Skip to main content

Clang Build Extension

Gitter Build Status Coverage Status

clang-build-ext Version clang-build-ext Python Versions clang-build-ext Downloads Per Day clang-build-ext Downloads Per Week clang-build-ext Downloads Per Month

Overview

clang-build-ext is a setuptools plugin that builds Python C/C++ extensions using the LLVM/Clang compiler toolchain instead of the system default compiler. Either a system-installed LLVM/Clang or the karellen-llvm-clang package can be used.

The plugin provides drop-in replacements for setuptools' build_ext and build_clib commands, along with features such as glob pattern expansion in source lists, Drakon IR bytecode embedding, and thin static library support.

Basic Usage

Add clang-build-ext to your build dependencies in pyproject.toml:

[build-system]
requires = ["setuptools", "clang-build-ext"]
build-backend = "setuptools.build_meta"

Register the custom build commands in setup.py:

from setuptools import setup, Extension
from karellen.clang_build_ext import ClangBuildExt, ClangBuildClib

setup(
    ...,
    ext_modules=[Extension("myext", ["src/*.c"])],
    cmdclass={
        "build_ext": ClangBuildExt,
        "build_clib": ClangBuildClib,
    },
)

Then build with:

python -m build

LLVM Toolchain

The compiler uses the full LLVM toolchain:

Tool Command
C compiler clang
C++ compiler clang++
Linker clang -fuse-ld=lld / clang++ -fuse-ld=lld
Archiver llvm-ar
Objcopy llvm-objcopy
Readelf llvm-readelf

lld is required, not merely preferred. Python's own build configuration supplies the linker command (LDSHARED, typically gcc -shared), and the plugin substitutes the toolchain's driver into it while re-applying -fuse-ld=lld, so every link goes through lld regardless of what the interpreter was built with — the same assumption already made for llvm-ar.

Features

Glob Pattern Expansion

Source file lists in both extensions and libraries support shell glob patterns. Patterns are expanded at build time, so you don't need to enumerate every source file in setup.py.

* matches within a single directory, while ** recurses to any depth. Given this tree:

src/module/
├── module.c
├── util/
│   └── util.c
└── codec/
    └── formats/
        └── png.c
Pattern Matches
src/module/*.c module.c
src/module/*/*.c util/util.c
src/module/**/*.c module.c, util/util.c, codec/formats/png.c

Use * when a flat directory is exactly what you want — it will not silently pick up sources added in a subdirectory later:

setup(
    ...,
    ext_modules=[
        Extension("myext", ["src/module/*.c"]),
    ],
    libraries=[
        ("mylib", {"sources": ["src/lib/*.c"]}),
    ],
    cmdclass={
        "build_ext": ClangBuildExt,
        "build_clib": ClangBuildClib,
    },
)

Use ** to pick up an entire source tree regardless of how deeply it is nested:

setup(
    ...,
    ext_modules=[
        Extension("myext", ["src/module/**/*.c"]),
    ],
    libraries=[
        ("mylib", {"sources": ["src/lib/**/*.c"]}),
    ],
    cmdclass={
        "build_ext": ClangBuildExt,
        "build_clib": ClangBuildClib,
    },
)

**/*.c already covers the top level, so pairing it with *.c is unnecessary. Should you list overlapping patterns anyway, each source is compiled once: duplicates are removed, keeping the order in which they were first matched.

Only the sources entry is glob-expanded. Every other build_info key of a build_clib library (macros, include_dirs, cflags, obj_deps) is passed through to setuptools unchanged.

C++ Extensions

Sources that setuptools detects as C++ (.cc, .cpp, .cxx) are compiled with the clang++ driver, and any extension containing one is linked with it too, so no extra configuration is required to mix C and C++ in one project:

setup(
    ...,
    ext_modules=[
        Extension("myext", ["src/module/*.cpp"],
                  include_dirs=["include"],
                  extra_compile_args=["-std=c++20"]),
    ],
    libraries=[
        ("mylib", {"sources": ["src/lib/*.cpp"],
                   "include_dirs": ["include"],
                   "macros": [("MYLIB_BUILD", "1")],
                   "cflags": ["-std=c++20"]}),
    ],
    cmdclass={
        "build_ext": ClangBuildExt,
        "build_clib": ClangBuildClib,
    },
)

Toolchain Location and RUNPATH

The two supported toolchains resolve their runtime libraries differently, and the build adapts to whichever one is in use.

A system LLVM/Clang keeps its libraries where the dynamic loader already looks, so nothing extra is required and nothing extra is emitted. Extensions link directly against the host's C++ runtime, which on most distributions is libstdc++ — distribution builds of Clang are generally configured to prefer it.

An in-environment LLVM/Clang, installed through karellen-llvm-clang, defaults to libc++ and ships its libraries inside the Python environment — libc++.so.1 and libc++abi.so.1 and libunwind.so.1 land in <prefix>/lib/<target-triple>/, which is on no default search path. Extensions built against it would record a bare DT_NEEDED and fail at import with:

ImportError: libc++.so.1: cannot open shared object file: No such file or directory

To prevent that, clang-build-ext detects karellen-llvm-core in the current environment and records an origin-relative RUNPATH on every extension it links:

${ORIGIN}/../../../lib:${ORIGIN}/../../../lib/x86_64-unknown-linux-gnu

This mirrors the RUNPATH karellen-llvm records in its own binaries. Because the offset is measured from the extension's install directory to the environment prefix, it holds in any environment with the same layout, so a wheel built this way imports without LD_LIBRARY_PATH wherever karellen-llvm-clang is installed alongside it.

Detection is on the distribution, not on PATH: if karellen-llvm-core is absent from the environment, the build is treated as a system-toolchain build and the link line is left alone. Any runtime_library_dirs you set on an Extension are preserved and come first.

Pinning the LLVM Runtime

A RUNPATH is only a path. Whichever karellen-llvm-core is installed where it points supplies the C++ runtime at import time, and that has to be the one the extensions were compiled and linked against — an upgrade or downgrade of LLVM in the environment otherwise repoints a built extension at a different runtime with nothing in the metadata to object.

So alongside the RUNPATH, and under exactly the same condition, the package being built gains a dependency on the LLVM it was built against:

Requires-Dist: karellen-llvm-core>=23.1.1.post6,<24

The floor is the installed version as it stands. A .postN in it is not packaging bookkeeping — it counts commits after the release tag, so two posts of one release are two different compilers. The ceiling is the next major version. Nothing is added to a package built against a system toolchain.

The dependency is contributed through a setuptools.finalize_distribution_options entry point, so it appears both in the wheel and in the metadata a front end prepares ahead of building, which is what pip resolves against. Dependencies you declare yourself, in setup.py or in [project], are untouched.

Drakon Enhancements

Drakon mode embeds LLVM intermediate representation (IR) bytecode into compiled binaries as custom ELF sections. This enables post-compilation IR analysis and transformation of the final binary.

When enabled, the build:

  1. Compiles with --save-temps=obj -fno-discard-value-names to produce .bc (bitcode) files alongside object files
  2. Includes .bc files in static libraries created via build_clib
  3. After linking, extracts .bc files from all linked objects and static libraries and embeds them into the output binary as .drakon.<name> ELF sections (marked noload,readonly)

Enable via environment variable or setup.cfg:

DRAKON=1 python -m build
# setup.cfg
[build_ext]
drakon = 1

Thin Static Libraries

Thin static libraries store references to object files rather than copies, reducing build artifact size during development.

Enable via environment variable or setup.cfg:

THIN=1 python -m build
# setup.cfg
[build_ext]
thin = 1

Both options can be combined:

DRAKON=1 THIN=1 python -m build
# setup.cfg
[build_ext]
drakon = 1
thin = 1

The build_clib command inherits drakon and thin settings from build_ext automatically.

Setuptools Compatibility

clang-build-ext supports setuptools 68 and newer, and is tested against the versions on either side of each API change that affects it:

setuptools Change
70.1 build_meta.get_requires_for_build_wheel stops requiring wheel; earlier versions need it installed for a --no-isolation build
72.2 UnixCCompiler gains the separate C++ executables (compiler_cxx, compiler_so_cxx, linker_so_cxx, linker_exe_cxx)
75.9 new_compiler moves from distutils.ccompiler to distutils.compilers.C.base
81.0 new_compiler drops the dry_run parameter

Release files for clang-build-ext 0.0.8

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for clang-build-ext 0.0.8
File Size Uploaded
clang_build_ext-0.0.8.tar.gz 20.4 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for clang-build-ext 0.0.8
File Interpreter ABI Platform
clang_build_ext-0.0.8-py3-none-any.whl Python 3 none any Details

Total release size: 33.8 kB

Release files / clang_build_ext-0.0.8.tar.gz

Download URL clang_build_ext-0.0.8.tar.gz
Size 20.4 kB
Tags Source
SHA-256 checksum
How to use checksums
a9ed5d7314b23cf6ebcf0d6e573e463378fc6c49ff5f26e1379c7de4a64dc851
BLAKE2b-256 checksum
How to use checksums
57852722af15674334d23d3c87af6e26b265ae2272d452d304423868b4b83431
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.7

Release files / clang_build_ext-0.0.8-py3-none-any.whl

Download URL clang_build_ext-0.0.8-py3-none-any.whl
Size 13.4 kB
Tags Python 3
SHA-256 checksum
How to use checksums
7f96a11b095e12dd7bd0d040cd3533a310eeec5d251e65ca19344f17b8150bc3
BLAKE2b-256 checksum
How to use checksums
0d0294e5c5724fddbf085cb5655cd327fe17f3903eebd0d38e4e2a2e55d1b782
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.7
Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page