Add serde-1.0.219-threads4 benchmark for the parallel frontend

[Kobzol]

I tried to compile a few of our benchmarks with -Zthreads=4 and serde got the largest improvement that I saw, even though it's quite short (~2s > ~1s). I would like to experiment with adding a benchmark with the parallel frontend, to see how well the monitoring works. It's a bit tricky in that icounts, cycles and task-clock are actually larger in the 4threads version, but wall-time is better. We'll see how well that works in practice, I guess.

Note that the collector currently has six CPU cores, that's why I used -Zthreads=4 as a conservative value.

$ hyperfine --warmup 0 --runs 2 --prepare "rm -rf target" "cargo +nightly build" "RUSTFLAGS='-Zthreads=4' cargo +nightly build"
Benchmark 1: cargo +nightly build
  Time (mean ± σ):      2.170 s ±  0.039 s    [User: 1.962 s, System: 0.385 s]
  Range (min … max):    2.143 s …  2.198 s    2 runs
 
Benchmark 2: RUSTFLAGS='-Zthreads=4' cargo +nightly build
  Time (mean ± σ):      1.249 s ±  0.079 s    [User: 2.528 s, System: 0.453 s]
  Range (min … max):    1.193 s …  1.304 s    2 runs
 
Summary
  RUSTFLAGS='-Zthreads=4' cargo +nightly build ran
    1.74 ± 0.11 times faster than cargo +nightly build

[lqd]

The parallel front end is not even tested on CI so benchmarking may be a bit early?

[Kobzol]

It is definitely an experiment. If something goes wrong, it's on @saethlin (#t-compiler/parallel-rustc > Simple parallel benchmark in rustc-perf @ 💬)... 😆

[nnethercote]

Seems fine except the 4threads in the name reads really strangely to me, probably because it's not a valid identifier. What about threads4 instead? That also matches -Zthreads=4 more closely.

[Kobzol]

Sure, renamed to -threads4.

[Kobzol]

Let's try and see what happens in practice. If this causes issues, we'll revert.