[unxed]

Step 5 of the split agreed in #76. Depends on #83. This branch contains #83's commit, so review only the last commit. Once #83 is merged I'll rebase and the diff will drop that commit.

NetBSD amd64/arm64, full tier. These reuse the System V and AAPCS64 backends unchanged. The NetBSD-specific parts are dlopen from libc.so and errno via __errno (internal/dl/dl_netbsd.go, internal/syscall/errno_netbsd.go), and netbsd is added to the Unix build constraints. Like FreeBSD, NetBSD needs -gcflags=github.com/go-webgpu/goffi/internal/fakecgo=-std.

One real fix is included. internal/fakecgo/netbsd.go defined environ, __progname and __ps_strings with //go:linkname only. That gives local symbols, but NetBSD's libc.so expects rtld to resolve them from the main program, so the process would die before main. They are now also exported with //go:cgo_export_dynamic, as freebsd.go already does.

scripts/check-platforms.sh replaces the cross-compile job. It covers every target purego supports under CGO_ENABLED=0. For each target it:

• builds ./... and links a consumer binary, and checks the tier the target claims (full, nocb, load or pending);
• on targets that use goffi's own assembly callback trampolines, requires ffi.callbackTrampoline to be a text symbol. This is fix: build callback_amd64.s on FreeBSD #83's check-callback-trampolines.sh merged into the new script, now also run on netbsd;
• expects pending targets (linux 386/arm/loong64/ppc64le/riscv64/s390x) to keep failing. If one starts building while the table still says pending, the script fails.

It also builds the examples.

purego-coexistence job. It builds a CGO_ENABLED=0 -tags nofakecgo program that imports both purego and goffi, calls getpid through each, and compares the results with os.Getpid().

Docs: docs/PLATFORMS.md covers the tiers, the build flags and the purego parity diff. It also records that windows/386 is loader-only (CallFunction returns ErrUnsupportedArchitecture); that target already built, but the docs never said so. Also updated: the README platform table, a ROADMAP section on the remaining Linux architectures, and CHANGELOG.

NetBSD support is checked by cross-compiling and linking only. It has not run on a NetBSD host.

The Android arm64 jobs fail in "Set up Android SDK" before any goffi code runs. That failure is unrelated to this PR and is fixed by #85.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[kolkov]

Thanks. A single script that owns the platform table and fails when a pending target starts building is a good idea, and the purego-coexistence job covers a combination users keep asking about (#22, #72). I reviewed only the last commit, as you suggested.

Verified locally (Go 1.27.1): scripts/check-platforms.sh passes, with all 12 supported targets building and linking and the six pending ones failing as expected. go vet and go test -count=1 ./... pass too.

Blocking

1. NetBSD is declared full tier on cross-compilation alone. #83 is the cautionary tale here: FreeBSD amd64 was "cross-compile verified" for months while every callback pointed into a data page. purego tests FreeBSD and NetBSD in a real VM on every PR (vmactions/freebsd-vm / vmactions/netbsd-vm, see their .github/workflows/test.yml bsd job and .github/scripts/bsd_tests.sh). Could you add the same for goffi, at least go test ./ffi with CGO_ENABLED=0 and the fakecgo -std flag on both BSDs? Without a runtime run, NetBSD should be listed as something like full (untested), not alongside the runtime-tested targets.

2. The environ/__progname/__ps_strings export: the claim is unverified and contradicted by purego. The CHANGELOG and docs/PLATFORMS.md say that without //go:cgo_export_dynamic "the process would have died before main", and PLATFORMS.md goes on to say goffi is "ahead of purego" because of it. purego's current internal/fakecgo/netbsd.go has only the //go:linkname lines, the same as Go's own runtime/cgo/netbsd.go, and purego's NetBSD VM job runs the whole test suite with CGO_ENABLED=0 and passes (for example the 2026-10-15 run on main: NetBSD on Go 1.25.8 and 1.26.1 both green). So either rtld resolves these differently than described, or something else covers it. The export may well be harmless, but please:

• drop the "ahead of purego … can actually start under rtld" sentence. We shouldn't publish a claim that a sibling project's tested platform is broken without evidence;
• reword the CHANGELOG entry so it describes the change, not a crash nobody has observed. Or back it with a NetBSD run with and without the export, which item 1 would provide.

3. RTLD_LOCAL is wrong for NetBSD. NetBSD's <dlfcn.h> has #define RTLD_LOCAL 0x200 (FreeBSD and Linux use 0). goffi itself doesn't pass RTLD_LOCAL on NetBSD today, but the constant and the comment ("match the FreeBSD/Linux numbering") are wrong. purego carries the same mistake, so that's probably where it came from.

4. Rebase onto current main before NetBSD goes in. Our internal/fakecgo defined SIG_BLOCK/SIG_UNBLOCK/SIG_SETMASK once with the Linux values 0/1/2. NetBSD (like Darwin and FreeBSD) uses 1/2/3 (sys/sys/signal.h:242-244), so _cgo_sys_thread_start would unblock every signal across pthread_create instead of blocking them. #90 (ported from purego#553, now merged) moved the constants into the per-OS files, including libcgo_netbsd.go, so NetBSD gets the right values once this branch is rebased.

#92 (also merged) went further on the fakecgo side of NetBSD, and part of this PR is now on main:

• trampolines_{amd64,arm64}.s build on NetBSD, the same tag change as here.
• Three layouts that cross-compiling can't catch are fixed. stack_t had the Linux field order (purego#433). pthread_mutex_t and pthread_cond_t were 8-byte integers, but on NetBSD they are structs with a magic field that libpthread checks, so going by NetBSD's libpthread source, locking the runtime-init mutex returns EINVAL without locking.
• The layouts are asserted at compile time, and CI vets internal/fakecgo for netbsd.

This branch doesn't have the layout fixes. That is item 1 again: nothing short of a run on NetBSD would have shown them. I haven't run NetBSD either, which is why main still doesn't list it as supported. After the rebase, the only fakecgo change left here is the netbsd.go export from item 2.

5. Required status check rename. cross-compile / "Cross-Compile" is a required check in branch protection, and this PR removes the job, so even after the rebase it would stay blocked waiting for a check that never reports. That's on our side. Just flagging that we'll swap it for "Platform Matrix" (and decide about "purego coexistence") when we merge, so nobody is surprised by the state. One rebase detail: on main, the cross-compile job that this PR replaces now also runs the FreeBSD test build with the libthr.so.3 DT_NEEDED check and the NetBSD fakecgo vet. Those steps need to move into platform-matrix (or stay in a job of their own). lint-targets (added in #92) would need netbsd rows once NetBSD becomes a supported target.

Non-blocking

6. Pin purego in purego-coexistence. go get github.com/ebitengine/purego takes whatever was released last, so a purego release can turn our CI red without any change here. Pin a version (and bump it deliberately).
7. The script checks compile + link + trampoline symbol, not tiers. Both the script header ("the script then proves the tier") and PLATFORMS.md ("a row here that disagrees with reality breaks the build") promise more. nocb and load are only built, and NewCallback failing or CallFunction returning ErrUnsupportedArchitecture is never asserted. Either tone the wording down or add small assertions.
8. README vs PLATFORMS.md disagree. Windows arm64 is "Tested (Snapdragon X)" in README but "Cross-compile verified" in PLATFORMS.md. macOS arm64 is "Tested (M3 Pro)" vs "Runtime tested (M-series)", although macos-latest is arm64 and it runs in CI. One table should be the source and the other should match it.
9. internal/arch/amd64/call_unix.s is an intentionally empty file; widening its constraint is harmless, it's just noise in the diff.

The ROADMAP "Architecture Expansion" section is good, and the ground rule there ("no architecture is promoted out of pending on cross-compilation alone") is the right one. Items 1–2 are really just that rule applied to NetBSD.

Once #83 is merged, a rebase will make this diff a lot easier to read.