https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117823
--- Comment #3 from Hongtao Liu <liuhongt at gcc dot gnu.org> --- > Whether it needs -ffast-math depends on how it behaves with respect to > rounding I guess. If (float)bf16 * (float)bf16 + (float)bf16 * (float)bf16 > performs the float add without intermediate rounding for the multiplications > it falls under the same umbrella as FMA - is there more actual difference? I think fassociative-math is needed.