End-to-end exercise of badc against the upstream Lua 5.5.0
interpreter and its official test suite. setup.py fetches the
release tarball plus the matching test suite, smoke.py builds
the interpreter with badc at both -O0 and -O, and runs a
curated subset of the test scripts against each lane.
The source distribution is not committed: setup.py pulls a
pinned mirror from the kromych/badc vendor-deps release and
verifies a sha256 before extracting under demos/lua/src/.
| File | Tracked? | Purpose |
|---|---|---|
setup.py |
yes | Fetch + extract the source and the test suite tarballs. Idempotent. |
smoke.py |
yes | Build the interpreter with badc + run the curated test subset. Returns 0 on success. |
src/* |
no | Lua interpreter sources from www.lua.org/ftp/. Produced by setup.py. |
tests/* |
no | Lua test scripts from www.lua.org/tests/. Produced by setup.py. |
.cache/ |
no | Cached tarballs. |
python demos/lua/setup.py # fetches sources + tests
python demos/lua/smoke.py # builds + runs the test subset
smoke.py returns 0 with smoke OK [no-O]: 19 scripts green /
smoke OK [-O]: 19 scripts green when both lanes pass. On a
regression it names the failing script, prints the last few
output lines, and returns 1.
smoke.py honours BADC=path/to/badc if you want a debug or
custom-built binary instead of target/release/badc.
24 of the upstream test scripts that exit 0 under
_port=true; _nomsg=true. The build threads
LUA_USER_H='"ltests.h"' into every TU and links
tests/ltests/ltests.c so the T internal-testing library
is registered on every state, which is what the api, gc,
gengc, memerr, and tracegc scripts assert against.
| Script | Coverage |
|---|---|
api.lua |
T internal API: stack manip, raw object access, GC peek |
bitwise.lua |
shift / and / or / xor / not at full 64-bit width |
calls.lua |
call semantics, __call, tail calls, binary chunks |
closure.lua |
upvalues, GC interaction, closure equality |
constructs.lua |
parser corner cases, statement layout |
coroutine.lua |
coroutine.create / resume / yield / wrap |
cstack.lua |
recursion depth, error overflow handling |
errors.lua |
pcall / xpcall, tokenised error messages |
events.lua |
metatables, __add / __index / friends |
gc.lua |
incremental GC sweeps + finalisers under T step control |
gengc.lua |
generational GC mode + minor/major cycle accounting |
goto.lua |
block-scope goto, global declarations |
literals.lua |
integer / hex / float / string escape coverage |
locals.lua |
to-be-closed variables, scope nesting |
math.lua |
math.huge, integer boundaries, rounding, random |
memerr.lua |
allocation-failure injection at every internal mem site |
nextvar.lua |
numeric for, generic for, table iteration |
pm.lua |
pattern matching (string.match / gmatch / gsub) |
sort.lua |
table.sort, table.move |
strings.lua |
string.format, string.pack / unpack |
tpack.lua |
binary pack / unpack across alignment shapes |
tracegc.lua |
per-step GC tracing via T.tracegc |
utf8.lua |
utf8.char / utf8.len / utf8.offset |
vararg.lua |
... propagation, table.pack / unpack |
Scripts outside this set are excluded for the reasons below; re-adding any of them is the natural follow-up once badc grows the missing surface:
attrib.lua, main.lua – shell out (os.execute) and load
compiled C extensions through package.loadlib. The dynamic-
link surface c5 ships today doesn’t cover the upstream test
helpers under tests/libs/.big.lua, heavy.lua, verybig.lua – 100 MB+ allocations
intended for soak runs; too slow for a per-PR smoke.db.lua – exercises debug.sethook line-counting at a
granularity that depends on the upstream VM’s instruction
layout.code.lua – introspects the compiled bytecode; only
meaningful under the upstream VM lowering.The same scripts run unchanged from .github/workflows/ci.yml
on every supported runner (ubuntu-latest, ubuntu-24.04-arm,
macos-latest, windows-latest, windows-11-arm) after the
existing demo smokes. actions/setup-python@v5 pins python to
the same 3.x interpreter on every lane.
Gating policy: macOS and the two Linux lanes are strict. The
two Windows lanes adopt the chibicc / tinycc convention –
smoke.py exits 2 when the build is green but specific tests
fail, and the workflow accepts that exit (|| test $? -eq 2)
while still failing on a real build error (exit 1).
math.lua:1137 fails on both Windows lanes. The assertion is
the random-float round-trip assert(tonumber(s) == n) where
s = tostring(n) over 400 random doubles, skipping the strings
that don’t look numeric via the regex ^%-?%d.
The seed is fixed (math.randomseed(1007) at line 858) so the
sequence is deterministic. On Windows one of the 400 floats
lands on a NaN or infinity bit pattern; msvcrt’s snprintf
prints those as 1.#INF / -1.#INF / 1.#IND / 1.#QNAN
rather than the C99 spellings (inf / nan) used by glibc and
Apple libm. The leading digit slips past the skip regex, the
string feeds into tonumber which returns nil, and the
assertion fires.
The Universal CRT switched to the C99 spellings in VS2015 and
the math entry points (frexp, pow) are exported directly
by ucrtbase.dll – the surgical ucrtbase::frexp and
ucrtbase::pow pins in libc/include/math.h are what
cleared the earlier math.lua:692 (frexp(+/-inf)) and
math.lua:1104 (pow(2, -1023) flushing to zero) divergences.
snprintf is harder. The C99 entry is an inline wrapper in
the MSVC headers around __stdio_common_vsnprintf; the
underscored legacy _snprintf is also not exported from
ucrtbase.dll directly on the Windows runners (both binding
attempts hit STATUS_ENTRYPOINT_NOT_FOUND at PE load). The
documented UCRT import path is via the
api-ms-win-crt-stdio-l1-1-0.dll API-set proxy DLL. Closing
this case needs either: (a) c5 support for binding through
api-ms-* proxies, or (b) binding __stdio_common_vsnprintf
(which is a direct ucrtbase export) plus an inline wrapper
that constructs a Windows-shaped va_list at the call site.
Tracked under the TODO marker.
VERSION, SRC_UPSTREAM_SHA, TESTS_UPSTREAM_SHA in
setup.py (and the matching constants in
scripts/vendor_deps/build_bundle.py).python scripts/vendor_deps/build_bundle.py -v, upload
the new tarballs to the vendor-deps-v1 release, refresh
manifest.json.python setup.py -v followed by python smoke.py.