https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92768
--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> --- For SSE2+, the code can of course use _mm_xor_si128 instead and _mm_castsi128_ps/_mm_castps_si128, but for SSE that is not an option. And not really sure what can be done there, the _mm_xor_ps arguments are given, and even if we e.g. change the xorps builtin to take integral vectors, there still would be casts to the floating point vectors and still risk of the optimizers with -fno-signed-zeros to optimize that away.
