I'm seeing a few regressions with baseline interpreter enabled for an ARM64 geckoview_example on the Pixel 2. I've attached the websites which are part of the tp6m p1 list. Unfortunately, I probably won't be able to provide much more information other than a gecko profile until I finish bug 1565788. I have not tried ARM32 yet.

Jan, would be able to look into this?

Denis, can you say more about how you measured this exactly?

We talked about this a bit on IRC. javascript.options.blinterp = false will disable all JITs now so this could be an Ion compile time issue or something...

I'm also working on some low-level interpreter loop optimizations for ARM64 in bug 1576567.

(In reply to Jan de Mooij [:jandem] from comment #3)

We talked about this a bit on IRC. javascript.options.blinterp = false will disable all JITs now so this could be an Ion compile time issue or something...

I'm also working on some low-level interpreter loop optimizations for ARM64 in bug 1576567.

Thanks for the clarification. Doing a more extensive run, it does indeed seem like these regressions are actually caused by a mix of baselinejit and Ion. I will open new bugs specific to those and close this one.

(In reply to Denis Palmeiro [:denispal] from comment #4)

Doing a more extensive run, it does indeed seem like these regressions are actually caused by a mix of baselinejit and Ion. I will open new bugs specific to those and close this one.

That's a relief :) Thanks for the update! For what it's worth, bug 1576567 just landed and it seems to be a small perf improvement when running the Baseline Interpreter on ARM64.