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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
|
/* Optimized memchr implementation for POWER10 LE.
Copyright (C) 2021-2024 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 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
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
# ifndef MEMCHR
# define MEMCHR __memchr
# endif
# define M_VREG_ZERO v20
# define M_OFF_START_LOOP 256
# define MEMCHR_SUBTRACT_VECTORS \
vsububm v4,v4,v18; \
vsububm v5,v5,v18; \
vsububm v6,v6,v18; \
vsububm v7,v7,v18;
# define M_TAIL(vreg,increment) \
vctzlsbb r4,vreg; \
cmpld r5,r4; \
ble L(null); \
addi r4,r4,increment; \
add r3,r6,r4; \
blr
/* TODO: Replace macros by the actual instructions when minimum binutils becomes
>= 2.35. This is used to keep compatibility with older versions. */
#define M_VEXTRACTBM(rt,vrb) \
.long(((4)<<(32-6)) \
| ((rt)<<(32-11)) \
| ((8)<<(32-16)) \
| ((vrb)<<(32-21)) \
| 1602)
#define M_LXVP(xtp,dq,ra) \
.long(((6)<<(32-6)) \
| ((((xtp)-32)>>1)<<(32-10)) \
| ((1)<<(32-11)) \
| ((ra)<<(32-16)) \
| dq)
#define CHECK16B(vreg,offset,addr,label) \
lxv vreg+32,offset(addr); \
vcmpequb. vreg,vreg,v18; \
bne cr6,L(label); \
cmpldi r5,16; \
ble L(null); \
addi r5,r5,-16;
/* Load 4 quadwords, merge into one VR for speed and check for NULLs. r6 has #
of bytes already checked. */
#define CHECK64B(offset,addr,label) \
M_LXVP(v4+32,offset,addr); \
M_LXVP(v6+32,offset+32,addr); \
MEMCHR_SUBTRACT_VECTORS; \
vminub v14,v4,v5; \
vminub v15,v6,v7; \
vminub v16,v14,v15; \
vcmpequb. v0,v16,M_VREG_ZERO; \
beq cr6,$+12; \
li r7,offset; \
b L(label); \
cmpldi r5,64; \
ble L(null); \
addi r5,r5,-64
/* Implements the function
void *[r3] memchr (const void *s [r3], int c [r4], size_t n [r5]). */
.machine power9
ENTRY_TOCLESS (MEMCHR)
CALL_MCOUNT 3
cmpldi r5,0
beq L(null)
mr r0,r5
xori r6,r4,0xff
mtvsrd v18+32,r4 /* matching char in v18 */
mtvsrd v19+32,r6 /* non matching char in v19 */
vspltb v18,v18,7 /* replicate */
vspltb v19,v19,7 /* replicate */
vspltisb M_VREG_ZERO,0
/* Next 16B-aligned address. Prepare address for L(aligned). */
addi r6,r3,16
clrrdi r6,r6,4
/* Align data and fill bytes not loaded with non matching char. */
lvx v0,0,r3
lvsr v1,0,r3
vperm v0,v19,v0,v1
vcmpequb. v6,v0,v18
bne cr6,L(found)
sub r4,r6,r3
cmpld r5,r4
ble L(null)
sub r5,r5,r4
/* Test up to OFF_START_LOOP-16 bytes in 16B chunks. The main loop is
optimized for longer strings, so checking the first bytes in 16B
chunks benefits a lot small strings. */
.p2align 5
L(aligned):
cmpldi r5,0
beq L(null)
CHECK16B(v0,0,r6,tail1)
CHECK16B(v1,16,r6,tail2)
CHECK16B(v2,32,r6,tail3)
CHECK16B(v3,48,r6,tail4)
CHECK16B(v4,64,r6,tail5)
CHECK16B(v5,80,r6,tail6)
CHECK16B(v6,96,r6,tail7)
CHECK16B(v7,112,r6,tail8)
CHECK16B(v8,128,r6,tail9)
CHECK16B(v9,144,r6,tail10)
CHECK16B(v10,160,r6,tail11)
CHECK16B(v0,176,r6,tail12)
CHECK16B(v1,192,r6,tail13)
CHECK16B(v2,208,r6,tail14)
CHECK16B(v3,224,r6,tail15)
cmpdi cr5,r4,0 /* Check if c == 0. This will be useful to
choose how we will perform the main loop. */
/* Prepare address for the loop. */
addi r4,r3,M_OFF_START_LOOP
clrrdi r4,r4,6
sub r6,r4,r3
sub r5,r0,r6
addi r6,r4,128
/* If c == 0, use the loop without the vsububm. */
beq cr5,L(loop)
/* This is very similar to the block after L(loop), the difference is
that here MEMCHR_SUBTRACT_VECTORS is not empty, and we subtract
each byte loaded by the char we are looking for, this way we can keep
using vminub to merge the results and checking for nulls. */
.p2align 5
L(memchr_loop):
CHECK64B(0,r4,pre_tail_64b)
CHECK64B(64,r4,pre_tail_64b)
addi r4,r4,256
CHECK64B(0,r6,tail_64b)
CHECK64B(64,r6,tail_64b)
addi r6,r6,256
CHECK64B(0,r4,pre_tail_64b)
CHECK64B(64,r4,pre_tail_64b)
addi r4,r4,256
CHECK64B(0,r6,tail_64b)
CHECK64B(64,r6,tail_64b)
addi r6,r6,256
b L(memchr_loop)
/* Switch to a more aggressive approach checking 64B each time. Use 2
pointers 128B apart and unroll the loop once to make the pointer
updates and usages separated enough to avoid stalls waiting for
address calculation. */
.p2align 5
L(loop):
#undef MEMCHR_SUBTRACT_VECTORS
#define MEMCHR_SUBTRACT_VECTORS /* nothing */
CHECK64B(0,r4,pre_tail_64b)
CHECK64B(64,r4,pre_tail_64b)
addi r4,r4,256
CHECK64B(0,r6,tail_64b)
CHECK64B(64,r6,tail_64b)
addi r6,r6,256
CHECK64B(0,r4,pre_tail_64b)
CHECK64B(64,r4,pre_tail_64b)
addi r4,r4,256
CHECK64B(0,r6,tail_64b)
CHECK64B(64,r6,tail_64b)
addi r6,r6,256
b L(loop)
.p2align 5
L(pre_tail_64b):
mr r6,r4
L(tail_64b):
/* OK, we found a null byte. Let's look for it in the current 64-byte
block and mark it in its corresponding VR. lxvp vx,0(ry) puts the
low 16B bytes into vx+1, and the high into vx, so the order here is
v5, v4, v7, v6. */
vcmpequb v1,v5,M_VREG_ZERO
vcmpequb v2,v4,M_VREG_ZERO
vcmpequb v3,v7,M_VREG_ZERO
vcmpequb v4,v6,M_VREG_ZERO
/* Take into account the other 64B blocks we had already checked. */
add r6,r6,r7
/* Extract first bit of each byte. */
M_VEXTRACTBM(r8,v1)
M_VEXTRACTBM(r9,v2)
M_VEXTRACTBM(r10,v3)
M_VEXTRACTBM(r11,v4)
/* Shift each value into their corresponding position. */
sldi r9,r9,16
sldi r10,r10,32
sldi r11,r11,48
/* Merge the results. */
or r8,r8,r9
or r9,r10,r11
or r11,r9,r8
cnttzd r0,r11 /* Count trailing zeros before the match. */
cmpld r5,r0
ble L(null)
add r3,r6,r0 /* Compute final address. */
blr
.p2align 5
L(tail1):
M_TAIL(v0,0)
.p2align 5
L(tail2):
M_TAIL(v1,16)
.p2align 5
L(tail3):
M_TAIL(v2,32)
.p2align 5
L(tail4):
M_TAIL(v3,48)
.p2align 5
L(tail5):
M_TAIL(v4,64)
.p2align 5
L(tail6):
M_TAIL(v5,80)
.p2align 5
L(tail7):
M_TAIL(v6,96)
.p2align 5
L(tail8):
M_TAIL(v7,112)
.p2align 5
L(tail9):
M_TAIL(v8,128)
.p2align 5
L(tail10):
M_TAIL(v9,144)
.p2align 5
L(tail11):
M_TAIL(v10,160)
.p2align 5
L(tail12):
M_TAIL(v0,176)
.p2align 5
L(tail13):
M_TAIL(v1,192)
.p2align 5
L(tail14):
M_TAIL(v2,208)
.p2align 5
L(tail15):
M_TAIL(v3,224)
.p2align 5
L(found):
vctzlsbb r7,v6
cmpld r5,r7
ble L(null)
add r3,r3,r7
blr
.p2align 5
L(null):
li r3,0
blr
END (MEMCHR)
weak_alias (__memchr, memchr)
libc_hidden_builtin_def (memchr)
|
