Thanks for all those comments! After mulling them I think I'll feel better if I can eliminate ssyntax in favor of infix operators. But there's two challenges to that:
I was trying to think of alternatives, thought "maybe a more complex symbol for one of the uses, like ..?", then wondered about a potential edge case. Really, I'm just thinking of the parsing algorithm---or, rather, lexing. If . was defined as the ssyntax is, would a..b expand into ((a) b)? Without spaces, it's fairly clear that certain arguments are "empty", since it could conceivably be tokenized as two .s. But a++b probably wouldn't tokenize to two +s. Suppose both . and .. were defined; how would a..b be resolved? Longest-operator-first?
f:g vs :syms
Could always go with another symbol for function composition. | comes to mind, but it's more like "reverse compose" at a Unix shell. On the other hand, as far as the function composition operator is concerned, I've seen mathematicians define it both ways---$(f \circ g)(x) = f(g(x))$ and $(f \circ g)(x) = g(f(x))$. No technical reason you couldn't use a pipe, just conventional.
Check out the details below, and give me your reactions. Is this too ugly to be worthwhile?
Excluding tests, this change reclaimed ~50 LoC. In all, this whole experiment has costed 225 LoC. I'm going to play with it for a bit before I decide to back it out.
---
Compromises:
1. In wart, unlike arc, you could mix unquote with ssyntax: car.,x, etc. This had to go.
2. You can no longer use ssyntax with operators: ++. used to expand to (++); now it's just a three-character op. Haskell's prefix syntax is now required to escape infix.
1. Periods can be part of operators, but dotted list syntax now uses ..., which is never an operator.
2. Period at end of symbol calls it. prn. is (prn), not (prn .)
3. Colon at start of symbol is part of the symbol. This was always true, for keyword args. It means I can't define :~ to support f:~g; it just didn't seem worth another special-case.
4. Bang at the end of a symbol is part of the symbol, for mac!, reverse!, etc.
5. Bang has another special-case. In keeping with left-associativity, prefix ops are always highest-precedence:
~odd?.n ; => (. (~ odd?) n)
However, ! at the start of a symbol is _lowest_ precedence:
1. Periods can be part of operators, but dotted list syntax now uses ..., which is never an operator.
Seems a worthwhile trade-off. Dotted lists are used infrequently enough, and an ellipsis does just as well as a single dot.
2. Period at end of symbol calls it. prn. is (prn), not (prn .)
Hm. So this is like a vestigial instance of ssyntax?
3. Colon at start of symbol is part of the symbol. This was always true, for keyword args. It means I can't define :~ to support f:~g; it just didn't seem worth another special-case.
Yeah, wouldn't want a special case on top of a special case! :)
4. Bang at the end of a symbol is part of the symbol, for mac!, reverse!, etc.
Have you considered a non-operator character for this use, to ditch the special case? I'm partial to mac=, reverse=, etc. I mean, since = isn't used for comparison anyway. And assuming that = is actually not an operator character. Did you ever decide if you wanted = to be an infix operator (and thus character)?
5. Bang has another special-case.
Whoa. Did I miss where this infix notation extended to prefix operators? Or does this work the same way ssyntax did? And if so, in what sense has ssyntax been removed? :)
Is this too ugly to be worthwhile?
Hm...parsing is getting too complicated for my tastes. But then, my taste is for parentheses. :P
Still, carving out special cases so ssyntax still "mostly works" isn't quite what I envision as a way to unify away ssyntax. Basically, is "traditional" (inasmuch as Arc establishes tradition) ssyntax prohibitively useful? Or can we do without some of its uses in the name of a more general infix notation without the complications of special symbol parsing?
Ok, I've tried to be ruthless and take the ssyntax stuff out. '!' is now always part of regular symbols, just like '?'. There's no prefix version like !a. And there's also no infix version, like f!sym.
It turns out this doesn't actually bring back any parens. !x becomes no.x or not.x. And in some situations I can replace a!b with a'b. (Maybe that's too clever.)
I've also dropped support for turning x. into (x). Not even traditional arc has that. Now x. turns into (x .).
Only remaining special-cases: '...', and ':' at start of sym is part of the sym.
Whoa. Did I miss where this infix notation extended to prefix operators?
Good point. This happened as part of the elimination of ssyntax, but I figured it was intuitive to extend left-associativity to prefix ops. However, now I see that:
(f a + b) => (f (+ a b))
but:
(- a + b) => (+ (- a) b)
Is that weird?
Thanks for the comments! This really seems like an improvement over my original idea.
Thanks for the comments! This really seems like an improvement over my original idea.
I'm glad you think so. I try to make my suggestions as nonprescriptive as possible, though (in full disclosure) I'm liable to lead you in circles back to prefix notation if you follow them too far. :P
It was that or lose <=, etc.
Oh, duh. Move along, nothing to see here!
Only remaining special-cases: '...', and ':' at start of sym is part of the sym.
I'm really okay with ..., because it doesn't feel like a "special case" as much as it does a built-in keyword; I wouldn't expect to be able to redefine fn or if, either. I don't really have an opinion on the :keyword symbols.
(f a + b) => (f (+ a b))
but:
(- a + b) => (+ (- a) b)
Is that weird?
Maybe, maybe not. It's not like every other language doesn't do mixfix with their "infix" notation. I just wasn't sure how it worked. Do you declare that certain operators are prefix? Or are they all potentially prefix, like
( mixfix a b ... ) --> ( ( mixfix a ) b ... )
where mixfix is any operator, a and b are any normal token, and ... is 0 or more tokens? Or something like that?
x-1.0
What's the intuitive way to parse this?
I'd say as subtraction of a float: (- x 1.0). If nothing else, I can't imagine a reason to do ((- x 1) 0).
Is it worth getting right, or should we just say, "don't use floats with infix"?
My gut reaction is that it's worth getting right, because programming languages shouldn't be ambiguous.
I notice that a lot of these problems seem to come from using the dot. Thinking back about ssyntax now, it occurs to me that the dot is probably the least-used among them, in Arc. If I were to guess from my own code, I'd rank their usage descending as ~, :, !, ., &. But hey, we can put numbers to that:
Mind you, it's been awhile, so I have no clue what all is in my personal ~/arc directory. Probably various experiments and output from things like sscontract (http://arclanguage.org/item?id=11179) and so on. All the same, the dot is low on the totem pole. I personally wouldn't be heartbroken to have to write (f x) instead of f.x, and you could reclaim the regular dotted list syntax. Would it be worthwhile to backtrack at this point and get !, :, and ~ functionality without worrying about .? There were some existing issues with ! and : (name collisions). ~ is prefix, but if you have a way of extending the infix notation for subtraction, surely it would apply to ~? Related thought: f ~ g could replace the f:~g you were worried about before.
Anyway, just some random thoughts off the top of my head. Do what you will with them.
Yeah you're right that using period as both an infix op and inside floats is kinda messy. I use it a lot more than you, so I'm still going through the five stages of grief in giving it up. In the meantime I've hacked together support for floats. Basically, the reader tries to greedily scan in a float everytime it encounters a sym-op boundary. Some increasingly weird examples:
Perhaps this is reasonable. We have a rule that's simple to explain, whose implications can be subtle to work out, but which programmers are expected to exercise taste in using. That could describe all of lisp.
"We have a rule that's simple to explain, whose implications can be subtle to work out, but which programmers are expected to exercise taste in using. That could describe all of lisp."
I don't think the syntax for numbers is very easy to explain. That's the weak link, IMO.
If it were me, I'd have no number literals, just a tool for translating number-like symbols into numbers. Of course, that approach would make arithmetic even less readable. :)
I think the next simplest option is to treat digits as a separate partition of characters like the partitions for infix and non-infix. Digits are sufficient to represent unsigned bigints with a simple syntax. Then most of the additional features of C's float syntax could be addressed by other operators:
-20.002e23
==>
(neg (20.@1002 * 10^23))
This hackish .@ operator, which simulates a decimal point, could be defined in Arc as follows:
(def dot-at (a b)
(while (<= 2 b)
(zap [/ _ 10] b))
(+ a (- b 1)))
You could avoid the need for this hack by treating . as a number character, but then you lose it as an operator.
"Basically, is "traditional" (inasmuch as Arc establishes tradition) ssyntax prohibitively useful?"
I don't think so. Nulan completely ditched Arc's ssyntax and only uses significant whitespace, ":" and ";". Yet, despite that, it's capable of getting rid of almost all parentheses.
Oh, by the way, ":" in Nulan has a completely different meaning from ":" in Arc. I just chose it because I thought it looked nice.
Actually, there's one major remaining question. This no longer works:
x-1.0
What's the intuitive way to parse this? Is it worth getting right, or should we just say, "don't use floats with infix"? Especially since wart recognizes whatever formats the underlying C recognizes:
-2.14e-3
It'll get complex fast to mix that with infix operators..
Did you ever decide if you wanted = to be an infix operator (and thus character)?
Yes, it's always been infix, so wart lost def= and function= when it gained infix ops. It was that or lose <=, etc. The question in my mind was more of style: even if assignment can be in infix, should I always use prefix?
