1320435 - don't reflow during paint suppression

Status: NEW

(Core :: Layout, defect, P3)

Description

[Timothy Nikkel (:tnikkel)]

Attachment: patch

Jeff noticed this. If we aren't painting then there is no point is doing reflows because we can't see the result of them. Adding some printfs lets one see that we do some reflows during paint suppression. We still need to resolve style so that we can start the loads of any CSS images. Hopefully there is nothing that happens as a result of or during reflow that we want to start as early as possible.

Looks like this is a ~4% win on tp5o non-e10s responsiveness, and about the same on the rest of talos.

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=8fec50b7c7f213c23ee4108965d0240c2e5a2890&newProject=try&newRevision=346c6d85ae03efe66e2e82a8b907b1a487f47b34&framework=1&showOnlyImportant=0

There are a few test failures

https://treeherder.mozilla.org/#/jobs?repo=try&revision=346c6d85ae03efe66e2e82a8b907b1a487f47b34

Comment 1

[Timothy Nikkel (:tnikkel)]

Couple of experiments. Only responsiveness changed in these, not the tp5 load score based on load time.

Going back to 250 ms (from 5 ms) paint suppression means this is a bigger win at about 8%

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=a70f2625969bc0fde2e7538a0a5072be738471b7&newProject=try&newRevision=19346ab3230000cac14a312e4f3958034d9b111a&framework=1&showOnlyImportant=0

Avoiding flushing style during paint suppression as well seems to be a 10% loss though

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=8fec50b7c7f213c23ee4108965d0240c2e5a2890&newProject=try&newRevision=becff4651c77fd5e2c8b62b25f3904303ff3b542&framework=1&showOnlyImportant=0

I guess we don't do any duplicate work with multiple style flushes, so this approach makes us do one big style flush.

Comment 2

[Astley Chen (inactive)]

Is it a perf work worth further investigation now?