It happens somewhere after the compiler, so officially not obvious.

My working hypothesis is that something goes wrong here:

		pciterinit(ctxt, &pcstmt, &sym.Pcdata{P: s.FuncInfo.IsStmtSym.P})

		if pcstmt.done != 0 {
			// Assembly files lack a pcstmt section, we assume that every instruction
			// is a valid statement.
			pcstmt.value = 1
		}

in cmd/link/internal/ld/dwarf.go
The iterator is empty, so is_stmt is always true.

Thanks. I suspect where that's going wrong is

		s.FuncInfo.IsStmtSym = r.syms.Lookup(dwarf.IsStmtPrefix+s.Name, int(s.Version))

in objReader.readSym. The IsStmt symbol is probably always declared as version 0, but s.Version may be 1 now for an ABIInternal symbol.

This was fixed by https://go-review.googlesource.com/c/go/+/149605
and I can confirm this by my failure to reproduce the bug now.

This appears to be failing again for linux-amd64-longtest.

Latest failure: https://build.golang.org/log/ad327f4df87dc312a4840aa82dccbaf832822120

@dr2chase Can you look into this and let us know if it should still be a release blocker?

It appears the latest breakage is from CL 152537, which is unrelated to function ABIs. I've moved this into a new issue: #29511.