While migrating our codebase from python 3.7 to python 3.8, one of our unit tests segfaulted and I’ve narrowed it down to the change of typeobject.c in:

  commit 351c67416ba4451eb3928fa0b2e933c2f25df1a3
  Author: Jeroen Demeyer <J.Demeyer@UGent.be>
  Date:   Fri May 10 19:21:11 2019 +0200
  bpo-35983: skip trashcan for subclasses (GH-11841)   

It seems that Py_TRASHCAN_SAFE_BEGIN/END are not backwards compatible (as claimed in the comment above their definition in object.h).

This patch has a unit test that currently fails, along with the change in typeobject.c that restores the code that broke it: 

https://github.com/iritkatriel/cpython/commit/d962dd7f800fdaaeacd5748c6e3d38bf1f5053c1

I believe that this change needs to be pushed until the Py_TRASHCAN_SAFE_BEGIN/END are removed and all users are forced to migrate to the new Py_TRASHCAN_BEGIN/END.

I need to check, but I think this is a duplicate of bpo-35983, which still has PR 12607 open.

Have you had a chance to look into this?

It took some debugging to get to the bottom of this problem in our system (nothing in the "what's new in python 3.8" doc gave any hint). 

The solution I went for was to replace the old macros with the new ones. If it's not easy or desirable to fix the old macros, should they not be removed?

Hi Irit,

Could you kindly check if PR 12607 solves this issue in any way?

I've confirmed that my test still segfaults on master, but if I take the Py_TRASHCAN_SAFE_BEGIN(op) from PR 12607 then it doesn't segfault.

I think that we should try to land PR 12607 first.

There has been no progress on PR12607. I think we should at least deprecate Py_TRASHCAN_SAFE_BEGIN/END in 3.10 and tell people to use Py_TRASHCAN_BEGIN/END instead.

Agreed with Irit. GH-12607 is conflicting with `main` and in an unclear review state. Irit's got both GH-20104 which fixes the segfault and GH-27693 which deprecates the old macros.

It's sad we missed the train for either GH-20104 or the original GH-12607 to get merged to 3.10 but it's a C API change so there's no discussion.

There's no hurry then in deciding for 3.11 if we leave the old macros as is and just deprecate them OR we both fix them using GH-20104 and deprecate them anyway. What are your thoughts, Pablo?

Re GH-20104 - What I did there was to simply revert the change that git bisect showed caused my segfault.

It's been a year since then, and I don't understand this stuff well enough to say that reverting this brings us to a good state, with respect to everything else that has changed. So let's be careful about that.

Superseeded by deprecation under Issue44874.