[ENH][VM] Improved code generation (hopefully ;)
Andreas Raab
andreas.raab at gmx.de
Wed Jul 9 14:51:03 UTC 2003
Hi Scott,
> > We discussed these changes but nobody (including me) was
> > able to recreate the improvement you reported.
>
> It should have been straightforward with the new methodcache. AFAIK,
> nobody else tested it. (Or at least, I saw no discussion other than
> plans to drop it because it was considered too invasive.)
No, actually I tested it but given my newly found knowledge about the
effects of branch-prediction it may well be that my tests were *way* off.
Considering that #benchFib always sends the same selector your scheme would
bring no improvement if the selector is in the first slot (aha!) and I think
this may have been what I based my evaluation upon.
> Did the root-table-overflow patch go in? (That one was where if the
> root table was about to overflow, immediately microGC, then tenure.)
Not sure - Tim?
> > Huh? I don't quite understand what you're saying here.
> > Where was that insn? What did it do? How was it related
> > to the mcache?
>
> It was in commonLookup, here's the code snippet out of interp.c
[snipped away]
Yes, if the miss-rate on the first branch is really around 66% that should
help. I'll give it another try.
> -- This comparison above is apples-oranges, but I can't easily go
> -- through my earlier posts to identify if I have a better one.
Fortunately I can (and I will ;-) my personal Squeak archives range _way_
back.
Cheers,
- Andreas
More information about the Squeak-dev
mailing list
|