An initial look shows us spending over 1.5s in minor collection because of constant overflows of the generic store buffer. These appear to be for the DenseRange barriers for array_splice. An initial look: the majority of these are sourced from an object in the nursery, so can simply be dropped. Checking before inserting reduces our total time in the minor GC down to <0.2s, an order of magnitude improvement: nice. This changes our score from on pdf.js from 7636 to 5040, a decrease in performance of -34%. It seems this test is too expensive and we need to do something more sophisticated. This bug is for doing something more sophisticated, whatever that turns out to be.
Bug 990336 got us close to what we had pre-ggc, but there is still a bit more to do before we can claim parity.
You need to log in before you can comment on or make changes to this bug.