ctypesgen (pypdfium2-team fork)

ctypesgen is a ctypes wrapper generator for Python.

This is a fork with the objective to better suit the needs of pypdfium2, and address some of the technical debt and (in our opinion) design issues that have accumulated due to highly conservative maintenance.

Here are some notes on our development intents:

• We do not mind API-breaking changes at this time.
• We endeavor to use plain ctypes as much as possible and keep the template lean.
• For now, we only envisage to work with ctypesgen's higher-level parts. The parser backend may be out of our scope.

System Dependencies

ctypesgen depends on the presence of an external C pre-processor, by default gcc or clang, as available. Alternatively, you may specify a custom pre-processor command using the --cpp option (e.g. --cpp "clang -E" to always use clang).

Tips & Tricks

• If you have multiple libraries that are supposed to interoperate with shared symbols, first create bindings to any shared headers and then use the -m / --link-modules option on dependants. (Otherwise, you'd create duplicate symbols that are formally different types, with need to cast between them.) If the module is not installed separately, you may prefix the module name with . for a relative import, and share boilerplate code using --no-embed-templates. Relative modules will be expected to be present in the output directory at compile time. Note, this strategy can also be used to bind to same-library headers separately; however, you'll need to resolve the dependency tree on your own.
• Extra include search paths can be provided using the -I option or by setting $CPATH/$C_INCLUDE_PATH. You could use this to add a header spoofing an external symbol via typedef void* SYMBOL; (c_void_p) that may be provided by a third-party binding at runtime.
• If building with --no-macro-guards and you encounter broken macros, you may use --symbol-rules (see below) or replace them manually. This can be necessary on C constructs like #define NAN (0.0f / 0.0f) that don't play well with python. In particular, you are likely to run into this with --all-headers.

Notes on symbol inclusion

• ctypesgen works with the following symbol rules:
  • yes: The symbol is eagerly included.
  • if_needed: The symbol is included if other included symbols depend on it (e.g. a type used in a function signature).
  • never: The symbol is always excluded, and implicitly all its dependants.
• Roughly speaking, symbols from caller-given headers get assigned the include rule yes, and any others if_needed. When building with --all-headers, all symbols default to yes regardless of their origin.
• --no-macros sets the include rule of all macro objects to never.
• Finally, the --symbol-rules option is applied, which can be used to assign symbol rules by regex fullmatch expressions, providing callers with powerful means of control over symbol inclusion.
• To filter out excess symbols, you'll usually want to use if_needed rather than never to avoid accidental exclusion of dependants. Use never only where this side effect is actually wanted, e.g. to exclude a broken symbol.

Binding against the Python API

cat >"overrides.py" <<END
import ctypes

class PyTypeObject (ctypes.Structure): pass
class PyObject (ctypes.Structure): pass

def POINTER(obj):
    if obj is PyObject: return ctypes.py_object
    return ctypes.POINTER(obj)
END

ctypesgen -l python --dllclass pythonapi --system-headers python3.X/Python.h --all-headers -m .overrides --linkage-anchor . -o ctypes_python.py

substituting 3.X with your system's python version.

Small test:

import sys
from ctypes import *
from ctypes_python import *

# Get a string from a Python C API function
v = Py_GetVersion()
v = cast(v, c_char_p).value.decode("utf-8")
print(v)
print(v == sys.version)  # True

# Convert back and forth between Native vs. C view of an object
class Test:
    def __init__(self, a):
        self.a = a

t = Test(a=123)
tc_ptr = cast(id(t), POINTER(PyObject_))
tc = tc_ptr.contents
print(tc.ob_refcnt)  # 1
Py_IncRef(t)
print(tc.ob_refcnt)  # 2 (incremented)
Py_DecRef(t)
print(tc.ob_refcnt)  # 1 (decremented)
t_back = cast(tc_ptr, py_object).value
print(t_back.a)
print(tc.ob_refcnt)  # 2 (new reference from t_back)

It should yield something like

3.11.6 (main, Oct  3 2023, 00:00:00) [GCC 12.3.1 20230508 (Red Hat 12.3.1-1)]
True
1
2
1
123
2

Known Limitations

ctypes

• Rare calling conventions other than cdecl or stdcall are not supported.
• Non-primitive return types in callbacks are not supported. An affected prototype wouldn't allow for the creation of a function instance, but not break the output as a whole.

pypdfium2-ctypesgen

• The DLL class is assumed to be CDLL, otherwise it needs to be given by the caller. We do not support mixed calling conventions, because it does not match the API layer of ctypes.
• We do not support binding to multiple binaries in the same output file. Instead, you'll want to create separate output files sharing the loader template, and possibly use module linking, as described above.

ctypesgen

• ctypesgen's parser was originally written for C99. Support for later standards (C11 etc.) is probably incomplete.
• The conflicting names resolver is largely untested, in particular the handling of dependants. Please report success or failure.
• Linked modules are naively prioritized in dependency resolver and conflicting names handler, i.e. intentional overrides are ignored. The position of includes is not honored; ctypesgen always imports linked modules at top level.

Fork rationale

Trying to get through changes upstream is tedious, with unclear outcome, and often not applicable due to mismatched intents (e.g. regarding backwards compatibility). Also consider that isolating commits in separate branches is not feasible anymore as merge conflicts arise (e.g. due to code cleanups and interfering changes).

Contrast this to a fork, which allows us to keep focused and effect improvements quickly, so as to invest developer time rationally.

However, we would be glad if our work could eventually be merged back upstream once the change set has matured, if upstream can arrange themselves with the radical changes. See ctypesgen#195 for discussion.

Syncing with upstream

• First, sync the fork's master branch using GitHub's web interface.
• View changes on GitHub's compare page.
• Pull and merge locally, then push the result.

Last time we had to do this, git merge origin/master -Xours did a good job. Changes to files we haven't really modified can usually just be pulled in as-is. Otherwise, you'll have to manually look through the changes and pick what you consider worthwhile on a case by case basis.

Note, it is important to verify the resulting merge commit for correctness - automatic merge strategies might produce mistakes!

Bugs

Oversights or unintentional breakage can happen at times. Feel free to file a bug report if you think a change introduces logical issues. However, please note our response policy below.

Contributions

We may accept contributions, but only if our code quality expectations are met.

Policy:

• We may not respond to your issue or PR.
• We may close an issue or PR without much feedback.
• We may lock discussions or contributions if our attention is getting DDOSed.
• We may not provide much usage support.