http://diana-adrianne.com/purecss-francine/ stresses WebRender

[jrmuizel]

We have some drawing bugs on this page and it doesn't run that fast.

That being said the page is kind of crazy

[kvark]

Here is some findings from GPU capturing:

• similarly to regression: low FPS on https://servo.org/bhtml-newtab/webrender-demos/moire.html #2706, there is a giant number of rounded clips. 14 layers, ~380 primitives for each. However, these don't appear symmetric here, so re-using one of the corners wouldn't work. Caching the clips between frame would help though.
• unlike the previous case, there are more clips being applied here on top (box_shadow), so we don't have a path of just exposing the rounded cornered rectangle primitive like we used to (a while ago).
• lots of batch breaks in transparency passes (in box_shadow shader in this case, but this is not relevant). This is a general problem we hit (e.g. WebRender has a bad time with https://jrmuizel.github.io/perf-tests/moving-balls.html #1501), and I have ideas to improve the situation, so I'll bump the priority on those ;)
• (minor) there is a few 0-instance calls (e.g. to B_LinearGradient), which we can eliminate
• there is overall a ton of gradients, borders, and clipped rectangles being drawn

[pcwalton]

Tessellating large rounded corners could help.

[jrmuizel]

Here is some more of this kind of thing:
http://diana-adrianne.com/purecss-zigario/
http://diana-adrianne.com/purecss-vignes/

[Zirro]

The performance for this page seems to have regressed further in recent versions of Nightly, with attempts to scroll repeatedly crashing the browser on my machine.

[kvark]

Looking into this. Confirmed the crash to be https://bugzilla.mozilla.org/show_bug.cgi?id=1489644

[gw3583]

@Zirro @kvark I'm working on a patch to fix a different performance issue - just mentioning it here as it might be relevant to performance on this page. The patch will allow us to treat transforms with scale + offset as the same coordinate system (previously we could only do this for simple translations). This will avoid the need for clip masks on any clip rectangles where the transform includes a scale + offset.

[gw3583]

Picture caching makes this quite usable, however when stuff does invalidate it's still very slow to draw. We can leave this open for the rasterization work to improve this - moving box shadows out of using clips would be the simplest fix for this.