[Haskell-cafe] know a workaround for greedy context reduction?

[Nicolas Frisby]

I revisited the Strongly typed Heterogeneous Lists paper and read
about the import hierarchy technique. The basic idea is to delay
importing the instances until as late as possible, which prevents the
context simplification. The instances are effectively imported in the
top, Main module.

I'm thinking of exporting a MyLibrary.Main or MyLibrary.Instances module.

Anyone have experience with this approach in a library design? Is it
worth the user's extra import? Any pitfalls?

On Sun, Dec 7, 2008 at 4:57 PM, Nicolas Frisby <nicolas.frisby at gmail.com> wrote:
> Seems I got ahead of myself with the bug search. I was thinking bug
> because when I ascribe a type, I expect the compiler to check and then
> respect it. With the "most general type" specification of the ":type"
> command in mind, this does make sense. Thanks for improving my
> internal notion of ":type".
>
> My grumble may seem more legitimate from a library perspective. I
> implement a type-level function Append with three (preferably hidden)
> ancillary classes and a single instance in order to support the
> multiple modalities (in the Mercury sense) of the Append logic
> function. When a user defines another function that uses the append
> method, it's obfuscating for the user to see the internal classes in
> the inferred type. That's what I would like to workaround.
>
> If we consider class C the internal and consider class D and the
> function f the library's exposed interface, then I'd like to see C
> instead of D in the context of f and any function the user defines
> with f, especially when I have supplied a preferred type for f.
>
>> f :: D a => () -> a
>> f () = d
>
>> *> :t f
>> f :: (C a) => () -> a
>
> No dice?
>
> Thanks again,
> Nick
>
> On Sun, Dec 7, 2008 at 2:34 PM, Simon Peyton-Jones
> <simonpj at microsoft.com> wrote:
>> This is perfectly reasonable behavior I'm afraid.  If you do ":info d" you'll get d's original type signature.  But ":type" takes an *arbitrary expression* (in this case a single variable 'd', and figures out its most general type.  You could have said ":t (3*3)" for example.
>>
>> In this case, when inferring the most general type of the expression "d", GHC tries to simplify the context (D a), and uses the instance declaration to reduce it to (C a).  And then it can't simplify it further.  But you *might* have had
>>        instance C a
>> somewhere, in which case it'd have been able to simplify the (C a) away.  So GHC must try that route.  If it fails, you want it to "back up" to a notationally more convenient type, but GHC can't do that, I'm afraid
>>
>> Simon
>>
>> | -----Original Message-----
>> | From: haskell-cafe-bounces at haskell.org [mailto:haskell-cafe-
>> | bounces at haskell.org] On Behalf Of Nicolas Frisby
>> | Sent: 06 December 2008 03:23
>> | To: haskell Cafe
>> | Subject: [Haskell-cafe] know a workaround for greedy context reduction?
>> |
>> | With these three declarations
>> |
>> |   {-# LANGUAGE FlexibleInstances #-}
>> |   {-# LANGUAGE UndecidableInstances #-}
>> |
>> |   class C a where c :: a
>> |   class C a => D a where d :: a
>> |   instance C a => D a where d = c
>> |
>> | ghci exhibits this behavior:
>> |
>> |   *> :t d
>> |   d :: (C a) => a
>> |
>> | Where I would prefer "d :: (D a) => a". In my actual examples, the
>> | context is much larger and I can't involve overlapping instances. Is
>> | there a known workaround? I didn't find a related bug on the GHC trac,
>> | and I don't know if other compilers behave in the same way.
>> | _______________________________________________
>> | Haskell-Cafe mailing list
>> | Haskell-Cafe at haskell.org
>> | http://www.haskell.org/mailman/listinfo/haskell-cafe
>>
>>
>