but that doesn't mean that program shouldn't be optimized to "mov eax, 5; ret". The hilarious bit is that the whole exception throw bullshit still happens, even though control flow analysis determined that "mov eax, 5; ret" is what the program actually is.
That's what's expected to happen, but there's no reason __cxa_throw et al have to do what they're supposed to, when they're being loaded from a shared object at run time. Case in point: the demo library mentioned at top of thread. Maybe if you compiled the whole program including libgcc_s (or equivalent) and statically linked them with flto? Admittedly maybe optimizing out eh table searching might be a bit of an ask.
Now, if you make compilers do even more stuff upon an exception throw, we'll get even further from having reasonable exceptions.
I'm sure they can optimize out an extra allocation. They've been able to do that for quite some time.