benchmark -C codegen-units=N

[llogiq]

Apparently, the default changed to codegen-units=16. However, when benchmarking I don't see a clear winner between N=1 or N=16 – it shows some small win on byte counting and some small loss on num_chars, but in both cases, it's a wash. Perhaps someone else can benchmark on a beefier machine?

[Deewiant]

I don't know what's "beefier" but I gave it an attempt with my i7-5930K @ 4.3 GHz. This is what cargo benchcmp says, excluding cases where the diff was either within 2 ns/iter or below 1.00%:

 name                               units1 ns/iter  units16 ns/iter  diff ns/iter  diff   speedup
 bench_count_00250_naive            69              74                          5   7.25%   x 0.93
 bench_count_01700_naive            463             458                        -5  -1.08%   x 1.01
 bench_count_02500_32               323             318                        -5  -1.55%   x 1.02
 bench_count_06000_32               764             748                       -16  -2.09%   x 1.02
 bench_count_09000_hyper            402             398                        -4  -1.00%   x 1.01
 bench_count_12000_32               1,521           1,490                     -31  -2.04%   x 1.02
 bench_count_12000_naive            3,231           3,191                     -40  -1.24%   x 1.01
 bench_count_17000_hyper            751             776                        25   3.33%   x 0.97
 bench_count_21000_hyper            919             964                        45   4.90%   x 0.95
 bench_count_25000_hyper            1,100           1,129                      29   2.64%   x 0.97
 bench_count_25000_naive            6,647           6,728                      81   1.22%   x 0.99
 bench_count_30000_hyper            1,305           1,369                      64   4.90%   x 0.95
 bench_num_chars_00900_naive        357             352                        -5  -1.40%   x 1.01
 bench_num_chars_01200_hyper        54              58                          4   7.41%   x 0.93
 bench_num_chars_01400_naive        543             549                         6   1.10%   x 0.99
 bench_num_chars_01700_naive        665             656                        -9  -1.35%   x 1.01
 bench_num_chars_02100_naive        819             808                       -11  -1.34%   x 1.01
 bench_num_chars_03000_naive        1,166           1,151                     -15  -1.29%   x 1.01
 bench_num_chars_07000_naive        2,707           2,673                     -34  -1.26%   x 1.01
 bench_num_chars_12000_naive        4,632           4,574                     -58  -1.25%   x 1.01
 bench_num_chars_17000_hyper        699             680                       -19  -2.72%   x 1.03
 bench_num_chars_21000_hyper        863             844                       -19  -2.20%   x 1.02
 bench_num_chars_21000_naive        7,993           8,098                     105   1.31%   x 0.99
 bench_num_chars_25000_hyper        1,024           1,001                     -23  -2.25%   x 1.02
 bench_num_chars_25000_naive        9,518           9,636                     118   1.24%   x 0.99
 bench_num_chars_30000_hyper        1,225           1,199                     -26  -2.12%   x 1.02
 bench_num_chars_30000_naive        11,418          11,584                    166   1.45%   x 0.99
 bench_num_chars_big_1000000_hyper  42,797          43,347                    550   1.29%   x 0.99

In short, it looks like a wash to me too. I also ran the comparison with the AVX options:

 name                               avx-units1 ns/iter  avx-units16 ns/iter  diff ns/iter   diff   speedup
 bench_count_00030_32               15                  12                             -3  -20.00%   x 1.25
 bench_count_00400_hyper            13                  16                              3   23.08%   x 0.81
 bench_count_01200_hyper            19                  22                              3   15.79%   x 0.86
 bench_count_01400_hyper            23                  26                              3   13.04%   x 0.88
 bench_count_01700_hyper            20                  24                              4   20.00%   x 0.83
 bench_count_02100_hyper            27                  30                              3   11.11%   x 0.90
 bench_count_02500_hyper            27                  30                              3   11.11%   x 0.90
 bench_count_02500_naive            395                 390                            -5   -1.27%   x 1.01
 bench_count_04000_hyper            35                  39                              4   11.43%   x 0.90
 bench_count_04000_naive            628                 620                            -8   -1.27%   x 1.01
 bench_count_05000_hyper            46                  50                              4    8.70%   x 0.92
 bench_count_05000_naive            787                 776                           -11   -1.40%   x 1.01
 bench_count_06000_32               477                 471                            -6   -1.26%   x 1.01
 bench_count_06000_hyper            54                  58                              4    7.41%   x 0.93
 bench_count_06000_naive            939                 928                           -11   -1.17%   x 1.01
 bench_count_08000_32               628                 621                            -7   -1.11%   x 1.01
 bench_count_08000_hyper            64                  69                              5    7.81%   x 0.93
 bench_count_08000_naive            1,253               1,238                         -15   -1.20%   x 1.01
 bench_count_09000_naive            1,410               1,393                         -17   -1.21%   x 1.01
 bench_count_10000_32               791                 780                           -11   -1.39%   x 1.01
 bench_count_10000_naive            1,563               1,545                         -18   -1.15%   x 1.01
 bench_count_12000_32               940                 929                           -11   -1.17%   x 1.01
 bench_count_12000_naive            1,876               1,854                         -22   -1.17%   x 1.01
 bench_count_14000_32               1,101               1,088                         -13   -1.18%   x 1.01
 bench_count_17000_naive            2,658               2,629                         -29   -1.09%   x 1.01
 bench_count_21000_32               1,645               1,626                         -19   -1.16%   x 1.01
 bench_count_21000_naive            3,282               3,244                         -38   -1.16%   x 1.01
 bench_count_25000_32               1,956               1,933                         -23   -1.18%   x 1.01
 bench_count_25000_naive            3,905               3,864                         -41   -1.05%   x 1.01
 bench_count_30000_32               2,348               2,324                         -24   -1.02%   x 1.01
 bench_count_big_0100000_hyper      1,095               1,071                         -24   -2.19%   x 1.02
 bench_count_big_1000000_hyper      25,438              24,621                       -817   -3.21%   x 1.03
 bench_num_chars_01200_naive        233                 230                            -3   -1.29%   x 1.01
 bench_num_chars_05000_naive        971                 960                           -11   -1.13%   x 1.01
 bench_num_chars_06000_naive        1,161               1,147                         -14   -1.21%   x 1.01
 bench_num_chars_10000_naive        1,933               1,910                         -23   -1.19%   x 1.01
 bench_num_chars_14000_naive        2,705               2,674                         -31   -1.15%   x 1.01
 bench_num_chars_17000_naive        3,287               3,253                         -34   -1.03%   x 1.01
 bench_num_chars_21000_hyper        238                 233                            -5   -2.10%   x 1.02
 bench_num_chars_25000_hyper        281                 277                            -4   -1.42%   x 1.01
 bench_num_chars_big_1000000_hyper  26,756              25,764                       -992   -3.71%   x 1.04

The differences are bigger but it still feels like noise to me.

[llogiq]

Thank you! That confirms my suspicion – there's either just not enough code to make a difference or thinLTO removes what difference is left.