void uni_vmovups_tail_avx512(const Operand &dst, const Operand &src) {
     if (dst.isMEM())
         vmovups(dst.getAddress() | ktail_mask, Vmm(src.getIdx()));
     else
         vmovups(Vmm(dst.getIdx()) | ktail_mask | T_z, src.getAddress());
 }
 void uni_vmovups_tail_avx2(const Operand &dst, const Operand &src) {
     if (dst.isMEM())
         vmaskmovps(dst.getAddress(), vtail_mask, Vmm(src.getIdx()));
     else
         vmaskmovps(Vmm(dst.getIdx()), vtail_mask, src.getAddress());
 }