[Haskell-cafe] Re: Map constructor in a DSL

[Emil Axelsson]

2010-10-28 12:09, Dupont Corentin skrev:
> I'm also looking at the Atom's DSL to get inspiration.
> Something I don't understand in it is that it has two languages, on typed:
>
> data  E  a  where
>    VRef     ::  V  a  ->  E  a
>    Const    ::  a  ->  E  a
>    Cast     ::  (NumE  a,  NumE  b)  =>  E  a  ->  E  b
>    Add      ::  NumE  a  =>  E  a  ->  E  a  ->  E  a
>
> etc.
>
> And, along with it, an untyped counterpart:
>
> -- | An untyped term.
> data  UE
>
>    =  UVRef  UV
>    |  UConst  Const
>
>    |  UCast   Type  UE
>    |  UAdd    UE  UE
>
> etc.
>
> What that for? What's the use of having beautiful GADT if you have to maintain an untyped ADT aside??

The general reason for this (I can't speak for Atom specifically) is 
that the typed representation can be quite hard to work with when you 
want to transform the expressions. But you can still often limit the use 
of the untyped representation to the last stages in the backend, which 
means that you can enjoy the typed representation in the initial stages.

In the development version of Feldspar, we use a typed representation 
combined with type-safe casting using Data.Typeable. Although not ideal, 
this seems to be better than having two different representations.

/ Emil