Nice find! I noticed the strange coloration with sub-pixel rendering (which seems worse with Vulkan), but never would have figured this was related to transparent backgrounds.

Very nice find indeed!

LGTM, didn't test.

Edit:

Currently, this is easy since the background is always a solid color. It may be a little more complex (but doable) if we implement gradients in future.

Lets deal with that when we actually do support gradients. We might also want to auto scale the gradients based on the output they are currently rendered on. So maybe we can just render the pango text directly on the gradient buffer.

Edit 2:
Actually, after reading though the linked wlroots issue:
We do create the cairo surface with CAIRO_FORMAT_ARGB32 and then set the wlroots buffer format to DRM_FORMAT_ARGB8888.
I don't know enough about color formats to verify if that is actually correct.

See src/buffer.c buffer_create_cairo().

Edit 2: Actually, after reading though the linked wlroots issue: We do create the cairo surface with CAIRO_FORMAT_ARGB32 and then set the wlroots buffer format to DRM_FORMAT_ARGB8888. I don't know enough about color formats to verify if that is actually correct.

For what it's worth, most of the discussion on the wlroots issue (before Félix weighed in) ended up being irrelevant. My original theories were wrong, and the pixel/color formats themselves are fine.

The issue is really just that Pango/Cairo does not behave properly (and in some cases actually produces out-of-range RGB values, which is a bug) when doing subpixel rendering over transparency. The behavior could be improved (e.g. limit the RGB values to be no greater than the alpha value -- I tried that as a post-processing step, and it helped) but the result would still be less than ideal. The subpixel rendering really needs to be done over an opaque background to work correctly. That's my understanding, anyway.

For what it's worth, most of the discussion on the wlroots issue (before Félix weighed in) ended up being irrelevant. My original theories were wrong, and the pixel/color formats themselves are fine.

Well it was interesting nevertheless. Even though I will likely have forgotten everything within two weeks again.
The intimate details of color formats are just not my cup of tea, I prefer things to be slightly higher level.

The subpixel rendering really needs to be done over an opaque background to work correctly. That's my understanding, anyway.

Should we add the cairo issue within the comment?

Anyway, feel free to go for the rebase and merge if you are happy with the PR. You have invested way more time in debugging and testing the solution / workaround than all of us and are thus the most qualified person to judge it.

Just tested this on my laptop, and the difference is immediately visible. I always assumed the color artifacts were caused by resampling issues with fractional scaling.

Thanks. Fantastic work.

This adds a limitation to how we composite SSD, but that's a discussion for a different day.

For what it's worth, I'm increasingly thinking that that we ought to consider assembling multiple "parts" in one cairo-surface anyway. This needs to experimentation, speed-testing and a strategy. For example, I think we ought to consider just creating one surface for the whole titlebar (incl. corners).