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