The reftest zwnj-01.xhtml fails on my Ubuntu 10.04 & 10.10a2 machines. (using a mozilla-central debug build) Reftest log is attached. The difference is in the 3rd line, with a ligature (I think) a little over halfway through the line. It's a very subtle difference -- only 3px according to the reftest output -- and it isn't visually noticeable unless you zoom in a *lot*.

Note that this reftest also apparently fails on OS X (it's marked as random-if-cocoa in reftest.list), and it also fails on WinCE Tegra (bug 514008 -- though in that case, it appears to be due to a lack of fonts on the device).

We already apply an SVG filter to the testcase to minimize differences in anti-aliasing. It looks as if some of the pixels here are still falling between two stools -- as it happens Jonathan was just telling me last week that that can happen occasionally.

Hmmmm... looking at the reftest log, it seems the SVG filter has NOT been applied successfully in this case, as it still shows all the antialiasing as many levels of gray. It should have been filtered to pure black-and-white.

Does dholbert have local changes that would break the SVG filter?

(In reply to comment #4) > Does dholbert have local changes that would break the SVG filter? Nope -- this is with a stock mozilla-central build. (In reply to comment #3) > Hmmmm... looking at the reftest log, it seems the SVG filter has NOT been > applied successfully in this case, as it still shows all the antialiasing as > many levels of gray. It should have been filtered to pure black-and-white. Actually, the filter *does* get applied, at least partially -- I confirmed that it converts text to black text. (I tried setting 'color: red' on the text in zwnj-01.xhtml, and it still produced the same snapshot with black text in the reftest output. And then if I remove the filter, I get red text in the output.) So, the filter's being applied and affects colors correctly, but it does _not_ affect shades of gray in the way that we're expecting it to.

(In reply to comment #5) > Actually, the filter *does* get applied, at least partially -- I confirmed that > it converts text to black text. oops, I meant to say "converts _red_ text to black text".

This is likely fixed by the patch in bug 581081.

(In reply to comment #5) > So, the filter's being applied and affects colors correctly, but it does _not_ > affect shades of gray in the way that we're expecting it to. As I pointed out in bug 581081, this is because the filter doesn't touch the alpha channel, so using it the way this test does requires putting a background on what's inside the filter; that's what the patch in bug 581081 does.

Should be fixed by landing of bug 581081, but would be good if you could confirm.

Confirmed fixed with current mozilla-central trunk (on an Ubuntu 10.04 machine). Thanks!