Add type refinement for abstract type bindings

[nicolasstucki]

No description provided.

[smarter]

If I can write any type on the rhs of Refined, isn't this completely unsound?

[nicolasstucki]

There are still some constraints that must be added.

[Blaisorblade]

But do we have a way to add constraints that makes this sound? I thought the plan was to warn the user that they use an escape hatch, just like asInstanceOf. Maybe call the member UnsafeRefined to make this more visible?

[Blaisorblade]

We currently prefer a different approach to the problem — we lose guarantees on the values, but we get type soundness back, and @nicolasstucki's on it. Up to him to close or update this PR.

[nicolasstucki]

@Blaisorblade the current implementation in this PR is a partial implementation of the new scheme.

Outdated

[nicolasstucki]

The unboxed result of RefinedScrutinee has the expected performance. Here is a small experiment https://gist.github.com/nicolasstucki/96347efecaa5f6fb71d7146635c079c3

[Blaisorblade]

FWIW: last I tried to write value class decorators with a single member, I ran into #4583, which seems hard to avoid.

[nicolasstucki]

Will reopen it when I'll have time to work on it