Constraints of function in record appears to nullifiy instance constraints.

[ben.kavanagh]

If an field of a record that has constraints in its type, defining a value for this field appears to cause the compiler to ignore outer level constraints declared for an instance.

This problem restricts generic functions written in syb-with-class and replib style to functions that do not have constraints.

[ben.kavanagh]

This is the code that fails.

[ben.kavanagh]

This is the same program with (Monad m) constraint removed. It compiles.

[simonpj]

Here are the key lines of code:

class Subst a t t' where
    subst :: (Monad m) => a -> t -> t' -> m t'

data SubstD a t t' 
  = SubstD (forall m. Monad m => a -> t -> t' -> m t')

instance Data (SubstD a t) t' => Subst a t t'          -- (1)

instance Subst a t t' => Sat (SubstD a t t') where     -- (2)
    dict = SubstD subst

The call to 'subst' on the last line gives rise to a constraint (Subst a t t'). But that constraint can be satisfied in two different ways:

1. Using the instance declaration for Subst (which matches anything!)
2. Using the context of the Sat (SubstD ..) instance declaration itself

If GHC uses (1) it gets into a corner it can't get out of, because now it needs (Data (SubstD a t) t'), and that it can't get. The error message is a bit misleading:

T3018.hs:29:28:
    Could not deduce (Data (SubstD a t) t') from the context (Monad m)
      arising from a use of `subst' at T3018.hs:29:28-32

it should really say

 ...from the context (Subst a t t', Monad m)

but that's a bit of a separate matter.

Now, you are hoping that (2) will happen, but I hope you can see that it's delicate. Adding the (Monad m) context just tips things over the edge so that GHC doesn't "see" the (Subst a t t') in the context until too late. But the real problem is that you are asking too much. Here is a simpler example:

f :: Eq [a] => a -> blah
f x = let g :: Int -> Int
          g = ....([x]==[x])...
      in ...

The use of == requires Eq [a], but GHC will probably use the list equality instance to simplify this to Eq a; and then it can't deduce Eq a from Eq [a]. Local constraints that shadow or override global instance declarations are extremely delicate.

All this is perhaps soluble if GHC were to be lazier about solving constraints, and only makes the attempt when it has all the evidence in hand. I'm thinking quite a bit about constraint solving at the moment and will bear that in mind. But I can't offer you an immediate solution. At least I hope I've explained the problem.