SWIG, Chicken and TinyCLOS

Note: this is a fairly technical post; if you have no interest in FFI's, you may still find the @ TinyCLOS macro useful.

When dealing with large C libraries, SWIG (the wrapper generator) can be a mixed blessing. On the one hand, it's a pleasure to work with wrapped C libraries from a dynamic language; on the other hand, generating the right wrappers can require significant time and effort, often with nothing to show for the plumbing work until the interface is complete.

In my case, the accessors and modifiers for C structures have been the most painful, initially at least. The library was full of complex, nested records of the following sort:

struct msg {
  int op;
  struct {
    char *name;
    union {
      int start;
      char *dst;
    } args;
  } req;

SWIG treats struct msg and its innards as separate objects; in Chicken, if you want to get to msg.args.start, you have to type a monstrosity like (msg-req-args-start-get (msg-req-args-get (msg-req-get msg))) (with bonus points for longer identifiers or deeper structures, of course).

The verbosity grows quadratically, and after a short while I started investigating the TinyCLOS mapping option. When invoked with the -proxy option, SWIG generates wrapper classes for C structures. This is enormously helpful: the previous incantation becomes (slot-ref (slot-ref (slot-ref msg 'req) 'args) 'start), which in real cases is a lot shorter due to, um, linear verbosity. To modify fields, you use slot-set!.

This was still too much typing, so I introduced the @ macro with which you can simply write (@ msg req args start), or (@ msg req name = "flush"):

 (define-syntax @
  (syntax-rules (=)
    ((_ o) o)
    ((_ o slot = v) (slot-set! o 'slot v))
    ((_ o slot . slots) (@ (slot-ref o 'slot) . slots))))

Finally, relief. In retrospect, I find it hard to believe that nobody solved this problem before; maybe there's some "standard" macro for this purpose, but I haven't found it.

This isn't the end of the saga, though. As soon as I moved back from experimenting to the actual library, I was hailed by a salvo of errors indicating that SWIG/TinyCLOS has probably never been used for large applications. Specifically, SWIG translates a composite structure name such as my_class into either <my-class> or <my_class> depending on the context. Presumably, SWIG/TinyCLOS was only tested for the traditional OOP toy examples (Shape, Pos etc.)

Fortunately this is easily fixed with perl -ne 'if (/<.*>/) { s/_/-/g; print } else { print }'. Older versions of SWIG also add an unnecessary (and harmful) (declare (uses tinyclos)) to the Scheme wrappers, but this is also easily excised.

The great news is that after all these machinations, as well as others not described here (involving callbacks and typemaps), SWIG/TinyCLOS seems to work without a hitch. I have had no problems using a large C library from a long-running Chicken program — writing the code was a lot of fun (compared to the SWIG saga), and, more importantly, there where no crashes. Has anybody else played with SWIG / Chicken / TinyCLOS?

Scheme object systems: POS

I'm no OOP fan (much less a fan of single-dispatch OOP), but sometimes I miss the implicit lexical scope that single-dispatch provides for methods. Take something as simple as

class Rect {
  int top, left, bottom, right;
  int area () const {
    return (top - bottom) * (right - left);

Most Scheme object systems (see for example the Chicken OOP section) turn the area() method body into something tedious along the lines of

(* (- (slot-ref self 'top) (slot-ref self 'bottom))
   (- (slot-ref self 'right) (slot-ref self 'left)))

or, with objects implemented as closures,

(* (- (self 'top) (self 'bottom) ...) ...)

Until recently, I thought that short of codewalker-based macros, nothing could restore the terseness of single-dispatch methods.

Well, I've discovered Blake McBride's POS (portable object system). With POS, you can write

(define-class Rect (ivars top left bottom right)
    (set-top! (self val) (set! top val)) ...
    (get-area (self) (* (- top bottom) (- right left)))))

POS is a set of pure R5RS macros, and correctly interacts with other syntax-rules macros (e.g. macros can appear within method bodies). The trick is not only to represent the objects as closures, but to also expand method bodies inside the closure:

;; not the actual POS expansion -- just an illustration
(define (make-rect)
  (let ((top #f) (left #f) (bottom #f) (right #f))
    (define self
      (lambda (meth-name . args)
        (case meth-name
           (apply (lambda (self val) (set! top val))
                  (cons self args))) ...
           (apply (lambda (self)
                    (* (- top bottom) (- right left))
                  (cons self args)))))))

This way, methods can access instance variables as simple literals. Each object is a dispatch function that closes over those variables.

POS is very useful, and I plan to add default getters and setters, as well as a way to convert between the closure representation and a-lists. This should help with persistence, among other things.

POS has a couple of extra features (inheritance, access to the parent object, class methods) but really is a light-weight system. The major downside is that methods (and instance variables) can no longer be added dynamically, since it's impossible to inject code (or data) into a closure.

Update: see the comments for yet another way of simulating an implicit "this" argument.