newlib/libc/machine/cris/memmove.c

   1 /* A memmove for CRIS.
   2    Copyright (C) 2000-2005 Axis Communications.
   3    All rights reserved.
   4
   5    Redistribution and use in source and binary forms, with or without
   6    modification, are permitted provided that the following conditions
   7    are met:
   8
   9    1. Redistributions of source code must retain the above copyright
  10       notice, this list of conditions and the following disclaimer.
  11
  12    2. Neither the name of Axis Communications nor the names of its
  13       contributors may be used to endorse or promote products derived
  14       from this software without specific prior written permission.
  15
  16    THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
  17    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  18    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  19    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
  20    COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
  21    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  22    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  23    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  25    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  26    IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  27    POSSIBILITY OF SUCH DAMAGE.  */
  28
  29 /* FIXME: This file should really only be used for reference, as the
  30    result is somewhat depending on gcc generating what we expect rather
  31    than what we describe.  An assembly file should be used instead.
  32
  33    Even worse, we base it on memcpy, on the assumption that overlapping
  34    moves are rare, and we will do no worse than the generic memmove.  */
  35
  36 #include <stddef.h>
  37
  38 /* Break even between movem and move16 is really at 38.7 * 2, but
  39    modulo 44, so up to the next multiple of 44, we use ordinary code.  */
  40 #define MEMMOVE_BY_BLOCK_THRESHOLD (44 * 2)
  41
  42 /* No name ambiguities in this file.  */
  43 __asm__ (".syntax no_register_prefix");
  44
  45 void *
  46 memmove(void *pdst, const void *psrc, size_t pn)
  47 {
  48   /* Now we want the parameters put in special registers.
  49      Make sure the compiler is able to make something useful of this.
  50      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
  51
  52      If gcc was allright, it really would need no temporaries, and no
  53      stack space to save stuff on.  */
  54
  55   register void *return_dst __asm__ ("r10") = pdst;
  56   register unsigned char *dst __asm__ ("r13") = pdst;
  57   register unsigned const char *src __asm__ ("r11") = psrc;
  58   register int n __asm__ ("r12") = pn;
  59
  60   /* Check and handle overlap.  */
  61   if (src < dst && dst < src + n)
  62     {
  63       /* Destructive overlap.  We could optimize this, but we don't (for
  64          the moment).  */
  65       src += n;
  66       dst += n;
  67       while (n--)
  68         {
  69           *--dst = *--src;
  70         }
  71
  72       return return_dst;
  73     }
  74   /* Whew, no overlap.  Proceed as with memcpy.  We could call it instead
  75      of having a copy here.  That would spoil some of the optimization, so
  76      we take the trouble with having two copies.  */
  77
  78   /* When src is aligned but not dst, this makes a few extra needless
  79      cycles.  I believe it would take as many to check that the
  80      re-alignment was unnecessary.  */
  81   if (((unsigned long) dst & 3) != 0
  82       /* Don't align if we wouldn't copy more than a few bytes; so we
  83          don't have to check further for overflows.  */
  84       && n >= 3)
  85   {
  86     if ((unsigned long) dst & 1)
  87       {
  88         n--;
  89         *dst = *src;
  90         src++;
  91         dst++;
  92       }
  93
  94     if ((unsigned long) dst & 2)
  95       {
  96         n -= 2;
  97         *(short *) dst = *(short *) src;
  98         src += 2;
  99         dst += 2;
 100       }
 101   }
 102
 103   /* Decide which copying method to use.  */
 104   if (n >= MEMMOVE_BY_BLOCK_THRESHOLD)
 105     {
 106       /* It is not optimal to tell the compiler about clobbering any
 107          registers; that will move the saving/restoring of those registers
 108          to the function prologue/epilogue, and make non-movem sizes
 109          suboptimal.  */
 110       __asm__ volatile
 111         ("\
 112          ;; GCC does promise correct register allocations, but let's    \n\
 113          ;; make sure it keeps its promises.                            \n\
 114          .ifnc %0-%1-%2,$r13-$r11-$r12                                  \n\
 115          .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\"       \n\
 116          .endif                                                         \n\
 117                                                                         \n\
 118          ;; Save the registers we'll use in the movem process           \n\
 119          ;; on the stack.                                               \n\
 120          subq   11*4,sp                                                 \n\
 121          movem  r10,[sp]                                                \n\
 122                                                                         \n\
 123          ;; Now we've got this:                                         \n\
 124          ;; r11 - src                                                   \n\
 125          ;; r13 - dst                                                   \n\
 126          ;; r12 - n                                                     \n\
 127                                                                         \n\
 128          ;; Update n for the first loop.                                \n\
 129          subq    44,r12                                                 \n\
 130 0:                                                                      \n\
 131 "
 132 #ifdef __arch_common_v10_v32
 133          /* Cater to branch offset difference between v32 and v10.  We
 134             assume the branch below has an 8-bit offset.  */
 135 "        setf\n"
 136 #endif
 137 "        movem  [r11+],r10                                              \n\
 138          subq   44,r12                                                  \n\
 139          bge     0b                                                     \n\
 140          movem  r10,[r13+]                                              \n\
 141                                                                         \n\
 142          ;; Compensate for last loop underflowing n.                    \n\
 143          addq   44,r12                                                  \n\
 144                                                                         \n\
 145          ;; Restore registers from stack.                               \n\
 146          movem [sp+],r10"
 147
 148          /* Outputs.  */
 149          : "=r" (dst), "=r" (src), "=r" (n)
 150
 151          /* Inputs.  */
 152          : "0" (dst), "1" (src), "2" (n));
 153     }
 154
 155   while (n >= 16)
 156     {
 157       *(long *) dst = *(long *) src; dst += 4; src += 4;
 158       *(long *) dst = *(long *) src; dst += 4; src += 4;
 159       *(long *) dst = *(long *) src; dst += 4; src += 4;
 160       *(long *) dst = *(long *) src; dst += 4; src += 4;
 161
 162       n -= 16;
 163     }
 164
 165   switch (n)
 166     {
 167     case 0:
 168       break;
 169
 170     case 1:
 171       *dst = *src;
 172       break;
 173
 174     case 2:
 175       *(short *) dst = *(short *) src;
 176       break;
 177
 178     case 3:
 179       *(short *) dst = *(short *) src; dst += 2; src += 2;
 180       *dst = *src;
 181       break;
 182
 183     case 4:
 184       *(long *) dst = *(long *) src;
 185       break;
 186
 187     case 5:
 188       *(long *) dst = *(long *) src; dst += 4; src += 4;
 189       *dst = *src;
 190       break;
 191
 192     case 6:
 193       *(long *) dst = *(long *) src; dst += 4; src += 4;
 194       *(short *) dst = *(short *) src;
 195       break;
 196
 197     case 7:
 198       *(long *) dst = *(long *) src; dst += 4; src += 4;
 199       *(short *) dst = *(short *) src; dst += 2; src += 2;
 200       *dst = *src;
 201       break;
 202
 203     case 8:
 204       *(long *) dst = *(long *) src; dst += 4; src += 4;
 205       *(long *) dst = *(long *) src;
 206       break;
 207
 208     case 9:
 209       *(long *) dst = *(long *) src; dst += 4; src += 4;
 210       *(long *) dst = *(long *) src; dst += 4; src += 4;
 211       *dst = *src;
 212       break;
 213
 214     case 10:
 215       *(long *) dst = *(long *) src; dst += 4; src += 4;
 216       *(long *) dst = *(long *) src; dst += 4; src += 4;
 217       *(short *) dst = *(short *) src;
 218       break;
 219
 220     case 11:
 221       *(long *) dst = *(long *) src; dst += 4; src += 4;
 222       *(long *) dst = *(long *) src; dst += 4; src += 4;
 223       *(short *) dst = *(short *) src; dst += 2; src += 2;
 224       *dst = *src;
 225       break;
 226
 227     case 12:
 228       *(long *) dst = *(long *) src; dst += 4; src += 4;
 229       *(long *) dst = *(long *) src; dst += 4; src += 4;
 230       *(long *) dst = *(long *) src;
 231       break;
 232
 233     case 13:
 234       *(long *) dst = *(long *) src; dst += 4; src += 4;
 235       *(long *) dst = *(long *) src; dst += 4; src += 4;
 236       *(long *) dst = *(long *) src; dst += 4; src += 4;
 237       *dst = *src;
 238       break;
 239
 240     case 14:
 241       *(long *) dst = *(long *) src; dst += 4; src += 4;
 242       *(long *) dst = *(long *) src; dst += 4; src += 4;
 243       *(long *) dst = *(long *) src; dst += 4; src += 4;
 244       *(short *) dst = *(short *) src;
 245       break;
 246
 247     case 15:
 248       *(long *) dst = *(long *) src; dst += 4; src += 4;
 249       *(long *) dst = *(long *) src; dst += 4; src += 4;
 250       *(long *) dst = *(long *) src; dst += 4; src += 4;
 251       *(short *) dst = *(short *) src; dst += 2; src += 2;
 252       *dst = *src;
 253       break;
 254     }
 255
 256   return return_dst;
 257 }