__cxa_finalize makes it hard to implement opposite __cxa_atexit order

[fweimer-rh]

The specification is pretty clear that __cxa_finalize should be called via the ELF destructor mechanism and invoke the registered __cxa_atexit callbacks for that particular shared object. But the ELF destructor order is not necessarily sequenced in a compatible way with the opposite __cxa_atexit order. This results in the wrong sequence of C++ destructor calls.

This specification seems to date back to an era where C++ implementations without a way to integrate with the system dynamic linker were common. Nowadays, I think we can do better by teaching the dynamic linker to invoke __cxa_atexit callbacks in the exact opposite order (regardless of DSOs), and do nothing in __cxa_finalize. But the current specification seems to preclude such a standards-conforming implementation, by explicitly tying the callbacks to the ELF destructor invocation.

(@jwakely brought this up recently.)