+Note [Dictionary-like types]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Being "dictionary-like" means either a dictionary type or a tuple thereof.
+In GHC 6.10 we build implication constraints which construct such tuples,
+and if we land up with a binding
+ t :: (C [a], Eq [a])
+ t = blah
+then we want to treat t as cheap under "-fdicts-cheap" for example.
+(Implication constraints are normally inlined, but sadly not if the
+occurrence is itself inside an INLINE function! Until we revise the
+handling of implication constraints, that is.) This turned out to
+be important in getting good arities in DPH code. Example:
+
+ class C a
+ class D a where { foo :: a -> a }
+ instance C a => D (Maybe a) where { foo x = x }
+
+ bar :: (C a, C b) => a -> b -> (Maybe a, Maybe b)
+ {-# INLINE bar #-}
+ bar x y = (foo (Just x), foo (Just y))
+
+Then 'bar' should jolly well have arity 4 (two dicts, two args), but
+we ended up with something like
+ bar = __inline_me__ (\d1,d2. let t :: (D (Maybe a), D (Maybe b)) = ...
+ in \x,y. <blah>)
+
+This is all a bit ad-hoc; eg it relies on knowing that implication
+constraints build tuples.
+