https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120598
--- Comment #7 from Segher Boessenkool <segher at gcc dot gnu.org> --- [I cannot read any of the attached code, but...] The proposed manually vectorised code converts 64-bit integers to IEEE SP floats, which is extremely lossy. I don't find it very surprising the compiler does not automatically do such things :-)
