Where the time goes, row by row

Where the loop body is arithmetic — integer or double — indexes a typed array or reads struct members, DominScript runs at a quarter to a half of CPython’s time on both platforms: the binder compiles provably-numeric statements and whole numeric loops into a small register machine, so those loops do not walk the tree. The four style rows are habits rather than algorithms — a for loop, a local declared inside the loop, a small helper, an integer in floating-point arithmetic — and they run at the same ratio; they are in the table so that a regression cannot hide.

Where the work is done inside a plugin, the plugin’s C code sets the pace: JSON and SHA-256 sit at 0.6–1.1×, map and regex at 0.6–0.8×, the one-pass SIMD string search at 0.1–0.2×. Python wins one row on each platform, and neither is the interpreter: SHA-256 on Linux at parity (the same C kernel on both sides, 11 ms each) and the regex row on Windows at 1.09× (the regex engine’s own allocations on the mingw allocator; the same row is 0.68× on Linux) — and SHA-256 at parity (the same C kernel on both sides, 11 ms each). On Windows every DominScript row is ahead of the Python column measured there, the regex row included (14 ms, 0.64×) — an earlier Windows run had it at 1.09×

The two workload shapes this language exists for: the packet data path — socket → buffer → struct → decision — runs at 0.25–0.30× in both its forms, the log-lines row at 0.55–0.63×. The two header-parsing rows read together: the same IPv4 + TCP headers parsed by hand sit at 0.88–0.92×, read through a packet big schema with bit fields at 0.48–0.50× — the readable form is the fast one.

The two tables are the same tree built with gcc on Linux and with mingw-w64 on the same machine booted into Windows, measured on the same day. The DominScript column is nearly the same on both (Fibonacci 9 vs 8 ms, the arithmetic loop 102 vs 102, N-body 32 vs 28); CPython is slower on Windows, so the ratios look better there; the string/heap rows are slower on Windows on the DominScript side (header parsing by hand 121 vs 100 ms, JSON 30 vs 22, regex 24 vs 13), the mingw allocator’s known cost. Twenty of the twenty-one rows are ahead of Python on both platforms. Earlier versions of this page published the same rows with worse ratios — several were behind Python, and they were shown as such; the rows never changed, the implementation did, and the set behind Python shrank with each round — today it is one row per platform. That is not cherry-picking: no row has ever been removed because it looked bad.