add simple allocation benchmarks

[vilterp]

cc @vchuravy @NHDaly

A simple-as-possible benchmark to help us see the impact of JuliaLang/julia#42768, JuliaLang/julia#43868 etc.

Was also thinking of using @timed and asserting against gc_stats.poolalloc to make sure we're actually measuring pool allocations.

Here it is running on JuliaLang/julia#43868:

julia> res_pr["strings"]
BenchmarkTools.Trial: 266 samples with 1 evaluation.
 Range (min … max):  16.121 ms … 42.947 ms  ┊ GC (min … max): 0.00% … 16.64%
 Time  (median):     16.783 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   18.576 ms ±  3.231 ms  ┊ GC (mean ± σ):  8.38% ± 10.70%

   ▅█▃                                                         
  ▆████▇▄▄▃▁▂▂▁▁▂▂▁▁▁▁▂▁▁▁▁▁▂▁▁▁▁▁▂▁▃▁▄▅▄▅▆▄▃▃▃▃▃▁▁▁▁▂▁▂▂▁▁▁▂ ▃
  16.1 ms         Histogram: frequency by time        25.3 ms <

 Memory estimate: 25.17 MiB, allocs estimate: 549000.

julia> res_pr["arrays"]
BenchmarkTools.Trial: 1504 samples with 1 evaluation.
 Range (min … max):  2.628 ms … 35.518 ms  ┊ GC (min … max): 0.00% … 76.78%
 Time  (median):     2.842 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   3.282 ms ±  1.582 ms  ┊ GC (mean ± σ):  9.75% ± 14.90%

  █▇▇▅▄▃▂▂ ▁                                     ▁            
  ████████████▇▆▅▅▅▆▁▄▅▄▅▅▄▁▁▁▁▁▄▄▁▄▁▄▁▄▁▄▁▁▁▁▅▇███▆▆▄▅▅▅▅▄▆ █
  2.63 ms      Histogram: log(frequency) by time     8.66 ms <

 Memory estimate: 5.35 MiB, allocs estimate: 100100.

julia> res_pr["structs"]
BenchmarkTools.Trial: 5105 samples with 1 evaluation.
 Range (min … max):  557.607 μs … 16.754 ms  ┊ GC (min … max):  0.00% … 93.99%
 Time  (median):     683.703 μs              ┊ GC (median):     0.00%
 Time  (mean ± σ):   963.428 μs ±  1.185 ms  ┊ GC (mean ± σ):  24.99% ± 17.89%

  █▇▆▄▂                                                        ▁
  ██████▇▇▅▃▄▅▃▁▃▄▃▅▁▁▃▁▁▁▁▃▁▁▁▁▁▃▁▁▁▁▁▁▁▃▁▃▁▆▇▇▆▇▆▆▆▆▇▇▇▇▆▆▇▅ █
  558 μs        Histogram: log(frequency) by time       6.8 ms <

 Memory estimate: 3.83 MiB, allocs estimate: 100100.

julia> res_pr["grow_array"]
BenchmarkTools.Trial: 1831 samples with 1 evaluation.
 Range (min … max):  2.166 ms … 19.977 ms  ┊ GC (min … max):  0.00% … 86.57%
 Time  (median):     2.353 ms              ┊ GC (median):     0.00%
 Time  (mean ± σ):   2.707 ms ±  1.839 ms  ┊ GC (mean ± σ):  11.24% ± 13.48%

  █▇▄▂                                                        
  ████▆▆▅▆▃▁▃▃▁▁▁▁▁▁▁▁▁▃▁▁▁▃▁▁▁▁▁▁▁▃▁▅▁▄▁▁▁▃▁▃▄▃▅▃▁▄▅▅▄▆▃▁▃▄ █
  2.17 ms      Histogram: log(frequency) by time     13.6 ms <

 Memory estimate: 2.57 MiB, allocs estimate: 48911.

[vtjnash]

Can we add a GC.safepoint() call and Core.Intrinsics.atomic_fence() (if available) also, so that the compiler is not permitted to defer the assignment to the loop footer and then skip the loop

[vilterp]

here's a gist of the results, running on JuliaLang/julia#43868 vs master. Shows no regressions

https://gist.github.com/vilterp/2444a133822427238e46eb23f5b60925

(though honestly running on my laptop is not consistent)

[vilterp]

Moved some work to inner loops to offset the impact of calling Threads.atomic_fence()

[NHDaly]

This LGTM. Thanks for setting it up! :) 👍 👍

[NHDaly]

Excellent, thanks!