PyPy's ctypes implementation
- application level code - code written in full Python
- interpreter level code - code written in RPython, compiled to something else, say C, part of the interpreter.
PyPy's ctypes implementation in its current state proves the feasibility of implementing a module with the same interface and behavior for PyPy as ctypes for CPython.
PyPy's implementation internally uses libffi like CPython's ctypes. In our implementation as much as possible of the code is written in full Python, not RPython. In CPython's situation, the equivalent would be to write as little as possible code in C. We essentially favored rapid experimentation over worrying about speed for this first trial implementation. This allowed to provide a working implementation with a large part of ctypes features in 2 months real time.
We reused the ctypes package version 1.0.2 as-is from CPython. We implemented _ctypes which is a C module in CPython mostly in pure Python based on a lower-level layer extension module _rawffi.
Low-level part: _rawffi
This PyPy extension module (pypy/module/_rawffi) exposes a simple interface to create C objects (arrays and structures) and calling functions in dynamic libraries through libffi. Freeing objects in most cases and making sure that objects referring to each other are kept alive is responsibility of the higher levels.
This module uses bindings to libffi which are defined in rpython/rlib/libffi.py.
We tried to keep this module as small as possible. It is conceivable that other implementations (e.g. Jython) could use our ctypes implementation by writing their version of _rawffi.
The reused ctypes package lives in lib_pypy/ctypes. _ctypes implementing the same interface as _ctypes in CPython is in lib_pypy/_ctypes.
Discussion and limitations
Reimplementing ctypes features was in general possible. PyPy supports pluggable garbage collectors, some of them are moving collectors, this means that the strategy of passing direct references inside Python objects to an external library is not feasible (unless the GCs support pinning, which is not the case right now). The consequence of this is that sometimes copying instead of sharing is required, this may result in some semantics differences. C objects created with _rawffi itself are allocated outside of the GC heap, such that they can be passed to external functions without worries.
Porting the implementation to interpreter-level should likely improve its speed. Furthermore the current layering and the current _rawffi interface require more object allocations and copying than strictly necessary; this too could be improved.
Here is a list of the limitations and missing features of the current implementation:
- ctypes.pythonapi lets you access the CPython C API emulation layer of PyPy, at your own risks and without doing anything sensible about the GIL. Since PyPy 2.3, these functions are also named with an extra "Py", for example PyPyInt_FromLong(). Basically, don't use this, but it might more or less work in simple cases if you do. (Obviously, assuming the PyObject pointers you get have any particular fields in any particular order is just going to crash.)
- We copy Python strings instead of having pointers to raw buffers
- Features we did not get to implement:
- custom alignment and bit-fields
- resizing (resize() function)
- non-native byte-order objects
- callbacks accepting by-value structures
- slight semantic differences that ctypes makes between its primitive types and user subclasses of its primitive types
Running application examples
pyglet is known to run. We also had some success with pygame-ctypes (which is no longer maintained) and with a snapshot of the experimental pysqlite-ctypes. We will only describe how to run the pyglet examples.
We tried pyglet checking it out from its repository at revision 1984.
From pyglet, the following examples are known to work:
The pypy-c translated to run the ctypes tests can be used to run the pyglet examples as well. They can be run like e.g.:
$ cd pyglet/ $ PYTHONPATH=. ../ctypes-stable/pypy/goal/pypy-c examples/opengl.py
they usually should be terminated with ctrl-c. Refer to the their doc strings for details about how they should behave.
The following examples don't work for reasons independent from ctypes:
- image_convert.py needs PIL
- image_display.py needs PIL
- astraea/astraea.py needs PIL
We did not try the following examples:
- media_player.py needs avbin or at least a proper sound card setup for .wav files
- video.py needs avbin
- soundscape needs avbin
We also released ctypes-configure, which is an experimental package trying to approach the portability issues of ctypes-based code.
One of ctypes problems is that ctypes programs are usually not very platform-independent. We created ctypes_configure, which invokes c compiler (via distutils) for various platform-dependent details like exact sizes of types (for example size_t), #defines, exact outline of structures etc. It replaces in this regard code generator (h2py).
`ctypes_configure/doc/sample.py`_ explains in details how to use it.