Regarding spilling and register allocation, IIRC[1] the unwind tables contain a simple bytecode that is interpreted by the undwinder to fixup stack and restore registers (indexed by the current ip address), so no actual registration of object to destroy need to be done, nor does the frame pointer need to exist at all: the unwinder interprets the bytecode to recursively adjust each stack frame and registers to the expected values.
[1] At some point I had actually learned to write unwind tables by hand, of course I completely forgot everything about that: look at trampoline_thunk at https://github.com/gpderetta/delimited/blob/master/delimited... even with extensive comments is still write-only code. It doesn't help that there aren't a lot of DWARF CFI tutorials around.
I think I'll test whether use of exceptions can affect optimizations concerning register allocation and spilling, like function prologues and epilogues. Unwinding will also need to call destructors, and for that you'll need to have the object pointer somewhere in (stack frame) memory. But it might be in just a register. Unless, of course, the unwinding engine can look into registers as well.
I guess it's not a huge deal either way, especially in 32-bit x86 code. For large register file systems it might be a different story, forcing otherwise unnecessary spills. Regardless, I'll find out when I got some time.
Anyways, my point was more generic. So-called zero-cost abstractions are my pet peeve, because they're often not zero cost precisely because of larger generated code size. Typical microbenchmarks don't capture this, because their hot path can usually fit comfortably in L1 code and data caches.