POPL

[Stephen Weeks]

Here's some stuff Henry sent me about POPL.

> MLton was discussed several times in POPL itself.  During Reppy's  talk,  and
> during  Wang's.   Off-line  I was talking to Reppy a bit and at some point in
> connection with something I mentioned how you have to decide when to write  a
> new  intermediate language.  He said that this wasn't such a problem any more
> for him (in Moby) because he uses your property lists now.

Cool.  I'm really thinking it would be nice to have support for them in the
language so that the compiler can do a better job (space wise) of implementing
them. 

> Wang's talk was also interesting: just using region stuff  to  prove  that  a
> copying  GC  is  safe.   He  produces  a  custom-GC  for  a program, which is
> certainly very much a MLton-style thing to do.  Interestingly he  found  that
> the  time  to copy a small object in his is faster than in the standard MLton
> one.  I suspect, and he agreed, that this is probably caused by the fact that
> it  is  straight-line  code  to  copy an object instead of a loop.  Still, an
> amusing number.  Mind you, all his test programs were very VERY small,  which
> means  not  many  kinds  of  objects.  If you have lots of different kinds of
> objects than you probably start ending up with I-cache problems with all  the
> dinky copy routines.

He did his custom-GC stuff within MLton, right?  So it was a very apples to
apples comparison.

How many different kinds of objects do you think there need to be before the
icache becomes a problem?  1,000? 10,000? 100,000?

I doubt there's more than a few thousand, even in a self compile.