Arc Forumnew | comments | leaders | submitlogin
Request for Design Discussion: Arc-F ssyntaxes
3 points by almkglor 5604 days ago | 7 comments
The primary design goal for Arc-F is to allow interoperability of diverse libraries written by diverse programmers.

In my opinion, this precludes the use of a global readtable or a global ssyntaxer. Instead, a readtable/ssyntaxer should be definable on a per-project (or similar finer grain) basis.

Since Arc-F already provides a "contexter" object which is typically used on a per-file basis, I propose that any ssyntax redefining scheme use the "contexter" object. In fact, there are currently two reserved unpackaged symbols for contexting which I intend to use for ssyntaxes, 'import-ssyntax and 'interface-ssyntax.

Here are some characteristics which I think are important for any proposed ssyntax redefinition scheme (in order of decreasing importance, as I see it):

1. Non-global.

This prevents many unexpected problems with multiple libraries defining and using the same ssyntax for different meanings. If ssyntaxes can be limited to the scope of a project, or the scope of some other finer grain other than the entire Arc process, then such problems are not going to exist at all.

This is assured by using contexts.

2. Attached to package interfaces.

If in the future, PG decides to add an ssyntax in Arc3, we should be able to release a similar ssyntax for Arc4F (or later). However, it would be preferable for a program written for Arc3F to run on Arc4F (or later) without excessive modification, even if that program uses that ssyntax for non-ssyntax purposes.

For example, if tomorrow PG releases the `$foo` syntax to mean `(fn () foo)`, then an existing Arc3F program which uses the variable `$-earned` should still run on Arc4F, preferably without modification. End of example

This means that unless the program is specifically modified to use <arc>v4 instead of <arc>v3, it will not use the new $foo syntax. Only if a program specifically uses <arc>v4 will the new ssyntax be enabled.

3. Allow mixing of ssyntax from different libraries.

This is a very complex requirement, but this is very desirable and a ssyntax redefinition scheme which can achieve this requirement would give us even better interoperability.

Specifically, it would allow two (or more) different libraries that define non-overlapping ssyntaxes to be joined into a single interface by a third library, which would effectively be our "merging library".

For example, consider two libraries, andf-ss and orf-ss.

andf-ss defines the ssyntax `foo&&bar` as `(andf foo bar)`, while orf-ss defines the ssyntax `foo//bar` as `(orf foo bar)`.

Since both libraries use nonconflicting definitions, it would be desirable to be able to use both their ssyntaxes in a single application or library. Preferably, this should be doable in a short declaration, but it is acceptable to have to do this in a separate small "merging" library. End of example

3.1. Allow user-controlled declarations of relative precedence.

Returning to the previous example, even if we were to allow both new ssyntaxes to be used in the same application or library, it would be preferable to define their precedence to each other. This is because something like `foo&&bar//quux` would be otherwise ambiguous: does this mean `(orf (andf foo bar) quux)` or `(andf foo (orf bar quux))`?

This is the problem of precedence. It would be desirable for the user to be able to specify the precedence.

For the sake of discussion, let us also suppose that the writer of andf-ss is different from the writer of orf-ss, and that neither of them are willing to communicate with the other (just because, okay? LOL). End of Example

This means that a number-based precedence level may not be fully general, because the chosen numbers may not reflect what the end user may actually desire; further, getting the library writers to change their libraries may not be possible.

My proposal is that instead of using numeric precedence, we instead use a system of defining an ssyntax as being "higher" or "lower" another ssyntax. This allows for generality, but it would have the cost of making the precedence description longer. It may be acceptable to use a system where the higher and lower declarations are optional, in which case the user is "in his or her own", so to speak, and using multiple different ssyntaxes in a single symbol will be ambiguous (preferably throwing an error).



4 points by cchooper 5604 days ago | link

I totally agree that these things should be non-global. In my opinion the grain should be as fine as possible, so that you can use embedded DSLs whenever possible (like w/html or LINQ).

But it's worth pointing out that there are work-arounds to a global read table. In CL, for example:

  #.(install-new-syntax)

  blah...blah...blah...code

  #.(revert-to-old-syntax)
Regarding the use of interfaces to specify the language version: have you ever considered making the language itself a package? The base language could just be a very primitive one, and then everything else could go in a package. It would mean there is no real distinction between 'the language' and 'a package that defines its own syntax'. That feels quite Lispy to me.

-----

4 points by almkglor 5604 days ago | link

> Regarding the use of interfaces to specify the language version: have you ever considered making the language itself a package?

Uh, yes. That's how it's already done. That's why arc.arc is itself a package. In fact, when you say '(in-package foo), the only thing that the contexter will add to the newly-defined package foo are the Arc axioms: '<axiom>fn, '<axiom>if, '<axiom>quote etc. That's why you have to declare '(using <arc>v3) for each package - because otherwise the language doesn't actually import arc.arc

Edit: Ultimately (and this is another disagreement with PG) arc.arc is a library, not part of the language. It's like the libc in this respect. Because you see, we already had the 150-year language 50 years ago.

Edit2: as an aside, the package system even lets you export a binding for 'quote, 'quasiquote, etc. In fact, once we have defined a formal method to add ssyntaxes, it would be possible to add support for `(let foo# ,foo (something foo#)) in a library, not in the language. How? By defining an export for 'quasiquote, say (interface v1 <auto-uniq>quasiquote), where <auto-uniq>quasiquote is a macro that expands to <axiom>quasiquote

> In my opinion the grain should be as fine as possible, so that you can use embedded DSLs whenever possible (like w/html or LINQ).

!!

Hmm. Expression level granularity? Hmm.

-----

1 point by cchooper 5604 days ago | link

Ah yes, I figured because it wasn't in lib/ then it wouldn't be a package.

-----

1 point by almkglor 5604 days ago | link

Hehe. It's in the arc-f/ directory more for historical reasons than because it's not a library, and also because it's treated specially (it gets precompiled to Scheme bytecode, for example).

-----

3 points by almkglor 5604 days ago | link

> But it's worth pointing out that there are work-arounds to a global read table.

How does it handle multiple threads loading code at the same time?

-----

5 points by cchooper 5604 days ago | link

Threads? In Common Lisp? You must be joking!

-----

3 points by almkglor 5604 days ago | link

!!

LOL! ^^ ^_^ >.< ^_^ ROFL!

As an aside, arc-f currently handles loading by wrapping a sizeable bit of 'load in 'atomic, but this is unnecessary, I think, if I use a thread-local for current-load-file* and friends.

-----