Since this afternoon, the page-loader test on btek has been crashing while loading the test copy of www.expedia.com. bryner and I have also reproduced this in linux builds, and me in win32. On linux, it crashes on the first load of that page, although on win2k, it actually made it through one complete cycle before crashing again. I'll attach stacks in a moment. bryner tried backing out waterson from this afternoon, and that apparently didn't fix it. I am building without hyatt's changes from this afternoon now and will see if that fixes it.

blocker

I have now fixed this by backing out my use of ReconstructStyleData in StyleSheetAdded. I did leave my use of this function in StyleSheetRemoved, which is called on a skin switch. So now we have reproducible crasher that will happen on a skin switch, and I can work to figure out why ReconstructStyleData can't be used everywhere.

hyatt's change turned btek green, marking fixed.

I'm keeping this open to cover wanting to switch over to the better API. Don't mark this fixed.

Comment on attachment 62133 [details] [diff] [review] Avoid reframing when style sheets are added, style rules are added, or alternate stylesheets are enabled/disabled. sr=waterson

Comment on attachment 62133 [details] [diff] [review] Avoid reframing when style sheets are added, style rules are added, or alternate stylesheets are enabled/disabled. r=jag with the comment change in PresShell::StyleSheetRemoved

Fixed.

So the first set of numbers off btek say that reductions for those pages that had <style> elements in the body of the content: avg. cached initial uncached www.msn.com 114% 113% www.expedia.com 126% 132% www.cnn.com 121% 120% (which is almost 500 msec better) The rest of the pages were neutral to favorable to this change. Overall, about ~2.5% win. Nice!

This fix caused the regression bug 129350. When a style sheet is added we need to reframe, at least that is the only thing that makes sense to me. Otherwise, existing frames will have bogus style data, based on the rules that were in the the old set of rules. Further analysis and work in bug 129350