Repro note: attached screenshot 1 shows a partly painted canvas. The paint rectangle is misplaced which is clearly seen upon enabling "Paint flashing" option in devtools - Rendering panel (screenshot 2).

Deleted: screenshot1.png 437 KB

	screenshot1.png 437 KB View Download

Deleted: screenshot2.png 546 KB

	screenshot2.png 546 KB View Download

Able to reproduce this issue on Win-10 using chrome reported version #56.0.2924.87 and latest canary #58.0.3003.0.
Unable to reproduce on Linux and Mac.

This is a non-regression issue as it is observed from M45 old builds. 

When checked using M-40 builds, the video at https://jsfiddle.net/gfgwdsho/ did not play.

Hence, marking it as untriaged to get more inputs from dev team.

Thanks...!!

Routing to canvas team - could anyone triage this?

It looks like the paint invalidations have a bad Offset.  The invalidation rectangles have the right size, but their top left corner reflects the top left corner of the clipped overflow content rather than the top left corner of the canvas element.

The reason this is Windows-specific probably has to do with the fact that ANGLE exposes a D3D feature that allows partial buffer swaps.  We use this to accelerate compositing by only updating invalidated area to the presentation surface.

The bug is canvas-specific. I tried using object-fit:cover with a video and with an img (animated gif), and they worked fine.

Object-fit has special code in CompositedLayerMapping for images IIRC. Probably
needs to be copied for canvas? Let me know if you need pointers. I think pdr
wrote the latest fixes (a year or so ago).

Junov, can you check contentsRect in CompositedLayermapping? I didn't verify if it's the culprit, but the code looks suspect.

Just noticed this bug. We are clearly not doing the right thing with object-fit for canvas (cover or content). Will address soon.

fserb@: I have another repro case directly from our layout tests. Please see the attached file. 

If you change the size of the canvas to be 288*144, which is smaller than 256*256, then the result is correct because we are not using GPU canvas. As long as the size of the canvas is larger than the threshold, we trigger the GPU canvas, and the result is wrong. Particularly for cover and none I think.

Also, looks like it is clipping that is wrong.

Deleted: circles-landscape.png 6.9 KB

	circles-landscape.png 6.9 KB View Download

Deleted: object-fit-canvas.html 1.9 KB

object-fit-canvas.html 1.9 KB View Download