As already noted in the code comment, we all knew that the in memory
conversion to a PNG stream is a cumbersome technique. But it was worth
to remove the GTK dependency.
After finding out, that gdk_pixbuf_new_from_data can actually use a
GVariant and share its memory without allocating new memory, I poked
the cairo docs for a similar technique.
There is cairo_image_surface_create_for_data, which would achieve the
same, but cairo uses another color byte order. So we have to manually
copy and reassemble the bytes.
It may look like a big rise in complexity, but actually, it's straight
forward copying pixel by pixel and changing the byte's order.
And the gdk_pixbuf_to_cairo_surface is 33x faster.
When using clang, this will create a race condition between
clang/gcc builds. Coveralls will take the first one. As GCC and clang
are counting line coverage differently, coveralls regularly fails.
While running under valgrind, it may take over 30 cycles in the loop to
wait for a running GTestDBus instance. So we have to increase the
waiting time.
Waiting a whole second will change the signal/noise ratio and will only
fail, if there's really a problem with the GTestDBus usage.
And as we're sleeping 500us in every loop, we usually get a super fast
response, too.
-j introduces race conditions, which lets the find -name ... -delete
commands fail, because those can't find specific files, which got
removed in the meantime.
The xctx color field is a full duplicate of the settings logic.
Only logic included in xctx color fields, are the frame colors, which
fall back to the global frame setting. So, this required to handle it
directly in settings.c
We haven't been using libdbus for a while so there is no use to
linking it with the build.
Unfortunately there is still a dependency on it as we depend on the dbus
package to tell us where to put our service files.
Alpine is running with Musl libc and musl uses extended precision
doubles, while valgrind can't handle extended precision,
2.3 == atof("2.3") won't be true under valgrind.
And therefore the option retrieval methods *_get_double would fail.
Also we have to increase the test verbosity, as `SKIPm` doesn't print
the message when skipping the tests.
See: silentbicycle/greatest#85
