Re: Talos Regression :( Dromaeo (jslib) decrease 7.31% on XP Firefox

[Henri Sivonen]

> On Fri, Sep 17, 2010 at 12:46 PM, Nobody
> <nob...@cruncher.build.mozilla.org> wrote:
> > Regression :( Dromaeo (jslib) decrease 7.31% on XP Firefox
> > ----------------------------------------------------------
> > Previous: avg 133.289 stddev 2.396 of 30 runs up to revision
> > 268ef4ccb5ff
> > New : avg 123.541 stddev 0.943 of 5 runs since revision bc15c280c430
> > Change : -9.748 (7.31% / z=4.068)
> > Graph : http://mzl.la/9hGqIo
> >
> > Changeset range:
> > http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=268ef4ccb5ff&to
> > change=bc15c280c430
> >
> > Changesets:
> > * http://hg.mozilla.org/mozilla-central/rev/0ab712643a66
> > : Henri Sivonen <hsiv...@iki.fi> - Bug 591981 - Make script-inserted
> > inline scripts run right away, make script-inserted external scripts behave
> > like async scripts and make document.write writing an inline script return
> > at a predictable time. r=jonas, a=blocking2.0-beta7.
> > : http://bugzilla.mozilla.org/show_bug.cgi?id=591981
> >
> > * http://hg.mozilla.org/mozilla-central/rev/bc15c280c430
> > : Henri Sivonen <hsiv...@iki.fi> - Bug 531056 - Make the HTML5 parser
> > unaware of <spacer>. r=jonas, a=jst.
> > : http://bugzilla.mozilla.org/show_bug.cgi?id=531056
> >
> > Bugs:
> > * http://bugzilla.mozilla.org/show_bug.cgi?id=531056 - [HTML5] Remove
> > <spacer> support
> > * http://bugzilla.mozilla.org/show_bug.cgi?id=591981 - Don't enforce
> > execute-in-insertion-order for DOM inserted <script>s

I won't be able to solve this mystery today. The regression is a Windows thing, so reproducing it locally using Mac or Linux is futile. I can't reproduce the problem on my Windows 7 VM using the Dromaeo Web harness. My XP VM isn't really suitable for running performance tests because it does its disk IO over Ethernet. I intend to continue investigation on Windows tomorrow.

I attached (https://bug591981.bugzilla.mozilla.org/attachment.cgi?id=476778) the only Gecko-side tweak that I could think of to the bug. However, that tweak didn't fix the Windows Dromaeo numbers on try. I verified on try that reverting nsScriptLoader.cpp and .h make a Dromaeo regression go away.

Should http://hg.mozilla.org/mozilla-central/rev/0ab712643a66 be backed out regressing site compat, HTML5-compliance and cross-browser compat? Or should it be left in keeping an ununderstood Dromaeo regression in the tree?

--
Henri Sivonen
hsiv...@iki.fi
http://hsivonen.iki.fi/

[Jonas Sicking]

I think at this point the regression has been in the tree long enough
that we need to back out unfortunately.

We'll have to figure out how to proceed from there once we know why
things regressed or otherwise have a patch that doesn't regress.

/ Jonas