1 /*
2 ;uInt longest_match_x64(
3 ;    deflate_state *s,
4 ;    IPos cur_match);                             // current match
5 
6 ; gvmat64.S -- Asm portion of the optimized longest_match for 32 bits x86_64
7 ;  (AMD64 on Athlon 64, Opteron, Phenom
8 ;     and Intel EM64T on Pentium 4 with EM64T, Pentium D, Core 2 Duo, Core I5/I7)
9 ; this file is translation from gvmat64.asm to GCC 4.x (for Linux, Mac XCode)
10 ; Copyright (C) 1995-2010 Jean-loup Gailly, Brian Raiter and Gilles Vollant.
11 ;
12 ; File written by Gilles Vollant, by converting to assembly the longest_match
13 ;  from Jean-loup Gailly in deflate.c of zLib and infoZip zip.
14 ;  and by taking inspiration on asm686 with masm, optimised assembly code
15 ;        from Brian Raiter, written 1998
16 ;
17 ;  This software is provided 'as-is', without any express or implied
18 ;  warranty.  In no event will the authors be held liable for any damages
19 ;  arising from the use of this software.
20 ;
21 ;  Permission is granted to anyone to use this software for any purpose,
22 ;  including commercial applications, and to alter it and redistribute it
23 ;  freely, subject to the following restrictions:
24 ;
25 ;  1. The origin of this software must not be misrepresented; you must not
26 ;     claim that you wrote the original software. If you use this software
27 ;     in a product, an acknowledgment in the product documentation would be
28 ;     appreciated but is not required.
29 ;  2. Altered source versions must be plainly marked as such, and must not be
30 ;     misrepresented as being the original software
31 ;  3. This notice may not be removed or altered from any source distribution.
32 ;
33 ;         http://www.zlib.net
34 ;         http://www.winimage.com/zLibDll
35 ;         http://www.muppetlabs.com/~breadbox/software/assembly.html
36 ;
37 ; to compile this file for zLib, I use option:
38 ;   gcc -c -arch x86_64 gvmat64.S
39 
40 
41 ;uInt longest_match(s, cur_match)
42 ;    deflate_state *s;
43 ;    IPos cur_match;                             // current match /
44 ;
45 ; with XCode for Mac, I had strange error with some jump on intel syntax
46 ; this is why BEFORE_JMP and AFTER_JMP are used
47  */
48 
49 
50 #define BEFORE_JMP .att_syntax
51 #define AFTER_JMP .intel_syntax noprefix
52 
53 #ifndef NO_UNDERLINE
54 #	define	match_init	_match_init
55 #	define	longest_match	_longest_match
56 #endif
57 
58 .intel_syntax noprefix
59 
60 .globl	match_init, longest_match
61 .text
62 longest_match:
63 
64 
65 
66 #define LocalVarsSize 96
67 /*
68 ; register used : rax,rbx,rcx,rdx,rsi,rdi,r8,r9,r10,r11,r12
69 ; free register :  r14,r15
70 ; register can be saved : rsp
71 */
72 
73 #define chainlenwmask     (rsp + 8 - LocalVarsSize)
74 #define nicematch         (rsp + 16 - LocalVarsSize)
75 
76 #define save_rdi        (rsp + 24 - LocalVarsSize)
77 #define save_rsi        (rsp + 32 - LocalVarsSize)
78 #define save_rbx        (rsp + 40 - LocalVarsSize)
79 #define save_rbp        (rsp + 48 - LocalVarsSize)
80 #define save_r12        (rsp + 56 - LocalVarsSize)
81 #define save_r13        (rsp + 64 - LocalVarsSize)
82 #define save_r14        (rsp + 72 - LocalVarsSize)
83 #define save_r15        (rsp + 80 - LocalVarsSize)
84 
85 
86 /*
87 ;  all the +4 offsets are due to the addition of pending_buf_size (in zlib
88 ;  in the deflate_state structure since the asm code was first written
89 ;  (if you compile with zlib 1.0.4 or older, remove the +4).
90 ;  Note : these value are good with a 8 bytes boundary pack structure
91 */
92 
93 #define    MAX_MATCH              258
94 #define    MIN_MATCH              3
95 #define    MIN_LOOKAHEAD          (MAX_MATCH+MIN_MATCH+1)
96 
97 /*
98 ;;; Offsets for fields in the deflate_state structure. These numbers
99 ;;; are calculated from the definition of deflate_state, with the
100 ;;; assumption that the compiler will dword-align the fields. (Thus,
101 ;;; changing the definition of deflate_state could easily cause this
102 ;;; program to crash horribly, without so much as a warning at
103 ;;; compile time. Sigh.)
104 
105 ;  all the +zlib1222add offsets are due to the addition of fields
106 ;  in zlib in the deflate_state structure since the asm code was first written
107 ;  (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)").
108 ;  (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0").
109 ;  if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8").
110 */
111 
112 
113 
114 /* you can check the structure offset by running
115 
116 #include <stdlib.h>
117 #include <stdio.h>
118 #include "deflate.h"
119 
120 void print_depl()
121 {
122 deflate_state ds;
123 deflate_state *s=&ds;
124 printf("size pointer=%u\n",(int)sizeof(void*));
125 
126 printf("#define dsWSize         %u\n",(int)(((char*)&(s->w_size))-((char*)s)));
127 printf("#define dsWMask         %u\n",(int)(((char*)&(s->w_mask))-((char*)s)));
128 printf("#define dsWindow        %u\n",(int)(((char*)&(s->window))-((char*)s)));
129 printf("#define dsPrev          %u\n",(int)(((char*)&(s->prev))-((char*)s)));
130 printf("#define dsMatchLen      %u\n",(int)(((char*)&(s->match_length))-((char*)s)));
131 printf("#define dsPrevMatch     %u\n",(int)(((char*)&(s->prev_match))-((char*)s)));
132 printf("#define dsStrStart      %u\n",(int)(((char*)&(s->strstart))-((char*)s)));
133 printf("#define dsMatchStart    %u\n",(int)(((char*)&(s->match_start))-((char*)s)));
134 printf("#define dsLookahead     %u\n",(int)(((char*)&(s->lookahead))-((char*)s)));
135 printf("#define dsPrevLen       %u\n",(int)(((char*)&(s->prev_length))-((char*)s)));
136 printf("#define dsMaxChainLen   %u\n",(int)(((char*)&(s->max_chain_length))-((char*)s)));
137 printf("#define dsGoodMatch     %u\n",(int)(((char*)&(s->good_match))-((char*)s)));
138 printf("#define dsNiceMatch     %u\n",(int)(((char*)&(s->nice_match))-((char*)s)));
139 }
140 */
141 
142 #define dsWSize          68
143 #define dsWMask          76
144 #define dsWindow         80
145 #define dsPrev           96
146 #define dsMatchLen       144
147 #define dsPrevMatch      148
148 #define dsStrStart       156
149 #define dsMatchStart     160
150 #define dsLookahead      164
151 #define dsPrevLen        168
152 #define dsMaxChainLen    172
153 #define dsGoodMatch      188
154 #define dsNiceMatch      192
155 
156 #define window_size      [ rcx + dsWSize]
157 #define WMask            [ rcx + dsWMask]
158 #define window_ad        [ rcx + dsWindow]
159 #define prev_ad          [ rcx + dsPrev]
160 #define strstart         [ rcx + dsStrStart]
161 #define match_start      [ rcx + dsMatchStart]
162 #define Lookahead        [ rcx + dsLookahead] //; 0ffffffffh on infozip
163 #define prev_length      [ rcx + dsPrevLen]
164 #define max_chain_length [ rcx + dsMaxChainLen]
165 #define good_match       [ rcx + dsGoodMatch]
166 #define nice_match       [ rcx + dsNiceMatch]
167 
168 /*
169 ; windows:
170 ; parameter 1 in rcx(deflate state s), param 2 in rdx (cur match)
171 
172 ; see http://weblogs.asp.net/oldnewthing/archive/2004/01/14/58579.aspx and
173 ; http://msdn.microsoft.com/library/en-us/kmarch/hh/kmarch/64bitAMD_8e951dd2-ee77-4728-8702-55ce4b5dd24a.xml.asp
174 ;
175 ; All registers must be preserved across the call, except for
176 ;   rax, rcx, rdx, r8, r9, r10, and r11, which are scratch.
177 
178 ;
179 ; gcc on macosx-linux:
180 ; see http://www.x86-64.org/documentation/abi-0.99.pdf
181 ; param 1 in rdi, param 2 in rsi
182 ; rbx, rsp, rbp, r12 to r15 must be preserved
183 
184 ;;; Save registers that the compiler may be using, and adjust esp to
185 ;;; make room for our stack frame.
186 
187 
188 ;;; Retrieve the function arguments. r8d will hold cur_match
189 ;;; throughout the entire function. edx will hold the pointer to the
190 ;;; deflate_state structure during the function's setup (before
191 ;;; entering the main loop.
192 
193 ; ms: parameter 1 in rcx (deflate_state* s), param 2 in edx -> r8 (cur match)
194 ; mac: param 1 in rdi, param 2 rsi
195 ; this clear high 32 bits of r8, which can be garbage in both r8 and rdx
196 */
197         mov [save_rbx],rbx
198         mov [save_rbp],rbp
199 
200 
201         mov rcx,rdi
202 
203         mov r8d,esi
204 
205 
206         mov [save_r12],r12
207         mov [save_r13],r13
208         mov [save_r14],r14
209         mov [save_r15],r15
210 
211 
212 //;;; uInt wmask = s->w_mask;
213 //;;; unsigned chain_length = s->max_chain_length;
214 //;;; if (s->prev_length >= s->good_match) {
215 //;;;     chain_length >>= 2;
216 //;;; }
217 
218 
219         mov edi, prev_length
220         mov esi, good_match
221         mov eax, WMask
222         mov ebx, max_chain_length
223         cmp edi, esi
224         jl  LastMatchGood
225         shr ebx, 2
226 LastMatchGood:
227 
228 //;;; chainlen is decremented once beforehand so that the function can
229 //;;; use the sign flag instead of the zero flag for the exit test.
230 //;;; It is then shifted into the high word, to make room for the wmask
231 //;;; value, which it will always accompany.
232 
233         dec ebx
234         shl ebx, 16
235         or  ebx, eax
236 
237 //;;; on zlib only
238 //;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;
239 
240 
241 
242         mov eax, nice_match
243         mov [chainlenwmask], ebx
244         mov r10d, Lookahead
245         cmp r10d, eax
246         cmovnl r10d, eax
247         mov [nicematch],r10d
248 
249 
250 
251 //;;; register Bytef *scan = s->window + s->strstart;
252         mov r10, window_ad
253         mov ebp, strstart
254         lea r13, [r10 + rbp]
255 
256 //;;; Determine how many bytes the scan ptr is off from being
257 //;;; dword-aligned.
258 
259          mov r9,r13
260          neg r13
261          and r13,3
262 
263 //;;; IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
264 //;;;     s->strstart - (IPos)MAX_DIST(s) : NIL;
265 
266 
267         mov eax, window_size
268         sub eax, MIN_LOOKAHEAD
269 
270 
271         xor edi,edi
272         sub ebp, eax
273 
274         mov r11d, prev_length
275 
276         cmovng ebp,edi
277 
278 //;;; int best_len = s->prev_length;
279 
280 
281 //;;; Store the sum of s->window + best_len in esi locally, and in esi.
282 
283        lea  rsi,[r10+r11]
284 
285 //;;; register ush scan_start = *(ushf*)scan;
286 //;;; register ush scan_end   = *(ushf*)(scan+best_len-1);
287 //;;; Posf *prev = s->prev;
288 
289         movzx r12d,word ptr [r9]
290         movzx ebx, word ptr [r9 + r11 - 1]
291 
292         mov rdi, prev_ad
293 
294 //;;; Jump into the main loop.
295 
296         mov edx, [chainlenwmask]
297 
298         cmp bx,word ptr [rsi + r8 - 1]
299         jz  LookupLoopIsZero
300 
301 
302 
303 LookupLoop1:
304         and r8d, edx
305 
306         movzx   r8d, word ptr [rdi + r8*2]
307         cmp r8d, ebp
308         jbe LeaveNow
309 
310 
311 
312         sub edx, 0x00010000
313 		BEFORE_JMP
314         js  LeaveNow
315 		AFTER_JMP
316 
317 LoopEntry1:
318         cmp bx,word ptr [rsi + r8 - 1]
319 		BEFORE_JMP
320         jz  LookupLoopIsZero
321 		AFTER_JMP
322 
323 LookupLoop2:
324         and r8d, edx
325 
326         movzx   r8d, word ptr [rdi + r8*2]
327         cmp r8d, ebp
328 		BEFORE_JMP
329         jbe LeaveNow
330 		AFTER_JMP
331         sub edx, 0x00010000
332 		BEFORE_JMP
333         js  LeaveNow
334 		AFTER_JMP
335 
336 LoopEntry2:
337         cmp bx,word ptr [rsi + r8 - 1]
338 		BEFORE_JMP
339         jz  LookupLoopIsZero
340 		AFTER_JMP
341 
342 LookupLoop4:
343         and r8d, edx
344 
345         movzx   r8d, word ptr [rdi + r8*2]
346         cmp r8d, ebp
347 		BEFORE_JMP
348         jbe LeaveNow
349 		AFTER_JMP
350         sub edx, 0x00010000
351 		BEFORE_JMP
352         js  LeaveNow
353 		AFTER_JMP
354 
355 LoopEntry4:
356 
357         cmp bx,word ptr [rsi + r8 - 1]
358 		BEFORE_JMP
359         jnz LookupLoop1
360         jmp LookupLoopIsZero
361 		AFTER_JMP
362 /*
363 ;;; do {
364 ;;;     match = s->window + cur_match;
365 ;;;     if (*(ushf*)(match+best_len-1) != scan_end ||
366 ;;;         *(ushf*)match != scan_start) continue;
367 ;;;     [...]
368 ;;; } while ((cur_match = prev[cur_match & wmask]) > limit
369 ;;;          && --chain_length != 0);
370 ;;;
371 ;;; Here is the inner loop of the function. The function will spend the
372 ;;; majority of its time in this loop, and majority of that time will
373 ;;; be spent in the first ten instructions.
374 ;;;
375 ;;; Within this loop:
376 ;;; ebx = scanend
377 ;;; r8d = curmatch
378 ;;; edx = chainlenwmask - i.e., ((chainlen << 16) | wmask)
379 ;;; esi = windowbestlen - i.e., (window + bestlen)
380 ;;; edi = prev
381 ;;; ebp = limit
382 */
383 .balign 16
384 LookupLoop:
385         and r8d, edx
386 
387         movzx   r8d, word ptr [rdi + r8*2]
388         cmp r8d, ebp
389 		BEFORE_JMP
390         jbe LeaveNow
391 		AFTER_JMP
392         sub edx, 0x00010000
393 		BEFORE_JMP
394         js  LeaveNow
395 		AFTER_JMP
396 
397 LoopEntry:
398 
399         cmp bx,word ptr [rsi + r8 - 1]
400 		BEFORE_JMP
401         jnz LookupLoop1
402 		AFTER_JMP
403 LookupLoopIsZero:
404         cmp     r12w, word ptr [r10 + r8]
405 		BEFORE_JMP
406         jnz LookupLoop1
407 		AFTER_JMP
408 
409 
410 //;;; Store the current value of chainlen.
411         mov [chainlenwmask], edx
412 /*
413 ;;; Point edi to the string under scrutiny, and esi to the string we
414 ;;; are hoping to match it up with. In actuality, esi and edi are
415 ;;; both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and edx is
416 ;;; initialized to -(MAX_MATCH_8 - scanalign).
417 */
418         lea rsi,[r8+r10]
419         mov rdx, 0xfffffffffffffef8 //; -(MAX_MATCH_8)
420         lea rsi, [rsi + r13 + 0x0108] //;MAX_MATCH_8]
421         lea rdi, [r9 + r13 + 0x0108] //;MAX_MATCH_8]
422 
423         prefetcht1 [rsi+rdx]
424         prefetcht1 [rdi+rdx]
425 
426 /*
427 ;;; Test the strings for equality, 8 bytes at a time. At the end,
428 ;;; adjust rdx so that it is offset to the exact byte that mismatched.
429 ;;;
430 ;;; We already know at this point that the first three bytes of the
431 ;;; strings match each other, and they can be safely passed over before
432 ;;; starting the compare loop. So what this code does is skip over 0-3
433 ;;; bytes, as much as necessary in order to dword-align the edi
434 ;;; pointer. (rsi will still be misaligned three times out of four.)
435 ;;;
436 ;;; It should be confessed that this loop usually does not represent
437 ;;; much of the total running time. Replacing it with a more
438 ;;; straightforward "rep cmpsb" would not drastically degrade
439 ;;; performance.
440 */
441 
442 LoopCmps:
443         mov rax, [rsi + rdx]
444         xor rax, [rdi + rdx]
445         jnz LeaveLoopCmps
446 
447         mov rax, [rsi + rdx + 8]
448         xor rax, [rdi + rdx + 8]
449         jnz LeaveLoopCmps8
450 
451 
452         mov rax, [rsi + rdx + 8+8]
453         xor rax, [rdi + rdx + 8+8]
454         jnz LeaveLoopCmps16
455 
456         add rdx,8+8+8
457 
458 		BEFORE_JMP
459         jnz  LoopCmps
460         jmp  LenMaximum
461 		AFTER_JMP
462 
463 LeaveLoopCmps16: add rdx,8
464 LeaveLoopCmps8: add rdx,8
465 LeaveLoopCmps:
466 
467         test    eax, 0x0000FFFF
468         jnz LenLower
469 
470         test eax,0xffffffff
471 
472         jnz LenLower32
473 
474         add rdx,4
475         shr rax,32
476         or ax,ax
477 		BEFORE_JMP
478         jnz LenLower
479 		AFTER_JMP
480 
481 LenLower32:
482         shr eax,16
483         add rdx,2
484 
485 LenLower:
486         sub al, 1
487         adc rdx, 0
488 //;;; Calculate the length of the match. If it is longer than MAX_MATCH,
489 //;;; then automatically accept it as the best possible match and leave.
490 
491         lea rax, [rdi + rdx]
492         sub rax, r9
493         cmp eax, MAX_MATCH
494 		BEFORE_JMP
495         jge LenMaximum
496 		AFTER_JMP
497 /*
498 ;;; If the length of the match is not longer than the best match we
499 ;;; have so far, then forget it and return to the lookup loop.
500 ;///////////////////////////////////
501 */
502         cmp eax, r11d
503         jg  LongerMatch
504 
505         lea rsi,[r10+r11]
506 
507         mov rdi, prev_ad
508         mov edx, [chainlenwmask]
509 		BEFORE_JMP
510         jmp LookupLoop
511 		AFTER_JMP
512 /*
513 ;;;         s->match_start = cur_match;
514 ;;;         best_len = len;
515 ;;;         if (len >= nice_match) break;
516 ;;;         scan_end = *(ushf*)(scan+best_len-1);
517 */
518 LongerMatch:
519         mov r11d, eax
520         mov match_start, r8d
521         cmp eax, [nicematch]
522 		BEFORE_JMP
523         jge LeaveNow
524 		AFTER_JMP
525 
526         lea rsi,[r10+rax]
527 
528         movzx   ebx, word ptr [r9 + rax - 1]
529         mov rdi, prev_ad
530         mov edx, [chainlenwmask]
531 		BEFORE_JMP
532         jmp LookupLoop
533 		AFTER_JMP
534 
535 //;;; Accept the current string, with the maximum possible length.
536 
537 LenMaximum:
538         mov r11d,MAX_MATCH
539         mov match_start, r8d
540 
541 //;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len;
542 //;;; return s->lookahead;
543 
544 LeaveNow:
545         mov eax, Lookahead
546         cmp r11d, eax
547         cmovng eax, r11d
548 
549 
550 
551 //;;; Restore the stack and return from whence we came.
552 
553 
554 //        mov rsi,[save_rsi]
555 //        mov rdi,[save_rdi]
556         mov rbx,[save_rbx]
557         mov rbp,[save_rbp]
558         mov r12,[save_r12]
559         mov r13,[save_r13]
560         mov r14,[save_r14]
561         mov r15,[save_r15]
562 
563 
564         ret 0
565 //; please don't remove this string !
566 //; Your can freely use gvmat64 in any free or commercial app
567 //; but it is far better don't remove the string in the binary!
568  //   db     0dh,0ah,"asm686 with masm, optimised assembly code from Brian Raiter, written 1998, converted to amd 64 by Gilles Vollant 2005",0dh,0ah,0
569 
570 
571 match_init:
572   ret 0
573 
574 
575