GeezerVison

I'm not sure what to make of the recommendation in the "what's new" paragraph. I specifically wouldn't recommend setting the size first, before setting the value. That seems incorrect and fragile. I'd remove it, but that'd leave us entirely without a recommendation.

I dislike PyList_New() + PyList_SET_ITEM() API since the list is immediately tracked by the GC and so calling gc.get_objects() can expose an invalid list object (ex: calling repr(list) can crash). See:

• Disallow creation of incomplete/inconsistent objects capi-workgroup/problems#56
• gh-107137: Add _PyTupleBuilder API to the internal C API #107139
• Avoid creating incomplete/invalid objects capi-workgroup/api-evolution#36

With this change, PyList_SetItem() and PyList_SET_ITEM() become inconsistent: PyList_SetItem() uses list->ob_size, whereas PyList_SET_ITEM() uses list->allocated.

PyList_SET_ITEM() is a strange beast: it doesn't DECREF the old value, it must only be used to initialize a list newly allocated by PyList_New(). But PyList_New(size) sets list->ob_size to size.

PyList_SetItem() is more regular: it calls DECREF on the old item.

Well, here the problem is not about designing a correct API, but to migrate Python 3.12 code to Python 3.13. Previously, PyList_SET_ITEM() could write safely in [0; allocated[ range, and in Python 3.13 it's limited to [0; ob_size[.

I'm not sure what to make of the recommendation in the "what's new" paragraph.

Since Python 3.13.0 final was not released, a Changelog entry is enough to document changes since 3.13 alpha1 (a NEWS.d entry added by the blurb tool).

I merged your change without Changelog entry, I don't think that it's needed for such subtile change. Thanks.

I created issue #111489: [C API] Make _PyList_FromArraySteal() function public.

without Changelog entry, I don't think that it's needed for such subtile change

I was referring to the existing whats-new paragraph, which says "if the assertion fails, set the size first". That is no longer true.

I dislike PyList_New() + PyList_SET_ITEM() API since the list is immediately tracked by the GC and so calling gc.get_objects() can expose an invalid list object (ex: calling repr(list) can crash)

That's why I prefer not requiring users to change the size before they set the value. Doing that would bring the list in an invalid state.

With this change, PyList_SetItem() and PyList_SET_ITEM() become inconsistent: PyList_SetItem() uses list->ob_size, whereas PyList_SET_ITEM() uses list->allocated.

I don't see an inconsistency here. They serve different needs, that's why we have two different functions. As you write, PyList_SetItem() decrefs the current value and replaces it with a new one. That can only safely be done within [0:size]. PyList_SET_ITEM() writes a value to an array position that does not yet have a value. That can only safely be done within [0:allocated]. Different needs, different functions, different boundary conditions.

You could rather argue that PyList_SET_ITEM() is only really valid within [size:allocated]. Within [0:size], it would overwrite existing, owned values. However, enforcing that would probably break even more code that might just look slightly brittle but is actually working perfectly fine (because it does what you suggested, it updates the size immediately before setting the value).

Cython uses the _SET_ITEM() functions to initialise freshly created lists and tuples, but it also calls PyList_SET_ITEM() in a fast, inlined PyList_Append() implementation:
https://github.com/cython/cython/blob/master/Cython/Utility/Optimize.c#L29-L77
We use this for list comprehensions, for example, where we control the list internally. As long as the next item fits into the already allocated array, we just put it there and increase the size.

I was referring to the existing whats-new paragraph, which says "if the assertion fails, set the size first". That is no longer true.

Ah oops, I tried to find a What's New entry, but I failed to find it, and so made the assumption that I didn't document the change. It should be updated, you're right.

Ah oops, I tried to find a What's New entry, but I failed to find it, and so made the assumption that I didn't document the change. It should be updated, you're right.

I wrote PR #111618 to update What's New in Python 3.13.