EmptyDataDecls and type context

[Paczesiowa]

the following code fails to parse correctly or typecheck under 6.12.1

{-# LANGUAGE EmptyDataDecls #-}
data Show a => Foo a

error:

    No context is allowed on a GADT-style data declaration
    (You can put a context on each contructor, though.)

it works with 6.10.

this is the reason of HList failure on 6.12.

[simonpj]

Why do you want this? It's utterly useless. Contexts on data declarations are a mis-feature of Haskell, and are especially useless when there are no constructors. Can't you just delete it?

Simon

[Paczesiowa]

I don't have anything to do with this, maybe you should ask Oleg about such definition for labels. I just wanted to fix HList (it fails to compile with 6.12) and I've discovered, that 6.12 works different than 6.10 and also reports really misleading error message. So, I'm not interested in making this usage work like in 6.10, but there should be a better error message and some info in docs, that the behavior changed.

Oleg's code:  http://old-darcs.well-typed.com/HList/Data/HList/Label2.hs

[guest]

I believe contexts on data declarations might not be totally useless, even if there are no constructors. The HList code in question had the following declaration

data HNat x => Label x ns desc -- labels are exclusively type-level entities

The context HNat is asserted not because we want to use any of HNat's methods. Rather, the constraint HNat was asserted as a kind predicate -- to show to the user that the type variable x is of a kind natural numbers, so to speak. Therefore, the type (Label HZero Foo Foo) is the correct instance of label. On the other hand, the type (Label Bool Foo Foo) is not a correct instance and will be flagged as such by the compiler.

Oleg

[simonpj]

What makes you think that GHC will reject the type (Label Bool Foo Foo)? I don't think it will.

The only effect of a context on a data decl is that a use of the constructor, either in pattern matching or in an expression, gives rise to that constraint. That's all.

So the constraint is no more than a comment. I suggest you use a comment!

I'm really want a proper kind system, though. It's silly that we can't give the kind of Label. But contexts are not the way to do it.

Simon

[simonpj]

OK well I still think it's stupid, but I accept that it's a needless change in behaviour from 6.10, so I'll make 6.12 accept the declaration just as 6.10 did. That way you won't have to change HList.

Mon Feb  8 16:37:51 GMT 2010  simonpj@microsoft.com                      
  * Fix Trac #3850
  
  This patch simply avoids a needless difference in behaviour from 
  6.10, and one that happens to affect HList. See Note [Stupid theta].

Please merge to 6.12 branch.

Simon

[igloo]

Merged.