parse PyCFunctions like pybind11 functions

[AaronOpfer]

PyCFunctions are surprisingly compatible with inspect.getfullargspec et. al., through a mechanism called __text_signature__. This means that pybind11-stubgen will not activate its docstring-parsing fallback. This is problematic for pybind11-stubgen however, because the __text_signature__ does not have any mechanism for including type hints or annotations.

This is relevant to pybind11-stubgen as I have been porting some C API functions away from pybind11 to CPython and hit a snag trying to do it piecemeal, as pybind11-stubgen won't generate stubs for functions created in the CPython way.

A test case could be as simple as:

namespace {

static PyObject* python_time_ns(PyObject* /*self*/, PyObject* /*args*/) {
    return PyLong_FromUnsignedLongLong(0);  // obviously just for the purpose of testing
}

static PyMethodDef memberdef_time_ns{.ml_name = "time_ns",
                                     .ml_meth = python_time_ns,
                                     .ml_flags = METH_NOARGS,
                                     .ml_doc = "time_ns() -> int\n"};

} // namespace

PyObject* create_time_ns_func(PyObject* modname) {
    return PyCFunction_NewEx(&memberdef_time_ns, NULL, modname);
}

And then installing this onto a pybind11 module:

    m.attr("time_ns") = create_time_ns_func(m.attr("__name__").ptr());

[AaronOpfer]

It was not clear to me how to run/augment the tests so I'm not sure if this PR breaks anything yet. I am also open to other suggestions or consideration of whether this change is appropriate since it's pretty clearly not pybind11-related.

[gentlegiantJGC]

Is there is reason why you are not using isinstance

[AaronOpfer]

Hmm, good question. it looks like using isinstance() would actually cover a few additional types, such as types.BuiltinMethodType, which is probably a good thing.

[gentlegiantJGC]

Adding your tests to the test suite would be helpful