AHashSet<usize> vs FnvHashSet<usize>

[eadf]

I've got an algorithm implementation I thought I should speed up by replacing FnvHashSet with AHashSet.
But the benchmark resulted in a +15% regression after the change:/

Is this to be expected? Could the problem be that I'm testing on a M1 Arm?

With AHashSet M1 (with +nightly-2021-05-10-aarch64-apple-darwin):

intersection_bench      time:   [38.696 ms 38.713 ms 38.730 ms]                               
                        change: [+15.275% +15.336% +15.395%] (p = 0.00 < 0.05)
                        Performance has regressed.

With FnvHashSet M1 (with +nightly-2021-05-10-aarch64-apple-darwin):

intersection_bench      time:   [33.544 ms 33.556 ms 33.571 ms]                               
                        change: [-13.369% -13.319% -13.269%] (p = 0.00 < 0.05)
                        Performance has improved.

I've got public snapshots if anyone wants to play along at home.

Edit:

Same test with x86_64 gives AHashSet a 2% advantage over FnvHashSet, but that's on a tired, heat-throttling laptop. So it's hard to draw any conclusions on those samples.

For good measure I ran the bench test in M1 again, and this time AHashSet repeatably wins by 10%, I'm utterly confused.

Update: Turns out the culprit was a rustup update that ran this morning. The numbers above was generated with an older compiler. With a fresh compiler I get these numbers:

With AHashSet M1 nightly-2021-05-30-aarch64-apple-darwin

intersection_bench      time:   [33.283 ms 33.298 ms 33.316 ms]                               
                        change: [-11.281% -11.221% -11.157%] (p = 0.00 < 0.05)
                        Performance has improved.

With FnvHashSet M1 (with +nightly-2021-05-30-aarch64-apple-darwin):

intersection_bench      time:   [37.606 ms 37.629 ms 37.655 ms]                               
                        change: [+12.910% +13.006% +13.102%] (p = 0.00 < 0.05)
                        Performance has regressed.

I quess we can close this Q now.