Nico Weber
Nico Weber
Seems like a legit fix, not like a hack. Together with clipping against the clip bounds, that should thoroughly fix this, right?
The 2nd commit in https://github.com/SerenityOS/serenity/compare/master...nico:serenity:deflate-peek?expand=1 does the first idea now. It's a 5% slow-down, but in return it's correct :/
With that change, UnsatisfiableReadBehavior can be removed – turns out webp doesn't need any special handling here. Regular deflate just happened to work before since it the superfluous bits came...
For fast deflate implementations, you normally do something like: * Peek 15 bits * Look into a table to see if the first ~8 bits are enough for at least...
#24567 uses the name `peek_bits_possibly_past_end()`, which seems possibly a good way forward.
Maybe one of you has full keyboard access enabled and the other doesn't?
I have half a local branch for this, I'll try to clean it up and publish it.
I remember intentionally not cherry picking this one a while ago, for the reasons you mention :)
> Note that Skia support quite a limited amount of color spaces, That sounds wrong to me: * IIRC I saw v4 code in skcms.cc (...maybe skia isn't configured right...
(FWIW, LTO is a much bigger win when combined with PGO. If you have just one of the two, PGO is a bigger win than LTO.)