http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57331
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Reduced testcase without missing return:
int
foo (int x)
{
void *p = x ? (void *) 1 : (void *) 0;
__INTPTR_TYPE__ b = (__INTPTR_TYPE__) p;
if (b)
return 0;
return 1;
}
The problem is that int_fits_type_p doesn't handle POINTER_TYPE_P on rhs.
range_fits_type_p should already ensure that innerop is INTEGER_TYPE_P or
POINTER_TYPE_P. So, either we can give up for POINTER_TYPE_P (TREE_TYPE
(innerop)) always, or we could handle just the most important pointer constant
(NULL), instead of the unconditional int_fits_type_p do something like
&& (POINTER_TYPE_P (TREE_TYPE (innerop))
? integer_zerop (op1)
: int_fits_type_p (op1, TREE_TYPE (innerop))))
(0 will fit in any pointer type), or we'd need to write larger code to do
essentially what int_fits_type_p does, but using lower_bound_in_type and
upper_bound_in_type instead of TYPE_{MIN,MAX}_VALUE. I'd say the last one
would be overkill.
