Short answer: There is no good way of doing what you want.
This is actually one of my biggest annoyances with haskell (right up there with disallowing infinite types). They are many techniques that work better or worse depending on the application, but non are very satisfactory IMO.
Your typeclass solution(or some variant of) is pretty much your best option. If you're careful about how you define your datatype and classes you can avoid the type families and such, but the whole point is to not have to be careful.
If your types are fixed (which is usually true as long as you're not using existentials) you might be able to get away with using -XDisambiguateFieldRecords

If you want anything better you're probably going to have to use some form of preprocessor (like OHaskell).


Supposedly OCaml has an OO feature that does this but I haven't tried it out.

I would suspect that the reason why haskell doesn't provide duck typeing on record fields is that analisys for optimizations is much more complicated (as it currently stands, records are nothing but sugar on top of algeraic datatypes).
You can end up with all sorts of weird things with duck typeing on record fields, like unnamed datatypes. For example:

(using class constraint style to inidicate a record field restriction for lack of a better syntax)
setPosition :: (position a) =>Vector -> a -> a
setPosition v x = x { position = v }

translate :: (position a) =>Vector -> a -> a
translate v x = x { position = v + (position x) }

getPosition :: (position a) => a -> Vector
getPosition x = position x

result :: Vector
result = getPosition $ translate someVector $ setPosition someOtherVector

The type variable 'a' in these functions is never fixed to a specific type, and it actually doesn't need to be. The compiler would just have to invent a suitable one (a type with only the field 'position' of type Vector).

Maybe someday haskell will finially implement good, clean, duck typeable, record functionality. I will be waiting...

- Job


On Fri, Sep 25, 2009 at 9:54 AM, Peter Verswyvelen <bugfact@gmail.com> wrote:
Haskell's records are a bit annoying, and type-classes often group together too many methods, which means you make early decisions about future unknown requirements, and IMO you always get it wrong :-)

After having read an email in the cafe about the Noop language & Self language, I realized that what I really would like to have is "strong duck typing" on records (or is it called structural subtyping? or prototype-based-objects? or something like that)

For example (silly example full of inaccuracies, but you get the picture):

class HasPosition a where
  position :: a -> Point
  withPosition :: Point -> a -> a

class HasVelocity a where
  velocity :: a -> Vector
  withVelocity :: Vector -> a -> a

which we really should write as

field HasPosition :: Point
field HasVelocity :: Vector


And then 

record IsKinetic :: HasPosition HasVelocity

suppose we write a function like

kineticEulerStep dt k = withPosition (position k .+^ dt *^ velocity k) k

kineticEulerStep will work on any type a that HasPosition and HasVelocity, and would get inferred signature

kineticEulerStep :: IsKinetic a => Float -> a -> a

which is identical to

kineticEulerStep :: (HasPosition a, HasVelocity a) => Float -> a -> a

So basically kineticEulerStep accepts anything that HasPosition and HasVelocity, whatever it is.

So if it walks like a duck and ..., then it is a duck, but statically known...

We could also do

field HasForce :: Vector
field HasMass :: Float

record IsDynamic :: IsKinetic HasForce HasMass

acceleration d = force d ^/ mass d
withAcceleration a d = withForce (a ^* mass d) d

dynamicEulerStep dt d = withVelocity (velocity d ^+^ dt *^ acceleration d)

Of course you would also need type families to be really correct since Vector, Point, etc should also be parametrized.

And really kineticEulerStep might also work on something that HasVelocity and HasAcceleration (since the code in dynamicEulerStep is almost the same as kineticEulerStep), so better abstraction might be needed.

I'm not sure what kind of overhead a system like this would have in Haskell, since I suspect the many dictionaries are often not optimized away.

I think for Haskell prime, something like this was suggested, but is was rejected?

Languages like OCaml and haXe also provide a similar feature?

I would like to collect ways of doing this in Haskell, without boilerplate, and preferably without runtime overhead.

I remember reading OOHaskell a while time ago, and while I didn't understand a lot of it, I recall it also was doing a similar thing, but since the compiler lacks native support, the error messages you get most likely make it impossible to figure out what is going wrong. I think Grapefruit's Records, HList, Data.Accessor, etc.. might also work.

Any guidelines and comments regarding "strong duck typing"/"structural subtyping" are very welcome, since the lack of this is the only reason why I would prefer a dynamic language over a static one.

Thanks a lot,
Peter Verswyvelen


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