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
|
/*#************************************************************************#*/
/*#-------------------------------------------------------------------------*/
/*# */
/*# FUNCTION NAME: memcpy() */
/*# */
/*# PARAMETERS: void* dst; Destination address. */
/*# void* src; Source address. */
/*# int len; Number of bytes to copy. */
/*# */
/*# RETURNS: dst. */
/*# */
/*# DESCRIPTION: Copies len bytes of memory from src to dst. No guarantees */
/*# about copying of overlapping memory areas. This routine is */
/*# very sensitive to compiler changes in register allocation. */
/*# Should really be rewritten to avoid this problem. */
/*# */
/*#-------------------------------------------------------------------------*/
/*# */
/*# HISTORY */
/*# */
/*# DATE NAME CHANGES */
/*# ---- ---- ------- */
/*# 941007 Kenny R Creation */
/*# 941011 Kenny R Lots of optimizations and inlining. */
/*# 941129 Ulf A Adapted for use in libc. */
/*# 950216 HP N==0 forgotten if non-aligned src/dst. */
/*# Added some optimizations. */
/*# 001025 HP Make src and dst char *. Align dst to */
/*# dword, not just word-if-both-src-and-dst- */
/*# are-misaligned. */
/*# */
/*#-------------------------------------------------------------------------*/
#include <linux/types.h>
void *memcpy(void *pdst,
const void *psrc,
size_t pn)
{
/* Ok. Now we want the parameters put in special registers.
Make sure the compiler is able to make something useful of this.
As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
If gcc was alright, it really would need no temporaries, and no
stack space to save stuff on. */
register void *return_dst __asm__ ("r10") = pdst;
register char *dst __asm__ ("r13") = pdst;
register const char *src __asm__ ("r11") = psrc;
register int n __asm__ ("r12") = pn;
/* When src is aligned but not dst, this makes a few extra needless
cycles. I believe it would take as many to check that the
re-alignment was unnecessary. */
if (((unsigned long) dst & 3) != 0
/* Don't align if we wouldn't copy more than a few bytes; so we
don't have to check further for overflows. */
&& n >= 3)
{
if ((unsigned long) dst & 1)
{
n--;
*(char*)dst = *(char*)src;
src++;
dst++;
}
if ((unsigned long) dst & 2)
{
n -= 2;
*(short*)dst = *(short*)src;
src += 2;
dst += 2;
}
}
/* Decide which copying method to use. */
if (n >= 44*2) /* Break even between movem and
move16 is at 38.7*2, but modulo 44. */
{
/* For large copies we use 'movem' */
/* It is not optimal to tell the compiler about clobbering any
registers; that will move the saving/restoring of those registers
to the function prologue/epilogue, and make non-movem sizes
suboptimal.
This method is not foolproof; it assumes that the "asm reg"
declarations at the beginning of the function really are used
here (beware: they may be moved to temporary registers).
This way, we do not have to save/move the registers around into
temporaries; we can safely use them straight away.
If you want to check that the allocation was right; then
check the equalities in the first comment. It should say
"r13=r13, r11=r11, r12=r12" */
__asm__ volatile ("\n\
;; Check that the following is true (same register names on \n\
;; both sides of equal sign, as in r8=r8): \n\
;; %0=r13, %1=r11, %2=r12 \n\
;; \n\
;; Save the registers we'll use in the movem process \n\
;; on the stack. \n\
subq 11*4,$sp \n\
movem $r10,[$sp] \n\
\n\
;; Now we've got this: \n\
;; r11 - src \n\
;; r13 - dst \n\
;; r12 - n \n\
\n\
;; Update n for the first loop \n\
subq 44,$r12 \n\
0: \n\
movem [$r11+],$r10 \n\
subq 44,$r12 \n\
bge 0b \n\
movem $r10,[$r13+] \n\
\n\
addq 44,$r12 ;; compensate for last loop underflowing n \n\
\n\
;; Restore registers from stack \n\
movem [$sp+],$r10"
/* Outputs */ : "=r" (dst), "=r" (src), "=r" (n)
/* Inputs */ : "0" (dst), "1" (src), "2" (n));
}
/* Either we directly starts copying, using dword copying
in a loop, or we copy as much as possible with 'movem'
and then the last block (<44 bytes) is copied here.
This will work since 'movem' will have updated src,dst,n. */
while ( n >= 16 )
{
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
n -= 16;
}
/* A switch() is definitely the fastest although it takes a LOT of code.
* Particularly if you inline code this.
*/
switch (n)
{
case 0:
break;
case 1:
*(char*)dst = *(char*)src;
break;
case 2:
*(short*)dst = *(short*)src;
break;
case 3:
*((short*)dst)++ = *((short*)src)++;
*(char*)dst = *(char*)src;
break;
case 4:
*((long*)dst)++ = *((long*)src)++;
break;
case 5:
*((long*)dst)++ = *((long*)src)++;
*(char*)dst = *(char*)src;
break;
case 6:
*((long*)dst)++ = *((long*)src)++;
*(short*)dst = *(short*)src;
break;
case 7:
*((long*)dst)++ = *((long*)src)++;
*((short*)dst)++ = *((short*)src)++;
*(char*)dst = *(char*)src;
break;
case 8:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
break;
case 9:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*(char*)dst = *(char*)src;
break;
case 10:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*(short*)dst = *(short*)src;
break;
case 11:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((short*)dst)++ = *((short*)src)++;
*(char*)dst = *(char*)src;
break;
case 12:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
break;
case 13:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*(char*)dst = *(char*)src;
break;
case 14:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*(short*)dst = *(short*)src;
break;
case 15:
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((long*)dst)++ = *((long*)src)++;
*((short*)dst)++ = *((short*)src)++;
*(char*)dst = *(char*)src;
break;
}
return return_dst; /* destination pointer. */
} /* memcpy() */
|
