Add PyObject_Vectorcall()

[vstinner]

No description provided.

[vstinner]

cc @jdemeyer @markshannon @encukou @erlend-aasland @corona10

Do you think that it makes sense to provide PyObject_Vectorcall() on Python 3.8 and older? This PR works on Python 2.7-3.13, PyPy 2 and PyPy 3 (see PR tests).

I chose to only provide PyObject_Vectorcall(). I didn't implement PyObject_VectorcallDict() nor PyObject_VectorcallMethod() (yet). I implemented PyVectorcall_NARGS() and PY_VECTORCALL_ARGUMENTS_OFFSET as well, but I'm not sure that it's worth it.

In Python 3.13, I removed the private alias _PyObject_Vectorcall (to PyObject_Vectorcall()): https://github.com/python/cpython/issues/106084 In Python 3.13, I removed 181 private functions in total. According to a code search on PyPI top 5,000 projects, _PyObject_Vectorcall`` is the mostly called functons of these removed function: python/cpython#106320 (comment)

I chose to write a simple but inefficient implementation just for backward compatibility. It should not be used for best performance: it creates a temporary tuple for positional arguments and a temporary dictionary for keyword arguments.

Writing an efficient implementation would require way more code, but Python 3.7 recently (at 2023-06-27) reached end of life. On CPython 3.8, the private _PyObject_Vectorcall() function is simply used. Only PyPy 3.8 should get bad performance: in exchange, you can use a single code base working on all Python versions and implementations.

[encukou]

Do you think that it makes sense to provide PyObject_Vectorcall() on Python 3.8 and older?

Yes, I thought the purpose of pythoncapi-compat is to provide backports like this.

I chose to write a simple but inefficient implementation just for backward compatibility.

Yup, being inefficient but correct is IMO a good trade-off for backports.

[vstinner]

Yes, I thought the purpose of pythoncapi-compat is to provide backports like this.

For me, this function is mostly written for performance. Using it on old Python versions make the code less efficient than using PyObject_Call(). But yeah, I concur that it's a nice trade-off.

[vstinner]

I merged my PR. Thanks for the review @encukou.