https://bugs.kde.org/show_bug.cgi?id=481127
--- Comment #6 from Bruno Lathuilière <bruno.lathuili...@edf.fr> --- (In reply to Paul Floyd from comment #5) > Should we also support F16? No, there are no Iop_MAddF16 or IopMSubF16. And to my knowledge AVX512 is the only way to generate half floating-point operations. And AVX512 is not yet supported by valgrind. > Does this also work with the other permutations 132 and 231? The patch concern only the backend so we can choose the permutation. The frontend has to treat correctly all the permutations, this already tested by none/tests/fma.c. > Lastly, do packed and scalar make any difference? In the amd64 frontend the vectorized fma are unvectorized. So in the backend we do not have to treat packed version of fma. I think it would be nice to introduce new IOp in VEX : Iop_MAdd_F64x2 > This will need a regression test as well. -- You are receiving this mail because: You are watching all bug changes.
