On 9/29/05, Angus Lees <[EMAIL PROTECTED]> wrote:
> At Tue, 27 Sep 2005 12:00:09 +1000, Bruce Badger wrote:
> > In fact, the very best of the JITing VMs can get performance that
> > exceeds that attainable by static compilation - because there is
> > more information available at run time to base the tuning decisions
> > upon.
>
> If a program's use changes over its invocation, and the JIT
> continually shifts its optimisation targets, then I can see the
> potential benefit of this approach. I don't believe, however, that
> there are many programs that have this dynamic behaviour.
I agree. It is only in very dynamic, high throughput and long-lived
services that one would see a measurable benefit in having such a
sophisticated VM, though I would not be surprised to see heavily used
web servers falling into this category.
> You can gain the same runtime knowledge in a statically compiled C
> program by compiling with gcc's -ffprofile-arcs, running over some
> typical use cases (will write a bunch of .gcda files) and then
> recompiling with -fbranch-probabilities.
Right. For static problems, or problems with a well understood number
of modes of operation static compilation can be superb. For each new
mode encountered in the wild, though, one would have to tweak the
compiler hints and rebuild to keep up with our imaginary perfect
JITer.
I think the key is your first point. The cases where absolute
performance is critical are very rare indeed. I'm happy, though, that
I am using an environment where I can focus on the problem at hand,
and delegate many low-level issues to the environment itself and at
the same time expect performance on a par with (and perhaps even
better than?) the best hand-crafted binaries.
Only real circumstances will tell. I'd love to work in an environment
that was sophisticated and high-load enough to put some of the
advanced JITing VMs to the test.
All the best,
Bruce
--
Make the most of your skills - with OpenSkills
http://www.openskills.org/
--
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html
