The idea can be extended to other situations where unsafeCoerce is
justified: empty types and phantom types. Given declarations
data Const a b = Const a
we can create O(1) wrappers:
f :: F Void -> F Void' -- empty datatypes are exchangeable
g :: F (Const a Int) -> F (Const a Bool) -- phantom parameter is insignificant
assuming they do not mess with invariants of F.
From this perspective I prefer proposal 2 (a new constraint): new types of
coercions might be added by adding special cases in the constraint solver.
It is also a reuse of syntax: we can write
foo = newtypeCast :: [Int] -> [Age]
instead of adding a new form of declaration.
As for safety, in some sense Map a b is parametric in a and b except
the Ord instance of a. We can coerce Map a b into Map c d if
we can coerce a into c, b into d and the Ord a instance matches
Ord c instance (for example, one is derived from the other).
This could be implemented via an instance like this:
instance (a ~~ c, b ~~ d, Ord a ~~ Ord c) => Map a b ~~ Map c d
where ~~ is kind-polymorphic "equality up to newtypes".
If Map was, say, indexed (nonparametric) in the first argument,
that would be: