1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
|
/* Copyright (c) 2009 CodeSourcery, Inc.
* 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 CodeSourcery 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 CODESOURCERY, INC. ``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 CODESOURCERY 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.
*/
#include "arm_asm.h"
#include <string.h>
#include <stdint.h>
#include <stddef.h>
/* Standard operations for word-sized values. */
#define WORD_REF(ADDRESS, OFFSET) \
*((WORD_TYPE*)((char*)(ADDRESS) + (OFFSET)))
#define WORD_COPY(OUT, IN, OFFSET) \
WORD_REF(OUT, OFFSET) = WORD_REF(IN, OFFSET)
/* On processors with NEON, we use 128-bit vectors. Also,
we need to include arm_neon.h to use these. */
#if defined(__ARM_NEON__)
#include <arm_neon.h>
#define WORD_TYPE uint8x16_t
#define WORD_SIZE 16
#define MAYBE_PREFETCH(IN) __builtin_prefetch((IN), 0, 0)
/* On ARM processors with 64-bit ldrd instructions, we use those,
except on Cortex-M* where benchmarking has shown them to
be slower. */
#elif defined(__ARM_ARCH_5E__) || defined(__ARM_ARCH_5TE__) \
|| defined(__ARM_ARCH_5TEJ__) || defined(_ISA_ARM_6)
#define WORD_TYPE uint64_t
#define WORD_SIZE 8
#define MAYBE_PREFETCH(IN) __builtin_prefetch((IN), 0, 0)
/* On everything else, we use 32-bit loads and stores, and
do not use prefetching. */
#else
#define WORD_TYPE uint32_t
#define WORD_SIZE 4
#define MAYBE_PREFETCH(IN)
#endif
/* On all ARM platforms, 'SHORTWORD' is a 32-bit value. */
#define SHORTWORD_TYPE uint32_t
#define SHORTWORD_SIZE 4
#define SHORTWORD_REF(ADDRESS, OFFSET) \
*((SHORTWORD_TYPE*)((char*)(ADDRESS) + (OFFSET)))
#define SHORTWORD_COPY(OUT, IN, OFFSET) \
SHORTWORD_REF(OUT, OFFSET) = SHORTWORD_REF(IN, OFFSET)
/* Shifting directionality depends on endianness. */
#ifdef __ARMEB__
#define SHORTWORD_SHIFT(IN0, IN1, OFFSET) \
((IN0) << ((OFFSET)*8)) | ((IN1) >> (SHORTWORD_SIZE*8 - (OFFSET)*8))
#else
#define SHORTWORD_SHIFT(IN0, IN1, OFFSET) \
((IN0) >> ((OFFSET)*8)) | ((IN1) << (SHORTWORD_SIZE*8 - (OFFSET)*8))
#endif
void *memcpy(void *OUT, const void *IN, size_t N)
{
void* OUT0 = OUT;
#if defined(PREFER_SIZE_OVER_SPEED) || defined(__OPTIMIZE_SIZE__)
const char* OUT_end = (char*)OUT + N;
while ((char*)OUT < OUT_end) {
*((char*)OUT) = *((char*)IN);
OUT++;
IN++;
}
return OUT0;
#else
/* Handle short strings and immediately return. */
if (__builtin_expect(N < SHORTWORD_SIZE, 1)) {
size_t i = 0;
while (i < N) {
((char*)OUT)[i] = ((char*)IN)[i];
i++;
}
return OUT;
}
const char* OUT_end = (char*)OUT + N;
/* Align OUT to SHORTWORD_SIZE. */
while ((uintptr_t)OUT % SHORTWORD_SIZE != 0) {
*(char*) (OUT++) = *(char*) (IN++);
}
if ((uintptr_t) IN % SHORTWORD_SIZE == 0) {
#if WORD_SIZE > SHORTWORD_SIZE
/* Align OUT to WORD_SIZE in steps of SHORTWORD_SIZE. */
if (__builtin_expect(OUT_end - (char*)OUT >= WORD_SIZE, 0)) {
while ((uintptr_t)OUT % WORD_SIZE != 0) {
SHORTWORD_COPY(OUT, IN, 0);
OUT += SHORTWORD_SIZE;
IN += SHORTWORD_SIZE;
}
if ((uintptr_t) IN % WORD_SIZE == 0) {
#endif /* WORD_SIZE > SHORTWORD_SIZE */
#if defined(__ARM_NEON__)
/* Testing on Cortex-A8 indicates that the following idiom
produces faster assembly code when doing vector copies,
but not when doing regular copies. */
size_t i = 0;
N = OUT_end - (char*)OUT;
MAYBE_PREFETCH(IN + 64);
MAYBE_PREFETCH(IN + 128);
MAYBE_PREFETCH(IN + 192);
if (N >= 640) {
MAYBE_PREFETCH(IN + 256);
MAYBE_PREFETCH(IN + 320);
MAYBE_PREFETCH(IN + 384);
MAYBE_PREFETCH(IN + 448);
MAYBE_PREFETCH(IN + 512);
MAYBE_PREFETCH(IN + 576);
MAYBE_PREFETCH(IN + 640);
MAYBE_PREFETCH(IN + 704);
/* We phrase the loop condition in this way so that the
i + WORD_SIZE * 16 value can be reused to increment i. */
while (i + WORD_SIZE * 16 <= N - 640) {
MAYBE_PREFETCH(IN + 768);
MAYBE_PREFETCH(IN + 832);
MAYBE_PREFETCH(IN + 896);
MAYBE_PREFETCH(IN + 960);
WORD_COPY(OUT, IN, i);
WORD_COPY(OUT, IN, i + WORD_SIZE * 1);
WORD_COPY(OUT, IN, i + WORD_SIZE * 2);
WORD_COPY(OUT, IN, i + WORD_SIZE * 3);
WORD_COPY(OUT, IN, i + WORD_SIZE * 4);
WORD_COPY(OUT, IN, i + WORD_SIZE * 5);
WORD_COPY(OUT, IN, i + WORD_SIZE * 6);
WORD_COPY(OUT, IN, i + WORD_SIZE * 7);
WORD_COPY(OUT, IN, i + WORD_SIZE * 8);
WORD_COPY(OUT, IN, i + WORD_SIZE * 9);
WORD_COPY(OUT, IN, i + WORD_SIZE * 10);
WORD_COPY(OUT, IN, i + WORD_SIZE * 11);
WORD_COPY(OUT, IN, i + WORD_SIZE * 12);
WORD_COPY(OUT, IN, i + WORD_SIZE * 13);
WORD_COPY(OUT, IN, i + WORD_SIZE * 14);
WORD_COPY(OUT, IN, i + WORD_SIZE * 15);
i += WORD_SIZE * 16;
}
}
while (i + WORD_SIZE * 16 <= N) {
WORD_COPY(OUT, IN, i);
WORD_COPY(OUT, IN, i + WORD_SIZE * 1);
WORD_COPY(OUT, IN, i + WORD_SIZE * 2);
WORD_COPY(OUT, IN, i + WORD_SIZE * 3);
WORD_COPY(OUT, IN, i + WORD_SIZE * 4);
WORD_COPY(OUT, IN, i + WORD_SIZE * 5);
WORD_COPY(OUT, IN, i + WORD_SIZE * 6);
WORD_COPY(OUT, IN, i + WORD_SIZE * 7);
WORD_COPY(OUT, IN, i + WORD_SIZE * 8);
WORD_COPY(OUT, IN, i + WORD_SIZE * 9);
WORD_COPY(OUT, IN, i + WORD_SIZE * 10);
WORD_COPY(OUT, IN, i + WORD_SIZE * 11);
WORD_COPY(OUT, IN, i + WORD_SIZE * 12);
WORD_COPY(OUT, IN, i + WORD_SIZE * 13);
WORD_COPY(OUT, IN, i + WORD_SIZE * 14);
WORD_COPY(OUT, IN, i + WORD_SIZE * 15);
i += WORD_SIZE * 16;
}
while (i + WORD_SIZE * 4 <= N) {
WORD_COPY(OUT, IN, i);
WORD_COPY(OUT, IN, i + WORD_SIZE * 1);
WORD_COPY(OUT, IN, i + WORD_SIZE * 2);
WORD_COPY(OUT, IN, i + WORD_SIZE * 3);
i += WORD_SIZE * 4;
}
while (i + WORD_SIZE <= N) {
WORD_COPY(OUT, IN, i);
i += WORD_SIZE;
}
OUT += i;
IN += i;
#else /* not defined(__ARM_NEON__) */
/* Note: 16-times unrolling is about 20% faster than 4-times
unrolling on both ARM Cortex-A8 and Cortex-M3. */
MAYBE_PREFETCH(IN + 64);
MAYBE_PREFETCH(IN + 128);
MAYBE_PREFETCH(IN + 192);
while (OUT_end - (char*)OUT >= WORD_SIZE * 16) {
MAYBE_PREFETCH(IN + 256);
MAYBE_PREFETCH(IN + 320);
WORD_COPY(OUT, IN, 0);
WORD_COPY(OUT, IN, WORD_SIZE * 1);
WORD_COPY(OUT, IN, WORD_SIZE * 2);
WORD_COPY(OUT, IN, WORD_SIZE * 3);
WORD_COPY(OUT, IN, WORD_SIZE * 4);
WORD_COPY(OUT, IN, WORD_SIZE * 5);
WORD_COPY(OUT, IN, WORD_SIZE * 6);
WORD_COPY(OUT, IN, WORD_SIZE * 7);
WORD_COPY(OUT, IN, WORD_SIZE * 8);
WORD_COPY(OUT, IN, WORD_SIZE * 9);
WORD_COPY(OUT, IN, WORD_SIZE * 10);
WORD_COPY(OUT, IN, WORD_SIZE * 11);
WORD_COPY(OUT, IN, WORD_SIZE * 12);
WORD_COPY(OUT, IN, WORD_SIZE * 13);
WORD_COPY(OUT, IN, WORD_SIZE * 14);
WORD_COPY(OUT, IN, WORD_SIZE * 15);
OUT += WORD_SIZE * 16;
IN += WORD_SIZE * 16;
}
while (WORD_SIZE * 4 <= OUT_end - (char*)OUT) {
WORD_COPY(OUT, IN, 0);
WORD_COPY(OUT, IN, WORD_SIZE * 1);
WORD_COPY(OUT, IN, WORD_SIZE * 2);
WORD_COPY(OUT, IN, WORD_SIZE * 3);
OUT += WORD_SIZE * 4;
IN += WORD_SIZE * 4;
}
while (WORD_SIZE <= OUT_end - (char*)OUT) {
WORD_COPY(OUT, IN, 0);
OUT += WORD_SIZE;
IN += WORD_SIZE;
}
#endif /* not defined(__ARM_NEON__) */
#if WORD_SIZE > SHORTWORD_SIZE
} else { /* if IN is not WORD_SIZE aligned */
while (SHORTWORD_SIZE * 4 <= OUT_end - (char*)OUT) {
SHORTWORD_COPY(OUT, IN, 0);
SHORTWORD_COPY(OUT, IN, SHORTWORD_SIZE * 1);
SHORTWORD_COPY(OUT, IN, SHORTWORD_SIZE * 2);
SHORTWORD_COPY(OUT, IN, SHORTWORD_SIZE * 3);
OUT += SHORTWORD_SIZE * 4;
IN += SHORTWORD_SIZE * 4;
}
} /* end if IN is not WORD_SIZE aligned */
} /* end if N >= WORD_SIZE */
while (SHORTWORD_SIZE <= OUT_end - (char*)OUT) {
SHORTWORD_COPY(OUT, IN, 0);
OUT += SHORTWORD_SIZE;
IN += SHORTWORD_SIZE;
}
#endif /* WORD_SIZE > SHORTWORD_SIZE */
} else { /* if IN is not SHORTWORD_SIZE aligned */
ptrdiff_t misalign = (uintptr_t)IN % SHORTWORD_SIZE;
SHORTWORD_TYPE temp1, temp2;
temp1 = SHORTWORD_REF(IN, -misalign);
/* Benchmarking indicates that unrolling this loop doesn't
produce a measurable performance improvement on ARM. */
while (SHORTWORD_SIZE <= OUT_end - (char*)OUT) {
IN += SHORTWORD_SIZE;
temp2 = SHORTWORD_REF(IN, -misalign);
SHORTWORD_REF(OUT, 0) = SHORTWORD_SHIFT(temp1, temp2, misalign);
temp1 = temp2;
OUT += SHORTWORD_SIZE;
}
} /* end if IN is not SHORTWORD_SIZE aligned */
while ((char*)OUT < OUT_end) {
*((char*)OUT) = *((char*)IN);
OUT++;
IN++;
}
return OUT0;
#endif
}
|