Re: [PATCH v2] aarch64: thunderx2 memcpy branches reordering

[Anton Youdkevitch <anton dot youdkevitch at bell-sw dot com>]

Adhemerval,

On 3/21/2019 23:38, Adhemerval Zanella wrote:
> If is just a non functional change, you should indicate it.  However it
> does change the code and thus for such refactor patches on arch-specific
> implementations we usually ask for, besides running the testcase for
> regression (which is also not indicate in your message in which platform
> you actually has tested), that at least some performance evaluation is
> provided (even if this shows no gain).
make check - no regressions
make bench - no performance regressions for memcpy

Should I resubmit the same patch with the description mentioning it is a
non functional change?


> On 21/03/2019 15:44, Anton Youdkevitch wrote:
> > Adhemaral,
> >
> > The reason I did the rewriting is two-fold. First, it
> > looks more clean as there aren't now two branches
> > immediately following each other. Second, it can be
> > performance beneficial as we are saving one branch
> > on entry for the most of the cases (we take it only if
> > the first iteration is a partial iteration). But, of course,
> > the performance benefit was not my concern here.
> > However, this is one instruction less anyway.
> >
> >
> > On 3/20/2019 19:29, Anton Youdkevitch wrote:
> > > Rewrote the branches in load and merge chunk
> > > so that the order is more in line with the
> > > most probable case.
> > >
> > > ChangeLog:
> > >      * sysdeps/aarch64/multiarch/memcpy_thunderx2.S:
> > >      branches reordering