Actually, after your email I played around with variants and it turns out they don't work at all. I'll rework the variants. The records work beatifully though.

> First, decomp is like lookup to me, so I expected the Right
> constructor for when that lookup succeeds.

Agreed!

> Second, (Label :: Label "x") is a pretty unacceptable way to write
> what is just x in other languages. One idea would be to
> follow <http://hackage.haskell.org/package/HList-0.3.0.1/docs/Data-HList-Labelable.html>
> and create values/labels which will do something like:
>
>  x .=. 1 -- call to inj perhaps?
>
>  v ^? x  -- call to decomp
>
> Another idea is to make `x stand for (Label :: Label "x"),
> much like 'x and ''x in template haskell. Trying out
> a good syntax by using a quasiquoter or preprocessor
> before getting something into ghc is probably worth doing.
> One example that has not been as useful as originally
> thought it would be is:
> <http://hackage.haskell.org/package/HList-0.3.0.1/docs/Data-HList-RecordPuns.html>

Yes, that is currently the most painful bit of the syntax. It should be possible to adopt HList labelable. I would like a small syntactic extension that allows 'x for (Label :: Label "x") indeed. I'll probably hack this up later.

> On a somewhat related note, would your strategy of
> having sorted labels give better compile times for
> for code which uses records that are a bit larger
> than a toy example:
> <http://code.haskell.org/~aavogt/xmonad-hlist/>

Depends, as far as I understand HList record sometimes require searching for a permutation of l such that l~l' which seems expensive to me. This is not necessary if we keep the row sorted. For projections and decompositions the performance is (theoretically) the same: linear searching in a list (sorted or unsorted list)  is O(n).

The real benefit of keeping the row sorted is that { x = 0 , y = 0 } and { y = 0, x = 0 } have the same type. When a row is not sorted, as in HList, then if we for example have an instance Eq for a row (because all elements support Eq) then for using (==) both arguments would have to the same order in the row or we need a manual call to a permutation function. When keeping the row ordered, this is not necessary. The same kind of problem occurs when we fix the type of a function to a specific row:
( using whishful syntax )
f :: Rec [ x = Int , y = Int ] -> Int

If the row is not ordered, then f { y = 0 , x = 0 } will not typecheck and will require a manual call to permute the row.

Cheers!

Atze

>
>
>
> Regards,
> Adam Vogt
>
> On Tue, Nov 26, 2013 at 12:09 PM, Atze van der Ploeg <atzeus@gmail.com> wrote:
> > Dear all,
> >
> > Extensible records have been a long outstanding feature request for GHC.
> > Using the new closed type families and type literals, it is actually
> > possible to implement Daan Leijen's "`Extensible Records with Scoped Labels"
> > system as a library.
> >
> > I have implemented this library at https://github.com/atzeus/openrec (the
> > documentation is at http://homepages.cwi.nl/~ploeg/openrecvardocs/). The
> > only thing it requires is the availability of a closed type family that
> > compares two type level symbols (i.e. the ordering on strings):
> > type family (m :: Symbol) <=.? (n :: Symbol) :: Bool
> > patches to GHC that implement this built-in closed type family are also at
> > the github site.
> >
> > I would like to generate some discussion about:
> > * Is this the interface we would like for open records and variants?
> > * Would it be worthwhile to invest in syntactic sugar for open record
> > operations?
> > * Any comments on the interface and or its implementation?
> >
> > Cheers!
> >
> > Atze van der Ploeg
> >
> > _______________________________________________
> > Haskell-Cafe mailing list
> > Haskell-Cafe@haskell.org
> > http://www.haskell.org/mailman/listinfo/haskell-cafe
> >

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe