This would allow mixing armel and softvfp ABIs, but not hard float ABIs, at least across compilation unit boundaries (that said, GCC never seems to optimize ABI bottlenecks within a compilation unit anyway)