[HarfBuzz] Mark zeroing mess

[Behdad Esfahbod]

Hi Jonathan,

To my surprise, I have produced evince that, contrary to our previous belief,
Uniscribe does NOT zero mark advances in the default (eg. Latin) shaper based
on Unicode.  In fact, this is my observation:


  - With GDEF:
    - If mark is class=3 in GDEF:
      -> advance zeroed regardless of whether mark attached in GPOS or not.
    - If mark is class != 0 in GDEF:
      -> advance NOT zeroed (and mark attachment obviously doesn't apply.)

 - No GDEF:
   - Mark advance is NOT zeroed; and mark attachment doesn't apply.

So, it looks like:

  1. Uniscribe is NOT synthesizing any GDEF,

  2. It's zeroing mark advance purely based on GDEF.

I think our GDEF synthesis is still a good idea.  But like to change advance
zeroing to work based on GDEF.  This will fix the following bug:

  https://github.com/googlei18n/noto-fonts/issues/472#issuecomment-147528411

as well as a bug Martin and I discovered today, which is: currently GC=Mc
marks retain their advance, even if they are attached.  That can't be right
either...

The Thai and Tibetan shapers also use BY_UNICODE_LATE mark-zeroing.  I'm going
to test those and change as well, to follow Uniscribe.

Comments?  Can you possibly get your monster test suite running after this change?

Thanks,
-- 
behdad
http://behdad.org/