Three correctness fixes to the --shared build path:
- Only skipped rebuilding a .pic.o when it existed at all, never checking
whether the source .cpp was newer - confirmed live: Tag.pic.o (and 6
other JetCore objects) sat uncompiled since 2026-08-25 while Tag.cpp
itself picked up same-day fixes, so every .jtc silently ran yesterday's
JetCore code.
- Only re-archive libCoreUtils_pic.a when a member object actually changed,
instead of unconditionally on every build.
- Link to a temp file in the same directory, then rename() it into place,
instead of linking directly onto the live .jtc path. A running APIServer
may already have that exact path dlopen()ed (hot-reload watches this
directory) - overwriting it in place corrupts the live mapping's code
pages out from under the process, causing a dlclose() SIGSEGV the next
time it reloads. rename() never touches the old inode, so anything with
it still mapped keeps seeing the unmodified original.
Purely a naming-convention change per user request, not a format
change - the output is still a real ELF shared object (-shared -fPIC),
still dlopen()-able exactly as before. .jtc ("jet compiled") marks it as
an APIServer-loadable content module rather than a generic library.
build-shared.sh: usage text and STEM derivation
(basename "$OUTPUT" .so -> .jtc) both updated - the STEM change is the
one functional piece here, everything else in this commit is comments/
usage strings. Doc comments in jetc.cpp/JetCodegen.h/host_test.cpp/
api_test.cpp updated to match.
Verified live: rebuilt jetc and all five demo*.jtc modules, reran
api_test/host_test against demo3.jtc (still resolves jetc_api_demo3/
jetc_render_demo3 and renders correctly) - the extension is purely
cosmetic to dlopen(), as expected.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
For the server you're about to build: `jetc --shared template.jet out.cpp`
now emits no main() at all - instead the render logic is exported as
`extern "C" void jetc_render_<template>(jet::Global &global,
coreutils::MString &out)`, one call per request, meant to be dlsym()'d
by a long-running host process instead of spawning a jet-2.0 process (or
even a separate compiled executable) per request. extern "C" here only
fixes the symbol's name mangling for a stable dlsym() target - the
parameter types are still real C++ classes, safe to pass across the
dlopen() boundary since the host and the .so share the same
CoreUtils/JetCore headers and toolchain (a same-ABI plugin boundary, not
a foreign C one). coreutils::Exception propagates out of the call
uncaught, so the host decides how to handle a render failure rather than
the module unilaterally deciding (the executable mode's try/catch prints
to stdout and exits, which is CGI-appropriate but wrong for a persistent
server). The exported symbol name is derived from the input filename
(jetc.cpp's baseNameNoExt+sanitizeIdent) so multiple compiled templates'
.so's stay collision-free even under dlopen(..., RTLD_GLOBAL).
jetc/build-shared.sh builds the .so: JetCore's 22 tag classes are still
force-linked directly (Tag::interpretSpan's fallback path can dispatch
to any of them at runtime, regardless of what a given template
statically references), but CoreUtils objects are compiled -fPIC and
archived into a throwaway libCoreUtils_pic.a instead of being linked
directly - libCoreUtils.a itself isn't position-independent, so linking
it into a .so fails outright ("relocation ... can not be used when
making a shared object"), and force-linking every CoreUtils .cpp
directly (rather than archiving, so the linker only pulls in what's
actually referenced) surfaced an unrelated, pre-existing incomplete
class - IMFMessage's vtable is never emitted anywhere, presumably a
partially-migrated class (see project memory on IMFMultipart/IMFText
still being fd-based) - as an unresolved dlopen()-time symbol, even
though nothing here uses it. A normal archive-based link (like the
existing libCoreUtils.a already does for jet-2.0) never pulls IMFMessage
in at all, so building libCoreUtils_pic.a the same way sidesteps it
without touching the unrelated bug itself.
jetc/host_test.cpp is a minimal stand-in for the real server: dlopen()s
a compiled .so, dlsym()s its entry point, and calls it twice with fresh
jet::Global objects (simulating two separate requests against one loaded
module) to prove the .so is genuinely reusable, not single-use. Verified
both demo templates (demo.jet, demo2.jet - the latter covering nested
if/for, a stateful while loop, and the interpretSpan fallback path) build
as .so's and produce the same byte-identical output as the executable
and interpreter paths, across two calls each. Also re-verified the
existing (non-shared) executable build path still matches jet-2.0's
output after the Codegen::generate() signature change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Brad Arant