1492859 - 3d transforms problem on https://youriguide.com/sample-full/

Status: RESOLVED DUPLICATE of bug 1415987

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

Description

[Darkspirit]

Attachment: 2018-09-20_16-33-33.mp4

Please see attached video. (Everything has been worse the days before.)

Comment 1

[Darkspirit]

Screencast = Try build from bug 1492566 comment 1:
mozregression --repo try --launch 576fe180d245f338bc522ed4ea346b0a22abfd2c --pref gfx.webrender.all:true -a https://youriguide.com/sample-full/

It looks the same on Win10 (GTX 1060) and looks fine without WebRender.

Comment 2

[Darkspirit]

The problem from attachment 9010649 [details] has been fixed:

https://hg.mozilla.org/integration/autoland/graph/76d218ea875e
> 76d218ea875e: Bug 1489337 - Don't set preserve-3d to true when creating WebRender commands for nsDisplayPerspective since it's not needed. r=jrmuizel 
mozregression --repo autoland --launch 76d218ea875e --pref gfx.webrender.all:true -a https://youriguide.com/sample-full/
first good

> 96059b4f6d8a: bug 1490119: Add build system bits for building dep updater. r=firefox-build-system-reviewers,mshal,rstrong 
mozregression --repo autoland --launch 96059b4f6d8a --pref gfx.webrender.all:true -a https://youriguide.com/sample-full/
last bad

Comment 3

[Darkspirit]

Attachment: Screenshot_20180921_211153.png

Comment 4

[Dzmitry Malyshau [:kvark]]

Awesome, thanks :darkspirit for confirming! The remaining issue appears to be simpler.

Comment 5

[Dzmitry Malyshau [:kvark]]

Reflecting my comment on Github, the remaining issue appears to be related to the image handling in Gecko: when the bug is visible on the corners, Gecko provides us with 1:1 image. Perhaps, it's an aggressive mipmapping gone wrong? Adding Andrew.

Comment 6

[Jeff Muizelaar [:jrmuizel]]

:kvark can you still reproduce this? I don't really understand what you think the problem was. Can you elaborate?

Comment 7

[Dzmitry Malyshau [:kvark]]

Yes, I'm still seeing it on 65.0a1 (2018-11-09) (64-bit)
What I said in Comment 5 is that Gecko provides us with bogus image data (size 1x1) in those buggy cases, I don't see WR doing anything wrong.

Comment 8

[Andrew Osmond [:aosmond] (he/him)]

Can still reproduce, I'll take a look.

Comment 9

[Andrew Osmond [:aosmond] (he/him)]

The stacking context has scale factors of (0, 0) which makes us request an image of 1x1 (since the clamp the width and height to be at least 1).

Comment 10

[Andrew Osmond [:aosmond] (he/him)]

We are missing:

https://searchfox.org/mozilla-central/rev/ff46b36ac2ebb243ec95fdab01340391963d62e5/layout/painting/FrameLayerBuilder.cpp#6114-6118

In the StackingContextHelper constructor. This would save us in this case.

Comment 11

[Andrew Osmond [:aosmond] (he/him)]

I can easily copy/paste the necessary code from comment 10, but it seems like bug 1415987 is going to fix it anyways since it will reuse that code path to calculate the scale.

Comment 12

[Darkspirit]

Fixed by bug 1415987 comment 17.