Specificity in contravariant implicit resolution

[smarter]

The Dotty resolution algorithm was recently ported to scalac by @milessabin in scala/scala#6037 (comment) where this came up:

trait Foo[-T]

trait Bar[-T] extends Foo[T]

object Test {
  implicit val fa: Foo[Any] = ???
  implicit val ba: Bar[Int] = ???

  def test: Unit = {
    implicitly[Foo[Int]]
  }
}

This fails because Bar[Int] is not considered more specific than Foo[Any], because the variance-flipping we do in isAsSpecificValueType is not thorough enough:

[log frontend] ==> isSubType Bar[Int] <:< Foo[Any]?
[log frontend]   ==> isSubType Bar[Int] <:< Foo?
[log frontend]     ==> isSubType Foo <:< Foo?
[log frontend]     <== isSubType Foo <:< Foo = true
[log frontend]   <== isSubType Bar[Int] <:< Foo = true
[log frontend]   ==> hasMatchingMember(Bar[Int] . Foo$$T :?  =+ Any), mbr:  =- Int?
[log frontend]     ==> isSubType  =- Int <:<  =+ Any?
[log frontend]       ==> isSubType Any <:< Int?
[log frontend]       <== isSubType Any <:< Int = false
[log frontend]     <== isSubType  =- Int <:<  =+ Any = false
[log frontend]   <== hasMatchingMember(Bar[Int] . Foo$$T :?  =+ Any), mbr:  =- Int = false
[log frontend] <== isSubType Bar[Int] <:< Foo[Any] = false

hasMatchingMember returns the original member T in Bar where the variance is not flipped (=- Int). We could add a special case in hasMatchingMember but then we might miss some other cases.
Personally, I've never really been convinced by the rationale for flipping only at the top-level, so I think the simplest solution would be to flip at all levels, then we could use a global mode.

[milessabin]

@odersky I'd appreciate some clarification on this ... it's unlikely that the Scala PR will be merged if the intended semantics in Dotty are unclear.

[smarter]

We've agreed in our internal meeting that the current Dotty behavior is wrong and should be changed so that Bar[Int] is more specific; we've also agreed to keep the current "flip variance only at the top-level" semantics. Of course all of this might still change in the future based on user feedback, etc, which hopefully Miles' implementation will help us get!

[milessabin]

Fixed in scalac ;-)

[smarter]

Fixed in #3061, test case: https://github.com/lampepfl/dotty/blob/master/tests/pos/i2974.scala