https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92747
--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> --- On Mon, 2 Dec 2019, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92747 > > Jakub Jelinek <jakub at gcc dot gnu.org> changed: > > What |Removed |Added > ---------------------------------------------------------------------------- > CC| |jakub at gcc dot gnu.org > > --- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > The generic builtins don't care about -fnon-call-exceptions though (e.g. sin > is > nothrow even with -fnon-call-exceptions), so for the start we might just use > nothrow/leaf for all x86 backend builtins for start. Though, eventually say > builtins that store something into memory (or load from memory) or might be > trapping should probably not be nothrow, though -fnon-call-exceptions is > Optimization flag and thus can differ between functions, so e.g. deciding > based > on the global flag_non_call_exceptions during construction of the builtin is > problematic. One trick would be to keep the default conservative and in ix86_gimple_fold_builtin adjust the behavior per call via gimple_call_set_nothrow. That would happen in the gimple lowering pass (before EH/CFG build).
