1
0
mirror of https://github.com/golang/go synced 2024-11-19 13:54:56 -07:00
go/src/runtime/memmove_arm64.s
Balaram Makam 213a75171d runtime: improve arm64 memmove implementation
Improve runtime memmove_arm64.s specializing for small copies and
processing 32 bytes per iteration for 32 bytes or more.

Benchmark results of runtime/Memmove on Amberwing:
name                      old time/op    new time/op     delta
Memmove/0                   7.61ns ± 0%     7.20ns ± 0%     ~     (p=0.053 n=5+7)
Memmove/1                   9.28ns ± 0%     8.80ns ± 0%   -5.17%  (p=0.000 n=4+8)
Memmove/2                   9.65ns ± 0%     9.20ns ± 0%   -4.68%  (p=0.000 n=5+8)
Memmove/3                   10.0ns ± 0%      9.2ns ± 0%   -7.83%  (p=0.000 n=5+8)
Memmove/4                   10.6ns ± 0%      9.2ns ± 0%  -13.21%  (p=0.000 n=5+8)
Memmove/5                   11.0ns ± 0%      9.2ns ± 0%  -16.36%  (p=0.000 n=5+8)
Memmove/6                   12.4ns ± 0%      9.2ns ± 0%  -25.81%  (p=0.000 n=5+8)
Memmove/7                   13.1ns ± 0%      9.2ns ± 0%  -29.56%  (p=0.000 n=5+8)
Memmove/8                   9.10ns ± 1%     9.20ns ± 0%   +1.08%  (p=0.002 n=5+8)
Memmove/9                   9.67ns ± 0%     9.20ns ± 0%   -4.88%  (p=0.000 n=5+8)
Memmove/10                  10.4ns ± 0%      9.2ns ± 0%  -11.54%  (p=0.000 n=5+8)
Memmove/11                  10.9ns ± 0%      9.2ns ± 0%  -15.60%  (p=0.000 n=5+8)
Memmove/12                  11.5ns ± 0%      9.2ns ± 0%  -20.00%  (p=0.000 n=5+8)
Memmove/13                  12.4ns ± 0%      9.2ns ± 0%  -25.81%  (p=0.000 n=5+8)
Memmove/14                  13.1ns ± 0%      9.2ns ± 0%  -29.77%  (p=0.000 n=5+8)
Memmove/15                  13.8ns ± 0%      9.2ns ± 0%  -33.33%  (p=0.000 n=5+8)
Memmove/16                  9.70ns ± 0%     9.20ns ± 0%   -5.19%  (p=0.000 n=5+8)
Memmove/32                  10.6ns ± 0%      9.2ns ± 0%  -13.21%  (p=0.000 n=4+8)
Memmove/64                  13.4ns ± 0%     10.2ns ± 0%  -23.88%  (p=0.000 n=4+8)
Memmove/128                 18.1ns ± 1%     13.2ns ± 0%  -26.99%  (p=0.000 n=5+8)
Memmove/256                 25.2ns ± 0%     16.4ns ± 0%  -34.92%  (p=0.000 n=5+8)
Memmove/512                 36.4ns ± 0%     22.8ns ± 0%  -37.36%  (p=0.000 n=5+8)
Memmove/1024                70.1ns ± 0%     36.8ns ±11%  -47.49%  (p=0.002 n=5+8)
Memmove/2048                 121ns ± 0%       61ns ± 0%     ~     (p=0.053 n=5+7)
Memmove/4096                 224ns ± 0%      120ns ± 0%  -46.43%  (p=0.000 n=5+8)
MemmoveUnalignedDst/0       8.40ns ± 0%     8.00ns ± 0%   -4.76%  (p=0.000 n=5+8)
MemmoveUnalignedDst/1       9.87ns ± 1%    10.00ns ± 0%     ~     (p=0.070 n=5+8)
MemmoveUnalignedDst/2       10.6ns ± 0%     10.4ns ± 0%   -1.89%  (p=0.000 n=5+8)
MemmoveUnalignedDst/3       10.8ns ± 0%     10.4ns ± 0%   -3.70%  (p=0.000 n=5+8)
MemmoveUnalignedDst/4       10.9ns ± 0%     10.3ns ± 0%     ~     (p=0.053 n=5+7)
MemmoveUnalignedDst/5       11.5ns ± 0%     10.3ns ± 1%  -10.22%  (p=0.000 n=4+8)
MemmoveUnalignedDst/6       13.2ns ± 0%     10.4ns ± 1%  -21.50%  (p=0.000 n=5+8)
MemmoveUnalignedDst/7       13.7ns ± 0%     10.3ns ± 1%  -24.64%  (p=0.000 n=4+8)
MemmoveUnalignedDst/8       10.1ns ± 0%     10.4ns ± 0%   +2.97%  (p=0.002 n=5+8)
MemmoveUnalignedDst/9       10.7ns ± 0%     10.4ns ± 0%   -2.80%  (p=0.000 n=5+8)
MemmoveUnalignedDst/10      11.2ns ± 1%     10.4ns ± 0%   -6.81%  (p=0.000 n=5+8)
MemmoveUnalignedDst/11      11.6ns ± 0%     10.4ns ± 0%  -10.34%  (p=0.000 n=5+8)
MemmoveUnalignedDst/12      12.5ns ± 2%     10.4ns ± 0%  -16.53%  (p=0.000 n=5+8)
MemmoveUnalignedDst/13      13.7ns ± 0%     10.4ns ± 0%  -24.09%  (p=0.000 n=5+8)
MemmoveUnalignedDst/14      14.0ns ± 0%     10.4ns ± 0%  -25.71%  (p=0.000 n=5+8)
MemmoveUnalignedDst/15      14.6ns ± 0%     10.4ns ± 0%  -28.77%  (p=0.000 n=5+8)
MemmoveUnalignedDst/16      10.5ns ± 0%     10.4ns ± 0%   -0.95%  (p=0.000 n=5+8)
MemmoveUnalignedDst/32      12.4ns ± 0%     11.6ns ± 0%   -6.05%  (p=0.000 n=5+8)
MemmoveUnalignedDst/64      15.2ns ± 0%     12.3ns ± 0%  -19.08%  (p=0.000 n=5+8)
MemmoveUnalignedDst/128     18.7ns ± 0%     15.2ns ± 0%  -18.72%  (p=0.000 n=5+8)
MemmoveUnalignedDst/256     25.1ns ± 0%     18.6ns ± 0%  -25.90%  (p=0.000 n=5+8)
MemmoveUnalignedDst/512     37.8ns ± 0%     24.4ns ± 0%  -35.45%  (p=0.000 n=5+8)
MemmoveUnalignedDst/1024    74.6ns ± 0%     40.4ns ± 0%     ~     (p=0.053 n=5+7)
MemmoveUnalignedDst/2048     133ns ± 0%       75ns ± 0%  -43.91%  (p=0.000 n=5+8)
MemmoveUnalignedDst/4096     247ns ± 0%      141ns ± 0%  -42.91%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/0       8.40ns ± 0%     8.00ns ± 0%   -4.76%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/1       9.81ns ± 0%    10.00ns ± 0%   +1.98%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/2       10.5ns ± 0%     10.0ns ± 0%   -4.76%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/3       10.7ns ± 1%     10.0ns ± 0%   -6.89%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/4       11.3ns ± 0%     10.0ns ± 0%  -11.50%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/5       11.6ns ± 0%     10.0ns ± 0%  -13.79%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/6       13.6ns ± 0%     10.0ns ± 0%  -26.47%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/7       14.4ns ± 0%     10.0ns ± 0%  -30.75%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/8       9.87ns ± 1%    10.00ns ± 0%     ~     (p=0.070 n=5+8)
MemmoveUnalignedSrc/9       10.4ns ± 0%     10.0ns ± 0%   -3.85%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/10      11.2ns ± 0%     10.0ns ± 0%  -10.71%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/11      11.8ns ± 0%     10.0ns ± 0%  -15.25%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/12      12.1ns ± 0%     10.0ns ± 0%  -17.36%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/13      13.6ns ± 0%     10.0ns ± 0%  -26.47%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/14      14.7ns ± 0%     10.0ns ± 0%  -31.79%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/15      14.4ns ± 0%     10.0ns ± 0%  -30.56%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/16      11.0ns ± 0%     10.0ns ± 0%   -9.09%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/32      11.5ns ± 0%     10.0ns ± 0%  -13.04%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/64      14.9ns ± 0%     11.2ns ± 0%  -24.83%  (p=0.000 n=4+8)
MemmoveUnalignedSrc/128     19.5ns ± 0%     15.2ns ± 0%  -22.05%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/256     27.3ns ± 2%     19.2ns ± 0%  -29.62%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/512     40.4ns ± 0%     27.2ns ± 0%  -32.67%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/1024    75.4ns ± 0%     44.4ns ± 0%  -41.15%  (p=0.000 n=5+8)
MemmoveUnalignedSrc/2048     131ns ± 0%       77ns ± 3%  -41.56%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/4096     248ns ± 0%      145ns ± 0%  -41.53%  (p=0.000 n=5+8)

name                      old speed      new speed       delta
Memmove/1                  108MB/s ± 0%    114MB/s ± 0%   +5.37%  (p=0.004 n=4+8)
Memmove/2                  207MB/s ± 0%    217MB/s ± 0%   +4.85%  (p=0.002 n=5+8)
Memmove/3                  301MB/s ± 0%    326MB/s ± 0%   +8.45%  (p=0.002 n=5+8)
Memmove/4                  377MB/s ± 0%    435MB/s ± 0%  +15.31%  (p=0.004 n=4+8)
Memmove/5                  455MB/s ± 0%    543MB/s ± 0%  +19.46%  (p=0.002 n=5+8)
Memmove/6                  483MB/s ± 0%    652MB/s ± 0%  +34.88%  (p=0.003 n=5+7)
Memmove/7                  537MB/s ± 0%    761MB/s ± 0%  +41.71%  (p=0.002 n=5+8)
Memmove/8                  879MB/s ± 1%    869MB/s ± 0%   -1.15%  (p=0.000 n=5+7)
Memmove/9                  931MB/s ± 0%    978MB/s ± 0%   +5.05%  (p=0.002 n=5+8)
Memmove/10                 960MB/s ± 0%   1086MB/s ± 0%  +13.13%  (p=0.002 n=5+8)
Memmove/11                1.00GB/s ± 0%   1.20GB/s ± 0%  +18.92%  (p=0.003 n=5+7)
Memmove/12                1.04GB/s ± 0%   1.30GB/s ± 0%  +25.40%  (p=0.002 n=5+8)
Memmove/13                1.05GB/s ± 0%   1.41GB/s ± 0%  +34.87%  (p=0.002 n=5+8)
Memmove/14                1.07GB/s ± 0%   1.52GB/s ± 0%  +42.14%  (p=0.002 n=5+8)
Memmove/15                1.09GB/s ± 0%   1.63GB/s ± 0%  +49.91%  (p=0.002 n=5+8)
Memmove/16                1.65GB/s ± 0%   1.74GB/s ± 0%   +5.40%  (p=0.003 n=5+7)
Memmove/32                3.01GB/s ± 0%   3.48GB/s ± 0%  +15.58%  (p=0.003 n=5+7)
Memmove/64                4.76GB/s ± 0%   6.27GB/s ± 0%  +31.75%  (p=0.003 n=5+7)
Memmove/128               7.08GB/s ± 1%   9.69GB/s ± 0%  +36.96%  (p=0.002 n=5+8)
Memmove/256               10.2GB/s ± 0%   15.6GB/s ± 0%  +53.58%  (p=0.002 n=5+8)
Memmove/512               14.1GB/s ± 0%   22.4GB/s ± 0%  +59.57%  (p=0.003 n=5+7)
Memmove/1024              14.6GB/s ± 0%   27.9GB/s ±10%  +91.00%  (p=0.002 n=5+8)
Memmove/2048              16.9GB/s ± 0%   33.4GB/s ± 0%  +98.32%  (p=0.003 n=5+7)
Memmove/4096              18.3GB/s ± 0%   33.9GB/s ± 0%  +85.80%  (p=0.002 n=5+8)
MemmoveUnalignedDst/1      101MB/s ± 1%    100MB/s ± 0%     ~     (p=0.586 n=5+8)
MemmoveUnalignedDst/2      189MB/s ± 0%    192MB/s ± 0%   +1.82%  (p=0.002 n=5+8)
MemmoveUnalignedDst/3      278MB/s ± 0%    288MB/s ± 0%   +3.88%  (p=0.003 n=5+7)
MemmoveUnalignedDst/4      368MB/s ± 0%    387MB/s ± 0%   +5.41%  (p=0.003 n=5+7)
MemmoveUnalignedDst/5      434MB/s ± 0%    484MB/s ± 0%  +11.52%  (p=0.002 n=5+8)
MemmoveUnalignedDst/6      454MB/s ± 0%    580MB/s ± 0%  +27.62%  (p=0.002 n=5+8)
MemmoveUnalignedDst/7      509MB/s ± 0%    677MB/s ± 0%  +33.01%  (p=0.002 n=5+8)
MemmoveUnalignedDst/8      792MB/s ± 0%    770MB/s ± 0%   -2.77%  (p=0.002 n=5+8)
MemmoveUnalignedDst/9      841MB/s ± 0%    866MB/s ± 0%   +2.92%  (p=0.002 n=5+8)
MemmoveUnalignedDst/10     896MB/s ± 0%    962MB/s ± 0%   +7.35%  (p=0.003 n=5+7)
MemmoveUnalignedDst/11     947MB/s ± 0%   1058MB/s ± 0%  +11.80%  (p=0.002 n=5+8)
MemmoveUnalignedDst/12     962MB/s ± 2%   1154MB/s ± 0%  +19.97%  (p=0.002 n=5+8)
MemmoveUnalignedDst/13     947MB/s ± 0%   1251MB/s ± 0%  +32.08%  (p=0.002 n=5+8)
MemmoveUnalignedDst/14    1.00GB/s ± 0%   1.35GB/s ± 0%  +34.55%  (p=0.002 n=5+8)
MemmoveUnalignedDst/15    1.03GB/s ± 0%   1.44GB/s ± 0%  +40.50%  (p=0.002 n=5+8)
MemmoveUnalignedDst/16    1.53GB/s ± 0%   1.54GB/s ± 0%   +0.77%  (p=0.002 n=5+8)
MemmoveUnalignedDst/32    2.58GB/s ± 0%   2.75GB/s ± 0%   +6.52%  (p=0.003 n=5+7)
MemmoveUnalignedDst/64    4.21GB/s ± 0%   5.19GB/s ± 0%  +23.40%  (p=0.004 n=5+6)
MemmoveUnalignedDst/128   6.86GB/s ± 0%   8.42GB/s ± 0%  +22.78%  (p=0.003 n=5+7)
MemmoveUnalignedDst/256   10.2GB/s ± 0%   13.8GB/s ± 0%  +35.15%  (p=0.002 n=5+8)
MemmoveUnalignedDst/512   13.5GB/s ± 0%   21.0GB/s ± 0%  +54.90%  (p=0.002 n=5+8)
MemmoveUnalignedDst/1024  13.7GB/s ± 0%   25.3GB/s ± 0%  +84.61%  (p=0.003 n=5+7)
MemmoveUnalignedDst/2048  15.3GB/s ± 0%   27.5GB/s ± 0%  +79.52%  (p=0.002 n=5+8)
MemmoveUnalignedDst/4096  16.5GB/s ± 0%   28.9GB/s ± 0%  +74.74%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/1      102MB/s ± 0%    100MB/s ± 0%   -2.02%  (p=0.000 n=5+7)
MemmoveUnalignedSrc/2      191MB/s ± 0%    200MB/s ± 0%   +4.78%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/3      279MB/s ± 0%    300MB/s ± 0%   +7.45%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/4      354MB/s ± 0%    400MB/s ± 0%  +13.10%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/5      431MB/s ± 0%    500MB/s ± 0%  +16.02%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/6      441MB/s ± 0%    600MB/s ± 0%  +36.03%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/7      485MB/s ± 0%    700MB/s ± 0%  +44.29%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/8      811MB/s ± 1%    800MB/s ± 0%   -1.36%  (p=0.016 n=5+8)
MemmoveUnalignedSrc/9      864MB/s ± 0%    900MB/s ± 0%   +4.07%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/10     893MB/s ± 0%    999MB/s ± 0%  +11.97%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/11     932MB/s ± 0%   1099MB/s ± 0%  +18.01%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/12     988MB/s ± 0%   1199MB/s ± 0%  +21.35%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/13     955MB/s ± 0%   1299MB/s ± 0%  +36.02%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/14     955MB/s ± 0%   1399MB/s ± 0%  +46.52%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/15    1.04GB/s ± 0%   1.50GB/s ± 0%  +44.18%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/16    1.45GB/s ± 0%   1.60GB/s ± 0%  +10.14%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/32    2.78GB/s ± 0%   3.20GB/s ± 0%  +15.16%  (p=0.003 n=5+7)
MemmoveUnalignedSrc/64    4.30GB/s ± 0%   5.72GB/s ± 0%  +32.90%  (p=0.003 n=5+7)
MemmoveUnalignedSrc/128   6.57GB/s ± 0%   8.42GB/s ± 0%  +28.06%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/256   9.39GB/s ± 1%  13.33GB/s ± 0%  +41.96%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/512   12.7GB/s ± 0%   18.8GB/s ± 0%  +48.53%  (p=0.003 n=5+7)
MemmoveUnalignedSrc/1024  13.6GB/s ± 0%   23.0GB/s ± 0%  +69.82%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/2048  15.6GB/s ± 0%   26.8GB/s ± 3%  +71.37%  (p=0.002 n=5+8)
MemmoveUnalignedSrc/4096  16.5GB/s ± 0%   28.2GB/s ± 0%  +71.40%  (p=0.002 n=5+8)

Fixes #22925

Change-Id: I38c1a9ad5c6e3f4f95fc521c4b7e3140b58b4737
Reviewed-on: https://go-review.googlesource.com/83799
Run-TryBot: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
2018-03-01 20:34:11 +00:00

124 lines
2.8 KiB
ArmAsm

// Copyright 2014 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include "textflag.h"
// void runtime·memmove(void*, void*, uintptr)
TEXT runtime·memmove(SB), NOSPLIT|NOFRAME, $0-24
MOVD to+0(FP), R3
MOVD from+8(FP), R4
MOVD n+16(FP), R5
CBNZ R5, check
RET
check:
CMP $16, R5
BLE copy16
AND $~31, R5, R7 // R7 is N&~31
SUB R7, R5, R6 // R6 is N&31
CMP R3, R4
BLT backward
// Copying forward proceeds by copying R7/8 words then copying R6 bytes.
// R3 and R4 are advanced as we copy.
// (There may be implementations of armv8 where copying by bytes until
// at least one of source or dest is word aligned is a worthwhile
// optimization, but the on the one tested so far (xgene) it did not
// make a significance difference.)
CBZ R7, noforwardlarge // Do we need to do any doubleword-by-doubleword copying?
ADD R3, R7, R9 // R9 points just past where we copy by word
forwardlargeloop:
LDP.P 32(R4), (R8, R10)
STP.P (R8, R10), 32(R3)
LDP -16(R4), (R11, R12)
STP (R11, R12), -16(R3)
SUB $32, R7, R7
CBNZ R7, forwardlargeloop
noforwardlarge:
CBNZ R6, forwardtail // Do we need to do any byte-by-byte copying?
RET
forwardtail:
ADD R3, R6, R9 // R9 points just past the destination memory
forwardtailloop:
MOVBU.P 1(R4), R8
MOVBU.P R8, 1(R3)
CMP R3, R9
BNE forwardtailloop
RET
// Small copies: 1..16 bytes.
copy16:
ADD R4, R5, R8 // R8 points just past the last source byte
ADD R3, R5, R9 // R9 points just past the last destination byte
CMP $8, R5
BLT copy7
MOVD (R4), R6
MOVD -8(R8), R7
MOVD R6, (R3)
MOVD R7, -8(R9)
RET
copy7:
TBZ $2, R5, copy3
MOVWU (R4), R6
MOVWU -4(R8), R7
MOVW R6, (R3)
MOVW R7, -4(R9)
RET
copy3:
TBZ $1, R5, copy1
MOVHU (R4), R6
MOVHU -2(R8), R7
MOVH R6, (R3)
MOVH R7, -2(R9)
RET
copy1:
MOVBU (R4), R6
MOVB R6, (R3)
RET
backward:
// Copying backwards proceeds by copying R6 bytes then copying R7/8 words.
// R3 and R4 are advanced to the end of the destination/source buffers
// respectively and moved back as we copy.
ADD R4, R5, R4 // R4 points just past the last source byte
ADD R3, R5, R3 // R3 points just past the last destination byte
CBZ R6, nobackwardtail // Do we need to do any byte-by-byte copying?
SUB R6, R3, R9 // R9 points at the lowest destination byte that should be copied by byte.
backwardtailloop:
MOVBU.W -1(R4), R8
MOVBU.W R8, -1(R3)
CMP R9, R3
BNE backwardtailloop
nobackwardtail:
CBNZ R7, backwardlarge // Do we need to do any doubleword-by-doubleword copying?
RET
backwardlarge:
SUB R7, R3, R9 // R9 points at the lowest destination byte
backwardlargeloop:
LDP -16(R4), (R8, R10)
STP (R8, R10), -16(R3)
LDP.W -32(R4), (R11, R12)
STP.W (R11, R12), -32(R3)
CMP R9, R3
BNE backwardlargeloop
RET