Nelua frontend benchmark, 2026-11-21 (x86-64) Baseline 5995eba598541d75d6a2557b73994a15c0922c58 Candidate 874acdd1049a35bc5d8d631d9ddb573d127fc32b Linux, Debian 12 userspace, Intel Core i7-11700F, CPU 3, performance governor. GCC 12.2.0 built the bundled Lua 5.4.6 interpreter; its binary was identical in both trees. The timer starts just before launching the nelua CLI and stops when it exits. -c includes parsing, analysis, generating C and writing the output file. The resulting C program is not compiled or executed inside the measurement. --lint parses the source and applies file-local syntax extensions. Each invocation starts a fresh compiler. Two independent batches each start with one untimed warm-up pair per input, followed by 15 measured pairs. Workload order rotates and the older/newer revision goes first on alternate rounds for each input. OS file caches and the explicit shared compiler cache are warm. No timing samples were removed. Revisions use the same absolute staging and output paths. Each emitted C file is hashed outside the timer and compared to the first result for that input. A mismatch aborts. The table divides the pooled median baseline time by the pooled median candidate time. Batch 1 used 0e1cb4a13ee24b7d4f03057c4f1dc6efb8223a25 and batch 2 used the candidate above. The sole intervening change was a spec assertion. All benchmark inputs, compiler sources, launchers and the harness were identical. Batch IDs in samples.csv preserve each run. Other project pages report independent experiments with different hardware or build profiles. Inputs: examples/helloworld.nelua, examples/matmul.nelua, tests/all_test.nelua, examples/syntax_tour.nelua from each revision, plus one generated function containing 2,048 comparisons against distinct negative integers. That last fixture intentionally requests a large number of diagnostic source locations. Reproduce on Linux in an environment with access to the Nelua fork: git clone https://github.com/Spaceghost/nelua.git frontend git -C frontend checkout 874acdd1049a35bc5d8d631d9ddb573d127fc32b git -C frontend worktree add --detach ../frontend-before 5995eba598541d75d6a2557b73994a15c0922c58 make -C frontend-before cp frontend-before/nelua-lua frontend/nelua-lua taskset -c 3 python3 frontend/tools/bench-frontend.py frontend-before frontend --rounds 15 > samples.csv Run the harness twice to reproduce the two batches, saving each CSV separately. Choose a CPU available to your process. Timings are nanoseconds in the CSV. Position is the execution order within each pair. Output bytes is zero for silent lint success. Times and file lengths can vary between machines and temporary directory paths.