[MIR2Vec] Add embedder for machine instructions (#162161)

Implement MIR2Vec embedder for generating vector representations of Machine IR instructions, basic blocks, and functions. This patch introduces changes necessary to *embed* machine opcodes. Machine operands would be handled incrementally in the upcoming patches.

Author: svkeerthy

llvm/include/llvm/CodeGen/MIR2Vec.h

@@ -52,11 +52,21 @@ class LLVMContext;
 class MIR2VecVocabLegacyAnalysis;
 class TargetInstrInfo;
 
+enum class MIR2VecKind { Symbolic };
+
 namespace mir2vec {
+
+// Forward declarations
+class MIREmbedder;
+class SymbolicMIREmbedder;
+
 extern llvm::cl::OptionCategory MIR2VecCategory;
 extern cl::opt<float> OpcWeight;
 
 using Embedding = ir2vec::Embedding;
+using MachineInstEmbeddingsMap = DenseMap<const MachineInstr *, Embedding>;
+using MachineBlockEmbeddingsMap =
+    DenseMap<const MachineBasicBlock *, Embedding>;
 
 /// Class for storing and accessing the MIR2Vec vocabulary.
 /// The MIRVocabulary class manages seed embeddings for LLVM Machine IR
@@ -107,19 +117,91 @@ class MIRVocabulary {
 
   const_iterator end() const { return Storage.end(); }
 
-  /// Total number of entries in the vocabulary
-  size_t getCanonicalSize() const { return Storage.size(); }
-
   MIRVocabulary() = delete;
 
   /// Factory method to create MIRVocabulary from vocabulary map
   static Expected<MIRVocabulary> create(VocabMap &&Entries,
                                         const TargetInstrInfo &TII);
 
+  /// Create a dummy vocabulary for testing purposes.
+  static Expected<MIRVocabulary>
+  createDummyVocabForTest(const TargetInstrInfo &TII, unsigned Dim = 1);
+
+  /// Total number of entries in the vocabulary
+  size_t getCanonicalSize() const { return Storage.size(); }
+
 private:
   MIRVocabulary(VocabMap &&Entries, const TargetInstrInfo &TII);
 };
 
+/// Base class for MIR embedders
+class MIREmbedder {
+protected:
+  const MachineFunction &MF;
+  const MIRVocabulary &Vocab;
+
+  /// Dimension of the embeddings; Captured from the vocabulary
+  const unsigned Dimension;
+
+  /// Weight for opcode embeddings
+  const float OpcWeight;
+
+  MIREmbedder(const MachineFunction &MF, const MIRVocabulary &Vocab)
+      : MF(MF), Vocab(Vocab), Dimension(Vocab.getDimension()),
+        OpcWeight(mir2vec::OpcWeight) {}
+
+  /// Function to compute embeddings.
+  Embedding computeEmbeddings() const;
+
+  /// Function to compute the embedding for a given machine basic block.
+  Embedding computeEmbeddings(const MachineBasicBlock &MBB) const;
+
+  /// Function to compute the embedding for a given machine instruction.
+  /// Specific to the kind of embeddings being computed.
+  virtual Embedding computeEmbeddings(const MachineInstr &MI) const = 0;
+
+public:
+  virtual ~MIREmbedder() = default;
+
+  /// Factory method to create an Embedder object of the specified kind
+  /// Returns nullptr if the requested kind is not supported.
+  static std::unique_ptr<MIREmbedder> create(MIR2VecKind Mode,
+                                             const MachineFunction &MF,
+                                             const MIRVocabulary &Vocab);
+
+  /// Computes and returns the embedding for a given machine instruction MI in
+  /// the machine function MF.
+  Embedding getMInstVector(const MachineInstr &MI) const {
+    return computeEmbeddings(MI);
+  }
+
+  /// Computes and returns the embedding for a given machine basic block in the
+  /// machine function MF.
+  Embedding getMBBVector(const MachineBasicBlock &MBB) const {
+    return computeEmbeddings(MBB);
+  }
+
+  /// Computes and returns the embedding for the current machine function.
+  Embedding getMFunctionVector() const {
+    // Currently, we always (re)compute the embeddings for the function. This is
+    // cheaper than caching the vector.
+    return computeEmbeddings();
+  }
+};
+
+/// Class for computing Symbolic embeddings
+/// Symbolic embeddings are constructed based on the entity-level
+/// representations obtained from the MIR Vocabulary.
+class SymbolicMIREmbedder : public MIREmbedder {
+private:
+  Embedding computeEmbeddings(const MachineInstr &MI) const override;
+
+public:
+  SymbolicMIREmbedder(const MachineFunction &F, const MIRVocabulary &Vocab);
+  static std::unique_ptr<SymbolicMIREmbedder>
+  create(const MachineFunction &MF, const MIRVocabulary &Vocab);
+};
+
 } // namespace mir2vec
 
 /// Pass to analyze and populate MIR2Vec vocabulary from a module
@@ -166,6 +248,31 @@ class MIR2VecVocabPrinterLegacyPass : public MachineFunctionPass {
   }
 };
 
+/// This pass prints the MIR2Vec embeddings for machine functions, basic blocks,
+/// and instructions
+class MIR2VecPrinterLegacyPass : public MachineFunctionPass {
+  raw_ostream &OS;
+
+public:
+  static char ID;
+  explicit MIR2VecPrinterLegacyPass(raw_ostream &OS)
+      : MachineFunctionPass(ID), OS(OS) {}
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<MIR2VecVocabLegacyAnalysis>();
+    AU.setPreservesAll();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  StringRef getPassName() const override {
+    return "MIR2Vec Embedder Printer Pass";
+  }
+};
+
+/// Create a machine pass that prints MIR2Vec embeddings
+MachineFunctionPass *createMIR2VecPrinterLegacyPass(raw_ostream &OS);
+
 } // namespace llvm
 
 #endif // LLVM_CODEGEN_MIR2VEC_H

llvm/include/llvm/CodeGen/Passes.h

@@ -93,6 +93,10 @@ createMachineFunctionPrinterPass(raw_ostream &OS,
 LLVM_ABI MachineFunctionPass *
 createMIR2VecVocabPrinterLegacyPass(raw_ostream &OS);
 
+/// MIR2VecPrinter pass - This pass prints out the MIR2Vec embeddings for
+/// machine functions, basic blocks and instructions.
+LLVM_ABI MachineFunctionPass *createMIR2VecPrinterLegacyPass(raw_ostream &OS);
+
 /// StackFramePrinter pass - This pass prints out the machine function's
 /// stack frame to the given stream as a debugging tool.
 LLVM_ABI MachineFunctionPass *createStackFrameLayoutAnalysisPass();

llvm/include/llvm/InitializePasses.h

@@ -222,6 +222,7 @@ LLVM_ABI void
 initializeMachineSanitizerBinaryMetadataLegacyPass(PassRegistry &);
 LLVM_ABI void initializeMIR2VecVocabLegacyAnalysisPass(PassRegistry &);
 LLVM_ABI void initializeMIR2VecVocabPrinterLegacyPassPass(PassRegistry &);
+LLVM_ABI void initializeMIR2VecPrinterLegacyPassPass(PassRegistry &);
 LLVM_ABI void initializeMachineSchedulerLegacyPass(PassRegistry &);
 LLVM_ABI void initializeMachineSinkingLegacyPass(PassRegistry &);
 LLVM_ABI void initializeMachineTraceMetricsWrapperPassPass(PassRegistry &);

llvm/lib/CodeGen/CodeGen.cpp

@@ -98,6 +98,7 @@ void llvm::initializeCodeGen(PassRegistry &Registry) {
   initializeMachineUniformityAnalysisPassPass(Registry);
   initializeMIR2VecVocabLegacyAnalysisPass(Registry);
   initializeMIR2VecVocabPrinterLegacyPassPass(Registry);
+  initializeMIR2VecPrinterLegacyPassPass(Registry);
   initializeMachineUniformityInfoPrinterPassPass(Registry);
   initializeMachineVerifierLegacyPassPass(Registry);
   initializeObjCARCContractLegacyPassPass(Registry);

llvm/lib/CodeGen/MIR2Vec.cpp

@@ -12,6 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/MIR2Vec.h"
+#include "llvm/ADT/DepthFirstIterator.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/IR/Module.h"
@@ -29,20 +30,30 @@ using namespace mir2vec;
 STATISTIC(MIRVocabMissCounter,
           "Number of lookups to MIR entities not present in the vocabulary");
 
-cl::OptionCategory llvm::mir2vec::MIR2VecCategory("MIR2Vec Options");
+namespace llvm {
+namespace mir2vec {
+cl::OptionCategory MIR2VecCategory("MIR2Vec Options");
 
 // FIXME: Use a default vocab when not specified
 static cl::opt<std::string>
     VocabFile("mir2vec-vocab-path", cl::Optional,
               cl::desc("Path to the vocabulary file for MIR2Vec"), cl::init(""),
               cl::cat(MIR2VecCategory));
-cl::opt<float>
-    llvm::mir2vec::OpcWeight("mir2vec-opc-weight", cl::Optional, cl::init(1.0),
-                             cl::desc("Weight for machine opcode embeddings"),
-                             cl::cat(MIR2VecCategory));
+cl::opt<float> OpcWeight("mir2vec-opc-weight", cl::Optional, cl::init(1.0),
+                         cl::desc("Weight for machine opcode embeddings"),
+                         cl::cat(MIR2VecCategory));
+cl::opt<MIR2VecKind> MIR2VecEmbeddingKind(
+    "mir2vec-kind", cl::Optional,
+    cl::values(clEnumValN(MIR2VecKind::Symbolic, "symbolic",
+                          "Generate symbolic embeddings for MIR")),
+    cl::init(MIR2VecKind::Symbolic), cl::desc("MIR2Vec embedding kind"),
+    cl::cat(MIR2VecCategory));
+
+} // namespace mir2vec
+} // namespace llvm
 
 //===----------------------------------------------------------------------===//
-// Vocabulary Implementation
+// Vocabulary
 //===----------------------------------------------------------------------===//
 
 MIRVocabulary::MIRVocabulary(VocabMap &&OpcodeEntries,
@@ -188,6 +199,28 @@ void MIRVocabulary::buildCanonicalOpcodeMapping() {
                     << " unique base opcodes\n");
 }
 
+Expected<MIRVocabulary>
+MIRVocabulary::createDummyVocabForTest(const TargetInstrInfo &TII,
+                                       unsigned Dim) {
+  assert(Dim > 0 && "Dimension must be greater than zero");
+
+  float DummyVal = 0.1f;
+
+  // Create dummy embeddings for all canonical opcode names
+  VocabMap DummyVocabMap;
+  for (unsigned Opcode = 0; Opcode < TII.getNumOpcodes(); ++Opcode) {
+    std::string BaseOpcode = extractBaseOpcodeName(TII.getName(Opcode));
+    if (DummyVocabMap.count(BaseOpcode) == 0) {
+      // Only add if not already present
+      DummyVocabMap[BaseOpcode] = Embedding(Dim, DummyVal);
+      DummyVal += 0.1f;
+    }
+  }
+
+  // Create and return vocabulary with dummy embeddings
+  return MIRVocabulary::create(std::move(DummyVocabMap), TII);
+}
+
 //===----------------------------------------------------------------------===//
 // MIR2VecVocabLegacyAnalysis Implementation
 //===----------------------------------------------------------------------===//
@@ -258,7 +291,73 @@ MIR2VecVocabLegacyAnalysis::getMIR2VecVocabulary(const Module &M) {
 }
 
 //===----------------------------------------------------------------------===//
-// Printer Passes Implementation
+// MIREmbedder and its subclasses
+//===----------------------------------------------------------------------===//
+
+std::unique_ptr<MIREmbedder> MIREmbedder::create(MIR2VecKind Mode,
+                                                 const MachineFunction &MF,
+                                                 const MIRVocabulary &Vocab) {
+  switch (Mode) {
+  case MIR2VecKind::Symbolic:
+    return std::make_unique<SymbolicMIREmbedder>(MF, Vocab);
+  }
+  return nullptr;
+}
+
+Embedding MIREmbedder::computeEmbeddings(const MachineBasicBlock &MBB) const {
+  Embedding MBBVector(Dimension, 0);
+
+  // Get instruction info for opcode name resolution
+  const auto &Subtarget = MF.getSubtarget();
+  const auto *TII = Subtarget.getInstrInfo();
+  if (!TII) {
+    MF.getFunction().getContext().emitError(
+        "MIR2Vec: No TargetInstrInfo available; cannot compute embeddings");
+    return MBBVector;
+  }
+
+  // Process each machine instruction in the basic block
+  for (const auto &MI : MBB) {
+    // Skip debug instructions and other metadata
+    if (MI.isDebugInstr())
+      continue;
+    MBBVector += computeEmbeddings(MI);
+  }
+
+  return MBBVector;
+}
+
+Embedding MIREmbedder::computeEmbeddings() const {
+  Embedding MFuncVector(Dimension, 0);
+
+  // Consider all reachable machine basic blocks in the function
+  for (const auto *MBB : depth_first(&MF))
+    MFuncVector += computeEmbeddings(*MBB);
+  return MFuncVector;
+}
+
+SymbolicMIREmbedder::SymbolicMIREmbedder(const MachineFunction &MF,
+                                         const MIRVocabulary &Vocab)
+    : MIREmbedder(MF, Vocab) {}
+
+std::unique_ptr<SymbolicMIREmbedder>
+SymbolicMIREmbedder::create(const MachineFunction &MF,
+                            const MIRVocabulary &Vocab) {
+  return std::make_unique<SymbolicMIREmbedder>(MF, Vocab);
+}
+
+Embedding SymbolicMIREmbedder::computeEmbeddings(const MachineInstr &MI) const {
+  // Skip debug instructions and other metadata
+  if (MI.isDebugInstr())
+    return Embedding(Dimension, 0);
+
+  // Todo: Add operand/argument contributions
+
+  return Vocab[MI.getOpcode()];
+}
+
+//===----------------------------------------------------------------------===//
+// Printer Passes
 //===----------------------------------------------------------------------===//
 
 char MIR2VecVocabPrinterLegacyPass::ID = 0;
@@ -297,3 +396,56 @@ MachineFunctionPass *
 llvm::createMIR2VecVocabPrinterLegacyPass(raw_ostream &OS) {
   return new MIR2VecVocabPrinterLegacyPass(OS);
 }
+
+char MIR2VecPrinterLegacyPass::ID = 0;
+INITIALIZE_PASS_BEGIN(MIR2VecPrinterLegacyPass, "print-mir2vec",
+                      "MIR2Vec Embedder Printer Pass", false, true)
+INITIALIZE_PASS_DEPENDENCY(MIR2VecVocabLegacyAnalysis)
+INITIALIZE_PASS_DEPENDENCY(MachineModuleInfoWrapperPass)
+INITIALIZE_PASS_END(MIR2VecPrinterLegacyPass, "print-mir2vec",
+                    "MIR2Vec Embedder Printer Pass", false, true)
+
+bool MIR2VecPrinterLegacyPass::runOnMachineFunction(MachineFunction &MF) {
+  auto &Analysis = getAnalysis<MIR2VecVocabLegacyAnalysis>();
+  auto VocabOrErr =
+      Analysis.getMIR2VecVocabulary(*MF.getFunction().getParent());
+  assert(VocabOrErr && "Failed to get MIR2Vec vocabulary");
+  auto &MIRVocab = *VocabOrErr;
+
+  auto Emb = mir2vec::MIREmbedder::create(MIR2VecEmbeddingKind, MF, MIRVocab);
+  if (!Emb) {
+    OS << "Error creating MIR2Vec embeddings for function " << MF.getName()
+       << "\n";
+    return false;
+  }
+
+  OS << "MIR2Vec embeddings for machine function " << MF.getName() << ":\n";
+  OS << "Machine Function vector: ";
+  Emb->getMFunctionVector().print(OS);
+
+  OS << "Machine basic block vectors:\n";
+  for (const MachineBasicBlock &MBB : MF) {
+    OS << "Machine basic block: " << MBB.getFullName() << ":\n";
+    Emb->getMBBVector(MBB).print(OS);
+  }
+
+  OS << "Machine instruction vectors:\n";
+  for (const MachineBasicBlock &MBB : MF) {
+    for (const MachineInstr &MI : MBB) {
+      // Skip debug instructions as they are not
+      // embedded
+      if (MI.isDebugInstr())
+        continue;
+
+      OS << "Machine instruction: ";
+      MI.print(OS);
+      Emb->getMInstVector(MI).print(OS);
+    }
+  }
+
+  return false;
+}
+
+MachineFunctionPass *llvm::createMIR2VecPrinterLegacyPass(raw_ostream &OS) {
+  return new MIR2VecPrinterLegacyPass(OS);
+}