[PATCH v3] eal: fix unaligned loads/stores in rte_memcpy_generic

[Morten Brørup]

> From: Stephen Hemminger [mailto:stephen at networkplumber.org]
> Sent: Sunday, 16 January 2022 17.34
> 
> On Sun, 16 Jan 2022 09:33:19 -0500
> Luc Pelletier <lucp.at.work at gmail.com> wrote:
> 
> > As a side note, and to follow up on Stephen's indication that this is
> > 'performance critical code', I think it might be worthwhile to
> > revisit/revalidate the current implementation of rte_memcpy. There's
> a
> > good thread here that mentions rte_memcpy, and its performance on at
> > least one platform/architecture combination is far from being the
> > best:
> >
> > https://github.com/microsoft/mimalloc/issues/201
> >
> > It seems like enhanced rep movsb could be faster on more recent CPUs,
> > but that's currently not being used in the current implementation of
> > rte_memcpy.
> >
> > I understand some of this may not be directly related to this patch,
> > but whoever looks at this patch might want to provide their thoughts
> > on whether updating rte_memcpy would be worthwhile? I suspect looking
> > at all current public implementations of memcpy (libc, microsoft,
> > compilers builtin implementations, etc.) might help in making
> > improvements.
> 
> I would prefer that rte_memcpy did not exist at all.
> Instead the system library should always be used.
> 
> It is only exists because some architectures have slower code
> in glibc.

I wonder if that is still the case?

Otherwise, DPDK is probably full of obsolete optimizations, which should be eliminated like this:

http://inbox.dpdk.org/dev/20210918114930.245387-1-mail@gms.tf/