I am taking the approach of putting the module in the capsule struct. This requires the client code to pass in the module (and will need some note in the changelog)

I created https://bugs.python.org/issue41798 "meta-issue" to discuss the usage of the PyCapsule C API with multi-phase init API (PEP 489).

I am taking the approach of putting the module in the capsule struct. This requires the client code to pass in the module (and will need some note in the changelog)

That sounds like a good plan, to pass the module through a state.

I would prefer with a first short PR to prepare the work to just pass a state to UCD_Check(), without all other changes.

This PR is too big, I cannot review it. GitHub even hides the diff by default because it is too large :-D

@vstinner Since the module state (which contains the capsule and a type) is new in this PR, it would be a struct with sizeof of 0. I can put an int in it just as a placeholder but I don't know if its a good idea to merge that in.

Most of the changes are breaking the functions (which are shared between the module and the type) into different entry points for either - so that I can get the module state from both flavors in the way that works for them. I propose to make the first stage just about this indirection, without switching to any heap types. What do you think?

@vstinner Since the module state (which contains the capsule and a type) is new in this PR, it would be a struct with sizeof of 0. I can put an int in it just as a placeholder but I don't know if its a good idea to merge that in.

You can try to work in multiple steps (multiple PRs):

• Add a ucd_type parameter to UCD_Check() macro
• Convert UCD_Type static type to a heap type without adding a module state (right now, it's not possible to create more than one instance of the unicodedata module)
• Add a module state

For the module state, it may be more flexible to pass ucnhash_CAPI to getname() and getcode() this API, rather than passing ucnhash_CAPI->module as the first parameter.

@vstinner step 1 is equivalent to Py_IS_TYPE, but sure I can do that first.

Are there any examples of heap types without module state that I can refer to?

For the module state, it may be more flexible to pass ucnhash_CAPI to getname() and getcode() this API, rather than passing ucnhash_CAPI->module as the first parameter.

@vstinner This is definitely pythonic 👍

This PR is outdated, unicodedata got many changes in the meanwhile. I proposed one approach to convert unicodedata to multi-phase init in https://bugs.python.org/issue42157 I close this PR. Once PR #22990 will be merged, I will propose a PR to finally convert the module to multi-phase init. Sorry for the misunderstanding, but this extension module is way more complex than other extensions, and I didn't spot all corner cases at the first review. See the bpo for the list of all issues and my proposal.