Implicit priority in companions of the ancestors of a type

[julienrf]

Consider the following cocde:

object Priority {

  class Ev[-A, +B]

  class Foo[A]

  class Quux

  // Foo is associated with Bar
  // Quux is part of Foo
  class Bar extends Foo[Quux]

  object Foo {
    implicit val ev: Ev[Any, Boolean] = new Ev
  }

  object Bar {
    implicit val ev: Ev[Any, String] = new Ev
  }

  object Quux {
    implicit val ev: Ev[Any, Int] = new Ev
  }

  // Here we require an implicit parameter of type `Ev[Bar, ?]`.
  // Candidates will be looked up from the companion objects of the
  // types associated with `Bar`: `Bar`, `Foo` and `Quux`
  def usage[X](implicit ev: Ev[Bar, X]): Unit = ()

  usage
}

We get ambiguous candidates: Quux.ev and Bar.ev. Note that Foo.ev has a lower priority and does not cause ambiguity. I would expect Quux.ev to also have a lower priority because it is defined in the companion of a type (Quux) which is part of a type (Foo[Quux]) which is associated with Bar.

[milessabin]

I agree that it's maybe a little counterintuitive that Quux.ev has the same priority as Bar.ev, but I think it's pretty clear from the spec why that is. Bar extends Foo[Quux] so Bar.ev has higher priority than Foo.ev, but Bar doesn't extend Quux, so Bar.ev and Quux.ev have the same priority.

In line with the spec you could reduce the priority of Quux.ev relative to Bar.ev by having Bar also extend Quux,

trait Quux
class Bar extends Foo[Quux] with Quux

but most likely that's not going to make much sense in your real context.

How would you revise the overloading rules to get the effect that you want?

[julienrf]

Yes, I agree that according to the spec what we observe is expected.

I think it would make sense to reduce the priority of implicits found in companion objects of types that are part of a type which is associated with the type looked up.

[som-snytt]

If C is derived from D, then it is also derived from the parts of D? for SIP language, in "overload resolution." Maybe it's time to rename the feature, too. "Cog selection from a box of gears."

[milessabin]

@som-snytt yup, that would be a route we could explore, but it's a semantic change rather than a bugfix.