Rev 750 | Rev 2288 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
431 | serge | 1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
2 | ;; ;; |
||
3 | ;; Copyright (C) KolibriOS team 2004-2007. All rights reserved. ;; |
||
4 | ;; Distributed under terms of the GNU General Public License ;; |
||
5 | ;; ;; |
||
6 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
7 | |||
593 | mikedld | 8 | $Revision: 1450 $ |
9 | |||
10 | |||
467 | mikedld | 11 | ; void __stdcall unpack(void* packed_data, void* unpacked_data); |
183 | diamond | 12 | unpack: |
13 | pushad |
||
14 | mov esi, [esp+32+4] |
||
15 | mov edi, [esp+32+8] |
||
16 | mov eax, [esi+8] |
||
17 | and al, 0xC0 |
||
18 | cmp al, 0xC0 |
||
19 | jz .failed |
||
20 | mov eax, [esi+8] |
||
21 | push eax |
||
22 | add esi, 12 |
||
23 | and al, not 0xC0 |
||
1450 | diamond | 24 | dec al |
183 | diamond | 25 | jz .lzma |
26 | .failed: |
||
27 | pop eax |
||
28 | popad |
||
29 | ret 8 |
||
30 | .lzma: |
||
31 | call .lzma_unpack |
||
32 | .common: |
||
33 | pop eax |
||
34 | test al, 0x80 |
||
35 | jnz .ctr1 |
||
36 | test al, 0x40 |
||
37 | jz .ok |
||
38 | lodsd |
||
39 | mov ecx, eax |
||
40 | jecxz .ok |
||
41 | mov dl, [esi] |
||
42 | mov esi, [esp+32+8] |
||
43 | .c1: |
||
44 | lodsb |
||
45 | sub al, 0E8h |
||
46 | cmp al, 1 |
||
47 | ja .c1 |
||
48 | cmp byte [esi], dl |
||
49 | jnz .c1 |
||
50 | lodsd |
||
51 | ; "bswap eax" is not supported on i386 |
||
52 | shr ax, 8 |
||
53 | ror eax, 16 |
||
54 | xchg al, ah |
||
55 | sub eax, esi |
||
56 | add eax, [esp+32+8] |
||
57 | mov [esi-4], eax |
||
58 | loop .c1 |
||
59 | .ok: |
||
60 | popad |
||
61 | ret 8 |
||
62 | .ctr1: |
||
63 | lodsd |
||
64 | mov ecx, eax |
||
65 | jecxz .ok |
||
66 | mov dl, [esi] |
||
67 | mov esi, [esp+32+8] |
||
68 | .c2: |
||
69 | lodsb |
||
70 | @@: |
||
71 | cmp al, 0xF |
||
72 | jnz .f |
||
73 | lodsb |
||
74 | cmp al, 80h |
||
75 | jb @b |
||
76 | cmp al, 90h |
||
77 | jb @f |
||
78 | .f: |
||
79 | sub al, 0E8h |
||
80 | cmp al, 1 |
||
81 | ja .c2 |
||
82 | @@: |
||
83 | cmp byte [esi], dl |
||
84 | jnz .c2 |
||
85 | lodsd |
||
86 | shr ax, 8 |
||
87 | ror eax, 16 |
||
88 | xchg al, ah |
||
89 | sub eax, esi |
||
90 | add eax, [esp+32+8] |
||
91 | mov [esi-4], eax |
||
92 | loop .c2 |
||
93 | jmp .ok |
||
94 | |||
95 | .lzma_unpack: |
||
96 | |||
97 | .pb = 2 ; pos state bits |
||
98 | .lp = 0 ; literal pos state bits |
||
99 | .lc = 3 ; literal context bits |
||
100 | .posStateMask = ((1 shl .pb)-1) |
||
101 | .literalPosMask = ((1 shl .lp)-1) |
||
102 | |||
103 | .kNumPosBitsMax = 4 |
||
104 | .kNumPosStatesMax = (1 shl .kNumPosBitsMax) |
||
105 | |||
106 | .kLenNumLowBits = 3 |
||
107 | .kLenNumLowSymbols = (1 shl .kLenNumLowBits) |
||
108 | .kLenNumMidBits = 3 |
||
109 | .kLenNumMidSymbols = (1 shl .kLenNumMidBits) |
||
110 | .kLenNumHighBits = 8 |
||
111 | .kLenNumHighSymbols = (1 shl .kLenNumHighBits) |
||
112 | |||
113 | .LenChoice = 0 |
||
114 | .LenChoice2 = 1 |
||
115 | .LenLow = 2 |
||
116 | .LenMid = (.LenLow + (.kNumPosStatesMax shl .kLenNumLowBits)) |
||
117 | .LenHigh = (.LenMid + (.kNumPosStatesMax shl .kLenNumMidBits)) |
||
118 | .kNumLenProbs = (.LenHigh + .kLenNumHighSymbols) |
||
119 | |||
120 | .kNumStates = 12 |
||
121 | .kNumLitStates = 7 |
||
122 | .kStartPosModelIndex = 4 |
||
123 | .kEndPosModelIndex = 14 |
||
124 | .kNumFullDistances = (1 shl (.kEndPosModelIndex/2)) |
||
125 | .kNumPosSlotBits = 6 |
||
126 | .kNumLenToPosStates = 4 |
||
127 | .kNumAlignBits = 4 |
||
128 | .kAlignTableSize = (1 shl .kNumAlignBits) |
||
129 | .kMatchMinLen = 2 |
||
130 | |||
131 | .IsMatch = 0 |
||
132 | .IsRep = (.IsMatch + (.kNumStates shl .kNumPosBitsMax)) |
||
133 | .IsRepG0 = (.IsRep + .kNumStates) |
||
134 | .IsRepG1 = (.IsRepG0 + .kNumStates) |
||
135 | .IsRepG2 = (.IsRepG1 + .kNumStates) |
||
136 | .IsRep0Long = (.IsRepG2 + .kNumStates) |
||
137 | .PosSlot = (.IsRep0Long + (.kNumStates shl .kNumPosBitsMax)) |
||
138 | .SpecPos = (.PosSlot + (.kNumLenToPosStates shl .kNumPosSlotBits)) |
||
139 | .Align_ = (.SpecPos + .kNumFullDistances - .kEndPosModelIndex) |
||
140 | .Lencoder = (.Align_ + .kAlignTableSize) |
||
141 | .RepLencoder = (.Lencoder + .kNumLenProbs) |
||
142 | .Literal = (.RepLencoder + .kNumLenProbs) |
||
143 | |||
144 | .LZMA_BASE_SIZE = 1846 ; must be ==Literal |
||
145 | .LZMA_LIT_SIZE = 768 |
||
146 | |||
147 | .kNumTopBits = 24 |
||
148 | .kTopValue = (1 shl .kNumTopBits) |
||
149 | |||
150 | .kNumBitModelTotalBits = 11 |
||
151 | .kBitModelTotal = (1 shl .kNumBitModelTotalBits) |
||
152 | .kNumMoveBits = 5 |
||
153 | |||
154 | push edi |
||
155 | ; int state=0; |
||
156 | xor ebx, ebx |
||
157 | mov [.previousByte], bl |
||
158 | ; unsigned rep0=1,rep1=1,rep2=1,rep3=1; |
||
159 | mov eax, 1 |
||
160 | mov edi, .rep0 |
||
161 | stosd |
||
162 | stosd |
||
163 | stosd |
||
164 | stosd |
||
165 | ; int len=0; |
||
166 | ; result=0; |
||
167 | mov ecx, .Literal + (.LZMA_LIT_SIZE shl (.lc+.lp)) |
||
168 | mov eax, .kBitModelTotal/2 |
||
709 | diamond | 169 | mov edi, [.p] |
183 | diamond | 170 | rep stosd |
171 | ; RangeDecoderInit |
||
172 | ; rd->ExtraBytes = 0 |
||
173 | ; rd->Buffer = stream |
||
174 | ; rd->BufferLim = stream+bufferSize |
||
175 | ; rd->Range = 0xFFFFFFFF |
||
176 | pop edi |
||
177 | mov ebp, [esi-8] ; dest_length |
||
178 | add ebp, edi ; ebp = destination limit |
||
179 | lodsd |
||
180 | ; rd->code_ = eax |
||
181 | mov [.code_], eax |
||
182 | or [.range], -1 |
||
183 | .main_loop: |
||
184 | cmp edi, ebp |
||
185 | jae .main_loop_done |
||
186 | mov edx, edi |
||
187 | and edx, .posStateMask |
||
188 | mov eax, ebx |
||
189 | shl eax, .kNumPosBitsMax+2 |
||
709 | diamond | 190 | lea eax, [.IsMatch*4 + eax + edx*4] |
191 | add eax, [.p] |
||
183 | diamond | 192 | call .RangeDecoderBitDecode |
193 | jc .1 |
||
194 | movzx eax, [.previousByte] |
||
195 | if .literalPosMask |
||
196 | mov ah, dl |
||
197 | and ah, .literalPosMask |
||
198 | end if |
||
199 | shr eax, 8-.lc |
||
200 | imul eax, .LZMA_LIT_SIZE*4 |
||
709 | diamond | 201 | add eax, .Literal*4 |
202 | add eax, [.p] |
||
183 | diamond | 203 | cmp ebx, .kNumLitStates |
204 | jb .literal |
||
205 | xor edx, edx |
||
206 | sub edx, [.rep0] |
||
207 | mov dl, [edi + edx] |
||
208 | call .LzmaLiteralDecodeMatch |
||
209 | jmp @f |
||
210 | .literal: |
||
211 | call .LzmaLiteralDecode |
||
212 | @@: |
||
213 | mov [.previousByte], al |
||
214 | stosb |
||
215 | mov al, bl |
||
216 | cmp bl, 4 |
||
217 | jb @f |
||
218 | mov al, 3 |
||
219 | cmp bl, 10 |
||
220 | jb @f |
||
221 | mov al, 6 |
||
222 | @@: sub bl, al |
||
223 | jmp .main_loop |
||
224 | .1: |
||
709 | diamond | 225 | lea eax, [.IsRep*4 + ebx*4] |
226 | add eax, [.p] |
||
183 | diamond | 227 | call .RangeDecoderBitDecode |
228 | jnc .10 |
||
709 | diamond | 229 | lea eax, [.IsRepG0*4 + ebx*4] |
230 | add eax, [.p] |
||
183 | diamond | 231 | call .RangeDecoderBitDecode |
232 | jc .111 |
||
233 | mov eax, ebx |
||
234 | shl eax, .kNumPosBitsMax+2 |
||
709 | diamond | 235 | lea eax, [.IsRep0Long*4 + eax + edx*4] |
236 | add eax, [.p] |
||
183 | diamond | 237 | call .RangeDecoderBitDecode |
238 | jc .1101 |
||
239 | cmp bl, 7 |
||
240 | setae bl |
||
241 | lea ebx, [9 + ebx + ebx] |
||
242 | xor edx, edx |
||
243 | sub edx, [.rep0] |
||
244 | mov al, [edi + edx] |
||
245 | stosb |
||
246 | mov [.previousByte], al |
||
247 | jmp .main_loop |
||
248 | .111: |
||
709 | diamond | 249 | lea eax, [.IsRepG1*4 + ebx*4] |
250 | add eax, [.p] |
||
183 | diamond | 251 | call .RangeDecoderBitDecode |
252 | mov eax, [.rep1] |
||
253 | jnc .l3 |
||
254 | .l1: |
||
709 | diamond | 255 | lea eax, [.IsRepG2*4 + ebx*4] |
256 | add eax, [.p] |
||
183 | diamond | 257 | call .RangeDecoderBitDecode |
258 | mov eax, [.rep2] |
||
259 | jnc .l2 |
||
260 | xchg [.rep3], eax |
||
261 | .l2: |
||
262 | push [.rep1] |
||
263 | pop [.rep2] |
||
264 | .l3: |
||
265 | xchg eax, [.rep0] |
||
266 | mov [.rep1], eax |
||
267 | .1101: |
||
709 | diamond | 268 | mov eax, .RepLencoder*4 |
269 | add eax, [.p] |
||
183 | diamond | 270 | call .LzmaLenDecode |
271 | cmp bl, 7 |
||
272 | setc bl |
||
273 | adc bl, bl |
||
274 | xor bl, 3 |
||
275 | add bl, 8 |
||
276 | jmp .repmovsb |
||
277 | .10: |
||
278 | mov eax, [.rep0] |
||
279 | xchg eax, [.rep1] |
||
280 | xchg eax, [.rep2] |
||
281 | xchg eax, [.rep3] |
||
282 | cmp bl, 7 |
||
283 | setc bl |
||
284 | adc bl, bl |
||
285 | xor bl, 3 |
||
286 | add bl, 7 |
||
709 | diamond | 287 | mov eax, .Lencoder*4 |
288 | add eax, [.p] |
||
183 | diamond | 289 | call .LzmaLenDecode |
290 | mov eax, .kNumLenToPosStates-1 |
||
291 | cmp eax, ecx |
||
292 | jb @f |
||
293 | mov eax, ecx |
||
294 | @@: |
||
295 | push ecx |
||
296 | mov ecx, .kNumPosSlotBits |
||
297 | shl eax, cl |
||
298 | shl eax, 2 |
||
709 | diamond | 299 | add eax, .PosSlot*4 |
300 | add eax, [.p] |
||
183 | diamond | 301 | call .RangeDecoderBitTreeDecode |
302 | mov [.rep0], ecx |
||
303 | cmp ecx, .kStartPosModelIndex |
||
304 | jb .l6 |
||
305 | push ecx |
||
306 | mov eax, ecx |
||
307 | and eax, 1 |
||
308 | shr ecx, 1 |
||
309 | or eax, 2 |
||
310 | dec ecx |
||
311 | shl eax, cl |
||
312 | mov [.rep0], eax |
||
313 | pop edx |
||
314 | cmp edx, .kEndPosModelIndex |
||
315 | jae .l5 |
||
316 | sub eax, edx |
||
317 | shl eax, 2 |
||
709 | diamond | 318 | add eax, (.SpecPos - 1)*4 |
319 | add eax, [.p] |
||
183 | diamond | 320 | call .RangeDecoderReverseBitTreeDecode |
321 | add [.rep0], ecx |
||
322 | jmp .l6 |
||
323 | .l5: |
||
324 | sub ecx, .kNumAlignBits |
||
325 | call .RangeDecoderDecodeDirectBits |
||
326 | mov ecx, .kNumAlignBits |
||
327 | shl eax, cl |
||
328 | add [.rep0], eax |
||
709 | diamond | 329 | mov eax, .Align_*4 |
330 | add eax, [.p] |
||
183 | diamond | 331 | call .RangeDecoderReverseBitTreeDecode |
332 | add [.rep0], ecx |
||
333 | .l6: |
||
334 | pop ecx |
||
335 | inc [.rep0] |
||
336 | jz .main_loop_done |
||
337 | .repmovsb: |
||
338 | add ecx, .kMatchMinLen |
||
339 | push esi |
||
340 | mov esi, edi |
||
341 | sub esi, [.rep0] |
||
342 | rep movsb |
||
343 | pop esi |
||
344 | mov al, [edi-1] |
||
345 | mov [.previousByte], al |
||
346 | jmp .main_loop |
||
347 | .main_loop_done: |
||
348 | ret |
||
349 | |||
350 | .RangeDecoderBitDecode: |
||
351 | ; in: eax->prob |
||
352 | ; out: CF=bit; destroys eax |
||
353 | push edx |
||
354 | mov edx, [.range] |
||
355 | shr edx, .kNumBitModelTotalBits |
||
356 | imul edx, [eax] |
||
357 | cmp [.code_], edx |
||
358 | jae .ae |
||
359 | mov [.range], edx |
||
360 | mov edx, .kBitModelTotal |
||
361 | sub edx, [eax] |
||
362 | shr edx, .kNumMoveBits |
||
363 | add [eax], edx |
||
364 | clc |
||
365 | .n: |
||
366 | lahf |
||
367 | cmp [.range], .kTopValue |
||
368 | jae @f |
||
369 | shl [.range], 8 |
||
370 | shl [.code_], 8 |
||
371 | lodsb |
||
372 | mov byte [.code_], al |
||
373 | @@: |
||
374 | sahf |
||
375 | pop edx |
||
376 | ret |
||
377 | .ae: |
||
378 | sub [.range], edx |
||
379 | sub [.code_], edx |
||
380 | mov edx, [eax] |
||
381 | shr edx, .kNumMoveBits |
||
382 | sub [eax], edx |
||
383 | stc |
||
384 | jmp .n |
||
385 | |||
386 | .RangeDecoderDecodeDirectBits: |
||
387 | ; in: ecx=numTotalBits |
||
388 | ; out: eax=result; destroys edx |
||
389 | xor eax, eax |
||
390 | .l: |
||
391 | shr [.range], 1 |
||
392 | shl eax, 1 |
||
393 | mov edx, [.code_] |
||
394 | sub edx, [.range] |
||
395 | jb @f |
||
396 | mov [.code_], edx |
||
397 | or eax, 1 |
||
398 | @@: |
||
399 | cmp [.range], .kTopValue |
||
400 | jae @f |
||
401 | shl [.range], 8 |
||
402 | shl [.code_], 8 |
||
403 | push eax |
||
404 | lodsb |
||
405 | mov byte [.code_], al |
||
406 | pop eax |
||
407 | @@: |
||
408 | loop .l |
||
409 | ret |
||
410 | |||
411 | .LzmaLiteralDecode: |
||
412 | ; in: eax->probs |
||
413 | ; out: al=byte; destroys edx |
||
414 | push ecx |
||
415 | mov ecx, 1 |
||
416 | @@: |
||
417 | push eax |
||
418 | lea eax, [eax+ecx*4] |
||
419 | call .RangeDecoderBitDecode |
||
420 | pop eax |
||
421 | adc cl, cl |
||
422 | jnc @b |
||
423 | .LzmaLiteralDecode.ret: |
||
424 | mov al, cl |
||
425 | pop ecx |
||
426 | ret |
||
427 | .LzmaLiteralDecodeMatch: |
||
428 | ; in: eax->probs, dl=matchByte |
||
429 | ; out: al=byte; destroys edx |
||
430 | push ecx |
||
431 | mov ecx, 1 |
||
432 | .LzmaLiteralDecodeMatch.1: |
||
433 | add dl, dl |
||
434 | setc ch |
||
435 | push eax |
||
436 | lea eax, [eax+ecx*4+0x100*4] |
||
437 | call .RangeDecoderBitDecode |
||
438 | pop eax |
||
439 | adc cl, cl |
||
440 | jc .LzmaLiteralDecode.ret |
||
441 | xor ch, cl |
||
442 | test ch, 1 |
||
443 | mov ch, 0 |
||
444 | jnz @b |
||
445 | jmp .LzmaLiteralDecodeMatch.1 |
||
446 | |||
447 | .LzmaLenDecode: |
||
448 | ; in: eax->prob, edx=posState |
||
449 | ; out: ecx=len |
||
450 | push eax |
||
451 | add eax, .LenChoice*4 |
||
452 | call .RangeDecoderBitDecode |
||
453 | pop eax |
||
454 | jnc .0 |
||
455 | push eax |
||
456 | add eax, .LenChoice2*4 |
||
457 | call .RangeDecoderBitDecode |
||
458 | pop eax |
||
459 | jc @f |
||
460 | mov ecx, .kLenNumMidBits |
||
461 | shl edx, cl |
||
462 | lea eax, [eax + .LenMid*4 + edx*4] |
||
463 | call .RangeDecoderBitTreeDecode |
||
464 | add ecx, .kLenNumLowSymbols |
||
465 | ret |
||
466 | @@: |
||
467 | add eax, .LenHigh*4 |
||
468 | mov ecx, .kLenNumHighBits |
||
469 | call .RangeDecoderBitTreeDecode |
||
470 | add ecx, .kLenNumLowSymbols + .kLenNumMidSymbols |
||
471 | ret |
||
472 | .0: |
||
473 | mov ecx, .kLenNumLowBits |
||
474 | shl edx, cl |
||
475 | lea eax, [eax + .LenLow*4 + edx*4] |
||
476 | .RangeDecoderBitTreeDecode: |
||
477 | ; in: eax->probs,ecx=numLevels |
||
478 | ; out: ecx=length; destroys edx |
||
479 | push ebx |
||
480 | mov edx, 1 |
||
481 | mov ebx, edx |
||
482 | @@: |
||
483 | push eax |
||
484 | lea eax, [eax+edx*4] |
||
485 | call .RangeDecoderBitDecode |
||
486 | pop eax |
||
487 | adc dl, dl |
||
488 | add bl, bl |
||
489 | loop @b |
||
490 | sub dl, bl |
||
491 | pop ebx |
||
492 | mov ecx, edx |
||
493 | ret |
||
494 | .RangeDecoderReverseBitTreeDecode: |
||
495 | ; in: eax->probs,ecx=numLevels |
||
496 | ; out: ecx=length; destroys edx |
||
497 | push ebx ecx |
||
498 | mov edx, 1 |
||
499 | xor ebx, ebx |
||
500 | @@: |
||
501 | push eax |
||
502 | lea eax, [eax+edx*4] |
||
503 | call .RangeDecoderBitDecode |
||
504 | lahf |
||
505 | adc edx, edx |
||
506 | sahf |
||
507 | rcr ebx, 1 |
||
508 | pop eax |
||
509 | loop @b |
||
510 | pop ecx |
||
511 | rol ebx, cl |
||
512 | mov ecx, ebx |
||
513 | pop ebx |
||
514 | ret |
||
515 | |||
516 | uglobal |
||
517 | align 4 |
||
709 | diamond | 518 | ;unpack.p rd unpack.LZMA_BASE_SIZE + (unpack.LZMA_LIT_SIZE shl (unpack.lc+unpack.lp)) |
519 | unpack.p dd ? |
||
183 | diamond | 520 | unpack.code_ dd ? |
521 | unpack.range dd ? |
||
522 | unpack.rep0 dd ? |
||
523 | unpack.rep1 dd ? |
||
524 | unpack.rep2 dd ? |
||
525 | unpack.rep3 dd ? |
||
526 | unpack.previousByte db ? |
||
527 | endg |