Dealias type when type parameters inference occurs

[Alex1005a]

The code below displays the type parameter [Res] =>> TypeAl[F, String, Res] for the toFunctorOps implicit conversion.

import cats.Functor
import cats.syntax.all.*

type Result[F[_], Err, Res] = F[Either[Err, Res]]

def test[F[_]: Functor](x: Result[F, String, Int]) = {
  x.map(identity) // error here
}

To fix this, you can do dealias if the type before dealias does not have the same arity as the constructor with which unification occurs, but the type after dealias does have the same arity. Then the type parameter F will be inferred.

[odersky]

I think @smarter would be the best person to review this. Guillaume, can you take a look? You have thought most about dealiasing and type inference.

[Gedochao]

@smarter @som-snytt
Have you maybe managed to take a look?

[som-snytt]

I'll learn something by taking a look. I recently removed dealiasing from Scala 2's typedNew and also contended in Scala 3 with the type of new A.C for C[T] where the unused check sees type [T] =>> this.A.C[T] at #23699

Edit: apparently I already have a branch review/23242...

[smarter]

Sorry for the late reply! So having type aliases influence type inference here was a conscious choice so that users could customize higher-order unification by defining their own type aliases. The logic proposed in this PR is sound, but I worry it's making the rules around type inference more complex. An alternative here would be to use a curried type alias to get unification to do exactly what you want:

//> using dep org.typelevel::cats-core:2.13.0

import cats.Functor
import cats.syntax.all.*

type Result[Err, Res] = [F[_]] =>> F[Either[Err, Res]]

def test[F[_]: Functor](x: Result[String, Int][F]) =
  x.map(identity) // ok

[som-snytt]

Thanks for the timely comment (just back from grocery shopping). This morning I got up to speed on the code and suspected there might be a user solution, but I'm not used to writing the funny arrow. I thought the point was that "unification", as a convention, already does what I would expect.

I saw that the PR flips canInstantiate in a way that seems obvious. I'll see if I have an opinion about the tweak. Rules for type inference only matter when it seems to infer the wrong thing (as in this case).