346641 - Error console performance

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect)

Description

[Jonathan Watt [:jwatt]]

Boris sent an email to mozilla.dev.performance regarding the unusable state of the Error Console on cairo builds. The text is copied below:

----------

So I discovered today that on trunk the error console takes _forever_ to open. To be precise, starting with a blank error console, then loading the following HTML:

<script>
for (var i = 0; i < 20; ++i) {
  setTimeout("throw 'This is a test of the emergency broadcast system; only a test and all that dudes'", 0);
}
</script>

then opening the error console, I see it take 15-20 seconds for the window to appear.

A profile shows that we're spending a _lot_ of time in reflow (pretty much all the time is in InitialReflow()) and of the reflow time about 90% is spent under nsTextFrame::MeasureText.  Almost all of that is under gfxPangoTextRun::Measure, gfxPangoTextRun::~gfxPangoTextRun, and gfxPangoFontGroup::MakeTextRun (with these last two accoundting for about 15% of the MeasureText time, with the rest under Measure).  The time under Measure is almost all gfxPangoTextRun::EnsurePangoLayout.

Of particular interest is pango_context_get_base_dir, which takes up about 25% of total layout time.  This seems to be doing a lot of mallocing, removing from font maps, creating new script iters, bsearch, hashtable lookups, utf8 to ucs4 conversion, etc...

I see this in both Firefox and Seamonkey cairo builds.

For comparison, the non-cairo builds open up the Error Console in about 4 seconds (which is _still_ ridiculous) when using the same testcase.  So there is probably also the problem of us doing a _lot_ of reflow here for some reason. And of cairo making that about 5x slower... :(

I'd be happy to mail the profile to anyone interested in it.

Also of interest, a reflow branch non-cairo build opens the console about twice as fast as a trunk non-cairo build.  I don't have a reflow branch cairo build on hand to test with, though.

-Boris

Comment 1

[James Ross]

Blocking the cross-platform cairo metabug for now. If the problem is only on Linux (unlikely, TBH) the correct blocking bug is bug 334720.

Comment 2

[Christian :Biesinger (don't email me, ping me on IRC)]

(In reply to comment #1)
> Blocking the cross-platform cairo metabug for now. If the problem is only on
> Linux (unlikely, TBH) the correct blocking bug is bug 334720.

seems likely to me, given:
> Almost all of that is under
> gfxPangoTextRun::Measure, gfxPangoTextRun::~gfxPangoTextRun, and
> gfxPangoFontGroup::MakeTextRun

Which is only used on Linux (well, the rest of X11 too, and BeOS, and maybe OS/2?)

Comment 3

[Ryan VanderMeulen [:RyanVM]]

Is this still an issue? At least on Windows, error console comes up after only a short delay.

Comment 4

[Boris Zbarsky [:bzbarsky]]

This does seem to be OK now with new-textframe and the reflow branch...

> At least on Windows,

This was a Linux-specific bug.

Comment 5

[Ryan VanderMeulen [:RyanVM]]

I came upon the bug when looking at bug 334719, so that's why I mentioned my platform :-)