[Haskell-cafe] Tricks with GMap -- question about conflicts w/ indexed type families

[Ryan Newton]

GMaps -- families of map implementations indexed by the key type --
are an example on the wiki:

  http://www.haskell.org/haskellwiki/GHC/Type_families

I've been using something like this myself.  It sure would be nice to
have a fully developed version on Hackage, and I may try to submit
this myself if no one else has already done it.  But I'm running into
a problem in achieving a particular optimization that I want --
namely, small tuple keys packed into Ints.

For example, with a key type (Int16,Int16) it's possible to pack the
key into a plain Int and use Data.IntMap.

In fact, in the rough/partial implementation here
(http://people.csail.mit.edu/newton/gmap/GMap.hs) I have a class for
things that fit in a word:

class FitInWord v where
  toWord   :: v -> Word
  fromWord :: Word -> v

(Or fitInInt for that matter.)  Which includes things like small tuples:

instance FitInWord (Int16,Int16) where
  toWord (a,b) = shiftL (fromIntegral a) 16 + (fromIntegral b)
  fromWord n = (fromIntegral$ shiftR n 16,
		fromIntegral$ n .&. 0xFFFF)

(If you know a better idea than using templates/program generation to
generate every combination of small scalar tuples that fit in a
word.... let me know.)
What I would next *like* to do is something like the following:

import qualified Data.IntMap as DI
instance FitInWord t => GMapKey t where
 data GMap t v           = GMapInt (DI.IntMap v) deriving Show
 empty                   = GMapInt DI.empty
 lookup k    (GMapInt m) = DI.lookup (wordToInt$ toWord k) m
 insert k v  (GMapInt m) = GMapInt (DI.insert (wordToInt$ toWord k) v m)
 alter  fn k (GMapInt m) = GMapInt (DI.alter fn (wordToInt$ toWord k) m)
 toList      (GMapInt m) = map (\ (i,v) -> (fromWord$ intToWord i, v)) $
   DI.toList m

The problem is that there's already a more general instance of GMapKey
that handles pairs by representing them as nested GMaps:

instance (GMapKey a, GMapKey b) => GMapKey (a, b) where
  data GMap (a, b) v            = GMapPair (GMap a (GMap b v))
  ....

Ideally, I want both of these to coexist (and to prioritize the more
specific one).  With normal type classes, OverlappingInstances can
handle this, but with type families I get an error like the following:

   Conflicting family instance declarations:
      data instance GMap t v -- Defined at Intel/GMap.hs:108:7-10
      data instance GMap (a, b) v -- Defined at Intel/GMap.hs:225:7-10

Any fixes?

Thanks,
-Ryan