On Thu, Jan 3, 2013 at 4:57 AM, Simon Peyton-Jones
When matching instances, GHC does not take account of the context of the instance. Say you have
data T a = ... data S1 = ... data S2 = ... [a] instance Num a => Num (T a) where ... [b] instance Show a => Num (T a) where ... instance Num S1
and suppose you need an instance of (Num (T S1)). Then although [a] and [b] overlap, you might say we should use [a], since S1 is an instance of Num, but not an instance of Show. But GHC does not do this. It matches instances only based on the bit after the "=>".
It seems you're making GHC seem more capricious than it is here. Even were GHC to consider contexts in instance selection, the choice of [a] over [b] would still be incoherent: the compiler has no way to prove that there is no later instance adding S1 to class Show. Indeed, the S1's and S2's are in some sense not relevant: because Haskell provides no mechanism to exclude types from classes, there is no way to ever coherently use instance [a] or [b], regardless of the argument to T. /g -- Sent from my mail client.