zstd: Less branching in asm version of adjustOffset

[greatroar]

We can do most of the work with conditional moves and we can handle the case offset != 0 with a single comparison against 2.

Also, shorten a conditional move in adjustOffsetInMemory from a MOVQ+CMOVQEQ to an ADCQ.

Benchmark results, Ivy Bridge:

                                       │     old      │                cmov              │
                                       │     B/s      │     B/s       vs base               │
Decoder_DecodeAll/kppkn.gtb.zst-8        447.7Mi ± 2%   445.0Mi ± 0%       ~ (p=0.355 n=20)
Decoder_DecodeAll/geo.protodata.zst-8    1.147Gi ± 0%   1.152Gi ± 1%  +0.39% (p=0.012 n=20)
Decoder_DecodeAll/plrabn12.txt.zst-8     354.0Mi ± 1%   351.6Mi ± 1%       ~ (p=0.076 n=20)
Decoder_DecodeAll/lcet10.txt.zst-8       424.9Mi ± 1%   420.2Mi ± 0%  -1.10% (p=0.000 n=20)
Decoder_DecodeAll/asyoulik.txt.zst-8     351.4Mi ± 1%   350.5Mi ± 0%       ~ (p=1.000 n=20)
Decoder_DecodeAll/alice29.txt.zst-8      350.5Mi ± 1%   349.5Mi ± 1%       ~ (p=0.525 n=20)
Decoder_DecodeAll/html_x_4.zst-8         1.437Gi ± 1%   1.440Gi ± 0%       ~ (p=0.607 n=20)
Decoder_DecodeAll/paper-100k.pdf.zst-8   4.193Gi ± 1%   4.219Gi ± 1%  +0.60% (p=0.007 n=20)
Decoder_DecodeAll/fireworks.jpeg.zst-8   8.878Gi ± 0%   8.798Gi ± 0%  -0.89% (p=0.000 n=20)
Decoder_DecodeAll/urls.10K.zst-8         596.6Mi ± 1%   596.0Mi ± 1%       ~ (p=0.337 n=20)
Decoder_DecodeAll/html.zst-8             933.2Mi ± 0%   936.1Mi ± 1%       ~ (p=0.659 n=20)
Decoder_DecodeAll/comp-data.bin.zst-8    394.0Mi ± 1%   390.9Mi ± 1%       ~ (p=0.149 n=20)
geomean                                  839.6Mi        837.4Mi       -0.26%

Tiger Lake (laptop):

                                        │     old      │                 cmov                 │
                                        │     B/s      │      B/s       vs base               │
Decoder_DecodeAll/kppkn.gtb.zst-16        408.2Mi ± 1%    416.9Mi ± 2%  +2.13% (p=0.000 n=20)
Decoder_DecodeAll/geo.protodata.zst-16    1.291Gi ± 1%    1.306Gi ± 2%       ~ (p=0.277 n=20)
Decoder_DecodeAll/plrabn12.txt.zst-16     332.9Mi ± 2%    339.3Mi ± 2%  +1.95% (p=0.030 n=20)
Decoder_DecodeAll/lcet10.txt.zst-16       394.1Mi ± 1%    398.6Mi ± 1%  +1.16% (p=0.006 n=20)
Decoder_DecodeAll/asyoulik.txt.zst-16     348.2Mi ± 1%    350.0Mi ± 0%       ~ (p=0.101 n=20)
Decoder_DecodeAll/alice29.txt.zst-16      320.1Mi ± 2%    319.5Mi ± 1%       ~ (p=0.265 n=20)
Decoder_DecodeAll/html_x_4.zst-16         2.109Gi ± 1%    2.138Gi ± 2%       ~ (p=0.068 n=20)
Decoder_DecodeAll/paper-100k.pdf.zst-16   5.258Gi ± 1%    5.291Gi ± 1%       ~ (p=0.121 n=20)
Decoder_DecodeAll/fireworks.jpeg.zst-16   12.40Gi ± 1%    12.54Gi ± 1%  +1.08% (p=0.004 n=20)
Decoder_DecodeAll/urls.10K.zst-16         553.6Mi ± 1%    563.5Mi ± 2%  +1.80% (p=0.000 n=20)
Decoder_DecodeAll/html.zst-16             987.2Mi ± 1%   1004.3Mi ± 2%       ~ (p=0.068 n=20)
Decoder_DecodeAll/comp-data.bin.zst-16    492.3Mi ± 2%    496.1Mi ± 1%       ~ (p=0.417 n=20)
geomean                                   907.9Mi         918.5Mi       +1.17%

[greatroar]

Hold on. I did something wrong with the benchmark figures on Ivy Bridge. Running some more.

[greatroar]

Corrected benchmark figures for Ivy Bridge posted above. The benefit of this is a bit dubious...

[klauspost]

Let me test it out. Generally I don't observe any benefit from fairly predictable CMOVs. But since you have older HW to test on we can take it in if there are no regressions.

[klauspost]

My benchmarks are far too noisy to see anything significant.