Rev 9029 | Go to most recent revision | Only display areas with differences | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed
Rev 9029 | Rev 9715 | ||
---|---|---|---|
1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
2 | ;; ;; |
2 | ;; ;; |
3 | ;; Copyright (C) KolibriOS team 2004-2015. All rights reserved. ;; |
3 | ;; Copyright (C) KolibriOS team 2004-2022. All rights reserved. ;; |
4 | ;; Distributed under terms of the GNU General Public License ;; |
4 | ;; Distributed under terms of the GNU General Public License ;; |
5 | ;; ;; |
5 | ;; ;; |
6 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
6 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
7 | 7 | ||
8 | $Revision: 9029 $ |
8 | $Revision: 9715 $ |
9 | 9 | ||
10 | 10 | ||
11 | ; @brief Unpack LZMA-compressed data. C-style declaration of the function. |
11 | ; @brief Unpack LZMA-compressed data. C-style declaration of the function. |
12 | ; |
12 | ; |
13 | ; ```void __stdcall unpack(void *packed_data, void *unpacked_data);``` |
13 | ; ```void __stdcall unpack(void *packed_data, void *unpacked_data);``` |
14 | ; |
14 | ; |
15 | ; @param packed_data Data to unpack |
15 | ; @param packed_data Data to unpack |
16 | ; @param unpacked_data Buffer to hold unpacked data |
16 | ; @param unpacked_data Buffer to hold unpacked data |
17 | ; @returns Nothing |
17 | ; @returns Nothing |
18 | unpack: |
18 | unpack: |
19 | pushad |
19 | pushad |
20 | mov esi, [esp+32+4] |
20 | mov esi, [esp+32+4] |
21 | mov edi, [esp+32+8] |
21 | mov edi, [esp+32+8] |
22 | mov eax, [esi+8] |
22 | mov eax, [esi+8] |
23 | and al, 0xC0 |
23 | and al, 0xC0 |
24 | cmp al, 0xC0 |
24 | cmp al, 0xC0 |
25 | jz .failed |
25 | jz .failed |
26 | mov eax, [esi+8] |
26 | mov eax, [esi+8] |
27 | push eax |
27 | push eax |
28 | add esi, 12 |
28 | add esi, 12 |
29 | and al, not 0xC0 |
29 | and al, not 0xC0 |
30 | dec al |
30 | dec al |
31 | jz .lzma |
31 | jz .lzma |
32 | .failed: |
32 | .failed: |
33 | pop eax |
33 | pop eax |
34 | popad |
34 | popad |
35 | ret 8 |
35 | ret 8 |
36 | .lzma: |
36 | .lzma: |
37 | call .lzma_unpack |
37 | call .lzma_unpack |
38 | .common: |
38 | .common: |
39 | pop eax |
39 | pop eax |
40 | test al, 0x80 |
40 | test al, 0x80 |
41 | jnz .ctr1 |
41 | jnz .ctr1 |
42 | test al, 0x40 |
42 | test al, 0x40 |
43 | jz .ok |
43 | jz .ok |
44 | lodsd |
44 | lodsd |
45 | mov ecx, eax |
45 | mov ecx, eax |
46 | jecxz .ok |
46 | jecxz .ok |
47 | mov dl, [esi] |
47 | mov dl, [esi] |
48 | mov esi, [esp+32+8] |
48 | mov esi, [esp+32+8] |
49 | .c1: |
49 | .c1: |
50 | lodsb |
50 | lodsb |
51 | sub al, 0E8h |
51 | sub al, 0E8h |
52 | cmp al, 1 |
52 | cmp al, 1 |
53 | ja .c1 |
53 | ja .c1 |
54 | cmp byte [esi], dl |
54 | cmp byte [esi], dl |
55 | jnz .c1 |
55 | jnz .c1 |
56 | lodsd |
56 | lodsd |
57 | ; "bswap eax" is not supported on i386 |
57 | ; "bswap eax" is not supported on i386 |
58 | shr ax, 8 |
58 | shr ax, 8 |
59 | ror eax, 16 |
59 | ror eax, 16 |
60 | xchg al, ah |
60 | xchg al, ah |
61 | sub eax, esi |
61 | sub eax, esi |
62 | add eax, [esp+32+8] |
62 | add eax, [esp+32+8] |
63 | mov [esi-4], eax |
63 | mov [esi-4], eax |
64 | loop .c1 |
64 | loop .c1 |
65 | .ok: |
65 | .ok: |
66 | popad |
66 | popad |
67 | ret 8 |
67 | ret 8 |
68 | .ctr1: |
68 | .ctr1: |
69 | lodsd |
69 | lodsd |
70 | mov ecx, eax |
70 | mov ecx, eax |
71 | jecxz .ok |
71 | jecxz .ok |
72 | mov dl, [esi] |
72 | mov dl, [esi] |
73 | mov esi, [esp+32+8] |
73 | mov esi, [esp+32+8] |
74 | .c2: |
74 | .c2: |
75 | lodsb |
75 | lodsb |
76 | @@: |
76 | @@: |
77 | cmp al, 0xF |
77 | cmp al, 0xF |
78 | jnz .f |
78 | jnz .f |
79 | lodsb |
79 | lodsb |
80 | cmp al, 80h |
80 | cmp al, 80h |
81 | jb @b |
81 | jb @b |
82 | cmp al, 90h |
82 | cmp al, 90h |
83 | jb @f |
83 | jb @f |
84 | .f: |
84 | .f: |
85 | sub al, 0E8h |
85 | sub al, 0E8h |
86 | cmp al, 1 |
86 | cmp al, 1 |
87 | ja .c2 |
87 | ja .c2 |
88 | @@: |
88 | @@: |
89 | cmp byte [esi], dl |
89 | cmp byte [esi], dl |
90 | jnz .c2 |
90 | jnz .c2 |
91 | lodsd |
91 | lodsd |
92 | shr ax, 8 |
92 | shr ax, 8 |
93 | ror eax, 16 |
93 | ror eax, 16 |
94 | xchg al, ah |
94 | xchg al, ah |
95 | sub eax, esi |
95 | sub eax, esi |
96 | add eax, [esp+32+8] |
96 | add eax, [esp+32+8] |
97 | mov [esi-4], eax |
97 | mov [esi-4], eax |
98 | loop .c2 |
98 | loop .c2 |
99 | jmp .ok |
99 | jmp .ok |
100 | 100 | ||
101 | .lzma_unpack: |
101 | .lzma_unpack: |
102 | 102 | ||
103 | .pb = 2 ; pos state bits |
103 | .pb = 2 ; pos state bits |
104 | .lp = 0 ; literal pos state bits |
104 | .lp = 0 ; literal pos state bits |
105 | .lc = 3 ; literal context bits |
105 | .lc = 3 ; literal context bits |
106 | .posStateMask = ((1 shl .pb)-1) |
106 | .posStateMask = ((1 shl .pb)-1) |
107 | .literalPosMask = ((1 shl .lp)-1) |
107 | .literalPosMask = ((1 shl .lp)-1) |
108 | 108 | ||
109 | .kNumPosBitsMax = 4 |
109 | .kNumPosBitsMax = 4 |
110 | .kNumPosStatesMax = (1 shl .kNumPosBitsMax) |
110 | .kNumPosStatesMax = (1 shl .kNumPosBitsMax) |
111 | 111 | ||
112 | .kLenNumLowBits = 3 |
112 | .kLenNumLowBits = 3 |
113 | .kLenNumLowSymbols = (1 shl .kLenNumLowBits) |
113 | .kLenNumLowSymbols = (1 shl .kLenNumLowBits) |
114 | .kLenNumMidBits = 3 |
114 | .kLenNumMidBits = 3 |
115 | .kLenNumMidSymbols = (1 shl .kLenNumMidBits) |
115 | .kLenNumMidSymbols = (1 shl .kLenNumMidBits) |
116 | .kLenNumHighBits = 8 |
116 | .kLenNumHighBits = 8 |
117 | .kLenNumHighSymbols = (1 shl .kLenNumHighBits) |
117 | .kLenNumHighSymbols = (1 shl .kLenNumHighBits) |
118 | 118 | ||
119 | .LenChoice = 0 |
119 | .LenChoice = 0 |
120 | .LenChoice2 = 1 |
120 | .LenChoice2 = 1 |
121 | .LenLow = 2 |
121 | .LenLow = 2 |
122 | .LenMid = (.LenLow + (.kNumPosStatesMax shl .kLenNumLowBits)) |
122 | .LenMid = (.LenLow + (.kNumPosStatesMax shl .kLenNumLowBits)) |
123 | .LenHigh = (.LenMid + (.kNumPosStatesMax shl .kLenNumMidBits)) |
123 | .LenHigh = (.LenMid + (.kNumPosStatesMax shl .kLenNumMidBits)) |
124 | .kNumLenProbs = (.LenHigh + .kLenNumHighSymbols) |
124 | .kNumLenProbs = (.LenHigh + .kLenNumHighSymbols) |
125 | 125 | ||
126 | .kNumStates = 12 |
126 | .kNumStates = 12 |
127 | .kNumLitStates = 7 |
127 | .kNumLitStates = 7 |
128 | .kStartPosModelIndex = 4 |
128 | .kStartPosModelIndex = 4 |
129 | .kEndPosModelIndex = 14 |
129 | .kEndPosModelIndex = 14 |
130 | .kNumFullDistances = (1 shl (.kEndPosModelIndex/2)) |
130 | .kNumFullDistances = (1 shl (.kEndPosModelIndex/2)) |
131 | .kNumPosSlotBits = 6 |
131 | .kNumPosSlotBits = 6 |
132 | .kNumLenToPosStates = 4 |
132 | .kNumLenToPosStates = 4 |
133 | .kNumAlignBits = 4 |
133 | .kNumAlignBits = 4 |
134 | .kAlignTableSize = (1 shl .kNumAlignBits) |
134 | .kAlignTableSize = (1 shl .kNumAlignBits) |
135 | .kMatchMinLen = 2 |
135 | .kMatchMinLen = 2 |
136 | 136 | ||
137 | .IsMatch = 0 |
137 | .IsMatch = 0 |
138 | .IsRep = (.IsMatch + (.kNumStates shl .kNumPosBitsMax)) |
138 | .IsRep = (.IsMatch + (.kNumStates shl .kNumPosBitsMax)) |
139 | .IsRepG0 = (.IsRep + .kNumStates) |
139 | .IsRepG0 = (.IsRep + .kNumStates) |
140 | .IsRepG1 = (.IsRepG0 + .kNumStates) |
140 | .IsRepG1 = (.IsRepG0 + .kNumStates) |
141 | .IsRepG2 = (.IsRepG1 + .kNumStates) |
141 | .IsRepG2 = (.IsRepG1 + .kNumStates) |
142 | .IsRep0Long = (.IsRepG2 + .kNumStates) |
142 | .IsRep0Long = (.IsRepG2 + .kNumStates) |
143 | .PosSlot = (.IsRep0Long + (.kNumStates shl .kNumPosBitsMax)) |
143 | .PosSlot = (.IsRep0Long + (.kNumStates shl .kNumPosBitsMax)) |
144 | .SpecPos = (.PosSlot + (.kNumLenToPosStates shl .kNumPosSlotBits)) |
144 | .SpecPos = (.PosSlot + (.kNumLenToPosStates shl .kNumPosSlotBits)) |
145 | .Align_ = (.SpecPos + .kNumFullDistances - .kEndPosModelIndex) |
145 | .Align_ = (.SpecPos + .kNumFullDistances - .kEndPosModelIndex) |
146 | .Lencoder = (.Align_ + .kAlignTableSize) |
146 | .Lencoder = (.Align_ + .kAlignTableSize) |
147 | .RepLencoder = (.Lencoder + .kNumLenProbs) |
147 | .RepLencoder = (.Lencoder + .kNumLenProbs) |
148 | .Literal = (.RepLencoder + .kNumLenProbs) |
148 | .Literal = (.RepLencoder + .kNumLenProbs) |
149 | 149 | ||
150 | .LZMA_BASE_SIZE = 1846 ; must be ==Literal |
150 | .LZMA_BASE_SIZE = 1846 ; must be ==Literal |
151 | .LZMA_LIT_SIZE = 768 |
151 | .LZMA_LIT_SIZE = 768 |
152 | 152 | ||
153 | .kNumTopBits = 24 |
153 | .kNumTopBits = 24 |
154 | .kTopValue = (1 shl .kNumTopBits) |
154 | .kTopValue = (1 shl .kNumTopBits) |
155 | 155 | ||
156 | .kNumBitModelTotalBits = 11 |
156 | .kNumBitModelTotalBits = 11 |
157 | .kBitModelTotal = (1 shl .kNumBitModelTotalBits) |
157 | .kBitModelTotal = (1 shl .kNumBitModelTotalBits) |
158 | .kNumMoveBits = 5 |
158 | .kNumMoveBits = 5 |
159 | 159 | ||
160 | push edi |
160 | push edi |
161 | ; int state=0; |
161 | ; int state=0; |
162 | xor ebx, ebx |
162 | xor ebx, ebx |
163 | mov [.previousByte], bl |
163 | mov [.previousByte], bl |
164 | ; unsigned rep0=1,rep1=1,rep2=1,rep3=1; |
164 | ; unsigned rep0=1,rep1=1,rep2=1,rep3=1; |
165 | mov eax, 1 |
165 | mov eax, 1 |
166 | mov edi, .rep0 |
166 | mov edi, .rep0 |
167 | stosd |
167 | stosd |
168 | stosd |
168 | stosd |
169 | stosd |
169 | stosd |
170 | stosd |
170 | stosd |
171 | ; int len=0; |
171 | ; int len=0; |
172 | ; result=0; |
172 | ; result=0; |
173 | mov ecx, .Literal + (.LZMA_LIT_SIZE shl (.lc+.lp)) |
173 | mov ecx, .Literal + (.LZMA_LIT_SIZE shl (.lc+.lp)) |
174 | mov eax, .kBitModelTotal/2 |
174 | mov eax, .kBitModelTotal/2 |
175 | mov edi, [.p] |
175 | mov edi, [.p] |
176 | rep stosd |
176 | rep stosd |
177 | ; RangeDecoderInit |
177 | ; RangeDecoderInit |
178 | ; rd->ExtraBytes = 0 |
178 | ; rd->ExtraBytes = 0 |
179 | ; rd->Buffer = stream |
179 | ; rd->Buffer = stream |
180 | ; rd->BufferLim = stream+bufferSize |
180 | ; rd->BufferLim = stream+bufferSize |
181 | ; rd->Range = 0xFFFFFFFF |
181 | ; rd->Range = 0xFFFFFFFF |
182 | pop edi |
182 | pop edi |
183 | mov ebp, [esi-8] ; dest_length |
183 | mov ebp, [esi-8] ; dest_length |
184 | add ebp, edi ; ebp = destination limit |
184 | add ebp, edi ; ebp = destination limit |
185 | lodsd |
185 | lodsd |
186 | ; rd->code_ = eax |
186 | ; rd->code_ = eax |
187 | mov [.code_], eax |
187 | mov [.code_], eax |
188 | or [.range], -1 |
188 | or [.range], -1 |
189 | .main_loop: |
189 | .main_loop: |
190 | cmp edi, ebp |
190 | cmp edi, ebp |
191 | jae .main_loop_done |
191 | jae .main_loop_done |
192 | mov edx, edi |
192 | mov edx, edi |
193 | and edx, .posStateMask |
193 | and edx, .posStateMask |
194 | mov eax, ebx |
194 | mov eax, ebx |
195 | shl eax, .kNumPosBitsMax+2 |
195 | shl eax, .kNumPosBitsMax+2 |
196 | lea eax, [.IsMatch*4 + eax + edx*4] |
196 | lea eax, [.IsMatch*4 + eax + edx*4] |
197 | add eax, [.p] |
197 | add eax, [.p] |
198 | call .RangeDecoderBitDecode |
198 | call .RangeDecoderBitDecode |
199 | jc .1 |
199 | jc .1 |
200 | movzx eax, [.previousByte] |
200 | movzx eax, [.previousByte] |
201 | if .literalPosMask |
201 | if .literalPosMask |
202 | mov ah, dl |
202 | mov ah, dl |
203 | and ah, .literalPosMask |
203 | and ah, .literalPosMask |
204 | end if |
204 | end if |
205 | shr eax, 8-.lc |
205 | shr eax, 8-.lc |
206 | imul eax, .LZMA_LIT_SIZE*4 |
206 | imul eax, .LZMA_LIT_SIZE*4 |
207 | add eax, .Literal*4 |
207 | add eax, .Literal*4 |
208 | add eax, [.p] |
208 | add eax, [.p] |
209 | cmp ebx, .kNumLitStates |
209 | cmp ebx, .kNumLitStates |
210 | jb .literal |
210 | jb .literal |
211 | xor edx, edx |
211 | xor edx, edx |
212 | sub edx, [.rep0] |
212 | sub edx, [.rep0] |
213 | mov dl, [edi + edx] |
213 | mov dl, [edi + edx] |
214 | call .LzmaLiteralDecodeMatch |
214 | call .LzmaLiteralDecodeMatch |
215 | jmp @f |
215 | jmp @f |
216 | .literal: |
216 | .literal: |
217 | call .LzmaLiteralDecode |
217 | call .LzmaLiteralDecode |
218 | @@: |
218 | @@: |
219 | mov [.previousByte], al |
219 | mov [.previousByte], al |
220 | stosb |
220 | stosb |
221 | mov al, bl |
221 | mov al, bl |
222 | cmp bl, 4 |
222 | cmp bl, 4 |
223 | jb @f |
223 | jb @f |
224 | mov al, 3 |
224 | mov al, 3 |
225 | cmp bl, 10 |
225 | cmp bl, 10 |
226 | jb @f |
226 | jb @f |
227 | mov al, 6 |
227 | mov al, 6 |
228 | @@: |
228 | @@: |
229 | sub bl, al |
229 | sub bl, al |
230 | jmp .main_loop |
230 | jmp .main_loop |
231 | .1: |
231 | .1: |
232 | lea eax, [.IsRep*4 + ebx*4] |
232 | lea eax, [.IsRep*4 + ebx*4] |
233 | add eax, [.p] |
233 | add eax, [.p] |
234 | call .RangeDecoderBitDecode |
234 | call .RangeDecoderBitDecode |
235 | jnc .10 |
235 | jnc .10 |
236 | lea eax, [.IsRepG0*4 + ebx*4] |
236 | lea eax, [.IsRepG0*4 + ebx*4] |
237 | add eax, [.p] |
237 | add eax, [.p] |
238 | call .RangeDecoderBitDecode |
238 | call .RangeDecoderBitDecode |
239 | jc .111 |
239 | jc .111 |
240 | mov eax, ebx |
240 | mov eax, ebx |
241 | shl eax, .kNumPosBitsMax+2 |
241 | shl eax, .kNumPosBitsMax+2 |
242 | lea eax, [.IsRep0Long*4 + eax + edx*4] |
242 | lea eax, [.IsRep0Long*4 + eax + edx*4] |
243 | add eax, [.p] |
243 | add eax, [.p] |
244 | call .RangeDecoderBitDecode |
244 | call .RangeDecoderBitDecode |
245 | jc .1101 |
245 | jc .1101 |
246 | cmp bl, 7 |
246 | cmp bl, 7 |
247 | setae bl |
247 | setae bl |
248 | lea ebx, [9 + ebx + ebx] |
248 | lea ebx, [9 + ebx + ebx] |
249 | xor edx, edx |
249 | xor edx, edx |
250 | sub edx, [.rep0] |
250 | sub edx, [.rep0] |
251 | mov al, [edi + edx] |
251 | mov al, [edi + edx] |
252 | stosb |
252 | stosb |
253 | mov [.previousByte], al |
253 | mov [.previousByte], al |
254 | jmp .main_loop |
254 | jmp .main_loop |
255 | .111: |
255 | .111: |
256 | lea eax, [.IsRepG1*4 + ebx*4] |
256 | lea eax, [.IsRepG1*4 + ebx*4] |
257 | add eax, [.p] |
257 | add eax, [.p] |
258 | call .RangeDecoderBitDecode |
258 | call .RangeDecoderBitDecode |
259 | mov eax, [.rep1] |
259 | mov eax, [.rep1] |
260 | jnc .l3 |
260 | jnc .l3 |
261 | .l1: |
261 | .l1: |
262 | lea eax, [.IsRepG2*4 + ebx*4] |
262 | lea eax, [.IsRepG2*4 + ebx*4] |
263 | add eax, [.p] |
263 | add eax, [.p] |
264 | call .RangeDecoderBitDecode |
264 | call .RangeDecoderBitDecode |
265 | mov eax, [.rep2] |
265 | mov eax, [.rep2] |
266 | jnc .l2 |
266 | jnc .l2 |
267 | xchg [.rep3], eax |
267 | xchg [.rep3], eax |
268 | .l2: |
268 | .l2: |
269 | push [.rep1] |
269 | push [.rep1] |
270 | pop [.rep2] |
270 | pop [.rep2] |
271 | .l3: |
271 | .l3: |
272 | xchg eax, [.rep0] |
272 | xchg eax, [.rep0] |
273 | mov [.rep1], eax |
273 | mov [.rep1], eax |
274 | .1101: |
274 | .1101: |
275 | mov eax, .RepLencoder*4 |
275 | mov eax, .RepLencoder*4 |
276 | add eax, [.p] |
276 | add eax, [.p] |
277 | call .LzmaLenDecode |
277 | call .LzmaLenDecode |
278 | cmp bl, 7 |
278 | cmp bl, 7 |
279 | setc bl |
279 | setc bl |
280 | adc bl, bl |
280 | adc bl, bl |
281 | xor bl, 3 |
281 | xor bl, 3 |
282 | add bl, 8 |
282 | add bl, 8 |
283 | jmp .repmovsb |
283 | jmp .repmovsb |
284 | .10: |
284 | .10: |
285 | mov eax, [.rep0] |
285 | mov eax, [.rep0] |
286 | xchg eax, [.rep1] |
286 | xchg eax, [.rep1] |
287 | xchg eax, [.rep2] |
287 | xchg eax, [.rep2] |
288 | xchg eax, [.rep3] |
288 | xchg eax, [.rep3] |
289 | cmp bl, 7 |
289 | cmp bl, 7 |
290 | setc bl |
290 | setc bl |
291 | adc bl, bl |
291 | adc bl, bl |
292 | xor bl, 3 |
292 | xor bl, 3 |
293 | add bl, 7 |
293 | add bl, 7 |
294 | mov eax, .Lencoder*4 |
294 | mov eax, .Lencoder*4 |
295 | add eax, [.p] |
295 | add eax, [.p] |
296 | call .LzmaLenDecode |
296 | call .LzmaLenDecode |
297 | mov eax, .kNumLenToPosStates-1 |
297 | mov eax, .kNumLenToPosStates-1 |
298 | cmp eax, ecx |
298 | cmp eax, ecx |
299 | jb @f |
299 | jb @f |
300 | mov eax, ecx |
300 | mov eax, ecx |
301 | @@: |
301 | @@: |
302 | push ecx |
302 | push ecx |
303 | mov ecx, .kNumPosSlotBits |
303 | mov ecx, .kNumPosSlotBits |
304 | shl eax, cl |
304 | shl eax, cl |
305 | shl eax, 2 |
305 | shl eax, 2 |
306 | add eax, .PosSlot*4 |
306 | add eax, .PosSlot*4 |
307 | add eax, [.p] |
307 | add eax, [.p] |
308 | call .RangeDecoderBitTreeDecode |
308 | call .RangeDecoderBitTreeDecode |
309 | mov [.rep0], ecx |
309 | mov [.rep0], ecx |
310 | cmp ecx, .kStartPosModelIndex |
310 | cmp ecx, .kStartPosModelIndex |
311 | jb .l6 |
311 | jb .l6 |
312 | push ecx |
312 | push ecx |
313 | mov eax, ecx |
313 | mov eax, ecx |
314 | and eax, 1 |
314 | and eax, 1 |
315 | shr ecx, 1 |
315 | shr ecx, 1 |
316 | or eax, 2 |
316 | or eax, 2 |
317 | dec ecx |
317 | dec ecx |
318 | shl eax, cl |
318 | shl eax, cl |
319 | mov [.rep0], eax |
319 | mov [.rep0], eax |
320 | pop edx |
320 | pop edx |
321 | cmp edx, .kEndPosModelIndex |
321 | cmp edx, .kEndPosModelIndex |
322 | jae .l5 |
322 | jae .l5 |
323 | sub eax, edx |
323 | sub eax, edx |
324 | shl eax, 2 |
324 | shl eax, 2 |
325 | add eax, (.SpecPos - 1)*4 |
325 | add eax, (.SpecPos - 1)*4 |
326 | add eax, [.p] |
326 | add eax, [.p] |
327 | call .RangeDecoderReverseBitTreeDecode |
327 | call .RangeDecoderReverseBitTreeDecode |
328 | add [.rep0], ecx |
328 | add [.rep0], ecx |
329 | jmp .l6 |
329 | jmp .l6 |
330 | .l5: |
330 | .l5: |
331 | sub ecx, .kNumAlignBits |
331 | sub ecx, .kNumAlignBits |
332 | call .RangeDecoderDecodeDirectBits |
332 | call .RangeDecoderDecodeDirectBits |
333 | mov ecx, .kNumAlignBits |
333 | mov ecx, .kNumAlignBits |
334 | shl eax, cl |
334 | shl eax, cl |
335 | add [.rep0], eax |
335 | add [.rep0], eax |
336 | mov eax, .Align_*4 |
336 | mov eax, .Align_*4 |
337 | add eax, [.p] |
337 | add eax, [.p] |
338 | call .RangeDecoderReverseBitTreeDecode |
338 | call .RangeDecoderReverseBitTreeDecode |
339 | add [.rep0], ecx |
339 | add [.rep0], ecx |
340 | .l6: |
340 | .l6: |
341 | pop ecx |
341 | pop ecx |
342 | inc [.rep0] |
342 | inc [.rep0] |
343 | jz .main_loop_done |
343 | jz .main_loop_done |
344 | .repmovsb: |
344 | .repmovsb: |
345 | add ecx, .kMatchMinLen |
345 | add ecx, .kMatchMinLen |
346 | push esi |
346 | push esi |
347 | mov esi, edi |
347 | mov esi, edi |
348 | sub esi, [.rep0] |
348 | sub esi, [.rep0] |
349 | rep movsb |
349 | rep movsb |
350 | pop esi |
350 | pop esi |
351 | mov al, [edi-1] |
351 | mov al, [edi-1] |
352 | mov [.previousByte], al |
352 | mov [.previousByte], al |
353 | jmp .main_loop |
353 | jmp .main_loop |
354 | .main_loop_done: |
354 | .main_loop_done: |
355 | ret |
355 | ret |
356 | 356 | ||
357 | .RangeDecoderBitDecode: |
357 | .RangeDecoderBitDecode: |
358 | ; in: eax->prob |
358 | ; in: eax->prob |
359 | ; out: CF=bit; destroys eax |
359 | ; out: CF=bit; destroys eax |
360 | push edx |
360 | push edx |
361 | mov edx, [.range] |
361 | mov edx, [.range] |
362 | shr edx, .kNumBitModelTotalBits |
362 | shr edx, .kNumBitModelTotalBits |
363 | imul edx, [eax] |
363 | imul edx, [eax] |
364 | cmp [.code_], edx |
364 | cmp [.code_], edx |
365 | jae .ae |
365 | jae .ae |
366 | mov [.range], edx |
366 | mov [.range], edx |
367 | mov edx, .kBitModelTotal |
367 | mov edx, .kBitModelTotal |
368 | sub edx, [eax] |
368 | sub edx, [eax] |
369 | shr edx, .kNumMoveBits |
369 | shr edx, .kNumMoveBits |
370 | add [eax], edx |
370 | add [eax], edx |
371 | clc |
371 | clc |
372 | .n: |
372 | .n: |
373 | lahf |
373 | lahf |
374 | cmp [.range], .kTopValue |
374 | cmp [.range], .kTopValue |
375 | jae @f |
375 | jae @f |
376 | shl [.range], 8 |
376 | shl [.range], 8 |
377 | shl [.code_], 8 |
377 | shl [.code_], 8 |
378 | lodsb |
378 | lodsb |
379 | mov byte [.code_], al |
379 | mov byte [.code_], al |
380 | @@: |
380 | @@: |
381 | sahf |
381 | sahf |
382 | pop edx |
382 | pop edx |
383 | ret |
383 | ret |
384 | .ae: |
384 | .ae: |
385 | sub [.range], edx |
385 | sub [.range], edx |
386 | sub [.code_], edx |
386 | sub [.code_], edx |
387 | mov edx, [eax] |
387 | mov edx, [eax] |
388 | shr edx, .kNumMoveBits |
388 | shr edx, .kNumMoveBits |
389 | sub [eax], edx |
389 | sub [eax], edx |
390 | stc |
390 | stc |
391 | jmp .n |
391 | jmp .n |
392 | 392 | ||
393 | .RangeDecoderDecodeDirectBits: |
393 | .RangeDecoderDecodeDirectBits: |
394 | ; in: ecx=numTotalBits |
394 | ; in: ecx=numTotalBits |
395 | ; out: eax=result; destroys edx |
395 | ; out: eax=result; destroys edx |
396 | xor eax, eax |
396 | xor eax, eax |
397 | .l: |
397 | .l: |
398 | shr [.range], 1 |
398 | shr [.range], 1 |
399 | shl eax, 1 |
399 | shl eax, 1 |
400 | mov edx, [.code_] |
400 | mov edx, [.code_] |
401 | sub edx, [.range] |
401 | sub edx, [.range] |
402 | jb @f |
402 | jb @f |
403 | mov [.code_], edx |
403 | mov [.code_], edx |
404 | or eax, 1 |
404 | or eax, 1 |
405 | @@: |
405 | @@: |
406 | cmp [.range], .kTopValue |
406 | cmp [.range], .kTopValue |
407 | jae @f |
407 | jae @f |
408 | shl [.range], 8 |
408 | shl [.range], 8 |
409 | shl [.code_], 8 |
409 | shl [.code_], 8 |
410 | push eax |
410 | push eax |
411 | lodsb |
411 | lodsb |
412 | mov byte [.code_], al |
412 | mov byte [.code_], al |
413 | pop eax |
413 | pop eax |
414 | @@: |
414 | @@: |
415 | loop .l |
415 | loop .l |
416 | ret |
416 | ret |
417 | 417 | ||
418 | .LzmaLiteralDecode: |
418 | .LzmaLiteralDecode: |
419 | ; in: eax->probs |
419 | ; in: eax->probs |
420 | ; out: al=byte; destroys edx |
420 | ; out: al=byte; destroys edx |
421 | push ecx |
421 | push ecx |
422 | mov ecx, 1 |
422 | mov ecx, 1 |
423 | @@: |
423 | @@: |
424 | push eax |
424 | push eax |
425 | lea eax, [eax+ecx*4] |
425 | lea eax, [eax+ecx*4] |
426 | call .RangeDecoderBitDecode |
426 | call .RangeDecoderBitDecode |
427 | pop eax |
427 | pop eax |
428 | adc cl, cl |
428 | adc cl, cl |
429 | jnc @b |
429 | jnc @b |
430 | .LzmaLiteralDecode.ret: |
430 | .LzmaLiteralDecode.ret: |
431 | mov al, cl |
431 | mov al, cl |
432 | pop ecx |
432 | pop ecx |
433 | ret |
433 | ret |
434 | .LzmaLiteralDecodeMatch: |
434 | .LzmaLiteralDecodeMatch: |
435 | ; in: eax->probs, dl=matchByte |
435 | ; in: eax->probs, dl=matchByte |
436 | ; out: al=byte; destroys edx |
436 | ; out: al=byte; destroys edx |
437 | push ecx |
437 | push ecx |
438 | mov ecx, 1 |
438 | mov ecx, 1 |
439 | .LzmaLiteralDecodeMatch.1: |
439 | .LzmaLiteralDecodeMatch.1: |
440 | add dl, dl |
440 | add dl, dl |
441 | setc ch |
441 | setc ch |
442 | push eax |
442 | push eax |
443 | lea eax, [eax+ecx*4+0x100*4] |
443 | lea eax, [eax+ecx*4+0x100*4] |
444 | call .RangeDecoderBitDecode |
444 | call .RangeDecoderBitDecode |
445 | pop eax |
445 | pop eax |
446 | adc cl, cl |
446 | adc cl, cl |
447 | jc .LzmaLiteralDecode.ret |
447 | jc .LzmaLiteralDecode.ret |
448 | xor ch, cl |
448 | xor ch, cl |
449 | test ch, 1 |
449 | test ch, 1 |
450 | mov ch, 0 |
450 | mov ch, 0 |
451 | jnz @b |
451 | jnz @b |
452 | jmp .LzmaLiteralDecodeMatch.1 |
452 | jmp .LzmaLiteralDecodeMatch.1 |
453 | 453 | ||
454 | .LzmaLenDecode: |
454 | .LzmaLenDecode: |
455 | ; in: eax->prob, edx=posState |
455 | ; in: eax->prob, edx=posState |
456 | ; out: ecx=len |
456 | ; out: ecx=len |
457 | push eax |
457 | push eax |
458 | add eax, .LenChoice*4 |
458 | add eax, .LenChoice*4 |
459 | call .RangeDecoderBitDecode |
459 | call .RangeDecoderBitDecode |
460 | pop eax |
460 | pop eax |
461 | jnc .0 |
461 | jnc .0 |
462 | push eax |
462 | push eax |
463 | add eax, .LenChoice2*4 |
463 | add eax, .LenChoice2*4 |
464 | call .RangeDecoderBitDecode |
464 | call .RangeDecoderBitDecode |
465 | pop eax |
465 | pop eax |
466 | jc @f |
466 | jc @f |
467 | mov ecx, .kLenNumMidBits |
467 | mov ecx, .kLenNumMidBits |
468 | shl edx, cl |
468 | shl edx, cl |
469 | lea eax, [eax + .LenMid*4 + edx*4] |
469 | lea eax, [eax + .LenMid*4 + edx*4] |
470 | call .RangeDecoderBitTreeDecode |
470 | call .RangeDecoderBitTreeDecode |
471 | add ecx, .kLenNumLowSymbols |
471 | add ecx, .kLenNumLowSymbols |
472 | ret |
472 | ret |
473 | @@: |
473 | @@: |
474 | add eax, .LenHigh*4 |
474 | add eax, .LenHigh*4 |
475 | mov ecx, .kLenNumHighBits |
475 | mov ecx, .kLenNumHighBits |
476 | call .RangeDecoderBitTreeDecode |
476 | call .RangeDecoderBitTreeDecode |
477 | add ecx, .kLenNumLowSymbols + .kLenNumMidSymbols |
477 | add ecx, .kLenNumLowSymbols + .kLenNumMidSymbols |
478 | ret |
478 | ret |
479 | .0: |
479 | .0: |
480 | mov ecx, .kLenNumLowBits |
480 | mov ecx, .kLenNumLowBits |
481 | shl edx, cl |
481 | shl edx, cl |
482 | lea eax, [eax + .LenLow*4 + edx*4] |
482 | lea eax, [eax + .LenLow*4 + edx*4] |
483 | .RangeDecoderBitTreeDecode: |
483 | .RangeDecoderBitTreeDecode: |
484 | ; in: eax->probs,ecx=numLevels |
484 | ; in: eax->probs,ecx=numLevels |
485 | ; out: ecx=length; destroys edx |
485 | ; out: ecx=length; destroys edx |
486 | push ebx |
486 | push ebx |
487 | mov edx, 1 |
487 | mov edx, 1 |
488 | mov ebx, edx |
488 | mov ebx, edx |
489 | @@: |
489 | @@: |
490 | push eax |
490 | push eax |
491 | lea eax, [eax+edx*4] |
491 | lea eax, [eax+edx*4] |
492 | call .RangeDecoderBitDecode |
492 | call .RangeDecoderBitDecode |
493 | pop eax |
493 | pop eax |
494 | adc dl, dl |
494 | adc dl, dl |
495 | add bl, bl |
495 | add bl, bl |
496 | loop @b |
496 | loop @b |
497 | sub dl, bl |
497 | sub dl, bl |
498 | pop ebx |
498 | pop ebx |
499 | mov ecx, edx |
499 | mov ecx, edx |
500 | ret |
500 | ret |
501 | .RangeDecoderReverseBitTreeDecode: |
501 | .RangeDecoderReverseBitTreeDecode: |
502 | ; in: eax->probs,ecx=numLevels |
502 | ; in: eax->probs,ecx=numLevels |
503 | ; out: ecx=length; destroys edx |
503 | ; out: ecx=length; destroys edx |
504 | push ebx ecx |
504 | push ebx ecx |
505 | mov edx, 1 |
505 | mov edx, 1 |
506 | xor ebx, ebx |
506 | xor ebx, ebx |
507 | @@: |
507 | @@: |
508 | push eax |
508 | push eax |
509 | lea eax, [eax+edx*4] |
509 | lea eax, [eax+edx*4] |
510 | call .RangeDecoderBitDecode |
510 | call .RangeDecoderBitDecode |
511 | lahf |
511 | lahf |
512 | adc edx, edx |
512 | adc edx, edx |
513 | sahf |
513 | sahf |
514 | rcr ebx, 1 |
514 | rcr ebx, 1 |
515 | pop eax |
515 | pop eax |
516 | loop @b |
516 | loop @b |
517 | pop ecx |
517 | pop ecx |
518 | rol ebx, cl |
518 | rol ebx, cl |
519 | mov ecx, ebx |
519 | mov ecx, ebx |
520 | pop ebx |
520 | pop ebx |
521 | ret |
521 | ret |
522 | 522 | ||
523 | uglobal |
523 | uglobal |
524 | align 4 |
524 | align 4 |
525 | ;unpack.p rd unpack.LZMA_BASE_SIZE + (unpack.LZMA_LIT_SIZE shl (unpack.lc+unpack.lp)) |
525 | ;unpack.p rd unpack.LZMA_BASE_SIZE + (unpack.LZMA_LIT_SIZE shl (unpack.lc+unpack.lp)) |
526 | unpack.p dd ? |
526 | unpack.p dd ? |
527 | unpack.code_ dd ? |
527 | unpack.code_ dd ? |
528 | unpack.range dd ? |
528 | unpack.range dd ? |
529 | unpack.rep0 dd ? |
529 | unpack.rep0 dd ? |
530 | unpack.rep1 dd ? |
530 | unpack.rep1 dd ? |
531 | unpack.rep2 dd ? |
531 | unpack.rep2 dd ? |
532 | unpack.rep3 dd ? |
532 | unpack.rep3 dd ? |
533 | unpack.previousByte db ? |
533 | unpack.previousByte db ? |
534 | endg |
534 | endg |