OS: Arch Linux
Graphics Hardware: AMD Radeon HD 4850 (RV770) (very old)

Example:
https://twitter.com/FirefoxNightly/status/1103593021122248707

With gfx.webrender.all enabled, videos on Twitter are not displayed properly sometimes, see the attached screenshot. To reproduce this bug, I often have to wait a few seconds or reload the page.

Firefox 66.0 is not affected, but 67.0b3 and current Nightly are affected, so this is a regression. Nightly build 20190227163104 is affected (I had this old one still installed for testing).

I wasn't able to reproduce this bug after downgrading the Mesa packages from 19.0.0 to 18.3.4.

Can you find a regression window using mozregression? https://mozilla.github.io/mozregression/

This was the first time I used mozregression and I hope that I did everything correctly. The result was:

Last good revision: 8c235a9b55a2b2d89115a7cc0f56a58c8883a980
First bad revision: 2853480ed90d0a995593e66fcd68a8bff1ad8d96

Pushlog:
https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=8c235a9b55a2b2d89115a7cc0f56a58c8883a980&tochange=2853480ed90d0a995593e66fcd68a8bff1ad8d96

This points to the patch for bug 1524385.

That's somewhat believable. I'll see if we can reproduce on a machine in the office.

Would it help if I bisect Mesa (probably between 18.3.0 and 19.0.0)? It would take some time because my hardware is very old, but I am curious why a change in Mesa triggers this bug.

(In reply to Viktor Jägersküpper from comment #5)

Would it help if I bisect Mesa (probably between 18.3.0 and 19.0.0)? It would take some time because my hardware is very old, but I am curious why a change in Mesa triggers this bug.

It would definitely be helpful to bisect mesa.

Or I should say if you find the commit in Mesa that causes that change that would be helpful for us in figuring out what's happening on our side. It may also be a Mesa bug in which case it's worth getting fixed upstream.

I finished the bisection of Mesa with the following result:

Last good commit:
6f4083143bb8c478ccfcaef034d183d89b471993 gallivm: use llvm jit code for decoding s3tc

First bad commit:
1250383e367fef6fdb954d69a7444634c6788f5d st/mesa: remove sampler associated with buffer texture in pbo logic

See the bad commit here:
https://gitlab.freedesktop.org/mesa/mesa/commit/1250383e367fef6fdb954d69a7444634c6788f5d

This is indeed between the branch points of 18.3.0 and 19.0.0.

I reproduced the bug with a recent 19.1.0-devel snapshot of the Mesa master branch. Furthermore I reproduced the bug with a current Nightly build on my laptop which has a similar graphics chip (AMD RV710), running Debian testing and after upgrading the Mesa packages from 18.3.4 to 19.0.0 (found in the experimental repository).

Does this help you? Or do you think that it's really a Mesa bug?

Clearing the needinfo request because this bug is now discussed on the mesa-dev mailing list, see here:
https://lists.freedesktop.org/archives/mesa-dev/2019-April/218511.html

It seems that there won't be a quick or even easy fix on the Mesa side. At least it seems that this bug only affects very old AMD hardware (DirectX 10 era).

This seems to be a regression because of a Mesa commit (see comment 8). Marking fix-optional for 67 beta because of external dependency.

According to Dave Airlie, this is really a bug in the r600 driver, which was uncovered by the mesa commit mentioned in comment 8, and then triggered by the WebRender commit mentioned in comment 3. So closing this bug as INVALID. Feel free to change that to WONTFIX if you think that this is be more appropriate.

Here's the patch to fix the mesa bug: https://patchwork.freedesktop.org/patch/302675/

The patch has been committed to Mesa master with commit e2fecf57e343961cd1db9251cef2c9eb95948d2f and has been included in the Mesa staging/19.0 branch (see https://gitlab.freedesktop.org/mesa/mesa/commit/602ca580486bd0f266ee742e7e9c2c1c4fd5cbf5), so it should be included in the next stable release.

The patch has been included in Mesa 19.0.4, I verified that this bug doesn't occur with that version any more.