1368866 - Clicking during window.requestAnimationFrame() loop kills the rendering performance when APZ is enabled

Status: RESOLVED WONTFIX

(Core :: Panning and Zooming, defect, P3)

Description

[Miko Mynttinen]

Attachment: testcase

Clicking during window.requestAnimationFrame() loop increases DL, FLB, and rasterizing time by 3-4x, when APZ is enabled. Disabling APZ will make the problem go away.

Steps to reproduce: Open the testcase and click anywhere in the window. The performance hit should be obvious.

Reproduced on FF53 and Nightly, on OSX and Windows 10.

Profile: https://perfht.ml/2rCIl7I (clicked at around t=3.4)

Regression window: https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=dd1abe874252e507b825a0a4e1063b0e13578288&tochange=946ed22cad04431c75ab5093989dfedf1bae5a3e

Comment 1

[Miko Mynttinen]

Kats: The regression range includes bug 1242690, which might be related. Could you please take a look?

Comment 2

[Kartikaya Gupta (email:[e-mail])]

If I follow the STR with apz.minimap.enabled=true, I can see the displayport expands on clicking, so that explains the slowdown in paint time. The displayport expanding is more or less expected behaviour - we keep it small initially for faster initial page load but as soon as you start scrolling (or apparently even if you click on the document) the displayport will expand to the "full" size that we use during steady state.

The alternative here is to find a way to keep the displayport small during a click - but it will still expand as soon as the user starts scrolling. I'm not sure how useful it would be to make this change - it seems a bit artificial unless this is reduced from a real-world test case?

Comment 3

[Kartikaya Gupta (email:[e-mail])]

ni for the question at the end of comment 2 ^

Comment 4

[Miko Mynttinen]

(In reply to Kartikaya Gupta (email:[e-mail]) from comment #3)
> ni for the question at the end of comment 2 ^

Thanks for the feedback. We are using this testcase for display list building benchmarking, so it is not from a real-world test case.

Comment 5

[Miko Mynttinen]

Resolving as WONTFIX as per comment 2.