sysdeps/sparc/sparc64/strncpy.S

   1 /* strncpy(DST, SRC, COUNT) - Copy no more than COUNT bytes of the
   2    null-terminated string from SRC to DST.  If SRC does not cover all of
   3    COUNT, the balance is zeroed.
   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(strncpy)
  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          mov            %o0, %g6                        /* IEU0         Group           */
  59         or              %g1, %lo(0x01010101), %g1       /* IEU1                         */
  60
  61         andcc           %o0, 7, %g0                     /* IEU1         Group           */
  62         sllx            %g1, 32, %g2                    /* IEU0                         */
  63         bne,pn          %icc, 26f                       /* CTI                          */
  64          or             %g1, %g2, %g1                   /* IEU0         Group           */
  65
  66         andcc           %o1, 7, %g3                     /* IEU1                         */
  67         bne,pn          %icc, 28f                       /* CTI                          */
  68          sllx           %g1, 7, %g2                     /* IEU0         Group           */
  69         ldx             [%o1], %o3                      /* Load                         */
  70
  71 1:      add             %o1, 8, %o1                     /* IEU1                         */
  72 2:      subcc           %o2, 8, %o2                     /* IEU1         Group           */
  73         bl,pn           %XCC, 18f                       /* CTI                          */
  74          sub            %o3, %g1, %o4                   /* IEU0                         */
  75
  76         add             %o0, 8, %o0                     /* IEU0         Group           */
  77 #ifdef EIGHTBIT_NOT_MORE
  78         andn            %o4, %o3, %o4                   /* IEU1                         */
  79 #endif
  80         mov             %o3, %g3                        /* IEU1                         */
  81         ldxa            [%o1] ASI_PNF, %o3              /* Load                         */
  82         add             %o1, 8, %o1                     /* IEU0         Group           */
  83
  84         andcc           %o4, %g2, %g0                   /* IEU1                         */
  85         be,a,pt         %xcc, 2b                        /* CTI                          */
  86          stx            %g3, [%o0-8]                    /* Store        Group           */
  87         srlx            %g3, 56, %g5                    /* IEU0         Group           */
  88
  89         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
  90         be,pn           %icc, 16f                       /* CTI                          */
  91          srlx           %g3, 48, %g4                    /* IEU0                         */
  92         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
  93
  94         be,pn           %icc, 15f                       /* CTI                          */
  95          srlx           %g3, 40, %g5                    /* IEU0                         */
  96         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
  97         be,pn           %icc, 14f                       /* CTI                          */
  98
  99          srlx           %g3, 32, %g4                    /* IEU0                         */
 100         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
 101         be,pn           %icc, 13f                       /* CTI                          */
 102          srlx           %g3, 24, %g5                    /* IEU0                         */
 103
 104         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 105         be,pn           %icc, 12f                       /* CTI                          */
 106          srlx           %g3, 16, %g4                    /* IEU0                         */
 107         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
 108
 109         be,pn           %icc, 11f                       /* CTI                          */
 110          srlx           %g3, 8, %g5                     /* IEU0                         */
 111         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 112         be,pn           %icc, 10f                       /* CTI                          */
 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
 158         .align          16
 159 10:     ba,pt           %xcc, 3b                        /* CTI                          */
 160          sllx           %g5, 8, %g3                     /* IEU0                         */
 161 11:     ba,pt           %xcc, 3b                        /* CTI          Group           */
 162          sllx           %g4, 16, %g3                    /* IEU0                         */
 163
 164 12:     ba,pt           %xcc, 3b                        /* CTI          Group           */
 165          sllx           %g5, 24, %g3                    /* IEU0                         */
 166 13:     ba,pt           %xcc, 3b                        /* CTI          Group           */
 167          sllx           %g4, 32, %g3                    /* IEU0                         */
 168
 169 14:     ba,pt           %xcc, 3b                        /* CTI          Group           */
 170          sllx           %g5, 40, %g3                    /* IEU0                         */
 171 15:     ba,pt           %xcc, 3b                        /* CTI          Group           */
 172          sllx           %g4, 48, %g3                    /* IEU0                         */
 173
 174 16:     ba,pt           %xcc, 3b                        /* CTI                          */
 175          sllx           %g5, 56, %g3                    /* IEU0                         */
 176 17:     or              %o3, %o4, %o3                   /* IEU0         Group           */
 177         sub             %o3, %g1, %o4                   /* IEU1                         */
 178
 179 18:     addcc           %o2, 8, %o2                     /* IEU1         Group           */
 180         be,pn           %XCC, 19f                       /* CTI                          */
 181          andcc          %o4, %g2, %g0                   /* IEU1         Group           */
 182         be,pt           %xcc, 21f                       /* CTI                          */
 183
 184          srlx           %o3, 56, %g5                    /* IEU0                         */
 185         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 186         be,pn           %icc, 20f                       /* CTI                          */
 187          stb            %g5, [%o0]                      /* Store                        */
 188
 189         add             %o0, 1, %o0                     /* IEU0         Group           */
 190         subcc           %o2, 1, %o2                     /* IEU1                         */
 191         be,pn           %XCC, 19f                       /* CTI                          */
 192          srlx           %o3, 48, %g5                    /* IEU0         Group           */
 193
 194         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 195         be,pn           %icc, 20f                       /* CTI                          */
 196          stb            %g5, [%o0]                      /* Store                        */
 197         add             %o0, 1, %o0                     /* IEU0         Group           */
 198
 199         subcc           %o2, 1, %o2                     /* IEU1                         */
 200         be,pn           %XCC, 19f                       /* CTI                          */
 201          srlx           %o3, 40, %g5                    /* IEU0         Group           */
 202         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 203
 204         be,pn           %icc, 20f                       /* CTI                          */
 205          stb            %g5, [%o0]                      /* Store                        */
 206         add             %o0, 1, %o0                     /* IEU0         Group           */
 207         subcc           %o2, 1, %o2                     /* IEU1                         */
 208
 209         be,pn           %XCC, 19f                       /* CTI                          */
 210          srlx           %o3, 32, %g5                    /* IEU0         Group           */
 211         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 212         be,pn           %icc, 20f                       /* CTI                          */
 213
 214          stb            %g5, [%o0]                      /* Store                        */
 215         add             %o0, 1, %o0                     /* IEU0         Group           */
 216         subcc           %o2, 1, %o2                     /* IEU1                         */
 217         be,pn           %XCC, 19f                       /* CTI                          */
 218
 219          srlx           %o3, 24, %g5                    /* IEU0         Group           */
 220         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 221         be,pn           %icc, 20f                       /* CTI                          */
 222          stb            %g5, [%o0]                      /* Store                        */
 223
 224         add             %o0, 1, %o0                     /* IEU0         Group           */
 225         subcc           %o2, 1, %o2                     /* IEU1                         */
 226         be,pn           %XCC, 19f                       /* CTI                          */
 227          srlx           %o3, 16, %g5                    /* IEU0         Group           */
 228
 229         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 230         be,pn           %icc, 20f                       /* CTI                          */
 231          stb            %g5, [%o0]                      /* Store                        */
 232         add             %o0, 1, %o0                     /* IEU0         Group           */
 233
 234         subcc           %o2, 1, %o2                     /* IEU1                         */
 235         be,pn           %XCC, 19f                       /* CTI                          */
 236          srlx           %o3, 8, %g5                     /* IEU0         Group           */
 237         stb             %g5, [%o0]                      /* Store                        */
 238
 239 19:     retl                                            /* CTI+IEU1     Group           */
 240          mov            %g6, %o0                        /* IEU0                         */
 241 50:     stb             %g0, [%o0]                      /* Store        Group           */
 242 20:     subcc           %o2, 1, %o2                     /* IEU1         Group           */
 243
 244         bne,pt          %XCC, 50b                       /* CTI                          */
 245          add            %o0, 1, %o0                     /* IEU0                         */
 246         retl                                            /* CTI+IEU1     Group           */
 247          mov            %g6, %o0                        /* IEU0                         */
 248
 249 21:     andcc           %o2, 4, %g0                     /* IEU1         Group           */
 250         be,pn           %icc, 22f                       /* CTI                          */
 251          srlx           %o3, 32, %g5                    /* IEU0                         */
 252         stw             %g5, [%o0]                      /* Store        Group           */
 253
 254         add             %o0, 4, %o0                     /* IEU0                         */
 255         mov             %o3, %g5                        /* IEU1                         */
 256 22:     andcc           %o2, 2, %g0                     /* IEU1         Group           */
 257         be,pn           %icc, 23f                       /* CTI                          */
 258
 259          srlx           %g5, 16, %g4                    /* IEU0                         */
 260         sth             %g4, [%o0]                      /* Store        Group           */
 261         add             %o0, 2, %o0                     /* IEU0                         */
 262         mov             %g5, %g4                        /* IEU1                         */
 263
 264 23:     srlx            %g4, 8, %g4                     /* IEU0         Group           */
 265         andcc           %o2, 1, %g0                     /* IEU1                         */
 266         bne,a,pn        %icc, 24f                       /* CTI                          */
 267          stb            %g4, [%o0]                      /* Store        Group           */
 268
 269 24:     retl                                            /* CTI+IEU1     Group           */
 270          mov            %g6, %o0                        /* IEU0                         */
 271 25:     andcc           %o0, 7, %g0                     /* IEU1         Group           */
 272         be,a,pn         %icc, 4b                        /* CTI                          */
 273
 274          andncc         %o2, 31, %g3                    /* IEU1         Group           */
 275         stb             %g0, [%o0]                      /* Store        Group           */
 276         subcc           %o2, 1, %o2                     /* IEU1                         */
 277         bne,pt          %XCC, 25b                       /* CTI                          */
 278
 279          add            %o0, 1, %o0                     /* IEU0         Group           */
 280         retl                                            /* CTI+IEU1     Group           */
 281          mov            %g6, %o0                        /* IEU0                         */
 282
 283         .align          16
 284 26:     ldub            [%o1], %o3                      /* Load                         */
 285         sllx            %g1, 7, %g2                     /* IEU0         Group           */
 286         stb             %o3, [%o0]                      /* Store                        */
 287 27:     subcc           %o2, 1, %o2                     /* IEU1                         */
 288
 289         be,pn           %XCC, 9b                        /* CTI                          */
 290          add            %o1, 1, %o1                     /* IEU0         Group           */
 291         add             %o0, 1, %o0                     /* IEU1                         */
 292         andcc           %o3, 0xff, %g0                  /* IEU1         Group           */
 293
 294         be,pn           %icc, 25b                       /* CTI                          */
 295          lduba          [%o1] ASI_PNF, %o3              /* Load                         */
 296         andcc           %o0, 7, %g0                     /* IEU1         Group           */
 297         bne,a,pt        %icc, 27b                       /* CTI                          */
 298
 299          stb            %o3, [%o0]                      /* Store                        */
 300         andcc           %o1, 7, %g3                     /* IEU1         Group           */
 301         be,a,pt         %icc, 1b                        /* CTI                          */
 302          ldx            [%o1], %o3                      /* Load                         */
 303
 304 28:     orcc            %g0, 64, %g4                    /* IEU1         Group           */
 305         sllx            %g3, 3, %g5                     /* IEU0                         */
 306         sub             %g4, %g5, %g4                   /* IEU0         Group           */
 307         sub             %o1, %g3, %o1                   /* IEU1                         */
 308                                                         /* %g1 = 0101010101010101
 309                                                            %g2 = 8080808080808080
 310                                                            %g3 = source alignment
 311                                                            %g5 = number of bits to shift left
 312                                                            %g4 = number of bits to shift right */
 313
 314         ldxa            [%o1] ASI_PNF, %o5              /* Load         Group           */
 315         addcc           %o1, 8, %o1                     /* IEU1                         */
 316
 317 29:     sllx            %o5, %g5, %o3                   /* IEU0         Group           */
 318         ldxa            [%o1] ASI_PNF, %o5              /* Load                         */
 319         subcc           %o2, 8, %o2                     /* IEU1                         */
 320         bl,pn           %XCC, 17b                       /* CTI                          */
 321
 322          srlx           %o5, %g4, %o4                   /* IEU0         Group           */
 323         add             %o1, 8, %o1                     /* IEU1                         */
 324         or              %o3, %o4, %o3                   /* IEU0         Group           */
 325         add             %o0, 8, %o0                     /* IEU1                         */
 326
 327         sub             %o3, %g1, %o4                   /* IEU0         Group           */
 328 #ifdef EIGHTBIT_NOT_RARE
 329         andn            %o4, %o3, %o4                   /* IEU0         Group           */
 330 #endif
 331         andcc           %o4, %g2, %g0                   /* IEU1         Group           */
 332         be,a,pt         %xcc, 29b                       /* CTI                          */
 333          stx            %o3, [%o0-8]                    /* Store                        */
 334
 335         srlx            %o3, 56, %o4                    /* IEU0         Group           */
 336         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 337         be,pn           %icc, 36f                       /* CTI                          */
 338          srlx           %o3, 48, %o4                    /* IEU0                         */
 339
 340         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 341         be,pn           %icc, 35f                       /* CTI                          */
 342          srlx           %o3, 40, %o4                    /* IEU0                         */
 343         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 344
 345         be,pn           %icc, 34f                       /* CTI                          */
 346          srlx           %o3, 32, %o4                    /* IEU0                         */
 347         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 348         be,pn           %icc, 33f                       /* CTI                          */
 349
 350          srlx           %o3, 24, %o4                    /* IEU0                         */
 351         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 352         be,pn           %icc, 32f                       /* CTI                          */
 353          srlx           %o3, 16, %o4                    /* IEU0                         */
 354
 355         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 356         be,pn           %icc, 31f                       /* CTI                          */
 357          srlx           %o3, 8, %o4                     /* IEU0                         */
 358         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 359
 360         be,pn           %icc, 30f                       /* CTI                          */
 361          andcc          %o3, 0xff, %g0                  /* IEU1         Group           */
 362         bne,pn          %icc, 29b                       /* CTI                          */
 363          stx            %o3, [%o0-8]                    /* Store                        */
 364
 365         ba,pt           %xcc, 4b                        /* CTI          Group           */
 366          andncc         %o2, 31, %g3                    /* IEU1                         */
 367 30:     srlx            %o3, 8, %o4                     /* IEU0                         */
 368         ba,pt           %xcc, 3b                        /* CTI                          */
 369
 370          sllx           %o4, 8, %g3                     /* IEU0         Group           */
 371 31:     srlx            %o3, 16, %o4                    /* IEU0         Group           */
 372         ba,pt           %xcc, 3b                        /* CTI                          */
 373          sllx           %o4, 16, %g3                    /* IEU0         Group           */
 374
 375 32:     srlx            %o3, 24, %o4                    /* IEU0         Group           */
 376         ba,pt           %xcc, 3b                        /* CTI                          */
 377          sllx           %o4, 24, %g3                    /* IEU0         Group           */
 378 33:     srlx            %o3, 32, %o4                    /* IEU0         Group           */
 379
 380         ba,pt           %xcc, 3b                        /* CTI                          */
 381          sllx           %o4, 32, %g3                    /* IEU0         Group           */
 382 34:     srlx            %o3, 40, %o4                    /* IEU0         Group           */
 383         ba,pt           %xcc, 3b                        /* CTI                          */
 384
 385          sllx           %o4, 40, %g3                    /* IEU0         Group           */
 386 35:     srlx            %o3, 48, %o4                    /* IEU0         Group           */
 387         ba,pt           %xcc, 3b                        /* CTI                          */
 388          sllx           %o4, 48, %g3                    /* IEU0         Group           */
 389
 390 36:     srlx            %o3, 56, %o4                    /* IEU0         Group           */
 391         ba,pt           %xcc, 3b                        /* CTI                          */
 392          sllx           %o4, 56, %g3                    /* IEU0         Group           */
 393 END(strncpy)
 394 libc_hidden_builtin_def (strncpy)