1 | /* stpncpy(DST, SRC, COUNT) - Copy no more than N characters of |
2 | SRC to DEST, returning the address of the terminating '\0' in |
3 | DEST, if any, or else DEST + N. |
4 | For SPARC v9. |
5 | Copyright (C) 1998-2024 Free Software Foundation, Inc. |
6 | This file is part of the GNU C Library. |
7 | |
8 | The GNU C Library is free software; you can redistribute it and/or |
9 | modify it under the terms of the GNU Lesser General Public |
10 | License as published by the Free Software Foundation; either |
11 | version 2.1 of the License, or (at your option) any later version. |
12 | |
13 | The GNU C Library is distributed in the hope that it will be useful, |
14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
16 | Lesser General Public License for more details. |
17 | |
18 | You should have received a copy of the GNU Lesser General Public |
19 | License along with the GNU C Library; if not, see |
20 | <https://www.gnu.org/licenses/>. */ |
21 | |
22 | #include <sysdep.h> |
23 | #include <asm/asi.h> |
24 | #ifndef XCC |
25 | #define XCC xcc |
26 | #define USE_BPR |
27 | .register %g2, #scratch |
28 | .register %g3, #scratch |
29 | .register %g6, #scratch |
30 | #endif |
31 | |
32 | /* Normally, this uses |
33 | ((xword - 0x0101010101010101) & 0x8080808080808080) test |
34 | to find out if any byte in xword could be zero. This is fast, but |
35 | also gives false alarm for any byte in range 0x81-0xff. It does |
36 | not matter for correctness, as if this test tells us there could |
37 | be some zero byte, we check it byte by byte, but if bytes with |
38 | high bits set are common in the strings, then this will give poor |
39 | performance. You can #define EIGHTBIT_NOT_RARE and the algorithm |
40 | will use one tick slower, but more precise test |
41 | ((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080), |
42 | which does not give any false alarms (but if some bits are set, |
43 | one cannot assume from it which bytes are zero and which are not). |
44 | It is yet to be measured, what is the correct default for glibc |
45 | in these days for an average user. |
46 | */ |
47 | |
48 | .text |
49 | .align 32 |
50 | ENTRY(__stpncpy) |
51 | sethi %hi(0x01010101), %g1 /* IEU0 Group */ |
52 | #ifdef USE_BPR |
53 | brz,pn %o2, 19f /* CTI+IEU1 */ |
54 | #else |
55 | tst %o2 /* IEU1 */ |
56 | be,pn %XCC, 19f /* CTI */ |
57 | #endif |
58 | or %g1, %lo(0x01010101), %g1 /* IEU1 */ |
59 | andcc %o0, 7, %g0 /* IEU1 Group */ |
60 | |
61 | sllx %g1, 32, %g2 /* IEU0 */ |
62 | bne,pn %icc, 26f /* CTI */ |
63 | or %g1, %g2, %g1 /* IEU0 Group */ |
64 | andcc %o1, 7, %g3 /* IEU1 */ |
65 | |
66 | bne,pn %icc, 28f /* CTI */ |
67 | sllx %g1, 7, %g2 /* IEU0 Group */ |
68 | ldx [%o1], %o3 /* Load */ |
69 | 1: add %o1, 8, %o1 /* IEU1 */ |
70 | |
71 | 2: subcc %o2, 8, %o2 /* IEU1 Group */ |
72 | bl,pn %XCC, 18f /* CTI */ |
73 | sub %o3, %g1, %o4 /* IEU0 */ |
74 | add %o0, 8, %o0 /* IEU0 Group */ |
75 | |
76 | #ifdef EIGHTBIT_NOT_MORE |
77 | andn %o4, %o3, %o4 /* IEU1 */ |
78 | #endif |
79 | mov %o3, %g3 /* IEU1 */ |
80 | ldxa [%o1] ASI_PNF, %o3 /* Load */ |
81 | add %o1, 8, %o1 /* IEU0 Group */ |
82 | andcc %o4, %g2, %g0 /* IEU1 */ |
83 | |
84 | be,a,pt %xcc, 2b /* CTI */ |
85 | stx %g3, [%o0-8] /* Store Group */ |
86 | srlx %g3, 56, %g5 /* IEU0 Group */ |
87 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
88 | |
89 | be,pn %icc, 16f /* CTI */ |
90 | srlx %g3, 48, %g4 /* IEU0 */ |
91 | andcc %g4, 0xff, %g0 /* IEU1 Group */ |
92 | be,pn %icc, 15f /* CTI */ |
93 | |
94 | srlx %g3, 40, %g5 /* IEU0 */ |
95 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
96 | be,pn %icc, 14f /* CTI */ |
97 | srlx %g3, 32, %g4 /* IEU0 */ |
98 | |
99 | andcc %g4, 0xff, %g0 /* IEU1 Group */ |
100 | be,pn %icc, 13f /* CTI */ |
101 | srlx %g3, 24, %g5 /* IEU0 */ |
102 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
103 | |
104 | be,pn %icc, 12f /* CTI */ |
105 | srlx %g3, 16, %g4 /* IEU0 */ |
106 | andcc %g4, 0xff, %g0 /* IEU1 Group */ |
107 | be,pn %icc, 11f /* CTI */ |
108 | |
109 | srlx %g3, 8, %g5 /* IEU0 */ |
110 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
111 | be,pn %icc, 10f /* CTI */ |
112 | sub %o0, 1, %g6 /* IEU0 */ |
113 | |
114 | andcc %g3, 0xff, %g0 /* IEU1 Group */ |
115 | bne,pt %icc, 2b /* CTI */ |
116 | 3: stx %g3, [%o0-8] /* Store */ |
117 | andncc %o2, 31, %g3 /* IEU1 Group */ |
118 | |
119 | 4: be,pn %XCC, 41f /* CTI */ |
120 | and %o2, 31, %o2 /* IEU1 Group */ |
121 | 40: stx %g0, [%o0] /* Store */ |
122 | stx %g0, [%o0 + 8] /* Store Group */ |
123 | |
124 | subcc %g3, 32, %g3 /* IEU1 */ |
125 | stx %g0, [%o0 + 16] /* Store Group */ |
126 | stx %g0, [%o0 + 24] /* Store Group */ |
127 | bne,pt %XCC, 40b /* CTI */ |
128 | |
129 | add %o0, 32, %o0 /* IEU0 */ |
130 | 41: subcc %o2, 8, %o2 /* IEU1 Group */ |
131 | bl,a,pn %XCC, 6f /* CTI */ |
132 | andcc %o2, 4, %g0 /* IEU1 Group */ |
133 | |
134 | 5: stx %g0, [%o0] /* Store */ |
135 | subcc %o2, 8, %o2 /* IEU1 Group */ |
136 | bge,pt %XCC, 5b /* CTI */ |
137 | add %o0, 8, %o0 /* IEU0 */ |
138 | |
139 | andcc %o2, 4, %g0 /* IEU1 Group */ |
140 | 6: be,a,pn %icc, 7f /* CTI */ |
141 | andcc %o2, 2, %g0 /* IEU1 Group */ |
142 | stw %g0, [%o0] /* Store */ |
143 | |
144 | add %o0, 4, %o0 /* IEU0 */ |
145 | andcc %o2, 2, %g0 /* IEU1 Group */ |
146 | 7: be,a,pn %icc, 8f /* CTI */ |
147 | andcc %o2, 1, %g0 /* IEU1 Group */ |
148 | |
149 | sth %g0, [%o0] /* Store */ |
150 | add %o0, 2, %o0 /* IEU0 */ |
151 | andcc %o2, 1, %g0 /* IEU1 Group */ |
152 | 8: bne,a,pn %icc, 9f /* CTI */ |
153 | |
154 | stb %g0, [%o0] /* Store */ |
155 | 9: retl /* CTI+IEU1 Group */ |
156 | mov %g6, %o0 /* IEU0 */ |
157 | 10: subcc %o0, 2, %g6 /* IEU1 Group */ |
158 | |
159 | ba,pt %xcc, 3b /* CTI */ |
160 | sllx %g5, 8, %g3 /* IEU0 */ |
161 | 11: subcc %o0, 3, %g6 /* IEU1 Group */ |
162 | ba,pt %xcc, 3b /* CTI */ |
163 | |
164 | sllx %g4, 16, %g3 /* IEU0 */ |
165 | 12: subcc %o0, 4, %g6 /* IEU1 Group */ |
166 | ba,pt %xcc, 3b /* CTI */ |
167 | sllx %g5, 24, %g3 /* IEU0 */ |
168 | |
169 | 13: subcc %o0, 5, %g6 /* IEU1 Group */ |
170 | ba,pt %xcc, 3b /* CTI */ |
171 | sllx %g4, 32, %g3 /* IEU0 */ |
172 | 14: subcc %o0, 6, %g6 /* IEU1 Group */ |
173 | |
174 | ba,pt %xcc, 3b /* CTI */ |
175 | sllx %g5, 40, %g3 /* IEU0 */ |
176 | 15: subcc %o0, 7, %g6 /* IEU1 Group */ |
177 | ba,pt %xcc, 3b /* CTI */ |
178 | |
179 | sllx %g4, 48, %g3 /* IEU0 */ |
180 | 16: subcc %o0, 8, %g6 /* IEU1 Group */ |
181 | ba,pt %xcc, 3b /* CTI */ |
182 | clr %g3 /* IEU0 */ |
183 | |
184 | .align 16 |
185 | 17: or %o3, %o4, %o3 /* IEU0 Group */ |
186 | sub %o3, %g1, %o4 /* IEU1 */ |
187 | 18: addcc %o2, 8, %o2 /* IEU1 Group */ |
188 | be,pn %XCC, 19f /* CTI */ |
189 | |
190 | andcc %o4, %g2, %g0 /* IEU1 Group */ |
191 | be,pt %xcc, 21f /* CTI */ |
192 | srlx %o3, 56, %g5 /* IEU0 */ |
193 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
194 | |
195 | be,pn %icc, 20f /* CTI */ |
196 | stb %g5, [%o0] /* Store */ |
197 | add %o0, 1, %o0 /* IEU0 Group */ |
198 | subcc %o2, 1, %o2 /* IEU1 */ |
199 | |
200 | be,pn %XCC, 19f /* CTI */ |
201 | srlx %o3, 48, %g5 /* IEU0 Group */ |
202 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
203 | be,pn %icc, 20f /* CTI */ |
204 | |
205 | stb %g5, [%o0] /* Store */ |
206 | add %o0, 1, %o0 /* IEU0 Group */ |
207 | subcc %o2, 1, %o2 /* IEU1 */ |
208 | be,pn %XCC, 19f /* CTI */ |
209 | |
210 | srlx %o3, 40, %g5 /* IEU0 Group */ |
211 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
212 | be,pn %icc, 20f /* CTI */ |
213 | stb %g5, [%o0] /* Store */ |
214 | |
215 | add %o0, 1, %o0 /* IEU0 Group */ |
216 | subcc %o2, 1, %o2 /* IEU1 */ |
217 | be,pn %XCC, 19f /* CTI */ |
218 | srlx %o3, 32, %g5 /* IEU0 Group */ |
219 | |
220 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
221 | be,pn %icc, 20f /* CTI */ |
222 | stb %g5, [%o0] /* Store */ |
223 | add %o0, 1, %o0 /* IEU0 Group */ |
224 | |
225 | subcc %o2, 1, %o2 /* IEU1 */ |
226 | be,pn %XCC, 19f /* CTI */ |
227 | srlx %o3, 24, %g5 /* IEU0 Group */ |
228 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
229 | |
230 | be,pn %icc, 20f /* CTI */ |
231 | stb %g5, [%o0] /* Store */ |
232 | add %o0, 1, %o0 /* IEU0 Group */ |
233 | subcc %o2, 1, %o2 /* IEU1 */ |
234 | |
235 | be,pn %XCC, 19f /* CTI */ |
236 | srlx %o3, 16, %g5 /* IEU0 Group */ |
237 | andcc %g5, 0xff, %g0 /* IEU1 Group */ |
238 | be,pn %icc, 20f /* CTI */ |
239 | |
240 | stb %g5, [%o0] /* Store */ |
241 | add %o0, 1, %o0 /* IEU0 Group */ |
242 | subcc %o2, 1, %o2 /* IEU1 */ |
243 | be,pn %XCC, 19f /* CTI */ |
244 | |
245 | srlx %o3, 8, %o3 /* IEU0 Group */ |
246 | stb %o3, [%o0] /* Store */ |
247 | 59: add %o0, 1, %o2 /* IEU1 */ |
248 | andcc %o3, 0xff, %g0 /* IEU1 Group */ |
249 | |
250 | retl /* CTI+IEU1 Group */ |
251 | movne %icc, %o2, %o0 /* Single Group */ |
252 | 19: retl /* CTI+IEU1 Group */ |
253 | nop /* IEU0 */ |
254 | |
255 | 20: mov %o0, %g6 /* IEU0 Group */ |
256 | subcc %o2, 1, %o2 /* IEU1 */ |
257 | be,pn %XCC, 51f /* CTI */ |
258 | add %o0, 1, %o0 /* IEU0 Group */ |
259 | |
260 | 50: stb %g0, [%o0] /* Store Group */ |
261 | subcc %o2, 1, %o2 /* IEU1 Group */ |
262 | bne,pt %XCC, 50b /* CTI */ |
263 | add %o0, 1, %o0 /* IEU0 */ |
264 | |
265 | 51: retl /* CTI+IEU1 Group */ |
266 | mov %g6, %o0 /* IEU0 */ |
267 | |
268 | .align 16 |
269 | 21: andcc %o2, 4, %g0 /* IEU1 Group */ |
270 | be,pn %icc, 22f /* CTI */ |
271 | srlx %o3, 32, %g5 /* IEU0 */ |
272 | stw %g5, [%o0] /* Store Group */ |
273 | |
274 | add %o0, 4, %o0 /* IEU0 */ |
275 | mov %o3, %g5 /* IEU1 */ |
276 | 22: andcc %o2, 2, %g0 /* IEU1 Group */ |
277 | be,pn %icc, 23f /* CTI */ |
278 | |
279 | srlx %g5, 16, %g4 /* IEU0 */ |
280 | sth %g4, [%o0] /* Store Group */ |
281 | add %o0, 2, %o0 /* IEU0 */ |
282 | mov %g5, %g4 /* IEU1 */ |
283 | |
284 | 23: srlx %g4, 8, %g4 /* IEU0 Group */ |
285 | andcc %o2, 1, %g0 /* IEU1 */ |
286 | bne,a,pn %icc, 24f /* CTI */ |
287 | stb %g4, [%o0] /* Store Group */ |
288 | |
289 | retl /* CTI+IEU1 Group */ |
290 | nop /* IEU0 */ |
291 | 24: retl /* CTI+IEU1 Group */ |
292 | add %o0, 1, %o0 /* IEU0 */ |
293 | |
294 | .align 16 |
295 | 55: sub %o0, 1, %g6 /* IEU0 Group */ |
296 | 25: andcc %o0, 7, %g0 /* IEU1 */ |
297 | be,a,pn %icc, 4b /* CTI */ |
298 | andncc %o2, 31, %g3 /* IEU1 Group */ |
299 | |
300 | stb %g0, [%o0] /* Store Group */ |
301 | subcc %o2, 1, %o2 /* IEU1 */ |
302 | bne,pt %XCC, 25b /* CTI */ |
303 | add %o0, 1, %o0 /* IEU0 Group */ |
304 | |
305 | retl /* CTI+IEU1 Group */ |
306 | mov %g6, %o0 /* IEU0 */ |
307 | |
308 | .align 16 |
309 | 26: ldub [%o1], %o3 /* Load */ |
310 | sllx %g1, 7, %g2 /* IEU0 Group */ |
311 | stb %o3, [%o0] /* Store */ |
312 | 27: subcc %o2, 1, %o2 /* IEU1 */ |
313 | |
314 | be,pn %XCC, 59b /* CTI */ |
315 | add %o1, 1, %o1 /* IEU0 Group */ |
316 | add %o0, 1, %o0 /* IEU1 */ |
317 | andcc %o3, 0xff, %g0 /* IEU1 Group */ |
318 | |
319 | be,pn %icc, 55b /* CTI */ |
320 | lduba [%o1] ASI_PNF, %o3 /* Load */ |
321 | andcc %o0, 7, %g0 /* IEU1 Group */ |
322 | bne,a,pt %icc, 27b /* CTI */ |
323 | |
324 | stb %o3, [%o0] /* Store */ |
325 | andcc %o1, 7, %g3 /* IEU1 Group */ |
326 | be,a,pt %icc, 1b /* CTI */ |
327 | ldx [%o1], %o3 /* Load */ |
328 | |
329 | 28: orcc %g0, 64, %g4 /* IEU1 Group */ |
330 | sllx %g3, 3, %g5 /* IEU0 */ |
331 | sub %g4, %g5, %g4 /* IEU0 Group */ |
332 | sub %o1, %g3, %o1 /* IEU1 */ |
333 | /* %g1 = 0101010101010101 |
334 | %g2 = 8080808080808080 |
335 | %g3 = source alignment |
336 | %g5 = number of bits to shift left |
337 | %g4 = number of bits to shift right */ |
338 | |
339 | ldxa [%o1] ASI_PNF, %o5 /* Load Group */ |
340 | addcc %o1, 8, %o1 /* IEU1 */ |
341 | 29: sllx %o5, %g5, %o3 /* IEU0 Group */ |
342 | ldxa [%o1] ASI_PNF, %o5 /* Load */ |
343 | |
344 | subcc %o2, 8, %o2 /* IEU1 */ |
345 | bl,pn %XCC, 17b /* CTI */ |
346 | srlx %o5, %g4, %o4 /* IEU0 Group */ |
347 | add %o1, 8, %o1 /* IEU1 */ |
348 | |
349 | or %o3, %o4, %o3 /* IEU0 Group */ |
350 | add %o0, 8, %o0 /* IEU1 */ |
351 | sub %o3, %g1, %o4 /* IEU0 Group */ |
352 | #ifdef EIGHTBIT_NOT_RARE |
353 | andn %o4, %o3, %o4 /* IEU0 Group */ |
354 | #endif |
355 | andcc %o4, %g2, %g0 /* IEU1 Group */ |
356 | |
357 | be,a,pt %xcc, 29b /* CTI */ |
358 | stx %o3, [%o0-8] /* Store */ |
359 | srlx %o3, 56, %o4 /* IEU0 Group */ |
360 | andcc %o4, 0xff, %g0 /* IEU1 Group */ |
361 | |
362 | be,pn %icc, 36f /* CTI */ |
363 | srlx %o3, 48, %g6 /* IEU0 */ |
364 | andcc %g6, 0xff, %g0 /* IEU1 Group */ |
365 | be,pn %icc, 35f /* CTI */ |
366 | |
367 | srlx %o3, 40, %o4 /* IEU0 */ |
368 | andcc %o4, 0xff, %g0 /* IEU1 Group */ |
369 | be,pn %icc, 34f /* CTI */ |
370 | srlx %o3, 32, %g6 /* IEU0 */ |
371 | |
372 | andcc %g6, 0xff, %g0 /* IEU1 Group */ |
373 | be,pn %icc, 33f /* CTI */ |
374 | srlx %o3, 24, %o4 /* IEU0 */ |
375 | andcc %o4, 0xff, %g0 /* IEU1 Group */ |
376 | |
377 | be,pn %icc, 32f /* CTI */ |
378 | srlx %o3, 16, %g6 /* IEU0 */ |
379 | andcc %g6, 0xff, %g0 /* IEU1 Group */ |
380 | be,pn %icc, 31f /* CTI */ |
381 | |
382 | srlx %o3, 8, %o4 /* IEU0 */ |
383 | andcc %o4, 0xff, %g0 /* IEU1 Group */ |
384 | be,pn %icc, 30f /* CTI */ |
385 | andcc %o3, 0xff, %g0 /* IEU1 Group */ |
386 | |
387 | bne,pn %icc, 29b /* CTI */ |
388 | stx %o3, [%o0-8] /* Store */ |
389 | sub %o0, 1, %g6 /* IEU0 Group */ |
390 | ba,pt %xcc, 4b /* CTI */ |
391 | |
392 | andncc %o2, 31, %g3 /* IEU1 */ |
393 | 30: subcc %o0, 2, %g6 /* IEU0 */ |
394 | ba,pt %xcc, 3b /* CTI */ |
395 | sllx %o4, 8, %g3 /* IEU0 Group */ |
396 | |
397 | 31: sllx %g6, 16, %g3 /* IEU0 Group */ |
398 | ba,pt %xcc, 3b /* CTI */ |
399 | sub %o0, 3, %g6 /* IEU1 */ |
400 | 32: subcc %o0, 4, %g6 /* IEU1 Group */ |
401 | |
402 | ba,pt %xcc, 3b /* CTI */ |
403 | sllx %o4, 24, %g3 /* IEU0 */ |
404 | 33: sllx %g6, 32, %g3 /* IEU0 Group */ |
405 | ba,pt %xcc, 3b /* CTI */ |
406 | |
407 | sub %o0, 5, %g6 /* IEU1 */ |
408 | 34: subcc %o0, 6, %g6 /* IEU1 Group */ |
409 | ba,pt %xcc, 3b /* CTI */ |
410 | sllx %o4, 40, %g3 /* IEU0 */ |
411 | |
412 | 35: sllx %g6, 48, %g3 /* IEU0 Group */ |
413 | ba,pt %xcc, 3b /* CTI */ |
414 | sub %o0, 7, %g6 /* IEU1 */ |
415 | 36: subcc %o0, 8, %g6 /* IEU1 Group */ |
416 | |
417 | ba,pt %xcc, 3b /* CTI */ |
418 | sllx %o4, 56, %g3 /* IEU0 */ |
419 | END(__stpncpy) |
420 | |
421 | libc_hidden_def (__stpncpy) |
422 | weak_alias (__stpncpy, stpncpy) |
423 | |