bpo-35081: Make some _PyGC macros internal

[vstinner]

• Move "GC" macros together:

  • PyObject_IS_GC()
  • _Py_AS_GC()
  • _PyObject_GC_IS_TRACKED()
  • _PyObject_GC_MAY_BE_TRACKED()

• Mark other GC macros as internal (#ifdef Py_BUILD_CORE)

• Replace _PyGC_generation0 with _PyRuntime.gc.generation0

• _queuemodule.c: replace _PyObject_GC_UNTRACK()
  with with PyObject_GC_Track()

https://bugs.python.org/issue35081

[vstinner]

I'm not really satisfied by my previous attempts to "Move _PyObject_GC_TRACK() to pycore_object.h": PR #10272 and PR #10276, so I created this PR to "prepare the code" for the second change.

[methane]

• _PyObject_GC_UNTRACK()
  with with PyObject_GC_Track()

UNTRACK and Track are opposite.

[vstinner]

UNTRACK and Track are opposite.

Ooooops :-) I fixed my PR.

I just noticed that _PyObject_GC_TRACK() _PyObject_GC_UNTRACK() macros are documented as "should not be used for extension modules". If I understand correctly, they macros have been broken in Python 3.7 since _PyObject_GC_TRACK() now requires Include/internal/pycore_pystate.h (called Include/internal/pystate.h in Python 3.7) to access _PyRuntime.gc.generation0. Include/internal/ headers are not installed by "make install", and so not usable by 3rd party extensions.

I understand that my PR is not really backward incompatible, since the macros have been broken since Python 3.7 already. Moreover, they were never intended to be used outside CPython... right?

[vstinner]

I used git push --force a second time to include the full list of macros removed from the public API in the commit message:

• _PyGCHead_NEXT(g), _PyGCHead_SET_NEXT(g, p)
• _PyGCHead_PREV(g), _PyGCHead_SET_PREV(g, p)
• _PyGCHead_FINALIZED(g), _PyGCHead_SET_FINALIZED(g)
• _PyGC_FINALIZED(o), _PyGC_SET_FINALIZED(o)
• _PyObject_GC_TRACK(o), _PyObject_GC_UNTRACK(o)
• _PyGC_PREV_MASK_FINALIZED
• _PyGC_PREV_MASK_COLLECTING
• _PyGC_PREV_SHIFT
• _PyGC_PREV_MASK

[scoder]

Making this internal broke Cython (cython/cython#2721). It's used in the finaliser implementation, to determine if an object for which tp_dealloc() is called has already been finalised or whether we have to do it. I'm not sure how to deal with this on our side now. Any clue?

[etejedor]

Will this be ported to 3.7 too? The _PyObject_GC_TRACK macro is broken in 3.7 too, which makes e.g. PyCFunction_New crash. Thank you!

[vstinner]

Will this be ported to 3.7 too? The _PyObject_GC_TRACK macro is broken in 3.7 too, which makes e.g. PyCFunction_New crash. Thank you!

I'm not aware of this issue. Please open a new bug report at https://bugs.python.org/

[etejedor]

Thanks for the quick reply @vstinner . Sure, I will open a new bug report.

On the other hand, I found that you already mention here:

https://bugs.python.org/issue35229

that the _PyObject_GC_TRACK() and _PyObject_GC_UNTRACK() have been broken in 3.7 and you explain the reason.

[vstinner]

which makes e.g. PyCFunction_New crash

I'm not aware of that. Does it crash at runtime? In a module part of the stdlib or in a third-party C extension? If it's third-party, how is the C extension compiled? Well, please explain all these stuff in your bug report :-)

I deprecated the functions in Python 3.6 doc:
60a6bae

[etejedor]

@vstinner please find here the details:
https://bugs.python.org/issue35408