https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114121
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Indeed. And -O2 -fno-tree-vectorize works.
I've changed it to
unsigned a, b, c, d, e;
unsigned _BitInt(256) f;
__attribute__((noipa)) unsigned short
bswap16 (int t)
{
return __builtin_bswap16 (t);
}
void
foo (unsigned z, unsigned _BitInt(512) y, unsigned *r)
{
unsigned t = __builtin_sub_overflow_p (0, y << 509, f);
z *= bswap16 (t);
d = __builtin_sub_overflow_p (c, 3, (unsigned _BitInt(512)) 0);
unsigned q = z + c + b;
unsigned short n = q >> (8 + a);
*r = b + e + n;
}
int
main ()
{
unsigned x;
foo (8, 2, &x);
if (x != 8)
__builtin_abort ();
}
and bswap16 is called with 1 with -O2 -fno-tree-vectorize and 0 with -O2, so
the problem is either during the computation of y << 509 (but that is fairly
simple thing out of bitintlower, set highest limb to the lowest limb << 61 and
clear all others),
or during the sub overflow.
