libc/string/sparc/sparc64/strcat.S

   1 /* strcat (dest, src) -- Append SRC on the end of DEST.
   2    For SPARC v9.
   3    Copyright (C) 1998, 1999, 2003 Free Software Foundation, Inc.
   4    This file is part of the GNU C Library.
   5    Contributed by Jakub Jelinek <jj@ultra.linux.cz> and
   6                   Jan Vondrak <jvon4518@ss1000.ms.mff.cuni.cz>.
   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    <http://www.gnu.org/licenses/>.  */
  21
  22 #include <asm/asi.h>
  23 #ifndef XCC
  24 #define XCC xcc
  25 #define USE_BPR
  26         .register       %g2, #scratch
  27         .register       %g3, #scratch
  28         .register       %g6, #scratch
  29 #endif
  30
  31         /* Normally, this uses
  32            ((xword - 0x0101010101010101) & 0x8080808080808080) test
  33            to find out if any byte in xword could be zero. This is fast, but
  34            also gives false alarm for any byte in range 0x81-0xff. It does
  35            not matter for correctness, as if this test tells us there could
  36            be some zero byte, we check it byte by byte, but if bytes with
  37            high bits set are common in the strings, then this will give poor
  38            performance. You can #define EIGHTBIT_NOT_RARE and the algorithm
  39            will use one tick slower, but more precise test
  40            ((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080),
  41            which does not give any false alarms (but if some bits are set,
  42            one cannot assume from it which bytes are zero and which are not).
  43            It is yet to be measured, what is the correct default for glibc
  44            in these days for an average user.
  45          */
  46
  47         .text
  48         .align          32
  49 ENTRY(strcat)
  50         sethi           %hi(0x01010101), %g1            /* IEU0         Group           */
  51         ldub            [%o0], %o3                      /* Load                         */
  52         or              %g1, %lo(0x01010101), %g1       /* IEU0         Group           */
  53         mov             %o0, %g6                        /* IEU1                         */
  54
  55         sllx            %g1, 32, %g2                    /* IEU0         Group           */
  56         andcc           %o0, 7, %g0                     /* IEU1                         */
  57         or              %g1, %g2, %g1                   /* IEU0         Group           */
  58         bne,pn          %icc, 32f                       /* CTI                          */
  59
  60          sllx           %g1, 7, %g2                     /* IEU0         Group           */
  61         brz,pn          %o3, 30f                        /* CTI+IEU1                     */
  62          ldx            [%o0], %o3                      /* Load                         */
  63 48:     add             %o0, 8, %o0                     /* IEU0         Group           */
  64
  65 49:     sub             %o3, %g1, %o2                   /* IEU0         Group           */
  66 #ifdef EIGHTBIT_NOT_RARE
  67         andn            %o2, %o3, %g5                   /* IEU0         Group           */
  68         ldxa            [%o0] ASI_PNF, %o3              /* Load                         */
  69         andcc           %g5, %g2, %g0                   /* IEU1         Group           */
  70 #else
  71         ldxa            [%o0] ASI_PNF, %o3              /* Load                         */
  72         andcc           %o2, %g2, %g0                   /* IEU1         Group           */
  73 #endif
  74         be,pt           %xcc, 49b                       /* CTI                          */
  75
  76          add            %o0, 8, %o0                     /* IEU0                         */
  77         addcc           %o2, %g1, %g3                   /* IEU1         Group           */
  78         srlx            %o2, 32, %o2                    /* IEU0                         */
  79 50:     andcc           %o2, %g2, %g0                   /* IEU1         Group           */
  80
  81         be,pn           %xcc, 51f                       /* CTI                          */
  82          srlx           %g3, 56, %o2                    /* IEU0                         */
  83         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
  84         be,pn           %icc, 29f                       /* CTI                          */
  85
  86          srlx           %g3, 48, %o2                    /* IEU0                         */
  87         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
  88         be,pn           %icc, 28f                       /* CTI                          */
  89          srlx           %g3, 40, %o2                    /* IEU0                         */
  90
  91         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
  92         be,pn           %icc, 27f                       /* CTI                          */
  93          srlx           %g3, 32, %o2                    /* IEU0                         */
  94         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
  95
  96         be,pn           %icc, 26f                       /* CTI                          */
  97 51:      srlx           %g3, 24, %o2                    /* IEU0                         */
  98         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
  99         be,pn           %icc, 25f                       /* CTI                          */
 100
 101          srlx           %g3, 16, %o2                    /* IEU0                         */
 102         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
 103         be,pn           %icc, 24f                       /* CTI                          */
 104          srlx           %g3, 8, %o2                     /* IEU0                         */
 105
 106         andcc           %o2, 0xff, %g0                  /* IEU1         Group           */
 107         be,pn           %icc, 23f                       /* CTI                          */
 108          sub            %o3, %g1, %o2                   /* IEU0                         */
 109         andcc           %g3, 0xff, %g0                  /* IEU1         Group           */
 110
 111         be,pn           %icc, 52f                       /* CTI                          */
 112          ldxa           [%o0] ASI_PNF, %o3              /* Load                         */
 113         andcc           %o2, %g2, %g0                   /* IEU1         Group           */
 114         be,pt           %xcc, 49b                       /* CTI                          */
 115
 116          add            %o0, 8, %o0                     /* IEU0                         */
 117         addcc           %o2, %g1, %g3                   /* IEU1         Group           */
 118         ba,pt           %xcc, 50b                       /* CTI                          */
 119          srlx           %o2, 32, %o2                    /* IEU0                         */
 120
 121         .align          16
 122 52:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 123          add            %o0, -9, %o0                    /* IEU0                         */
 124 23:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 125          add            %o0, -10, %o0                   /* IEU0                         */
 126
 127 24:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 128          add            %o0, -11, %o0                   /* IEU0                         */
 129 25:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 130          add            %o0, -12, %o0                   /* IEU0                         */
 131
 132 26:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 133          add            %o0, -13, %o0                   /* IEU0                         */
 134 27:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 135          add            %o0, -14, %o0                   /* IEU0                         */
 136
 137 28:     ba,pt           %xcc, 12f                       /* CTI          Group           */
 138          add            %o0, -15, %o0                   /* IEU0                         */
 139 29:     add             %o0, -16, %o0                   /* IEU0         Group           */
 140 30:     andcc           %o1, 7, %g3                     /* IEU1                         */
 141
 142 31:     bne,pn          %icc, 14f                       /* CTI                          */
 143          orcc           %g0, 64, %g4                    /* IEU1         Group           */
 144 1:      ldx             [%o1], %o3                      /* Load                         */
 145         add             %o1, 8, %o1                     /* IEU1                         */
 146
 147 2:      mov             %o3, %g3                        /* IEU0         Group           */
 148 3:      sub             %o3, %g1, %o2                   /* IEU1                         */
 149         ldxa            [%o1] ASI_PNF, %o3              /* Load                         */
 150 #ifdef EIGHTBIT_NOT_RARE
 151         andn            %o2, %g3, %o2                   /* IEU0         Group           */
 152 #endif
 153         add             %o0, 8, %o0                     /* IEU0         Group           */
 154
 155         andcc           %o2, %g2, %g0                   /* IEU1                         */
 156         add             %o1, 8, %o1                     /* IEU0         Group           */
 157         be,a,pt         %xcc, 2b                        /* CTI                          */
 158          stx            %g3, [%o0 - 8]                  /* Store                        */
 159
 160         srlx            %g3, 56, %g5                    /* IEU0         Group           */
 161         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 162         be,pn           %icc, 11f                       /* CTI                          */
 163          srlx           %g3, 48, %g4                    /* IEU0                         */
 164
 165         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
 166         be,pn           %icc, 10f                       /* CTI                          */
 167          srlx           %g3, 40, %g5                    /* IEU0                         */
 168         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 169
 170         be,pn           %icc, 9f                        /* CTI                          */
 171          srlx           %g3, 32, %g4                    /* IEU0                         */
 172         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
 173         be,pn           %icc, 8f                        /* CTI                          */
 174
 175          srlx           %g3, 24, %g5                    /* IEU0                         */
 176         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 177         be,pn           %icc, 7f                        /* CTI                          */
 178          srlx           %g3, 16, %g4                    /* IEU0                         */
 179
 180         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
 181         be,pn           %icc, 6f                        /* CTI                          */
 182          srlx           %g3, 8, %g5                     /* IEU0                         */
 183         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 184
 185         be,pn           %icc, 5f                        /* CTI                          */
 186          sub            %o3, %g1, %o2                   /* IEU0                         */
 187         stx             %g3, [%o0 - 8]                  /* Store        Group           */
 188         andcc           %g3, 0xff, %g0                  /* IEU1                         */
 189
 190         bne,pt          %icc, 3b                        /* CTI                          */
 191          mov            %o3, %g3                        /* IEU0         Group           */
 192 4:      retl                                            /* CTI+IEU1     Group           */
 193          mov            %g6, %o0                        /* IEU0                         */
 194
 195         .align          16
 196 5:      stb             %g5, [%o0 - 2]                  /* Store        Group           */
 197         srlx            %g3, 16, %g4                    /* IEU0                         */
 198 6:      sth             %g4, [%o0 - 4]                  /* Store        Group           */
 199         srlx            %g3, 32, %g4                    /* IEU0                         */
 200
 201         stw             %g4, [%o0 - 8]                  /* Store        Group           */
 202         retl                                            /* CTI+IEU1     Group           */
 203          mov            %g6, %o0                        /* IEU0                         */
 204 7:      stb             %g5, [%o0 - 4]                  /* Store        Group           */
 205
 206         srlx            %g3, 32, %g4                    /* IEU0                         */
 207 8:      stw             %g4, [%o0 - 8]                  /* Store        Group           */
 208         retl                                            /* CTI+IEU1     Group           */
 209          mov            %g6, %o0                        /* IEU0                         */
 210
 211 9:      stb             %g5, [%o0 - 6]                  /* Store        Group           */
 212         srlx            %g3, 48, %g4                    /* IEU0                         */
 213 10:     sth             %g4, [%o0 - 8]                  /* Store        Group           */
 214         retl                                            /* CTI+IEU1     Group           */
 215
 216          mov            %g6, %o0                        /* IEU0                         */
 217 11:     stb             %g5, [%o0 - 8]                  /* Store        Group           */
 218         retl                                            /* CTI+IEU1     Group           */
 219          mov            %g6, %o0                        /* IEU0                         */
 220
 221         .align          16
 222 32:     andcc           %o0, 7, %g0                     /* IEU1         Group           */
 223         be,a,pn         %icc, 48b                       /* CTI                          */
 224          ldx            [%o0], %o3                      /* Load                         */
 225         add             %o0, 1, %o0                     /* IEU0         Group           */
 226
 227         brnz,a,pt       %o3, 32b                        /* CTI+IEU1                     */
 228          lduba          [%o0] ASI_PNF, %o3              /* Load                         */
 229         add             %o0, -1, %o0                    /* IEU0         Group           */
 230         andcc           %o0, 7, %g0                     /* IEU1         Group           */
 231
 232         be,a,pn         %icc, 31b                       /* CTI                          */
 233          andcc          %o1, 7, %g3                     /* IEU1         Group           */
 234 12:     ldub            [%o1], %o3                      /* Load                         */
 235         stb             %o3, [%o0]                      /* Store        Group           */
 236
 237 13:     add             %o0, 1, %o0                     /* IEU0                         */
 238         add             %o1, 1, %o1                     /* IEU1                         */
 239         andcc           %o3, 0xff, %g0                  /* IEU1         Group           */
 240         be,pn           %icc, 4b                        /* CTI                          */
 241
 242          lduba          [%o1] ASI_PNF, %o3              /* Load                         */
 243         andcc           %o0, 7, %g0                     /* IEU1         Group           */
 244         bne,a,pt        %icc, 13b                       /* CTI                          */
 245          stb            %o3, [%o0]                      /* Store                        */
 246
 247         andcc           %o1, 7, %g3                     /* IEU1         Group           */
 248         be,a,pt         %icc, 1b                        /* CTI                          */
 249          ldx            [%o1], %o3                      /* Load                         */
 250         orcc            %g0, 64, %g4                    /* IEU1         Group           */
 251
 252 14:     sllx            %g3, 3, %g5                     /* IEU0                         */
 253         sub             %o1, %g3, %o1                   /* IEU0         Group           */
 254         sub             %g4, %g5, %g4                   /* IEU1                         */
 255                                                         /* %g1 = 0101010101010101       *
 256                                                          * %g2 = 8080808080808080       *
 257                                                          * %g3 = source alignment       *
 258                                                          * %g5 = number of bits to shift left  *
 259                                                          * %g4 = number of bits to shift right */
 260         ldxa            [%o1] ASI_PNF, %o5              /* Load         Group           */
 261
 262         addcc           %o1, 8, %o1                     /* IEU1                         */
 263 15:     sllx            %o5, %g5, %o3                   /* IEU0         Group           */
 264         ldxa            [%o1] ASI_PNF, %o5              /* Load                         */
 265         srlx            %o5, %g4, %o4                   /* IEU0         Group           */
 266
 267         add             %o0, 8, %o0                     /* IEU1                         */
 268         or              %o3, %o4, %o3                   /* IEU0         Group           */
 269         add             %o1, 8, %o1                     /* IEU1                         */
 270         sub             %o3, %g1, %o4                   /* IEU0         Group           */
 271
 272 #ifdef EIGHTBIT_NOT_RARE
 273         andn            %o4, %o3, %o4                   /* IEU0         Group           */
 274 #endif
 275         andcc           %o4, %g2, %g0                   /* IEU1         Group           */
 276         be,a,pt         %xcc, 15b                       /* CTI                          */
 277          stx            %o3, [%o0 - 8]                  /* Store                        */
 278         srlx            %o3, 56, %o4                    /* IEU0         Group           */
 279
 280         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 281         be,pn           %icc, 22f                       /* CTI                          */
 282          srlx           %o3, 48, %o4                    /* IEU0                         */
 283         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 284
 285         be,pn           %icc, 21f                       /* CTI                          */
 286          srlx           %o3, 40, %o4                    /* IEU0                         */
 287         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 288         be,pn           %icc, 20f                       /* CTI                          */
 289
 290          srlx           %o3, 32, %o4                    /* IEU0                         */
 291         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 292         be,pn           %icc, 19f                       /* CTI                          */
 293          srlx           %o3, 24, %o4                    /* IEU0                         */
 294
 295         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 296         be,pn           %icc, 18f                       /* CTI                          */
 297          srlx           %o3, 16, %o4                    /* IEU0                         */
 298         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 299
 300         be,pn           %icc, 17f                       /* CTI                          */
 301          srlx           %o3, 8, %o4                     /* IEU0                         */
 302         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 303         be,pn           %icc, 16f                       /* CTI                          */
 304
 305          andcc          %o3, 0xff, %g0                  /* IEU1         Group           */
 306         bne,pn          %icc, 15b                       /* CTI                          */
 307          stx            %o3, [%o0 - 8]                  /* Store                        */
 308         retl                                            /* CTI+IEU1     Group           */
 309
 310          mov            %g6, %o0                        /* IEU0                         */
 311
 312         .align          16
 313 16:     srlx            %o3, 8, %o4                     /* IEU0         Group           */
 314         stb             %o4, [%o0 - 2]                  /* Store                        */
 315 17:     srlx            %o3, 16, %o4                    /* IEU0         Group           */
 316         stb             %o4, [%o0 - 3]                  /* Store                        */
 317
 318 18:     srlx            %o3, 24, %o4                    /* IEU0         Group           */
 319         stb             %o4, [%o0 - 4]                  /* Store                        */
 320 19:     srlx            %o3, 32, %o4                    /* IEU0         Group           */
 321         stw             %o4, [%o0 - 8]                  /* Store                        */
 322
 323         retl                                            /* CTI+IEU1     Group           */
 324          mov            %g6, %o0                        /* IEU0                         */
 325         nop
 326         nop
 327
 328 20:     srlx            %o3, 40, %o4                    /* IEU0         Group           */
 329         stb             %o4, [%o0 - 6]                  /* Store                        */
 330 21:     srlx            %o3, 48, %o4                    /* IEU0         Group           */
 331         stb             %o4, [%o0 - 7]                  /* Store                        */
 332
 333 22:     srlx            %o3, 56, %o4                    /* IEU0         Group           */
 334         stb             %o4, [%o0 - 8]                  /* Store                        */
 335         retl                                            /* CTI+IEU1     Group           */
 336          mov            %g6, %o0                        /* IEU0                         */
 337 END(strcat)
 338 libc_hidden_def(strcat)