> | > On 1/3/07, Roberto Zunino <zun...@di.unipi.it> wrote:
> | >> 1) Why the first version did not typececk?
> | >
> | > 1) Class constraints can't be used on pattern matching. They ARE
> | > restrictive on construction, however. This is arguably bug in the
> | > Haskell standard. It is fixed in GHC HEAD for datatypes declared
> | > in the GADT way, so as not to break H98 code:
> |
> | http://article.gmane.org/gmane.comp.lang.haskell.cvs.all/29458/matc
> |h=gadt+class+context
> |
> | To quote from there: "I think this is stupid, but it's what H98
> | says."
> |
> | Maybe it is time to consider it deprecated to follow the Haskell 98
> | standard /to the letter/.

> GHC follows this strange standard when you write data type decls in
> H98 form

>         data Eq a => T a = C a Int  |  D

> Here, pattern-matching on either C or D will cause an (Eq a)
> constraint to be required.

> However, GHC does *not* follow this convention when you write the
> data type decl in GADT style syntax:

>         data T a where
>           C :: Eq a => a -> Int -> T a
>           D :: T a

> Here, (a) you can be selective; in this case, C has the context but D
> does not. And (b) GHC treats the constraints sensibly:
>         - (Eq a) is *required* when C is used to construct a value
>         - (Eq a) is *provided* when C is used in pattern matching

> In short, in GADT-style syntax, GHC is free to do as it pleases, so
> it does the "right" thing.  In this case, then, you can avoid the H98
> "bug" by using GADT-style syntax.

> All of this is documented in the user manual.  If it's not clear,
> please help me improve it.