glibc/sysdeps/aarch64/memset.S
Wilco Dijkstra a8c5a2a952 This is an optimized memset for AArch64. Memset is split into 4 main cases:
small sets of up to 16 bytes, medium of 16..96 bytes which are fully unrolled.
Large memsets of more than 96 bytes align the destination and use an unrolled
loop processing 64 bytes per iteration.  Memsets of zero of more than 256 use
the dc zva instruction, and there are faster versions for the common ZVA sizes
64 or 128.  STP of Q registers is used to reduce codesize without loss of
performance.

The speedup on test-memset is 1% on Cortex-A57 and 8% on Cortex-A53.

	* sysdeps/aarch64/memset.S (__memset):
	Rewrite of optimized memset.
2016-05-12 16:44:53 +01:00

193 lines
4.2 KiB
ArmAsm

/* Copyright (C) 2012-2016 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
<http://www.gnu.org/licenses/>. */
#include <sysdep.h>
/* Assumptions:
*
* ARMv8-a, AArch64, unaligned accesses
*
*/
#define dstin x0
#define val x1
#define valw w1
#define count x2
#define dst x3
#define dstend x4
#define tmp1 x5
#define tmp1w w5
#define tmp2 x6
#define tmp2w w6
#define zva_len x7
#define zva_lenw w7
ENTRY_ALIGN (__memset, 6)
dup v0.16B, valw
add dstend, dstin, count
cmp count, 96
b.hi L(set_long)
cmp count, 16
b.hs L(set_medium)
mov val, v0.D[0]
/* Set 0..15 bytes. */
tbz count, 3, 1f
str val, [dstin]
str val, [dstend, -8]
ret
nop
1: tbz count, 2, 2f
str valw, [dstin]
str valw, [dstend, -4]
ret
2: cbz count, 3f
strb valw, [dstin]
tbz count, 1, 3f
strh valw, [dstend, -2]
3: ret
/* Set 17..96 bytes. */
L(set_medium):
str q0, [dstin]
tbnz count, 6, L(set96)
str q0, [dstend, -16]
tbz count, 5, 1f
str q0, [dstin, 16]
str q0, [dstend, -32]
1: ret
.p2align 4
/* Set 64..96 bytes. Write 64 bytes from the start and
32 bytes from the end. */
L(set96):
str q0, [dstin, 16]
stp q0, q0, [dstin, 32]
stp q0, q0, [dstend, -32]
ret
.p2align 3
nop
L(set_long):
and valw, valw, 255
bic dst, dstin, 15
str q0, [dstin]
cmp count, 256
ccmp valw, 0, 0, cs
b.eq L(try_zva)
L(no_zva):
sub count, dstend, dst /* Count is 16 too large. */
add dst, dst, 16
sub count, count, 64 + 16 /* Adjust count and bias for loop. */
1: stp q0, q0, [dst], 64
stp q0, q0, [dst, -32]
L(tail64):
subs count, count, 64
b.hi 1b
2: stp q0, q0, [dstend, -64]
stp q0, q0, [dstend, -32]
ret
.p2align 3
L(try_zva):
mrs tmp1, dczid_el0
tbnz tmp1w, 4, L(no_zva)
and tmp1w, tmp1w, 15
cmp tmp1w, 4 /* ZVA size is 64 bytes. */
b.ne L(zva_128)
/* Write the first and last 64 byte aligned block using stp rather
than using DC ZVA. This is faster on some cores.
*/
L(zva_64):
str q0, [dst, 16]
stp q0, q0, [dst, 32]
bic dst, dst, 63
stp q0, q0, [dst, 64]
stp q0, q0, [dst, 96]
sub count, dstend, dst /* Count is now 128 too large. */
sub count, count, 128+64+64 /* Adjust count and bias for loop. */
add dst, dst, 128
nop
1: dc zva, dst
add dst, dst, 64
subs count, count, 64
b.hi 1b
stp q0, q0, [dst, 0]
stp q0, q0, [dst, 32]
stp q0, q0, [dstend, -64]
stp q0, q0, [dstend, -32]
ret
.p2align 3
L(zva_128):
cmp tmp1w, 5 /* ZVA size is 128 bytes. */
b.ne L(zva_other)
str q0, [dst, 16]
stp q0, q0, [dst, 32]
stp q0, q0, [dst, 64]
stp q0, q0, [dst, 96]
bic dst, dst, 127
sub count, dstend, dst /* Count is now 128 too large. */
sub count, count, 128+128 /* Adjust count and bias for loop. */
add dst, dst, 128
1: dc zva, dst
add dst, dst, 128
subs count, count, 128
b.hi 1b
stp q0, q0, [dstend, -128]
stp q0, q0, [dstend, -96]
stp q0, q0, [dstend, -64]
stp q0, q0, [dstend, -32]
ret
L(zva_other):
mov tmp2w, 4
lsl zva_lenw, tmp2w, tmp1w
add tmp1, zva_len, 64 /* Max alignment bytes written. */
cmp count, tmp1
blo L(no_zva)
sub tmp2, zva_len, 1
add tmp1, dst, zva_len
add dst, dst, 16
subs count, tmp1, dst /* Actual alignment bytes to write. */
bic tmp1, tmp1, tmp2 /* Aligned dc zva start address. */
beq 2f
1: stp q0, q0, [dst], 64
stp q0, q0, [dst, -32]
subs count, count, 64
b.hi 1b
2: mov dst, tmp1
sub count, dstend, tmp1 /* Remaining bytes to write. */
subs count, count, zva_len
b.lo 4f
3: dc zva, dst
add dst, dst, zva_len
subs count, count, zva_len
b.hs 3b
4: add count, count, zva_len
b L(tail64)
END (__memset)
weak_alias (__memset, memset)
libc_hidden_builtin_def (memset)