buffer: improve copy() performance

[mscdex]

Results of included benchmark:

                                                               confidence improvement accuracy (*)   (**)   (***)
 buffers/buffer-copy.js n=6000000 partial='false' bytes=0            ***    580.96 %       ±3.05% ±4.06%  ±5.30%
 buffers/buffer-copy.js n=6000000 partial='false' bytes=8            ***    179.90 %       ±5.71% ±7.69% ±10.20%
 buffers/buffer-copy.js n=6000000 partial='false' bytes=128          ***    174.81 %       ±1.17% ±1.56%  ±2.05%
 buffers/buffer-copy.js n=6000000 partial='false' bytes=32768          *      3.62 %       ±3.30% ±4.44%  ±5.89%
 buffers/buffer-copy.js n=6000000 partial='true' bytes=0             ***    566.67 %       ±3.63% ±4.87%  ±6.42%
 buffers/buffer-copy.js n=6000000 partial='true' bytes=8             ***     24.66 %       ±1.17% ±1.57%  ±2.05%
 buffers/buffer-copy.js n=6000000 partial='true' bytes=128           ***     19.53 %       ±0.61% ±0.81%  ±1.06%
 buffers/buffer-copy.js n=6000000 partial='true' bytes=32768                 -0.52 %       ±0.98% ±1.31%  ±1.70%

Checklist

• make -j4 test (UNIX), or vcbuild test (Windows) passes
• tests and/or benchmarks are included
• commit message follows commit guidelines

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/24900/

[YurySolovyov]

any results for >32k?

[mscdex]

@YurySolovyov No, is there a particular value you have in mind that you would like to see?

[YurySolovyov]

I just wonder if it actually degrades for larger inputs

[mscdex]

FWIW I just tried with 1MB and it seems only then is there a slight regression (~1%) when doing a partial copy. The funny thing is that even deferring to C++ with no parameter/value validation (basically just a memset()) in that particular case, it slows things down significantly more than the original regression.

My guess is that there are missing optimizations in V8 for some larger values?

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/24959/

[Trott]

@nodejs/buffer

[Trott]

Landed in 6d351d4