Does it work ok on M63?

Repros on a Pixel 1 with M63. Does not repro on Linux with or without gpu
raster.

High DPI issue then?

Did we enable zoom-for-highdpi on ChromeOS or something like that? I wasn't aware of it.

--force-device-scale-factor=2 repros on Linux.
--prefer-compositing-to-lcd-text does not reproduce on Linux.

To be more exact: reproduces on ChromeOS Beta 63.0.3239.70
Also reproduces on Linux with  --force-device-scale-factor=2, on Linux 63.0.3239.108. 

Note that it reproduces on Linux both with hardware and software rasterization.

Assigning for bisect.

Reproduces on Mac Retina.

This is broken at least as far back as r464641, or M-59. So not a regression at this point. I'll go further back.

Amazing nobody noticed.

And broken in M-55. I think it's fair to say this is not a regression.

I'll still own it because I'm refactoring this code.

This might be a site bug, at least for stockplanconnect. They have a ::before psuedo element to display the image, and they have

@media only screen and (-webkit-min-device-pixel-ratio: 2),not all

with a modified top property. If you remove all that crap it works just fine. Maybe they are trying to work around a Safari bug?

Could someone with mint access see if the CSS is the same?

More details on stockplanconnect: When the page see a scale factor of 2 it shifts to a sprite atlas with icons at double the size. It also adjusts the top of the pseudo element presumably to index into the right place in the new atlas.

But the CSS size of the pseudo element is auto, so does not change with the dpi setting. Trying to draw an image that is twice as big into an element that is the same size will crop the image, as one would expect. The offset also seems to be incorrect.

So site bug.

Broken in Safari on Retina screens, looks the same as Chrome. Works in Firefox because they ignore the media selector.

Site bug.