Idiomatic Koru Kernels Match Hand-Specialized C (korulang.org)

Koru’s developers report that fused “kernel” code written in idiomatic Koru compiles to performance within about 1% of hand-specialized, fixed-size scalarized C for a 5-body n-body benchmark, outperforming plain reference C, Rust, and Zig by roughly 12–17%. They argue the key is that kernel code preserves optimization-relevant structure (fixed pairwise interactions, explicit aliasing constraints) so the compiler can generate fast machine code without programmer “smuggling” tricks via specialized C rewrites or fast-math. A follow-up Common Lisp kernel DSL can match C on x86-64 with explicit SIMD, but lags on arm64 where it falls back to scalar code, while Koru’s compiler-to-Zig/LLVM approach relies on platform-appropriate auto-vectorization.