des-586.pl 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453
  1. #!/usr/local/bin/perl
  2. #
  3. # The inner loop instruction sequence and the IP/FP modifications are from
  4. # Svend Olaf Mikkelsen <[email protected]>
  5. #
  6. $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  7. push(@INC,"${dir}","${dir}../../perlasm");
  8. require "x86asm.pl";
  9. require "cbc.pl";
  10. require "desboth.pl";
  11. # base code is in microsft
  12. # op dest, source
  13. # format.
  14. #
  15. &asm_init($ARGV[0],"des-586.pl");
  16. $L="edi";
  17. $R="esi";
  18. $trans="ebp";
  19. $small_footprint=1 if (grep(/\-DOPENSSL_SMALL_FOOTPRINT/,@ARGV));
  20. # one can discuss setting this variable to 1 unconditionally, as
  21. # the folded loop is only 3% slower than unrolled, but >7 times smaller
  22. &public_label("DES_SPtrans");
  23. &DES_encrypt_internal();
  24. &DES_decrypt_internal();
  25. &DES_encrypt("DES_encrypt1",1);
  26. &DES_encrypt("DES_encrypt2",0);
  27. &DES_encrypt3("DES_encrypt3",1);
  28. &DES_encrypt3("DES_decrypt3",0);
  29. &cbc("DES_ncbc_encrypt","DES_encrypt1","DES_encrypt1",0,4,5,3,5,-1);
  30. &cbc("DES_ede3_cbc_encrypt","DES_encrypt3","DES_decrypt3",0,6,7,3,4,5);
  31. &DES_SPtrans();
  32. &asm_finish();
  33. sub DES_encrypt_internal()
  34. {
  35. &function_begin_B("_x86_DES_encrypt");
  36. if ($small_footprint)
  37. {
  38. &lea("edx",&DWP(128,"ecx"));
  39. &push("edx");
  40. &push("ecx");
  41. &set_label("eloop");
  42. &D_ENCRYPT(0,$L,$R,0,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  43. &comment("");
  44. &D_ENCRYPT(1,$R,$L,2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  45. &comment("");
  46. &add("ecx",16);
  47. &cmp("ecx",&swtmp(1));
  48. &mov(&swtmp(0),"ecx");
  49. &jb(&label("eloop"));
  50. &add("esp",8);
  51. }
  52. else
  53. {
  54. &push("ecx");
  55. for ($i=0; $i<16; $i+=2)
  56. {
  57. &comment("Round $i");
  58. &D_ENCRYPT($i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  59. &comment("Round ".sprintf("%d",$i+1));
  60. &D_ENCRYPT($i+1,$R,$L,($i+1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  61. }
  62. &add("esp",4);
  63. }
  64. &ret();
  65. &function_end_B("_x86_DES_encrypt");
  66. }
  67. sub DES_decrypt_internal()
  68. {
  69. &function_begin_B("_x86_DES_decrypt");
  70. if ($small_footprint)
  71. {
  72. &push("ecx");
  73. &lea("ecx",&DWP(128,"ecx"));
  74. &push("ecx");
  75. &set_label("dloop");
  76. &D_ENCRYPT(0,$L,$R,-2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  77. &comment("");
  78. &D_ENCRYPT(1,$R,$L,-4,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  79. &comment("");
  80. &sub("ecx",16);
  81. &cmp("ecx",&swtmp(1));
  82. &mov(&swtmp(0),"ecx");
  83. &ja(&label("dloop"));
  84. &add("esp",8);
  85. }
  86. else
  87. {
  88. &push("ecx");
  89. for ($i=15; $i>0; $i-=2)
  90. {
  91. &comment("Round $i");
  92. &D_ENCRYPT(15-$i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  93. &comment("Round ".sprintf("%d",$i-1));
  94. &D_ENCRYPT(15-$i+1,$R,$L,($i-1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  95. }
  96. &add("esp",4);
  97. }
  98. &ret();
  99. &function_end_B("_x86_DES_decrypt");
  100. }
  101. sub DES_encrypt
  102. {
  103. local($name,$do_ip)=@_;
  104. &function_begin_B($name);
  105. &push("esi");
  106. &push("edi");
  107. &comment("");
  108. &comment("Load the 2 words");
  109. if ($do_ip)
  110. {
  111. &mov($R,&wparam(0));
  112. &xor( "ecx", "ecx" );
  113. &push("ebx");
  114. &push("ebp");
  115. &mov("eax",&DWP(0,$R,"",0));
  116. &mov("ebx",&wparam(2)); # get encrypt flag
  117. &mov($L,&DWP(4,$R,"",0));
  118. &comment("");
  119. &comment("IP");
  120. &IP_new("eax",$L,$R,3);
  121. }
  122. else
  123. {
  124. &mov("eax",&wparam(0));
  125. &xor( "ecx", "ecx" );
  126. &push("ebx");
  127. &push("ebp");
  128. &mov($R,&DWP(0,"eax","",0));
  129. &mov("ebx",&wparam(2)); # get encrypt flag
  130. &rotl($R,3);
  131. &mov($L,&DWP(4,"eax","",0));
  132. &rotl($L,3);
  133. }
  134. # PIC-ification:-)
  135. &call (&label("pic_point"));
  136. &set_label("pic_point");
  137. &blindpop($trans);
  138. &lea ($trans,&DWP(&label("DES_SPtrans")."-".&label("pic_point"),$trans));
  139. &mov( "ecx", &wparam(1) );
  140. &cmp("ebx","0");
  141. &je(&label("decrypt"));
  142. &call("_x86_DES_encrypt");
  143. &jmp(&label("done"));
  144. &set_label("decrypt");
  145. &call("_x86_DES_decrypt");
  146. &set_label("done");
  147. if ($do_ip)
  148. {
  149. &comment("");
  150. &comment("FP");
  151. &mov("edx",&wparam(0));
  152. &FP_new($L,$R,"eax",3);
  153. &mov(&DWP(0,"edx","",0),"eax");
  154. &mov(&DWP(4,"edx","",0),$R);
  155. }
  156. else
  157. {
  158. &comment("");
  159. &comment("Fixup");
  160. &rotr($L,3); # r
  161. &mov("eax",&wparam(0));
  162. &rotr($R,3); # l
  163. &mov(&DWP(0,"eax","",0),$L);
  164. &mov(&DWP(4,"eax","",0),$R);
  165. }
  166. &pop("ebp");
  167. &pop("ebx");
  168. &pop("edi");
  169. &pop("esi");
  170. &ret();
  171. &function_end_B($name);
  172. }
  173. sub D_ENCRYPT
  174. {
  175. local($r,$L,$R,$S,$trans,$u,$tmp1,$tmp2,$t,$wp1)=@_;
  176. &mov( $u, &DWP(&n2a($S*4),$tmp2,"",0));
  177. &xor( $tmp1, $tmp1);
  178. &mov( $t, &DWP(&n2a(($S+1)*4),$tmp2,"",0));
  179. &xor( $u, $R);
  180. &xor( $tmp2, $tmp2);
  181. &xor( $t, $R);
  182. &and( $u, "0xfcfcfcfc" );
  183. &and( $t, "0xcfcfcfcf" );
  184. &movb( &LB($tmp1), &LB($u) );
  185. &movb( &LB($tmp2), &HB($u) );
  186. &rotr( $t, 4 );
  187. &xor( $L, &DWP(" ",$trans,$tmp1,0));
  188. &movb( &LB($tmp1), &LB($t) );
  189. &xor( $L, &DWP("0x200",$trans,$tmp2,0));
  190. &movb( &LB($tmp2), &HB($t) );
  191. &shr( $u, 16);
  192. &xor( $L, &DWP("0x100",$trans,$tmp1,0));
  193. &movb( &LB($tmp1), &HB($u) );
  194. &shr( $t, 16);
  195. &xor( $L, &DWP("0x300",$trans,$tmp2,0));
  196. &movb( &LB($tmp2), &HB($t) );
  197. &and( $u, "0xff" );
  198. &and( $t, "0xff" );
  199. &xor( $L, &DWP("0x600",$trans,$tmp1,0));
  200. &xor( $L, &DWP("0x700",$trans,$tmp2,0));
  201. &mov( $tmp2, $wp1 );
  202. &xor( $L, &DWP("0x400",$trans,$u,0));
  203. &xor( $L, &DWP("0x500",$trans,$t,0));
  204. }
  205. sub n2a
  206. {
  207. sprintf("%d",$_[0]);
  208. }
  209. # now has a side affect of rotating $a by $shift
  210. sub R_PERM_OP
  211. {
  212. local($a,$b,$tt,$shift,$mask,$last)=@_;
  213. &rotl( $a, $shift ) if ($shift != 0);
  214. &mov( $tt, $a );
  215. &xor( $a, $b );
  216. &and( $a, $mask );
  217. # This can never succeed, and besides it is difficult to see what the
  218. # idea was - Ben 13 Feb 99
  219. if (!$last eq $b)
  220. {
  221. &xor( $b, $a );
  222. &xor( $tt, $a );
  223. }
  224. else
  225. {
  226. &xor( $tt, $a );
  227. &xor( $b, $a );
  228. }
  229. &comment("");
  230. }
  231. sub IP_new
  232. {
  233. local($l,$r,$tt,$lr)=@_;
  234. &R_PERM_OP($l,$r,$tt, 4,"0xf0f0f0f0",$l);
  235. &R_PERM_OP($r,$tt,$l,20,"0xfff0000f",$l);
  236. &R_PERM_OP($l,$tt,$r,14,"0x33333333",$r);
  237. &R_PERM_OP($tt,$r,$l,22,"0x03fc03fc",$r);
  238. &R_PERM_OP($l,$r,$tt, 9,"0xaaaaaaaa",$r);
  239. if ($lr != 3)
  240. {
  241. if (($lr-3) < 0)
  242. { &rotr($tt, 3-$lr); }
  243. else { &rotl($tt, $lr-3); }
  244. }
  245. if ($lr != 2)
  246. {
  247. if (($lr-2) < 0)
  248. { &rotr($r, 2-$lr); }
  249. else { &rotl($r, $lr-2); }
  250. }
  251. }
  252. sub FP_new
  253. {
  254. local($l,$r,$tt,$lr)=@_;
  255. if ($lr != 2)
  256. {
  257. if (($lr-2) < 0)
  258. { &rotl($r, 2-$lr); }
  259. else { &rotr($r, $lr-2); }
  260. }
  261. if ($lr != 3)
  262. {
  263. if (($lr-3) < 0)
  264. { &rotl($l, 3-$lr); }
  265. else { &rotr($l, $lr-3); }
  266. }
  267. &R_PERM_OP($l,$r,$tt, 0,"0xaaaaaaaa",$r);
  268. &R_PERM_OP($tt,$r,$l,23,"0x03fc03fc",$r);
  269. &R_PERM_OP($l,$r,$tt,10,"0x33333333",$l);
  270. &R_PERM_OP($r,$tt,$l,18,"0xfff0000f",$l);
  271. &R_PERM_OP($l,$tt,$r,12,"0xf0f0f0f0",$r);
  272. &rotr($tt , 4);
  273. }
  274. sub DES_SPtrans
  275. {
  276. &set_label("DES_SPtrans",64);
  277. &data_word(0x02080800, 0x00080000, 0x02000002, 0x02080802);
  278. &data_word(0x02000000, 0x00080802, 0x00080002, 0x02000002);
  279. &data_word(0x00080802, 0x02080800, 0x02080000, 0x00000802);
  280. &data_word(0x02000802, 0x02000000, 0x00000000, 0x00080002);
  281. &data_word(0x00080000, 0x00000002, 0x02000800, 0x00080800);
  282. &data_word(0x02080802, 0x02080000, 0x00000802, 0x02000800);
  283. &data_word(0x00000002, 0x00000800, 0x00080800, 0x02080002);
  284. &data_word(0x00000800, 0x02000802, 0x02080002, 0x00000000);
  285. &data_word(0x00000000, 0x02080802, 0x02000800, 0x00080002);
  286. &data_word(0x02080800, 0x00080000, 0x00000802, 0x02000800);
  287. &data_word(0x02080002, 0x00000800, 0x00080800, 0x02000002);
  288. &data_word(0x00080802, 0x00000002, 0x02000002, 0x02080000);
  289. &data_word(0x02080802, 0x00080800, 0x02080000, 0x02000802);
  290. &data_word(0x02000000, 0x00000802, 0x00080002, 0x00000000);
  291. &data_word(0x00080000, 0x02000000, 0x02000802, 0x02080800);
  292. &data_word(0x00000002, 0x02080002, 0x00000800, 0x00080802);
  293. # nibble 1
  294. &data_word(0x40108010, 0x00000000, 0x00108000, 0x40100000);
  295. &data_word(0x40000010, 0x00008010, 0x40008000, 0x00108000);
  296. &data_word(0x00008000, 0x40100010, 0x00000010, 0x40008000);
  297. &data_word(0x00100010, 0x40108000, 0x40100000, 0x00000010);
  298. &data_word(0x00100000, 0x40008010, 0x40100010, 0x00008000);
  299. &data_word(0x00108010, 0x40000000, 0x00000000, 0x00100010);
  300. &data_word(0x40008010, 0x00108010, 0x40108000, 0x40000010);
  301. &data_word(0x40000000, 0x00100000, 0x00008010, 0x40108010);
  302. &data_word(0x00100010, 0x40108000, 0x40008000, 0x00108010);
  303. &data_word(0x40108010, 0x00100010, 0x40000010, 0x00000000);
  304. &data_word(0x40000000, 0x00008010, 0x00100000, 0x40100010);
  305. &data_word(0x00008000, 0x40000000, 0x00108010, 0x40008010);
  306. &data_word(0x40108000, 0x00008000, 0x00000000, 0x40000010);
  307. &data_word(0x00000010, 0x40108010, 0x00108000, 0x40100000);
  308. &data_word(0x40100010, 0x00100000, 0x00008010, 0x40008000);
  309. &data_word(0x40008010, 0x00000010, 0x40100000, 0x00108000);
  310. # nibble 2
  311. &data_word(0x04000001, 0x04040100, 0x00000100, 0x04000101);
  312. &data_word(0x00040001, 0x04000000, 0x04000101, 0x00040100);
  313. &data_word(0x04000100, 0x00040000, 0x04040000, 0x00000001);
  314. &data_word(0x04040101, 0x00000101, 0x00000001, 0x04040001);
  315. &data_word(0x00000000, 0x00040001, 0x04040100, 0x00000100);
  316. &data_word(0x00000101, 0x04040101, 0x00040000, 0x04000001);
  317. &data_word(0x04040001, 0x04000100, 0x00040101, 0x04040000);
  318. &data_word(0x00040100, 0x00000000, 0x04000000, 0x00040101);
  319. &data_word(0x04040100, 0x00000100, 0x00000001, 0x00040000);
  320. &data_word(0x00000101, 0x00040001, 0x04040000, 0x04000101);
  321. &data_word(0x00000000, 0x04040100, 0x00040100, 0x04040001);
  322. &data_word(0x00040001, 0x04000000, 0x04040101, 0x00000001);
  323. &data_word(0x00040101, 0x04000001, 0x04000000, 0x04040101);
  324. &data_word(0x00040000, 0x04000100, 0x04000101, 0x00040100);
  325. &data_word(0x04000100, 0x00000000, 0x04040001, 0x00000101);
  326. &data_word(0x04000001, 0x00040101, 0x00000100, 0x04040000);
  327. # nibble 3
  328. &data_word(0x00401008, 0x10001000, 0x00000008, 0x10401008);
  329. &data_word(0x00000000, 0x10400000, 0x10001008, 0x00400008);
  330. &data_word(0x10401000, 0x10000008, 0x10000000, 0x00001008);
  331. &data_word(0x10000008, 0x00401008, 0x00400000, 0x10000000);
  332. &data_word(0x10400008, 0x00401000, 0x00001000, 0x00000008);
  333. &data_word(0x00401000, 0x10001008, 0x10400000, 0x00001000);
  334. &data_word(0x00001008, 0x00000000, 0x00400008, 0x10401000);
  335. &data_word(0x10001000, 0x10400008, 0x10401008, 0x00400000);
  336. &data_word(0x10400008, 0x00001008, 0x00400000, 0x10000008);
  337. &data_word(0x00401000, 0x10001000, 0x00000008, 0x10400000);
  338. &data_word(0x10001008, 0x00000000, 0x00001000, 0x00400008);
  339. &data_word(0x00000000, 0x10400008, 0x10401000, 0x00001000);
  340. &data_word(0x10000000, 0x10401008, 0x00401008, 0x00400000);
  341. &data_word(0x10401008, 0x00000008, 0x10001000, 0x00401008);
  342. &data_word(0x00400008, 0x00401000, 0x10400000, 0x10001008);
  343. &data_word(0x00001008, 0x10000000, 0x10000008, 0x10401000);
  344. # nibble 4
  345. &data_word(0x08000000, 0x00010000, 0x00000400, 0x08010420);
  346. &data_word(0x08010020, 0x08000400, 0x00010420, 0x08010000);
  347. &data_word(0x00010000, 0x00000020, 0x08000020, 0x00010400);
  348. &data_word(0x08000420, 0x08010020, 0x08010400, 0x00000000);
  349. &data_word(0x00010400, 0x08000000, 0x00010020, 0x00000420);
  350. &data_word(0x08000400, 0x00010420, 0x00000000, 0x08000020);
  351. &data_word(0x00000020, 0x08000420, 0x08010420, 0x00010020);
  352. &data_word(0x08010000, 0x00000400, 0x00000420, 0x08010400);
  353. &data_word(0x08010400, 0x08000420, 0x00010020, 0x08010000);
  354. &data_word(0x00010000, 0x00000020, 0x08000020, 0x08000400);
  355. &data_word(0x08000000, 0x00010400, 0x08010420, 0x00000000);
  356. &data_word(0x00010420, 0x08000000, 0x00000400, 0x00010020);
  357. &data_word(0x08000420, 0x00000400, 0x00000000, 0x08010420);
  358. &data_word(0x08010020, 0x08010400, 0x00000420, 0x00010000);
  359. &data_word(0x00010400, 0x08010020, 0x08000400, 0x00000420);
  360. &data_word(0x00000020, 0x00010420, 0x08010000, 0x08000020);
  361. # nibble 5
  362. &data_word(0x80000040, 0x00200040, 0x00000000, 0x80202000);
  363. &data_word(0x00200040, 0x00002000, 0x80002040, 0x00200000);
  364. &data_word(0x00002040, 0x80202040, 0x00202000, 0x80000000);
  365. &data_word(0x80002000, 0x80000040, 0x80200000, 0x00202040);
  366. &data_word(0x00200000, 0x80002040, 0x80200040, 0x00000000);
  367. &data_word(0x00002000, 0x00000040, 0x80202000, 0x80200040);
  368. &data_word(0x80202040, 0x80200000, 0x80000000, 0x00002040);
  369. &data_word(0x00000040, 0x00202000, 0x00202040, 0x80002000);
  370. &data_word(0x00002040, 0x80000000, 0x80002000, 0x00202040);
  371. &data_word(0x80202000, 0x00200040, 0x00000000, 0x80002000);
  372. &data_word(0x80000000, 0x00002000, 0x80200040, 0x00200000);
  373. &data_word(0x00200040, 0x80202040, 0x00202000, 0x00000040);
  374. &data_word(0x80202040, 0x00202000, 0x00200000, 0x80002040);
  375. &data_word(0x80000040, 0x80200000, 0x00202040, 0x00000000);
  376. &data_word(0x00002000, 0x80000040, 0x80002040, 0x80202000);
  377. &data_word(0x80200000, 0x00002040, 0x00000040, 0x80200040);
  378. # nibble 6
  379. &data_word(0x00004000, 0x00000200, 0x01000200, 0x01000004);
  380. &data_word(0x01004204, 0x00004004, 0x00004200, 0x00000000);
  381. &data_word(0x01000000, 0x01000204, 0x00000204, 0x01004000);
  382. &data_word(0x00000004, 0x01004200, 0x01004000, 0x00000204);
  383. &data_word(0x01000204, 0x00004000, 0x00004004, 0x01004204);
  384. &data_word(0x00000000, 0x01000200, 0x01000004, 0x00004200);
  385. &data_word(0x01004004, 0x00004204, 0x01004200, 0x00000004);
  386. &data_word(0x00004204, 0x01004004, 0x00000200, 0x01000000);
  387. &data_word(0x00004204, 0x01004000, 0x01004004, 0x00000204);
  388. &data_word(0x00004000, 0x00000200, 0x01000000, 0x01004004);
  389. &data_word(0x01000204, 0x00004204, 0x00004200, 0x00000000);
  390. &data_word(0x00000200, 0x01000004, 0x00000004, 0x01000200);
  391. &data_word(0x00000000, 0x01000204, 0x01000200, 0x00004200);
  392. &data_word(0x00000204, 0x00004000, 0x01004204, 0x01000000);
  393. &data_word(0x01004200, 0x00000004, 0x00004004, 0x01004204);
  394. &data_word(0x01000004, 0x01004200, 0x01004000, 0x00004004);
  395. # nibble 7
  396. &data_word(0x20800080, 0x20820000, 0x00020080, 0x00000000);
  397. &data_word(0x20020000, 0x00800080, 0x20800000, 0x20820080);
  398. &data_word(0x00000080, 0x20000000, 0x00820000, 0x00020080);
  399. &data_word(0x00820080, 0x20020080, 0x20000080, 0x20800000);
  400. &data_word(0x00020000, 0x00820080, 0x00800080, 0x20020000);
  401. &data_word(0x20820080, 0x20000080, 0x00000000, 0x00820000);
  402. &data_word(0x20000000, 0x00800000, 0x20020080, 0x20800080);
  403. &data_word(0x00800000, 0x00020000, 0x20820000, 0x00000080);
  404. &data_word(0x00800000, 0x00020000, 0x20000080, 0x20820080);
  405. &data_word(0x00020080, 0x20000000, 0x00000000, 0x00820000);
  406. &data_word(0x20800080, 0x20020080, 0x20020000, 0x00800080);
  407. &data_word(0x20820000, 0x00000080, 0x00800080, 0x20020000);
  408. &data_word(0x20820080, 0x00800000, 0x20800000, 0x20000080);
  409. &data_word(0x00820000, 0x00020080, 0x20020080, 0x20800000);
  410. &data_word(0x00000080, 0x20820000, 0x00820080, 0x00000000);
  411. &data_word(0x20000000, 0x20800080, 0x00020000, 0x00820080);
  412. }