use `FunctionalInterface` to indicate `trait` cannot have a constructor, or violate other SAM type conditions

[adriaanm]

to deal with no-op constructors gratuitously being added, ruining a trait's potential as a SAM

[retronym]

Noting some overlap with universal traits, in which extends Any prevents constructors.

scala> trait T extends Any { def foo = 14 }
defined trait T

scala> :javap -private T
Compiled from "<console>"
public interface T {
  public abstract int foo();
}

scala> trait T extends Any { def foo = 14; ??? }
<console>:11: error: this statement is not allowed in universal trait extending from class Any
       trait T extends Any { def foo = 14; ??? }
                                           ^

[adriaanm]

ref #191

[lrytz]

what's the proposal exactly? that the compiler automatically adds @FI to interface classfiles that have only methods (no fields, init stats)?

[retronym]

I take it to mean that if a trait is so annotated, we'll a) error if a constructor is required, b) error if any ancestor has a constructor and c) allow LMF to use it as a lambda target type.

[retronym]

Maybe FI is the wrong annotation to mark this, though

[lrytz]

It seems unfortunate to me that the user is required add an annotation in order to get the LMF translation.

[retronym]

Separate compilation and the generality of traits are the thorns in our side.

We might be able to do better down the track with a ScalaLambdaMetafactory that checks at link time which ancestor traits actually have mixin constructors, and generates a suitable constructor for the generated anonymous class. We'd need to pass make the linearization order known the the bootstrap method to call them in the right order.

You could even implement fields in the same manner.

[lrytz]

That sounds interesting! Maybe at some point having our own LMF will solve enough issues for us to consider it.

But back to the proposal. Suppose I write @FunctionalInterface trait B extends A { def sam(): Int } with A coming from an external library. If A changes things may break, but that's irrespective of whether the @FI annotation is user-written or automatically added.

[retronym]

Let's call the annotation @NoInit (named after the dotty flag of the same name), rather than @FunctionaInterface. (The latter has the extra meaning of "has exactly one abstract method" which is a separate concern.)

@NoInit would mean: this trait has not initializer today, and all its Scala defined ancestors are must also have this annotation. By using this annotation, I'm saying that intend to keep things that way, so clients of the class can omit the call to $init$.

Traits are super fragile with separate compilation, but the current scheme allows you to add side effects to the constructor of a trait without recompiling subclasses.

There is just the one special case that if a trait has no method bodies in its decls, the $init$ method and the impl class are elided. That inconsistency brings separate compilation problems: https://gist.github.com/retronym/af3037454509de9412fc8e1d5d38d43c

The closest thing we have to this today is trait T extends AnyVal, although that brings the addtional restriction that you can't extends arbitrary Java interfaces and that you can't refer to AnyRef methods.

[DarkDimius]

For the record, Dotty uses NoInits flag to represent similar notion, with only difference that it does not speak about ancestors. Why would you need transitivity?

[retronym]

With dotty, compile:

trait T
trait U extends T { def apply(a: Any): Any }
object Test {
  def main(args: Array[String]): Unit = {
    val u: U = x => x
    assert(u("") == "")
  }
}

Then, recompile T as:

trait T { throw new Exception("hello?") }

Rerunning Test, we don't get the Exception("hello"), nor do we get a LinkageError.

Dotty does seem to respect ancestors when computing whether to use indy or not:

trait T { ??? }
trait U extends T { def apply(a: Any): Any }
object Test {
  def main(args: Array[String]): Unit = {
    val u: U = x => x // new Test$$anonfun$2
    assert(u("") == "")
  }
}

[retronym]

This is conceptually the same as in Java:

// run 1
interface I {}
class C implements I {}

class I

This would result in an IncompatibleClassChangeError when loading C.