https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70216
--- Comment #10 from Oleg Endo <olegendo at gcc dot gnu.org> --- (In reply to Rich Felker from comment #9) > From a Linux standpoint, there is no trapa trap number defined that would > cause a fatal signal. The ones that are defined are for syscalls and debug > breakpoints. On the other hand, a permanently-undefined opcode would cause > SIGILL. So for Linux targets with aim of generating code that does what you > intend for it to do on both existing and future kernels, I think it makes > much more sense to use FFFD. This also aligns with the behavior of > __builtin_trap() on other targets, which is almost always to execute an > instruction that causes SIGILL (or, if unavailable, SIGSEGV). > > I'm not opposed to supporting a -m option to choose a trapa trap instead, > and/or making something with trapa the default for freestanding/non-linux > targets. So my patch from Comment 6 could be extended to handle the option -mbuiltin-trap=insn-FFFD and define that as the default for linux configs. Would that do the job? > FYI this issue is currently a regression that prevents building Linux with > gcc7, since gcc7 introduced an optimization that transforms x/0 to > __builtin_trap() rather than calling the libgcc div function. It's OK to add __builtin_trap to GCC 7. Could you have a look and try the patch in Comment 6? I don't have so much time for SH stuff these days...
