An operator precedence error from bug 605043 means that the y-axis glyph offsets are not correctly inverted, so GPOS offsets get applied in the wrong direction when the HarfBuzz shaper is enabled on Linux.

I don't understand why the bug is not showing up in the same way on other platforms. With GDI and DWrite, fixing the error changes the positions in layout/reftests/text/475092-pos.html by only one pixel, which makes it pass. (The reftest still doesn't pass on Linux or Mac - see bug 482596 - but is much closer on Linux at least.)

(In reply to comment #1) > Created attachment 493586 [details] [diff] [review] > patch > > I don't understand why the bug is not showing up in the same way on other > platforms. It does... I just noticed it a couple of days ago on OS X, when testing a complex Arabic-script font that used to work; was intending to file and investigate, but you got there first - thanks!

We should add a testcase to catch this, especially as y-axis sign errors are so easy to make, and not always obvious at first because vertical deltas are rarely used. I think we can do something using @font-face with a font that's known to support vertically-stacked marks...

Here are a couple of reftests using the Scheherazade font that we have in the reftest fonts directory; they will fail if we make a sign error handling the glyph y-offsets.

Comment on attachment 493667 [details] [diff] [review] reftests for direction of vertical glyph offsets Nice, thank you!

Pushed this to trunk: http://hg.mozilla.org/mozilla-central/rev/3e3726fc8083 (patch) http://hg.mozilla.org/mozilla-central/rev/a5aafda069f7 (tests)