Thanks for the reference. Espresso doesn't support recursive records. Does system in your paper allow for those? Is there anything particularly tricky about supporting them?
Hi, author here, yes your paper was a huge inspiration and I really should add a references section in the readme. The row types in Expresso use very simple lacks constraints, similar to the old Hugs TREX system, and don't permit duplicate/scoped labels. I felt this was also a pretty good sweet spot.
Thanks :-)
TREX (by Mark Jones and Benedict Gaster) is great. The beauty is that a lacks constraints can get translated to fields offsets at runtime making it super efficient!
E.g. a type like:
foo :: r/x => { x :: int | r } -> int
foo r = r.x
Gets translated at runtime to a function where the lacks constraint r/x becomes an actual offset parameter.
I’ve toyed with the idea of doing a language like this. With a small difference, records would be the only way to express values and bindings.
There would be no need to think of currying and partial application of functions as it would naturally follow from partially evaluating incomplete record inputs.
Another (nice?) property would be that a value always comes with a name binding. This no need to encode semantics of values purely as positions in lists and tuples.
So a question then to an implementer, is this something you considered? Is feasible?
I've felt for a while that something like this would be superior to the positional way of working, in most cases. But I don't know whether this would result in an elegant system in reality, or not. What if positional were just sugar for records indexed by natural numbers?
There are some precedence in SQL. Which seems to work out ok for the most part. But then in SQL functions are not relations, and do use positional parameters, so not entirely pure either.
Som help from type/name inference might help in wiring things togetger.
Not much. But basically a bit inspired by Cat (a stack based, typed, concatenative language), and a thought that stacks could be replaced by first class environments and rewriting, so you could have a concatenative language with partial evaluation. Instead of the usual ordered parameter lists, and/or curried functions with partial application, functions take environments and produce environments. A program like "{a=1}{b=2}{c=a+b}" would evaluate to "{c=3}". While "{b=2}{c=a+b}" evaluates to "{c=a+2}"
It seems like there's just a global namespace of labels. This reminds me of the situation in Javascript before symbols were introduced. I'm curious why this choice was made?
Labels in Expresso are just type-level symbols. Label names can be re-used over and over again in different records and variants. This works fine because a name is not tired to any particular nominal type (like a constructor in Haskell or Java).
Yes because duplicate labels are not allowed within the scope of a single record (you can specify which one overrides the other, but perhaps we also need a convenient rename syntax). This point was that different records can happily contain the same field names. The situation is the same as SQL.
21 comments
[ 2.1 ms ] story [ 64.4 ms ] threadFor those interested in more of the beautiful theory on extensible row types, this project seems to be based on an earlier paper I wrote on scoped labels: https://www.microsoft.com/en-us/research/wp-content/uploads/...
E.g. a type like:
foo :: r/x => { x :: int | r } -> int
foo r = r.x
Gets translated at runtime to a function where the lacks constraint r/x becomes an actual offset parameter.
There would be no need to think of currying and partial application of functions as it would naturally follow from partially evaluating incomplete record inputs.
Another (nice?) property would be that a value always comes with a name binding. This no need to encode semantics of values purely as positions in lists and tuples.
So a question then to an implementer, is this something you considered? Is feasible?
Som help from type/name inference might help in wiring things togetger.
For example, one can already introduce all the fields of a record as local bindings, e.g.
let {..} = import "List.x"
or just a subset using, e.g.
let {reverse} = import "List.x"
Similarly for function argument bindings, e.g.
f {x, y} = x - y
The above is just syntactic sugar for:
f r = r.x - r.y
Such named arguments of course prevent arguments with the same type being passed in the wrong order, e.g.
f {x=2, y=1}
There is still some minor work to be done to better support inline type annotations on such patterns to make them more usable.
Your idea of unifying bindings with records is interesting and not something I have considered.
Couldn’t find a syntax that made sense though...
(http://www.nuprl.org/documents/Constable/Automath-35years.pd...)
RECENT RESULTS IN TYPE THEORY AND THEIR RELATIONSHIP TO AUTOMATH, ROBERT L. CONSTABLE