libc: scalar strlen() in RISC-V assembly

[strajabot]

Summary

Includes a scalar implementation of strlen() for the RISC-V
architecture and changes to the corresponding manpage.

Performance was benchamarked using before and after:
https://github.com/clausecker/strperf

os: FreeBSD
arch: riscv
        │ strlen_baseline │             strlen_scalar              │
        │     sec/op      │   sec/op     vs base                   │
Short        541.2µ ± 17%   401.6µ ± 0%  -25.78% (p=0.000 n=21+20)
Mid          249.6µ ±  3%   191.9µ ± 0%  -23.13% (p=0.000 n=21+20)
Long         124.6µ ±  0%   110.7µ ± 0%  -11.13% (p=0.000 n=21+20)
geomean      256.3µ         204.3µ       -20.26%

        │ strlen_baseline │              strlen_scalar               │
        │       B/s       │      B/s       vs base                   │
Short       220.3Mi ± 14%    296.8Mi ± 0%  +34.74% (p=0.000 n=21+20)
Mid         477.6Mi ±  3%    621.3Mi ± 0%  +30.09% (p=0.000 n=21+20)
Long        956.9Mi ±  0%   1076.7Mi ± 0%  +12.52% (p=0.000 n=21+20)
geomean     465.2Mi          583.4Mi       +25.40%

Test Plan

Passes the tests in the test suite

[fuz]

LGTM with indicated possibilities of improvement.

It might be possible to improve performance further by unrolling the main loop once or by processing 16 bytes per iteration.

lib/libc/riscv/string/strlen.S

25

Instead of loading -1 here, you could simply reuse the mask in t0. It doesn't matter what you or in, as long as it's not zeros.

[mhorne]

The new .S file needs a license.

I trust @fuz and the test suite to provide proper review of the assembly; more than I can.

What hardware was the benchmarking performed on? This is worth noting in the commit description.

lib/libc/riscv/string/Makefile.inc

2–5

[mhorne]

lib/libc/riscv/string/strlen.S

5–8

One more thing: we generally prefer C-style comments in assembly sources (/* comment */).

[strajabot]

Add license, remove unnecessary li, reformat comments

[mhorne]

Thanks a lot!

[fuz]

We will bulk commit these following an exp-run of the ports tree to give these changes some more field testing.