https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63594
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Created attachment 33761 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=33761&action=edit WIP patch for discussions >From what I see, if TARGET_AVX512BW is not defined, then we obviously can't use ix86_vector_duplicate_value, but need two instructions (either it can be QI->V32QI / HI->V16HI broadcast followed by concat of the two parts, or QI->V16QI / HI->V8HI broadcast followed by concat of the 4 parts together). But, it seems even for -mavx2 or -mavx we actually generate terrible code, for -mavx2 there is no point in using 2 instructions when in theory vpbroadcast{b,w} should handle it alone just fine. The patch enables all of that, but unfortunately we generate perhaps not so good code with it, e.g. for -mavx2 in testchar32, we spill the argument always to memory, and then broadcast it from memory, even when vmovd + broadcast from register could have been used. And in testchar16, for some reason we spill into memory, and broadcast from vmovd result (so the spill is totally useless). Uros/Kyrill, any thoughts on this?
