zlib/contrib/asm386/gvmat32.asm

   1 ;
   2 ; gvmat32.asm -- Asm portion of the optimized longest_match for 32 bits x86
   3 ; Copyright (C) 1995-1996 Jean-loup Gailly and Gilles Vollant.
   4 ; File written by Gilles Vollant, by modifiying the longest_match
   5 ;  from Jean-loup Gailly in deflate.c
   6 ; It need wmask == 0x7fff
   7 ;     (assembly code is faster with a fixed wmask)
   8 ;
   9 ; For Visual C++ 4.2 and ML 6.11c (version in directory \MASM611C of Win95 DDK)
  10 ;   I compile with : "ml /coff /Zi /c gvmat32.asm"
  11 ;
  12
  13 ;uInt longest_match_7fff(s, cur_match)
  14 ;    deflate_state *s;
  15 ;    IPos cur_match;                             /* current match */
  16
  17         NbStack         equ     76
  18         cur_match       equ     dword ptr[esp+NbStack-0]
  19         str_s           equ     dword ptr[esp+NbStack-4]
  20 ; 5 dword on top (ret,ebp,esi,edi,ebx)
  21         adrret          equ     dword ptr[esp+NbStack-8]
  22         pushebp         equ     dword ptr[esp+NbStack-12]
  23         pushedi         equ     dword ptr[esp+NbStack-16]
  24         pushesi         equ     dword ptr[esp+NbStack-20]
  25         pushebx         equ     dword ptr[esp+NbStack-24]
  26
  27         chain_length    equ     dword ptr [esp+NbStack-28]
  28         limit           equ     dword ptr [esp+NbStack-32]
  29         best_len        equ     dword ptr [esp+NbStack-36]
  30         window          equ     dword ptr [esp+NbStack-40]
  31         prev            equ     dword ptr [esp+NbStack-44]
  32         scan_start      equ      word ptr [esp+NbStack-48]
  33         wmask           equ     dword ptr [esp+NbStack-52]
  34         match_start_ptr equ     dword ptr [esp+NbStack-56]
  35         nice_match      equ     dword ptr [esp+NbStack-60]
  36         scan            equ     dword ptr [esp+NbStack-64]
  37
  38         windowlen       equ     dword ptr [esp+NbStack-68]
  39         match_start     equ     dword ptr [esp+NbStack-72]
  40         strend          equ     dword ptr [esp+NbStack-76]
  41         NbStackAdd      equ     (NbStack-24)
  42
  43     .386p
  44
  45     name    gvmatch
  46     .MODEL  FLAT
  47
  48
  49
  50 ;  all the +4 offsets are due to the addition of pending_buf_size (in zlib
  51 ;  in the deflate_state structure since the asm code was first written
  52 ;  (if you compile with zlib 1.0.4 or older, remove the +4).
  53 ;  Note : these value are good with a 8 bytes boundary pack structure
  54     dep_chain_length    equ     70h+4
  55     dep_window          equ     2ch+4
  56     dep_strstart        equ     60h+4
  57     dep_prev_length     equ     6ch+4
  58     dep_nice_match      equ     84h+4
  59     dep_w_size          equ     20h+4
  60     dep_prev            equ     34h+4
  61     dep_w_mask          equ     28h+4
  62     dep_good_match      equ     80h+4
  63     dep_match_start     equ     64h+4
  64     dep_lookahead       equ     68h+4
  65
  66
  67 _TEXT                   segment
  68
  69 IFDEF NOUNDERLINE
  70                         public  longest_match_7fff
  71 ;                        public  match_init
  72 ELSE
  73                         public  _longest_match_7fff
  74 ;                        public  _match_init
  75 ENDIF
  76
  77     MAX_MATCH           equ     258
  78     MIN_MATCH           equ     3
  79     MIN_LOOKAHEAD       equ     (MAX_MATCH+MIN_MATCH+1)
  80
  81
  82
  83 IFDEF NOUNDERLINE
  84 ;match_init      proc near
  85 ;                ret
  86 ;match_init      endp
  87 ELSE
  88 ;_match_init     proc near
  89 ;                ret
  90 ;_match_init     endp
  91 ENDIF
  92
  93
  94 IFDEF NOUNDERLINE
  95 longest_match_7fff   proc near
  96 ELSE
  97 _longest_match_7fff  proc near
  98 ENDIF
  99
 100         mov     edx,[esp+4]
 101
 102
 103
 104         push    ebp
 105         push    edi
 106         push    esi
 107         push    ebx
 108
 109         sub     esp,NbStackAdd
 110
 111 ; initialize or check the variables used in match.asm.
 112         mov     ebp,edx
 113
 114 ; chain_length = s->max_chain_length
 115 ; if (prev_length>=good_match) chain_length >>= 2
 116         mov     edx,[ebp+dep_chain_length]
 117         mov     ebx,[ebp+dep_prev_length]
 118         cmp     [ebp+dep_good_match],ebx
 119         ja      noshr
 120         shr     edx,2
 121 noshr:
 122 ; we increment chain_length because in the asm, the --chain_lenght is in the beginning of the loop
 123         inc     edx
 124         mov     edi,[ebp+dep_nice_match]
 125         mov     chain_length,edx
 126         mov     eax,[ebp+dep_lookahead]
 127         cmp     eax,edi
 128 ; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;
 129         jae     nolookaheadnicematch
 130         mov     edi,eax
 131 nolookaheadnicematch:
 132 ; best_len = s->prev_length
 133         mov     best_len,ebx
 134
 135 ; window = s->window
 136         mov     esi,[ebp+dep_window]
 137         mov     ecx,[ebp+dep_strstart]
 138         mov     window,esi
 139
 140         mov     nice_match,edi
 141 ; scan = window + strstart
 142         add     esi,ecx
 143         mov     scan,esi
 144 ; dx = *window
 145         mov     dx,word ptr [esi]
 146 ; bx = *(window+best_len-1)
 147         mov     bx,word ptr [esi+ebx-1]
 148         add     esi,MAX_MATCH-1
 149 ; scan_start = *scan
 150         mov     scan_start,dx
 151 ; strend = scan + MAX_MATCH-1
 152         mov     strend,esi
 153 ; bx = scan_end = *(window+best_len-1)
 154
 155 ;    IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
 156 ;        s->strstart - (IPos)MAX_DIST(s) : NIL;
 157
 158         mov     esi,[ebp+dep_w_size]
 159         sub     esi,MIN_LOOKAHEAD
 160 ; here esi = MAX_DIST(s)
 161         sub     ecx,esi
 162         ja      nodist
 163         xor     ecx,ecx
 164 nodist:
 165         mov     limit,ecx
 166
 167 ; prev = s->prev
 168         mov     edx,[ebp+dep_prev]
 169         mov     prev,edx
 170
 171 ;
 172         mov     edx,dword ptr [ebp+dep_match_start]
 173         mov     bp,scan_start
 174         mov     eax,cur_match
 175         mov     match_start,edx
 176
 177         mov     edx,window
 178         mov     edi,edx
 179         add     edi,best_len
 180         mov     esi,prev
 181         dec     edi
 182 ; windowlen = window + best_len -1
 183         mov     windowlen,edi
 184
 185         jmp     beginloop2
 186         align   4
 187
 188 ; here, in the loop
 189 ;       eax = ax = cur_match
 190 ;       ecx = limit
 191 ;        bx = scan_end
 192 ;        bp = scan_start
 193 ;       edi = windowlen (window + best_len -1)
 194 ;       esi = prev
 195
 196
 197 ;// here; chain_length <=16
 198 normalbeg0add16:
 199         add     chain_length,16
 200         jz      exitloop
 201 normalbeg0:
 202         cmp     word ptr[edi+eax],bx
 203         je      normalbeg2noroll
 204 rcontlabnoroll:
 205 ; cur_match = prev[cur_match & wmask]
 206         and     eax,7fffh
 207         mov     ax,word ptr[esi+eax*2]
 208 ; if cur_match > limit, go to exitloop
 209         cmp     ecx,eax
 210         jnb     exitloop
 211 ; if --chain_length != 0, go to exitloop
 212         dec     chain_length
 213         jnz     normalbeg0
 214         jmp     exitloop
 215
 216 normalbeg2noroll:
 217 ; if (scan_start==*(cur_match+window)) goto normalbeg2
 218         cmp     bp,word ptr[edx+eax]
 219         jne     rcontlabnoroll
 220         jmp     normalbeg2
 221
 222 contloop3:
 223         mov     edi,windowlen
 224
 225 ; cur_match = prev[cur_match & wmask]
 226         and     eax,7fffh
 227         mov     ax,word ptr[esi+eax*2]
 228 ; if cur_match > limit, go to exitloop
 229         cmp     ecx,eax
 230 jnbexitloopshort1:
 231         jnb     exitloop
 232 ; if --chain_length != 0, go to exitloop
 233
 234
 235 ; begin the main loop
 236 beginloop2:
 237         sub     chain_length,16+1
 238 ; if chain_length <=16, don't use the unrolled loop
 239         jna     normalbeg0add16
 240
 241 do16:
 242         cmp     word ptr[edi+eax],bx
 243         je      normalbeg2dc0
 244
 245 maccn   MACRO   lab
 246         and     eax,7fffh
 247         mov     ax,word ptr[esi+eax*2]
 248         cmp     ecx,eax
 249         jnb     exitloop
 250         cmp     word ptr[edi+eax],bx
 251         je      lab
 252         ENDM
 253
 254 rcontloop0:
 255         maccn   normalbeg2dc1
 256
 257 rcontloop1:
 258         maccn   normalbeg2dc2
 259
 260 rcontloop2:
 261         maccn   normalbeg2dc3
 262
 263 rcontloop3:
 264         maccn   normalbeg2dc4
 265
 266 rcontloop4:
 267         maccn   normalbeg2dc5
 268
 269 rcontloop5:
 270         maccn   normalbeg2dc6
 271
 272 rcontloop6:
 273         maccn   normalbeg2dc7
 274
 275 rcontloop7:
 276         maccn   normalbeg2dc8
 277
 278 rcontloop8:
 279         maccn   normalbeg2dc9
 280
 281 rcontloop9:
 282         maccn   normalbeg2dc10
 283
 284 rcontloop10:
 285         maccn   short normalbeg2dc11
 286
 287 rcontloop11:
 288         maccn   short normalbeg2dc12
 289
 290 rcontloop12:
 291         maccn   short normalbeg2dc13
 292
 293 rcontloop13:
 294         maccn   short normalbeg2dc14
 295
 296 rcontloop14:
 297         maccn   short normalbeg2dc15
 298
 299 rcontloop15:
 300         and     eax,7fffh
 301         mov     ax,word ptr[esi+eax*2]
 302         cmp     ecx,eax
 303         jnb     exitloop
 304
 305         sub     chain_length,16
 306         ja      do16
 307         jmp     normalbeg0add16
 308
 309 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 310
 311 normbeg MACRO   rcontlab,valsub
 312 ; if we are here, we know that *(match+best_len-1) == scan_end
 313         cmp     bp,word ptr[edx+eax]
 314 ; if (match != scan_start) goto rcontlab
 315         jne     rcontlab
 316 ; calculate the good chain_length, and we'll compare scan and match string
 317         add     chain_length,16-valsub
 318         jmp     iseq
 319         ENDM
 320
 321
 322 normalbeg2dc11:
 323         normbeg rcontloop11,11
 324
 325 normalbeg2dc12:
 326         normbeg short rcontloop12,12
 327
 328 normalbeg2dc13:
 329         normbeg short rcontloop13,13
 330
 331 normalbeg2dc14:
 332         normbeg short rcontloop14,14
 333
 334 normalbeg2dc15:
 335         normbeg short rcontloop15,15
 336
 337 normalbeg2dc10:
 338         normbeg rcontloop10,10
 339
 340 normalbeg2dc9:
 341         normbeg rcontloop9,9
 342
 343 normalbeg2dc8:
 344         normbeg rcontloop8,8
 345
 346 normalbeg2dc7:
 347         normbeg rcontloop7,7
 348
 349 normalbeg2dc6:
 350         normbeg rcontloop6,6
 351
 352 normalbeg2dc5:
 353         normbeg rcontloop5,5
 354
 355 normalbeg2dc4:
 356         normbeg rcontloop4,4
 357
 358 normalbeg2dc3:
 359         normbeg rcontloop3,3
 360
 361 normalbeg2dc2:
 362         normbeg rcontloop2,2
 363
 364 normalbeg2dc1:
 365         normbeg rcontloop1,1
 366
 367 normalbeg2dc0:
 368         normbeg rcontloop0,0
 369
 370
 371 ; we go in normalbeg2 because *(ushf*)(match+best_len-1) == scan_end
 372
 373 normalbeg2:
 374         mov     edi,window
 375
 376         cmp     bp,word ptr[edi+eax]
 377         jne     contloop3                   ; if *(ushf*)match != scan_start, continue
 378
 379 iseq:
 380 ; if we are here, we know that *(match+best_len-1) == scan_end
 381 ; and (match == scan_start)
 382
 383         mov     edi,edx
 384         mov     esi,scan                    ; esi = scan
 385         add     edi,eax                     ; edi = window + cur_match = match
 386
 387         mov     edx,[esi+3]                 ; compare manually dword at match+3
 388         xor     edx,[edi+3]                 ; and scan +3
 389
 390         jz      begincompare                ; if equal, go to long compare
 391
 392 ; we will determine the unmatch byte and calculate len (in esi)
 393         or      dl,dl
 394         je      eq1rr
 395         mov     esi,3
 396         jmp     trfinval
 397 eq1rr:
 398         or      dx,dx
 399         je      eq1
 400
 401         mov     esi,4
 402         jmp     trfinval
 403 eq1:
 404         and     edx,0ffffffh
 405         jz      eq11
 406         mov     esi,5
 407         jmp     trfinval
 408 eq11:
 409         mov     esi,6
 410         jmp     trfinval
 411
 412 begincompare:
 413         ; here we now scan and match begin same
 414         add     edi,6
 415         add     esi,6
 416         mov     ecx,(MAX_MATCH-(2+4))/4     ; scan for at most MAX_MATCH bytes
 417         repe    cmpsd                       ; loop until mismatch
 418
 419         je      trfin                       ; go to trfin if not unmatch
 420 ; we determine the unmatch byte
 421         sub     esi,4
 422         mov     edx,[edi-4]
 423         xor     edx,[esi]
 424
 425         or      dl,dl
 426         jnz     trfin
 427         inc     esi
 428
 429         or      dx,dx
 430         jnz     trfin
 431         inc     esi
 432
 433         and     edx,0ffffffh
 434         jnz     trfin
 435         inc     esi
 436
 437 trfin:
 438         sub     esi,scan          ; esi = len
 439 trfinval:
 440 ; here we have finised compare, and esi contain len of equal string
 441         cmp     esi,best_len        ; if len > best_len, go newbestlen
 442         ja      short newbestlen
 443 ; now we restore edx, ecx and esi, for the big loop
 444         mov     esi,prev
 445         mov     ecx,limit
 446         mov     edx,window
 447         jmp     contloop3
 448
 449 newbestlen:
 450         mov     best_len,esi        ; len become best_len
 451
 452         mov     match_start,eax     ; save new position as match_start
 453         cmp     esi,nice_match      ; if best_len >= nice_match, exit
 454         jae     exitloop
 455         mov     ecx,scan
 456         mov     edx,window          ; restore edx=window
 457         add     ecx,esi
 458         add     esi,edx
 459
 460         dec     esi
 461         mov     windowlen,esi       ; windowlen = window + best_len-1
 462         mov     bx,[ecx-1]          ; bx = *(scan+best_len-1) = scan_end
 463
 464 ; now we restore ecx and esi, for the big loop :
 465         mov     esi,prev
 466         mov     ecx,limit
 467         jmp     contloop3
 468
 469 exitloop:
 470 ; exit : s->match_start=match_start
 471         mov     ebx,match_start
 472         mov     ebp,str_s
 473         mov     ecx,best_len
 474         mov     dword ptr [ebp+dep_match_start],ebx
 475         mov     eax,dword ptr [ebp+dep_lookahead]
 476         cmp     ecx,eax
 477         ja      minexlo
 478         mov     eax,ecx
 479 minexlo:
 480 ; return min(best_len,s->lookahead)
 481
 482 ; restore stack and register ebx,esi,edi,ebp
 483         add     esp,NbStackAdd
 484
 485         pop     ebx
 486         pop     esi
 487         pop     edi
 488         pop     ebp
 489         ret
 490 InfoAuthor:
 491 ; please don't remove this string !
 492 ; Your are free use gvmat32 in any fre or commercial apps if you don't remove the string in the binary!
 493         db     0dh,0ah,"GVMat32 optimised assembly code written 1996-98 by Gilles Vollant",0dh,0ah
 494
 495
 496
 497 IFDEF NOUNDERLINE
 498 longest_match_7fff   endp
 499 ELSE
 500 _longest_match_7fff  endp
 501 ENDIF
 502
 503
 504 IFDEF NOUNDERLINE
 505 cpudetect32     proc near
 506 ELSE
 507 _cpudetect32    proc near
 508 ENDIF
 509
 510
 511         pushfd                  ; push original EFLAGS
 512         pop     eax             ; get original EFLAGS
 513         mov     ecx, eax        ; save original EFLAGS
 514         xor     eax, 40000h     ; flip AC bit in EFLAGS
 515         push    eax             ; save new EFLAGS value on stack
 516         popfd                   ; replace current EFLAGS value
 517         pushfd                  ; get new EFLAGS
 518         pop     eax             ; store new EFLAGS in EAX
 519         xor     eax, ecx        ; can’t toggle AC bit, processor=80386
 520         jz      end_cpu_is_386  ; jump if 80386 processor
 521         push    ecx
 522         popfd                   ; restore AC bit in EFLAGS first
 523
 524         pushfd
 525         pushfd
 526         pop     ecx
 527
 528         mov     eax, ecx        ; get original EFLAGS
 529         xor     eax, 200000h    ; flip ID bit in EFLAGS
 530         push    eax             ; save new EFLAGS value on stack
 531         popfd                   ; replace current EFLAGS value
 532         pushfd                  ; get new EFLAGS
 533         pop             eax                 ; store new EFLAGS in EAX
 534         popfd                   ; restore original EFLAGS
 535         xor             eax, ecx        ; can’t toggle ID bit,
 536         je              is_old_486              ; processor=old
 537
 538         mov     eax,1
 539         db      0fh,0a2h        ;CPUID
 540
 541 exitcpudetect:
 542         ret
 543
 544 end_cpu_is_386:
 545         mov     eax,0300h
 546         jmp     exitcpudetect
 547
 548 is_old_486:
 549         mov     eax,0400h
 550         jmp     exitcpudetect
 551
 552 IFDEF NOUNDERLINE
 553 cpudetect32     endp
 554 ELSE
 555 _cpudetect32    endp
 556 ENDIF
 557
 558 _TEXT   ends
 559 end