Arc Forumnew | comments | leaders | submitlogin
2 points by shader 5822 days ago | link | parent

It seems to me that, however we do it, it would probably be slow. Maybe, as the source is being read in, the compiler/interpreter could somehow flag each symbol as bound or unbound? Then you could have your ofns just inspect all of the symbols below them, and check whether they're bound or not.

About nested ofns, how would you expect them to work? That should be the real test of what method to use in determining which ofns bind which vars. Obviously they're bound in alphabetical order. How about having the outer ofn bind as many vars as it has args, and leaving the rest unbound for the next layer? This would make it impossible to tell which ofn would bind which var at compile time (unless you know the arity of the outer function based on it's environment), but it would (I think) make the most sense. It does leave room for confusing circumstances though ;)

What would happen if there weren't enough arguments to fill in all of the vars? Return a fn? Even more fun!

If the ofns bind inwards, you would need to do your example backwards: [map [map (+ d c) b] a]

You could however try and bind outwards, so the inner ofn binds first, but I don't see how that would be a good idea.

What do you think? Any better ideas? Any glaring flaws with my idea? (I don't know lisp or arc that well)



1 point by Darmani 5821 days ago | link

Source-code traversals are O(n), and compile-speed is not that important below a certain point. And finding all the globals only needs to be done once.

At first I thought it was important to have this conventionalize-ofns function run inside every def and every mac, so that, if I wished to use ofns inside a macro-expansion, which symbols are arguments would not depend on the context. I then realized this would break for virtually any method of generating macro-expansions other than quasiquote, and that, since, under the curretn implementation, two pseudo-gensyms have the same alphabetical ordering as the order they were created in unless they're of different length, it would still be workable (just do a (w/uniq (a b c) ...) and I'll hardly notice the difference -- saves fewer characters overall, but I find it a little more readable; plus, these can be reused if the macro-expansion contains multiple ofns).

If we are to do ofns at runtime, they would be much more workable, but also more unpredictable. I've said once before that a runtime-list of all bound symbols would be desirable for other reasons, but, if I'm testing code in the REPL and set a to some value, I don't want half the functions I call breaking for some mysterious reason. Especially considering I haven't found a way to unbind variables.

-----

1 point by shader 5821 days ago | link

Well, I think that the ofns should probably be closures, at which point the binding of the variables only matters when they are defined, after that defining a out of context wouldn't overwrite the a in the ofn. If not, we could have the ofn replace each symbol with a gensym tagged with 1st, 2nd, 3rd, etc. Then it wouldn't matter what you did with the original name. The gensym replacement only happens, of course, if the var is unbound during definition.

Here's a question: how does anarki's [ _1 _2 _3 ] form work when nested?

-----