• Sources: primary, discussion
  • Summary: The Go blog post covers the simd and archsimd packages behind GOEXPERIMENT=simd, which now reach AVX through AVX-512 on amd64, NEON on arm64, and the 128-bit wasm vector set, with one method name shared across widths and architectures and masks as opaque types the compiler fuses into single instructions where the hardware allows. The post is explicit that CPU feature checks are required to avoid SIGILL and also act as optimization hints, and that loop bounds have to be written so the prove pass can drop bounds checks. The API stays experimental, and SVE, SVE2, riscv64, ppc64, s390x, and loong64 are listed as future work.
  • Why it matters: Writing SIMD in Go has required hand-written assembly, and a portable experimental API changes what a Go hot loop can reach without dropping out of the language.
  • Follow-up: Watch whether the simd and archsimd packages leave GOEXPERIMENT and which of SVE, riscv64, and the remaining architectures land first.