Re: [Qemu-devel] [RFC PATCH v2 1/2] utils: Add helper to read arm MIDR_EL1 register

[Richard Henderson]

On 08/18/2016 02:39 AM, Paolo Bonzini wrote:
> On 18/08/2016 11:01, Vijay Kilari wrote:
> > On Thu, Aug 18, 2016 at 2:20 PM, Paolo Bonzini <address@hidden> wrote:
> > > On 18/08/2016 09:56, Vijay Kilari wrote:
> > > > The get_aarch_cpu_id() has check " if (unlikely(!cpu_info_read)) ".
> > > > If we call get_aarch_cpu_id() from is_thunderx_pass2_cpu() which is
> > > > called from inside the loop, we will be adding one additional check.
> > >
> > > On the other hand, you are making an assumption that the caller of
> > > is_thunderx_pass2_cpu() calls get_aarch64_cpu_id() first, and not
> > > documenting it anywhere.
> > >
> > > And given that you shouldn't call _any_ function from inside such a hot
> > > loop, your solution is inferior on both counts.
> >
> > Yes, but I could not think of better way to get rid of this check.
>
>     bool need_aa64_prefetch = is_thunderx_pass2();
>     for (...) {
>          if (need_aa64_prefetch) {
>              ...
>          }
>     }
>
> The check on cpu_info_read is done just once.

Supposing a check is required at all, this is still inferior to either

(1) If completely outside the loop,

  if (is_thunderx_pass2()) {
      for (...)
        ...
  } else {
      for (...)
  }

or (2) ifunc, so that we only check once, not every invocation.


r~