Egor,
On 17 Jul 2007 02:44:25 +0400, Egor Pasko <[EMAIL PROTECTED]> wrote:
Guys,
I am not very stuck with ABCD. I am not. Not at all. Not am I.
Ehm.
Lickily, I finished digging into the implementation and making sure it
is correct. Now I am pretty confident that classic_abcd does the right
thing! (no guarantees, you know, it's software) Had to refactor the
code a bit to fill in the gaps of my poor understanding. I think, we
should commit the changes...
featuring:
* two-state Inequality Graph, dot printing is just beautiful
* better readability
* unit-like-tests against the new functionality
* option: -XX:jit.arg.dump_abcd_stats=true to dump stats
(total/eliminated)
* same amount of checks eliminated as before
* well-known tests breaking oldish ABCD _passed_, of course
in all, HARMONY-4476 (more details in JIRA)
Given that Windows is not what I am lucky with today, if a JIT guru
(Pavel, Mikhail, George?) had time to take a look at the patch and run
'build test' on Windows that would be really-really great!
I've commented in JIRA and started "build test" with your patch on Win32 -
goes fine so far.
And now the ugly porn:
1. I could not run almost all of DaCapo benches for various reasons, so
tested only on hsqldb, wow, anybody aware of it or is it just me
ugly little creature? had little time, sorry
I also ran into the several issues trying DaCapo benchmarks. They are caused
by recent commits in JIT and verifier. The JIT issues will be fixed soon,
but feel free to report the so we don't miss one. I'm also looking into the
verifier failure.
2. ~10% bounds checks removed in hot methods and 8% in total, with
ABCD innocent and many other optimizations very very guilty or
absent. OMG!
Did you compare the numbers before/after your patch? I remember we regressed
a bit in terms of performance after accepting last set of ABCD patches -
this was because of fixing too optimistic bound check removals.
3. "memopt" is probably the ugliest!! In my example:
1. array A lies in a non-volatile field
2. A.length is computed right in the method entrance
3. A is then loaded via "ldfield" bytecode instruction in the
middle of the method
4. nothing writes to the field, just accesses elements of A
And who could imagine that A.length would be computed twice with
"memopt" having nothing to do with second appearance A.length?
Thus, not eliminating the second appearance of this sequence:
I9:ldflda [t1.BidirectionalBubbleSort2::a] -) t4:ref:int32[]
I10:ldind.unc:[] [t4] ((t2,t3)) -) t5:int32[]
I11:chknull t5 -) t6:tau
I12:tauhastype t5,int32[] -) t7:tau
I13:arraylen t5 ((t6,t7)) -) t8:int32
this is what should be optimized out definitely! and the thing that
breaks two A.length-s apart killing the idea of ABCD.
Not good. But if this code is eligible for optimization we should optimize
it. Please file a JIRA issue, the patch would be even better! :)
4. "loop versioning" is not implemented, and this is what I would like
to take. I already wrote some 2-component-nullstone-like
performance tests to detect how good a JVM deals with loop
versioning. Will post the bunch of them soon.
JIT gurus,
given the ugly (1) - (4), rather critical for performance do you like
the idea to fight them in a high priority? Could you share your
vision, please?
Assuming we agreed to have a Harmony milstone every 2 month, the next (M3)
should be at the end of August. Now I"m going to fix as many bugs as
possible to raise Linux stability to the same level as Windows before the
milestone. Then I'll get back to performance tasks, including those you
listed above. Still I'm ready to discuss and test any performance patches
from you any time.
Thanks,
Pavel
--
Egor Pasko
