gh-46376: minimal fix for setting pointers via pointers in ctypes

[code-of-kpp]

Previous attempts to address the oldest open issue (gh-46376) were a bit too ambitious.

Here, I tackle half of the problem: the segfaults.

There’s still a soft memory leak (objects previously assigned to pointer contents are kept in _objects). It’s completely unclear how to fix these and if it’s even possible at all. This situation is similar to cast(), where we don’t know anymore if the memory can be cleaned if the object was casted at least once.

A new idea is that in some cases, specifically when we have pointer, structure, array, or union assignment to pointer’s contents, we will save the RHS value object into pointer’s b_base. We only need to do this when b_base is NULL/None or when we are in a chain of pointers that start from another pointer created directly with a pointer() call (so NOT as part of structure, union, etc.).

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[code-of-kpp]

/cc @ambv and @encukou