necc generated code and compiler field, 21 November 2026 Previous compiler b4d0ead81489a0e814110806e0fb482318784667 Current compiler 2ebbb9d95777071d1209e65c50e6178b40ada980 Intel Core i7-11700F, Linux x86-64, Debian 12, CPU 5. Both necc executables were built by the same Nelua frontend, GCC 12.2 at -O3 and nelinker. Both compilers use the same staged SDK and TinyCC preprocessor with the unlocked output fix. The compiler and executable SHA-256 digests and fixture hashes are in checks.json. Arithmetic workload One externally called function takes a run-time count and seed. Starting with x=19, the executable computes x=(33*x+i+17)&1048575 for i from 0 to 5,999,999. The startup assembly writes its 64-bit result. Each admitted executable must emit exactly d3 86 0f 00 00 00 00 00 and exit successfully. The timer includes process launch and exit. Compilation and linking happen outside the runtime measurement. Zig cc 0.17.0 uses -O2 -mcpu=x86_64 -g0 -fno-sanitize=all. Its executable and the previous necc executable match the SHA-256 of the binaries from the earlier same-day SDK comparison that exposed the runtime gap. Full comparison 23 interleaved rounds plus an untimed round, shuffled with seed 60491. The same six compile cases and two runtime cases ran against 13 external configurations. The compiler/SDK input and output digests are recorded in checks.json. Five compile cases and the Lua VM probe produced byte-identical necc objects before and after. The arithmetic kernel produced different machine code and the same runtime result. summary.csv records median milliseconds, quartiles and sizes. samples.csv records all 2,760 wall times in nanoseconds. A compile is measured through the CLI producing its object. Each compile round has a unique inert macro to defeat object caching. All 120 configurations/workloads in the timing table qualified. Compile probes and the full eight-case fixture set use the same controls as the earlier SDK comparison. The Lua VM probe is separate from the timed field. It failed for cproc/QBE (volatile store unsupported) and PCC -O (20 s qualification timeout). Full-field arithmetic medians: previous 9.318 ms, current 1.626 ms, Zig -O2 2.979 ms. The necc kernel's machine code changed from 69 to 154 bytes. Focused control A separate 31-round repeat randomized all three executables in each round and checked all 93 returns. Median times: previous 9.990 ms, current 1.760 ms, Zig -O2 3.155 ms. focused-samples.csv records every sample. No timings were discarded. Independent validation 26,299 inputs across 57 separately compiled loop shapes matched a scalar Python model and GCC with undefined-behavior sanitization. 23 of those routines changed machine code. The full Lua 5.4.6 string implementation produced an identical object, and its hash routine matched 263,688 expected results. A balanced 42-round repeat measured 9.959 ms for the previous necc, 1.722 ms for the new necc, and 3.207 ms for Zig. The new executable finished ahead of the frozen Zig executable in all 42 rounds. A separate Lua string-hash control measured 42.302 ms before and 42.487 ms after. independent-checks.tar.gz includes the drivers, raw samples and per-case checks. Archive SHA-256: 5b3f1aebc7a43bebf23f1b1303c8294ea429ecfe7918ba2e67d949fd4fff6b23 Repeat the runtime test using Python 3.11 or newer, x86-64 Linux binutils, the two necc builds above, their shared SDK and the pinned Zig compiler: NECC_SYSROOT=/path/to/sdk python3 runtime-replay.py --before /path/to/old/necc --after /path/to/new/necc --zig /path/to/zig --cpu 5 --rounds 31 > samples.csv Select a CPU available on the host. The script includes the exact synthetic C kernel, the common assembly entry point and an independent scalar check of the result.