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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
|
/* Optimized version of the standard memcpy() function.
This file is part of the GNU C Library.
Copyright (C) 1991,92,93,97,98,99 Free Software Foundation, Inc.
Contributed by Dan Pop <Dan.Pop@cern.ch>.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Library General Public License as
published by the Free Software Foundation; either version 2 of the
License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Library General Public License for more details.
You should have received a copy of the GNU Library General Public
License along with the GNU C Library; see the file COPYING.LIB. If not,
write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
Boston, MA 02111-1307, USA. */
/* Return: dest
Inputs:
in0: dest
in1: src
in2: byte count
An assembly implementation of the algorithm used by the generic C
version from glibc. The case when all three arguments are multiples
of 8 is treated separatedly, for extra performance.
In this form, it assumes little endian mode. For big endian mode,
sh1 must be computed using an extra instruction: sub sh1 = 64, sh1
and the order of r[MEMLAT] and r[MEMLAT+1] must be reverted in the
shrp instruction. */
#include <sysdep.h>
#undef ret
#define OP_T_THRES 16
#define OPSIZ 8
#define saved_pfs r14
#define sf r15
#define rescnt r16
#define saved_pr r17
#define saved_lc r18
#define dest r19
#define src r20
#define len r21
#define asrc r22
#define tmp2 r23
#define tmp3 r24
#define tmp4 r25
#define ptable r26
#define ploop56 r27
#define loopaddr r28
#define sh1 r29
#define loopcnt r30
#define value r31
#define dl0 r22
#define dh0 r23
#define dl1 r24
#define dh1 r25
#define dl2 r26
#define dh2 r27
#define dl3 r28
#define dh3 r29
#define LOOP(shift) \
.align 32 ; \
.loop##shift##: \
(p[0]) ld8 r[0] = [asrc], 8 ; /* w1 */ \
(p[MEMLAT+1]) st8 [dest] = value, 8 ; \
(p[MEMLAT]) shrp value = r[MEMLAT], r[MEMLAT+1], shift ; \
nop.b 0 ; \
nop.b 0 ; \
br.ctop.sptk .loop##shift ; \
br.cond.sptk .cpyfew ; /* deal with the remaining bytes */
ENTRY(memcpy)
alloc saved_pfs = ar.pfs, 3, 40-3, 0, 40
#include "softpipe.h"
.rotr r[MEMLAT + 2], q[MEMLAT + 1], s0[2], s1[2], s2[2], s3[2]
.rotf tl0[5], th0[5], tl1[5], th1[5], tl2[5], th2[5], tl3[5], th3[5]
.rotp p[MEMLAT + 2]
mov ret0 = in0 // return value = dest
mov saved_pr = pr // save the predicate registers
brp.loop.many.tk.tk.imp .l0, .done - 16
mov saved_lc = ar.lc // save the loop counter
or tmp3 = in0, in1 ;; // tmp3 = dest | src
or tmp3 = tmp3, in2 // tmp3 = dest | src | len
mov dest = in0 // dest
mov src = in1 // src
mov len = in2 // len
sub tmp2 = r0, in0 // tmp2 = -dest
cmp.eq p6, p0 = in2, r0 // if (len == 0)
(p6) br.cond.spnt .restore_and_exit;;// return dest;
and tmp4 = 7, tmp3 // tmp4 = (dest | src | len) & 7
tbit.nz p8, p0 = src, 3 ;; // test for 16-byte boundary align
cmp.ne p6, p0 = tmp4, r0 // if ((dest | src | len) & 7 != 0)
(p6) br.cond.sptk .next // goto next;
// The optimal case, when dest, src and len are all multiples of 8
(p8) ld8 value = [src], 8 // align src if necessary
(p8) adds len = -8, len ;; // adjust len accordingly
shr.u loopcnt = len, 6 // loopcnt = len / 64
shr.u rescnt = len, 3 // rescnt = len / 8
mov pr.rot = 1 << 16 // set rotating predicates
mov ar.ec = 4 + 1 ;; // set the epilog counter
cmp.eq p6, p0 = loopcnt, r0
and rescnt = 7, rescnt // resnt = residual word count
adds loopcnt = -1, loopcnt // --loopcnt
(p8) st8 [dest] = value, 8 // copy one word if aligning
(p6) br.cond.spnt .epilog;; // there are < 8 words to copy
add sf = 64 * 4, src
mov ar.lc = loopcnt // set the loop counter
mov s0[1] = src
add s1[1] = 16*1, src
add s2[1] = 16*2, src
add s3[1] = 16*3, src
;;
mov dl0 = dest
add dh0 = 8 * 1, dest
add dl1 = 8 * 2, dest
add dh1 = 8 * 3, dest
add dl2 = 8 * 4, dest
add dh2 = 8 * 5, dest
add dl3 = 8 * 6, dest
add dh3 = 8 * 7, dest
;;
.l0:
(p[0]) lfetch.nta [sf], 64
(p[0]) ldfp8 tl0[0], th0[0] = [s0[1]]
(p[0]) ldfp8 tl1[0], th1[0] = [s1[1]]
(p[0]) ldfp8 tl2[0], th2[0] = [s2[1]]
(p[0]) ldfp8 tl3[0], th3[0] = [s3[1]]
(p[0]) add s0[0] = 64, s0[1]
(p[0]) add s1[0] = 64, s1[1]
(p[0]) add s2[0] = 64, s2[1]
(p[0]) add s3[0] = 64, s3[1]
(p[1]) mov src = s0[1] // for the epilog code
(p[4]) stf8 [dl0] = tl0[4], 64
(p[4]) stf8 [dh0] = th0[4], 64
(p[4]) stf8 [dl1] = tl1[4], 64
(p[4]) stf8 [dh1] = th1[4], 64
(p[4]) stf8 [dl2] = tl2[4], 64
(p[4]) stf8 [dh2] = th2[4], 64
(p[4]) stf8 [dl3] = tl3[4], 64
(p[4]) stf8 [dh3] = th3[4], 64
br.ctop.sptk.many .l0
.done:
mov dest = dl0
.epilog:
cmp.eq p6, p0 = rescnt, r0 // are there any words left to copy?
tbit.nz p10, p0 = rescnt, 0
(p6) br.cond.spnt .restore_and_exit ;;
(p10) ld8 r[0] = [src], 8
tbit.nz p11, p0 = rescnt, 1 ;;
(p11) ld8 r[1] = [src], 8
(p10) st8 [dest] = r[0], 8 ;;
(p11) ld8 r[2] = [src], 8
(p11) st8 [dest] = r[1], 8
tbit.nz p12, p0 = rescnt, 2 ;;
(p12) ld8 r[3] = [src], 8
(p11) st8 [dest] = r[2], 8 ;;
(p12) ld8 r[4] = [src], 8
(p12) st8 [dest] = r[3], 8 ;;
(p12) ld8 r[5] = [src], 8
(p12) st8 [dest] = r[4], 8
mov ar.lc = saved_lc ;; // restore the loop counter
(p12) ld8 r[6] = [src], 8
(p12) st8 [dest] = r[5], 8
mov ar.pfs = saved_pfs;; // restore the PFS
(p12) st8 [dest] = r[6]
mov pr = saved_pr, -1 // restore the predicate registers
br.ret.sptk.many b0
.next:
cmp.ge p6, p0 = OP_T_THRES, len // is len <= OP_T_THRES
and loopcnt = 7, tmp2 // loopcnt = -dest % 8
(p6) br.cond.spnt .cpyfew // copy byte by byte
;;
cmp.eq p6, p0 = loopcnt, r0
(p6) br.cond.sptk .dest_aligned
sub len = len, loopcnt // len -= -dest % 8
adds loopcnt = -1, loopcnt // --loopcnt
;;
mov ar.lc = loopcnt
.l1: // copy -dest % 8 bytes
ld1 value = [src], 1 // value = *src++
;;
st1 [dest] = value, 1 // *dest++ = value
br.cloop.dptk .l1
.dest_aligned:
and sh1 = 7, src // sh1 = src % 8
and tmp2 = -8, len // tmp2 = len & -OPSIZ
and asrc = -8, src // asrc = src & -OPSIZ -- align src
shr.u loopcnt = len, 3 // loopcnt = len / 8
and len = 7, len;; // len = len % 8
adds loopcnt = -1, loopcnt // --loopcnt
addl tmp4 = @ltoff(.table), gp
addl tmp3 = @ltoff(.loop56), gp
mov ar.ec = MEMLAT + 1 // set EC
mov pr.rot = 1 << 16;; // set rotating predicates
mov ar.lc = loopcnt // set LC
cmp.eq p6, p0 = sh1, r0 // is the src aligned?
(p6) br.cond.sptk .src_aligned
add src = src, tmp2 // src += len & -OPSIZ
shl sh1 = sh1, 3 // sh1 = 8 * (src % 8)
ld8 ploop56 = [tmp3] // ploop56 = &loop56
ld8 ptable = [tmp4];; // ptable = &table
add tmp3 = ptable, sh1;; // tmp3 = &table + sh1
mov ar.ec = MEMLAT + 1 + 1 // one more pass needed
ld8 tmp4 = [tmp3];; // tmp4 = loop offset
sub loopaddr = ploop56,tmp4 // loopadd = &loop56 - loop offset
ld8 r[1] = [asrc], 8;; // w0
mov b6 = loopaddr;;
br b6 // jump to the appropriate loop
LOOP(8)
LOOP(16)
LOOP(24)
LOOP(32)
LOOP(40)
LOOP(48)
LOOP(56)
.src_aligned:
.l3:
(p[0]) ld8 r[0] = [src], 8
(p[MEMLAT]) st8 [dest] = r[MEMLAT], 8
br.ctop.dptk .l3
.cpyfew:
cmp.eq p6, p0 = len, r0 // is len == 0 ?
adds len = -1, len // --len;
(p6) br.cond.spnt .restore_and_exit ;;
mov ar.lc = len
.l4:
ld1 value = [src], 1
;;
st1 [dest] = value, 1
br.cloop.dptk .l4 ;;
.restore_and_exit:
mov ar.pfs = saved_pfs // restore the PFS
mov pr = saved_pr, -1 // restore the predicate registers
mov ar.lc = saved_lc // restore the loop counter
br.ret.sptk.many b0
.align 8
.table:
data8 0 // dummy entry
data8 .loop56 - .loop8
data8 .loop56 - .loop16
data8 .loop56 - .loop24
data8 .loop56 - .loop32
data8 .loop56 - .loop40
data8 .loop56 - .loop48
data8 .loop56 - .loop56
END(memcpy)
|
