nx_bit.nx

nx_bit.nx -- low-level bit primitives. FOUNDATION LAYER. Every crypto module, every bitmap, every CRC-style checksum, every compressed-encoding parser eventually wants: popcount, count-leading-zeros, count-trailing-zeros, byte-swap, rotate-left/right. Today every module re-implements these via shift+mask loops. Centralising matches: * R1 cache-conscious: one canonical implementation per primitive means one inlining decision the optimiser sees everywhere * R2 attestable: SHA-256 + ChaCha20 + others share the same few primitives -- audit once instead of N copies * Future: lower to RV64 Zbb (bit-manipulation extension) when the target supports it; ~2-5x speedup over scalar fallback Bit primitives shipped: nx_popcount32 / nx_popcount64 Hamming weight nx_clz32 / nx_clz64 leading zero count nx_ctz32 / nx_ctz64 trailing zero count nx_bswap32 / nx_bswap64 byte-reverse nx_rotl32 / nx_rotl64 rotate left nx_rotr32 / nx_rotr64 rotate right nx_log2_floor log2 floor (= 63 - clz) nx_is_pow2 power-of-two check nx_align_up round up to align (must be POT) Algorithm references: * popcount: Hamming weight via parallel-prefix (folklore; Knuth TAoCP Vol 4A 7.1.3). Constant 12 ops for 64-bit. * clz / ctz: branch-free de Bruijn sequence multiply (Leiserson et al, "Using De Bruijn sequences to index a 1 in a computer word", 1998 unpublished). Constant 8-10 ops. * bswap: 4-step shuffle (Schneier, Practical Cryptography 2003). Forward path: when nxc2 grows the Zbb intrinsic, lower these to `cpop`, `clz`, `ctz`, `rev8`, `rori`, `rorw` instructions. Estimated 3-5x speedup on hot crypto / hashing paths.