1/*
2  Copyright (c) 1990-2007 Info-ZIP.  All rights reserved.
3
4  See the accompanying file LICENSE, version 2000-Apr-09 or later
5  (the contents of which are also included in zip.h) for terms of use.
6  If, for some reason, all these files are missing, the Info-ZIP license
7  also may be found at:  ftp://ftp.info-zip.org/pub/infozip/license.html
8*/
9/*
10 * crc_i386.S, optimized CRC calculation function for Zip and UnZip,
11 * created by Paul Kienitz and Christian Spieler.  Last revised 07 Jan 2007.
12 *
13 * GRR 961110:  incorporated Scott Field optimizations from win32/crc_i386.asm
14 *              => overall 6% speedup in "unzip -tq" on 9MB zipfile (486-66)
15 *
16 * SPC 970402:  revised for Rodney Brown's optimizations (32-bit-wide
17 *              aligned reads for most of the data from buffer), can be
18 *              disabled by defining the macro NO_32_BIT_LOADS
19 *
20 * SPC 971012:  added Rodney Brown's additional tweaks for 32-bit-optimized
21 *              CPUs (like the Pentium Pro, Pentium II, and probably some
22 *              Pentium clones). This optimization is controlled by the
23 *              preprocessor switch "__686" and is disabled by default.
24 *              (This default is based on the assumption that most users
25 *              do not yet work on a Pentium Pro or Pentium II machine ...)
26 *
27 * COS 050116:  Enabled the 686 build by default, because there are hardly any
28 *              pre-686 CPUs in serious use nowadays. (See SPC 970402 above.)
29 *
30 * SPC 060103:  Updated code to incorporate newer optimizations found in zlib.
31 *
32 * SPC 070107:  Added conditional switch to deactivate crc32() compilation.
33 *
34 * FLAT memory model assumed.  Calling interface:
35 *   - args are pushed onto the stack from right to left,
36 *   - return value is given in the EAX register,
37 *   - all other registers (with exception of EFLAGS) are preserved. (With
38 *     GNU C 2.7.x, %edx and %ecx are `scratch' registers, but preserving
39 *     them nevertheless adds only 4 single byte instructions.)
40 *
41 * This source generates the function
42 * ulg crc32(ulg crc, ZCONST uch *buf, extent len).
43 *
44 * Loop unrolling can be disabled by defining the macro NO_UNROLLED_LOOPS.
45 * This results in shorter code at the expense of reduced performance.
46 */
47
48/* This file is NOT used in conjunction with zlib, or when only creation of
49 * the basic CRC_32_Table (for other purpose) is requested.
50 */
51#if !defined(USE_ZLIB) && !defined(CRC_TABLE_ONLY)
52
53/* Preprocess with -DNO_UNDERLINE if your C compiler does not prefix
54 * external symbols with an underline character '_'.
55 */
56#if defined(NO_UNDERLINE) || defined(__ELF__)
57#  define _crc32            crc32
58#  define _get_crc_table    get_crc_table
59#endif
60/* Use 16-byte alignment if your assembler supports it. Warning: gas
61 * uses a log(x) parameter (.align 4 means 16-byte alignment). On SVR4
62 * the parameter is a number of bytes.
63 */
64#ifndef ALIGNMENT
65#  define ALIGNMENT .align 4,0x90
66#endif
67
68#if defined(i386) || defined(_i386) || defined(_I386) || defined(__i386)
69
70/* This version is for 386 Unix, OS/2, MSDOS in 32 bit mode (gcc & gas).
71 * Warning: it uses the AT&T syntax: mov source,dest
72 * This file is only optional. If you want to use the C version,
73 * remove -DASM_CRC from CFLAGS in Makefile and set OBJA to an empty string.
74 */
75
76                .file   "crc_i386.S"
77
78#if !defined(PRE_686) && !defined(__686)
79   /* Optimize for Pentium Pro and compatible CPUs by default. */
80#  define __686
81#endif
82
83#if defined(NO_STD_STACKFRAME) && defined(USE_STD_STACKFRAME)
84#  undef USE_STACKFRAME
85#else
86   /* The default is to use standard stack frame entry, because it
87    * results in smaller code!
88    */
89#  ifndef USE_STD_STACKFRAME
90#    define USE_STD_STACKFRAME
91#  endif
92#endif
93
94#ifdef USE_STD_STACKFRAME
95#  define _STD_ENTRY    pushl   %ebp ; movl   %esp,%ebp
96#  define arg1  8(%ebp)
97#  define arg2  12(%ebp)
98#  define arg3  16(%ebp)
99#  define _STD_LEAVE    popl    %ebp
100#else /* !USE_STD_STACKFRAME */
101#  define _STD_ENTRY
102#  define arg1  24(%esp)
103#  define arg2  28(%esp)
104#  define arg3  32(%esp)
105#  define _STD_LEAVE
106#endif /* ?USE_STD_STACKFRAME */
107
108/*
109 * These two (three) macros make up the loop body of the CRC32 cruncher.
110 * registers modified:
111 *   eax  : crc value "c"
112 *   esi  : pointer to next data byte (or lword) "buf++"
113 * registers read:
114 *   edi  : pointer to base of crc_table array
115 * scratch registers:
116 *   ebx  : index into crc_table array
117 *          (requires upper three bytes = 0 when __686 is undefined)
118 */
119#ifndef __686   /* optimize for 386, 486, Pentium */
120#define Do_CRC          /* c = (c >> 8) ^ table[c & 0xFF] */\
121                movb    %al, %bl                ;/* tmp = c & 0xFF  */\
122                shrl    $8, %eax                ;/* c = (c >> 8)    */\
123                xorl    (%edi, %ebx, 4), %eax   ;/* c ^= table[tmp] */
124#else   /* __686 : optimize for Pentium Pro and compatible CPUs */
125#define Do_CRC          /* c = (c >> 8) ^ table[c & 0xFF] */\
126                movzbl  %al, %ebx               ;/* tmp = c & 0xFF  */\
127                shrl    $8, %eax                ;/* c = (c >> 8)    */\
128                xorl    (%edi, %ebx, 4), %eax   ;/* c ^=table[tmp]  */
129#endif  /* ?__686 */
130
131#define Do_CRC_byte             /* c = (c >> 8) ^ table[(c^*buf++)&0xFF] */\
132                xorb    (%esi), %al     ;/* c ^= *buf  */\
133                incl    %esi            ;/* buf++      */\
134                Do_CRC
135
136#define Do_CRC_byteof(ofs)      /* c = (c >> 8) ^ table[(c^*buf++)&0xFF] */\
137                xorb    ofs(%esi), %al  ;/* c ^= *buf  */\
138                incl    %esi            ;/* buf++      */\
139                Do_CRC
140
141#ifndef  NO_32_BIT_LOADS
142# ifdef IZ_CRCOPTIM_UNFOLDTBL
143   /* the edx register is needed in crc calculation */
144#  define SavLen arg3
145#  define UpdCRC_lword \
146                movzbl  %al, %ebx               ; \
147                movl    3072(%edi,%ebx,4), %edx ; \
148                movzbl  %ah, %ebx               ; \
149                shrl    $16, %eax               ; \
150                xor     2048(%edi,%ebx,4), %edx ; \
151                movzbl  %al, %ebx               ; \
152                shrl    $8,%eax                 ; \
153                xorl    1024(%edi,%ebx,4), %edx ; \
154                movl    (%edi,%eax,4), %eax     ; \
155                xorl    %edx,%eax               ;
156#  define UpdCRC_lword_sh(dwPtrIncr) \
157                movzbl  %al, %ebx               ; \
158                movl    3072(%edi,%ebx,4), %edx ; \
159                movzbl  %ah, %ebx               ; \
160                shrl    $16, %eax               ; \
161                xor     2048(%edi,%ebx,4), %edx ; \
162                movzbl  %al, %ebx               ; \
163                addl    $4*(dwPtrIncr), %esi    ;/* ((ulg *)buf)+=dwPtrIncr */\
164                shrl    $8,%eax                 ; \
165                xorl    1024(%edi,%ebx,4), %edx ; \
166                movl    (%edi,%eax,4),%eax      ; \
167                xorl    %edx,%eax               ;
168# else /* !IZ_CRCOPTIM_UNFOLDTBL */
169   /* the edx register is not needed anywhere else */
170#  define SavLen %edx
171#  define UpdCRC_lword \
172                Do_CRC \
173                Do_CRC \
174                Do_CRC \
175                Do_CRC
176#  define UpdCRC_lword_sh(dwPtrIncr) \
177                Do_CRC \
178                Do_CRC \
179                addl    $4*(dwPtrIncr), %esi    ;/* ((ulg *)buf)++   */\
180                Do_CRC \
181                Do_CRC
182# endif /* ?IZ_CRCOPTIM_UNFOLDTBL */
183#define Do_CRC_lword \
184                xorl    (%esi), %eax           ;/* c ^= *(ulg *)buf */\
185                UpdCRC_lword_sh(1)              /* ... ((ulg *)buf)++ */
186#define Do_CRC_4lword \
187                xorl    (%esi), %eax           ;/* c ^= *(ulg *)buf */\
188                UpdCRC_lword \
189                xorl    4(%esi), %eax          ;/* c ^= *((ulg *)buf+1) */\
190                UpdCRC_lword \
191                xorl    8(%esi), %eax          ;/* c ^= *((ulg *)buf+2) */\
192                UpdCRC_lword \
193                xorl    12(%esi), %eax         ;/* c ^= *((ulg *)buf]+3 */\
194                UpdCRC_lword_sh(4)              /* ... ((ulg *)buf)+=4 */
195#endif  /* !NO_32_BIT_LOADS */
196
197
198                .text
199
200                .globl  _crc32
201
202_crc32:                         /* ulg crc32(ulg crc, uch *buf, extent len) */
203                _STD_ENTRY
204                pushl   %edi
205                pushl   %esi
206                pushl   %ebx
207                pushl   %edx
208                pushl   %ecx
209
210                movl    arg2, %esi           /* 2nd arg: uch *buf            */
211                subl    %eax, %eax           /* > if (!buf)                  */
212                testl   %esi, %esi           /* >   return 0;                */
213                jz      .L_fine              /* > else {                     */
214                call    _get_crc_table
215                movl    %eax, %edi
216                movl    arg1, %eax           /* 1st arg: ulg crc             */
217#ifndef __686
218                subl    %ebx, %ebx           /* ebx=0; bl usable as dword    */
219#endif
220                movl    arg3, %ecx           /* 3rd arg: extent len          */
221                notl    %eax                 /* >   c = ~crc;                */
222
223                testl   %ecx, %ecx
224#ifndef  NO_UNROLLED_LOOPS
225                jz      .L_bail
226#  ifndef  NO_32_BIT_LOADS
227                /* Assert now have positive length */
228.L_align_loop:
229                testl   $3, %esi        /* Align buf on lword boundary */
230                jz      .L_aligned_now
231                Do_CRC_byte
232                decl    %ecx
233                jnz     .L_align_loop
234.L_aligned_now:
235#  endif  /* !NO_32_BIT_LOADS */
236                movl    %ecx, SavLen         /* save current value of len */
237                shrl    $4, %ecx             /* ecx = len / 16   */
238                jz      .L_No_Sixteens
239/*  align loop head at start of 486 internal cache line !! */
240                ALIGNMENT
241.L_Next_Sixteen:
242#  ifndef NO_32_BIT_LOADS
243                 Do_CRC_4lword
244#  else   /* NO_32_BIT_LOADS */
245                 Do_CRC_byteof(0)
246                 Do_CRC_byteof(1)
247                 Do_CRC_byteof(2)
248                 Do_CRC_byteof(3)
249                 Do_CRC_byteof(4)
250                 Do_CRC_byteof(5)
251                 Do_CRC_byteof(6)
252                 Do_CRC_byteof(7)
253                 Do_CRC_byteof(8)
254                 Do_CRC_byteof(9)
255                 Do_CRC_byteof(10)
256                 Do_CRC_byteof(11)
257                 Do_CRC_byteof(12)
258                 Do_CRC_byteof(13)
259                 Do_CRC_byteof(14)
260                 Do_CRC_byteof(15)
261                 addl    $16,%esi        ;/* buf += 16 */
262#  endif  /* ?NO_32_BIT_LOADS */
263                decl    %ecx
264                jnz     .L_Next_Sixteen
265
266.L_No_Sixteens:
267                movl    SavLen, %ecx
268                andl    $15, %ecx         /* ecx = len % 16   */
269# ifndef NO_32_BIT_LOADS
270                shrl    $2,%ecx           /* ecx = len / 4    */
271                jz      .L_No_Fours
272.L_Next_Four:
273                Do_CRC_lword
274                decl    %ecx
275                jnz     .L_Next_Four
276.L_No_Fours:
277                movl    SavLen,%ecx
278                andl    $3,%ecx          /* ecx = len % 4 */
279# endif /* !NO_32_BIT_LOADS */
280#endif /* !NO_UNROLLED_LOOPS */
281                jz      .L_bail          /* > if (len)                       */
282/* align loop head at start of 486 internal cache line !! */
283                ALIGNMENT
284.L_loupe:                                /* >   do {                         */
285                 Do_CRC_byte             /*       c = CRC32(c,*buf++,crctab);*/
286                decl    %ecx             /* >   } while (--len);             */
287                jnz     .L_loupe
288
289.L_bail:                                 /* > }                              */
290                notl    %eax             /* > return ~c;                     */
291.L_fine:
292                popl    %ecx
293                popl    %edx
294                popl    %ebx
295                popl    %esi
296                popl    %edi
297                _STD_LEAVE
298                ret
299
300#else
301 error: this asm version is for 386 only
302#endif /* i386 || _i386 || _I386 || __i386 */
303
304#endif /* !USE_ZLIB && !CRC_TABLE_ONLY */
305