"Monolithic" x86 assembler replacement for aes_core.c. Up to +15% better
[openssl.git] / crypto / aes / asm / aes-586.pl
1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
8 #
9 # Version 2.0.
10 #
11 # You might fail to appreciate this module performance from the first
12 # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
13 # to be *the* best Intel C compiler without -KPIC, performance appears
14 # to be virtually identical... But try to re-configure with shared
15 # library support... Aha! Intel compiler "suddenly" lags behind by 30%
16 # [on P4, more on others]:-) And if compared to position-independent
17 # code generated by GNU C, this code performs *more* than *twice* as
18 # fast! Yes, all this buzz about PIC means that unlike other hand-
19 # coded implementations, this one was explicitly designed to be safe
20 # to use even in shared library context... This also means that this
21 # code isn't necessarily absolutely fastest "ever," because in order
22 # to achieve position independence an extra register has to be
23 # off-loaded to stack, which affects the benchmark result.
24 #
25 # Special note about instruction choice. Do you recall RC4_INT code
26 # performing poorly on P4? It might be the time to figure out why.
27 # RC4_INT code implies effective address calculations in base+offset*4
28 # form. Trouble is that it seems that offset scaling turned to be
29 # critical path... At least eliminating scaling resulted in 2.8x RC4
30 # performance improvement [as you might recall]. As AES code is hungry
31 # for scaling too, I [try to] avoid the latter by favoring off-by-2
32 # shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF.
33 #
34 # As was shown by Dean Gaudet <dean@arctic.org>, the above note turned
35 # void. Performance improvement with off-by-2 shifts was observed on
36 # intermediate implementation, which was spilling yet another register
37 # to stack... Final offset*4 code below runs just a tad faster on P4,
38 # but exhibits up to 10% improvement on other cores.
39 #
40 # Second version is "monolithic" replacement for aes_core.c, which in
41 # addition to AES_[de|en]crypt implements AES_set_[de|en]cryption_key.
42 # This made it possible to implement little-endian variant of the
43 # algorithm without modifying the base C code. Motivating factor for
44 # the undertaken effort was that it appeared that in tight IA-32
45 # register window little-endian flavor could achieve slightly higher
46 # Instruction Level Parallelism, and it indeed resulted in up to 15%
47 # better performance on most recent µ-archs...
48 #
49 # Current ECB performance numbers for 128-bit key in cycles per byte
50 # [measure commonly used by AES benchmarkers] are:
51 #
52 #               small footprint         fully unrolled
53 # P4[-3]        23[24]                  22[23]
54 # AMD K8        19                      18
55 # PIII          26(*)                   23
56 # Pentium       63(*)                   52
57 #
58 # (*)   Performance difference between small footprint code and fully
59 #       unrolled in more commonly used CBC mode is not as big, 7% for
60 #       PIII and 15% for Pentium, which I consider tolerable.
61
62 push(@INC,"perlasm","../../perlasm");
63 require "x86asm.pl";
64
65 &asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386");
66
67 $small_footprint=1;     # $small_footprint=1 code is ~5% slower [on
68                         # recent µ-archs], but ~5 times smaller!
69                         # I favor compact code, because it minimizes
70                         # cache contention...
71 $vertical_spin=0;       # shift "verticaly" defaults to 0, because of
72                         # its proof-of-concept status, see below...
73
74 $s0="eax";
75 $s1="ebx";
76 $s2="ecx";
77 $s3="edx";
78 $key="esi";
79 $acc="edi";
80
81 if ($vertical_spin) {
82         # I need high parts of volatile registers to be accessible...
83         $s1="esi";      $key="ebx";
84         $s2="edi";      $acc="ecx";
85 }
86 # Note that there is no decvert(), as well as last encryption round is
87 # performed with "horizontal" shifts. This is because this "vertical"
88 # implementation [one which groups shifts on a given $s[i] to form a
89 # "column," unlike "horizontal" one, which groups shifts on different
90 # $s[i] to form a "row"] is work in progress. It was observed to run
91 # few percents faster on Intel cores, but not AMD. On AMD K8 core it's
92 # whole 12% slower:-( So we face a trade-off... Shall it be resolved
93 # some day? Till then the code is considered experimental and by
94 # default remains dormant...
95
96 sub encvert()
97 { my ($te,@s) = @_;
98   my $v0 = $acc, $v1 = $key;
99
100         &mov    ($v0,$s[3]);                            # copy s3
101         &mov    (&DWP(0,"esp"),$s[2]);                  # save s2
102         &mov    ($v1,$s[0]);                            # copy s0
103         &mov    (&DWP(4,"esp"),$s[1]);                  # save s1
104
105         &movz   ($s[2],&HB($s[0]));
106         &and    ($s[0],0xFF);
107         &mov    ($s[0],&DWP(1024*0,$te,$s[0],4));       # s0>>0
108         &shr    ($v1,16);
109         &mov    ($s[3],&DWP(1024*1,$te,$s[2],4));       # s0>>8
110         &movz   ($s[1],&HB($v1));
111         &and    ($v1,0xFF);
112         &mov    ($s[2],&DWP(1024*2,$te,$v1,4));         # s0>>16
113          &mov   ($v1,$v0);
114         &mov    ($s[1],&DWP(1024*3,$te,$s[1],4));       # s0>>24
115
116         &and    ($v0,0xFF);
117         &xor    ($s[3],&DWP(1024*0,$te,$v0,4));         # s3>>0
118         &movz   ($v0,&HB($v1));
119         &shr    ($v1,16);
120         &xor    ($s[2],&DWP(1024*1,$te,$v0,4));         # s3>>8
121         &movz   ($v0,&HB($v1));
122         &and    ($v1,0xFF);
123         &xor    ($s[1],&DWP(1024*2,$te,$v1,4));         # s3>>16
124          &mov   ($v1,&DWP(0,"esp"));                    # restore s2
125         &xor    ($s[0],&DWP(1024*3,$te,$v0,4));         # s3>>24
126
127         &mov    ($v0,$v1);
128         &and    ($v1,0xFF);
129         &xor    ($s[2],&DWP(1024*0,$te,$v1,4));         # s2>>0
130         &movz   ($v1,&HB($v0));
131         &shr    ($v0,16);
132         &xor    ($s[1],&DWP(1024*1,$te,$v1,4));         # s2>>8
133         &movz   ($v1,&HB($v0));
134         &and    ($v0,0xFF);
135         &xor    ($s[0],&DWP(1024*2,$te,$v0,4));         # s2>>16
136          &mov   ($v0,&DWP(4,"esp"));                    # restore s1
137         &xor    ($s[3],&DWP(1024*3,$te,$v1,4));         # s2>>24
138
139         &mov    ($v1,$v0);
140         &and    ($v0,0xFF);
141         &xor    ($s[1],&DWP(1024*0,$te,$v0,4));         # s1>>0
142         &movz   ($v0,&HB($v1));
143         &shr    ($v1,16);
144         &xor    ($s[0],&DWP(1024*1,$te,$v0,4));         # s1>>8
145         &movz   ($v0,&HB($v1));
146         &and    ($v1,0xFF);
147         &xor    ($s[3],&DWP(1024*2,$te,$v1,4));         # s1>>16
148          &mov   ($key,&DWP(12,"esp"));                  # reincarnate v1 as key
149         &xor    ($s[2],&DWP(1024*3,$te,$v0,4));         # s1>>24
150 }
151
152 sub encstep()
153 { my ($i,$te,@s) = @_;
154   my $tmp = $key;
155   my $out = $i==3?$s[0]:$acc;
156
157         # lines marked with #%e?x[i] denote "reordered" instructions...
158         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
159         else        {   &mov    ($out,$s[0]);
160                         &and    ($out,0xFF);                    }
161         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
162         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
163                         &mov    ($out,&DWP(1024*0,$te,$out,4));
164
165         if ($i==3)  {   $tmp=$s[1];                             }##%eax
166                         &movz   ($tmp,&HB($s[1]));
167                         &xor    ($out,&DWP(1024*1,$te,$tmp,4));
168
169         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(0,"esp")); }##%ebx
170         else        {   &mov    ($tmp,$s[2]);
171                         &shr    ($tmp,16);                      }
172         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
173                         &and    ($tmp,0xFF);
174                         &xor    ($out,&DWP(1024*2,$te,$tmp,4));
175
176         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }##%ecx
177         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
178         else        {   &mov    ($tmp,$s[3]); 
179                         &shr    ($tmp,24)                       }
180                         &xor    ($out,&DWP(1024*3,$te,$tmp,4));
181         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
182         if ($i==3)  {   &mov    ($s[3],$acc);                   }
183                         &comment();
184 }
185
186 sub enclast()
187 { my ($i,$te,@s)=@_;
188   my $tmp = $key;
189   my $out = $i==3?$s[0]:$acc;
190
191         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
192         else        {   &mov    ($out,$s[0]);                   }
193                         &and    ($out,0xFF);
194         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
195         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
196                         &mov    ($out,&DWP(1024*0,$te,$out,4));
197                         &and    ($out,0x000000ff);
198
199         if ($i==3)  {   $tmp=$s[1];                             }##%eax
200                         &movz   ($tmp,&HB($s[1]));
201                         &mov    ($tmp,&DWP(0,$te,$tmp,4));
202                         &and    ($tmp,0x0000ff00);
203                         &xor    ($out,$tmp);
204
205         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(0,"esp")); }##%ebx
206         else        {   mov     ($tmp,$s[2]);
207                         &shr    ($tmp,16);                      }
208         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
209                         &and    ($tmp,0xFF);
210                         &mov    ($tmp,&DWP(0,$te,$tmp,4));
211                         &and    ($tmp,0x00ff0000);
212                         &xor    ($out,$tmp);
213
214         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }##%ecx
215         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
216         else        {   &mov    ($tmp,$s[3]);
217                         &shr    ($tmp,24);                      }
218                         &mov    ($tmp,&DWP(0,$te,$tmp,4));
219                         &and    ($tmp,0xff000000);
220                         &xor    ($out,$tmp);
221         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
222         if ($i==3)  {   &mov    ($s[3],$acc);                   }
223 }
224
225 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
226 &public_label("AES_Te");
227 &function_begin("AES_encrypt");
228         &mov    ($acc,&wparam(0));              # load inp
229         &mov    ($key,&wparam(2));              # load key
230
231         &call   (&label("pic_point"));          # make it PIC!
232         &set_label("pic_point");
233         &blindpop("ebp");
234         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
235
236         # allocate aligned stack frame
237         &mov    ($s0,"esp");
238         &sub    ("esp",20);
239         &and    ("esp",-16);
240
241         &mov    (&DWP(12,"esp"),$key);          # save key
242         &mov    (&DWP(16,"esp"),$s0);           # save %esp
243
244         &mov    ($s0,&DWP(0,$acc));             # load input data
245         &mov    ($s1,&DWP(4,$acc));
246         &mov    ($s2,&DWP(8,$acc));
247         &mov    ($s3,&DWP(12,$acc));
248
249         &xor    ($s0,&DWP(0,$key));
250         &xor    ($s1,&DWP(4,$key));
251         &xor    ($s2,&DWP(8,$key));
252         &xor    ($s3,&DWP(12,$key));
253
254         &mov    ($acc,&DWP(240,$key));          # load key->rounds
255
256         if ($small_footprint) {
257             &lea        ($acc,&DWP(-2,$acc,$acc));
258             &lea        ($acc,&DWP(0,$key,$acc,8));
259             &mov        (&DWP(8,"esp"),$acc);   # end of key schedule
260             &align      (4);
261             &set_label("loop");
262                 if ($vertical_spin) {
263                     &encvert("ebp",$s0,$s1,$s2,$s3);
264                 } else {
265                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
266                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
267                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
268                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
269                 }
270                 &add    ($key,16);                      # advance rd_key
271                 &xor    ($s0,&DWP(0,$key));
272                 &xor    ($s1,&DWP(4,$key));
273                 &xor    ($s2,&DWP(8,$key));
274                 &xor    ($s3,&DWP(12,$key));
275             &cmp        ($key,&DWP(8,"esp"));
276             &mov        (&DWP(12,"esp"),$key);
277             &jb         (&label("loop"));
278         }
279         else {
280             &cmp        ($acc,10);
281             &jle        (&label("10rounds"));
282             &cmp        ($acc,12);
283             &jle        (&label("12rounds"));
284
285         &set_label("14rounds");
286             for ($i=1;$i<3;$i++) {
287                 if ($vertical_spin) {
288                     &encvert("ebp",$s0,$s1,$s2,$s3);
289                 } else {
290                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
291                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
292                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
293                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
294                 }
295                 &xor    ($s0,&DWP(16*$i+0,$key));
296                 &xor    ($s1,&DWP(16*$i+4,$key));
297                 &xor    ($s2,&DWP(16*$i+8,$key));
298                 &xor    ($s3,&DWP(16*$i+12,$key));
299             }
300             &add        ($key,32);
301             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
302         &set_label("12rounds");
303             for ($i=1;$i<3;$i++) {
304                 if ($vertical_spin) {
305                     &encvert("ebp",$s0,$s1,$s2,$s3);
306                 } else {
307                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
308                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
309                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
310                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
311                 }
312                 &xor    ($s0,&DWP(16*$i+0,$key));
313                 &xor    ($s1,&DWP(16*$i+4,$key));
314                 &xor    ($s2,&DWP(16*$i+8,$key));
315                 &xor    ($s3,&DWP(16*$i+12,$key));
316             }
317             &add        ($key,32);
318             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
319         &set_label("10rounds");
320             for ($i=1;$i<10;$i++) {
321                 if ($vertical_spin) {
322                     &encvert("ebp",$s0,$s1,$s2,$s3);
323                 } else {
324                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
325                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
326                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
327                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
328                 }
329                 &xor    ($s0,&DWP(16*$i+0,$key));
330                 &xor    ($s1,&DWP(16*$i+4,$key));
331                 &xor    ($s2,&DWP(16*$i+8,$key));
332                 &xor    ($s3,&DWP(16*$i+12,$key));
333             }
334         }
335
336         &add    ("ebp",4*1024);                 # skip to Te4
337         if ($vertical_spin) {
338             # "reincarnate" some registers for "horizontal" spin...
339             &mov        ($s1="ebx",$key="esi");
340             &mov        ($s2="ecx",$acc="edi");
341         }
342         &enclast(0,"ebp",$s0,$s1,$s2,$s3);
343         &enclast(1,"ebp",$s1,$s2,$s3,$s0);
344         &enclast(2,"ebp",$s2,$s3,$s0,$s1);
345         &enclast(3,"ebp",$s3,$s0,$s1,$s2);
346
347         &mov    ("esp",&DWP(16,"esp"));         # restore %esp
348         &add    ($key,$small_footprint?16:160);
349         &xor    ($s0,&DWP(0,$key));
350         &xor    ($s1,&DWP(4,$key));
351         &xor    ($s2,&DWP(8,$key));
352         &xor    ($s3,&DWP(12,$key));
353
354         &mov    ($acc,&wparam(1));              # load out
355         &mov    (&DWP(0,$acc),$s0);             # write output data
356         &mov    (&DWP(4,$acc),$s1);
357         &mov    (&DWP(8,$acc),$s2);
358         &mov    (&DWP(12,$acc),$s3);
359
360         &pop    ("edi");
361         &pop    ("esi");
362         &pop    ("ebx");
363         &pop    ("ebp");
364         &ret    ();
365
366 &set_label("AES_Te",64);        # Yes! I keep it in the code segment!
367         &data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
368         &data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
369         &data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
370         &data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
371         &data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
372         &data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
373         &data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
374         &data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
375         &data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
376         &data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
377         &data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
378         &data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
379         &data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
380         &data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
381         &data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
382         &data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
383         &data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
384         &data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
385         &data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
386         &data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
387         &data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
388         &data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
389         &data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
390         &data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
391         &data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
392         &data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
393         &data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
394         &data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
395         &data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
396         &data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
397         &data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
398         &data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
399         &data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
400         &data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
401         &data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
402         &data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
403         &data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
404         &data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
405         &data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
406         &data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
407         &data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
408         &data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
409         &data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
410         &data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
411         &data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
412         &data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
413         &data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
414         &data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
415         &data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
416         &data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
417         &data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
418         &data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
419         &data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
420         &data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
421         &data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
422         &data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
423         &data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
424         &data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
425         &data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
426         &data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
427         &data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
428         &data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
429         &data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
430         &data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
431 #Te1:
432         &data_word(0x6363c6a5, 0x7c7cf884, 0x7777ee99, 0x7b7bf68d);
433         &data_word(0xf2f2ff0d, 0x6b6bd6bd, 0x6f6fdeb1, 0xc5c59154);
434         &data_word(0x30306050, 0x01010203, 0x6767cea9, 0x2b2b567d);
435         &data_word(0xfefee719, 0xd7d7b562, 0xabab4de6, 0x7676ec9a);
436         &data_word(0xcaca8f45, 0x82821f9d, 0xc9c98940, 0x7d7dfa87);
437         &data_word(0xfafaef15, 0x5959b2eb, 0x47478ec9, 0xf0f0fb0b);
438         &data_word(0xadad41ec, 0xd4d4b367, 0xa2a25ffd, 0xafaf45ea);
439         &data_word(0x9c9c23bf, 0xa4a453f7, 0x7272e496, 0xc0c09b5b);
440         &data_word(0xb7b775c2, 0xfdfde11c, 0x93933dae, 0x26264c6a);
441         &data_word(0x36366c5a, 0x3f3f7e41, 0xf7f7f502, 0xcccc834f);
442         &data_word(0x3434685c, 0xa5a551f4, 0xe5e5d134, 0xf1f1f908);
443         &data_word(0x7171e293, 0xd8d8ab73, 0x31316253, 0x15152a3f);
444         &data_word(0x0404080c, 0xc7c79552, 0x23234665, 0xc3c39d5e);
445         &data_word(0x18183028, 0x969637a1, 0x05050a0f, 0x9a9a2fb5);
446         &data_word(0x07070e09, 0x12122436, 0x80801b9b, 0xe2e2df3d);
447         &data_word(0xebebcd26, 0x27274e69, 0xb2b27fcd, 0x7575ea9f);
448         &data_word(0x0909121b, 0x83831d9e, 0x2c2c5874, 0x1a1a342e);
449         &data_word(0x1b1b362d, 0x6e6edcb2, 0x5a5ab4ee, 0xa0a05bfb);
450         &data_word(0x5252a4f6, 0x3b3b764d, 0xd6d6b761, 0xb3b37dce);
451         &data_word(0x2929527b, 0xe3e3dd3e, 0x2f2f5e71, 0x84841397);
452         &data_word(0x5353a6f5, 0xd1d1b968, 0x00000000, 0xededc12c);
453         &data_word(0x20204060, 0xfcfce31f, 0xb1b179c8, 0x5b5bb6ed);
454         &data_word(0x6a6ad4be, 0xcbcb8d46, 0xbebe67d9, 0x3939724b);
455         &data_word(0x4a4a94de, 0x4c4c98d4, 0x5858b0e8, 0xcfcf854a);
456         &data_word(0xd0d0bb6b, 0xefefc52a, 0xaaaa4fe5, 0xfbfbed16);
457         &data_word(0x434386c5, 0x4d4d9ad7, 0x33336655, 0x85851194);
458         &data_word(0x45458acf, 0xf9f9e910, 0x02020406, 0x7f7ffe81);
459         &data_word(0x5050a0f0, 0x3c3c7844, 0x9f9f25ba, 0xa8a84be3);
460         &data_word(0x5151a2f3, 0xa3a35dfe, 0x404080c0, 0x8f8f058a);
461         &data_word(0x92923fad, 0x9d9d21bc, 0x38387048, 0xf5f5f104);
462         &data_word(0xbcbc63df, 0xb6b677c1, 0xdadaaf75, 0x21214263);
463         &data_word(0x10102030, 0xffffe51a, 0xf3f3fd0e, 0xd2d2bf6d);
464         &data_word(0xcdcd814c, 0x0c0c1814, 0x13132635, 0xececc32f);
465         &data_word(0x5f5fbee1, 0x979735a2, 0x444488cc, 0x17172e39);
466         &data_word(0xc4c49357, 0xa7a755f2, 0x7e7efc82, 0x3d3d7a47);
467         &data_word(0x6464c8ac, 0x5d5dbae7, 0x1919322b, 0x7373e695);
468         &data_word(0x6060c0a0, 0x81811998, 0x4f4f9ed1, 0xdcdca37f);
469         &data_word(0x22224466, 0x2a2a547e, 0x90903bab, 0x88880b83);
470         &data_word(0x46468cca, 0xeeeec729, 0xb8b86bd3, 0x1414283c);
471         &data_word(0xdedea779, 0x5e5ebce2, 0x0b0b161d, 0xdbdbad76);
472         &data_word(0xe0e0db3b, 0x32326456, 0x3a3a744e, 0x0a0a141e);
473         &data_word(0x494992db, 0x06060c0a, 0x2424486c, 0x5c5cb8e4);
474         &data_word(0xc2c29f5d, 0xd3d3bd6e, 0xacac43ef, 0x6262c4a6);
475         &data_word(0x919139a8, 0x959531a4, 0xe4e4d337, 0x7979f28b);
476         &data_word(0xe7e7d532, 0xc8c88b43, 0x37376e59, 0x6d6ddab7);
477         &data_word(0x8d8d018c, 0xd5d5b164, 0x4e4e9cd2, 0xa9a949e0);
478         &data_word(0x6c6cd8b4, 0x5656acfa, 0xf4f4f307, 0xeaeacf25);
479         &data_word(0x6565caaf, 0x7a7af48e, 0xaeae47e9, 0x08081018);
480         &data_word(0xbaba6fd5, 0x7878f088, 0x25254a6f, 0x2e2e5c72);
481         &data_word(0x1c1c3824, 0xa6a657f1, 0xb4b473c7, 0xc6c69751);
482         &data_word(0xe8e8cb23, 0xdddda17c, 0x7474e89c, 0x1f1f3e21);
483         &data_word(0x4b4b96dd, 0xbdbd61dc, 0x8b8b0d86, 0x8a8a0f85);
484         &data_word(0x7070e090, 0x3e3e7c42, 0xb5b571c4, 0x6666ccaa);
485         &data_word(0x484890d8, 0x03030605, 0xf6f6f701, 0x0e0e1c12);
486         &data_word(0x6161c2a3, 0x35356a5f, 0x5757aef9, 0xb9b969d0);
487         &data_word(0x86861791, 0xc1c19958, 0x1d1d3a27, 0x9e9e27b9);
488         &data_word(0xe1e1d938, 0xf8f8eb13, 0x98982bb3, 0x11112233);
489         &data_word(0x6969d2bb, 0xd9d9a970, 0x8e8e0789, 0x949433a7);
490         &data_word(0x9b9b2db6, 0x1e1e3c22, 0x87871592, 0xe9e9c920);
491         &data_word(0xcece8749, 0x5555aaff, 0x28285078, 0xdfdfa57a);
492         &data_word(0x8c8c038f, 0xa1a159f8, 0x89890980, 0x0d0d1a17);
493         &data_word(0xbfbf65da, 0xe6e6d731, 0x424284c6, 0x6868d0b8);
494         &data_word(0x414182c3, 0x999929b0, 0x2d2d5a77, 0x0f0f1e11);
495         &data_word(0xb0b07bcb, 0x5454a8fc, 0xbbbb6dd6, 0x16162c3a);
496 #Te2:
497         &data_word(0x63c6a563, 0x7cf8847c, 0x77ee9977, 0x7bf68d7b);
498         &data_word(0xf2ff0df2, 0x6bd6bd6b, 0x6fdeb16f, 0xc59154c5);
499         &data_word(0x30605030, 0x01020301, 0x67cea967, 0x2b567d2b);
500         &data_word(0xfee719fe, 0xd7b562d7, 0xab4de6ab, 0x76ec9a76);
501         &data_word(0xca8f45ca, 0x821f9d82, 0xc98940c9, 0x7dfa877d);
502         &data_word(0xfaef15fa, 0x59b2eb59, 0x478ec947, 0xf0fb0bf0);
503         &data_word(0xad41ecad, 0xd4b367d4, 0xa25ffda2, 0xaf45eaaf);
504         &data_word(0x9c23bf9c, 0xa453f7a4, 0x72e49672, 0xc09b5bc0);
505         &data_word(0xb775c2b7, 0xfde11cfd, 0x933dae93, 0x264c6a26);
506         &data_word(0x366c5a36, 0x3f7e413f, 0xf7f502f7, 0xcc834fcc);
507         &data_word(0x34685c34, 0xa551f4a5, 0xe5d134e5, 0xf1f908f1);
508         &data_word(0x71e29371, 0xd8ab73d8, 0x31625331, 0x152a3f15);
509         &data_word(0x04080c04, 0xc79552c7, 0x23466523, 0xc39d5ec3);
510         &data_word(0x18302818, 0x9637a196, 0x050a0f05, 0x9a2fb59a);
511         &data_word(0x070e0907, 0x12243612, 0x801b9b80, 0xe2df3de2);
512         &data_word(0xebcd26eb, 0x274e6927, 0xb27fcdb2, 0x75ea9f75);
513         &data_word(0x09121b09, 0x831d9e83, 0x2c58742c, 0x1a342e1a);
514         &data_word(0x1b362d1b, 0x6edcb26e, 0x5ab4ee5a, 0xa05bfba0);
515         &data_word(0x52a4f652, 0x3b764d3b, 0xd6b761d6, 0xb37dceb3);
516         &data_word(0x29527b29, 0xe3dd3ee3, 0x2f5e712f, 0x84139784);
517         &data_word(0x53a6f553, 0xd1b968d1, 0x00000000, 0xedc12ced);
518         &data_word(0x20406020, 0xfce31ffc, 0xb179c8b1, 0x5bb6ed5b);
519         &data_word(0x6ad4be6a, 0xcb8d46cb, 0xbe67d9be, 0x39724b39);
520         &data_word(0x4a94de4a, 0x4c98d44c, 0x58b0e858, 0xcf854acf);
521         &data_word(0xd0bb6bd0, 0xefc52aef, 0xaa4fe5aa, 0xfbed16fb);
522         &data_word(0x4386c543, 0x4d9ad74d, 0x33665533, 0x85119485);
523         &data_word(0x458acf45, 0xf9e910f9, 0x02040602, 0x7ffe817f);
524         &data_word(0x50a0f050, 0x3c78443c, 0x9f25ba9f, 0xa84be3a8);
525         &data_word(0x51a2f351, 0xa35dfea3, 0x4080c040, 0x8f058a8f);
526         &data_word(0x923fad92, 0x9d21bc9d, 0x38704838, 0xf5f104f5);
527         &data_word(0xbc63dfbc, 0xb677c1b6, 0xdaaf75da, 0x21426321);
528         &data_word(0x10203010, 0xffe51aff, 0xf3fd0ef3, 0xd2bf6dd2);
529         &data_word(0xcd814ccd, 0x0c18140c, 0x13263513, 0xecc32fec);
530         &data_word(0x5fbee15f, 0x9735a297, 0x4488cc44, 0x172e3917);
531         &data_word(0xc49357c4, 0xa755f2a7, 0x7efc827e, 0x3d7a473d);
532         &data_word(0x64c8ac64, 0x5dbae75d, 0x19322b19, 0x73e69573);
533         &data_word(0x60c0a060, 0x81199881, 0x4f9ed14f, 0xdca37fdc);
534         &data_word(0x22446622, 0x2a547e2a, 0x903bab90, 0x880b8388);
535         &data_word(0x468cca46, 0xeec729ee, 0xb86bd3b8, 0x14283c14);
536         &data_word(0xdea779de, 0x5ebce25e, 0x0b161d0b, 0xdbad76db);
537         &data_word(0xe0db3be0, 0x32645632, 0x3a744e3a, 0x0a141e0a);
538         &data_word(0x4992db49, 0x060c0a06, 0x24486c24, 0x5cb8e45c);
539         &data_word(0xc29f5dc2, 0xd3bd6ed3, 0xac43efac, 0x62c4a662);
540         &data_word(0x9139a891, 0x9531a495, 0xe4d337e4, 0x79f28b79);
541         &data_word(0xe7d532e7, 0xc88b43c8, 0x376e5937, 0x6ddab76d);
542         &data_word(0x8d018c8d, 0xd5b164d5, 0x4e9cd24e, 0xa949e0a9);
543         &data_word(0x6cd8b46c, 0x56acfa56, 0xf4f307f4, 0xeacf25ea);
544         &data_word(0x65caaf65, 0x7af48e7a, 0xae47e9ae, 0x08101808);
545         &data_word(0xba6fd5ba, 0x78f08878, 0x254a6f25, 0x2e5c722e);
546         &data_word(0x1c38241c, 0xa657f1a6, 0xb473c7b4, 0xc69751c6);
547         &data_word(0xe8cb23e8, 0xdda17cdd, 0x74e89c74, 0x1f3e211f);
548         &data_word(0x4b96dd4b, 0xbd61dcbd, 0x8b0d868b, 0x8a0f858a);
549         &data_word(0x70e09070, 0x3e7c423e, 0xb571c4b5, 0x66ccaa66);
550         &data_word(0x4890d848, 0x03060503, 0xf6f701f6, 0x0e1c120e);
551         &data_word(0x61c2a361, 0x356a5f35, 0x57aef957, 0xb969d0b9);
552         &data_word(0x86179186, 0xc19958c1, 0x1d3a271d, 0x9e27b99e);
553         &data_word(0xe1d938e1, 0xf8eb13f8, 0x982bb398, 0x11223311);
554         &data_word(0x69d2bb69, 0xd9a970d9, 0x8e07898e, 0x9433a794);
555         &data_word(0x9b2db69b, 0x1e3c221e, 0x87159287, 0xe9c920e9);
556         &data_word(0xce8749ce, 0x55aaff55, 0x28507828, 0xdfa57adf);
557         &data_word(0x8c038f8c, 0xa159f8a1, 0x89098089, 0x0d1a170d);
558         &data_word(0xbf65dabf, 0xe6d731e6, 0x4284c642, 0x68d0b868);
559         &data_word(0x4182c341, 0x9929b099, 0x2d5a772d, 0x0f1e110f);
560         &data_word(0xb07bcbb0, 0x54a8fc54, 0xbb6dd6bb, 0x162c3a16);
561 #Te3:
562         &data_word(0xc6a56363, 0xf8847c7c, 0xee997777, 0xf68d7b7b);
563         &data_word(0xff0df2f2, 0xd6bd6b6b, 0xdeb16f6f, 0x9154c5c5);
564         &data_word(0x60503030, 0x02030101, 0xcea96767, 0x567d2b2b);
565         &data_word(0xe719fefe, 0xb562d7d7, 0x4de6abab, 0xec9a7676);
566         &data_word(0x8f45caca, 0x1f9d8282, 0x8940c9c9, 0xfa877d7d);
567         &data_word(0xef15fafa, 0xb2eb5959, 0x8ec94747, 0xfb0bf0f0);
568         &data_word(0x41ecadad, 0xb367d4d4, 0x5ffda2a2, 0x45eaafaf);
569         &data_word(0x23bf9c9c, 0x53f7a4a4, 0xe4967272, 0x9b5bc0c0);
570         &data_word(0x75c2b7b7, 0xe11cfdfd, 0x3dae9393, 0x4c6a2626);
571         &data_word(0x6c5a3636, 0x7e413f3f, 0xf502f7f7, 0x834fcccc);
572         &data_word(0x685c3434, 0x51f4a5a5, 0xd134e5e5, 0xf908f1f1);
573         &data_word(0xe2937171, 0xab73d8d8, 0x62533131, 0x2a3f1515);
574         &data_word(0x080c0404, 0x9552c7c7, 0x46652323, 0x9d5ec3c3);
575         &data_word(0x30281818, 0x37a19696, 0x0a0f0505, 0x2fb59a9a);
576         &data_word(0x0e090707, 0x24361212, 0x1b9b8080, 0xdf3de2e2);
577         &data_word(0xcd26ebeb, 0x4e692727, 0x7fcdb2b2, 0xea9f7575);
578         &data_word(0x121b0909, 0x1d9e8383, 0x58742c2c, 0x342e1a1a);
579         &data_word(0x362d1b1b, 0xdcb26e6e, 0xb4ee5a5a, 0x5bfba0a0);
580         &data_word(0xa4f65252, 0x764d3b3b, 0xb761d6d6, 0x7dceb3b3);
581         &data_word(0x527b2929, 0xdd3ee3e3, 0x5e712f2f, 0x13978484);
582         &data_word(0xa6f55353, 0xb968d1d1, 0x00000000, 0xc12ceded);
583         &data_word(0x40602020, 0xe31ffcfc, 0x79c8b1b1, 0xb6ed5b5b);
584         &data_word(0xd4be6a6a, 0x8d46cbcb, 0x67d9bebe, 0x724b3939);
585         &data_word(0x94de4a4a, 0x98d44c4c, 0xb0e85858, 0x854acfcf);
586         &data_word(0xbb6bd0d0, 0xc52aefef, 0x4fe5aaaa, 0xed16fbfb);
587         &data_word(0x86c54343, 0x9ad74d4d, 0x66553333, 0x11948585);
588         &data_word(0x8acf4545, 0xe910f9f9, 0x04060202, 0xfe817f7f);
589         &data_word(0xa0f05050, 0x78443c3c, 0x25ba9f9f, 0x4be3a8a8);
590         &data_word(0xa2f35151, 0x5dfea3a3, 0x80c04040, 0x058a8f8f);
591         &data_word(0x3fad9292, 0x21bc9d9d, 0x70483838, 0xf104f5f5);
592         &data_word(0x63dfbcbc, 0x77c1b6b6, 0xaf75dada, 0x42632121);
593         &data_word(0x20301010, 0xe51affff, 0xfd0ef3f3, 0xbf6dd2d2);
594         &data_word(0x814ccdcd, 0x18140c0c, 0x26351313, 0xc32fecec);
595         &data_word(0xbee15f5f, 0x35a29797, 0x88cc4444, 0x2e391717);
596         &data_word(0x9357c4c4, 0x55f2a7a7, 0xfc827e7e, 0x7a473d3d);
597         &data_word(0xc8ac6464, 0xbae75d5d, 0x322b1919, 0xe6957373);
598         &data_word(0xc0a06060, 0x19988181, 0x9ed14f4f, 0xa37fdcdc);
599         &data_word(0x44662222, 0x547e2a2a, 0x3bab9090, 0x0b838888);
600         &data_word(0x8cca4646, 0xc729eeee, 0x6bd3b8b8, 0x283c1414);
601         &data_word(0xa779dede, 0xbce25e5e, 0x161d0b0b, 0xad76dbdb);
602         &data_word(0xdb3be0e0, 0x64563232, 0x744e3a3a, 0x141e0a0a);
603         &data_word(0x92db4949, 0x0c0a0606, 0x486c2424, 0xb8e45c5c);
604         &data_word(0x9f5dc2c2, 0xbd6ed3d3, 0x43efacac, 0xc4a66262);
605         &data_word(0x39a89191, 0x31a49595, 0xd337e4e4, 0xf28b7979);
606         &data_word(0xd532e7e7, 0x8b43c8c8, 0x6e593737, 0xdab76d6d);
607         &data_word(0x018c8d8d, 0xb164d5d5, 0x9cd24e4e, 0x49e0a9a9);
608         &data_word(0xd8b46c6c, 0xacfa5656, 0xf307f4f4, 0xcf25eaea);
609         &data_word(0xcaaf6565, 0xf48e7a7a, 0x47e9aeae, 0x10180808);
610         &data_word(0x6fd5baba, 0xf0887878, 0x4a6f2525, 0x5c722e2e);
611         &data_word(0x38241c1c, 0x57f1a6a6, 0x73c7b4b4, 0x9751c6c6);
612         &data_word(0xcb23e8e8, 0xa17cdddd, 0xe89c7474, 0x3e211f1f);
613         &data_word(0x96dd4b4b, 0x61dcbdbd, 0x0d868b8b, 0x0f858a8a);
614         &data_word(0xe0907070, 0x7c423e3e, 0x71c4b5b5, 0xccaa6666);
615         &data_word(0x90d84848, 0x06050303, 0xf701f6f6, 0x1c120e0e);
616         &data_word(0xc2a36161, 0x6a5f3535, 0xaef95757, 0x69d0b9b9);
617         &data_word(0x17918686, 0x9958c1c1, 0x3a271d1d, 0x27b99e9e);
618         &data_word(0xd938e1e1, 0xeb13f8f8, 0x2bb39898, 0x22331111);
619         &data_word(0xd2bb6969, 0xa970d9d9, 0x07898e8e, 0x33a79494);
620         &data_word(0x2db69b9b, 0x3c221e1e, 0x15928787, 0xc920e9e9);
621         &data_word(0x8749cece, 0xaaff5555, 0x50782828, 0xa57adfdf);
622         &data_word(0x038f8c8c, 0x59f8a1a1, 0x09808989, 0x1a170d0d);
623         &data_word(0x65dabfbf, 0xd731e6e6, 0x84c64242, 0xd0b86868);
624         &data_word(0x82c34141, 0x29b09999, 0x5a772d2d, 0x1e110f0f);
625         &data_word(0x7bcbb0b0, 0xa8fc5454, 0x6dd6bbbb, 0x2c3a1616);
626 #Te4:
627         &data_word(0x63636363, 0x7c7c7c7c, 0x77777777, 0x7b7b7b7b);
628         &data_word(0xf2f2f2f2, 0x6b6b6b6b, 0x6f6f6f6f, 0xc5c5c5c5);
629         &data_word(0x30303030, 0x01010101, 0x67676767, 0x2b2b2b2b);
630         &data_word(0xfefefefe, 0xd7d7d7d7, 0xabababab, 0x76767676);
631         &data_word(0xcacacaca, 0x82828282, 0xc9c9c9c9, 0x7d7d7d7d);
632         &data_word(0xfafafafa, 0x59595959, 0x47474747, 0xf0f0f0f0);
633         &data_word(0xadadadad, 0xd4d4d4d4, 0xa2a2a2a2, 0xafafafaf);
634         &data_word(0x9c9c9c9c, 0xa4a4a4a4, 0x72727272, 0xc0c0c0c0);
635         &data_word(0xb7b7b7b7, 0xfdfdfdfd, 0x93939393, 0x26262626);
636         &data_word(0x36363636, 0x3f3f3f3f, 0xf7f7f7f7, 0xcccccccc);
637         &data_word(0x34343434, 0xa5a5a5a5, 0xe5e5e5e5, 0xf1f1f1f1);
638         &data_word(0x71717171, 0xd8d8d8d8, 0x31313131, 0x15151515);
639         &data_word(0x04040404, 0xc7c7c7c7, 0x23232323, 0xc3c3c3c3);
640         &data_word(0x18181818, 0x96969696, 0x05050505, 0x9a9a9a9a);
641         &data_word(0x07070707, 0x12121212, 0x80808080, 0xe2e2e2e2);
642         &data_word(0xebebebeb, 0x27272727, 0xb2b2b2b2, 0x75757575);
643         &data_word(0x09090909, 0x83838383, 0x2c2c2c2c, 0x1a1a1a1a);
644         &data_word(0x1b1b1b1b, 0x6e6e6e6e, 0x5a5a5a5a, 0xa0a0a0a0);
645         &data_word(0x52525252, 0x3b3b3b3b, 0xd6d6d6d6, 0xb3b3b3b3);
646         &data_word(0x29292929, 0xe3e3e3e3, 0x2f2f2f2f, 0x84848484);
647         &data_word(0x53535353, 0xd1d1d1d1, 0x00000000, 0xedededed);
648         &data_word(0x20202020, 0xfcfcfcfc, 0xb1b1b1b1, 0x5b5b5b5b);
649         &data_word(0x6a6a6a6a, 0xcbcbcbcb, 0xbebebebe, 0x39393939);
650         &data_word(0x4a4a4a4a, 0x4c4c4c4c, 0x58585858, 0xcfcfcfcf);
651         &data_word(0xd0d0d0d0, 0xefefefef, 0xaaaaaaaa, 0xfbfbfbfb);
652         &data_word(0x43434343, 0x4d4d4d4d, 0x33333333, 0x85858585);
653         &data_word(0x45454545, 0xf9f9f9f9, 0x02020202, 0x7f7f7f7f);
654         &data_word(0x50505050, 0x3c3c3c3c, 0x9f9f9f9f, 0xa8a8a8a8);
655         &data_word(0x51515151, 0xa3a3a3a3, 0x40404040, 0x8f8f8f8f);
656         &data_word(0x92929292, 0x9d9d9d9d, 0x38383838, 0xf5f5f5f5);
657         &data_word(0xbcbcbcbc, 0xb6b6b6b6, 0xdadadada, 0x21212121);
658         &data_word(0x10101010, 0xffffffff, 0xf3f3f3f3, 0xd2d2d2d2);
659         &data_word(0xcdcdcdcd, 0x0c0c0c0c, 0x13131313, 0xecececec);
660         &data_word(0x5f5f5f5f, 0x97979797, 0x44444444, 0x17171717);
661         &data_word(0xc4c4c4c4, 0xa7a7a7a7, 0x7e7e7e7e, 0x3d3d3d3d);
662         &data_word(0x64646464, 0x5d5d5d5d, 0x19191919, 0x73737373);
663         &data_word(0x60606060, 0x81818181, 0x4f4f4f4f, 0xdcdcdcdc);
664         &data_word(0x22222222, 0x2a2a2a2a, 0x90909090, 0x88888888);
665         &data_word(0x46464646, 0xeeeeeeee, 0xb8b8b8b8, 0x14141414);
666         &data_word(0xdededede, 0x5e5e5e5e, 0x0b0b0b0b, 0xdbdbdbdb);
667         &data_word(0xe0e0e0e0, 0x32323232, 0x3a3a3a3a, 0x0a0a0a0a);
668         &data_word(0x49494949, 0x06060606, 0x24242424, 0x5c5c5c5c);
669         &data_word(0xc2c2c2c2, 0xd3d3d3d3, 0xacacacac, 0x62626262);
670         &data_word(0x91919191, 0x95959595, 0xe4e4e4e4, 0x79797979);
671         &data_word(0xe7e7e7e7, 0xc8c8c8c8, 0x37373737, 0x6d6d6d6d);
672         &data_word(0x8d8d8d8d, 0xd5d5d5d5, 0x4e4e4e4e, 0xa9a9a9a9);
673         &data_word(0x6c6c6c6c, 0x56565656, 0xf4f4f4f4, 0xeaeaeaea);
674         &data_word(0x65656565, 0x7a7a7a7a, 0xaeaeaeae, 0x08080808);
675         &data_word(0xbabababa, 0x78787878, 0x25252525, 0x2e2e2e2e);
676         &data_word(0x1c1c1c1c, 0xa6a6a6a6, 0xb4b4b4b4, 0xc6c6c6c6);
677         &data_word(0xe8e8e8e8, 0xdddddddd, 0x74747474, 0x1f1f1f1f);
678         &data_word(0x4b4b4b4b, 0xbdbdbdbd, 0x8b8b8b8b, 0x8a8a8a8a);
679         &data_word(0x70707070, 0x3e3e3e3e, 0xb5b5b5b5, 0x66666666);
680         &data_word(0x48484848, 0x03030303, 0xf6f6f6f6, 0x0e0e0e0e);
681         &data_word(0x61616161, 0x35353535, 0x57575757, 0xb9b9b9b9);
682         &data_word(0x86868686, 0xc1c1c1c1, 0x1d1d1d1d, 0x9e9e9e9e);
683         &data_word(0xe1e1e1e1, 0xf8f8f8f8, 0x98989898, 0x11111111);
684         &data_word(0x69696969, 0xd9d9d9d9, 0x8e8e8e8e, 0x94949494);
685         &data_word(0x9b9b9b9b, 0x1e1e1e1e, 0x87878787, 0xe9e9e9e9);
686         &data_word(0xcececece, 0x55555555, 0x28282828, 0xdfdfdfdf);
687         &data_word(0x8c8c8c8c, 0xa1a1a1a1, 0x89898989, 0x0d0d0d0d);
688         &data_word(0xbfbfbfbf, 0xe6e6e6e6, 0x42424242, 0x68686868);
689         &data_word(0x41414141, 0x99999999, 0x2d2d2d2d, 0x0f0f0f0f);
690         &data_word(0xb0b0b0b0, 0x54545454, 0xbbbbbbbb, 0x16161616);
691 #rcon:
692         &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
693         &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
694         &data_word(0x0000001b, 0x00000036);
695 &function_end_B("AES_encrypt");
696
697 #------------------------------------------------------------------#
698
699 $s0="eax";
700 $s1="ebx";
701 $s2="ecx";
702 $s3="edx";
703 $key="edi";
704 $acc="esi";
705
706 sub decstep()
707 { my ($i,$td,@s) = @_;
708   my $tmp = $key;
709   my $out = $i==3?$s[0]:$acc;
710
711         # no instructions are reordered, as performance appears
712         # optimal... or rather that all attempts to reorder didn't
713         # result in better performance [which by the way is not a
714         # bit lower than ecryption].
715         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
716         else        {   &mov    ($out,$s[0]);                   }
717                         &and    ($out,0xFF);
718                         &mov    ($out,&DWP(1024*0,$td,$out,4));
719
720         if ($i==3)  {   $tmp=$s[1];                             }
721                         &movz   ($tmp,&HB($s[1]));
722                         &xor    ($out,&DWP(1024*1,$td,$tmp,4));
723
724         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
725         else        {   &mov    ($tmp,$s[2]);                   }
726                         &shr    ($tmp,16);
727                         &and    ($tmp,0xFF);
728                         &xor    ($out,&DWP(1024*2,$td,$tmp,4));
729
730         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }
731         else        {   &mov    ($tmp,$s[3]);                   }
732                         &shr    ($tmp,24);
733                         &xor    ($out,&DWP(1024*3,$td,$tmp,4));
734         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
735         if ($i==3)  {   &mov    ($s[3],&DWP(0,"esp"));          }
736                         &comment();
737 }
738
739 sub declast()
740 { my ($i,$td,@s)=@_;
741   my $tmp = $key;
742   my $out = $i==3?$s[0]:$acc;
743
744         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
745         else        {   &mov    ($out,$s[0]);                   }
746                         &and    ($out,0xFF);
747                         &mov    ($out,&DWP(0,$td,$out,4));
748                         &and    ($out,0x000000ff);
749
750         if ($i==3)  {   $tmp=$s[1];                             }
751                         &movz   ($tmp,&HB($s[1]));
752                         &mov    ($tmp,&DWP(0,$td,$tmp,4));
753                         &and    ($tmp,0x0000ff00);
754                         &xor    ($out,$tmp);
755
756         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
757         else        {   mov     ($tmp,$s[2]);                   }
758                         &shr    ($tmp,16);
759                         &and    ($tmp,0xFF);
760                         &mov    ($tmp,&DWP(0,$td,$tmp,4));
761                         &and    ($tmp,0x00ff0000);
762                         &xor    ($out,$tmp);
763
764         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }
765         else        {   &mov    ($tmp,$s[3]);                   }
766                         &shr    ($tmp,24);
767                         &mov    ($tmp,&DWP(0,$td,$tmp,4));
768                         &and    ($tmp,0xff000000);
769                         &xor    ($out,$tmp);
770         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
771         if ($i==3)  {   &mov    ($s[3],&DWP(0,"esp"));          }
772 }
773
774 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
775 &public_label("AES_Td");
776 &function_begin("AES_decrypt");
777         &mov    ($acc,&wparam(0));              # load inp
778         &mov    ($key,&wparam(2));              # load key
779
780         &call   (&label("pic_point"));          # make it PIC!
781         &set_label("pic_point");
782         &blindpop("ebp");
783         &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
784
785         # allocate aligned stack frame
786         &mov    ($s0,"esp");
787         &sub    ("esp",20);
788         &and    ("esp",-16);
789
790         &mov    (&DWP(12,"esp"),$key);          # save key
791         &mov    (&DWP(16,"esp"),$s0);           # save %esp
792
793         &mov    ($s0,&DWP(0,$acc));             # load input data
794         &mov    ($s1,&DWP(4,$acc));
795         &mov    ($s2,&DWP(8,$acc));
796         &mov    ($s3,&DWP(12,$acc));
797
798         &xor    ($s0,&DWP(0,$key));
799         &xor    ($s1,&DWP(4,$key));
800         &xor    ($s2,&DWP(8,$key));
801         &xor    ($s3,&DWP(12,$key));
802
803         &mov    ($acc,&DWP(240,$key));          # load key->rounds
804
805         if ($small_footprint) {
806             &lea        ($acc,&DWP(-2,$acc,$acc));
807             &lea        ($acc,&DWP(0,$key,$acc,8));
808             &mov        (&DWP(8,"esp"),$acc);   # end of key schedule
809             &align      (4);
810             &set_label("loop");
811                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
812                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
813                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
814                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
815                 &add    ($key,16);                      # advance rd_key
816                 &xor    ($s0,&DWP(0,$key));
817                 &xor    ($s1,&DWP(4,$key));
818                 &xor    ($s2,&DWP(8,$key));
819                 &xor    ($s3,&DWP(12,$key));
820             &cmp        ($key,&DWP(8,"esp"));
821             &mov        (&DWP(12,"esp"),$key);
822             &jb         (&label("loop"));
823         }
824         else {
825             &cmp        ($acc,10);
826             &jle        (&label("10rounds"));
827             &cmp        ($acc,12);
828             &jle        (&label("12rounds"));
829
830         &set_label("14rounds");
831             for ($i=1;$i<3;$i++) {
832                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
833                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
834                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
835                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
836                 &xor    ($s0,&DWP(16*$i+0,$key));
837                 &xor    ($s1,&DWP(16*$i+4,$key));
838                 &xor    ($s2,&DWP(16*$i+8,$key));
839                 &xor    ($s3,&DWP(16*$i+12,$key));
840             }
841             &add        ($key,32);
842             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
843         &set_label("12rounds");
844             for ($i=1;$i<3;$i++) {
845                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
846                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
847                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
848                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
849                 &xor    ($s0,&DWP(16*$i+0,$key));
850                 &xor    ($s1,&DWP(16*$i+4,$key));
851                 &xor    ($s2,&DWP(16*$i+8,$key));
852                 &xor    ($s3,&DWP(16*$i+12,$key));
853             }
854             &add        ($key,32);
855             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
856         &set_label("10rounds");
857             for ($i=1;$i<10;$i++) {
858                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
859                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
860                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
861                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
862                 &xor    ($s0,&DWP(16*$i+0,$key));
863                 &xor    ($s1,&DWP(16*$i+4,$key));
864                 &xor    ($s2,&DWP(16*$i+8,$key));
865                 &xor    ($s3,&DWP(16*$i+12,$key));
866             }
867         }
868
869         &add    ("ebp",4*1024);                 # skip to Te4
870         &declast(0,"ebp",$s0,$s3,$s2,$s1);
871         &declast(1,"ebp",$s1,$s0,$s3,$s2);
872         &declast(2,"ebp",$s2,$s1,$s0,$s3);
873         &declast(3,"ebp",$s3,$s2,$s1,$s0);
874
875         &mov    ("esp",&DWP(16,"esp"));         # restore %esp
876         &add    ($key,$small_footprint?16:160);
877         &xor    ($s0,&DWP(0,$key));
878         &xor    ($s1,&DWP(4,$key));
879         &xor    ($s2,&DWP(8,$key));
880         &xor    ($s3,&DWP(12,$key));
881
882         &mov    ($key,&wparam(1));              # load out
883         &mov    (&DWP(0,$key),$s0);             # write output data
884         &mov    (&DWP(4,$key),$s1);
885         &mov    (&DWP(8,$key),$s2);
886         &mov    (&DWP(12,$key),$s3);
887
888         &pop    ("edi");
889         &pop    ("esi");
890         &pop    ("ebx");
891         &pop    ("ebp");
892         &ret    ();
893
894 &set_label("AES_Td",64);        # Yes! I keep it in the code segment!
895         &data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
896         &data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
897         &data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
898         &data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
899         &data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
900         &data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
901         &data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
902         &data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
903         &data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
904         &data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
905         &data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
906         &data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
907         &data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
908         &data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
909         &data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
910         &data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
911         &data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
912         &data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
913         &data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
914         &data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
915         &data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
916         &data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
917         &data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
918         &data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
919         &data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
920         &data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
921         &data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
922         &data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
923         &data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
924         &data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
925         &data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
926         &data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
927         &data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
928         &data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
929         &data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
930         &data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
931         &data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
932         &data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
933         &data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
934         &data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
935         &data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
936         &data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
937         &data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
938         &data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
939         &data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
940         &data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
941         &data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
942         &data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
943         &data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
944         &data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
945         &data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
946         &data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
947         &data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
948         &data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
949         &data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
950         &data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
951         &data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
952         &data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
953         &data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
954         &data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
955         &data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
956         &data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
957         &data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
958         &data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
959 #Td1:
960         &data_word(0xa7f45150, 0x65417e53, 0xa4171ac3, 0x5e273a96);
961         &data_word(0x6bab3bcb, 0x459d1ff1, 0x58faacab, 0x03e34b93);
962         &data_word(0xfa302055, 0x6d76adf6, 0x76cc8891, 0x4c02f525);
963         &data_word(0xd7e54ffc, 0xcb2ac5d7, 0x44352680, 0xa362b58f);
964         &data_word(0x5ab1de49, 0x1bba2567, 0x0eea4598, 0xc0fe5de1);
965         &data_word(0x752fc302, 0xf04c8112, 0x97468da3, 0xf9d36bc6);
966         &data_word(0x5f8f03e7, 0x9c921595, 0x7a6dbfeb, 0x595295da);
967         &data_word(0x83bed42d, 0x217458d3, 0x69e04929, 0xc8c98e44);
968         &data_word(0x89c2756a, 0x798ef478, 0x3e58996b, 0x71b927dd);
969         &data_word(0x4fe1beb6, 0xad88f017, 0xac20c966, 0x3ace7db4);
970         &data_word(0x4adf6318, 0x311ae582, 0x33519760, 0x7f536245);
971         &data_word(0x7764b1e0, 0xae6bbb84, 0xa081fe1c, 0x2b08f994);
972         &data_word(0x68487058, 0xfd458f19, 0x6cde9487, 0xf87b52b7);
973         &data_word(0xd373ab23, 0x024b72e2, 0x8f1fe357, 0xab55662a);
974         &data_word(0x28ebb207, 0xc2b52f03, 0x7bc5869a, 0x0837d3a5);
975         &data_word(0x872830f2, 0xa5bf23b2, 0x6a0302ba, 0x8216ed5c);
976         &data_word(0x1ccf8a2b, 0xb479a792, 0xf207f3f0, 0xe2694ea1);
977         &data_word(0xf4da65cd, 0xbe0506d5, 0x6234d11f, 0xfea6c48a);
978         &data_word(0x532e349d, 0x55f3a2a0, 0xe18a0532, 0xebf6a475);
979         &data_word(0xec830b39, 0xef6040aa, 0x9f715e06, 0x106ebd51);
980         &data_word(0x8a213ef9, 0x06dd963d, 0x053eddae, 0xbde64d46);
981         &data_word(0x8d5491b5, 0x5dc47105, 0xd406046f, 0x155060ff);
982         &data_word(0xfb981924, 0xe9bdd697, 0x434089cc, 0x9ed96777);
983         &data_word(0x42e8b0bd, 0x8b890788, 0x5b19e738, 0xeec879db);
984         &data_word(0x0a7ca147, 0x0f427ce9, 0x1e84f8c9, 0x00000000);
985         &data_word(0x86800983, 0xed2b3248, 0x70111eac, 0x725a6c4e);
986         &data_word(0xff0efdfb, 0x38850f56, 0xd5ae3d1e, 0x392d3627);
987         &data_word(0xd90f0a64, 0xa65c6821, 0x545b9bd1, 0x2e36243a);
988         &data_word(0x670a0cb1, 0xe757930f, 0x96eeb4d2, 0x919b1b9e);
989         &data_word(0xc5c0804f, 0x20dc61a2, 0x4b775a69, 0x1a121c16);
990         &data_word(0xba93e20a, 0x2aa0c0e5, 0xe0223c43, 0x171b121d);
991         &data_word(0x0d090e0b, 0xc78bf2ad, 0xa8b62db9, 0xa91e14c8);
992         &data_word(0x19f15785, 0x0775af4c, 0xdd99eebb, 0x607fa3fd);
993         &data_word(0x2601f79f, 0xf5725cbc, 0x3b6644c5, 0x7efb5b34);
994         &data_word(0x29438b76, 0xc623cbdc, 0xfcedb668, 0xf1e4b863);
995         &data_word(0xdc31d7ca, 0x85634210, 0x22971340, 0x11c68420);
996         &data_word(0x244a857d, 0x3dbbd2f8, 0x32f9ae11, 0xa129c76d);
997         &data_word(0x2f9e1d4b, 0x30b2dcf3, 0x52860dec, 0xe3c177d0);
998         &data_word(0x16b32b6c, 0xb970a999, 0x489411fa, 0x64e94722);
999         &data_word(0x8cfca8c4, 0x3ff0a01a, 0x2c7d56d8, 0x903322ef);
1000         &data_word(0x4e4987c7, 0xd138d9c1, 0xa2ca8cfe, 0x0bd49836);
1001         &data_word(0x81f5a6cf, 0xde7aa528, 0x8eb7da26, 0xbfad3fa4);
1002         &data_word(0x9d3a2ce4, 0x9278500d, 0xcc5f6a9b, 0x467e5462);
1003         &data_word(0x138df6c2, 0xb8d890e8, 0xf7392e5e, 0xafc382f5);
1004         &data_word(0x805d9fbe, 0x93d0697c, 0x2dd56fa9, 0x1225cfb3);
1005         &data_word(0x99acc83b, 0x7d1810a7, 0x639ce86e, 0xbb3bdb7b);
1006         &data_word(0x7826cd09, 0x18596ef4, 0xb79aec01, 0x9a4f83a8);
1007         &data_word(0x6e95e665, 0xe6ffaa7e, 0xcfbc2108, 0xe815efe6);
1008         &data_word(0x9be7bad9, 0x366f4ace, 0x099fead4, 0x7cb029d6);
1009         &data_word(0xb2a431af, 0x233f2a31, 0x94a5c630, 0x66a235c0);
1010         &data_word(0xbc4e7437, 0xca82fca6, 0xd090e0b0, 0xd8a73315);
1011         &data_word(0x9804f14a, 0xdaec41f7, 0x50cd7f0e, 0xf691172f);
1012         &data_word(0xd64d768d, 0xb0ef434d, 0x4daacc54, 0x0496e4df);
1013         &data_word(0xb5d19ee3, 0x886a4c1b, 0x1f2cc1b8, 0x5165467f);
1014         &data_word(0xea5e9d04, 0x358c015d, 0x7487fa73, 0x410bfb2e);
1015         &data_word(0x1d67b35a, 0xd2db9252, 0x5610e933, 0x47d66d13);
1016         &data_word(0x61d79a8c, 0x0ca1377a, 0x14f8598e, 0x3c13eb89);
1017         &data_word(0x27a9ceee, 0xc961b735, 0xe51ce1ed, 0xb1477a3c);
1018         &data_word(0xdfd29c59, 0x73f2553f, 0xce141879, 0x37c773bf);
1019         &data_word(0xcdf753ea, 0xaafd5f5b, 0x6f3ddf14, 0xdb447886);
1020         &data_word(0xf3afca81, 0xc468b93e, 0x3424382c, 0x40a3c25f);
1021         &data_word(0xc31d1672, 0x25e2bc0c, 0x493c288b, 0x950dff41);
1022         &data_word(0x01a83971, 0xb30c08de, 0xe4b4d89c, 0xc1566490);
1023         &data_word(0x84cb7b61, 0xb632d570, 0x5c6c4874, 0x57b8d042);
1024 #Td2:
1025         &data_word(0xf45150a7, 0x417e5365, 0x171ac3a4, 0x273a965e);
1026         &data_word(0xab3bcb6b, 0x9d1ff145, 0xfaacab58, 0xe34b9303);
1027         &data_word(0x302055fa, 0x76adf66d, 0xcc889176, 0x02f5254c);
1028         &data_word(0xe54ffcd7, 0x2ac5d7cb, 0x35268044, 0x62b58fa3);
1029         &data_word(0xb1de495a, 0xba25671b, 0xea45980e, 0xfe5de1c0);
1030         &data_word(0x2fc30275, 0x4c8112f0, 0x468da397, 0xd36bc6f9);
1031         &data_word(0x8f03e75f, 0x9215959c, 0x6dbfeb7a, 0x5295da59);
1032         &data_word(0xbed42d83, 0x7458d321, 0xe0492969, 0xc98e44c8);
1033         &data_word(0xc2756a89, 0x8ef47879, 0x58996b3e, 0xb927dd71);
1034         &data_word(0xe1beb64f, 0x88f017ad, 0x20c966ac, 0xce7db43a);
1035         &data_word(0xdf63184a, 0x1ae58231, 0x51976033, 0x5362457f);
1036         &data_word(0x64b1e077, 0x6bbb84ae, 0x81fe1ca0, 0x08f9942b);
1037         &data_word(0x48705868, 0x458f19fd, 0xde94876c, 0x7b52b7f8);
1038         &data_word(0x73ab23d3, 0x4b72e202, 0x1fe3578f, 0x55662aab);
1039         &data_word(0xebb20728, 0xb52f03c2, 0xc5869a7b, 0x37d3a508);
1040         &data_word(0x2830f287, 0xbf23b2a5, 0x0302ba6a, 0x16ed5c82);
1041         &data_word(0xcf8a2b1c, 0x79a792b4, 0x07f3f0f2, 0x694ea1e2);
1042         &data_word(0xda65cdf4, 0x0506d5be, 0x34d11f62, 0xa6c48afe);
1043         &data_word(0x2e349d53, 0xf3a2a055, 0x8a0532e1, 0xf6a475eb);
1044         &data_word(0x830b39ec, 0x6040aaef, 0x715e069f, 0x6ebd5110);
1045         &data_word(0x213ef98a, 0xdd963d06, 0x3eddae05, 0xe64d46bd);
1046         &data_word(0x5491b58d, 0xc471055d, 0x06046fd4, 0x5060ff15);
1047         &data_word(0x981924fb, 0xbdd697e9, 0x4089cc43, 0xd967779e);
1048         &data_word(0xe8b0bd42, 0x8907888b, 0x19e7385b, 0xc879dbee);
1049         &data_word(0x7ca1470a, 0x427ce90f, 0x84f8c91e, 0x00000000);
1050         &data_word(0x80098386, 0x2b3248ed, 0x111eac70, 0x5a6c4e72);
1051         &data_word(0x0efdfbff, 0x850f5638, 0xae3d1ed5, 0x2d362739);
1052         &data_word(0x0f0a64d9, 0x5c6821a6, 0x5b9bd154, 0x36243a2e);
1053         &data_word(0x0a0cb167, 0x57930fe7, 0xeeb4d296, 0x9b1b9e91);
1054         &data_word(0xc0804fc5, 0xdc61a220, 0x775a694b, 0x121c161a);
1055         &data_word(0x93e20aba, 0xa0c0e52a, 0x223c43e0, 0x1b121d17);
1056         &data_word(0x090e0b0d, 0x8bf2adc7, 0xb62db9a8, 0x1e14c8a9);
1057         &data_word(0xf1578519, 0x75af4c07, 0x99eebbdd, 0x7fa3fd60);
1058         &data_word(0x01f79f26, 0x725cbcf5, 0x6644c53b, 0xfb5b347e);
1059         &data_word(0x438b7629, 0x23cbdcc6, 0xedb668fc, 0xe4b863f1);
1060         &data_word(0x31d7cadc, 0x63421085, 0x97134022, 0xc6842011);
1061         &data_word(0x4a857d24, 0xbbd2f83d, 0xf9ae1132, 0x29c76da1);
1062         &data_word(0x9e1d4b2f, 0xb2dcf330, 0x860dec52, 0xc177d0e3);
1063         &data_word(0xb32b6c16, 0x70a999b9, 0x9411fa48, 0xe9472264);
1064         &data_word(0xfca8c48c, 0xf0a01a3f, 0x7d56d82c, 0x3322ef90);
1065         &data_word(0x4987c74e, 0x38d9c1d1, 0xca8cfea2, 0xd498360b);
1066         &data_word(0xf5a6cf81, 0x7aa528de, 0xb7da268e, 0xad3fa4bf);
1067         &data_word(0x3a2ce49d, 0x78500d92, 0x5f6a9bcc, 0x7e546246);
1068         &data_word(0x8df6c213, 0xd890e8b8, 0x392e5ef7, 0xc382f5af);
1069         &data_word(0x5d9fbe80, 0xd0697c93, 0xd56fa92d, 0x25cfb312);
1070         &data_word(0xacc83b99, 0x1810a77d, 0x9ce86e63, 0x3bdb7bbb);
1071         &data_word(0x26cd0978, 0x596ef418, 0x9aec01b7, 0x4f83a89a);
1072         &data_word(0x95e6656e, 0xffaa7ee6, 0xbc2108cf, 0x15efe6e8);
1073         &data_word(0xe7bad99b, 0x6f4ace36, 0x9fead409, 0xb029d67c);
1074         &data_word(0xa431afb2, 0x3f2a3123, 0xa5c63094, 0xa235c066);
1075         &data_word(0x4e7437bc, 0x82fca6ca, 0x90e0b0d0, 0xa73315d8);
1076         &data_word(0x04f14a98, 0xec41f7da, 0xcd7f0e50, 0x91172ff6);
1077         &data_word(0x4d768dd6, 0xef434db0, 0xaacc544d, 0x96e4df04);
1078         &data_word(0xd19ee3b5, 0x6a4c1b88, 0x2cc1b81f, 0x65467f51);
1079         &data_word(0x5e9d04ea, 0x8c015d35, 0x87fa7374, 0x0bfb2e41);
1080         &data_word(0x67b35a1d, 0xdb9252d2, 0x10e93356, 0xd66d1347);
1081         &data_word(0xd79a8c61, 0xa1377a0c, 0xf8598e14, 0x13eb893c);
1082         &data_word(0xa9ceee27, 0x61b735c9, 0x1ce1ede5, 0x477a3cb1);
1083         &data_word(0xd29c59df, 0xf2553f73, 0x141879ce, 0xc773bf37);
1084         &data_word(0xf753eacd, 0xfd5f5baa, 0x3ddf146f, 0x447886db);
1085         &data_word(0xafca81f3, 0x68b93ec4, 0x24382c34, 0xa3c25f40);
1086         &data_word(0x1d1672c3, 0xe2bc0c25, 0x3c288b49, 0x0dff4195);
1087         &data_word(0xa8397101, 0x0c08deb3, 0xb4d89ce4, 0x566490c1);
1088         &data_word(0xcb7b6184, 0x32d570b6, 0x6c48745c, 0xb8d04257);
1089 #Td3:
1090         &data_word(0x5150a7f4, 0x7e536541, 0x1ac3a417, 0x3a965e27);
1091         &data_word(0x3bcb6bab, 0x1ff1459d, 0xacab58fa, 0x4b9303e3);
1092         &data_word(0x2055fa30, 0xadf66d76, 0x889176cc, 0xf5254c02);
1093         &data_word(0x4ffcd7e5, 0xc5d7cb2a, 0x26804435, 0xb58fa362);
1094         &data_word(0xde495ab1, 0x25671bba, 0x45980eea, 0x5de1c0fe);
1095         &data_word(0xc302752f, 0x8112f04c, 0x8da39746, 0x6bc6f9d3);
1096         &data_word(0x03e75f8f, 0x15959c92, 0xbfeb7a6d, 0x95da5952);
1097         &data_word(0xd42d83be, 0x58d32174, 0x492969e0, 0x8e44c8c9);
1098         &data_word(0x756a89c2, 0xf478798e, 0x996b3e58, 0x27dd71b9);
1099         &data_word(0xbeb64fe1, 0xf017ad88, 0xc966ac20, 0x7db43ace);
1100         &data_word(0x63184adf, 0xe582311a, 0x97603351, 0x62457f53);
1101         &data_word(0xb1e07764, 0xbb84ae6b, 0xfe1ca081, 0xf9942b08);
1102         &data_word(0x70586848, 0x8f19fd45, 0x94876cde, 0x52b7f87b);
1103         &data_word(0xab23d373, 0x72e2024b, 0xe3578f1f, 0x662aab55);
1104         &data_word(0xb20728eb, 0x2f03c2b5, 0x869a7bc5, 0xd3a50837);
1105         &data_word(0x30f28728, 0x23b2a5bf, 0x02ba6a03, 0xed5c8216);
1106         &data_word(0x8a2b1ccf, 0xa792b479, 0xf3f0f207, 0x4ea1e269);
1107         &data_word(0x65cdf4da, 0x06d5be05, 0xd11f6234, 0xc48afea6);
1108         &data_word(0x349d532e, 0xa2a055f3, 0x0532e18a, 0xa475ebf6);
1109         &data_word(0x0b39ec83, 0x40aaef60, 0x5e069f71, 0xbd51106e);
1110         &data_word(0x3ef98a21, 0x963d06dd, 0xddae053e, 0x4d46bde6);
1111         &data_word(0x91b58d54, 0x71055dc4, 0x046fd406, 0x60ff1550);
1112         &data_word(0x1924fb98, 0xd697e9bd, 0x89cc4340, 0x67779ed9);
1113         &data_word(0xb0bd42e8, 0x07888b89, 0xe7385b19, 0x79dbeec8);
1114         &data_word(0xa1470a7c, 0x7ce90f42, 0xf8c91e84, 0x00000000);
1115         &data_word(0x09838680, 0x3248ed2b, 0x1eac7011, 0x6c4e725a);
1116         &data_word(0xfdfbff0e, 0x0f563885, 0x3d1ed5ae, 0x3627392d);
1117         &data_word(0x0a64d90f, 0x6821a65c, 0x9bd1545b, 0x243a2e36);
1118         &data_word(0x0cb1670a, 0x930fe757, 0xb4d296ee, 0x1b9e919b);
1119         &data_word(0x804fc5c0, 0x61a220dc, 0x5a694b77, 0x1c161a12);
1120         &data_word(0xe20aba93, 0xc0e52aa0, 0x3c43e022, 0x121d171b);
1121         &data_word(0x0e0b0d09, 0xf2adc78b, 0x2db9a8b6, 0x14c8a91e);
1122         &data_word(0x578519f1, 0xaf4c0775, 0xeebbdd99, 0xa3fd607f);
1123         &data_word(0xf79f2601, 0x5cbcf572, 0x44c53b66, 0x5b347efb);
1124         &data_word(0x8b762943, 0xcbdcc623, 0xb668fced, 0xb863f1e4);
1125         &data_word(0xd7cadc31, 0x42108563, 0x13402297, 0x842011c6);
1126         &data_word(0x857d244a, 0xd2f83dbb, 0xae1132f9, 0xc76da129);
1127         &data_word(0x1d4b2f9e, 0xdcf330b2, 0x0dec5286, 0x77d0e3c1);
1128         &data_word(0x2b6c16b3, 0xa999b970, 0x11fa4894, 0x472264e9);
1129         &data_word(0xa8c48cfc, 0xa01a3ff0, 0x56d82c7d, 0x22ef9033);
1130         &data_word(0x87c74e49, 0xd9c1d138, 0x8cfea2ca, 0x98360bd4);
1131         &data_word(0xa6cf81f5, 0xa528de7a, 0xda268eb7, 0x3fa4bfad);
1132         &data_word(0x2ce49d3a, 0x500d9278, 0x6a9bcc5f, 0x5462467e);
1133         &data_word(0xf6c2138d, 0x90e8b8d8, 0x2e5ef739, 0x82f5afc3);
1134         &data_word(0x9fbe805d, 0x697c93d0, 0x6fa92dd5, 0xcfb31225);
1135         &data_word(0xc83b99ac, 0x10a77d18, 0xe86e639c, 0xdb7bbb3b);
1136         &data_word(0xcd097826, 0x6ef41859, 0xec01b79a, 0x83a89a4f);
1137         &data_word(0xe6656e95, 0xaa7ee6ff, 0x2108cfbc, 0xefe6e815);
1138         &data_word(0xbad99be7, 0x4ace366f, 0xead4099f, 0x29d67cb0);
1139         &data_word(0x31afb2a4, 0x2a31233f, 0xc63094a5, 0x35c066a2);
1140         &data_word(0x7437bc4e, 0xfca6ca82, 0xe0b0d090, 0x3315d8a7);
1141         &data_word(0xf14a9804, 0x41f7daec, 0x7f0e50cd, 0x172ff691);
1142         &data_word(0x768dd64d, 0x434db0ef, 0xcc544daa, 0xe4df0496);
1143         &data_word(0x9ee3b5d1, 0x4c1b886a, 0xc1b81f2c, 0x467f5165);
1144         &data_word(0x9d04ea5e, 0x015d358c, 0xfa737487, 0xfb2e410b);
1145         &data_word(0xb35a1d67, 0x9252d2db, 0xe9335610, 0x6d1347d6);
1146         &data_word(0x9a8c61d7, 0x377a0ca1, 0x598e14f8, 0xeb893c13);
1147         &data_word(0xceee27a9, 0xb735c961, 0xe1ede51c, 0x7a3cb147);
1148         &data_word(0x9c59dfd2, 0x553f73f2, 0x1879ce14, 0x73bf37c7);
1149         &data_word(0x53eacdf7, 0x5f5baafd, 0xdf146f3d, 0x7886db44);
1150         &data_word(0xca81f3af, 0xb93ec468, 0x382c3424, 0xc25f40a3);
1151         &data_word(0x1672c31d, 0xbc0c25e2, 0x288b493c, 0xff41950d);
1152         &data_word(0x397101a8, 0x08deb30c, 0xd89ce4b4, 0x6490c156);
1153         &data_word(0x7b6184cb, 0xd570b632, 0x48745c6c, 0xd04257b8);
1154 #Td4:
1155         &data_word(0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5);
1156         &data_word(0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838);
1157         &data_word(0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e);
1158         &data_word(0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb);
1159         &data_word(0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282);
1160         &data_word(0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787);
1161         &data_word(0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444);
1162         &data_word(0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb);
1163         &data_word(0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232);
1164         &data_word(0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d);
1165         &data_word(0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b);
1166         &data_word(0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e);
1167         &data_word(0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666);
1168         &data_word(0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2);
1169         &data_word(0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949);
1170         &data_word(0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525);
1171         &data_word(0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464);
1172         &data_word(0x86868686, 0x68686868, 0x98989898, 0x16161616);
1173         &data_word(0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc);
1174         &data_word(0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292);
1175         &data_word(0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050);
1176         &data_word(0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada);
1177         &data_word(0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757);
1178         &data_word(0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484);
1179         &data_word(0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000);
1180         &data_word(0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a);
1181         &data_word(0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505);
1182         &data_word(0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606);
1183         &data_word(0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f);
1184         &data_word(0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202);
1185         &data_word(0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303);
1186         &data_word(0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b);
1187         &data_word(0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141);
1188         &data_word(0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea);
1189         &data_word(0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece);
1190         &data_word(0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373);
1191         &data_word(0x96969696, 0xacacacac, 0x74747474, 0x22222222);
1192         &data_word(0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585);
1193         &data_word(0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8);
1194         &data_word(0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e);
1195         &data_word(0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171);
1196         &data_word(0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989);
1197         &data_word(0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e);
1198         &data_word(0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b);
1199         &data_word(0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b);
1200         &data_word(0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020);
1201         &data_word(0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe);
1202         &data_word(0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4);
1203         &data_word(0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333);
1204         &data_word(0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131);
1205         &data_word(0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959);
1206         &data_word(0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f);
1207         &data_word(0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9);
1208         &data_word(0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d);
1209         &data_word(0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f);
1210         &data_word(0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef);
1211         &data_word(0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d);
1212         &data_word(0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0);
1213         &data_word(0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c);
1214         &data_word(0x83838383, 0x53535353, 0x99999999, 0x61616161);
1215         &data_word(0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e);
1216         &data_word(0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626);
1217         &data_word(0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363);
1218         &data_word(0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d);
1219 &function_end_B("AES_decrypt");
1220
1221 sub enckey()
1222 {
1223         &movz   ("esi",&LB("edx"));             # rk[i]>>0
1224         &mov    ("ebx",&DWP(0,"ebp","esi",4));
1225         &movz   ("esi",&HB("edx"));             # rk[i]>>8
1226         &and    ("ebx",0xFF000000);
1227         &xor    ("eax","ebx");
1228
1229         &mov    ("ebx",&DWP(0,"ebp","esi",4));
1230         &shr    ("edx",16);
1231         &and    ("ebx",0x000000FF);
1232         &movz   ("esi",&LB("edx"));             # rk[i]>>16
1233         &xor    ("eax","ebx");
1234
1235         &mov    ("ebx",&DWP(0,"ebp","esi",4));
1236         &movz   ("esi",&HB("edx"));             # rk[i]>>24
1237         &and    ("ebx",0x0000FF00);
1238         &xor    ("eax","ebx");
1239
1240         &mov    ("ebx",&DWP(0,"ebp","esi",4));
1241         &and    ("ebx",0x00FF0000);
1242         &xor    ("eax","ebx");
1243
1244         &xor    ("eax",&DWP(1024,"ebp","ecx",4));       # rcon
1245 }
1246
1247 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
1248 #                        AES_KEY *key)
1249 &public_label("AES_Te");
1250 &function_begin("AES_set_encrypt_key");
1251         &mov    ("esi",&wparam(0));             # user supplied key
1252         &mov    ("edi",&wparam(2));             # private key schedule
1253
1254         &test   ("esi",-1);
1255         &jz     (&label("badpointer"));
1256         &test   ("edi",-1);
1257         &jz     (&label("badpointer"));
1258
1259         &call   (&label("pic_point"));
1260         &set_label("pic_point");
1261         &blindpop("ebp");
1262         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1263         &add    ("ebp",1024*4);                 # skip to Te4
1264
1265         &mov    ("ecx",&wparam(1));             # number of bits in key
1266         &cmp    ("ecx",128);
1267         &je     (&label("10rounds"));
1268         &cmp    ("ecx",192);
1269         &je     (&label("12rounds"));
1270         &cmp    ("ecx",256);
1271         &je     (&label("14rounds"));
1272         &mov    ("eax",-2);                     # invalid number of bits
1273         &jmp    (&label("exit"));
1274
1275     &set_label("10rounds");
1276         &mov    ("eax",&DWP(0,"esi"));          # copy first 4 dwords
1277         &mov    ("ebx",&DWP(4,"esi"));
1278         &mov    ("ecx",&DWP(8,"esi"));
1279         &mov    ("edx",&DWP(12,"esi"));
1280         &mov    (&DWP(0,"edi"),"eax");
1281         &mov    (&DWP(4,"edi"),"ebx");
1282         &mov    (&DWP(8,"edi"),"ecx");
1283         &mov    (&DWP(12,"edi"),"edx");
1284
1285         &xor    ("ecx","ecx");
1286         &jmp    (&label("10shortcut"));
1287
1288         &align  (4);
1289         &set_label("10loop");
1290                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1291                 &mov    ("edx",&DWP(12,"edi"));         # rk[3]
1292         &set_label("10shortcut");
1293                 &enckey ();
1294
1295                 &mov    (&DWP(16,"edi"),"eax");         # rk[4]
1296                 &xor    ("eax",&DWP(4,"edi"));
1297                 &mov    (&DWP(20,"edi"),"eax");         # rk[5]
1298                 &xor    ("eax",&DWP(8,"edi"));
1299                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1300                 &xor    ("eax",&DWP(12,"edi"));
1301                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1302                 &inc    ("ecx");
1303                 &add    ("edi",16);
1304                 &cmp    ("ecx",10);
1305         &jl     (&label("10loop"));
1306
1307         &mov    (&DWP(80,"edi"),10);            # setup number of rounds
1308         &xor    ("eax","eax");
1309         &jmp    (&label("exit"));
1310                 
1311     &set_label("12rounds");
1312         &mov    ("eax",&DWP(0,"esi"));          # copy first 6 dwords
1313         &mov    ("ebx",&DWP(4,"esi"));
1314         &mov    ("ecx",&DWP(8,"esi"));
1315         &mov    ("edx",&DWP(12,"esi"));
1316         &mov    (&DWP(0,"edi"),"eax");
1317         &mov    (&DWP(4,"edi"),"ebx");
1318         &mov    (&DWP(8,"edi"),"ecx");
1319         &mov    (&DWP(12,"edi"),"edx");
1320         &mov    ("ecx",&DWP(16,"esi"));
1321         &mov    ("edx",&DWP(20,"esi"));
1322         &mov    (&DWP(16,"edi"),"ecx");
1323         &mov    (&DWP(20,"edi"),"edx");
1324
1325         &xor    ("ecx","ecx");
1326         &jmp    (&label("12shortcut"));
1327
1328         &align  (4);
1329         &set_label("12loop");
1330                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1331                 &mov    ("edx",&DWP(20,"edi"));         # rk[5]
1332         &set_label("12shortcut");
1333                 &enckey ();
1334
1335                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1336                 &xor    ("eax",&DWP(4,"edi"));
1337                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1338                 &xor    ("eax",&DWP(8,"edi"));
1339                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1340                 &xor    ("eax",&DWP(12,"edi"));
1341                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1342
1343                 &cmp    ("ecx",7);
1344                 &je     (&label("12break"));
1345                 &inc    ("ecx");
1346
1347                 &xor    ("eax",&DWP(16,"edi"));
1348                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1349                 &xor    ("eax",&DWP(20,"edi"));
1350                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1351
1352                 &add    ("edi",24);
1353         &jmp    (&label("12loop"));
1354
1355         &set_label("12break");
1356         &mov    (&DWP(72,"edi"),12);            # setup number of rounds
1357         &xor    ("eax","eax");
1358         &jmp    (&label("exit"));
1359
1360     &set_label("14rounds");
1361         &mov    ("eax",&DWP(0,"esi"));          # copy first 8 dwords
1362         &mov    ("ebx",&DWP(4,"esi"));
1363         &mov    ("ecx",&DWP(8,"esi"));
1364         &mov    ("edx",&DWP(12,"esi"));
1365         &mov    (&DWP(0,"edi"),"eax");
1366         &mov    (&DWP(4,"edi"),"ebx");
1367         &mov    (&DWP(8,"edi"),"ecx");
1368         &mov    (&DWP(12,"edi"),"edx");
1369         &mov    ("eax",&DWP(16,"esi"));
1370         &mov    ("ebx",&DWP(20,"esi"));
1371         &mov    ("ecx",&DWP(24,"esi"));
1372         &mov    ("edx",&DWP(28,"esi"));
1373         &mov    (&DWP(16,"edi"),"eax");
1374         &mov    (&DWP(20,"edi"),"ebx");
1375         &mov    (&DWP(24,"edi"),"ecx");
1376         &mov    (&DWP(28,"edi"),"edx");
1377
1378         &xor    ("ecx","ecx");
1379         &jmp    (&label("14shortcut"));
1380
1381         &align  (4);
1382         &set_label("14loop");
1383                 &mov    ("edx",&DWP(28,"edi"));         # rk[7]
1384         &set_label("14shortcut");
1385                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1386
1387                 &enckey ();
1388
1389                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1390                 &xor    ("eax",&DWP(4,"edi"));
1391                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1392                 &xor    ("eax",&DWP(8,"edi"));
1393                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1394                 &xor    ("eax",&DWP(12,"edi"));
1395                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1396
1397                 &cmp    ("ecx",6);
1398                 &je     (&label("14break"));
1399                 &inc    ("ecx");
1400
1401                 &mov    ("edx","eax");
1402                 &mov    ("eax",&DWP(16,"edi"));         # rk[4]
1403                 &movz   ("esi",&LB("edx"));             # rk[11]>>0
1404                 &mov    ("ebx",&DWP(0,"ebp","esi",4));
1405                 &movz   ("esi",&HB("edx"));             # rk[11]>>8
1406                 &and    ("ebx",0x000000FF);
1407                 &xor    ("eax","ebx");
1408
1409                 &mov    ("ebx",&DWP(0,"ebp","esi",4));
1410                 &shr    ("edx",16);
1411                 &and    ("ebx",0x0000FF00);
1412                 &movz   ("esi",&LB("edx"));             # rk[11]>>16
1413                 &xor    ("eax","ebx");
1414
1415                 &mov    ("ebx",&DWP(0,"ebp","esi",4));
1416                 &movz   ("esi",&HB("edx"));             # rk[11]>>24
1417                 &and    ("ebx",0x00FF0000);
1418                 &xor    ("eax","ebx");
1419
1420                 &mov    ("ebx",&DWP(0,"ebp","esi",4));
1421                 &and    ("ebx",0xFF000000);
1422                 &xor    ("eax","ebx");
1423
1424                 &mov    (&DWP(48,"edi"),"eax");         # rk[12]
1425                 &xor    ("eax",&DWP(20,"edi"));
1426                 &mov    (&DWP(52,"edi"),"eax");         # rk[13]
1427                 &xor    ("eax",&DWP(24,"edi"));
1428                 &mov    (&DWP(56,"edi"),"eax");         # rk[14]
1429                 &xor    ("eax",&DWP(28,"edi"));
1430                 &mov    (&DWP(60,"edi"),"eax");         # rk[15]
1431
1432                 &add    ("edi",32);
1433         &jmp    (&label("14loop"));
1434
1435         &set_label("14break");
1436         &mov    (&DWP(48,"edi"),14);            # setup number of rounds
1437         &xor    ("eax","eax");
1438         &jmp    (&label("exit"));
1439
1440     &set_label("badpointer");
1441         &mov    ("eax",-1);
1442     &set_label("exit");
1443 &function_end("AES_set_encrypt_key");
1444
1445 sub deckey()
1446 { my ($i,$ptr,$te4,$td) = @_;
1447
1448         &mov    ("eax",&DWP($i,$ptr));
1449         &mov    ("edx","eax");
1450         &movz   ("ebx",&HB("eax"));
1451         &shr    ("edx",16);
1452         &and    ("eax",0xFF);
1453         &movz   ("eax",&BP(0,$te4,"eax",4));
1454         &movz   ("ebx",&BP(0,$te4,"ebx",4));
1455         &mov    ("eax",&DWP(1024*0,$td,"eax",4));
1456         &xor    ("eax",&DWP(1024*1,$td,"ebx",4));
1457         &movz   ("ebx",&HB("edx"));
1458         &and    ("edx",0xFF);
1459         &movz   ("edx",&BP(0,$te4,"edx",4));
1460         &movz   ("ebx",&BP(0,$te4,"ebx",4));
1461         &xor    ("eax",&DWP(1024*2,$td,"edx",4));
1462         &xor    ("eax",&DWP(1024*3,$td,"ebx",4));
1463         &mov    (&DWP($i,$ptr),"eax");
1464 }
1465
1466 # int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
1467 #                        AES_KEY *key)
1468 &public_label("AES_Td");
1469 &public_label("AES_Te");
1470 &function_begin_B("AES_set_decrypt_key");
1471         &mov    ("eax",&wparam(0));
1472         &mov    ("ecx",&wparam(1));
1473         &mov    ("edx",&wparam(2));
1474         &sub    ("esp",12);
1475         &mov    (&DWP(0,"esp"),"eax");
1476         &mov    (&DWP(4,"esp"),"ecx");
1477         &mov    (&DWP(8,"esp"),"edx");
1478         &call   ("AES_set_encrypt_key");
1479         &add    ("esp",12);
1480         &cmp    ("eax",0);
1481         &je     (&label("proceed"));
1482         &ret    ();
1483
1484     &set_label("proceed");
1485         &push   ("ebp");
1486         &push   ("ebx");
1487         &push   ("esi");
1488         &push   ("edi");
1489
1490         &mov    ("esi",&wparam(2));
1491         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1492         &lea    ("ecx",&DWP(0,"","ecx",4));
1493         &lea    ("edi",&DWP(0,"esi","ecx",4));  # pointer to last chunk
1494
1495         &align  (4);
1496         &set_label("invert");                   # invert order of chunks
1497                 &mov    ("eax",&DWP(0,"esi"));
1498                 &mov    ("ebx",&DWP(4,"esi"));
1499                 &mov    ("ecx",&DWP(0,"edi"));
1500                 &mov    ("edx",&DWP(4,"edi"));
1501                 &mov    (&DWP(0,"edi"),"eax");
1502                 &mov    (&DWP(4,"edi"),"ebx");
1503                 &mov    (&DWP(0,"esi"),"ecx");
1504                 &mov    (&DWP(4,"esi"),"edx");
1505                 &mov    ("eax",&DWP(8,"esi"));
1506                 &mov    ("ebx",&DWP(12,"esi"));
1507                 &mov    ("ecx",&DWP(8,"edi"));
1508                 &mov    ("edx",&DWP(12,"edi"));
1509                 &mov    (&DWP(8,"edi"),"eax");
1510                 &mov    (&DWP(12,"edi"),"ebx");
1511                 &mov    (&DWP(8,"esi"),"ecx");
1512                 &mov    (&DWP(12,"esi"),"edx");
1513                 &add    ("esi",16);
1514                 &sub    ("edi",16);
1515                 &cmp    ("esi","edi");
1516         &jne    (&label("invert"));
1517
1518         &call   (&label("pic_point"));
1519         &set_label("pic_point");
1520         blindpop("ebp");
1521         &lea    ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
1522         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1523         &add    ("ebp",1024*4);                 # skip to Te4
1524
1525         &mov    ("esi",&wparam(2));
1526         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1527         &dec    ("ecx");
1528         &align  (4);
1529         &set_label("permute");                  # permute the key schedule
1530                 &add    ("esi",16);
1531                 &deckey (0,"esi","ebp","edi");
1532                 &deckey (4,"esi","ebp","edi");
1533                 &deckey (8,"esi","ebp","edi");
1534                 &deckey (12,"esi","ebp","edi");
1535                 &dec    ("ecx");
1536         &jnz    (&label("permute"));
1537
1538         &xor    ("eax","eax");                  # return success
1539 &function_end("AES_set_decrypt_key");
1540
1541 &asm_finish();