It doesn't seem right to infer bfd_arch_mips_per_address when trying
to figure out whether the ABI uses 32- or 64-bit addresses.  Back when
we rejected inter-linking of code for 32-bit and 64-bit ISAs, we'd
probably always get it right, but this is now allowed, so this could
possibly break o32 (and perhaps other 32-bit arches).  I unfortunately
have no idea of how to trigger this particular code chunk (defining a
function with attribute constructor didn't do it), so this hasn't got
much testing.  Ok to install anyway?