On 2025-06-01 15:46, Adam D. Barratt wrote:
> On Thu, 2025-05-29 at 15:41 +0200, Aurelien Jarno wrote:
> > - Add an SVE implementation of memset on aarch64.
> >
> > => This improves the performances of memset by up to 20% on arm64
> > hardware with SVE support. This change is in trixie/sid for more than
> > 2 months.
> >
> > - Improve generic implementation of memset on aarch64.
> >
> > => This improves the performances of memset by up to 24% on arm64
> > hardware by avoiding branches and using overlapping store. This
> > change is in trixie/sid for more than 2 months.
>
> Unfortunately one or other of the above changes (I'm guessing the
> first) is causing a build failure on arm64:
>
> ../sysdeps/aarch64/multiarch/memset.c: In function ‘select_memset_ifunc’:
> ../sysdeps/aarch64/multiarch/memset.c:52:11: error: ‘prefer_sve_ifuncs’
> undeclared (first use in this function)
> 52 | if (prefer_sve_ifuncs && zva_size == 64)
> | ^~~~~~~~~~~~~~~~~
> ../sysdeps/aarch64/multiarch/memset.c:52:11: note: each undeclared identifier
> is reported only once for each function it appears in
Oops I missed that we revert the prefer_sve_ifuncs definition due to
#1083095. This means we also need to revert this:
commit 4f3b4e7b52b14dcc4633f319bd4087af76674ae6
Author: Wilco Dijkstra <wilco.dijks...@arm.com>
Date: Thu Feb 27 16:28:52 2025 +0000
AArch64: Use prefer_sve_ifuncs for SVE memset
Use prefer_sve_ifuncs for SVE memset just like memcpy.
Reviewed-by: Yury Khrustalev <yury.khrusta...@arm.com>
(cherry picked from commit 0f044be1dae5169d0e57f8d487b427863aeadab4)
diff --git a/sysdeps/aarch64/multiarch/memset.c
b/sysdeps/aarch64/multiarch/memset.c
index d5b7665949..6c9bb910c6 100644
--- a/sysdeps/aarch64/multiarch/memset.c
+++ b/sysdeps/aarch64/multiarch/memset.c
@@ -49,7 +49,7 @@ select_memset_ifunc (void)
if (IS_A64FX (midr) && zva_size == 256)
return __memset_a64fx;
- if (zva_size == 64)
+ if (prefer_sve_ifuncs && zva_size == 64)
return __memset_sve_zva64;
}
I'll work on an update, this time after more testing.
Regards
Aurelien
--
Aurelien Jarno GPG: 4096R/1DDD8C9B
aurel...@aurel32.net http://aurel32.net
