https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122103
--- Comment #27 from Richard Biener <rguenth at gcc dot gnu.org> ---
There's definitely a gap in gimple_could_trap_p_1 handling of calls. Any
builtin is handled as
tree type = TREE_TYPE (gimple_call_fntype (s));
bool fp_operation = FLOAT_TYPE_P (type);
return fp_operation && flag_trapping_math;
so the float-to-integer builtins can be handled explicitly. But then any
call to a regular function is also classified as not trapping which is
of course not true either (but usually harmless unless the function is
also const or pure in which case it's unspecified whether setting IEEE
flags would be considered making const/pure an invalid specification).
It might be good to raise this in a separate bug.
