This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [Patch] aarch64: Thunderx specific memcpy and memmove
- From: Szabolcs Nagy <szabolcs dot nagy at arm dot com>
- To: Steve Ellcey <sellcey at caviumnetworks dot com>, libc-alpha <libc-alpha at sourceware dot org>
- Cc: <nd at arm dot com>, Siddhesh Poyarekar <siddhesh at sourceware dot org>, Adhemerval Zanella <adhemerval dot zanella at linaro dot org>
- Date: Mon, 27 Mar 2017 11:45:12 +0100
- Subject: Re: [Patch] aarch64: Thunderx specific memcpy and memmove
- Authentication-results: sourceware.org; auth=none
- Authentication-results: linaro.org; dkim=none (message not signed) header.d=none;linaro.org; dmarc=none action=none header.from=arm.com;
- Nodisclaimer: True
- References: <1490397926.19074.73.camel@caviumnetworks.com>
- Spamdiagnosticmetadata: NSPM
- Spamdiagnosticoutput: 1:99
On 24/03/17 23:25, Steve Ellcey wrote:
> Now that the IFUNC infrastructure for aarch64 is in place, here is a
> patch to use it to create ThunderX specific versions of memcpy and
> memmove.
>
> This was part of my original patch before it was split in two and a
> couple of issues were raised at that time.
>
> Siddhesh Poyarekar wanted to separate the generic and thunderx copies
> of memcpy/memmove instead of using ifdefs in a combined source file.
> I prefer the ifdef version as a cleaner implementation with less code
> duplication but I can change it if that is the consensus.
>
both are fine with me.
> Also Adhemerval Zanella did some benchmarking that showed the
> prefetching done in the thunderx version might be appropriate for the
> generic version. However if you look at the prefetching we only do it
> every other time through the loop. This is because the loop copies 64
> bytes and the ThunderX cache line size is 128 bytes. If other aarch64
> chips have a 64 byte cache line they might want a different prefetching
> setup.
>
> If people think we should use the ThunderX version of memcpy for all
> aarch64 systems I am happy to drop this patch and create one that just
> changes memcpy.S to do the ThunderX style prefetches for all aarch64
> systems.
>
adding prefetches to the generic code is preferable
if it can make both thunderx and generic users happy.
we need to find what's the best way to add the prefetches,
the new memcpy benchmarks may help here.
> Steve Ellcey
> sellcey@cavium.com
>
>
> 2017-03-24 Steve Ellcey <sellcey@caviumnetworks.com>
>
> * sysdeps/aarch64/memcpy.S (MEMMOVE, MEMCPY): New macros.
> (memmove): Use MEMMOVE for name.
> (memcpy): Use MEMCPY for name. Add loop with prefetching
> under USE_THUNDERX macro.
> * sysdeps/aarch64/multiarch/Makefile: New file.
> * sysdeps/aarch64/multiarch/ifunc-impl-list.c: Likewise.
> * sysdeps/aarch64/multiarch/init-arch.h: Likewise.
> * sysdeps/aarch64/multiarch/memcpy.c: Likewise.
> * sysdeps/aarch64/multiarch/memcpy_generic.S: Likewise.
> * sysdeps/aarch64/multiarch/memcpy_thunderx.S: Likewise.
> * sysdeps/aarch64/multiarch/memmove.c: Likewise.
>