2.9
[glibc/nacl-glibc.git] / sysdeps / sparc / sparc64 / stpncpy.S
blobc1ea820a9b6c0cb290e10543359d4e50220e970f
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, 1999, 2002, 2003 Free Software Foundation, Inc.
6    This file is part of the GNU C Library.
7    Contributed by Jakub Jelinek <jj@ultra.linux.cz> and
8                   Jan Vondrak <jvon4518@ss1000.ms.mff.cuni.cz>.
10    The GNU C Library is free software; you can redistribute it and/or
11    modify it under the terms of the GNU Lesser General Public
12    License as published by the Free Software Foundation; either
13    version 2.1 of the License, or (at your option) any later version.
15    The GNU C Library is distributed in the hope that it will be useful,
16    but WITHOUT ANY WARRANTY; without even the implied warranty of
17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18    Lesser General Public License for more details.
20    You should have received a copy of the GNU Lesser General Public
21    License along with the GNU C Library; if not, write to the Free
22    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
23    02111-1307 USA.  */
25 #include <sysdep.h>
26 #include <asm/asi.h>
27 #ifndef XCC
28 #define XCC xcc
29 #define USE_BPR
30         .register       %g2, #scratch
31         .register       %g3, #scratch
32         .register       %g6, #scratch
33 #endif
35         /* Normally, this uses
36            ((xword - 0x0101010101010101) & 0x8080808080808080) test
37            to find out if any byte in xword could be zero. This is fast, but
38            also gives false alarm for any byte in range 0x81-0xff. It does
39            not matter for correctness, as if this test tells us there could
40            be some zero byte, we check it byte by byte, but if bytes with
41            high bits set are common in the strings, then this will give poor
42            performance. You can #define EIGHTBIT_NOT_RARE and the algorithm
43            will use one tick slower, but more precise test
44            ((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080),
45            which does not give any false alarms (but if some bits are set,
46            one cannot assume from it which bytes are zero and which are not).
47            It is yet to be measured, what is the correct default for glibc
48            in these days for an average user.
49          */
51         .text
52         .align          32
53 ENTRY(__stpncpy)
54         sethi           %hi(0x01010101), %g1            /* IEU0         Group           */
55 #ifdef USE_BPR
56         brz,pn          %o2, 19f                        /* CTI+IEU1                     */
57 #else
58         tst             %o2                             /* IEU1                         */
59         be,pn           %XCC, 19f                       /* CTI                          */
60 #endif
61          or             %g1, %lo(0x01010101), %g1       /* IEU1                         */
62         andcc           %o0, 7, %g0                     /* IEU1         Group           */
64         sllx            %g1, 32, %g2                    /* IEU0                         */
65         bne,pn          %icc, 26f                       /* CTI                          */
66          or             %g1, %g2, %g1                   /* IEU0         Group           */
67         andcc           %o1, 7, %g3                     /* IEU1                         */
69         bne,pn          %icc, 28f                       /* CTI                          */
70          sllx           %g1, 7, %g2                     /* IEU0         Group           */
71         ldx             [%o1], %o3                      /* Load                         */
72 1:      add             %o1, 8, %o1                     /* IEU1                         */
74 2:      subcc           %o2, 8, %o2                     /* IEU1         Group           */
75         bl,pn           %XCC, 18f                       /* CTI                          */
76          sub            %o3, %g1, %o4                   /* IEU0                         */
77         add             %o0, 8, %o0                     /* IEU0         Group           */
79 #ifdef EIGHTBIT_NOT_MORE
80         andn            %o4, %o3, %o4                   /* IEU1                         */
81 #endif
82         mov             %o3, %g3                        /* IEU1                         */
83         ldxa            [%o1] ASI_PNF, %o3              /* Load                         */
84         add             %o1, 8, %o1                     /* IEU0         Group           */
85         andcc           %o4, %g2, %g0                   /* IEU1                         */
87         be,a,pt         %xcc, 2b                        /* CTI                          */
88          stx            %g3, [%o0-8]                    /* Store        Group           */
89         srlx            %g3, 56, %g5                    /* IEU0         Group           */
90         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
92         be,pn           %icc, 16f                       /* CTI                          */
93          srlx           %g3, 48, %g4                    /* IEU0                         */
94         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
95         be,pn           %icc, 15f                       /* CTI                          */
97          srlx           %g3, 40, %g5                    /* IEU0                         */
98         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
99         be,pn           %icc, 14f                       /* CTI                          */
100          srlx           %g3, 32, %g4                    /* IEU0                         */
102         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
103         be,pn           %icc, 13f                       /* CTI                          */
104          srlx           %g3, 24, %g5                    /* IEU0                         */
105         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
107         be,pn           %icc, 12f                       /* CTI                          */
108          srlx           %g3, 16, %g4                    /* IEU0                         */
109         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
110         be,pn           %icc, 11f                       /* CTI                          */
112          srlx           %g3, 8, %g5                     /* IEU0                         */
113         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
114         be,pn           %icc, 10f                       /* CTI                          */
115          sub            %o0, 1, %g6                     /* IEU0                         */
117         andcc           %g3, 0xff, %g0                  /* IEU1         Group           */
118         bne,pt          %icc, 2b                        /* CTI                          */
119 3:       stx            %g3, [%o0-8]                    /* Store                        */
120         andncc          %o2, 31, %g3                    /* IEU1         Group           */
122 4:      be,pn           %XCC, 41f                       /* CTI                          */
123          and            %o2, 31, %o2                    /* IEU1         Group           */
124 40:     stx             %g0, [%o0]                      /* Store                        */
125         stx             %g0, [%o0 + 8]                  /* Store        Group           */
127         subcc           %g3, 32, %g3                    /* IEU1                         */
128         stx             %g0, [%o0 + 16]                 /* Store        Group           */
129         stx             %g0, [%o0 + 24]                 /* Store        Group           */
130         bne,pt          %XCC, 40b                       /* CTI                          */
132          add            %o0, 32, %o0                    /* IEU0                         */
133 41:     subcc           %o2, 8, %o2                     /* IEU1         Group           */
134         bl,a,pn         %XCC, 6f                        /* CTI                          */
135          andcc          %o2, 4, %g0                     /* IEU1         Group           */
137 5:      stx             %g0, [%o0]                      /* Store                        */
138         subcc           %o2, 8, %o2                     /* IEU1         Group           */
139         bge,pt          %XCC, 5b                        /* CTI                          */
140          add            %o0, 8, %o0                     /* IEU0                         */
142         andcc           %o2, 4, %g0                     /* IEU1         Group           */
143 6:      be,a,pn         %icc, 7f                        /* CTI                          */
144          andcc          %o2, 2, %g0                     /* IEU1         Group           */
145         stw             %g0, [%o0]                      /* Store                        */
147         add             %o0, 4, %o0                     /* IEU0                         */
148         andcc           %o2, 2, %g0                     /* IEU1         Group           */
149 7:      be,a,pn         %icc, 8f                        /* CTI                          */
150          andcc          %o2, 1, %g0                     /* IEU1         Group           */
152         sth             %g0, [%o0]                      /* Store                        */
153         add             %o0, 2, %o0                     /* IEU0                         */
154         andcc           %o2, 1, %g0                     /* IEU1         Group           */
155 8:      bne,a,pn        %icc, 9f                        /* CTI                          */
157          stb            %g0, [%o0]                      /* Store                        */
158 9:      retl                                            /* CTI+IEU1     Group           */
159          mov            %g6, %o0                        /* IEU0                         */
160 10:     subcc           %o0, 2, %g6                     /* IEU1         Group           */
162         ba,pt           %xcc, 3b                        /* CTI                          */
163          sllx           %g5, 8, %g3                     /* IEU0                         */
164 11:     subcc           %o0, 3, %g6                     /* IEU1         Group           */
165         ba,pt           %xcc, 3b                        /* CTI                          */
167          sllx           %g4, 16, %g3                    /* IEU0                         */
168 12:     subcc           %o0, 4, %g6                     /* IEU1         Group           */
169         ba,pt           %xcc, 3b                        /* CTI                          */
170          sllx           %g5, 24, %g3                    /* IEU0                         */
172 13:     subcc           %o0, 5, %g6                     /* IEU1         Group           */
173         ba,pt           %xcc, 3b                        /* CTI                          */
174          sllx           %g4, 32, %g3                    /* IEU0                         */
175 14:     subcc           %o0, 6, %g6                     /* IEU1         Group           */
177         ba,pt           %xcc, 3b                        /* CTI                          */
178          sllx           %g5, 40, %g3                    /* IEU0                         */
179 15:     subcc           %o0, 7, %g6                     /* IEU1         Group           */
180         ba,pt           %xcc, 3b                        /* CTI                          */
182          sllx           %g4, 48, %g3                    /* IEU0                         */
183 16:     subcc           %o0, 8, %g6                     /* IEU1         Group           */
184         ba,pt           %xcc, 3b                        /* CTI                          */
185          clr            %g3                             /* IEU0                         */
187         .align          16
188 17:     or              %o3, %o4, %o3                   /* IEU0         Group           */
189         sub             %o3, %g1, %o4                   /* IEU1                         */
190 18:     addcc           %o2, 8, %o2                     /* IEU1         Group           */
191         be,pn           %XCC, 19f                       /* CTI                          */
193          andcc          %o4, %g2, %g0                   /* IEU1         Group           */
194         be,pt           %xcc, 21f                       /* CTI                          */
195          srlx           %o3, 56, %g5                    /* IEU0                         */
196         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
198         be,pn           %icc, 20f                       /* CTI                          */
199          stb            %g5, [%o0]                      /* Store                        */
200         add             %o0, 1, %o0                     /* IEU0         Group           */
201         subcc           %o2, 1, %o2                     /* IEU1                         */
203         be,pn           %XCC, 19f                       /* CTI                          */
204          srlx           %o3, 48, %g5                    /* IEU0         Group           */
205         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
206         be,pn           %icc, 20f                       /* CTI                          */
208          stb            %g5, [%o0]                      /* Store                        */
209         add             %o0, 1, %o0                     /* IEU0         Group           */
210         subcc           %o2, 1, %o2                     /* IEU1                         */
211         be,pn           %XCC, 19f                       /* CTI                          */
213          srlx           %o3, 40, %g5                    /* IEU0         Group           */
214         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
215         be,pn           %icc, 20f                       /* CTI                          */
216          stb            %g5, [%o0]                      /* Store                        */
218         add             %o0, 1, %o0                     /* IEU0         Group           */
219         subcc           %o2, 1, %o2                     /* IEU1                         */
220         be,pn           %XCC, 19f                       /* CTI                          */
221          srlx           %o3, 32, %g5                    /* IEU0         Group           */
223         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
224         be,pn           %icc, 20f                       /* CTI                          */
225          stb            %g5, [%o0]                      /* Store                        */
226         add             %o0, 1, %o0                     /* IEU0         Group           */
228         subcc           %o2, 1, %o2                     /* IEU1                         */
229         be,pn           %XCC, 19f                       /* CTI                          */
230          srlx           %o3, 24, %g5                    /* IEU0         Group           */
231         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
233         be,pn           %icc, 20f                       /* CTI                          */
234          stb            %g5, [%o0]                      /* Store                        */
235         add             %o0, 1, %o0                     /* IEU0         Group           */
236         subcc           %o2, 1, %o2                     /* IEU1                         */
238         be,pn           %XCC, 19f                       /* CTI                          */
239          srlx           %o3, 16, %g5                    /* IEU0         Group           */
240         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
241         be,pn           %icc, 20f                       /* CTI                          */
243          stb            %g5, [%o0]                      /* Store                        */
244         add             %o0, 1, %o0                     /* IEU0         Group           */
245         subcc           %o2, 1, %o2                     /* IEU1                         */
246         be,pn           %XCC, 19f                       /* CTI                          */
248          srlx           %o3, 8, %o3                     /* IEU0         Group           */
249         stb             %o3, [%o0]                      /* Store                        */ 
250 59:     add             %o0, 1, %o2                     /* IEU1                         */
251         andcc           %o3, 0xff, %g0                  /* IEU1         Group           */
253         retl                                            /* CTI+IEU1     Group           */
254          movne          %icc, %o2, %o0                  /* Single       Group           */
255 19:     retl                                            /* CTI+IEU1     Group           */
256          nop                                            /* IEU0                         */
258 20:     mov             %o0, %g6                        /* IEU0         Group           */
259         subcc           %o2, 1, %o2                     /* IEU1                         */
260         be,pn           %XCC, 51f                       /* CTI                          */
261          add            %o0, 1, %o0                     /* IEU0         Group           */
263 50:     stb             %g0, [%o0]                      /* Store        Group           */
264         subcc           %o2, 1, %o2                     /* IEU1         Group           */
265         bne,pt          %XCC, 50b                       /* CTI                          */
266          add            %o0, 1, %o0                     /* IEU0                         */
268 51:     retl                                            /* CTI+IEU1     Group           */
269          mov            %g6, %o0                        /* IEU0                         */
271         .align          16
272 21:     andcc           %o2, 4, %g0                     /* IEU1         Group           */
273         be,pn           %icc, 22f                       /* CTI                          */
274          srlx           %o3, 32, %g5                    /* IEU0                         */
275         stw             %g5, [%o0]                      /* Store        Group           */
277         add             %o0, 4, %o0                     /* IEU0                         */
278         mov             %o3, %g5                        /* IEU1                         */
279 22:     andcc           %o2, 2, %g0                     /* IEU1         Group           */
280         be,pn           %icc, 23f                       /* CTI                          */
282          srlx           %g5, 16, %g4                    /* IEU0                         */
283         sth             %g4, [%o0]                      /* Store        Group           */
284         add             %o0, 2, %o0                     /* IEU0                         */
285         mov             %g5, %g4                        /* IEU1                         */
287 23:     srlx            %g4, 8, %g4                     /* IEU0         Group           */
288         andcc           %o2, 1, %g0                     /* IEU1                         */
289         bne,a,pn        %icc, 24f                       /* CTI                          */
290          stb            %g4, [%o0]                      /* Store        Group           */
292         retl                                            /* CTI+IEU1     Group           */
293          nop                                            /* IEU0                         */
294 24:     retl                                            /* CTI+IEU1     Group           */
295          add            %o0, 1, %o0                     /* IEU0                         */
297         .align          16
298 55:     sub             %o0, 1, %g6                     /* IEU0         Group           */
299 25:     andcc           %o0, 7, %g0                     /* IEU1                         */
300         be,a,pn         %icc, 4b                        /* CTI                          */
301          andncc         %o2, 31, %g3                    /* IEU1         Group           */
303         stb             %g0, [%o0]                      /* Store        Group           */
304         subcc           %o2, 1, %o2                     /* IEU1                         */
305         bne,pt          %XCC, 25b                       /* CTI                          */
306          add            %o0, 1, %o0                     /* IEU0         Group           */
308         retl                                            /* CTI+IEU1     Group           */
309          mov            %g6, %o0                        /* IEU0                         */
311         .align          16
312 26:     ldub            [%o1], %o3                      /* Load                         */
313         sllx            %g1, 7, %g2                     /* IEU0         Group           */
314         stb             %o3, [%o0]                      /* Store                        */
315 27:     subcc           %o2, 1, %o2                     /* IEU1                         */
317         be,pn           %XCC, 59b                       /* CTI                          */
318          add            %o1, 1, %o1                     /* IEU0         Group           */
319         add             %o0, 1, %o0                     /* IEU1                         */
320         andcc           %o3, 0xff, %g0                  /* IEU1         Group           */
322         be,pn           %icc, 55b                       /* CTI                          */
323          lduba          [%o1] ASI_PNF, %o3              /* Load                         */
324         andcc           %o0, 7, %g0                     /* IEU1         Group           */
325         bne,a,pt        %icc, 27b                       /* CTI                          */
327          stb            %o3, [%o0]                      /* Store                        */
328         andcc           %o1, 7, %g3                     /* IEU1         Group           */
329         be,a,pt         %icc, 1b                        /* CTI                          */
330          ldx            [%o1], %o3                      /* Load                         */
332 28:     orcc            %g0, 64, %g4                    /* IEU1         Group           */
333         sllx            %g3, 3, %g5                     /* IEU0                         */
334         sub             %g4, %g5, %g4                   /* IEU0         Group           */
335         sub             %o1, %g3, %o1                   /* IEU1                         */
336                                                         /* %g1 = 0101010101010101
337                                                            %g2 = 8080808080808080
338                                                            %g3 = source alignment
339                                                            %g5 = number of bits to shift left
340                                                            %g4 = number of bits to shift right */
342         ldxa            [%o1] ASI_PNF, %o5              /* Load         Group           */
343         addcc           %o1, 8, %o1                     /* IEU1                         */
344 29:     sllx            %o5, %g5, %o3                   /* IEU0         Group           */
345         ldxa            [%o1] ASI_PNF, %o5              /* Load                         */
347         subcc           %o2, 8, %o2                     /* IEU1                         */
348         bl,pn           %XCC, 17b                       /* CTI                          */
349          srlx           %o5, %g4, %o4                   /* IEU0         Group           */
350         add             %o1, 8, %o1                     /* IEU1                         */
352         or              %o3, %o4, %o3                   /* IEU0         Group           */
353         add             %o0, 8, %o0                     /* IEU1                         */
354         sub             %o3, %g1, %o4                   /* IEU0         Group           */
355 #ifdef EIGHTBIT_NOT_RARE
356         andn            %o4, %o3, %o4                   /* IEU0         Group           */
357 #endif
358         andcc           %o4, %g2, %g0                   /* IEU1         Group           */
360         be,a,pt         %xcc, 29b                       /* CTI                          */
361          stx            %o3, [%o0-8]                    /* Store                        */
362         srlx            %o3, 56, %o4                    /* IEU0         Group           */
363         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
365         be,pn           %icc, 36f                       /* CTI                          */
366          srlx           %o3, 48, %g6                    /* IEU0                         */
367         andcc           %g6, 0xff, %g0                  /* IEU1         Group           */
368         be,pn           %icc, 35f                       /* CTI                          */
370          srlx           %o3, 40, %o4                    /* IEU0                         */
371         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
372         be,pn           %icc, 34f                       /* CTI                          */
373          srlx           %o3, 32, %g6                    /* IEU0                         */
375         andcc           %g6, 0xff, %g0                  /* IEU1         Group           */
376         be,pn           %icc, 33f                       /* CTI                          */
377          srlx           %o3, 24, %o4                    /* IEU0                         */
378         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
380         be,pn           %icc, 32f                       /* CTI                          */
381          srlx           %o3, 16, %g6                    /* IEU0                         */
382         andcc           %g6, 0xff, %g0                  /* IEU1         Group           */
383         be,pn           %icc, 31f                       /* CTI                          */
385          srlx           %o3, 8, %o4                     /* IEU0                         */
386         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
387         be,pn           %icc, 30f                       /* CTI                          */
388          andcc          %o3, 0xff, %g0                  /* IEU1         Group           */
390         bne,pn          %icc, 29b                       /* CTI                          */
391          stx            %o3, [%o0-8]                    /* Store                        */
392         sub             %o0, 1, %g6                     /* IEU0         Group           */
393         ba,pt           %xcc, 4b                        /* CTI                          */
395          andncc         %o2, 31, %g3                    /* IEU1                         */
396 30:     subcc           %o0, 2, %g6                     /* IEU0                         */
397         ba,pt           %xcc, 3b                        /* CTI                          */
398          sllx           %o4, 8, %g3                     /* IEU0         Group           */
400 31:     sllx            %g6, 16, %g3                    /* IEU0         Group           */
401         ba,pt           %xcc, 3b                        /* CTI                          */
402          sub            %o0, 3, %g6                     /* IEU1                         */
403 32:     subcc           %o0, 4, %g6                     /* IEU1         Group           */
405         ba,pt           %xcc, 3b                        /* CTI                          */
406          sllx           %o4, 24, %g3                    /* IEU0                         */
407 33:     sllx            %g6, 32, %g3                    /* IEU0         Group           */
408         ba,pt           %xcc, 3b                        /* CTI                          */
410          sub            %o0, 5, %g6                     /* IEU1                         */
411 34:     subcc           %o0, 6, %g6                     /* IEU1         Group           */
412         ba,pt           %xcc, 3b                        /* CTI                          */
413          sllx           %o4, 40, %g3                    /* IEU0                         */
415 35:     sllx            %g6, 48, %g3                    /* IEU0         Group           */
416         ba,pt           %xcc, 3b                        /* CTI                          */
417          sub            %o0, 7, %g6                     /* IEU1                         */
418 36:     subcc           %o0, 8, %g6                     /* IEU1         Group           */
420         ba,pt           %xcc, 3b                        /* CTI                          */
421          sllx           %o4, 56, %g3                    /* IEU0                         */
422 END(__stpncpy)
424 libc_hidden_def (__stpncpy)
425 weak_alias (__stpncpy, stpncpy)