RE: [PATCH v8] eal/x86: improve rte_memcpy const size 16 performance

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Konstantin Ananyev [mailto:konstantin.ananyev@huawei.com]
> Sent: Monday, 10 June 2024 15.40
> 
> > When the rte_memcpy() size is 16, the same 16 bytes are copied twice.
> > In the case where the size is known to be 16 at build time, omit the
> > duplicate copy.
> >
> > Reduced the amount of effectively copy-pasted code by using #ifdef
> > inside functions instead of outside functions.
> >
> > Suggested-by: Stephen Hemminger <stephen@networkplumber.org>
> > Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> > Acked-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> > Depends-on: series-31578 ("provide toolchain abstracted
> __builtin_constant_p")
> >
> > v8:
> > * Keep trying to fix that CI does not understand the dependency...
> >   Depend on series instead of patch. Github only understands series.
> > * Fix typo in patch description.
> > v7:
> > * Keep trying to fix that CI does not understand the dependency...
> >   Depend on patch instead of series.
> >   Move dependency out of the patch description itself, and down to the
> >   version log.
> > v6:
> > * Trying to fix CI not understanding dependency...
> >   Don't wrap dependency line.
> > v5:
> > * Fix for building with MSVC:
> >   Use __rte_constant() instead of __builtin_constant_p().
> >   Add dependency on patch providing __rte_constant().
> > v4:
> > * There are no problems compiling AVX2, only AVX. (Bruce Richardson)
> > v3:
> > * AVX2 is a superset of AVX;
> >   for a block of AVX code, testing for AVX suffices. (Bruce Richardson)
> > * Define RTE_MEMCPY_AVX if AVX is available, to avoid copy-pasting the
> >   check for older GCC version. (Bruce Richardson)
> > v2:
> > * For GCC, version 11 is required for proper AVX handling;
> >   if older GCC version, treat AVX as SSE.
> >   Clang does not have this issue.
> >   Note: Original code always treated AVX as SSE, regardless of compiler.
> > * Do not add copyright. (Stephen Hemminger)
> 
> Acked-by: Konstantin Ananyev <konstantin.ananyev@huawei.com>
> 
> The code change itself -  LGTM.
> Out of interest - do you expect any perf diff with these changes?

I don't expect a significant perf diff with these changes, not even for the affected special cases. But the generated code (for the affected cases) is smaller.

Stephen noticed that the code generated from rte_memcpy() was inefficient in some cases [1], so I decided to fix it.

[1]: https://inbox.dpdk.org/dev/20240302090207.428d4853@hermes.local/

The code generated from rte_memcpy() was not incorrect, only slightly inefficient (for the affected cases), so the patch is not a bugfix in need of backporting.

> On my box I didn’t see any with 'memcpy_perf_autotest'.
> Konstantin
>