https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88541
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Because AVX was what is in i386-builtins.def already and as the testcase shows, it works with just -mavx. You can not just load/store those vectors, you can also do logical operations on them etc. (through the floating point vector ops). With GFNI I believe the point was that the corresponding mode wasn't allowed in the registers without SSE2, so you'd get ICEs.
