1
/* Copyright (C) 1991, 1993, 1995, 1997, 1998, 2003, 2006 Free Software
4
Contributed by Torbjorn Granlund (tege@sics.se).
6
NOTE: The canonical source of this file is maintained with the GNU C Library.
7
Bugs can be reported to bug-glibc@prep.ai.mit.edu.
9
This program is free software; you can redistribute it and/or modify it
10
under the terms of the GNU General Public License as published by the
11
Free Software Foundation; either version 2, or (at your option) any
14
This program is distributed in the hope that it will be useful,
15
but WITHOUT ANY WARRANTY; without even the implied warranty of
16
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17
GNU General Public License for more details.
19
You should have received a copy of the GNU General Public License
20
along with this program; if not, write to the Free Software
21
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301,
37
# if __BYTE_ORDER == __BIG_ENDIAN
38
# define WORDS_BIGENDIAN
41
#else /* Not in the GNU C library. */
43
# include <sys/types.h>
45
/* Type to use for aligned memory operations.
46
This should normally be the biggest type supported by a single load
47
and store. Must be an unsigned type. */
48
# define op_t unsigned long int
49
# define OPSIZ (sizeof(op_t))
51
/* Threshold value for when to enter the unrolled loops. */
52
# define OP_T_THRES 16
54
/* Type to use for unaligned operations. */
55
typedef unsigned char byte;
57
# ifndef WORDS_BIGENDIAN
58
# define MERGE(w0, sh_1, w1, sh_2) (((w0) >> (sh_1)) | ((w1) << (sh_2)))
60
# define MERGE(w0, sh_1, w1, sh_2) (((w0) << (sh_1)) | ((w1) >> (sh_2)))
63
#endif /* In the GNU C library. */
65
#ifdef WORDS_BIGENDIAN
66
# define CMP_LT_OR_GT(a, b) ((a) > (b) ? 1 : -1)
68
# define CMP_LT_OR_GT(a, b) memcmp_bytes ((a), (b))
71
/* BE VERY CAREFUL IF YOU CHANGE THIS CODE! */
73
/* The strategy of this memcmp is:
75
1. Compare bytes until one of the block pointers is aligned.
77
2. Compare using memcmp_common_alignment or
78
memcmp_not_common_alignment, regarding the alignment of the other
79
block after the initial byte operations. The maximum number of
80
full words (of type op_t) are compared in this way.
82
3. Compare the few remaining bytes. */
84
#ifndef WORDS_BIGENDIAN
85
/* memcmp_bytes -- Compare A and B bytewise in the byte order of the machine.
86
A and B are known to be different.
87
This is needed only on little-endian machines. */
93
memcmp_bytes (long unsigned int a, long unsigned int b)
95
long int srcp1 = (long int) &a;
96
long int srcp2 = (long int) &b;
101
a0 = ((byte *) srcp1)[0];
102
b0 = ((byte *) srcp2)[0];
111
/* memcmp_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN `op_t'
112
objects (not LEN bytes!). Both SRCP1 and SRCP2 should be aligned for
113
memory operations on `op_t's. */
118
memcmp_common_alignment (long int srcp1, long int srcp2, size_t len)
125
default: /* Avoid warning about uninitialized local variables. */
127
a0 = ((op_t *) srcp1)[0];
128
b0 = ((op_t *) srcp2)[0];
134
a1 = ((op_t *) srcp1)[0];
135
b1 = ((op_t *) srcp2)[0];
141
if (OP_T_THRES <= 3 * OPSIZ && len == 0)
143
a0 = ((op_t *) srcp1)[0];
144
b0 = ((op_t *) srcp2)[0];
147
a1 = ((op_t *) srcp1)[0];
148
b1 = ((op_t *) srcp2)[0];
152
if (OP_T_THRES <= 3 * OPSIZ && len == 0)
159
a0 = ((op_t *) srcp1)[0];
160
b0 = ((op_t *) srcp2)[0];
162
return CMP_LT_OR_GT (a1, b1);
165
a1 = ((op_t *) srcp1)[1];
166
b1 = ((op_t *) srcp2)[1];
168
return CMP_LT_OR_GT (a0, b0);
171
a0 = ((op_t *) srcp1)[2];
172
b0 = ((op_t *) srcp2)[2];
174
return CMP_LT_OR_GT (a1, b1);
177
a1 = ((op_t *) srcp1)[3];
178
b1 = ((op_t *) srcp2)[3];
180
return CMP_LT_OR_GT (a0, b0);
188
/* This is the right position for do0. Please don't move
192
return CMP_LT_OR_GT (a1, b1);
196
/* memcmp_not_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN
197
`op_t' objects (not LEN bytes!). SRCP2 should be aligned for memory
198
operations on `op_t', but SRCP1 *should be unaligned*. */
203
memcmp_not_common_alignment (long int srcp1, long int srcp2, size_t len)
210
/* Calculate how to shift a word read at the memory operation
211
aligned srcp1 to make it aligned for comparison. */
213
shl = 8 * (srcp1 % OPSIZ);
214
shr = 8 * OPSIZ - shl;
216
/* Make SRCP1 aligned by rounding it down to the beginning of the `op_t'
217
it points in the middle of. */
222
default: /* Avoid warning about uninitialized local variables. */
224
a1 = ((op_t *) srcp1)[0];
225
a2 = ((op_t *) srcp1)[1];
226
b2 = ((op_t *) srcp2)[0];
232
a0 = ((op_t *) srcp1)[0];
233
a1 = ((op_t *) srcp1)[1];
234
b1 = ((op_t *) srcp2)[0];
239
if (OP_T_THRES <= 3 * OPSIZ && len == 0)
241
a3 = ((op_t *) srcp1)[0];
242
a0 = ((op_t *) srcp1)[1];
243
b0 = ((op_t *) srcp2)[0];
247
a2 = ((op_t *) srcp1)[0];
248
a3 = ((op_t *) srcp1)[1];
249
b3 = ((op_t *) srcp2)[0];
253
if (OP_T_THRES <= 3 * OPSIZ && len == 0)
260
a0 = ((op_t *) srcp1)[0];
261
b0 = ((op_t *) srcp2)[0];
262
x = MERGE(a2, shl, a3, shr);
264
return CMP_LT_OR_GT (x, b3);
267
a1 = ((op_t *) srcp1)[1];
268
b1 = ((op_t *) srcp2)[1];
269
x = MERGE(a3, shl, a0, shr);
271
return CMP_LT_OR_GT (x, b0);
274
a2 = ((op_t *) srcp1)[2];
275
b2 = ((op_t *) srcp2)[2];
276
x = MERGE(a0, shl, a1, shr);
278
return CMP_LT_OR_GT (x, b1);
281
a3 = ((op_t *) srcp1)[3];
282
b3 = ((op_t *) srcp2)[3];
283
x = MERGE(a1, shl, a2, shr);
285
return CMP_LT_OR_GT (x, b2);
293
/* This is the right position for do0. Please don't move
296
x = MERGE(a2, shl, a3, shr);
298
return CMP_LT_OR_GT (x, b3);
303
rpl_memcmp (const void *s1, const void *s2, size_t len)
307
long int srcp1 = (long int) s1;
308
long int srcp2 = (long int) s2;
311
if (len >= OP_T_THRES)
313
/* There are at least some bytes to compare. No need to test
314
for LEN == 0 in this alignment loop. */
315
while (srcp2 % OPSIZ != 0)
317
a0 = ((byte *) srcp1)[0];
318
b0 = ((byte *) srcp2)[0];
327
/* SRCP2 is now aligned for memory operations on `op_t'.
328
SRCP1 alignment determines if we can do a simple,
329
aligned compare or need to shuffle bits. */
331
if (srcp1 % OPSIZ == 0)
332
res = memcmp_common_alignment (srcp1, srcp2, len / OPSIZ);
334
res = memcmp_not_common_alignment (srcp1, srcp2, len / OPSIZ);
338
/* Number of bytes remaining in the interval [0..OPSIZ-1]. */
339
srcp1 += len & -OPSIZ;
340
srcp2 += len & -OPSIZ;
344
/* There are just a few bytes to compare. Use byte memory operations. */
347
a0 = ((byte *) srcp1)[0];
348
b0 = ((byte *) srcp2)[0];
362
weak_alias (memcmp, bcmp)