[Haskell-cafe] bug in ghci ?

[Kevin Quick]

On Thu, 08 Jul 2010 09:48:34 -0700, Daniel Fischer <daniel.is.fischer at web.de> wrote:

> On Thursday 08 July 2010 18:24:05, Ben Millwood wrote:
>> On Thu, Jul 8, 2010 at 3:45 PM, Daniel Fischer <daniel.is.fischer at web.de>
> wrote:
>> > Well, I made the suggestion of emitting a warning on instance
>> > declarations without method definitions. That would be comparatively
>> > easy to implement (even with an additional check to only emit the
>> > warning if the class defines any methods) and catch many (if not most)
>> > cases.
>>
>> Unfortunately, it would catch some perfectly valid cases, see the list
>> of instances for flat datatypes here:
>>
>> http://hackage.haskell.org/packages/archive/deepseq/1.1.0.0/doc/html/src
>> /Control-DeepSeq.html
>>
>> This demonstrates that there is at least one (admittedly probably not
>> much more than one) case where a class with methods would have a
>> default implementation that was total and valid in some cases.
>
> Good point.
> So one should check for more than one class-method [then defining no
> methods in the instance declaration is likely to lead to a default-method
> loop if there are default methods for all, otherwise GHC will warn
> already].
> That can of course still give rise to spurious warnings, but is less likely
> to.

I would think that only mutually recursive default methods would require respecification and that there could be any number of default methods that were reasonable as is.  Since it's probably quite difficult for the Haskell compiler to analytically detect non-terminating v.s. terminating mutual recursion it may be useful to define an explicit comment flag for this case.

For example:

    class Show a where
       shows = showsPrec 5
       showsPrec _ = shows
       {-# REDEFINE_ONE: shows showsPrec #-}

This would fairly simply allow a warning to be generated for an instance which did not redefine one of the identified methods; it would capture that requirement in the same place the recursive definition was defined, it would avoid false warnings, and it would be backward compatible (and it might be Haddock-able as well).

-- 
-KQ