improve chunk throughput with elide array bound checks

[wjiec]

Gain about ~3% speedup by inlining the slide method and reducing one assignment statement.

goos: linux
goarch: amd64
pkg: github.com/restic/chunker
cpu: Intel(R) Xeon(R) Platinum 8269CY CPU @ 2.50GHz
                    │   old.txt    │               new.txt               │
                    │    sec/op    │    sec/op     vs base               │
ChunkerWithSHA256-8    158.7m ± 0%    154.3m ± 0%  -2.79% (p=0.000 n=10)
Chunker-8              65.72m ± 1%    61.49m ± 0%  -6.44% (p=0.000 n=10)
NewChunker-8           136.4µ ± 9%    131.3µ ± 4%       ~ (p=0.063 n=10)
PolDivMod-8            2.877n ± 0%    2.876n ± 0%       ~ (p=0.490 n=10)
PolDiv-8               2.890n ± 0%    2.891n ± 0%       ~ (p=0.952 n=10)
PolMod-8               2.894n ± 1%    2.894n ± 0%       ~ (p=0.809 n=10)
PolDeg-8              0.3148n ± 0%   0.3144n ± 0%       ~ (p=0.075 n=10)
RandomPolynomial-8     2.017m ± 3%    2.054m ± 4%       ~ (p=0.684 n=10)
PolIrreducible-8       1.146m ± 0%    1.146m ± 0%       ~ (p=0.971 n=10)
geomean                6.634µ         6.550µ       -1.28%

                    │   old.txt    │               new.txt               │
                    │     B/s      │     B/s       vs base               │
ChunkerWithSHA256-8   201.7Mi ± 0%   207.4Mi ± 0%  +2.87% (p=0.000 n=10)
Chunker-8             486.9Mi ± 1%   520.4Mi ± 0%  +6.88% (p=0.000 n=10)
geomean               313.4Mi        328.6Mi       +4.86%

[MichaelEischer]

The change to use

		out := win[wpos%windowSize]
		win[wpos%windowSize] = b

yields the following speedup for me:

goos: linux
goarch: amd64
pkg: github.com/restic/chunker
cpu: AMD Ryzen 5 5600X 6-Core Processor
name                  old time/op   new time/op   delta
ChunkerWithSHA256-12   56.0ms ± 0%   54.5ms ± 0%  -2.60%  (p=0.000 n=8+9)
Chunker-12             41.5ms ± 0%   40.0ms ± 0%  -3.58%  (p=0.000 n=9+9)

name                  old speed     new speed     delta
ChunkerWithSHA256-12  599MB/s ± 0%  615MB/s ± 0%  +2.67%  (p=0.000 n=8+9)
Chunker-12            809MB/s ± 0%  840MB/s ± 0%  +3.71%  (p=0.000 n=9+9)

The other changes have no measurable effect. Regarding readability I'd prefer to keep the slide() method. Thus, please limit the changes in the PR to the just mentioned code snippet and update the corresponding comment to read // limit wpos to elide array bound checks

[wjiec]

done

[MichaelEischer]

LGTM. Thanks a lot for discovering the performance improvement. And sorry for completely dissecting your PR.

I'm not sure when I'll get around to take a look at the polynomials PR.

[wjiec]

Cool. It's all about making the program better. 😁