https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83641
--- Comment #3 from Jeffrey A. Law <law at redhat dot com> --- So the issue here is when we have a noreturn function we use a push/pop sequence to probe the top of the stack. The generic dwarf2 CFI bits interpret the pop as restoring the value of the popped register. But that's not strictly correct if the register is also callee-saved (which implies that it'll be set later by the body of the function). Of course the push/pop sequence is not needed in this case since the callee register save acts as a probe. This is a known, relatively minor, inefficiency. The good news is the bogus CFI issue is a strict subset of the performance issue and thus if we fix the performance issue, the codegen issue gets resolved at the same time. I expect to have this fixed within the next couple hours.
