On Fri, Oct 4, 2019 at 5:36 PM Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote: > > Just checking for Cortex-A7 being the boot CPU is probably > > sufficient, that takes care of the common case of all the > > A7-only embedded chips that people definitely are going to care > > about for a long time. > > > > But do you agree that disabling kernel mode NEON altogether for these > systems is probably more sensible than testing for CPU part IDs in an > arbitrary crypto driver? No. That NEON code is _still_ faster than the generic C code. But it is not as fast as the scalar code. There might be another primitive that has a fast NEON implementation but does not have a fast scalar implementation. The choice there would be between fast NEON and slow generic. In that case, we want fast NEON. Also, different algorithms lend themselves to different implementation strategies. Leave this up to the chacha code, as Zinc does it, since this is the place that has the most information to decide what it should be running.
