`just slow it`....

Raab, Andreas Andreas.Raab at disney.com
Sat Feb 10 22:03:47 UTC 2001


Stephan,

> >Quite a difference I'd say. Must be Bowie ;-)
> 
> Okay, I rechecked while playing "Rosenstolz" and it's still the same 
> performance... :-)

Ah! Give that man ten extra MHz for good taste :-)

> Ahem, but here's more serious observation: I recompiled the 
> Solaris VM using Sun's CC and got a 13% performance improvement. 
> I used CC=cc, CFLAGS=-xO5 vs. CC=gcc, CFLAGS=-O3. Then I also
> checked the generated SPARC assembler code (weird stuff) for
> interp.c and noticed also SUN's compiler generates the superfluous
> "if byte >= 255" test.  So it could be even faster if you'd help
> it with some assembler, similar as the gnuify 
> script rewrites the switch to allow gcc better optimization.

You might experiment with how much of the gnuification Sun's CC can deal
with. It could be that it has some of those features.

> From the windows benchmark, it seems that only ~8% are used 
> to force the display to the screen.

Yes, but you have to keep in mind that those functions use highly optimized
unrolled code. I checked the code generated by GCC and with a little help it
did generate the absolutely fastest code for generic Pentium architectures
(that is excluding the use of MMX or similar).

> 6% of processing time are used (wasted) to round the corners.

Turn it off, turn it off ;-)

> 18% are spent for layouting the browser.

Yes, I'm looking into this particular issue.

> > >[message tally]
> >This looks like you might try to disable #smartUpdating of 
> >browsers (it's a preference). But everything else looks 
> >quite reasonable to me.
> 
> It's nice that you and Bob and others find places where you 
> can save some time by disabling features.  However, I want
> all the features and still a fast system.

I know and we're trying to give it to you. But these numbers you're sending
are really useful since they do allow us to look at specific places. The
thing is just that most of us run high end machines (my daily machine is a
700MHz P3) and so some things just slip by. The more you tell us where the
bottle necks are, the more we can do.

Cheers,
  - Andreas





More information about the Squeak-dev mailing list