https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94715
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The GIMPLE transformation is correct and should stay as is.
User could have written int t = x * x; return t; instead.
What you are asking for is a new RTL optimization somewhere (dunno if in
expander, or where else), where e.g. using value ranges from the GIMPLE opts
the middle-end? sees that both sign and zero extension do what needs to be
performed because the value doesn't have msb set, and uses costs to decide
which one is faster.
# RANGE [0, 2147483647] NONZERO 2147483645
t_2 = x_1(D) * x_1(D);
# RANGE [0, 2147483647] NONZERO 2147483645
_3 = (long long unsigned int) t_2;
So, in this case it could be an optimization done in expand_expr_real_2 in the
CASE_CONVERT: case after the
else if (modifier == EXPAND_INITIALIZER)
case or so.
