[Haskell-cafe] Cons of -XUndecidableInstances

[James Cook]

On Jun 7, 2011, at 12:43 PM, MigMit wrote:

>
>> One particularly trivial example that comes to mind is:
>>
>>   newtype Mu f = Mu (f (Mu f))
>>
>>   instance Show (f (Mu f)) => Show (Mu f) where
>>       show (Mu x) = "Mu (" ++ show x ++ ")"
>>       -- Or however you'd like to show it
>
> Ehm, that does look like poor design.
>
> Sure you don't mean "Mu f can be printed if and only if f (Mu f) can  
> be printed". What you probably mean is "if f transforms printable  
> things to printable things, then Mu f is a printable thing". And you  
> CAN express just that:
>

Actually, I would argue that the former _is_ what is meant.  It's a  
weaker condition than the latter and it is the necessary and  
sufficient condition to define the instance - one of the steps  
involved in formatting a value of type "Mu f" is to format a value of  
type "f (Mu f)".  It doesn't actually matter whether "forall x. Show x  
=> Show (f x)" holds in general.

> type ShowD a = forall p. (forall x. Show x => p x) -> p a
>
> showD :: Show a => ShowD a
> showD px = px
>
> class ShowF f where showF :: Show a => ShowD (f a)
>
> instance Show a => Show (F a) where... -- here goes your "f"
>
> instance ShowF F where showF = showD -- and that is the only line of  
> boilerplate
>
> instance ShowF f => Show (Mu f) where
>  show (Mu fm) = "Mu (" ++ runShowHelper (showF (ShowHelper show)) fm  
> ++ ")"
>
> newtype ShowHelper x = ShowHelper {runShowHelper :: x -> String}
>
> Sorry for possible bugs — I don't have ghc anywhere near me at the  
> moment, but the idea is clear, I guess.
>

I don't really see how this is preferable when the compiler can solve  
the equation automatically.  All that is needed is to tell it to try.   
If portability is a concern then I could see going through the  
gymnastics (and also eliminating the use of higher-rank types), but  
that's the only case in which I would consider it the preferred option.

-- James