4
0
mirror of git://sourceware.org/git/newlib-cygwin.git synced 2025-02-14 21:19:17 +08:00
Sebastian Pop 9938a64ca9 aarch64: optimize the unaligned case of memcmp
This brings to newlib a performance improvement that we developed in Bionic
libc.  That change has been submitted for review to Bionic libc:
https://android-review.googlesource.com/418279

A similar patch has been submitted for review in glibc:
https://sourceware.org/ml/libc-alpha/2017-06/msg01143.html

Patch written by Vikas Sinha and Sebastian Pop.

The performance was measured on the bionic-benchmarks on a hikey (aarch64 8xA53)
board. There was no performance change to the existing benchmark
and a performance improvement on the new benchmark for memcmp
on the unaligned side. The new benchmark has been submitted for
review at https://android-review.googlesource.com/414860

The overall performance improves by 18% for the small data set 8
and the performance improves by 450% for the large data set 64k.

The base is with the libc from /system/lib64. The bionic libc
with this patch is in /data.

hikey:/data # export LD_LIBRARY_PATH=/system/lib64
hikey:/data # ./bionic-benchmarks --benchmark_filter='BM_string_memcmp*'
Run on (8 X 2.4 MHz CPU s)
Benchmark                                Time           CPU Iterations
----------------------------------------------------------------------
BM_string_memcmp/8                      30 ns         30 ns   22955680    251.07MB/s
BM_string_memcmp/64                     57 ns         57 ns   12349184   1076.99MB/s
BM_string_memcmp/512                   305 ns        305 ns    2297163   1.56496GB/s
BM_string_memcmp/1024                  571 ns        571 ns    1225211   1.66912GB/s
BM_string_memcmp/8k                   4307 ns       4306 ns     162562   1.77177GB/s
BM_string_memcmp/16k                  8676 ns       8675 ns      80676   1.75887GB/s
BM_string_memcmp/32k                 19233 ns      19230 ns      36394   1.58695GB/s
BM_string_memcmp/64k                 36986 ns      36984 ns      18952   1.65029GB/s
BM_string_memcmp_aligned/8             199 ns        199 ns    3519166   38.3336MB/s
BM_string_memcmp_aligned/64            386 ns        386 ns    1810734   158.073MB/s
BM_string_memcmp_aligned/512          1735 ns       1734 ns     403981   281.525MB/s
BM_string_memcmp_aligned/1024         3200 ns       3200 ns     218838   305.151MB/s
BM_string_memcmp_aligned/8k          25084 ns      25080 ns      28180   311.507MB/s
BM_string_memcmp_aligned/16k         51730 ns      51729 ns      13521   302.057MB/s
BM_string_memcmp_aligned/32k        103228 ns     103228 ns       6782   302.727MB/s
BM_string_memcmp_aligned/64k        207117 ns     207087 ns       3450   301.806MB/s
BM_string_memcmp_unaligned/8           339 ns        339 ns    2070998   22.5302MB/s
BM_string_memcmp_unaligned/64         1392 ns       1392 ns     502796   43.8454MB/s
BM_string_memcmp_unaligned/512        9194 ns       9194 ns      76133   53.1104MB/s
BM_string_memcmp_unaligned/1024      18325 ns      18323 ns      38206   53.2963MB/s
BM_string_memcmp_unaligned/8k       148579 ns     148574 ns       4713   52.5831MB/s
BM_string_memcmp_unaligned/16k      298169 ns     298120 ns       2344   52.4118MB/s
BM_string_memcmp_unaligned/32k      598813 ns     598797 ns       1085    52.188MB/s
BM_string_memcmp_unaligned/64k     1196079 ns    1196083 ns        540   52.2539MB/s

hikey:/data # export LD_LIBRARY_PATH=/data
hikey:/data # ./bionic-benchmarks --benchmark_filter='BM_string_memcmp*'
Run on (8 X 2.4 MHz CPU s)
Benchmark                                Time           CPU Iterations
----------------------------------------------------------------------
BM_string_memcmp/8                      30 ns         30 ns   23209918   252.802MB/s
BM_string_memcmp/64                     57 ns         57 ns   12348447   1076.95MB/s
BM_string_memcmp/512                   305 ns        305 ns    2296878   1.56471GB/s
BM_string_memcmp/1024                  572 ns        571 ns    1224426    1.6689GB/s
BM_string_memcmp/8k                   4309 ns       4308 ns     162491   1.77109GB/s
BM_string_memcmp/16k                  9348 ns       9345 ns      74894   1.63285GB/s
BM_string_memcmp/32k                 18329 ns      18322 ns      38249    1.6656GB/s
BM_string_memcmp/64k                 36992 ns      36981 ns      18952   1.65045GB/s
BM_string_memcmp_aligned/8             199 ns        199 ns    3513925   38.3162MB/s
BM_string_memcmp_aligned/64            386 ns        386 ns    1814038   158.192MB/s
BM_string_memcmp_aligned/512          1735 ns       1735 ns     402279   281.502MB/s
BM_string_memcmp_aligned/1024         3204 ns       3202 ns     218761   304.941MB/s
BM_string_memcmp_aligned/8k          25577 ns      25569 ns      27406   305.548MB/s
BM_string_memcmp_aligned/16k         52143 ns      52123 ns      13522   299.769MB/s
BM_string_memcmp_aligned/32k        105169 ns     105127 ns       6637    297.26MB/s
BM_string_memcmp_aligned/64k        206508 ns     206383 ns       3417   302.835MB/s
BM_string_memcmp_unaligned/8           282 ns        282 ns    2482953    27.062MB/s
BM_string_memcmp_unaligned/64          542 ns        541 ns    1298317    112.77MB/s
BM_string_memcmp_unaligned/512        2152 ns       2152 ns     325267   226.915MB/s
BM_string_memcmp_unaligned/1024       4025 ns       4025 ns     173904   242.622MB/s
BM_string_memcmp_unaligned/8k        32276 ns      32271 ns      21818    242.09MB/s
BM_string_memcmp_unaligned/16k       65970 ns      65970 ns      10554   236.851MB/s
BM_string_memcmp_unaligned/32k      131241 ns     131242 ns       5129    238.11MB/s
BM_string_memcmp_unaligned/64k      266159 ns     266160 ns       2661   234.821MB/s
2017-06-26 10:22:40 +02:00

221 lines
6.0 KiB
ArmAsm

/* memcmp - compare memory
Copyright (c) 2013, Linaro Limited
Copyright (c) 2017, Samsung Austin R&D Center
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
* Neither the name of the Linaro nor the
names of its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
#if (defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED))
/* See memcmp-stub.c */
#else
/* Assumptions:
*
* ARMv8-a, AArch64
*/
.macro def_fn f p2align=0
.text
.p2align \p2align
.global \f
.type \f, %function
\f:
.endm
/* Parameters and result. */
#define src1 x0
#define src2 x1
#define limit x2
#define result x0
/* Internal variables. */
#define data1 x3
#define data1w w3
#define data2 x4
#define data2w w4
#define has_nul x5
#define diff x6
#define endloop x7
#define tmp1 x8
#define tmp2 x9
#define tmp3 x10
#define pos x11
#define limit_wd x12
#define mask x13
def_fn memcmp p2align=6
cbz limit, .Lret0
eor tmp1, src1, src2
tst tmp1, #7
b.ne .Lmisaligned8
ands tmp1, src1, #7
b.ne .Lmutual_align
add limit_wd, limit, #7
lsr limit_wd, limit_wd, #3
/* Start of performance-critical section -- one 64B cache line. */
.Lloop_aligned:
ldr data1, [src1], #8
ldr data2, [src2], #8
.Lstart_realigned:
subs limit_wd, limit_wd, #1
eor diff, data1, data2 /* Non-zero if differences found. */
csinv endloop, diff, xzr, ne /* Last Dword or differences. */
cbz endloop, .Lloop_aligned
/* End of performance-critical section -- one 64B cache line. */
/* Not reached the limit, must have found a diff. */
cbnz limit_wd, .Lnot_limit
/* Limit % 8 == 0 => all bytes significant. */
ands limit, limit, #7
b.eq .Lnot_limit
lsl limit, limit, #3 /* Bits -> bytes. */
mov mask, #~0
#ifdef __AARCH64EB__
lsr mask, mask, limit
#else
lsl mask, mask, limit
#endif
bic data1, data1, mask
bic data2, data2, mask
orr diff, diff, mask
.Lnot_limit:
#ifndef __AARCH64EB__
rev diff, diff
rev data1, data1
rev data2, data2
#endif
/* The MS-non-zero bit of DIFF marks either the first bit
that is different, or the end of the significant data.
Shifting left now will bring the critical information into the
top bits. */
clz pos, diff
lsl data1, data1, pos
lsl data2, data2, pos
/* But we need to zero-extend (char is unsigned) the value and then
perform a signed 32-bit subtraction. */
lsr data1, data1, #56
sub result, data1, data2, lsr #56
ret
.Lmutual_align:
/* Sources are mutually aligned, but are not currently at an
alignment boundary. Round down the addresses and then mask off
the bytes that precede the start point. */
bic src1, src1, #7
bic src2, src2, #7
add limit, limit, tmp1 /* Adjust the limit for the extra. */
lsl tmp1, tmp1, #3 /* Bytes beyond alignment -> bits. */
ldr data1, [src1], #8
neg tmp1, tmp1 /* Bits to alignment -64. */
ldr data2, [src2], #8
mov tmp2, #~0
#ifdef __AARCH64EB__
/* Big-endian. Early bytes are at MSB. */
lsl tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */
#else
/* Little-endian. Early bytes are at LSB. */
lsr tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */
#endif
add limit_wd, limit, #7
orr data1, data1, tmp2
orr data2, data2, tmp2
lsr limit_wd, limit_wd, #3
b .Lstart_realigned
.Lret0:
mov result, #0
ret
.p2align 6
.Lmisaligned8:
cmp limit, #8
b.lo .LmisalignedLt8
.LunalignedGe8 :
/* Load the first dword with both src potentially unaligned. */
ldr data1, [src1]
ldr data2, [src2]
eor diff, data1, data2 /* Non-zero if differences found. */
cbnz diff, .Lnot_limit
/* Sources are not aligned: align one of the sources. */
and tmp1, src1, #0x7
orr tmp3, xzr, #0x8
sub pos, tmp3, tmp1
/* Increment SRC pointers by POS so SRC1 is word-aligned. */
add src1, src1, pos
add src2, src2, pos
sub limit, limit, pos
lsr limit_wd, limit, #3
cmp limit_wd, #0
/* save #bytes to go back to be able to read 8byte at end
pos=negative offset position to read 8 bytes when len%8 != 0 */
and limit, limit, #7
sub pos, limit, #8
b .Lstart_part_realigned
.p2align 5
.Lloop_part_aligned:
ldr data1, [src1], #8
ldr data2, [src2], #8
subs limit_wd, limit_wd, #1
.Lstart_part_realigned:
eor diff, data1, data2 /* Non-zero if differences found. */
cbnz diff, .Lnot_limit
b.ne .Lloop_part_aligned
/* process leftover bytes: read the leftover bytes, starting with
negative offset - so we can load 8 bytes. */
ldr data1, [src1, pos]
ldr data2, [src2, pos]
eor diff, data1, data2 /* Non-zero if differences found. */
b .Lnot_limit
.LmisalignedLt8:
sub limit, limit, #1
1:
ldrb data1w, [src1], #1
ldrb data2w, [src2], #1
subs limit, limit, #1
ccmp data1w, data2w, #0, cs /* NZCV = 0b0000. */
b.eq 1b
sub result, data1, data2
ret
.size memcmp, . - memcmp
#endif