Re: [RFC 1/8] net/ena: fix warnings related to rte_memcpy and gcc-12

[Stephen Hemminger <stephen@networkplumber.org>]

On Wed, 8 Jun 2022 21:18:15 +0200
Michał Krawczyk <mk@semihalf.com> wrote:

> śr., 8 cze 2022 o 17:32 Stephen Hemminger <stephen@networkplumber.org>
> napisał(a):
> >
> > On Wed, 8 Jun 2022 14:29:58 +0200
> > Michał Krawczyk <mk@semihalf.com> wrote:
> >  
> > > wt., 7 cze 2022 o 19:17 Stephen Hemminger <stephen@networkplumber.org>
> > > napisał(a):  
> > > >
> > > > Rte_memcpy is not needed for small objects only used on control
> > > > path. Regular memcpy is as fast or faster and there is more
> > > > robust since static analysis etc knows what it does.
> > > >
> > > > In this driver it was redefining all memcpy as rte_memcpy
> > > > which is even worse.  
> > >
> > > Hi Stephen,
> > >
> > > I would like to shed some light on why we're redefining all the memcpy
> > > as rte_memcpy. The ENA HAL is unmodifiable, as it's shared across many
> > > platforms and we cannot simply adjust it for the DPDK. We can use the
> > > ena_plat_dpdk.h to change the ena_com (HAL) definitions, and that's
> > > what we're doing with memcpy. It's being used on the data path for the
> > > Tx, to copy the bounce buffers. Following the recommendations in [1]
> > > plus the results from [2], we wanted to make use of the optimized
> > > memcpy on the ENA's data path as well to reduce the CPU time spent in
> > > the PMD. I'm worried that removing rte_memcpy from the ena_plat_dpdk.h
> > > will result in some performance degradation for the ENA data path.
> > > However I understand your concerns for the control path and I'm ok
> > > with it.
> > >
> > > [1] https://doc.dpdk.org/guides/prog_guide/writing_efficient_code.html#memory
> > > [2] https://www.intel.com/content/www/us/en/developer/articles/technical/performance-optimization-of-memcpy-in-dpdk.html
> > >
> > > Thanks,
> > > Michal
> > >  
> >
> >
> > I admit to having little sympathy unfixable for base/ style code.
> > You could have just replaced memcpy() in their with an abstraction layer
> > like other drivers.
> >  
> 
> We'll probably end up with the solution you're suggesting. For now
> let's remove the memcpy redefinition at all to suppress the warnings.
> 
> Acked-by: Michal Krawczyk <mk@semiahalf.com>

Lets see if we can fix rte_memcpy() on x86 first.

It seems to me that rte_memcpy() should be an inline that only handles variable
size data, and use __builtin_memcpy() automatically for fixed size values.