linux/lib/zlib/match586.S

   1 /* match.s -- Pentium-optimized version of longest_match()
   2  * Written for zlib 1.1.2
   3  * Copyright (C) 1998 Brian Raiter <breadbox@muppetlabs.com>
   4  *
   5  * This is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License.
   7  */
   8
   9 #ifndef NO_UNDERLINE
  10 #define match_init      _ipcomp_match_init
  11 #define longest_match   _ipcomp_longest_match
  12 #else
  13 #define match_init      ipcomp_match_init
  14 #define longest_match   ipcomp_longest_match
  15 #endif
  16
  17 #define MAX_MATCH       (258)
  18 #define MIN_MATCH       (3)
  19 #define MIN_LOOKAHEAD   (MAX_MATCH + MIN_MATCH + 1)
  20 #define MAX_MATCH_8     ((MAX_MATCH + 7) & ~7)
  21
  22 /* stack frame offsets */
  23
  24 #define wmask                   0       /* local copy of s->wmask       */
  25 #define window                  4       /* local copy of s->window      */
  26 #define windowbestlen           8       /* s->window + bestlen          */
  27 #define chainlenscanend         12      /* high word: current chain len */
  28                                         /* low word: last bytes sought  */
  29 #define scanstart               16      /* first two bytes of string    */
  30 #define scanalign               20      /* dword-misalignment of string */
  31 #define nicematch               24      /* a good enough match size     */
  32 #define bestlen                 28      /* size of best match so far    */
  33 #define scan                    32      /* ptr to string wanting match  */
  34
  35 #define LocalVarsSize           (36)
  36 /*      saved ebx               36 */
  37 /*      saved edi               40 */
  38 /*      saved esi               44 */
  39 /*      saved ebp               48 */
  40 /*      return address          52 */
  41 #define deflatestate            56      /* the function arguments       */
  42 #define curmatch                60
  43
  44 /* Offsets for fields in the deflate_state structure. These numbers
  45  * are calculated from the definition of deflate_state, with the
  46  * assumption that the compiler will dword-align the fields. (Thus,
  47  * changing the definition of deflate_state could easily cause this
  48  * program to crash horribly, without so much as a warning at
  49  * compile time. Sigh.)
  50  */
  51 #define dsWSize                 36
  52 #define dsWMask                 44
  53 #define dsWindow                48
  54 #define dsPrev                  56
  55 #define dsMatchLen              88
  56 #define dsPrevMatch             92
  57 #define dsStrStart              100
  58 #define dsMatchStart            104
  59 #define dsLookahead             108
  60 #define dsPrevLen               112
  61 #define dsMaxChainLen           116
  62 #define dsGoodMatch             132
  63 #define dsNiceMatch             136
  64
  65
  66 .file "match.S"
  67
  68 .globl  match_init, longest_match
  69
  70 .text
  71
  72 /* uInt longest_match(deflate_state *deflatestate, IPos curmatch) */
  73
  74 longest_match:
  75
  76 /* Save registers that the compiler may be using, and adjust %esp to    */
  77 /* make room for our stack frame.                                       */
  78
  79                 pushl   %ebp
  80                 pushl   %edi
  81                 pushl   %esi
  82                 pushl   %ebx
  83                 subl    $LocalVarsSize, %esp
  84
  85 /* Retrieve the function arguments. %ecx will hold cur_match            */
  86 /* throughout the entire function. %edx will hold the pointer to the    */
  87 /* deflate_state structure during the function's setup (before          */
  88 /* entering the main loop).                                             */
  89
  90                 movl    deflatestate(%esp), %edx
  91                 movl    curmatch(%esp), %ecx
  92
  93 /* if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;      */
  94
  95                 movl    dsNiceMatch(%edx), %eax
  96                 movl    dsLookahead(%edx), %ebx
  97                 cmpl    %eax, %ebx
  98                 jl      LookaheadLess
  99                 movl    %eax, %ebx
 100 LookaheadLess:  movl    %ebx, nicematch(%esp)
 101
 102 /* register Bytef *scan = s->window + s->strstart;                      */
 103
 104                 movl    dsWindow(%edx), %esi
 105                 movl    %esi, window(%esp)
 106                 movl    dsStrStart(%edx), %ebp
 107                 lea     (%esi,%ebp), %edi
 108                 movl    %edi, scan(%esp)
 109
 110 /* Determine how many bytes the scan ptr is off from being              */
 111 /* dword-aligned.                                                       */
 112
 113                 movl    %edi, %eax
 114                 negl    %eax
 115                 andl    $3, %eax
 116                 movl    %eax, scanalign(%esp)
 117
 118 /* IPos limit = s->strstart > (IPos)MAX_DIST(s) ?                       */
 119 /*     s->strstart - (IPos)MAX_DIST(s) : NIL;                           */
 120
 121                 movl    dsWSize(%edx), %eax
 122                 subl    $MIN_LOOKAHEAD, %eax
 123                 subl    %eax, %ebp
 124                 jg      LimitPositive
 125                 xorl    %ebp, %ebp
 126 LimitPositive:
 127
 128 /* unsigned chain_length = s->max_chain_length;                         */
 129 /* if (s->prev_length >= s->good_match) {                               */
 130 /*     chain_length >>= 2;                                              */
 131 /* }                                                                    */
 132
 133                 movl    dsPrevLen(%edx), %eax
 134                 movl    dsGoodMatch(%edx), %ebx
 135                 cmpl    %ebx, %eax
 136                 movl    dsMaxChainLen(%edx), %ebx
 137                 jl      LastMatchGood
 138                 shrl    $2, %ebx
 139 LastMatchGood:
 140
 141 /* chainlen is decremented once beforehand so that the function can     */
 142 /* use the sign flag instead of the zero flag for the exit test.        */
 143 /* It is then shifted into the high word, to make room for the scanend  */
 144 /* scanend value, which it will always accompany.                       */
 145
 146                 decl    %ebx
 147                 shll    $16, %ebx
 148
 149 /* int best_len = s->prev_length;                                       */
 150
 151                 movl    dsPrevLen(%edx), %eax
 152                 movl    %eax, bestlen(%esp)
 153
 154 /* Store the sum of s->window + best_len in %esi locally, and in %esi.  */
 155
 156                 addl    %eax, %esi
 157                 movl    %esi, windowbestlen(%esp)
 158
 159 /* register ush scan_start = *(ushf*)scan;                              */
 160 /* register ush scan_end   = *(ushf*)(scan+best_len-1);                 */
 161
 162                 movw    (%edi), %bx
 163                 movw    %bx, scanstart(%esp)
 164                 movw    -1(%edi,%eax), %bx
 165                 movl    %ebx, chainlenscanend(%esp)
 166
 167 /* Posf *prev = s->prev;                                                */
 168 /* uInt wmask = s->w_mask;                                              */
 169
 170                 movl    dsPrev(%edx), %edi
 171                 movl    dsWMask(%edx), %edx
 172                 mov     %edx, wmask(%esp)
 173
 174 /* Jump into the main loop.                                             */
 175
 176                 jmp     LoopEntry
 177
 178 .balign 16
 179
 180 /* do {
 181  *     match = s->window + cur_match;
 182  *     if (*(ushf*)(match+best_len-1) != scan_end ||
 183  *         *(ushf*)match != scan_start) continue;
 184  *     [...]
 185  * } while ((cur_match = prev[cur_match & wmask]) > limit
 186  *          && --chain_length != 0);
 187  *
 188  * Here is the inner loop of the function. The function will spend the
 189  * majority of its time in this loop, and majority of that time will
 190  * be spent in the first ten instructions.
 191  *
 192  * Within this loop:
 193  * %ebx = chainlenscanend - i.e., ((chainlen << 16) | scanend)
 194  * %ecx = curmatch
 195  * %edx = curmatch & wmask
 196  * %esi = windowbestlen - i.e., (window + bestlen)
 197  * %edi = prev
 198  * %ebp = limit
 199  *
 200  * Two optimization notes on the choice of instructions:
 201  *
 202  * The first instruction uses a 16-bit address, which costs an extra,
 203  * unpairable cycle. This is cheaper than doing a 32-bit access and
 204  * zeroing the high word, due to the 3-cycle misalignment penalty which
 205  * would occur half the time. This also turns out to be cheaper than
 206  * doing two separate 8-bit accesses, as the memory is so rarely in the
 207  * L1 cache.
 208  *
 209  * The window buffer, however, apparently spends a lot of time in the
 210  * cache, and so it is faster to retrieve the word at the end of the
 211  * match string with two 8-bit loads. The instructions that test the
 212  * word at the beginning of the match string, however, are executed
 213  * much less frequently, and there it was cheaper to use 16-bit
 214  * instructions, which avoided the necessity of saving off and
 215  * subsequently reloading one of the other registers.
 216  */
 217 LookupLoop:
 218                                                         /* 1 U & V  */
 219                 movw    (%edi,%edx,2), %cx              /* 2 U pipe */
 220                 movl    wmask(%esp), %edx               /* 2 V pipe */
 221                 cmpl    %ebp, %ecx                      /* 3 U pipe */
 222                 jbe     LeaveNow                        /* 3 V pipe */
 223                 subl    $0x00010000, %ebx               /* 4 U pipe */
 224                 js      LeaveNow                        /* 4 V pipe */
 225 LoopEntry:      movb    -1(%esi,%ecx), %al              /* 5 U pipe */
 226                 andl    %ecx, %edx                      /* 5 V pipe */
 227                 cmpb    %bl, %al                        /* 6 U pipe */
 228                 jnz     LookupLoop                      /* 6 V pipe */
 229                 movb    (%esi,%ecx), %ah
 230                 cmpb    %bh, %ah
 231                 jnz     LookupLoop
 232                 movl    window(%esp), %eax
 233                 movw    (%eax,%ecx), %ax
 234                 cmpw    scanstart(%esp), %ax
 235                 jnz     LookupLoop
 236
 237 /* Store the current value of chainlen.                                 */
 238
 239                 movl    %ebx, chainlenscanend(%esp)
 240
 241 /* Point %edi to the string under scrutiny, and %esi to the string we   */
 242 /* are hoping to match it up with. In actuality, %esi and %edi are      */
 243 /* both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and %edx is      */
 244 /* initialized to -(MAX_MATCH_8 - scanalign).                           */
 245
 246                 movl    window(%esp), %esi
 247                 movl    scan(%esp), %edi
 248                 addl    %ecx, %esi
 249                 movl    scanalign(%esp), %eax
 250                 movl    $(-MAX_MATCH_8), %edx
 251                 lea     MAX_MATCH_8(%edi,%eax), %edi
 252                 lea     MAX_MATCH_8(%esi,%eax), %esi
 253
 254 /* Test the strings for equality, 8 bytes at a time. At the end,
 255  * adjust %edx so that it is offset to the exact byte that mismatched.
 256  *
 257  * We already know at this point that the first three bytes of the
 258  * strings match each other, and they can be safely passed over before
 259  * starting the compare loop. So what this code does is skip over 0-3
 260  * bytes, as much as necessary in order to dword-align the %edi
 261  * pointer. (%esi will still be misaligned three times out of four.)
 262  *
 263  * It should be confessed that this loop usually does not represent
 264  * much of the total running time. Replacing it with a more
 265  * straightforward "rep cmpsb" would not drastically degrade
 266  * performance.
 267  */
 268 LoopCmps:
 269                 movl    (%esi,%edx), %eax
 270                 movl    (%edi,%edx), %ebx
 271                 xorl    %ebx, %eax
 272                 jnz     LeaveLoopCmps
 273                 movl    4(%esi,%edx), %eax
 274                 movl    4(%edi,%edx), %ebx
 275                 xorl    %ebx, %eax
 276                 jnz     LeaveLoopCmps4
 277                 addl    $8, %edx
 278                 jnz     LoopCmps
 279                 jmp     LenMaximum
 280 LeaveLoopCmps4: addl    $4, %edx
 281 LeaveLoopCmps:  testl   $0x0000FFFF, %eax
 282                 jnz     LenLower
 283                 addl    $2, %edx
 284                 shrl    $16, %eax
 285 LenLower:       subb    $1, %al
 286                 adcl    $0, %edx
 287
 288 /* Calculate the length of the match. If it is longer than MAX_MATCH,   */
 289 /* then automatically accept it as the best possible match and leave.   */
 290
 291                 lea     (%edi,%edx), %eax
 292                 movl    scan(%esp), %edi
 293                 subl    %edi, %eax
 294                 cmpl    $MAX_MATCH, %eax
 295                 jge     LenMaximum
 296
 297 /* If the length of the match is not longer than the best match we      */
 298 /* have so far, then forget it and return to the lookup loop.           */
 299
 300                 movl    deflatestate(%esp), %edx
 301                 movl    bestlen(%esp), %ebx
 302                 cmpl    %ebx, %eax
 303                 jg      LongerMatch
 304                 movl    chainlenscanend(%esp), %ebx
 305                 movl    windowbestlen(%esp), %esi
 306                 movl    dsPrev(%edx), %edi
 307                 movl    wmask(%esp), %edx
 308                 andl    %ecx, %edx
 309                 jmp     LookupLoop
 310
 311 /*         s->match_start = cur_match;                                  */
 312 /*         best_len = len;                                              */
 313 /*         if (len >= nice_match) break;                                */
 314 /*         scan_end = *(ushf*)(scan+best_len-1);                        */
 315
 316 LongerMatch:    movl    nicematch(%esp), %ebx
 317                 movl    %eax, bestlen(%esp)
 318                 movl    %ecx, dsMatchStart(%edx)
 319                 cmpl    %ebx, %eax
 320                 jge     LeaveNow
 321                 movl    window(%esp), %esi
 322                 addl    %eax, %esi
 323                 movl    %esi, windowbestlen(%esp)
 324                 movl    chainlenscanend(%esp), %ebx
 325                 movw    -1(%edi,%eax), %bx
 326                 movl    dsPrev(%edx), %edi
 327                 movl    %ebx, chainlenscanend(%esp)
 328                 movl    wmask(%esp), %edx
 329                 andl    %ecx, %edx
 330                 jmp     LookupLoop
 331
 332 /* Accept the current string, with the maximum possible length.         */
 333
 334 LenMaximum:     movl    deflatestate(%esp), %edx
 335                 movl    $MAX_MATCH, bestlen(%esp)
 336                 movl    %ecx, dsMatchStart(%edx)
 337
 338 /* if ((uInt)best_len <= s->lookahead) return (uInt)best_len;           */
 339 /* return s->lookahead;                                                 */
 340
 341 LeaveNow:
 342                 movl    deflatestate(%esp), %edx
 343                 movl    bestlen(%esp), %ebx
 344                 movl    dsLookahead(%edx), %eax
 345                 cmpl    %eax, %ebx
 346                 jg      LookaheadRet
 347                 movl    %ebx, %eax
 348 LookaheadRet:
 349
 350 /* Restore the stack and return from whence we came.                    */
 351
 352                 addl    $LocalVarsSize, %esp
 353                 popl    %ebx
 354                 popl    %esi
 355                 popl    %edi
 356                 popl    %ebp
 357 match_init:     ret