internal/fakecgo: preserve frame pointer across mstart call in threadentry

[gdams]

Fixes #485.

Problem

internal/fakecgo's threadentry calls runtime.mstart directly. For a fakecgo-created thread the M's stack is system-allocated, so when the M exits the runtime does mexit(osStack=true), which returns from mstart back into threadentry (by design, so the pthread can unwind and exit). But mstart runs the scheduler and comes back with BP = 0.

On Go 1.28 the amd64 compiler emits a frame-pointer LEAVE epilogue (MOV BP,SP; POP BP) for threadentry. With BP = 0, LEAVE sets SP = 0 and the following RET dereferences a NULL pointer:

#0  fakecgo.threadentry (...) at internal/fakecgo/go_linux.go:56    return nil
    => leave        rbp = 0x0        // SIGSEGV, si_addr = NULL
#1  threadentry_trampoline (trampolines_amd64.s)
#2  start_thread

Real cgo doesn't hit this because its x_cgo_threadentry invokes mstart through crosscall_amd64, which saves/restores the callee-saved registers (including BP) on the stack around the call. fakecgo's threadentry skipped that step.

This is mostly latent because Ms usually park and get reused rather than exiting. It surfaces whenever an M actually exits — e.g. a runtime.LockOSThread goroutine that finishes (which is how the Go standard library's os.TestProgWideChdir tripped it under CGO_ENABLED=0).

Fix

Route the ts.fn (runtime.mstart) call through a small assembly shim, callThreadEntryFn, that saves and restores the callee-saved registers on the stack around the call (using the existing PUSH_REGS_HOST_TO_ABI0 / POP_REGS_HOST_TO_ABI0 macros), mirroring crosscall_amd64. The existing call5 helper can't be reused because it stashes SP in BP (MOV SP,BP), which mstart zeroes.

• go_linux.go, go_darwin.go, go_freebsd.go, go_netbsd.go: call callThreadEntryFn(ts.fn).
• trampolines_amd64.s: the shim.
• threadentry_amd64.go: asm function declaration.
• threadentry_noasm.go: plain-call fallback for non-amd64 (unaffected — those epilogues don't rely on BP).
• threadentry_test.go: regression test that forces locked OS threads to exit.

Testing

• New TestThreadEntryReturn crashes reliably before this change and passes after.
• With the fix vendored into a CGO_ENABLED=0 + Go 1.28-devel toolchain, the standard library's os, syscall and runtime suites pass again.
• Cross-builds verified for linux/amd64, linux/arm64, linux/386, darwin/amd64, netbsd/amd64; gofmt clean.

[Copilot]

Pull request overview

This PR fixes an amd64-specific crash in internal/fakecgo when a fakecgo-created OS thread exits and runtime.mstart returns back into threadentry with BP clobbered, which can break the frame-pointer epilogue on newer Go toolchains.

Changes:

• Routes threadentry’s call to ts.fn (runtime.mstart) through a new amd64 assembly shim that saves/restores callee-saved registers (notably BP) around the call.
• Adds Go declarations / non-amd64 fallback for callThreadEntryFn.
• Adds a regression test that forces locked OS threads to exit to exercise the mstart return path.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
internal/fakecgo/trampolines_amd64.s	Adds callThreadEntryFn asm shim to preserve callee-saved regs (incl. frame pointer) across mstart.
internal/fakecgo/threadentry_repro_test.go	Adds regression test that reproduces the prior crash by forcing locked threads to exit.
internal/fakecgo/threadentry_noasm.go	Adds non-amd64 Go fallback implementation for callThreadEntryFn.
internal/fakecgo/threadentry_amd64.go	Adds amd64 Go declaration for the asm-implemented callThreadEntryFn.
internal/fakecgo/go_linux.go	Updates threadentry to call callThreadEntryFn(ts.fn) instead of directly invoking the raw PC.
internal/fakecgo/go_darwin.go	Same threadentry update for Darwin.
internal/fakecgo/go_freebsd.go	Same threadentry update for FreeBSD.
internal/fakecgo/go_netbsd.go	Same threadentry update for NetBSD.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[TotallyGamerJet]

LGTM

[hajimehoshi]

LGTM % comments, thanks!

[gdams]

@hajimehoshi updated PTAL

[hajimehoshi]

Still LGTM