Iterate in reverse when comparing Stacks

[saethlin]

I was originally going to propose this after I (eventually, sorry) am done with #1935 but @Nilstrieb independently identified this code path as hot and came up with a different optimization, which is basically this: rust-lang/rust#49892 (comment)

Anecdotally, I've found this optimization to be much more effective but of course that is probably workload-dependent.

[RalfJung]

Thanks, this makes sense!

Bummer that we don't have perf tracking infrastructure... I'll just trust that this works.

[saethlin]

We could also wait for @Nilstrieb to confirm that this is at least as impactful as they optimization they were considering.

[RalfJung]

Well, have you tested this PR and confirmed that it helps on some benchmarks?

[Noratrieb]

Hmm, that is interesting. Running the serde1 benchmark with hyperfine results in this:

faster-tag-partial-eq (local)
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):      5.644 s ±  0.216 s    [User: 5.559 s, System: 0.078 s]
  Range (min … max):    5.542 s …  6.244 s    10 runs

master
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):      6.160 s ±  0.069 s    [User: 6.078 s, System: 0.076 s]
  Range (min … max):    6.080 s …  6.313 s    10 runs

faster-stack-cmp
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):      6.202 s ±  0.087 s    [User: 6.114 s, System: 0.080 s]
  Range (min … max):    6.042 s …  6.332 s    10 runs

[Noratrieb]

Even more interesting, I ran the mse test as well

faster-tag-partial-eq
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):     29.984 s ±  0.170 s    [User: 29.650 s, System: 0.314 s]
  Range (min … max):   29.788 s … 30.355 s    10 runs

master
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):     34.884 s ±  0.130 s    [User: 34.583 s, System: 0.294 s]
  Range (min … max):   34.743 s … 35.135 s    10 runs

faster-stack-cmp
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):     35.406 s ±  0.128 s    [User: 35.090 s, System: 0.311 s]
  Range (min … max):   35.227 s … 35.634 s    10 runs

[RalfJung]

So on its own this seems to make it worse, if it has an effect at all.
Does rebasing on top of #2218 change anything?

[Noratrieb]

Rebasing the two and comparing them with master again

both serde1
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):      5.260 s ±  0.106 s    [User: 5.158 s, System: 0.079 s]
  Range (min … max):    5.201 s …  5.554 s    10 runs

both mse
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):     30.479 s ±  0.226 s    [User: 30.155 s, System: 0.317 s]
  Range (min … max):   30.340 s … 31.095 s    10 runs

master serde1
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):      6.152 s ±  0.062 s    [User: 6.071 s, System: 0.074 s]
  Range (min … max):    6.069 s …  6.268 s    10 runs

master mse
Benchmark 1: cargo +miri miri run
  Time (mean ± σ):     35.022 s ±  0.408 s    [User: 34.702 s, System: 0.313 s]
  Range (min … max):   34.658 s … 36.133 s    10 runs

[RalfJung]

Those are nice wins, but what I meant is comparing this PR with master after #2218 landed. The wins you are showing might be entirely due to #2218?

[saethlin]

Based on some experimentation I just did on my end, I think the thing to do here is land a few of my local test cases as bench-cargo-miri programs as well as land #1935 then re-evaluate micro-optimizations in this area.

[saethlin]

I'll re-open this if it still seems interesting after making all the decisions on #2315

[saethlin]

Since #2315, the stack comparison code path is now optimized to a memcmp, and in combination with fact that we are comparing less bits makes this code path a lot less interesting. Awesome!