gh-140382: Add support for __getattr__ method to Argument Clinic

[StanFromIreland]

• Issue: Support __getattr__ method in AC #140382

[merwok]

Could you create a new issue for this, and tag people interested in AC?

[StanFromIreland]

I can create a new issue, but all CODEOWNERs have already been notified. Do you have anyone else in mind?

[ZeroIntensity]

Wouldn't it make more sense to just support arbitrary dunder methods? __getattr__ isn't very common in C code, among other things.

[AA-Turner]

I'm also not convinced this is worth it. __getattr__() really ought to be implemented at the Python level, modules implemented entirely in C (like zlib) are very much a special case, and aren't used for new stdlib modules.

A

[ZeroIntensity]

On the other hand, I would be very supportive of allowing arbitrary dunders in AC, though I suspect that would be a lot of extra work (because we would need to map dunders to PyTypeObject fields).

[merwok]

I can create a new issue, but all CODEOWNERs have already been notified. Do you have anyone else in mind?

No, it’s just the general principle that «should we do?» discussions happen in issues (and when needed on the forums), then the PR can host «how to do?» discussions.

[skirpichev]

Wouldn't it make more sense to just support arbitrary dunder methods?

I think it's ok to add support of specific dunders on-demand. It's not too helpful in general, because sometimes we can do too much with slots, even override the docstring.

But please think about more generic case. For example, for most dunders we have very specific function signatures. It will be redundant and error-prone to specify this stuff by hand.

[vstinner]

Wouldn't it make more sense to just support arbitrary dunder methods? getattr isn't very common in C code, among other things.

I would prefer to only add support when it will really be needed, so we can test on actual code rather than unit tests. In short, I'm supportive of only adding support for __getattr__() for now.

[ZeroIntensity]

I don't think that will work well in practice. Judging by this PR, adding support for a new dunder method seems non-trivial. People will probably decide to just write things the old way instead of updating clinic, which we don't want.