xtc-build is a small Python library for describing C object files, static
archives, and shared libraries as an explicit dependency graph. The same graph
can be inspected, built immediately from Python, or written as a GNU Makefile
or Ninja build file.
It deliberately has no mandatory project object. Artifacts describe what must
be built; a lightweight BuildContext supplies the toolchain and output
directory.
Status: early alpha. The initial backends target GCC/Clang-compatible C toolchains, GNU Make, and Ninja on Unix-like systems.
Install the latest release from PyPI:
python -m pip install xtc-buildDevelopment versions built from main are available from
TestPyPI for testing:
python -m pip install --pre --upgrade \
--index-url https://test.pypi.org/simple/ xtc-buildDevelopment requires cc, ar, and make; uv sync installs Ninja. On
macOS, the system-provided Apple Clang toolchain is sufficient; Homebrew is not
required. Missing tools
are reported as test errors rather than skipped tests. Install uv from the
official installation page,
or directly with:
curl -LsSf https://astral.sh/uv/install.sh | shThen clone the repository and set up the development environment:
git clone git@github.com:xtc-tools/xtc-build.git
cd xtc-build
uv sync
source .venv/bin/activate
make checkuv sync creates the virtual environment, installs the project in editable
mode, installs the default dev dependency group, and synchronizes everything
with the committed uv.lock. Alternatively, run checks without activating the
environment:
uv run make checkThe full check requires 100% line and branch coverage and writes an HTML report
to htmlcov/. Use make check-pytest to run pytest without collecting
coverage.
import ctypes
from xtc_build import (
Archive,
BuildContext,
ExternalLibrary,
Object,
SharedLibrary,
)
ctx = BuildContext(
build_dir="build",
compile_flags="-O2 -Wall",
defines=["PROJECT_BUILD=1"],
)
helper = Object(
"helper",
source="src/helper.c",
inputs=["include/helper.h"],
includes=["include"],
compile_flags=["-Wconversion"],
pic=True,
)
api = Object("api", source="src/api.c", includes=["include"], pic=True)
core = Archive("core", objects=[helper])
math = ExternalLibrary("math", link_flags=["-lm"])
library = SharedLibrary(
"example",
objects=[api],
archives=[core],
libraries=[math],
link_flags=["-Wl,--no-undefined"],
)Flag fields such as compile_flags, archive_flags, and link_flags accept
either a sequence of arguments or a shell-like string. Strings are parsed with
shlex.split(), including quote handling, but commands are still executed
without a shell.
BuildContext provides default compile_flags, link_flags, and defines.
Artifact-specific values are appended after these defaults. Set
override_flags=True on an Object or SharedLibrary to ignore the matching
context flags, or override_defines=True on an Object to ignore context
definitions.
The default GnuToolchain uses the system cc and ar commands. On macOS,
cc selects Apple Clang and shared libraries use the native .dylib format.
A specific compatible compiler can still be selected explicitly:
from xtc_build import GnuToolchain
ctx = BuildContext(toolchain=GnuToolchain(cc="gcc"))Every artifact is independently buildable. Building a composite artifact also builds its stale transitive dependencies:
helper.build(ctx) # build/example.o-style object output
core.build(ctx) # builds its object dependencies first
path = library.build(ctx)
loaded = ctypes.CDLL(str(path))Artifact.build() returns its output path. BuildContext.build() instead
returns the tuple of artifacts rebuilt by that invocation:
rebuilt = ctx.build(library)The immediate executor compares input and output timestamps and stores each command's deterministic JSON state, including its ordered input and output paths, beside the output. The stored JSON is compared directly and can be diffed when diagnosing a rebuild. Changing compiler flags or declared inputs therefore causes a rebuild even when source timestamps did not change.
Object files and static archives produced outside the graph can be declared as input-only artifacts:
from xtc_build import (
Archive,
ExternalArchive,
ExternalObject,
ExternalSharedLibrary,
SharedLibrary,
)
prebuilt_object = ExternalObject("generated/helper.o", pic=True)
archive = Archive("core", objects=[prebuilt_object])
prebuilt_archive = ExternalArchive("vendor/libsupport.a", pic=True)
prebuilt_shared = ExternalSharedLibrary("vendor/libdependency.so")
library = SharedLibrary(
"example",
objects=[api],
archives=[prebuilt_archive],
libraries=[prebuilt_shared],
)ExternalObject, ExternalArchive, and ExternalSharedLibrary have no build
command and are not graph nodes. Their exact paths, including extensions, are
explicit command inputs and Make or Ninja prerequisites. They must exist before
immediate execution or be produced independently before a generated backend
needs them. Set
pic=True when an object or archive is suitable for linking into a shared
library. On Linux and macOS, each external shared library's resolved parent
directory is added to the linked library's runtime search path. Windows does
not support embedded rpaths, so its DLL search path must be configured at
runtime.
graph = ctx.graph(library)
for node in graph.topological_order():
command = graph.command(node)
print(node.name)
print("dependencies:", graph.dependencies(node))
print("argv:", command.argv)
print("inputs:", command.inputs)
print("outputs:", command.outputs)Commands are structured argument vectors, not shell strings. This allows the
immediate executor to invoke them without shell=True and lets output backends
perform their own quoting.
ctx.write_makefile("build/Makefile", targets=[core, library])Run it from the directory against which source paths were declared:
make -f build/Makefile
make -f build/Makefile cleanObject rules emit GCC-compatible dependency files (-MMD -MP). Declare
non-discoverable inputs such as generated headers and linker scripts explicitly
with Object(inputs=[...]).
ctx.write_ninja("build/build.ninja", targets=[core, library])Run Ninja from the directory against which source paths were declared:
ninja -f build/build.ninja
ninja -f build/build.ninja -t cleanNinja consumes compiler depfiles with deps = gcc, so GCC and Clang discovered
headers participate in incremental rebuilding. Explicit and external inputs are
emitted as build-edge dependencies. The build file is rewritten only when its
content changes.
- Dependencies between built artifacts are explicit and cycle checked.
- Output collisions are rejected while constructing a graph.
- Link order follows declaration order.
- Objects linked into a shared library must declare
pic=Trueby default. Setrequire_pic=FalseonSharedLibraryonly when the target platform permits it. ExternalLibraryis not a build node; it is an explicit collection of link flags for a library built outside the graph.- A project/grouping abstraction may be added later as an optional convenience, but is not required by the core model.
Package versions are derived from Git tags by setuptools-scm. Release tags
must use the xtc-build-vX.Y.Z form. Create an annotated tag for each release:
git tag -a xtc-build-v1.2.3 -m "xtc-build 1.2.3"A build at that tag has version 1.2.3; later commits receive a development
version derived from the tag and Git revision. Builds without Git metadata use
0.0.0 as a fallback.
After all checks pass, pushes to main publish development distributions to
TestPyPI, while
xtc-build-vX.Y.Z tags publish releases to
PyPI. Both publication jobs in
.github/workflows/ci.yml use trusted publishing through the testpypi and
pypi GitHub environments; those trusted publishers must be configured on the
corresponding package indexes before the jobs can authenticate.
Published versions can be inspected through the package-index JSON APIs:
curl -fsSL https://pypi.org/pypi/xtc-build/json \
| jq -r '.releases | keys[]'
curl -fsSL https://test.pypi.org/pypi/xtc-build/json \
| jq -r '.releases | keys[]'BSD 3-Clause. See LICENSE.