Additional OS targets (BSD / macOS via syscall mapping)

• Type: feature
• Status: rainy-day
• Owner: —
• Opened: 2026-06-16

Premise

PXX emits static, syscall-only binaries (no libc, no ld.so, no shared libs; ldd → "not a dynamic executable"). The only host contract is the kernel syscall ABI. So porting to another Unix-like OS is mostly mapping out its syscall numbers + ABI, not a userland/libc port.

Scope

• BSD (Free/Open/Net): already understood to be "a matter of mapping syscalls" — different numbers, same static-binary model. Bounded.
• macOS: expected to be the same shape — map the Mach/BSD syscall layer + the Mach-O object format (PXX currently emits ELF; macOS needs Mach-O, so a second object writer alongside elfwriter.inc). The codegen/backends are unchanged; it's an OS-ABI + container-format port.
• Windows: a real future target, tracked separately — [[feature-port-windows-pe]] (Track W). Different shape from BSD/macOS: not a static syscall-only binary (PE + Win32/NT syscalls aren't a stable public contract), but a CRT-free thin binding against kernel32/ntdll — Windows' own stable standard library, the same role libc.so plays for OpenBSD. This note previously said "deprioritized, not on the roadmap" (2026-06-16), predating and contradicting the detailed Windows ticket opened a month later; corrected 2026-08-01.

Notes

• NOT only syscall numbers — a few Linux /proc-based readbacks need OS equivalents too. The static-binary model maps cleanly, but some runtime code reads Linux procfs text, which BSD/macOS do not have:

  • CPU load sampler (lib/rtl/palparallel.pas, PXXQueryFreeCores for the parallel-for pwLoadOnce/pwLoadCont policies) reads /proc/stat. BSD → sysctl kern.cp_time (per-CPU jiffies); macOS → host_processor_info / sysctlbyname. No /proc/stat there. It already fails SAFE (falls back to the fixed worker count), so this is a feature-parity item, not a blocker.
  • CPU count uses sched_getaffinity (palparallel); BSD/macOS → sysctl hw.ncpu / hw.activecpu.
  • Environment variables (GetEnvironmentVariable in sysutils, NilPy os.environ/os.getenv, C getenv — landed 2026-07-31) all read /proc/self/environ. BSD/macOS have no /proc by default either — needs the envp-off-the-initial-stack route the environment-variable ticket considered and skipped for Linux (/proc was simpler there), or a per-OS equivalent. Also note: /proc isn't guaranteed mounted even on Linux (containers, minimal installs) — this is a live gap there too, not only a porting concern.
  • Audit for any other /proc/* or Linux-specific sysfs reads when porting (grep '/proc/' across lib/rtl/**). Treat the sampler/affinity as part of the per-OS port, alongside the syscall table.

• This is about the compiler binary's host portability and what users can target. The static-syscall-only design is what makes it cheap (see the kernel-only portability property).

• Per-OS work = syscall table + entry/exit ABI + object/container writer (ELF done; Mach-O / PE per target). Backends (x86-64/aarch64/arm32/...) are reused as-is.

• Distro packaging stays trivial: source + a static binary + a man page.