Specialize generics for value classes

[japgolly]

Using a fixpoint types, every node in the tree becomes boxed.

It would be great if Fix could be made a value class.

final case class Fix[F[_]](unfix: F[Fix[F]]) extends AnyVal

[odersky]

The line as given is compiled fine by dotty. Is there something else that does not work?

[DarkDimius]

I've added a test that will make sure that this code keep compiling.

[japgolly]

I just tested this with:

final case class Fix[F[_]](unfix: F[Fix[F]]) extends AnyVal

sealed trait Calc[+A]
case class Num(n: Int) extends Calc[Nothing]
case class Add[A](a: A, b: A) extends Calc[A]

object X {
  def c: Fix[Calc] =
    Fix[Calc](Add(
      Fix[Calc](Num(1)),
      Fix[Calc](Num(3))))
}

and the bytecode is:

  public java.lang.Object c();
    Code:
       0: getstatic     #23                 // Field Fix$.MODULE$:LFix$;
       3: getstatic     #28                 // Field Add$.MODULE$:LAdd$;
       6: new           #30                 // class Fix
       9: dup
      10: getstatic     #23                 // Field Fix$.MODULE$:LFix$;
      13: getstatic     #35                 // Field Num$.MODULE$:LNum$;
      16: iconst_1
      17: invokevirtual #39                 // Method Num$.apply:(I)LNum;
      20: invokevirtual #42                 // Method Fix$.apply:(Ljava/lang/Object;)Ljava/lang/Object;
      23: invokespecial #45                 // Method Fix."<init>":(Ljava/lang/Object;)V
      26: new           #30                 // class Fix
      29: dup
      30: getstatic     #23                 // Field Fix$.MODULE$:LFix$;
      33: getstatic     #35                 // Field Num$.MODULE$:LNum$;
      36: iconst_3
      37: invokevirtual #39                 // Method Num$.apply:(I)LNum;
      40: invokevirtual #42                 // Method Fix$.apply:(Ljava/lang/Object;)Ljava/lang/Object;
      43: invokespecial #45                 // Method Fix."<init>":(Ljava/lang/Object;)V
      46: invokevirtual #48                 // Method Add$.apply:(Ljava/lang/Object;Ljava/lang/Object;)LAdd;
      49: invokevirtual #42                 // Method Fix$.apply:(Ljava/lang/Object;)Ljava/lang/Object;
      52: areturn

The value class is being instantiated.
It would be awesome if the resulting bytecode were:

  public Calc c();
    Code:
       0: getstatic     #23                 // Field Add$.MODULE$:LAdd$;
       3: getstatic     #28                 // Field Num$.MODULE$:LNum$;
       6: iconst_1
       7: invokevirtual #32                 // Method Num$.apply:(I)LNum;
      10: getstatic     #28                 // Field Num$.MODULE$:LNum$;
      13: iconst_3
      14: invokevirtual #32                 // Method Num$.apply:(I)LNum;
      17: invokevirtual #35                 // Method Add$.apply:(Ljava/lang/Object;Ljava/lang/Object;)LAdd;
      20: areturn

[japgolly]

@odersky @DarkDimius should we reopen this ticket for the above, or should I create a separate one?

[DarkDimius]

@japgolly, The feature that your example needs to compile to efficient code is value-class specialization of Add. We have discussed a possible implementation idea in linker but it is very far from our current roadmap.

Currently, without value-class specialization, your optimized code has different semantics.
Let me illustrate this by having a slightly modified versions of add and Fix:

case class Add[A](a: A, b: A) extends Calc[A]{
 println(a.toString) // change
}

final case class Fix[F[_]](unfix: F[Fix[F]]) extends AnyVal {
  def toString() = "fix" // change
}

Which toString should Add call? in the bytecode you're suggesting the constructor argument of Add a will be passed unboxed and toString will be called on the a.unfix instead, which will not behave as Fix.toString.

[japgolly]

@DarkDimius Ah of course, I didn't even think about specialisation! So this will most likely be solved by Linker and WPO later?

[smarter]

Closing with stat:revisit since specialization isn't likely to come to Dotty soon.