[DanielLuning]

What changed

A new bench, bench/elixir_mode.exs, times each of five kernels three ways at 1 B to 1 KiB: the native kernel, its Elixir mode (__defk_elixir__/2) and the example's optimised Elixir. Two read-only hash kernels (xxh64, SipHash-2-4 from hashes.exs), one kernel that allocates and writes (ws_mask), and two kernels with no loop (add, clamp). The result is recorded as docs/performance.md 5.14, with a row in the myths table.

The question (defk-vcx.2): defk generates Elixir beside the C for each kernel, so could a call switch to that Elixir below the crossover size? Answer: no. Elixir mode's rule (test oracle, never production, never a baseline) stands; the entry says so.

Results

Two quiet runs (other CPU 0.3-0.9 cores), medians of 7 interleaved rounds of 20,000 calls, ns per call:

kernel	size	kernel	Elixir mode	Elixir (opt.)	mode/kernel	mode/Elixir
xxh64	1 B	36	603	178	17x	3.4x
xxh64	1 KiB	80	18,100	9,437	227x	1.9x
siphash24	1 B	40	3,002	1,717	76x	1.7x
ws_mask	1 B	83	566	82	6.8x	6.9x
ws_mask	1 KiB	233	145,500	707	624x	206x
add (no loop)		17	248	3	15x	79x
clamp (no loop)		19	222	5	12x	48x

Elixir mode's entry alone costs ~230 ns (argument checks, the __defk_stats__ NIF call for the budget, the process-dictionary budget), 10x a kernel call. Its loops are 1.6-3.4x the optimised Elixir on reads (a helper call per element) and 100-200x on writes (a new binary per write).

The ticket's threshold for a decision ticket (Elixir mode within ~2x of the optimised Elixir) is met by the two read-only hashes from 64 B up, but a switch would replace the kernel call, and the kernel is 7-17x faster than Elixir mode at 1 B and 12-15x on a kernel with no loop at all. There is no size where the generated Elixir is the fast side, so no decision ticket is filed. Owner's call if you read the threshold differently.

How it was checked

mix ci on the M-series laptop: 2305 passed, 815 excluded. The bench checks agreement of the three sides on every input before timing. No test added: the bench is not under test/.

• Touches a promised surface? No: a bench script and the performance guide.
• Touches test/stability/? No.