summary refs log tree commit diff
path: root/sysdeps/x86_64/multiarch/strrchr-evex.S
blob: 26b34578751f1d8e2dc499b8a14a08fdb6a340c8 (plain) (blame)
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
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
/* strrchr/wcsrchr optimized with 256-bit EVEX instructions.
   Copyright (C) 2021-2022 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>

# ifndef STRRCHR
#  define STRRCHR	__strrchr_evex
# endif

# include "x86-evex256-vecs.h"

# ifdef USE_AS_WCSRCHR
#  define SHIFT_REG	rsi
#  define kunpck_2x	kunpckbw
#  define kmov_2x	kmovd
#  define maskz_2x	ecx
#  define maskm_2x	eax
#  define CHAR_SIZE	4
#  define VPMIN	vpminud
#  define VPTESTN	vptestnmd
#  define VPTEST	vptestmd
#  define VPBROADCAST	vpbroadcastd
#  define VPCMPEQ	vpcmpeqd
#  define VPCMP	vpcmpd

#  define USE_WIDE_CHAR
# else
#  define SHIFT_REG	rdi
#  define kunpck_2x	kunpckdq
#  define kmov_2x	kmovq
#  define maskz_2x	rcx
#  define maskm_2x	rax

#  define CHAR_SIZE	1
#  define VPMIN	vpminub
#  define VPTESTN	vptestnmb
#  define VPTEST	vptestmb
#  define VPBROADCAST	vpbroadcastb
#  define VPCMPEQ	vpcmpeqb
#  define VPCMP	vpcmpb
# endif

# include "reg-macros.h"

# define VMATCH	VMM(0)
# define CHAR_PER_VEC	(VEC_SIZE / CHAR_SIZE)
# define PAGE_SIZE	4096

	.section SECTION(.text), "ax", @progbits
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)
L(page_cross_continue):
	VMOVU	(%rdi), %VMM(1)
	/* k0 has a 1 for each zero CHAR in VEC(1).  */
	VPTESTN	%VMM(1), %VMM(1), %k0
	KMOV	%k0, %VRSI
	test	%VRSI, %VRSI
	jz	L(aligned_more)
	/* fallthrough: zero CHAR in first VEC.  */
	/* K1 has a 1 for each search CHAR match in VEC(1).  */
	VPCMPEQ	%VMATCH, %VMM(1), %k1
	KMOV	%k1, %VRAX
	/* Build mask up until first zero CHAR (used to mask of
	   potential search CHAR matches past the end of the string).
	 */
	blsmsk	%VRSI, %VRSI
	and	%VRSI, %VRAX
	jz	L(ret0)
	/* Get last match (the `and` removed any out of bounds matches).
	 */
	bsr	%VRAX, %VRAX
# 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, %VRAX
	blsmsk	%VRCX, %VRCX
	/* eax non-zero if search CHAR in range.  */
	and	%VRCX, %VRAX
	jnz	L(first_vec_x1_return)

	/* fallthrough: no match in VEC(2) then need to check for
	   earlier matches (in VEC(1)).  */
	.p2align 4,, 4
L(first_vec_x0_test):
	VPCMPEQ	%VMATCH, %VMM(1), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jz	L(ret1)
	bsr	%VRAX, %VRAX
# ifdef USE_AS_WCSRCHR
	leaq	(%rsi, %rax, CHAR_SIZE), %rax
# else
	addq	%rsi, %rax
# endif
L(ret1):
	ret

	.p2align 4,, 10
L(first_vec_x1_or_x2):
	VPCMPEQ	%VMM(3), %VMATCH, %k3
	VPCMPEQ	%VMM(2), %VMATCH, %k2
	/* K2 and K3 have 1 for any search CHAR match. Test if any
	   matches between either of them. Otherwise check VEC(1).  */
	KORTEST %k2, %k3
	jz	L(first_vec_x0_test)

	/* Guranteed that VEC(2) and VEC(3) are within range so merge
	   the two bitmasks then get last result.  */
	kunpck_2x %k2, %k3, %k3
	kmov_2x	%k3, %maskm_2x
	bsr	%maskm_2x, %maskm_2x
	leaq	(VEC_SIZE * 1)(%r8, %rax, CHAR_SIZE), %rax
	ret

	.p2align 4,, 7
L(first_vec_x3):
	VPCMPEQ	%VMATCH, %VMM(4), %k1
	KMOV	%k1, %VRAX
	blsmsk	%VRCX, %VRCX
	/* If no search CHAR match in range check VEC(1)/VEC(2)/VEC(3).
	 */
	and	%VRCX, %VRAX
	jz	L(first_vec_x1_or_x2)
	bsr	%VRAX, %VRAX
	leaq	(VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
	ret


	.p2align 4,, 6
L(first_vec_x0_x1_test):
	VPCMPEQ	%VMATCH, %VMM(2), %k1
	KMOV	%k1, %VRAX
	/* Check VEC(2) for last match first. If no match try VEC(1).
	 */
	test	%VRAX, %VRAX
	jz	L(first_vec_x0_test)
	.p2align 4,, 4
L(first_vec_x1_return):
	bsr	%VRAX, %VRAX
	leaq	(VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %rax
	ret


	.p2align 4,, 10
L(first_vec_x2):
	VPCMPEQ	%VMATCH, %VMM(3), %k1
	KMOV	%k1, %VRAX
	blsmsk	%VRCX, %VRCX
	/* Check VEC(3) for last match first. If no match try
	   VEC(2)/VEC(1).  */
	and	%VRCX, %VRAX
	jz	L(first_vec_x0_x1_test)
	bsr	%VRAX, %VRAX
	leaq	(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
	ret


	.p2align 4,, 12
L(aligned_more):
	/* 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

	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
	movq	%rdi, %r8
	test	%VRCX, %VRCX
	jnz	L(first_vec_x3)

	andq	$-(VEC_SIZE * 2), %rdi
	.p2align 4,, 10
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)

	VPCMPEQ	%VMM(5), %VMATCH, %k2
	vpxord	%VMM(6), %VMATCH, %VMM(7)

	VPMIN	%VMM(5), %VMM(6), %VMM(8)
	VPMIN	%VMM(8), %VMM(7), %VMM(7)

	VPTESTN	%VMM(7), %VMM(7), %k1
	subq	$(VEC_SIZE * -2), %rdi
	KORTEST %k1, %k2
	jz	L(first_aligned_loop)

	VPCMPEQ	%VMM(6), %VMATCH, %k3
	VPTESTN	%VMM(8), %VMM(8), %k1

	/* If k1 is zero, then we found a CHAR match but no null-term.
	   We can now safely throw out VEC1-4.  */
	KTEST	%k1, %k1
	jz	L(second_aligned_loop_prep)

	KORTEST %k2, %k3
	jnz	L(return_first_aligned_loop)


	.p2align 4,, 6
L(first_vec_x1_or_x2_or_x3):
	VPCMPEQ	%VMM(4), %VMATCH, %k4
	KMOV	%k4, %VRAX
	bsr	%VRAX, %VRAX
	jz	L(first_vec_x1_or_x2)
	leaq	(VEC_SIZE * 3)(%r8, %rax, CHAR_SIZE), %rax
	ret


	.p2align 4,, 8
L(return_first_aligned_loop):
	VPTESTN	%VMM(5), %VMM(5), %k0

	/* Combined results from VEC5/6.  */
	kunpck_2x %k0, %k1, %k0
	kmov_2x	%k0, %maskz_2x

	blsmsk	%maskz_2x, %maskz_2x
	kunpck_2x %k2, %k3, %k3
	kmov_2x	%k3, %maskm_2x
	and	%maskz_2x, %maskm_2x
	jz	L(first_vec_x1_or_x2_or_x3)

	bsr	%maskm_2x, %maskm_2x
	leaq	(VEC_SIZE * 2)(%rdi, %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
	/* Ideally we would safe k2/k3 but `kmov/kunpck` take uops on
	   port0 and have noticable overhead in the loop.  */
	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)
	VPCMPEQ	%VMM(5), %VMATCH, %k2
	vpxord	%VMM(6), %VMATCH, %VMM(3)

	VPMIN	%VMM(5), %VMM(6), %VMM(4)
	VPMIN	%VMM(3), %VMM(4), %VMM(3)

	VPTESTN	%VMM(3), %VMM(3), %k1
	subq	$(VEC_SIZE * -2), %rdi
	KORTEST %k1, %k2
	jz	L(second_aligned_loop)
	VPCMPEQ	%VMM(6), %VMATCH, %k3
	VPTESTN	%VMM(4), %VMM(4), %k1
	KTEST	%k1, %k1
	jz	L(second_aligned_loop_set_furthest_match)

	/* branch here because we know we have a match in VEC7/8 but
	   might not in VEC5/6 so the latter is expected to be less
	   likely.  */
	KORTEST %k2, %k3
	jnz	L(return_new_match)

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

	.p2align 4,, 10
L(return_new_match):
	VPTESTN	%VMM(5), %VMM(5), %k0

	/* Combined results from VEC5/6.  */
	kunpck_2x %k0, %k1, %k0
	kmov_2x	%k0, %maskz_2x

	blsmsk	%maskz_2x, %maskz_2x
	kunpck_2x %k2, %k3, %k3
	kmov_2x	%k3, %maskm_2x

	/* Match at end was out-of-bounds so use last known match.  */
	and	%maskz_2x, %maskm_2x
	jz	L(return_old_match)

	bsr	%maskm_2x, %maskm_2x
	leaq	(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
	ret

L(cross_page_boundary):
	/* eax contains all the page offset bits of src (rdi). `xor rdi,
	   rax` sets pointer will all page offset bits cleared so
	   offset of (PAGE_SIZE - VEC_SIZE) will get last aligned VEC
	   before page cross (guranteed to be safe to read). Doing this
	   as opposed to `movq %rdi, %rax; andq $-VEC_SIZE, %rax` saves
	   a bit of code size.  */
	xorq	%rdi, %rax
	VMOVU	(PAGE_SIZE - VEC_SIZE)(%rax), %VMM(1)
	VPTESTN	%VMM(1), %VMM(1), %k0
	KMOV	%k0, %VRCX

	/* Shift out zero CHAR matches that are before the begining of
	   src (rdi).  */
# ifdef USE_AS_WCSRCHR
	movl	%edi, %esi
	andl	$(VEC_SIZE - 1), %esi
	shrl	$2, %esi
# endif
	shrx	%VGPR(SHIFT_REG), %VRCX, %VRCX

	test	%VRCX, %VRCX
	jz	L(page_cross_continue)

	/* Found zero CHAR so need to test for search CHAR.  */
	VPCMP	$0, %VMATCH, %VMM(1), %k1
	KMOV	%k1, %VRAX
	/* Shift out search CHAR matches that are before the begining of
	   src (rdi).  */
	shrx	%VGPR(SHIFT_REG), %VRAX, %VRAX

	/* Check if any search CHAR match in range.  */
	blsmsk	%VRCX, %VRCX
	and	%VRCX, %VRAX
	jz	L(ret3)
	bsr	%VRAX, %VRAX
# ifdef USE_AS_WCSRCHR
	leaq	(%rdi, %rax, CHAR_SIZE), %rax
# else
	addq	%rdi, %rax
# endif
L(ret3):
	ret
END(STRRCHR)
#endif