1251488 - ArenaHeader bool bit to record whether the arena contains any weakmap keys

Status: ASSIGNED

Description

[Steve Fink [:sfink] [:s:]]

Rather than looking up every possible weak key while doing the weak marking phase, we could set a bit on the arena saying whether it contains any weakmap keys.

Comment 1

[Steve Fink [:sfink] [:s:]]

Created attachment 8723888 [details] [diff] [review]
ArenaHeader bool bit to record whether the arena contains any weakmap keys

I implemented this, and saw basically no difference in timing. I even added measurements of how many times we stopped after checking the bit, and it happened 20 million times (out of about 20.1 million total lookups).

Without optimization: 1:52.81 (1 minute 52 seconds)
With optimization: 1:50.87

saves=20468588 hits=13870 misses=132400
  saves: the arena bit was unset
  hits: it was a weak map key
  misses: arena bit was set but it wasn't a weak map key

Perhaps the lookup is fast? Or perhaps mapping a GCCellPtr to an arena header is slow? I don't know. Leaving this here for posterity.

Comment 2

[Jon Coppeard (:jonco)]

That's a shame.  I guess lookup is just really quick...

Comment 3

[Steve Fink [:sfink] [:s:]]

Created attachment 8724151 [details]
weakstress.js

For the record, this is the benchmark test script I used to measure the behavior. It's a little crazy -- it creates a bit of a tangled-up graph and then GCs. Although I reported the total times above, I was mostly looking at the 'Weak Marking' output in MOZ_GCTIMER. And there was a lot of time being spent in there during the graph construction.

For the final GCs, I see:

original:

          Mark Weak: 2205.529ms
          Mark Weak: 2209.159ms
 
with optimization:

          Mark Weak: 2131.709ms
          Mark Weak: 2179.285ms

So (assuming these single runs are to be believed; I haven't looked at the variance) it's a 3% improvement on a workload that's heavy on weak marking, but only for this single phase.