x86: Use AVX2 memcpy/memset on Skylake server [BZ #21396]
commit4cb334c4d6249686653137ec273d081371b3672d
authorH.J. Lu <hjl.tools@gmail.com>
Tue, 18 Apr 2017 21:01:45 +0000 (18 14:01 -0700)
committerH.J. Lu <hjl.tools@gmail.com>
Tue, 18 Apr 2017 21:01:45 +0000 (18 14:01 -0700)
tree0ddce19851595aeccad1caf222c91014f7817de2
parent1c53cb49de6d82d9469ccbd5aa0c55924502bd8b
x86: Use AVX2 memcpy/memset on Skylake server [BZ #21396]

On Skylake server, AVX512 load/store instructions in memcpy/memset may
lead to lower CPU turbo frequency in certain situations.  Use of AVX2
in memcpy/memset has been observed to have improved overall performance
in many workloads due to the higher frequency.

Since AVX512ER is unique to Xeon Phi, this patch sets Prefer_No_AVX512
if AVX512ER isn't available so that AVX2 versions of memcpy/memset are
used on Skylake server.

[BZ #21396]
* sysdeps/x86/cpu-features.c (init_cpu_features): Set
Prefer_No_AVX512 if AVX512ER isn't available.
* sysdeps/x86/cpu-features.h (bit_arch_Prefer_No_AVX512): New.
(index_arch_Prefer_No_AVX512): Likewise.
* sysdeps/x86_64/multiarch/memcpy.S (__new_memcpy): Don't use
AVX512 version if Prefer_No_AVX512 is set.
* sysdeps/x86_64/multiarch/memcpy_chk.S (__memcpy_chk):
Likewise.
* sysdeps/x86_64/multiarch/memmove.S (__libc_memmove): Likewise.
* sysdeps/x86_64/multiarch/memmove_chk.S (__memmove_chk):
Likewise.
* sysdeps/x86_64/multiarch/mempcpy.S (__mempcpy): Likewise.
* sysdeps/x86_64/multiarch/mempcpy_chk.S (__mempcpy_chk):
Likewise.
* sysdeps/x86_64/multiarch/memset.S (memset): Likewise.
* sysdeps/x86_64/multiarch/memset_chk.S (__memset_chk):
Likewise.
ChangeLog
sysdeps/x86/cpu-features.c
sysdeps/x86/cpu-features.h
sysdeps/x86_64/multiarch/memcpy.S
sysdeps/x86_64/multiarch/memcpy_chk.S
sysdeps/x86_64/multiarch/memmove.S
sysdeps/x86_64/multiarch/memmove_chk.S
sysdeps/x86_64/multiarch/mempcpy.S
sysdeps/x86_64/multiarch/mempcpy_chk.S
sysdeps/x86_64/multiarch/memset.S
sysdeps/x86_64/multiarch/memset_chk.S