[aat] Tamil shaping fails to combine glyphs

[ebraminio]

I guess something is wrong here, https://chromium-review.googlesource.com/c/chromium/src/+/1275945/33/third_party/WebKit/LayoutTests/platform/mac-mac10.10/fast/text/international/complex-character-based-fallback-expected.png

I don't have access to 10.10-11 where the issue shows itself more but I see something on 10.13 also,

bash-3.2$ util/hb-view /Library/Fonts/Tamil\ MN.ttc --shaper coretext தொகுக்கப்பட்டு

bash-3.2$ util/hb-view /Library/Fonts/Tamil\ MN.ttc --shaper ot தொகுக்கப்பட்டு

Maybe two separate issues, mort on first and morx on second.

[drott]

bash-3.2$ util/hb-view /Library/Fonts/Tamil\ MN.ttc --shaper ot தொகுக்கப்பட்டு

Do you see the same result when you use --font-funcs=ot

[ebraminio]

Oh, font funcs let me see

[ebraminio]

btw can you see which font on 10.10 and 10.11 is used actually on that test, it shouldn't be Latha, or, is it? Exact version will helpful I guess

[drott]

I'll try to figure out which font is used on the 10.10 bot. On 10.13 the browser fallback code, calling AppKit finds "/Library/Fonts/Tamil Sangam MN.ttc", in which it looks correct and not disjointed.

[ebraminio]

No difference with --font-funcs=ot

[drott]

However, we do have a reproduction on 10.13 with /Library/Fonts/Tamil\ MN.ttc - I am assuming Tamil Sangam MN might not be available on the older bots / Mac OS versions.

[ebraminio]

Ah, I didn't notice it was in commit queue :/ I wished it was just merged.

[drott]

Well, emotionally, I would agree :) But it's better to work these out first instead of facing a revert or these issue reaching users. Good that you spotted these.

[drott]

I ran a short simulation of the fallback code on Mac OS 10.10 and it comes back with
"Tamil Sangam MN" on the 10.10 bot as well.

[drott]

With the Tamil Sangam MN font from the 10.10 bot:

CoreText:

$ ../test/shaping/hb-unicode-encode U+0BA4,U+0BCA,U+0B95,U+0BC1,U+0B95,U+0BCD,U+0B95,U+0BAA,U+0BCD,U+0BAA,U+0B9F,U+0BCD,U+0B9F,U+0BC1 | ./hb-shape ~/dev/tamil_sangam/Tamil_Sangam_MN_10_10.ttc --shaper=coretext
[tgm_e=0+734|tgc_ta=0+678|tgm_aa=0+534|tgc_ku=2+845|tgc_k=4+709|tgc_ka=6+709|tgc_p=7+585|tgc_pa=9+585|tgc_tt=10+763|tgc_ttu=12+669]

HB AAT:

$ ../test/shaping/hb-unicode-encode U+0BA4,U+0BCA,U+0B95,U+0BC1,U+0B95,U+0BCD,U+0B95,U+0BAA,U+0BCD,U+0BAA,U+0B9F,U+0BCD,U+0B9F,U+0BC1 | ./hb-shape ~/dev/tamil_sangam/Tamil_Sangam_MN_10_10.ttc --shaper=ot --font-funcs=ot
[tgm_e=0+734|tgc_ta=0+679|tgm_aa=0+492|tgc_ka=2+710|tgm_u=2+558|tgc_ka=4+710|tgm_pulli=4+234|tgc_ka=6+710|tgc_pa=7+585|tgm_pulli=7+234|tgc_pa=9+585|tgc_tta=10+773|tgm_pulli=10+234|tgc_tta=12+773|tgm_u=12+558]

[drott]

Images

CoreText

HB AAT

[behdad]

Ok, this is an easy and hard one at the same time. The font has broken GSUB/GPOS tables. So, it's our priority that is getting in the way. At the same time, I'm not sure I want to switch the priority around.

So, if a font has both tables, it will behave differently on Mac compared to other OSes. The Mac behavior might still be inferior to the GSUB/GPOS behavior but have been done because older Mac didn't support full Indic2... So, by always prefering Mac tables we will be doing such fonts a disservice, even though that's what you'd get with them on Mac still.

Not sure. @jfkthame ?

[behdad]

Other idea was to prefer AAT for Apple fonts. But checking OS/2 vendor ID for Tamil Sangam MN, it's "MUTF". So, that wouldn't work.

Next idea, we can blacklist their GSUB/GPOS tables same way we blacklist GDEF table of many fonts. I like this best. I just need the GSUB/GPOS/GDEF lengths of affected fonts. Maybe someone can run a check to see which fonts on OS X 10.10 have both sets of tables.