Defining types is messy

[encukou]

I had several false starts on this issue and always got too deep in describing the current implementation. So I'll post a half-baked list to get something down faster.

• How are/should fields handled, esp. special ones like type, dict, and superclass ones:
  • What should tp_traverse visit?
  • What code is in charge of initializing them? Zeroing them before initialization? Clearing them?
• Can an object outlive tp_clear?
• Can I rely on tp_clear being called?
• Which initialization/finalization APIs call a type's slots, and which ones bypass them?

etc.

[gvanrossum]

See for example python/cpython#111339

[encukou]

That one is about defining internal static types, which are quite special (shareable across subinterpreters, deep-freezeable). Defining those isn't (yet) part of the public API; extensions need to either use heap types or stick to a single interpreter.

[encukou]

FWIW, here are some discussions about solutions:

• Design type creation API markshannon/New-C-API-for-Python#4
• Grand Unified Python Object Layout faster-cpython/ideas#553