KarpelesLab / anyhash / 24334979488

88%

push

github

Optimize whirlpool, crc32, xxh3/xxh128, tiger throughput

whirlpool (132 → 307 MB/s, 2.3x):
  Unroll the 10-round key schedule and state update loops with 16 named
  scalar locals (K0..K7, s0..s7), eliminating (i+k)%8 modulo ops and the
  two 128-byte copy() calls per round. The compiler keeps all state in
  registers.

crc32 MSB-first (700 → 2854 MB/s, 4.1x):
  Implement slicing-by-8: process 8 bytes per iteration using 8
  precomputed 256-entry uint32 tables (computed from the existing single
  table at init), with a scalar tail loop for the remainder.

xxh3/xxh128 (4.1 → 5.3 GB/s, 28%):
  Add bounds-check-elimination hints (_ = input[63]; _ = secret[63]) in
  xxh3AccumulateStripe and xxh3ScrambleAcc, removing per-iteration BCE
  overhead that was 12% of profile time.

tiger (809 → 810 MB/s, marginal):
  Add _ = block[63] BCE hint in tigerCompress. Tiger's cost is dominated
  by the 8-table-lookup round function which is already inlined and
  at the algorithmic limit for pure Go.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

87 of 90 new or added lines in 4 files covered. (96.67%)

2891 of 3311 relevant lines covered (87.32%)

681.69 hits per line