72ca9705118383335aa94c523f0d3bb2cce03714
[openssl.git] / crypto / des / asm / des-586.pl
1 #!/usr/local/bin/perl
2 #
3 # The inner loop instruction sequence and the IP/FP modifications are from
4 # Svend Olaf Mikkelsen <svolaf@inet.uni-c.dk>
5 #
6
7 push(@INC,"perlasm","../../perlasm");
8 require "x86asm.pl";
9 require "cbc.pl";
10 require "desboth.pl";
11
12 # base code is in microsft
13 # op dest, source
14 # format.
15 #
16
17 &asm_init($ARGV[0],"des-586.pl");
18
19 $L="edi";
20 $R="esi";
21
22 &external_label("DES_SPtrans");
23 &DES_encrypt("DES_encrypt1",1);
24 &DES_encrypt("DES_encrypt2",0);
25 &DES_encrypt3("DES_encrypt3",1);
26 &DES_encrypt3("DES_decrypt3",0);
27 &cbc("DES_ncbc_encrypt","DES_encrypt1","DES_encrypt1",0,4,5,3,5,-1);
28 &cbc("DES_ede3_cbc_encrypt","DES_encrypt3","DES_decrypt3",0,6,7,3,4,5);
29 &DES_SPtrans();
30
31 &asm_finish();
32
33 sub DES_encrypt
34         {
35         local($name,$do_ip)=@_;
36
37         &function_begin_B($name,"EXTRN   _DES_SPtrans:DWORD");
38
39         &push("esi");
40         &push("edi");
41
42         &comment("");
43         &comment("Load the 2 words");
44         $trans="ebp";
45
46         if ($do_ip)
47                 {
48                 &mov($R,&wparam(0));
49                  &xor(  "ecx",          "ecx"           );
50
51                 &push("ebx");
52                 &push("ebp");
53
54                 &mov("eax",&DWP(0,$R,"",0));
55                  &mov("ebx",&wparam(2));        # get encrypt flag
56                 &mov($L,&DWP(4,$R,"",0));
57                 &comment("");
58                 &comment("IP");
59                 &IP_new("eax",$L,$R,3);
60                 }
61         else
62                 {
63                 &mov("eax",&wparam(0));
64                  &xor(  "ecx",          "ecx"           );
65
66                 &push("ebx");
67                 &push("ebp");
68
69                 &mov($R,&DWP(0,"eax","",0));
70                  &mov("ebx",&wparam(2));        # get encrypt flag
71                 &rotl($R,3);
72                 &mov($L,&DWP(4,"eax","",0));
73                 &rotl($L,3);
74                 }
75
76         # PIC-ification:-)
77         &picmeup($trans,"DES_SPtrans");
78         #if ($cpp)      { &picmeup($trans,"DES_SPtrans");   }
79         #else           { &lea($trans,&DWP("DES_SPtrans")); }
80
81         &mov(   "ecx",  &wparam(1)      );
82         &cmp("ebx","0");
83         &je(&label("start_decrypt"));
84
85         for ($i=0; $i<16; $i+=2)
86                 {
87                 &comment("");
88                 &comment("Round $i");
89                 &D_ENCRYPT($i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx");
90
91                 &comment("");
92                 &comment("Round ".sprintf("%d",$i+1));
93                 &D_ENCRYPT($i+1,$R,$L,($i+1)*2,$trans,"eax","ebx","ecx","edx");
94                 }
95         &jmp(&label("end"));
96
97         &set_label("start_decrypt");
98
99         for ($i=15; $i>0; $i-=2)
100                 {
101                 &comment("");
102                 &comment("Round $i");
103                 &D_ENCRYPT(15-$i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx");
104                 &comment("");
105                 &comment("Round ".sprintf("%d",$i-1));
106                 &D_ENCRYPT(15-$i+1,$R,$L,($i-1)*2,$trans,"eax","ebx","ecx","edx");
107                 }
108
109         &set_label("end");
110
111         if ($do_ip)
112                 {
113                 &comment("");
114                 &comment("FP");
115                 &mov("edx",&wparam(0));
116                 &FP_new($L,$R,"eax",3);
117
118                 &mov(&DWP(0,"edx","",0),"eax");
119                 &mov(&DWP(4,"edx","",0),$R);
120                 }
121         else
122                 {
123                 &comment("");
124                 &comment("Fixup");
125                 &rotr($L,3);            # r
126                  &mov("eax",&wparam(0));
127                 &rotr($R,3);            # l
128                  &mov(&DWP(0,"eax","",0),$L);
129                  &mov(&DWP(4,"eax","",0),$R);
130                 }
131
132         &pop("ebp");
133         &pop("ebx");
134         &pop("edi");
135         &pop("esi");
136         &ret();
137
138         &function_end_B($name);
139         }
140
141 sub D_ENCRYPT
142         {
143         local($r,$L,$R,$S,$trans,$u,$tmp1,$tmp2,$t)=@_;
144
145          &mov(  $u,             &DWP(&n2a($S*4),$tmp2,"",0));
146         &xor(   $tmp1,          $tmp1);
147          &mov(  $t,             &DWP(&n2a(($S+1)*4),$tmp2,"",0));
148         &xor(   $u,             $R);
149         &xor(   $tmp2,          $tmp2);
150          &xor(  $t,             $R);
151         &and(   $u,             "0xfcfcfcfc"    );
152          &and(  $t,             "0xcfcfcfcf"    );
153         &movb(  &LB($tmp1),     &LB($u) );
154          &movb( &LB($tmp2),     &HB($u) );
155         &rotr(  $t,             4               );
156         &xor(   $L,             &DWP("     ",$trans,$tmp1,0));
157          &movb( &LB($tmp1),     &LB($t) );
158          &xor(  $L,             &DWP("0x200",$trans,$tmp2,0));
159          &movb( &LB($tmp2),     &HB($t) );
160         &shr(   $u,             16);
161          &xor(  $L,             &DWP("0x100",$trans,$tmp1,0));
162          &movb( &LB($tmp1),     &HB($u) );
163         &shr(   $t,             16);
164          &xor(  $L,             &DWP("0x300",$trans,$tmp2,0));
165         &movb(  &LB($tmp2),     &HB($t) );
166          &and(  $u,             "0xff"  );
167         &and(   $t,             "0xff"  );
168          &xor(  $L,             &DWP("0x600",$trans,$tmp1,0));
169          &xor(  $L,             &DWP("0x700",$trans,$tmp2,0));
170         &mov(   $tmp2,          &wparam(1)      );
171          &xor(  $L,             &DWP("0x400",$trans,$u,0));
172          &xor(  $L,             &DWP("0x500",$trans,$t,0));
173         }
174
175 sub n2a
176         {
177         sprintf("%d",$_[0]);
178         }
179
180 # now has a side affect of rotating $a by $shift
181 sub R_PERM_OP
182         {
183         local($a,$b,$tt,$shift,$mask,$last)=@_;
184
185         &rotl(  $a,             $shift          ) if ($shift != 0);
186         &mov(   $tt,            $a              );
187         &xor(   $a,             $b              );
188         &and(   $a,             $mask           );
189         # This can never succeed, and besides it is difficult to see what the
190         # idea was - Ben 13 Feb 99
191         if (!$last eq $b)
192                 {
193                 &xor(   $b,             $a              );
194                 &xor(   $tt,            $a              );
195                 }
196         else
197                 {
198                 &xor(   $tt,            $a              );
199                 &xor(   $b,             $a              );
200                 }
201         &comment("");
202         }
203
204 sub IP_new
205         {
206         local($l,$r,$tt,$lr)=@_;
207
208         &R_PERM_OP($l,$r,$tt, 4,"0xf0f0f0f0",$l);
209         &R_PERM_OP($r,$tt,$l,20,"0xfff0000f",$l);
210         &R_PERM_OP($l,$tt,$r,14,"0x33333333",$r);
211         &R_PERM_OP($tt,$r,$l,22,"0x03fc03fc",$r);
212         &R_PERM_OP($l,$r,$tt, 9,"0xaaaaaaaa",$r);
213         
214         if ($lr != 3)
215                 {
216                 if (($lr-3) < 0)
217                         { &rotr($tt,    3-$lr); }
218                 else    { &rotl($tt,    $lr-3); }
219                 }
220         if ($lr != 2)
221                 {
222                 if (($lr-2) < 0)
223                         { &rotr($r,     2-$lr); }
224                 else    { &rotl($r,     $lr-2); }
225                 }
226         }
227
228 sub FP_new
229         {
230         local($l,$r,$tt,$lr)=@_;
231
232         if ($lr != 2)
233                 {
234                 if (($lr-2) < 0)
235                         { &rotl($r,     2-$lr); }
236                 else    { &rotr($r,     $lr-2); }
237                 }
238         if ($lr != 3)
239                 {
240                 if (($lr-3) < 0)
241                         { &rotl($l,     3-$lr); }
242                 else    { &rotr($l,     $lr-3); }
243                 }
244
245         &R_PERM_OP($l,$r,$tt, 0,"0xaaaaaaaa",$r);
246         &R_PERM_OP($tt,$r,$l,23,"0x03fc03fc",$r);
247         &R_PERM_OP($l,$r,$tt,10,"0x33333333",$l);
248         &R_PERM_OP($r,$tt,$l,18,"0xfff0000f",$l);
249         &R_PERM_OP($l,$tt,$r,12,"0xf0f0f0f0",$r);
250         &rotr($tt       , 4);
251         }
252
253 sub DES_SPtrans
254         {
255         &public_label("DES_SPtrans");
256         &set_label("DES_SPtrans",64);
257         &data_word(0x02080800, 0x00080000, 0x02000002, 0x02080802);
258         &data_word(0x02000000, 0x00080802, 0x00080002, 0x02000002);
259         &data_word(0x00080802, 0x02080800, 0x02080000, 0x00000802);
260         &data_word(0x02000802, 0x02000000, 0x00000000, 0x00080002);
261         &data_word(0x00080000, 0x00000002, 0x02000800, 0x00080800);
262         &data_word(0x02080802, 0x02080000, 0x00000802, 0x02000800);
263         &data_word(0x00000002, 0x00000800, 0x00080800, 0x02080002);
264         &data_word(0x00000800, 0x02000802, 0x02080002, 0x00000000);
265         &data_word(0x00000000, 0x02080802, 0x02000800, 0x00080002);
266         &data_word(0x02080800, 0x00080000, 0x00000802, 0x02000800);
267         &data_word(0x02080002, 0x00000800, 0x00080800, 0x02000002);
268         &data_word(0x00080802, 0x00000002, 0x02000002, 0x02080000);
269         &data_word(0x02080802, 0x00080800, 0x02080000, 0x02000802);
270         &data_word(0x02000000, 0x00000802, 0x00080002, 0x00000000);
271         &data_word(0x00080000, 0x02000000, 0x02000802, 0x02080800);
272         &data_word(0x00000002, 0x02080002, 0x00000800, 0x00080802);
273         # nibble 1
274         &data_word(0x40108010, 0x00000000, 0x00108000, 0x40100000);
275         &data_word(0x40000010, 0x00008010, 0x40008000, 0x00108000);
276         &data_word(0x00008000, 0x40100010, 0x00000010, 0x40008000);
277         &data_word(0x00100010, 0x40108000, 0x40100000, 0x00000010);
278         &data_word(0x00100000, 0x40008010, 0x40100010, 0x00008000);
279         &data_word(0x00108010, 0x40000000, 0x00000000, 0x00100010);
280         &data_word(0x40008010, 0x00108010, 0x40108000, 0x40000010);
281         &data_word(0x40000000, 0x00100000, 0x00008010, 0x40108010);
282         &data_word(0x00100010, 0x40108000, 0x40008000, 0x00108010);
283         &data_word(0x40108010, 0x00100010, 0x40000010, 0x00000000);
284         &data_word(0x40000000, 0x00008010, 0x00100000, 0x40100010);
285         &data_word(0x00008000, 0x40000000, 0x00108010, 0x40008010);
286         &data_word(0x40108000, 0x00008000, 0x00000000, 0x40000010);
287         &data_word(0x00000010, 0x40108010, 0x00108000, 0x40100000);
288         &data_word(0x40100010, 0x00100000, 0x00008010, 0x40008000);
289         &data_word(0x40008010, 0x00000010, 0x40100000, 0x00108000);
290         # nibble 2
291         &data_word(0x04000001, 0x04040100, 0x00000100, 0x04000101);
292         &data_word(0x00040001, 0x04000000, 0x04000101, 0x00040100);
293         &data_word(0x04000100, 0x00040000, 0x04040000, 0x00000001);
294         &data_word(0x04040101, 0x00000101, 0x00000001, 0x04040001);
295         &data_word(0x00000000, 0x00040001, 0x04040100, 0x00000100);
296         &data_word(0x00000101, 0x04040101, 0x00040000, 0x04000001);
297         &data_word(0x04040001, 0x04000100, 0x00040101, 0x04040000);
298         &data_word(0x00040100, 0x00000000, 0x04000000, 0x00040101);
299         &data_word(0x04040100, 0x00000100, 0x00000001, 0x00040000);
300         &data_word(0x00000101, 0x00040001, 0x04040000, 0x04000101);
301         &data_word(0x00000000, 0x04040100, 0x00040100, 0x04040001);
302         &data_word(0x00040001, 0x04000000, 0x04040101, 0x00000001);
303         &data_word(0x00040101, 0x04000001, 0x04000000, 0x04040101);
304         &data_word(0x00040000, 0x04000100, 0x04000101, 0x00040100);
305         &data_word(0x04000100, 0x00000000, 0x04040001, 0x00000101);
306         &data_word(0x04000001, 0x00040101, 0x00000100, 0x04040000);
307         # nibble 3
308         &data_word(0x00401008, 0x10001000, 0x00000008, 0x10401008);
309         &data_word(0x00000000, 0x10400000, 0x10001008, 0x00400008);
310         &data_word(0x10401000, 0x10000008, 0x10000000, 0x00001008);
311         &data_word(0x10000008, 0x00401008, 0x00400000, 0x10000000);
312         &data_word(0x10400008, 0x00401000, 0x00001000, 0x00000008);
313         &data_word(0x00401000, 0x10001008, 0x10400000, 0x00001000);
314         &data_word(0x00001008, 0x00000000, 0x00400008, 0x10401000);
315         &data_word(0x10001000, 0x10400008, 0x10401008, 0x00400000);
316         &data_word(0x10400008, 0x00001008, 0x00400000, 0x10000008);
317         &data_word(0x00401000, 0x10001000, 0x00000008, 0x10400000);
318         &data_word(0x10001008, 0x00000000, 0x00001000, 0x00400008);
319         &data_word(0x00000000, 0x10400008, 0x10401000, 0x00001000);
320         &data_word(0x10000000, 0x10401008, 0x00401008, 0x00400000);
321         &data_word(0x10401008, 0x00000008, 0x10001000, 0x00401008);
322         &data_word(0x00400008, 0x00401000, 0x10400000, 0x10001008);
323         &data_word(0x00001008, 0x10000000, 0x10000008, 0x10401000);
324         # nibble 4
325         &data_word(0x08000000, 0x00010000, 0x00000400, 0x08010420);
326         &data_word(0x08010020, 0x08000400, 0x00010420, 0x08010000);
327         &data_word(0x00010000, 0x00000020, 0x08000020, 0x00010400);
328         &data_word(0x08000420, 0x08010020, 0x08010400, 0x00000000);
329         &data_word(0x00010400, 0x08000000, 0x00010020, 0x00000420);
330         &data_word(0x08000400, 0x00010420, 0x00000000, 0x08000020);
331         &data_word(0x00000020, 0x08000420, 0x08010420, 0x00010020);
332         &data_word(0x08010000, 0x00000400, 0x00000420, 0x08010400);
333         &data_word(0x08010400, 0x08000420, 0x00010020, 0x08010000);
334         &data_word(0x00010000, 0x00000020, 0x08000020, 0x08000400);
335         &data_word(0x08000000, 0x00010400, 0x08010420, 0x00000000);
336         &data_word(0x00010420, 0x08000000, 0x00000400, 0x00010020);
337         &data_word(0x08000420, 0x00000400, 0x00000000, 0x08010420);
338         &data_word(0x08010020, 0x08010400, 0x00000420, 0x00010000);
339         &data_word(0x00010400, 0x08010020, 0x08000400, 0x00000420);
340         &data_word(0x00000020, 0x00010420, 0x08010000, 0x08000020);
341         # nibble 5
342         &data_word(0x80000040, 0x00200040, 0x00000000, 0x80202000);
343         &data_word(0x00200040, 0x00002000, 0x80002040, 0x00200000);
344         &data_word(0x00002040, 0x80202040, 0x00202000, 0x80000000);
345         &data_word(0x80002000, 0x80000040, 0x80200000, 0x00202040);
346         &data_word(0x00200000, 0x80002040, 0x80200040, 0x00000000);
347         &data_word(0x00002000, 0x00000040, 0x80202000, 0x80200040);
348         &data_word(0x80202040, 0x80200000, 0x80000000, 0x00002040);
349         &data_word(0x00000040, 0x00202000, 0x00202040, 0x80002000);
350         &data_word(0x00002040, 0x80000000, 0x80002000, 0x00202040);
351         &data_word(0x80202000, 0x00200040, 0x00000000, 0x80002000);
352         &data_word(0x80000000, 0x00002000, 0x80200040, 0x00200000);
353         &data_word(0x00200040, 0x80202040, 0x00202000, 0x00000040);
354         &data_word(0x80202040, 0x00202000, 0x00200000, 0x80002040);
355         &data_word(0x80000040, 0x80200000, 0x00202040, 0x00000000);
356         &data_word(0x00002000, 0x80000040, 0x80002040, 0x80202000);
357         &data_word(0x80200000, 0x00002040, 0x00000040, 0x80200040);
358         # nibble 6
359         &data_word(0x00004000, 0x00000200, 0x01000200, 0x01000004);
360         &data_word(0x01004204, 0x00004004, 0x00004200, 0x00000000);
361         &data_word(0x01000000, 0x01000204, 0x00000204, 0x01004000);
362         &data_word(0x00000004, 0x01004200, 0x01004000, 0x00000204);
363         &data_word(0x01000204, 0x00004000, 0x00004004, 0x01004204);
364         &data_word(0x00000000, 0x01000200, 0x01000004, 0x00004200);
365         &data_word(0x01004004, 0x00004204, 0x01004200, 0x00000004);
366         &data_word(0x00004204, 0x01004004, 0x00000200, 0x01000000);
367         &data_word(0x00004204, 0x01004000, 0x01004004, 0x00000204);
368         &data_word(0x00004000, 0x00000200, 0x01000000, 0x01004004);
369         &data_word(0x01000204, 0x00004204, 0x00004200, 0x00000000);
370         &data_word(0x00000200, 0x01000004, 0x00000004, 0x01000200);
371         &data_word(0x00000000, 0x01000204, 0x01000200, 0x00004200);
372         &data_word(0x00000204, 0x00004000, 0x01004204, 0x01000000);
373         &data_word(0x01004200, 0x00000004, 0x00004004, 0x01004204);
374         &data_word(0x01000004, 0x01004200, 0x01004000, 0x00004004);
375         # nibble 7
376         &data_word(0x20800080, 0x20820000, 0x00020080, 0x00000000);
377         &data_word(0x20020000, 0x00800080, 0x20800000, 0x20820080);
378         &data_word(0x00000080, 0x20000000, 0x00820000, 0x00020080);
379         &data_word(0x00820080, 0x20020080, 0x20000080, 0x20800000);
380         &data_word(0x00020000, 0x00820080, 0x00800080, 0x20020000);
381         &data_word(0x20820080, 0x20000080, 0x00000000, 0x00820000);
382         &data_word(0x20000000, 0x00800000, 0x20020080, 0x20800080);
383         &data_word(0x00800000, 0x00020000, 0x20820000, 0x00000080);
384         &data_word(0x00800000, 0x00020000, 0x20000080, 0x20820080);
385         &data_word(0x00020080, 0x20000000, 0x00000000, 0x00820000);
386         &data_word(0x20800080, 0x20020080, 0x20020000, 0x00800080);
387         &data_word(0x20820000, 0x00000080, 0x00800080, 0x20020000);
388         &data_word(0x20820080, 0x00800000, 0x20800000, 0x20000080);
389         &data_word(0x00820000, 0x00020080, 0x20020080, 0x20800000);
390         &data_word(0x00000080, 0x20820000, 0x00820080, 0x00000000);
391         &data_word(0x20000000, 0x20800080, 0x00020000, 0x00820080);
392         }