• Some tests fail on Travis because the font cache is not reset. I will manually delete the font cache for a CI re-run if there is general agreement over the PR.

• Supporting multiple fonts from a single ttc file is going to be a bit complicated to do properly, not so much on the ft2font/_ft2 side (it's just passing an additional index to the ctor), but because many places right now assume that you can use a font's pathname as cache key.

Could one use something like pathname[1] as caching key if there are multiple fonts in a file?

I'd rather just do it properly and switch everything to (pathname, index) (which is the standard representation FreeType uses, see https://www.freetype.org/freetype2/docs/reference/ft2-base_interface.html#FT_New_Face (pathname, face_index)). But that's out of scope for this PR.

What's blocking? May I help here?

We need a second positive review from one of the matplotlib reviewers before the PR can be merged.

@anntzer Travis Python 3.7 now fails with:

________________________________ test_find_ttc _________________________________
[gw1] linux -- Python 3.7.0 /home/travis/virtualenv/python3.7.0/bin/python
    @pytest.mark.xfail(not os.environ.get("TRAVIS"), reason="Font may be missing.")
    def test_find_ttc():
        fp = FontProperties(family=["WenQuanYi Zen Hei"])
        font = findfont(fp)
>       assert os.path.basename(font) == "wqy-zenhei.ttc"
E       AssertionError: assert 'DejaVuSans.ttf' == 'wqy-zenhei.ttc'
E         - DejaVuSans.ttf
E         + wqy-zenhei.ttc
lib/matplotlib/tests/test_font_manager.py:106: AssertionError

This has to be investigated before merging. Is the font missing there?

This has to be investigated before merging. Is the font missing there?

That's because I need to temporarily force the use of new font caches. Added a second commit that does so; then we can later revert that commit.

Some of the metadata seems to end up wrong when using this. FT2Font.get_sfnt() gives back garbage on most (with a few odd exceptions) .TTC fonts. This makes certain fonts inaccessible - in my specific case, I'm asking for Microsoft YaHei and ending up with Microsoft YaHei light, and there's no way to specifically get the normal weight version.

I guess that may be because YaHei has the light font at index 0 and the regular one at index 1? (I don't have access to it.)
In any case this PR never pretended to support fonts after the first one.

No, light and normal are separate TTCs (the whole point of .ttc is that it efficiently stores fonts that have many of the same glyphs; packing completely different versions of the same font doesn't do you any good.) The problem is that MatPlotLib can't load their metadata aside from the font family and sees them as identical.

YaHei should come preinstalled on every Windows machine with 8 or later. I see similar issues with most .ttcs on my system, though maybe it's something unique Microsoft is doing. I only saw it work correctly with non-CJK fonts though I imagine that's coincidence.

Oh, I see. I think it's basically the same issue as #8550? (in which case once again it's orthogonal to ttc support)

No; it's the same end result, but in that case the issue was failure to correctly interpret the metadata, whereas here the metadata comes back as junk. The font is named "YaHei Light;" if it could read the name it would assign it the correct weight, it just can't read the name.

With a conditional breakpoint on font_manager.py:342 with the condition set to 'ttc' in font.fname, I see sfnt full of unreadable bytes, as opposed to when it loads a .ttf font where it's full of ASCII strings. Do you not see this same behavior?

No, the family name (as loaded by FreeType into the family_name field of FT_FaceRec) is 'Microsoft YaHei'.
Now there are additional names that are available via the sfnt table mechanism, but we don't load them because they are in a bunch of random encodings (specifically, here they are in some MS encodings which are UTF-16-BE) (https://github.com/anntzer/freetypybind/blob/master/src/_ft2.cpp#L317 for some common cases...). For fonts that have sfnt entries encoded as mac_roman (that's the 1, 0, 0, 2/4 at the top of the function) (I guess these are the ttfs you loaded) we load that metadata (and even then we don't use the family name entry from it), but otherwise we don't bother.

Ah, I get it now. Okay, this is a Microsoft being weird issue, not a .ttc issue. Weird how it worked before the .ttc switch, but it's not your problem. Thanks for the info.

@anntzer Looks like you need to do a commit/revert dance on this PR because of the cache? If so, I think you can self-merge, and then do the revert commit soon after? (not sure how big a deal it is to leave master in an uncahched state)

I'll open a PR to restore the cache (that's what tests that the cache actually works...) immediately after this is merged (e.g. by you? :)).

Sure, I just wanted to make sure you were around to do it. If now is good, here it goes...