(In reply to Ionuț Goldan [:igoldan], Performance Sheriffing from comment #3)
I suspect this is the same issue we had recently: assorted-dom appears to be bimodal, sometimes it regresses for no apparent reason and then later it goes back to normal again.
How certain are you about this? I looked over the graphs (from Linux PGO builds), but they don't seem to have returned to normal.
Very certain. I investigated this more and I think we should WONTFIX/INVALID this because:
(1) Bug 1527843 should only affect things involving |new| or cross-realm functions and these tests don't have anything like that. It's just plain getters that are inlined and optimized to the point that we have an (almost?) empty loop (it's very sensitive to random code/cache changes).
(2) When this landed cross-realm functions weren't enabled for content code.
(3) We've seen very similar assorted-dom regressions and improvements lately. I was wondering about OPT and PGO both regressing, but that happened also in bug 1524482 for example.
Also note that the worker-getname-performance-getter subtest regressed by 18.33% on the "opt" build according to comment 0. In bug 1524482, which turned out to be a temporary regression/fluke, we had an 18.48% regression so it's the same order of magnitude as well.