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
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
|
/* Implementation for strrchr using evex256 and evex512.
Copyright (C) 2022-2023 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 <isa-level.h>
#if ISA_SHOULD_BUILD (4)
# include <sysdep.h>
# ifdef USE_AS_WCSRCHR
# if VEC_SIZE == 64
# define RCX_M cx
# define KORTEST_M kortestw
# else
# define RCX_M cl
# define KORTEST_M kortestb
# endif
# define SHIFT_REG VRCX
# define CHAR_SIZE 4
# define VPCMP vpcmpd
# define VPMIN vpminud
# define VPCOMPRESS vpcompressd
# define VPTESTN vptestnmd
# define VPTEST vptestmd
# define VPBROADCAST vpbroadcastd
# define VPCMPEQ vpcmpeqd
# else
# define SHIFT_REG VRDI
# define CHAR_SIZE 1
# define VPCMP vpcmpb
# define VPMIN vpminub
# define VPCOMPRESS vpcompressb
# define VPTESTN vptestnmb
# define VPTEST vptestmb
# define VPBROADCAST vpbroadcastb
# define VPCMPEQ vpcmpeqb
# define RCX_M VRCX
# define KORTEST_M KORTEST
# endif
# define VMATCH VMM(0)
# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
# define PAGE_SIZE 4096
.section SECTION(.text), "ax", @progbits
/* Aligning entry point to 64 byte, provides better performance for
one vector length string. */
ENTRY_P2ALIGN(STRRCHR, 6)
movl %edi, %eax
/* Broadcast CHAR to VMATCH. */
VPBROADCAST %esi, %VMATCH
andl $(PAGE_SIZE - 1), %eax
cmpl $(PAGE_SIZE - VEC_SIZE), %eax
jg L(cross_page_boundary)
VMOVU (%rdi), %VMM(1)
/* k0 has a 1 for each zero CHAR in YMM1. */
VPTESTN %VMM(1), %VMM(1), %k0
KMOV %k0, %VGPR(rsi)
test %VGPR(rsi), %VGPR(rsi)
jz L(aligned_more)
/* fallthrough: zero CHAR in first VEC. */
L(page_cross_return):
/* K1 has a 1 for each search CHAR match in VEC(1). */
VPCMPEQ %VMATCH, %VMM(1), %k1
KMOV %k1, %VGPR(rax)
/* Build mask up until first zero CHAR (used to mask of
potential search CHAR matches past the end of the string). */
blsmsk %VGPR(rsi), %VGPR(rsi)
/* Use `and` here to remove any out of bounds matches so we can
do a reverse scan on `rax` to find the last match. */
and %VGPR(rsi), %VGPR(rax)
jz L(ret0)
/* Get last match. */
bsr %VGPR(rax), %VGPR(rax)
# ifdef USE_AS_WCSRCHR
leaq (%rdi, %rax, CHAR_SIZE), %rax
# else
addq %rdi, %rax
# endif
L(ret0):
ret
/* Returns for first vec x1/x2/x3 have hard coded backward
search path for earlier matches. */
.p2align 4,, 6
L(first_vec_x1):
VPCMPEQ %VMATCH, %VMM(2), %k1
KMOV %k1, %VGPR(rax)
blsmsk %VGPR(rcx), %VGPR(rcx)
/* eax non-zero if search CHAR in range. */
and %VGPR(rcx), %VGPR(rax)
jnz L(first_vec_x1_return)
/* fallthrough: no match in YMM2 then need to check for earlier
matches (in YMM1). */
.p2align 4,, 4
L(first_vec_x0_test):
VPCMPEQ %VMATCH, %VMM(1), %k1
KMOV %k1, %VGPR(rax)
test %VGPR(rax), %VGPR(rax)
jz L(ret1)
bsr %VGPR(rax), %VGPR(rax)
# ifdef USE_AS_WCSRCHR
leaq (%rsi, %rax, CHAR_SIZE), %rax
# else
addq %rsi, %rax
# endif
L(ret1):
ret
.p2align 4,, 10
L(first_vec_x3):
VPCMPEQ %VMATCH, %VMM(4), %k1
KMOV %k1, %VGPR(rax)
blsmsk %VGPR(rcx), %VGPR(rcx)
/* If no search CHAR match in range check YMM1/YMM2/YMM3. */
and %VGPR(rcx), %VGPR(rax)
jz L(first_vec_x1_or_x2)
bsr %VGPR(rax), %VGPR(rax)
leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 4
L(first_vec_x2):
VPCMPEQ %VMATCH, %VMM(3), %k1
KMOV %k1, %VGPR(rax)
blsmsk %VGPR(rcx), %VGPR(rcx)
/* Check YMM3 for last match first. If no match try YMM2/YMM1. */
and %VGPR(rcx), %VGPR(rax)
jz L(first_vec_x0_x1_test)
bsr %VGPR(rax), %VGPR(rax)
leaq (VEC_SIZE * 2)(%r8, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 6
L(first_vec_x0_x1_test):
VPCMPEQ %VMATCH, %VMM(2), %k1
KMOV %k1, %VGPR(rax)
/* Check YMM2 for last match first. If no match try YMM1. */
test %VGPR(rax), %VGPR(rax)
jz L(first_vec_x0_test)
.p2align 4,, 4
L(first_vec_x1_return):
bsr %VGPR(rax), %VGPR(rax)
leaq (VEC_SIZE)(%r8, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 12
L(aligned_more):
L(page_cross_continue):
/* Need to keep original pointer incase VEC(1) has last match. */
movq %rdi, %rsi
andq $-VEC_SIZE, %rdi
VMOVU VEC_SIZE(%rdi), %VMM(2)
VPTESTN %VMM(2), %VMM(2), %k0
KMOV %k0, %VRCX
movq %rdi, %r8
test %VRCX, %VRCX
jnz L(first_vec_x1)
VMOVU (VEC_SIZE * 2)(%rdi), %VMM(3)
VPTESTN %VMM(3), %VMM(3), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jnz L(first_vec_x2)
VMOVU (VEC_SIZE * 3)(%rdi), %VMM(4)
VPTESTN %VMM(4), %VMM(4), %k0
KMOV %k0, %VRCX
/* Intentionally use 64-bit here. EVEX256 version needs 1-byte
padding for efficient nop before loop alignment. */
test %rcx, %rcx
jnz L(first_vec_x3)
andq $-(VEC_SIZE * 2), %rdi
.p2align 4
L(first_aligned_loop):
/* Preserve VEC(1), VEC(2), VEC(3), and VEC(4) until we can
gurantee they don't store a match. */
VMOVA (VEC_SIZE * 4)(%rdi), %VMM(5)
VMOVA (VEC_SIZE * 5)(%rdi), %VMM(6)
VPCMP $4, %VMM(5), %VMATCH, %k2
VPCMP $4, %VMM(6), %VMATCH, %k3{%k2}
VPMIN %VMM(5), %VMM(6), %VMM(7)
VPTEST %VMM(7), %VMM(7), %k1{%k3}
subq $(VEC_SIZE * -2), %rdi
KORTEST_M %k1, %k1
jc L(first_aligned_loop)
VPTESTN %VMM(7), %VMM(7), %k1
KMOV %k1, %VRDX
test %VRDX, %VRDX
jz L(second_aligned_loop_prep)
KORTEST_M %k3, %k3
jnc L(return_first_aligned_loop)
.p2align 4,, 6
L(first_vec_x1_or_x2_or_x3):
VPCMPEQ %VMM(4), %VMATCH, %k4
KMOV %k4, %VRAX
test %VRAX, %VRAX
jz L(first_vec_x1_or_x2)
bsr %VRAX, %VRAX
leaq (VEC_SIZE * 3)(%r8, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 8
L(return_first_aligned_loop):
VPTESTN %VMM(5), %VMM(5), %k0
KMOV %k0, %VRCX
blsmsk %VRCX, %VRCX
jnc L(return_first_new_match_first)
blsmsk %VRDX, %VRDX
VPCMPEQ %VMM(6), %VMATCH, %k0
KMOV %k0, %VRAX
addq $VEC_SIZE, %rdi
and %VRDX, %VRAX
jnz L(return_first_new_match_ret)
subq $VEC_SIZE, %rdi
L(return_first_new_match_first):
KMOV %k2, %VRAX
# ifdef USE_AS_WCSRCHR
xorl $((1 << CHAR_PER_VEC)- 1), %VRAX
and %VRCX, %VRAX
# else
andn %VRCX, %VRAX, %VRAX
# endif
jz L(first_vec_x1_or_x2_or_x3)
L(return_first_new_match_ret):
bsr %VRAX, %VRAX
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 10
L(first_vec_x1_or_x2):
VPCMPEQ %VMM(3), %VMATCH, %k3
KMOV %k3, %VRAX
test %VRAX, %VRAX
jz L(first_vec_x0_x1_test)
bsr %VRAX, %VRAX
leaq (VEC_SIZE * 2)(%r8, %rax, CHAR_SIZE), %rax
ret
.p2align 4
/* We can throw away the work done for the first 4x checks here
as we have a later match. This is the 'fast' path persay. */
L(second_aligned_loop_prep):
L(second_aligned_loop_set_furthest_match):
movq %rdi, %rsi
VMOVA %VMM(5), %VMM(7)
VMOVA %VMM(6), %VMM(8)
.p2align 4
L(second_aligned_loop):
VMOVU (VEC_SIZE * 4)(%rdi), %VMM(5)
VMOVU (VEC_SIZE * 5)(%rdi), %VMM(6)
VPCMP $4, %VMM(5), %VMATCH, %k2
VPCMP $4, %VMM(6), %VMATCH, %k3{%k2}
VPMIN %VMM(5), %VMM(6), %VMM(4)
VPTEST %VMM(4), %VMM(4), %k1{%k3}
subq $(VEC_SIZE * -2), %rdi
KMOV %k1, %VRCX
inc %RCX_M
jz L(second_aligned_loop)
VPTESTN %VMM(4), %VMM(4), %k1
KMOV %k1, %VRDX
test %VRDX, %VRDX
jz L(second_aligned_loop_set_furthest_match)
KORTEST_M %k3, %k3
jnc L(return_new_match)
/* branch here because there is a significant advantage interms
of output dependency chance in using edx. */
L(return_old_match):
VPCMPEQ %VMM(8), %VMATCH, %k0
KMOV %k0, %VRCX
bsr %VRCX, %VRCX
jnz L(return_old_match_ret)
VPCMPEQ %VMM(7), %VMATCH, %k0
KMOV %k0, %VRCX
bsr %VRCX, %VRCX
subq $VEC_SIZE, %rsi
L(return_old_match_ret):
leaq (VEC_SIZE * 3)(%rsi, %rcx, CHAR_SIZE), %rax
ret
L(return_new_match):
VPTESTN %VMM(5), %VMM(5), %k0
KMOV %k0, %VRCX
blsmsk %VRCX, %VRCX
jnc L(return_new_match_first)
dec %VRDX
VPCMPEQ %VMM(6), %VMATCH, %k0
KMOV %k0, %VRAX
addq $VEC_SIZE, %rdi
and %VRDX, %VRAX
jnz L(return_new_match_ret)
subq $VEC_SIZE, %rdi
L(return_new_match_first):
KMOV %k2, %VRAX
# ifdef USE_AS_WCSRCHR
xorl $((1 << CHAR_PER_VEC)- 1), %VRAX
and %VRCX, %VRAX
# else
andn %VRCX, %VRAX, %VRAX
# endif
jz L(return_old_match)
L(return_new_match_ret):
bsr %VRAX, %VRAX
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 4,, 4
L(cross_page_boundary):
xorq %rdi, %rax
mov $-1, %VRDX
VMOVU (PAGE_SIZE - VEC_SIZE)(%rax), %VMM(6)
VPTESTN %VMM(6), %VMM(6), %k0
KMOV %k0, %VRSI
# ifdef USE_AS_WCSRCHR
movl %edi, %ecx
and $(VEC_SIZE - 1), %ecx
shrl $2, %ecx
# endif
shlx %SHIFT_REG, %VRDX, %VRDX
# ifdef USE_AS_WCSRCHR
kmovw %edx, %k1
# else
KMOV %VRDX, %k1
# endif
VPCOMPRESS %VMM(6), %VMM(1){%k1}{z}
/* We could technically just jmp back after the vpcompress but
it doesn't save any 16-byte blocks. */
shrx %SHIFT_REG, %VRSI, %VRSI
test %VRSI, %VRSI
jnz L(page_cross_return)
jmp L(page_cross_continue)
/* 1-byte from cache line. */
END(STRRCHR)
#endif
|