[llvm] [llvm-ir2vec] adding function-embedding map API to ir2vec python bindings (PR #177166)
Nishant Sachdeva via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 28 02:05:15 PST 2026
https://github.com/nishant-sachdeva updated https://github.com/llvm/llvm-project/pull/177166
>From daa538fae9666837e03a31e628d0e70cce52264d Mon Sep 17 00:00:00 2001
From: nishant-sachdeva <nishant.sachdeva at research.iiit.ac.in>
Date: Wed, 21 Jan 2026 10:47:10 +0530
Subject: [PATCH 1/5] Adding an initEmbedding API to ir2vec python bindings
---
.../llvm-ir2vec/bindings/ir2vec-bindings.py | 21 ++++-
.../tools/llvm-ir2vec/Bindings/CMakeLists.txt | 3 +-
llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp | 87 ++++++++++++++++++-
3 files changed, 106 insertions(+), 5 deletions(-)
diff --git a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
index 459f353a478cb..e6734d2055cd8 100644
--- a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
+++ b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
@@ -1,7 +1,22 @@
-# RUN: env PYTHONPATH=%llvm_lib_dir %python %s
+# RUN: rm -rf %t.ll
+# RUN: echo "define i32 @add(i32 %%a, i32 %%b) {" > %t.ll
+# RUN: echo "entry:" >> %t.ll
+# RUN: echo " %%sum = add i32 %%a, %%b" >> %t.ll
+# RUN: echo " ret i32 %%sum" >> %t.ll
+# RUN: echo "}" >> %t.ll
+# RUN: env PYTHONPATH=%llvm_lib_dir %python %s %t.ll %ir2vec_test_vocab_dir/dummy_3D_nonzero_opc_vocab.json | FileCheck %s
+import sys
import ir2vec
-print("SUCCESS: Module imported")
+ll_file = sys.argv[1]
+vocab_path = sys.argv[2]
-# CHECK: SUCCESS: Module imported
+tool = ir2vec.initEmbedding(filename=ll_file, mode="sym", vocab_path=vocab_path)
+
+if tool is not None:
+ print("SUCCESS: Tool initialized")
+ print(f"Tool type: {type(tool).__name__}")
+
+# CHECK: SUCCESS: Tool initialized
+# CHECK: Tool type: IR2VecTool
diff --git a/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt b/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
index 5df3720a24777..677208774f5a1 100644
--- a/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
+++ b/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
@@ -1,4 +1,4 @@
-find_package(Python COMPONENTS Interpreter Development.Module REQUIRED)
+find_package(Python ${Python3_VERSION} EXACT COMPONENTS Interpreter Development.Module REQUIRED)
execute_process(
COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
@@ -10,5 +10,6 @@ set_target_properties(LLVMEmbUtils PROPERTIES POSITION_INDEPENDENT_CODE ON)
nanobind_add_module(ir2vec MODULE PyIR2Vec.cpp)
target_link_libraries(ir2vec PRIVATE LLVMEmbUtils)
+target_include_directories(ir2vec PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/..)
message(STATUS "Python bindings for llvm-ir2vec will be built with nanobind")
diff --git a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
index b3a46d429b6d4..24297d15caaf1 100644
--- a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
+++ b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
@@ -7,7 +7,92 @@
//===----------------------------------------------------------------------===//
#include <nanobind/nanobind.h>
+#include <nanobind/stl/string.h>
+#include <nanobind/stl/unique_ptr.h>
+
+#include "lib/Utils.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IRReader/IRReader.h"
+#include "llvm/Support/SourceMgr.h"
+
+#include <fstream>
+#include <memory>
+#include <string>
namespace nb = nanobind;
+using namespace llvm;
+using namespace llvm::ir2vec;
+
+namespace llvm {
+namespace ir2vec {
+void setIR2VecVocabPath(StringRef Path);
+StringRef getIR2VecVocabPath();
+} // namespace ir2vec
+} // namespace llvm
+
+namespace {
+
+bool fileNotValid(const std::string &Filename) {
+ std::ifstream F(Filename, std::ios_base::in | std::ios_base::binary);
+ return !F.good();
+}
+
+std::unique_ptr<Module> getLLVMIR(const std::string &Filename,
+ LLVMContext &Context) {
+ SMDiagnostic Err;
+ auto M = parseIRFile(Filename, Err, Context);
+ if (!M)
+ throw std::runtime_error("Failed to parse IR file.");
+ return M;
+}
+
+class PyIR2VecTool {
+private:
+ std::unique_ptr<LLVMContext> Ctx;
+ std::unique_ptr<Module> M;
+ std::unique_ptr<IR2VecTool> Tool;
+
+public:
+ PyIR2VecTool(const std::string &Filename, const std::string &Mode,
+ const std::string &VocabPath) {
+ if (fileNotValid(Filename))
+ throw std::runtime_error("Invalid file path");
+
+ if (Mode != "sym" && Mode != "fa")
+ throw std::runtime_error("Invalid mode. Use 'sym' or 'fa'");
+
+ if (VocabPath.empty())
+ throw std::runtime_error("Error - Empty Vocab Path not allowed");
+
+ setIR2VecVocabPath(VocabPath);
+
+ Ctx = std::make_unique<LLVMContext>();
+ M = getLLVMIR(Filename, *Ctx);
+ Tool = std::make_unique<IR2VecTool>(*M);
+
+ bool Ok = Tool->initializeVocabulary();
+ if (!Ok)
+ throw std::runtime_error("Failed to initialize IR2Vec vocabulary");
+ }
+};
+
+} // namespace
+
+NB_MODULE(ir2vec, m) {
+ m.doc() = std::string("Python bindings for ") + ToolName;
+
+ nb::class_<PyIR2VecTool>(m, "IR2VecTool")
+ .def(nb::init<const std::string &, const std::string &,
+ const std::string &>(),
+ nb::arg("filename"), nb::arg("mode"), nb::arg("vocab_path"));
-NB_MODULE(ir2vec, m) { m.doc() = "Python bindings for IR2Vec"; }
+ m.def(
+ "initEmbedding",
+ [](const std::string &filename, const std::string &mode,
+ const std::string &vocab_path) {
+ return std::make_unique<PyIR2VecTool>(filename, mode, vocab_path);
+ },
+ nb::arg("filename"), nb::arg("mode") = "sym", nb::arg("vocab_path"),
+ nb::rv_policy::take_ownership);
+}
>From 025fd318cdf72040cc77de3b6776a427fc06556a Mon Sep 17 00:00:00 2001
From: nishant-sachdeva <nishant.sachdeva at research.iiit.ac.in>
Date: Tue, 27 Jan 2026 14:57:43 +0530
Subject: [PATCH 2/5] initEmbedding API prepared with rebase on main
---
llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp | 18 +++++-------------
1 file changed, 5 insertions(+), 13 deletions(-)
diff --git a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
index 24297d15caaf1..f2eac46eb5f02 100644
--- a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
+++ b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
@@ -24,13 +24,6 @@ namespace nb = nanobind;
using namespace llvm;
using namespace llvm::ir2vec;
-namespace llvm {
-namespace ir2vec {
-void setIR2VecVocabPath(StringRef Path);
-StringRef getIR2VecVocabPath();
-} // namespace ir2vec
-} // namespace llvm
-
namespace {
bool fileNotValid(const std::string &Filename) {
@@ -63,17 +56,16 @@ class PyIR2VecTool {
throw std::runtime_error("Invalid mode. Use 'sym' or 'fa'");
if (VocabPath.empty())
- throw std::runtime_error("Error - Empty Vocab Path not allowed");
-
- setIR2VecVocabPath(VocabPath);
+ throw std::runtime_error("Empty Vocab Path not allowed");
Ctx = std::make_unique<LLVMContext>();
M = getLLVMIR(Filename, *Ctx);
Tool = std::make_unique<IR2VecTool>(*M);
- bool Ok = Tool->initializeVocabulary();
- if (!Ok)
- throw std::runtime_error("Failed to initialize IR2Vec vocabulary");
+ if (auto Err = Tool->initializeVocabulary(VocabPath)) {
+ throw std::runtime_error("Failed to initialize IR2Vec vocabulary: " +
+ toString(std::move(Err)));
+ }
}
};
>From ca2a4d749205b984cc4f9ea1a3e89f3a3cd644cd Mon Sep 17 00:00:00 2001
From: nishant-sachdeva <nishant.sachdeva at research.iiit.ac.in>
Date: Wed, 28 Jan 2026 15:34:51 +0530
Subject: [PATCH 3/5] Refining version setting and requirements for python3
around nanobind
---
llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt b/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
index 677208774f5a1..efa890f5025dc 100644
--- a/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
+++ b/llvm/tools/llvm-ir2vec/Bindings/CMakeLists.txt
@@ -1,4 +1,5 @@
-find_package(Python ${Python3_VERSION} EXACT COMPONENTS Interpreter Development.Module REQUIRED)
+set(BINDINGS_MINIMUM_PYTHON_VERSION 3.10)
+find_package(Python ${BINDINGS_MINIMUM_PYTHON_VERSION} EXACT COMPONENTS Interpreter Development.Module REQUIRED)
execute_process(
COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
>From 00bcc13ebb3a83ac2d0067f9680bbf322d693103 Mon Sep 17 00:00:00 2001
From: nishant-sachdeva <nishant.sachdeva at research.iiit.ac.in>
Date: Wed, 21 Jan 2026 18:20:16 +0530
Subject: [PATCH 4/5] Adding getFuncEmbMap functionality to ir2vec python
bindings
---
.../llvm-ir2vec/bindings/ir2vec-bindings.py | 41 +++++++++++++++++++
llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp | 38 ++++++++++++++++-
llvm/tools/llvm-ir2vec/lib/Utils.cpp | 35 ++++++++++++++++
llvm/tools/llvm-ir2vec/lib/Utils.h | 9 ++++
4 files changed, 121 insertions(+), 2 deletions(-)
diff --git a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
index e6734d2055cd8..be25284a70fff 100644
--- a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
+++ b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
@@ -4,6 +4,30 @@
# RUN: echo " %%sum = add i32 %%a, %%b" >> %t.ll
# RUN: echo " ret i32 %%sum" >> %t.ll
# RUN: echo "}" >> %t.ll
+# RUN: echo "" >> %t.ll
+# RUN: echo "define i32 @multiply(i32 %%x, i32 %%y) {" >> %t.ll
+# RUN: echo "entry:" >> %t.ll
+# RUN: echo " %%prod = mul i32 %%x, %%y" >> %t.ll
+# RUN: echo " ret i32 %%prod" >> %t.ll
+# RUN: echo "}" >> %t.ll
+# RUN: echo "" >> %t.ll
+# RUN: echo "define i32 @conditional(i32 %%n) {" >> %t.ll
+# RUN: echo "entry:" >> %t.ll
+# RUN: echo " %%cmp = icmp sgt i32 %%n, 0" >> %t.ll
+# RUN: echo " br i1 %%cmp, label %%positive, label %%negative" >> %t.ll
+# RUN: echo "" >> %t.ll
+# RUN: echo "positive:" >> %t.ll
+# RUN: echo " %%pos_val = add i32 %%n, 10" >> %t.ll
+# RUN: echo " br label %%exit" >> %t.ll
+# RUN: echo "" >> %t.ll
+# RUN: echo "negative:" >> %t.ll
+# RUN: echo " %%neg_val = sub i32 %%n, 10" >> %t.ll
+# RUN: echo " br label %%exit" >> %t.ll
+# RUN: echo "" >> %t.ll
+# RUN: echo "exit:" >> %t.ll
+# RUN: echo " %%result = phi i32 [ %%pos_val, %%positive ], [ %%neg_val, %%negative ]" >> %t.ll
+# RUN: echo " ret i32 %%result" >> %t.ll
+# RUN: echo "}" >> %t.ll
# RUN: env PYTHONPATH=%llvm_lib_dir %python %s %t.ll %ir2vec_test_vocab_dir/dummy_3D_nonzero_opc_vocab.json | FileCheck %s
import sys
@@ -18,5 +42,22 @@
print("SUCCESS: Tool initialized")
print(f"Tool type: {type(tool).__name__}")
+ # Test getFuncEmbMap
+ func_emb_map = tool.getFuncEmbMap()
+ print(f"Number of functions: {len(func_emb_map)}")
+
+ # Check that all three functions are present
+ expected_funcs = ["add", "multiply", "conditional"]
+ for func_name in expected_funcs:
+ if func_name in func_emb_map:
+ emb = func_emb_map[func_name]
+ print(f"Function '{func_name}': embedding shape = {emb.shape}")
+ else:
+ print(f"ERROR: Function '{func_name}' not found")
+
# CHECK: SUCCESS: Tool initialized
# CHECK: Tool type: IR2VecTool
+# CHECK: Number of functions: 3
+# CHECK: Function 'add': embedding shape =
+# CHECK: Function 'multiply': embedding shape =
+# CHECK: Function 'conditional': embedding shape =
diff --git a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
index f2eac46eb5f02..6307af8dc2baa 100644
--- a/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
+++ b/llvm/tools/llvm-ir2vec/Bindings/PyIR2Vec.cpp
@@ -7,6 +7,9 @@
//===----------------------------------------------------------------------===//
#include <nanobind/nanobind.h>
+#include <nanobind/ndarray.h>
+#include <nanobind/stl/map.h>
+#include <nanobind/stl/pair.h>
#include <nanobind/stl/string.h>
#include <nanobind/stl/unique_ptr.h>
@@ -45,6 +48,7 @@ class PyIR2VecTool {
std::unique_ptr<LLVMContext> Ctx;
std::unique_ptr<Module> M;
std::unique_ptr<IR2VecTool> Tool;
+ IR2VecKind EmbKind;
public:
PyIR2VecTool(const std::string &Filename, const std::string &Mode,
@@ -52,8 +56,13 @@ class PyIR2VecTool {
if (fileNotValid(Filename))
throw std::runtime_error("Invalid file path");
- if (Mode != "sym" && Mode != "fa")
+ EmbKind = [](const std::string &Mode) -> IR2VecKind {
+ if (Mode == "sym")
+ return IR2VecKind::Symbolic;
+ if (Mode == "fa")
+ return IR2VecKind::FlowAware;
throw std::runtime_error("Invalid mode. Use 'sym' or 'fa'");
+ }(Mode);
if (VocabPath.empty())
throw std::runtime_error("Empty Vocab Path not allowed");
@@ -67,6 +76,27 @@ class PyIR2VecTool {
toString(std::move(Err)));
}
}
+
+ nb::dict getFuncEmbMap() {
+ auto result = Tool->getFunctionEmbeddings(EmbKind);
+ nb::dict nb_result;
+
+ for (const auto &[func_ptr, embedding] : result) {
+ std::string func_name = func_ptr->getName().str();
+ auto data = embedding.getData();
+ size_t shape[1] = {data.size()};
+ double *data_ptr = new double[data.size()];
+ std::copy(data.data(), data.data() + data.size(), data_ptr);
+
+ auto nb_array = nb::ndarray<nb::numpy, double>(
+ data_ptr, {data.size()}, nb::capsule(data_ptr, [](void *p) noexcept {
+ delete[] static_cast<double *>(p);
+ }));
+ nb_result[nb::str(func_name.c_str())] = nb_array;
+ }
+
+ return nb_result;
+ }
};
} // namespace
@@ -77,7 +107,11 @@ NB_MODULE(ir2vec, m) {
nb::class_<PyIR2VecTool>(m, "IR2VecTool")
.def(nb::init<const std::string &, const std::string &,
const std::string &>(),
- nb::arg("filename"), nb::arg("mode"), nb::arg("vocab_path"));
+ nb::arg("filename"), nb::arg("mode"), nb::arg("vocab_path"))
+ .def("getFuncEmbMap", &PyIR2VecTool::getFuncEmbMap,
+ "Generate function-level embeddings for all functions\n"
+ "Returns: dict[str, ndarray[float64]] - "
+ "{function_name: embedding}");
m.def(
"initEmbedding",
diff --git a/llvm/tools/llvm-ir2vec/lib/Utils.cpp b/llvm/tools/llvm-ir2vec/lib/Utils.cpp
index 190d9259e45b3..4e8589885e019 100644
--- a/llvm/tools/llvm-ir2vec/lib/Utils.cpp
+++ b/llvm/tools/llvm-ir2vec/lib/Utils.cpp
@@ -151,6 +151,41 @@ void IR2VecTool::writeEntitiesToStream(raw_ostream &OS) {
OS << Entities[EntityID] << '\t' << EntityID << '\n';
}
+std::pair<const Function *, Embedding>
+IR2VecTool::getFunctionEmbedding(const Function &F, IR2VecKind Kind) const {
+ assert(Vocab && Vocab->isValid() && "Vocabulary not initialized");
+
+ if (F.isDeclaration())
+ return {nullptr, Embedding()};
+
+ auto Emb = Embedder::create(Kind, F, *Vocab);
+ if (!Emb) {
+ return {nullptr, Embedding()};
+ }
+
+ auto FuncVec = Emb->getFunctionVector();
+
+ return {&F, std::move(FuncVec)};
+}
+
+FuncEmbMap IR2VecTool::getFunctionEmbeddings(IR2VecKind Kind) const {
+ assert(Vocab && Vocab->isValid() && "Vocabulary not initialized");
+
+ FuncEmbMap Result;
+
+ for (const Function &F : M) {
+ if (F.isDeclaration())
+ continue;
+
+ auto Emb = getFunctionEmbedding(F, Kind);
+ if (Emb.first != nullptr) {
+ Result.try_emplace(Emb.first, std::move(Emb.second));
+ }
+ }
+
+ return Result;
+}
+
void IR2VecTool::writeEmbeddingsToStream(raw_ostream &OS,
EmbeddingLevel Level) const {
if (!Vocab->isValid()) {
diff --git a/llvm/tools/llvm-ir2vec/lib/Utils.h b/llvm/tools/llvm-ir2vec/lib/Utils.h
index d9715b03c3082..d115d9a26ca90 100644
--- a/llvm/tools/llvm-ir2vec/lib/Utils.h
+++ b/llvm/tools/llvm-ir2vec/lib/Utils.h
@@ -16,6 +16,7 @@
#define LLVM_TOOLS_LLVM_IR2VEC_UTILS_UTILS_H
#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/DenseMap.h"
#include "llvm/Analysis/IR2Vec.h"
#include "llvm/CodeGen/MIR2Vec.h"
#include "llvm/CodeGen/MIRParser/MIRParser.h"
@@ -72,6 +73,7 @@ struct TripletResult {
/// Entity mappings: [entity_name]
using EntityList = std::vector<std::string>;
+using FuncEmbMap = DenseMap<const Function *, ir2vec::Embedding>;
namespace ir2vec {
@@ -112,6 +114,13 @@ class IR2VecTool {
/// Returns EntityList containing all entity strings
static EntityList collectEntityMappings();
+ // Get embedding for a single function
+ std::pair<const Function *, Embedding>
+ getFunctionEmbedding(const Function &F, IR2VecKind Kind) const;
+
+ /// Get embeddings for all functions in the module
+ FuncEmbMap getFunctionEmbeddings(IR2VecKind Kind) const;
+
/// Dump entity ID to string mappings
static void writeEntitiesToStream(raw_ostream &OS);
>From 5ab0fff5e32517fbd3391b5cb1c4dd8495d05cb5 Mon Sep 17 00:00:00 2001
From: nishant-sachdeva <nishant.sachdeva at research.iiit.ac.in>
Date: Wed, 21 Jan 2026 19:36:48 +0530
Subject: [PATCH 5/5] Changing unit-test structure for function embeddings
---
.../llvm-ir2vec/bindings/ir2vec-bindings.py | 28 +++++++++----------
1 file changed, 13 insertions(+), 15 deletions(-)
diff --git a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
index be25284a70fff..986cc2979bd9e 100644
--- a/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
+++ b/llvm/test/tools/llvm-ir2vec/bindings/ir2vec-bindings.py
@@ -40,24 +40,22 @@
if tool is not None:
print("SUCCESS: Tool initialized")
- print(f"Tool type: {type(tool).__name__}")
# Test getFuncEmbMap
+ print("\n=== Function Embeddings ===")
func_emb_map = tool.getFuncEmbMap()
- print(f"Number of functions: {len(func_emb_map)}")
- # Check that all three functions are present
- expected_funcs = ["add", "multiply", "conditional"]
- for func_name in expected_funcs:
- if func_name in func_emb_map:
- emb = func_emb_map[func_name]
- print(f"Function '{func_name}': embedding shape = {emb.shape}")
- else:
- print(f"ERROR: Function '{func_name}' not found")
+ # Sorting the function names for deterministic output
+ for func_name in sorted(func_emb_map.keys()):
+ emb = func_emb_map[func_name]
+ print(f"Function: {func_name}")
+ print(f" Embedding: {emb.tolist()}")
# CHECK: SUCCESS: Tool initialized
-# CHECK: Tool type: IR2VecTool
-# CHECK: Number of functions: 3
-# CHECK: Function 'add': embedding shape =
-# CHECK: Function 'multiply': embedding shape =
-# CHECK: Function 'conditional': embedding shape =
+# CHECK: === Function Embeddings ===
+# CHECK: Function: add
+# CHECK-NEXT: Embedding: [38.0, 40.0, 42.0]
+# CHECK: Function: conditional
+# CHECK-NEXT: Embedding: [413.20000000298023, 421.20000000298023, 429.20000000298023]
+# CHECK: Function: multiply
+# CHECK-NEXT: Embedding: [50.0, 52.0, 54.0]
More information about the llvm-commits
mailing list