Upgrade CPython to the new PyMemberDef API: names starting with Py_ prefix (Py_T_INT instead of T_INT)

[vstinner]

The CPython code base was written for the old "structmember.h" API with names like READONLY and T_INT. I propose to upgrade the code base to new public names like Py_READONLY and Py_T_INT.

#include "structmember.h" should be removed, but #include <stddef.h> should maybe be added in exchange to get the standard offsetof() function.

Linked PRs

• gh-106869: Use new PyMemberDef constant names #106871

[vstinner]

I don't get why 3 constants are private, name starting with _Py prefix:

• _Py_T_OBJECT: replace T_OBJECT. It's still documented as T_OBJECT: https://docs.python.org/dev/c-api/structures.html#c.T_OBJECT "This results in surprising behavior in Python: deleting the attribute effectively sets it to None." There is Py_T_OBJECT_EX, but it's a different API: it raises an exception if the pointer is NULL, rather than returning None.
• _Py_T_NONE: replace T_NONE -- it's still documented as T_NONE: https://docs.python.org/dev/c-api/structures.html#c.T_NONE
• _Py_WRITE_RESTRICTED: replace PY_WRITE_RESTRICTED

By the way, (Py_AUDIT_READ | _Py_WRITE_RESTRICTED) replaces RESTRICTED, it uses _Py_WRITE_RESTRICTED.

[vstinner]

_Py_WRITE_RESTRICTED and its deprecated alias WRITE_RESTRICTED do nothing. Maybe their usage should just be removed in the code.

[vstinner]

I generated a big PR making all mechanical name changes at once: PR #106871. Maybe this PR is too big (ex: to be reviewed), I can slice it into smaller PRs (ex: 20 files max per PR).

I created this PR to see how it would look, so we can agree on how the change should be done.

cc @encukou @serhiy-storchaka

[vstinner]

Ok, I updated 78 files to use the new public API rather than "structmember.h" API which doesn't use the Py_ prefix.

Thanks @encukou for the new API :-)

About questions on the new API: well, they can be addressed separately.

For now, I replaced T_OBJECT with _Py_T_OBJECT. Maybe some files should be updated to use Py_T_OBJECT_EX and so raise AttributeError if the attribute was deleted (set to None), rather than returning None.

[serhiy-storchaka]

Maybe some files should be updated to use Py_T_OBJECT_EX and so raise AttributeError if the attribute was deleted (set to None), rather than returning None.

I am doing this right now. 😉

The main effect not in deletion, but that these attributes are None by default.

[encukou]

Usage of the old names meant that they were tested, however indirectly. Now the old names have no tests. Could you add tests?