1 /* A memcpy for CRIS.
2    Copyright (C) 1994-2005 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 #include <picolibc.h>
34 
35 #include <stddef.h>
36 #include "../../string/local.h"
37 
38 /* Break even between movem and move16 is really at 38.7 * 2, but
39    modulo 44, so up to the next multiple of 44, we use ordinary code.  */
40 #define MEMCPY_BY_BLOCK_THRESHOLD (44 * 2)
41 
42 /* No name ambiguities in this file.  */
43 __asm__ (".syntax no_register_prefix");
44 
45 void *
46 __inhibit_loop_to_libcall
memcpy(void * __restrict pdst,const void * __restrict psrc,size_t pn)47 memcpy(void *__restrict pdst, const void *__restrict psrc, size_t pn)
48 {
49   /* Now we want the parameters put in special registers.
50      Make sure the compiler is able to make something useful of this.
51      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
52 
53      If gcc was allright, it really would need no temporaries, and no
54      stack space to save stuff on.  */
55 
56   register void *return_dst __asm__ ("r10") = pdst;
57   register unsigned char *dst __asm__ ("r13") = pdst;
58   register unsigned const char *src __asm__ ("r11") = psrc;
59   register int n __asm__ ("r12") = pn;
60 
61   /* When src is aligned but not dst, this makes a few extra needless
62      cycles.  I believe it would take as many to check that the
63      re-alignment was unnecessary.  */
64   if (((unsigned long) dst & 3) != 0
65       /* Don't align if we wouldn't copy more than a few bytes; so we
66 	 don't have to check further for overflows.  */
67       && n >= 3)
68   {
69     if ((unsigned long) dst & 1)
70       {
71 	n--;
72 	*dst = *src;
73 	src++;
74 	dst++;
75       }
76 
77     if ((unsigned long) dst & 2)
78       {
79 	n -= 2;
80 	*(short *) dst = *(short *) src;
81 	src += 2;
82 	dst += 2;
83       }
84   }
85 
86   /* Decide which copying method to use.  */
87   if (n >= MEMCPY_BY_BLOCK_THRESHOLD)
88     {
89       /* It is not optimal to tell the compiler about clobbering any
90 	 registers; that will move the saving/restoring of those registers
91 	 to the function prologue/epilogue, and make non-movem sizes
92 	 suboptimal.  */
93       __asm__ volatile
94 	("\
95 	 ;; GCC does promise correct register allocations, but let's	\n\
96 	 ;; make sure it keeps its promises.				\n\
97 	 .ifnc %0-%1-%2,$r13-$r11-$r12					\n\
98 	 .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\"	\n\
99 	 .endif								\n\
100 									\n\
101 	 ;; Save the registers we'll use in the movem process		\n\
102 	 ;; on the stack.						\n\
103 	 subq	11*4,sp							\n\
104 	 movem	r10,[sp]						\n\
105 									\n\
106 	 ;; Now we've got this:						\n\
107 	 ;; r11 - src							\n\
108 	 ;; r13 - dst							\n\
109 	 ;; r12 - n							\n\
110 									\n\
111 	 ;; Update n for the first loop.				\n\
112 	 subq	 44,r12							\n\
113 0:									\n\
114 "
115 #ifdef __arch_common_v10_v32
116 	 /* Cater to branch offset difference between v32 and v10.  We
117 	    assume the branch below has an 8-bit offset.  */
118 "	 setf\n"
119 #endif
120 "	 movem	[r11+],r10						\n\
121 	 subq	44,r12							\n\
122 	 bge	 0b							\n\
123 	 movem	r10,[r13+]						\n\
124 									\n\
125 	 ;; Compensate for last loop underflowing n.			\n\
126 	 addq	44,r12							\n\
127 									\n\
128 	 ;; Restore registers from stack.				\n\
129 	 movem [sp+],r10"
130 
131 	 /* Outputs.  */
132 	 : "=r" (dst), "=r" (src), "=r" (n)
133 
134 	 /* Inputs.  */
135 	 : "0" (dst), "1" (src), "2" (n));
136     }
137 
138   while (n >= 16)
139     {
140       *(long *) dst = *(long *) src; dst += 4; src += 4;
141       *(long *) dst = *(long *) src; dst += 4; src += 4;
142       *(long *) dst = *(long *) src; dst += 4; src += 4;
143       *(long *) dst = *(long *) src; dst += 4; src += 4;
144 
145       n -= 16;
146     }
147 
148   switch (n)
149     {
150     case 0:
151       break;
152 
153     case 1:
154       *dst = *src;
155       break;
156 
157     case 2:
158       *(short *) dst = *(short *) src;
159       break;
160 
161     case 3:
162       *(short *) dst = *(short *) src; dst += 2; src += 2;
163       *dst = *src;
164       break;
165 
166     case 4:
167       *(long *) dst = *(long *) src;
168       break;
169 
170     case 5:
171       *(long *) dst = *(long *) src; dst += 4; src += 4;
172       *dst = *src;
173       break;
174 
175     case 6:
176       *(long *) dst = *(long *) src; dst += 4; src += 4;
177       *(short *) dst = *(short *) src;
178       break;
179 
180     case 7:
181       *(long *) dst = *(long *) src; dst += 4; src += 4;
182       *(short *) dst = *(short *) src; dst += 2; src += 2;
183       *dst = *src;
184       break;
185 
186     case 8:
187       *(long *) dst = *(long *) src; dst += 4; src += 4;
188       *(long *) dst = *(long *) src;
189       break;
190 
191     case 9:
192       *(long *) dst = *(long *) src; dst += 4; src += 4;
193       *(long *) dst = *(long *) src; dst += 4; src += 4;
194       *dst = *src;
195       break;
196 
197     case 10:
198       *(long *) dst = *(long *) src; dst += 4; src += 4;
199       *(long *) dst = *(long *) src; dst += 4; src += 4;
200       *(short *) dst = *(short *) src;
201       break;
202 
203     case 11:
204       *(long *) dst = *(long *) src; dst += 4; src += 4;
205       *(long *) dst = *(long *) src; dst += 4; src += 4;
206       *(short *) dst = *(short *) src; dst += 2; src += 2;
207       *dst = *src;
208       break;
209 
210     case 12:
211       *(long *) dst = *(long *) src; dst += 4; src += 4;
212       *(long *) dst = *(long *) src; dst += 4; src += 4;
213       *(long *) dst = *(long *) src;
214       break;
215 
216     case 13:
217       *(long *) dst = *(long *) src; dst += 4; src += 4;
218       *(long *) dst = *(long *) src; dst += 4; src += 4;
219       *(long *) dst = *(long *) src; dst += 4; src += 4;
220       *dst = *src;
221       break;
222 
223     case 14:
224       *(long *) dst = *(long *) src; dst += 4; src += 4;
225       *(long *) dst = *(long *) src; dst += 4; src += 4;
226       *(long *) dst = *(long *) src; dst += 4; src += 4;
227       *(short *) dst = *(short *) src;
228       break;
229 
230     case 15:
231       *(long *) dst = *(long *) src; dst += 4; src += 4;
232       *(long *) dst = *(long *) src; dst += 4; src += 4;
233       *(long *) dst = *(long *) src; dst += 4; src += 4;
234       *(short *) dst = *(short *) src; dst += 2; src += 2;
235       *dst = *src;
236       break;
237     }
238 
239   return return_dst;
240 }
241