@ai-ecoverse/wasm-emscripten

Emscripten 6.0.9-git for the SLICC wasm realm. Launchers run emcc.py (and friends) on @ai-ecoverse/wasix-python (python3 -S) with EM_CONFIG pointing at this package's config.

emcc does not run on slicc-kernel yet (exec of the glue by path, and no node), see slicc-kernel#390. Since 6.0.9-18 the companions are found beside this package (npm, slicc's ipk add -g, pnpm's virtual store) instead of only under /shared/lib/node_modules.

Companions

Default link (wasm realm)

Executable links (not -c / -r / -shared) inject homescoop_em_cli_ldflags (-sENVIRONMENT=web,worker,node -sEXIT_RUNTIME=1 -sALLOW_MEMORY_GROWTH=1 -sFORCE_FILESYSTEM=1) and lib/slicc/{slicc_spawn,slicc_exec,slicc_signals,slicc_select,slicc_libc_gaps,slicc_jobs}.o and the archives lib/slicc/libslicc_signals.a (kill), lib/slicc/libslicc_select.a (poll, pselect), lib/slicc/libslicc_libc_gaps.a (nanosleep and the 8 set*id), lib/slicc/libslicc_jobs.a (tcgetpgrp, tcsetpgrp) and lib/slicc/libslicc_realpath.a, with the JS libraries lib/slicc/{slicc-writev,slicc-linkat,slicc-uname,slicc-env}.js. slicc-env.js (since 6.0.9-27) gives a program only the environment the kernel passes: none of Emscripten's placeholders (USER/LOGNAME=web_user, HOME=/home/web_user, PATH=/, PWD=/), so under env -i they are unset. Opt out with SLICC_EMCC_PLAIN=1 for plain web glue.

-sMAIN_MODULE=1 and -sLINKABLE=1 links (since 6.0.9-30, homescoop#1678) wrap every input in --whole-archive, so all of libc.a is linked, and every slicc override of a libc function would be a duplicate symbol. Those links take lib/slicc/linkable/*.o instead: the same six objects compiled with -fPIC (a main module needs it), with kill, poll, pselect, nanosleep, setuid, setgid, seteuid, setegid, setreuid, setregid, setresuid, setresgid, tcgetpgrp and tcsetpgrp renamed out (-D<name>=slicc_linkable_<name>, so the shim's own calls inside an object still reach its version), and without the override archives. Those links get libc's versions of these functions and musl's realpath (ELOOP at 40 symlink follows): no kernel signals through kill, no kernel poll, pselect or credentials, and tcgetpgrp/tcsetpgrp answer 0 and EINVAL on a terminal. Everything else in the slicc defaults (spawn, exec, the signal mask and raise, uname, the environment, the kernel's process groups and sessions) stays. -sMAIN_MODULE=2 and -sSIDE_MODULE are not affected: the first links like any executable, the second gets no slicc defaults. Every -sMAIN_MODULE link needs the -fPIC system libraries in sysroot/lib/wasm32-emscripten/pic; emscripten-cache ships them from 6.0.9-7 (homescoop#1699), so these links need nothing built and work from a read-only global install. With emscripten-cache 6.0.9-6 (wasm-emscripten 6.0.9-30) emcc tried to build them into the package and failed with PermissionError. lib/slicc is compiled from homescoop's slicc shims (shims/slicc, Apache-2.0); THIRD-PARTY-NOTICES.md carries that licence. lib/slicc/slicc_main_envp.o (make: -Dmain=slicc_tool_main) and lib/slicc/slicc_fork.o + lib/slicc/slicc-fork.js (needs ASYNCIFY) are shipped but not auto-linked. Fork opt-in:

emcc … $SLICC_EM_LIBDIR/slicc_fork.o \
  --js-library $SLICC_EM_LIBDIR/slicc-fork.js \
  -sASYNCIFY -sASYNCIFY_STACK_SIZE=1048576 \
  -sEXPORTED_RUNTIME_METHODS=FS,ENV,callMain,sliccRunMain,sliccForkChild

Root $PKG/emcc (cmake toolchain) and $PKG/bin/emcc are the same self-sufficient launcher.

CACHE model

• EM_CACHE = $XDG_CACHE_HOME/emscripten or ~/.cache/emscripten (writable, user-owned)
• On first use, bin/em-ensure-cache creates $EM_CACHE/sysroot as a symlink into @ai-ecoverse/emscripten-cache/sysroot — no 100MB copy
• It also copies stamps/sysroot_install.stamp as a real file into $EM_CACHE so emcc never reinstalls headers through the symlink
• js_output/ and symbol_lists/ warm in the user dir (survive ipk reinstall)
• FROZEN_CACHE = False so warm caches work; --clear-cache skips a symlinked sysroot and the stamp
• Override with EM_CACHE if needed

SLICC patches (September + homescoop)

1. Wasm-realm CLI link defaults — see above; optional SLICC_VFS_PRE_JS only when set (node-realm)
2. Configure patch — autoconf/conftest skips NODERAWFS under CLI defaults
3. ac_cv_build seeding — emconfigure sets ac_cv_build=x86_64-pc-linux-gnu (and host/target) so config.guess is not required in-realm
4. Closure unsupported — -sCLOSURE=1 / --closure fail with: Closure Compiler is not available in SLICC's Emscripten package
5. Cache erase — will not delete a symlinked sysroot
6. execvpe — tools/utils.py passes os.environ so emconfigure/emmake keep CC=emcc

Package-only patches (package-patches/, since 6.0.9-27)

On top of the shared patches/ (which scripts/emscripten-tree.sh also applies for emscripten-cache, llvm-emscripten, wasm-clang, wasm-binaryen and cmake), this package ships five patches from package-patches/ and one from package-patches/libc/. They are kept out of patches/ so they don't change the published glue of those packages without their own release:

• 0014, the symlink bound (homescoop#611): FS.lookupPath allows 40 symlink follows, as Linux does, and the 41st is ELOOP. Before it, a path of exactly 40 follows was ELOOP in every program this emcc linked.

• 0015, slicc-env.js in the default link (homescoop#1531), as homescoop_em_cli_ldflags has it.

• 0016, libslicc_realpath.a in the default link (homescoop#1631, #1664): the default link adds lib/slicc/libslicc_realpath.a, an archive whose one member, slicc_realpath.o, is musl's realpath.c with libc patch 0017, compiled with libc's flags. It comes after the program's own inputs and before libc.a, so the linker pulls it only when realpath is still undefined, and then libc.a's own realpath member is never pulled in; emscripten-cache's prebuilt libc.a stays as it is. A program that defines its own realpath (in its sources, an object or an archive or -l on its command line) keeps it, as it did before 6.0.9-28. 6.0.9-28 added slicc_realpath.o as a plain object, which is always linked, so such a program got duplicate symbol: realpath (#1664). SLICC_EMCC_PLAIN=1 links skip the archive and get musl's bound.

• 0018, the override archives and -sMAIN_MODULE=1 (homescoop#1678): the default link adds lib/slicc/libslicc_jobs.a beside libslicc_realpath.a, and for -sMAIN_MODULE=1 and -sLINKABLE links takes lib/slicc/linkable/<obj> for each default object and leaves both archives out. slicc_jobs.o is compiled with tcgetpgrp and tcsetpgrp renamed out; libslicc_jobs.a has one member per override, slicc_tcgetpgrp.o and slicc_tcsetpgrp.o, each slicc_jobs.c with every other global renamed, so the linker pulls each only while its symbol is undefined, and a program's own tcsetpgrp wins. libc's tcsetpgrp is not enough: it goes through emscripten's TIOCSPGRP ioctl, which is EINVAL on every terminal and never reaches the kernel. build.sh fails if a member defines a global the base object also defines, or if a linkable object still defines one of libc's.

• 0019, the other override archives (homescoop#1724): the default link also adds lib/slicc/libslicc_signals.a, libslicc_select.a and libslicc_libc_gaps.a. slicc_signals.o, slicc_select.o and slicc_libc_gaps.o are compiled with the 12 libc functions they override renamed <object>_<name> (slicc_signals_kill, slicc_select_poll, slicc_libc_gaps_setuid, …). Each archive member, slicc_<name>.o, is a one-line function <name> that calls that renamed definition, so the linker pulls a member only while the program leaves <name> undefined, and a program that defines its own kill, poll, pselect, nanosleep or set*id links and runs its own; until 6.0.9-30 those 12 were strong definitions in the always-linked objects, so it got duplicate symbol. The members depend only on the always-linked objects, never on each other, so a program's own poll leaves the slicc pselect in place. The shims' own calls (slicc_exec.o calls kill) go through the symbol, so they reach the program's version when it has one. build.sh fails if a plain object still defines an override, or a member defines anything but its own symbol.

• libc/0017, musl's realpath bound (homescoop#1631): musl stops at ++cnt == SYMLOOP_MAX, so realpath() of a path of exactly 40 symlink follows was ELOOP even after 0014 fixed open() and stat(). 0017 makes it cnt++ == SYMLOOP_MAX: 40 follows resolve and the 41st is ELOOP, as in glibc.

build.sh applies 0014-0016, 0018 and 0019 to the shipped tree and to the tree it links its PRESTAGE programs with. package-patches/libc/ is applied to the shipped tree only, so the PRESTAGE host tree builds the same upstream libc.a that emscripten-cache ships; PRESTAGE then checks open, stat and realpath at 40 and 41 follows (absolute and relative chains) with the default link, and that without libslicc_realpath.a realpath is still ELOOP at 40. wasm-clang, wasm-binaryen and cmake pick them up at their own next release, or a later wave folds them into patches/.

Since 6.0.9-29 PRESTAGE also links programs that define their own realpath: in the program's source, in a later object, in an archive and through -l. Each must link with the default link and run its own realpath. It also checks that 6.0.9-28's shape, slicc_realpath.o as a plain object beside a program's own realpath, still gives duplicate symbol: realpath (homescoop#1664).

Since 6.0.9-30 PRESTAGE links jobs.c (realpath, uname, tcgetpgrp(0) and tcsetpgrp(0, getpid())) with the default link (tcgetpgrp self tcsetpgrp 0 ok; without libslicc_jobs.a it is libc's 0 and EINVAL), with -sMAIN_MODULE=1 and with -sLINKABLE=1, and runs each; the last two must answer libc's 0 and EINVAL and export slicc_linkable_kill, which the default link's wasm must not have. It links own-tc.c, which defines its own tcsetpgrp, and runs its own. 6.0.9-29's lib/slicc (slicc_jobs.o with both overrides, no linkable/) must give duplicate symbol: tcsetpgrp for own-tc.c and fail the -sMAIN_MODULE=1 and -sLINKABLE=1 links (homescoop#1678).

Since 6.0.9-31 PRESTAGE links own-libc.c, which calls all 12 of kill, poll, pselect, nanosleep and the 8 set*id, with a --pre-js stand-in kernel, and prints for each whether the slicc, libc or the program's own version answered. The default link must print slicc for all 12 (and libc for all 12 without the three archives). Defining any one of them in the program's source, or poll and nanosleep in its own -l archive, must link and print own for those and slicc for the rest. With 6.0.9-30's plain objects each of the 12 gives duplicate symbol: <name> (homescoop#1724). priv.c checks that the privilege drop really happens: as root, with a stand-in kernel that keeps credentials, setgid(5), setresuid(6, 0, 6) and setuid(5) must change what getgid, getresuid, getuid and geteuid report. Without libslicc_libc_gaps.a, libc's stubs return -1 and leave every id at 0, so the row fails.

Commands

emcc, em++, emar, emranlib, emcmake, emconfigure, emmake, embuilder, em-config, emscan-deps

emcc hello.c -o hello.js    # runnable in the wasm realm
emcmake cmake -S . -B build
emconfigure ./configure --host=wasm32-unknown-emscripten