Nelua ecosystem measurements ยท 21 November 2026 Read summary.csv for medians, inclusive quartiles, minimum, maximum and units. The raw CSVs include the warm-up where it was recorded (round zero). Each comparison used fresh, interleaved processes on the same host. These are warm filesystem results. Other host activity contributed to wide wall-time ranges. No installation telemetry or workload-frequency counts were available. Priorities and coverage * Cascadia Core. Boot-to-command runs once per VM launch. Changing the local compression of the release initramfs saved 2.27 s at the median in TCG. The direct-shell control also used the same kernel and cpio. It bypasses service startup, including the fixture's requested 350 ms failed readiness check. It is a diagnostic lower bound, not an equivalent production boot. Theseus Init is a related project outside Nelua core. * Generated Lua VM. Switch dispatch occurs on every interpreted bytecode. The table improvement saved 108 ms per million-call run and 93 ms during a fresh Nelua matrix C emission with this interpreter configuration. TinyCC still led by 329 ms on the million-call input and 238 ms on the frontend workload. This comparison applies to an interpreter whose VM was compiled by necc; the system Nelua compiler uses its own interpreter. * Nelibc. Search is a library primitive and the distribution matters. The alternating 64 KiB memmem miss saved 215 us/call. The interior-byte miss saved 126 us/call. The 63-byte-run miss already took about 1 us. Early long-needle memmem hits became more expensive. See all 33 rows per round for text, seeded random data, binary data, hits and misses. * Native Nelua and necc compilation. Building the real Lua VM translation unit still took 138 ms with necc versus 24 ms with the SDK TinyCC. Fork versus upstream release compilation differed by workload. Trying -O3 for the bundled Lua interpreter gave mixed results; frontend-flags.csv compares it with the default -O2 in seven interleaved rounds, including byte-identical C checks. No compiler flag change was shipped. * nelinker. The current standalone build linked the full Lua interpreter in 10.71 ms versus TinyCC's 16.31 ms. Their peak RSS differed by 4 KiB. nebuild was competitive on unchanged targets; Ninja led by 0.81 ms on the larger manifest. These timings do not establish a scheduling gap. * Inventory. The Core index covers the compiler, Nelibc, necc, nelinker, nebuild, neboot and the Nebox userland including nsh. The established compiler frontend, C-toolchain, linker/archive, unchanged-build and libc tables remain available on their project pages. The neboot page records an OVMF boot qualification. Direct kernel boot here exercises no EFI menu. The Workshop also lists Drift, Atlas firmware, melodica.h, libc.bend and Writ. They supply no matched timing table on the index. Substring search Before 70bcd89457f2b1433916747df715ea8546e54ffb After 84267ab9185f25c88eca4cd10366207a5028ecc7 Reference glibc 2.36; gcc 12.2.0 Debian 12; Linux 6.18.54-0-lts Intel Core i5-10400; pinned logical CPU 0; GCC/glibc hybrid Nelua /opt/nelua-dabbd9ef/nelua; Zig 0.17.0 /opt/zig/zig In separate current and previous Nelibc trees on the designated builder, build each GCC hybrid archive as the unprivileged build user. Give each tree its own compiler cache. Example for either tree: make lib-gcc "NELUA=/opt/nelua-dabbd9ef/nelua --cache-dir $PWD/cache" ZIG=/opt/zig/zig In an empty working directory download substring-rounds.sh alongside this file and run: INPUT=$AFTER_TREE/tools/substring_cases.c BEFORE=$BEFORE_TREE/build/libnelibc-gcc.a \ AFTER=$AFTER_TREE/build/libnelibc-gcc.a CPU=0 sh substring-rounds.sh >repeat.csv The driver links the same GCC -O2 -fno-builtin consumer three ways. The executable calls the functions through volatile function pointers. Each sample is ns/call over 256 searches. The C source defines all bytes, the independent first-match oracle and guard mappings for both inputs. For the early 63-byte-run fixture, inserting 64 bytes at 8 completes the earlier run at offset 0. substring.csv records the actual first offset. The binary miss and hit cases use an eight-byte needle containing NUL. Round zero warms each implementation. Every measured round rotates order. Hybrid archive SHA-256: before 7b49ca2d980184a62927f12fe9189b2f7da29374ae15d409a8bbbb43423a246d after c041b9f73446ee7e181cb7bf6a229892a0a9b7765fd2d7dd2423f09c4a483efa. Toolchain and build-tool reproduction Download replay.tar.gz from the same directory. In an empty work directory extract it, then obtain these sibling checkouts. Repository names are under github.com/Spaceghost except for nelua-reference and tcc-source. The sibling directory nelua is Spaceghost/nelua. The public necc runtime fixture is loaded from nelua-site by the replay. nelua 874acdd1049a35bc5d8d631d9ddb573d127fc32b nelua-reference edubart/nelua-lang a58450563e2d2ec49bff499865c8b5cfdf6ff81a necc e1eab18d53e763693ca91cef30b476c6c29f45b0 necc-before 2ebbb9d95777071d1209e65c50e6178b40ada980 (necc) nelinker a349c2a9177aeeeae01078c1632d5a36cb276559 nebuild 60d1200805b0529720fdea2a21b68c1156816340 nebuild-before 0441e93 (nebuild) nelua-site 5f61fa3c44c53920401e0b578af4ff49e5234672 tcc-source TinyCC/tinycc 32d1c0a5614b9979768b3711726db3d601750ed3 These experiments used x86-64 Fedora 44, Linux 7.1.10-200.fc44.x86_64, Intel Core i7-9700K, GCC 16.2.1, GNU binutils 2.46.1, GNU make and Ninja 1.13.2. The compiler and linker processes used CPU 2; nebuild used CPU 3. Tests include process startup. Use an available CPU when replaying. For the compiler and linker process timings, round zero is excluded in summary.csv. The boot and EFI samples include every launch. The nebuild harness discards its own three warm-ups and numbers the twelve measured rounds from 1. mkdir -p timings/{necc,link,vm} necc/build nelinker/build sdkroot make -C nelua && make -C nelua-reference taskset -c 2 python3 timings/compare-nelua.py > nelua.csv For the -O3 experiment use another tree at 874acdd, build it with make OPTCFLAGS=-O3, then run the fork's tools/bench-frontend.py on both trees with --rounds 7 on one CPU. The regular release compilation above used -C -r -b --cc gcc. Generated programs ran outside the compilation timer. They printed "hello world" and "-18.8963499125" respectively. Use Cascadia release core-dev-2026.11.20 as the immutable bootstrap SDK. Obtain the image with repository access and verify its published SHA256SUMS: archive=$PWD/cascadia-core-x86_64.cpio.gz (cd sdkroot && gzip -dc "$archive" | cpio -id --no-absolute-filenames 'bin/necc' 'bin/nelinker' 'usr/lib/necc/*') export NECC_SYSROOT=$PWD/sdkroot/usr/lib/necc export NECC_NELINKER=$PWD/sdkroot/bin/nelinker (cd necc-before && ../sdkroot/bin/necc --release necc.nelua -o ../necc/build/necc) (cd nelinker/src && NECC_NELINKER= ../../sdkroot/bin/necc --release nelinker.nelua -o ../build/nelinker) For the C compile and VM workloads rebuild the SDK's single-threaded TinyCC with the current necc recipe. Apply the Nelibc pinned x87 fix with its zero-context patch option in the TinyCC source checkout. Confirm the added assignment is directly after the fstpt encoding in the VT_LDOUBLE case. git -C tcc-source apply --unidiff-zero "$PWD/nelibc/tools/toolchains/tinycc-x87-argument-pop.patch" (cd tcc-source && ./configure && make tccdefs_.h) necc/rebuild-tcc.sh "$NECC_SYSROOT" "$PWD/tcc-source" "$PWD/timings/new-tcc" cp timings/new-tcc "$NECC_SYSROOT/tcc/tcc" (cd necc && ../sdkroot/bin/necc --release necc.nelua -o build/candidate) unset NECC_NELINKER bash timings/build_vm.sh ln -s ../../nelua/lualib timings/vm/lualib ln -s ../../nelua/lib timings/vm/lib taskset -c 2 python3 timings/compare-necc.py > necc-final.csv taskset -c 2 python3 timings/vmbench.py > vm.csv taskset -c 2 python3 timings/frontend-vmbench.py > frontend-vm.csv taskset -c 2 python3 timings/link_bench.py > link.csv taskset -c 2 python3 timings/link-vm.py > link-vm.csv The source for all interpreter objects is the fork's bundled Lua. The VM dispatch comparison compiles lvm.c with LUA_USE_JUMPTABLE=0 for all three builds so the compiler receives the same portable C switch. All its other objects use the same TinyCC, headers and ABI. GNU as and ld.bfd build the separate arithmetic runtime control. compare-necc.py verifies its eight output bytes. The arithmetic necc objects had identical SHA-256 before and after: 3b94f54406b4ab1188281294c1536109ae4b384b3e6fc9ba622462256ba2e157. The arithmetic timings on these identical objects varied. Compile time, standalone program execution and execution of the mixed interpreter have separate rows. Both previous-revision compiler builds (before and after the SDK preprocessor update) produced identical executables on the measurement host. Link time measures process launch through exit. GNU time -f %M measures peak RSS in a second fresh invocation, excluding the Python harness. Both outputs must execute successfully. For the real SDK archive-member input the check requires exit 42. For the full Lua input, both linked programs must evaluate an expression and print 42. for d in nebuild-before nebuild; do (cd "$d/src" && NECC_NELINKER= ../../sdkroot/bin/necc --release nebuild.nelua -o ../nebuild) done python3 nelua-site/public/nelua/assets/build-tools-20261121/bench.py incremental \ nebuild-before/nebuild nebuild/nebuild --rounds 12 --cpu 3 --output timings/nebuild Boot reproduction and effect of the host Theseus Init checkout: 72095fd70e77da93fada45686459985cee5429d3. The source revision, conversion commands, first-command timing harness and release asset names are in Cascadia 3fa30c5703ef7801b9ebd327f29d359f903026eb, docs/CORE-DEVELOPER.md, section "Faster repeated boots under software emulation". The timing harness uses the same graph for both codecs. boot.csv has the individual prompt and successful command times. QEMU 10.0.13 (Debian), one pinned i7-9700K logical CPU, one TCG vCPU, 512 MiB, serial stdio, local storage, no network. The archive sizes were 47,925,071 and 55,997,154 bytes before and after. Each round alternated order. The full release smoke checked the readiness ordering, skipped dependency, interactive console, supervisor exit and poweroff on the LZ4 image. EFI menu profile The Core EFI disk from the Cascadia core-native-services-e15751b artifacts uses neboot b8c56a815f6e07beb7abdad9339568790594d4d4. Four fresh boots on the same i7-9700K with QEMU 10.0.13, TCG, one vCPU, 1 GiB and OVMF split launch to completed menu, immediate option 1 selection to the "Starting" marker immediately before EFI StartImage, and handoff to a working shell checking /proc/1/exe. All four powered off on request. The timings in efi.csv include every run. This is a separate EFI disk experiment from the gzip/LZ4 direct-kernel comparison. Uncompress cascadia-core-uefi-x86_64.img.gz as core.img in an empty directory. Copy OVMF_CODE.fd and a clean OVMF_VARS.fd template into it. The measured disk SHA-256 was 50ed4d8452ff4925522dcba33eb08fb24bbf3c53e137454913e482845f4e9630. Firmware SHA-256 was 904bfa3e0d966372b43b804c4fe323ae63751566687c2bfdf52ca947f47eb13a and vars SHA-256 was 6ed987af3a3c155be71665f510eae3e007eda9b8b94afd59d45e91c4a11565cc. Run efi.py in that directory with the neboot checkout mounted at /tools, or update its /tools path for the checkout. It uses the repository's QEMU console, fresh EFI variables per round, q35, -cpu max and removable USB. Additional substring controls readme-search.c and readme-search.sh exercise the unmodified 5,334-byte Nelibc README in the updated checkout. Copy both files into an empty directory and run sh readme-search.sh PREVIOUS UPDATED. Supply absolute paths to the two GCC hybrid checkouts with built archives. The six rounds after warm-up verified offsets on all 512 calls per cell. readme-search.csv records every case, including process warm-up. The anchor control archive contains the original tools/strstr_bench.c --anchor-controls samples. Build each using gcc -O2 -fno-builtin with the corresponding archive (or no archive for glibc), -pthread -lm. Run in rotating order on CPU 0. The CSV filename identifies the implementation. Across six rounds the 64-byte equal-frequency control rose from 83 to 101 ns, and the equivalent 64 KiB case moved from 66.85 to 66.48 us (glibc 107.28 us). Nebox word count Before 94fac62b7ad738620ffc5c9028bac2029a1efa7d After 339eae37a65f6e93ddea6e57275b6815714f8545 GNU coreutils 9.1, LC_ALL=C, same i5-10400 guest as Nelibc. Build both revisions with the Spaceghost Nelua fork at 874acdd, Zig, and the same static Zig archive from Nelibc 84267ab (SHA-256 5b01a7c67d20968a6dec81c44518751b27c6fcc9298f8ff7dddd9bfc167d05f2). Use the repository Makefile, specifying NELUA, ZIG and NELIBC, then copy the binaries to previous/nebox and updated/nebox in a scratch directory. With the Nelibc checkout at 84267ab, generate the source corpus: (cd nelibc && git ls-files 'src/*.nelua' | LC_ALL=C sort | xargs cat) > corpus.nelua LC_ALL=C python3 wc.py > wc.csv The corpus SHA-256 is edc71e2ae6ec70c67eddf93ad5422b4d3527292f2f3f6af50c200bba02f007df. For the shell control, place a Nebox checkout at nebox/ and run LC_ALL=C python3 shell.py. Debian dash 0.5.12-2 ran the same scripts. The catalog fixture and interpreter arguments are recorded in shell.py. The 16 rotating word-count rounds include one warm-up. Every process checked its counts. For wc -l, 15 measured times spanned 4.64โ€“5.00 ms before and 1.59โ€“1.68 ms after, compared with 0.95โ€“1.35 ms for GNU. All counted 47,135 lines. Qualification Nelibc manual gate results for 84267ab are recorded in nelibc-gates.csv. All twelve targets exited zero as an unprivileged build user. The separate Fedora job completed successfully, including both Bend smoke targets: https://github.com/Spaceghost/nelibc/actions/runs/44938088906 necc passed check.sh, check-driver.sh, the SysV and Microsoft ABI checks, and the runtime check. Cascadia passed its repository check and boot smoke. Nebox's unprivileged conformance suite recorded 1,570 passes. The wc/id differential suite matched all 446 cases. Another 70 previous/updated comparisons exercised empty, binary, UTF-8 and cross-buffer inputs in C and C.UTF-8 locales. The options -l, -lc and -lm also matched GNU wc. On Debian 12, make test stopped at factor's directory-read comparison: GNU coreutils 9.1 returned zero while both Nebox revisions returned one. The complete factor comparison on the same updated binary with Ubuntu 24.04's GNU coreutils 9.4 passed all 677 cases. The subsequent Debian Makefile comparisons were invoked individually on the same updated binary. All 21 exited zero. Nebox shipped on its default branch, master. Its hosted push workflow selects main, and this master push did not trigger it.