On Nov 18, 3:43 pm, Simon Peyton-Jones
But anyway, the original TDNR thing is perfectly well defined. It might occasionally be surprising. But that doesn't stop the OO folk from loving it.
Not only OO folks but I think anybody who works with many records having similar selectors. In the past I had hopes for Daan Leijen's "Extensible records with scoped labels" [1] to have some impact on Haskell's record system but that didn't happen. Module-scoped selectors + the fact that you need one file per module are just not very convenient. Type classes only lead to more boilerplate in this area. If I compare for instance: data D1 = D1 { d1_p :: Int } data D2 = D2 { d2_p :: Int } class P a where p :: a -> Int withP :: a -> Int -> a instance P D1 where p = d1_p withP p x = p { d1_p = x } instance P D2 where p = d2_p withP p x = p { d2_p = x } with the TDNR solution: data D1 = D1 { p :: Int } data D2 = D2 { p :: Int } I have a clear preference. I think the important issue is not so much that one can access p with OO-like notation, but that the scope of p is effectively restricted. And very often one is not done with just one type class but instead one writes many for the different record fields. It was nice to see that DDC also has something similar to TDNR [2]. I would be happy if someone corrects me and points out an easy solution for this problem, but I fail to see one. Cheers, Levi --- [1] http://legacy.cs.uu.nl/daan/pubs.html#scopedlabels [2] http://www.haskell.org/haskellwiki/DDC/FieldProjections