llvm-nanobind 21.1.6.2

LLVM-C Python bindings with nanobind.

llvm-nanobind

Python bindings for the LLVM-C API using nanobind.

This project provides a Pythonic interface to LLVM's compiler infrastructure, enabling you to build compilers, analyzers, and code transformation tools in Python.

Current status: Experimental but usable. The bindings target LLVM 21.1.6 and are published as llvm-nanobind on PyPI (import name: llvm). Wheels bundle LLVM and are built/tested in CI for Linux x86_64/aarch64, macOS arm64, and Windows x86_64. The API is still evolving; expect breaking changes until a stable release.

Version numbers track LLVM: 21.1.6.1 means LLVM 21.1.6 plus binding/package revision 1.

Note: This project is 90%+ vibe coded. It is mostly an experiment to see what LLMs can do when you set things up properly.

Installation

Released wheels bundle LLVM for supported platforms, so normal installation is:

pip install llvm-nanobind
python -c "import llvm; print(llvm)"

For source/development builds, see Development.

See llvm-nanobind-example for a simple example project.

Quick Start

import llvm

# Create a simple function that returns 42.
with llvm.create_context() as ctx:
    i32 = ctx.types.i32
    fn_type = ctx.types.function(i32, [])

    with ctx.create_module("example") as mod:
        fn = mod.add_function("get_answer", fn_type)
        entry = fn.append_basic_block("entry")

        with entry.create_builder() as builder:
            builder.ret(i32.constant(42))

        assert mod.verify(), mod.verification_error
        print(mod)

This exact snippet is available as examples/quick_start.py.

Examples

Runnable examples live in examples/ and are covered by tests/test_examples.py.

uv run python examples/quick_start.py
uv run python examples/transform_replace_add.py

More examples worth browsing:

• examples/intrinsic_memcpy.py - call an LLVM intrinsic by name with Builder.intrinsic(...)
• examples/optimize_module.py - optimize a module with a PassBuilder pipeline string
• examples/optimize_function.py - optimize one function with a function-level PassBuilder pipeline string
• examples/emit_object_assembly.py - emit host object code and assembly from a module
• examples/jit_add.py - JIT-compile IR, call it through ctypes, and register a Python callback
• examples/instruction_metadata.py - attach custom metadata to an instruction and print the IR
• examples/named_metadata.py - create module named metadata and print the IR
• examples/metadata_debug_info.py - attach metadata, create debug info, and use debug-location scopes
• examples/transform_replace_add.py - simple IR transformation using operands, RAUW, and instruction deletion
• examples/bc-stats.py - print per-function instruction histograms from LLVM IR
• examples/bc-graphviz.py - emit a GraphViz-style control-flow graph from LLVM IR
• examples/bc-profile.py - instrument LLVM IR with profiling hooks

Current capabilities

• Pythonic wrappers for contexts, modules, types, values, functions, basic blocks, builders, metadata, debug info, object files, targets, target machines, and pass builder options
• IR construction, traversal, and transformation helpers: constants, globals, PHI/control flow/memory/cast/cmp instructions, operands, predecessors, RAUW, split blocks, move/clone/erase instructions
• IR and bitcode parsing/writing, lazy bitcode modules, module cloning/linking, diagnostics, attributes, COMDATs, calling conventions, linkage/visibility/storage controls
• Metadata/debug-info APIs including named metadata views, module flag views, instruction/global metadata mappings, DIBuilder recipes, and debug-location scopes
• Target lookup, data layouts, Module.emit_object(), Module.emit_assembly(), target-machine emission, and PassBuilder pipeline execution via Module.optimize(), Function.optimize(), and Module.run_passes()
• Generic intrinsic calls with Builder.intrinsic(...) and in-process JIT execution through the LLVM-C ORC LLJIT API
• Lifetime/validity guards that turn many use-after-free, disposed-object, null-reference, and wrong-kind mistakes into Python exceptions instead of hard crashes
• Auto-generated typed .pyi stubs for IDEs/type checkers
• Golden-master tests, Python regression scripts, examples, and vendored llvm-c-test lit tests, including CI coverage against installed wheels

Documentation

Type stubs are auto-generated and provide IDE intellisense. After building, find them at:

.venv/lib/python3.*/site-packages/llvm/__init__.pyi

For development documentation, see devdocs/README.md.

Known limitations

• The API is not stable yet; method names and ownership rules may still change.
• Scope follows the LLVM-C API, not the full LLVM C++ API. JIT support is intentionally limited to what is exposed cleanly through LLVM-C ORC LLJIT.
• Docs are currently the README, examples, devdocs/, and the generated .pyi stub; there is no hosted API reference yet.
• Prebuilt wheels currently target CPython 3.12+ stable ABI on Linux x86_64/aarch64, macOS arm64/x86_64, and Windows x86_64. Other platforms require a source build.

Development

Setup

Source/development builds need LLVM 21.1.6 available to CMake. The LLVM archives used for packaging are published at LLVMParty/llvm-builds v21.1.6.

From a checkout, download the matching LLVM archive and then run uv sync:

# Choose the archive matching your platform:
#   llvm-21.1.6-linux-x86_64.zip
#   llvm-21.1.6-linux-aarch64.zip
#   llvm-21.1.6-macos-arm64.zip
#   llvm-21.1.6-macos-x86_64.zip
#   llvm-21.1.6-windows-x86_64.zip
python tools/ci/install_llvm.py \
  --version 21.1.6 \
  --archive llvm-21.1.6-linux-x86_64.zip \
  --dest .llvm \
  --prefix-file .llvm-prefix

uv sync --verbose

If you are using your own LLVM install instead, set LLVM_ROOT to its install prefix:

export LLVM_ROOT=/path/to/llvm
uv sync --verbose

If .llvm-prefix already points at another LLVM install, delete it first; CMake may reuse it from a previous configure.

For offline builds:

uv sync --offline --no-build-isolation --verbose

For Windows C++/LSP setup, see the Windows development section below.

Testing

# Main golden-master suite:
# - runs C++ test executables from build/
# - runs paired Python scripts
# - compares Python output against stored C++ behavior
uv run run_tests.py

# Python-only regression scripts in tests/regressions/
uv run run_tests.py --regressions

# Vendored llvm-c-test lit suite against the C test binary
# Rebuilds the vendored C binary before running lit.
uv run run_llvm_c_tests.py
uv run run_llvm_c_tests.py -v

# Vendored llvm-c-test lit suite against the Python implementation
uv run run_llvm_c_tests.py --use-python

# Run the Python llvm-c-test port directly during development
uv run python -m llvm_c_test --targets-list

# Type checking (not a test suite, but commonly run in CI/dev)
uvx ty check

If you want the closest thing to “run everything in this repo”, use:

uv run run_tests.py
uv run run_tests.py --regressions
uv run run_llvm_c_tests.py
uv run run_llvm_c_tests.py --use-python

Python tests here are intended to be executable as standalone scripts (e.g. uv run tests/test_module.py or uv run tests/regressions/test_const_bytes.py). They are generally pytest-compatible too, but direct script execution is the historical/default style used by run_tests.py and for one-off debugging.

pytest is still useful for targeting specific regression files or subsets, but it is not our complete top-level test entrypoint by itself.

Coverage

# Run with coverage
uv run coverage run run_llvm_c_tests.py --use-python
uv run coverage combine
uv run coverage report --include="llvm_c_test/*"

Windows source builds and C++ IntelliSense

The wheel is the easiest path on Windows. If you need a source/development build, install Visual Studio or Visual Studio Build Tools with the C++ workload, then fetch LLVM and run uv sync:

py -3.12 tools\ci\install_llvm.py `
  --version 21.1.6 `
  --archive llvm-21.1.6-windows-x86_64.zip `
  --dest .llvm `
  --prefix-file .llvm-prefix

uv sync --verbose

That command downloads from LLVMParty/llvm-builds v21.1.6, installs into .llvm, and writes .llvm-prefix for CMake.

If you are using your own LLVM install instead:

$env:LLVM_ROOT = "C:\path\to\llvm"
uv sync --verbose

For C++ local development with an LSP, create a local CMakeUserPresets.json at the repository root. CMake gets LLVM from .llvm-prefix created above, or from LLVM_ROOT on first configure. The compiler paths below assume the usual LLVM installer location; adjust them if your clang-cl.exe is elsewhere:

{
    "version": 3,
    "configurePresets": [
        {
            "name": "clang-cl",
            "displayName": "Ninja with clang-cl",
            "generator": "Ninja",
            "binaryDir": "${sourceDir}/build",
            "cacheVariables": {
                "CMAKE_C_COMPILER": "C:/Program Files/LLVM/bin/clang-cl.exe",
                "CMAKE_CXX_COMPILER": "C:/Program Files/LLVM/bin/clang-cl.exe",
                "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
                "CMAKE_BUILD_TYPE": "RelWithDebInfo"
            }
        }
    ],
    "buildPresets": [
        {
            "name": "clang-cl",
            "configurePreset": "clang-cl"
        }
    ]
}

Then configure/build with:

cmake --preset clang-cl
cmake --build --preset clang-cl

If you move or replace the LLVM install, delete or regenerate .llvm-prefix from any previous configure.

License

This project is licensed under the MIT License. See LICENSE for details.

LLVM is licensed under the Apache License v2.0 with LLVM Exceptions.