Writing extension modules for pypy
This document tries to explain how to interface the PyPy python interpreter with any external library.
Note: We try to describe state-of-the art, but it might fade out of date as this is the front on which things are changing in pypy rapidly.
Right now, there are three possibilities of providing third-party modules for the PyPy python interpreter (in order of usefulness):
The ctypes module in PyPy is ready to use. It's goal is to be as-compatible-as-possible with the CPython ctypes version. Right now it's able to support large examples, such as pyglet. PyPy is planning to have a 100% compatible ctypes implementation, without the CPython C-level API bindings (so it is very unlikely that direct object-manipulation trickery through this API will work).
We also provide a ctypes-configure for overcoming the platform dependencies, not relying on the ctypes codegen. This tool works by querying gcc about platform-dependent details (compiling small snippets of C code and running them), so it'll benefit not pypy-related ctypes-based modules as well.
ctypes call are optimized by the JIT and the resulting machine code contains a direct call to the target C function. However, due to the very dynamic nature of ctypes, some overhead over a bare C call is still present, in particular to check/convert the types of the parameters. Moreover, even if most calls are optimized, some cannot and thus need to follow the slow path, not optimized by the JIT.
Stable, CPython-compatible API. Most calls are fast, optimized by JIT.
Problems with platform-dependency (although we partially solve those). Although the JIT optimizes ctypes calls, some overhead is still present. The slow-path is very slow.
Mostly in order to be able to write a ctypes module, we developed a very low-level libffi bindings called _ffi. (libffi is a C-level library for dynamic calling, which is used by CPython ctypes). This library provides stable and usable API, although it's API is a very low-level one. It does not contain any magic. It is also optimized by the JIT, but has much less overhead than ctypes.
It Works. Probably more suitable for a delicate code where ctypes magic goes in a way. All calls are optimized by the JIT, there is no slow path as in ctypes.
It combines disadvantages of using ctypes with disadvantages of using mixed modules. CPython-incompatible API, very rough and low-level.
This is the most advanced and powerful way of writing extension modules. It has some serious disadvantages:
- a mixed module needs to be written in RPython, which is far more complicated than Python (XXX link)
- due to lack of separate compilation (as of July 2011), each compilation-check requires to recompile whole PyPy python interpreter, which takes 0.5-1h. We plan to solve this at some point in near future.
- although rpython is a garbage-collected language, the border between C and RPython needs to be managed by hand (each object that goes into the C level must be explicitly freed).
Some documentation is available here
- XXX we should provide detailed docs about lltype and rffi, especially if we
- want people to follow that way.
This method is still experimental and is being exercised on a branch, reflex-support, which adds the cppyy module. The method works by using the Reflex package to provide reflection information of the C++ code, which is then used to automatically generate bindings at runtime. From a python standpoint, there is no difference between generating bindings at runtime, or having them "statically" generated and available in scripts or compiled into extension modules: python classes and functions are always runtime structures, created when a script or module loads. However, if the backend itself is capable of dynamic behavior, it is a much better functional match to python, allowing tighter integration and more natural language mappings. Full details are available here.
The cppyy module is written in RPython, which makes it possible to keep the code execution visible to the JIT all the way to the actual point of call into C++, thus allowing for a very fast interface. Reflex is currently in use in large software environments in High Energy Physics (HEP), across many different projects and packages, and its use can be virtually completely automated in a production environment. One of its uses in HEP is in providing language bindings for CPython. Thus, it is possible to use Reflex to have bound code work on both CPython and on PyPy. In the medium-term, Reflex will be replaced by cling, which is based on llvm. This will affect the backend only; the python-side interface is expected to remain the same, except that cling adds a lot of dynamic behavior to C++, enabling further language integration.
C++ is a large language, and cppyy is not yet feature-complete. Still, the experience gained in developing the equivalent bindings for CPython means that adding missing features is a simple matter of engineering, not a question of research. The module is written so that currently missing features should do no harm if you don't use them, if you do need a particular feature, it may be necessary to work around it in python or with a C++ helper function. Although Reflex works on various platforms, the bindings with PyPy have only been tested on Linux.