451028 - Bug 447739 regressed performance on 64bit Linux

Status: RESOLVED FIXED

(Core :: Layout, defect)

Description

[Arpad Borsos [:Swatinem]]

With the patch from bug 447739 applied, the performance of the scrollmark test drops to ~122280ms, which is even below 1.9.0 level (Which is ~72532ms)
Without the patch, the performance was roughly 20s with the iframe fully visible.

System: Ubuntu Hardy 64bit

Comment 1

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Interesting, I can actually reproduce this on Mac (debug). QuartzDebug shows that the problem is we're painting too large an area, we seem to basically be painting two lines instead of one at every scroll operation.

Comment 2

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Attachment: fix

This fixes it for me. Arpad, if you're able to test this, I'd appreciate it.

Sorry David, I know you asked me to check those additions, and I did, but I still got it wrong. The excluded area should not be translated, we've already intersected r with r+aDelta so that's the absolute area (relative to the reference frame) that we don't need to invalidate and no further translation is necessary. Additional translation means we've started using r+2*aDelta which is obviously wrong.

I checked this does not regress bug 447739.

Comment 3

[Arpad Borsos [:Swatinem]]

This patch got me back to 20437ms on the scrollmark in the case of a fully visible iframe, but I get 126106ms with a partially offscreen one.
Thats weird.

Comment 4

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Pushed 40bd4f8a989d.

Arpad, I'm afraid that if you still see the problem in a nightly build you'd better file yet another bug.