[llvm] [IR2Vec] Initial infrastructure for MIR2Vec (PR #161463)
Mircea Trofin via llvm-commits
llvm-commits at lists.llvm.org
Wed Oct 1 17:50:55 PDT 2025
================
@@ -0,0 +1,300 @@
+//===- MIR2Vec.cpp - Implementation of MIR2Vec ---------------------------===//
+//
+// 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 implements the MIR2Vec algorithm for Machine IR embeddings.
+///
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/MIR2Vec.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/TargetInstrInfo.h"
+#include "llvm/IR/Module.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/Errc.h"
+#include "llvm/Support/MemoryBuffer.h"
+#include "llvm/Support/Regex.h"
+
+using namespace llvm;
+using namespace mir2vec;
+
+#define DEBUG_TYPE "mir2vec"
+
+STATISTIC(MIRVocabMissCounter,
+ "Number of lookups to MIR entities not present in the vocabulary");
+
+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> OpcWeight("mir2vec-opc-weight", cl::Optional, cl::init(1.0),
+ cl::desc("Weight for machine opcode embeddings"),
+ cl::cat(MIR2VecCategory));
+} // namespace mir2vec
+} // namespace llvm
+
+//===----------------------------------------------------------------------===//
+// Vocabulary Implementation
+//===----------------------------------------------------------------------===//
+
+MIRVocabulary::MIRVocabulary(VocabMap &&OpcodeEntries,
+ const TargetInstrInfo *TII) {
+ // Early return for invalid inputs - creates empty/invalid vocabulary
+ if (!TII || OpcodeEntries.empty())
+ return;
+
+ buildCanonicalOpcodeMapping(*TII);
+
+ unsigned CanonicalOpcodeCount = UniqueBaseOpcodeNames.size();
+ assert(CanonicalOpcodeCount > 0 &&
+ "No canonical opcodes found for target - invalid vocabulary");
+ Layout.OperandBase = CanonicalOpcodeCount;
+ generateStorage(OpcodeEntries, *TII);
+ Layout.TotalEntries = Storage.size();
+}
+
+std::string MIRVocabulary::extractBaseOpcodeName(StringRef InstrName) {
+ // Extract base instruction name using regex to capture letters and
+ // underscores Examples: "ADD32rr" -> "ADD", "ARITH_FENCE" -> "ARITH_FENCE"
+ //
+ // TODO: Consider more sophisticated extraction:
+ // - Handle complex prefixes like "AVX1_SETALLONES" correctly (Currently, it
+ // would naively map to "AVX")
+ // - Extract width suffixes (8,16,32,64) as separate features
+ // - Capture addressing mode suffixes (r,i,m,ri,etc.) for better analysis
+ // (Currently, instances like "MOV32mi" map to "MOV", but "ADDPDrr" would map
+ // to "ADDPDrr")
+
+ assert(!InstrName.empty() && "Instruction name should not be empty");
+
+ // Use regex to extract initial sequence of letters and underscores
+ static const Regex BaseOpcodeRegex("([a-zA-Z_]+)");
+ SmallVector<StringRef, 2> Matches;
+
+ if (BaseOpcodeRegex.match(InstrName, &Matches) && Matches.size() > 1) {
+ StringRef Match = Matches[1];
+ // Trim trailing underscores
+ while (!Match.empty() && Match.back() == '_')
+ Match = Match.drop_back();
+ return Match.str();
+ }
+
+ // Fallback to original name if no pattern matches
+ return InstrName.str();
+}
+
+unsigned MIRVocabulary::getCanonicalIndexForBaseName(StringRef BaseName) const {
+ assert(!UniqueBaseOpcodeNames.empty() && "Canonical mapping not built");
+ auto It = std::find(UniqueBaseOpcodeNames.begin(),
+ UniqueBaseOpcodeNames.end(), BaseName.str());
+ assert(It != UniqueBaseOpcodeNames.end() &&
+ "Base name not found in unique opcodes");
+ return std::distance(UniqueBaseOpcodeNames.begin(), It);
+}
+
+std::string MIRVocabulary::getStringKey(unsigned Pos) const {
+ assert(isValid() && "MIR2Vec Vocabulary is invalid");
+ assert(Pos < Layout.TotalEntries && "Position out of bounds in vocabulary");
+
+ // For now, all entries are opcodes since we only have one section
+ if (Pos < Layout.OperandBase && Pos < UniqueBaseOpcodeNames.size()) {
+ // Convert canonical index back to base opcode name
+ auto It = UniqueBaseOpcodeNames.begin();
+ std::advance(It, Pos);
+ return *It;
+ }
+
+ llvm_unreachable("Invalid position in vocabulary");
+ return "";
+}
+
+void MIRVocabulary::generateStorage(const VocabMap &OpcodeMap,
+ const TargetInstrInfo &TII) {
+
+ // Helper for handling missing entities in the vocabulary.
+ // Currently, we use a zero vector. In the future, we will throw an error to
+ // ensure that *all* known entities are present in the vocabulary.
+ auto handleMissingEntity = [](StringRef Key) {
+ LLVM_DEBUG(errs() << "MIR2Vec: Missing vocabulary entry for " << Key
+ << "; using zero vector. This will result in an error "
+ "in the future.\n");
+ ++MIRVocabMissCounter;
+ };
+
+ // Initialize opcode embeddings section
+ unsigned EmbeddingDim = OpcodeMap.begin()->second.size();
+ std::vector<Embedding> OpcodeEmbeddings(Layout.OperandBase,
+ Embedding(EmbeddingDim));
+
+ // Populate opcode embeddings using canonical mapping
+ for (auto COpcodeName : UniqueBaseOpcodeNames) {
+ if (auto It = OpcodeMap.find(COpcodeName); It != OpcodeMap.end()) {
+ auto COpcodeIndex = getCanonicalIndexForBaseName(COpcodeName);
+ assert(COpcodeIndex < Layout.OperandBase &&
+ "Canonical index out of bounds");
+ OpcodeEmbeddings[COpcodeIndex] = It->second;
+ } else {
+ handleMissingEntity(COpcodeName);
+ }
+ }
+
+ // TODO: Add operand/argument embeddings as additional sections
+ // This will require extending the vocabulary format and layout
+
+ // Scale the vocabulary sections based on the provided weights
+ auto scaleVocabSection = [](std::vector<Embedding> &Embeddings,
+ double Weight) {
+ for (auto &Embedding : Embeddings)
+ Embedding *= Weight;
+ };
+ scaleVocabSection(OpcodeEmbeddings, OpcWeight);
+
+ std::vector<std::vector<Embedding>> Sections(1);
+ Sections[0] = std::move(OpcodeEmbeddings);
+
+ // (Re-)Initialize storage using placement new (assignment operator is
+ // deleted)
+ new (&Storage) ir2vec::VocabStorage(std::move(Sections));
+}
+
+void MIRVocabulary::buildCanonicalOpcodeMapping(const TargetInstrInfo &TII) {
+ // Check if already built
+ if (!UniqueBaseOpcodeNames.empty())
+ return;
+
+ // Build mapping from opcodes to canonical base opcode indices
+ for (unsigned Opcode = 0; Opcode < TII.getNumOpcodes(); ++Opcode) {
+ std::string BaseOpcode = extractBaseOpcodeName(TII.getName(Opcode));
+ UniqueBaseOpcodeNames.insert(BaseOpcode);
+ }
+
+ LLVM_DEBUG(dbgs() << "MIR2Vec: Built canonical mapping for target with "
+ << UniqueBaseOpcodeNames.size()
+ << " unique base opcodes\n");
+}
+
+//===----------------------------------------------------------------------===//
+// MIR2VecVocabLegacyAnalysis Implementation
+//===----------------------------------------------------------------------===//
+
+char MIR2VecVocabLegacyAnalysis::ID = 0;
+INITIALIZE_PASS_BEGIN(MIR2VecVocabLegacyAnalysis, "mir2vec-vocab",
+ "MIR2Vec Vocabulary Analysis", false, true)
+INITIALIZE_PASS_DEPENDENCY(MachineModuleInfoWrapperPass)
+INITIALIZE_PASS_END(MIR2VecVocabLegacyAnalysis, "mir2vec-vocab",
+ "MIR2Vec Vocabulary Analysis", false, true)
+
+StringRef MIR2VecVocabLegacyAnalysis::getPassName() const {
+ return "MIR2Vec Vocabulary Analysis";
+}
+
+Error MIR2VecVocabLegacyAnalysis::readVocabulary() {
+ // TODO: Extend vocabulary format to support multiple sections
+ // (opcodes, operands, etc.) similar to IR2Vec structure
+ if (VocabFile.empty())
+ return createStringError(
+ errc::invalid_argument,
+ "MIR2Vec vocabulary file path not specified; set it "
+ "using --mir2vec-vocab-path");
+
+ auto BufOrError = MemoryBuffer::getFileOrSTDIN(VocabFile, /*IsText=*/true);
+ if (!BufOrError)
+ return createFileError(VocabFile, BufOrError.getError());
+
+ auto Content = BufOrError.get()->getBuffer();
+
+ Expected<json::Value> ParsedVocabValue = json::parse(Content);
+ if (!ParsedVocabValue)
+ return ParsedVocabValue.takeError();
+
+ unsigned Dim = 0;
+ if (auto Err = ir2vec::VocabStorage::parseVocabSection(
+ "entities", *ParsedVocabValue, StrVocabMap, Dim))
+ return Err;
+
+ return Error::success();
+}
+
+void MIR2VecVocabLegacyAnalysis::emitError(Error Err, LLVMContext &Ctx) {
+ Ctx.emitError(toString(std::move(Err)));
+}
+
+mir2vec::MIRVocabulary
+MIR2VecVocabLegacyAnalysis::getMIR2VecVocabulary(const Module &M) {
+ if (StrVocabMap.empty()) {
+ if (Error Err = readVocabulary()) {
+ emitError(std::move(Err), M.getContext());
+ return mir2vec::MIRVocabulary(std::move(StrVocabMap), nullptr);
+ }
+ }
+
+ // Get machine module info to access machine functions and target info
+ MachineModuleInfo &MMI = getAnalysis<MachineModuleInfoWrapperPass>().getMMI();
+
+ // Find first available machine function to get target instruction info
+ for (const auto &F : M) {
+ if (F.isDeclaration())
+ continue;
+
+ if (auto *MF = MMI.getMachineFunction(F)) {
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
+ return mir2vec::MIRVocabulary(std::move(StrVocabMap), TII);
+ }
+ }
+
+ // No machine functions available - return invalid vocabulary
+ emitError(make_error<StringError>("No machine functions found in module",
+ inconvertibleErrorCode()),
+ M.getContext());
+ return mir2vec::MIRVocabulary(std::move(StrVocabMap), nullptr);
+}
+
+//===----------------------------------------------------------------------===//
+// Printer Passes Implementation
+//===----------------------------------------------------------------------===//
+
+char MIR2VecVocabPrinterLegacyPass::ID = 0;
+INITIALIZE_PASS_BEGIN(MIR2VecVocabPrinterLegacyPass, "print-mir2vec-vocab",
+ "MIR2Vec Vocabulary Printer Pass", false, true)
+INITIALIZE_PASS_DEPENDENCY(MIR2VecVocabLegacyAnalysis)
+INITIALIZE_PASS_DEPENDENCY(MachineModuleInfoWrapperPass)
+INITIALIZE_PASS_END(MIR2VecVocabPrinterLegacyPass, "print-mir2vec-vocab",
+ "MIR2Vec Vocabulary Printer Pass", false, true)
+
+bool MIR2VecVocabPrinterLegacyPass::runOnMachineFunction(MachineFunction &MF) {
+ return false;
+}
+
+bool MIR2VecVocabPrinterLegacyPass::doFinalization(Module &M) {
+ auto &Analysis = getAnalysis<MIR2VecVocabLegacyAnalysis>();
+ auto MIR2VecVocab = Analysis.getMIR2VecVocabulary(M);
+
+ if (!MIR2VecVocab.isValid()) {
+ OS << "MIR2Vec Vocabulary Printer: Invalid vocabulary\n";
----------------
mtrofin wrote:
Guarded by LLVM_DEBUG maybe?
https://github.com/llvm/llvm-project/pull/161463
More information about the llvm-commits
mailing list