sysdeps/sparc/sparc64/memchr.S

   1 /* memchr (str, ch, n) -- Return pointer to first occurrence of CH in STR less
   2    than N.
   3    For SPARC v9.
   4    Copyright (C) 1998, 1999, 2000, 2003 Free Software Foundation, Inc.
   5    This file is part of the GNU C Library.
   6    Contributed by Jan Vondrak <jvon4518@ss1000.ms.mff.cuni.cz> and
   7                   Jakub Jelinek <jj@ultra.linux.cz>.
   8    This version is developed using the same algorithm as the fast C
   9    version which carries the following introduction:
  10    Based on strlen implementation by Torbjorn Granlund (tege@sics.se),
  11    with help from Dan Sahlin (dan@sics.se) and
  12    commentary by Jim Blandy (jimb@ai.mit.edu);
  13    adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu),
  14    and implemented by Roland McGrath (roland@ai.mit.edu).
  15
  16    The GNU C Library is free software; you can redistribute it and/or
  17    modify it under the terms of the GNU Lesser General Public
  18    License as published by the Free Software Foundation; either
  19    version 2.1 of the License, or (at your option) any later version.
  20
  21    The GNU C Library is distributed in the hope that it will be useful,
  22    but WITHOUT ANY WARRANTY; without even the implied warranty of
  23    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  24    Lesser General Public License for more details.
  25
  26    You should have received a copy of the GNU Lesser General Public
  27    License along with the GNU C Library; if not, write to the Free
  28    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  29    02111-1307 USA.  */
  30
  31 #include <sysdep.h>
  32 #include <asm/asi.h>
  33 #ifndef XCC
  34 #define XCC xcc
  35 #define USE_BPR
  36         .register       %g2, #scratch
  37         .register       %g3, #scratch
  38 #endif
  39
  40         /* Normally, this uses
  41            ((xword - 0x0101010101010101) & 0x8080808080808080) test
  42            to find out if any byte in xword could be zero. This is fast, but
  43            also gives false alarm for any byte in range 0x81-0xff. It does
  44            not matter for correctness, as if this test tells us there could
  45            be some zero byte, we check it byte by byte, but if bytes with
  46            high bits set are common in the strings, then this will give poor
  47            performance. You can #define EIGHTBIT_NOT_RARE and the algorithm
  48            will use one tick slower, but more precise test
  49            ((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080),
  50            which does not give any false alarms (but if some bits are set,
  51            one cannot assume from it which bytes are zero and which are not).
  52            It is yet to be measured, what is the correct default for glibc
  53            in these days for an average user.
  54          */
  55
  56         .text
  57         .align          32
  58 ENTRY(__memchr)
  59         and             %o1, 0xff, %o1                  /* IEU0         Group           */
  60 #ifdef USE_BPR
  61         brz,pn          %o2, 12f                        /* CTI+IEU1                     */
  62 #else
  63         tst             %o2                             /* IEU1                         */
  64         be,pn           %XCC, 12f                       /* CTI                          */
  65 #endif
  66          sll            %o1, 8, %g3                     /* IEU0         Group           */
  67         add             %o0, %o2, %o2                   /* IEU1                         */
  68
  69         sethi           %hi(0x01010101), %g1            /* IEU0         Group           */
  70         or              %g3, %o1, %g3                   /* IEU1                         */
  71         ldub            [%o0], %o3                      /* Load                         */
  72         sllx            %g3, 16, %g5                    /* IEU0         Group           */
  73
  74         or              %g1, %lo(0x01010101), %g1       /* IEU1                         */
  75         sllx            %g1, 32, %g2                    /* IEU0         Group           */
  76         or              %g3, %g5, %g3                   /* IEU1                         */
  77         sllx            %g3, 32, %g5                    /* IEU0         Group           */
  78
  79         cmp             %o3, %o1                        /* IEU1                         */
  80         be,pn           %xcc, 13f                       /* CTI                          */
  81          or             %g1, %g2, %g1                   /* IEU0         Group           */
  82         andcc           %o0, 7, %g0                     /* IEU1                         */
  83
  84         bne,a,pn        %icc, 21f                       /* CTI                          */
  85          add            %o0, 1, %o0                     /* IEU0         Group           */
  86         ldx             [%o0], %o3                      /* Load         Group           */
  87         sllx            %g1, 7, %g2                     /* IEU0                         */
  88
  89         or              %g3, %g5, %g3                   /* IEU1                         */
  90 1:      add             %o0, 8, %o0                     /* IEU0         Group           */
  91         xor             %o3, %g3, %o4                   /* IEU1                         */
  92                                                         /* %g1 = 0101010101010101       *
  93                                                          * %g2 = 8080088080808080       *
  94                                                          * %g3 =  c c c c c c c c       *
  95                                                          * %o3 =      value             *
  96                                                          * %o4 =   value XOR c          */
  97 2:      cmp             %o0, %o2                        /* IEU1         Group           */
  98
  99         bg,pn           %XCC, 11f                       /* CTI                          */
 100          ldxa           [%o0] ASI_PNF, %o3              /* Load                         */
 101         sub             %o4, %g1, %o5                   /* IEU0         Group           */
 102         add             %o0, 8, %o0                     /* IEU1                         */
 103 #ifdef EIGHTBIT_NOT_RARE
 104         andn            %o5, %o4, %o5                   /* IEU0         Group           */
 105 #endif
 106
 107         andcc           %o5, %g2, %g0                   /* IEU1         Group           */
 108         be,a,pt         %xcc, 2b                        /* CTI                          */
 109          xor            %o3, %g3, %o4                   /* IEU0                         */
 110         srlx            %o4, 56, %g5                    /* IEU0                         */
 111
 112         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 113         be,pn           %icc, 3f                        /* CTI                          */
 114          srlx           %o4, 48, %g5                    /* IEU0                         */
 115         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 116
 117         be,pn           %icc, 4f                        /* CTI                          */
 118          srlx           %o4, 40, %g5                    /* IEU0                         */
 119         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 120         be,pn           %icc, 5f                        /* CTI                          */
 121
 122          srlx           %o4, 32, %g5                    /* IEU0                         */
 123         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 124         be,pn           %icc, 6f                        /* CTI                          */
 125          srlx           %o4, 24, %g5                    /* IEU0                         */
 126
 127         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 128         be,pn           %icc, 7f                        /* CTI                          */
 129          srlx           %o4, 16, %g5                    /* IEU0                         */
 130         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 131
 132         be,pn           %icc, 8f                        /* CTI                          */
 133          srlx           %o4, 8, %g5                     /* IEU0                         */
 134         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 135         be,pn           %icc, 9f                        /* CTI                          */
 136
 137          andcc          %o4, 0xff, %g0                  /* IEU1         Group           */
 138         bne,pt          %icc, 2b                        /* CTI                          */
 139          xor            %o3, %g3, %o4                   /* IEU0                         */
 140         retl                                            /* CTI+IEU1     Group           */
 141
 142          add            %o0, -9, %o0                    /* IEU0                         */
 143
 144         .align          16
 145 3:      retl                                            /* CTI+IEU1     Group           */
 146          add            %o0, -16, %o0                   /* IEU0                         */
 147 4:      retl                                            /* CTI+IEU1     Group           */
 148          add            %o0, -15, %o0                   /* IEU0                         */
 149
 150 5:      retl                                            /* CTI+IEU1     Group           */
 151          add            %o0, -14, %o0                   /* IEU0                         */
 152 6:      retl                                            /* CTI+IEU1     Group           */
 153          add            %o0, -13, %o0                   /* IEU0                         */
 154
 155 7:      retl                                            /* CTI+IEU1     Group           */
 156          add            %o0, -12, %o0                   /* IEU0                         */
 157 8:      retl                                            /* CTI+IEU1     Group           */
 158          add            %o0, -11, %o0                   /* IEU0                         */
 159
 160 9:      retl                                            /* CTI+IEU1     Group           */
 161          add            %o0, -10, %o0                   /* IEU0                         */
 162 11:     sub             %o4, %g1, %o5                   /* IEU0         Group           */
 163         sub             %o0, 8, %o0                     /* IEU1                         */
 164
 165         andcc           %o5, %g2, %g0                   /* IEU1         Group           */
 166         be,pt           %xcc, 12f                       /* CTI                          */
 167          sub            %o2, %o0, %o2                   /* IEU0                         */
 168         tst             %o2                             /* IEU1         Group           */
 169
 170         be,pn           %XCC, 12f                       /* CTI                          */
 171          srlx           %o4, 56, %g5                    /* IEU0                         */
 172         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 173         be,pn           %icc, 13f                       /* CTI                          */
 174
 175          cmp            %o2, 1                          /* IEU0                         */
 176         be,pn           %XCC, 12f                       /* CTI          Group           */
 177          srlx           %o4, 48, %g5                    /* IEU0                         */
 178         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 179
 180         be,pn           %icc, 14f                       /* CTI                          */
 181          cmp            %o2, 2                          /* IEU1         Group           */
 182         be,pn           %XCC, 12f                       /* CTI                          */
 183          srlx           %o4, 40, %g5                    /* IEU0                         */
 184
 185         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 186         be,pn           %icc, 15f                       /* CTI                          */
 187          cmp            %o2, 3                          /* IEU1         Group           */
 188         be,pn           %XCC, 12f                       /* CTI                          */
 189
 190          srlx           %o4, 32, %g5                    /* IEU0                         */
 191         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 192         be,pn           %icc, 16f                       /* CTI                          */
 193          cmp            %o2, 4                          /* IEU1         Group           */
 194
 195         be,pn           %XCC, 12f                       /* CTI                          */
 196          srlx           %o4, 24, %g5                    /* IEU0                         */
 197         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 198         be,pn           %icc, 17f                       /* CTI                          */
 199
 200          cmp            %o2, 5                          /* IEU1         Group           */
 201         be,pn           %XCC, 12f                       /* CTI                          */
 202          srlx           %o4, 16, %g5                    /* IEU0                         */
 203         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 204
 205         be,pn           %icc, 18f                       /* CTI                          */
 206          cmp            %o2, 6                          /* IEU1         Group           */
 207         be,pn           %XCC, 12f                       /* CTI                          */
 208          srlx           %o4, 8, %g5                     /* IEU0                         */
 209
 210         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 211         be,pn           %icc, 19f                       /* CTI                          */
 212          nop                                            /* IEU0                         */
 213 12:     retl                                            /* CTI+IEU1     Group           */
 214
 215          clr            %o0                             /* IEU0                         */
 216         nop                                             /* Stub                         */
 217 13:     retl                                            /* CTI+IEU1     Group           */
 218          nop                                            /* IEU0                         */
 219
 220 14:     retl                                            /* CTI+IEU1     Group           */
 221          add            %o0, 1, %o0                     /* IEU0                         */
 222 15:     retl                                            /* CTI+IEU1     Group           */
 223          add            %o0, 2, %o0                     /* IEU0                         */
 224
 225 16:     retl                                            /* CTI+IEU1     Group           */
 226          add            %o0, 3, %o0                     /* IEU0                         */
 227 17:     retl                                            /* CTI+IEU1     Group           */
 228          add            %o0, 4, %o0                     /* IEU0                         */
 229
 230 18:     retl                                            /* CTI+IEU1     Group           */
 231          add            %o0, 5, %o0                     /* IEU0                         */
 232 19:     retl                                            /* CTI+IEU1     Group           */
 233          add            %o0, 6, %o0                     /* IEU0                         */
 234
 235 21:     cmp             %o0, %o2                        /* IEU1                         */
 236         be,pn           %XCC, 12b                       /* CTI                          */
 237          sllx           %g1, 7, %g2                     /* IEU0         Group           */
 238         ldub            [%o0], %o3                      /* Load                         */
 239
 240         or              %g3, %g5, %g3                   /* IEU1                         */
 241 22:     andcc           %o0, 7, %g0                     /* IEU1         Group           */
 242         be,a,pn         %icc, 1b                        /* CTI                          */
 243          ldx            [%o0], %o3                      /* Load                         */
 244
 245         cmp             %o3, %o1                        /* IEU1         Group           */
 246         be,pn           %xcc, 23f                       /* CTI                          */
 247          add            %o0, 1, %o0                     /* IEU0                         */
 248         cmp             %o0, %o2                        /* IEU1         Group           */
 249
 250         bne,a,pt        %XCC, 22b                       /* CTI                          */
 251          ldub           [%o0], %o3                      /* Load                         */
 252         retl                                            /* CTI+IEU1     Group           */
 253          clr            %o0                             /* IEU0                         */
 254
 255 23:     retl                                            /* CTI+IEU1     Group           */
 256          add            %o0, -1, %o0                    /* IEU0                         */
 257 END(__memchr)
 258
 259 weak_alias (__memchr, memchr)
 260 #if !__BOUNDED_POINTERS__
 261 weak_alias (__memchr, __ubp_memchr)
 262 #endif
 263 libc_hidden_builtin_def (memchr)