diff options
author | Mark Adler <madler@alumni.caltech.edu> | 2011-09-09 23:21:57 -0700 |
---|---|---|
committer | Mark Adler <madler@alumni.caltech.edu> | 2011-09-09 23:21:57 -0700 |
commit | 13a294f044ef0a89b2dcbfbb5d4d4c792673348e (patch) | |
tree | d9b377b4d8c00633c3da4e96659bfba9b08287f2 /contrib/inflate86 | |
parent | 7c2a874e50b871d04fbd19501f7b42cff55e5abc (diff) | |
download | zlib-1.2.0.1.tar.gz zlib-1.2.0.1.tar.bz2 zlib-1.2.0.1.zip |
zlib 1.2.0.1v1.2.0.1
Diffstat (limited to 'contrib/inflate86')
-rw-r--r-- | contrib/inflate86/inffas86.c | 783 | ||||
-rw-r--r-- | contrib/inflate86/inffast.S | 2472 |
2 files changed, 2160 insertions, 1095 deletions
diff --git a/contrib/inflate86/inffas86.c b/contrib/inflate86/inffas86.c new file mode 100644 index 0000000..4534693 --- /dev/null +++ b/contrib/inflate86/inffas86.c | |||
@@ -0,0 +1,783 @@ | |||
1 | /* inffas86.c is a hand tuned assembler version of | ||
2 | * | ||
3 | * inffast.c -- fast decoding | ||
4 | * Copyright (C) 1995-2003 Mark Adler | ||
5 | * For conditions of distribution and use, see copyright notice in zlib.h | ||
6 | * | ||
7 | * Copyright (C) 2003 Chris Anderson <christop@charm.net> | ||
8 | * Please use the copyright conditions above. | ||
9 | * | ||
10 | * Mar-13-2003 -- Most of this is derived from inffast.S which is derived from | ||
11 | * the gcc -S output of zlib-1.2.0/inffast.c. Zlib-1.2.0 is in beta release at | ||
12 | * the moment. I have successfully compiled and tested this code with gcc2.96, | ||
13 | * gcc3.2, icc5.0, msvc6.0. It is very close to the speed of inffast.S | ||
14 | * compiled with gcc -DNO_MMX, but inffast.S is still faster on the P3 with MMX | ||
15 | * enabled. I will attempt to merge the MMX code into this version. Newer | ||
16 | * versions of this and inffast.S can be found at | ||
17 | * http://www.eetbeetee.com/zlib/ and http://www.charm.net/~christop/zlib/ | ||
18 | */ | ||
19 | |||
20 | #include "zutil.h" | ||
21 | #include "inftrees.h" | ||
22 | #include "inflate.h" | ||
23 | #include "inffast.h" | ||
24 | |||
25 | /* Mark Adler's comments from inffast.c: */ | ||
26 | |||
27 | /* | ||
28 | Decode literal, length, and distance codes and write out the resulting | ||
29 | literal and match bytes until either not enough input or output is | ||
30 | available, an end-of-block is encountered, or a data error is encountered. | ||
31 | When large enough input and output buffers are supplied to inflate(), for | ||
32 | example, a 16K input buffer and a 64K output buffer, more than 95% of the | ||
33 | inflate execution time is spent in this routine. | ||
34 | |||
35 | Entry assumptions: | ||
36 | |||
37 | state->mode == LEN | ||
38 | strm->avail_in >= 6 | ||
39 | strm->avail_out >= 258 | ||
40 | start >= strm->avail_out | ||
41 | state->bits < 8 | ||
42 | |||
43 | On return, state->mode is one of: | ||
44 | |||
45 | LEN -- ran out of enough output space or enough available input | ||
46 | TYPE -- reached end of block code, inflate() to interpret next block | ||
47 | BAD -- error in block data | ||
48 | |||
49 | Notes: | ||
50 | |||
51 | - The maximum input bits used by a length/distance pair is 15 bits for the | ||
52 | length code, 5 bits for the length extra, 15 bits for the distance code, | ||
53 | and 13 bits for the distance extra. This totals 48 bits, or six bytes. | ||
54 | Therefore if strm->avail_in >= 6, then there is enough input to avoid | ||
55 | checking for available input while decoding. | ||
56 | |||
57 | - The maximum bytes that a single length/distance pair can output is 258 | ||
58 | bytes, which is the maximum length that can be coded. inflate_fast() | ||
59 | requires strm->avail_out >= 258 for each loop to avoid checking for | ||
60 | output space. | ||
61 | */ | ||
62 | void inflate_fast(strm, start) | ||
63 | z_streamp strm; | ||
64 | unsigned start; /* inflate()'s starting value for strm->avail_out */ | ||
65 | { | ||
66 | struct inflate_state FAR *state; | ||
67 | struct inffast_ar { | ||
68 | void *esp; /* esp save */ | ||
69 | unsigned char FAR *in; /* local strm->next_in */ | ||
70 | unsigned char FAR *last; /* while in < last, enough input available */ | ||
71 | unsigned char FAR *out; /* local strm->next_out */ | ||
72 | unsigned char FAR *beg; /* inflate()'s initial strm->next_out */ | ||
73 | unsigned char FAR *end; /* while out < end, enough space available */ | ||
74 | unsigned wsize; /* window size or zero if not using window */ | ||
75 | unsigned write; /* window write index */ | ||
76 | unsigned char FAR *window; /* allocated sliding window, if wsize != 0 */ | ||
77 | unsigned long hold; /* local strm->hold */ | ||
78 | unsigned bits; /* local strm->bits */ | ||
79 | code const FAR *lcode; /* local strm->lencode */ | ||
80 | code const FAR *dcode; /* local strm->distcode */ | ||
81 | unsigned lmask; /* mask for first level of length codes */ | ||
82 | unsigned dmask; /* mask for first level of distance codes */ | ||
83 | unsigned len; /* match length, unused bytes */ | ||
84 | unsigned dist; /* match distance */ | ||
85 | unsigned status; /* this is set when state changes */ | ||
86 | } ar; | ||
87 | |||
88 | /* copy state to local variables */ | ||
89 | state = (struct inflate_state FAR *)strm->state; | ||
90 | ar.in = strm->next_in; | ||
91 | ar.last = ar.in + (strm->avail_in - 5); | ||
92 | ar.out = strm->next_out; | ||
93 | ar.beg = ar.out - (start - strm->avail_out); | ||
94 | ar.end = ar.out + (strm->avail_out - 257); | ||
95 | ar.wsize = state->wsize; | ||
96 | ar.write = state->write; | ||
97 | ar.window = state->window; | ||
98 | ar.hold = state->hold; | ||
99 | ar.bits = state->bits; | ||
100 | ar.lcode = state->lencode; | ||
101 | ar.dcode = state->distcode; | ||
102 | ar.lmask = (1U << state->lenbits) - 1; | ||
103 | ar.dmask = (1U << state->distbits) - 1; | ||
104 | |||
105 | /* decode literals and length/distances until end-of-block or not enough | ||
106 | input data or output space */ | ||
107 | |||
108 | /* align in on 2 byte boundary */ | ||
109 | if (((unsigned long)(void *)ar.in & 0x1) != 0) { | ||
110 | ar.hold += (unsigned long)*ar.in++ << ar.bits; | ||
111 | ar.bits += 8; | ||
112 | } | ||
113 | |||
114 | #if defined( __GNUC__ ) || defined( __ICC ) | ||
115 | __asm__ __volatile__ ( | ||
116 | " leal %0, %%eax\n" | ||
117 | " pushf\n" | ||
118 | " pushl %%ebp\n" | ||
119 | " movl %%esp, (%%eax)\n" | ||
120 | " movl %%eax, %%esp\n" | ||
121 | " movl 4(%%esp), %%esi\n" /* esi = in */ | ||
122 | " movl 12(%%esp), %%edi\n" /* edi = out */ | ||
123 | " movl 36(%%esp), %%edx\n" /* edx = hold */ | ||
124 | " movl 40(%%esp), %%ebx\n" /* ebx = bits */ | ||
125 | " movl 44(%%esp), %%ebp\n" /* ebp = lcode */ | ||
126 | |||
127 | " cld\n" | ||
128 | " jmp .L_do_loop\n" | ||
129 | |||
130 | ".L_while_test:\n" | ||
131 | " cmpl %%edi, 20(%%esp)\n" | ||
132 | " jbe .L_break_loop\n" | ||
133 | " cmpl %%esi, 8(%%esp)\n" | ||
134 | " jbe .L_break_loop\n" | ||
135 | |||
136 | ".L_do_loop:\n" | ||
137 | " cmpb $15, %%bl\n" | ||
138 | " ja .L_get_length_code\n" /* if (15 < bits) */ | ||
139 | |||
140 | " xorl %%eax, %%eax\n" | ||
141 | " lodsw\n" /* al = *(ushort *)in++ */ | ||
142 | " movb %%bl, %%cl\n" /* cl = bits, needs it for shifting */ | ||
143 | " addb $16, %%bl\n" /* bits += 16 */ | ||
144 | " shll %%cl, %%eax\n" | ||
145 | " orl %%eax, %%edx\n" /* hold |= *((ushort *)in)++ << bits */ | ||
146 | |||
147 | ".L_get_length_code:\n" | ||
148 | " movl 52(%%esp), %%eax\n" /* eax = lmask */ | ||
149 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
150 | " movl (%%ebp,%%eax,4), %%eax\n" /* eax = lcode[hold & lmask] */ | ||
151 | |||
152 | ".L_dolen:\n" | ||
153 | " movb %%ah, %%cl\n" /* cl = this.bits */ | ||
154 | " subb %%ah, %%bl\n" /* bits -= this.bits */ | ||
155 | " shrl %%cl, %%edx\n" /* hold >>= this.bits */ | ||
156 | |||
157 | " testb %%al, %%al\n" | ||
158 | " jnz .L_test_for_length_base\n" /* if (op != 0) 45.7% */ | ||
159 | |||
160 | " shrl $16, %%eax\n" /* output this.val char */ | ||
161 | " stosb\n" | ||
162 | " jmp .L_while_test\n" | ||
163 | |||
164 | ".L_test_for_length_base:\n" | ||
165 | " movl %%eax, %%ecx\n" /* len = this */ | ||
166 | " shrl $16, %%ecx\n" /* len = this.val */ | ||
167 | " movl %%ecx, 60(%%esp)\n" /* len = this */ | ||
168 | " movb %%al, %%cl\n" | ||
169 | |||
170 | " testb $16, %%al\n" | ||
171 | " jz .L_test_for_second_level_length\n" /* if ((op & 16) == 0) 8% */ | ||
172 | " andb $15, %%cl\n" /* op &= 15 */ | ||
173 | " jz .L_decode_distance\n" /* if (!op) */ | ||
174 | " cmpb %%cl, %%bl\n" | ||
175 | " jae .L_add_bits_to_len\n" /* if (op <= bits) */ | ||
176 | |||
177 | " movb %%cl, %%ch\n" /* stash op in ch, freeing cl */ | ||
178 | " xorl %%eax, %%eax\n" | ||
179 | " lodsw\n" /* al = *(ushort *)in++ */ | ||
180 | " movb %%bl, %%cl\n" /* cl = bits, needs it for shifting */ | ||
181 | " addb $16, %%bl\n" /* bits += 16 */ | ||
182 | " shll %%cl, %%eax\n" | ||
183 | " orl %%eax, %%edx\n" /* hold |= *((ushort *)in)++ << bits */ | ||
184 | " movb %%ch, %%cl\n" /* move op back to ecx */ | ||
185 | |||
186 | ".L_add_bits_to_len:\n" | ||
187 | " movl $1, %%eax\n" | ||
188 | " shll %%cl, %%eax\n" | ||
189 | " decl %%eax\n" | ||
190 | " subb %%cl, %%bl\n" | ||
191 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
192 | " shrl %%cl, %%edx\n" | ||
193 | " addl %%eax, 60(%%esp)\n" /* len += hold & mask[op] */ | ||
194 | |||
195 | ".L_decode_distance:\n" | ||
196 | " cmpb $15, %%bl\n" | ||
197 | " ja .L_get_distance_code\n" /* if (15 < bits) */ | ||
198 | |||
199 | " xorl %%eax, %%eax\n" | ||
200 | " lodsw\n" /* al = *(ushort *)in++ */ | ||
201 | " movb %%bl, %%cl\n" /* cl = bits, needs it for shifting */ | ||
202 | " addb $16, %%bl\n" /* bits += 16 */ | ||
203 | " shll %%cl, %%eax\n" | ||
204 | " orl %%eax, %%edx\n" /* hold |= *((ushort *)in)++ << bits */ | ||
205 | |||
206 | ".L_get_distance_code:\n" | ||
207 | " movl 56(%%esp), %%eax\n" /* eax = dmask */ | ||
208 | " movl 48(%%esp), %%ecx\n" /* ecx = dcode */ | ||
209 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
210 | " movl (%%ecx,%%eax,4), %%eax\n"/* eax = dcode[hold & dmask] */ | ||
211 | |||
212 | ".L_dodist:\n" | ||
213 | " movl %%eax, %%ebp\n" /* dist = this */ | ||
214 | " shrl $16, %%ebp\n" /* dist = this.val */ | ||
215 | " movb %%ah, %%cl\n" | ||
216 | " subb %%ah, %%bl\n" /* bits -= this.bits */ | ||
217 | " shrl %%cl, %%edx\n" /* hold >>= this.bits */ | ||
218 | " movb %%al, %%cl\n" /* cl = this.op */ | ||
219 | |||
220 | " testb $16, %%al\n" /* if ((op & 16) == 0) */ | ||
221 | " jz .L_test_for_second_level_dist\n" | ||
222 | " andb $15, %%cl\n" /* op &= 15 */ | ||
223 | " jz .L_check_dist_one\n" | ||
224 | " cmpb %%cl, %%bl\n" | ||
225 | " jae .L_add_bits_to_dist\n" /* if (op <= bits) 97.6% */ | ||
226 | |||
227 | " movb %%cl, %%ch\n" /* stash op in ch, freeing cl */ | ||
228 | " xorl %%eax, %%eax\n" | ||
229 | " lodsw\n" /* al = *(ushort *)in++ */ | ||
230 | " movb %%bl, %%cl\n" /* cl = bits, needs it for shifting */ | ||
231 | " addb $16, %%bl\n" /* bits += 16 */ | ||
232 | " shll %%cl, %%eax\n" | ||
233 | " orl %%eax, %%edx\n" /* hold |= *((ushort *)in)++ << bits */ | ||
234 | " movb %%ch, %%cl\n" /* move op back to ecx */ | ||
235 | |||
236 | ".L_add_bits_to_dist:\n" | ||
237 | " movl $1, %%eax\n" | ||
238 | " shll %%cl, %%eax\n" | ||
239 | " decl %%eax\n" /* (1 << op) - 1 */ | ||
240 | " subb %%cl, %%bl\n" | ||
241 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
242 | " shrl %%cl, %%edx\n" | ||
243 | " addl %%eax, %%ebp\n" /* dist += hold & ((1 << op) - 1) */ | ||
244 | |||
245 | ".L_check_window:\n" | ||
246 | " movl %%esi, 4(%%esp)\n" /* save in so from can use it's reg */ | ||
247 | " movl %%edi, %%eax\n" | ||
248 | " subl 16(%%esp), %%eax\n" /* nbytes = out - beg */ | ||
249 | |||
250 | " cmpl %%ebp, %%eax\n" | ||
251 | " jb .L_clip_window\n" /* if (dist > nbytes) 4.2% */ | ||
252 | |||
253 | " movl 60(%%esp), %%ecx\n" | ||
254 | " movl %%edi, %%esi\n" | ||
255 | " subl %%ebp, %%esi\n" /* from = out - dist */ | ||
256 | |||
257 | " subl $3, %%ecx\n" /* copy from to out */ | ||
258 | " movb (%%esi), %%al\n" | ||
259 | " movb %%al, (%%edi)\n" | ||
260 | " movb 1(%%esi), %%al\n" | ||
261 | " movb 2(%%esi), %%ah\n" | ||
262 | " addl $3, %%esi\n" | ||
263 | " movb %%al, 1(%%edi)\n" | ||
264 | " movb %%ah, 2(%%edi)\n" | ||
265 | " addl $3, %%edi\n" | ||
266 | " rep movsb\n" | ||
267 | |||
268 | " movl 4(%%esp), %%esi\n" /* move in back to %esi, toss from */ | ||
269 | " movl 44(%%esp), %%ebp\n" /* ebp = lcode */ | ||
270 | " jmp .L_while_test\n" | ||
271 | |||
272 | ".L_check_dist_one:\n" | ||
273 | " cmpl $1, %%ebp\n" /* if dist 1, is a memset */ | ||
274 | " jne .L_check_window\n" | ||
275 | " cmpl %%edi, 16(%%esp)\n" | ||
276 | " je .L_check_window\n" | ||
277 | |||
278 | " decl %%edi\n" | ||
279 | " movl 60(%%esp), %%ecx\n" | ||
280 | " movb (%%edi), %%al\n" | ||
281 | " subl $3, %%ecx\n" | ||
282 | |||
283 | " movb %%al, 1(%%edi)\n" /* memset out with from[-1] */ | ||
284 | " movb %%al, 2(%%edi)\n" | ||
285 | " movb %%al, 3(%%edi)\n" | ||
286 | " addl $4, %%edi\n" | ||
287 | " rep stosb\n" | ||
288 | " movl 44(%%esp), %%ebp\n" /* ebp = lcode */ | ||
289 | " jmp .L_while_test\n" | ||
290 | |||
291 | ".L_test_for_second_level_length:\n" | ||
292 | " testb $64, %%al\n" | ||
293 | " jnz .L_test_for_end_of_block\n" /* if ((op & 64) != 0) */ | ||
294 | |||
295 | " movl $1, %%eax\n" | ||
296 | " shll %%cl, %%eax\n" | ||
297 | " decl %%eax\n" | ||
298 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
299 | " addl 60(%%esp), %%eax\n" /* eax += this.val */ | ||
300 | " movl (%%ebp,%%eax,4), %%eax\n" /* eax = lcode[val+(hold&mask[op])]*/ | ||
301 | " jmp .L_dolen\n" | ||
302 | |||
303 | ".L_test_for_second_level_dist:\n" | ||
304 | " testb $64, %%al\n" | ||
305 | " jnz .L_invalid_distance_code\n" /* if ((op & 64) != 0) */ | ||
306 | |||
307 | " movl $1, %%eax\n" | ||
308 | " shll %%cl, %%eax\n" | ||
309 | " decl %%eax\n" | ||
310 | " andl %%edx, %%eax\n" /* eax &= hold */ | ||
311 | " addl %%ebp, %%eax\n" /* eax += this.val */ | ||
312 | " movl 48(%%esp), %%ecx\n" /* ecx = dcode */ | ||
313 | " movl (%%ecx,%%eax,4), %%eax\n" /* eax = dcode[val+(hold&mask[op])]*/ | ||
314 | " jmp .L_dodist\n" | ||
315 | |||
316 | ".L_clip_window:\n" | ||
317 | " movl %%eax, %%ecx\n" | ||
318 | " movl 24(%%esp), %%eax\n" /* prepare for dist compare */ | ||
319 | " negl %%ecx\n" /* nbytes = -nbytes */ | ||
320 | " movl 32(%%esp), %%esi\n" /* from = window */ | ||
321 | |||
322 | " cmpl %%ebp, %%eax\n" | ||
323 | " jb .L_invalid_distance_too_far\n" /* if (dist > wsize) */ | ||
324 | |||
325 | " addl %%ebp, %%ecx\n" /* nbytes = dist - nbytes */ | ||
326 | " cmpl $0, 28(%%esp)\n" | ||
327 | " jne .L_wrap_around_window\n" /* if (write != 0) */ | ||
328 | |||
329 | " subl %%ecx, %%eax\n" | ||
330 | " addl %%eax, %%esi\n" /* from += wsize - nbytes */ | ||
331 | |||
332 | " movl 60(%%esp), %%eax\n" | ||
333 | " cmpl %%ecx, %%eax\n" | ||
334 | " jbe .L_do_copy1\n" /* if (nbytes >= len) */ | ||
335 | |||
336 | " subl %%ecx, %%eax\n" /* len -= nbytes */ | ||
337 | " rep movsb\n" | ||
338 | " movl %%edi, %%esi\n" | ||
339 | " subl %%ebp, %%esi\n" /* from = out - dist */ | ||
340 | " jmp .L_do_copy1\n" | ||
341 | |||
342 | " cmpl %%ecx, %%eax\n" | ||
343 | " jbe .L_do_copy1\n" /* if (nbytes >= len) */ | ||
344 | |||
345 | " subl %%ecx, %%eax\n" /* len -= nbytes */ | ||
346 | " rep movsb\n" | ||
347 | " movl %%edi, %%esi\n" | ||
348 | " subl %%ebp, %%esi\n" /* from = out - dist */ | ||
349 | " jmp .L_do_copy1\n" | ||
350 | |||
351 | ".L_wrap_around_window:\n" | ||
352 | " movl 28(%%esp), %%eax\n" | ||
353 | " cmpl %%eax, %%ecx\n" | ||
354 | " jbe .L_contiguous_in_window\n" /* if (write >= nbytes) */ | ||
355 | |||
356 | " addl 24(%%esp), %%esi\n" | ||
357 | " addl %%eax, %%esi\n" | ||
358 | " subl %%ecx, %%esi\n" /* from += wsize + write - nbytes */ | ||
359 | " subl %%eax, %%ecx\n" /* nbytes -= write */ | ||
360 | |||
361 | " movl 60(%%esp), %%eax\n" | ||
362 | " cmpl %%ecx, %%eax\n" | ||
363 | " jbe .L_do_copy1\n" /* if (nbytes >= len) */ | ||
364 | |||
365 | " subl %%ecx, %%eax\n" /* len -= nbytes */ | ||
366 | " rep movsb\n" | ||
367 | " movl 32(%%esp), %%esi\n" /* from = window */ | ||
368 | " movl 28(%%esp), %%ecx\n" /* nbytes = write */ | ||
369 | " cmpl %%ecx, %%eax\n" | ||
370 | " jbe .L_do_copy1\n" /* if (nbytes >= len) */ | ||
371 | |||
372 | " subl %%ecx, %%eax\n" /* len -= nbytes */ | ||
373 | " rep movsb\n" | ||
374 | " movl %%edi, %%esi\n" | ||
375 | " subl %%ebp, %%esi\n" /* from = out - dist */ | ||
376 | " jmp .L_do_copy1\n" | ||
377 | |||
378 | ".L_contiguous_in_window:\n" | ||
379 | " addl %%eax, %%esi\n" | ||
380 | " subl %%ecx, %%esi\n" /* from += write - nbytes */ | ||
381 | |||
382 | " movl 60(%%esp), %%eax\n" | ||
383 | " cmpl %%ecx, %%eax\n" | ||
384 | " jbe .L_do_copy1\n" /* if (nbytes >= len) */ | ||
385 | |||
386 | " subl %%ecx, %%eax\n" /* len -= nbytes */ | ||
387 | " rep movsb\n" | ||
388 | " movl %%edi, %%esi\n" | ||
389 | " subl %%ebp, %%esi\n" /* from = out - dist */ | ||
390 | |||
391 | ".L_do_copy1:\n" | ||
392 | " movl %%eax, %%ecx\n" | ||
393 | " rep movsb\n" | ||
394 | |||
395 | " movl 4(%%esp), %%esi\n" /* move in back to %esi, toss from */ | ||
396 | " movl 44(%%esp), %%ebp\n" /* ebp = lcode */ | ||
397 | " jmp .L_while_test\n" | ||
398 | |||
399 | ".L_test_for_end_of_block:\n" | ||
400 | " testb $32, %%al\n" | ||
401 | " jz .L_invalid_literal_length_code\n" | ||
402 | " movl $1, 68(%%esp)\n" | ||
403 | " jmp .L_break_loop_with_status\n" | ||
404 | |||
405 | ".L_invalid_literal_length_code:\n" | ||
406 | " movl $2, 68(%%esp)\n" | ||
407 | " jmp .L_break_loop_with_status\n" | ||
408 | |||
409 | ".L_invalid_distance_code:\n" | ||
410 | " movl $3, 68(%%esp)\n" | ||
411 | " jmp .L_break_loop_with_status\n" | ||
412 | |||
413 | ".L_invalid_distance_too_far:\n" | ||
414 | " movl 4(%%esp), %%esi\n" | ||
415 | " movl $4, 68(%%esp)\n" | ||
416 | " jmp .L_break_loop_with_status\n" | ||
417 | |||
418 | ".L_break_loop:\n" | ||
419 | " movl $0, 68(%%esp)\n" | ||
420 | |||
421 | ".L_break_loop_with_status:\n" | ||
422 | /* put in, out, bits, and hold back into ar and pop esp */ | ||
423 | " movl %%esi, 4(%%esp)\n" | ||
424 | " movl %%edi, 12(%%esp)\n" | ||
425 | " movl %%ebx, 40(%%esp)\n" | ||
426 | " movl %%edx, 36(%%esp)\n" | ||
427 | " movl (%%esp), %%esp\n" | ||
428 | " popl %%ebp\n" | ||
429 | " popf\n" | ||
430 | : | ||
431 | : "m" (ar) | ||
432 | : "memory", "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" | ||
433 | ); | ||
434 | #elif defined( _MSC_VER ) | ||
435 | __asm { | ||
436 | lea eax, ar | ||
437 | pushfd | ||
438 | push ebp | ||
439 | mov [eax], esp | ||
440 | mov esp, eax | ||
441 | mov esi, [esp+4] /* esi = in */ | ||
442 | mov edi, [esp+12] /* edi = out */ | ||
443 | mov edx, [esp+36] /* edx = hold */ | ||
444 | mov ebx, [esp+40] /* ebx = bits */ | ||
445 | mov ebp, [esp+44] /* ebp = lcode */ | ||
446 | |||
447 | cld | ||
448 | jmp L_do_loop | ||
449 | |||
450 | L_while_test: | ||
451 | cmp [esp+20], edi | ||
452 | jbe L_break_loop | ||
453 | cmp [esp+8], esi | ||
454 | jbe L_break_loop | ||
455 | |||
456 | L_do_loop: | ||
457 | cmp bl, 15 | ||
458 | ja L_get_length_code /* if (15 < bits) */ | ||
459 | |||
460 | xor eax, eax | ||
461 | lodsw /* al = *(ushort *)in++ */ | ||
462 | mov cl, bl /* cl = bits, needs it for shifting */ | ||
463 | add bl, 16 /* bits += 16 */ | ||
464 | shl eax, cl | ||
465 | or edx, eax /* hold |= *((ushort *)in)++ << bits */ | ||
466 | |||
467 | L_get_length_code: | ||
468 | mov eax, [esp+52] /* eax = lmask */ | ||
469 | and eax, edx /* eax &= hold */ | ||
470 | mov eax, [ebp+eax*4] /* eax = lcode[hold & lmask] */ | ||
471 | |||
472 | L_dolen: | ||
473 | mov cl, ah /* cl = this.bits */ | ||
474 | sub bl, ah /* bits -= this.bits */ | ||
475 | shr edx, cl /* hold >>= this.bits */ | ||
476 | |||
477 | test al, al | ||
478 | jnz L_test_for_length_base /* if (op != 0) 45.7% */ | ||
479 | |||
480 | shr eax, 16 /* output this.val char */ | ||
481 | stosb | ||
482 | jmp L_while_test | ||
483 | |||
484 | L_test_for_length_base: | ||
485 | mov ecx, eax /* len = this */ | ||
486 | shr ecx, 16 /* len = this.val */ | ||
487 | mov [esp+60], ecx /* len = this */ | ||
488 | mov cl, al | ||
489 | |||
490 | test al, 16 | ||
491 | jz L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ | ||
492 | and cl, 15 /* op &= 15 */ | ||
493 | jz L_decode_distance /* if (!op) */ | ||
494 | cmp bl, cl | ||
495 | jae L_add_bits_to_len /* if (op <= bits) */ | ||
496 | |||
497 | mov ch, cl /* stash op in ch, freeing cl */ | ||
498 | xor eax, eax | ||
499 | lodsw /* al = *(ushort *)in++ */ | ||
500 | mov cl, bl /* cl = bits, needs it for shifting */ | ||
501 | add bl, 16 /* bits += 16 */ | ||
502 | shl eax, cl | ||
503 | or edx, eax /* hold |= *((ushort *)in)++ << bits */ | ||
504 | mov cl, ch /* move op back to ecx */ | ||
505 | |||
506 | L_add_bits_to_len: | ||
507 | mov eax, 1 | ||
508 | shl eax, cl | ||
509 | dec eax | ||
510 | sub bl, cl | ||
511 | and eax, edx /* eax &= hold */ | ||
512 | shr edx, cl | ||
513 | add [esp+60], eax /* len += hold & mask[op] */ | ||
514 | |||
515 | L_decode_distance: | ||
516 | cmp bl, 15 | ||
517 | ja L_get_distance_code /* if (15 < bits) */ | ||
518 | |||
519 | xor eax, eax | ||
520 | lodsw /* al = *(ushort *)in++ */ | ||
521 | mov cl, bl /* cl = bits, needs it for shifting */ | ||
522 | add bl, 16 /* bits += 16 */ | ||
523 | shl eax, cl | ||
524 | or edx, eax /* hold |= *((ushort *)in)++ << bits */ | ||
525 | |||
526 | L_get_distance_code: | ||
527 | mov eax, [esp+56] /* eax = dmask */ | ||
528 | mov ecx, [esp+48] /* ecx = dcode */ | ||
529 | and eax, edx /* eax &= hold */ | ||
530 | mov eax, [ecx+eax*4]/* eax = dcode[hold & dmask] */ | ||
531 | |||
532 | L_dodist: | ||
533 | mov ebp, eax /* dist = this */ | ||
534 | shr ebp, 16 /* dist = this.val */ | ||
535 | mov cl, ah | ||
536 | sub bl, ah /* bits -= this.bits */ | ||
537 | shr edx, cl /* hold >>= this.bits */ | ||
538 | mov cl, al /* cl = this.op */ | ||
539 | |||
540 | test al, 16 /* if ((op & 16) == 0) */ | ||
541 | jz L_test_for_second_level_dist | ||
542 | and cl, 15 /* op &= 15 */ | ||
543 | jz L_check_dist_one | ||
544 | cmp bl, cl | ||
545 | jae L_add_bits_to_dist /* if (op <= bits) 97.6% */ | ||
546 | |||
547 | mov ch, cl /* stash op in ch, freeing cl */ | ||
548 | xor eax, eax | ||
549 | lodsw /* al = *(ushort *)in++ */ | ||
550 | mov cl, bl /* cl = bits, needs it for shifting */ | ||
551 | add bl, 16 /* bits += 16 */ | ||
552 | shl eax, cl | ||
553 | or edx, eax /* hold |= *((ushort *)in)++ << bits */ | ||
554 | mov cl, ch /* move op back to ecx */ | ||
555 | |||
556 | L_add_bits_to_dist: | ||
557 | mov eax, 1 | ||
558 | shl eax, cl | ||
559 | dec eax /* (1 << op) - 1 */ | ||
560 | sub bl, cl | ||
561 | and eax, edx /* eax &= hold */ | ||
562 | shr edx, cl | ||
563 | add ebp, eax /* dist += hold & ((1 << op) - 1) */ | ||
564 | |||
565 | L_check_window: | ||
566 | mov [esp+4], esi /* save in so from can use it's reg */ | ||
567 | mov eax, edi | ||
568 | sub eax, [esp+16] /* nbytes = out - beg */ | ||
569 | |||
570 | cmp eax, ebp | ||
571 | jb L_clip_window /* if (dist > nbytes) 4.2% */ | ||
572 | |||
573 | mov ecx, [esp+60] | ||
574 | mov esi, edi | ||
575 | sub esi, ebp /* from = out - dist */ | ||
576 | |||
577 | sub ecx, 3 /* copy from to out */ | ||
578 | mov al, [esi] | ||
579 | mov [edi], al | ||
580 | mov al, [esi+1] | ||
581 | mov ah, [esi+2] | ||
582 | add esi, 3 | ||
583 | mov [edi+1], al | ||
584 | mov [edi+2], ah | ||
585 | add edi, 3 | ||
586 | rep movsb | ||
587 | |||
588 | mov esi, [esp+4] /* move in back to %esi, toss from */ | ||
589 | mov ebp, [esp+44] /* ebp = lcode */ | ||
590 | jmp L_while_test | ||
591 | |||
592 | L_check_dist_one: | ||
593 | cmp ebp, 1 /* if dist 1, is a memset */ | ||
594 | jne L_check_window | ||
595 | cmp [esp+16], edi | ||
596 | je L_check_window | ||
597 | |||
598 | dec edi | ||
599 | mov ecx, [esp+60] | ||
600 | mov al, [edi] | ||
601 | sub ecx, 3 | ||
602 | |||
603 | mov [edi+1], al /* memset out with from[-1] */ | ||
604 | mov [edi+2], al | ||
605 | mov [edi+3], al | ||
606 | add edi, 4 | ||
607 | rep stosb | ||
608 | mov ebp, [esp+44] /* ebp = lcode */ | ||
609 | jmp L_while_test | ||
610 | |||
611 | L_test_for_second_level_length: | ||
612 | test al, 64 | ||
613 | jnz L_test_for_end_of_block /* if ((op & 64) != 0) */ | ||
614 | |||
615 | mov eax, 1 | ||
616 | shl eax, cl | ||
617 | dec eax | ||
618 | and eax, edx /* eax &= hold */ | ||
619 | add eax, [esp+60] /* eax += this.val */ | ||
620 | mov eax, [ebp+eax*4] /* eax = lcode[val+(hold&mask[op])]*/ | ||
621 | jmp L_dolen | ||
622 | |||
623 | L_test_for_second_level_dist: | ||
624 | test al, 64 | ||
625 | jnz L_invalid_distance_code /* if ((op & 64) != 0) */ | ||
626 | |||
627 | mov eax, 1 | ||
628 | shl eax, cl | ||
629 | dec eax | ||
630 | and eax, edx /* eax &= hold */ | ||
631 | add eax, ebp /* eax += this.val */ | ||
632 | mov ecx, [esp+48] /* ecx = dcode */ | ||
633 | mov eax, [ecx+eax*4] /* eax = dcode[val+(hold&mask[op])]*/ | ||
634 | jmp L_dodist | ||
635 | |||
636 | L_clip_window: | ||
637 | mov ecx, eax | ||
638 | mov eax, [esp+24] /* prepare for dist compare */ | ||
639 | neg ecx /* nbytes = -nbytes */ | ||
640 | mov esi, [esp+32] /* from = window */ | ||
641 | |||
642 | cmp eax, ebp | ||
643 | jb L_invalid_distance_too_far /* if (dist > wsize) */ | ||
644 | |||
645 | add ecx, ebp /* nbytes = dist - nbytes */ | ||
646 | cmp dword ptr [esp+28], 0 | ||
647 | jne L_wrap_around_window /* if (write != 0) */ | ||
648 | |||
649 | sub eax, ecx | ||
650 | add esi, eax /* from += wsize - nbytes */ | ||
651 | |||
652 | mov eax, [esp+60] | ||
653 | cmp eax, ecx | ||
654 | jbe L_do_copy1 /* if (nbytes >= len) */ | ||
655 | |||
656 | sub eax, ecx /* len -= nbytes */ | ||
657 | rep movsb | ||
658 | mov esi, edi | ||
659 | sub esi, ebp /* from = out - dist */ | ||
660 | jmp L_do_copy1 | ||
661 | |||
662 | cmp eax, ecx | ||
663 | jbe L_do_copy1 /* if (nbytes >= len) */ | ||
664 | |||
665 | sub eax, ecx /* len -= nbytes */ | ||
666 | rep movsb | ||
667 | mov esi, edi | ||
668 | sub esi, ebp /* from = out - dist */ | ||
669 | jmp L_do_copy1 | ||
670 | |||
671 | L_wrap_around_window: | ||
672 | mov eax, [esp+28] | ||
673 | cmp ecx, eax | ||
674 | jbe L_contiguous_in_window /* if (write >= nbytes) */ | ||
675 | |||
676 | add esi, [esp+24] | ||
677 | add esi, eax | ||
678 | sub esi, ecx /* from += wsize + write - nbytes */ | ||
679 | sub ecx, eax /* nbytes -= write */ | ||
680 | |||
681 | mov eax, [esp+60] | ||
682 | cmp eax, ecx | ||
683 | jbe L_do_copy1 /* if (nbytes >= len) */ | ||
684 | |||
685 | sub eax, ecx /* len -= nbytes */ | ||
686 | rep movsb | ||
687 | mov esi, [esp+32] /* from = window */ | ||
688 | mov ecx, [esp+28] /* nbytes = write */ | ||
689 | cmp eax, ecx | ||
690 | jbe L_do_copy1 /* if (nbytes >= len) */ | ||
691 | |||
692 | sub eax, ecx /* len -= nbytes */ | ||
693 | rep movsb | ||
694 | mov esi, edi | ||
695 | sub esi, ebp /* from = out - dist */ | ||
696 | jmp L_do_copy1 | ||
697 | |||
698 | L_contiguous_in_window: | ||
699 | add esi, eax | ||
700 | sub esi, ecx /* from += write - nbytes */ | ||
701 | |||
702 | mov eax, [esp+60] | ||
703 | cmp eax, ecx | ||
704 | jbe L_do_copy1 /* if (nbytes >= len) */ | ||
705 | |||
706 | sub eax, ecx /* len -= nbytes */ | ||
707 | rep movsb | ||
708 | mov esi, edi | ||
709 | sub esi, ebp /* from = out - dist */ | ||
710 | |||
711 | L_do_copy1: | ||
712 | mov ecx, eax | ||
713 | rep movsb | ||
714 | |||
715 | mov esi, [esp+4] /* move in back to %esi, toss from */ | ||
716 | mov ebp, [esp+44] /* ebp = lcode */ | ||
717 | jmp L_while_test | ||
718 | |||
719 | L_test_for_end_of_block: | ||
720 | test al, 32 | ||
721 | jz L_invalid_literal_length_code | ||
722 | mov dword ptr [esp+68], 1 | ||
723 | jmp L_break_loop_with_status | ||
724 | |||
725 | L_invalid_literal_length_code: | ||
726 | mov dword ptr [esp+68], 2 | ||
727 | jmp L_break_loop_with_status | ||
728 | |||
729 | L_invalid_distance_code: | ||
730 | mov dword ptr [esp+68], 3 | ||
731 | jmp L_break_loop_with_status | ||
732 | |||
733 | L_invalid_distance_too_far: | ||
734 | mov esi, [esp+4] | ||
735 | mov dword ptr [esp+68], 4 | ||
736 | jmp L_break_loop_with_status | ||
737 | |||
738 | L_break_loop: | ||
739 | mov dword ptr [esp+68], 0 | ||
740 | |||
741 | L_break_loop_with_status: | ||
742 | /* put in, out, bits, and hold back into ar and pop esp */ | ||
743 | mov [esp+4], esi | ||
744 | mov [esp+12], edi | ||
745 | mov [esp+40], ebx | ||
746 | mov [esp+36], edx | ||
747 | mov esp, [esp] | ||
748 | pop ebp | ||
749 | popfd | ||
750 | } | ||
751 | #endif | ||
752 | |||
753 | if (ar.status > 1) { | ||
754 | if (ar.status == 2) | ||
755 | strm->msg = "invalid literal/length code"; | ||
756 | else if (ar.status == 3) | ||
757 | strm->msg = "invalid distance code"; | ||
758 | else | ||
759 | strm->msg = "invalid distance too far back"; | ||
760 | state->mode = BAD; | ||
761 | } | ||
762 | else if ( ar.status == 1 ) { | ||
763 | state->mode = TYPE; | ||
764 | } | ||
765 | |||
766 | /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ | ||
767 | ar.len = ar.bits >> 3; | ||
768 | ar.in -= ar.len; | ||
769 | ar.bits -= ar.len << 3; | ||
770 | ar.hold &= (1U << ar.bits) - 1; | ||
771 | |||
772 | /* update state and return */ | ||
773 | strm->next_in = ar.in; | ||
774 | strm->next_out = ar.out; | ||
775 | strm->avail_in = (unsigned)(ar.in < ar.last ? 5 + (ar.last - ar.in) : | ||
776 | 5 - (ar.in - ar.last)); | ||
777 | strm->avail_out = (unsigned)(ar.out < ar.end ? 257 + (ar.end - ar.out) : | ||
778 | 257 - (ar.out - ar.end)); | ||
779 | state->hold = ar.hold; | ||
780 | state->bits = ar.bits; | ||
781 | return; | ||
782 | } | ||
783 | |||
diff --git a/contrib/inflate86/inffast.S b/contrib/inflate86/inffast.S index d1e80ef..3602907 100644 --- a/contrib/inflate86/inffast.S +++ b/contrib/inflate86/inffast.S | |||
@@ -1,1095 +1,1377 @@ | |||
1 | /* | 1 | /* |
2 | * inffast.S is a hand tuned assembler version of: | 2 | * inffast.S is a hand tuned assembler version of: |
3 | * | 3 | * |
4 | * inffast.c -- fast decoding | 4 | * inffast.c -- fast decoding |
5 | * Copyright (C) 1995-2003 Mark Adler | 5 | * Copyright (C) 1995-2003 Mark Adler |
6 | * For conditions of distribution and use, see copyright notice in zlib.h | 6 | * For conditions of distribution and use, see copyright notice in zlib.h |
7 | * | 7 | * |
8 | * Copyright (C) 2003 Chris Anderson <christop@charm.net> | 8 | * Copyright (C) 2003 Chris Anderson <christop@charm.net> |
9 | * Please use the copyright conditions above. | 9 | * Please use the copyright conditions above. |
10 | * | 10 | * |
11 | * This version (Jan-23-2003) of inflate_fast was coded and tested under | 11 | * This version (Jan-23-2003) of inflate_fast was coded and tested under |
12 | * GNU/Linux on a pentium 3, using the gcc-3.2 compiler distribution. On that | 12 | * GNU/Linux on a pentium 3, using the gcc-3.2 compiler distribution. On that |
13 | * machine, I found that gzip style archives decompressed about 20% faster than | 13 | * machine, I found that gzip style archives decompressed about 20% faster than |
14 | * the gcc-3.2 -O3 -fomit-frame-pointer compiled version. Your results will | 14 | * the gcc-3.2 -O3 -fomit-frame-pointer compiled version. Your results will |
15 | * depend on how large of a buffer is used for z_stream.next_in & next_out | 15 | * depend on how large of a buffer is used for z_stream.next_in & next_out |
16 | * (8K-32K worked best for my 256K cpu cache) and how much overhead there is in | 16 | * (8K-32K worked best for my 256K cpu cache) and how much overhead there is in |
17 | * stream processing I/O and crc32/addler32. In my case, this routine used | 17 | * stream processing I/O and crc32/addler32. In my case, this routine used |
18 | * 70% of the cpu time and crc32 used 20%. | 18 | * 70% of the cpu time and crc32 used 20%. |
19 | * | 19 | * |
20 | * I am confident that this version will work in the general case, but I have | 20 | * I am confident that this version will work in the general case, but I have |
21 | * not tested a wide variety of datasets or a wide variety of platforms. | 21 | * not tested a wide variety of datasets or a wide variety of platforms. |
22 | * | 22 | * |
23 | * Jan-24-2003 -- Added -DUSE_MMX define for slightly faster inflating. | 23 | * Jan-24-2003 -- Added -DUSE_MMX define for slightly faster inflating. |
24 | * It should be a runtime flag instead of compile time flag... | 24 | * It should be a runtime flag instead of compile time flag... |
25 | */ | 25 | * |
26 | 26 | * Jan-26-2003 -- Added runtime check for MMX support with cpuid instruction. | |
27 | .file "inffast.S" | 27 | * With -DUSE_MMX, only MMX code is compiled. With -DNO_MMX, only non-MMX code |
28 | 28 | * is compiled. Without either option, runtime detection is enabled. Runtime | |
29 | .globl inflate_fast | 29 | * detection should work on all modern cpus and the recomended algorithm (flip |
30 | 30 | * ID bit on eflags and then use the cpuid instruction) is used in many | |
31 | .text | 31 | * multimedia applications. Tested under win2k with gcc-2.95 and gas-2.12 |
32 | .align 4,0 | 32 | * distributed with cygwin3. Compiling with gcc-2.95 -c inffast.S -o |
33 | .L_invalid_literal_length_code_msg: | 33 | * inffast.obj generates a COFF object which can then be linked with MSVC++ |
34 | .string "invalid literal/length code" | 34 | * compiled code. Tested under FreeBSD 4.7 with gcc-2.95. |
35 | 35 | * | |
36 | .align 4,0 | 36 | * Jan-28-2003 -- Tested Athlon XP... MMX mode is slower than no MMX (and |
37 | .L_invalid_distance_code_msg: | 37 | * slower than compiler generated code). Adjusted cpuid check to use the MMX |
38 | .string "invalid distance code" | 38 | * code only for Pentiums < P4 until I have more data on the P4. Speed |
39 | 39 | * improvment is only about 15% on the Athlon when compared with code generated | |
40 | .align 4,0 | 40 | * with MSVC++. Not sure yet, but I think the P4 will also be slower using the |
41 | .L_invalid_distance_too_far_msg: | 41 | * MMX mode because many of it's x86 ALU instructions execute in .5 cycles and |
42 | .string "invalid distance too far back" | 42 | * have less latency than MMX ops. Added code to buffer the last 11 bytes of |
43 | 43 | * the input stream since the MMX code grabs bits in chunks of 32, which | |
44 | #if defined( USE_MMX ) | 44 | * differs from the inffast.c algorithm. I don't think there would have been |
45 | .align 4,0 | 45 | * read overruns where a page boundary was crossed (a segfault), but there |
46 | .L_mask: /* mask[N] = ( 1 << N ) - 1 */ | 46 | * could have been overruns when next_in ends on unaligned memory (unintialized |
47 | .long 0 | 47 | * memory read). |
48 | .long 1 | 48 | * |
49 | .long 3 | 49 | * Mar-13-2003 -- P4 MMX is slightly slower than P4 NO_MMX. I created a C |
50 | .long 7 | 50 | * version of the non-MMX code so that it doesn't depend on zstrm and zstate |
51 | .long 15 | 51 | * structure offsets which are hard coded in this file. This was last tested |
52 | .long 31 | 52 | * with zlib-1.2.0 which is currently in beta testing, newer versions of this |
53 | .long 63 | 53 | * and inffas86.c can be found at http://www.eetbeetee.com/zlib/ and |
54 | .long 127 | 54 | * http://www.charm.net/~christop/zlib/ |
55 | .long 255 | 55 | */ |
56 | .long 511 | 56 | |
57 | .long 1023 | 57 | |
58 | .long 2047 | 58 | /* |
59 | .long 4095 | 59 | * if you have underscore linking problems (_inflate_fast undefined), try |
60 | .long 8191 | 60 | * using -DGAS_COFF |
61 | .long 16383 | 61 | */ |
62 | .long 32767 | 62 | #if ! defined( GAS_COFF ) && ! defined( GAS_ELF ) |
63 | .long 65535 | 63 | |
64 | .long 131071 | 64 | #if defined( __CYGWIN__ ) |
65 | .long 262143 | 65 | #define GAS_COFF /* windows object format */ |
66 | .long 524287 | 66 | #else |
67 | .long 1048575 | 67 | #define GAS_ELF |
68 | .long 2097151 | 68 | #endif |
69 | .long 4194303 | 69 | |
70 | .long 8388607 | 70 | #endif /* ! GAS_COFF && ! GAS_ELF */ |
71 | .long 16777215 | 71 | |
72 | .long 33554431 | 72 | |
73 | .long 67108863 | 73 | #if defined( GAS_COFF ) |
74 | .long 134217727 | 74 | |
75 | .long 268435455 | 75 | /* coff externals have underscores */ |
76 | .long 536870911 | 76 | #define inflate_fast _inflate_fast |
77 | .long 1073741823 | 77 | #define inflate_fast_use_mmx _inflate_fast_use_mmx |
78 | .long 2147483647 | 78 | |
79 | .long 4294967295 | 79 | #endif /* GAS_COFF */ |
80 | #endif | 80 | |
81 | 81 | ||
82 | .text | 82 | .file "inffast.S" |
83 | 83 | ||
84 | /* | 84 | .globl inflate_fast |
85 | * struct z_stream offsets, in zlib.h | 85 | |
86 | */ | 86 | .text |
87 | #define next_in_strm 0 /* strm->next_in */ | 87 | .align 4,0 |
88 | #define avail_in_strm 4 /* strm->avail_in */ | 88 | .L_invalid_literal_length_code_msg: |
89 | #define next_out_strm 12 /* strm->next_out */ | 89 | .string "invalid literal/length code" |
90 | #define avail_out_strm 16 /* strm->avail_out */ | 90 | |
91 | #define msg_strm 24 /* strm->msg */ | 91 | .align 4,0 |
92 | #define state_strm 28 /* strm->state */ | 92 | .L_invalid_distance_code_msg: |
93 | 93 | .string "invalid distance code" | |
94 | /* | 94 | |
95 | * struct inflate_state offsets, in inflate.h | 95 | .align 4,0 |
96 | */ | 96 | .L_invalid_distance_too_far_msg: |
97 | #define mode_state 0 /* state->mode */ | 97 | .string "invalid distance too far back" |
98 | #define wsize_state 32 /* state->wsize */ | 98 | |
99 | #define write_state 36 /* state->write */ | 99 | #if ! defined( NO_MMX ) |
100 | #define window_state 40 /* state->window */ | 100 | .align 4,0 |
101 | #define hold_state 44 /* state->hold */ | 101 | .L_mask: /* mask[N] = ( 1 << N ) - 1 */ |
102 | #define bits_state 48 /* state->bits */ | 102 | .long 0 |
103 | #define lencode_state 64 /* state->lencode */ | 103 | .long 1 |
104 | #define distcode_state 68 /* state->distcode */ | 104 | .long 3 |
105 | #define lenbits_state 72 /* state->lenbits */ | 105 | .long 7 |
106 | #define distbits_state 76 /* state->distbits */ | 106 | .long 15 |
107 | 107 | .long 31 | |
108 | /* | 108 | .long 63 |
109 | * inflate_fast's activation record | 109 | .long 127 |
110 | */ | 110 | .long 255 |
111 | #define local_var_size 56 /* how much local space for vars */ | 111 | .long 511 |
112 | #define strm_sp 80 /* first arg: z_stream * (local_var_size + 24) */ | 112 | .long 1023 |
113 | #define start_sp 84 /* second arg: unsigned int (local_var_size + 28) */ | 113 | .long 2047 |
114 | 114 | .long 4095 | |
115 | /* | 115 | .long 8191 |
116 | * offsets for local vars on stack | 116 | .long 16383 |
117 | */ | 117 | .long 32767 |
118 | #define out 52 /* unsigned char* */ | 118 | .long 65535 |
119 | #define window 48 /* unsigned char* */ | 119 | .long 131071 |
120 | #define wsize 44 /* unsigned int */ | 120 | .long 262143 |
121 | #define write 40 /* unsigned int */ | 121 | .long 524287 |
122 | #define in 36 /* unsigned char* */ | 122 | .long 1048575 |
123 | #define beg 32 /* unsigned char* */ | 123 | .long 2097151 |
124 | #define dist 28 /* unsigned int */ | 124 | .long 4194303 |
125 | #define len 24 /* unsigned int */ | 125 | .long 8388607 |
126 | #define last 20 /* unsigned char* */ | 126 | .long 16777215 |
127 | #define end 16 /* unsigned char* */ | 127 | .long 33554431 |
128 | #define dcode 12 /* code* */ | 128 | .long 67108863 |
129 | #define lcode 8 /* code* */ | 129 | .long 134217727 |
130 | #define dmask 4 /* unsigned int */ | 130 | .long 268435455 |
131 | #define lmask 0 /* unsigned int */ | 131 | .long 536870911 |
132 | 132 | .long 1073741823 | |
133 | /* | 133 | .long 2147483647 |
134 | * typedef enum inflate_mode consts, in inflate.h | 134 | .long 4294967295 |
135 | */ | 135 | #endif /* NO_MMX */ |
136 | #ifndef NO_GUNZIP | 136 | |
137 | #define GUNZIP | 137 | .text |
138 | #endif | 138 | |
139 | 139 | /* | |
140 | #ifdef GUNZIP | 140 | * struct z_stream offsets, in zlib.h |
141 | #define INFLATE_MODE_TYPE 11 /* state->mode flags enum-ed in inflate.h */ | 141 | */ |
142 | #define INFLATE_MODE_BAD 26 | 142 | #define next_in_strm 0 /* strm->next_in */ |
143 | #else | 143 | #define avail_in_strm 4 /* strm->avail_in */ |
144 | #define INFLATE_MODE_TYPE 3 | 144 | #define next_out_strm 12 /* strm->next_out */ |
145 | #define INFLATE_MODE_BAD 17 | 145 | #define avail_out_strm 16 /* strm->avail_out */ |
146 | #endif | 146 | #define msg_strm 24 /* strm->msg */ |
147 | 147 | #define state_strm 28 /* strm->state */ | |
148 | 148 | ||
149 | .align 16,0x90 | 149 | /* |
150 | inflate_fast: | 150 | * struct inflate_state offsets, in inflate.h |
151 | pushl %edi | 151 | */ |
152 | pushl %esi | 152 | #define mode_state 0 /* state->mode */ |
153 | pushl %ebp | 153 | #define wsize_state 32 /* state->wsize */ |
154 | pushl %ebx | 154 | #define write_state 36 /* state->write */ |
155 | pushf /* save eflags (strm_sp, state_sp assumes this is 32 bits) */ | 155 | #define window_state 40 /* state->window */ |
156 | subl $local_var_size, %esp | 156 | #define hold_state 44 /* state->hold */ |
157 | cld | 157 | #define bits_state 48 /* state->bits */ |
158 | #if defined( USE_MMX ) | 158 | #define lencode_state 64 /* state->lencode */ |
159 | emms | 159 | #define distcode_state 68 /* state->distcode */ |
160 | #endif | 160 | #define lenbits_state 72 /* state->lenbits */ |
161 | 161 | #define distbits_state 76 /* state->distbits */ | |
162 | #define strm_r %esi | 162 | |
163 | #define state_r %edi | 163 | /* |
164 | 164 | * inflate_fast's activation record | |
165 | movl strm_sp(%esp), strm_r | 165 | */ |
166 | movl state_strm(strm_r), state_r | 166 | #define local_var_size 64 /* how much local space for vars */ |
167 | 167 | #define strm_sp 88 /* first arg: z_stream * (local_var_size + 24) */ | |
168 | /* in = strm->next_in; | 168 | #define start_sp 92 /* second arg: unsigned int (local_var_size + 28) */ |
169 | * out = strm->next_out; | 169 | |
170 | * last = in + strm->avail_in - 5; | 170 | /* |
171 | * beg = out - (start - strm->avail_out); | 171 | * offsets for local vars on stack |
172 | * end = out + (strm->avail_out - 257); | 172 | */ |
173 | */ | 173 | #define out 60 /* unsigned char* */ |
174 | movl next_in_strm(strm_r), %eax | 174 | #define window 56 /* unsigned char* */ |
175 | movl next_out_strm(strm_r), %ebx | 175 | #define wsize 52 /* unsigned int */ |
176 | movl avail_in_strm(strm_r), %edx | 176 | #define write 48 /* unsigned int */ |
177 | movl avail_out_strm(strm_r), %ecx | 177 | #define in 44 /* unsigned char* */ |
178 | movl start_sp(%esp), %ebp | 178 | #define beg 40 /* unsigned char* */ |
179 | 179 | #define buf 28 /* char[ 12 ] */ | |
180 | addl %eax, %edx /* avail_in += next_in */ | 180 | #define len 24 /* unsigned int */ |
181 | subl $5, %edx /* avail_in -= 5 */ | 181 | #define last 20 /* unsigned char* */ |
182 | 182 | #define end 16 /* unsigned char* */ | |
183 | subl %ecx, %ebp /* start -= avail_out */ | 183 | #define dcode 12 /* code* */ |
184 | negl %ebp /* start = -start */ | 184 | #define lcode 8 /* code* */ |
185 | addl %ebx, %ebp /* start += next_out */ | 185 | #define dmask 4 /* unsigned int */ |
186 | 186 | #define lmask 0 /* unsigned int */ | |
187 | subl $257, %ecx /* avail_out -= 257 */ | 187 | |
188 | addl %ebx, %ecx /* avail_out += out */ | 188 | /* |
189 | 189 | * typedef enum inflate_mode consts, in inflate.h | |
190 | movl %eax, in(%esp) | 190 | */ |
191 | movl %ebx, out(%esp) | 191 | #ifndef NO_GUNZIP |
192 | movl %edx, last(%esp) | 192 | #define GUNZIP |
193 | movl %ebp, beg(%esp) | 193 | #endif |
194 | movl %ecx, end(%esp) | 194 | |
195 | 195 | #ifdef GUNZIP | |
196 | /* wsize = state->wsize; | 196 | #define INFLATE_MODE_TYPE 11 /* state->mode flags enum-ed in inflate.h */ |
197 | * write = state->write; | 197 | #define INFLATE_MODE_BAD 26 |
198 | * window = state->window; | 198 | #else |
199 | * hold = state->hold; | 199 | #define INFLATE_MODE_TYPE 3 |
200 | * bits = state->bits; | 200 | #define INFLATE_MODE_BAD 17 |
201 | * lcode = state->lencode; | 201 | #endif |
202 | * dcode = state->distcode; | 202 | |
203 | * lmask = ( 1 << state->lenbits ) - 1; | 203 | |
204 | * dmask = ( 1 << state->distbits ) - 1; | 204 | #if ! defined( USE_MMX ) && ! defined( NO_MMX ) |
205 | */ | 205 | |
206 | 206 | #define RUN_TIME_MMX | |
207 | movl lencode_state(state_r), %eax | 207 | |
208 | movl distcode_state(state_r), %ecx | 208 | #define CHECK_MMX 1 |
209 | 209 | #define DO_USE_MMX 2 | |
210 | movl %eax, lcode(%esp) | 210 | #define DONT_USE_MMX 3 |
211 | movl %ecx, dcode(%esp) | 211 | |
212 | 212 | .globl inflate_fast_use_mmx | |
213 | movl $1, %eax | 213 | |
214 | movl lenbits_state(state_r), %ecx | 214 | .data |
215 | shll %cl, %eax | 215 | |
216 | decl %eax | 216 | .align 4,0 |
217 | movl %eax, lmask(%esp) | 217 | inflate_fast_use_mmx: /* integer flag for run time control 1=check,2=mmx,3=no */ |
218 | 218 | .long CHECK_MMX | |
219 | movl $1, %eax | 219 | |
220 | movl distbits_state(state_r), %ecx | 220 | #if defined( GAS_ELF ) |
221 | shll %cl, %eax | 221 | /* elf info */ |
222 | decl %eax | 222 | .type inflate_fast_use_mmx,@object |
223 | movl %eax, dmask(%esp) | 223 | .size inflate_fast_use_mmx,4 |
224 | 224 | #endif | |
225 | movl wsize_state(state_r), %eax | 225 | |
226 | movl write_state(state_r), %ecx | 226 | #endif /* RUN_TIME_MMX */ |
227 | movl window_state(state_r), %edx | 227 | |
228 | 228 | #if defined( GAS_COFF ) | |
229 | movl %eax, wsize(%esp) | 229 | /* coff info: scl 2 = extern, type 32 = function */ |
230 | movl %ecx, write(%esp) | 230 | .def inflate_fast; .scl 2; .type 32; .endef |
231 | movl %edx, window(%esp) | 231 | #endif |
232 | 232 | ||
233 | #if ! defined( USE_MMX ) | 233 | .text |
234 | 234 | ||
235 | #define hold_r %ebp | 235 | .align 32,0x90 |
236 | #define bits_r %bl | 236 | inflate_fast: |
237 | #define bitslong_r %ebx | 237 | pushl %edi |
238 | 238 | pushl %esi | |
239 | movl hold_state(state_r), hold_r | 239 | pushl %ebp |
240 | movl bits_state(state_r), bitslong_r | 240 | pushl %ebx |
241 | 241 | pushf /* save eflags (strm_sp, state_sp assumes this is 32 bits) */ | |
242 | #else /* USE_MMX */ | 242 | subl $local_var_size, %esp |
243 | 243 | cld | |
244 | #define hold_mm %mm0 | 244 | |
245 | #define bits_r %ebp | 245 | #define strm_r %esi |
246 | #define bitslong_r %ebp | 246 | #define state_r %edi |
247 | 247 | ||
248 | movl hold_state(state_r), %ebx | 248 | movl strm_sp(%esp), strm_r |
249 | movl bits_state(state_r), bitslong_r | 249 | movl state_strm(strm_r), state_r |
250 | 250 | ||
251 | #endif | 251 | /* in = strm->next_in; |
252 | 252 | * out = strm->next_out; | |
253 | #undef strm_r | 253 | * last = in + strm->avail_in - 11; |
254 | #undef state_r | 254 | * beg = out - (start - strm->avail_out); |
255 | #define in_r %esi | 255 | * end = out + (strm->avail_out - 257); |
256 | #define from_r %esi | 256 | */ |
257 | #define out_r %edi | 257 | movl avail_in_strm(strm_r), %edx |
258 | 258 | movl next_in_strm(strm_r), %eax | |
259 | movl in(%esp), in_r | 259 | |
260 | 260 | addl %eax, %edx /* avail_in += next_in */ | |
261 | #if ! defined ( USE_MMX ) | 261 | subl $11, %edx /* avail_in -= 11 */ |
262 | 262 | ||
263 | /* align in_r on word boundary */ | 263 | movl %eax, in(%esp) |
264 | testl $1, in_r | 264 | movl %edx, last(%esp) |
265 | jz .L_is_aligned | 265 | |
266 | xorl %eax, %eax | 266 | movl start_sp(%esp), %ebp |
267 | movb (in_r), %al | 267 | movl avail_out_strm(strm_r), %ecx |
268 | incl in_r | 268 | movl next_out_strm(strm_r), %ebx |
269 | movb bits_r, %cl | 269 | |
270 | addb $8, bits_r | 270 | subl %ecx, %ebp /* start -= avail_out */ |
271 | shll %cl, %eax | 271 | negl %ebp /* start = -start */ |
272 | orl %eax, hold_r | 272 | addl %ebx, %ebp /* start += next_out */ |
273 | 273 | ||
274 | #else | 274 | subl $257, %ecx /* avail_out -= 257 */ |
275 | /* align in_r on long boundary */ | 275 | addl %ebx, %ecx /* avail_out += out */ |
276 | .L_align_long: | 276 | |
277 | testl $3, in_r | 277 | movl %ebx, out(%esp) |
278 | jz .L_is_aligned | 278 | movl %ebp, beg(%esp) |
279 | xorl %eax, %eax | 279 | movl %ecx, end(%esp) |
280 | movb (in_r), %al | 280 | |
281 | incl in_r | 281 | /* wsize = state->wsize; |
282 | movl bits_r, %ecx | 282 | * write = state->write; |
283 | addl $8, bits_r | 283 | * window = state->window; |
284 | shll %cl, %eax | 284 | * hold = state->hold; |
285 | orl %eax, %ebx | 285 | * bits = state->bits; |
286 | jmp .L_align_long | 286 | * lcode = state->lencode; |
287 | 287 | * dcode = state->distcode; | |
288 | #endif | 288 | * lmask = ( 1 << state->lenbits ) - 1; |
289 | 289 | * dmask = ( 1 << state->distbits ) - 1; | |
290 | .L_is_aligned: | 290 | */ |
291 | movl out(%esp), out_r | 291 | |
292 | 292 | movl lencode_state(state_r), %eax | |
293 | #if defined ( USE_MMX ) | 293 | movl distcode_state(state_r), %ecx |
294 | 294 | ||
295 | #define used_mm %mm1 | 295 | movl %eax, lcode(%esp) |
296 | #define dmask2_mm %mm2 | 296 | movl %ecx, dcode(%esp) |
297 | #define lmask2_mm %mm3 | 297 | |
298 | #define lmask_mm %mm4 | 298 | movl $1, %eax |
299 | #define dmask_mm %mm5 | 299 | movl lenbits_state(state_r), %ecx |
300 | #define tmp_mm %mm6 | 300 | shll %cl, %eax |
301 | 301 | decl %eax | |
302 | movl out(%esp), out_r | 302 | movl %eax, lmask(%esp) |
303 | movd lmask(%esp), lmask_mm | 303 | |
304 | movq lmask_mm, lmask2_mm | 304 | movl $1, %eax |
305 | movd dmask(%esp), dmask_mm | 305 | movl distbits_state(state_r), %ecx |
306 | movq dmask_mm, dmask2_mm | 306 | shll %cl, %eax |
307 | movd %ebx, hold_mm | 307 | decl %eax |
308 | pxor used_mm, used_mm | 308 | movl %eax, dmask(%esp) |
309 | movl lcode(%esp), %ebx /* ebx = lcode */ | 309 | |
310 | #endif | 310 | movl wsize_state(state_r), %eax |
311 | 311 | movl write_state(state_r), %ecx | |
312 | jmp .L_do_loop | 312 | movl window_state(state_r), %edx |
313 | 313 | ||
314 | .align 16,0x90 | 314 | movl %eax, wsize(%esp) |
315 | 315 | movl %ecx, write(%esp) | |
316 | #if ! defined ( USE_MMX ) | 316 | movl %edx, window(%esp) |
317 | 317 | ||
318 | .L_do_loop: | 318 | movl hold_state(state_r), %ebp |
319 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out | 319 | movl bits_state(state_r), %ebx |
320 | * | 320 | |
321 | * do { | 321 | #undef strm_r |
322 | * if (bits < 15) { | 322 | #undef state_r |
323 | * hold |= *((unsigned short *)in)++ << bits; | 323 | |
324 | * bits += 16 | 324 | #define in_r %esi |
325 | * } | 325 | #define from_r %esi |
326 | * this = lcode[hold & lmask] | 326 | #define out_r %edi |
327 | */ | 327 | |
328 | cmpb $15, bits_r | 328 | movl in(%esp), in_r |
329 | ja .L_get_length_code /* if (15 < bits) */ | 329 | movl last(%esp), %ecx |
330 | 330 | cmpl in_r, %ecx | |
331 | xorl %eax, %eax | 331 | ja .L_align_long /* if in < last */ |
332 | lodsw /* al = *(ushort *)in++ */ | 332 | |
333 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | 333 | addl $11, %ecx /* ecx = &in[ avail_in ] */ |
334 | addb $16, bits_r /* bits += 16 */ | 334 | subl in_r, %ecx /* ecx = avail_in */ |
335 | shll %cl, %eax | 335 | movl $12, %eax |
336 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | 336 | subl %ecx, %eax /* eax = 12 - avail_in */ |
337 | 337 | leal buf(%esp), %edi | |
338 | .L_get_length_code: | 338 | rep movsb /* memcpy( buf, in, avail_in ) */ |
339 | movl lmask(%esp), %edx /* edx = lmask */ | 339 | movl %eax, %ecx |
340 | movl lcode(%esp), %ecx /* ecx = lcode */ | 340 | xorl %eax, %eax |
341 | andl hold_r, %edx /* edx &= hold */ | 341 | rep stosb /* memset( &buf[ avail_in ], 0, 12 - avail_in ) */ |
342 | movl (%ecx,%edx,4), %eax /* eax = lcode[hold & lmask] */ | 342 | leal buf(%esp), in_r /* in = buf */ |
343 | 343 | movl in_r, last(%esp) /* last = in, do just one iteration */ | |
344 | #else /* USE_MMX */ | 344 | jmp .L_is_aligned |
345 | 345 | ||
346 | .L_do_loop: | 346 | /* align in_r on long boundary */ |
347 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 347 | .L_align_long: |
348 | 348 | testl $3, in_r | |
349 | cmpl $32, bits_r | 349 | jz .L_is_aligned |
350 | ja .L_get_length_code /* if (32 < bits) */ | 350 | xorl %eax, %eax |
351 | 351 | movb (in_r), %al | |
352 | movd bits_r, tmp_mm | 352 | incl in_r |
353 | movd (in_r), %mm7 | 353 | movl %ebx, %ecx |
354 | addl $4, in_r | 354 | addl $8, %ebx |
355 | psllq tmp_mm, %mm7 | 355 | shll %cl, %eax |
356 | addl $32, bits_r | 356 | orl %eax, %ebp |
357 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ | 357 | jmp .L_align_long |
358 | 358 | ||
359 | .L_get_length_code: | 359 | .L_is_aligned: |
360 | pand hold_mm, lmask_mm | 360 | movl out(%esp), out_r |
361 | movd lmask_mm, %eax | 361 | |
362 | movq lmask2_mm, lmask_mm | 362 | #if defined( NO_MMX ) |
363 | movl (%ebx,%eax,4), %eax /* eax = lcode[hold & lmask] */ | 363 | jmp .L_do_loop |
364 | 364 | #endif | |
365 | #endif | 365 | |
366 | 366 | #if defined( USE_MMX ) | |
367 | #if ! defined( USE_MMX ) | 367 | jmp .L_init_mmx |
368 | 368 | #endif | |
369 | .L_dolen: | 369 | |
370 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out | 370 | /*** Runtime MMX check ***/ |
371 | * | 371 | |
372 | * dolen: | 372 | #if defined( RUN_TIME_MMX ) |
373 | * bits -= this.bits; | 373 | .L_check_mmx: |
374 | * hold >>= this.bits | 374 | cmpl $DO_USE_MMX, inflate_fast_use_mmx |
375 | */ | 375 | je .L_init_mmx |
376 | movb %ah, %cl /* cl = this.bits */ | 376 | ja .L_do_loop /* > 2 */ |
377 | subb %ah, bits_r /* bits -= this.bits */ | 377 | |
378 | shrl %cl, hold_r /* hold >>= this.bits */ | 378 | pushl %eax |
379 | 379 | pushl %ebx | |
380 | /* check if op is a literal | 380 | pushl %ecx |
381 | * if (op == 0) { | 381 | pushl %edx |
382 | * PUP(out) = this.val; | 382 | pushf |
383 | * } | 383 | movl (%esp), %eax /* copy eflags to eax */ |
384 | */ | 384 | xorl $0x200000, (%esp) /* try toggling ID bit of eflags (bit 21) |
385 | testb %al, %al | 385 | * to see if cpu supports cpuid... |
386 | jnz .L_test_for_length_base /* if (op != 0) 45.7% */ | 386 | * ID bit method not supported by NexGen but |
387 | 387 | * bios may load a cpuid instruction and | |
388 | shrl $16, %eax /* output this.val char */ | 388 | * cpuid may be disabled on Cyrix 5-6x86 */ |
389 | stosb | 389 | popf |
390 | 390 | pushf | |
391 | #else /* USE_MMX */ | 391 | popl %edx /* copy new eflags to edx */ |
392 | 392 | xorl %eax, %edx /* test if ID bit is flipped */ | |
393 | #define len_r %edx | 393 | jz .L_dont_use_mmx /* not flipped if zero */ |
394 | 394 | xorl %eax, %eax | |
395 | .L_dolen: | 395 | cpuid |
396 | movzbl %ah, %ecx /* ecx = this.bits */ | 396 | cmpl $0x756e6547, %ebx /* check for GenuineIntel in ebx,ecx,edx */ |
397 | movl %eax, len_r /* len = this */ | 397 | jne .L_dont_use_mmx |
398 | shrl $16, len_r /* len = this.val */ | 398 | cmpl $0x6c65746e, %ecx |
399 | movd %ecx, used_mm | 399 | jne .L_dont_use_mmx |
400 | subl %ecx, bits_r /* bits -= this.bits */ | 400 | cmpl $0x49656e69, %edx |
401 | 401 | jne .L_dont_use_mmx | |
402 | testb %al, %al | 402 | movl $1, %eax |
403 | jnz .L_test_for_length_base /* if (op != 0) 45.7% */ | 403 | cpuid /* get cpu features */ |
404 | 404 | shrl $8, %eax | |
405 | movb %dl, (out_r) | 405 | andl $15, %eax |
406 | incl out_r | 406 | cmpl $6, %eax /* check for Pentium family, is 0xf for P4 */ |
407 | 407 | jne .L_dont_use_mmx | |
408 | #endif | 408 | testl $0x800000, %edx /* test if MMX feature is set (bit 23) */ |
409 | 409 | jnz .L_use_mmx | |
410 | .L_while_test: | 410 | jmp .L_dont_use_mmx |
411 | /* while (in < last && out < end) | 411 | .L_use_mmx: |
412 | */ | 412 | movl $DO_USE_MMX, inflate_fast_use_mmx |
413 | cmpl out_r, end(%esp) | 413 | jmp .L_check_mmx_pop |
414 | jbe .L_break_loop /* if (out >= end) */ | 414 | .L_dont_use_mmx: |
415 | 415 | movl $DONT_USE_MMX, inflate_fast_use_mmx | |
416 | cmpl in_r, last(%esp) | 416 | .L_check_mmx_pop: |
417 | ja .L_do_loop /* if (in < last) */ | 417 | popl %edx |
418 | jmp .L_break_loop | 418 | popl %ecx |
419 | 419 | popl %ebx | |
420 | #if ! defined( USE_MMX ) | 420 | popl %eax |
421 | 421 | jmp .L_check_mmx | |
422 | .L_test_for_length_base: | 422 | #endif |
423 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = len | 423 | |
424 | * | 424 | |
425 | * else if (op & 16) { | 425 | /*** Non-MMX code ***/ |
426 | * len = this.val | 426 | |
427 | * op &= 15 | 427 | #if defined ( NO_MMX ) || defined( RUN_TIME_MMX ) |
428 | * if (op) { | 428 | |
429 | * if (op > bits) { | 429 | #define hold_r %ebp |
430 | * hold |= *((unsigned short *)in)++ << bits; | 430 | #define bits_r %bl |
431 | * bits += 16 | 431 | #define bitslong_r %ebx |
432 | * } | 432 | |
433 | * len += hold & mask[op]; | 433 | .align 32,0x90 |
434 | * bits -= op; | 434 | .L_while_test: |
435 | * hold >>= op; | 435 | /* while (in < last && out < end) |
436 | * } | 436 | */ |
437 | */ | 437 | cmpl out_r, end(%esp) |
438 | #define len_r %edx | 438 | jbe .L_break_loop /* if (out >= end) */ |
439 | movl %eax, len_r /* len = this */ | 439 | |
440 | shrl $16, len_r /* len = this.val */ | 440 | cmpl in_r, last(%esp) |
441 | movb %al, %cl | 441 | jbe .L_break_loop |
442 | 442 | ||
443 | testb $16, %al | 443 | .L_do_loop: |
444 | jz .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ | 444 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out |
445 | andb $15, %cl /* op &= 15 */ | 445 | * |
446 | jz .L_save_len /* if (!op) */ | 446 | * do { |
447 | cmpb %cl, bits_r | 447 | * if (bits < 15) { |
448 | jae .L_add_bits_to_len /* if (op <= bits) */ | 448 | * hold |= *((unsigned short *)in)++ << bits; |
449 | 449 | * bits += 16 | |
450 | movb %cl, %ch /* stash op in ch, freeing cl */ | 450 | * } |
451 | xorl %eax, %eax | 451 | * this = lcode[hold & lmask] |
452 | lodsw /* al = *(ushort *)in++ */ | 452 | */ |
453 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | 453 | cmpb $15, bits_r |
454 | addb $16, bits_r /* bits += 16 */ | 454 | ja .L_get_length_code /* if (15 < bits) */ |
455 | shll %cl, %eax | 455 | |
456 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | 456 | xorl %eax, %eax |
457 | movb %ch, %cl /* move op back to ecx */ | 457 | lodsw /* al = *(ushort *)in++ */ |
458 | 458 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | |
459 | .L_add_bits_to_len: | 459 | addb $16, bits_r /* bits += 16 */ |
460 | movl $1, %eax | 460 | shll %cl, %eax |
461 | shll %cl, %eax | 461 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ |
462 | decl %eax | 462 | |
463 | subb %cl, bits_r | 463 | .L_get_length_code: |
464 | andl hold_r, %eax /* eax &= hold */ | 464 | movl lmask(%esp), %edx /* edx = lmask */ |
465 | shrl %cl, hold_r | 465 | movl lcode(%esp), %ecx /* ecx = lcode */ |
466 | addl %eax, len_r /* len += hold & mask[op] */ | 466 | andl hold_r, %edx /* edx &= hold */ |
467 | 467 | movl (%ecx,%edx,4), %eax /* eax = lcode[hold & lmask] */ | |
468 | .L_save_len: | 468 | |
469 | movl len_r, len(%esp) /* save len */ | 469 | .L_dolen: |
470 | #undef len_r | 470 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out |
471 | 471 | * | |
472 | .L_decode_distance: | 472 | * dolen: |
473 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 473 | * bits -= this.bits; |
474 | * | 474 | * hold >>= this.bits |
475 | * if (bits < 15) { | 475 | */ |
476 | * hold |= *((unsigned short *)in)++ << bits; | 476 | movb %ah, %cl /* cl = this.bits */ |
477 | * bits += 16 | 477 | subb %ah, bits_r /* bits -= this.bits */ |
478 | * } | 478 | shrl %cl, hold_r /* hold >>= this.bits */ |
479 | * this = dcode[hold & dmask]; | 479 | |
480 | * dodist: | 480 | /* check if op is a literal |
481 | * bits -= this.bits; | 481 | * if (op == 0) { |
482 | * hold >>= this.bits; | 482 | * PUP(out) = this.val; |
483 | * op = this.op; | 483 | * } |
484 | */ | 484 | */ |
485 | 485 | testb %al, %al | |
486 | cmpb $15, bits_r | 486 | jnz .L_test_for_length_base /* if (op != 0) 45.7% */ |
487 | ja .L_get_distance_code /* if (15 < bits) */ | 487 | |
488 | 488 | shrl $16, %eax /* output this.val char */ | |
489 | xorl %eax, %eax | 489 | stosb |
490 | lodsw /* al = *(ushort *)in++ */ | 490 | jmp .L_while_test |
491 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | 491 | |
492 | addb $16, bits_r /* bits += 16 */ | 492 | .L_test_for_length_base: |
493 | shll %cl, %eax | 493 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = len |
494 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | 494 | * |
495 | 495 | * else if (op & 16) { | |
496 | .L_get_distance_code: | 496 | * len = this.val |
497 | movl dmask(%esp), %edx /* edx = dmask */ | 497 | * op &= 15 |
498 | movl dcode(%esp), %ecx /* ecx = dcode */ | 498 | * if (op) { |
499 | andl hold_r, %edx /* edx &= hold */ | 499 | * if (op > bits) { |
500 | movl (%ecx,%edx,4), %eax /* eax = dcode[hold & dmask] */ | 500 | * hold |= *((unsigned short *)in)++ << bits; |
501 | 501 | * bits += 16 | |
502 | #else /* USE_MMX */ | 502 | * } |
503 | 503 | * len += hold & mask[op]; | |
504 | .L_test_for_length_base: | 504 | * bits -= op; |
505 | testb $16, %al | 505 | * hold >>= op; |
506 | jz .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ | 506 | * } |
507 | andl $15, %eax /* op &= 15 */ | 507 | */ |
508 | jz .L_decode_distance /* if (!op) */ | 508 | #define len_r %edx |
509 | 509 | movl %eax, len_r /* len = this */ | |
510 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 510 | shrl $16, len_r /* len = this.val */ |
511 | movd %eax, used_mm | 511 | movb %al, %cl |
512 | movd hold_mm, %ecx | 512 | |
513 | subl %eax, bits_r | 513 | testb $16, %al |
514 | andl .L_mask(,%eax,4), %ecx | 514 | jz .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ |
515 | addl %ecx, len_r /* len += hold & mask[op] */ | 515 | andb $15, %cl /* op &= 15 */ |
516 | 516 | jz .L_save_len /* if (!op) */ | |
517 | .L_decode_distance: | 517 | cmpb %cl, bits_r |
518 | 518 | jae .L_add_bits_to_len /* if (op <= bits) */ | |
519 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 519 | |
520 | 520 | movb %cl, %ch /* stash op in ch, freeing cl */ | |
521 | cmpl $32, bits_r | 521 | xorl %eax, %eax |
522 | ja .L_get_dist_code /* if (32 < bits) */ | 522 | lodsw /* al = *(ushort *)in++ */ |
523 | 523 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | |
524 | movd bits_r, tmp_mm | 524 | addb $16, bits_r /* bits += 16 */ |
525 | movd (in_r), %mm7 | 525 | shll %cl, %eax |
526 | addl $4, in_r | 526 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ |
527 | psllq tmp_mm, %mm7 | 527 | movb %ch, %cl /* move op back to ecx */ |
528 | addl $32, bits_r | 528 | |
529 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ | 529 | .L_add_bits_to_len: |
530 | 530 | movl $1, %eax | |
531 | .L_get_dist_code: | 531 | shll %cl, %eax |
532 | movl dcode(%esp), %ebx /* ebx = dcode */ | 532 | decl %eax |
533 | pand hold_mm, dmask_mm | 533 | subb %cl, bits_r |
534 | movd dmask_mm, %eax | 534 | andl hold_r, %eax /* eax &= hold */ |
535 | movq dmask2_mm, dmask_mm | 535 | shrl %cl, hold_r |
536 | movl (%ebx,%eax,4), %eax /* eax = dcode[hold & lmask] */ | 536 | addl %eax, len_r /* len += hold & mask[op] */ |
537 | 537 | ||
538 | #endif | 538 | .L_save_len: |
539 | 539 | movl len_r, len(%esp) /* save len */ | |
540 | #if ! defined( USE_MMX ) | 540 | #undef len_r |
541 | 541 | ||
542 | #define dist_r %edx | 542 | .L_decode_distance: |
543 | .L_dodist: | 543 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
544 | movl %eax, dist_r /* dist = this */ | 544 | * |
545 | shrl $16, dist_r /* dist = this.val */ | 545 | * if (bits < 15) { |
546 | movb %ah, %cl | 546 | * hold |= *((unsigned short *)in)++ << bits; |
547 | subb %ah, bits_r /* bits -= this.bits */ | 547 | * bits += 16 |
548 | shrl %cl, hold_r /* hold >>= this.bits */ | 548 | * } |
549 | 549 | * this = dcode[hold & dmask]; | |
550 | /* if (op & 16) { | 550 | * dodist: |
551 | * dist = this.val | 551 | * bits -= this.bits; |
552 | * op &= 15 | 552 | * hold >>= this.bits; |
553 | * if (op > bits) { | 553 | * op = this.op; |
554 | * hold |= *((unsigned short *)in)++ << bits; | 554 | */ |
555 | * bits += 16 | 555 | |
556 | * } | 556 | cmpb $15, bits_r |
557 | * dist += hold & mask[op]; | 557 | ja .L_get_distance_code /* if (15 < bits) */ |
558 | * bits -= op; | 558 | |
559 | * hold >>= op; | 559 | xorl %eax, %eax |
560 | */ | 560 | lodsw /* al = *(ushort *)in++ */ |
561 | movb %al, %cl /* cl = this.op */ | 561 | movb bits_r, %cl /* cl = bits, needs it for shifting */ |
562 | 562 | addb $16, bits_r /* bits += 16 */ | |
563 | testb $16, %al /* if ((op & 16) == 0) */ | 563 | shll %cl, %eax |
564 | jz .L_test_for_second_level_dist | 564 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ |
565 | andb $15, %cl /* op &= 15 */ | 565 | |
566 | jz .L_check_dist_one | 566 | .L_get_distance_code: |
567 | cmpb %cl, bits_r | 567 | movl dmask(%esp), %edx /* edx = dmask */ |
568 | jae .L_add_bits_to_dist /* if (op <= bits) 97.6% */ | 568 | movl dcode(%esp), %ecx /* ecx = dcode */ |
569 | 569 | andl hold_r, %edx /* edx &= hold */ | |
570 | movb %cl, %ch /* stash op in ch, freeing cl */ | 570 | movl (%ecx,%edx,4), %eax /* eax = dcode[hold & dmask] */ |
571 | xorl %eax, %eax | 571 | |
572 | lodsw /* al = *(ushort *)in++ */ | 572 | #define dist_r %edx |
573 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | 573 | .L_dodist: |
574 | addb $16, bits_r /* bits += 16 */ | 574 | movl %eax, dist_r /* dist = this */ |
575 | shll %cl, %eax | 575 | shrl $16, dist_r /* dist = this.val */ |
576 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | 576 | movb %ah, %cl |
577 | movb %ch, %cl /* move op back to ecx */ | 577 | subb %ah, bits_r /* bits -= this.bits */ |
578 | 578 | shrl %cl, hold_r /* hold >>= this.bits */ | |
579 | .L_add_bits_to_dist: | 579 | |
580 | movl $1, %eax | 580 | /* if (op & 16) { |
581 | shll %cl, %eax | 581 | * dist = this.val |
582 | decl %eax /* (1 << op) - 1 */ | 582 | * op &= 15 |
583 | subb %cl, bits_r | 583 | * if (op > bits) { |
584 | andl hold_r, %eax /* eax &= hold */ | 584 | * hold |= *((unsigned short *)in)++ << bits; |
585 | shrl %cl, hold_r | 585 | * bits += 16 |
586 | addl %eax, dist_r /* dist += hold & ((1 << op) - 1) */ | 586 | * } |
587 | jmp .L_check_window | 587 | * dist += hold & mask[op]; |
588 | 588 | * bits -= op; | |
589 | #else /* USE_MMX */ | 589 | * hold >>= op; |
590 | 590 | */ | |
591 | #define dist_r %ebx | 591 | movb %al, %cl /* cl = this.op */ |
592 | .L_dodist: | 592 | |
593 | movzbl %ah, %ecx /* ecx = this.bits */ | 593 | testb $16, %al /* if ((op & 16) == 0) */ |
594 | movl %eax, dist_r | 594 | jz .L_test_for_second_level_dist |
595 | shrl $16, dist_r /* dist = this.val */ | 595 | andb $15, %cl /* op &= 15 */ |
596 | subl %ecx, bits_r /* bits -= this.bits */ | 596 | jz .L_check_dist_one |
597 | movd %ecx, used_mm | 597 | cmpb %cl, bits_r |
598 | 598 | jae .L_add_bits_to_dist /* if (op <= bits) 97.6% */ | |
599 | testb $16, %al /* if ((op & 16) == 0) */ | 599 | |
600 | jz .L_test_for_second_level_dist | 600 | movb %cl, %ch /* stash op in ch, freeing cl */ |
601 | andl $15, %eax /* op &= 15 */ | 601 | xorl %eax, %eax |
602 | jz .L_check_dist_one | 602 | lodsw /* al = *(ushort *)in++ */ |
603 | 603 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | |
604 | .L_add_bits_to_dist: | 604 | addb $16, bits_r /* bits += 16 */ |
605 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 605 | shll %cl, %eax |
606 | movd %eax, used_mm /* save bit length of current op */ | 606 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ |
607 | movd hold_mm, %ecx /* get the next bits on input stream */ | 607 | movb %ch, %cl /* move op back to ecx */ |
608 | subl %eax, bits_r /* bits -= op bits */ | 608 | |
609 | andl .L_mask(,%eax,4), %ecx /* ecx = hold & mask[op] */ | 609 | .L_add_bits_to_dist: |
610 | addl %ecx, dist_r /* dist += hold & mask[op] */ | 610 | movl $1, %eax |
611 | jmp .L_check_window | 611 | shll %cl, %eax |
612 | 612 | decl %eax /* (1 << op) - 1 */ | |
613 | #endif | 613 | subb %cl, bits_r |
614 | 614 | andl hold_r, %eax /* eax &= hold */ | |
615 | .align 16,0x90 | 615 | shrl %cl, hold_r |
616 | 616 | addl %eax, dist_r /* dist += hold & ((1 << op) - 1) */ | |
617 | .L_check_dist_one: | 617 | jmp .L_check_window |
618 | cmpl $1, dist_r | 618 | |
619 | jne .L_check_window | 619 | .L_check_window: |
620 | cmpl out_r, beg(%esp) | 620 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
621 | je .L_check_window | 621 | * %ecx = nbytes |
622 | 622 | * | |
623 | decl out_r | 623 | * nbytes = out - beg; |
624 | #if ! defined( USE_MMX ) | 624 | * if (dist <= nbytes) { |
625 | movl len(%esp), %ecx | 625 | * from = out - dist; |
626 | #else | 626 | * do { |
627 | movl len_r, %ecx | 627 | * PUP(out) = PUP(from); |
628 | #endif | 628 | * } while (--len > 0) { |
629 | movb (out_r), %al | 629 | * } |
630 | subl $3, %ecx | 630 | */ |
631 | 631 | ||
632 | movb %al, 1(out_r) | 632 | movl in_r, in(%esp) /* save in so from can use it's reg */ |
633 | movb %al, 2(out_r) | 633 | movl out_r, %eax |
634 | movb %al, 3(out_r) | 634 | subl beg(%esp), %eax /* nbytes = out - beg */ |
635 | addl $4, out_r | 635 | |
636 | rep stosb | 636 | cmpl dist_r, %eax |
637 | 637 | jb .L_clip_window /* if (dist > nbytes) 4.2% */ | |
638 | #if defined( USE_MMX ) | 638 | |
639 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | 639 | movl len(%esp), %ecx |
640 | #endif | 640 | movl out_r, from_r |
641 | jmp .L_while_test | 641 | subl dist_r, from_r /* from = out - dist */ |
642 | 642 | ||
643 | .align 16,0x90 | 643 | subl $3, %ecx |
644 | 644 | movb (from_r), %al | |
645 | .L_check_window: | 645 | movb %al, (out_r) |
646 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 646 | movb 1(from_r), %al |
647 | * %ecx = nbytes | 647 | movb 2(from_r), %dl |
648 | * | 648 | addl $3, from_r |
649 | * nbytes = out - beg; | 649 | movb %al, 1(out_r) |
650 | * if (dist <= nbytes) { | 650 | movb %dl, 2(out_r) |
651 | * from = out - dist; | 651 | addl $3, out_r |
652 | * do { | 652 | rep movsb |
653 | * PUP(out) = PUP(from); | 653 | |
654 | * } while (--len > 0) { | 654 | movl in(%esp), in_r /* move in back to %esi, toss from */ |
655 | * } | 655 | jmp .L_while_test |
656 | */ | 656 | |
657 | 657 | .align 16,0x90 | |
658 | movl in_r, in(%esp) /* save in so from can use it's reg */ | 658 | .L_check_dist_one: |
659 | movl out_r, %eax | 659 | cmpl $1, dist_r |
660 | subl beg(%esp), %eax /* nbytes = out - beg */ | 660 | jne .L_check_window |
661 | 661 | cmpl out_r, beg(%esp) | |
662 | cmpl dist_r, %eax | 662 | je .L_check_window |
663 | jb .L_clip_window /* if (dist > nbytes) 4.2% */ | 663 | |
664 | 664 | decl out_r | |
665 | #if ! defined( USE_MMX ) | 665 | movl len(%esp), %ecx |
666 | movl len(%esp), %ecx | 666 | movb (out_r), %al |
667 | #else | 667 | subl $3, %ecx |
668 | movl len_r, %ecx | 668 | |
669 | #endif | 669 | movb %al, 1(out_r) |
670 | movl out_r, from_r | 670 | movb %al, 2(out_r) |
671 | subl dist_r, from_r /* from = out - dist */ | 671 | movb %al, 3(out_r) |
672 | 672 | addl $4, out_r | |
673 | subl $3, %ecx | 673 | rep stosb |
674 | movb (from_r), %al | 674 | |
675 | movb %al, (out_r) | 675 | jmp .L_while_test |
676 | movb 1(from_r), %al | 676 | |
677 | movb 2(from_r), %dl | 677 | .align 16,0x90 |
678 | addl $3, from_r | 678 | .L_test_for_second_level_length: |
679 | movb %al, 1(out_r) | 679 | /* else if ((op & 64) == 0) { |
680 | movb %dl, 2(out_r) | 680 | * this = lcode[this.val + (hold & mask[op])]; |
681 | addl $3, out_r | 681 | * } |
682 | rep movsb | 682 | */ |
683 | 683 | testb $64, %al | |
684 | movl in(%esp), in_r /* move in back to %esi, toss from */ | 684 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ |
685 | #if defined( USE_MMX ) | 685 | |
686 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | 686 | movl $1, %eax |
687 | #endif | 687 | shll %cl, %eax |
688 | jmp .L_while_test | 688 | decl %eax |
689 | 689 | andl hold_r, %eax /* eax &= hold */ | |
690 | .align 16,0x90 | 690 | addl %edx, %eax /* eax += this.val */ |
691 | 691 | movl lcode(%esp), %edx /* edx = lcode */ | |
692 | #if ! defined( USE_MMX ) | 692 | movl (%edx,%eax,4), %eax /* eax = lcode[val + (hold&mask[op])] */ |
693 | 693 | jmp .L_dolen | |
694 | .L_test_for_second_level_length: | 694 | |
695 | /* else if ((op & 64) == 0) { | 695 | .align 16,0x90 |
696 | * this = lcode[this.val + (hold & mask[op])]; | 696 | .L_test_for_second_level_dist: |
697 | * } | 697 | /* else if ((op & 64) == 0) { |
698 | */ | 698 | * this = dcode[this.val + (hold & mask[op])]; |
699 | testb $64, %al | 699 | * } |
700 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ | 700 | */ |
701 | 701 | testb $64, %al | |
702 | movl $1, %eax | 702 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ |
703 | shll %cl, %eax | 703 | |
704 | decl %eax | 704 | movl $1, %eax |
705 | andl hold_r, %eax /* eax &= hold */ | 705 | shll %cl, %eax |
706 | addl %edx, %eax /* eax += this.val */ | 706 | decl %eax |
707 | movl lcode(%esp), %edx /* edx = lcode */ | 707 | andl hold_r, %eax /* eax &= hold */ |
708 | movl (%edx,%eax,4), %eax /* eax = lcode[val + (hold&mask[op])] */ | 708 | addl %edx, %eax /* eax += this.val */ |
709 | jmp .L_dolen | 709 | movl dcode(%esp), %edx /* edx = dcode */ |
710 | 710 | movl (%edx,%eax,4), %eax /* eax = dcode[val + (hold&mask[op])] */ | |
711 | #else /* USE_MMX */ | 711 | jmp .L_dodist |
712 | 712 | ||
713 | .L_test_for_second_level_length: | 713 | .align 16,0x90 |
714 | testb $64, %al | 714 | .L_clip_window: |
715 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ | 715 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
716 | 716 | * %ecx = nbytes | |
717 | andl $15, %eax | 717 | * |
718 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 718 | * else { |
719 | movd hold_mm, %ecx | 719 | * if (dist > wsize) { |
720 | andl .L_mask(,%eax,4), %ecx | 720 | * invalid distance |
721 | addl len_r, %ecx | 721 | * } |
722 | movl (%ebx,%ecx,4), %eax /* eax = lcode[hold & lmask] */ | 722 | * from = window; |
723 | jmp .L_dolen | 723 | * nbytes = dist - nbytes; |
724 | 724 | * if (write == 0) { | |
725 | #endif | 725 | * from += wsize - nbytes; |
726 | 726 | */ | |
727 | .align 16,0x90 | 727 | #define nbytes_r %ecx |
728 | 728 | movl %eax, nbytes_r | |
729 | #if ! defined( USE_MMX ) | 729 | movl wsize(%esp), %eax /* prepare for dist compare */ |
730 | 730 | negl nbytes_r /* nbytes = -nbytes */ | |
731 | .L_test_for_second_level_dist: | 731 | movl window(%esp), from_r /* from = window */ |
732 | /* else if ((op & 64) == 0) { | 732 | |
733 | * this = dcode[this.val + (hold & mask[op])]; | 733 | cmpl dist_r, %eax |
734 | * } | 734 | jb .L_invalid_distance_too_far /* if (dist > wsize) */ |
735 | */ | 735 | |
736 | testb $64, %al | 736 | addl dist_r, nbytes_r /* nbytes = dist - nbytes */ |
737 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ | 737 | cmpl $0, write(%esp) |
738 | 738 | jne .L_wrap_around_window /* if (write != 0) */ | |
739 | movl $1, %eax | 739 | |
740 | shll %cl, %eax | 740 | subl nbytes_r, %eax |
741 | decl %eax | 741 | addl %eax, from_r /* from += wsize - nbytes */ |
742 | andl hold_r, %eax /* eax &= hold */ | 742 | |
743 | addl %edx, %eax /* eax += this.val */ | 743 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
744 | movl dcode(%esp), %edx /* edx = dcode */ | 744 | * %ecx = nbytes, %eax = len |
745 | movl (%edx,%eax,4), %eax /* eax = dcode[val + (hold&mask[op])] */ | 745 | * |
746 | jmp .L_dodist | 746 | * if (nbytes < len) { |
747 | 747 | * len -= nbytes; | |
748 | #else /* USE_MMX */ | 748 | * do { |
749 | 749 | * PUP(out) = PUP(from); | |
750 | .L_test_for_second_level_dist: | 750 | * } while (--nbytes); |
751 | testb $64, %al | 751 | * from = out - dist; |
752 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ | 752 | * } |
753 | 753 | * } | |
754 | andl $15, %eax | 754 | */ |
755 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 755 | #define len_r %eax |
756 | movd hold_mm, %ecx | 756 | movl len(%esp), len_r |
757 | andl .L_mask(,%eax,4), %ecx | 757 | cmpl nbytes_r, len_r |
758 | movl dcode(%esp), %eax /* ecx = dcode */ | 758 | jbe .L_do_copy1 /* if (nbytes >= len) */ |
759 | addl dist_r, %ecx | 759 | |
760 | movl (%eax,%ecx,4), %eax /* eax = lcode[hold & lmask] */ | 760 | subl nbytes_r, len_r /* len -= nbytes */ |
761 | jmp .L_dodist | 761 | rep movsb |
762 | 762 | movl out_r, from_r | |
763 | #endif | 763 | subl dist_r, from_r /* from = out - dist */ |
764 | 764 | jmp .L_do_copy1 | |
765 | .align 16,0x90 | 765 | |
766 | .L_clip_window: | 766 | cmpl nbytes_r, len_r |
767 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 767 | jbe .L_do_copy1 /* if (nbytes >= len) */ |
768 | * %ecx = nbytes | 768 | |
769 | * | 769 | subl nbytes_r, len_r /* len -= nbytes */ |
770 | * else { | 770 | rep movsb |
771 | * if (dist > wsize) { | 771 | movl out_r, from_r |
772 | * invalid distance | 772 | subl dist_r, from_r /* from = out - dist */ |
773 | * } | 773 | jmp .L_do_copy1 |
774 | * from = window; | 774 | |
775 | * nbytes = dist - nbytes; | 775 | .L_wrap_around_window: |
776 | * if (write == 0) { | 776 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
777 | * from += wsize - nbytes; | 777 | * %ecx = nbytes, %eax = write, %eax = len |
778 | */ | 778 | * |
779 | #define nbytes_r %ecx | 779 | * else if (write < nbytes) { |
780 | 780 | * from += wsize + write - nbytes; | |
781 | movl %eax, nbytes_r | 781 | * nbytes -= write; |
782 | movl wsize(%esp), %eax /* prepare for dist compare */ | 782 | * if (nbytes < len) { |
783 | negl nbytes_r /* nbytes = -nbytes */ | 783 | * len -= nbytes; |
784 | movl window(%esp), from_r /* from = window */ | 784 | * do { |
785 | 785 | * PUP(out) = PUP(from); | |
786 | cmpl dist_r, %eax | 786 | * } while (--nbytes); |
787 | jb .L_invalid_distance_too_far /* if (dist > wsize) */ | 787 | * from = window; |
788 | 788 | * nbytes = write; | |
789 | addl dist_r, nbytes_r /* nbytes = dist - nbytes */ | 789 | * if (nbytes < len) { |
790 | cmpl $0, write(%esp) | 790 | * len -= nbytes; |
791 | jne .L_wrap_around_window /* if (write != 0) */ | 791 | * do { |
792 | 792 | * PUP(out) = PUP(from); | |
793 | subl nbytes_r, %eax | 793 | * } while(--nbytes); |
794 | addl %eax, from_r /* from += wsize - nbytes */ | 794 | * from = out - dist; |
795 | 795 | * } | |
796 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 796 | * } |
797 | * %ecx = nbytes, %eax = len | 797 | * } |
798 | * | 798 | */ |
799 | * if (nbytes < len) { | 799 | #define write_r %eax |
800 | * len -= nbytes; | 800 | movl write(%esp), write_r |
801 | * do { | 801 | cmpl write_r, nbytes_r |
802 | * PUP(out) = PUP(from); | 802 | jbe .L_contiguous_in_window /* if (write >= nbytes) */ |
803 | * } while (--nbytes); | 803 | |
804 | * from = out - dist; | 804 | addl wsize(%esp), from_r |
805 | * } | 805 | addl write_r, from_r |
806 | * } | 806 | subl nbytes_r, from_r /* from += wsize + write - nbytes */ |
807 | */ | 807 | subl write_r, nbytes_r /* nbytes -= write */ |
808 | 808 | #undef write_r | |
809 | #if ! defined( USE_MMX ) | 809 | |
810 | #define len_r %eax | 810 | movl len(%esp), len_r |
811 | movl len(%esp), len_r | 811 | cmpl nbytes_r, len_r |
812 | #endif | 812 | jbe .L_do_copy1 /* if (nbytes >= len) */ |
813 | cmpl nbytes_r, len_r | 813 | |
814 | jbe .L_do_copy1 /* if (nbytes >= len) */ | 814 | subl nbytes_r, len_r /* len -= nbytes */ |
815 | 815 | rep movsb | |
816 | subl nbytes_r, len_r /* len -= nbytes */ | 816 | movl window(%esp), from_r /* from = window */ |
817 | rep movsb | 817 | movl write(%esp), nbytes_r /* nbytes = write */ |
818 | movl out_r, from_r | 818 | cmpl nbytes_r, len_r |
819 | subl dist_r, from_r /* from = out - dist */ | 819 | jbe .L_do_copy1 /* if (nbytes >= len) */ |
820 | jmp .L_do_copy1 | 820 | |
821 | 821 | subl nbytes_r, len_r /* len -= nbytes */ | |
822 | cmpl nbytes_r, len_r | 822 | rep movsb |
823 | jbe .L_do_copy1 /* if (nbytes >= len) */ | 823 | movl out_r, from_r |
824 | 824 | subl dist_r, from_r /* from = out - dist */ | |
825 | subl nbytes_r, len_r /* len -= nbytes */ | 825 | jmp .L_do_copy1 |
826 | rep movsb | 826 | |
827 | movl out_r, from_r | 827 | .L_contiguous_in_window: |
828 | subl dist_r, from_r /* from = out - dist */ | 828 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist |
829 | jmp .L_do_copy1 | 829 | * %ecx = nbytes, %eax = write, %eax = len |
830 | 830 | * | |
831 | .L_wrap_around_window: | 831 | * else { |
832 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 832 | * from += write - nbytes; |
833 | * %ecx = nbytes, %eax = write, %eax = len | 833 | * if (nbytes < len) { |
834 | * | 834 | * len -= nbytes; |
835 | * else if (write < nbytes) { | 835 | * do { |
836 | * from += wsize + write - nbytes; | 836 | * PUP(out) = PUP(from); |
837 | * nbytes -= write; | 837 | * } while (--nbytes); |
838 | * if (nbytes < len) { | 838 | * from = out - dist; |
839 | * len -= nbytes; | 839 | * } |
840 | * do { | 840 | * } |
841 | * PUP(out) = PUP(from); | 841 | */ |
842 | * } while (--nbytes); | 842 | #define write_r %eax |
843 | * from = window; | 843 | addl write_r, from_r |
844 | * nbytes = write; | 844 | subl nbytes_r, from_r /* from += write - nbytes */ |
845 | * if (nbytes < len) { | 845 | #undef write_r |
846 | * len -= nbytes; | 846 | |
847 | * do { | 847 | movl len(%esp), len_r |
848 | * PUP(out) = PUP(from); | 848 | cmpl nbytes_r, len_r |
849 | * } while(--nbytes); | 849 | jbe .L_do_copy1 /* if (nbytes >= len) */ |
850 | * from = out - dist; | 850 | |
851 | * } | 851 | subl nbytes_r, len_r /* len -= nbytes */ |
852 | * } | 852 | rep movsb |
853 | * } | 853 | movl out_r, from_r |
854 | */ | 854 | subl dist_r, from_r /* from = out - dist */ |
855 | #define write_r %eax | 855 | |
856 | 856 | .L_do_copy1: | |
857 | movl write(%esp), write_r | 857 | /* regs: %esi = from, %esi = in, %ebp = hold, %bl = bits, %edi = out |
858 | cmpl write_r, nbytes_r | 858 | * %eax = len |
859 | jbe .L_contiguous_in_window /* if (write >= nbytes) */ | 859 | * |
860 | 860 | * while (len > 0) { | |
861 | addl wsize(%esp), from_r | 861 | * PUP(out) = PUP(from); |
862 | addl write_r, from_r | 862 | * len--; |
863 | subl nbytes_r, from_r /* from += wsize + write - nbytes */ | 863 | * } |
864 | subl write_r, nbytes_r /* nbytes -= write */ | 864 | * } |
865 | #undef write_r | 865 | * } while (in < last && out < end); |
866 | 866 | */ | |
867 | #if ! defined( USE_MMX ) | 867 | #undef nbytes_r |
868 | movl len(%esp), len_r | 868 | #define in_r %esi |
869 | #endif | 869 | movl len_r, %ecx |
870 | cmpl nbytes_r, len_r | 870 | rep movsb |
871 | jbe .L_do_copy1 /* if (nbytes >= len) */ | 871 | |
872 | 872 | movl in(%esp), in_r /* move in back to %esi, toss from */ | |
873 | subl nbytes_r, len_r /* len -= nbytes */ | 873 | jmp .L_while_test |
874 | rep movsb | 874 | |
875 | movl window(%esp), from_r /* from = window */ | 875 | #undef len_r |
876 | movl write(%esp), nbytes_r /* nbytes = write */ | 876 | #undef dist_r |
877 | cmpl nbytes_r, len_r | 877 | |
878 | jbe .L_do_copy1 /* if (nbytes >= len) */ | 878 | #endif /* NO_MMX || RUN_TIME_MMX */ |
879 | 879 | ||
880 | subl nbytes_r, len_r /* len -= nbytes */ | 880 | |
881 | rep movsb | 881 | /*** MMX code ***/ |
882 | movl out_r, from_r | 882 | |
883 | subl dist_r, from_r /* from = out - dist */ | 883 | #if defined( USE_MMX ) || defined( RUN_TIME_MMX ) |
884 | jmp .L_do_copy1 | 884 | |
885 | 885 | .align 32,0x90 | |
886 | .L_contiguous_in_window: | 886 | .L_init_mmx: |
887 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | 887 | emms |
888 | * %ecx = nbytes, %eax = write, %eax = len | 888 | |
889 | * | 889 | #undef bits_r |
890 | * else { | 890 | #undef bitslong_r |
891 | * from += write - nbytes; | 891 | #define bitslong_r %ebp |
892 | * if (nbytes < len) { | 892 | #define hold_mm %mm0 |
893 | * len -= nbytes; | 893 | movd %ebp, hold_mm |
894 | * do { | 894 | movl %ebx, bitslong_r |
895 | * PUP(out) = PUP(from); | 895 | |
896 | * } while (--nbytes); | 896 | #define used_mm %mm1 |
897 | * from = out - dist; | 897 | #define dmask2_mm %mm2 |
898 | * } | 898 | #define lmask2_mm %mm3 |
899 | * } | 899 | #define lmask_mm %mm4 |
900 | */ | 900 | #define dmask_mm %mm5 |
901 | #define write_r %eax | 901 | #define tmp_mm %mm6 |
902 | 902 | ||
903 | addl write_r, from_r | 903 | movd lmask(%esp), lmask_mm |
904 | subl nbytes_r, from_r /* from += write - nbytes */ | 904 | movq lmask_mm, lmask2_mm |
905 | #undef write_r | 905 | movd dmask(%esp), dmask_mm |
906 | 906 | movq dmask_mm, dmask2_mm | |
907 | #if ! defined( USE_MMX ) | 907 | pxor used_mm, used_mm |
908 | movl len(%esp), len_r | 908 | movl lcode(%esp), %ebx /* ebx = lcode */ |
909 | #endif | 909 | jmp .L_do_loop_mmx |
910 | cmpl nbytes_r, len_r | 910 | |
911 | jbe .L_do_copy1 /* if (nbytes >= len) */ | 911 | .align 32,0x90 |
912 | 912 | .L_while_test_mmx: | |
913 | subl nbytes_r, len_r /* len -= nbytes */ | 913 | /* while (in < last && out < end) |
914 | rep movsb | 914 | */ |
915 | movl out_r, from_r | 915 | cmpl out_r, end(%esp) |
916 | subl dist_r, from_r /* from = out - dist */ | 916 | jbe .L_break_loop /* if (out >= end) */ |
917 | 917 | ||
918 | .L_do_copy1: | 918 | cmpl in_r, last(%esp) |
919 | /* regs: %esi = from, %esi = in, %ebp = hold, %bl = bits, %edi = out | 919 | jbe .L_break_loop |
920 | * %eax = len | 920 | |
921 | * | 921 | .L_do_loop_mmx: |
922 | * while (len > 0) { | 922 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
923 | * PUP(out) = PUP(from); | 923 | |
924 | * len--; | 924 | cmpl $32, bitslong_r |
925 | * } | 925 | ja .L_get_length_code_mmx /* if (32 < bits) */ |
926 | * } | 926 | |
927 | * } while (in < last && out < end); | 927 | movd bitslong_r, tmp_mm |
928 | */ | 928 | movd (in_r), %mm7 |
929 | #undef nbytes_r | 929 | addl $4, in_r |
930 | #define in_r %esi | 930 | psllq tmp_mm, %mm7 |
931 | 931 | addl $32, bitslong_r | |
932 | movl len_r, %ecx | 932 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ |
933 | rep movsb | 933 | |
934 | 934 | .L_get_length_code_mmx: | |
935 | movl in(%esp), in_r /* move in back to %esi, toss from */ | 935 | pand hold_mm, lmask_mm |
936 | #if defined( USE_MMX ) | 936 | movd lmask_mm, %eax |
937 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | 937 | movq lmask2_mm, lmask_mm |
938 | #endif | 938 | movl (%ebx,%eax,4), %eax /* eax = lcode[hold & lmask] */ |
939 | jmp .L_while_test | 939 | |
940 | 940 | .L_dolen_mmx: | |
941 | #undef len_r | 941 | movzbl %ah, %ecx /* ecx = this.bits */ |
942 | #undef from_r | 942 | movd %ecx, used_mm |
943 | #undef dist_r | 943 | subl %ecx, bitslong_r /* bits -= this.bits */ |
944 | 944 | ||
945 | .L_invalid_distance_code: | 945 | testb %al, %al |
946 | /* else { | 946 | jnz .L_test_for_length_base_mmx /* if (op != 0) 45.7% */ |
947 | * strm->msg = "invalid distance code"; | 947 | |
948 | * state->mode = BAD; | 948 | shrl $16, %eax /* output this.val char */ |
949 | * } | 949 | stosb |
950 | */ | 950 | jmp .L_while_test_mmx |
951 | movl $.L_invalid_distance_code_msg, %ecx | 951 | |
952 | movl $INFLATE_MODE_BAD, %edx | 952 | .L_test_for_length_base_mmx: |
953 | jmp .L_update_stream_state | 953 | #define len_r %edx |
954 | 954 | movl %eax, len_r /* len = this */ | |
955 | .L_test_for_end_of_block: | 955 | shrl $16, len_r /* len = this.val */ |
956 | /* else if (op & 32) { | 956 | |
957 | * state->mode = TYPE; | 957 | testb $16, %al |
958 | * break; | 958 | jz .L_test_for_second_level_length_mmx /* if ((op & 16) == 0) 8% */ |
959 | * } | 959 | andl $15, %eax /* op &= 15 */ |
960 | */ | 960 | jz .L_decode_distance_mmx /* if (!op) */ |
961 | testb $32, %al | 961 | |
962 | jz .L_invalid_literal_length_code /* if ((op & 32) == 0) */ | 962 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
963 | 963 | movd %eax, used_mm | |
964 | movl $0, %ecx | 964 | movd hold_mm, %ecx |
965 | movl $INFLATE_MODE_TYPE, %edx | 965 | subl %eax, bitslong_r |
966 | jmp .L_update_stream_state | 966 | andl .L_mask(,%eax,4), %ecx |
967 | 967 | addl %ecx, len_r /* len += hold & mask[op] */ | |
968 | .L_invalid_literal_length_code: | 968 | |
969 | /* else { | 969 | .L_decode_distance_mmx: |
970 | * strm->msg = "invalid literal/length code"; | 970 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
971 | * state->mode = BAD; | 971 | |
972 | * } | 972 | cmpl $32, bitslong_r |
973 | */ | 973 | ja .L_get_dist_code_mmx /* if (32 < bits) */ |
974 | movl $.L_invalid_literal_length_code_msg, %ecx | 974 | |
975 | movl $INFLATE_MODE_BAD, %edx | 975 | movd bitslong_r, tmp_mm |
976 | jmp .L_update_stream_state | 976 | movd (in_r), %mm7 |
977 | 977 | addl $4, in_r | |
978 | .L_invalid_distance_too_far: | 978 | psllq tmp_mm, %mm7 |
979 | /* strm->msg = "invalid distance too far back"; | 979 | addl $32, bitslong_r |
980 | * state->mode = BAD; | 980 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ |
981 | */ | 981 | |
982 | movl in(%esp), in_r /* from_r has in's reg, put in back */ | 982 | .L_get_dist_code_mmx: |
983 | movl $.L_invalid_distance_too_far_msg, %ecx | 983 | movl dcode(%esp), %ebx /* ebx = dcode */ |
984 | movl $INFLATE_MODE_BAD, %edx | 984 | pand hold_mm, dmask_mm |
985 | jmp .L_update_stream_state | 985 | movd dmask_mm, %eax |
986 | 986 | movq dmask2_mm, dmask_mm | |
987 | .L_update_stream_state: | 987 | movl (%ebx,%eax,4), %eax /* eax = dcode[hold & lmask] */ |
988 | /* set strm->msg = %ecx, strm->state->mode = %edx */ | 988 | |
989 | movl strm_sp(%esp), %eax | 989 | .L_dodist_mmx: |
990 | testl %ecx, %ecx /* if (msg != NULL) */ | 990 | #define dist_r %ebx |
991 | jz .L_skip_msg | 991 | movzbl %ah, %ecx /* ecx = this.bits */ |
992 | movl %ecx, msg_strm(%eax) /* strm->msg = msg */ | 992 | movl %eax, dist_r |
993 | .L_skip_msg: | 993 | shrl $16, dist_r /* dist = this.val */ |
994 | movl state_strm(%eax), %eax /* state = strm->state */ | 994 | subl %ecx, bitslong_r /* bits -= this.bits */ |
995 | movl %edx, mode_state(%eax) /* state->mode = edx (BAD | TYPE) */ | 995 | movd %ecx, used_mm |
996 | 996 | ||
997 | .L_break_loop: | 997 | testb $16, %al /* if ((op & 16) == 0) */ |
998 | 998 | jz .L_test_for_second_level_dist_mmx | |
999 | #define strm_r %eax | 999 | andl $15, %eax /* op &= 15 */ |
1000 | #define state_r %edx | 1000 | jz .L_check_dist_one_mmx |
1001 | 1001 | ||
1002 | /* len = bits >> 3; | 1002 | .L_add_bits_to_dist_mmx: |
1003 | * in -= len; | 1003 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
1004 | * bits -= len << 3; | 1004 | movd %eax, used_mm /* save bit length of current op */ |
1005 | * hold &= (1U << bits) - 1; | 1005 | movd hold_mm, %ecx /* get the next bits on input stream */ |
1006 | * state->hold = hold; | 1006 | subl %eax, bitslong_r /* bits -= op bits */ |
1007 | * state->bits = bits; | 1007 | andl .L_mask(,%eax,4), %ecx /* ecx = hold & mask[op] */ |
1008 | * strm->next_in = in; | 1008 | addl %ecx, dist_r /* dist += hold & mask[op] */ |
1009 | * strm->next_out = out; | 1009 | |
1010 | */ | 1010 | .L_check_window_mmx: |
1011 | movl strm_sp(%esp), strm_r | 1011 | movl in_r, in(%esp) /* save in so from can use it's reg */ |
1012 | movl bitslong_r, %ecx | 1012 | movl out_r, %eax |
1013 | movl state_strm(strm_r), state_r | 1013 | subl beg(%esp), %eax /* nbytes = out - beg */ |
1014 | shrl $3, %ecx | 1014 | |
1015 | subl %ecx, in_r | 1015 | cmpl dist_r, %eax |
1016 | shll $3, %ecx | 1016 | jb .L_clip_window_mmx /* if (dist > nbytes) 4.2% */ |
1017 | subl %ecx, bitslong_r | 1017 | |
1018 | movl out_r, next_out_strm(strm_r) | 1018 | movl len_r, %ecx |
1019 | movl in_r, next_in_strm(strm_r) | 1019 | movl out_r, from_r |
1020 | movl bitslong_r, bits_state(state_r) | 1020 | subl dist_r, from_r /* from = out - dist */ |
1021 | 1021 | ||
1022 | movl bitslong_r, %ecx | 1022 | subl $3, %ecx |
1023 | movl $1, %ebx /* overwrites bitslong_r, %bl */ | 1023 | movb (from_r), %al |
1024 | shll %cl, %ebx | 1024 | movb %al, (out_r) |
1025 | decl %ebx | 1025 | movb 1(from_r), %al |
1026 | 1026 | movb 2(from_r), %dl | |
1027 | #undef bits_r | 1027 | addl $3, from_r |
1028 | #undef bitslong_r | 1028 | movb %al, 1(out_r) |
1029 | 1029 | movb %dl, 2(out_r) | |
1030 | #if ! defined( USE_MMX ) | 1030 | addl $3, out_r |
1031 | 1031 | rep movsb | |
1032 | andl %ebx, hold_r | 1032 | |
1033 | movl hold_r, hold_state(state_r) | 1033 | movl in(%esp), in_r /* move in back to %esi, toss from */ |
1034 | 1034 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | |
1035 | #else /* USE_MMX */ | 1035 | jmp .L_while_test_mmx |
1036 | 1036 | ||
1037 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | 1037 | .align 16,0x90 |
1038 | movd hold_mm, %ecx | 1038 | .L_check_dist_one_mmx: |
1039 | andl %ebx, %ecx | 1039 | cmpl $1, dist_r |
1040 | movl %ecx, hold_state(state_r) | 1040 | jne .L_check_window_mmx |
1041 | 1041 | cmpl out_r, beg(%esp) | |
1042 | #endif | 1042 | je .L_check_window_mmx |
1043 | 1043 | ||
1044 | #define last_r %ebx | 1044 | decl out_r |
1045 | 1045 | movl len_r, %ecx | |
1046 | /* strm->avail_in = in < last ? 5 + (last - in) : 5 - (in - last) */ | 1046 | movb (out_r), %al |
1047 | movl last(%esp), last_r | 1047 | subl $3, %ecx |
1048 | cmpl in_r, last_r | 1048 | |
1049 | jbe .L_last_is_smaller /* if (in >= last) */ | 1049 | movb %al, 1(out_r) |
1050 | 1050 | movb %al, 2(out_r) | |
1051 | subl in_r, last_r /* last -= in */ | 1051 | movb %al, 3(out_r) |
1052 | addl $5, last_r /* last += 5 */ | 1052 | addl $4, out_r |
1053 | movl last_r, avail_in_strm(strm_r) | 1053 | rep stosb |
1054 | jmp .L_fixup_out | 1054 | |
1055 | .L_last_is_smaller: | 1055 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ |
1056 | subl last_r, in_r /* in -= last */ | 1056 | jmp .L_while_test_mmx |
1057 | negl in_r /* in = -in */ | 1057 | |
1058 | addl $5, in_r /* in += 5 */ | 1058 | .align 16,0x90 |
1059 | movl in_r, avail_in_strm(strm_r) | 1059 | .L_test_for_second_level_length_mmx: |
1060 | 1060 | testb $64, %al | |
1061 | #undef last_r | 1061 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ |
1062 | #define end_r %ebx | 1062 | |
1063 | 1063 | andl $15, %eax | |
1064 | .L_fixup_out: | 1064 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
1065 | /* strm->avail_out = out < end ? 257 + (end - out) : 257 - (out - end)*/ | 1065 | movd hold_mm, %ecx |
1066 | movl end(%esp), end_r | 1066 | andl .L_mask(,%eax,4), %ecx |
1067 | cmpl out_r, end_r | 1067 | addl len_r, %ecx |
1068 | jbe .L_end_is_smaller /* if (out >= end) */ | 1068 | movl (%ebx,%ecx,4), %eax /* eax = lcode[hold & lmask] */ |
1069 | 1069 | jmp .L_dolen_mmx | |
1070 | subl out_r, end_r /* end -= out */ | 1070 | |
1071 | addl $257, end_r /* end += 257 */ | 1071 | .align 16,0x90 |
1072 | movl end_r, avail_out_strm(strm_r) | 1072 | .L_test_for_second_level_dist_mmx: |
1073 | jmp .L_done | 1073 | testb $64, %al |
1074 | .L_end_is_smaller: | 1074 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ |
1075 | subl end_r, out_r /* out -= end */ | 1075 | |
1076 | negl out_r /* out = -out */ | 1076 | andl $15, %eax |
1077 | addl $257, out_r /* out += 257 */ | 1077 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ |
1078 | movl out_r, avail_out_strm(strm_r) | 1078 | movd hold_mm, %ecx |
1079 | 1079 | andl .L_mask(,%eax,4), %ecx | |
1080 | #undef end_r | 1080 | movl dcode(%esp), %eax /* ecx = dcode */ |
1081 | 1081 | addl dist_r, %ecx | |
1082 | .L_done: | 1082 | movl (%eax,%ecx,4), %eax /* eax = lcode[hold & lmask] */ |
1083 | #if defined( USE_MMX ) | 1083 | jmp .L_dodist_mmx |
1084 | emms | 1084 | |
1085 | #endif | 1085 | .align 16,0x90 |
1086 | addl $local_var_size, %esp | 1086 | .L_clip_window_mmx: |
1087 | popf | 1087 | #define nbytes_r %ecx |
1088 | popl %ebx | 1088 | movl %eax, nbytes_r |
1089 | popl %ebp | 1089 | movl wsize(%esp), %eax /* prepare for dist compare */ |
1090 | popl %esi | 1090 | negl nbytes_r /* nbytes = -nbytes */ |
1091 | popl %edi | 1091 | movl window(%esp), from_r /* from = window */ |
1092 | ret | 1092 | |
1093 | 1093 | cmpl dist_r, %eax | |
1094 | .type inflate_fast,@function | 1094 | jb .L_invalid_distance_too_far /* if (dist > wsize) */ |
1095 | .size inflate_fast,.-inflate_fast | 1095 | |
1096 | addl dist_r, nbytes_r /* nbytes = dist - nbytes */ | ||
1097 | cmpl $0, write(%esp) | ||
1098 | jne .L_wrap_around_window_mmx /* if (write != 0) */ | ||
1099 | |||
1100 | subl nbytes_r, %eax | ||
1101 | addl %eax, from_r /* from += wsize - nbytes */ | ||
1102 | |||
1103 | cmpl nbytes_r, len_r | ||
1104 | jbe .L_do_copy1_mmx /* if (nbytes >= len) */ | ||
1105 | |||
1106 | subl nbytes_r, len_r /* len -= nbytes */ | ||
1107 | rep movsb | ||
1108 | movl out_r, from_r | ||
1109 | subl dist_r, from_r /* from = out - dist */ | ||
1110 | jmp .L_do_copy1_mmx | ||
1111 | |||
1112 | cmpl nbytes_r, len_r | ||
1113 | jbe .L_do_copy1_mmx /* if (nbytes >= len) */ | ||
1114 | |||
1115 | subl nbytes_r, len_r /* len -= nbytes */ | ||
1116 | rep movsb | ||
1117 | movl out_r, from_r | ||
1118 | subl dist_r, from_r /* from = out - dist */ | ||
1119 | jmp .L_do_copy1_mmx | ||
1120 | |||
1121 | .L_wrap_around_window_mmx: | ||
1122 | #define write_r %eax | ||
1123 | movl write(%esp), write_r | ||
1124 | cmpl write_r, nbytes_r | ||
1125 | jbe .L_contiguous_in_window_mmx /* if (write >= nbytes) */ | ||
1126 | |||
1127 | addl wsize(%esp), from_r | ||
1128 | addl write_r, from_r | ||
1129 | subl nbytes_r, from_r /* from += wsize + write - nbytes */ | ||
1130 | subl write_r, nbytes_r /* nbytes -= write */ | ||
1131 | #undef write_r | ||
1132 | |||
1133 | cmpl nbytes_r, len_r | ||
1134 | jbe .L_do_copy1_mmx /* if (nbytes >= len) */ | ||
1135 | |||
1136 | subl nbytes_r, len_r /* len -= nbytes */ | ||
1137 | rep movsb | ||
1138 | movl window(%esp), from_r /* from = window */ | ||
1139 | movl write(%esp), nbytes_r /* nbytes = write */ | ||
1140 | cmpl nbytes_r, len_r | ||
1141 | jbe .L_do_copy1_mmx /* if (nbytes >= len) */ | ||
1142 | |||
1143 | subl nbytes_r, len_r /* len -= nbytes */ | ||
1144 | rep movsb | ||
1145 | movl out_r, from_r | ||
1146 | subl dist_r, from_r /* from = out - dist */ | ||
1147 | jmp .L_do_copy1_mmx | ||
1148 | |||
1149 | .L_contiguous_in_window_mmx: | ||
1150 | #define write_r %eax | ||
1151 | addl write_r, from_r | ||
1152 | subl nbytes_r, from_r /* from += write - nbytes */ | ||
1153 | #undef write_r | ||
1154 | |||
1155 | cmpl nbytes_r, len_r | ||
1156 | jbe .L_do_copy1_mmx /* if (nbytes >= len) */ | ||
1157 | |||
1158 | subl nbytes_r, len_r /* len -= nbytes */ | ||
1159 | rep movsb | ||
1160 | movl out_r, from_r | ||
1161 | subl dist_r, from_r /* from = out - dist */ | ||
1162 | |||
1163 | .L_do_copy1_mmx: | ||
1164 | #undef nbytes_r | ||
1165 | #define in_r %esi | ||
1166 | movl len_r, %ecx | ||
1167 | rep movsb | ||
1168 | |||
1169 | movl in(%esp), in_r /* move in back to %esi, toss from */ | ||
1170 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | ||
1171 | jmp .L_while_test_mmx | ||
1172 | |||
1173 | #undef hold_r | ||
1174 | #undef bitslong_r | ||
1175 | |||
1176 | #endif /* USE_MMX || RUN_TIME_MMX */ | ||
1177 | |||
1178 | |||
1179 | /*** USE_MMX, NO_MMX, and RUNTIME_MMX from here on ***/ | ||
1180 | |||
1181 | .L_invalid_distance_code: | ||
1182 | /* else { | ||
1183 | * strm->msg = "invalid distance code"; | ||
1184 | * state->mode = BAD; | ||
1185 | * } | ||
1186 | */ | ||
1187 | movl $.L_invalid_distance_code_msg, %ecx | ||
1188 | movl $INFLATE_MODE_BAD, %edx | ||
1189 | jmp .L_update_stream_state | ||
1190 | |||
1191 | .L_test_for_end_of_block: | ||
1192 | /* else if (op & 32) { | ||
1193 | * state->mode = TYPE; | ||
1194 | * break; | ||
1195 | * } | ||
1196 | */ | ||
1197 | testb $32, %al | ||
1198 | jz .L_invalid_literal_length_code /* if ((op & 32) == 0) */ | ||
1199 | |||
1200 | movl $0, %ecx | ||
1201 | movl $INFLATE_MODE_TYPE, %edx | ||
1202 | jmp .L_update_stream_state | ||
1203 | |||
1204 | .L_invalid_literal_length_code: | ||
1205 | /* else { | ||
1206 | * strm->msg = "invalid literal/length code"; | ||
1207 | * state->mode = BAD; | ||
1208 | * } | ||
1209 | */ | ||
1210 | movl $.L_invalid_literal_length_code_msg, %ecx | ||
1211 | movl $INFLATE_MODE_BAD, %edx | ||
1212 | jmp .L_update_stream_state | ||
1213 | |||
1214 | .L_invalid_distance_too_far: | ||
1215 | /* strm->msg = "invalid distance too far back"; | ||
1216 | * state->mode = BAD; | ||
1217 | */ | ||
1218 | movl in(%esp), in_r /* from_r has in's reg, put in back */ | ||
1219 | movl $.L_invalid_distance_too_far_msg, %ecx | ||
1220 | movl $INFLATE_MODE_BAD, %edx | ||
1221 | jmp .L_update_stream_state | ||
1222 | |||
1223 | .L_update_stream_state: | ||
1224 | /* set strm->msg = %ecx, strm->state->mode = %edx */ | ||
1225 | movl strm_sp(%esp), %eax | ||
1226 | testl %ecx, %ecx /* if (msg != NULL) */ | ||
1227 | jz .L_skip_msg | ||
1228 | movl %ecx, msg_strm(%eax) /* strm->msg = msg */ | ||
1229 | .L_skip_msg: | ||
1230 | movl state_strm(%eax), %eax /* state = strm->state */ | ||
1231 | movl %edx, mode_state(%eax) /* state->mode = edx (BAD | TYPE) */ | ||
1232 | jmp .L_break_loop | ||
1233 | |||
1234 | .align 32,0x90 | ||
1235 | .L_break_loop: | ||
1236 | |||
1237 | /* | ||
1238 | * Regs: | ||
1239 | * | ||
1240 | * bits = %ebp when mmx, and in %ebx when non-mmx | ||
1241 | * hold = %hold_mm when mmx, and in %ebp when non-mmx | ||
1242 | * in = %esi | ||
1243 | * out = %edi | ||
1244 | */ | ||
1245 | |||
1246 | #if defined( USE_MMX ) || defined( RUN_TIME_MMX ) | ||
1247 | |||
1248 | #if defined( RUN_TIME_MMX ) | ||
1249 | |||
1250 | cmpl $DO_USE_MMX, inflate_fast_use_mmx | ||
1251 | jne .L_update_next_in | ||
1252 | |||
1253 | #endif /* RUN_TIME_MMX */ | ||
1254 | |||
1255 | movl %ebp, %ebx | ||
1256 | |||
1257 | .L_update_next_in: | ||
1258 | |||
1259 | #endif | ||
1260 | |||
1261 | #define strm_r %eax | ||
1262 | #define state_r %edx | ||
1263 | |||
1264 | /* len = bits >> 3; | ||
1265 | * in -= len; | ||
1266 | * bits -= len << 3; | ||
1267 | * hold &= (1U << bits) - 1; | ||
1268 | * state->hold = hold; | ||
1269 | * state->bits = bits; | ||
1270 | * strm->next_in = in; | ||
1271 | * strm->next_out = out; | ||
1272 | */ | ||
1273 | movl strm_sp(%esp), strm_r | ||
1274 | movl %ebx, %ecx | ||
1275 | movl state_strm(strm_r), state_r | ||
1276 | shrl $3, %ecx | ||
1277 | subl %ecx, in_r | ||
1278 | shll $3, %ecx | ||
1279 | subl %ecx, %ebx | ||
1280 | movl out_r, next_out_strm(strm_r) | ||
1281 | movl %ebx, bits_state(state_r) | ||
1282 | movl %ebx, %ecx | ||
1283 | |||
1284 | leal buf(%esp), %ebx | ||
1285 | cmpl %ebx, last(%esp) | ||
1286 | jne .L_buf_not_used /* if buf != last */ | ||
1287 | |||
1288 | subl %ebx, in_r /* in -= buf */ | ||
1289 | movl next_in_strm(strm_r), %ebx | ||
1290 | movl %ebx, last(%esp) /* last = strm->next_in */ | ||
1291 | addl %ebx, in_r /* in += strm->next_in */ | ||
1292 | movl avail_in_strm(strm_r), %ebx | ||
1293 | subl $11, %ebx | ||
1294 | addl %ebx, last(%esp) /* last = &strm->next_in[ avail_in - 11 ] */ | ||
1295 | |||
1296 | .L_buf_not_used: | ||
1297 | movl in_r, next_in_strm(strm_r) | ||
1298 | |||
1299 | movl $1, %ebx | ||
1300 | shll %cl, %ebx | ||
1301 | decl %ebx | ||
1302 | |||
1303 | #if defined( USE_MMX ) || defined( RUN_TIME_MMX ) | ||
1304 | |||
1305 | #if defined( RUN_TIME_MMX ) | ||
1306 | |||
1307 | cmpl $DO_USE_MMX, inflate_fast_use_mmx | ||
1308 | jne .L_update_hold | ||
1309 | |||
1310 | #endif /* RUN_TIME_MMX */ | ||
1311 | |||
1312 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
1313 | movd hold_mm, %ebp | ||
1314 | |||
1315 | emms | ||
1316 | |||
1317 | .L_update_hold: | ||
1318 | |||
1319 | #endif /* USE_MMX || RUN_TIME_MMX */ | ||
1320 | |||
1321 | andl %ebx, %ebp | ||
1322 | movl %ebp, hold_state(state_r) | ||
1323 | |||
1324 | #define last_r %ebx | ||
1325 | |||
1326 | /* strm->avail_in = in < last ? 11 + (last - in) : 11 - (in - last) */ | ||
1327 | movl last(%esp), last_r | ||
1328 | cmpl in_r, last_r | ||
1329 | jbe .L_last_is_smaller /* if (in >= last) */ | ||
1330 | |||
1331 | subl in_r, last_r /* last -= in */ | ||
1332 | addl $11, last_r /* last += 11 */ | ||
1333 | movl last_r, avail_in_strm(strm_r) | ||
1334 | jmp .L_fixup_out | ||
1335 | .L_last_is_smaller: | ||
1336 | subl last_r, in_r /* in -= last */ | ||
1337 | negl in_r /* in = -in */ | ||
1338 | addl $11, in_r /* in += 11 */ | ||
1339 | movl in_r, avail_in_strm(strm_r) | ||
1340 | |||
1341 | #undef last_r | ||
1342 | #define end_r %ebx | ||
1343 | |||
1344 | .L_fixup_out: | ||
1345 | /* strm->avail_out = out < end ? 257 + (end - out) : 257 - (out - end)*/ | ||
1346 | movl end(%esp), end_r | ||
1347 | cmpl out_r, end_r | ||
1348 | jbe .L_end_is_smaller /* if (out >= end) */ | ||
1349 | |||
1350 | subl out_r, end_r /* end -= out */ | ||
1351 | addl $257, end_r /* end += 257 */ | ||
1352 | movl end_r, avail_out_strm(strm_r) | ||
1353 | jmp .L_done | ||
1354 | .L_end_is_smaller: | ||
1355 | subl end_r, out_r /* out -= end */ | ||
1356 | negl out_r /* out = -out */ | ||
1357 | addl $257, out_r /* out += 257 */ | ||
1358 | movl out_r, avail_out_strm(strm_r) | ||
1359 | |||
1360 | #undef end_r | ||
1361 | #undef strm_r | ||
1362 | #undef state_r | ||
1363 | |||
1364 | .L_done: | ||
1365 | addl $local_var_size, %esp | ||
1366 | popf | ||
1367 | popl %ebx | ||
1368 | popl %ebp | ||
1369 | popl %esi | ||
1370 | popl %edi | ||
1371 | ret | ||
1372 | |||
1373 | #if defined( GAS_ELF ) | ||
1374 | /* elf info */ | ||
1375 | .type inflate_fast,@function | ||
1376 | .size inflate_fast,.-inflate_fast | ||
1377 | #endif | ||