Skip to main content

Will be stack-allocation based on escape analysis integrated in Mustang?

3 replies [Last post]
linuxhippy
Offline
Joined: 2004-01-07
Points: 0

Hi there!

I found the comment that escape-analysis based optimizations will be integrated into Mustang but till now nothing has happend :-(
However it seems from the performance-side not a lot has happend till Tiger...

This would open the door for at least 2 (especially on server) very cool optimizations:
* stack-allocation
* lock elimination

Or will we have to wait another 24 months till we ~maybe see something working :-(

lg Clemens

Reply viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.
linuxhippy
Offline
Joined: 2004-01-07
Points: 0

Hi again,

this bug-log made me a bit enthusiastic:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6325352

Could you please explain what the following line means:
* call record_for_igvn(phi) only when escape analysis is enabled.

Does that mean Mustang does already this sort of optimization, or is it already implemented but not enabled by default?

Thanks you in advance, lg Clemens

tmarble
Offline
Joined: 2003-08-22
Points: 0

Ig:

Are there cool optimizations to be had through
escape analysis? It certainly seems likely and
we have several efforts currently underway.

We usually publish results of performance
improvements from release to release
(for example, see the J2SE 5.0 Performance Whitepaper [1]).
Although we have given some early indications
of Mustang performance [2], we don't publish
performance results prior to a release.
Understandably some propsective optimizations work and
some don't but rest assured that we now have
quite an impressive battery of automated performance
testing and, for most platforms and most benchmarks,
we are seeing nice performance improvements in Mustang
relative to Tiger. ;-)

What's fundamentally new about Mustang is you don't
have to wait 24 months "to see something working".
You can get binary and source snapshots now [3].
You can see that a lot has happened since Tiger [4].
And if you have some passion and ability around
techniques such as escape analysis [b]you[/b] can make
contributions to Mustang [5].

HTH,

--Tom

[1] http://java.sun.com/performance/reference/whitepapers/5.0_performance.html
[2] see the slides from "TS-7984 Java Platform Performance" from JavaOne San Francisco 2005.
[3] https://mustang.dev.java.net/
[4] https://mustang.dev.java.net/MustangCommittedBugs.html
[5] https://mustang.dev.java.net/collaborate.html

linuxhippy
Offline
Joined: 2004-01-07
Points: 0

> Are there cool optimizations to be had through
> escape analysis? It certainly seems likely and
> we have several efforts currently underway.
Great to here something is happening on that side.

> We usually publish results of performance
> improvements from release to release
> (for example, see the J2SE 5.0 Performance Whitepaper
Always found SUN's benchmark-results very interresting - more or less unbelieveable.
Especially the performance-whitepapers looked like every new jvm-version is at least 50% faster the the previous - like the results published for the aggressiv.gc flag - more a joke than reality.
Always thought microbenchmarks are bad ;-)

> And if you have some passion and ability around
> techniques such as escape analysis [b]you[/b] can
> make
> contributions to Mustang [5].
I am currently contributing to one free java implementation and whatching the development of several free jvms, but give my at least ONE reason why I should give Sun my source and all my rights to that source.
Its absolutly no problem for me that Sun would use my source, but the fact that I loose all my rights is quite problematic - Sun may decide to no more distribute java for free and what happens with my source - it helps something I never wanted to see happen.
The problem seems that Sun does not give any type of agreement - its their source and therefor of course they may do with it whatever they want to do. No luck for me ;-)
Furthermore I am quite in doubt I could implement some JIT-optimizations, I am quite not educated enought in this field ;-)
And again furthermore Nokia gives me a 770 for porting a free jvm to their platform so no luck with a t-shirt ;-)

> What's fundamentally new about Mustang is you don't
> have to wait 24 months "to see something working".
Yes, thats really great - almost unbelievable SUN's management gave its ok to this ;-)
I think both, SUN and the java-developers can benefit of this situation - mustang bugs can be found during development whereas devs see whats going on on future java releases. Again, really great!

lg Clemens

Message was edited by: linuxhippy

Message was edited by: linuxhippy