Initial port for AIX

[mehendarkarprajwal]

Recently llama.cpp started using upstream subprocess.h for launching llama-server in router mode.
Here -> ggml-org/llama.cpp@ec18edf
And some of the APIs used in this file are not present on AIX natively. So using combination of other primitive APIs like fork() and exec() I have replicated the functionality of the posix_spawn_file_actions_addchdir
With these changes llama.cpp now works smoothly on AIX.

Please let me know if any changes required
thank you!!

[sheredom]

Is there anyway to test for AIX? Does #104 not 'solve' the cwd issue?

[bernardladenthin]

Before I get to AIX: this currently doesn't build on the non-AIX path with the project's
own warning settings.

subprocess.h:1137: error: unused variable 'actions_created' [-Werror=unused-variable]
subprocess.h:1138: error: unused variable 'posix_error'
subprocess.h:1139: error: unused variable 'actions'

They are still declared at function scope in the #else branch, but the posix_spawn code
now redeclares them in the new inner scope, and the if (actions_created) cleanup is gone.
Dropping the outer three should be enough.

The AIX path itself looks sound. Forcing -D_AIX on Linux builds and passes the full
suite, 432/432 — chdir takes effect in the child and a missing binary reports
subprocess_error_not_found. (Rebased onto main you also need #104's SUBPROCESS_HAVE_CWD,
otherwise the 9 cwd tests are silently skipped.)

That is also a cheap answer to the CI question: the AIX branch can be compiled and
exercised on any Linux runner via -D_AIX, which would have caught the build error above.

One thing to expect on real AIX, unrelated to this PR: create_subprocess_fail_divzero
will fail in all 9 language modes. PowerPC does not trap on integer division by zero, so
the child never dies from SIGFPE. Measured on ppc64le Linux — x86 exits 136, POWER returns
a value and exits 0. Worth guarding the test by architecture rather than debugging it.

I'm trying to get access to an AIX machine to confirm the AIX-specific parts (execvpe,
posix_spawn behaviour) and will report back.

[bernardladenthin]

Follow-up: I now have AIX 7.2 TL04 (7200-04-02-2027) running on POWER8 with
GCC 13.3.0, so the AIX-specific parts are confirmed on a real system rather than
by forcing -D_AIX on Linux.

Both premises of this PR hold. posix_spawn_file_actions_addchdir_np is
absent from spawn.h and from libc.a, and execvpe is exported but declared
in no header — the comment in the diff is accurate word for word.

@sheredom, to your two questions:

Does #104 solve the cwd issue? No. #104 made the probe correct for old glibc
and old macOS, but AIX defines neither __GLIBC__ nor __APPLE__, so it falls
through to the #else and is told SUBPROCESS_HAVE_CWD 1. On AIX that is wrong,
and the build fails at link time:

subprocess.h:1265:19: warning: implicit declaration of function
  'posix_spawn_file_actions_addchdir_np'; did you mean 'posix_spawn_file_actions_addopen'?
ld: 0711-317 ERROR: Undefined symbol: .posix_spawn_file_actions_addchdir_np

The compiler's suggestion is worth noting: spawn.h is being read and the
sibling functions are declared. Exactly one is missing.

Is there any way to test for AIX? Yes, and cheaply. Select the
implementation through a macro rather than _AIX, and any Linux runner can
exercise the fork path:

-DSUBPROCESS_SPAWN_VIA_FORK=1   ->  441/441 on glibc

That is a better CI lever than -D_AIX, and it would have caught the build
error I reported earlier, which is still open on this branch.

---

What real AIX shows

On unpatched main, AIX needs -DSUBPROCESS_HAVE_CWD=0 merely to link, and
then fails 27 of 441 tests. Nine are a PowerPC property unrelated to this PR
(see below). The other 18 are these two, in all nine language modes:

create_ex_subprocess_create_failure_preserves_error
create_ex_subprocess_create_failure_does_not_leak_resources

  Expected : (subprocess_error_not_found) == (subprocess_create_ex(...))
    Actual : -4 vs 0

posix_spawn on AIX returns success for a binary that does not exist. So
SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS guesses wrong on AIX too, for the same
structural reason as SUBPROCESS_HAVE_CWD: the #else hands 1 to every
platform nobody thought about. This PR's error pipe fixes exactly those 18.

AIX is not the only one

Taken from each vendor's own spawn.h and libc, not from documentation:

	declares
FreeBSD 13.1–15.0	..._addchdir_np	fine
NetBSD 10.0, 10.1	..._addchdir, no _np	handled by #102
NetBSD 9.3, 9.4	neither	#102's guard has no version test
OpenBSD 7.7–7.9	neither	unhandled
AIX 7.2	neither	this PR

Where I got to

I have opened #112 with your commit e6cec1b included unchanged. It starts from
current main and resolves the conflict against #102, #104, #106 and #110
additively rather than replaying the restructuring: main's posix_spawn path
stays byte-identical and the fork path sits beside it, +183/-7 where this PR
is +227/-97, so no other platform can regress. On top of that the probe now asks
two questions — whether a chdir file action exists at all, and which spelling —
which covers OpenBSD and NetBSD 9 as well.

On AIX the result passes 441 of 441 with no override beyond -maix64.
Unchanged elsewhere: Linux glibc, musl and i386 441/441, Windows MSVC 387/387 in
x86 and x64 against both CRTs, clang with ASan and UBSan 441/441, and NetBSD
9.4/10.1, OpenBSD 7.9, FreeBSD 14.3 all compile clean against sysroots built from
the vendors' release sets. A test spawning from eight threads, 640 spawns, passes
on both paths under ASan — worth having, since fork() in a threaded process is
the risky part and the suite is single-threaded.

Two things that are not this PR's problem but block AIX in practice:

• utest.h has no AIX branch and stops at #error Unsupported platform!, and it
  does not build 32-bit at all, which is GCC's default on AIX. Fixed in
  Support AIX in utest_ns() utest.h#189 and Fix 32-bit builds under -Wpedantic utest.h#188.
• create_subprocess_fail_divzero cannot hold on PowerPC: divw/divd leave
  the result undefined and raise no exception, so the child exits 0. The test
  already excludes AArch64 for the same reason; RISC-V belongs there too. Also
  measured on ppc64le Linux, so it is architecture, not AIX.

[sheredom]

Closing in favour of #112