Add unlifted versions of other primitive things

[treeowl]

The ones that come to mind off the top of my head:

• MVar
• MutVar
• Weak
• StablePtr
• TVar
• TMVar, using the "null pointer" trick.

and Ptr.
As I recall, there's a classic MVar-based concurrent queue implementation that involves MVar (MVar ...); we could offer an improved version.

[andrewthad]

Not Ptr and StablePtr. Those are both backed by Addr#. But all those other ones, yes.

[treeowl]

You're probably right about Ptr, but StablePtr is another story. That's backed by StablePtr#, which is a pointer into (or index in? Don't remember) the stable pointer table. It makes perfect sense to store unlifted things in that table, I believe.

[treeowl]

Oh, and StableName too! The type story there is stupid, so we would probably have something like

-- No type argument!
newtype StableName#
  = StableName (Exts.StableName# Any)

then just wrap the primops with unsafe coercions to make them work with unlifted things. We can make the same type work with lifted things while we're at it.

I believe another thing that makes sense is a version of StableName for singletons (or better, parameterized by value in a DependentHaskell future), but that doesn't seem like a good fit for this package.

[andrewthad]

StablePtr cannot be made to work. We had to remove the PrimUnlifted instance for StablePtr two years ago, and nothing about the representation of StablePtr# has changed since then. You wouldn't even be able to write a PrimUnlifted instance for it without unsafeCoerce#, and unsafeCoerce# between types of boxed and unboxed representation is never sound.

[treeowl]

I think you're mixing up concerns. StablePtr can't have a PrimUnlifted instance, but we can have an UnliftedStablePtr that points to unlifted objects.

[andrewthad]

You're right. I had misread. Yes, that should work, and I've actually done before in an application, so I'm pretty confident that it does work.