Re: early termination for `map'

[Andy Wingo <wingo@pobox.com>]

On Thu 05 May 2011 18:26, ludo@gnu.org (Ludovic Courtès) writes:

> Andy Wingo <wingo@pobox.com> writes:
>
>> If you call `map' or `for-each' with more than one list, our versions of
>> these operators will detect if the lists are of unequal length, and
>> throw an error in that case.
>>
>> However, SRFI-1 has long provided an extension to this, to allow for
>> early termination when any of the lists runs out.  R6RS adopted this,
>> and it looks like R7RS will ratify that.
>>
>> So perhaps it's time for us to change as well.
>
> To change the default ‘map’ & ‘for-each’ to do like SRFI-1’s, right?

Yeah, that was the proposal; but the argument is a bit weaker, now that
I found that the R6RS did not go with this change.  So I don't really
know.

The reason I was thinking of doing this is because it turns out to help
performance to have map in scheme, at this point; or at least not hurt
it, and things will get better when we grow an optimizer.

So I implemented map in Scheme, with circularity detection and all, only
to find strange errors in the ecmascript compiler (!).  Turns out those
errors happened when loading goops, because it tried to extend a
primitive generic, but map wasn't a primitive any more, and instead of
failing nicely it corrupted memory.

>> This would also allow us to get rid of the hack in srfi-1.c in which,
>> when and if GOOPS gets loaded, srfi-1 extends the `map' and `for-each'
>> primitive generics with its own early-termination code, which in effect
>> gives early termination to every `map' user, regardless of whether that
>> module has imported srfi-1 or goops.  Sometimes I think that Mikael put
>> the Oops in Goops for a reason ;-)
>
> Heh.  :-)
>
> Code not using GOOPS can tell the difference between the default and
> SRFI-1 versions, and it can arrange to explicitly use one or the other.
> So changing the default versions would be an incompatibility—arguably
> not an important one, but still maybe enough to wait until 2.2?

Yeah, perhaps you are right, and the srfi-1 / goops interaction should
probably get some other resolution -- like merge-generics on import, or
something.

Thanks for your thoughts,

Andy
-- 
http://wingolog.org/