Mercurial > hg > sv-dependency-builds

annotate src/zlib-1.2.7/contrib/masmx86/inffas32.asm @ 168:ceec0dd9ec9c

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Replace these with versions built using an older toolset (so as to avoid ABI compatibilities when linking on Ubuntu 14.04 for packaging purposes)
author Chris Cannam <cannam@all-day-breakfast.com>
date Fri, 07 Feb 2020 11:51:13 +0000
parents 8a15ff55d9af
children
rev   line source
cannam@89 1 ;/* inffas32.asm is a hand tuned assembler version of inffast.c -- fast decoding
cannam@89 2 ; *
cannam@89 3 ; * inffas32.asm is derivated from inffas86.c, with translation of assembly code
cannam@89 4 ; *
cannam@89 5 ; * Copyright (C) 1995-2003 Mark Adler
cannam@89 6 ; * For conditions of distribution and use, see copyright notice in zlib.h
cannam@89 7 ; *
cannam@89 8 ; * Copyright (C) 2003 Chris Anderson <christop@charm.net>
cannam@89 9 ; * Please use the copyright conditions above.
cannam@89 10 ; *
cannam@89 11 ; * Mar-13-2003 -- Most of this is derived from inffast.S which is derived from
cannam@89 12 ; * the gcc -S output of zlib-1.2.0/inffast.c. Zlib-1.2.0 is in beta release at
cannam@89 13 ; * the moment. I have successfully compiled and tested this code with gcc2.96,
cannam@89 14 ; * gcc3.2, icc5.0, msvc6.0. It is very close to the speed of inffast.S
cannam@89 15 ; * compiled with gcc -DNO_MMX, but inffast.S is still faster on the P3 with MMX
cannam@89 16 ; * enabled. I will attempt to merge the MMX code into this version. Newer
cannam@89 17 ; * versions of this and inffast.S can be found at
cannam@89 18 ; * http://www.eetbeetee.com/zlib/ and http://www.charm.net/~christop/zlib/
cannam@89 19 ; *
cannam@89 20 ; * 2005 : modification by Gilles Vollant
cannam@89 21 ; */
cannam@89 22 ; For Visual C++ 4.x and higher and ML 6.x and higher
cannam@89 23 ; ml.exe is in directory \MASM611C of Win95 DDK
cannam@89 24 ; ml.exe is also distributed in http://www.masm32.com/masmdl.htm
cannam@89 25 ; and in VC++2003 toolkit at http://msdn.microsoft.com/visualc/vctoolkit2003/
cannam@89 26 ;
cannam@89 27 ;
cannam@89 28 ; compile with command line option
cannam@89 29 ; ml /coff /Zi /c /Flinffas32.lst inffas32.asm
cannam@89 30
cannam@89 31 ; if you define NO_GZIP (see inflate.h), compile with
cannam@89 32 ; ml /coff /Zi /c /Flinffas32.lst /DNO_GUNZIP inffas32.asm
cannam@89 33
cannam@89 34
cannam@89 35 ; zlib122sup is 0 fort zlib 1.2.2.1 and lower
cannam@89 36 ; zlib122sup is 8 fort zlib 1.2.2.2 and more (with addition of dmax and head
cannam@89 37 ; in inflate_state in inflate.h)
cannam@89 38 zlib1222sup equ 8
cannam@89 39
cannam@89 40
cannam@89 41 IFDEF GUNZIP
cannam@89 42 INFLATE_MODE_TYPE equ 11
cannam@89 43 INFLATE_MODE_BAD equ 26
cannam@89 44 ELSE
cannam@89 45 IFNDEF NO_GUNZIP
cannam@89 46 INFLATE_MODE_TYPE equ 11
cannam@89 47 INFLATE_MODE_BAD equ 26
cannam@89 48 ELSE
cannam@89 49 INFLATE_MODE_TYPE equ 3
cannam@89 50 INFLATE_MODE_BAD equ 17
cannam@89 51 ENDIF
cannam@89 52 ENDIF
cannam@89 53
cannam@89 54
cannam@89 55 ; 75 "inffast.S"
cannam@89 56 ;FILE "inffast.S"
cannam@89 57
cannam@89 58 ;;;GLOBAL _inflate_fast
cannam@89 59
cannam@89 60 ;;;SECTION .text
cannam@89 61
cannam@89 62
cannam@89 63
cannam@89 64 .586p
cannam@89 65 .mmx
cannam@89 66
cannam@89 67 name inflate_fast_x86
cannam@89 68 .MODEL FLAT
cannam@89 69
cannam@89 70 _DATA segment
cannam@89 71 inflate_fast_use_mmx:
cannam@89 72 dd 1
cannam@89 73
cannam@89 74
cannam@89 75 _TEXT segment
cannam@89 76
cannam@89 77
cannam@89 78
cannam@89 79 ALIGN 4
cannam@89 80 db 'Fast decoding Code from Chris Anderson'
cannam@89 81 db 0
cannam@89 82
cannam@89 83 ALIGN 4
cannam@89 84 invalid_literal_length_code_msg:
cannam@89 85 db 'invalid literal/length code'
cannam@89 86 db 0
cannam@89 87
cannam@89 88 ALIGN 4
cannam@89 89 invalid_distance_code_msg:
cannam@89 90 db 'invalid distance code'
cannam@89 91 db 0
cannam@89 92
cannam@89 93 ALIGN 4
cannam@89 94 invalid_distance_too_far_msg:
cannam@89 95 db 'invalid distance too far back'
cannam@89 96 db 0
cannam@89 97
cannam@89 98
cannam@89 99 ALIGN 4
cannam@89 100 inflate_fast_mask:
cannam@89 101 dd 0
cannam@89 102 dd 1
cannam@89 103 dd 3
cannam@89 104 dd 7
cannam@89 105 dd 15
cannam@89 106 dd 31
cannam@89 107 dd 63
cannam@89 108 dd 127
cannam@89 109 dd 255
cannam@89 110 dd 511
cannam@89 111 dd 1023
cannam@89 112 dd 2047
cannam@89 113 dd 4095
cannam@89 114 dd 8191
cannam@89 115 dd 16383
cannam@89 116 dd 32767
cannam@89 117 dd 65535
cannam@89 118 dd 131071
cannam@89 119 dd 262143
cannam@89 120 dd 524287
cannam@89 121 dd 1048575
cannam@89 122 dd 2097151
cannam@89 123 dd 4194303
cannam@89 124 dd 8388607
cannam@89 125 dd 16777215
cannam@89 126 dd 33554431
cannam@89 127 dd 67108863
cannam@89 128 dd 134217727
cannam@89 129 dd 268435455
cannam@89 130 dd 536870911
cannam@89 131 dd 1073741823
cannam@89 132 dd 2147483647
cannam@89 133 dd 4294967295
cannam@89 134
cannam@89 135
cannam@89 136 mode_state equ 0 ;/* state->mode */
cannam@89 137 wsize_state equ (32+zlib1222sup) ;/* state->wsize */
cannam@89 138 write_state equ (36+4+zlib1222sup) ;/* state->write */
cannam@89 139 window_state equ (40+4+zlib1222sup) ;/* state->window */
cannam@89 140 hold_state equ (44+4+zlib1222sup) ;/* state->hold */
cannam@89 141 bits_state equ (48+4+zlib1222sup) ;/* state->bits */
cannam@89 142 lencode_state equ (64+4+zlib1222sup) ;/* state->lencode */
cannam@89 143 distcode_state equ (68+4+zlib1222sup) ;/* state->distcode */
cannam@89 144 lenbits_state equ (72+4+zlib1222sup) ;/* state->lenbits */
cannam@89 145 distbits_state equ (76+4+zlib1222sup) ;/* state->distbits */
cannam@89 146
cannam@89 147
cannam@89 148 ;;SECTION .text
cannam@89 149 ; 205 "inffast.S"
cannam@89 150 ;GLOBAL inflate_fast_use_mmx
cannam@89 151
cannam@89 152 ;SECTION .data
cannam@89 153
cannam@89 154
cannam@89 155 ; GLOBAL inflate_fast_use_mmx:object
cannam@89 156 ;.size inflate_fast_use_mmx, 4
cannam@89 157 ; 226 "inffast.S"
cannam@89 158 ;SECTION .text
cannam@89 159
cannam@89 160 ALIGN 4
cannam@89 161 _inflate_fast proc near
cannam@89 162 .FPO (16, 4, 0, 0, 1, 0)
cannam@89 163 push edi
cannam@89 164 push esi
cannam@89 165 push ebp
cannam@89 166 push ebx
cannam@89 167 pushfd
cannam@89 168 sub esp,64
cannam@89 169 cld
cannam@89 170
cannam@89 171
cannam@89 172
cannam@89 173
cannam@89 174 mov esi, [esp+88]
cannam@89 175 mov edi, [esi+28]
cannam@89 176
cannam@89 177
cannam@89 178
cannam@89 179
cannam@89 180
cannam@89 181
cannam@89 182
cannam@89 183 mov edx, [esi+4]
cannam@89 184 mov eax, [esi+0]
cannam@89 185
cannam@89 186 add edx,eax
cannam@89 187 sub edx,11
cannam@89 188
cannam@89 189 mov [esp+44],eax
cannam@89 190 mov [esp+20],edx
cannam@89 191
cannam@89 192 mov ebp, [esp+92]
cannam@89 193 mov ecx, [esi+16]
cannam@89 194 mov ebx, [esi+12]
cannam@89 195
cannam@89 196 sub ebp,ecx
cannam@89 197 neg ebp
cannam@89 198 add ebp,ebx
cannam@89 199
cannam@89 200 sub ecx,257
cannam@89 201 add ecx,ebx
cannam@89 202
cannam@89 203 mov [esp+60],ebx
cannam@89 204 mov [esp+40],ebp
cannam@89 205 mov [esp+16],ecx
cannam@89 206 ; 285 "inffast.S"
cannam@89 207 mov eax, [edi+lencode_state]
cannam@89 208 mov ecx, [edi+distcode_state]
cannam@89 209
cannam@89 210 mov [esp+8],eax
cannam@89 211 mov [esp+12],ecx
cannam@89 212
cannam@89 213 mov eax,1
cannam@89 214 mov ecx, [edi+lenbits_state]
cannam@89 215 shl eax,cl
cannam@89 216 dec eax
cannam@89 217 mov [esp+0],eax
cannam@89 218
cannam@89 219 mov eax,1
cannam@89 220 mov ecx, [edi+distbits_state]
cannam@89 221 shl eax,cl
cannam@89 222 dec eax
cannam@89 223 mov [esp+4],eax
cannam@89 224
cannam@89 225 mov eax, [edi+wsize_state]
cannam@89 226 mov ecx, [edi+write_state]
cannam@89 227 mov edx, [edi+window_state]
cannam@89 228
cannam@89 229 mov [esp+52],eax
cannam@89 230 mov [esp+48],ecx
cannam@89 231 mov [esp+56],edx
cannam@89 232
cannam@89 233 mov ebp, [edi+hold_state]
cannam@89 234 mov ebx, [edi+bits_state]
cannam@89 235 ; 321 "inffast.S"
cannam@89 236 mov esi, [esp+44]
cannam@89 237 mov ecx, [esp+20]
cannam@89 238 cmp ecx,esi
cannam@89 239 ja L_align_long
cannam@89 240
cannam@89 241 add ecx,11
cannam@89 242 sub ecx,esi
cannam@89 243 mov eax,12
cannam@89 244 sub eax,ecx
cannam@89 245 lea edi, [esp+28]
cannam@89 246 rep movsb
cannam@89 247 mov ecx,eax
cannam@89 248 xor eax,eax
cannam@89 249 rep stosb
cannam@89 250 lea esi, [esp+28]
cannam@89 251 mov [esp+20],esi
cannam@89 252 jmp L_is_aligned
cannam@89 253
cannam@89 254
cannam@89 255 L_align_long:
cannam@89 256 test esi,3
cannam@89 257 jz L_is_aligned
cannam@89 258 xor eax,eax
cannam@89 259 mov al, [esi]
cannam@89 260 inc esi
cannam@89 261 mov ecx,ebx
cannam@89 262 add ebx,8
cannam@89 263 shl eax,cl
cannam@89 264 or ebp,eax
cannam@89 265 jmp L_align_long
cannam@89 266
cannam@89 267 L_is_aligned:
cannam@89 268 mov edi, [esp+60]
cannam@89 269 ; 366 "inffast.S"
cannam@89 270 L_check_mmx:
cannam@89 271 cmp dword ptr [inflate_fast_use_mmx],2
cannam@89 272 je L_init_mmx
cannam@89 273 ja L_do_loop
cannam@89 274
cannam@89 275 push eax
cannam@89 276 push ebx
cannam@89 277 push ecx
cannam@89 278 push edx
cannam@89 279 pushfd
cannam@89 280 mov eax, [esp]
cannam@89 281 xor dword ptr [esp],0200000h
cannam@89 282
cannam@89 283
cannam@89 284
cannam@89 285
cannam@89 286 popfd
cannam@89 287 pushfd
cannam@89 288 pop edx
cannam@89 289 xor edx,eax
cannam@89 290 jz L_dont_use_mmx
cannam@89 291 xor eax,eax
cannam@89 292 cpuid
cannam@89 293 cmp ebx,0756e6547h
cannam@89 294 jne L_dont_use_mmx
cannam@89 295 cmp ecx,06c65746eh
cannam@89 296 jne L_dont_use_mmx
cannam@89 297 cmp edx,049656e69h
cannam@89 298 jne L_dont_use_mmx
cannam@89 299 mov eax,1
cannam@89 300 cpuid
cannam@89 301 shr eax,8
cannam@89 302 and eax,15
cannam@89 303 cmp eax,6
cannam@89 304 jne L_dont_use_mmx
cannam@89 305 test edx,0800000h
cannam@89 306 jnz L_use_mmx
cannam@89 307 jmp L_dont_use_mmx
cannam@89 308 L_use_mmx:
cannam@89 309 mov dword ptr [inflate_fast_use_mmx],2
cannam@89 310 jmp L_check_mmx_pop
cannam@89 311 L_dont_use_mmx:
cannam@89 312 mov dword ptr [inflate_fast_use_mmx],3
cannam@89 313 L_check_mmx_pop:
cannam@89 314 pop edx
cannam@89 315 pop ecx
cannam@89 316 pop ebx
cannam@89 317 pop eax
cannam@89 318 jmp L_check_mmx
cannam@89 319 ; 426 "inffast.S"
cannam@89 320 ALIGN 4
cannam@89 321 L_do_loop:
cannam@89 322 ; 437 "inffast.S"
cannam@89 323 cmp bl,15
cannam@89 324 ja L_get_length_code
cannam@89 325
cannam@89 326 xor eax,eax
cannam@89 327 lodsw
cannam@89 328 mov cl,bl
cannam@89 329 add bl,16
cannam@89 330 shl eax,cl
cannam@89 331 or ebp,eax
cannam@89 332
cannam@89 333 L_get_length_code:
cannam@89 334 mov edx, [esp+0]
cannam@89 335 mov ecx, [esp+8]
cannam@89 336 and edx,ebp
cannam@89 337 mov eax, [ecx+edx*4]
cannam@89 338
cannam@89 339 L_dolen:
cannam@89 340
cannam@89 341
cannam@89 342
cannam@89 343
cannam@89 344
cannam@89 345
cannam@89 346 mov cl,ah
cannam@89 347 sub bl,ah
cannam@89 348 shr ebp,cl
cannam@89 349
cannam@89 350
cannam@89 351
cannam@89 352
cannam@89 353
cannam@89 354
cannam@89 355 test al,al
cannam@89 356 jnz L_test_for_length_base
cannam@89 357
cannam@89 358 shr eax,16
cannam@89 359 stosb
cannam@89 360
cannam@89 361 L_while_test:
cannam@89 362
cannam@89 363
cannam@89 364 cmp [esp+16],edi
cannam@89 365 jbe L_break_loop
cannam@89 366
cannam@89 367 cmp [esp+20],esi
cannam@89 368 ja L_do_loop
cannam@89 369 jmp L_break_loop
cannam@89 370
cannam@89 371 L_test_for_length_base:
cannam@89 372 ; 502 "inffast.S"
cannam@89 373 mov edx,eax
cannam@89 374 shr edx,16
cannam@89 375 mov cl,al
cannam@89 376
cannam@89 377 test al,16
cannam@89 378 jz L_test_for_second_level_length
cannam@89 379 and cl,15
cannam@89 380 jz L_save_len
cannam@89 381 cmp bl,cl
cannam@89 382 jae L_add_bits_to_len
cannam@89 383
cannam@89 384 mov ch,cl
cannam@89 385 xor eax,eax
cannam@89 386 lodsw
cannam@89 387 mov cl,bl
cannam@89 388 add bl,16
cannam@89 389 shl eax,cl
cannam@89 390 or ebp,eax
cannam@89 391 mov cl,ch
cannam@89 392
cannam@89 393 L_add_bits_to_len:
cannam@89 394 mov eax,1
cannam@89 395 shl eax,cl
cannam@89 396 dec eax
cannam@89 397 sub bl,cl
cannam@89 398 and eax,ebp
cannam@89 399 shr ebp,cl
cannam@89 400 add edx,eax
cannam@89 401
cannam@89 402 L_save_len:
cannam@89 403 mov [esp+24],edx
cannam@89 404
cannam@89 405
cannam@89 406 L_decode_distance:
cannam@89 407 ; 549 "inffast.S"
cannam@89 408 cmp bl,15
cannam@89 409 ja L_get_distance_code
cannam@89 410
cannam@89 411 xor eax,eax
cannam@89 412 lodsw
cannam@89 413 mov cl,bl
cannam@89 414 add bl,16
cannam@89 415 shl eax,cl
cannam@89 416 or ebp,eax
cannam@89 417
cannam@89 418 L_get_distance_code:
cannam@89 419 mov edx, [esp+4]
cannam@89 420 mov ecx, [esp+12]
cannam@89 421 and edx,ebp
cannam@89 422 mov eax, [ecx+edx*4]
cannam@89 423
cannam@89 424
cannam@89 425 L_dodist:
cannam@89 426 mov edx,eax
cannam@89 427 shr edx,16
cannam@89 428 mov cl,ah
cannam@89 429 sub bl,ah
cannam@89 430 shr ebp,cl
cannam@89 431 ; 584 "inffast.S"
cannam@89 432 mov cl,al
cannam@89 433
cannam@89 434 test al,16
cannam@89 435 jz L_test_for_second_level_dist
cannam@89 436 and cl,15
cannam@89 437 jz L_check_dist_one
cannam@89 438 cmp bl,cl
cannam@89 439 jae L_add_bits_to_dist
cannam@89 440
cannam@89 441 mov ch,cl
cannam@89 442 xor eax,eax
cannam@89 443 lodsw
cannam@89 444 mov cl,bl
cannam@89 445 add bl,16
cannam@89 446 shl eax,cl
cannam@89 447 or ebp,eax
cannam@89 448 mov cl,ch
cannam@89 449
cannam@89 450 L_add_bits_to_dist:
cannam@89 451 mov eax,1
cannam@89 452 shl eax,cl
cannam@89 453 dec eax
cannam@89 454 sub bl,cl
cannam@89 455 and eax,ebp
cannam@89 456 shr ebp,cl
cannam@89 457 add edx,eax
cannam@89 458 jmp L_check_window
cannam@89 459
cannam@89 460 L_check_window:
cannam@89 461 ; 625 "inffast.S"
cannam@89 462 mov [esp+44],esi
cannam@89 463 mov eax,edi
cannam@89 464 sub eax, [esp+40]
cannam@89 465
cannam@89 466 cmp eax,edx
cannam@89 467 jb L_clip_window
cannam@89 468
cannam@89 469 mov ecx, [esp+24]
cannam@89 470 mov esi,edi
cannam@89 471 sub esi,edx
cannam@89 472
cannam@89 473 sub ecx,3
cannam@89 474 mov al, [esi]
cannam@89 475 mov [edi],al
cannam@89 476 mov al, [esi+1]
cannam@89 477 mov dl, [esi+2]
cannam@89 478 add esi,3
cannam@89 479 mov [edi+1],al
cannam@89 480 mov [edi+2],dl
cannam@89 481 add edi,3
cannam@89 482 rep movsb
cannam@89 483
cannam@89 484 mov esi, [esp+44]
cannam@89 485 jmp L_while_test
cannam@89 486
cannam@89 487 ALIGN 4
cannam@89 488 L_check_dist_one:
cannam@89 489 cmp edx,1
cannam@89 490 jne L_check_window
cannam@89 491 cmp [esp+40],edi
cannam@89 492 je L_check_window
cannam@89 493
cannam@89 494 dec edi
cannam@89 495 mov ecx, [esp+24]
cannam@89 496 mov al, [edi]
cannam@89 497 sub ecx,3
cannam@89 498
cannam@89 499 mov [edi+1],al
cannam@89 500 mov [edi+2],al
cannam@89 501 mov [edi+3],al
cannam@89 502 add edi,4
cannam@89 503 rep stosb
cannam@89 504
cannam@89 505 jmp L_while_test
cannam@89 506
cannam@89 507 ALIGN 4
cannam@89 508 L_test_for_second_level_length:
cannam@89 509
cannam@89 510
cannam@89 511
cannam@89 512
cannam@89 513 test al,64
cannam@89 514 jnz L_test_for_end_of_block
cannam@89 515
cannam@89 516 mov eax,1
cannam@89 517 shl eax,cl
cannam@89 518 dec eax
cannam@89 519 and eax,ebp
cannam@89 520 add eax,edx
cannam@89 521 mov edx, [esp+8]
cannam@89 522 mov eax, [edx+eax*4]
cannam@89 523 jmp L_dolen
cannam@89 524
cannam@89 525 ALIGN 4
cannam@89 526 L_test_for_second_level_dist:
cannam@89 527
cannam@89 528
cannam@89 529
cannam@89 530
cannam@89 531 test al,64
cannam@89 532 jnz L_invalid_distance_code
cannam@89 533
cannam@89 534 mov eax,1
cannam@89 535 shl eax,cl
cannam@89 536 dec eax
cannam@89 537 and eax,ebp
cannam@89 538 add eax,edx
cannam@89 539 mov edx, [esp+12]
cannam@89 540 mov eax, [edx+eax*4]
cannam@89 541 jmp L_dodist
cannam@89 542
cannam@89 543 ALIGN 4
cannam@89 544 L_clip_window:
cannam@89 545 ; 721 "inffast.S"
cannam@89 546 mov ecx,eax
cannam@89 547 mov eax, [esp+52]
cannam@89 548 neg ecx
cannam@89 549 mov esi, [esp+56]
cannam@89 550
cannam@89 551 cmp eax,edx
cannam@89 552 jb L_invalid_distance_too_far
cannam@89 553
cannam@89 554 add ecx,edx
cannam@89 555 cmp dword ptr [esp+48],0
cannam@89 556 jne L_wrap_around_window
cannam@89 557
cannam@89 558 sub eax,ecx
cannam@89 559 add esi,eax
cannam@89 560 ; 749 "inffast.S"
cannam@89 561 mov eax, [esp+24]
cannam@89 562 cmp eax,ecx
cannam@89 563 jbe L_do_copy1
cannam@89 564
cannam@89 565 sub eax,ecx
cannam@89 566 rep movsb
cannam@89 567 mov esi,edi
cannam@89 568 sub esi,edx
cannam@89 569 jmp L_do_copy1
cannam@89 570
cannam@89 571 cmp eax,ecx
cannam@89 572 jbe L_do_copy1
cannam@89 573
cannam@89 574 sub eax,ecx
cannam@89 575 rep movsb
cannam@89 576 mov esi,edi
cannam@89 577 sub esi,edx
cannam@89 578 jmp L_do_copy1
cannam@89 579
cannam@89 580 L_wrap_around_window:
cannam@89 581 ; 793 "inffast.S"
cannam@89 582 mov eax, [esp+48]
cannam@89 583 cmp ecx,eax
cannam@89 584 jbe L_contiguous_in_window
cannam@89 585
cannam@89 586 add esi, [esp+52]
cannam@89 587 add esi,eax
cannam@89 588 sub esi,ecx
cannam@89 589 sub ecx,eax
cannam@89 590
cannam@89 591
cannam@89 592 mov eax, [esp+24]
cannam@89 593 cmp eax,ecx
cannam@89 594 jbe L_do_copy1
cannam@89 595
cannam@89 596 sub eax,ecx
cannam@89 597 rep movsb
cannam@89 598 mov esi, [esp+56]
cannam@89 599 mov ecx, [esp+48]
cannam@89 600 cmp eax,ecx
cannam@89 601 jbe L_do_copy1
cannam@89 602
cannam@89 603 sub eax,ecx
cannam@89 604 rep movsb
cannam@89 605 mov esi,edi
cannam@89 606 sub esi,edx
cannam@89 607 jmp L_do_copy1
cannam@89 608
cannam@89 609 L_contiguous_in_window:
cannam@89 610 ; 836 "inffast.S"
cannam@89 611 add esi,eax
cannam@89 612 sub esi,ecx
cannam@89 613
cannam@89 614
cannam@89 615 mov eax, [esp+24]
cannam@89 616 cmp eax,ecx
cannam@89 617 jbe L_do_copy1
cannam@89 618
cannam@89 619 sub eax,ecx
cannam@89 620 rep movsb
cannam@89 621 mov esi,edi
cannam@89 622 sub esi,edx
cannam@89 623
cannam@89 624 L_do_copy1:
cannam@89 625 ; 862 "inffast.S"
cannam@89 626 mov ecx,eax
cannam@89 627 rep movsb
cannam@89 628
cannam@89 629 mov esi, [esp+44]
cannam@89 630 jmp L_while_test
cannam@89 631 ; 878 "inffast.S"
cannam@89 632 ALIGN 4
cannam@89 633 L_init_mmx:
cannam@89 634 emms
cannam@89 635
cannam@89 636
cannam@89 637
cannam@89 638
cannam@89 639
cannam@89 640 movd mm0,ebp
cannam@89 641 mov ebp,ebx
cannam@89 642 ; 896 "inffast.S"
cannam@89 643 movd mm4,dword ptr [esp+0]
cannam@89 644 movq mm3,mm4
cannam@89 645 movd mm5,dword ptr [esp+4]
cannam@89 646 movq mm2,mm5
cannam@89 647 pxor mm1,mm1
cannam@89 648 mov ebx, [esp+8]
cannam@89 649 jmp L_do_loop_mmx
cannam@89 650
cannam@89 651 ALIGN 4
cannam@89 652 L_do_loop_mmx:
cannam@89 653 psrlq mm0,mm1
cannam@89 654
cannam@89 655 cmp ebp,32
cannam@89 656 ja L_get_length_code_mmx
cannam@89 657
cannam@89 658 movd mm6,ebp
cannam@89 659 movd mm7,dword ptr [esi]
cannam@89 660 add esi,4
cannam@89 661 psllq mm7,mm6
cannam@89 662 add ebp,32
cannam@89 663 por mm0,mm7
cannam@89 664
cannam@89 665 L_get_length_code_mmx:
cannam@89 666 pand mm4,mm0
cannam@89 667 movd eax,mm4
cannam@89 668 movq mm4,mm3
cannam@89 669 mov eax, [ebx+eax*4]
cannam@89 670
cannam@89 671 L_dolen_mmx:
cannam@89 672 movzx ecx,ah
cannam@89 673 movd mm1,ecx
cannam@89 674 sub ebp,ecx
cannam@89 675
cannam@89 676 test al,al
cannam@89 677 jnz L_test_for_length_base_mmx
cannam@89 678
cannam@89 679 shr eax,16
cannam@89 680 stosb
cannam@89 681
cannam@89 682 L_while_test_mmx:
cannam@89 683
cannam@89 684
cannam@89 685 cmp [esp+16],edi
cannam@89 686 jbe L_break_loop
cannam@89 687
cannam@89 688 cmp [esp+20],esi
cannam@89 689 ja L_do_loop_mmx
cannam@89 690 jmp L_break_loop
cannam@89 691
cannam@89 692 L_test_for_length_base_mmx:
cannam@89 693
cannam@89 694 mov edx,eax
cannam@89 695 shr edx,16
cannam@89 696
cannam@89 697 test al,16
cannam@89 698 jz L_test_for_second_level_length_mmx
cannam@89 699 and eax,15
cannam@89 700 jz L_decode_distance_mmx
cannam@89 701
cannam@89 702 psrlq mm0,mm1
cannam@89 703 movd mm1,eax
cannam@89 704 movd ecx,mm0
cannam@89 705 sub ebp,eax
cannam@89 706 and ecx, [inflate_fast_mask+eax*4]
cannam@89 707 add edx,ecx
cannam@89 708
cannam@89 709 L_decode_distance_mmx:
cannam@89 710 psrlq mm0,mm1
cannam@89 711
cannam@89 712 cmp ebp,32
cannam@89 713 ja L_get_dist_code_mmx
cannam@89 714
cannam@89 715 movd mm6,ebp
cannam@89 716 movd mm7,dword ptr [esi]
cannam@89 717 add esi,4
cannam@89 718 psllq mm7,mm6
cannam@89 719 add ebp,32
cannam@89 720 por mm0,mm7
cannam@89 721
cannam@89 722 L_get_dist_code_mmx:
cannam@89 723 mov ebx, [esp+12]
cannam@89 724 pand mm5,mm0
cannam@89 725 movd eax,mm5
cannam@89 726 movq mm5,mm2
cannam@89 727 mov eax, [ebx+eax*4]
cannam@89 728
cannam@89 729 L_dodist_mmx:
cannam@89 730
cannam@89 731 movzx ecx,ah
cannam@89 732 mov ebx,eax
cannam@89 733 shr ebx,16
cannam@89 734 sub ebp,ecx
cannam@89 735 movd mm1,ecx
cannam@89 736
cannam@89 737 test al,16
cannam@89 738 jz L_test_for_second_level_dist_mmx
cannam@89 739 and eax,15
cannam@89 740 jz L_check_dist_one_mmx
cannam@89 741
cannam@89 742 L_add_bits_to_dist_mmx:
cannam@89 743 psrlq mm0,mm1
cannam@89 744 movd mm1,eax
cannam@89 745 movd ecx,mm0
cannam@89 746 sub ebp,eax
cannam@89 747 and ecx, [inflate_fast_mask+eax*4]
cannam@89 748 add ebx,ecx
cannam@89 749
cannam@89 750 L_check_window_mmx:
cannam@89 751 mov [esp+44],esi
cannam@89 752 mov eax,edi
cannam@89 753 sub eax, [esp+40]
cannam@89 754
cannam@89 755 cmp eax,ebx
cannam@89 756 jb L_clip_window_mmx
cannam@89 757
cannam@89 758 mov ecx,edx
cannam@89 759 mov esi,edi
cannam@89 760 sub esi,ebx
cannam@89 761
cannam@89 762 sub ecx,3
cannam@89 763 mov al, [esi]
cannam@89 764 mov [edi],al
cannam@89 765 mov al, [esi+1]
cannam@89 766 mov dl, [esi+2]
cannam@89 767 add esi,3
cannam@89 768 mov [edi+1],al
cannam@89 769 mov [edi+2],dl
cannam@89 770 add edi,3
cannam@89 771 rep movsb
cannam@89 772
cannam@89 773 mov esi, [esp+44]
cannam@89 774 mov ebx, [esp+8]
cannam@89 775 jmp L_while_test_mmx
cannam@89 776
cannam@89 777 ALIGN 4
cannam@89 778 L_check_dist_one_mmx:
cannam@89 779 cmp ebx,1
cannam@89 780 jne L_check_window_mmx
cannam@89 781 cmp [esp+40],edi
cannam@89 782 je L_check_window_mmx
cannam@89 783
cannam@89 784 dec edi
cannam@89 785 mov ecx,edx
cannam@89 786 mov al, [edi]
cannam@89 787 sub ecx,3
cannam@89 788
cannam@89 789 mov [edi+1],al
cannam@89 790 mov [edi+2],al
cannam@89 791 mov [edi+3],al
cannam@89 792 add edi,4
cannam@89 793 rep stosb
cannam@89 794
cannam@89 795 mov ebx, [esp+8]
cannam@89 796 jmp L_while_test_mmx
cannam@89 797
cannam@89 798 ALIGN 4
cannam@89 799 L_test_for_second_level_length_mmx:
cannam@89 800 test al,64
cannam@89 801 jnz L_test_for_end_of_block
cannam@89 802
cannam@89 803 and eax,15
cannam@89 804 psrlq mm0,mm1
cannam@89 805 movd ecx,mm0
cannam@89 806 and ecx, [inflate_fast_mask+eax*4]
cannam@89 807 add ecx,edx
cannam@89 808 mov eax, [ebx+ecx*4]
cannam@89 809 jmp L_dolen_mmx
cannam@89 810
cannam@89 811 ALIGN 4
cannam@89 812 L_test_for_second_level_dist_mmx:
cannam@89 813 test al,64
cannam@89 814 jnz L_invalid_distance_code
cannam@89 815
cannam@89 816 and eax,15
cannam@89 817 psrlq mm0,mm1
cannam@89 818 movd ecx,mm0
cannam@89 819 and ecx, [inflate_fast_mask+eax*4]
cannam@89 820 mov eax, [esp+12]
cannam@89 821 add ecx,ebx
cannam@89 822 mov eax, [eax+ecx*4]
cannam@89 823 jmp L_dodist_mmx
cannam@89 824
cannam@89 825 ALIGN 4
cannam@89 826 L_clip_window_mmx:
cannam@89 827
cannam@89 828 mov ecx,eax
cannam@89 829 mov eax, [esp+52]
cannam@89 830 neg ecx
cannam@89 831 mov esi, [esp+56]
cannam@89 832
cannam@89 833 cmp eax,ebx
cannam@89 834 jb L_invalid_distance_too_far
cannam@89 835
cannam@89 836 add ecx,ebx
cannam@89 837 cmp dword ptr [esp+48],0
cannam@89 838 jne L_wrap_around_window_mmx
cannam@89 839
cannam@89 840 sub eax,ecx
cannam@89 841 add esi,eax
cannam@89 842
cannam@89 843 cmp edx,ecx
cannam@89 844 jbe L_do_copy1_mmx
cannam@89 845
cannam@89 846 sub edx,ecx
cannam@89 847 rep movsb
cannam@89 848 mov esi,edi
cannam@89 849 sub esi,ebx
cannam@89 850 jmp L_do_copy1_mmx
cannam@89 851
cannam@89 852 cmp edx,ecx
cannam@89 853 jbe L_do_copy1_mmx
cannam@89 854
cannam@89 855 sub edx,ecx
cannam@89 856 rep movsb
cannam@89 857 mov esi,edi
cannam@89 858 sub esi,ebx
cannam@89 859 jmp L_do_copy1_mmx
cannam@89 860
cannam@89 861 L_wrap_around_window_mmx:
cannam@89 862
cannam@89 863 mov eax, [esp+48]
cannam@89 864 cmp ecx,eax
cannam@89 865 jbe L_contiguous_in_window_mmx
cannam@89 866
cannam@89 867 add esi, [esp+52]
cannam@89 868 add esi,eax
cannam@89 869 sub esi,ecx
cannam@89 870 sub ecx,eax
cannam@89 871
cannam@89 872
cannam@89 873 cmp edx,ecx
cannam@89 874 jbe L_do_copy1_mmx
cannam@89 875
cannam@89 876 sub edx,ecx
cannam@89 877 rep movsb
cannam@89 878 mov esi, [esp+56]
cannam@89 879 mov ecx, [esp+48]
cannam@89 880 cmp edx,ecx
cannam@89 881 jbe L_do_copy1_mmx
cannam@89 882
cannam@89 883 sub edx,ecx
cannam@89 884 rep movsb
cannam@89 885 mov esi,edi
cannam@89 886 sub esi,ebx
cannam@89 887 jmp L_do_copy1_mmx
cannam@89 888
cannam@89 889 L_contiguous_in_window_mmx:
cannam@89 890
cannam@89 891 add esi,eax
cannam@89 892 sub esi,ecx
cannam@89 893
cannam@89 894
cannam@89 895 cmp edx,ecx
cannam@89 896 jbe L_do_copy1_mmx
cannam@89 897
cannam@89 898 sub edx,ecx
cannam@89 899 rep movsb
cannam@89 900 mov esi,edi
cannam@89 901 sub esi,ebx
cannam@89 902
cannam@89 903 L_do_copy1_mmx:
cannam@89 904
cannam@89 905
cannam@89 906 mov ecx,edx
cannam@89 907 rep movsb
cannam@89 908
cannam@89 909 mov esi, [esp+44]
cannam@89 910 mov ebx, [esp+8]
cannam@89 911 jmp L_while_test_mmx
cannam@89 912 ; 1174 "inffast.S"
cannam@89 913 L_invalid_distance_code:
cannam@89 914
cannam@89 915
cannam@89 916
cannam@89 917
cannam@89 918
cannam@89 919 mov ecx, invalid_distance_code_msg
cannam@89 920 mov edx,INFLATE_MODE_BAD
cannam@89 921 jmp L_update_stream_state
cannam@89 922
cannam@89 923 L_test_for_end_of_block:
cannam@89 924
cannam@89 925
cannam@89 926
cannam@89 927
cannam@89 928
cannam@89 929 test al,32
cannam@89 930 jz L_invalid_literal_length_code
cannam@89 931
cannam@89 932 mov ecx,0
cannam@89 933 mov edx,INFLATE_MODE_TYPE
cannam@89 934 jmp L_update_stream_state
cannam@89 935
cannam@89 936 L_invalid_literal_length_code:
cannam@89 937
cannam@89 938
cannam@89 939
cannam@89 940
cannam@89 941
cannam@89 942 mov ecx, invalid_literal_length_code_msg
cannam@89 943 mov edx,INFLATE_MODE_BAD
cannam@89 944 jmp L_update_stream_state
cannam@89 945
cannam@89 946 L_invalid_distance_too_far:
cannam@89 947
cannam@89 948
cannam@89 949
cannam@89 950 mov esi, [esp+44]
cannam@89 951 mov ecx, invalid_distance_too_far_msg
cannam@89 952 mov edx,INFLATE_MODE_BAD
cannam@89 953 jmp L_update_stream_state
cannam@89 954
cannam@89 955 L_update_stream_state:
cannam@89 956
cannam@89 957 mov eax, [esp+88]
cannam@89 958 test ecx,ecx
cannam@89 959 jz L_skip_msg
cannam@89 960 mov [eax+24],ecx
cannam@89 961 L_skip_msg:
cannam@89 962 mov eax, [eax+28]
cannam@89 963 mov [eax+mode_state],edx
cannam@89 964 jmp L_break_loop
cannam@89 965
cannam@89 966 ALIGN 4
cannam@89 967 L_break_loop:
cannam@89 968 ; 1243 "inffast.S"
cannam@89 969 cmp dword ptr [inflate_fast_use_mmx],2
cannam@89 970 jne L_update_next_in
cannam@89 971
cannam@89 972
cannam@89 973
cannam@89 974 mov ebx,ebp
cannam@89 975
cannam@89 976 L_update_next_in:
cannam@89 977 ; 1266 "inffast.S"
cannam@89 978 mov eax, [esp+88]
cannam@89 979 mov ecx,ebx
cannam@89 980 mov edx, [eax+28]
cannam@89 981 shr ecx,3
cannam@89 982 sub esi,ecx
cannam@89 983 shl ecx,3
cannam@89 984 sub ebx,ecx
cannam@89 985 mov [eax+12],edi
cannam@89 986 mov [edx+bits_state],ebx
cannam@89 987 mov ecx,ebx
cannam@89 988
cannam@89 989 lea ebx, [esp+28]
cannam@89 990 cmp [esp+20],ebx
cannam@89 991 jne L_buf_not_used
cannam@89 992
cannam@89 993 sub esi,ebx
cannam@89 994 mov ebx, [eax+0]
cannam@89 995 mov [esp+20],ebx
cannam@89 996 add esi,ebx
cannam@89 997 mov ebx, [eax+4]
cannam@89 998 sub ebx,11
cannam@89 999 add [esp+20],ebx
cannam@89 1000
cannam@89 1001 L_buf_not_used:
cannam@89 1002 mov [eax+0],esi
cannam@89 1003
cannam@89 1004 mov ebx,1
cannam@89 1005 shl ebx,cl
cannam@89 1006 dec ebx
cannam@89 1007
cannam@89 1008
cannam@89 1009
cannam@89 1010
cannam@89 1011
cannam@89 1012 cmp dword ptr [inflate_fast_use_mmx],2
cannam@89 1013 jne L_update_hold
cannam@89 1014
cannam@89 1015
cannam@89 1016
cannam@89 1017 psrlq mm0,mm1
cannam@89 1018 movd ebp,mm0
cannam@89 1019
cannam@89 1020 emms
cannam@89 1021
cannam@89 1022 L_update_hold:
cannam@89 1023
cannam@89 1024
cannam@89 1025
cannam@89 1026 and ebp,ebx
cannam@89 1027 mov [edx+hold_state],ebp
cannam@89 1028
cannam@89 1029
cannam@89 1030
cannam@89 1031
cannam@89 1032 mov ebx, [esp+20]
cannam@89 1033 cmp ebx,esi
cannam@89 1034 jbe L_last_is_smaller
cannam@89 1035
cannam@89 1036 sub ebx,esi
cannam@89 1037 add ebx,11
cannam@89 1038 mov [eax+4],ebx
cannam@89 1039 jmp L_fixup_out
cannam@89 1040 L_last_is_smaller:
cannam@89 1041 sub esi,ebx
cannam@89 1042 neg esi
cannam@89 1043 add esi,11
cannam@89 1044 mov [eax+4],esi
cannam@89 1045
cannam@89 1046
cannam@89 1047
cannam@89 1048
cannam@89 1049 L_fixup_out:
cannam@89 1050
cannam@89 1051 mov ebx, [esp+16]
cannam@89 1052 cmp ebx,edi
cannam@89 1053 jbe L_end_is_smaller
cannam@89 1054
cannam@89 1055 sub ebx,edi
cannam@89 1056 add ebx,257
cannam@89 1057 mov [eax+16],ebx
cannam@89 1058 jmp L_done
cannam@89 1059 L_end_is_smaller:
cannam@89 1060 sub edi,ebx
cannam@89 1061 neg edi
cannam@89 1062 add edi,257
cannam@89 1063 mov [eax+16],edi
cannam@89 1064
cannam@89 1065
cannam@89 1066
cannam@89 1067
cannam@89 1068
cannam@89 1069 L_done:
cannam@89 1070 add esp,64
cannam@89 1071 popfd
cannam@89 1072 pop ebx
cannam@89 1073 pop ebp
cannam@89 1074 pop esi
cannam@89 1075 pop edi
cannam@89 1076 ret
cannam@89 1077 _inflate_fast endp
cannam@89 1078
cannam@89 1079 _TEXT ends
cannam@89 1080 end