diff options
author | Mark Adler <madler@alumni.caltech.edu> | 2011-09-09 23:21:47 -0700 |
---|---|---|
committer | Mark Adler <madler@alumni.caltech.edu> | 2011-09-09 23:21:47 -0700 |
commit | 7c2a874e50b871d04fbd19501f7b42cff55e5abc (patch) | |
tree | 1879cd29182ababb17cde77cee5ce74505db4006 /contrib/inflate86 | |
parent | a383133c4e7b93113cee912f213cf9502d785fa7 (diff) | |
download | zlib-1.2.0.tar.gz zlib-1.2.0.tar.bz2 zlib-1.2.0.zip |
zlib 1.2.0v1.2.0
Diffstat (limited to 'contrib/inflate86')
-rw-r--r-- | contrib/inflate86/inffast.S | 1095 |
1 files changed, 1095 insertions, 0 deletions
diff --git a/contrib/inflate86/inffast.S b/contrib/inflate86/inffast.S new file mode 100644 index 0000000..d1e80ef --- /dev/null +++ b/contrib/inflate86/inffast.S | |||
@@ -0,0 +1,1095 @@ | |||
1 | /* | ||
2 | * inffast.S is a hand tuned assembler version of: | ||
3 | * | ||
4 | * inffast.c -- fast decoding | ||
5 | * Copyright (C) 1995-2003 Mark Adler | ||
6 | * For conditions of distribution and use, see copyright notice in zlib.h | ||
7 | * | ||
8 | * Copyright (C) 2003 Chris Anderson <christop@charm.net> | ||
9 | * Please use the copyright conditions above. | ||
10 | * | ||
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 | ||
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 | ||
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 | ||
17 | * stream processing I/O and crc32/addler32. In my case, this routine used | ||
18 | * 70% of the cpu time and crc32 used 20%. | ||
19 | * | ||
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. | ||
22 | * | ||
23 | * Jan-24-2003 -- Added -DUSE_MMX define for slightly faster inflating. | ||
24 | * It should be a runtime flag instead of compile time flag... | ||
25 | */ | ||
26 | |||
27 | .file "inffast.S" | ||
28 | |||
29 | .globl inflate_fast | ||
30 | |||
31 | .text | ||
32 | .align 4,0 | ||
33 | .L_invalid_literal_length_code_msg: | ||
34 | .string "invalid literal/length code" | ||
35 | |||
36 | .align 4,0 | ||
37 | .L_invalid_distance_code_msg: | ||
38 | .string "invalid distance code" | ||
39 | |||
40 | .align 4,0 | ||
41 | .L_invalid_distance_too_far_msg: | ||
42 | .string "invalid distance too far back" | ||
43 | |||
44 | #if defined( USE_MMX ) | ||
45 | .align 4,0 | ||
46 | .L_mask: /* mask[N] = ( 1 << N ) - 1 */ | ||
47 | .long 0 | ||
48 | .long 1 | ||
49 | .long 3 | ||
50 | .long 7 | ||
51 | .long 15 | ||
52 | .long 31 | ||
53 | .long 63 | ||
54 | .long 127 | ||
55 | .long 255 | ||
56 | .long 511 | ||
57 | .long 1023 | ||
58 | .long 2047 | ||
59 | .long 4095 | ||
60 | .long 8191 | ||
61 | .long 16383 | ||
62 | .long 32767 | ||
63 | .long 65535 | ||
64 | .long 131071 | ||
65 | .long 262143 | ||
66 | .long 524287 | ||
67 | .long 1048575 | ||
68 | .long 2097151 | ||
69 | .long 4194303 | ||
70 | .long 8388607 | ||
71 | .long 16777215 | ||
72 | .long 33554431 | ||
73 | .long 67108863 | ||
74 | .long 134217727 | ||
75 | .long 268435455 | ||
76 | .long 536870911 | ||
77 | .long 1073741823 | ||
78 | .long 2147483647 | ||
79 | .long 4294967295 | ||
80 | #endif | ||
81 | |||
82 | .text | ||
83 | |||
84 | /* | ||
85 | * struct z_stream offsets, in zlib.h | ||
86 | */ | ||
87 | #define next_in_strm 0 /* strm->next_in */ | ||
88 | #define avail_in_strm 4 /* strm->avail_in */ | ||
89 | #define next_out_strm 12 /* strm->next_out */ | ||
90 | #define avail_out_strm 16 /* strm->avail_out */ | ||
91 | #define msg_strm 24 /* strm->msg */ | ||
92 | #define state_strm 28 /* strm->state */ | ||
93 | |||
94 | /* | ||
95 | * struct inflate_state offsets, in inflate.h | ||
96 | */ | ||
97 | #define mode_state 0 /* state->mode */ | ||
98 | #define wsize_state 32 /* state->wsize */ | ||
99 | #define write_state 36 /* state->write */ | ||
100 | #define window_state 40 /* state->window */ | ||
101 | #define hold_state 44 /* state->hold */ | ||
102 | #define bits_state 48 /* state->bits */ | ||
103 | #define lencode_state 64 /* state->lencode */ | ||
104 | #define distcode_state 68 /* state->distcode */ | ||
105 | #define lenbits_state 72 /* state->lenbits */ | ||
106 | #define distbits_state 76 /* state->distbits */ | ||
107 | |||
108 | /* | ||
109 | * inflate_fast's activation record | ||
110 | */ | ||
111 | #define local_var_size 56 /* how much local space for vars */ | ||
112 | #define strm_sp 80 /* first arg: z_stream * (local_var_size + 24) */ | ||
113 | #define start_sp 84 /* second arg: unsigned int (local_var_size + 28) */ | ||
114 | |||
115 | /* | ||
116 | * offsets for local vars on stack | ||
117 | */ | ||
118 | #define out 52 /* unsigned char* */ | ||
119 | #define window 48 /* unsigned char* */ | ||
120 | #define wsize 44 /* unsigned int */ | ||
121 | #define write 40 /* unsigned int */ | ||
122 | #define in 36 /* unsigned char* */ | ||
123 | #define beg 32 /* unsigned char* */ | ||
124 | #define dist 28 /* unsigned int */ | ||
125 | #define len 24 /* unsigned int */ | ||
126 | #define last 20 /* unsigned char* */ | ||
127 | #define end 16 /* unsigned char* */ | ||
128 | #define dcode 12 /* code* */ | ||
129 | #define lcode 8 /* code* */ | ||
130 | #define dmask 4 /* unsigned int */ | ||
131 | #define lmask 0 /* unsigned int */ | ||
132 | |||
133 | /* | ||
134 | * typedef enum inflate_mode consts, in inflate.h | ||
135 | */ | ||
136 | #ifndef NO_GUNZIP | ||
137 | #define GUNZIP | ||
138 | #endif | ||
139 | |||
140 | #ifdef GUNZIP | ||
141 | #define INFLATE_MODE_TYPE 11 /* state->mode flags enum-ed in inflate.h */ | ||
142 | #define INFLATE_MODE_BAD 26 | ||
143 | #else | ||
144 | #define INFLATE_MODE_TYPE 3 | ||
145 | #define INFLATE_MODE_BAD 17 | ||
146 | #endif | ||
147 | |||
148 | |||
149 | .align 16,0x90 | ||
150 | inflate_fast: | ||
151 | pushl %edi | ||
152 | pushl %esi | ||
153 | pushl %ebp | ||
154 | pushl %ebx | ||
155 | pushf /* save eflags (strm_sp, state_sp assumes this is 32 bits) */ | ||
156 | subl $local_var_size, %esp | ||
157 | cld | ||
158 | #if defined( USE_MMX ) | ||
159 | emms | ||
160 | #endif | ||
161 | |||
162 | #define strm_r %esi | ||
163 | #define state_r %edi | ||
164 | |||
165 | movl strm_sp(%esp), strm_r | ||
166 | movl state_strm(strm_r), state_r | ||
167 | |||
168 | /* in = strm->next_in; | ||
169 | * out = strm->next_out; | ||
170 | * last = in + strm->avail_in - 5; | ||
171 | * beg = out - (start - strm->avail_out); | ||
172 | * end = out + (strm->avail_out - 257); | ||
173 | */ | ||
174 | movl next_in_strm(strm_r), %eax | ||
175 | movl next_out_strm(strm_r), %ebx | ||
176 | movl avail_in_strm(strm_r), %edx | ||
177 | movl avail_out_strm(strm_r), %ecx | ||
178 | movl start_sp(%esp), %ebp | ||
179 | |||
180 | addl %eax, %edx /* avail_in += next_in */ | ||
181 | subl $5, %edx /* avail_in -= 5 */ | ||
182 | |||
183 | subl %ecx, %ebp /* start -= avail_out */ | ||
184 | negl %ebp /* start = -start */ | ||
185 | addl %ebx, %ebp /* start += next_out */ | ||
186 | |||
187 | subl $257, %ecx /* avail_out -= 257 */ | ||
188 | addl %ebx, %ecx /* avail_out += out */ | ||
189 | |||
190 | movl %eax, in(%esp) | ||
191 | movl %ebx, out(%esp) | ||
192 | movl %edx, last(%esp) | ||
193 | movl %ebp, beg(%esp) | ||
194 | movl %ecx, end(%esp) | ||
195 | |||
196 | /* wsize = state->wsize; | ||
197 | * write = state->write; | ||
198 | * window = state->window; | ||
199 | * hold = state->hold; | ||
200 | * bits = state->bits; | ||
201 | * lcode = state->lencode; | ||
202 | * dcode = state->distcode; | ||
203 | * lmask = ( 1 << state->lenbits ) - 1; | ||
204 | * dmask = ( 1 << state->distbits ) - 1; | ||
205 | */ | ||
206 | |||
207 | movl lencode_state(state_r), %eax | ||
208 | movl distcode_state(state_r), %ecx | ||
209 | |||
210 | movl %eax, lcode(%esp) | ||
211 | movl %ecx, dcode(%esp) | ||
212 | |||
213 | movl $1, %eax | ||
214 | movl lenbits_state(state_r), %ecx | ||
215 | shll %cl, %eax | ||
216 | decl %eax | ||
217 | movl %eax, lmask(%esp) | ||
218 | |||
219 | movl $1, %eax | ||
220 | movl distbits_state(state_r), %ecx | ||
221 | shll %cl, %eax | ||
222 | decl %eax | ||
223 | movl %eax, dmask(%esp) | ||
224 | |||
225 | movl wsize_state(state_r), %eax | ||
226 | movl write_state(state_r), %ecx | ||
227 | movl window_state(state_r), %edx | ||
228 | |||
229 | movl %eax, wsize(%esp) | ||
230 | movl %ecx, write(%esp) | ||
231 | movl %edx, window(%esp) | ||
232 | |||
233 | #if ! defined( USE_MMX ) | ||
234 | |||
235 | #define hold_r %ebp | ||
236 | #define bits_r %bl | ||
237 | #define bitslong_r %ebx | ||
238 | |||
239 | movl hold_state(state_r), hold_r | ||
240 | movl bits_state(state_r), bitslong_r | ||
241 | |||
242 | #else /* USE_MMX */ | ||
243 | |||
244 | #define hold_mm %mm0 | ||
245 | #define bits_r %ebp | ||
246 | #define bitslong_r %ebp | ||
247 | |||
248 | movl hold_state(state_r), %ebx | ||
249 | movl bits_state(state_r), bitslong_r | ||
250 | |||
251 | #endif | ||
252 | |||
253 | #undef strm_r | ||
254 | #undef state_r | ||
255 | #define in_r %esi | ||
256 | #define from_r %esi | ||
257 | #define out_r %edi | ||
258 | |||
259 | movl in(%esp), in_r | ||
260 | |||
261 | #if ! defined ( USE_MMX ) | ||
262 | |||
263 | /* align in_r on word boundary */ | ||
264 | testl $1, in_r | ||
265 | jz .L_is_aligned | ||
266 | xorl %eax, %eax | ||
267 | movb (in_r), %al | ||
268 | incl in_r | ||
269 | movb bits_r, %cl | ||
270 | addb $8, bits_r | ||
271 | shll %cl, %eax | ||
272 | orl %eax, hold_r | ||
273 | |||
274 | #else | ||
275 | /* align in_r on long boundary */ | ||
276 | .L_align_long: | ||
277 | testl $3, in_r | ||
278 | jz .L_is_aligned | ||
279 | xorl %eax, %eax | ||
280 | movb (in_r), %al | ||
281 | incl in_r | ||
282 | movl bits_r, %ecx | ||
283 | addl $8, bits_r | ||
284 | shll %cl, %eax | ||
285 | orl %eax, %ebx | ||
286 | jmp .L_align_long | ||
287 | |||
288 | #endif | ||
289 | |||
290 | .L_is_aligned: | ||
291 | movl out(%esp), out_r | ||
292 | |||
293 | #if defined ( USE_MMX ) | ||
294 | |||
295 | #define used_mm %mm1 | ||
296 | #define dmask2_mm %mm2 | ||
297 | #define lmask2_mm %mm3 | ||
298 | #define lmask_mm %mm4 | ||
299 | #define dmask_mm %mm5 | ||
300 | #define tmp_mm %mm6 | ||
301 | |||
302 | movl out(%esp), out_r | ||
303 | movd lmask(%esp), lmask_mm | ||
304 | movq lmask_mm, lmask2_mm | ||
305 | movd dmask(%esp), dmask_mm | ||
306 | movq dmask_mm, dmask2_mm | ||
307 | movd %ebx, hold_mm | ||
308 | pxor used_mm, used_mm | ||
309 | movl lcode(%esp), %ebx /* ebx = lcode */ | ||
310 | #endif | ||
311 | |||
312 | jmp .L_do_loop | ||
313 | |||
314 | .align 16,0x90 | ||
315 | |||
316 | #if ! defined ( USE_MMX ) | ||
317 | |||
318 | .L_do_loop: | ||
319 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out | ||
320 | * | ||
321 | * do { | ||
322 | * if (bits < 15) { | ||
323 | * hold |= *((unsigned short *)in)++ << bits; | ||
324 | * bits += 16 | ||
325 | * } | ||
326 | * this = lcode[hold & lmask] | ||
327 | */ | ||
328 | cmpb $15, bits_r | ||
329 | ja .L_get_length_code /* if (15 < bits) */ | ||
330 | |||
331 | xorl %eax, %eax | ||
332 | lodsw /* al = *(ushort *)in++ */ | ||
333 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | ||
334 | addb $16, bits_r /* bits += 16 */ | ||
335 | shll %cl, %eax | ||
336 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | ||
337 | |||
338 | .L_get_length_code: | ||
339 | movl lmask(%esp), %edx /* edx = lmask */ | ||
340 | movl lcode(%esp), %ecx /* ecx = lcode */ | ||
341 | andl hold_r, %edx /* edx &= hold */ | ||
342 | movl (%ecx,%edx,4), %eax /* eax = lcode[hold & lmask] */ | ||
343 | |||
344 | #else /* USE_MMX */ | ||
345 | |||
346 | .L_do_loop: | ||
347 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
348 | |||
349 | cmpl $32, bits_r | ||
350 | ja .L_get_length_code /* if (32 < bits) */ | ||
351 | |||
352 | movd bits_r, tmp_mm | ||
353 | movd (in_r), %mm7 | ||
354 | addl $4, in_r | ||
355 | psllq tmp_mm, %mm7 | ||
356 | addl $32, bits_r | ||
357 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ | ||
358 | |||
359 | .L_get_length_code: | ||
360 | pand hold_mm, lmask_mm | ||
361 | movd lmask_mm, %eax | ||
362 | movq lmask2_mm, lmask_mm | ||
363 | movl (%ebx,%eax,4), %eax /* eax = lcode[hold & lmask] */ | ||
364 | |||
365 | #endif | ||
366 | |||
367 | #if ! defined( USE_MMX ) | ||
368 | |||
369 | .L_dolen: | ||
370 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out | ||
371 | * | ||
372 | * dolen: | ||
373 | * bits -= this.bits; | ||
374 | * hold >>= this.bits | ||
375 | */ | ||
376 | movb %ah, %cl /* cl = this.bits */ | ||
377 | subb %ah, bits_r /* bits -= this.bits */ | ||
378 | shrl %cl, hold_r /* hold >>= this.bits */ | ||
379 | |||
380 | /* check if op is a literal | ||
381 | * if (op == 0) { | ||
382 | * PUP(out) = this.val; | ||
383 | * } | ||
384 | */ | ||
385 | testb %al, %al | ||
386 | jnz .L_test_for_length_base /* if (op != 0) 45.7% */ | ||
387 | |||
388 | shrl $16, %eax /* output this.val char */ | ||
389 | stosb | ||
390 | |||
391 | #else /* USE_MMX */ | ||
392 | |||
393 | #define len_r %edx | ||
394 | |||
395 | .L_dolen: | ||
396 | movzbl %ah, %ecx /* ecx = this.bits */ | ||
397 | movl %eax, len_r /* len = this */ | ||
398 | shrl $16, len_r /* len = this.val */ | ||
399 | movd %ecx, used_mm | ||
400 | subl %ecx, bits_r /* bits -= this.bits */ | ||
401 | |||
402 | testb %al, %al | ||
403 | jnz .L_test_for_length_base /* if (op != 0) 45.7% */ | ||
404 | |||
405 | movb %dl, (out_r) | ||
406 | incl out_r | ||
407 | |||
408 | #endif | ||
409 | |||
410 | .L_while_test: | ||
411 | /* while (in < last && out < end) | ||
412 | */ | ||
413 | cmpl out_r, end(%esp) | ||
414 | jbe .L_break_loop /* if (out >= end) */ | ||
415 | |||
416 | cmpl in_r, last(%esp) | ||
417 | ja .L_do_loop /* if (in < last) */ | ||
418 | jmp .L_break_loop | ||
419 | |||
420 | #if ! defined( USE_MMX ) | ||
421 | |||
422 | .L_test_for_length_base: | ||
423 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = len | ||
424 | * | ||
425 | * else if (op & 16) { | ||
426 | * len = this.val | ||
427 | * op &= 15 | ||
428 | * if (op) { | ||
429 | * if (op > bits) { | ||
430 | * hold |= *((unsigned short *)in)++ << bits; | ||
431 | * bits += 16 | ||
432 | * } | ||
433 | * len += hold & mask[op]; | ||
434 | * bits -= op; | ||
435 | * hold >>= op; | ||
436 | * } | ||
437 | */ | ||
438 | #define len_r %edx | ||
439 | movl %eax, len_r /* len = this */ | ||
440 | shrl $16, len_r /* len = this.val */ | ||
441 | movb %al, %cl | ||
442 | |||
443 | testb $16, %al | ||
444 | jz .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ | ||
445 | andb $15, %cl /* op &= 15 */ | ||
446 | jz .L_save_len /* if (!op) */ | ||
447 | cmpb %cl, bits_r | ||
448 | jae .L_add_bits_to_len /* if (op <= bits) */ | ||
449 | |||
450 | movb %cl, %ch /* stash op in ch, freeing cl */ | ||
451 | xorl %eax, %eax | ||
452 | lodsw /* al = *(ushort *)in++ */ | ||
453 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | ||
454 | addb $16, bits_r /* bits += 16 */ | ||
455 | shll %cl, %eax | ||
456 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | ||
457 | movb %ch, %cl /* move op back to ecx */ | ||
458 | |||
459 | .L_add_bits_to_len: | ||
460 | movl $1, %eax | ||
461 | shll %cl, %eax | ||
462 | decl %eax | ||
463 | subb %cl, bits_r | ||
464 | andl hold_r, %eax /* eax &= hold */ | ||
465 | shrl %cl, hold_r | ||
466 | addl %eax, len_r /* len += hold & mask[op] */ | ||
467 | |||
468 | .L_save_len: | ||
469 | movl len_r, len(%esp) /* save len */ | ||
470 | #undef len_r | ||
471 | |||
472 | .L_decode_distance: | ||
473 | /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
474 | * | ||
475 | * if (bits < 15) { | ||
476 | * hold |= *((unsigned short *)in)++ << bits; | ||
477 | * bits += 16 | ||
478 | * } | ||
479 | * this = dcode[hold & dmask]; | ||
480 | * dodist: | ||
481 | * bits -= this.bits; | ||
482 | * hold >>= this.bits; | ||
483 | * op = this.op; | ||
484 | */ | ||
485 | |||
486 | cmpb $15, bits_r | ||
487 | ja .L_get_distance_code /* if (15 < bits) */ | ||
488 | |||
489 | xorl %eax, %eax | ||
490 | lodsw /* al = *(ushort *)in++ */ | ||
491 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | ||
492 | addb $16, bits_r /* bits += 16 */ | ||
493 | shll %cl, %eax | ||
494 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | ||
495 | |||
496 | .L_get_distance_code: | ||
497 | movl dmask(%esp), %edx /* edx = dmask */ | ||
498 | movl dcode(%esp), %ecx /* ecx = dcode */ | ||
499 | andl hold_r, %edx /* edx &= hold */ | ||
500 | movl (%ecx,%edx,4), %eax /* eax = dcode[hold & dmask] */ | ||
501 | |||
502 | #else /* USE_MMX */ | ||
503 | |||
504 | .L_test_for_length_base: | ||
505 | testb $16, %al | ||
506 | jz .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */ | ||
507 | andl $15, %eax /* op &= 15 */ | ||
508 | jz .L_decode_distance /* if (!op) */ | ||
509 | |||
510 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
511 | movd %eax, used_mm | ||
512 | movd hold_mm, %ecx | ||
513 | subl %eax, bits_r | ||
514 | andl .L_mask(,%eax,4), %ecx | ||
515 | addl %ecx, len_r /* len += hold & mask[op] */ | ||
516 | |||
517 | .L_decode_distance: | ||
518 | |||
519 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
520 | |||
521 | cmpl $32, bits_r | ||
522 | ja .L_get_dist_code /* if (32 < bits) */ | ||
523 | |||
524 | movd bits_r, tmp_mm | ||
525 | movd (in_r), %mm7 | ||
526 | addl $4, in_r | ||
527 | psllq tmp_mm, %mm7 | ||
528 | addl $32, bits_r | ||
529 | por %mm7, hold_mm /* hold_mm |= *((uint *)in)++ << bits */ | ||
530 | |||
531 | .L_get_dist_code: | ||
532 | movl dcode(%esp), %ebx /* ebx = dcode */ | ||
533 | pand hold_mm, dmask_mm | ||
534 | movd dmask_mm, %eax | ||
535 | movq dmask2_mm, dmask_mm | ||
536 | movl (%ebx,%eax,4), %eax /* eax = dcode[hold & lmask] */ | ||
537 | |||
538 | #endif | ||
539 | |||
540 | #if ! defined( USE_MMX ) | ||
541 | |||
542 | #define dist_r %edx | ||
543 | .L_dodist: | ||
544 | movl %eax, dist_r /* dist = this */ | ||
545 | shrl $16, dist_r /* dist = this.val */ | ||
546 | movb %ah, %cl | ||
547 | subb %ah, bits_r /* bits -= this.bits */ | ||
548 | shrl %cl, hold_r /* hold >>= this.bits */ | ||
549 | |||
550 | /* if (op & 16) { | ||
551 | * dist = this.val | ||
552 | * op &= 15 | ||
553 | * if (op > bits) { | ||
554 | * hold |= *((unsigned short *)in)++ << bits; | ||
555 | * bits += 16 | ||
556 | * } | ||
557 | * dist += hold & mask[op]; | ||
558 | * bits -= op; | ||
559 | * hold >>= op; | ||
560 | */ | ||
561 | movb %al, %cl /* cl = this.op */ | ||
562 | |||
563 | testb $16, %al /* if ((op & 16) == 0) */ | ||
564 | jz .L_test_for_second_level_dist | ||
565 | andb $15, %cl /* op &= 15 */ | ||
566 | jz .L_check_dist_one | ||
567 | cmpb %cl, bits_r | ||
568 | jae .L_add_bits_to_dist /* if (op <= bits) 97.6% */ | ||
569 | |||
570 | movb %cl, %ch /* stash op in ch, freeing cl */ | ||
571 | xorl %eax, %eax | ||
572 | lodsw /* al = *(ushort *)in++ */ | ||
573 | movb bits_r, %cl /* cl = bits, needs it for shifting */ | ||
574 | addb $16, bits_r /* bits += 16 */ | ||
575 | shll %cl, %eax | ||
576 | orl %eax, hold_r /* hold |= *((ushort *)in)++ << bits */ | ||
577 | movb %ch, %cl /* move op back to ecx */ | ||
578 | |||
579 | .L_add_bits_to_dist: | ||
580 | movl $1, %eax | ||
581 | shll %cl, %eax | ||
582 | decl %eax /* (1 << op) - 1 */ | ||
583 | subb %cl, bits_r | ||
584 | andl hold_r, %eax /* eax &= hold */ | ||
585 | shrl %cl, hold_r | ||
586 | addl %eax, dist_r /* dist += hold & ((1 << op) - 1) */ | ||
587 | jmp .L_check_window | ||
588 | |||
589 | #else /* USE_MMX */ | ||
590 | |||
591 | #define dist_r %ebx | ||
592 | .L_dodist: | ||
593 | movzbl %ah, %ecx /* ecx = this.bits */ | ||
594 | movl %eax, dist_r | ||
595 | shrl $16, dist_r /* dist = this.val */ | ||
596 | subl %ecx, bits_r /* bits -= this.bits */ | ||
597 | movd %ecx, used_mm | ||
598 | |||
599 | testb $16, %al /* if ((op & 16) == 0) */ | ||
600 | jz .L_test_for_second_level_dist | ||
601 | andl $15, %eax /* op &= 15 */ | ||
602 | jz .L_check_dist_one | ||
603 | |||
604 | .L_add_bits_to_dist: | ||
605 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
606 | movd %eax, used_mm /* save bit length of current op */ | ||
607 | movd hold_mm, %ecx /* get the next bits on input stream */ | ||
608 | subl %eax, bits_r /* bits -= op bits */ | ||
609 | andl .L_mask(,%eax,4), %ecx /* ecx = hold & mask[op] */ | ||
610 | addl %ecx, dist_r /* dist += hold & mask[op] */ | ||
611 | jmp .L_check_window | ||
612 | |||
613 | #endif | ||
614 | |||
615 | .align 16,0x90 | ||
616 | |||
617 | .L_check_dist_one: | ||
618 | cmpl $1, dist_r | ||
619 | jne .L_check_window | ||
620 | cmpl out_r, beg(%esp) | ||
621 | je .L_check_window | ||
622 | |||
623 | decl out_r | ||
624 | #if ! defined( USE_MMX ) | ||
625 | movl len(%esp), %ecx | ||
626 | #else | ||
627 | movl len_r, %ecx | ||
628 | #endif | ||
629 | movb (out_r), %al | ||
630 | subl $3, %ecx | ||
631 | |||
632 | movb %al, 1(out_r) | ||
633 | movb %al, 2(out_r) | ||
634 | movb %al, 3(out_r) | ||
635 | addl $4, out_r | ||
636 | rep stosb | ||
637 | |||
638 | #if defined( USE_MMX ) | ||
639 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | ||
640 | #endif | ||
641 | jmp .L_while_test | ||
642 | |||
643 | .align 16,0x90 | ||
644 | |||
645 | .L_check_window: | ||
646 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
647 | * %ecx = nbytes | ||
648 | * | ||
649 | * nbytes = out - beg; | ||
650 | * if (dist <= nbytes) { | ||
651 | * from = out - dist; | ||
652 | * do { | ||
653 | * PUP(out) = PUP(from); | ||
654 | * } while (--len > 0) { | ||
655 | * } | ||
656 | */ | ||
657 | |||
658 | movl in_r, in(%esp) /* save in so from can use it's reg */ | ||
659 | movl out_r, %eax | ||
660 | subl beg(%esp), %eax /* nbytes = out - beg */ | ||
661 | |||
662 | cmpl dist_r, %eax | ||
663 | jb .L_clip_window /* if (dist > nbytes) 4.2% */ | ||
664 | |||
665 | #if ! defined( USE_MMX ) | ||
666 | movl len(%esp), %ecx | ||
667 | #else | ||
668 | movl len_r, %ecx | ||
669 | #endif | ||
670 | movl out_r, from_r | ||
671 | subl dist_r, from_r /* from = out - dist */ | ||
672 | |||
673 | subl $3, %ecx | ||
674 | movb (from_r), %al | ||
675 | movb %al, (out_r) | ||
676 | movb 1(from_r), %al | ||
677 | movb 2(from_r), %dl | ||
678 | addl $3, from_r | ||
679 | movb %al, 1(out_r) | ||
680 | movb %dl, 2(out_r) | ||
681 | addl $3, out_r | ||
682 | rep movsb | ||
683 | |||
684 | movl in(%esp), in_r /* move in back to %esi, toss from */ | ||
685 | #if defined( USE_MMX ) | ||
686 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | ||
687 | #endif | ||
688 | jmp .L_while_test | ||
689 | |||
690 | .align 16,0x90 | ||
691 | |||
692 | #if ! defined( USE_MMX ) | ||
693 | |||
694 | .L_test_for_second_level_length: | ||
695 | /* else if ((op & 64) == 0) { | ||
696 | * this = lcode[this.val + (hold & mask[op])]; | ||
697 | * } | ||
698 | */ | ||
699 | testb $64, %al | ||
700 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ | ||
701 | |||
702 | movl $1, %eax | ||
703 | shll %cl, %eax | ||
704 | decl %eax | ||
705 | andl hold_r, %eax /* eax &= hold */ | ||
706 | addl %edx, %eax /* eax += this.val */ | ||
707 | movl lcode(%esp), %edx /* edx = lcode */ | ||
708 | movl (%edx,%eax,4), %eax /* eax = lcode[val + (hold&mask[op])] */ | ||
709 | jmp .L_dolen | ||
710 | |||
711 | #else /* USE_MMX */ | ||
712 | |||
713 | .L_test_for_second_level_length: | ||
714 | testb $64, %al | ||
715 | jnz .L_test_for_end_of_block /* if ((op & 64) != 0) */ | ||
716 | |||
717 | andl $15, %eax | ||
718 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
719 | movd hold_mm, %ecx | ||
720 | andl .L_mask(,%eax,4), %ecx | ||
721 | addl len_r, %ecx | ||
722 | movl (%ebx,%ecx,4), %eax /* eax = lcode[hold & lmask] */ | ||
723 | jmp .L_dolen | ||
724 | |||
725 | #endif | ||
726 | |||
727 | .align 16,0x90 | ||
728 | |||
729 | #if ! defined( USE_MMX ) | ||
730 | |||
731 | .L_test_for_second_level_dist: | ||
732 | /* else if ((op & 64) == 0) { | ||
733 | * this = dcode[this.val + (hold & mask[op])]; | ||
734 | * } | ||
735 | */ | ||
736 | testb $64, %al | ||
737 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ | ||
738 | |||
739 | movl $1, %eax | ||
740 | shll %cl, %eax | ||
741 | decl %eax | ||
742 | andl hold_r, %eax /* eax &= hold */ | ||
743 | addl %edx, %eax /* eax += this.val */ | ||
744 | movl dcode(%esp), %edx /* edx = dcode */ | ||
745 | movl (%edx,%eax,4), %eax /* eax = dcode[val + (hold&mask[op])] */ | ||
746 | jmp .L_dodist | ||
747 | |||
748 | #else /* USE_MMX */ | ||
749 | |||
750 | .L_test_for_second_level_dist: | ||
751 | testb $64, %al | ||
752 | jnz .L_invalid_distance_code /* if ((op & 64) != 0) */ | ||
753 | |||
754 | andl $15, %eax | ||
755 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
756 | movd hold_mm, %ecx | ||
757 | andl .L_mask(,%eax,4), %ecx | ||
758 | movl dcode(%esp), %eax /* ecx = dcode */ | ||
759 | addl dist_r, %ecx | ||
760 | movl (%eax,%ecx,4), %eax /* eax = lcode[hold & lmask] */ | ||
761 | jmp .L_dodist | ||
762 | |||
763 | #endif | ||
764 | |||
765 | .align 16,0x90 | ||
766 | .L_clip_window: | ||
767 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
768 | * %ecx = nbytes | ||
769 | * | ||
770 | * else { | ||
771 | * if (dist > wsize) { | ||
772 | * invalid distance | ||
773 | * } | ||
774 | * from = window; | ||
775 | * nbytes = dist - nbytes; | ||
776 | * if (write == 0) { | ||
777 | * from += wsize - nbytes; | ||
778 | */ | ||
779 | #define nbytes_r %ecx | ||
780 | |||
781 | movl %eax, nbytes_r | ||
782 | movl wsize(%esp), %eax /* prepare for dist compare */ | ||
783 | negl nbytes_r /* nbytes = -nbytes */ | ||
784 | movl window(%esp), from_r /* from = window */ | ||
785 | |||
786 | cmpl dist_r, %eax | ||
787 | jb .L_invalid_distance_too_far /* if (dist > wsize) */ | ||
788 | |||
789 | addl dist_r, nbytes_r /* nbytes = dist - nbytes */ | ||
790 | cmpl $0, write(%esp) | ||
791 | jne .L_wrap_around_window /* if (write != 0) */ | ||
792 | |||
793 | subl nbytes_r, %eax | ||
794 | addl %eax, from_r /* from += wsize - nbytes */ | ||
795 | |||
796 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
797 | * %ecx = nbytes, %eax = len | ||
798 | * | ||
799 | * if (nbytes < len) { | ||
800 | * len -= nbytes; | ||
801 | * do { | ||
802 | * PUP(out) = PUP(from); | ||
803 | * } while (--nbytes); | ||
804 | * from = out - dist; | ||
805 | * } | ||
806 | * } | ||
807 | */ | ||
808 | |||
809 | #if ! defined( USE_MMX ) | ||
810 | #define len_r %eax | ||
811 | movl len(%esp), len_r | ||
812 | #endif | ||
813 | cmpl nbytes_r, len_r | ||
814 | jbe .L_do_copy1 /* if (nbytes >= len) */ | ||
815 | |||
816 | subl nbytes_r, len_r /* len -= nbytes */ | ||
817 | rep movsb | ||
818 | movl out_r, from_r | ||
819 | subl dist_r, from_r /* from = out - dist */ | ||
820 | jmp .L_do_copy1 | ||
821 | |||
822 | cmpl nbytes_r, len_r | ||
823 | jbe .L_do_copy1 /* if (nbytes >= len) */ | ||
824 | |||
825 | subl nbytes_r, len_r /* len -= nbytes */ | ||
826 | rep movsb | ||
827 | movl out_r, from_r | ||
828 | subl dist_r, from_r /* from = out - dist */ | ||
829 | jmp .L_do_copy1 | ||
830 | |||
831 | .L_wrap_around_window: | ||
832 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
833 | * %ecx = nbytes, %eax = write, %eax = len | ||
834 | * | ||
835 | * else if (write < nbytes) { | ||
836 | * from += wsize + write - nbytes; | ||
837 | * nbytes -= write; | ||
838 | * if (nbytes < len) { | ||
839 | * len -= nbytes; | ||
840 | * do { | ||
841 | * PUP(out) = PUP(from); | ||
842 | * } while (--nbytes); | ||
843 | * from = window; | ||
844 | * nbytes = write; | ||
845 | * if (nbytes < len) { | ||
846 | * len -= nbytes; | ||
847 | * do { | ||
848 | * PUP(out) = PUP(from); | ||
849 | * } while(--nbytes); | ||
850 | * from = out - dist; | ||
851 | * } | ||
852 | * } | ||
853 | * } | ||
854 | */ | ||
855 | #define write_r %eax | ||
856 | |||
857 | movl write(%esp), write_r | ||
858 | cmpl write_r, nbytes_r | ||
859 | jbe .L_contiguous_in_window /* if (write >= nbytes) */ | ||
860 | |||
861 | addl wsize(%esp), from_r | ||
862 | addl write_r, from_r | ||
863 | subl nbytes_r, from_r /* from += wsize + write - nbytes */ | ||
864 | subl write_r, nbytes_r /* nbytes -= write */ | ||
865 | #undef write_r | ||
866 | |||
867 | #if ! defined( USE_MMX ) | ||
868 | movl len(%esp), len_r | ||
869 | #endif | ||
870 | cmpl nbytes_r, len_r | ||
871 | jbe .L_do_copy1 /* if (nbytes >= len) */ | ||
872 | |||
873 | subl nbytes_r, len_r /* len -= nbytes */ | ||
874 | rep movsb | ||
875 | movl window(%esp), from_r /* from = window */ | ||
876 | movl write(%esp), nbytes_r /* nbytes = write */ | ||
877 | cmpl nbytes_r, len_r | ||
878 | jbe .L_do_copy1 /* if (nbytes >= len) */ | ||
879 | |||
880 | subl nbytes_r, len_r /* len -= nbytes */ | ||
881 | rep movsb | ||
882 | movl out_r, from_r | ||
883 | subl dist_r, from_r /* from = out - dist */ | ||
884 | jmp .L_do_copy1 | ||
885 | |||
886 | .L_contiguous_in_window: | ||
887 | /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist | ||
888 | * %ecx = nbytes, %eax = write, %eax = len | ||
889 | * | ||
890 | * else { | ||
891 | * from += write - nbytes; | ||
892 | * if (nbytes < len) { | ||
893 | * len -= nbytes; | ||
894 | * do { | ||
895 | * PUP(out) = PUP(from); | ||
896 | * } while (--nbytes); | ||
897 | * from = out - dist; | ||
898 | * } | ||
899 | * } | ||
900 | */ | ||
901 | #define write_r %eax | ||
902 | |||
903 | addl write_r, from_r | ||
904 | subl nbytes_r, from_r /* from += write - nbytes */ | ||
905 | #undef write_r | ||
906 | |||
907 | #if ! defined( USE_MMX ) | ||
908 | movl len(%esp), len_r | ||
909 | #endif | ||
910 | cmpl nbytes_r, len_r | ||
911 | jbe .L_do_copy1 /* if (nbytes >= len) */ | ||
912 | |||
913 | subl nbytes_r, len_r /* len -= nbytes */ | ||
914 | rep movsb | ||
915 | movl out_r, from_r | ||
916 | subl dist_r, from_r /* from = out - dist */ | ||
917 | |||
918 | .L_do_copy1: | ||
919 | /* regs: %esi = from, %esi = in, %ebp = hold, %bl = bits, %edi = out | ||
920 | * %eax = len | ||
921 | * | ||
922 | * while (len > 0) { | ||
923 | * PUP(out) = PUP(from); | ||
924 | * len--; | ||
925 | * } | ||
926 | * } | ||
927 | * } while (in < last && out < end); | ||
928 | */ | ||
929 | #undef nbytes_r | ||
930 | #define in_r %esi | ||
931 | |||
932 | movl len_r, %ecx | ||
933 | rep movsb | ||
934 | |||
935 | movl in(%esp), in_r /* move in back to %esi, toss from */ | ||
936 | #if defined( USE_MMX ) | ||
937 | movl lcode(%esp), %ebx /* move lcode back to %ebx, toss dist */ | ||
938 | #endif | ||
939 | jmp .L_while_test | ||
940 | |||
941 | #undef len_r | ||
942 | #undef from_r | ||
943 | #undef dist_r | ||
944 | |||
945 | .L_invalid_distance_code: | ||
946 | /* else { | ||
947 | * strm->msg = "invalid distance code"; | ||
948 | * state->mode = BAD; | ||
949 | * } | ||
950 | */ | ||
951 | movl $.L_invalid_distance_code_msg, %ecx | ||
952 | movl $INFLATE_MODE_BAD, %edx | ||
953 | jmp .L_update_stream_state | ||
954 | |||
955 | .L_test_for_end_of_block: | ||
956 | /* else if (op & 32) { | ||
957 | * state->mode = TYPE; | ||
958 | * break; | ||
959 | * } | ||
960 | */ | ||
961 | testb $32, %al | ||
962 | jz .L_invalid_literal_length_code /* if ((op & 32) == 0) */ | ||
963 | |||
964 | movl $0, %ecx | ||
965 | movl $INFLATE_MODE_TYPE, %edx | ||
966 | jmp .L_update_stream_state | ||
967 | |||
968 | .L_invalid_literal_length_code: | ||
969 | /* else { | ||
970 | * strm->msg = "invalid literal/length code"; | ||
971 | * state->mode = BAD; | ||
972 | * } | ||
973 | */ | ||
974 | movl $.L_invalid_literal_length_code_msg, %ecx | ||
975 | movl $INFLATE_MODE_BAD, %edx | ||
976 | jmp .L_update_stream_state | ||
977 | |||
978 | .L_invalid_distance_too_far: | ||
979 | /* strm->msg = "invalid distance too far back"; | ||
980 | * state->mode = BAD; | ||
981 | */ | ||
982 | movl in(%esp), in_r /* from_r has in's reg, put in back */ | ||
983 | movl $.L_invalid_distance_too_far_msg, %ecx | ||
984 | movl $INFLATE_MODE_BAD, %edx | ||
985 | jmp .L_update_stream_state | ||
986 | |||
987 | .L_update_stream_state: | ||
988 | /* set strm->msg = %ecx, strm->state->mode = %edx */ | ||
989 | movl strm_sp(%esp), %eax | ||
990 | testl %ecx, %ecx /* if (msg != NULL) */ | ||
991 | jz .L_skip_msg | ||
992 | movl %ecx, msg_strm(%eax) /* strm->msg = msg */ | ||
993 | .L_skip_msg: | ||
994 | movl state_strm(%eax), %eax /* state = strm->state */ | ||
995 | movl %edx, mode_state(%eax) /* state->mode = edx (BAD | TYPE) */ | ||
996 | |||
997 | .L_break_loop: | ||
998 | |||
999 | #define strm_r %eax | ||
1000 | #define state_r %edx | ||
1001 | |||
1002 | /* len = bits >> 3; | ||
1003 | * in -= len; | ||
1004 | * bits -= len << 3; | ||
1005 | * hold &= (1U << bits) - 1; | ||
1006 | * state->hold = hold; | ||
1007 | * state->bits = bits; | ||
1008 | * strm->next_in = in; | ||
1009 | * strm->next_out = out; | ||
1010 | */ | ||
1011 | movl strm_sp(%esp), strm_r | ||
1012 | movl bitslong_r, %ecx | ||
1013 | movl state_strm(strm_r), state_r | ||
1014 | shrl $3, %ecx | ||
1015 | subl %ecx, in_r | ||
1016 | shll $3, %ecx | ||
1017 | subl %ecx, bitslong_r | ||
1018 | movl out_r, next_out_strm(strm_r) | ||
1019 | movl in_r, next_in_strm(strm_r) | ||
1020 | movl bitslong_r, bits_state(state_r) | ||
1021 | |||
1022 | movl bitslong_r, %ecx | ||
1023 | movl $1, %ebx /* overwrites bitslong_r, %bl */ | ||
1024 | shll %cl, %ebx | ||
1025 | decl %ebx | ||
1026 | |||
1027 | #undef bits_r | ||
1028 | #undef bitslong_r | ||
1029 | |||
1030 | #if ! defined( USE_MMX ) | ||
1031 | |||
1032 | andl %ebx, hold_r | ||
1033 | movl hold_r, hold_state(state_r) | ||
1034 | |||
1035 | #else /* USE_MMX */ | ||
1036 | |||
1037 | psrlq used_mm, hold_mm /* hold_mm >>= last bit length */ | ||
1038 | movd hold_mm, %ecx | ||
1039 | andl %ebx, %ecx | ||
1040 | movl %ecx, hold_state(state_r) | ||
1041 | |||
1042 | #endif | ||
1043 | |||
1044 | #define last_r %ebx | ||
1045 | |||
1046 | /* strm->avail_in = in < last ? 5 + (last - in) : 5 - (in - last) */ | ||
1047 | movl last(%esp), last_r | ||
1048 | cmpl in_r, last_r | ||
1049 | jbe .L_last_is_smaller /* if (in >= last) */ | ||
1050 | |||
1051 | subl in_r, last_r /* last -= in */ | ||
1052 | addl $5, last_r /* last += 5 */ | ||
1053 | movl last_r, avail_in_strm(strm_r) | ||
1054 | jmp .L_fixup_out | ||
1055 | .L_last_is_smaller: | ||
1056 | subl last_r, in_r /* in -= last */ | ||
1057 | negl in_r /* in = -in */ | ||
1058 | addl $5, in_r /* in += 5 */ | ||
1059 | movl in_r, avail_in_strm(strm_r) | ||
1060 | |||
1061 | #undef last_r | ||
1062 | #define end_r %ebx | ||
1063 | |||
1064 | .L_fixup_out: | ||
1065 | /* strm->avail_out = out < end ? 257 + (end - out) : 257 - (out - end)*/ | ||
1066 | movl end(%esp), end_r | ||
1067 | cmpl out_r, end_r | ||
1068 | jbe .L_end_is_smaller /* if (out >= end) */ | ||
1069 | |||
1070 | subl out_r, end_r /* end -= out */ | ||
1071 | addl $257, end_r /* end += 257 */ | ||
1072 | movl end_r, avail_out_strm(strm_r) | ||
1073 | jmp .L_done | ||
1074 | .L_end_is_smaller: | ||
1075 | subl end_r, out_r /* out -= end */ | ||
1076 | negl out_r /* out = -out */ | ||
1077 | addl $257, out_r /* out += 257 */ | ||
1078 | movl out_r, avail_out_strm(strm_r) | ||
1079 | |||
1080 | #undef end_r | ||
1081 | |||
1082 | .L_done: | ||
1083 | #if defined( USE_MMX ) | ||
1084 | emms | ||
1085 | #endif | ||
1086 | addl $local_var_size, %esp | ||
1087 | popf | ||
1088 | popl %ebx | ||
1089 | popl %ebp | ||
1090 | popl %esi | ||
1091 | popl %edi | ||
1092 | ret | ||
1093 | |||
1094 | .type inflate_fast,@function | ||
1095 | .size inflate_fast,.-inflate_fast | ||