fix: use underlying of VCs to compute generic signature

[hamzaremmal]

Have the correct generic signature generated (à la Scala 2) for value classes while at the same time preserving the fix in #20463

Closes #24276
Closes #22140

[lrytz]

(there seems to be some failed test, but how on do you find the failure in this huge log file..?)

[hamzaremmal]

So far, one test had to be disabled on Scala.js because of reflection and the other i8001 was just wrong.

[hamzaremmal]

Another error here is the following:

class Box[A](val value: A) extends AnyVal
class ArrayBox[A](val value: Array[A]) extends AnyVal

class Foo:
  def foo(x: Box[Int]): Box[Int] = Box(x.value)

Now, foo is erased to (java.lang.Integer): java.lang.Integer and the generated generic signature is (int): int. I know erasure cannot change but I'm wondering why we erase to java.lang.Integer instead of int? @sjrd any ideas?

[hamzaremmal]

Same for this method:

class ArrayBox[A](val value: Array[A]) extends AnyVal
class Foo:
  def foo(x: ArrayBox[Int]): ArrayBox[Int] = ArrayBox(x.value)

foo erased to (java.lang.Object): java.lang.Object rather than (int[]): int[].

[sjrd]

Probably historical. There is often not a good explanation for things involving value classes. They are full of weird non-optimal stuff.

[sjrd]

Now, foo is erased to (java.lang.Integer): java.lang.Integer and the generated generic signature is (int): int.

That is a real problem, though. We need the Java generic signature to be such that its (Java) erasure corresponds to the erased signature that we emit.

[lrytz]

Now, foo is erased to (java.lang.Integer): java.lang.Integer and the generated generic signature is (int): int.

That is a real problem, though. We need the Java generic signature to be such that its (Java) erasure corresponds to the erased signature that we emit.

Scala 2 seems to do the same. ASM output:

  // signature (I)I
  public foo(Ljava/lang/Integer;)Ljava/lang/Integer;

With 3.7.3 I get

  // signature (Ljava/lang/Integer;)Ljava/lang/Integer;
  public foo(Ljava/lang/Integer;)Ljava/lang/Integer;

So in the current form, this PR would align with the wrong behavior of Scala 2, IIUC?