Some CoreFoundation objects, such as those produced by CoreText, have
expensive callbacks that run when they are released. By offloading the
CFRelease calls to another thread, we can avoid important threads being
blocked by unexpectedly expensive callbacks.
This commit also changes the way that the coretext shaper's run iterator
builds its string. Rather than using a CFMutableString, an ArrayList of
unichars is built which is passed to CFStringCreateWithCharactersNoCopy,
which is a lot more efficient since it avoids all the CoreFoundation
overhead.
See: https://github.com/ziglang/zig/issues/17851
Users were noticing that frame render times got slower over time. I
believe (thanks to community for pointing it out) that this is the
culprit.
This works around this issue by clearing and reinitializing the LRU
after a certain number of evictions. When the Zig issue has a better
resolution (either rehash() as a workaround or a better hash
implementation overall) we can change this.
Now that we're padding the cells with blanks if we have shaped ligatures
we don't actually know the exact count based on the CoreText APIs, so we
should just dynamically add.
We previously split text runs for shaping on bg color changes. As
pointed out in Discord, this is not necessary, since we can always color
cells according to their desired background even if the text in the cell
shapes to something else.
Fixes#1708
Harfbuzz does this automatically. Our tests for harfbuzz test this. We
had a todo in CoreText to mimic this but wasn't sure if it was useful.
Turns out, it is important (see bug!)
Fixes#1664
I previously asserted that we got exactly one run from CoreText because
I assumed that our run iterator was perfectly splitting runs for
CoreText. This assumption appears to be false and that seems okay.
The test case in this commit produces two runs that are directly next to
each other and there's no downside to simply iterating over them. So
this commit changes to iterate over the runs.
