mirror/glibc
2
0
Fork 0
mirror of git://sourceware.org/git/glibc.git synced 2024-11-21 01:12:26 +08:00
Code Issues Packages Projects Releases Wiki Activity
72739c79f6
glibc/sysdeps/aarch64/multiarch/memcpy_falkor.S
Krzysztof Koch d1f75e9644 AArch64: Merge Falkor memcpy and memmove implementations 
Falkor's memcpy and memmove share some implementation details,
therefore, the two routines are moved to a single source file
for code reuse.

The two routines now share code for small and medium copies
(up to and including 128 bytes). Large copies in memcpy do not
handle overlap correctly, consequently, the loops for
moving/copying more than 128 bytes stay separate for memcpy
and memmove.

To increase code reuse a number of small modifications were made:

1. The old implementation of memcpy copied the first 16-bytes as
   soon as the size of data was determined to be greater than 32 bytes.
   For memcpy code to also work when copying small/medium overlapping
   data, the first load and store was moved to the large copy case.
2. Medium memcpy case no longer assumes that 16 bytes were already
   copied and uses 8 registers to copy up to 128 bytes.
3. Small case for memmove was enlarged to that of memcpy, which is
   less than or equal to 32 bytes.
4. Medium case for memmove was enlarged to that of memcpy, which is
   less than or equal to 128 bytes.

Other changes include:

1. Improve alignment of existing loop bodies.
2. 'Delouse' memmove and memcpy input arguments. Make sure that
   upper 32-bits of input registers are zeroed if unused.
3. Do one more iteration in memmove loops and reduce the number of
   copies made from the start/end of the buffer, depending on
   the direction of the memmove loop.

Benchmarking:

Looking at the results from bench-memcpy-random.out, we can see that
now memmove_falkor is about 5% faster than memcpy_falkor_old, while
memmove_falkor_old was more than 15% slower. The memcpy implementation
remained largely unmodified, so there is no significant performance
change.

The reason for such a significant memmove performance gain is the
increase of the upper bound on the small copy case to 32 bytes and
the increase of the upper bound on the medium copy case to 128 bytes.

Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>
2020-06-08 14:13:05 +01:00

316 lines
7.4 KiB
ArmAsm
Raw Blame History

/* Optimized memcpy for Qualcomm Falkor processor.
Copyright (C) 2017-2020 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>
/* Assumptions:
ARMv8-a, AArch64, falkor, unaligned accesses. */
#define dstin x0
#define src x1
#define count x2
#define dst x3
#define srcend x4
#define dstend x5
#define tmp1 x14
#define A_x x6
#define B_x x7
#define A_w w6
#define B_w w7
#define A_q q0
#define B_q q1
#define C_q q2
#define D_q q3
#define E_q q4
#define F_q q5
#define G_q q6
#define H_q q7
#define Q_q q6
#define S_q q22
/* Copies are split into 3 main cases:
1. Small copies of up to 32 bytes
2. Medium copies of 33..128 bytes which are fully unrolled
3. Large copies of more than 128 bytes.
Large copies align the source to a quad word and use an unrolled loop
processing 64 bytes per iteration.
FALKOR-SPECIFIC DESIGN:
The smallest copies (32 bytes or less) focus on optimal pipeline usage,
which is why the redundant copies of 0-3 bytes have been replaced with
conditionals, since the former would unnecessarily break across multiple
issue groups. The medium copy group has been enlarged to 128 bytes since
bumping up the small copies up to 32 bytes allows us to do that without
cost and also allows us to reduce the size of the prep code before loop64.
The copy loop uses only one register q0. This is to ensure that all loads
hit a single hardware prefetcher which can get correctly trained to prefetch
a single stream.
The non-temporal stores help optimize cache utilization. */
#if IS_IN (libc)
ENTRY_ALIGN (__memcpy_falkor, 6)
DELOUSE (0)
DELOUSE (1)
DELOUSE (2)
cmp count, 32
add srcend, src, count
add dstend, dstin, count
b.ls L(copy32)
cmp count, 128
b.hi L(copy_long)
/* Medium copies: 33..128 bytes. */
L(copy128):
sub tmp1, count, 1
ldr A_q, [src]
ldr B_q, [src, 16]
ldr C_q, [srcend, -32]
ldr D_q, [srcend, -16]
tbz tmp1, 6, 1f
ldr E_q, [src, 32]
ldr F_q, [src, 48]
ldr G_q, [srcend, -64]
ldr H_q, [srcend, -48]
str G_q, [dstend, -64]
str H_q, [dstend, -48]
str E_q, [dstin, 32]
str F_q, [dstin, 48]
1:
str A_q, [dstin]
str B_q, [dstin, 16]
str C_q, [dstend, -32]
str D_q, [dstend, -16]
ret
.p2align 4
/* Small copies: 0..32 bytes. */
L(copy32):
/* 16-32 */
cmp count, 16
b.lo 1f
ldr A_q, [src]
ldr B_q, [srcend, -16]
str A_q, [dstin]
str B_q, [dstend, -16]
ret
.p2align 4
1:
/* 8-15 */
tbz count, 3, 1f
ldr A_x, [src]
ldr B_x, [srcend, -8]
str A_x, [dstin]
str B_x, [dstend, -8]
ret
.p2align 4
1:
/* 4-7 */
tbz count, 2, 1f
ldr A_w, [src]
ldr B_w, [srcend, -4]
str A_w, [dstin]
str B_w, [dstend, -4]
ret
.p2align 4
1:
/* 2-3 */
tbz count, 1, 1f
ldrh A_w, [src]
ldrh B_w, [srcend, -2]
strh A_w, [dstin]
strh B_w, [dstend, -2]
ret
.p2align 4
1:
/* 0-1 */
tbz count, 0, 1f
ldrb A_w, [src]
strb A_w, [dstin]
1:
ret
/* Align SRC to 16 bytes and copy; that way at least one of the
accesses is aligned throughout the copy sequence.
The count is off by 0 to 15 bytes, but this is OK because we trim
off the last 64 bytes to copy off from the end. Due to this the
loop never runs out of bounds. */
.p2align 4
nop /* Align loop64 below. */
L(copy_long):
ldr A_q, [src]
sub count, count, 64 + 16
and tmp1, src, 15
str A_q, [dstin]
bic src, src, 15
sub dst, dstin, tmp1
add count, count, tmp1
L(loop64):
ldr A_q, [src, 16]!
str A_q, [dst, 16]
ldr A_q, [src, 16]!
subs count, count, 64
str A_q, [dst, 32]
ldr A_q, [src, 16]!
str A_q, [dst, 48]
ldr A_q, [src, 16]!
str A_q, [dst, 64]!
b.hi L(loop64)
/* Write the last full set of 64 bytes. The remainder is at most 64
bytes, so it is safe to always copy 64 bytes from the end even if
there is just 1 byte left. */
ldr E_q, [srcend, -64]
str E_q, [dstend, -64]
ldr D_q, [srcend, -48]
str D_q, [dstend, -48]
ldr C_q, [srcend, -32]
str C_q, [dstend, -32]
ldr B_q, [srcend, -16]
str B_q, [dstend, -16]
ret
END (__memcpy_falkor)
libc_hidden_builtin_def (__memcpy_falkor)
/* RATIONALE:
The move has 4 distinct parts:
* Small moves of 32 bytes and under.
* Medium sized moves of 33-128 bytes (fully unrolled).
* Large moves where the source address is higher than the destination
(forward copies)
* Large moves where the destination address is higher than the source
(copy backward, or move).
We use only two registers q6 and q22 for the moves and move 32 bytes at a
time to correctly train the hardware prefetcher for better throughput.
For small and medium cases memcpy is used. */
ENTRY_ALIGN (__memmove_falkor, 6)
DELOUSE (0)
DELOUSE (1)
DELOUSE (2)
cmp count, 32
add srcend, src, count
add dstend, dstin, count
b.ls L(copy32)
cmp count, 128
b.ls L(copy128)
sub tmp1, dstin, src
ccmp tmp1, count, 2, hi
b.lo L(move_long)
/* CASE: Copy Forwards
Align src to 16 byte alignment so that we don't cross cache line
boundaries on both loads and stores. There are at least 128 bytes
to copy, so copy 16 bytes unaligned and then align. The loop
copies 32 bytes per iteration and prefetches one iteration ahead. */
ldr S_q, [src]
and tmp1, src, 15
bic src, src, 15
sub dst, dstin, tmp1
add count, count, tmp1 /* Count is now 16 too large. */
ldr Q_q, [src, 16]!
str S_q, [dstin]
ldr S_q, [src, 16]!
sub count, count, 32 + 32 + 16 /* Test and readjust count. */
.p2align 4
1:
subs count, count, 32
str Q_q, [dst, 16]
ldr Q_q, [src, 16]!
str S_q, [dst, 32]!
ldr S_q, [src, 16]!
b.hi 1b
/* Copy 32 bytes from the end before writing the data prefetched in the
last loop iteration. */
2:
ldr B_q, [srcend, -32]
ldr C_q, [srcend, -16]
str Q_q, [dst, 16]
str S_q, [dst, 32]
str B_q, [dstend, -32]
str C_q, [dstend, -16]
ret
/* CASE: Copy Backwards
Align srcend to 16 byte alignment so that we don't cross cache line
boundaries on both loads and stores. There are at least 128 bytes
to copy, so copy 16 bytes unaligned and then align. The loop
copies 32 bytes per iteration and prefetches one iteration ahead. */
.p2align 4
nop
nop
L(move_long):
cbz tmp1, 3f /* Return early if src == dstin */
ldr S_q, [srcend, -16]
and tmp1, srcend, 15
sub srcend, srcend, tmp1
ldr Q_q, [srcend, -16]!
str S_q, [dstend, -16]
sub count, count, tmp1
ldr S_q, [srcend, -16]!
sub dstend, dstend, tmp1
sub count, count, 32 + 32
1:
subs count, count, 32
str Q_q, [dstend, -16]
ldr Q_q, [srcend, -16]!
str S_q, [dstend, -32]!
ldr S_q, [srcend, -16]!
b.hi 1b
/* Copy 32 bytes from the start before writing the data prefetched in the
last loop iteration. */
ldr B_q, [src, 16]
ldr C_q, [src]
str Q_q, [dstend, -16]
str S_q, [dstend, -32]
str B_q, [dstin, 16]
str C_q, [dstin]
3: ret
END (__memmove_falkor)
libc_hidden_builtin_def (__memmove_falkor)
#endif
Reference in New Issue View Git Blame Copy Permalink