Context-bound on unnamed type parameter becomes Any

[dwijnand]

Welcome to Scala 2.12.8 (OpenJDK 64-Bit Server VM, Java 11.0.2).
Type in expressions for evaluation. Or try :help.

scala> class F[_]
defined class F

scala> class A[_: F]
defined class A

scala> implicit val x: F[Int] = new F[Int]
x: F[Int] = F@217235f5

scala> new A[Int]
<console>:14: error: could not find implicit value for evidence parameter of type F[Any]
       new A[Int]
       ^

[szeiger]

Not limited to classes:

scala> def f[_ : F] = ()
[[syntax trees at end of                     typer]] // <console>
...
        def f[_](implicit evidence$1: F[Any]): Unit = ()

The relevant section of the spec is https://www.scala-lang.org/files/archive/spec/2.11/07-implicits.html#context-bounds-and-view-bounds. It does not mention wildcards explicitly but the current behavior violates this part:

A type parameter A of a method or non-trait class may also have one or more context bounds A : T. In this case the type parameter may be instantiated to any type S for which evidence exists at the instantiation point that S satisfies the bound T. Such evidence consists of an implicit value with type T[S].

[szeiger]

Let met try my best @smarter impersonation: "Works in Dotty"

Well, sort of. Dotty doesn't allow wildcard type parameters anymore so the problem doesn't come up in the first place. If that's the way forward then we should do the same in 2.14.

[smarter]

If that's the way forward then we should do the same in 2.14.

I believe so, I'd say the fact that wildcards can be used in type parameter position is just an accident of the parser implementation and not a real feature.

[dwijnand]

Is there a problem is supporting this pattern?

On the other side of the equation: can we find a good reason why to support this? Maybe we should run the community build against the removal to see how/why it's used.

[adriaanm]

It's not intended to be a wildcard, but an unnamed type parameter. I think that's that the bug is.

[adriaanm]

Just another use of _ 🤷‍♂️

[szeiger]

I'd say the fact that wildcards can be used in type parameter position is just an accident of the parser implementation

The syntax is officially documented in the spec though. Just not the semantics, which shouldn't be a problem if they are simply unnamed type parameters. I also assumed that's what they are. In this case it should be easy to fix by synthesizing a name very early (parsers or namers) before the context bound desugaring happens.

Good reasons for disallowing this syntax:

• It's inconsistent with value parameters:

  scala> def f(_: Int) = ()
  <console>:1: error: identifier expected but '_' found.
  def f(_: Int) = ()
        ^
  
  scala> def f[_] = ()
  f: [_]=> Unit
  

• It's yet another use for _ in types that can be confused with type constructors and existentials.

[szeiger]

OTOH it is very similar to the use of _ in type members where it also stands for an unnamed type parameter:

scala> type F[_] = Int
// defined alias type F = [_$11] => Int

[dwijnand]

If we don't drop unnamed type parameters, we might want to lift the restriction on naming multiple proper type parameters _:

scala> class G[_]
defined class G

scala> class H[_]
defined class H

scala> class A[_: G, _: H]
<console>:13: error: _ is already defined as type _
       class A[_: G, _: H]
                     ^

like you can for the parameters of a type constructor:

scala> class F[_[_, _]]
defined class F

• It's inconsistent with value parameters:

  scala> def f(_: Int) = ()
  <console>:1: error: identifier expected but '_' found.
  def f(_: Int) = ()
        ^
  
  scala> def f[_] = ()
  f: [_]=> Unit
  

It would be nice to be able to name a parameter _, so to avoid triggering the unused lint:

15:20:13 $ sc -Ywarn-unused
Welcome to Scala 2.12.8 (OpenJDK 64-Bit Server VM, Java 11.0.2).
Type in expressions for evaluation. Or try :help.

scala> class A { def foo(x: Int) = () }; final class B extends A { override def foo(x: Int) = println(x) }
<console>:11: warning: parameter value x in method foo is never used
       class A { def foo(x: Int) = () }; final class B extends A { override def foo(x: Int) = println(x) }
                         ^
defined class A
defined class B

[smarter]

Please no, no extra meaning for _.

[dwijnand]

I wouldn't classify it as extra. It's the same meaning that already exists: the parameter is being (un)named _. It would be removing an inconsistency (which, I thought generally was well appreciated.)

[som-snytt]

Deprecated but supported by scala/scala#9712