Make opaque type refinements of inline proxy objects abstract type constructors

[EugeneFlesselle]

Change the rhs of the added refinements from TypeAliases to RealTypeBoundss. This allows the opaque types to remain abstract type constructors, which is necessary for the MatchTypeCasePattern.AbstractTypeConstructor logic.

In particular, an AppliedType where the tycon was a reference to the refinement, used to dealias before proceeding with the comparison of the tycons, which is a requirement of the aforementioned AbstractTypeConstructor case of the MatchReducer.

Fixes #20427 Alternative to #20457

[sjrd]

This bound (ah ah!) to be break something. There are very real things we can do with type aliases that we cannot do with abstract type members of equal bounds. For example, manipulating arrays:

scala> class Foo { type T = Int }
// defined class Foo

scala> val foo = new Foo
val foo: Foo = Foo@2ca54da9

scala> val a: Array[Int] = { val b = new Array[foo.T](1); b }
val a: Array[Int] = Array(0)

scala> class Foo { type T >: Int <: Int }
// defined class Foo

scala> val foo = new Foo
val foo: Foo = Foo@2bfc2f8b

scala> val a: Array[Int] = { val b = new Array[foo.T](1); b }
-- [E172] Type Error: ----------------------------------------------------------
1 |val a: Array[Int] = { val b = new Array[foo.T](1); b }
  |                                                 ^
  |                                           No ClassTag available for foo.T
1 error found

[EugeneFlesselle]

@sjrd Indeed, but that (specific issue) should not be a problem for inlining right ? since implicit resolution is already done by that point. i.e. val b = Array[foo.T](1) is an error, but val b = Array[foo.T](1)(using evidence$1$proxy1 : ClassTag[Int]) is ok.

I am by no means claiming there is nothing else that could be impacted though.

[sjrd]

Perhaps. There are definitely other things that require an actual type alias. Extending or instantiating an aliased class, for example. And that will remain through inlining.

[EugeneFlesselle]

Again, just an empirical observation, but they both seem to work in the following simple test:

object A:
  opaque type W[T] = T

  inline def foo[T: ClassTag](x: T): Array[W[T]] =
    Array[W[T]](x)

  class Bar
  inline def bar =
    class Baz extends W[Bar]
    new W[Bar]

def Test =
  A.foo[Int](0) // ok
  A.bar // ok

[EugeneFlesselle]

It would be great to know if the changes break anything in the openCB, @WojciechMazur would it be possible to test that ?

[WojciechMazur]

I've started the OpenCB, I'll report the results when it's done

[EugeneFlesselle]

I've started the OpenCB, I'll report the results when it's done

Cool, thank you!

[WojciechMazur]

We've tested 1591 projects, 61 of them were already failing in last week nightly (3.5.1-RC1-bin-20240626-41f1489-NIGHTLY) and there are no new regressions found since that nightly version