sysdeps/sparc/sparc64/strcpy.S

   1 /* Copy SRC to DEST returning 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 Jan Vondrak <jvon4518@ss1000.ms.mff.cuni.cz> and
   6                   Jakub Jelinek <jj@ultra.linux.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 <sysdep.h>
  23 #include <asm/asi.h>
  24 #ifndef XCC
  25         .register       %g2, #scratch
  26         .register       %g3, #scratch
  27         .register       %g6, #scratch
  28 #endif
  29
  30         /* Normally, this uses
  31            ((xword - 0x0101010101010101) & 0x8080808080808080) test
  32            to find out if any byte in xword could be zero. This is fast, but
  33            also gives false alarm for any byte in range 0x81-0xff. It does
  34            not matter for correctness, as if this test tells us there could
  35            be some zero byte, we check it byte by byte, but if bytes with
  36            high bits set are common in the strings, then this will give poor
  37            performance. You can #define EIGHTBIT_NOT_RARE and the algorithm
  38            will use one tick slower, but more precise test
  39            ((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080),
  40            which does not give any false alarms (but if some bits are set,
  41            one cannot assume from it which bytes are zero and which are not).
  42            It is yet to be measured, what is the correct default for glibc
  43            in these days for an average user.
  44          */
  45
  46         .text
  47         .align          32
  48 ENTRY(strcpy)
  49         sethi           %hi(0x01010101), %g1            /* IEU0         Group           */
  50         mov             %o0, %g6                        /* IEU1                         */
  51         or              %g1, %lo(0x01010101), %g1       /* IEU0         Group           */
  52         andcc           %o0, 7, %g0                     /* IEU1                         */
  53
  54         sllx            %g1, 32, %g2                    /* IEU0         Group           */
  55         bne,pn          %icc, 12f                       /* CTI                          */
  56          andcc          %o1, 7, %g3                     /* IEU1                         */
  57         or              %g1, %g2, %g1                   /* IEU0         Group           */
  58
  59         bne,pn          %icc, 14f                       /* CTI                          */
  60          sllx           %g1, 7, %g2                     /* IEU0         Group           */
  61 1:      ldx             [%o1], %o3                      /* Load                         */
  62         add             %o1, 8, %o1                     /* IEU1                         */
  63
  64 2:      mov             %o3, %g3                        /* IEU0         Group           */
  65 3:      sub             %o3, %g1, %o2                   /* IEU1                         */
  66         ldxa            [%o1] ASI_PNF, %o3              /* Load                         */
  67 #ifdef EIGHTBIT_NOT_RARE
  68         andn            %o2, %g3, %o2                   /* IEU0         Group           */
  69 #endif
  70         add             %o0, 8, %o0                     /* IEU0         Group           */
  71
  72         andcc           %o2, %g2, %g0                   /* IEU1                         */
  73         add             %o1, 8, %o1                     /* IEU0         Group           */
  74         be,a,pt         %xcc, 2b                        /* CTI                          */
  75          stx            %g3, [%o0 - 8]                  /* Store                        */
  76
  77         srlx            %g3, 56, %g5                    /* IEU0         Group           */
  78         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
  79         be,pn           %icc, 11f                       /* CTI                          */
  80          srlx           %g3, 48, %g4                    /* IEU0                         */
  81
  82         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
  83         be,pn           %icc, 10f                       /* CTI                          */
  84          srlx           %g3, 40, %g5                    /* IEU0                         */
  85         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
  86
  87         be,pn           %icc, 9f                        /* CTI                          */
  88          srlx           %g3, 32, %g4                    /* IEU0                         */
  89         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
  90         be,pn           %icc, 8f                        /* CTI                          */
  91
  92          srlx           %g3, 24, %g5                    /* IEU0                         */
  93         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
  94         be,pn           %icc, 7f                        /* CTI                          */
  95          srlx           %g3, 16, %g4                    /* IEU0                         */
  96
  97         andcc           %g4, 0xff, %g0                  /* IEU1         Group           */
  98         be,pn           %icc, 6f                        /* CTI                          */
  99          srlx           %g3, 8, %g5                     /* IEU0                         */
 100         andcc           %g5, 0xff, %g0                  /* IEU1         Group           */
 101
 102         be,pn           %icc, 5f                        /* CTI                          */
 103          sub            %o3, %g1, %o2                   /* IEU0                         */
 104         stx             %g3, [%o0 - 8]                  /* Store        Group           */
 105         andcc           %g3, 0xff, %g0                  /* IEU1                         */
 106
 107         bne,pt          %icc, 3b                        /* CTI                          */
 108          mov            %o3, %g3                        /* IEU0         Group           */
 109 4:      retl                                            /* CTI+IEU1     Group           */
 110          mov            %g6, %o0                        /* IEU0                         */
 111
 112         .align          16
 113 5:      stb             %g5, [%o0 - 2]                  /* Store        Group           */
 114         srlx            %g3, 16, %g4                    /* IEU0                         */
 115 6:      sth             %g4, [%o0 - 4]                  /* Store        Group           */
 116         srlx            %g3, 32, %g4                    /* IEU0                         */
 117
 118         stw             %g4, [%o0 - 8]                  /* Store        Group           */
 119         retl                                            /* CTI+IEU1     Group           */
 120          mov            %g6, %o0                        /* IEU0                         */
 121 7:      stb             %g5, [%o0 - 4]                  /* Store        Group           */
 122
 123         srlx            %g3, 32, %g4                    /* IEU0                         */
 124 8:      stw             %g4, [%o0 - 8]                  /* Store        Group           */
 125         retl                                            /* CTI+IEU1     Group           */
 126          mov            %g6, %o0                        /* IEU0                         */
 127
 128 9:      stb             %g5, [%o0 - 6]                  /* Store        Group           */
 129         srlx            %g3, 48, %g4                    /* IEU0                         */
 130 10:     sth             %g4, [%o0 - 8]                  /* Store        Group           */
 131         retl                                            /* CTI+IEU1     Group           */
 132
 133          mov            %g6, %o0                        /* IEU0                         */
 134 11:     stb             %g5, [%o0 - 8]                  /* Store        Group           */
 135         retl                                            /* CTI+IEU1     Group           */
 136          mov            %g6, %o0                        /* IEU0                         */
 137
 138 12:     or              %g1, %g2, %g1                   /* IEU0         Group           */
 139         ldub            [%o1], %o3                      /* Load                         */
 140         sllx            %g1, 7, %g2                     /* IEU0         Group           */
 141         stb             %o3, [%o0]                      /* Store        Group           */
 142
 143 13:     add             %o0, 1, %o0                     /* IEU0                         */
 144         add             %o1, 1, %o1                     /* IEU1                         */
 145         andcc           %o3, 0xff, %g0                  /* IEU1         Group           */
 146         be,pn           %icc, 4b                        /* CTI                          */
 147
 148          lduba          [%o1] ASI_PNF, %o3              /* Load                         */
 149         andcc           %o0, 7, %g0                     /* IEU1         Group           */
 150         bne,a,pt        %icc, 13b                       /* CTI                          */
 151          stb            %o3, [%o0]                      /* Store                        */
 152
 153         andcc           %o1, 7, %g3                     /* IEU1         Group           */
 154         be,a,pt         %icc, 1b                        /* CTI                          */
 155          ldx            [%o1], %o3                      /* Load                         */
 156 14:     orcc            %g0, 64, %g4                    /* IEU1         Group           */
 157
 158         sllx            %g3, 3, %g5                     /* IEU0                         */
 159         sub             %o1, %g3, %o1                   /* IEU0         Group           */
 160         sub             %g4, %g5, %g4                   /* IEU1                         */
 161                                                         /* %g1 = 0101010101010101       *
 162                                                          * %g2 = 8080808080808080       *
 163                                                          * %g3 = source alignment       *
 164                                                          * %g5 = number of bits to shift left  *
 165                                                          * %g4 = number of bits to shift right */
 166         ldxa            [%o1] ASI_PNF, %o5              /* Load         Group           */
 167
 168         addcc           %o1, 8, %o1                     /* IEU1                         */
 169 15:     sllx            %o5, %g5, %o3                   /* IEU0         Group           */
 170         ldxa            [%o1] ASI_PNF, %o5              /* Load                         */
 171         srlx            %o5, %g4, %o4                   /* IEU0         Group           */
 172
 173         add             %o0, 8, %o0                     /* IEU1                         */
 174         or              %o3, %o4, %o3                   /* IEU0         Group           */
 175         add             %o1, 8, %o1                     /* IEU1                         */
 176         sub             %o3, %g1, %o4                   /* IEU0         Group           */
 177
 178 #ifdef EIGHTBIT_NOT_RARE
 179         andn            %o4, %o3, %o4                   /* IEU0         Group           */
 180 #endif
 181         andcc           %o4, %g2, %g0                   /* IEU1         Group           */
 182         be,a,pt         %xcc, 15b                       /* CTI                          */
 183          stx            %o3, [%o0 - 8]                  /* Store                        */
 184         srlx            %o3, 56, %o4                    /* IEU0         Group           */
 185
 186         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 187         be,pn           %icc, 22f                       /* CTI                          */
 188          srlx           %o3, 48, %o4                    /* IEU0                         */
 189         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 190
 191         be,pn           %icc, 21f                       /* CTI                          */
 192          srlx           %o3, 40, %o4                    /* IEU0                         */
 193         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 194         be,pn           %icc, 20f                       /* CTI                          */
 195
 196          srlx           %o3, 32, %o4                    /* IEU0                         */
 197         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 198         be,pn           %icc, 19f                       /* CTI                          */
 199          srlx           %o3, 24, %o4                    /* IEU0                         */
 200
 201         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 202         be,pn           %icc, 18f                       /* CTI                          */
 203          srlx           %o3, 16, %o4                    /* IEU0                         */
 204         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 205
 206         be,pn           %icc, 17f                       /* CTI                          */
 207          srlx           %o3, 8, %o4                     /* IEU0                         */
 208         andcc           %o4, 0xff, %g0                  /* IEU1         Group           */
 209         be,pn           %icc, 16f                       /* CTI                          */
 210
 211          andcc          %o3, 0xff, %g0                  /* IEU1         Group           */
 212         bne,pn          %icc, 15b                       /* CTI                          */
 213          stx            %o3, [%o0 - 8]                  /* Store                        */
 214         retl                                            /* CTI+IEU1     Group           */
 215
 216          mov            %g6, %o0                        /* IEU0                         */
 217
 218         .align          16
 219 16:     srlx            %o3, 8, %o4                     /* IEU0         Group           */
 220         stb             %o4, [%o0 - 2]                  /* Store                        */
 221 17:     srlx            %o3, 16, %o4                    /* IEU0         Group           */
 222         stb             %o4, [%o0 - 3]                  /* Store                        */
 223
 224 18:     srlx            %o3, 24, %o4                    /* IEU0         Group           */
 225         stb             %o4, [%o0 - 4]                  /* Store                        */
 226 19:     srlx            %o3, 32, %o4                    /* IEU0         Group           */
 227         stw             %o4, [%o0 - 8]                  /* Store                        */
 228
 229         retl                                            /* CTI+IEU1     Group           */
 230          mov            %g6, %o0                        /* IEU0                         */
 231         nop
 232         nop
 233
 234 20:     srlx            %o3, 40, %o4                    /* IEU0         Group           */
 235         stb             %o4, [%o0 - 6]                  /* Store                        */
 236 21:     srlx            %o3, 48, %o4                    /* IEU0         Group           */
 237         stb             %o4, [%o0 - 7]                  /* Store                        */
 238
 239 22:     srlx            %o3, 56, %o4                    /* IEU0         Group           */
 240         stb             %o4, [%o0 - 8]                  /* Store                        */
 241         retl                                            /* CTI+IEU1     Group           */
 242          mov            %g6, %o0                        /* IEU0                         */
 243 END(strcpy)
 244 libc_hidden_builtin_def (strcpy)