zstd: Write table clearing in a way that the compiler recognizes

[greatroar]

Benchmark results on amd64 below. These do not take into account #701. They're for Go 1.19; Go 1.20 produces slightly better asm for the old code, but still produces pretty bad asm, esp. on 32-bit platforms.

See also golang/go#56954.

name                                 old speed      new speed       delta
Encoder_EncodeAllXML-8                283MB/s ± 1%    284MB/s ± 0%     ~     (p=0.026 n=30+20)
Encoder_EncodeAllSimple/fastest-8     111MB/s ± 0%    111MB/s ± 1%     ~     (p=0.011 n=28+20)
Encoder_EncodeAllSimple/default-8    78.4MB/s ± 1%   78.3MB/s ± 1%     ~     (p=0.572 n=30+19)
Encoder_EncodeAllSimple/better-8     65.9MB/s ± 1%   66.2MB/s ± 1%   +0.53%  (p=0.009 n=30+20)
Encoder_EncodeAllSimple/best-8       11.1MB/s ± 1%   11.6MB/s ± 3%   +4.42%  (p=0.000 n=27+28)
Encoder_EncodeAllSimple4K/fastest-8   911MB/s ± 1%    914MB/s ± 1%   +0.31%  (p=0.004 n=29+20)
Encoder_EncodeAllSimple4K/default-8  73.1MB/s ± 1%   73.6MB/s ± 1%   +0.67%  (p=0.000 n=29+20)
Encoder_EncodeAllSimple4K/better-8   60.5MB/s ± 1%   62.7MB/s ± 1%   +3.64%  (p=0.000 n=29+17)
Encoder_EncodeAllSimple4K/best-8     8.62MB/s ± 3%  10.11MB/s ± 1%  +17.24%  (p=0.000 n=30+27)
Encoder_EncodeAllHTML-8               133MB/s ± 1%    133MB/s ± 1%     ~     (p=0.101 n=30+19)
Encoder_EncodeAllTwain-8             84.8MB/s ± 1%   86.2MB/s ± 3%   +1.63%  (p=0.000 n=24+20)
Encoder_EncodeAllPi-8                62.6MB/s ± 1%   62.7MB/s ± 0%     ~     (p=0.102 n=30+20)
Random4KEncodeAllFastest-8           2.50GB/s ± 1%   2.50GB/s ± 1%     ~     (p=0.449 n=29+20)
Random10MBEncodeAllFastest-8         2.39GB/s ± 2%   2.52GB/s ± 6%   +5.23%  (p=0.000 n=27+20)

The speedup for the best encoder on short inputs is expected, since that spends much of its time clearing hash tables. The speedup for Random10MBEncodeAllFastest may have to do with stack allocations, given that allocs/op is always zero but alloc/op is as follows:

name                                 old alloc/op   new alloc/op    delta
Encoder_EncodeAllXML-8                  0.00B           0.00B          ~     (all equal)
Encoder_EncodeAllSimple/fastest-8       2.73B ±27%      3.00B ± 0%     ~     (p=0.018 n=30+18)
Encoder_EncodeAllSimple/default-8       4.00B ± 0%      4.00B ± 0%     ~     (all equal)
Encoder_EncodeAllSimple/better-8        5.00B ± 0%      5.00B ± 0%     ~     (all equal)
Encoder_EncodeAllSimple/best-8          19.5B ± 3%      19.0B ± 0%   -2.40%  (p=0.000 n=30+24)
Encoder_EncodeAllSimple4K/fastest-8     0.00B           0.00B          ~     (all equal)
Encoder_EncodeAllSimple4K/default-8     0.00B           0.00B          ~     (all equal)
Encoder_EncodeAllSimple4K/better-8      0.00B           0.00B          ~     (all equal)
Encoder_EncodeAllSimple4K/best-8        2.00B ± 0%      1.43B ±40%  -28.33%  (p=0.000 n=30+30)
Encoder_EncodeAllHTML-8                 2.37B ±27%      2.25B ±33%     ~     (p=0.398 n=30+20)
Encoder_EncodeAllTwain-8                0.00B           0.00B          ~     (all equal)
Encoder_EncodeAllPi-8                   12.4B ± 5%      12.2B ± 6%     ~     (p=0.283 n=30+20)
Random4KEncodeAllFastest-8              0.00B           0.00B          ~     (all equal)
Random10MBEncodeAllFastest-8           31.9kB ± 2%     30.5kB ± 9%   -4.27%  (p=0.002 n=28+20)

[klauspost]

This shouldn't really be called often enough to be measurable. I will check the cause of that.

[greatroar]

FWIW, I've seen this before in pierrec/lz4#169, so I wasn't too surprised. The size of the hash table for the best encoder is

(1<<22 + 1<<18) * 2 * sizeof(int32) = 34 MiB.

[klauspost]

Sure, but it shouldn't really be called that often.

[klauspost]

LGTM. Found a way to reduce resets somewhat. I will add that as well.