Closed Bug 481103 Opened 15 years ago Closed 14 years ago

TM: Dromaeo v3 Sunspider massive slowdown after 2-28 merge

Categories

(Core :: JavaScript Engine, defect, P2)

x86
Windows Vista
defect

Tracking

()

RESOLVED INCOMPLETE

People

(Reporter: RyanVM, Unassigned)

References

()

Details

(Keywords: perf, regression)

See the URL above. This regressed on m-c with the Sat Feb 28 11:57:55 2009 -0800 TM merge. It regressed in the TM nightlies between 2-26 and 2-27.

Regression range:
http://hg.mozilla.org/tracemonkey/pushloghtml?fromchange=aea34f524423&tochange=737ca70d654f
Flags: blocking1.9.1?
Keywords: perf, regression
Summary: Dromaeo v3 Sunspider massive slowdown after 2-28 TM merge → TM: Dromaeo v3 Sunspider massive slowdown after 2-28 merge
Today's TM merge didn't help this, FWIW.
http://dromaeo.com/?id=61342,61343,61567

First two are the same as the URL from the bug. Third is today's build. It looks like some of the tests are back to where they were, but some are still in bad shape.
It appears that whatever caused this regression is fixed on the Tracemonkey branch. The link below shows the following TM nightlies: 2-26, 3-3, 3-4, 3-5, and 3-6
http://dromaeo.com/?id=61794,61788,61786,61791,61792

The 2-26 build shows the numbers I got with the final TM nightly before the regression hit. The 3-3 build shows the regression fully. The 3-4 and 3-5 builds show some improvement in the tests, such as AES. The 3-6 shows us fully returning to pre-regression performance levels and then some.

I'll confirm that this is fixed when the next TM merge occurs, but this certainly looks promising.
Is bug 480494 related to this? 

It does have an earlier regression range though, http://hg.mozilla.org/tracemonkey/pushloghtml?fromchange=4ada519b5c97&tochange=5de9f1e51c68
http://dromaeo.com/?id=61342,61928

The first is the original final pre-regression build from comment #0. The second is current tip after today's m-c merge. As you can see, things are looking very good, except for the two DNA tests. The rest look to be even or faster with today's build, and the final score is higher with today's build.

So, it looks like this is mostly fixed, but not all the way yet.
Flags: blocking1.9.1? → blocking1.9.1+
http://dromaeo.com/?id=61342,62632

Pre-regression build and m-c tip including today's m-c merge. DNA is clearly still showing problems, but everything else looks good (or close enough to be considered equal)
Priority: -- → P2
http://dromaeo.com/?id=61342,62632,63738

Running a build off the current tip (post upvar2 landing), overall speed is definitely up. However, that mainly seems to be due to a massive speedup of Bitwise And. Otherwise, some of the other tests (like Compute Bits in Byte) are showing regressions.
upvar2 had nothing to do with any change to bitops-3bit-bits-in-byte.js (if that is the "Compute Bits in Byte" test) since that test has two top-level functions, no closures. Either some other patch made for a change, or this is noise (which could be a bug in itself).

Dromaeo is not always measuring what the individual SunSpider tests were intended to measure, and do measure when run via SunSpider. See bug 480494 comment 24.

/be
Yeah, it could very well have been another one of the changes from the TM merges that happened over the last 48 hours or so. I doubt a change of an order of magnitude is noise, though, especially given that the numbers have been otherwise stable on that test to this point.
If DNA is the main issue, I'm already working on that in bug 480494.
Depends on: 480494
It's more likely to be Andreas' type mismatch patch that improved this.
Who owns this?
I intend to work on this, but I have 4.5 blockers ahead of it in my queue, so I didn't want to take it until I cleared those out.
given these results, I don't think we can block on this:

http://dromaeo.com/?id=65550,65551
Flags: blocking1.9.1+ → blocking1.9.1-
(In reply to comment #14)
> given these results, I don't think we can block on this:
> 
> http://dromaeo.com/?id=65550,65551

What versions are being compared in that view?
Fx3 (the slower one) vs. Fx3.5
We're not going back to this one any time soon, I have confidence.
Status: NEW → RESOLVED
Closed: 14 years ago
Resolution: --- → INCOMPLETE
You need to log in before you can comment on or make changes to this bug.