Skip to content

Latest commit

 

History

History
198 lines (155 loc) · 9.87 KB

Architecture.md

File metadata and controls

198 lines (155 loc) · 9.87 KB
Raw

PyO3: Architecture

This document roughly describes the high-level architecture of PyO3. If you want to become familiar with the codebase you are in the right place!

Overview

PyO3 provides a bridge between Rust and Python, based on the Python/C API. Thus, PyO3 has low-level bindings of these API as its core. On top of that, we have higher-level bindings to operate Python objects safely. Also, to define Python classes and functions in Rust code, we have trait PyClass and a set of protocol traits (e.g., PyIterProtocol) for supporting object protocols (i.e., __dunder__ methods). Since implementing PyClass requires lots of boilerplate, we have a proc-macro #[pyclass].

To summarize, there are six main parts to the PyO3 codebase.

  1. Low-level bindings of Python/C API.
  2. Bindings to Python objects.
  3. PyClass and related functionalities.
  4. Procedural macros to simplify usage for users.
  5. build.rs and pyo3-build-config

1. Low-level bindings of Python/C API

pyo3-ffi contains wrappers of the Python/C API. This is currently done by hand rather than automated tooling because:

  • it gives us best control about how to adapt C conventions to Rust, and
  • there are many Python interpreter versions we support in a single set of files.

We aim to provide straight-forward Rust wrappers resembling the file structure of cpython/Include.

We are continuously updating the module to match the latest CPython version which PyO3 supports (i.e. as of time of writing Python 3.13). The tracking issue is #1289, and contribution is welcome.

In the pyo3-ffi crate, there is lots of conditional compilation such as #[cfg(Py_LIMITED_API)], #[cfg(Py_3_7)], and #[cfg(PyPy)]. Py_LIMITED_API corresponds to #define Py_LIMITED_API macro in Python/C API. With Py_LIMITED_API, we can build a Python-version-agnostic binary called an abi3 wheel. Py_3_7 means that the API is available from Python >= 3.7. There are also Py_3_8, Py_3_9, and so on. PyPy means that the API definition is for PyPy. Those flags are set in build.rs.

2. Bindings to Python objects

src/types contains bindings to built-in types of Python, such as dict and list. For historical reasons, Python's object is called PyAny in PyO3 and located in src/types/any.rs.

Currently, PyAny is a straightforward wrapper of ffi::PyObject, defined as:

#[repr(transparent)]
pub struct PyAny(UnsafeCell<ffi::PyObject>);

Concrete Python objects are implemented by wrapping PyAny, e.g.,:

#[repr(transparent)]
pub struct PyDict(PyAny);

These types are not intended to be accessed directly, and instead are used through the Py<T> and Bound<T> smart pointers.

We have some macros in src/types/mod.rs which make it easier to implement APIs for concrete Python types.

3. PyClass and related functionalities

src/pycell.rs, src/pyclass.rs, and src/type_object.rs contain types and traits to make #[pyclass] work. Also, src/pyclass_init.rs and [src/impl_/pyclass.rs] have related functionalities.

To realize object-oriented programming in C, all Python objects have ob_base: PyObject as their first field in their structure definition. Thanks to this guarantee, casting *mut A to *mut PyObject is valid if A is a Python object.

To ensure this guarantee, we have a wrapper struct PyCell<T> in src/pycell.rs which is roughly:

#[repr(C)]
pub struct PyCell<T: PyClass> {
    ob_base: crate::ffi::PyObject,
    inner: T,
}

Thus, when copying a Rust struct to a Python object, we first allocate PyCell on the Python heap and then move T into it. Also, PyCell provides RefCell-like methods to ensure Rust's borrow rules. See the documentation for more.

PyCell<T> requires that T implements PyClass. This trait is somewhat complex and derives many traits, but the most important one is PyTypeInfo in src/type_object.rs. PyTypeInfo is also implemented for built-in types. In Python, all objects have their types, and types are also objects of type. For example, you can see type({}) shows dict and type(type({})) shows type in Python REPL. T: PyTypeInfo implies that T has a corresponding type object.

Protocol methods

Python has some built-in special methods called dunder methods, such as __iter__. They are called "slots" in the abstract objects layer in Python/C API. We provide a way to implement those protocols similarly, by recognizing special names in #[pymethods], with a few new ones for slots that can not be implemented in Python, such as GC support.

4. Procedural macros to simplify usage for users.

pyo3-macros provides five proc-macro APIs: pymodule, pyfunction, pyclass, pymethods, and #[derive(FromPyObject)]. pyo3-macros-backend has the actual implementations of these APIs. src/impl_ contains #[doc(hidden)] functionality used in code generated by these proc-macros, such as parsing function arguments.

5. build.rs and pyo3-build-config

PyO3 supports a wide range of OSes, interpreters and use cases. The correct environment must be detected at build time in order to set up relevant conditional compilation correctly. This logic is captured in the pyo3-build-config crate, which is a build-dependency of pyo3 and pyo3-macros, and can also be used by downstream users in the same way.

In pyo3-build-config's build.rs the build environment is detected and inlined into the crate as a "config file". This works in all cases except for cross-compiling, where it is necessary to capture this from the pyo3 build.rs to get some extra environment variables that Cargo doesn't set for build dependencies.

The pyo3 build.rs also runs some safety checks such as ensuring the Python version detected is actually supported.

Some of the functionality of pyo3-build-config:

  • Find the interpreter for build and detect the Python version.
    • We have to set some version flags like #[cfg(Py_3_7)].
    • If the interpreter is PyPy, we set #[cfg(PyPy).
    • If the PYO3_CONFIG_FILE environment variable is set then that file's contents will be used instead of any detected configuration.
    • If the PYO3_NO_PYTHON environment variable is set then the interpreter detection is bypassed entirely and only abi3 extensions can be built.
  • Check if we are building a Python extension.
    • If we are building an extension (e.g., Python library installable by pip), we don't link libpython. Currently we use the extension-module feature for this purpose. This may change in the future. See #1123.
  • Cross-compiling configuration
    • If TARGET architecture and HOST architecture differ, we can find cross compile information from environment variables (PYO3_CROSS_LIB_DIR, PYO3_CROSS_PYTHON_VERSION and PYO3_CROSS_PYTHON_IMPLEMENTATION) or system files. When cross compiling extension modules it is often possible to make it work without any additional user input.
    • When an experimental feature generate-import-lib is enabled, the pyo3-ffi build script can generate python3.dll import libraries for Windows targets automatically via an external python3-dll-a crate. This enables the users to cross compile Python extensions for Windows without having to install any Windows Python libraries.
  翻译：