1821937 - CanvasRenderingContext2D::BeginPath() show up in Speedometer 3 canvas (Charts-chartjs, Performance-Dashboard) profiles

Status: NEW

(Core :: Graphics: Canvas2D, defect, P2)

Description

[Olli Pettay [:smaug][[e-mail]]]

Looks like at least mPath = nullptr; is causing quite a bit allocator traffic.
Could we reuse some of the objects?
https://searchfox.org/mozilla-central/rev/2ede53e39955988f98db4369f7ce09614b22104a/dom/canvas/CanvasRenderingContext2D.cpp#2842-2843

I used https://speedometer-preview.netlify.app/interactiverunner for running the test, and it is ZoomTheChart-sync part where the BeginPath shows up.

Comment 1

[Bob Hood [:bhood]]

Jeff, would this be at all impacting SP3?

Comment 2

[Olli Pettay [:smaug][[e-mail]]]

This is all about sp3 ;) React-Stockcharts is part of that.

Comment 3

[Jeff Muizelaar [:jrmuizel]]

Can you include a link to the profile? Path objects are immutable so reuse is not so easy or pleasant.

Comment 4

[Olli Pettay [:smaug][[e-mail]]]

https://share.firefox.dev/3YTzPN0
That is on Windows.

Comment 5

[Jeff Muizelaar [:jrmuizel]]

PathRecording currently keeps two copies of the Path data: mPathOps and mPath. mPath is used for implementing ContainsPoint(), StrokeContainsPoint(), GetBounds() and GetStrokedBounds() however we rarely use these methods. It's probably reasonable to make conversion to mPath on demand instead of eager. That would make recording cheaper and reduce the amount of time spent freeing. It should also help SVG blob performance.

Comment 6

[Jeff Muizelaar [:jrmuizel]]

Olli, can you get a new profile now that bug 1447839 has landed?

Comment 7

[Olli Pettay [:smaug][[e-mail]]]

It is better, but still quite (de)allocation heavy: https://share.firefox.dev/400eB1m

Without looking at the code, just the profile:
Would it make sense to not use std::vector, but AutoTArray?
I see also std::unordered_set. Is that possibly more allocator heavy than the hashtables we normally use?

Comment 8

[Markus Stange [:mstange]]

This also shows up on the Charts-chartjs and Perf-Dashboard subtests. Fixing the performance difference with Chrome would improve Firefox's score on these subtests by 0.9% and 0.5% respectively.

Firefox: https://share.firefox.dev/44EoIv0
Chrome: https://share.firefox.dev/3Y5OOV8