1 | /* Vector optimized 32/64 bit S/390 version of wcspbrk. |
2 | Copyright (C) 2015-2022 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | |
5 | The GNU C Library is free software; you can redistribute it and/or |
6 | modify it under the terms of the GNU Lesser General Public |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the License, or (at your option) any later version. |
9 | |
10 | The GNU C Library is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | Lesser General Public License for more details. |
14 | |
15 | You should have received a copy of the GNU Lesser General Public |
16 | License along with the GNU C Library; if not, see |
17 | <https://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <ifunc-wcspbrk.h> |
20 | #if HAVE_WCSPBRK_Z13 |
21 | |
22 | # include "sysdep.h" |
23 | # include "asm-syntax.h" |
24 | |
25 | .text |
26 | |
27 | /* wchar_t *wcspbrk (const wchar_t *s, const wchar_t * accept) |
28 | The wcspbrk() function locates the first occurrence in the string s |
29 | of any of the characters in the string accept and returns a pointer |
30 | to that character or NULL if not found. |
31 | |
32 | This method checks the length of accept string. If it fits entirely |
33 | in one vector register, a fast algorithm is used, which does not need |
34 | to check multiple parts of accept-string. Otherwise a slower full |
35 | check of accept-string is used. |
36 | |
37 | register overview: |
38 | r3: pointer to start of accept-string |
39 | r2: pointer to start of search-string |
40 | r0: loaded byte count of vlbb search-string (32bit unsigned) |
41 | r4: found byte index (32bit unsigned) |
42 | r1: current return len (64bit unsigned) |
43 | v16: search-string |
44 | v17: accept-string |
45 | v18: temp-vreg |
46 | |
47 | ONLY FOR SLOW: |
48 | v19: first accept-string |
49 | v20: zero for preparing acc-vector |
50 | v21: global mask; 1 indicates a match between |
51 | search-string-vreg and any accept-character |
52 | v22: current mask; 1 indicates a match between |
53 | search-string-vreg and any accept-character in current acc-vreg |
54 | v24: one for result-checking of former string-part |
55 | v30, v31: for re-/storing registers r6, r8, r9 |
56 | r5: current len of accept-string |
57 | r6: zero-index in search-string or 16 if no zero |
58 | or min(zero-index, loaded byte count) |
59 | r8: >0, if former accept-string-part contains a zero, |
60 | otherwise =0; |
61 | r9: loaded byte count of vlbb accept-string |
62 | */ |
63 | ENTRY(WCSPBRK_Z13) |
64 | .machine "z13" |
65 | .machinemode "zarch_nohighgprs" |
66 | |
67 | tmll %r2,3 /* Test if s is 4-byte aligned? */ |
68 | jne .Lfallback /* And use common-code variant if not. */ |
69 | |
70 | /* |
71 | Check if accept-string fits in one vreg: |
72 | ---------------------------------------- |
73 | */ |
74 | vlbb %v17,0(%r3),6 /* Load accept. */ |
75 | lcbb %r0,0(%r3),6 |
76 | jo .Lcheck_onbb /* Special case if accept lays |
77 | on block-boundary. */ |
78 | |
79 | .Lcheck_notonbb: |
80 | lghi %r1,0 /* Zero out current len. */ |
81 | vlgvf %r0,%v17,0 /* Get first element. */ |
82 | clije %r0,0,.Lfast_end_null /* Return null if accept is empty. */ |
83 | |
84 | vistrfs %v17,%v17 /* Fill with zeros after first zero. */ |
85 | je .Lfast /* Zero found -> accept fits in one vreg. */ |
86 | j .Lslow /* No zero -> accept exceeds one vreg */ |
87 | |
88 | |
89 | .Lcheck_onbb: |
90 | /* Accept lays on block-boundary. */ |
91 | nill %r0,65532 /* Recognize only fully loaded characters. */ |
92 | je .Lcheck_onbb2 /* Reload vr, if we loaded no full wchar_t. */ |
93 | vfenezf %v18,%v17,%v17 /* Search zero in loaded accept bytes. */ |
94 | vlgvb %r4,%v18,7 /* Get index of zero or 16 if not found. */ |
95 | clrjl %r4,%r0,.Lcheck_notonbb /* Zero index < loaded bytes count -> |
96 | accept fits in one vreg; |
97 | Fill with zeros and proceed |
98 | with FAST. */ |
99 | .Lcheck_onbb2: |
100 | vl %v17,0(%r3) /* Load accept, which exceeds loaded bytes. */ |
101 | j .Lcheck_notonbb /* Check if accept fits in one vreg. */ |
102 | |
103 | |
104 | /* |
105 | Search s for accept in one vreg |
106 | ------------------------------- |
107 | */ |
108 | .Lfast: |
109 | /* Complete accept-string in v17 and remaining bytes are zero. */ |
110 | |
111 | vlbb %v16,0(%r2),6 /* Load s until next 4k-byte boundary. */ |
112 | lcbb %r0,0(%r2),6 /* Get bytes to 4k-byte boundary or 16. */ |
113 | |
114 | vfaezfs %v18,%v16,%v17,0 /* Find first element in v16 unequal to any |
115 | in v17 or first zero element. */ |
116 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ |
117 | /* If found index is within loaded bytes, return with found |
118 | element index (=equal count). */ |
119 | clrjl %r4,%r0,.Lfast_loop_found2 |
120 | |
121 | /* Align s to 16 byte. */ |
122 | risbgn %r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15. */ |
123 | lghi %r1,16 /* current_len = 16. */ |
124 | slr %r1,%r4 /* Compute bytes to 16bytes boundary. */ |
125 | |
126 | .Lfast_loop: |
127 | vl %v16,0(%r1,%r2) /* Load search-string. */ |
128 | vfaezfs %v18,%v16,%v17,0 /* Find first element in v16 equal to any |
129 | in v17 or first zero element. */ |
130 | jno .Lfast_loop_found |
131 | |
132 | vl %v16,16(%r1,%r2) |
133 | vfaezfs %v18,%v16,%v17,0 |
134 | jno .Lfast_loop_found16 |
135 | |
136 | vl %v16,32(%r1,%r2) |
137 | vfaezfs %v18,%v16,%v17,0 |
138 | jno .Lfast_loop_found32 |
139 | |
140 | vl %v16,48(%r1,%r2) |
141 | vfaezfs %v18,%v16,%v17,0 |
142 | jno .Lfast_loop_found48 |
143 | |
144 | aghi %r1,64 |
145 | j .Lfast_loop /* Loop if no element was unequal to accept |
146 | and not zero. */ |
147 | |
148 | /* Found equal or zero element. */ |
149 | .Lfast_loop_found48: |
150 | aghi %r1,16 |
151 | .Lfast_loop_found32: |
152 | aghi %r1,16 |
153 | .Lfast_loop_found16: |
154 | aghi %r1,16 |
155 | .Lfast_loop_found: |
156 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ |
157 | .Lfast_loop_found2: |
158 | srlg %r5,%r4,2 /* Convert byte-index to character-index. */ |
159 | vlgvf %r0,%v16,0(%r5) /* Get found element. */ |
160 | clije %r0,0,.Lfast_end_null /* Return null if no accept-char found */ |
161 | algfr %r1,%r4 /* Add found index of char to current len. */ |
162 | la %r2,0(%r1,%r2) /* And return pointer to first equal char. */ |
163 | br %r14 |
164 | |
165 | .Lfast_end_null: |
166 | lghi %r2,0 /* Return null if no character is equal. */ |
167 | br %r14 |
168 | |
169 | |
170 | |
171 | |
172 | /* |
173 | Search s for accept in multiple vregs |
174 | ------------------------------------- |
175 | */ |
176 | .Lslow: |
177 | /* Save registers. */ |
178 | vlvgg %v30,%r6,0 |
179 | vlvgp %v31,%r8,%r9 |
180 | |
181 | /* Accept in v17 without zero */ |
182 | vlr %v19,%v17 /* Save first acc-part for a fast reload. */ |
183 | vzero %v20 /* Zero for preparing acc-vector. */ |
184 | vone %v24 /* One for checking result of former string. */ |
185 | |
186 | /* Align s to 16 byte. */ |
187 | risbg %r4,%r2,60,128+63,0 /* Test if s is aligned and |
188 | %r4 = bits 60-63 'and' 15. */ |
189 | je .Lslow_loop_str /* If s is aligned, loop aligned. */ |
190 | lghi %r0,15 |
191 | slr %r0,%r4 /* Compute highest index to load (15-x). */ |
192 | vll %v16,%r0,0(%r2) /* Load up to 16byte boundary; |
193 | needs highest index, left bytes are 0. */ |
194 | ahi %r0,1 /* Work with loaded byte count. */ |
195 | vzero %v21 /* Zero out global mask. */ |
196 | lghi %r5,0 /* Set current len of accept-string to zero. */ |
197 | vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */ |
198 | lghi %r8,0 /* There is no zero in first accept-part. */ |
199 | vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */ |
200 | clije %r6,0,.Lslow_end_null /* If first element is zero |
201 | (end of string) -> return null */ |
202 | clr %r0,%r6 /* cc==1 if loaded byte count < zero-index. */ |
203 | locrl %r6,%r0 /* Load on cc==1; zero-index = lbc. */ |
204 | j .Lslow_loop_acc |
205 | |
206 | |
207 | /* Process s in 16byte aligned loop. */ |
208 | .Lslow_next_str: |
209 | /* Check results of former processed str-part. */ |
210 | vfeef %v18,%v21,%v24 /* Find first equal match in global mask |
211 | (ones in element). */ |
212 | vlgvb %r4,%v18,7 /* Get index of first one (=equal) |
213 | or 16 if no match. */ |
214 | /* Equal-index < min(zero-index, loaded byte count) |
215 | -> return pointer to equal element. */ |
216 | clrjl %r4,%r6,.Lslow_index_found |
217 | /* Zero-index < loaded byte count |
218 | -> former str-part was last str-part |
219 | -> return null */ |
220 | clrjl %r6,%r0,.Lslow_end_null |
221 | /* All elements are zero (=no match) -> proceed with next str-part. */ |
222 | |
223 | vlr %v17,%v19 /* Load first part of accept (no zero). */ |
224 | algfr %r1,%r0 /* Add loaded byte count to current len. */ |
225 | |
226 | .Lslow_loop_str: |
227 | vl %v16,0(%r1,%r2) /* Load search-string */ |
228 | lghi %r0,16 /* Loaded byte count is 16. */ |
229 | vzero %v21 /* Zero out global mask. */ |
230 | lghi %r5,0 /* Set current len of accept to zero. */ |
231 | vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */ |
232 | lghi %r8,0 /* There is no zero in first accept-part. */ |
233 | vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */ |
234 | clije %r6,0,.Lslow_end_null /* If first element is zero |
235 | (end of string) -> return null. */ |
236 | |
237 | .Lslow_loop_acc: |
238 | vfaef %v22,%v16,%v17,4 /* Create matching-mask (1 in mask -> |
239 | Character matches any accepted character in |
240 | this accept-string-part) IN=0, RT=1. */ |
241 | vlgvf %r4,%v22,0 /* Get result of first element. */ |
242 | /* First element is equal to any accepted characters |
243 | (all other parts of accept cannot lead to a match before this one) |
244 | -> current len is pointing to first element |
245 | -> return found */ |
246 | clijh %r4,0,.Lslow_end_found |
247 | vo %v21,%v21,%v22 /* Global-mask = global-|matching-mask. */ |
248 | /* Proceed with next acc until end of acc is reached. */ |
249 | |
250 | |
251 | .Lslow_next_acc: |
252 | clijh %r8,0,.Lslow_next_str /* There was a zero in the last acc-part |
253 | -> add index to current len and |
254 | end. */ |
255 | vlbb %v17,16(%r5,%r3),6 /* Load next accept part. */ |
256 | aghi %r5,16 /* Increment current len of accept-string. */ |
257 | lcbb %r9,0(%r5,%r3),6 /* Get loaded byte count of accept-string. */ |
258 | jo .Lslow_next_acc_onbb /* Jump away ifaccept-string is |
259 | on block-boundary. */ |
260 | .Lslow_next_acc_notonbb: |
261 | vistrfs %v17,%v17 /* Fill with zeros after first zero. */ |
262 | jo .Lslow_loop_acc /* No zero found -> no preparation needed. */ |
263 | |
264 | .Lslow_next_acc_prepare_zero: |
265 | /* Zero in accept-part: fill zeros with first-accept-character. */ |
266 | vlgvf %r8,%v17,0 /* Load first element of acc-part. */ |
267 | clije %r8,0,.Lslow_next_str /* Proceed with next string-part, |
268 | If first char in this part of accept |
269 | is a zero. */ |
270 | /* r8>0 -> zero found in this acc-part. */ |
271 | vrepf %v18,%v17,0 /* Replicate first char accross all chars. */ |
272 | vceqf %v22,%v20,%v17 /* Create a mask (v22) of null chars |
273 | by comparing with 0 (v20). */ |
274 | vsel %v17,%v18,%v17,%v22 /* Replace null chars with first char. */ |
275 | j .Lslow_loop_acc /* Accept part is prepared -> process. */ |
276 | |
277 | .Lslow_next_acc_onbb: |
278 | nill %r9,65532 /* Recognize only fully loaded characters. */ |
279 | je .Lslow_next_acc_onbb2 /* Reload vr, if no full wchar_t. */ |
280 | vfenezf %v18,%v17,%v17 /* Find zero in loaded bytes of accept part. */ |
281 | vlgvb %r8,%v18,7 /* Load byte index of zero. */ |
282 | clrjl %r8,%r9,.Lslow_next_acc_notonbb /* Found a zero in loaded bytes |
283 | -> Prepare vreg. */ |
284 | .Lslow_next_acc_onbb2: |
285 | vl %v17,0(%r5,%r3) /* Load over boundary ... */ |
286 | lghi %r8,0 /* r8=0 -> no zero in this part of acc, |
287 | check for zero is in jump-target. */ |
288 | j .Lslow_next_acc_notonbb /* ... and search for zero in |
289 | fully loaded vreg again. */ |
290 | |
291 | .Lslow_end_null: |
292 | lghi %r1,0 /* Return null if no character is equal. */ |
293 | j .Lslow_end |
294 | |
295 | .Lslow_loop_found: |
296 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ |
297 | srlg %r5,%r4,2 /* Convert byte-index to character-index. */ |
298 | vlgvf %r0,%v16,0(%r5) /* Get found element. */ |
299 | clije %r0,0,.Lslow_end_null /* Return null if no acc-char found. */ |
300 | |
301 | .Lslow_index_found: |
302 | algfr %r1,%r4 /* Add found index of char to current len. */ |
303 | .Lslow_end_found: |
304 | la %r1,0(%r1,%r2) /* And return pointer to first equal char. */ |
305 | |
306 | .Lslow_end: |
307 | /* Restore registers. */ |
308 | vlgvg %r6,%v30,0 |
309 | vlgvg %r8,%v31,0 |
310 | vlgvg %r9,%v31,1 |
311 | lgr %r2,%r1 |
312 | br %r14 |
313 | .Lfallback: |
314 | jg WCSPBRK_C |
315 | END(WCSPBRK_Z13) |
316 | |
317 | # if ! HAVE_WCSPBRK_IFUNC |
318 | strong_alias (WCSPBRK_Z13, wcspbrk) |
319 | # endif |
320 | |
321 | # if defined HAVE_S390_MIN_Z13_ZARCH_ASM_SUPPORT \ |
322 | && defined SHARED && IS_IN (libc) |
323 | strong_alias (WCSPBRK_Z13, __GI_wcspbrk) |
324 | # endif |
325 | #endif |
326 | |