Unable to reproduce the issue on MacBook Air 10.11.4 using chrome latest stable M50-50.0.2661.94. Observed the fixed header is seen and no artifacts while scrolling the compositor_bug.html file.

joel@ - Could you please recheck this issue on latest chrome stable by creating a new profile under chrome://settings with no apps or extensions in your browser. If issue still persists on latest stable please try it on incognito mode as well. Attaching screen-cast from our end, please have a look on it and let us know if we are missing anything from our end.

Thanks!

Deleted: Recording #7.mp4 3.0 MB

Recording #7.mp4 3.0 MB Download

Thanks for looking into this. I'm starting to fear it may be a GPU-specific issue. I was running on the previous stable release, but I just updated to 50.0.2661.94, and the issue still appears (see attached .mov).

Note that I'm on a Late 2013 15" MacBook Pro with an nVidia GeForce GT 750M. This issue was reported to me by a customer who's also running Chrome/Mac, but unfortunately I don't have any information on his GPU.

Deleted: 607565.mov 23.5 MB

607565.mov 23.5 MB Download

Thank you for providing more feedback. Adding requester "brajkumar@chromium.org" for another review and adding "Needs-Review" label for tracking.

For more details visit https://www.chromium.org/issue-tracking/autotriage - Your friendly Sheriffbot

I can confirm this is happening.

Interesting enough, the issue happens only when input has left position that is < -3996px! 

I suspect this may be because CC thinks that the layer is outside of the region where it has to be painted. This can explain why it happens only when the page is scrolled.

3996px is 4096px - 100px. While I don't believe the original repro I posted had anything precisely 100px wide, that still seems suspicious. Might that have some subtle interaction with the size of tiles, or an assumed maximum resolution? I haven't looked at the Chrome rendering code in ages, or I'd go look for that myself.

It's probably the interest rect calculation being wrong. Will look.

https://code.google.com/p/chromium/codesearch#chromium/src/third_party/WebKit/Source/core/layout/compositing/CompositedLayerMapping.cpp&q=compositedlayermapping.cpp&sq=package:chromium&type=cs&l=2189

It's not the interest rect.

Take that back, it is the interest rect. If I adjust the kPixelDistanceToRecord
constant, the magic left: padding value for the input changes.

Fix in progress.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/2ea2924dc159391166a1b42008f576a501e42ea3

commit 2ea2924dc159391166a1b42008f576a501e42ea3
Author: chrishtr <chrishtr@chromium.org>
Date: Fri May 20 00:17:25 2016

Apply scroll offset and content box offset before recursing in mapAncestorToLocal

The way mapAncestorToLocal works is to compute a localToAncestor transform, then
invert it at the end. Thus the incremental pieces of the transforms must be applied
bottom-up.

Also clean up a few places where scrollOffset() is simpler to use than scrollPosition.

BUG= 607565 

Review-Url: https://codereview.chromium.org/1994203003
Cr-Commit-Position: refs/heads/master@{#394922}

[modify] https://crrev.com/2ea2924dc159391166a1b42008f576a501e42ea3/third_party/WebKit/Source/core/dom/IntersectionObservation.cpp
[modify] https://crrev.com/2ea2924dc159391166a1b42008f576a501e42ea3/third_party/WebKit/Source/core/layout/LayoutView.cpp
[modify] https://crrev.com/2ea2924dc159391166a1b42008f576a501e42ea3/third_party/WebKit/Source/core/layout/MapCoordinatesTest.cpp

Thanks for fixing this so quickly!