[issue10977] Concrete object C API needs abstract path for subclasses of builtin types

[Nick Coghlan]

Nick Coghlan <ncoghlan at gmail.com> added the comment:

1. It's an external API. We *don't control* most of the potentially broken code, so saying "just fix the call sites" effectively solves nothing and leaves classes like OrderedDict at risk of subtle silent corruption whenever they are passed to a 3rd party C extension.

2. The alternate "fix" is to tighten up the Py<X>_Check calls to Py<X>_CheckExact in order to reject subclasses. That would be a huge pain from a backwards compatibility point of view, so it isn't a realistic option.

3. Another alternate "fix" would be to exclude the concrete API from the stable ABI. That was already done for the macro interfaces, but it's too late for the functions - they were included in the stable interface for the 3.2 release.

4. The fact that the macros are irredeemably broken when used without ensuring that the supplied object is exactly the expected type is a very poor excuse for not fixing the function calls. It sounds like "The taillight is busted too, so let's not bother fixing the brakes".

5. There's a reason I put the "fast path" point second - that benefit is just an added bonus that comes from fixing the real design flaw that Raymond pointed out.

All that said, there really is one specific case where fixing this problem could cause a serious backwards compatibility issue: subclasses of builtin types that are *themselves* written in C. In such cases, the method implementations would quite likely use the concrete API to handle the base class state, then do their own thing to handle the rest of the update. Adding an implicit fallback to the concrete API implementations would cause such cases to trigger infinite recursion at the C level.

----------

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue10977>
_______________________________________