Fix desugaring of context bounds in extensions

[rjolly]

Fixes #11586
Unfixes #11583

[neko-kai]

I guess the change should avoid breaking #11583 to have a chance...

[rjolly]

@neko-kai see my comment in #11584

[bishabosha]

I find the main issue here is that this change results in a worse search for eligible extensions (from the same source) than the status quo.
If the compiler never improved search for extension methods, (where there is a context parameter after the extended parameter), would it be worth all the manual boilerplate to replicate the previous desugaring and get the old behaviour?

[neko-kai]

@bishabosha I think the desugaring is not at fault here since you're going to get the same type inference behavior if you were to write using Functor[F] instead of : Functor.

Previous desugaring would merely mask the issue by making context bounds infer "better" by making them more than just a short-hand which they are usually, then cause surprising change in behavior if a context bound was replaced with an explicit using/implicit clause.

[bishabosha]

@bishabosha I think the desugaring is not at fault here since you're going to get the same type inference behavior if you were to write using Functor[F] instead of : Functor.

Previous desugaring would merely mask the issue by making context bounds infer "better" by making them more than just a short-hand which they are usually, then cause surprising change in behavior if a context bound was replaced with an explicit using/implicit clause.

Sorry, I should perhaps rephrase my original comment, I am in agreement in principle with the new proposed desugaring. However I am concerned that perhaps it should not go forward right now, purely because an extension such as

extension [F[_], A](f: F[A])(using Functor[F]) def ...

can not be found as easily by the compiler (therefore leading to more boilerplate to put all using clauses at the end of extension methods, a.k.a - how context bounds desugar before this PR).

[neko-kai]

However I am concerned that perhaps it should not go forward right now, purely because an extension such as can not be found as easily by the compiler (therefore leading to more boilerplate to put all using clauses at the end of extension methods, a.k.a - how context bounds desugar before this PR).

Yeah, but you'd lose that property immediately if you wrote an explicit using instead of a context bound anyway. I think ripping the band-aid early, while there are still very few users that would be impacted is better than waiting until there are even further impediments to this, such as binary compatibility, that would cause a temporary change 'for convenience' to be stuck in the language forever and become another embarassing inconsistency that has to be explained over and over to anyone who encounters it...

[bishabosha]

LGTM

[bishabosha]

@rjolly this PR was reverted after merging because it was decided to wait to fix issues like #11713 first

[rjolly]

Yes, I saw that, thanks

[griggt]

Should this be revisited now that #11713 is fixed?

[bishabosha]

would that be a scala 4 change?