aes-parisc.pl 29 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041
  1. #! /usr/bin/env perl
  2. # Copyright 2009-2020 The OpenSSL Project Authors. All Rights Reserved.
  3. #
  4. # Licensed under the Apache License 2.0 (the "License"). You may not use
  5. # this file except in compliance with the License. You can obtain a copy
  6. # in the file LICENSE in the source distribution or at
  7. # https://www.openssl.org/source/license.html
  8. # ====================================================================
  9. # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  10. # project. The module is, however, dual licensed under OpenSSL and
  11. # CRYPTOGAMS licenses depending on where you obtain it. For further
  12. # details see http://www.openssl.org/~appro/cryptogams/.
  13. # ====================================================================
  14. # AES for PA-RISC.
  15. #
  16. # June 2009.
  17. #
  18. # The module is mechanical transliteration of aes-sparcv9.pl, but with
  19. # a twist: S-boxes are compressed even further down to 1K+256B. On
  20. # PA-7100LC performance is ~40% better than gcc 3.2 generated code and
  21. # is about 33 cycles per byte processed with 128-bit key. Newer CPUs
  22. # perform at 16 cycles per byte. It's not faster than code generated
  23. # by vendor compiler, but recall that it has compressed S-boxes, which
  24. # requires extra processing.
  25. #
  26. # Special thanks to polarhome.com for providing HP-UX account.
  27. # $output is the last argument if it looks like a file (it has an extension)
  28. # $flavour is the first argument if it doesn't look like a file
  29. $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
  30. $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
  31. $output and open STDOUT,">$output";
  32. if ($flavour =~ /64/) {
  33. $LEVEL ="2.0W";
  34. $SIZE_T =8;
  35. $FRAME_MARKER =80;
  36. $SAVED_RP =16;
  37. $PUSH ="std";
  38. $PUSHMA ="std,ma";
  39. $POP ="ldd";
  40. $POPMB ="ldd,mb";
  41. } else {
  42. $LEVEL ="1.0";
  43. $SIZE_T =4;
  44. $FRAME_MARKER =48;
  45. $SAVED_RP =20;
  46. $PUSH ="stw";
  47. $PUSHMA ="stwm";
  48. $POP ="ldw";
  49. $POPMB ="ldwm";
  50. }
  51. $FRAME=16*$SIZE_T+$FRAME_MARKER;# 16 saved regs + frame marker
  52. # [+ argument transfer]
  53. $inp="%r26"; # arg0
  54. $out="%r25"; # arg1
  55. $key="%r24"; # arg2
  56. ($s0,$s1,$s2,$s3) = ("%r1","%r2","%r3","%r4");
  57. ($t0,$t1,$t2,$t3) = ("%r5","%r6","%r7","%r8");
  58. ($acc0, $acc1, $acc2, $acc3, $acc4, $acc5, $acc6, $acc7,
  59. $acc8, $acc9,$acc10,$acc11,$acc12,$acc13,$acc14,$acc15) =
  60. ("%r9","%r10","%r11","%r12","%r13","%r14","%r15","%r16",
  61. "%r17","%r18","%r19","%r20","%r21","%r22","%r23","%r26");
  62. $tbl="%r28";
  63. $rounds="%r29";
  64. $code=<<___;
  65. .LEVEL $LEVEL
  66. .SPACE \$TEXT\$
  67. .SUBSPA \$CODE\$,QUAD=0,ALIGN=8,ACCESS=0x2C,CODE_ONLY
  68. .EXPORT AES_encrypt,ENTRY,ARGW0=GR,ARGW1=GR,ARGW2=GR
  69. .ALIGN 64
  70. AES_encrypt
  71. .PROC
  72. .CALLINFO FRAME=`$FRAME-16*$SIZE_T`,NO_CALLS,SAVE_RP,ENTRY_GR=18
  73. .ENTRY
  74. $PUSH %r2,-$SAVED_RP(%sp) ; standard prologue
  75. $PUSHMA %r3,$FRAME(%sp)
  76. $PUSH %r4,`-$FRAME+1*$SIZE_T`(%sp)
  77. $PUSH %r5,`-$FRAME+2*$SIZE_T`(%sp)
  78. $PUSH %r6,`-$FRAME+3*$SIZE_T`(%sp)
  79. $PUSH %r7,`-$FRAME+4*$SIZE_T`(%sp)
  80. $PUSH %r8,`-$FRAME+5*$SIZE_T`(%sp)
  81. $PUSH %r9,`-$FRAME+6*$SIZE_T`(%sp)
  82. $PUSH %r10,`-$FRAME+7*$SIZE_T`(%sp)
  83. $PUSH %r11,`-$FRAME+8*$SIZE_T`(%sp)
  84. $PUSH %r12,`-$FRAME+9*$SIZE_T`(%sp)
  85. $PUSH %r13,`-$FRAME+10*$SIZE_T`(%sp)
  86. $PUSH %r14,`-$FRAME+11*$SIZE_T`(%sp)
  87. $PUSH %r15,`-$FRAME+12*$SIZE_T`(%sp)
  88. $PUSH %r16,`-$FRAME+13*$SIZE_T`(%sp)
  89. $PUSH %r17,`-$FRAME+14*$SIZE_T`(%sp)
  90. $PUSH %r18,`-$FRAME+15*$SIZE_T`(%sp)
  91. blr %r0,$tbl
  92. ldi 3,$t0
  93. L\$enc_pic
  94. andcm $tbl,$t0,$tbl
  95. ldo L\$AES_Te-L\$enc_pic($tbl),$tbl
  96. and $inp,$t0,$t0
  97. sub $inp,$t0,$inp
  98. ldw 0($inp),$s0
  99. ldw 4($inp),$s1
  100. ldw 8($inp),$s2
  101. comib,= 0,$t0,L\$enc_inp_aligned
  102. ldw 12($inp),$s3
  103. sh3addl $t0,%r0,$t0
  104. subi 32,$t0,$t0
  105. mtctl $t0,%cr11
  106. ldw 16($inp),$t1
  107. vshd $s0,$s1,$s0
  108. vshd $s1,$s2,$s1
  109. vshd $s2,$s3,$s2
  110. vshd $s3,$t1,$s3
  111. L\$enc_inp_aligned
  112. bl _parisc_AES_encrypt,%r31
  113. nop
  114. extru,<> $out,31,2,%r0
  115. b L\$enc_out_aligned
  116. nop
  117. _srm $s0,24,$acc0
  118. _srm $s0,16,$acc1
  119. stb $acc0,0($out)
  120. _srm $s0,8,$acc2
  121. stb $acc1,1($out)
  122. _srm $s1,24,$acc4
  123. stb $acc2,2($out)
  124. _srm $s1,16,$acc5
  125. stb $s0,3($out)
  126. _srm $s1,8,$acc6
  127. stb $acc4,4($out)
  128. _srm $s2,24,$acc0
  129. stb $acc5,5($out)
  130. _srm $s2,16,$acc1
  131. stb $acc6,6($out)
  132. _srm $s2,8,$acc2
  133. stb $s1,7($out)
  134. _srm $s3,24,$acc4
  135. stb $acc0,8($out)
  136. _srm $s3,16,$acc5
  137. stb $acc1,9($out)
  138. _srm $s3,8,$acc6
  139. stb $acc2,10($out)
  140. stb $s2,11($out)
  141. stb $acc4,12($out)
  142. stb $acc5,13($out)
  143. stb $acc6,14($out)
  144. b L\$enc_done
  145. stb $s3,15($out)
  146. L\$enc_out_aligned
  147. stw $s0,0($out)
  148. stw $s1,4($out)
  149. stw $s2,8($out)
  150. stw $s3,12($out)
  151. L\$enc_done
  152. $POP `-$FRAME-$SAVED_RP`(%sp),%r2 ; standard epilogue
  153. $POP `-$FRAME+1*$SIZE_T`(%sp),%r4
  154. $POP `-$FRAME+2*$SIZE_T`(%sp),%r5
  155. $POP `-$FRAME+3*$SIZE_T`(%sp),%r6
  156. $POP `-$FRAME+4*$SIZE_T`(%sp),%r7
  157. $POP `-$FRAME+5*$SIZE_T`(%sp),%r8
  158. $POP `-$FRAME+6*$SIZE_T`(%sp),%r9
  159. $POP `-$FRAME+7*$SIZE_T`(%sp),%r10
  160. $POP `-$FRAME+8*$SIZE_T`(%sp),%r11
  161. $POP `-$FRAME+9*$SIZE_T`(%sp),%r12
  162. $POP `-$FRAME+10*$SIZE_T`(%sp),%r13
  163. $POP `-$FRAME+11*$SIZE_T`(%sp),%r14
  164. $POP `-$FRAME+12*$SIZE_T`(%sp),%r15
  165. $POP `-$FRAME+13*$SIZE_T`(%sp),%r16
  166. $POP `-$FRAME+14*$SIZE_T`(%sp),%r17
  167. $POP `-$FRAME+15*$SIZE_T`(%sp),%r18
  168. bv (%r2)
  169. .EXIT
  170. $POPMB -$FRAME(%sp),%r3
  171. .PROCEND
  172. .ALIGN 16
  173. _parisc_AES_encrypt
  174. .PROC
  175. .CALLINFO MILLICODE
  176. .ENTRY
  177. ldw 240($key),$rounds
  178. ldw 0($key),$t0
  179. ldw 4($key),$t1
  180. ldw 8($key),$t2
  181. _srm $rounds,1,$rounds
  182. xor $t0,$s0,$s0
  183. ldw 12($key),$t3
  184. _srm $s0,24,$acc0
  185. xor $t1,$s1,$s1
  186. ldw 16($key),$t0
  187. _srm $s1,16,$acc1
  188. xor $t2,$s2,$s2
  189. ldw 20($key),$t1
  190. xor $t3,$s3,$s3
  191. ldw 24($key),$t2
  192. ldw 28($key),$t3
  193. L\$enc_loop
  194. _srm $s2,8,$acc2
  195. ldwx,s $acc0($tbl),$acc0
  196. _srm $s3,0,$acc3
  197. ldwx,s $acc1($tbl),$acc1
  198. _srm $s1,24,$acc4
  199. ldwx,s $acc2($tbl),$acc2
  200. _srm $s2,16,$acc5
  201. ldwx,s $acc3($tbl),$acc3
  202. _srm $s3,8,$acc6
  203. ldwx,s $acc4($tbl),$acc4
  204. _srm $s0,0,$acc7
  205. ldwx,s $acc5($tbl),$acc5
  206. _srm $s2,24,$acc8
  207. ldwx,s $acc6($tbl),$acc6
  208. _srm $s3,16,$acc9
  209. ldwx,s $acc7($tbl),$acc7
  210. _srm $s0,8,$acc10
  211. ldwx,s $acc8($tbl),$acc8
  212. _srm $s1,0,$acc11
  213. ldwx,s $acc9($tbl),$acc9
  214. _srm $s3,24,$acc12
  215. ldwx,s $acc10($tbl),$acc10
  216. _srm $s0,16,$acc13
  217. ldwx,s $acc11($tbl),$acc11
  218. _srm $s1,8,$acc14
  219. ldwx,s $acc12($tbl),$acc12
  220. _srm $s2,0,$acc15
  221. ldwx,s $acc13($tbl),$acc13
  222. ldwx,s $acc14($tbl),$acc14
  223. ldwx,s $acc15($tbl),$acc15
  224. addib,= -1,$rounds,L\$enc_last
  225. ldo 32($key),$key
  226. _ror $acc1,8,$acc1
  227. xor $acc0,$t0,$t0
  228. ldw 0($key),$s0
  229. _ror $acc2,16,$acc2
  230. xor $acc1,$t0,$t0
  231. ldw 4($key),$s1
  232. _ror $acc3,24,$acc3
  233. xor $acc2,$t0,$t0
  234. ldw 8($key),$s2
  235. _ror $acc5,8,$acc5
  236. xor $acc3,$t0,$t0
  237. ldw 12($key),$s3
  238. _ror $acc6,16,$acc6
  239. xor $acc4,$t1,$t1
  240. _ror $acc7,24,$acc7
  241. xor $acc5,$t1,$t1
  242. _ror $acc9,8,$acc9
  243. xor $acc6,$t1,$t1
  244. _ror $acc10,16,$acc10
  245. xor $acc7,$t1,$t1
  246. _ror $acc11,24,$acc11
  247. xor $acc8,$t2,$t2
  248. _ror $acc13,8,$acc13
  249. xor $acc9,$t2,$t2
  250. _ror $acc14,16,$acc14
  251. xor $acc10,$t2,$t2
  252. _ror $acc15,24,$acc15
  253. xor $acc11,$t2,$t2
  254. xor $acc12,$acc14,$acc14
  255. xor $acc13,$t3,$t3
  256. _srm $t0,24,$acc0
  257. xor $acc14,$t3,$t3
  258. _srm $t1,16,$acc1
  259. xor $acc15,$t3,$t3
  260. _srm $t2,8,$acc2
  261. ldwx,s $acc0($tbl),$acc0
  262. _srm $t3,0,$acc3
  263. ldwx,s $acc1($tbl),$acc1
  264. _srm $t1,24,$acc4
  265. ldwx,s $acc2($tbl),$acc2
  266. _srm $t2,16,$acc5
  267. ldwx,s $acc3($tbl),$acc3
  268. _srm $t3,8,$acc6
  269. ldwx,s $acc4($tbl),$acc4
  270. _srm $t0,0,$acc7
  271. ldwx,s $acc5($tbl),$acc5
  272. _srm $t2,24,$acc8
  273. ldwx,s $acc6($tbl),$acc6
  274. _srm $t3,16,$acc9
  275. ldwx,s $acc7($tbl),$acc7
  276. _srm $t0,8,$acc10
  277. ldwx,s $acc8($tbl),$acc8
  278. _srm $t1,0,$acc11
  279. ldwx,s $acc9($tbl),$acc9
  280. _srm $t3,24,$acc12
  281. ldwx,s $acc10($tbl),$acc10
  282. _srm $t0,16,$acc13
  283. ldwx,s $acc11($tbl),$acc11
  284. _srm $t1,8,$acc14
  285. ldwx,s $acc12($tbl),$acc12
  286. _srm $t2,0,$acc15
  287. ldwx,s $acc13($tbl),$acc13
  288. _ror $acc1,8,$acc1
  289. ldwx,s $acc14($tbl),$acc14
  290. _ror $acc2,16,$acc2
  291. xor $acc0,$s0,$s0
  292. ldwx,s $acc15($tbl),$acc15
  293. _ror $acc3,24,$acc3
  294. xor $acc1,$s0,$s0
  295. ldw 16($key),$t0
  296. _ror $acc5,8,$acc5
  297. xor $acc2,$s0,$s0
  298. ldw 20($key),$t1
  299. _ror $acc6,16,$acc6
  300. xor $acc3,$s0,$s0
  301. ldw 24($key),$t2
  302. _ror $acc7,24,$acc7
  303. xor $acc4,$s1,$s1
  304. ldw 28($key),$t3
  305. _ror $acc9,8,$acc9
  306. xor $acc5,$s1,$s1
  307. ldw 1024+0($tbl),%r0 ; prefetch te4
  308. _ror $acc10,16,$acc10
  309. xor $acc6,$s1,$s1
  310. ldw 1024+32($tbl),%r0 ; prefetch te4
  311. _ror $acc11,24,$acc11
  312. xor $acc7,$s1,$s1
  313. ldw 1024+64($tbl),%r0 ; prefetch te4
  314. _ror $acc13,8,$acc13
  315. xor $acc8,$s2,$s2
  316. ldw 1024+96($tbl),%r0 ; prefetch te4
  317. _ror $acc14,16,$acc14
  318. xor $acc9,$s2,$s2
  319. ldw 1024+128($tbl),%r0 ; prefetch te4
  320. _ror $acc15,24,$acc15
  321. xor $acc10,$s2,$s2
  322. ldw 1024+160($tbl),%r0 ; prefetch te4
  323. _srm $s0,24,$acc0
  324. xor $acc11,$s2,$s2
  325. ldw 1024+192($tbl),%r0 ; prefetch te4
  326. xor $acc12,$acc14,$acc14
  327. xor $acc13,$s3,$s3
  328. ldw 1024+224($tbl),%r0 ; prefetch te4
  329. _srm $s1,16,$acc1
  330. xor $acc14,$s3,$s3
  331. b L\$enc_loop
  332. xor $acc15,$s3,$s3
  333. .ALIGN 16
  334. L\$enc_last
  335. ldo 1024($tbl),$rounds
  336. _ror $acc1,8,$acc1
  337. xor $acc0,$t0,$t0
  338. ldw 0($key),$s0
  339. _ror $acc2,16,$acc2
  340. xor $acc1,$t0,$t0
  341. ldw 4($key),$s1
  342. _ror $acc3,24,$acc3
  343. xor $acc2,$t0,$t0
  344. ldw 8($key),$s2
  345. _ror $acc5,8,$acc5
  346. xor $acc3,$t0,$t0
  347. ldw 12($key),$s3
  348. _ror $acc6,16,$acc6
  349. xor $acc4,$t1,$t1
  350. _ror $acc7,24,$acc7
  351. xor $acc5,$t1,$t1
  352. _ror $acc9,8,$acc9
  353. xor $acc6,$t1,$t1
  354. _ror $acc10,16,$acc10
  355. xor $acc7,$t1,$t1
  356. _ror $acc11,24,$acc11
  357. xor $acc8,$t2,$t2
  358. _ror $acc13,8,$acc13
  359. xor $acc9,$t2,$t2
  360. _ror $acc14,16,$acc14
  361. xor $acc10,$t2,$t2
  362. _ror $acc15,24,$acc15
  363. xor $acc11,$t2,$t2
  364. xor $acc12,$acc14,$acc14
  365. xor $acc13,$t3,$t3
  366. _srm $t0,24,$acc0
  367. xor $acc14,$t3,$t3
  368. _srm $t1,16,$acc1
  369. xor $acc15,$t3,$t3
  370. _srm $t2,8,$acc2
  371. ldbx $acc0($rounds),$acc0
  372. _srm $t1,24,$acc4
  373. ldbx $acc1($rounds),$acc1
  374. _srm $t2,16,$acc5
  375. _srm $t3,0,$acc3
  376. ldbx $acc2($rounds),$acc2
  377. ldbx $acc3($rounds),$acc3
  378. _srm $t3,8,$acc6
  379. ldbx $acc4($rounds),$acc4
  380. _srm $t2,24,$acc8
  381. ldbx $acc5($rounds),$acc5
  382. _srm $t3,16,$acc9
  383. _srm $t0,0,$acc7
  384. ldbx $acc6($rounds),$acc6
  385. ldbx $acc7($rounds),$acc7
  386. _srm $t0,8,$acc10
  387. ldbx $acc8($rounds),$acc8
  388. _srm $t3,24,$acc12
  389. ldbx $acc9($rounds),$acc9
  390. _srm $t0,16,$acc13
  391. _srm $t1,0,$acc11
  392. ldbx $acc10($rounds),$acc10
  393. _srm $t1,8,$acc14
  394. ldbx $acc11($rounds),$acc11
  395. ldbx $acc12($rounds),$acc12
  396. ldbx $acc13($rounds),$acc13
  397. _srm $t2,0,$acc15
  398. ldbx $acc14($rounds),$acc14
  399. dep $acc0,7,8,$acc3
  400. ldbx $acc15($rounds),$acc15
  401. dep $acc4,7,8,$acc7
  402. dep $acc1,15,8,$acc3
  403. dep $acc5,15,8,$acc7
  404. dep $acc2,23,8,$acc3
  405. dep $acc6,23,8,$acc7
  406. xor $acc3,$s0,$s0
  407. xor $acc7,$s1,$s1
  408. dep $acc8,7,8,$acc11
  409. dep $acc12,7,8,$acc15
  410. dep $acc9,15,8,$acc11
  411. dep $acc13,15,8,$acc15
  412. dep $acc10,23,8,$acc11
  413. dep $acc14,23,8,$acc15
  414. xor $acc11,$s2,$s2
  415. bv (%r31)
  416. .EXIT
  417. xor $acc15,$s3,$s3
  418. .PROCEND
  419. .ALIGN 64
  420. L\$AES_Te
  421. .WORD 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
  422. .WORD 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
  423. .WORD 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
  424. .WORD 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
  425. .WORD 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
  426. .WORD 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
  427. .WORD 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
  428. .WORD 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
  429. .WORD 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
  430. .WORD 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
  431. .WORD 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
  432. .WORD 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
  433. .WORD 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
  434. .WORD 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
  435. .WORD 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
  436. .WORD 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
  437. .WORD 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
  438. .WORD 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
  439. .WORD 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
  440. .WORD 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
  441. .WORD 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
  442. .WORD 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
  443. .WORD 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
  444. .WORD 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
  445. .WORD 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
  446. .WORD 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
  447. .WORD 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
  448. .WORD 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
  449. .WORD 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
  450. .WORD 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
  451. .WORD 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
  452. .WORD 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
  453. .WORD 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
  454. .WORD 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
  455. .WORD 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
  456. .WORD 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
  457. .WORD 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
  458. .WORD 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
  459. .WORD 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
  460. .WORD 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
  461. .WORD 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
  462. .WORD 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
  463. .WORD 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
  464. .WORD 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
  465. .WORD 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
  466. .WORD 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
  467. .WORD 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
  468. .WORD 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
  469. .WORD 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
  470. .WORD 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
  471. .WORD 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
  472. .WORD 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
  473. .WORD 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
  474. .WORD 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
  475. .WORD 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
  476. .WORD 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
  477. .WORD 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
  478. .WORD 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
  479. .WORD 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
  480. .WORD 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
  481. .WORD 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
  482. .WORD 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
  483. .WORD 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
  484. .WORD 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
  485. .BYTE 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
  486. .BYTE 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
  487. .BYTE 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
  488. .BYTE 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
  489. .BYTE 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
  490. .BYTE 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
  491. .BYTE 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
  492. .BYTE 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
  493. .BYTE 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
  494. .BYTE 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
  495. .BYTE 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
  496. .BYTE 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
  497. .BYTE 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
  498. .BYTE 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
  499. .BYTE 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
  500. .BYTE 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
  501. .BYTE 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
  502. .BYTE 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
  503. .BYTE 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
  504. .BYTE 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
  505. .BYTE 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
  506. .BYTE 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
  507. .BYTE 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
  508. .BYTE 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
  509. .BYTE 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
  510. .BYTE 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
  511. .BYTE 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
  512. .BYTE 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
  513. .BYTE 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
  514. .BYTE 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
  515. .BYTE 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
  516. .BYTE 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
  517. ___
  518. $code.=<<___;
  519. .EXPORT AES_decrypt,ENTRY,ARGW0=GR,ARGW1=GR,ARGW2=GR
  520. .ALIGN 16
  521. AES_decrypt
  522. .PROC
  523. .CALLINFO FRAME=`$FRAME-16*$SIZE_T`,NO_CALLS,SAVE_RP,ENTRY_GR=18
  524. .ENTRY
  525. $PUSH %r2,-$SAVED_RP(%sp) ; standard prologue
  526. $PUSHMA %r3,$FRAME(%sp)
  527. $PUSH %r4,`-$FRAME+1*$SIZE_T`(%sp)
  528. $PUSH %r5,`-$FRAME+2*$SIZE_T`(%sp)
  529. $PUSH %r6,`-$FRAME+3*$SIZE_T`(%sp)
  530. $PUSH %r7,`-$FRAME+4*$SIZE_T`(%sp)
  531. $PUSH %r8,`-$FRAME+5*$SIZE_T`(%sp)
  532. $PUSH %r9,`-$FRAME+6*$SIZE_T`(%sp)
  533. $PUSH %r10,`-$FRAME+7*$SIZE_T`(%sp)
  534. $PUSH %r11,`-$FRAME+8*$SIZE_T`(%sp)
  535. $PUSH %r12,`-$FRAME+9*$SIZE_T`(%sp)
  536. $PUSH %r13,`-$FRAME+10*$SIZE_T`(%sp)
  537. $PUSH %r14,`-$FRAME+11*$SIZE_T`(%sp)
  538. $PUSH %r15,`-$FRAME+12*$SIZE_T`(%sp)
  539. $PUSH %r16,`-$FRAME+13*$SIZE_T`(%sp)
  540. $PUSH %r17,`-$FRAME+14*$SIZE_T`(%sp)
  541. $PUSH %r18,`-$FRAME+15*$SIZE_T`(%sp)
  542. blr %r0,$tbl
  543. ldi 3,$t0
  544. L\$dec_pic
  545. andcm $tbl,$t0,$tbl
  546. ldo L\$AES_Td-L\$dec_pic($tbl),$tbl
  547. and $inp,$t0,$t0
  548. sub $inp,$t0,$inp
  549. ldw 0($inp),$s0
  550. ldw 4($inp),$s1
  551. ldw 8($inp),$s2
  552. comib,= 0,$t0,L\$dec_inp_aligned
  553. ldw 12($inp),$s3
  554. sh3addl $t0,%r0,$t0
  555. subi 32,$t0,$t0
  556. mtctl $t0,%cr11
  557. ldw 16($inp),$t1
  558. vshd $s0,$s1,$s0
  559. vshd $s1,$s2,$s1
  560. vshd $s2,$s3,$s2
  561. vshd $s3,$t1,$s3
  562. L\$dec_inp_aligned
  563. bl _parisc_AES_decrypt,%r31
  564. nop
  565. extru,<> $out,31,2,%r0
  566. b L\$dec_out_aligned
  567. nop
  568. _srm $s0,24,$acc0
  569. _srm $s0,16,$acc1
  570. stb $acc0,0($out)
  571. _srm $s0,8,$acc2
  572. stb $acc1,1($out)
  573. _srm $s1,24,$acc4
  574. stb $acc2,2($out)
  575. _srm $s1,16,$acc5
  576. stb $s0,3($out)
  577. _srm $s1,8,$acc6
  578. stb $acc4,4($out)
  579. _srm $s2,24,$acc0
  580. stb $acc5,5($out)
  581. _srm $s2,16,$acc1
  582. stb $acc6,6($out)
  583. _srm $s2,8,$acc2
  584. stb $s1,7($out)
  585. _srm $s3,24,$acc4
  586. stb $acc0,8($out)
  587. _srm $s3,16,$acc5
  588. stb $acc1,9($out)
  589. _srm $s3,8,$acc6
  590. stb $acc2,10($out)
  591. stb $s2,11($out)
  592. stb $acc4,12($out)
  593. stb $acc5,13($out)
  594. stb $acc6,14($out)
  595. b L\$dec_done
  596. stb $s3,15($out)
  597. L\$dec_out_aligned
  598. stw $s0,0($out)
  599. stw $s1,4($out)
  600. stw $s2,8($out)
  601. stw $s3,12($out)
  602. L\$dec_done
  603. $POP `-$FRAME-$SAVED_RP`(%sp),%r2 ; standard epilogue
  604. $POP `-$FRAME+1*$SIZE_T`(%sp),%r4
  605. $POP `-$FRAME+2*$SIZE_T`(%sp),%r5
  606. $POP `-$FRAME+3*$SIZE_T`(%sp),%r6
  607. $POP `-$FRAME+4*$SIZE_T`(%sp),%r7
  608. $POP `-$FRAME+5*$SIZE_T`(%sp),%r8
  609. $POP `-$FRAME+6*$SIZE_T`(%sp),%r9
  610. $POP `-$FRAME+7*$SIZE_T`(%sp),%r10
  611. $POP `-$FRAME+8*$SIZE_T`(%sp),%r11
  612. $POP `-$FRAME+9*$SIZE_T`(%sp),%r12
  613. $POP `-$FRAME+10*$SIZE_T`(%sp),%r13
  614. $POP `-$FRAME+11*$SIZE_T`(%sp),%r14
  615. $POP `-$FRAME+12*$SIZE_T`(%sp),%r15
  616. $POP `-$FRAME+13*$SIZE_T`(%sp),%r16
  617. $POP `-$FRAME+14*$SIZE_T`(%sp),%r17
  618. $POP `-$FRAME+15*$SIZE_T`(%sp),%r18
  619. bv (%r2)
  620. .EXIT
  621. $POPMB -$FRAME(%sp),%r3
  622. .PROCEND
  623. .ALIGN 16
  624. _parisc_AES_decrypt
  625. .PROC
  626. .CALLINFO MILLICODE
  627. .ENTRY
  628. ldw 240($key),$rounds
  629. ldw 0($key),$t0
  630. ldw 4($key),$t1
  631. ldw 8($key),$t2
  632. ldw 12($key),$t3
  633. _srm $rounds,1,$rounds
  634. xor $t0,$s0,$s0
  635. ldw 16($key),$t0
  636. xor $t1,$s1,$s1
  637. ldw 20($key),$t1
  638. _srm $s0,24,$acc0
  639. xor $t2,$s2,$s2
  640. ldw 24($key),$t2
  641. xor $t3,$s3,$s3
  642. ldw 28($key),$t3
  643. _srm $s3,16,$acc1
  644. L\$dec_loop
  645. _srm $s2,8,$acc2
  646. ldwx,s $acc0($tbl),$acc0
  647. _srm $s1,0,$acc3
  648. ldwx,s $acc1($tbl),$acc1
  649. _srm $s1,24,$acc4
  650. ldwx,s $acc2($tbl),$acc2
  651. _srm $s0,16,$acc5
  652. ldwx,s $acc3($tbl),$acc3
  653. _srm $s3,8,$acc6
  654. ldwx,s $acc4($tbl),$acc4
  655. _srm $s2,0,$acc7
  656. ldwx,s $acc5($tbl),$acc5
  657. _srm $s2,24,$acc8
  658. ldwx,s $acc6($tbl),$acc6
  659. _srm $s1,16,$acc9
  660. ldwx,s $acc7($tbl),$acc7
  661. _srm $s0,8,$acc10
  662. ldwx,s $acc8($tbl),$acc8
  663. _srm $s3,0,$acc11
  664. ldwx,s $acc9($tbl),$acc9
  665. _srm $s3,24,$acc12
  666. ldwx,s $acc10($tbl),$acc10
  667. _srm $s2,16,$acc13
  668. ldwx,s $acc11($tbl),$acc11
  669. _srm $s1,8,$acc14
  670. ldwx,s $acc12($tbl),$acc12
  671. _srm $s0,0,$acc15
  672. ldwx,s $acc13($tbl),$acc13
  673. ldwx,s $acc14($tbl),$acc14
  674. ldwx,s $acc15($tbl),$acc15
  675. addib,= -1,$rounds,L\$dec_last
  676. ldo 32($key),$key
  677. _ror $acc1,8,$acc1
  678. xor $acc0,$t0,$t0
  679. ldw 0($key),$s0
  680. _ror $acc2,16,$acc2
  681. xor $acc1,$t0,$t0
  682. ldw 4($key),$s1
  683. _ror $acc3,24,$acc3
  684. xor $acc2,$t0,$t0
  685. ldw 8($key),$s2
  686. _ror $acc5,8,$acc5
  687. xor $acc3,$t0,$t0
  688. ldw 12($key),$s3
  689. _ror $acc6,16,$acc6
  690. xor $acc4,$t1,$t1
  691. _ror $acc7,24,$acc7
  692. xor $acc5,$t1,$t1
  693. _ror $acc9,8,$acc9
  694. xor $acc6,$t1,$t1
  695. _ror $acc10,16,$acc10
  696. xor $acc7,$t1,$t1
  697. _ror $acc11,24,$acc11
  698. xor $acc8,$t2,$t2
  699. _ror $acc13,8,$acc13
  700. xor $acc9,$t2,$t2
  701. _ror $acc14,16,$acc14
  702. xor $acc10,$t2,$t2
  703. _ror $acc15,24,$acc15
  704. xor $acc11,$t2,$t2
  705. xor $acc12,$acc14,$acc14
  706. xor $acc13,$t3,$t3
  707. _srm $t0,24,$acc0
  708. xor $acc14,$t3,$t3
  709. xor $acc15,$t3,$t3
  710. _srm $t3,16,$acc1
  711. _srm $t2,8,$acc2
  712. ldwx,s $acc0($tbl),$acc0
  713. _srm $t1,0,$acc3
  714. ldwx,s $acc1($tbl),$acc1
  715. _srm $t1,24,$acc4
  716. ldwx,s $acc2($tbl),$acc2
  717. _srm $t0,16,$acc5
  718. ldwx,s $acc3($tbl),$acc3
  719. _srm $t3,8,$acc6
  720. ldwx,s $acc4($tbl),$acc4
  721. _srm $t2,0,$acc7
  722. ldwx,s $acc5($tbl),$acc5
  723. _srm $t2,24,$acc8
  724. ldwx,s $acc6($tbl),$acc6
  725. _srm $t1,16,$acc9
  726. ldwx,s $acc7($tbl),$acc7
  727. _srm $t0,8,$acc10
  728. ldwx,s $acc8($tbl),$acc8
  729. _srm $t3,0,$acc11
  730. ldwx,s $acc9($tbl),$acc9
  731. _srm $t3,24,$acc12
  732. ldwx,s $acc10($tbl),$acc10
  733. _srm $t2,16,$acc13
  734. ldwx,s $acc11($tbl),$acc11
  735. _srm $t1,8,$acc14
  736. ldwx,s $acc12($tbl),$acc12
  737. _srm $t0,0,$acc15
  738. ldwx,s $acc13($tbl),$acc13
  739. _ror $acc1,8,$acc1
  740. ldwx,s $acc14($tbl),$acc14
  741. _ror $acc2,16,$acc2
  742. xor $acc0,$s0,$s0
  743. ldwx,s $acc15($tbl),$acc15
  744. _ror $acc3,24,$acc3
  745. xor $acc1,$s0,$s0
  746. ldw 16($key),$t0
  747. _ror $acc5,8,$acc5
  748. xor $acc2,$s0,$s0
  749. ldw 20($key),$t1
  750. _ror $acc6,16,$acc6
  751. xor $acc3,$s0,$s0
  752. ldw 24($key),$t2
  753. _ror $acc7,24,$acc7
  754. xor $acc4,$s1,$s1
  755. ldw 28($key),$t3
  756. _ror $acc9,8,$acc9
  757. xor $acc5,$s1,$s1
  758. ldw 1024+0($tbl),%r0 ; prefetch td4
  759. _ror $acc10,16,$acc10
  760. xor $acc6,$s1,$s1
  761. ldw 1024+32($tbl),%r0 ; prefetch td4
  762. _ror $acc11,24,$acc11
  763. xor $acc7,$s1,$s1
  764. ldw 1024+64($tbl),%r0 ; prefetch td4
  765. _ror $acc13,8,$acc13
  766. xor $acc8,$s2,$s2
  767. ldw 1024+96($tbl),%r0 ; prefetch td4
  768. _ror $acc14,16,$acc14
  769. xor $acc9,$s2,$s2
  770. ldw 1024+128($tbl),%r0 ; prefetch td4
  771. _ror $acc15,24,$acc15
  772. xor $acc10,$s2,$s2
  773. ldw 1024+160($tbl),%r0 ; prefetch td4
  774. _srm $s0,24,$acc0
  775. xor $acc11,$s2,$s2
  776. ldw 1024+192($tbl),%r0 ; prefetch td4
  777. xor $acc12,$acc14,$acc14
  778. xor $acc13,$s3,$s3
  779. ldw 1024+224($tbl),%r0 ; prefetch td4
  780. xor $acc14,$s3,$s3
  781. xor $acc15,$s3,$s3
  782. b L\$dec_loop
  783. _srm $s3,16,$acc1
  784. .ALIGN 16
  785. L\$dec_last
  786. ldo 1024($tbl),$rounds
  787. _ror $acc1,8,$acc1
  788. xor $acc0,$t0,$t0
  789. ldw 0($key),$s0
  790. _ror $acc2,16,$acc2
  791. xor $acc1,$t0,$t0
  792. ldw 4($key),$s1
  793. _ror $acc3,24,$acc3
  794. xor $acc2,$t0,$t0
  795. ldw 8($key),$s2
  796. _ror $acc5,8,$acc5
  797. xor $acc3,$t0,$t0
  798. ldw 12($key),$s3
  799. _ror $acc6,16,$acc6
  800. xor $acc4,$t1,$t1
  801. _ror $acc7,24,$acc7
  802. xor $acc5,$t1,$t1
  803. _ror $acc9,8,$acc9
  804. xor $acc6,$t1,$t1
  805. _ror $acc10,16,$acc10
  806. xor $acc7,$t1,$t1
  807. _ror $acc11,24,$acc11
  808. xor $acc8,$t2,$t2
  809. _ror $acc13,8,$acc13
  810. xor $acc9,$t2,$t2
  811. _ror $acc14,16,$acc14
  812. xor $acc10,$t2,$t2
  813. _ror $acc15,24,$acc15
  814. xor $acc11,$t2,$t2
  815. xor $acc12,$acc14,$acc14
  816. xor $acc13,$t3,$t3
  817. _srm $t0,24,$acc0
  818. xor $acc14,$t3,$t3
  819. xor $acc15,$t3,$t3
  820. _srm $t3,16,$acc1
  821. _srm $t2,8,$acc2
  822. ldbx $acc0($rounds),$acc0
  823. _srm $t1,24,$acc4
  824. ldbx $acc1($rounds),$acc1
  825. _srm $t0,16,$acc5
  826. _srm $t1,0,$acc3
  827. ldbx $acc2($rounds),$acc2
  828. ldbx $acc3($rounds),$acc3
  829. _srm $t3,8,$acc6
  830. ldbx $acc4($rounds),$acc4
  831. _srm $t2,24,$acc8
  832. ldbx $acc5($rounds),$acc5
  833. _srm $t1,16,$acc9
  834. _srm $t2,0,$acc7
  835. ldbx $acc6($rounds),$acc6
  836. ldbx $acc7($rounds),$acc7
  837. _srm $t0,8,$acc10
  838. ldbx $acc8($rounds),$acc8
  839. _srm $t3,24,$acc12
  840. ldbx $acc9($rounds),$acc9
  841. _srm $t2,16,$acc13
  842. _srm $t3,0,$acc11
  843. ldbx $acc10($rounds),$acc10
  844. _srm $t1,8,$acc14
  845. ldbx $acc11($rounds),$acc11
  846. ldbx $acc12($rounds),$acc12
  847. ldbx $acc13($rounds),$acc13
  848. _srm $t0,0,$acc15
  849. ldbx $acc14($rounds),$acc14
  850. dep $acc0,7,8,$acc3
  851. ldbx $acc15($rounds),$acc15
  852. dep $acc4,7,8,$acc7
  853. dep $acc1,15,8,$acc3
  854. dep $acc5,15,8,$acc7
  855. dep $acc2,23,8,$acc3
  856. dep $acc6,23,8,$acc7
  857. xor $acc3,$s0,$s0
  858. xor $acc7,$s1,$s1
  859. dep $acc8,7,8,$acc11
  860. dep $acc12,7,8,$acc15
  861. dep $acc9,15,8,$acc11
  862. dep $acc13,15,8,$acc15
  863. dep $acc10,23,8,$acc11
  864. dep $acc14,23,8,$acc15
  865. xor $acc11,$s2,$s2
  866. bv (%r31)
  867. .EXIT
  868. xor $acc15,$s3,$s3
  869. .PROCEND
  870. .ALIGN 64
  871. L\$AES_Td
  872. .WORD 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
  873. .WORD 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
  874. .WORD 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
  875. .WORD 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
  876. .WORD 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
  877. .WORD 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
  878. .WORD 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
  879. .WORD 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
  880. .WORD 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
  881. .WORD 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
  882. .WORD 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
  883. .WORD 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
  884. .WORD 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
  885. .WORD 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
  886. .WORD 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
  887. .WORD 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
  888. .WORD 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
  889. .WORD 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
  890. .WORD 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
  891. .WORD 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
  892. .WORD 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
  893. .WORD 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
  894. .WORD 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
  895. .WORD 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
  896. .WORD 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
  897. .WORD 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
  898. .WORD 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
  899. .WORD 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
  900. .WORD 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
  901. .WORD 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
  902. .WORD 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
  903. .WORD 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
  904. .WORD 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
  905. .WORD 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
  906. .WORD 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
  907. .WORD 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
  908. .WORD 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
  909. .WORD 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
  910. .WORD 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
  911. .WORD 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
  912. .WORD 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
  913. .WORD 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
  914. .WORD 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
  915. .WORD 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
  916. .WORD 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
  917. .WORD 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
  918. .WORD 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
  919. .WORD 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
  920. .WORD 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
  921. .WORD 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
  922. .WORD 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
  923. .WORD 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
  924. .WORD 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
  925. .WORD 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
  926. .WORD 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
  927. .WORD 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
  928. .WORD 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
  929. .WORD 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
  930. .WORD 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
  931. .WORD 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
  932. .WORD 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
  933. .WORD 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
  934. .WORD 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
  935. .WORD 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
  936. .BYTE 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
  937. .BYTE 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
  938. .BYTE 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
  939. .BYTE 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
  940. .BYTE 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
  941. .BYTE 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
  942. .BYTE 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
  943. .BYTE 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
  944. .BYTE 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
  945. .BYTE 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
  946. .BYTE 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
  947. .BYTE 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
  948. .BYTE 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
  949. .BYTE 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
  950. .BYTE 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
  951. .BYTE 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
  952. .BYTE 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
  953. .BYTE 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
  954. .BYTE 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
  955. .BYTE 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
  956. .BYTE 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
  957. .BYTE 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
  958. .BYTE 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
  959. .BYTE 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
  960. .BYTE 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
  961. .BYTE 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
  962. .BYTE 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
  963. .BYTE 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
  964. .BYTE 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
  965. .BYTE 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
  966. .BYTE 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
  967. .BYTE 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
  968. .STRINGZ "AES for PA-RISC, CRYPTOGAMS by <appro\@openssl.org>"
  969. ___
  970. if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  971. =~ /GNU assembler/) {
  972. $gnuas = 1;
  973. }
  974. foreach (split("\n",$code)) {
  975. s/\`([^\`]*)\`/eval $1/ge;
  976. # translate made up instructions: _ror, _srm
  977. s/_ror(\s+)(%r[0-9]+),/shd$1$2,$2,/ or
  978. s/_srm(\s+%r[0-9]+),([0-9]+),/
  979. $SIZE_T==4 ? sprintf("extru%s,%d,8,",$1,31-$2)
  980. : sprintf("extrd,u%s,%d,8,",$1,63-$2)/e;
  981. s/(\.LEVEL\s+2\.0)W/$1w/ if ($gnuas && $SIZE_T==8);
  982. s/\.SPACE\s+\$TEXT\$/.text/ if ($gnuas && $SIZE_T==8);
  983. s/\.SUBSPA.*// if ($gnuas && $SIZE_T==8);
  984. s/,\*/,/ if ($SIZE_T==4);
  985. s/\bbv\b(.*\(%r2\))/bve$1/ if ($SIZE_T==8);
  986. print $_,"\n";
  987. }
  988. close STDOUT or die "error closing STDOUT: $!";