Post #1610612
2026-03-31 14:07 UTC
@menio @Sobex Those are for exact integer multiplication of double-precision mantissas. Maybe Karatsuba? But my bet is for crypto. Modular exponentiation, elliptic curve scalar multiplication, and polynomial multiplication[1] are usually bottlenecked by big-integer multiply; so two SIMD lanes doing 53x53-bit multiplies per cycle is way faster than scalar umulh/umull sequences.
[1] RSA, DH, GCM, and post-quantum lattice schemes requires that.
+
Replies (1)
-
@vito@ruby.social 2026-03-31 14:09
@menio@an.ti.social.br @Sobex@social.sciences.re That could make way more sense than just JavaScriptCore, as that would affect the Secure Enclave and kernel crypto paths (CommonCrypto/corecrypto), as squeezing more multiply throughput out of the NEON pipeline without going through the scalar integer unit could make a potentially big difference. 🤔