aarch64: Optimized memset for Kunpeng processor.

Due to the branch prediction issue of Kunpeng processor, we found
memset_generic has poor performance on middle sizes setting, and so
we reconstructed the logic, expanded the loop by 4 times in set_long
to solve the problem, even when setting below 1K sizes have benefit.

Another change is that DZ_ZVA seems no work when setting zero, so we
discarded it and used set_long to set zero instead. Fewer branches and
predictions also make the zero case have slightly improvement.

Checked on aarch64-linux-gnu.

Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
This commit is contained in:
Xuelei Zhang 2019-12-19 15:26:33 +00:00 committed by Adhemerval Zanella
parent c2150769d0
commit 525de033a9
4 changed files with 119 additions and 2 deletions

View File

@ -1,7 +1,7 @@
ifeq ($(subdir),string)
sysdep_routines += memcpy_generic memcpy_thunderx memcpy_thunderx2 \
memcpy_falkor memmove_falkor \
memset_generic memset_falkor memset_emag \
memset_generic memset_falkor memset_emag memset_kunpeng \
memchr_generic memchr_nosimd \
strlen_generic strlen_asimd
endif

View File

@ -53,6 +53,7 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
can do a comparative analysis with __memset_generic. */
IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_falkor)
IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_emag)
IFUNC_IMPL_ADD (array, i, memset, 1, __memset_kunpeng)
IFUNC_IMPL_ADD (array, i, memset, 1, __memset_generic))
IFUNC_IMPL (i, name, memchr,
IFUNC_IMPL_ADD (array, i, memchr, 1, __memchr_nosimd)

View File

@ -30,10 +30,13 @@ extern __typeof (__redirect_memset) __libc_memset;
extern __typeof (__redirect_memset) __memset_falkor attribute_hidden;
extern __typeof (__redirect_memset) __memset_emag attribute_hidden;
extern __typeof (__redirect_memset) __memset_kunpeng attribute_hidden;
extern __typeof (__redirect_memset) __memset_generic attribute_hidden;
libc_ifunc (__libc_memset,
((IS_FALKOR (midr) || IS_PHECDA (midr)) && zva_size == 64
IS_KUNPENG (midr)
?__memset_kunpeng
: ((IS_FALKOR (midr) || IS_PHECDA (midr)) && zva_size == 64
? __memset_falkor
: (IS_EMAG (midr) && zva_size == 64
? __memset_emag

View File

@ -0,0 +1,113 @@
/* Optimized memset for Huawei Kunpeng processor.
Copyright (C) 2012-2019 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library. If not, see
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#include <sysdeps/aarch64/memset-reg.h>
#if IS_IN (libc)
# define MEMSET __memset_kunpeng
/* Assumptions:
*
* ARMv8-a, AArch64, unaligned accesses
*
*/
ENTRY_ALIGN (MEMSET, 6)
DELOUSE (0)
DELOUSE (2)
dup v0.16B, valw
add dstend, dstin, count
cmp count, 128
b.hs L(set_long)
cmp count, 16
b.lo L(less16)
/* Set 16..127 bytes. */
str q0, [dstin]
tbnz count, 6, L(set127)
str q0, [dstend, -16]
tbz count, 5, 1f
str q0, [dstin, 16]
str q0, [dstend, -32]
1: ret
.p2align 4
/* Set 64..127 bytes. Write 64 bytes from the start and
64 bytes from the end. */
L(set127):
stp q0, q0, [dstin, 16]
str q0, [dstin, 48]
stp q0, q0, [dstend, -64]
stp q0, q0, [dstend, -32]
ret
.p2align 4
/* Set 0..15 bytes. */
L(less16):
tbz count, 3, L(less8)
str d0, [dstin]
str d0, [dstend, -8]
ret
L(less8):
tbz count, 2, 2f
str s0, [dstin]
str s0, [dstend, -4]
ret
2: cbz count, 3f
str b0, [dstin]
tbz count, 1, 3f
str h0, [dstend, -2]
3: ret
.p2align 4
L(set_long):
bic dst, dstin, 15
str q0, [dstin]
sub count, dstend, dst /* Count is 16 too large. */
sub dst, dst, 16 /* Dst is biased by -32. */
sub count, count, 64 + 16 + 1 /* Adjust count and bias for loop. */
1: stp q0, q0, [dst, 32]
stp q0, q0, [dst, 64]!
subs count, count, 64
b.lo 1f
stp q0, q0, [dst, 32]
stp q0, q0, [dst, 64]!
subs count, count, 64
b.lo 1f
stp q0, q0, [dst, 32]
stp q0, q0, [dst, 64]!
subs count, count, 64
b.lo 1f
stp q0, q0, [dst, 32]
stp q0, q0, [dst, 64]!
subs count, count, 64
b.hs 1b
1: stp q0, q0, [dstend, -64]
stp q0, q0, [dstend, -32]
ret
END (MEMSET)
libc_hidden_builtin_def (MEMSET)
#endif