TestEquality for references

[Edward Kmett]

I'd like to propose adding a bunch of instances for TestEquality and
TestCoercion to base and primitive types such as: IORef, STRef s, MVar as
well as MutVar and any appropriately uncoercible array types we have in
primitive.

With these you can learn about the equality of the types of elements of an
STRef when you go to

     testEquality :: STRef s a -> STRef s b -> Maybe (a :~: b)

I've been using an ad hoc versions of this on my own for some time, across
a wide array of packages, based on Atze van der Ploeg's paper:
https://dl.acm.org/citation.cfm?id=2976008 and currently I get by by
unsafeCoercing reallyUnsafePointerEquality# and unsafeCoercing the witness
that I get back in turn. =/

With this the notion of a "Key" introduced there can be safely modeled with
an STRef s (Proxy a).

This would make it {-# LANGUAGE Safe #-} for users to construct
heterogeneous container types that don't need Typeable information about
the values.

Implementation wise, these can either use the value equality of those
underlying primitive types and then produce a witness either by
unsafeCoerce, or by adding new stronger primitives in ghc-prim to produce
the witness in a type-safe manner, giving us well typed core all the way
down.

-Edward
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/libraries/attachments/20181202/beb88254/attachment.html>