IR2Vec.h

Go to the documentation of this file.

//===- IR2Vec.h - Implementation of IR2Vec ----------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM
// Exceptions. See the LICENSE file for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file defines the IR2Vec vocabulary analysis(IR2VecVocabAnalysis),
/// the core ir2vec::Embedder interface for generating IR embeddings,
/// and related utilities like the IR2VecPrinterPass.
///
/// Program Embeddings are typically or derived-from a learned
/// representation of the program. Such embeddings are used to represent the
/// programs as input to machine learning algorithms. IR2Vec represents the
/// LLVM IR as embeddings.
///
/// The IR2Vec algorithm is described in the following paper:
///
///   IR2Vec: LLVM IR Based Scalable Program Embeddings, S. VenkataKeerthy,
///   Rohit Aggarwal, Shalini Jain, Maunendra Sankar Desarkar, Ramakrishna
///   Upadrasta, and Y. N. Srikant, ACM Transactions on Architecture and
///   Code Optimization (TACO), 2020. https://doi.org/10.1145/3418463.
///   https://arxiv.org/abs/1909.06228
///
//===----------------------------------------------------------------------===//
 
#ifndef LLVM_ANALYSIS_IR2VEC_H
#define LLVM_ANALYSIS_IR2VEC_H
 
#include "llvm/ADT/DenseMap.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ErrorOr.h"
#include "llvm/Support/JSON.h"
#include <array>
#include <map>
 
namespace llvm {
 
class Module;
class BasicBlock;
class Instruction;
class Function;
class Value;
class raw_ostream;
class LLVMContext;
class IR2VecVocabAnalysis;
 
/// IR2Vec computes two kinds of embeddings: Symbolic and Flow-aware.
/// Symbolic embeddings capture the "syntactic" and "statistical correlation"
/// of the IR entities. Flow-aware embeddings build on top of symbolic
/// embeddings and additionally capture the flow information in the IR.
/// IR2VecKind is used to specify the type of embeddings to generate.
/// Note: Implementation of FlowAware embeddings is not same as the one
/// described in the paper. The current implementation is a simplified version
/// that captures the flow information (SSA-based use-defs) without tracing
/// through memory level use-defs in the embedding computation described in the
/// paper.
enum class IR2VecKind { Symbolic, FlowAware };
 
namespace ir2vec {
 
extern llvm::cl::OptionCategory IR2VecCategory;
LLVM_ABI extern cl::opt<float> OpcWeight;
LLVM_ABI extern cl::opt<float> TypeWeight;
LLVM_ABI extern cl::opt<float> ArgWeight;
LLVM_ABI extern cl::opt<IR2VecKind> IR2VecEmbeddingKind;
 
/// Embedding is a datatype that wraps std::vector<double>. It provides
/// additional functionality for arithmetic and comparison operations.
/// It is meant to be used *like* std::vector<double> but is more restrictive
/// in the sense that it does not allow the user to change the size of the
/// embedding vector. The dimension of the embedding is fixed at the time of
/// construction of Embedding object. But the elements can be modified in-place.
struct Embedding {
private:
  std::vector<double> Data;
 
public:
  Embedding() = default;
  Embedding(const std::vector<double> &V) : Data(V) {}
  Embedding(std::vector<double> &&V) : Data(std::move(V)) {}
  Embedding(std::initializer_list<double> IL) : Data(IL) {}
 
  explicit Embedding(size_t Size) : Data(Size) {}
  Embedding(size_t Size, double InitialValue) : Data(Size, InitialValue) {}
 
  size_t size() const { return Data.size(); }
  bool empty() const { return Data.empty(); }
 
  double &operator[](size_t Itr) {
    assert(Itr < Data.size() && "Index out of bounds");
    return Data[Itr];
  }
 
  const double &operator[](size_t Itr) const {
    assert(Itr < Data.size() && "Index out of bounds");
    return Data[Itr];
  }
 
  using iterator = typename std::vector<double>::iterator;
  using const_iterator = typename std::vector<double>::const_iterator;
 
  iterator begin() { return Data.begin(); }
  iterator end() { return Data.end(); }
  const_iterator begin() const { return Data.begin(); }
  const_iterator end() const { return Data.end(); }
  const_iterator cbegin() const { return Data.cbegin(); }
  const_iterator cend() const { return Data.cend(); }
 
  const std::vector<double> &getData() const { return Data; }
 
  /// Arithmetic operators
  LLVM_ABI Embedding &operator+=(const Embedding &RHS);
  LLVM_ABI Embedding operator+(const Embedding &RHS) const;
  LLVM_ABI Embedding &operator-=(const Embedding &RHS);
  LLVM_ABI Embedding operator-(const Embedding &RHS) const;
  LLVM_ABI Embedding &operator*=(double Factor);
  LLVM_ABI Embedding operator*(double Factor) const;
 
  /// Adds Src Embedding scaled by Factor with the called Embedding.
  /// Called_Embedding += Src * Factor
  LLVM_ABI Embedding &scaleAndAdd(const Embedding &Src, float Factor);
 
  /// Returns true if the embedding is approximately equal to the RHS embedding
  /// within the specified tolerance.
  LLVM_ABI bool approximatelyEquals(const Embedding &RHS,
                                    double Tolerance = 1e-4) const;
 
  LLVM_ABI void print(raw_ostream &OS) const;
};
 
using InstEmbeddingsMap = DenseMap<const Instruction *, Embedding>;
using BBEmbeddingsMap = DenseMap<const BasicBlock *, Embedding>;
 
/// Class for storing and accessing the IR2Vec vocabulary.
/// The Vocabulary class manages seed embeddings for LLVM IR entities. The
/// seed embeddings are the initial learned representations of the entities
/// of LLVM IR. The IR2Vec representation for a given IR is derived from these
/// seed embeddings.
///
/// The vocabulary contains the seed embeddings for three types of entities:
/// instruction opcodes, types, and operands. Types are grouped/canonicalized
/// for better learning (e.g., all float variants map to FloatTy). The
/// vocabulary abstracts away the canonicalization effectively, the exposed APIs
/// handle all the known LLVM IR opcodes, types and operands.
///
/// This class helps populate the seed embeddings in an internal vector-based
/// ADT. It provides logic to map every IR entity to a specific slot index or
/// position in this vector, enabling O(1) embedding lookup while avoiding
/// unnecessary computations involving string based lookups while generating the
/// embeddings.
class Vocabulary {
  friend class llvm::IR2VecVocabAnalysis;
  using VocabVector = std::vector<ir2vec::Embedding>;
  VocabVector Vocab;
  bool Valid = false;
 
public:
  // Slot layout:
  // [0 .. MaxOpcodes-1]               => Instruction opcodes
  // [MaxOpcodes .. MaxOpcodes+MaxCanonicalTypeIDs-1] => Canonicalized types
  // [MaxOpcodes+MaxCanonicalTypeIDs .. NumCanonicalEntries-1] => Operand kinds
 
  /// Canonical type IDs supported by IR2Vec Vocabulary
  enum class CanonicalTypeID : unsigned {
    FloatTy,
    VoidTy,
    LabelTy,
    MetadataTy,
    VectorTy,
    TokenTy,
    IntegerTy,
    FunctionTy,
    PointerTy,
    StructTy,
    ArrayTy,
    UnknownTy,
    MaxCanonicalType
  };
 
  /// Operand kinds supported by IR2Vec Vocabulary
  enum class OperandKind : unsigned {
    FunctionID,
    PointerID,
    ConstantID,
    VariableID,
    MaxOperandKind
  };
 
  /// Vocabulary layout constants
#define LAST_OTHER_INST(NUM) static constexpr unsigned MaxOpcodes = NUM;
#include "llvm/IR/Instruction.def"
#undef LAST_OTHER_INST
 
  static constexpr unsigned MaxTypeIDs = Type::TypeID::TargetExtTyID + 1;
  static constexpr unsigned MaxCanonicalTypeIDs =
      static_cast<unsigned>(CanonicalTypeID::MaxCanonicalType);
  static constexpr unsigned MaxOperandKinds =
      static_cast<unsigned>(OperandKind::MaxOperandKind);
 
  Vocabulary() = default;
  LLVM_ABI Vocabulary(VocabVector &&Vocab);
 
  LLVM_ABI bool isValid() const;
  LLVM_ABI unsigned getDimension() const;
  /// Total number of entries (opcodes + canonicalized types + operand kinds)
  static constexpr size_t getCanonicalSize() { return NumCanonicalEntries; }
 
  /// Function to get vocabulary key for a given Opcode
  LLVM_ABI static StringRef getVocabKeyForOpcode(unsigned Opcode);
 
  /// Function to get vocabulary key for a given TypeID
  LLVM_ABI static StringRef getVocabKeyForTypeID(Type::TypeID TypeID);
 
  /// Function to get vocabulary key for a given OperandKind
  LLVM_ABI static StringRef getVocabKeyForOperandKind(OperandKind Kind);
 
  /// Function to classify an operand into OperandKind
  LLVM_ABI static OperandKind getOperandKind(const Value *Op);
 
  /// Functions to return the slot index or position of a given Opcode, TypeID,
  /// or OperandKind in the vocabulary.
  LLVM_ABI static unsigned getSlotIndex(unsigned Opcode);
  LLVM_ABI static unsigned getSlotIndex(Type::TypeID TypeID);
  LLVM_ABI static unsigned getSlotIndex(const Value &Op);
 
  /// Accessors to get the embedding for a given entity.
  LLVM_ABI const ir2vec::Embedding &operator[](unsigned Opcode) const;
  LLVM_ABI const ir2vec::Embedding &operator[](Type::TypeID TypeId) const;
  LLVM_ABI const ir2vec::Embedding &operator[](const Value &Arg) const;
 
  /// Const Iterator type aliases
  using const_iterator = VocabVector::const_iterator;
  const_iterator begin() const {
    assert(Valid && "IR2Vec Vocabulary is invalid");
    return Vocab.begin();
  }
 
  const_iterator cbegin() const {
    assert(Valid && "IR2Vec Vocabulary is invalid");
    return Vocab.cbegin();
  }
 
  const_iterator end() const {
    assert(Valid && "IR2Vec Vocabulary is invalid");
    return Vocab.end();
  }
 
  const_iterator cend() const {
    assert(Valid && "IR2Vec Vocabulary is invalid");
    return Vocab.cend();
  }
 
  /// Returns the string key for a given index position in the vocabulary.
  /// This is useful for debugging or printing the vocabulary. Do not use this
  /// for embedding generation as string based lookups are inefficient.
  LLVM_ABI static StringRef getStringKey(unsigned Pos);
 
  /// Create a dummy vocabulary for testing purposes.
  LLVM_ABI static VocabVector createDummyVocabForTest(unsigned Dim = 1);
 
  LLVM_ABI bool invalidate(Module &M, const PreservedAnalyses &PA,
                           ModuleAnalysisManager::Invalidator &Inv) const;
 
private:
  constexpr static unsigned NumCanonicalEntries =
      MaxOpcodes + MaxCanonicalTypeIDs + MaxOperandKinds;
 
  /// String mappings for CanonicalTypeID values
  static constexpr StringLiteral CanonicalTypeNames[] = {
      "FloatTy",   "VoidTy",   "LabelTy",   "MetadataTy",
      "VectorTy",  "TokenTy",  "IntegerTy", "FunctionTy",
      "PointerTy", "StructTy", "ArrayTy",   "UnknownTy"};
  static_assert(std::size(CanonicalTypeNames) ==
                    static_cast<unsigned>(CanonicalTypeID::MaxCanonicalType),
                "CanonicalTypeNames array size must match MaxCanonicalType");
 
  /// String mappings for OperandKind values
  static constexpr StringLiteral OperandKindNames[] = {"Function", "Pointer",
                                                       "Constant", "Variable"};
  static_assert(std::size(OperandKindNames) ==
                    static_cast<unsigned>(OperandKind::MaxOperandKind),
                "OperandKindNames array size must match MaxOperandKind");
 
  /// Every known TypeID defined in llvm/IR/Type.h is expected to have a
  /// corresponding mapping here in the same order as enum Type::TypeID.
  static constexpr std::array<CanonicalTypeID, MaxTypeIDs> TypeIDMapping = {{
      CanonicalTypeID::FloatTy,    // HalfTyID = 0
      CanonicalTypeID::FloatTy,    // BFloatTyID
      CanonicalTypeID::FloatTy,    // FloatTyID
      CanonicalTypeID::FloatTy,    // DoubleTyID
      CanonicalTypeID::FloatTy,    // X86_FP80TyID
      CanonicalTypeID::FloatTy,    // FP128TyID
      CanonicalTypeID::FloatTy,    // PPC_FP128TyID
      CanonicalTypeID::VoidTy,     // VoidTyID
      CanonicalTypeID::LabelTy,    // LabelTyID
      CanonicalTypeID::MetadataTy, // MetadataTyID
      CanonicalTypeID::VectorTy,   // X86_AMXTyID
      CanonicalTypeID::TokenTy,    // TokenTyID
      CanonicalTypeID::IntegerTy,  // IntegerTyID
      CanonicalTypeID::FunctionTy, // FunctionTyID
      CanonicalTypeID::PointerTy,  // PointerTyID
      CanonicalTypeID::StructTy,   // StructTyID
      CanonicalTypeID::ArrayTy,    // ArrayTyID
      CanonicalTypeID::VectorTy,   // FixedVectorTyID
      CanonicalTypeID::VectorTy,   // ScalableVectorTyID
      CanonicalTypeID::PointerTy,  // TypedPointerTyID
      CanonicalTypeID::UnknownTy   // TargetExtTyID
  }};
  static_assert(TypeIDMapping.size() == MaxTypeIDs,
                "TypeIDMapping must cover all Type::TypeID values");
 
  /// Function to get vocabulary key for canonical type by enum
  LLVM_ABI static StringRef
  getVocabKeyForCanonicalTypeID(CanonicalTypeID CType);
 
  /// Function to convert TypeID to CanonicalTypeID
  LLVM_ABI static CanonicalTypeID getCanonicalTypeID(Type::TypeID TypeID);
};
 
/// Embedder provides the interface to generate embeddings (vector
/// representations) for instructions, basic blocks, and functions. The
/// vector representations are generated using IR2Vec algorithms.
///
/// The Embedder class is an abstract class and it is intended to be
/// subclassed for different IR2Vec algorithms like Symbolic and Flow-aware.
class Embedder {
protected:
  const Function &F;
  const Vocabulary &Vocab;
 
  /// Dimension of the vector representation; captured from the input vocabulary
  const unsigned Dimension;
 
  /// Weights for different entities (like opcode, arguments, types)
  /// in the IR instructions to generate the vector representation.
  const float OpcWeight, TypeWeight, ArgWeight;
 
  // Utility maps - these are used to store the vector representations of
  // instructions, basic blocks and functions.
  mutable Embedding FuncVector;
  mutable BBEmbeddingsMap BBVecMap;
  mutable InstEmbeddingsMap InstVecMap;
 
  LLVM_ABI Embedder(const Function &F, const Vocabulary &Vocab);
 
  /// Function to compute embeddings. It generates embeddings for all
  /// the instructions and basic blocks in the function F.
  void computeEmbeddings() const;
 
  /// Function to compute the embedding for a given basic block.
  /// Specific to the kind of embeddings being computed.
  virtual void computeEmbeddings(const BasicBlock &BB) const = 0;
 
public:
  virtual ~Embedder() = default;
 
  /// Factory method to create an Embedder object.
  LLVM_ABI static std::unique_ptr<Embedder>
  create(IR2VecKind Mode, const Function &F, const Vocabulary &Vocab);
 
  /// Returns a map containing instructions and the corresponding embeddings for
  /// the function F if it has been computed. If not, it computes the embeddings
  /// for the function and returns the map.
  LLVM_ABI const InstEmbeddingsMap &getInstVecMap() const;
 
  /// Returns a map containing basic block and the corresponding embeddings for
  /// the function F if it has been computed. If not, it computes the embeddings
  /// for the function and returns the map.
  LLVM_ABI const BBEmbeddingsMap &getBBVecMap() const;
 
  /// Returns the embedding for a given basic block in the function F if it has
  /// been computed. If not, it computes the embedding for the basic block and
  /// returns it.
  LLVM_ABI const Embedding &getBBVector(const BasicBlock &BB) const;
 
  /// Computes and returns the embedding for the current function.
  LLVM_ABI const Embedding &getFunctionVector() const;
};
 
/// Class for computing the Symbolic embeddings of IR2Vec.
/// Symbolic embeddings are constructed based on the entity-level
/// representations obtained from the Vocabulary.
class LLVM_ABI SymbolicEmbedder : public Embedder {
private:
  void computeEmbeddings(const BasicBlock &BB) const override;
 
public:
  SymbolicEmbedder(const Function &F, const Vocabulary &Vocab)
      : Embedder(F, Vocab) {}
};
 
/// Class for computing the Flow-aware embeddings of IR2Vec.
/// Flow-aware embeddings build on the vocabulary, just like Symbolic
/// embeddings, and additionally capture the flow information in the IR.
class LLVM_ABI FlowAwareEmbedder : public Embedder {
private:
  void computeEmbeddings(const BasicBlock &BB) const override;
 
public:
  FlowAwareEmbedder(const Function &F, const Vocabulary &Vocab)
      : Embedder(F, Vocab) {}
};
 
} // namespace ir2vec
 
/// This analysis provides the vocabulary for IR2Vec. The vocabulary provides a
/// mapping between an entity of the IR (like opcode, type, argument, etc.) and
/// its corresponding embedding.
class IR2VecVocabAnalysis : public AnalysisInfoMixin<IR2VecVocabAnalysis> {
  using VocabVector = std::vector<ir2vec::Embedding>;
  using VocabMap = std::map<std::string, ir2vec::Embedding>;
  VocabMap OpcVocab, TypeVocab, ArgVocab;
  VocabVector Vocab;
 
  Error readVocabulary();
  Error parseVocabSection(StringRef Key, const json::Value &ParsedVocabValue,
                          VocabMap &TargetVocab, unsigned &Dim);
  void generateNumMappedVocab();
  void emitError(Error Err, LLVMContext &Ctx);
 
public:
  LLVM_ABI static AnalysisKey Key;
  IR2VecVocabAnalysis() = default;
  LLVM_ABI explicit IR2VecVocabAnalysis(const VocabVector &Vocab);
  LLVM_ABI explicit IR2VecVocabAnalysis(VocabVector &&Vocab);
  using Result = ir2vec::Vocabulary;
  LLVM_ABI Result run(Module &M, ModuleAnalysisManager &MAM);
};
 
/// This pass prints the IR2Vec embeddings for instructions, basic blocks, and
/// functions.
class IR2VecPrinterPass : public PassInfoMixin<IR2VecPrinterPass> {
  raw_ostream &OS;
 
public:
  explicit IR2VecPrinterPass(raw_ostream &OS) : OS(OS) {}
  LLVM_ABI PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
  static bool isRequired() { return true; }
};
 
/// This pass prints the embeddings in the vocabulary
class IR2VecVocabPrinterPass : public PassInfoMixin<IR2VecVocabPrinterPass> {
  raw_ostream &OS;
 
public:
  explicit IR2VecVocabPrinterPass(raw_ostream &OS) : OS(OS) {}
  LLVM_ABI PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
  static bool isRequired() { return true; }
};
 
} // namespace llvm
 
#endif // LLVM_ANALYSIS_IR2VEC_H