vms-40b: Alpha struct stat/sigaction layouts + stat64 wrappers

[baron-3dl]

Summary

Rung A1 of the OVMX-on-Alpha epic (rd vms-40b). A0 landed the Alpha freestanding backend (src/libvmssys/arch/alpha/{crt0,syscall,sigreturn}.S + the __alpha__ syscall table) but deliberately left struct vms_stat and struct vms_sigaction undefined for Alpha, so any code path touching them failed loud (compiler warning) rather than silently reading a wrong layout. This PR defines both, and proves the whole tests/libvmssys suite green under qemu-alpha.

Struct layouts — provenance

Both verified byte-for-byte by compiling offsetof/sizeof probes against the verbatim struct definitions from the real kernel headers and running them under qemu-alpha (not derived from memory):

• struct vms_stat ← arch/alpha/include/uapi/asm/stat.h, struct stat64, Linux 6.6.52 source (same header shipped by the gcc-alpha-linux-gnu/libc6-dev-alpha-cross cross toolchain). This is what fstat64/fstatat64 fill.
• struct vms_sigaction ← arch/alpha/include/uapi/asm/signal.h, struct sigaction (same shape in glibc's alpha bits/sigaction.h). Order is handler/mask/flags — not flags-then-mask like x86_64/aarch64 — and there is no sa_restorer field.

Wrapper fixes in vms_syscall.h:

• vms_sys_fstat now calls __NR_fstat64 (427) instead of the old __NR_fstat (91, a narrower non-stat64 struct) — the known trap from the rung brief.
• vms_sys_rt_sigaction gets an Alpha-specific 5-arg body: Alpha's rt_sigaction is SYSCALL_DEFINE5(rt_sigaction, sig, act, oact, sigsetsize, restorer) — the restorer is an explicit 5th syscall argument, not a struct field.

Four more pre-existing gaps found proving the gate

None of these are caused by the struct work above — all were already blocking tests/libvmssys under qemu-alpha before this PR, just invisible because nothing had tried to run the suite there yet.

1. No hardware integer divide/remainder. Alpha has no div/idiv-equivalent; GCC calls __divqu/__remqu. Added as hand-written assembly (arch/alpha/softdiv.S) because these use a non-standard calling convention (args in $24/$25, return address in $23 not $26, result in $27, and — the part that cost real debugging time — the callee must preserve $1-$8 because GCC's div+mod fusion (e.g. vms_snprintf.c's uint_to_str) keeps the original dividend live across the call in an otherwise-volatile register).
2. No inline __builtin_sqrt/floor/ceil lowering on Alpha's GCC backend — always a libcall (confirmed via -S output even with -mieee). Added freestanding implementations (arch/alpha/softmath.c): schoolbook division, Newton-Raphson sqrt from a bit-halved exponent guess, IEEE-754 mantissa-mask floor/ceil. Both this and authenticity: INV-1 system-identity SSOT + logical-driven login banners (vms-e652) #1 must avoid glibc/libgcc (Rule 3), hence hand-rolled rather than linked from the cross sysroot.
3. Alpha's mmap flag bits are not the generic Linux ones (arch/alpha/include/uapi/asm/mman.h): MAP_FIXED/MAP_ANONYMOUS/MAP_NORESERVE all differ. With the generic values, test_vmssys_syscall's mmap test got a small negative errno back that didn't equal VMS_MAP_FAILED, so the test dereferenced it and segfaulted.
4. No TLS thread-pointer setup in crt0. Alpha sets the TLS base via PALcode (rduniq/wruniq, not a syscall). The TCB header offset (16 bytes, Variant I) isn't published in the toolchain headers, and installed glibc's own __tls_init_tp never calls wruniq at all (it assumes the kernel already set "unique" for the initial thread — an assumption our freestanding CRT can't rely on). Pinned it empirically: a probe filled a scratch buffer with a marker at a known offset, swept wruniq() across candidate base addresses, and asked a real __thread variable (in a separate TU, so the compiler couldn't constant-fold it away) which value it read back. Exactly one offset matched.
5. Alpha's signal numbers are OSF/1, not generic Linux (e.g. SIGCHLD=20, not 17; several others renumbered). test_vmssys_futex's clone()-based fork emulation passed the generic VMS_SIGCHLD (17, which is Alpha's real SIGSTOP) as the clone exit signal and got EINVAL back every time; Alpha's real SIGCHLD (20) succeeds. Also found and removed a stale duplicate #define VMS_SIGCHLD 17 elsewhere in vms_types.h that was silently clobbering the per-arch value via macro redefinition.

Gate — qemu-alpha evidence

$ cmake -S src/libvmssys -B build-alpha -DCMAKE_SYSTEM_NAME=Linux -DCMAKE_SYSTEM_PROCESSOR=alpha \
    -DCMAKE_C_COMPILER=alpha-linux-gnu-gcc -DCMAKE_CROSSCOMPILING_EMULATOR=qemu-alpha
$ cmake --build build-alpha && ctest --test-dir build-alpha --output-on-failure

Test project .../build-alpha
    Start 1: test_vmssys_syscall
1/7 Test #1: test_vmssys_syscall ..............   Passed    0.00 sec
    Start 2: test_vmssys_string
2/7 Test #2: test_vmssys_string ...............   Passed    0.00 sec
    Start 3: test_vmssys_snprintf
3/7 Test #3: test_vmssys_snprintf .............   Passed    0.00 sec
    Start 4: test_vmssys_futex
4/7 Test #4: test_vmssys_futex ................   Passed    0.01 sec
    Start 5: test_vmssys_stdio
5/7 Test #5: test_vmssys_stdio ................   Passed    0.00 sec
    Start 6: test_vmssys_math
6/7 Test #6: test_vmssys_math .................   Passed    0.01 sec
    Start 7: test_vmssys_crt
7/7 Test #7: test_vmssys_crt ..................   Passed    0.00 sec

100% tests passed, 0 tests failed out of 7

The Alpha "declared inside parameter list" warnings are gone (confirmed clean -Wall -Wextra -ffreestanding -fno-builtin compile of every src/libvmssys/*.c).

Also needed (unrelated to the struct/ABI work, but required to make the qemu-alpha gate runnable at all): enable_language(ASM) and enable_testing() when src/libvmssys/CMakeLists.txt is configured standalone (cmake -S src/libvmssys), which the alpha gate does — previously only exercised by the netbsd-vax path, which needs neither.

x86_64 / aarch64 unchanged

• x86_64: full-tree cmake -B build -DCMAKE_BUILD_TYPE=Debug -DBUILD_TESTS=ON -DBUILD_TOOLS=ON configures clean; ctest -R vmssys → 7/7 pass.
• aarch64: standalone cross-compile (-DCMAKE_SYSTEM_PROCESSOR=aarch64 -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc) builds clean, no warnings.

Neither arch touches softdiv.S/softmath.c (Alpha-only in CMakeLists) or the Alpha branches of vms_types.h/vms_syscall.h/vms_runtime_init.c.

Test plan

• tests/libvmssys cross-compiles for alpha-linux-gnu with zero warnings/errors
• All 7 tests/libvmssys suites pass under qemu-alpha (shown above)
• x86_64 full-tree configure + ctest -R vmssys unaffected (7/7 pass)
• aarch64 standalone cross-compile unaffected (clean build)

🤖 Generated with Claude Code