@ai-ecoverse/wasix-perl

Perl 5.42 for slicc WASIX (perl, prove, pod2man, cpan, perldoc).

• Cross-built with perl-cross + wasixcc against wasix-sysroot
• Built on wasix-sysroot 2025.9.30-42 (the asyncify tree, for real fork)
• From wasix-sysroot and slicc-kernel (≥ 1.120.0): user ids, %SIG (inherited IGNORE too), signal masks, flock (per open file description) and fcntl record locks, mkfifo and times; file modes via slicc_fs; DynaLoader and all of Encode statically linked
• Core XS (POSIX, Fcntl, Cwd, Digest::*, Encode, Socket, …) statically linked
• fork / system / backticks via WASIX proc_fork (asyncify)
• PERL5LIB defaults to package lib/perl5 plus ~/.local/share/perl5
• use re 'debug' / 'eval' and XS Cwd (getcwd, abs_path) work since 5.42.0-7: these modules only load their XS when DynaLoader::boot_DynaLoader exists (-7 to -12 provided a no-op one; -13 links the real DynaLoader). In -6, perl -Mre=debug died ("Undefined subroutine &re::install") and getcwd inside a mount returned an empty string.
• File modes on slicc-kernel ≥ 1.35.1: umask, chmod, mkdir modes and File::Temp's 600/700 hold, and stat reports them (wasix-sysroot slicc_fs, since -17); older kernels leave files 644 and directories 755

5.42.0-18: Unix-domain socket addresses

IO::Socket::UNIX's hostpath and peerpath, and Socket::unpack_sockaddr_un on what getsockname, getpeername or accept return, give the socket's path. Since wasix-sysroot -31 these return a Linux-length address (the family, the path and its NUL), and up to 5.42.0-17 Socket.xs accepted only the full 112-byte sockaddr_un there and died with Bad arg length for Socket::unpack_sockaddr_un; the sockets themselves worked (homescoop#767).

5.42.0-19: real flock and mkfifo (wasix-sysroot -34; slicc-kernel ≥ 1.120.0)

• flock is libc's own flock(2), through slicc-kernel's slicc.flock. As on Linux, a lock belongs to the open file description, on any handle: a file, a pipe, a terminal, a socket or a directory.
  • A duped handle and one a forked child inherits share the lock. A second open of the same file, or the other end of a pipe, conflicts.
  • LOCK_EX on a read-only handle excludes, so the flock(DATA, LOCK_EX|LOCK_NB) single-instance idiom refuses a second instance (homescoop#758). Locks on pipes and STDOUT exclude too (homescoop#765).
  • A handled signal can end a blocking flock with EINTR, as on Linux.
  • flock and fcntl record locks are separate, as on Linux.
• mkfifo: POSIX::mkfifo and mknod make real FIFOs (slicc_fs.path_mkfifo, homescoop#878). Up to -18 they failed with EOPNOTSUPP.
• Kernel: this needs slicc-kernel ≥ 1.120.0 (engines).

5.42.0-20: Linux file-type bits for FIFOs and sockets

stat and lstat report a FIFO as 0010000 and a socket as 0140000 in (stat)[2], as Linux and coreutils stat -c %f do, and Fcntl's S_IFMT, S_IFIFO and S_IFSOCK are Linux's (0170000, 0010000, 0140000), so S_ISFIFO/S_ISSOCK agree with them (homescoop#1015). Up to 5.42.0-19 these were wasix-libc's values (0140000 for a FIFO, 0160000 for a socket, S_IFMT 0160000), so code comparing raw modes with Linux literals was wrong. -p and -S are unchanged. The build checks that package.json's description names the wasix-sysroot it was built on (homescoop#752).

5.42.0-22: a 64 MiB initial heap; wasix-sysroot -42

• Sysroot: built on wasix-sysroot 2025.9.30-42. Since -40 its malloc allocates from the initial memory, since -41 relative paths and getcwd go through the process's cwd descriptor (a removed cwd gives ENOENT, as on Linux, on slicc-kernel ≥ 1.135.6), and since -42 gethostname/uname read /etc/hostname first (homescoop#1462).
• Initial heap: perl.wasm links a 64 MiB initial memory (homescoop#1287, slicc-kernel#1172). Loading core modules, a 100000-key hash and an automake-style parse peak at 22-44 MiB, so they never grow the shared memory; bigger jobs grow past it, which -39's allocator refresh keeps safe. Up to 5.42.0-21 perl started at 13 MiB and grew for each of them.
• Kernel floor: engines asks for slicc-kernel ≥ 1.120.0. Kernels from 1.106.2 on refuse to install the package on an older kernel (slicc-kernel#797); older kernels install it without a warning, and there flock, mkfifo and the signal features above aren't real (homescoop#651).

5.42.0-21: both licence texts; wasix-sysroot -39

• Licence: perl is Artistic-1.0-Perl OR GPL-1.0-or-later, and LICENSE now carries both texts, perl's Artistic and Copying (GPL version 1). Up to 5.42.0-20 it held the Artistic text only (homescoop#995).
• Sysroot: built on wasix-sysroot 2025.9.30-39, whose malloc refreshes a thread's view of grown shared memory (homescoop#1287). perl is single-threaded, so nothing changes for scripts. Since -37 libc's S_IF* bits are Linux's too, the values patches/0009 already gives.
• flock lifetimes: a lock belongs to its open file description (since -19). It stays held when a forked child that inherited the handle exits, and is freed when the holder closes its last handle on that description or is killed. cert/flock-lifetime.mjs checks each case and EINTR (homescoop#1024).
• Config: no %Config value names the build machine's paths (since -19; homescoop#848, #850).

5.42.0-15 / -16: flock excludes other processes (up to -18)

flock takes whole-file fcntl record locks (LOCK_SH → F_RDLCK, LOCK_EX → F_WRLCK, LOCK_UN → F_UNLCK; LOCK_NB → F_SETLK, else F_SETLKW), which slicc-kernel enforces between processes (homescoop#631). Up to -14 every flock succeeded and excluded nobody, so two autom4te runs on one autom4te.cache raced ("cannot rename autom4te.cache/output.0t"). A refused LOCK_NB sets EWOULDBLOCK. These are fcntl semantics, as perl's own flock emulation has them: a lock belongs to the process (a forked child doesn't inherit it, and closing any handle on the file drops it). fcntl wants a writable handle for a write lock and a readable one for a read lock, so LOCK_EX on a read-only handle takes a read lock (other writers stay out, other read-only LOCK_EX holders don't) and LOCK_SH on a write-only one a write lock. So two processes that each LOCK_EX a read-only handle (the flock(DATA, LOCK_EX|LOCK_NB) single-instance idiom) are both granted: open the lock file read-write for exclusion (homescoop#758). flock on something that isn't a regular file (a pipe, STDOUT on a terminal or pipe, /dev/null, a directory) succeeds and excludes nobody, since slicc-kernel keeps record locks on regular files only (5.42.0-16; -15 failed it with EBADF, homescoop#757). lockf is still an advisory no-op.

5.42.0-14: needs slicc-kernel ≥ 1.100.0

A packaging-only release of 5.42.0-13 (the same files): engines now asks for slicc-kernel ≥ 1.100.0, the first kernel with slicc.fcntl_lock. On an older kernel the libc grants every fcntl record lock locally, so two processes could both hold a write lock on the same bytes and lose data (homescoop#552). Up to -13 engines said ≥ 1.44.0.

The floor is enforced from slicc-kernel 1.106.2 on: its command catalog refuses a package whose engines.slicc-kernel the running kernel doesn't meet (ai-ecoverse/slicc-kernel#797). Older kernels install the package without a warning, and there fcntl locks, and so flock, aren't real: use slicc-kernel ≥ 1.100.0 (homescoop#651).

From slicc-kernel 1.106.2 the commands' /usr/bin entries are the scripts themselves (slicc-kernel#762), so pod2man --help and pod2text --help, which read their usage from $0, print it (homescoop#491); older kernels print nothing there.

5.42.0-13: wasix-sysroot 2025.9.30-30

• POSIX record locks: Fcntl exports F_GETLK, F_SETLK, F_SETLKW and F_RDLCK/F_WRLCK/F_UNLCK (up to -12, use Fcntl qw(F_SETLK) died with "Your vendor has not defined Fcntl macro F_GETLK"), and fcntl takes them to slicc-kernel (≥ 1.100.0), which keeps the locks per process, as Linux does: a forked child asking for a write lock on a range the parent holds gets EAGAIN, F_SETLKW waits for the unlock, and F_GETLK names the holder. On an older kernel every lock is granted locally and F_GETLK answers F_UNLCK, so a script that probes F_SETLK and falls back now thinks it has locking there (hence -14's engines floor). The struct is pack('s s x4 q q l x4', $type, $whence, $start, $len, $pid) (32 bytes on wasm32). flock is unchanged: an advisory no-op that succeeds on an open handle.
• The real DynaLoader, as upstream builds it for a static perl (dl_none.xs): DynaLoader::dl_error says "Not implemented", and loading a module that isn't linked in croaks with "Can't load module …, dynamic loading not available in this perl." Up to -12 boot_DynaLoader was a no-op marker and the dl_* functions were missing.
• All of Encode: its XS sub-extensions (Encode::Unicode, Byte, Symbol, EBCDIC, CN, JP, KR, TW) are linked in, so encode("UTF-16LE", …), cp1252, koi8-r, shiftjis, euc-kr, big5-eten and the rest of Encode->encodings(":all") work. Up to -12 they died with "Unknown encoding".
• wasix-sysroot -29 and -30 also bring poll with events 0, WASIX epoll with Linux op numbers and named semaphores (no perl builtin uses them).

5.42.0-12: wasix-sysroot 2025.9.30-28

• Signal masks are real: POSIX::sigprocmask blocks and unblocks through slicc-kernel (≥ 1.88.0). A blocked signal stays pending (POSIX::sigpending) without running its %SIG handler or cutting a select sleep short, and the handler runs once when it is unblocked. Up to 5.42.0-11 the mask was ignored: the handler ran at once and nothing was ever pending. Kernels before 1.88.0 take wasix-sysroot -28's ENOSYS fallback (its README).
• bin/perl.wasm is mode 755 like the other command files (5.42.0-11 shipped it 644).

5.42.0-11

• utime $atime, $mtime, $fh (a filehandle) works: futimes is libc's futimens; up to 5.42.0-10 it was a stub that failed with EOPNOTSUPP (homescoop#307). utime undef, undef, $fh sets now.
• flock succeeds on an open handle and fails with EBADF on a closed one. slicc-kernel has no advisory locks, so a lock is never refused (as in wasix-python and wasix-ruby); up to -10 every flock failed with EOPNOTSUPP, which modules that lock their files (DB_File-style caches, File::Temp with :flock users) treated as fatal. lockf the same.
• use filetest 'access' makes -r/-w/-x ask libc's eaccess (wasix-sysroot -25); up to -10 eaccess was a stub, so they were false.
• prove and TAP::Harness work, and so does 4-argument select (and IO::Select) on more than one handle: WASI's fd_set is a list of fds, not a bit mask, and perl handed select() its bit strings as fd_sets (patches/0005-sselect-wasi-fd-set.patch). Up to -10 prove reported "No plan found in TAP output" for every test (homescoop#306) and IO::Select->can_read could trap.
• The build is reproducible: perl -V's "Compiled at" is perl 5.42.0's release date (SOURCE_DATE_EPOCH), archive dates are zero, and a second build of the same commit is byte-identical (checked by ladder-pr's reproducible (wasix-perl) job, reproducible: true in recipe.yaml).

5.42.0-10: wasix-sysroot 2025.9.30-27

• utime undef, undef, $file (and anything that touches a file through it) sets the times to now; up to 5.42.0-9 it set them to 0 (1970), from wasix-sysroot's broken utimensat(NULL) (homescoop#294).
• times reports the process CPU time slicc-kernel keeps; system and children's times are 0. $^T/time are unchanged.
• An empty path is ENOENT (-e "", open(my $f, "<", "")), as on Linux.
• getcwd after a chdir through a symlink is the physical directory (since wasix-sysroot -18; 5.42.0-8 and older gave the logical one).
• Files have owners on slicc-kernel ≥ 1.59.0 with users on: chown, -o, -O and stat's uid/gid are the kernel's, and -r/-w/-x ask it (access, slicc-kernel ≥ 1.63.0).

5.42.0-9: wasix-sysroot 2025.9.30-20

• %SIG handlers fire on slicc-kernel ≥ 1.47.1: the libc reports every disposition to the kernel (slicc.sigaction_set), so $SIG{CHLD} reaping and $SIG{WINCH} work. alarm, sleep interrupted by a signal and kill $$ behave as on Linux.
• $<, $>, $(, getpwuid and getgrgid come from slicc-kernel's process credentials (slicc-kernel ≥ 1.44.0, no fallback). Root is uid 0 with home /root; a kernel.users.add user gets its own ids.