + + // Generate a sequence of selects of the form: + // SELECT(Mask3, In3, + // SELECT(Mask2, In2, + // ( ...))) + for (unsigned In = 0; In < NumIncoming; In++) { + VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), + P->getParent()); + VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); + + for (unsigned part = 0; part < UF; ++part) { + // We don't need to 'select' the first PHI operand because it is + // the default value if all of the other masks don't match. + if (In == 0) + Entry[part] = In0[part]; + else + // Select between the current value and the previous incoming edge + // based on the incoming mask. + Entry[part] = Builder.CreateSelect(Cond[part], In0[part], + Entry[part], "predphi"); + } + } continue; } @@ -990,36 +2154,43 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, Legal->getInductionVars()->lookup(P); switch (II.IK) { - case LoopVectorizationLegality::NoInduction: + case LoopVectorizationLegality::IK_NoInduction: llvm_unreachable("Unknown induction"); - case LoopVectorizationLegality::IntInduction: { - assert(P == OldInduction && "Unexpected PHI"); - Value *Broadcasted = getBroadcastInstrs(Induction); - // After broadcasting the induction variable we need to make the - // vector consecutive by adding 0, 1, 2 ... - Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted); - WidenMap[OldInduction] = ConsecutiveInduction; + case LoopVectorizationLegality::IK_IntInduction: { + assert(P->getType() == II.StartValue->getType() && "Types must match"); + Type *PhiTy = P->getType(); + Value *Broadcasted; + if (P == OldInduction) { + // Handle the canonical induction variable. We might have had to + // extend the type. + Broadcasted = Builder.CreateTrunc(Induction, PhiTy); + } else { + // Handle other induction variables that are now based on the + // canonical one. + Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, + "normalized.idx"); + NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); + Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx, + "offset.idx"); + } + Broadcasted = getBroadcastInstrs(Broadcasted); + // After broadcasting the induction variable we need to make the vector + // consecutive by adding 0, 1, 2, etc. + for (unsigned part = 0; part < UF; ++part) + Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); continue; } - case LoopVectorizationLegality::ReverseIntInduction: - case LoopVectorizationLegality::PtrInduction: + case LoopVectorizationLegality::IK_ReverseIntInduction: + case LoopVectorizationLegality::IK_PtrInduction: + case LoopVectorizationLegality::IK_ReversePtrInduction: // Handle reverse integer and pointer inductions. - Value *StartIdx = 0; - // If we have a single integer induction variable then use it. - // Otherwise, start counting at zero. - if (OldInduction) { - LoopVectorizationLegality::InductionInfo OldII = - Legal->getInductionVars()->lookup(OldInduction); - StartIdx = OldII.StartValue; - } else { - StartIdx = ConstantInt::get(Induction->getType(), 0); - } + Value *StartIdx = ExtendedIdx; // This is the normalized GEP that starts counting at zero. Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, "normalized.idx"); // Handle the reverse integer induction variable case. - if (LoopVectorizationLegality::ReverseIntInduction == II.IK) { + if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { IntegerType *DstTy = cast(II.StartValue->getType()); Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, "resize.norm.idx"); @@ -1030,30 +2201,40 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, Value *Broadcasted = getBroadcastInstrs(ReverseInd); // After broadcasting the induction variable we need to make the // vector consecutive by adding ... -3, -2, -1, 0. - Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted, - true); - WidenMap[it] = ConsecutiveInduction; + for (unsigned part = 0; part < UF; ++part) + Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part, + true); continue; } // Handle the pointer induction variable case. assert(P->getType()->isPointerTy() && "Unexpected type."); + // Is this a reverse induction ptr or a consecutive induction ptr. + bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == + II.IK); + // This is the vector of results. Notice that we don't generate // vector geps because scalar geps result in better code. - Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); - for (unsigned int i = 0; i < VF; ++i) { - Constant *Idx = ConstantInt::get(Induction->getType(), i); - Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, - "gep.idx"); - Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, - "next.gep"); - VecVal = Builder.CreateInsertElement(VecVal, SclrGep, - Builder.getInt32(i), - "insert.gep"); + for (unsigned part = 0; part < UF; ++part) { + Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); + for (unsigned int i = 0; i < VF; ++i) { + int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); + Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); + Value *GlobalIdx; + if (!Reverse) + GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); + else + GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); + + Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, + "next.gep"); + VecVal = Builder.CreateInsertElement(VecVal, SclrGep, + Builder.getInt32(i), + "insert.gep"); + } + Entry[part] = VecVal; } - - WidenMap[it] = VecVal; continue; } @@ -1079,41 +2260,48 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, case Instruction::Xor: { // Just widen binops. BinaryOperator *BinOp = dyn_cast(it); - Value *A = getVectorValue(it->getOperand(0)); - Value *B = getVectorValue(it->getOperand(1)); + VectorParts &A = getVectorValue(it->getOperand(0)); + VectorParts &B = getVectorValue(it->getOperand(1)); // Use this vector value for all users of the original instruction. - Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B); - WidenMap[it] = V; - - // Update the NSW, NUW and Exact flags. - BinaryOperator *VecOp = cast(V); - if (isa(BinOp)) { - VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); - VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); + for (unsigned Part = 0; Part < UF; ++Part) { + Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); + + // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. + BinaryOperator *VecOp = dyn_cast(V); + if (VecOp && isa(BinOp)) { + VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); + VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); + } + if (VecOp && isa(VecOp)) + VecOp->setIsExact(BinOp->isExact()); + + Entry[Part] = V; } - if (isa(VecOp)) - VecOp->setIsExact(BinOp->isExact()); break; } case Instruction::Select: { // Widen selects. // If the selector is loop invariant we can create a select // instruction with a scalar condition. Otherwise, use vector-select. - Value *Cond = it->getOperand(0); - bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop); + bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), + OrigLoop); // 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. - Cond = getVectorValue(Cond); - if (InvariantCond) - Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0)); - - Value *Op0 = getVectorValue(it->getOperand(1)); - Value *Op1 = getVectorValue(it->getOperand(2)); - WidenMap[it] = Builder.CreateSelect(Cond, Op0, Op1); + VectorParts &Cond = getVectorValue(it->getOperand(0)); + VectorParts &Op0 = getVectorValue(it->getOperand(1)); + VectorParts &Op1 = getVectorValue(it->getOperand(2)); + Value *ScalarCond = Builder.CreateExtractElement(Cond[0], + Builder.getInt32(0)); + for (unsigned Part = 0; Part < UF; ++Part) { + Entry[Part] = Builder.CreateSelect( + InvariantCond ? ScalarCond : Cond[Part], + Op0[Part], + Op1[Part]); + } break; } @@ -1122,94 +2310,23 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, // Widen compares. Generate vector compares. bool FCmp = (it->getOpcode() == Instruction::FCmp); CmpInst *Cmp = dyn_cast(it); - Value *A = getVectorValue(it->getOperand(0)); - Value *B = getVectorValue(it->getOperand(1)); - if (FCmp) - WidenMap[it] = Builder.CreateFCmp(Cmp->getPredicate(), A, B); - else - WidenMap[it] = Builder.CreateICmp(Cmp->getPredicate(), A, B); - break; - } - - case Instruction::Store: { - // Attempt to issue a wide store. - StoreInst *SI = dyn_cast(it); - Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF); - Value *Ptr = SI->getPointerOperand(); - unsigned Alignment = SI->getAlignment(); - - assert(!Legal->isUniform(Ptr) && - "We do not allow storing to uniform addresses"); - - GetElementPtrInst *Gep = dyn_cast(Ptr); - - // This store does not use GEPs. - if (!Legal->isConsecutivePtr(Ptr)) { - scalarizeInstruction(it); - break; - } - - if (Gep) { - // The last index does not have to be the induction. It can be - // consecutive and be a function of the index. For example A[I+1]; - unsigned NumOperands = Gep->getNumOperands(); - Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1)); - LastIndex = Builder.CreateExtractElement(LastIndex, Zero); - - // Create the new GEP with the new induction variable. - GetElementPtrInst *Gep2 = cast(Gep->clone()); - Gep2->setOperand(NumOperands - 1, LastIndex); - Ptr = Builder.Insert(Gep2); - } else { - // Use the induction element ptr. - assert(isa(Ptr) && "Invalid induction ptr"); - Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero); + VectorParts &A = getVectorValue(it->getOperand(0)); + VectorParts &B = getVectorValue(it->getOperand(1)); + for (unsigned Part = 0; Part < UF; ++Part) { + Value *C = 0; + if (FCmp) + C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); + else + C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); + Entry[Part] = C; } - Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo()); - Value *Val = getVectorValue(SI->getValueOperand()); - Builder.CreateStore(Val, Ptr)->setAlignment(Alignment); break; } - case Instruction::Load: { - // Attempt to issue a wide load. - LoadInst *LI = dyn_cast(it); - Type *RetTy = VectorType::get(LI->getType(), VF); - Value *Ptr = LI->getPointerOperand(); - unsigned Alignment = LI->getAlignment(); - GetElementPtrInst *Gep = dyn_cast(Ptr); - - // If the pointer is loop invariant or if it is non consecutive, - // scalarize the load. - bool Con = Legal->isConsecutivePtr(Ptr); - if (Legal->isUniform(Ptr) || !Con) { - scalarizeInstruction(it); - break; - } - - if (Gep) { - // The last index does not have to be the induction. It can be - // consecutive and be a function of the index. For example A[I+1]; - unsigned NumOperands = Gep->getNumOperands(); - Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1)); - LastIndex = Builder.CreateExtractElement(LastIndex, Zero); - - // Create the new GEP with the new induction variable. - GetElementPtrInst *Gep2 = cast(Gep->clone()); - Gep2->setOperand(NumOperands - 1, LastIndex); - Ptr = Builder.Insert(Gep2); - } else { - // Use the induction element ptr. - assert(isa(Ptr) && "Invalid induction ptr"); - Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero); - } - Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo()); - LI = Builder.CreateLoad(Ptr); - LI->setAlignment(Alignment); - // Use this vector value for all users of the load. - WidenMap[it] = LI; - break; - } + case Instruction::Store: + case Instruction::Load: + vectorizeMemoryInstruction(it, Legal); + break; case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: @@ -1232,27 +2349,38 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, CI->getType()); Value *Broadcasted = getBroadcastInstrs(ScalarCast); - WidenMap[it] = getConsecutiveVector(Broadcasted); + for (unsigned Part = 0; Part < UF; ++Part) + Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); break; } /// Vectorize casts. - Value *A = getVectorValue(it->getOperand(0)); Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); - WidenMap[it] = Builder.CreateCast(CI->getOpcode(), A, DestTy); + + VectorParts &A = getVectorValue(it->getOperand(0)); + for (unsigned Part = 0; Part < UF; ++Part) + Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); break; } case Instruction::Call: { - assert(isTriviallyVectorizableIntrinsic(it)); + // Ignore dbg intrinsics. + if (isa(it)) + break; + Module *M = BB->getParent()->getParent(); - IntrinsicInst *II = cast(it); - Intrinsic::ID ID = II->getIntrinsicID(); - SmallVector Args; - for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) - Args.push_back(getVectorValue(II->getArgOperand(i))); - Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) }; - Function *F = Intrinsic::getDeclaration(M, ID, Tys); - WidenMap[it] = Builder.CreateCall(F, Args); + CallInst *CI = cast(it); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); + assert(ID && "Not an intrinsic call!"); + for (unsigned Part = 0; Part < UF; ++Part) { + SmallVector Args; + for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { + VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); + Args.push_back(Arg[Part]); + } + Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) }; + Function *F = Intrinsic::getDeclaration(M, ID, Tys); + Entry[Part] = Builder.CreateCall(F, Args); + } break; } @@ -1269,12 +2397,14 @@ void InnerLoopVectorizer::updateAnalysis() { SE->forgetLoop(OrigLoop); // Update the dominator tree information. - assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) && + assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && "Entry does not dominate exit."); - DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock); + for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) + DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); + DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); - DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock); + DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); @@ -1297,12 +2427,6 @@ bool LoopVectorizationLegality::canVectorizeWithIfConvert() { if (!isa(BB->getTerminator())) return false; - // We must have at most two predecessors because we need to convert - // all PHIs to selects. - unsigned Preds = std::distance(pred_begin(BB), pred_end(BB)); - if (Preds > 2) - return false; - // We must be able to predicate all blocks that need to be predicated. if (blockNeedsPredication(BB) && !blockCanBePredicated(BB)) return false; @@ -1349,7 +2473,7 @@ bool LoopVectorizationLegality::canVectorize() { // Do not loop-vectorize loops with a tiny trip count. unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); - if (TC > 0u && TC < TinyTripCountThreshold) { + if (TC > 0u && TC < TinyTripCountVectorThreshold) { DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is not worth vectorizing.\n"); return false; @@ -1380,10 +2504,38 @@ bool LoopVectorizationLegality::canVectorize() { return true; } +static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) { + if (Ty->isPointerTy()) + return DL.getIntPtrType(Ty->getContext()); + return Ty; +} + +static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) { + Ty0 = convertPointerToIntegerType(DL, Ty0); + Ty1 = convertPointerToIntegerType(DL, Ty1); + if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) + return Ty0; + return Ty1; +} + bool LoopVectorizationLegality::canVectorizeInstrs() { BasicBlock *PreHeader = TheLoop->getLoopPreheader(); BasicBlock *Header = TheLoop->getHeader(); + // If we marked the scalar loop as "already vectorized" then no need + // to vectorize it again. + if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) { + DEBUG(dbgs() << "LV: This loop was vectorized before\n"); + return false; + } + + // Look for the attribute signaling the absence of NaNs. + Function &F = *Header->getParent(); + if (F.hasFnAttribute("no-nans-fp-math")) + HasFunNoNaNAttr = F.getAttributes().getAttribute( + AttributeSet::FunctionIndex, + "no-nans-fp-math").getValueAsString() == "true"; + // For each block in the loop. for (Loop::block_iterator bb = TheLoop->block_begin(), be = TheLoop->block_end(); bb != be; ++bb) { @@ -1393,15 +2545,11 @@ bool LoopVectorizationLegality::canVectorizeInstrs() { ++it) { if (PHINode *Phi = dyn_cast(it)) { - // This should not happen because the loop should be normalized. - if (Phi->getNumIncomingValues() != 2) { - DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); - return false; - } - + Type *PhiTy = Phi->getType(); // Check that this PHI type is allowed. - if (!Phi->getType()->isIntegerTy() && - !Phi->getType()->isPointerTy()) { + if (!PhiTy->isIntegerTy() && + !PhiTy->isFloatingPointTy() && + !PhiTy->isPointerTy()) { DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); return false; } @@ -1412,19 +2560,31 @@ bool LoopVectorizationLegality::canVectorizeInstrs() { if (*bb != Header) continue; + // We only allow if-converted PHIs with more than two incoming values. + if (Phi->getNumIncomingValues() != 2) { + DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); + return false; + } + // This is the value coming from the preheader. Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); // Check if this is an induction variable. InductionKind IK = isInductionVariable(Phi); - if (NoInduction != IK) { + if (IK_NoInduction != IK) { + // Get the widest type. + if (!WidestIndTy) + WidestIndTy = convertPointerToIntegerType(*DL, PhiTy); + else + WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy); + // Int inductions are special because we only allow one IV. - if (IK == IntInduction) { - if (Induction) { - DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); - return false; - } - Induction = Phi; + if (IK == IK_IntInduction) { + // Use the phi node with the widest type as induction. Use the last + // one if there are multiple (no good reason for doing this other + // than it is expedient). + if (!Induction || PhiTy == WidestIndTy) + Induction = Phi; } DEBUG(dbgs() << "LV: Found an induction variable.\n"); @@ -1432,34 +2592,51 @@ bool LoopVectorizationLegality::canVectorizeInstrs() { continue; } - if (AddReductionVar(Phi, IntegerAdd)) { + if (AddReductionVar(Phi, RK_IntegerAdd)) { DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); continue; } - if (AddReductionVar(Phi, IntegerMult)) { + if (AddReductionVar(Phi, RK_IntegerMult)) { DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); continue; } - if (AddReductionVar(Phi, IntegerOr)) { + if (AddReductionVar(Phi, RK_IntegerOr)) { DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); continue; } - if (AddReductionVar(Phi, IntegerAnd)) { + if (AddReductionVar(Phi, RK_IntegerAnd)) { DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); continue; } - if (AddReductionVar(Phi, IntegerXor)) { + if (AddReductionVar(Phi, RK_IntegerXor)) { DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); continue; } + if (AddReductionVar(Phi, RK_IntegerMinMax)) { + DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, RK_FloatMult)) { + DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, RK_FloatAdd)) { + DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, RK_FloatMinMax)) { + DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n"); + continue; + } DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); return false; }// end of PHI handling - // We still don't handle functions. + // We still don't handle functions. However, we can ignore dbg intrinsic + // calls and we do handle certain intrinsic and libm functions. CallInst *CI = dyn_cast(it); - if (CI && !isTriviallyVectorizableIntrinsic(it)) { + if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa(CI)) { DEBUG(dbgs() << "LV: Found a call site.\n"); return false; } @@ -1497,7 +2674,8 @@ bool LoopVectorizationLegality::canVectorizeInstrs() { if (!Induction) { DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); - assert(getInductionVars()->size() && "No induction variables"); + if (Inductions.empty()) + return false; } return true; @@ -1532,7 +2710,44 @@ void LoopVectorizationLegality::collectLoopUniforms() { } } +AliasAnalysis::Location +LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) { + if (StoreInst *Store = dyn_cast(Inst)) + return AA->getLocation(Store); + else if (LoadInst *Load = dyn_cast(Inst)) + return AA->getLocation(Load); + + llvm_unreachable("Should be either load or store instruction"); +} + +bool +LoopVectorizationLegality::hasPossibleGlobalWriteReorder( + Value *Object, + Instruction *Inst, + AliasMultiMap& WriteObjects, + unsigned MaxByteWidth) { + + AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst); + + std::vector::iterator + it = WriteObjects[Object].begin(), + end = WriteObjects[Object].end(); + + for (; it != end; ++it) { + Instruction* I = *it; + if (I == Inst) + continue; + + AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I); + if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth), + ThatLoc.getWithNewSize(MaxByteWidth))) + return true; + } + return false; +} + bool LoopVectorizationLegality::canVectorizeMemory() { + typedef SmallVector ValueVector; typedef SmallPtrSet ValueSet; // Holds the Load and Store *instructions*. @@ -1541,6 +2756,8 @@ bool LoopVectorizationLegality::canVectorizeMemory() { PtrRtCheck.Pointers.clear(); PtrRtCheck.Need = false; + const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); + // For each block. for (Loop::block_iterator bb = TheLoop->block_begin(), be = TheLoop->block_end(); bb != be; ++bb) { @@ -1555,7 +2772,7 @@ bool LoopVectorizationLegality::canVectorizeMemory() { if (it->mayReadFromMemory()) { LoadInst *Ld = dyn_cast(it); if (!Ld) return false; - if (!Ld->isSimple()) { + if (!Ld->isSimple() && !IsAnnotatedParallel) { DEBUG(dbgs() << "LV: Found a non-simple load.\n"); return false; } @@ -1567,7 +2784,7 @@ bool LoopVectorizationLegality::canVectorizeMemory() { if (it->mayWriteToMemory()) { StoreInst *St = dyn_cast(it); if (!St) return false; - if (!St->isSimple()) { + if (!St->isSimple() && !IsAnnotatedParallel) { DEBUG(dbgs() << "LV: Found a non-simple store.\n"); return false; } @@ -1586,9 +2803,10 @@ bool LoopVectorizationLegality::canVectorizeMemory() { return true; } - // Holds the read and read-write *pointers* that we find. - ValueVector Reads; - ValueVector ReadWrites; + // Holds the read and read-write *pointers* that we find. These maps hold + // unique values for pointers (so no need for multi-map). + AliasMap Reads; + AliasMap ReadWrites; // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects // multiple times on the same object. If the ptr is accessed twice, once @@ -1610,7 +2828,14 @@ bool LoopVectorizationLegality::canVectorizeMemory() { // If we did *not* see this pointer before, insert it to // the read-write list. At this phase it is only a 'write' list. if (Seen.insert(Ptr)) - ReadWrites.push_back(Ptr); + ReadWrites.insert(std::make_pair(Ptr, ST)); + } + + if (IsAnnotatedParallel) { + DEBUG(dbgs() + << "LV: A loop annotated parallel, ignore memory dependency " + << "checks.\n"); + return true; } for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { @@ -1624,8 +2849,8 @@ bool LoopVectorizationLegality::canVectorizeMemory() { // If the address of i is unknown (for example A[B[i]]) then we may // read a few words, modify, and write a few words, and some of the // words may be written to the same address. - if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr)) - Reads.push_back(Ptr); + if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr)) + Reads.insert(std::make_pair(Ptr, LD)); } // If we write (or read-write) to a single destination and there are no @@ -1635,29 +2860,41 @@ bool LoopVectorizationLegality::canVectorizeMemory() { return true; } + unsigned NumReadPtrs = 0; + unsigned NumWritePtrs = 0; + // Find pointers with computable bounds. We are going to use this information // to place a runtime bound check. bool CanDoRT = true; - for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) - if (hasComputableBounds(*I)) { - PtrRtCheck.insert(SE, TheLoop, *I); - DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); + AliasMap::iterator MI, ME; + for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { + Value *V = (*MI).first; + if (hasComputableBounds(V)) { + PtrRtCheck.insert(SE, TheLoop, V, true); + NumWritePtrs++; + DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); } else { CanDoRT = false; break; } - for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) - if (hasComputableBounds(*I)) { - PtrRtCheck.insert(SE, TheLoop, *I); - DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); + } + for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { + Value *V = (*MI).first; + if (hasComputableBounds(V)) { + PtrRtCheck.insert(SE, TheLoop, V, false); + NumReadPtrs++; + DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); } else { CanDoRT = false; break; } + } // Check that we did not collect too many pointers or found a // unsizeable pointer. - if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) { + unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1)); + DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n"); + if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { PtrRtCheck.reset(); CanDoRT = false; } @@ -1668,43 +2905,104 @@ bool LoopVectorizationLegality::canVectorizeMemory() { bool NeedRTCheck = false; + // Biggest vectorized access possible, vector width * unroll factor. + // TODO: We're being very pessimistic here, find a way to know the + // real access width before getting here. + unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) * + TTI->getMaximumUnrollFactor(); // Now that the pointers are in two lists (Reads and ReadWrites), we // can check that there are no conflicts between each of the writes and // between the writes to the reads. - ValueSet WriteObjects; + // Note that WriteObjects duplicates the stores (indexed now by underlying + // objects) to avoid pointing to elements inside ReadWrites. + // TODO: Maybe create a new type where they can interact without duplication. + AliasMultiMap WriteObjects; ValueVector TempObjects; // Check that the read-writes do not conflict with other read-write // pointers. - for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) { - GetUnderlyingObjects(*I, TempObjects, DL); - for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); - it != e; ++it) { - if (!isIdentifiedObject(*it)) { - DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n"); + bool AllWritesIdentified = true; + for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { + Value *Val = (*MI).first; + Instruction *Inst = (*MI).second; + + GetUnderlyingObjects(Val, TempObjects, DL); + for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); + UI != UE; ++UI) { + if (!isIdentifiedObject(*UI)) { + DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n"); NeedRTCheck = true; + AllWritesIdentified = false; } - if (!WriteObjects.insert(*it)) { + + // Never seen it before, can't alias. + if (WriteObjects[*UI].empty()) { + DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n"); + WriteObjects[*UI].push_back(Inst); + continue; + } + // Direct alias found. + if (!AA || dyn_cast(*UI) == NULL) { DEBUG(dbgs() << "LV: Found a possible write-write reorder:" - << **it <<"\n"); + << **UI <<"\n"); return false; } + DEBUG(dbgs() << "LV: Found a conflicting global value:" + << **UI <<"\n"); + DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n"); + DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); + + // If global alias, make sure they do alias. + if (hasPossibleGlobalWriteReorder(*UI, + Inst, + WriteObjects, + MaxByteWidth)) { + DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI + << "\n"); + return false; + } + + // Didn't alias, insert into map for further reference. + WriteObjects[*UI].push_back(Inst); } TempObjects.clear(); } /// Check that the reads don't conflict with the read-writes. - for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) { - GetUnderlyingObjects(*I, TempObjects, DL); - for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); - it != e; ++it) { - if (!isIdentifiedObject(*it)) { - DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n"); + for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { + Value *Val = (*MI).first; + GetUnderlyingObjects(Val, TempObjects, DL); + for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); + UI != UE; ++UI) { + // If all of the writes are identified then we don't care if the read + // pointer is identified or not. + if (!AllWritesIdentified && !isIdentifiedObject(*UI)) { + DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n"); NeedRTCheck = true; } - if (WriteObjects.count(*it)) { - DEBUG(dbgs() << "LV: Found a possible read/write reorder:" - << **it <<"\n"); + + // Never seen it before, can't alias. + if (WriteObjects[*UI].empty()) + continue; + // Direct alias found. + if (!AA || dyn_cast(*UI) == NULL) { + DEBUG(dbgs() << "LV: Found a possible write-write reorder:" + << **UI <<"\n"); + return false; + } + DEBUG(dbgs() << "LV: Found a global value: " + << **UI <<"\n"); + Instruction *Inst = (*MI).second; + DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n"); + DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); + + // If global alias, make sure they do alias. + if (hasPossibleGlobalWriteReorder(*UI, + Inst, + WriteObjects, + MaxByteWidth)) { + DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI + << "\n"); return false; } } @@ -1724,6 +3022,26 @@ bool LoopVectorizationLegality::canVectorizeMemory() { return true; } +static bool hasMultipleUsesOf(Instruction *I, + SmallPtrSet &Insts) { + unsigned NumUses = 0; + for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) { + if (Insts.count(dyn_cast(*Use))) + ++NumUses; + if (NumUses > 1) + return true; + } + + return false; +} + +static bool areAllUsesIn(Instruction *I, SmallPtrSet &Set) { + for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) + if (!Set.count(dyn_cast(*Use))) + return false; + return true; +} + bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, ReductionKind Kind) { if (Phi->getNumIncomingValues() != 2) @@ -1742,102 +3060,234 @@ bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, // This includes users of the reduction, variables (which form a cycle // which ends in the phi node). Instruction *ExitInstruction = 0; + // Indicates that we found a reduction operation in our scan. + bool FoundReduxOp = false; + + // We start with the PHI node and scan for all of the users of this + // instruction. All users must be instructions that can be used as reduction + // variables (such as ADD). We must have a single out-of-block user. The cycle + // must include the original PHI. + bool FoundStartPHI = false; + + // To recognize min/max patterns formed by a icmp select sequence, we store + // the number of instruction we saw from the recognized min/max pattern, + // to make sure we only see exactly the two instructions. + unsigned NumCmpSelectPatternInst = 0; + ReductionInstDesc ReduxDesc(false, 0); + + SmallPtrSet VisitedInsts; + SmallVector Worklist; + Worklist.push_back(Phi); + VisitedInsts.insert(Phi); + + // A value in the reduction can be used: + // - By the reduction: + // - Reduction operation: + // - One use of reduction value (safe). + // - Multiple use of reduction value (not safe). + // - PHI: + // - All uses of the PHI must be the reduction (safe). + // - Otherwise, not safe. + // - By one instruction outside of the loop (safe). + // - By further instructions outside of the loop (not safe). + // - By an instruction that is not part of the reduction (not safe). + // This is either: + // * An instruction type other than PHI or the reduction operation. + // * A PHI in the header other than the initial PHI. + while (!Worklist.empty()) { + Instruction *Cur = Worklist.back(); + Worklist.pop_back(); - // Iter is our iterator. We start with the PHI node and scan for all of the - // users of this instruction. All users must be instructions that can be - // used as reduction variables (such as ADD). We may have a single - // out-of-block user. The cycle must end with the original PHI. - Instruction *Iter = Phi; - while (true) { - // If the instruction has no users then this is a broken - // chain and can't be a reduction variable. - if (Iter->use_empty()) + // No Users. + // If the instruction has no users then this is a broken chain and can't be + // a reduction variable. + if (Cur->use_empty()) return false; - // Any reduction instr must be of one of the allowed kinds. - if (!isReductionInstr(Iter, Kind)) + bool IsAPhi = isa(Cur); + + // A header PHI use other than the original PHI. + if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) return false; - // Did we find a user inside this loop already ? - bool FoundInBlockUser = false; - // Did we reach the initial PHI node already ? - bool FoundStartPHI = false; - - // For each of the *users* of iter. - for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end(); - it != e; ++it) { - Instruction *U = cast(*it); - // We already know that the PHI is a user. - if (U == Phi) { - FoundStartPHI = true; - continue; - } + // Reductions of instructions such as Div, and Sub is only possible if the + // LHS is the reduction variable. + if (!Cur->isCommutative() && !IsAPhi && !isa(Cur) && + !isa(Cur) && !isa(Cur) && + !VisitedInsts.count(dyn_cast(Cur->getOperand(0)))) + return false; + + // Any reduction instruction must be of one of the allowed kinds. + ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc); + if (!ReduxDesc.IsReduction) + return false; + + // A reduction operation must only have one use of the reduction value. + if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax && + hasMultipleUsesOf(Cur, VisitedInsts)) + return false; + + // All inputs to a PHI node must be a reduction value. + if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) + return false; + + if (Kind == RK_IntegerMinMax && (isa(Cur) || + isa(Cur))) + ++NumCmpSelectPatternInst; + if (Kind == RK_FloatMinMax && (isa(Cur) || + isa(Cur))) + ++NumCmpSelectPatternInst; + + // Check whether we found a reduction operator. + FoundReduxOp |= !IsAPhi; + + // Process users of current instruction. Push non PHI nodes after PHI nodes + // onto the stack. This way we are going to have seen all inputs to PHI + // nodes once we get to them. + SmallVector NonPHIs; + SmallVector PHIs; + for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E; + ++UI) { + Instruction *Usr = cast(*UI); // Check if we found the exit user. - BasicBlock *Parent = U->getParent(); + BasicBlock *Parent = Usr->getParent(); if (!TheLoop->contains(Parent)) { // Exit if you find multiple outside users. if (ExitInstruction != 0) return false; - ExitInstruction = Iter; - } - - // We allow in-loop PHINodes which are not the original reduction PHI - // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE - // structure) then don't skip this PHI. - if (isa(Iter) && isa(U) && - U->getParent() != TheLoop->getHeader() && - TheLoop->contains(U) && - Iter->getNumUses() > 1) + ExitInstruction = Cur; continue; + } - // We can't have multiple inside users. - if (FoundInBlockUser) - return false; - FoundInBlockUser = true; - Iter = U; + // Process instructions only once (termination). + if (VisitedInsts.insert(Usr)) { + if (isa(Usr)) + PHIs.push_back(Usr); + else + NonPHIs.push_back(Usr); + } + // Remember that we completed the cycle. + if (Usr == Phi) + FoundStartPHI = true; } + Worklist.append(PHIs.begin(), PHIs.end()); + Worklist.append(NonPHIs.begin(), NonPHIs.end()); + } + + // This means we have seen one but not the other instruction of the + // pattern or more than just a select and cmp. + if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && + NumCmpSelectPatternInst != 2) + return false; - // We found a reduction var if we have reached the original - // phi node and we only have a single instruction with out-of-loop - // users. - if (FoundStartPHI && ExitInstruction) { - // This instruction is allowed to have out-of-loop users. - AllowedExit.insert(ExitInstruction); + if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) + return false; - // Save the description of this reduction variable. - ReductionDescriptor RD(RdxStart, ExitInstruction, Kind); - Reductions[Phi] = RD; - return true; - } + // We found a reduction var if we have reached the original phi node and we + // only have a single instruction with out-of-loop users. - // If we've reached the start PHI but did not find an outside user then - // this is dead code. Abort. - if (FoundStartPHI) - return false; + // This instruction is allowed to have out-of-loop users. + AllowedExit.insert(ExitInstruction); + + // Save the description of this reduction variable. + ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, + ReduxDesc.MinMaxKind); + Reductions[Phi] = RD; + // We've ended the cycle. This is a reduction variable if we have an + // outside user and it has a binary op. + + return true; +} + +/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction +/// pattern corresponding to a min(X, Y) or max(X, Y). +LoopVectorizationLegality::ReductionInstDesc +LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, + ReductionInstDesc &Prev) { + + assert((isa(I) || isa(I) || isa(I)) && + "Expect a select instruction"); + Instruction *Cmp = 0; + SelectInst *Select = 0; + + // We must handle the select(cmp()) as a single instruction. Advance to the + // select. + if ((Cmp = dyn_cast(I)) || (Cmp = dyn_cast(I))) { + if (!Cmp->hasOneUse() || !(Select = dyn_cast(*I->use_begin()))) + return ReductionInstDesc(false, I); + return ReductionInstDesc(Select, Prev.MinMaxKind); } + + // Only handle single use cases for now. + if (!(Select = dyn_cast(I))) + return ReductionInstDesc(false, I); + if (!(Cmp = dyn_cast(I->getOperand(0))) && + !(Cmp = dyn_cast(I->getOperand(0)))) + return ReductionInstDesc(false, I); + if (!Cmp->hasOneUse()) + return ReductionInstDesc(false, I); + + Value *CmpLeft; + Value *CmpRight; + + // Look for a min/max pattern. + if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_UIntMin); + else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_UIntMax); + else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_SIntMax); + else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_SIntMin); + else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_FloatMin); + else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_FloatMax); + else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_FloatMin); + else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) + return ReductionInstDesc(Select, MRK_FloatMax); + + return ReductionInstDesc(false, I); } -bool +LoopVectorizationLegality::ReductionInstDesc LoopVectorizationLegality::isReductionInstr(Instruction *I, - ReductionKind Kind) { + ReductionKind Kind, + ReductionInstDesc &Prev) { + bool FP = I->getType()->isFloatingPointTy(); + bool FastMath = (FP && I->isCommutative() && I->isAssociative()); switch (I->getOpcode()) { default: - return false; + return ReductionInstDesc(false, I); case Instruction::PHI: - // possibly. - return true; - case Instruction::Add: + if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && + Kind != RK_FloatMinMax)) + return ReductionInstDesc(false, I); + return ReductionInstDesc(I, Prev.MinMaxKind); case Instruction::Sub: - return Kind == IntegerAdd; + case Instruction::Add: + return ReductionInstDesc(Kind == RK_IntegerAdd, I); case Instruction::Mul: - return Kind == IntegerMult; + return ReductionInstDesc(Kind == RK_IntegerMult, I); case Instruction::And: - return Kind == IntegerAnd; + return ReductionInstDesc(Kind == RK_IntegerAnd, I); case Instruction::Or: - return Kind == IntegerOr; + return ReductionInstDesc(Kind == RK_IntegerOr, I); case Instruction::Xor: - return Kind == IntegerXor; + return ReductionInstDesc(Kind == RK_IntegerXor, I); + case Instruction::FMul: + return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); + case Instruction::FAdd: + return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); + case Instruction::FCmp: + case Instruction::ICmp: + case Instruction::Select: + if (Kind != RK_IntegerMinMax && + (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) + return ReductionInstDesc(false, I); + return isMinMaxSelectCmpPattern(I, Prev); } } @@ -1846,37 +3296,39 @@ LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { Type *PhiTy = Phi->getType(); // We only handle integer and pointer inductions variables. if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) - return NoInduction; + return IK_NoInduction; - // Check that the PHI is consecutive and starts at zero. + // Check that the PHI is consecutive. const SCEV *PhiScev = SE->getSCEV(Phi); const SCEVAddRecExpr *AR = dyn_cast(PhiScev); if (!AR) { DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); - return NoInduction; + return IK_NoInduction; } const SCEV *Step = AR->getStepRecurrence(*SE); // Integer inductions need to have a stride of one. if (PhiTy->isIntegerTy()) { if (Step->isOne()) - return IntInduction; + return IK_IntInduction; if (Step->isAllOnesValue()) - return ReverseIntInduction; - return NoInduction; + return IK_ReverseIntInduction; + return IK_NoInduction; } // Calculate the pointer stride and check if it is consecutive. const SCEVConstant *C = dyn_cast(Step); if (!C) - return NoInduction; + return IK_NoInduction; assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); if (C->getValue()->equalsInt(Size)) - return PtrInduction; + return IK_PtrInduction; + else if (C->getValue()->equalsInt(0 - Size)) + return IK_ReversePtrInduction; - return NoInduction; + return IK_NoInduction; } bool LoopVectorizationLegality::isInductionVariable(const Value *V) { @@ -1898,8 +3350,12 @@ bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { - // We don't predicate loads/stores at the moment. - if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow()) + // We might be able to hoist the load. + if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it)) + return false; + + // We predicate stores at the moment. + if (it->mayWriteToMemory() || it->mayThrow()) return false; // The instructions below can trap. @@ -1913,6 +3369,10 @@ bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { } } + // Check that we can actually speculate the hoistable loads. + if (!LoadSpeculation.canHoistAllLoads()) + return false; + return true; } @@ -1925,18 +3385,34 @@ bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { return AR->isAffine(); } -unsigned +LoopVectorizationCostModel::VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, - unsigned UserVF) { + unsigned UserVF) { + // Width 1 means no vectorize + VectorizationFactor Factor = { 1U, 0U }; if (OptForSize && Legal->getRuntimePointerCheck()->Need) { DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); - return 1; + return Factor; } // Find the trip count. unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); DEBUG(dbgs() << "LV: Found trip count:"<block_begin(), + be = TheLoop->block_end(); bb != be; ++bb) { + BasicBlock *BB = *bb; + + // For each instruction in the loop. + for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { + Type *T = it->getType(); + + // Only examine Loads, Stores and PHINodes. + if (!isa(it) && !isa(it) && !isa(it)) + continue; + + // Examine PHI nodes that are reduction variables. + if (PHINode *PN = dyn_cast(it)) + if (!Legal->getReductionVars()->count(PN)) + continue; + + // Examine the stored values. + if (StoreInst *ST = dyn_cast(it)) + T = ST->getValueOperand()->getType(); + + // Ignore loaded pointer types and stored pointer types that are not + // consecutive. However, we do want to take consecutive stores/loads of + // pointer vectors into account. + if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) + continue; + + MaxWidth = std::max(MaxWidth, + (unsigned)DL->getTypeSizeInBits(T->getScalarType())); + } + } + + return MaxWidth; +} + +unsigned +LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, + unsigned UserUF, + unsigned VF, + unsigned LoopCost) { + + // -- The unroll heuristics -- + // We unroll the loop in order to expose ILP and reduce the loop overhead. + // There are many micro-architectural considerations that we can't predict + // at this level. For example frontend pressure (on decode or fetch) due to + // code size, or the number and capabilities of the execution ports. + // + // We use the following heuristics to select the unroll factor: + // 1. If the code has reductions the we unroll in order to break the cross + // iteration dependency. + // 2. If the loop is really small then we unroll in order to reduce the loop + // overhead. + // 3. We don't unroll if we think that we will spill registers to memory due + // to the increased register pressure. + + // Use the user preference, unless 'auto' is selected. + if (UserUF != 0) + return UserUF; + + // When we optimize for size we don't unroll. + if (OptForSize) + return 1; + + // Do not unroll loops with a relatively small trip count. + unsigned TC = SE->getSmallConstantTripCount(TheLoop, + TheLoop->getLoopLatch()); + if (TC > 1 && TC < TinyTripCountUnrollThreshold) + return 1; + + unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); + DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << + " vector registers\n"); + + LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); + // We divide by these constants so assume that we have at least one + // instruction that uses at least one register. + R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); + R.NumInstructions = std::max(R.NumInstructions, 1U); + + // We calculate the unroll factor using the following formula. + // Subtract the number of loop invariants from the number of available + // registers. These registers are used by all of the unrolled instances. + // Next, divide the remaining registers by the number of registers that is + // required by the loop, in order to estimate how many parallel instances + // fit without causing spills. + unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; + + // Clamp the unroll factor ranges to reasonable factors. + unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); + + // 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) + LoopCost = expectedCost(VF); + + // Clamp the calculated UF to be between the 1 and the max unroll factor + // that the target allows. + if (UF > MaxUnrollSize) + UF = MaxUnrollSize; + else if (UF < 1) + UF = 1; + + if (Legal->getReductionVars()->size()) { + DEBUG(dbgs() << "LV: Unrolling because of reductions. \n"); + return UF; + } + + // We want to unroll tiny loops in order to reduce the loop overhead. + // We assume that the cost overhead is 1 and we use the cost model + // to estimate the cost of the loop and unroll until the cost of the + // loop overhead is about 5% of the cost of the loop. + DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n"); + if (LoopCost < 20) { + DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n"); + unsigned NewUF = 20/LoopCost + 1; + return std::min(NewUF, UF); + } + + DEBUG(dbgs() << "LV: Not Unrolling. \n"); + return 1; +} + +LoopVectorizationCostModel::RegisterUsage +LoopVectorizationCostModel::calculateRegisterUsage() { + // This function calculates the register usage by measuring the highest number + // of values that are alive at a single location. Obviously, this is a very + // rough estimation. We scan the loop in a topological order in order and + // assign a number to each instruction. We use RPO to ensure that defs are + // met before their users. We assume that each instruction that has in-loop + // users starts an interval. We record every time that an in-loop value is + // used, so we have a list of the first and last occurrences of each + // instruction. Next, we transpose this data structure into a multi map that + // holds the list of intervals that *end* at a specific location. This multi + // map allows us to perform a linear search. We scan the instructions linearly + // and record each time that a new interval starts, by placing it in a set. + // If we find this value in the multi-map then we remove it from the set. + // The max register usage is the maximum size of the set. + // We also search for instructions that are defined outside the loop, but are + // used inside the loop. We need this number separately from the max-interval + // usage number because when we unroll, loop-invariant values do not take + // more register. + LoopBlocksDFS DFS(TheLoop); + DFS.perform(LI); + + RegisterUsage R; + R.NumInstructions = 0; + + // Each 'key' in the map opens a new interval. The values + // of the map are the index of the 'last seen' usage of the + // instruction that is the key. + typedef DenseMap IntervalMap; + // Maps instruction to its index. + DenseMap IdxToInstr; + // Marks the end of each interval. + IntervalMap EndPoint; + // Saves the list of instruction indices that are used in the loop. + SmallSet Ends; + // Saves the list of values that are used in the loop but are + // defined outside the loop, such as arguments and constants. + SmallPtrSet LoopInvariants; + + unsigned Index = 0; + for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), + be = DFS.endRPO(); bb != be; ++bb) { + R.NumInstructions += (*bb)->size(); + for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; + ++it) { + Instruction *I = it; + IdxToInstr[Index++] = I; + + // Save the end location of each USE. + for (unsigned i = 0; i < I->getNumOperands(); ++i) { + Value *U = I->getOperand(i); + Instruction *Instr = dyn_cast(U); + + // Ignore non-instruction values such as arguments, constants, etc. + if (!Instr) continue; + + // If this instruction is outside the loop then record it and continue. + if (!TheLoop->contains(Instr)) { + LoopInvariants.insert(Instr); + continue; + } + + // Overwrite previous end points. + EndPoint[Instr] = Index; + Ends.insert(Instr); + } + } + } + + // Saves the list of intervals that end with the index in 'key'. + typedef SmallVector InstrList; + DenseMap TransposeEnds; + + // Transpose the EndPoints to a list of values that end at each index. + for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); + it != e; ++it) + TransposeEnds[it->second].push_back(it->first); + + SmallSet OpenIntervals; + unsigned MaxUsage = 0; + + + DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); + for (unsigned int i = 0; i < Index; ++i) { + Instruction *I = IdxToInstr[i]; + // Ignore instructions that are never used within the loop. + if (!Ends.count(I)) continue; + + // Remove all of the instructions that end at this location. + InstrList &List = TransposeEnds[i]; + for (unsigned int j=0, e = List.size(); j < e; ++j) + OpenIntervals.erase(List[j]); + + // Count the number of live interals. + MaxUsage = std::max(MaxUsage, OpenIntervals.size()); + + DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << + OpenIntervals.size() <<"\n"); + + // Add the current instruction to the list of open intervals. + OpenIntervals.insert(I); + } + + unsigned Invariant = LoopInvariants.size(); + DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n"); + DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n"); + DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n"); + + R.LoopInvariantRegs = Invariant; + R.MaxLocalUsers = MaxUsage; + return R; } unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { @@ -2004,6 +3717,10 @@ unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { // For each instruction in the old loop. for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { + // Skip dbg intrinsics. + if (isa(it)) + continue; + unsigned C = getInstructionCost(it, VF); Cost += C; DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " << @@ -2024,8 +3741,6 @@ unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { - assert(VTTI && "Invalid vector target transformation info"); - // If we know that this instruction will remain uniform, check the cost of // the scalar version. if (Legal->isUniformAfterVectorization(I)) @@ -2037,12 +3752,13 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { // TODO: We need to estimate the cost of intrinsic calls. switch (I->getOpcode()) { case Instruction::GetElementPtr: - // We mark this instruction as zero-cost because scalar GEPs are usually - // lowered to the intruction addressing mode. At the moment we don't - // generate vector geps. + // We mark this instruction as zero-cost because the cost of GEPs in + // vectorized code depends on whether the corresponding memory instruction + // is scalarized or not. Therefore, we handle GEPs with the memory + // instruction cost. return 0; case Instruction::Br: { - return VTTI->getCFInstrCost(I->getOpcode()); + return TTI.getCFInstrCost(I->getOpcode()); } case Instruction::PHI: //TODO: IF-converted IFs become selects. @@ -2064,93 +3780,89 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { case Instruction::AShr: case Instruction::And: case Instruction::Or: - case Instruction::Xor: - return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy); + case Instruction::Xor: { + // Certain instructions can be cheaper to vectorize if they have a constant + // second vector operand. One example of this are shifts on x86. + TargetTransformInfo::OperandValueKind Op1VK = + TargetTransformInfo::OK_AnyValue; + TargetTransformInfo::OperandValueKind Op2VK = + TargetTransformInfo::OK_AnyValue; + + if (isa(I->getOperand(1))) + Op2VK = TargetTransformInfo::OK_UniformConstantValue; + + return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); + } case Instruction::Select: { SelectInst *SI = cast(I); const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); Type *CondTy = SI->getCondition()->getType(); - if (ScalarCond) + if (!ScalarCond) CondTy = VectorType::get(CondTy, VF); - return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); + return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); } case Instruction::ICmp: case Instruction::FCmp: { Type *ValTy = I->getOperand(0)->getType(); VectorTy = ToVectorTy(ValTy, VF); - return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy); + return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); } - case Instruction::Store: { - StoreInst *SI = cast(I); - Type *ValTy = SI->getValueOperand()->getType(); + case Instruction::Store: + case Instruction::Load: { + StoreInst *SI = dyn_cast(I); + LoadInst *LI = dyn_cast(I); + Type *ValTy = (SI ? SI->getValueOperand()->getType() : + LI->getType()); VectorTy = ToVectorTy(ValTy, VF); + unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); + unsigned AS = SI ? SI->getPointerAddressSpace() : + LI->getPointerAddressSpace(); + Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); + // We add the cost of address computation here instead of with the gep + // instruction because only here we know whether the operation is + // scalarized. if (VF == 1) - return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, - SI->getAlignment(), - SI->getPointerAddressSpace()); - - // Scalarized stores. - if (!Legal->isConsecutivePtr(SI->getPointerOperand())) { + return TTI.getAddressComputationCost(VectorTy) + + TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); + + // Scalarized loads/stores. + int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); + bool Reverse = ConsecutiveStride < 0; + unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); + unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; + if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { unsigned Cost = 0; - // The cost of extracting from the value vector and pointer vector. - Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF); + Type *PtrTy = ToVectorTy(Ptr->getType(), VF); for (unsigned i = 0; i < VF; ++i) { - Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement, - VectorTy, i); - Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement, - PtrTy, i); + // The cost of extracting the pointer operand. + Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); + // In case of STORE, the cost of ExtractElement from the vector. + // In case of LOAD, the cost of InsertElement into the returned + // vector. + Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : + Instruction::InsertElement, + VectorTy, i); } - // The cost of the scalar stores. - Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(), - ValTy->getScalarType(), - SI->getAlignment(), - SI->getPointerAddressSpace()); + // The cost of the scalar loads/stores. + Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType()); + Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), + Alignment, AS); return Cost; } - // Wide stores. - return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(), - SI->getPointerAddressSpace()); - } - case Instruction::Load: { - LoadInst *LI = cast(I); - - if (VF == 1) - return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, - LI->getAlignment(), - LI->getPointerAddressSpace()); - - // Scalarized loads. - if (!Legal->isConsecutivePtr(LI->getPointerOperand())) { - unsigned Cost = 0; - Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF); - - // The cost of extracting from the pointer vector. - for (unsigned i = 0; i < VF; ++i) - Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement, - PtrTy, i); - - // The cost of inserting data to the result vector. - for (unsigned i = 0; i < VF; ++i) - Cost += VTTI->getVectorInstrCost(Instruction::InsertElement, - VectorTy, i); - - // The cost of the scalar stores. - Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(), - RetTy->getScalarType(), - LI->getAlignment(), - LI->getPointerAddressSpace()); - return Cost; - } + // Wide load/stores. + unsigned Cost = TTI.getAddressComputationCost(VectorTy); + Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); - // Wide loads. - return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(), - LI->getPointerAddressSpace()); + if (Reverse) + Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, + VectorTy, 0); + return Cost; } case Instruction::ZExt: case Instruction::SExt: @@ -2168,20 +3880,21 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { // The cost of these is the same as the scalar operation. if (I->getOpcode() == Instruction::Trunc && Legal->isInductionVariable(I->getOperand(0))) - return VTTI->getCastInstrCost(I->getOpcode(), I->getType(), - I->getOperand(0)->getType()); + return TTI.getCastInstrCost(I->getOpcode(), I->getType(), + I->getOperand(0)->getType()); Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); - return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); + return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); } case Instruction::Call: { - assert(isTriviallyVectorizableIntrinsic(I)); - IntrinsicInst *II = cast(I); - Type *RetTy = ToVectorTy(II->getType(), VF); + CallInst *CI = cast(I); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); + assert(ID && "Not an intrinsic call!"); + Type *RetTy = ToVectorTy(CI->getType(), VF); SmallVector Tys; - for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) - Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF)); - return VTTI->getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys); + for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) + Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); + return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); } default: { // We are scalarizing the instruction. Return the cost of the scalar @@ -2190,10 +3903,10 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { unsigned Cost = 0; if (!RetTy->isVoidTy() && VF != 1) { - unsigned InsCost = VTTI->getVectorInstrCost(Instruction::InsertElement, - VectorTy); - unsigned ExtCost = VTTI->getVectorInstrCost(Instruction::ExtractElement, - VectorTy); + unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, + VectorTy); + unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, + VectorTy); // The cost of inserting the results plus extracting each one of the // operands. @@ -2202,7 +3915,7 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { // The cost of executing VF copies of the scalar instruction. This opcode // is unknown. Assume that it is the same as 'mul'. - Cost += VF * VTTI->getArithmeticInstrCost(Instruction::Mul, VectorTy); + Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); return Cost; } }// end of switch. @@ -2218,6 +3931,7 @@ char LoopVectorize::ID = 0; static const char lv_name[] = "Loop Vectorization"; INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) @@ -2228,4 +3942,14 @@ namespace llvm { } } +bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { + // Check for a store. + if (StoreInst *ST = dyn_cast(Inst)) + return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; + // Check for a load. + if (LoadInst *LI = dyn_cast(Inst)) + return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; + + return false; +}