GeezerVison

Points taken. Sorry X[int, P] slipped my mind since I hadn't fixed __parameters__ yet.

D'oh, I just realised I answered my own question since we need __parameters__ to implement X[int, P] in the first place.

Let's wait until #23060 has landed and then see what adjustments this might need.

Finally looking at this...

Actually I just realized PEP 612 doesn't explicitly say that ParamSpec should support TypeVar-like substitution. Maybe I over complicated things a little by allowing that in types.GenericAlias (just for collections.abc.Callable)?

Yeah, ParamSpec is super special, and you can't do anything with it that isn't explicitly mentioned in the PEP. So it's debatable whether Callable[P, int] should even have P in its __parameters__. Then again, for the simple case A = Callable[P, int] there's no reason why we couldn't allow A[[str, str]] and get Callable[[str, str], int] as the result. But type checkers wouldn't know what to do with that -- ParamSpec is a magic cookie.

So let's not worry about those clearly wrong cases.

Sorry, submitted too soon. (UPDATE: Not, I made a mistake -- or GitHub did. :-)

Guido, thank you so much for all the time you've put into reviewing this PR! I learnt quite a bit about the typing module in the process :).

Re: the 3.9 issues that you discovered: maybe create a PR for those first, land that in master and 3.9 branch, then merge master into this PR?

Yeah I agree, I created #23915 for that. Once that one lands, I'll merge the changes in along with some of your suggestions.

I think it’s ready!

@Fidget-Spinner Let me know if you agree.

@Fidget-Spinner Let me know if you agree.

Yes! Sorry, I thought you were going to merge it in so I was waiting too 😆 .

I figured that was the case. :-)

@gvanrossum: Please replace # with GH- in the commit message next time. Thanks!

@gvanrossum Oh dang I knew I had a hunch I missed something :/. I forgot to convert all lists to tuples in ga_new/setup_ga too (I did it only for getitem and co.). Well on the bright side, having that as a separate PR makes backporting to 3.9 easier if anything :), and anyways lists aren't valid in genericalias other than the Callable case (which we already have covered), so we won't have anything broken even without that in right now.

I'll fix that some other time, right now I'm celebrating 🥳