(In reply to Kacper Michajłow [:kasper93] from comment #2)
Yes, your new build resolves the main issue. Here is the profile if you are interested https://perfht.ml/2RhF1cc
Do you have some reference benchmark you would like me to test against? To see if there are maybe more areas to improve?
Thanks for testing this, it's also been confirmed in another bug.
I don't have anything specific benchmark-wise, we have some of our own benchmark testing, but they didn't seem to pick this up.
The changes for bug 1464032 are actually for moving the accelerated canvas drawing to the GPU thread and are mainly behind a pref.
Obviously they do touch some existing code-paths and I mistakenly introduced this bug in a fairly late refactor during reviews.
Part of that was to avoid getting data back from the GPU process off the content main-thread, so I'll have to come up with a different way of doing that.