1 /* Copyright (C) 1991, 1993, 1995, 1997, 1998 Free Software Foundation, Inc.
2 Contributed by Torbjorn Granlund (tege@sics.se).
4 NOTE: The canonical source of this file is maintained with the GNU C Library.
5 Bugs can be reported to bug-glibc@prep.ai.mit.edu.
7 This program is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 2, or (at your option) any
12 This program is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with this program; if not, write to the Free Software
19 Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307,
27 #if defined __cplusplus || (defined __STDC__ && __STDC__)
28 # define __ptr_t void *
29 #else /* Not C++ or ANSI C. */
32 # define __ptr_t char *
33 #endif /* C++ or ANSI C. */
36 # if defined __GNUC__ || (defined __STDC__ && __STDC__)
37 # define __P(args) args
43 #if defined HAVE_STRING_H || defined _LIBC
53 #else /* Not in the GNU C library. */
55 # include <sys/types.h>
57 /* Type to use for aligned memory operations.
58 This should normally be the biggest type supported by a single load
59 and store. Must be an unsigned type. */
60 # define op_t unsigned long int
61 # define OPSIZ (sizeof(op_t))
63 /* Threshold value for when to enter the unrolled loops. */
64 # define OP_T_THRES 16
66 /* Type to use for unaligned operations. */
67 typedef unsigned char byte;
69 # ifndef WORDS_BIGENDIAN
70 # define MERGE(w0, sh_1, w1, sh_2) (((w0) >> (sh_1)) | ((w1) << (sh_2)))
72 # define MERGE(w0, sh_1, w1, sh_2) (((w0) << (sh_1)) | ((w1) >> (sh_2)))
75 #endif /* In the GNU C library. */
77 #ifdef WORDS_BIGENDIAN
78 # define CMP_LT_OR_GT(a, b) ((a) > (b) ? 1 : -1)
80 # define CMP_LT_OR_GT(a, b) memcmp_bytes ((a), (b))
83 /* BE VERY CAREFUL IF YOU CHANGE THIS CODE! */
85 /* The strategy of this memcmp is:
87 1. Compare bytes until one of the block pointers is aligned.
89 2. Compare using memcmp_common_alignment or
90 memcmp_not_common_alignment, regarding the alignment of the other
91 block after the initial byte operations. The maximum number of
92 full words (of type op_t) are compared in this way.
94 3. Compare the few remaining bytes. */
96 #ifndef WORDS_BIGENDIAN
97 /* memcmp_bytes -- Compare A and B bytewise in the byte order of the machine.
98 A and B are known to be different.
99 This is needed only on little-endian machines. */
101 static int memcmp_bytes __P((op_t, op_t));
107 memcmp_bytes (long unsigned int a, long unsigned int b)
109 long int srcp1 = (long int) &a;
110 long int srcp2 = (long int) &b;
115 a0 = ((byte *) srcp1)[0];
116 b0 = ((byte *) srcp2)[0];
125 static int memcmp_common_alignment __P((long, long, size_t));
127 /* memcmp_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN `op_t'
128 objects (not LEN bytes!). Both SRCP1 and SRCP2 should be aligned for
129 memory operations on `op_t's. */
134 memcmp_common_alignment (long int srcp1, long int srcp2, size_t len)
141 default: /* Avoid warning about uninitialized local variables. */
143 a0 = ((op_t *) srcp1)[0];
144 b0 = ((op_t *) srcp2)[0];
150 a1 = ((op_t *) srcp1)[0];
151 b1 = ((op_t *) srcp2)[0];
157 if (OP_T_THRES <= 3 * OPSIZ && len == 0)
159 a0 = ((op_t *) srcp1)[0];
160 b0 = ((op_t *) srcp2)[0];
163 a1 = ((op_t *) srcp1)[0];
164 b1 = ((op_t *) srcp2)[0];
168 if (OP_T_THRES <= 3 * OPSIZ && len == 0)
175 a0 = ((op_t *) srcp1)[0];
176 b0 = ((op_t *) srcp2)[0];
178 return CMP_LT_OR_GT (a1, b1);
181 a1 = ((op_t *) srcp1)[1];
182 b1 = ((op_t *) srcp2)[1];
184 return CMP_LT_OR_GT (a0, b0);
187 a0 = ((op_t *) srcp1)[2];
188 b0 = ((op_t *) srcp2)[2];
190 return CMP_LT_OR_GT (a1, b1);
193 a1 = ((op_t *) srcp1)[3];
194 b1 = ((op_t *) srcp2)[3];
196 return CMP_LT_OR_GT (a0, b0);
204 /* This is the right position for do0. Please don't move
208 return CMP_LT_OR_GT (a1, b1);
212 static int memcmp_not_common_alignment __P((long, long, size_t));
214 /* memcmp_not_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN
215 `op_t' objects (not LEN bytes!). SRCP2 should be aligned for memory
216 operations on `op_t', but SRCP1 *should be unaligned*. */
221 memcmp_not_common_alignment (long int srcp1, long int srcp2, size_t len)
228 /* Calculate how to shift a word read at the memory operation
229 aligned srcp1 to make it aligned for comparison. */
231 shl = 8 * (srcp1 % OPSIZ);
232 shr = 8 * OPSIZ - shl;
234 /* Make SRCP1 aligned by rounding it down to the beginning of the `op_t'
235 it points in the middle of. */
240 default: /* Avoid warning about uninitialized local variables. */
242 a1 = ((op_t *) srcp1)[0];
243 a2 = ((op_t *) srcp1)[1];
244 b2 = ((op_t *) srcp2)[0];
250 a0 = ((op_t *) srcp1)[0];
251 a1 = ((op_t *) srcp1)[1];
252 b1 = ((op_t *) srcp2)[0];
257 if (OP_T_THRES <= 3 * OPSIZ && len == 0)
259 a3 = ((op_t *) srcp1)[0];
260 a0 = ((op_t *) srcp1)[1];
261 b0 = ((op_t *) srcp2)[0];
265 a2 = ((op_t *) srcp1)[0];
266 a3 = ((op_t *) srcp1)[1];
267 b3 = ((op_t *) srcp2)[0];
271 if (OP_T_THRES <= 3 * OPSIZ && len == 0)
278 a0 = ((op_t *) srcp1)[0];
279 b0 = ((op_t *) srcp2)[0];
280 x = MERGE(a2, shl, a3, shr);
282 return CMP_LT_OR_GT (x, b3);
285 a1 = ((op_t *) srcp1)[1];
286 b1 = ((op_t *) srcp2)[1];
287 x = MERGE(a3, shl, a0, shr);
289 return CMP_LT_OR_GT (x, b0);
292 a2 = ((op_t *) srcp1)[2];
293 b2 = ((op_t *) srcp2)[2];
294 x = MERGE(a0, shl, a1, shr);
296 return CMP_LT_OR_GT (x, b1);
299 a3 = ((op_t *) srcp1)[3];
300 b3 = ((op_t *) srcp2)[3];
301 x = MERGE(a1, shl, a2, shr);
303 return CMP_LT_OR_GT (x, b2);
311 /* This is the right position for do0. Please don't move
314 x = MERGE(a2, shl, a3, shr);
316 return CMP_LT_OR_GT (x, b3);
321 rpl_memcmp (const void *s1, const void *s2, size_t len)
325 long int srcp1 = (long int) s1;
326 long int srcp2 = (long int) s2;
329 if (len >= OP_T_THRES)
331 /* There are at least some bytes to compare. No need to test
332 for LEN == 0 in this alignment loop. */
333 while (srcp2 % OPSIZ != 0)
335 a0 = ((byte *) srcp1)[0];
336 b0 = ((byte *) srcp2)[0];
345 /* SRCP2 is now aligned for memory operations on `op_t'.
346 SRCP1 alignment determines if we can do a simple,
347 aligned compare or need to shuffle bits. */
349 if (srcp1 % OPSIZ == 0)
350 res = memcmp_common_alignment (srcp1, srcp2, len / OPSIZ);
352 res = memcmp_not_common_alignment (srcp1, srcp2, len / OPSIZ);
356 /* Number of bytes remaining in the interval [0..OPSIZ-1]. */
357 srcp1 += len & -OPSIZ;
358 srcp2 += len & -OPSIZ;
362 /* There are just a few bytes to compare. Use byte memory operations. */
365 a0 = ((byte *) srcp1)[0];
366 b0 = ((byte *) srcp2)[0];
380 weak_alias (memcmp, bcmp)