GitLab

fc: Fix an ordering issue

Closes #537

See merge request !290

With the recently introduced threading for fontconfig
calls, we introduced an ordering issue where the fontmap
may die before an outstanding thread returns, and then
the code that unrefs the pattern object tries to remove
it from the fontmap cache.

Prevent that by giving each thread a strong ref on the
fontmap while it runs.

Fixes: #537

Speed up font sorting

See merge request !221

Add sysprof marks around the expensive fontconfig
calls, and for when we are waiting on them.

With a big fontconfig configuration, FcInit takes some time
(60-100ms on my system). Doing this work off the main thread
can potentially avoid blocking other work.

To take advantage of this, GTK calls pango_cairo_font_map_get_default()
early to initiate the creation of a fontmap, thereby triggering the
FcInit() call.

fontconfig is thread-safe, so we can do calls that
can take several milliseconds to complete in a thread.
This patch does that for FcFontSetMatch and FcFontSetSort.
It is a win, at least for FcFontSetSort, since it
reduces the time that the main thread spends in
pango_fc_patterns_get_font_pattern for i > 0 from
around 10 to 6 milliseconds.

We have filtered-by-format lists of fonts available
now, so we should use them to ensure we always operate
on the same set of fonts. Also, fix another case of
passing NULL for the config.

The fontconfig configuration changes rarely, so we can avoid
unnecessary work by only doing the filtering by supported formats
once, and keeping the result.

Fix FcConfig uses

See merge request !288

Passing NULL for a config argument in a fontconfig
api means we are using the default configuration,
which may be different from the one the fontmap
is supposed to be using.

Fix an instance of this in pango_fc_face_list_sizes.

Passing NULL for a config argument in a fontconfig
api means we are using the default configuration,
which may be different from the one the fontmap
is supposed to be using.

Fix an instance of this in pango_fc_fontset_load_next_font.

Ci fixes

See merge request !286

Remove the dashes that cause the runs to break, and seem
to be causing issues on both Debian and newer Fedora.

Switch to Fedora 33, and add the few extra python
packages that gi-docgen needs.

Gir fixes

See merge request !285

PANGO_VERSION_MIN_REQUIRED is a build utility, and
is not useful as a constant in the gir. Leave it out.

PANGO_MAJOR/MINOR/MICRO_VERSION are useful to have
in the gir.

Add casts to get some defined constants to have
the right type in the gir. The games we play...

g-ir-scanner's handling of #defines is really rudimentary,
so we have to trick it to pick up PANGO_ATTR_TO_TEXT_END
as a constant with value 0, and then override the value
with an annotation.

See gobject-introspection#369

Revert "ci: Temporarily disable macOS CI job as runner is offline"

See merge request !284

This reverts commit 59cdceff.

The macOS runner is back.

The reprovisioned runner has Python 3.9 installed and includes
required change in its bash_profile.

docs: Remove references to gdk_pango_context_get_for_screen

See merge request !283

This reverts commit 59cdceff.

The macOS runner is back.

Remove references to a function that no longer exists in gtk.

ci: Temporarily disable macOS CI job as runner is offline

See merge request !282

See Infrastructure/GitLab#495

Fix font roundtrip

Closes #530

See merge request !280

Test that round-tripping through pango_font_describe
works. This is currently broken for scalable bitmap
fonts, such as color Emoji fonts.

We skip the test on OS X where we are most likely
missing the fonts.

We are using the size from the FcPattern. For scalable
bitmap fonts, this has been scaled to match the requested
pixel size. To make a font description that can be turned
back into a FcPattern and roundtrip successfully, we need
to undo that scaling. Thankfully, fontconfig leaves the
pixelsizefixupfactor in the pattern, so it is easy to do.

Fixes: #530

pango_cairo_font_map_get_default is transfer none.

tests: Avoid locale dependency

Closes #474

See merge request !279

When producing output, use the C locale for formatting
floating point numbers, to avoid locale dependent output.

Fixes: #474

Suggested by Philip Withnall.

Fixes: #415

Attr list overflow

Closes #455

See merge request !278

Avoid overflow when updating the end_index of
attributes in pango_attr_list_update. This is
a real risk, because end_index is commonly set
to G_MAXUINT to mean 'until the very end'.

Test included.

Fixes: #455