Develop and Download Open Source Software

プロジェクト・ナニ屋

Browse Subversion Repository

/[naniya]/branches/mty-makai/crypt64.S

Contents of /branches/mty-makai/crypt64.S

Parent Directory Parent Directory | Revision Log Revision Log

Revision 192 - (show annotations) (download)
Mon May 26 06:29:35 2008 UTC (15 years, 10 months ago) by notanpe
File size: 7501 byte(s) 
シーズン 2 に向けて
1 /******************************************************-*-fundamental-*-
2 *
3 * $Id$
4 */
5
6 #ifndef DEBUG
7 #define DEBUG 0
8 #endif
9
10 #if defined(USE_MMX)
11
12 #define USE_T 1
13
14 .if !USE_T // to suppress code generated
15 #include "x86-mmx.S"
16 .endif
17 #undef K
18
19 #define PTR_T .long
20
21 #define W 8
22 #define O 1
23 #define R(n) %mm##n
24 #define MOV movq
25 #define AND pand
26 #define ANDN pandn
27 #define OR por
28 #define XOR pxor
29 #define OX 0x00
30
31 #define AX %eax
32 #define CNT %ax
33 #define KSI %dl
34 #define PARAM %edx
35 #define TX %ebx
36 #define K %ecx
37 #define T /* +16 */ %esi
38 #define LR /* +16 */ %ebp
39 #define RL /* +16 */ %edi
40 #define H %ecx
41
42 #elif defined(USE_64)
43
44 #define USE_T 0
45
46 #define PTR_T .quad
47
48 #define W 8
49 #define O 1
50 #define R(n) %r##n
51 #define MOV mov
52 #define NOT not
53 #define AND and
54 #define OR or
55 #define XOR xor
56 #define OX 0x00
57
58 #define AX %rax
59 #define CNT 4(%rsp)
60 #define KSI 6(%rsp)
61 #define PARAM %rdx
62 #define TX %rdx
63 #define K %rcx
64 #define T /* +16 */ %rsi
65 #define LR /* +16 */ %rbp
66 #define RL /* +16 */ %rdi
67 #define H %rcx
68
69 #define WK(n) W*(n+16+8)(%rsp)
70
71 #define r4 rax
72 #define r5 rbx
73 #define r6 rdx
74 #define r7 rsi
75 #define r0 r8
76 #define r1 r9
77 #define r2 r10
78 #define r3 r11
79
80 #elif defined(USE_64_XMM)
81
82 #define USE_T 1
83
84 .if !USE_T // to suppress code generated
85 //#include "x64-sse.S"
86 .endif
87 #undef K
88
89 #define PTR_T .quad
90
91 #define W 16
92 #define O 4
93 #define R(n) %xmm##n
94 #define MOV movdqa
95 #define AND pand
96 #define ANDN pandn
97 #define OR por
98 #define XOR pxor
99 /* disp8 ���������������������Adisp32 �����������K�v���� */
100 #define OX 0x100000
101
102 #define AX %rax
103 #define CNT %ax
104 #define KSI %dl
105 #define PARAM %rdx
106 #define TX %rbx
107 #define K %rcx
108 #define T /* +16 */ %rsi
109 #define LR /* +16 */ %rbp
110 #define RL /* +16 */ %rdi
111 #define H %rcx
112
113 #else /* XMM */
114
115 #define USE_T 1 /* TLS �����������������c */
116
117 .if 0 // to suppress code generated
118 #include "x86-sse.S"
119 .endif
120 #undef K
121 #undef tmp_at
122 #ifdef USE_REGPARM
123 #define tmp_at(n) param64+W*(((n+16)&31)+64)
124 #else /* USE_REGPARM */
125 #define tmp_at(n) _param64+W*(((n+16)&31)+64)
126 #endif /* USE_REGPARM */
127
128 #define PTR_T .long
129
130 #define W 16
131 #define O 4
132 #define R(n) %xmm##n
133 #define MOV movdqa
134 #define AND pand
135 #define ANDN pandn
136 #define OR por
137 #define XOR pxor
138 /* disp8 ���������������������Adisp32 �����������K�v���� */
139 #define OX 0x100000
140
141 #define AX %eax
142 #define CNT %ax
143 #define KSI %dl
144 #define PARAM %edx
145 #define TX %ebx
146 #define K %ecx
147 #define T /* +16 */ %esi
148 #define LR /* +16 */ %ebp
149 #define RL /* +16 */ %edi
150 #define H %ecx
151
152 #endif
153
154 #ifndef WK
155 #if USE_T
156 #define WK(n) W*(n)(T)
157 #endif
158 #endif
159
160 #define LOADI(v,n,d,s) MOV W*((s)-16)+OX(LR),R(d);Ln##n:
161 #define XORK(v,n,d,s) XOR W*s(K),R(d)
162
163 #define RMOVE(v,n,d,s) MOV R(s),R(d)
164 #define RANDN(v,n,d,s) ANDN R(s),R(d)
165 #define RAND(v,n,d,s) AND R(s),R(d)
166 #define ROR(v,n,d,s) OR R(s),R(d)
167 #define RXOR(v,n,d,s) XOR R(s),R(d)
168
169 #define NOP
170 #define RXORN RXOR
171 #define MXORN MXOR
172
173 #define XORR(v,n,d,s) XOR W*((s)-16)(RL),R(d)
174 #define STORR(v,n,d,s) MOV R(d),W*((s)-16)(RL)
175
176 #define MXORR(v,n,d,s) XOR R(d),W*((s)-16)(RL)
177
178 #define STORE(v,n,d,o) MOV R(d),WK((o)-16)
179
180 #define MMOVE(v,n,d,o) MOV WK((o)-16),R(d)
181 #define MXOR(v,n,d,o) XOR WK((o)-16),R(d)
182 #define MOR(v,n,d,o) OR WK((o)-16),R(d)
183 #define MANDN(v,n,d,o) ANDN WK((o)-16),R(d)
184 #define MAND(v,n,d,o) AND WK((o)-16),R(d)
185
186 #ifdef NOT
187 #define RNOT(v,n,d,x) not R(d)
188 #else
189 #define RNOT(v,n,d,x) XOR (T),R(d)
190 #endif
191
192 // (ecx, edx)
193
194 // ecx k
195 // edx t+0
196 // esi t+16
197 // ebp lr+16
198 // edi rl+16
199
200 .text
201 .align 16
202 .globl crypt64_desc
203 crypt64_desc:
204 .globl _crypt64_desc
205 _crypt64_desc:
206 PTR_T crypt64_pro
207 PTR_T crypt64_crypt
208 PTR_T crypt64_cmp_pro
209 PTR_T crypt64_cmp_ep
210 PTR_T crypt64_ep
211 PTR_T crypt64_ep_end
212 .byte 16/W
213 .byte O
214 .word (loe-los)/(2*48)
215
216 /* disp(%reg) disp �����|�C���^ */
217 #define LOF(n) (Ln##n-crypt64_crypt-O)
218 los:
219 .word LOF(00),LOF(10),LOF(20),LOF(30),LOF(40),LOF(50)
220 .word LOF(01),LOF(11),LOF(21),LOF(31),LOF(41),LOF(51)
221 .word LOF(02),LOF(12),LOF(22),LOF(32),LOF(42),LOF(52)
222 .word LOF(03),LOF(13),LOF(23),LOF(33),LOF(43),LOF(53)
223 .word LOF(04),LOF(14),LOF(24),LOF(34),LOF(44),LOF(54)
224 .word LOF(05),LOF(15),LOF(25),LOF(35),LOF(45),LOF(55)
225 .word LOF(06),LOF(16),LOF(26),LOF(36),LOF(46),LOF(56)
226 .word LOF(07),LOF(17),LOF(27),LOF(37),LOF(47),LOF(57)
227 loe:
228 .align 64
229
230 .globl _crypt64
231 _crypt64:
232 crypt64_pro:
233 push LR
234 push RL
235 push TX
236 push T
237 #ifdef USE_64
238 push %rbx
239 push %r12
240 push %r13
241 push %r14
242 push %r15
243 sub $0x110,%rsp
244 #endif
245
246 #if 0&&DEBUG>=1
247 rdtsc
248 #endif
249 push AX // t0 ���������_�~�[
250
251 crypt64_crypt:
252
253 // �����l�N���A
254 // �����h������������������ memset(3) �g�������S����(��)
255 XOR R(7),R(7)
256 mov $-W*64,AX
257 clr_lr:
258 MOV R(7),W*64(PARAM,AX)
259 add $W,AX
260 jne clr_lr
261
262 // �����������t���n�����B���[�����������B
263 lea W*(16)(PARAM),LR /* XXX */
264 lea W*(32+16)(PARAM),RL
265 #ifndef USE_64
266 lea W*(64+16)(PARAM),T
267 #endif
268
269 movb $0,KSI
270 movw $16*25,CNT
271
272 .align 64
273 loo:
274 add W*2*56(K),K
275
276 #ifdef S1
277
278 #if USE_T
279 #undef tmp_at
280 #define tmp_at(n) W*(n)(T)
281 #endif
282
283 #undef pnot
284 #define pnot (T)
285
286 #ifdef a6_p
287 #define S_1(o1,o2,o3,o4) S1(o1,o2,o3,o4,)
288 #define S_2(o1,o2,o3,o4) S2(o1,o2,o3,o4,)
289 #define S_3(o1,o2,o3,o4) S3(o1,o2,o3,o4,)
290 #define S_4(o1,o2,o3,o4) S4(o1,o2,o3,o4,)
291 #define S_5(o1,o2,o3,o4) S5(o1,o2,o3,o4,)
292 #define S_6(o1,o2,o3,o4) S6(o1,o2,o3,o4,)
293 #define S_7(o1,o2,o3,o4) S7(o1,o2,o3,o4,)
294 #define S_8(o1,o2,o3,o4) S8(o1,o2,o3,o4,)
295 #else
296 #define S_1(o1,o2,o3,o4) S1(o1,o2,o3,o4)
297 #define S_2(o1,o2,o3,o4) S2(o1,o2,o3,o4)
298 #define S_3(o1,o2,o3,o4) S3(o1,o2,o3,o4)
299 #define S_4(o1,o2,o3,o4) S4(o1,o2,o3,o4)
300 #define S_5(o1,o2,o3,o4) S5(o1,o2,o3,o4)
301 #define S_6(o1,o2,o3,o4) S6(o1,o2,o3,o4)
302 #define S_7(o1,o2,o3,o4) S7(o1,o2,o3,o4)
303 #define S_8(o1,o2,o3,o4) S8(o1,o2,o3,o4)
304 #endif
305
306 #undef B
307 #define B(n) W*(n-16)(RL)
308
309 #define EK(e,k1,k2,k3,k4,k5,k6) \
310 LOADI(1,0##e,0,(31&(4*(e)-1))); XORK(1,0##e,0,k1); \
311 LOADI(2,1##e,1,(31&(4*(e)+0))); XORK(2,1##e,1,k2); \
312 LOADI(3,2##e,2,(31&(4*(e)+1))); XORK(3,2##e,2,k3); \
313 LOADI(4,3##e,3,(31&(4*(e)+2))); XORK(4,3##e,3,k4); \
314 LOADI(5,4##e,4,(31&(4*(e)+3))); XORK(5,4##e,4,k5); \
315 LOADI(6,5##e,5,(31&(4*(e)+4))); XORK(6,5##e,5,k6)
316
317 /* use John */
318 EK(0,13,16,10,23,0,4);S_1(B(8),B(16),B(22),B(30))
319 EK(1,2,27,14,5,20,9);S_2(B(12),B(27),B(1),B(17))
320 EK(2,22,18,11,3,25,7);S_3(B(23),B(15),B(29),B(5))
321 EK(3,15,6,26,19,12,1);S_4(B(25),B(19),B(9),B(0))
322 EK(4,68,79,58,64,74,82);S_5(B(7),B(13),B(24),B(2))
323 EK(5,57,67,78,72,60,75);S_6(B(3),B(28),B(10),B(18))
324 EK(6,71,76,66,83,61,80);S_7(B(31),B(11),B(21),B(6))
325 EK(7,73,69,77,63,56,59);S_8(B(4),B(26),B(14),B(20))
326
327 #elif defined(USE_64_XMM)
328 #include "x64-xmm.inc"
329 #elif defined(USE_64)
330 #include "x64-alu.inc"
331 #else
332 #include "crypt64.inc"
333 #endif
334
335 mov LR,TX
336 addb $(256/16),KSI
337 cmovne RL,LR
338 cmovne TX,RL
339 decw CNT
340 jne loo
341
342 // ���i���������|�C���^����������
343 lea W*(32+16)(LR),H
344
345 // �������������������}��
346 // �������������R�[�h���W�J���������L���`���������������B
347 lea W*(-16)(RL),PARAM
348 lea W*(-16)(H),T
349
350 crypt64_cmp_pro:
351
352 #if 0&&DEBUG>=1
353 // ���r�����v��
354 rdtsc
355 push AX // t1
356 #endif
357 // �������������������v�������������A���������B
358 lea W*(16)(PARAM),LR
359 lea W*(32+16)(PARAM),RL
360 lea W*(64+16)(PARAM),T
361
362 // ���r��������������������������
363
364 crypt64_cmp_ep:
365
366 crypt64_ep:
367
368 pop LR // t1 ���������_�~�[
369
370 #if 0&&DEBUG>=1
371 // �v�����W�v
372 rdtsc // t2
373 mov AX,PARAM
374 sub LR,PARAM // t2 - t1
375 mov LR,AX
376 pop LR // t0
377 sub LR,AX // t1 - t0
378 #if defined(USE_64) || defined(USE_64_XMM)
379 shl $32,PARAM
380 and $0xFFFFFFFF,%rax
381 or PARAM,%rax
382 #endif
383 #endif
384
385 #ifdef USE_64
386 add $0x110,%rsp
387 pop %r15
388 pop %r14
389 pop %r13
390 pop %r12
391 pop %rbx
392 #endif
393 pop T
394 pop TX
395 pop RL
396 pop LR
397
398 ret
399
400 crypt64_ep_end:
401
402 //EOF

Properties

Name Value
svn:eol-style native
svn:keywords Author Date Id Rev URL
Back to OSDN">Back to OSDN
 ViewVC Help
Powered by ViewVC 1.1.26  
About OSDN
Find Software
Develop Software
Help