On 21/01/15 21:51, Simon Marlow wrote:
Adam, do you have any measurements for how much code gets generated for a record declaration with ORF, compared with a record declaration in GHC right now? That's one thing that has been a nagging worry for me with ORF, but I just don't have any idea if it's a real problem or not.
Yes, that was something that was a bit unsatisfying about the original implementation, though unfortunately I don't have hard numbers comparing the relative code sizes. But Simon PJ and I have realised that we can be much more efficient: the only things that need to be generated for record declarations are selector functions (as at present) and updater functions (one per field, per type). Everything else (typeclass dfuns, type family axioms) can be made up on-the-fly in the typechecker. So I don't think it will make a practical difference.
Under Nikita's proposal, zero code is generated for a record "declaration" (since there isn't one), and the lenses are tiny expressions too. There's some boilerplate in the library, but it's generated once and for all, and isn't that huge anyway. The lightweightness of it from a code-size standpoint is very attractive.
Agreed. I'm coming to see how much of a virtue it is to be lightweight! I'm a bit worried, however, that if we want newtype T = MkT {| foo :: Int |} x = view [l|foo|] (MkT 42) -- whatever our syntax for [l|...|] is to be well-typed, we need an instance for FieldOwner "foo" Int T to be generated somewhere (perhaps via GND), so I suspect the code generation cost for non-anonymous overloaded records is about the same as with ORF. Nikita's proposal would let you choose whether to pay that cost at declaration time, which has advantages and disadvantages, of course. Adam -- Adam Gundry, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/