You're right, of course. I'm now thinking we really do need ~R# to make this work. That's annoying, but not technically difficult. I'll continue to think about this. Richard On 2013-07-23 10:00, Joachim Breitner wrote:
Hi,
Am Dienstag, den 23.07.2013, 09:51 +0100 schrieb Richard Eisenberg:
A few responses:
- As Simon said, there is no great reason we don't have ~R# in Core. It's just that we looked through GHC and, without newtype coercions, there is no need for ~R#, so we opted not to include it. I'm still not convinced we need it for newtype coercions, though. What if we have this in Core?
class NT a b = MkNT { castNT :: a -> b ; uncastNT :: b -> a } NTCo:Age :: Age ~R# Int -- see [1] NTAgeInst :: NT Age Int NTAgeInst = MkNT { castNT = \(a :: Age). a |> NTCo:Age ; uncastNT = \(x :: Int). x |> sym NTCo:Age }
I thought about this class definition, and it is has the nice property that we can actually implement the methods by hand (without the zero-cost of course), which would be a good lint-like check that we do not generate illegal instances. The problem is that It it would not allow deriving instance NT a b => NT [a] [b] as there is no way to extract the coercion that was used in the implementation of NT a b. Hence the need to expose (to Core, not to tue user) the coercion in the class: The cast operations do not compose well.
Greetings, Joachim
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs