WP38: SIMD
State: Landed · Done: 10 of 10 · Design note: docs/wp38-simd.md
Stages
- S0: the baselines (#517)
- S1: the
indexOfAnykernel and itsstd/textsurface - S2:
--cpu x86-64-v3(#660) - S3a and S3: the bar kernels measured, then
nish:simdwith its bar met (#1064) - S5: AES-GCM on AES-NI and PMULL
- S6a and S6b: byte-run kernels for JSON and Eisel–Lemire in
numberF64 - S7: crediting an offset index from the test that bounds it
- S9 and S10: SHA-256 on SHA-NI and ARMv8, ChaCha20 on AVX2 and NEON; CRC-32 on the same seam
- S8: the SSSE3 and NEON Adler-32 kernels, measured past their bar (#1079)
-
--cpulevels for AArch64 andx86-64-v4, each on its own measurement: measured, not built (#1432)
Next
Nothing is scheduled. x86-64-v4 is kept out of --cpu (decided 2026-11-12): it clears S2's bar against the baseline but gains nothing over x86-64-v3 (§7).
Not planned
The lexer on indexOfAny (S4), the structural index (S6c) and per-value cuts (S6d), both measured out of reach, target intrinsics.

