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