https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97459
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Or: unsigned r3_128u_v4 (__uint128_t n) { unsigned long a; a = (n >> 96); a += (n >> 64) & 0xffffffffULL; a += (n >> 32) & 0xffffffffULL; a += (n & 0xffffffffULL); return a % 3; } if the target doesn't have (efficient) multi-word shifts.