https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93742
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|normal |enhancement
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
One thing which could be done is take:
_1 = a_10(D) > 0;
_2 = a_10(D) < 0;
_3 = 2147483647 - a_10(D); // INT_MAX - a
_4 = _3 < b_11(D);
_9 = _1 & _4;
_5 = -2147483648 - a_10(D); // INT_MIN - a
_6 = _5 > b_11(D);
_13 = _2 & _6;
_14 = _9 | _13;
And recognize that as
_5 = .ADD_OVERFLOW (a_3(D), b_4(D));
_1 = IMAGPART_EXPR <_5>;
But that won't catch the original testcase right away though as the incoming
gimple has a lot of branches.
The above was provided by:
bool add_overflow(int a, int b, int *res) {
int t= 0;
int t1;
int ty = a > 0;
int ty1 = a < 0;
int tt = ty & b > INT_MAX - a;
int tt1 = ty1 & b < INT_MIN - a;
if (tt | tt1) {
t1= 1;
} else {
*res = a + b;
t1= 0;
}
return t1;
}
Note not even clang is able to detect the above really.
