https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115242
--- Comment #6 from Florian Weimer <fw at gcc dot gnu.org> --- (In reply to Alisa Sireneva from comment #4) > I think this is a wider issue. The root of the problem is that > __builtin_unwind_init() affects what it _thinks_ are _callee-saved_ > registers. I think this should be fine. The functions use the platform calling convention. If they are called from a function that preserves additional registers, the compiler needs to emit code to spill those registers before calling those functions. As far as I understand it, the unwinder code tracks a larger register set than just the callee-saved registers, so this should work. > In this bug, the compiler thinks a register doesn't exist when it does. But > it's also possible that the compiler thinks a register is caller-saved while > it's actually callee-saved, due to differences in ABI between libgcc_s and > user code (as reproduced e.g. on x64 by throwing from > __attribute__((ms_abi)), but should be possible with plain forced unwinding > too). According to a quick test, GCC handles this correctly and spills the additional registers when calling a non-ms_abi function from an ms_abi function. I expect that it's also possible to compensate for the lack of vector register tracking in the GNU/Linux x86-64 unwinder through appropriate code generation, but I haven't got a good test case for that. The issue on POWER is different because the ABI was enhanced retroactively, supposedly in a transparent way.
