]>
Commit | Line | Data |
---|---|---|
f0ba6598 | 1 | /* Vector optimized 32/64 bit S/390 version of wcspbrk. |
bfff8b1b | 2 | Copyright (C) 2015-2017 Free Software Foundation, Inc. |
f0ba6598 SL |
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 | <http://www.gnu.org/licenses/>. */ | |
18 | ||
19 | #if defined HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc) | |
20 | ||
21 | # include "sysdep.h" | |
22 | # include "asm-syntax.h" | |
23 | ||
24 | .text | |
25 | ||
26 | /* wchar_t *wcspbrk (const wchar_t *s, const wchar_t * accept) | |
27 | The wcspbrk() function locates the first occurrence in the string s | |
28 | of any of the characters in the string accept and returns a pointer | |
29 | to that character or NULL if not found. | |
30 | ||
31 | This method checks the length of accept string. If it fits entirely | |
32 | in one vector register, a fast algorithm is used, which does not need | |
33 | to check multiple parts of accept-string. Otherwise a slower full | |
34 | check of accept-string is used. | |
35 | ||
36 | register overview: | |
37 | r3: pointer to start of accept-string | |
38 | r2: pointer to start of search-string | |
39 | r0: loaded byte count of vlbb search-string (32bit unsigned) | |
40 | r4: found byte index (32bit unsigned) | |
41 | r1: current return len (64bit unsigned) | |
42 | v16: search-string | |
43 | v17: accept-string | |
44 | v18: temp-vreg | |
45 | ||
46 | ONLY FOR SLOW: | |
47 | v19: first accept-string | |
48 | v20: zero for preparing acc-vector | |
49 | v21: global mask; 1 indicates a match between | |
50 | search-string-vreg and any accept-character | |
51 | v22: current mask; 1 indicates a match between | |
52 | search-string-vreg and any accept-character in current acc-vreg | |
53 | v24: one for result-checking of former string-part | |
54 | v30, v31: for re-/storing registers r6, r8, r9 | |
55 | r5: current len of accept-string | |
56 | r6: zero-index in search-string or 16 if no zero | |
57 | or min(zero-index, loaded byte count) | |
58 | r8: >0, if former accept-string-part contains a zero, | |
59 | otherwise =0; | |
60 | r9: loaded byte count of vlbb accept-string | |
61 | */ | |
62 | ENTRY(__wcspbrk_vx) | |
63 | .machine "z13" | |
64 | .machinemode "zarch_nohighgprs" | |
65 | ||
66 | tmll %r2,3 /* Test if s is 4-byte aligned? */ | |
67 | jne .Lfallback /* And use common-code variant if not. */ | |
68 | ||
69 | /* | |
70 | Check if accept-string fits in one vreg: | |
71 | ---------------------------------------- | |
72 | */ | |
73 | vlbb %v17,0(%r3),6 /* Load accept. */ | |
74 | lcbb %r0,0(%r3),6 | |
75 | jo .Lcheck_onbb /* Special case if accept lays | |
76 | on block-boundary. */ | |
77 | ||
78 | .Lcheck_notonbb: | |
79 | lghi %r1,0 /* Zero out current len. */ | |
80 | vlgvf %r0,%v17,0 /* Get first element. */ | |
81 | clije %r0,0,.Lfast_end_null /* Return null if accept is empty. */ | |
82 | ||
83 | vistrfs %v17,%v17 /* Fill with zeros after first zero. */ | |
84 | je .Lfast /* Zero found -> accept fits in one vreg. */ | |
85 | j .Lslow /* No zero -> accept exceeds one vreg */ | |
86 | ||
87 | ||
88 | .Lcheck_onbb: | |
89 | /* Accept lays on block-boundary. */ | |
90 | nill %r0,65532 /* Recognize only fully loaded characters. */ | |
91 | je .Lcheck_onbb2 /* Reload vr, if we loaded no full wchar_t. */ | |
92 | vfenezf %v18,%v17,%v17 /* Search zero in loaded accept bytes. */ | |
93 | vlgvb %r4,%v18,7 /* Get index of zero or 16 if not found. */ | |
94 | clrjl %r4,%r0,.Lcheck_notonbb /* Zero index < loaded bytes count -> | |
95 | accept fits in one vreg; | |
96 | Fill with zeros and proceed | |
97 | with FAST. */ | |
98 | .Lcheck_onbb2: | |
99 | vl %v17,0(%r3) /* Load accept, which exceeds loaded bytes. */ | |
100 | j .Lcheck_notonbb /* Check if accept fits in one vreg. */ | |
101 | ||
102 | ||
103 | /* | |
104 | Search s for accept in one vreg | |
105 | ------------------------------- | |
106 | */ | |
107 | .Lfast: | |
108 | /* Complete accept-string in v17 and remaining bytes are zero. */ | |
109 | ||
110 | vlbb %v16,0(%r2),6 /* Load s until next 4k-byte boundary. */ | |
111 | lcbb %r0,0(%r2),6 /* Get bytes to 4k-byte boundary or 16. */ | |
112 | ||
113 | vfaezfs %v18,%v16,%v17,0 /* Find first element in v16 unequal to any | |
114 | in v17 or first zero element. */ | |
115 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ | |
116 | /* If found index is within loaded bytes, return with found | |
117 | element index (=equal count). */ | |
118 | clrjl %r4,%r0,.Lfast_loop_found2 | |
119 | ||
120 | /* Align s to 16 byte. */ | |
121 | risbgn %r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15. */ | |
122 | lghi %r1,16 /* current_len = 16. */ | |
123 | slr %r1,%r4 /* Compute bytes to 16bytes boundary. */ | |
124 | ||
125 | .Lfast_loop: | |
126 | vl %v16,0(%r1,%r2) /* Load search-string. */ | |
127 | vfaezfs %v18,%v16,%v17,0 /* Find first element in v16 equal to any | |
128 | in v17 or first zero element. */ | |
129 | jno .Lfast_loop_found | |
130 | ||
131 | vl %v16,16(%r1,%r2) | |
132 | vfaezfs %v18,%v16,%v17,0 | |
133 | jno .Lfast_loop_found16 | |
134 | ||
135 | vl %v16,32(%r1,%r2) | |
136 | vfaezfs %v18,%v16,%v17,0 | |
137 | jno .Lfast_loop_found32 | |
138 | ||
139 | vl %v16,48(%r1,%r2) | |
140 | vfaezfs %v18,%v16,%v17,0 | |
141 | jno .Lfast_loop_found48 | |
142 | ||
143 | aghi %r1,64 | |
144 | j .Lfast_loop /* Loop if no element was unequal to accept | |
145 | and not zero. */ | |
146 | ||
147 | /* Found equal or zero element. */ | |
148 | .Lfast_loop_found48: | |
149 | aghi %r1,16 | |
150 | .Lfast_loop_found32: | |
151 | aghi %r1,16 | |
152 | .Lfast_loop_found16: | |
153 | aghi %r1,16 | |
154 | .Lfast_loop_found: | |
155 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ | |
156 | .Lfast_loop_found2: | |
157 | srlg %r5,%r4,2 /* Convert byte-index to character-index. */ | |
158 | vlgvf %r0,%v16,0(%r5) /* Get found element. */ | |
159 | clije %r0,0,.Lfast_end_null /* Return null if no accept-char found */ | |
160 | algfr %r1,%r4 /* Add found index of char to current len. */ | |
161 | la %r2,0(%r1,%r2) /* And return pointer to first equal char. */ | |
162 | br %r14 | |
163 | ||
164 | .Lfast_end_null: | |
165 | lghi %r2,0 /* Return null if no character is equal. */ | |
166 | br %r14 | |
167 | ||
168 | ||
169 | ||
170 | ||
171 | /* | |
172 | Search s for accept in multiple vregs | |
173 | ------------------------------------- | |
174 | */ | |
175 | .Lslow: | |
176 | /* Save registers. */ | |
177 | vlvgg %v30,%r6,0 | |
178 | vlvgp %v31,%r8,%r9 | |
179 | ||
180 | /* Accept in v17 without zero */ | |
181 | vlr %v19,%v17 /* Save first acc-part for a fast reload. */ | |
182 | vzero %v20 /* Zero for preparing acc-vector. */ | |
183 | vone %v24 /* One for checking result of former string. */ | |
184 | ||
185 | /* Align s to 16 byte. */ | |
186 | risbg %r4,%r2,60,128+63,0 /* Test if s is aligned and | |
187 | %r4 = bits 60-63 'and' 15. */ | |
188 | je .Lslow_loop_str /* If s is aligned, loop aligned. */ | |
189 | lghi %r0,15 | |
190 | slr %r0,%r4 /* Compute highest index to load (15-x). */ | |
191 | vll %v16,%r0,0(%r2) /* Load up to 16byte boundary; | |
192 | needs highest index, left bytes are 0. */ | |
193 | ahi %r0,1 /* Work with loaded byte count. */ | |
194 | vzero %v21 /* Zero out global mask. */ | |
195 | lghi %r5,0 /* Set current len of accept-string to zero. */ | |
196 | vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */ | |
197 | lghi %r8,0 /* There is no zero in first accept-part. */ | |
198 | vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */ | |
199 | clije %r6,0,.Lslow_end_null /* If first element is zero | |
200 | (end of string) -> return null */ | |
201 | clr %r0,%r6 /* cc==1 if loaded byte count < zero-index. */ | |
202 | locrl %r6,%r0 /* Load on cc==1; zero-index = lbc. */ | |
203 | j .Lslow_loop_acc | |
204 | ||
205 | ||
206 | /* Process s in 16byte aligned loop. */ | |
207 | .Lslow_next_str: | |
208 | /* Check results of former processed str-part. */ | |
209 | vfeef %v18,%v21,%v24 /* Find first equal match in global mask | |
210 | (ones in element). */ | |
211 | vlgvb %r4,%v18,7 /* Get index of first one (=equal) | |
212 | or 16 if no match. */ | |
213 | /* Equal-index < min(zero-index, loaded byte count) | |
214 | -> return pointer to equal element. */ | |
215 | clrjl %r4,%r6,.Lslow_index_found | |
216 | /* Zero-index < loaded byte count | |
217 | -> former str-part was last str-part | |
218 | -> return null */ | |
219 | clrjl %r6,%r0,.Lslow_end_null | |
220 | /* All elements are zero (=no match) -> proceed with next str-part. */ | |
221 | ||
222 | vlr %v17,%v19 /* Load first part of accept (no zero). */ | |
223 | algfr %r1,%r0 /* Add loaded byte count to current len. */ | |
224 | ||
225 | .Lslow_loop_str: | |
226 | vl %v16,0(%r1,%r2) /* Load search-string */ | |
227 | lghi %r0,16 /* Loaded byte count is 16. */ | |
228 | vzero %v21 /* Zero out global mask. */ | |
229 | lghi %r5,0 /* Set current len of accept to zero. */ | |
230 | vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */ | |
231 | lghi %r8,0 /* There is no zero in first accept-part. */ | |
232 | vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */ | |
233 | clije %r6,0,.Lslow_end_null /* If first element is zero | |
234 | (end of string) -> return null. */ | |
235 | ||
236 | .Lslow_loop_acc: | |
237 | vfaef %v22,%v16,%v17,4 /* Create matching-mask (1 in mask -> | |
238 | Character matches any accepted character in | |
239 | this accept-string-part) IN=0, RT=1. */ | |
240 | vlgvf %r4,%v22,0 /* Get result of first element. */ | |
241 | /* First element is equal to any accepted characters | |
242 | (all other parts of accept cannot lead to a match before this one) | |
243 | -> current len is pointing to first element | |
244 | -> return found */ | |
245 | clijh %r4,0,.Lslow_end_found | |
246 | vo %v21,%v21,%v22 /* Global-mask = global-|matching-mask. */ | |
247 | /* Proceed with next acc until end of acc is reached. */ | |
248 | ||
249 | ||
250 | .Lslow_next_acc: | |
251 | clijh %r8,0,.Lslow_next_str /* There was a zero in the last acc-part | |
252 | -> add index to current len and | |
253 | end. */ | |
254 | vlbb %v17,16(%r5,%r3),6 /* Load next accept part. */ | |
255 | aghi %r5,16 /* Increment current len of accept-string. */ | |
256 | lcbb %r9,0(%r5,%r3),6 /* Get loaded byte count of accept-string. */ | |
257 | jo .Lslow_next_acc_onbb /* Jump away ifaccept-string is | |
258 | on block-boundary. */ | |
259 | .Lslow_next_acc_notonbb: | |
260 | vistrfs %v17,%v17 /* Fill with zeros after first zero. */ | |
261 | jo .Lslow_loop_acc /* No zero found -> no preparation needed. */ | |
262 | ||
263 | .Lslow_next_acc_prepare_zero: | |
264 | /* Zero in accept-part: fill zeros with first-accept-character. */ | |
265 | vlgvf %r8,%v17,0 /* Load first element of acc-part. */ | |
266 | clije %r8,0,.Lslow_next_str /* Proceed with next string-part, | |
267 | If first char in this part of accept | |
268 | is a zero. */ | |
269 | /* r8>0 -> zero found in this acc-part. */ | |
270 | vrepf %v18,%v17,0 /* Replicate first char accross all chars. */ | |
271 | vceqf %v22,%v20,%v17 /* Create a mask (v22) of null chars | |
272 | by comparing with 0 (v20). */ | |
273 | vsel %v17,%v18,%v17,%v22 /* Replace null chars with first char. */ | |
274 | j .Lslow_loop_acc /* Accept part is prepared -> process. */ | |
275 | ||
276 | .Lslow_next_acc_onbb: | |
277 | nill %r9,65532 /* Recognize only fully loaded characters. */ | |
278 | je .Lslow_next_acc_onbb2 /* Reload vr, if no full wchar_t. */ | |
279 | vfenezf %v18,%v17,%v17 /* Find zero in loaded bytes of accept part. */ | |
280 | vlgvb %r8,%v18,7 /* Load byte index of zero. */ | |
281 | clrjl %r8,%r9,.Lslow_next_acc_notonbb /* Found a zero in loaded bytes | |
282 | -> Prepare vreg. */ | |
283 | .Lslow_next_acc_onbb2: | |
284 | vl %v17,0(%r5,%r3) /* Load over boundary ... */ | |
285 | lghi %r8,0 /* r8=0 -> no zero in this part of acc, | |
286 | check for zero is in jump-target. */ | |
287 | j .Lslow_next_acc_notonbb /* ... and search for zero in | |
288 | fully loaded vreg again. */ | |
289 | ||
290 | .Lslow_end_null: | |
291 | lghi %r1,0 /* Return null if no character is equal. */ | |
292 | j .Lslow_end | |
293 | ||
294 | .Lslow_loop_found: | |
295 | vlgvb %r4,%v18,7 /* Load byte index of found element. */ | |
296 | srlg %r5,%r4,2 /* Convert byte-index to character-index. */ | |
297 | vlgvf %r0,%v16,0(%r5) /* Get found element. */ | |
298 | clije %r0,0,.Lslow_end_null /* Return null if no acc-char found. */ | |
299 | ||
300 | .Lslow_index_found: | |
301 | algfr %r1,%r4 /* Add found index of char to current len. */ | |
302 | .Lslow_end_found: | |
303 | la %r1,0(%r1,%r2) /* And return pointer to first equal char. */ | |
304 | ||
305 | .Lslow_end: | |
306 | /* Restore registers. */ | |
307 | vlgvg %r6,%v30,0 | |
308 | vlgvg %r8,%v31,0 | |
309 | vlgvg %r9,%v31,1 | |
310 | lgr %r2,%r1 | |
311 | br %r14 | |
312 | .Lfallback: | |
313 | jg __wcspbrk_c | |
314 | END(__wcspbrk_vx) | |
315 | #endif /* HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc) */ |