Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Adding IR2Vec as an analysis pass #134004

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 13 commits into
base: main
Choose a base branch
from
Open

Adding IR2Vec as an analysis pass #134004

wants to merge 13 commits into from

Conversation

svkeerthy
Copy link

@svkeerthy svkeerthy commented Apr 1, 2025

This PR introduces IR2Vec as an analysis pass. The changes include:

  • Logic for generating Symbolic encodings.
  • 75D learned vocabulary.
  • lit tests.

Here is the link to the RFC - https://discourse.llvm.org/t/rfc-enhancing-mlgo-inlining-with-ir2vec-embeddings

Acknowledgements: contributors - https://github.com/IITH-Compilers/IR2Vec/graphs/contributors

Copy link

github-actions bot commented Apr 1, 2025

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

@svkeerthy svkeerthy marked this pull request as ready for review April 7, 2025 20:22
@llvmbot
Copy link
Member

llvmbot commented Apr 7, 2025

@llvm/pr-subscribers-llvm-analysis

Author: S. VenkataKeerthy (svkeerthy)

Changes

This PR introduces IR2Vec as an analysis pass. The changes include:

  • Logic for generating Symbolic encodings.
  • 75D learned vocabulary.
  • lit tests.

(Planning to post an RFC; Will update the PR with RFC link)

Acknowledgements: contributors - https://github.com/IITH-Compilers/IR2Vec/graphs/contributors


Patch is 88.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/134004.diff

10 Files Affected:

  • (added) llvm/include/llvm/Analysis/IR2VecAnalysis.h (+131)
  • (modified) llvm/lib/Analysis/CMakeLists.txt (+1)
  • (added) llvm/lib/Analysis/IR2VecAnalysis.cpp (+429)
  • (added) llvm/lib/Analysis/models/seedEmbeddingVocab75D.json (+65)
  • (modified) llvm/lib/Passes/PassBuilder.cpp (+1)
  • (modified) llvm/lib/Passes/PassRegistry.def (+3)
  • (added) llvm/test/Analysis/IR2Vec/Inputs/dummy_3D_vocab.json (+7)
  • (added) llvm/test/Analysis/IR2Vec/Inputs/dummy_5D_vocab.json (+11)
  • (added) llvm/test/Analysis/IR2Vec/basic.ll (+50)
  • (added) llvm/test/Analysis/IR2Vec/if-else.ll (+38)
diff --git a/llvm/include/llvm/Analysis/IR2VecAnalysis.h b/llvm/include/llvm/Analysis/IR2VecAnalysis.h
new file mode 100644
index 0000000000000..dd5c00a1168b8
--- /dev/null
+++ b/llvm/include/llvm/Analysis/IR2VecAnalysis.h
@@ -0,0 +1,131 @@
+//===- IR2VecAnalysis.h - IR2Vec Analysis Implementation -------*- 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 contains the declaration of IR2VecAnalysis that computes
+/// IR2Vec Embeddings of the program.
+///
+/// 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_IR2VECANALYSIS_H
+#define LLVM_ANALYSIS_IR2VECANALYSIS_H
+
+#include "llvm/ADT/MapVector.h"
+#include "llvm/IR/PassManager.h"
+#include <map>
+
+namespace llvm {
+
+class Module;
+class BasicBlock;
+class Instruction;
+class Function;
+
+namespace ir2vec {
+using Embedding = std::vector<double>;
+// FIXME: Current the keys are strings. This can be changed to
+// use integers for cheaper lookups.
+using Vocab = std::map<std::string, Embedding>;
+} // namespace ir2vec
+
+class IR2VecVocabResult;
+class IR2VecResult;
+
+/// 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> {
+  unsigned DIM = 0;
+  ir2vec::Vocab Vocabulary;
+  Error readVocabulary();
+
+public:
+  static AnalysisKey Key;
+  IR2VecVocabAnalysis() = default;
+  using Result = IR2VecVocabResult;
+  Result run(Module &M, ModuleAnalysisManager &MAM);
+};
+
+class IR2VecVocabResult {
+  ir2vec::Vocab Vocabulary;
+  bool Valid = false;
+  unsigned DIM = 0;
+
+public:
+  IR2VecVocabResult() = default;
+  IR2VecVocabResult(ir2vec::Vocab &&Vocabulary, unsigned Dim);
+
+  // Helper functions
+  bool isValid() const { return Valid; }
+  const ir2vec::Vocab &getVocabulary() const;
+  unsigned getDimension() const { return DIM; }
+  bool invalidate(Module &M, const PreservedAnalyses &PA,
+                  ModuleAnalysisManager::Invalidator &Inv);
+};
+
+class IR2VecResult {
+  SmallMapVector<const Instruction *, ir2vec::Embedding, 128> InstVecMap;
+  SmallMapVector<const BasicBlock *, ir2vec::Embedding, 16> BBVecMap;
+  ir2vec::Embedding FuncVector;
+  unsigned DIM = 0;
+  bool Valid = false;
+
+public:
+  IR2VecResult() = default;
+  IR2VecResult(
+      SmallMapVector<const Instruction *, ir2vec::Embedding, 128> &&InstMap,
+      SmallMapVector<const BasicBlock *, ir2vec::Embedding, 16> &&BBMap,
+      ir2vec::Embedding &&FuncVector, unsigned Dim);
+  bool isValid() const { return Valid; }
+
+  const SmallMapVector<const Instruction *, ir2vec::Embedding, 128> &
+  getInstVecMap() const;
+  const SmallMapVector<const BasicBlock *, ir2vec::Embedding, 16> &
+  getBBVecMap() const;
+  const ir2vec::Embedding &getFunctionVector() const;
+  unsigned getDimension() const;
+};
+
+/// This analysis provides the IR2Vec embeddings for instructions, basic blocks,
+/// and functions.
+class IR2VecAnalysis : public AnalysisInfoMixin<IR2VecAnalysis> {
+public:
+  IR2VecAnalysis() = default;
+  static AnalysisKey Key;
+  using Result = IR2VecResult;
+  Result run(Function &F, FunctionAnalysisManager &FAM);
+};
+
+/// This pass prints the IR2Vec embeddings for instructions, basic blocks, and
+/// functions.
+class IR2VecPrinterPass : public PassInfoMixin<IR2VecPrinterPass> {
+  raw_ostream &OS;
+  void printVector(const ir2vec::Embedding &Vec) const;
+
+public:
+  explicit IR2VecPrinterPass(raw_ostream &OS) : OS(OS) {}
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
+  static bool isRequired() { return true; }
+};
+
+} // namespace llvm
+
+#endif // LLVM_ANALYSIS_IR2VECANALYSIS_H
diff --git a/llvm/lib/Analysis/CMakeLists.txt b/llvm/lib/Analysis/CMakeLists.txt
index fbf3b587d6bd2..8a6399f756f27 100644
--- a/llvm/lib/Analysis/CMakeLists.txt
+++ b/llvm/lib/Analysis/CMakeLists.txt
@@ -67,6 +67,7 @@ add_llvm_component_library(LLVMAnalysis
   GlobalsModRef.cpp
   GuardUtils.cpp
   HeatUtils.cpp
+  IR2VecAnalysis.cpp
   IRSimilarityIdentifier.cpp
   IVDescriptors.cpp
   IVUsers.cpp
diff --git a/llvm/lib/Analysis/IR2VecAnalysis.cpp b/llvm/lib/Analysis/IR2VecAnalysis.cpp
new file mode 100644
index 0000000000000..1ff233145769f
--- /dev/null
+++ b/llvm/lib/Analysis/IR2VecAnalysis.cpp
@@ -0,0 +1,429 @@
+//===- IR2VecAnalysis.cpp - IR2Vec Analysis Implementation ----------------===//
+//
+// 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 IR2Vec algorithm.
+///
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/IR2VecAnalysis.h"
+
+#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PassManager.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/Errc.h"
+#include "llvm/Support/Error.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/Format.h"
+#include "llvm/Support/JSON.h"
+#include "llvm/Support/MemoryBuffer.h"
+
+using namespace llvm;
+using namespace ir2vec;
+
+#define DEBUG_TYPE "ir2vec"
+
+STATISTIC(DataMissCounter, "Number of data misses in the vocabulary");
+
+/// 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.
+// FIXME: Currently we support only Symbolic.  Add support for
+// Flow-aware in upcoming patches.
+enum class IR2VecKind { Symbolic, Flowaware };
+
+static cl::OptionCategory IR2VecAnalysisCategory("IR2Vec Analysis Options");
+
+cl::opt<IR2VecKind>
+    IR2VecMode("ir2vec-mode",
+               cl::desc("Choose type of embeddings to generate:"),
+               cl::values(clEnumValN(IR2VecKind::Symbolic, "symbolic",
+                                     "Generates symbolic embeddings"),
+                          clEnumValN(IR2VecKind::Flowaware, "flowaware",
+                                     "Generates flow-aware embeddings")),
+               cl::init(IR2VecKind::Symbolic), cl::cat(IR2VecAnalysisCategory));
+
+// FIXME: Use a default vocab when not specified
+static cl::opt<std::string>
+    VocabFile("ir2vec-vocab-path", cl::Optional,
+              cl::desc("Path to the vocabulary file for IR2Vec"), cl::init(""),
+              cl::cat(IR2VecAnalysisCategory));
+
+AnalysisKey IR2VecVocabAnalysis::Key;
+AnalysisKey IR2VecAnalysis::Key;
+
+// ==----------------------------------------------------------------------===//
+// Embeddings and its subclasses
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// Embeddings provides the interface to generate vector representations for
+/// instructions, basic blocks, and functions. The vector
+/// representations are generated using IR2Vec algorithms.
+///
+/// The Embeddings class is an abstract class and it is intended to be
+/// subclassed for different IR2Vec algorithms like Symbolic and Flow-aware.
+class Embeddings {
+protected:
+  const Function &F;
+  Vocab Vocabulary;
+
+  /// Weights for different entities (like opcode, arguments, types)
+  /// in the IR instructions to generate the vector representation.
+  // FIXME: Defaults to the values used in the original algorithm. Can be
+  // parameterized later.
+  float WO = 1.0, WT = 0.5, WA = 0.2;
+
+  /// Dimension of the vector representation; captured from the input vocabulary
+  unsigned DIM = 300;
+
+  // Utility maps - these are used to store the vector representations of
+  // instructions, basic blocks and functions.
+  Embedding FuncVector;
+  SmallMapVector<const BasicBlock *, Embedding, 16> BBVecMap;
+  SmallMapVector<const Instruction *, Embedding, 128> InstVecMap;
+
+  Embeddings(const Function &F, const Vocab &Vocabulary, unsigned DIM)
+      : F(F), Vocabulary(Vocabulary), DIM(DIM) {}
+
+  /// Lookup vocabulary for a given Key. If the key is not found, it returns a
+  /// zero vector.
+  Embedding lookupVocab(const std::string &Key);
+
+public:
+  virtual ~Embeddings() = default;
+
+  /// Top level function to compute embeddings. Given a function, it
+  /// generates embeddings for all the instructions and basic blocks in that
+  /// function. Logic of computing the embeddings is specific to the kind of
+  /// embeddings being computed.
+  virtual void computeEmbeddings() = 0;
+
+  /// Returns a map containing instructions and the corresponding vector
+  /// representations for a given module corresponding to the IR2Vec
+  /// algorithm.
+  const SmallMapVector<const Instruction *, Embedding, 128> &
+  getInstVecMap() const {
+    return InstVecMap;
+  }
+
+  /// Returns a map containing basic block and the corresponding vector
+  /// representations for a given module corresponding to the IR2Vec
+  /// algorithm.
+  const SmallMapVector<const BasicBlock *, Embedding, 16> &getBBVecMap() const {
+    return BBVecMap;
+  }
+
+  /// Returns the vector representation for a given function corresponding to
+  /// the IR2Vec algorithm.
+  const Embedding &getFunctionVector() const { return FuncVector; }
+};
+
+/// Class for computing the Symbolic embeddings of IR2Vec
+class Symbolic : public Embeddings {
+private:
+  /// Utility function to compute the vector representation for a given basic
+  /// block.
+  Embedding computeBB2Vec(const BasicBlock &BB);
+
+  /// Utility function to compute the vector representation for a given
+  /// function.
+  Embedding computeFunc2Vec();
+
+public:
+  Symbolic(const Function &F, const Vocab &Vocabulary, unsigned DIM)
+      : Embeddings(F, Vocabulary, DIM) {
+    FuncVector = Embedding(DIM, 0);
+  }
+  void computeEmbeddings() override;
+};
+
+/// Scales the vector Vec by Factor
+void scaleVector(Embedding &Vec, const float Factor) {
+  std::transform(Vec.begin(), Vec.end(), Vec.begin(),
+                 [Factor](double X) { return X * Factor; });
+}
+
+/// Adds two vectors: Vec += Vec2
+void addVectors(Embedding &Vec, const Embedding &Vec2) {
+  std::transform(Vec.begin(), Vec.end(), Vec2.begin(), Vec.begin(),
+                 std::plus<double>());
+}
+
+// FIXME: Currently lookups are string based. Use numeric Keys
+// for efficiency.
+Embedding Embeddings::lookupVocab(const std::string &Key) {
+  Embedding Vec(DIM, 0);
+  // FIXME: Use zero vectors in vocab and assert failure for
+  // unknown entities rather than silently returning zeroes here.
+  if (Vocabulary.find(Key) == Vocabulary.end()) {
+    LLVM_DEBUG(errs() << "cannot find key in map : " << Key << "\n");
+    DataMissCounter++;
+  } else {
+    Vec = Vocabulary[Key];
+  }
+  return Vec;
+}
+
+void Symbolic::computeEmbeddings() {
+  if (F.isDeclaration())
+    return;
+  for (auto &BB : F) {
+    auto It = BBVecMap.find(&BB);
+    if (It != BBVecMap.end())
+      continue;
+    BBVecMap[&BB] = computeBB2Vec(BB);
+    addVectors(FuncVector, BBVecMap[&BB]);
+  }
+}
+
+Embedding Symbolic::computeBB2Vec(const BasicBlock &BB) {
+  Embedding BBVector(DIM, 0);
+
+  for (auto &I : BB) {
+    Embedding InstVector(DIM, 0);
+
+    auto Vec = lookupVocab(I.getOpcodeName());
+    scaleVector(Vec, WO);
+    addVectors(InstVector, Vec);
+
+    auto Type = I.getType();
+    if (Type->isVoidTy()) {
+      Vec = lookupVocab("voidTy");
+    } else if (Type->isFloatingPointTy()) {
+      Vec = lookupVocab("floatTy");
+    } else if (Type->isIntegerTy()) {
+      Vec = lookupVocab("integerTy");
+    } else if (Type->isFunctionTy()) {
+      Vec = lookupVocab("functionTy");
+    } else if (Type->isStructTy()) {
+      Vec = lookupVocab("structTy");
+    } else if (Type->isArrayTy()) {
+      Vec = lookupVocab("arrayTy");
+    } else if (Type->isPointerTy()) {
+      Vec = lookupVocab("pointerTy");
+    } else if (Type->isVectorTy()) {
+      Vec = lookupVocab("vectorTy");
+    } else if (Type->isEmptyTy()) {
+      Vec = lookupVocab("emptyTy");
+    } else if (Type->isLabelTy()) {
+      Vec = lookupVocab("labelTy");
+    } else if (Type->isTokenTy()) {
+      Vec = lookupVocab("tokenTy");
+    } else if (Type->isMetadataTy()) {
+      Vec = lookupVocab("metadataTy");
+    } else {
+      Vec = lookupVocab("unknownTy");
+    }
+    scaleVector(Vec, WT);
+    addVectors(InstVector, Vec);
+
+    for (auto &Op : I.operands()) {
+      Embedding Vec;
+      if (isa<Function>(Op)) {
+        Vec = lookupVocab("function");
+      } else if (isa<PointerType>(Op->getType())) {
+        Vec = lookupVocab("pointer");
+      } else if (isa<Constant>(Op)) {
+        Vec = lookupVocab("constant");
+      } else {
+        Vec = lookupVocab("variable");
+      }
+      scaleVector(Vec, WA);
+      addVectors(InstVector, Vec);
+    }
+    InstVecMap[&I] = InstVector;
+    addVectors(BBVector, InstVector);
+  }
+  return BBVector;
+}
+} // namespace
+
+// ==----------------------------------------------------------------------===//
+// IR2VecVocabResult and IR2VecVocabAnalysis
+//===----------------------------------------------------------------------===//
+
+IR2VecVocabResult::IR2VecVocabResult(ir2vec::Vocab &&Vocabulary, unsigned Dim)
+    : Vocabulary(std::move(Vocabulary)), Valid(true), DIM(Dim) {}
+
+const ir2vec::Vocab &IR2VecVocabResult::getVocabulary() const {
+  assert(Valid);
+  return Vocabulary;
+}
+
+// For now, assume vocabulary is stable unless explicitly invalidated.
+bool IR2VecVocabResult::invalidate(Module &M, const PreservedAnalyses &PA,
+                                   ModuleAnalysisManager::Invalidator &Inv) {
+  auto PAC = PA.getChecker<IR2VecVocabAnalysis>();
+  return !(PAC.preservedWhenStateless());
+}
+
+// FIXME: Make this optional. We can avoid file reads
+// by auto-generating the vocabulary during the build time.
+Error IR2VecVocabAnalysis::readVocabulary() {
+  auto BufOrError = MemoryBuffer::getFileOrSTDIN(VocabFile, /*IsText=*/true);
+  if (!BufOrError) {
+    return createFileError(VocabFile, BufOrError.getError());
+  }
+  auto Content = BufOrError.get()->getBuffer();
+  json::Path::Root Path("");
+  Expected<json::Value> ParsedVocabValue = json::parse(Content);
+  if (!ParsedVocabValue)
+    return ParsedVocabValue.takeError();
+
+  bool Res = json::fromJSON(*ParsedVocabValue, Vocabulary, Path);
+  if (!Res) {
+    return createStringError(errc::illegal_byte_sequence,
+                             "Unable to parse the vocabulary");
+  }
+  assert(Vocabulary.size() > 0 && "Vocabulary is empty");
+
+  unsigned Dim = Vocabulary.begin()->second.size();
+  assert(Dim > 0 && "Dimension of vocabulary is zero");
+  assert(std::all_of(Vocabulary.begin(), Vocabulary.end(),
+                     [Dim](const std::pair<StringRef, Embedding> &Entry) {
+                       return Entry.second.size() == Dim;
+                     }) &&
+         "All vectors in the vocabulary are not of the same dimension");
+  this->DIM = Dim;
+  return Error::success();
+}
+
+IR2VecVocabAnalysis::Result
+IR2VecVocabAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
+  auto Ctx = &M.getContext();
+  if (VocabFile.empty()) {
+    // FIXME: Use default vocabulary
+    Ctx->emitError("IR2Vec vocabulary file path not specified");
+    return IR2VecVocabResult(); // Return invalid result
+  }
+  if (auto Err = readVocabulary()) {
+    handleAllErrors(std::move(Err), [&](const ErrorInfoBase &EI) {
+      Ctx->emitError("Error reading vocabulary: " + EI.message());
+    });
+    return IR2VecVocabResult();
+  }
+  return IR2VecVocabResult(std::move(Vocabulary), DIM);
+}
+
+// ==----------------------------------------------------------------------===//
+// IR2VecResult and IR2VecAnalysis
+//===----------------------------------------------------------------------===//
+
+IR2VecResult::IR2VecResult(
+    SmallMapVector<const Instruction *, Embedding, 128> &&InstMap,
+    SmallMapVector<const BasicBlock *, Embedding, 16> &&BBMap,
+    Embedding &&FuncVector, unsigned Dim)
+    : InstVecMap(std::move(InstMap)), BBVecMap(std::move(BBMap)),
+      FuncVector(std::move(FuncVector)), DIM(Dim), Valid(true) {}
+
+const SmallMapVector<const Instruction *, Embedding, 128> &
+IR2VecResult::getInstVecMap() const {
+  assert(Valid);
+  return InstVecMap;
+}
+const SmallMapVector<const BasicBlock *, Embedding, 16> &
+IR2VecResult::getBBVecMap() const {
+  assert(Valid);
+  return BBVecMap;
+}
+const Embedding &IR2VecResult::getFunctionVector() const {
+  assert(Valid);
+  return FuncVector;
+}
+unsigned IR2VecResult::getDimension() const { return DIM; }
+
+IR2VecAnalysis::Result IR2VecAnalysis::run(Function &F,
+                                           FunctionAnalysisManager &FAM) {
+  auto *VocabRes = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F)
+                       .getCachedResult<IR2VecVocabAnalysis>(*F.getParent());
+  auto Ctx = &F.getContext();
+  if (!VocabRes->isValid()) {
+    Ctx->emitError("IR2Vec vocabulary is invalid");
+    return IR2VecResult();
+  }
+
+  auto Dim = VocabRes->getDimension();
+  if (Dim <= 0) {
+    Ctx->emitError("IR2Vec vocabulary dimension is zero");
+    return IR2VecResult();
+  }
+
+  auto Vocabulary = VocabRes->getVocabulary();
+  std::unique_ptr<Embeddings> Emb;
+  switch (IR2VecMode) {
+  case IR2VecKind::Symbolic:
+    Emb = std::make_unique<Symbolic>(F, Vocabulary, Dim);
+    break;
+  case IR2VecKind::Flowaware:
+    // FIXME: Add support for flow-aware embeddings
+    llvm_unreachable("Flow-aware embeddings are not supported yet");
+    break;
+  default:
+    llvm_unreachable("Invalid IR2Vec mode");
+  }
+  Emb->computeEmbeddings();
+  auto InstMap = Emb->getInstVecMap();
+  auto BBMap = Emb->getBBVecMap();
+  auto FuncVec = Emb->getFunctionVector();
+  return IR2VecResult(std::move(InstMap), std::move(BBMap), std::move(FuncVec),
+                      Dim);
+}
+
+// ==----------------------------------------------------------------------===//
+// IR2VecPrinterPass
+//===----------------------------------------------------------------------===//
+
+void IR2VecPrinterPass::printVector(const Embedding &Vec) const {
+  OS << " [";
+  for (auto &Elem : Vec)
+    OS << " " << format("%.2f", Elem) << " ";
+  OS << "]\n";
+}
+
+PreservedAnalyses IR2VecPrinterPass::run(Module &M,
+                                         ModuleAnalysisManager &MAM) {
+  auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(M);
+  assert(IR2VecVocabResult.isValid() && "Vocab is invalid");
+
+  for (Function &F : M) {
+    auto &FAM =
+        MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
+
+    auto IR2VecRes = FAM.getResult<IR2VecAnalysis>(F);
+
+    if (!IR2VecRes.isValid()) {
+      auto Ctx = &F.getContext();
+      Ctx->emitError("IR2Vec embeddings are invalid");
+      return PreservedAnalyses::all();
+    }
+
+    OS << "IR2Vec embeddings for function " << F.getName() << ":\n";
+    OS << "Function vector: ";
+    printVector(IR2VecRes.getFunctionVector());
+
+    OS << "Basic block vectors:\n";
+    for (const auto &BBVector : IR2VecRes.getBBVecMap()) {
+      OS << "Basic block: " << BBVector.first...
[truncated]

Copy link
Contributor

@nikic nikic left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Marking this as changes requested, as this still needs an RFC on discourse before being merged.

@svkeerthy svkeerthy requested a review from nikic May 11, 2025 23:24
@svkeerthy
Copy link
Author

@efriedma-quic, Added the documentation as per your suggestion in the RFC. Please have a look.
(Tagging here as I am unable to make changes or add reviewers yet)

Copy link
Member

@mtrofin mtrofin left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

some doc nits

// FIXME: Currently lookups are string based. Use numeric Keys
// for efficiency.
auto Type = I.getType();
if (Type->isVoidTy()) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider using TypeSwitch?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If the string were slightly changed, one could at least make them one line per type with a macro.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Created a macro and refactored the code. About half of the types that are checked here are not classes. Hence did not use TypeSwitch.

Comment on lines 90 to 94
IR2VecResult(
const SmallMapVector<const Instruction *, ir2vec::Embedding, 128>
&&InstMap,
const SmallMapVector<const BasicBlock *, ir2vec::Embedding, 16> &&BBMap,
const ir2vec::Embedding &&FuncVector);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This constructor is simply to accept an "unpacked" Embeddings object. You can also just accept an Embeddings& and eliminate all the unpacking. Or, make Embeddings an IR2VecResult, so you don't need to translate (and duplicate) stuff at all.

Copy link
Author

@svkeerthy svkeerthy May 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Removed IR2VecResult completely. Please see my comment on removing IR2VecAnalysis.

@svkeerthy
Copy link
Author

Currently, we have simplified the code by removing IR2VecAnalysis and directly using the ir2vec::Embedder.

Motivation

  • Our initial intention for IR2VecAnalysis was to leverage LLVM's analysis caching. However, recent experiments indicate that for our current use cases, recomputing embedding vectors is often cheaper than the overhead of caching them.
  • This change also addresses earlier review feedback (related to simplifying how Embeddings objects are handled) by moving towards using the ir2vec::Embedder more directly as the primary container for embedding results.

@svkeerthy svkeerthy requested review from snehasish and jdoerfert May 15, 2025 15:49
Copy link
Contributor

@boomanaiden154 boomanaiden154 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A couple nits, otherwise LGTM. Please wait for approval from others who have commented on the patch before merging.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

7 participants