SIMD vectorization, io_uring, and syscall batching (2x throughput with Claude Code Opus 4.6)#354

Open

slartibardfast wants to merge 49 commits into

wangyu-:branch_libevfrom

slartibardfast:branch_libev

Commits on Feb 20, 2026

Add benchmark/test harness and CI for FEC and CRC32 vectorization

connollydavid
and
claude
committed
Replace CRC32 (zlib) with CRC32C (Castagnoli) in production packet path

connollydavid
and
claude
committed
Vectorize addmul1 with SSSE3 (x86_64) and NEON (aarch64) nibble lookup

connollydavid
and
claude
committed
Document scalar fallback as deliberately not auto-vectorizable

connollydavid
and
claude
committed
Add AVX2 addmul1 path with runtime CPUID dispatch

connollydavid
and
claude
committed
Refactor packet cook pipeline into context struct for benchmarking

connollydavid
and
claude
committed
Vectorize XOR loops in packet cook with SSE2 and NEON

connollydavid
and
claude
committed
Eliminate per-call malloc in fec_decode

connollydavid
and
claude
committed
Add OPTIMIZATION.md documenting SIMD vectorization results

connollydavid
and
claude
committed
Add CI static builds and profiling script for target hardware

connollydavid
and
claude
committed

Commits on Feb 25, 2026

Commits on Feb 26, 2026

Commits on Feb 27, 2026