1 /* A memmove for CRIS.
2 Copyright (C) 2000-2005 Axis Communications.
3 All rights reserved.
4
5 Redistribution and use in source and binary forms, with or without
6 modification, are permitted provided that the following conditions
7 are met:
8
9 1. Redistributions of source code must retain the above copyright
10 notice, this list of conditions and the following disclaimer.
11
12 2. Neither the name of Axis Communications nor the names of its
13 contributors may be used to endorse or promote products derived
14 from this software without specific prior written permission.
15
16 THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
17 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
20 COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
21 INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22 (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
25 STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
26 IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27 POSSIBILITY OF SUCH DAMAGE. */
28
29 /* FIXME: This file should really only be used for reference, as the
30 result is somewhat depending on gcc generating what we expect rather
31 than what we describe. An assembly file should be used instead.
32
33 Even worse, we base it on memcpy, on the assumption that overlapping
34 moves are rare, and we will do no worse than the generic memmove. */
35
36 #include <picolibc.h>
37
38 #include <stddef.h>
39
40 /* Break even between movem and move16 is really at 38.7 * 2, but
41 modulo 44, so up to the next multiple of 44, we use ordinary code. */
42 #define MEMMOVE_BY_BLOCK_THRESHOLD (44 * 2)
43
44 /* No name ambiguities in this file. */
45 __asm__ (".syntax no_register_prefix");
46
47 void *
memmove(void * pdst,const void * psrc,size_t pn)48 memmove(void *pdst, const void *psrc, size_t pn)
49 {
50 /* Now we want the parameters put in special registers.
51 Make sure the compiler is able to make something useful of this.
52 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
53
54 If gcc was allright, it really would need no temporaries, and no
55 stack space to save stuff on. */
56
57 register void *return_dst __asm__ ("r10") = pdst;
58 register unsigned char *dst __asm__ ("r13") = pdst;
59 register unsigned const char *src __asm__ ("r11") = psrc;
60 register int n __asm__ ("r12") = pn;
61
62 /* Check and handle overlap. */
63 if (src < dst && dst < src + n)
64 {
65 /* Destructive overlap. We could optimize this, but we don't (for
66 the moment). */
67 src += n;
68 dst += n;
69 while (n--)
70 {
71 *--dst = *--src;
72 }
73
74 return return_dst;
75 }
76 /* Whew, no overlap. Proceed as with memcpy. We could call it instead
77 of having a copy here. That would spoil some of the optimization, so
78 we take the trouble with having two copies. */
79
80 /* When src is aligned but not dst, this makes a few extra needless
81 cycles. I believe it would take as many to check that the
82 re-alignment was unnecessary. */
83 if (((unsigned long) dst & 3) != 0
84 /* Don't align if we wouldn't copy more than a few bytes; so we
85 don't have to check further for overflows. */
86 && n >= 3)
87 {
88 if ((unsigned long) dst & 1)
89 {
90 n--;
91 *dst = *src;
92 src++;
93 dst++;
94 }
95
96 if ((unsigned long) dst & 2)
97 {
98 n -= 2;
99 *(short *) dst = *(short *) src;
100 src += 2;
101 dst += 2;
102 }
103 }
104
105 /* Decide which copying method to use. */
106 if (n >= MEMMOVE_BY_BLOCK_THRESHOLD)
107 {
108 /* It is not optimal to tell the compiler about clobbering any
109 registers; that will move the saving/restoring of those registers
110 to the function prologue/epilogue, and make non-movem sizes
111 suboptimal. */
112 __asm__ volatile
113 ("\
114 ;; GCC does promise correct register allocations, but let's \n\
115 ;; make sure it keeps its promises. \n\
116 .ifnc %0-%1-%2,$r13-$r11-$r12 \n\
117 .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\" \n\
118 .endif \n\
119 \n\
120 ;; Save the registers we'll use in the movem process \n\
121 ;; on the stack. \n\
122 subq 11*4,sp \n\
123 movem r10,[sp] \n\
124 \n\
125 ;; Now we've got this: \n\
126 ;; r11 - src \n\
127 ;; r13 - dst \n\
128 ;; r12 - n \n\
129 \n\
130 ;; Update n for the first loop. \n\
131 subq 44,r12 \n\
132 0: \n\
133 "
134 #ifdef __arch_common_v10_v32
135 /* Cater to branch offset difference between v32 and v10. We
136 assume the branch below has an 8-bit offset. */
137 " setf\n"
138 #endif
139 " movem [r11+],r10 \n\
140 subq 44,r12 \n\
141 bge 0b \n\
142 movem r10,[r13+] \n\
143 \n\
144 ;; Compensate for last loop underflowing n. \n\
145 addq 44,r12 \n\
146 \n\
147 ;; Restore registers from stack. \n\
148 movem [sp+],r10"
149
150 /* Outputs. */
151 : "=r" (dst), "=r" (src), "=r" (n)
152
153 /* Inputs. */
154 : "0" (dst), "1" (src), "2" (n));
155 }
156
157 while (n >= 16)
158 {
159 *(long *) dst = *(long *) src; dst += 4; src += 4;
160 *(long *) dst = *(long *) src; dst += 4; src += 4;
161 *(long *) dst = *(long *) src; dst += 4; src += 4;
162 *(long *) dst = *(long *) src; dst += 4; src += 4;
163
164 n -= 16;
165 }
166
167 switch (n)
168 {
169 case 0:
170 break;
171
172 case 1:
173 *dst = *src;
174 break;
175
176 case 2:
177 *(short *) dst = *(short *) src;
178 break;
179
180 case 3:
181 *(short *) dst = *(short *) src; dst += 2; src += 2;
182 *dst = *src;
183 break;
184
185 case 4:
186 *(long *) dst = *(long *) src;
187 break;
188
189 case 5:
190 *(long *) dst = *(long *) src; dst += 4; src += 4;
191 *dst = *src;
192 break;
193
194 case 6:
195 *(long *) dst = *(long *) src; dst += 4; src += 4;
196 *(short *) dst = *(short *) src;
197 break;
198
199 case 7:
200 *(long *) dst = *(long *) src; dst += 4; src += 4;
201 *(short *) dst = *(short *) src; dst += 2; src += 2;
202 *dst = *src;
203 break;
204
205 case 8:
206 *(long *) dst = *(long *) src; dst += 4; src += 4;
207 *(long *) dst = *(long *) src;
208 break;
209
210 case 9:
211 *(long *) dst = *(long *) src; dst += 4; src += 4;
212 *(long *) dst = *(long *) src; dst += 4; src += 4;
213 *dst = *src;
214 break;
215
216 case 10:
217 *(long *) dst = *(long *) src; dst += 4; src += 4;
218 *(long *) dst = *(long *) src; dst += 4; src += 4;
219 *(short *) dst = *(short *) src;
220 break;
221
222 case 11:
223 *(long *) dst = *(long *) src; dst += 4; src += 4;
224 *(long *) dst = *(long *) src; dst += 4; src += 4;
225 *(short *) dst = *(short *) src; dst += 2; src += 2;
226 *dst = *src;
227 break;
228
229 case 12:
230 *(long *) dst = *(long *) src; dst += 4; src += 4;
231 *(long *) dst = *(long *) src; dst += 4; src += 4;
232 *(long *) dst = *(long *) src;
233 break;
234
235 case 13:
236 *(long *) dst = *(long *) src; dst += 4; src += 4;
237 *(long *) dst = *(long *) src; dst += 4; src += 4;
238 *(long *) dst = *(long *) src; dst += 4; src += 4;
239 *dst = *src;
240 break;
241
242 case 14:
243 *(long *) dst = *(long *) src; dst += 4; src += 4;
244 *(long *) dst = *(long *) src; dst += 4; src += 4;
245 *(long *) dst = *(long *) src; dst += 4; src += 4;
246 *(short *) dst = *(short *) src;
247 break;
248
249 case 15:
250 *(long *) dst = *(long *) src; dst += 4; src += 4;
251 *(long *) dst = *(long *) src; dst += 4; src += 4;
252 *(long *) dst = *(long *) src; dst += 4; src += 4;
253 *(short *) dst = *(short *) src; dst += 2; src += 2;
254 *dst = *src;
255 break;
256 }
257
258 return return_dst;
259 }
260
