Re: (set! (@@ MOD NAME) EXP) considered harmful

[Neil Jerram]

Andy Wingo <address@hidden> writes:

> Hi Mark,
>
> On Wed 02 Sep 2009 18:17, Mark H Weaver <address@hidden> writes:
>
>> The ability to set! arbitrary module top-level variables from outside
>> the module, using the syntax (set! (@@ MOD NAME) EXP), destroys our
>> ability to several important optimizations.
>>
>> As long as such ability exists, we must pessimistically assume that
>> any module top-level variable might change at any time, which means,
>> for example, that we cannot inline top-level procedure applications
>> from within the module itself.  This could be a HUGE efficiency win in
>> the common case where such top-level procedures are never set! from
>> within the module.
>
> I disagree, heartily.
>
> Guile is a dynamic system, one where you can get into modules and
> redefine functions. All you have to do is (set-current-module!
> (resolve-module '(foo))) and (define bar #f) to do the equivalent of
> (set! (@@ (foo) bar) #f). It's just like that.
>
> And we should be proud of that, IMO! Your program might be a reified gem
> once you're finished with it, but it normally doesn't arrive there
> without much debugging.
>
> Furthermore, (set! (@ ..) ..) is a /consequence/ of hygiene. A hygienic
> macro that expands out to a set! of a variable in its context will
> compile to a (set! (@@ ...) ...). You can always get at the
> module-private vars, even with hygienic systems -- see the
> eval-in-module hacks from ikarus if you don't believe me there.
>
> Now, regarding optimizations. We can optimize whatever we like ;)
>
> More seriously, there are many possibilities there. One is to define
> your modules in lexical forms, as Chez does it -- so private functions
> will always be inlined. See the end of section 2.4 on
> http://www.scheme.com/csug7/use.html.
>
> Another is to always present module-private bindings as candidates for
> inlining.
>
> Another is to allow the user to mark certain procedures as inlinable.
> That will certainly be the case with `map', for example.
>
> Another is to have an optimization knob you can crank up, understanding
> that higher levels of optimization prevent some runtime modifiability.
>
> In short, @ and @@ are lovely and useful, and we should be happy with
> them. (set! (@@ ...) ...) is not an expected operation from user code,
> and may reasonably be discounted even at low optimization levels. I
> don't see the problem :)

I agree with all that.  Even as we do more static-feeling things like
compilation, we'd like to keep all the dynamic possibilities too.

Regarding optimization, maybe another possibility would be to have
detailed tracking of the underlying definitions that each piece of
optimized+compiled code depends on, and to recompile+reoptimize
whenever one of the underlying definitions changes.

Or is that just cloud cuckoo land?

(Ludovic recently suggested an optimization for constant additions, and
I assumed that would require this kind of rollback, if the definition of
`+' was later changed.)

      Neil