11f13597dSJung-uk Kim#!/usr/bin/env perl 21f13597dSJung-uk Kim 31f13597dSJung-uk Kim# ==================================================================== 41f13597dSJung-uk Kim# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL 51f13597dSJung-uk Kim# project. The module is, however, dual licensed under OpenSSL and 61f13597dSJung-uk Kim# CRYPTOGAMS licenses depending on where you obtain it. For further 71f13597dSJung-uk Kim# details see http://www.openssl.org/~appro/cryptogams/. 81f13597dSJung-uk Kim# ==================================================================== 91f13597dSJung-uk Kim 101f13597dSJung-uk Kim# Needs more work: key setup, CBC routine... 111f13597dSJung-uk Kim# 121f13597dSJung-uk Kim# ppc_AES_[en|de]crypt perform at 18 cycles per byte processed with 131f13597dSJung-uk Kim# 128-bit key, which is ~40% better than 64-bit code generated by gcc 141f13597dSJung-uk Kim# 4.0. But these are not the ones currently used! Their "compact" 151f13597dSJung-uk Kim# counterparts are, for security reason. ppc_AES_encrypt_compact runs 161f13597dSJung-uk Kim# at 1/2 of ppc_AES_encrypt speed, while ppc_AES_decrypt_compact - 171f13597dSJung-uk Kim# at 1/3 of ppc_AES_decrypt. 181f13597dSJung-uk Kim 191f13597dSJung-uk Kim# February 2010 201f13597dSJung-uk Kim# 211f13597dSJung-uk Kim# Rescheduling instructions to favour Power6 pipeline gave 10% 221f13597dSJung-uk Kim# performance improvement on the platfrom in question (and marginal 231f13597dSJung-uk Kim# improvement even on others). It should be noted that Power6 fails 241f13597dSJung-uk Kim# to process byte in 18 cycles, only in 23, because it fails to issue 251f13597dSJung-uk Kim# 4 load instructions in two cycles, only in 3. As result non-compact 261f13597dSJung-uk Kim# block subroutines are 25% slower than one would expect. Compact 271f13597dSJung-uk Kim# functions scale better, because they have pure computational part, 281f13597dSJung-uk Kim# which scales perfectly with clock frequency. To be specific 291f13597dSJung-uk Kim# ppc_AES_encrypt_compact operates at 42 cycles per byte, while 301f13597dSJung-uk Kim# ppc_AES_decrypt_compact - at 55 (in 64-bit build). 311f13597dSJung-uk Kim 321f13597dSJung-uk Kim$flavour = shift; 331f13597dSJung-uk Kim 341f13597dSJung-uk Kimif ($flavour =~ /64/) { 351f13597dSJung-uk Kim $SIZE_T =8; 361f13597dSJung-uk Kim $LRSAVE =2*$SIZE_T; 371f13597dSJung-uk Kim $STU ="stdu"; 381f13597dSJung-uk Kim $POP ="ld"; 391f13597dSJung-uk Kim $PUSH ="std"; 401f13597dSJung-uk Kim} elsif ($flavour =~ /32/) { 411f13597dSJung-uk Kim $SIZE_T =4; 421f13597dSJung-uk Kim $LRSAVE =$SIZE_T; 431f13597dSJung-uk Kim $STU ="stwu"; 441f13597dSJung-uk Kim $POP ="lwz"; 451f13597dSJung-uk Kim $PUSH ="stw"; 461f13597dSJung-uk Kim} else { die "nonsense $flavour"; } 471f13597dSJung-uk Kim 487bded2dbSJung-uk Kim$LITTLE_ENDIAN = ($flavour=~/le$/) ? $SIZE_T : 0; 497bded2dbSJung-uk Kim 501f13597dSJung-uk Kim$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 511f13597dSJung-uk Kim( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or 521f13597dSJung-uk Kim( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or 531f13597dSJung-uk Kimdie "can't locate ppc-xlate.pl"; 541f13597dSJung-uk Kim 551f13597dSJung-uk Kimopen STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!"; 561f13597dSJung-uk Kim 571f13597dSJung-uk Kim$FRAME=32*$SIZE_T; 581f13597dSJung-uk Kim 591f13597dSJung-uk Kimsub _data_word() 601f13597dSJung-uk Kim{ my $i; 611f13597dSJung-uk Kim while(defined($i=shift)) { $code.=sprintf"\t.long\t0x%08x,0x%08x\n",$i,$i; } 621f13597dSJung-uk Kim} 631f13597dSJung-uk Kim 641f13597dSJung-uk Kim$sp="r1"; 651f13597dSJung-uk Kim$toc="r2"; 661f13597dSJung-uk Kim$inp="r3"; 671f13597dSJung-uk Kim$out="r4"; 681f13597dSJung-uk Kim$key="r5"; 691f13597dSJung-uk Kim 701f13597dSJung-uk Kim$Tbl0="r3"; 711f13597dSJung-uk Kim$Tbl1="r6"; 721f13597dSJung-uk Kim$Tbl2="r7"; 737bded2dbSJung-uk Kim$Tbl3=$out; # stay away from "r2"; $out is offloaded to stack 741f13597dSJung-uk Kim 751f13597dSJung-uk Kim$s0="r8"; 761f13597dSJung-uk Kim$s1="r9"; 771f13597dSJung-uk Kim$s2="r10"; 781f13597dSJung-uk Kim$s3="r11"; 791f13597dSJung-uk Kim 801f13597dSJung-uk Kim$t0="r12"; 817bded2dbSJung-uk Kim$t1="r0"; # stay away from "r13"; 821f13597dSJung-uk Kim$t2="r14"; 831f13597dSJung-uk Kim$t3="r15"; 841f13597dSJung-uk Kim 851f13597dSJung-uk Kim$acc00="r16"; 861f13597dSJung-uk Kim$acc01="r17"; 871f13597dSJung-uk Kim$acc02="r18"; 881f13597dSJung-uk Kim$acc03="r19"; 891f13597dSJung-uk Kim 901f13597dSJung-uk Kim$acc04="r20"; 911f13597dSJung-uk Kim$acc05="r21"; 921f13597dSJung-uk Kim$acc06="r22"; 931f13597dSJung-uk Kim$acc07="r23"; 941f13597dSJung-uk Kim 951f13597dSJung-uk Kim$acc08="r24"; 961f13597dSJung-uk Kim$acc09="r25"; 971f13597dSJung-uk Kim$acc10="r26"; 981f13597dSJung-uk Kim$acc11="r27"; 991f13597dSJung-uk Kim 1001f13597dSJung-uk Kim$acc12="r28"; 1011f13597dSJung-uk Kim$acc13="r29"; 1021f13597dSJung-uk Kim$acc14="r30"; 1031f13597dSJung-uk Kim$acc15="r31"; 1041f13597dSJung-uk Kim 1051f13597dSJung-uk Kim$mask80=$Tbl2; 1061f13597dSJung-uk Kim$mask1b=$Tbl3; 1071f13597dSJung-uk Kim 1081f13597dSJung-uk Kim$code.=<<___; 1091f13597dSJung-uk Kim.machine "any" 1101f13597dSJung-uk Kim.text 1111f13597dSJung-uk Kim 1121f13597dSJung-uk Kim.align 7 1131f13597dSJung-uk KimLAES_Te: 1141f13597dSJung-uk Kim mflr r0 1151f13597dSJung-uk Kim bcl 20,31,\$+4 1161f13597dSJung-uk Kim mflr $Tbl0 ; vvvvv "distance" between . and 1st data entry 1171f13597dSJung-uk Kim addi $Tbl0,$Tbl0,`128-8` 1181f13597dSJung-uk Kim mtlr r0 1191f13597dSJung-uk Kim blr 1201f13597dSJung-uk Kim .long 0 1211f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 1221f13597dSJung-uk Kim .space `64-9*4` 1231f13597dSJung-uk KimLAES_Td: 1241f13597dSJung-uk Kim mflr r0 1251f13597dSJung-uk Kim bcl 20,31,\$+4 1261f13597dSJung-uk Kim mflr $Tbl0 ; vvvvvvvv "distance" between . and 1st data entry 1271f13597dSJung-uk Kim addi $Tbl0,$Tbl0,`128-64-8+2048+256` 1281f13597dSJung-uk Kim mtlr r0 1291f13597dSJung-uk Kim blr 1301f13597dSJung-uk Kim .long 0 1311f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 1321f13597dSJung-uk Kim .space `128-64-9*4` 1331f13597dSJung-uk Kim___ 1341f13597dSJung-uk Kim&_data_word( 1351f13597dSJung-uk Kim 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, 1361f13597dSJung-uk Kim 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554, 1371f13597dSJung-uk Kim 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, 1381f13597dSJung-uk Kim 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a, 1391f13597dSJung-uk Kim 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, 1401f13597dSJung-uk Kim 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b, 1411f13597dSJung-uk Kim 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, 1421f13597dSJung-uk Kim 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b, 1431f13597dSJung-uk Kim 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, 1441f13597dSJung-uk Kim 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f, 1451f13597dSJung-uk Kim 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, 1461f13597dSJung-uk Kim 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f, 1471f13597dSJung-uk Kim 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, 1481f13597dSJung-uk Kim 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5, 1491f13597dSJung-uk Kim 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, 1501f13597dSJung-uk Kim 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f, 1511f13597dSJung-uk Kim 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, 1521f13597dSJung-uk Kim 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb, 1531f13597dSJung-uk Kim 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, 1541f13597dSJung-uk Kim 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497, 1551f13597dSJung-uk Kim 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, 1561f13597dSJung-uk Kim 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed, 1571f13597dSJung-uk Kim 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, 1581f13597dSJung-uk Kim 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a, 1591f13597dSJung-uk Kim 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, 1601f13597dSJung-uk Kim 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594, 1611f13597dSJung-uk Kim 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, 1621f13597dSJung-uk Kim 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3, 1631f13597dSJung-uk Kim 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, 1641f13597dSJung-uk Kim 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504, 1651f13597dSJung-uk Kim 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, 1661f13597dSJung-uk Kim 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d, 1671f13597dSJung-uk Kim 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, 1681f13597dSJung-uk Kim 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739, 1691f13597dSJung-uk Kim 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, 1701f13597dSJung-uk Kim 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395, 1711f13597dSJung-uk Kim 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, 1721f13597dSJung-uk Kim 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883, 1731f13597dSJung-uk Kim 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, 1741f13597dSJung-uk Kim 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76, 1751f13597dSJung-uk Kim 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, 1761f13597dSJung-uk Kim 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4, 1771f13597dSJung-uk Kim 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, 1781f13597dSJung-uk Kim 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b, 1791f13597dSJung-uk Kim 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, 1801f13597dSJung-uk Kim 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0, 1811f13597dSJung-uk Kim 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, 1821f13597dSJung-uk Kim 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818, 1831f13597dSJung-uk Kim 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, 1841f13597dSJung-uk Kim 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651, 1851f13597dSJung-uk Kim 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, 1861f13597dSJung-uk Kim 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85, 1871f13597dSJung-uk Kim 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, 1881f13597dSJung-uk Kim 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12, 1891f13597dSJung-uk Kim 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, 1901f13597dSJung-uk Kim 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9, 1911f13597dSJung-uk Kim 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, 1921f13597dSJung-uk Kim 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7, 1931f13597dSJung-uk Kim 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, 1941f13597dSJung-uk Kim 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a, 1951f13597dSJung-uk Kim 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, 1961f13597dSJung-uk Kim 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8, 1971f13597dSJung-uk Kim 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, 1981f13597dSJung-uk Kim 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a); 1991f13597dSJung-uk Kim$code.=<<___; 2001f13597dSJung-uk Kim.byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 2011f13597dSJung-uk Kim.byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 2021f13597dSJung-uk Kim.byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 2031f13597dSJung-uk Kim.byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 2041f13597dSJung-uk Kim.byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc 2051f13597dSJung-uk Kim.byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 2061f13597dSJung-uk Kim.byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a 2071f13597dSJung-uk Kim.byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 2081f13597dSJung-uk Kim.byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 2091f13597dSJung-uk Kim.byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 2101f13597dSJung-uk Kim.byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b 2111f13597dSJung-uk Kim.byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf 2121f13597dSJung-uk Kim.byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 2131f13597dSJung-uk Kim.byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 2141f13597dSJung-uk Kim.byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 2151f13597dSJung-uk Kim.byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 2161f13597dSJung-uk Kim.byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 2171f13597dSJung-uk Kim.byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 2181f13597dSJung-uk Kim.byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 2191f13597dSJung-uk Kim.byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb 2201f13597dSJung-uk Kim.byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c 2211f13597dSJung-uk Kim.byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 2221f13597dSJung-uk Kim.byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 2231f13597dSJung-uk Kim.byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 2241f13597dSJung-uk Kim.byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 2251f13597dSJung-uk Kim.byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a 2261f13597dSJung-uk Kim.byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e 2271f13597dSJung-uk Kim.byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e 2281f13597dSJung-uk Kim.byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 2291f13597dSJung-uk Kim.byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf 2301f13597dSJung-uk Kim.byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 2311f13597dSJung-uk Kim.byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 2321f13597dSJung-uk Kim___ 2331f13597dSJung-uk Kim&_data_word( 2341f13597dSJung-uk Kim 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, 2351f13597dSJung-uk Kim 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393, 2361f13597dSJung-uk Kim 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, 2371f13597dSJung-uk Kim 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f, 2381f13597dSJung-uk Kim 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, 2391f13597dSJung-uk Kim 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6, 2401f13597dSJung-uk Kim 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, 2411f13597dSJung-uk Kim 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844, 2421f13597dSJung-uk Kim 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, 2431f13597dSJung-uk Kim 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4, 2441f13597dSJung-uk Kim 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, 2451f13597dSJung-uk Kim 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94, 2461f13597dSJung-uk Kim 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, 2471f13597dSJung-uk Kim 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a, 2481f13597dSJung-uk Kim 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, 2491f13597dSJung-uk Kim 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c, 2501f13597dSJung-uk Kim 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, 2511f13597dSJung-uk Kim 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a, 2521f13597dSJung-uk Kim 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, 2531f13597dSJung-uk Kim 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051, 2541f13597dSJung-uk Kim 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, 2551f13597dSJung-uk Kim 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff, 2561f13597dSJung-uk Kim 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, 2571f13597dSJung-uk Kim 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb, 2581f13597dSJung-uk Kim 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, 2591f13597dSJung-uk Kim 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e, 2601f13597dSJung-uk Kim 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, 2611f13597dSJung-uk Kim 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a, 2621f13597dSJung-uk Kim 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, 2631f13597dSJung-uk Kim 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16, 2641f13597dSJung-uk Kim 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, 2651f13597dSJung-uk Kim 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8, 2661f13597dSJung-uk Kim 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, 2671f13597dSJung-uk Kim 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34, 2681f13597dSJung-uk Kim 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, 2691f13597dSJung-uk Kim 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120, 2701f13597dSJung-uk Kim 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, 2711f13597dSJung-uk Kim 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0, 2721f13597dSJung-uk Kim 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, 2731f13597dSJung-uk Kim 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef, 2741f13597dSJung-uk Kim 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, 2751f13597dSJung-uk Kim 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4, 2761f13597dSJung-uk Kim 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, 2771f13597dSJung-uk Kim 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5, 2781f13597dSJung-uk Kim 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, 2791f13597dSJung-uk Kim 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b, 2801f13597dSJung-uk Kim 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, 2811f13597dSJung-uk Kim 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6, 2821f13597dSJung-uk Kim 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, 2831f13597dSJung-uk Kim 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0, 2841f13597dSJung-uk Kim 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, 2851f13597dSJung-uk Kim 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f, 2861f13597dSJung-uk Kim 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, 2871f13597dSJung-uk Kim 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f, 2881f13597dSJung-uk Kim 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, 2891f13597dSJung-uk Kim 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713, 2901f13597dSJung-uk Kim 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, 2911f13597dSJung-uk Kim 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c, 2921f13597dSJung-uk Kim 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, 2931f13597dSJung-uk Kim 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86, 2941f13597dSJung-uk Kim 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, 2951f13597dSJung-uk Kim 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541, 2961f13597dSJung-uk Kim 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, 2971f13597dSJung-uk Kim 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742); 2981f13597dSJung-uk Kim$code.=<<___; 2991f13597dSJung-uk Kim.byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 3001f13597dSJung-uk Kim.byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb 3011f13597dSJung-uk Kim.byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 3021f13597dSJung-uk Kim.byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb 3031f13597dSJung-uk Kim.byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d 3041f13597dSJung-uk Kim.byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e 3051f13597dSJung-uk Kim.byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 3061f13597dSJung-uk Kim.byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 3071f13597dSJung-uk Kim.byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 3081f13597dSJung-uk Kim.byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 3091f13597dSJung-uk Kim.byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda 3101f13597dSJung-uk Kim.byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 3111f13597dSJung-uk Kim.byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a 3121f13597dSJung-uk Kim.byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 3131f13597dSJung-uk Kim.byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 3141f13597dSJung-uk Kim.byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b 3151f13597dSJung-uk Kim.byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea 3161f13597dSJung-uk Kim.byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 3171f13597dSJung-uk Kim.byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 3181f13597dSJung-uk Kim.byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e 3191f13597dSJung-uk Kim.byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 3201f13597dSJung-uk Kim.byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b 3211f13597dSJung-uk Kim.byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 3221f13597dSJung-uk Kim.byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 3231f13597dSJung-uk Kim.byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 3241f13597dSJung-uk Kim.byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f 3251f13597dSJung-uk Kim.byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d 3261f13597dSJung-uk Kim.byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef 3271f13597dSJung-uk Kim.byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 3281f13597dSJung-uk Kim.byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 3291f13597dSJung-uk Kim.byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 3301f13597dSJung-uk Kim.byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d 3311f13597dSJung-uk Kim 3321f13597dSJung-uk Kim 3331f13597dSJung-uk Kim.globl .AES_encrypt 3341f13597dSJung-uk Kim.align 7 3351f13597dSJung-uk Kim.AES_encrypt: 3361f13597dSJung-uk Kim $STU $sp,-$FRAME($sp) 3371f13597dSJung-uk Kim mflr r0 3381f13597dSJung-uk Kim 3397bded2dbSJung-uk Kim $PUSH $out,`$FRAME-$SIZE_T*19`($sp) 3401f13597dSJung-uk Kim $PUSH r14,`$FRAME-$SIZE_T*18`($sp) 3411f13597dSJung-uk Kim $PUSH r15,`$FRAME-$SIZE_T*17`($sp) 3421f13597dSJung-uk Kim $PUSH r16,`$FRAME-$SIZE_T*16`($sp) 3431f13597dSJung-uk Kim $PUSH r17,`$FRAME-$SIZE_T*15`($sp) 3441f13597dSJung-uk Kim $PUSH r18,`$FRAME-$SIZE_T*14`($sp) 3451f13597dSJung-uk Kim $PUSH r19,`$FRAME-$SIZE_T*13`($sp) 3461f13597dSJung-uk Kim $PUSH r20,`$FRAME-$SIZE_T*12`($sp) 3471f13597dSJung-uk Kim $PUSH r21,`$FRAME-$SIZE_T*11`($sp) 3481f13597dSJung-uk Kim $PUSH r22,`$FRAME-$SIZE_T*10`($sp) 3491f13597dSJung-uk Kim $PUSH r23,`$FRAME-$SIZE_T*9`($sp) 3501f13597dSJung-uk Kim $PUSH r24,`$FRAME-$SIZE_T*8`($sp) 3511f13597dSJung-uk Kim $PUSH r25,`$FRAME-$SIZE_T*7`($sp) 3521f13597dSJung-uk Kim $PUSH r26,`$FRAME-$SIZE_T*6`($sp) 3531f13597dSJung-uk Kim $PUSH r27,`$FRAME-$SIZE_T*5`($sp) 3541f13597dSJung-uk Kim $PUSH r28,`$FRAME-$SIZE_T*4`($sp) 3551f13597dSJung-uk Kim $PUSH r29,`$FRAME-$SIZE_T*3`($sp) 3561f13597dSJung-uk Kim $PUSH r30,`$FRAME-$SIZE_T*2`($sp) 3571f13597dSJung-uk Kim $PUSH r31,`$FRAME-$SIZE_T*1`($sp) 3581f13597dSJung-uk Kim $PUSH r0,`$FRAME+$LRSAVE`($sp) 3591f13597dSJung-uk Kim 3601f13597dSJung-uk Kim andi. $t0,$inp,3 3611f13597dSJung-uk Kim andi. $t1,$out,3 3621f13597dSJung-uk Kim or. $t0,$t0,$t1 3631f13597dSJung-uk Kim bne Lenc_unaligned 3641f13597dSJung-uk Kim 3651f13597dSJung-uk KimLenc_unaligned_ok: 3667bded2dbSJung-uk Kim___ 3677bded2dbSJung-uk Kim$code.=<<___ if (!$LITTLE_ENDIAN); 3681f13597dSJung-uk Kim lwz $s0,0($inp) 3691f13597dSJung-uk Kim lwz $s1,4($inp) 3701f13597dSJung-uk Kim lwz $s2,8($inp) 3711f13597dSJung-uk Kim lwz $s3,12($inp) 3727bded2dbSJung-uk Kim___ 3737bded2dbSJung-uk Kim$code.=<<___ if ($LITTLE_ENDIAN); 3747bded2dbSJung-uk Kim lwz $t0,0($inp) 3757bded2dbSJung-uk Kim lwz $t1,4($inp) 3767bded2dbSJung-uk Kim lwz $t2,8($inp) 3777bded2dbSJung-uk Kim lwz $t3,12($inp) 3787bded2dbSJung-uk Kim rotlwi $s0,$t0,8 3797bded2dbSJung-uk Kim rotlwi $s1,$t1,8 3807bded2dbSJung-uk Kim rotlwi $s2,$t2,8 3817bded2dbSJung-uk Kim rotlwi $s3,$t3,8 3827bded2dbSJung-uk Kim rlwimi $s0,$t0,24,0,7 3837bded2dbSJung-uk Kim rlwimi $s1,$t1,24,0,7 3847bded2dbSJung-uk Kim rlwimi $s2,$t2,24,0,7 3857bded2dbSJung-uk Kim rlwimi $s3,$t3,24,0,7 3867bded2dbSJung-uk Kim rlwimi $s0,$t0,24,16,23 3877bded2dbSJung-uk Kim rlwimi $s1,$t1,24,16,23 3887bded2dbSJung-uk Kim rlwimi $s2,$t2,24,16,23 3897bded2dbSJung-uk Kim rlwimi $s3,$t3,24,16,23 3907bded2dbSJung-uk Kim___ 3917bded2dbSJung-uk Kim$code.=<<___; 3921f13597dSJung-uk Kim bl LAES_Te 3931f13597dSJung-uk Kim bl Lppc_AES_encrypt_compact 3947bded2dbSJung-uk Kim $POP $out,`$FRAME-$SIZE_T*19`($sp) 3957bded2dbSJung-uk Kim___ 3967bded2dbSJung-uk Kim$code.=<<___ if ($LITTLE_ENDIAN); 3977bded2dbSJung-uk Kim rotlwi $t0,$s0,8 3987bded2dbSJung-uk Kim rotlwi $t1,$s1,8 3997bded2dbSJung-uk Kim rotlwi $t2,$s2,8 4007bded2dbSJung-uk Kim rotlwi $t3,$s3,8 4017bded2dbSJung-uk Kim rlwimi $t0,$s0,24,0,7 4027bded2dbSJung-uk Kim rlwimi $t1,$s1,24,0,7 4037bded2dbSJung-uk Kim rlwimi $t2,$s2,24,0,7 4047bded2dbSJung-uk Kim rlwimi $t3,$s3,24,0,7 4057bded2dbSJung-uk Kim rlwimi $t0,$s0,24,16,23 4067bded2dbSJung-uk Kim rlwimi $t1,$s1,24,16,23 4077bded2dbSJung-uk Kim rlwimi $t2,$s2,24,16,23 4087bded2dbSJung-uk Kim rlwimi $t3,$s3,24,16,23 4097bded2dbSJung-uk Kim stw $t0,0($out) 4107bded2dbSJung-uk Kim stw $t1,4($out) 4117bded2dbSJung-uk Kim stw $t2,8($out) 4127bded2dbSJung-uk Kim stw $t3,12($out) 4137bded2dbSJung-uk Kim___ 4147bded2dbSJung-uk Kim$code.=<<___ if (!$LITTLE_ENDIAN); 4151f13597dSJung-uk Kim stw $s0,0($out) 4161f13597dSJung-uk Kim stw $s1,4($out) 4171f13597dSJung-uk Kim stw $s2,8($out) 4181f13597dSJung-uk Kim stw $s3,12($out) 4197bded2dbSJung-uk Kim___ 4207bded2dbSJung-uk Kim$code.=<<___; 4211f13597dSJung-uk Kim b Lenc_done 4221f13597dSJung-uk Kim 4231f13597dSJung-uk KimLenc_unaligned: 4241f13597dSJung-uk Kim subfic $t0,$inp,4096 4251f13597dSJung-uk Kim subfic $t1,$out,4096 4261f13597dSJung-uk Kim andi. $t0,$t0,4096-16 4271f13597dSJung-uk Kim beq Lenc_xpage 4281f13597dSJung-uk Kim andi. $t1,$t1,4096-16 4291f13597dSJung-uk Kim bne Lenc_unaligned_ok 4301f13597dSJung-uk Kim 4311f13597dSJung-uk KimLenc_xpage: 4321f13597dSJung-uk Kim lbz $acc00,0($inp) 4331f13597dSJung-uk Kim lbz $acc01,1($inp) 4341f13597dSJung-uk Kim lbz $acc02,2($inp) 4351f13597dSJung-uk Kim lbz $s0,3($inp) 4361f13597dSJung-uk Kim lbz $acc04,4($inp) 4371f13597dSJung-uk Kim lbz $acc05,5($inp) 4381f13597dSJung-uk Kim lbz $acc06,6($inp) 4391f13597dSJung-uk Kim lbz $s1,7($inp) 4401f13597dSJung-uk Kim lbz $acc08,8($inp) 4411f13597dSJung-uk Kim lbz $acc09,9($inp) 4421f13597dSJung-uk Kim lbz $acc10,10($inp) 4431f13597dSJung-uk Kim insrwi $s0,$acc00,8,0 4441f13597dSJung-uk Kim lbz $s2,11($inp) 4451f13597dSJung-uk Kim insrwi $s1,$acc04,8,0 4461f13597dSJung-uk Kim lbz $acc12,12($inp) 4471f13597dSJung-uk Kim insrwi $s0,$acc01,8,8 4481f13597dSJung-uk Kim lbz $acc13,13($inp) 4491f13597dSJung-uk Kim insrwi $s1,$acc05,8,8 4501f13597dSJung-uk Kim lbz $acc14,14($inp) 4511f13597dSJung-uk Kim insrwi $s0,$acc02,8,16 4521f13597dSJung-uk Kim lbz $s3,15($inp) 4531f13597dSJung-uk Kim insrwi $s1,$acc06,8,16 4541f13597dSJung-uk Kim insrwi $s2,$acc08,8,0 4551f13597dSJung-uk Kim insrwi $s3,$acc12,8,0 4561f13597dSJung-uk Kim insrwi $s2,$acc09,8,8 4571f13597dSJung-uk Kim insrwi $s3,$acc13,8,8 4581f13597dSJung-uk Kim insrwi $s2,$acc10,8,16 4591f13597dSJung-uk Kim insrwi $s3,$acc14,8,16 4601f13597dSJung-uk Kim 4611f13597dSJung-uk Kim bl LAES_Te 4621f13597dSJung-uk Kim bl Lppc_AES_encrypt_compact 4637bded2dbSJung-uk Kim $POP $out,`$FRAME-$SIZE_T*19`($sp) 4641f13597dSJung-uk Kim 4651f13597dSJung-uk Kim extrwi $acc00,$s0,8,0 4661f13597dSJung-uk Kim extrwi $acc01,$s0,8,8 4671f13597dSJung-uk Kim stb $acc00,0($out) 4681f13597dSJung-uk Kim extrwi $acc02,$s0,8,16 4691f13597dSJung-uk Kim stb $acc01,1($out) 4701f13597dSJung-uk Kim stb $acc02,2($out) 4711f13597dSJung-uk Kim extrwi $acc04,$s1,8,0 4721f13597dSJung-uk Kim stb $s0,3($out) 4731f13597dSJung-uk Kim extrwi $acc05,$s1,8,8 4741f13597dSJung-uk Kim stb $acc04,4($out) 4751f13597dSJung-uk Kim extrwi $acc06,$s1,8,16 4761f13597dSJung-uk Kim stb $acc05,5($out) 4771f13597dSJung-uk Kim stb $acc06,6($out) 4781f13597dSJung-uk Kim extrwi $acc08,$s2,8,0 4791f13597dSJung-uk Kim stb $s1,7($out) 4801f13597dSJung-uk Kim extrwi $acc09,$s2,8,8 4811f13597dSJung-uk Kim stb $acc08,8($out) 4821f13597dSJung-uk Kim extrwi $acc10,$s2,8,16 4831f13597dSJung-uk Kim stb $acc09,9($out) 4841f13597dSJung-uk Kim stb $acc10,10($out) 4851f13597dSJung-uk Kim extrwi $acc12,$s3,8,0 4861f13597dSJung-uk Kim stb $s2,11($out) 4871f13597dSJung-uk Kim extrwi $acc13,$s3,8,8 4881f13597dSJung-uk Kim stb $acc12,12($out) 4891f13597dSJung-uk Kim extrwi $acc14,$s3,8,16 4901f13597dSJung-uk Kim stb $acc13,13($out) 4911f13597dSJung-uk Kim stb $acc14,14($out) 4921f13597dSJung-uk Kim stb $s3,15($out) 4931f13597dSJung-uk Kim 4941f13597dSJung-uk KimLenc_done: 4951f13597dSJung-uk Kim $POP r0,`$FRAME+$LRSAVE`($sp) 4961f13597dSJung-uk Kim $POP r14,`$FRAME-$SIZE_T*18`($sp) 4971f13597dSJung-uk Kim $POP r15,`$FRAME-$SIZE_T*17`($sp) 4981f13597dSJung-uk Kim $POP r16,`$FRAME-$SIZE_T*16`($sp) 4991f13597dSJung-uk Kim $POP r17,`$FRAME-$SIZE_T*15`($sp) 5001f13597dSJung-uk Kim $POP r18,`$FRAME-$SIZE_T*14`($sp) 5011f13597dSJung-uk Kim $POP r19,`$FRAME-$SIZE_T*13`($sp) 5021f13597dSJung-uk Kim $POP r20,`$FRAME-$SIZE_T*12`($sp) 5031f13597dSJung-uk Kim $POP r21,`$FRAME-$SIZE_T*11`($sp) 5041f13597dSJung-uk Kim $POP r22,`$FRAME-$SIZE_T*10`($sp) 5051f13597dSJung-uk Kim $POP r23,`$FRAME-$SIZE_T*9`($sp) 5061f13597dSJung-uk Kim $POP r24,`$FRAME-$SIZE_T*8`($sp) 5071f13597dSJung-uk Kim $POP r25,`$FRAME-$SIZE_T*7`($sp) 5081f13597dSJung-uk Kim $POP r26,`$FRAME-$SIZE_T*6`($sp) 5091f13597dSJung-uk Kim $POP r27,`$FRAME-$SIZE_T*5`($sp) 5101f13597dSJung-uk Kim $POP r28,`$FRAME-$SIZE_T*4`($sp) 5111f13597dSJung-uk Kim $POP r29,`$FRAME-$SIZE_T*3`($sp) 5121f13597dSJung-uk Kim $POP r30,`$FRAME-$SIZE_T*2`($sp) 5131f13597dSJung-uk Kim $POP r31,`$FRAME-$SIZE_T*1`($sp) 5141f13597dSJung-uk Kim mtlr r0 5151f13597dSJung-uk Kim addi $sp,$sp,$FRAME 5161f13597dSJung-uk Kim blr 5171f13597dSJung-uk Kim .long 0 5181f13597dSJung-uk Kim .byte 0,12,4,1,0x80,18,3,0 5191f13597dSJung-uk Kim .long 0 5201f13597dSJung-uk Kim 5211f13597dSJung-uk Kim.align 5 5221f13597dSJung-uk KimLppc_AES_encrypt: 5231f13597dSJung-uk Kim lwz $acc00,240($key) 5241f13597dSJung-uk Kim addi $Tbl1,$Tbl0,3 5251f13597dSJung-uk Kim lwz $t0,0($key) 5261f13597dSJung-uk Kim addi $Tbl2,$Tbl0,2 5271f13597dSJung-uk Kim lwz $t1,4($key) 5281f13597dSJung-uk Kim addi $Tbl3,$Tbl0,1 5291f13597dSJung-uk Kim lwz $t2,8($key) 5301f13597dSJung-uk Kim addi $acc00,$acc00,-1 5311f13597dSJung-uk Kim lwz $t3,12($key) 5321f13597dSJung-uk Kim addi $key,$key,16 5331f13597dSJung-uk Kim xor $s0,$s0,$t0 5341f13597dSJung-uk Kim xor $s1,$s1,$t1 5351f13597dSJung-uk Kim xor $s2,$s2,$t2 5361f13597dSJung-uk Kim xor $s3,$s3,$t3 5371f13597dSJung-uk Kim mtctr $acc00 5381f13597dSJung-uk Kim.align 4 5391f13597dSJung-uk KimLenc_loop: 5401f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24+3`,21,28 5411f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24+3`,21,28 5421f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24+3`,21,28 5431f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24+3`,21,28 5441f13597dSJung-uk Kim lwz $t0,0($key) 5451f13597dSJung-uk Kim rlwinm $acc04,$s1,`32-16+3`,21,28 5461f13597dSJung-uk Kim lwz $t1,4($key) 5471f13597dSJung-uk Kim rlwinm $acc05,$s2,`32-16+3`,21,28 5481f13597dSJung-uk Kim lwz $t2,8($key) 5491f13597dSJung-uk Kim rlwinm $acc06,$s3,`32-16+3`,21,28 5501f13597dSJung-uk Kim lwz $t3,12($key) 5511f13597dSJung-uk Kim rlwinm $acc07,$s0,`32-16+3`,21,28 5521f13597dSJung-uk Kim lwzx $acc00,$Tbl0,$acc00 5531f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8+3`,21,28 5541f13597dSJung-uk Kim lwzx $acc01,$Tbl0,$acc01 5551f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8+3`,21,28 5561f13597dSJung-uk Kim lwzx $acc02,$Tbl0,$acc02 5571f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8+3`,21,28 5581f13597dSJung-uk Kim lwzx $acc03,$Tbl0,$acc03 5591f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8+3`,21,28 5601f13597dSJung-uk Kim lwzx $acc04,$Tbl1,$acc04 5611f13597dSJung-uk Kim rlwinm $acc12,$s3,`0+3`,21,28 5621f13597dSJung-uk Kim lwzx $acc05,$Tbl1,$acc05 5631f13597dSJung-uk Kim rlwinm $acc13,$s0,`0+3`,21,28 5641f13597dSJung-uk Kim lwzx $acc06,$Tbl1,$acc06 5651f13597dSJung-uk Kim rlwinm $acc14,$s1,`0+3`,21,28 5661f13597dSJung-uk Kim lwzx $acc07,$Tbl1,$acc07 5671f13597dSJung-uk Kim rlwinm $acc15,$s2,`0+3`,21,28 5681f13597dSJung-uk Kim lwzx $acc08,$Tbl2,$acc08 5691f13597dSJung-uk Kim xor $t0,$t0,$acc00 5701f13597dSJung-uk Kim lwzx $acc09,$Tbl2,$acc09 5711f13597dSJung-uk Kim xor $t1,$t1,$acc01 5721f13597dSJung-uk Kim lwzx $acc10,$Tbl2,$acc10 5731f13597dSJung-uk Kim xor $t2,$t2,$acc02 5741f13597dSJung-uk Kim lwzx $acc11,$Tbl2,$acc11 5751f13597dSJung-uk Kim xor $t3,$t3,$acc03 5761f13597dSJung-uk Kim lwzx $acc12,$Tbl3,$acc12 5771f13597dSJung-uk Kim xor $t0,$t0,$acc04 5781f13597dSJung-uk Kim lwzx $acc13,$Tbl3,$acc13 5791f13597dSJung-uk Kim xor $t1,$t1,$acc05 5801f13597dSJung-uk Kim lwzx $acc14,$Tbl3,$acc14 5811f13597dSJung-uk Kim xor $t2,$t2,$acc06 5821f13597dSJung-uk Kim lwzx $acc15,$Tbl3,$acc15 5831f13597dSJung-uk Kim xor $t3,$t3,$acc07 5841f13597dSJung-uk Kim xor $t0,$t0,$acc08 5851f13597dSJung-uk Kim xor $t1,$t1,$acc09 5861f13597dSJung-uk Kim xor $t2,$t2,$acc10 5871f13597dSJung-uk Kim xor $t3,$t3,$acc11 5881f13597dSJung-uk Kim xor $s0,$t0,$acc12 5891f13597dSJung-uk Kim xor $s1,$t1,$acc13 5901f13597dSJung-uk Kim xor $s2,$t2,$acc14 5911f13597dSJung-uk Kim xor $s3,$t3,$acc15 5921f13597dSJung-uk Kim addi $key,$key,16 593*b8721c16SJung-uk Kim bdnz Lenc_loop 5941f13597dSJung-uk Kim 5951f13597dSJung-uk Kim addi $Tbl2,$Tbl0,2048 5961f13597dSJung-uk Kim nop 5971f13597dSJung-uk Kim lwz $t0,0($key) 5981f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24`,24,31 5991f13597dSJung-uk Kim lwz $t1,4($key) 6001f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24`,24,31 6011f13597dSJung-uk Kim lwz $t2,8($key) 6021f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24`,24,31 6031f13597dSJung-uk Kim lwz $t3,12($key) 6041f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24`,24,31 6051f13597dSJung-uk Kim lwz $acc08,`2048+0`($Tbl0) ! prefetch Te4 6061f13597dSJung-uk Kim rlwinm $acc04,$s1,`32-16`,24,31 6071f13597dSJung-uk Kim lwz $acc09,`2048+32`($Tbl0) 6081f13597dSJung-uk Kim rlwinm $acc05,$s2,`32-16`,24,31 6091f13597dSJung-uk Kim lwz $acc10,`2048+64`($Tbl0) 6101f13597dSJung-uk Kim rlwinm $acc06,$s3,`32-16`,24,31 6111f13597dSJung-uk Kim lwz $acc11,`2048+96`($Tbl0) 6121f13597dSJung-uk Kim rlwinm $acc07,$s0,`32-16`,24,31 6131f13597dSJung-uk Kim lwz $acc12,`2048+128`($Tbl0) 6141f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8`,24,31 6151f13597dSJung-uk Kim lwz $acc13,`2048+160`($Tbl0) 6161f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8`,24,31 6171f13597dSJung-uk Kim lwz $acc14,`2048+192`($Tbl0) 6181f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8`,24,31 6191f13597dSJung-uk Kim lwz $acc15,`2048+224`($Tbl0) 6201f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8`,24,31 6211f13597dSJung-uk Kim lbzx $acc00,$Tbl2,$acc00 6221f13597dSJung-uk Kim rlwinm $acc12,$s3,`0`,24,31 6231f13597dSJung-uk Kim lbzx $acc01,$Tbl2,$acc01 6241f13597dSJung-uk Kim rlwinm $acc13,$s0,`0`,24,31 6251f13597dSJung-uk Kim lbzx $acc02,$Tbl2,$acc02 6261f13597dSJung-uk Kim rlwinm $acc14,$s1,`0`,24,31 6271f13597dSJung-uk Kim lbzx $acc03,$Tbl2,$acc03 6281f13597dSJung-uk Kim rlwinm $acc15,$s2,`0`,24,31 6291f13597dSJung-uk Kim lbzx $acc04,$Tbl2,$acc04 6301f13597dSJung-uk Kim rlwinm $s0,$acc00,24,0,7 6311f13597dSJung-uk Kim lbzx $acc05,$Tbl2,$acc05 6321f13597dSJung-uk Kim rlwinm $s1,$acc01,24,0,7 6331f13597dSJung-uk Kim lbzx $acc06,$Tbl2,$acc06 6341f13597dSJung-uk Kim rlwinm $s2,$acc02,24,0,7 6351f13597dSJung-uk Kim lbzx $acc07,$Tbl2,$acc07 6361f13597dSJung-uk Kim rlwinm $s3,$acc03,24,0,7 6371f13597dSJung-uk Kim lbzx $acc08,$Tbl2,$acc08 6381f13597dSJung-uk Kim rlwimi $s0,$acc04,16,8,15 6391f13597dSJung-uk Kim lbzx $acc09,$Tbl2,$acc09 6401f13597dSJung-uk Kim rlwimi $s1,$acc05,16,8,15 6411f13597dSJung-uk Kim lbzx $acc10,$Tbl2,$acc10 6421f13597dSJung-uk Kim rlwimi $s2,$acc06,16,8,15 6431f13597dSJung-uk Kim lbzx $acc11,$Tbl2,$acc11 6441f13597dSJung-uk Kim rlwimi $s3,$acc07,16,8,15 6451f13597dSJung-uk Kim lbzx $acc12,$Tbl2,$acc12 6461f13597dSJung-uk Kim rlwimi $s0,$acc08,8,16,23 6471f13597dSJung-uk Kim lbzx $acc13,$Tbl2,$acc13 6481f13597dSJung-uk Kim rlwimi $s1,$acc09,8,16,23 6491f13597dSJung-uk Kim lbzx $acc14,$Tbl2,$acc14 6501f13597dSJung-uk Kim rlwimi $s2,$acc10,8,16,23 6511f13597dSJung-uk Kim lbzx $acc15,$Tbl2,$acc15 6521f13597dSJung-uk Kim rlwimi $s3,$acc11,8,16,23 6531f13597dSJung-uk Kim or $s0,$s0,$acc12 6541f13597dSJung-uk Kim or $s1,$s1,$acc13 6551f13597dSJung-uk Kim or $s2,$s2,$acc14 6561f13597dSJung-uk Kim or $s3,$s3,$acc15 6571f13597dSJung-uk Kim xor $s0,$s0,$t0 6581f13597dSJung-uk Kim xor $s1,$s1,$t1 6591f13597dSJung-uk Kim xor $s2,$s2,$t2 6601f13597dSJung-uk Kim xor $s3,$s3,$t3 6611f13597dSJung-uk Kim blr 6621f13597dSJung-uk Kim .long 0 6631f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 6641f13597dSJung-uk Kim 6651f13597dSJung-uk Kim.align 4 6661f13597dSJung-uk KimLppc_AES_encrypt_compact: 6671f13597dSJung-uk Kim lwz $acc00,240($key) 6681f13597dSJung-uk Kim addi $Tbl1,$Tbl0,2048 6691f13597dSJung-uk Kim lwz $t0,0($key) 6701f13597dSJung-uk Kim lis $mask80,0x8080 6711f13597dSJung-uk Kim lwz $t1,4($key) 6721f13597dSJung-uk Kim lis $mask1b,0x1b1b 6731f13597dSJung-uk Kim lwz $t2,8($key) 6741f13597dSJung-uk Kim ori $mask80,$mask80,0x8080 6751f13597dSJung-uk Kim lwz $t3,12($key) 6761f13597dSJung-uk Kim ori $mask1b,$mask1b,0x1b1b 6771f13597dSJung-uk Kim addi $key,$key,16 6781f13597dSJung-uk Kim mtctr $acc00 6791f13597dSJung-uk Kim.align 4 6801f13597dSJung-uk KimLenc_compact_loop: 6811f13597dSJung-uk Kim xor $s0,$s0,$t0 6821f13597dSJung-uk Kim xor $s1,$s1,$t1 6831f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24`,24,31 6841f13597dSJung-uk Kim xor $s2,$s2,$t2 6851f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24`,24,31 6861f13597dSJung-uk Kim xor $s3,$s3,$t3 6871f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24`,24,31 6881f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24`,24,31 6891f13597dSJung-uk Kim rlwinm $acc04,$s1,`32-16`,24,31 6901f13597dSJung-uk Kim rlwinm $acc05,$s2,`32-16`,24,31 6911f13597dSJung-uk Kim rlwinm $acc06,$s3,`32-16`,24,31 6921f13597dSJung-uk Kim rlwinm $acc07,$s0,`32-16`,24,31 6931f13597dSJung-uk Kim lbzx $acc00,$Tbl1,$acc00 6941f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8`,24,31 6951f13597dSJung-uk Kim lbzx $acc01,$Tbl1,$acc01 6961f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8`,24,31 6971f13597dSJung-uk Kim lbzx $acc02,$Tbl1,$acc02 6981f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8`,24,31 6991f13597dSJung-uk Kim lbzx $acc03,$Tbl1,$acc03 7001f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8`,24,31 7011f13597dSJung-uk Kim lbzx $acc04,$Tbl1,$acc04 7021f13597dSJung-uk Kim rlwinm $acc12,$s3,`0`,24,31 7031f13597dSJung-uk Kim lbzx $acc05,$Tbl1,$acc05 7041f13597dSJung-uk Kim rlwinm $acc13,$s0,`0`,24,31 7051f13597dSJung-uk Kim lbzx $acc06,$Tbl1,$acc06 7061f13597dSJung-uk Kim rlwinm $acc14,$s1,`0`,24,31 7071f13597dSJung-uk Kim lbzx $acc07,$Tbl1,$acc07 7081f13597dSJung-uk Kim rlwinm $acc15,$s2,`0`,24,31 7091f13597dSJung-uk Kim lbzx $acc08,$Tbl1,$acc08 7101f13597dSJung-uk Kim rlwinm $s0,$acc00,24,0,7 7111f13597dSJung-uk Kim lbzx $acc09,$Tbl1,$acc09 7121f13597dSJung-uk Kim rlwinm $s1,$acc01,24,0,7 7131f13597dSJung-uk Kim lbzx $acc10,$Tbl1,$acc10 7141f13597dSJung-uk Kim rlwinm $s2,$acc02,24,0,7 7151f13597dSJung-uk Kim lbzx $acc11,$Tbl1,$acc11 7161f13597dSJung-uk Kim rlwinm $s3,$acc03,24,0,7 7171f13597dSJung-uk Kim lbzx $acc12,$Tbl1,$acc12 7181f13597dSJung-uk Kim rlwimi $s0,$acc04,16,8,15 7191f13597dSJung-uk Kim lbzx $acc13,$Tbl1,$acc13 7201f13597dSJung-uk Kim rlwimi $s1,$acc05,16,8,15 7211f13597dSJung-uk Kim lbzx $acc14,$Tbl1,$acc14 7221f13597dSJung-uk Kim rlwimi $s2,$acc06,16,8,15 7231f13597dSJung-uk Kim lbzx $acc15,$Tbl1,$acc15 7241f13597dSJung-uk Kim rlwimi $s3,$acc07,16,8,15 7251f13597dSJung-uk Kim rlwimi $s0,$acc08,8,16,23 7261f13597dSJung-uk Kim rlwimi $s1,$acc09,8,16,23 7271f13597dSJung-uk Kim rlwimi $s2,$acc10,8,16,23 7281f13597dSJung-uk Kim rlwimi $s3,$acc11,8,16,23 7291f13597dSJung-uk Kim lwz $t0,0($key) 7301f13597dSJung-uk Kim or $s0,$s0,$acc12 7311f13597dSJung-uk Kim lwz $t1,4($key) 7321f13597dSJung-uk Kim or $s1,$s1,$acc13 7331f13597dSJung-uk Kim lwz $t2,8($key) 7341f13597dSJung-uk Kim or $s2,$s2,$acc14 7351f13597dSJung-uk Kim lwz $t3,12($key) 7361f13597dSJung-uk Kim or $s3,$s3,$acc15 7371f13597dSJung-uk Kim 7381f13597dSJung-uk Kim addi $key,$key,16 7391f13597dSJung-uk Kim bdz Lenc_compact_done 7401f13597dSJung-uk Kim 7411f13597dSJung-uk Kim and $acc00,$s0,$mask80 # r1=r0&0x80808080 7421f13597dSJung-uk Kim and $acc01,$s1,$mask80 7431f13597dSJung-uk Kim and $acc02,$s2,$mask80 7441f13597dSJung-uk Kim and $acc03,$s3,$mask80 7451f13597dSJung-uk Kim srwi $acc04,$acc00,7 # r1>>7 7461f13597dSJung-uk Kim andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f 7471f13597dSJung-uk Kim srwi $acc05,$acc01,7 7481f13597dSJung-uk Kim andc $acc09,$s1,$mask80 7491f13597dSJung-uk Kim srwi $acc06,$acc02,7 7501f13597dSJung-uk Kim andc $acc10,$s2,$mask80 7511f13597dSJung-uk Kim srwi $acc07,$acc03,7 7521f13597dSJung-uk Kim andc $acc11,$s3,$mask80 7531f13597dSJung-uk Kim sub $acc00,$acc00,$acc04 # r1-(r1>>7) 7541f13597dSJung-uk Kim sub $acc01,$acc01,$acc05 7551f13597dSJung-uk Kim sub $acc02,$acc02,$acc06 7561f13597dSJung-uk Kim sub $acc03,$acc03,$acc07 7571f13597dSJung-uk Kim add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 7581f13597dSJung-uk Kim add $acc09,$acc09,$acc09 7591f13597dSJung-uk Kim add $acc10,$acc10,$acc10 7601f13597dSJung-uk Kim add $acc11,$acc11,$acc11 7611f13597dSJung-uk Kim and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 7621f13597dSJung-uk Kim and $acc01,$acc01,$mask1b 7631f13597dSJung-uk Kim and $acc02,$acc02,$mask1b 7641f13597dSJung-uk Kim and $acc03,$acc03,$mask1b 7651f13597dSJung-uk Kim xor $acc00,$acc00,$acc08 # r2 7661f13597dSJung-uk Kim xor $acc01,$acc01,$acc09 7671f13597dSJung-uk Kim rotlwi $acc12,$s0,16 # ROTATE(r0,16) 7681f13597dSJung-uk Kim xor $acc02,$acc02,$acc10 7691f13597dSJung-uk Kim rotlwi $acc13,$s1,16 7701f13597dSJung-uk Kim xor $acc03,$acc03,$acc11 7711f13597dSJung-uk Kim rotlwi $acc14,$s2,16 7721f13597dSJung-uk Kim 7731f13597dSJung-uk Kim xor $s0,$s0,$acc00 # r0^r2 7741f13597dSJung-uk Kim rotlwi $acc15,$s3,16 7751f13597dSJung-uk Kim xor $s1,$s1,$acc01 7761f13597dSJung-uk Kim rotrwi $s0,$s0,24 # ROTATE(r2^r0,24) 7771f13597dSJung-uk Kim xor $s2,$s2,$acc02 7781f13597dSJung-uk Kim rotrwi $s1,$s1,24 7791f13597dSJung-uk Kim xor $s3,$s3,$acc03 7801f13597dSJung-uk Kim rotrwi $s2,$s2,24 7811f13597dSJung-uk Kim xor $s0,$s0,$acc00 # ROTATE(r2^r0,24)^r2 7821f13597dSJung-uk Kim rotrwi $s3,$s3,24 7831f13597dSJung-uk Kim xor $s1,$s1,$acc01 7841f13597dSJung-uk Kim xor $s2,$s2,$acc02 7851f13597dSJung-uk Kim xor $s3,$s3,$acc03 7861f13597dSJung-uk Kim rotlwi $acc08,$acc12,8 # ROTATE(r0,24) 7871f13597dSJung-uk Kim xor $s0,$s0,$acc12 # 7881f13597dSJung-uk Kim rotlwi $acc09,$acc13,8 7891f13597dSJung-uk Kim xor $s1,$s1,$acc13 7901f13597dSJung-uk Kim rotlwi $acc10,$acc14,8 7911f13597dSJung-uk Kim xor $s2,$s2,$acc14 7921f13597dSJung-uk Kim rotlwi $acc11,$acc15,8 7931f13597dSJung-uk Kim xor $s3,$s3,$acc15 7941f13597dSJung-uk Kim xor $s0,$s0,$acc08 # 7951f13597dSJung-uk Kim xor $s1,$s1,$acc09 7961f13597dSJung-uk Kim xor $s2,$s2,$acc10 7971f13597dSJung-uk Kim xor $s3,$s3,$acc11 7981f13597dSJung-uk Kim 7991f13597dSJung-uk Kim b Lenc_compact_loop 8001f13597dSJung-uk Kim.align 4 8011f13597dSJung-uk KimLenc_compact_done: 8021f13597dSJung-uk Kim xor $s0,$s0,$t0 8031f13597dSJung-uk Kim xor $s1,$s1,$t1 8041f13597dSJung-uk Kim xor $s2,$s2,$t2 8051f13597dSJung-uk Kim xor $s3,$s3,$t3 8061f13597dSJung-uk Kim blr 8071f13597dSJung-uk Kim .long 0 8081f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 8097bded2dbSJung-uk Kim.size .AES_encrypt,.-.AES_encrypt 8101f13597dSJung-uk Kim 8111f13597dSJung-uk Kim.globl .AES_decrypt 8121f13597dSJung-uk Kim.align 7 8131f13597dSJung-uk Kim.AES_decrypt: 8141f13597dSJung-uk Kim $STU $sp,-$FRAME($sp) 8151f13597dSJung-uk Kim mflr r0 8161f13597dSJung-uk Kim 8177bded2dbSJung-uk Kim $PUSH $out,`$FRAME-$SIZE_T*19`($sp) 8181f13597dSJung-uk Kim $PUSH r14,`$FRAME-$SIZE_T*18`($sp) 8191f13597dSJung-uk Kim $PUSH r15,`$FRAME-$SIZE_T*17`($sp) 8201f13597dSJung-uk Kim $PUSH r16,`$FRAME-$SIZE_T*16`($sp) 8211f13597dSJung-uk Kim $PUSH r17,`$FRAME-$SIZE_T*15`($sp) 8221f13597dSJung-uk Kim $PUSH r18,`$FRAME-$SIZE_T*14`($sp) 8231f13597dSJung-uk Kim $PUSH r19,`$FRAME-$SIZE_T*13`($sp) 8241f13597dSJung-uk Kim $PUSH r20,`$FRAME-$SIZE_T*12`($sp) 8251f13597dSJung-uk Kim $PUSH r21,`$FRAME-$SIZE_T*11`($sp) 8261f13597dSJung-uk Kim $PUSH r22,`$FRAME-$SIZE_T*10`($sp) 8271f13597dSJung-uk Kim $PUSH r23,`$FRAME-$SIZE_T*9`($sp) 8281f13597dSJung-uk Kim $PUSH r24,`$FRAME-$SIZE_T*8`($sp) 8291f13597dSJung-uk Kim $PUSH r25,`$FRAME-$SIZE_T*7`($sp) 8301f13597dSJung-uk Kim $PUSH r26,`$FRAME-$SIZE_T*6`($sp) 8311f13597dSJung-uk Kim $PUSH r27,`$FRAME-$SIZE_T*5`($sp) 8321f13597dSJung-uk Kim $PUSH r28,`$FRAME-$SIZE_T*4`($sp) 8331f13597dSJung-uk Kim $PUSH r29,`$FRAME-$SIZE_T*3`($sp) 8341f13597dSJung-uk Kim $PUSH r30,`$FRAME-$SIZE_T*2`($sp) 8351f13597dSJung-uk Kim $PUSH r31,`$FRAME-$SIZE_T*1`($sp) 8361f13597dSJung-uk Kim $PUSH r0,`$FRAME+$LRSAVE`($sp) 8371f13597dSJung-uk Kim 8381f13597dSJung-uk Kim andi. $t0,$inp,3 8391f13597dSJung-uk Kim andi. $t1,$out,3 8401f13597dSJung-uk Kim or. $t0,$t0,$t1 8411f13597dSJung-uk Kim bne Ldec_unaligned 8421f13597dSJung-uk Kim 8431f13597dSJung-uk KimLdec_unaligned_ok: 8447bded2dbSJung-uk Kim___ 8457bded2dbSJung-uk Kim$code.=<<___ if (!$LITTLE_ENDIAN); 8461f13597dSJung-uk Kim lwz $s0,0($inp) 8471f13597dSJung-uk Kim lwz $s1,4($inp) 8481f13597dSJung-uk Kim lwz $s2,8($inp) 8491f13597dSJung-uk Kim lwz $s3,12($inp) 8507bded2dbSJung-uk Kim___ 8517bded2dbSJung-uk Kim$code.=<<___ if ($LITTLE_ENDIAN); 8527bded2dbSJung-uk Kim lwz $t0,0($inp) 8537bded2dbSJung-uk Kim lwz $t1,4($inp) 8547bded2dbSJung-uk Kim lwz $t2,8($inp) 8557bded2dbSJung-uk Kim lwz $t3,12($inp) 8567bded2dbSJung-uk Kim rotlwi $s0,$t0,8 8577bded2dbSJung-uk Kim rotlwi $s1,$t1,8 8587bded2dbSJung-uk Kim rotlwi $s2,$t2,8 8597bded2dbSJung-uk Kim rotlwi $s3,$t3,8 8607bded2dbSJung-uk Kim rlwimi $s0,$t0,24,0,7 8617bded2dbSJung-uk Kim rlwimi $s1,$t1,24,0,7 8627bded2dbSJung-uk Kim rlwimi $s2,$t2,24,0,7 8637bded2dbSJung-uk Kim rlwimi $s3,$t3,24,0,7 8647bded2dbSJung-uk Kim rlwimi $s0,$t0,24,16,23 8657bded2dbSJung-uk Kim rlwimi $s1,$t1,24,16,23 8667bded2dbSJung-uk Kim rlwimi $s2,$t2,24,16,23 8677bded2dbSJung-uk Kim rlwimi $s3,$t3,24,16,23 8687bded2dbSJung-uk Kim___ 8697bded2dbSJung-uk Kim$code.=<<___; 8701f13597dSJung-uk Kim bl LAES_Td 8711f13597dSJung-uk Kim bl Lppc_AES_decrypt_compact 8727bded2dbSJung-uk Kim $POP $out,`$FRAME-$SIZE_T*19`($sp) 8737bded2dbSJung-uk Kim___ 8747bded2dbSJung-uk Kim$code.=<<___ if ($LITTLE_ENDIAN); 8757bded2dbSJung-uk Kim rotlwi $t0,$s0,8 8767bded2dbSJung-uk Kim rotlwi $t1,$s1,8 8777bded2dbSJung-uk Kim rotlwi $t2,$s2,8 8787bded2dbSJung-uk Kim rotlwi $t3,$s3,8 8797bded2dbSJung-uk Kim rlwimi $t0,$s0,24,0,7 8807bded2dbSJung-uk Kim rlwimi $t1,$s1,24,0,7 8817bded2dbSJung-uk Kim rlwimi $t2,$s2,24,0,7 8827bded2dbSJung-uk Kim rlwimi $t3,$s3,24,0,7 8837bded2dbSJung-uk Kim rlwimi $t0,$s0,24,16,23 8847bded2dbSJung-uk Kim rlwimi $t1,$s1,24,16,23 8857bded2dbSJung-uk Kim rlwimi $t2,$s2,24,16,23 8867bded2dbSJung-uk Kim rlwimi $t3,$s3,24,16,23 8877bded2dbSJung-uk Kim stw $t0,0($out) 8887bded2dbSJung-uk Kim stw $t1,4($out) 8897bded2dbSJung-uk Kim stw $t2,8($out) 8907bded2dbSJung-uk Kim stw $t3,12($out) 8917bded2dbSJung-uk Kim___ 8927bded2dbSJung-uk Kim$code.=<<___ if (!$LITTLE_ENDIAN); 8931f13597dSJung-uk Kim stw $s0,0($out) 8941f13597dSJung-uk Kim stw $s1,4($out) 8951f13597dSJung-uk Kim stw $s2,8($out) 8961f13597dSJung-uk Kim stw $s3,12($out) 8977bded2dbSJung-uk Kim___ 8987bded2dbSJung-uk Kim$code.=<<___; 8991f13597dSJung-uk Kim b Ldec_done 9001f13597dSJung-uk Kim 9011f13597dSJung-uk KimLdec_unaligned: 9021f13597dSJung-uk Kim subfic $t0,$inp,4096 9031f13597dSJung-uk Kim subfic $t1,$out,4096 9041f13597dSJung-uk Kim andi. $t0,$t0,4096-16 9051f13597dSJung-uk Kim beq Ldec_xpage 9061f13597dSJung-uk Kim andi. $t1,$t1,4096-16 9071f13597dSJung-uk Kim bne Ldec_unaligned_ok 9081f13597dSJung-uk Kim 9091f13597dSJung-uk KimLdec_xpage: 9101f13597dSJung-uk Kim lbz $acc00,0($inp) 9111f13597dSJung-uk Kim lbz $acc01,1($inp) 9121f13597dSJung-uk Kim lbz $acc02,2($inp) 9131f13597dSJung-uk Kim lbz $s0,3($inp) 9141f13597dSJung-uk Kim lbz $acc04,4($inp) 9151f13597dSJung-uk Kim lbz $acc05,5($inp) 9161f13597dSJung-uk Kim lbz $acc06,6($inp) 9171f13597dSJung-uk Kim lbz $s1,7($inp) 9181f13597dSJung-uk Kim lbz $acc08,8($inp) 9191f13597dSJung-uk Kim lbz $acc09,9($inp) 9201f13597dSJung-uk Kim lbz $acc10,10($inp) 9211f13597dSJung-uk Kim insrwi $s0,$acc00,8,0 9221f13597dSJung-uk Kim lbz $s2,11($inp) 9231f13597dSJung-uk Kim insrwi $s1,$acc04,8,0 9241f13597dSJung-uk Kim lbz $acc12,12($inp) 9251f13597dSJung-uk Kim insrwi $s0,$acc01,8,8 9261f13597dSJung-uk Kim lbz $acc13,13($inp) 9271f13597dSJung-uk Kim insrwi $s1,$acc05,8,8 9281f13597dSJung-uk Kim lbz $acc14,14($inp) 9291f13597dSJung-uk Kim insrwi $s0,$acc02,8,16 9301f13597dSJung-uk Kim lbz $s3,15($inp) 9311f13597dSJung-uk Kim insrwi $s1,$acc06,8,16 9321f13597dSJung-uk Kim insrwi $s2,$acc08,8,0 9331f13597dSJung-uk Kim insrwi $s3,$acc12,8,0 9341f13597dSJung-uk Kim insrwi $s2,$acc09,8,8 9351f13597dSJung-uk Kim insrwi $s3,$acc13,8,8 9361f13597dSJung-uk Kim insrwi $s2,$acc10,8,16 9371f13597dSJung-uk Kim insrwi $s3,$acc14,8,16 9381f13597dSJung-uk Kim 9391f13597dSJung-uk Kim bl LAES_Td 9401f13597dSJung-uk Kim bl Lppc_AES_decrypt_compact 9417bded2dbSJung-uk Kim $POP $out,`$FRAME-$SIZE_T*19`($sp) 9421f13597dSJung-uk Kim 9431f13597dSJung-uk Kim extrwi $acc00,$s0,8,0 9441f13597dSJung-uk Kim extrwi $acc01,$s0,8,8 9451f13597dSJung-uk Kim stb $acc00,0($out) 9461f13597dSJung-uk Kim extrwi $acc02,$s0,8,16 9471f13597dSJung-uk Kim stb $acc01,1($out) 9481f13597dSJung-uk Kim stb $acc02,2($out) 9491f13597dSJung-uk Kim extrwi $acc04,$s1,8,0 9501f13597dSJung-uk Kim stb $s0,3($out) 9511f13597dSJung-uk Kim extrwi $acc05,$s1,8,8 9521f13597dSJung-uk Kim stb $acc04,4($out) 9531f13597dSJung-uk Kim extrwi $acc06,$s1,8,16 9541f13597dSJung-uk Kim stb $acc05,5($out) 9551f13597dSJung-uk Kim stb $acc06,6($out) 9561f13597dSJung-uk Kim extrwi $acc08,$s2,8,0 9571f13597dSJung-uk Kim stb $s1,7($out) 9581f13597dSJung-uk Kim extrwi $acc09,$s2,8,8 9591f13597dSJung-uk Kim stb $acc08,8($out) 9601f13597dSJung-uk Kim extrwi $acc10,$s2,8,16 9611f13597dSJung-uk Kim stb $acc09,9($out) 9621f13597dSJung-uk Kim stb $acc10,10($out) 9631f13597dSJung-uk Kim extrwi $acc12,$s3,8,0 9641f13597dSJung-uk Kim stb $s2,11($out) 9651f13597dSJung-uk Kim extrwi $acc13,$s3,8,8 9661f13597dSJung-uk Kim stb $acc12,12($out) 9671f13597dSJung-uk Kim extrwi $acc14,$s3,8,16 9681f13597dSJung-uk Kim stb $acc13,13($out) 9691f13597dSJung-uk Kim stb $acc14,14($out) 9701f13597dSJung-uk Kim stb $s3,15($out) 9711f13597dSJung-uk Kim 9721f13597dSJung-uk KimLdec_done: 9731f13597dSJung-uk Kim $POP r0,`$FRAME+$LRSAVE`($sp) 9741f13597dSJung-uk Kim $POP r14,`$FRAME-$SIZE_T*18`($sp) 9751f13597dSJung-uk Kim $POP r15,`$FRAME-$SIZE_T*17`($sp) 9761f13597dSJung-uk Kim $POP r16,`$FRAME-$SIZE_T*16`($sp) 9771f13597dSJung-uk Kim $POP r17,`$FRAME-$SIZE_T*15`($sp) 9781f13597dSJung-uk Kim $POP r18,`$FRAME-$SIZE_T*14`($sp) 9791f13597dSJung-uk Kim $POP r19,`$FRAME-$SIZE_T*13`($sp) 9801f13597dSJung-uk Kim $POP r20,`$FRAME-$SIZE_T*12`($sp) 9811f13597dSJung-uk Kim $POP r21,`$FRAME-$SIZE_T*11`($sp) 9821f13597dSJung-uk Kim $POP r22,`$FRAME-$SIZE_T*10`($sp) 9831f13597dSJung-uk Kim $POP r23,`$FRAME-$SIZE_T*9`($sp) 9841f13597dSJung-uk Kim $POP r24,`$FRAME-$SIZE_T*8`($sp) 9851f13597dSJung-uk Kim $POP r25,`$FRAME-$SIZE_T*7`($sp) 9861f13597dSJung-uk Kim $POP r26,`$FRAME-$SIZE_T*6`($sp) 9871f13597dSJung-uk Kim $POP r27,`$FRAME-$SIZE_T*5`($sp) 9881f13597dSJung-uk Kim $POP r28,`$FRAME-$SIZE_T*4`($sp) 9891f13597dSJung-uk Kim $POP r29,`$FRAME-$SIZE_T*3`($sp) 9901f13597dSJung-uk Kim $POP r30,`$FRAME-$SIZE_T*2`($sp) 9911f13597dSJung-uk Kim $POP r31,`$FRAME-$SIZE_T*1`($sp) 9921f13597dSJung-uk Kim mtlr r0 9931f13597dSJung-uk Kim addi $sp,$sp,$FRAME 9941f13597dSJung-uk Kim blr 9951f13597dSJung-uk Kim .long 0 9961f13597dSJung-uk Kim .byte 0,12,4,1,0x80,18,3,0 9971f13597dSJung-uk Kim .long 0 9981f13597dSJung-uk Kim 9991f13597dSJung-uk Kim.align 5 10001f13597dSJung-uk KimLppc_AES_decrypt: 10011f13597dSJung-uk Kim lwz $acc00,240($key) 10021f13597dSJung-uk Kim addi $Tbl1,$Tbl0,3 10031f13597dSJung-uk Kim lwz $t0,0($key) 10041f13597dSJung-uk Kim addi $Tbl2,$Tbl0,2 10051f13597dSJung-uk Kim lwz $t1,4($key) 10061f13597dSJung-uk Kim addi $Tbl3,$Tbl0,1 10071f13597dSJung-uk Kim lwz $t2,8($key) 10081f13597dSJung-uk Kim addi $acc00,$acc00,-1 10091f13597dSJung-uk Kim lwz $t3,12($key) 10101f13597dSJung-uk Kim addi $key,$key,16 10111f13597dSJung-uk Kim xor $s0,$s0,$t0 10121f13597dSJung-uk Kim xor $s1,$s1,$t1 10131f13597dSJung-uk Kim xor $s2,$s2,$t2 10141f13597dSJung-uk Kim xor $s3,$s3,$t3 10151f13597dSJung-uk Kim mtctr $acc00 10161f13597dSJung-uk Kim.align 4 10171f13597dSJung-uk KimLdec_loop: 10181f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24+3`,21,28 10191f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24+3`,21,28 10201f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24+3`,21,28 10211f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24+3`,21,28 10221f13597dSJung-uk Kim lwz $t0,0($key) 10231f13597dSJung-uk Kim rlwinm $acc04,$s3,`32-16+3`,21,28 10241f13597dSJung-uk Kim lwz $t1,4($key) 10251f13597dSJung-uk Kim rlwinm $acc05,$s0,`32-16+3`,21,28 10261f13597dSJung-uk Kim lwz $t2,8($key) 10271f13597dSJung-uk Kim rlwinm $acc06,$s1,`32-16+3`,21,28 10281f13597dSJung-uk Kim lwz $t3,12($key) 10291f13597dSJung-uk Kim rlwinm $acc07,$s2,`32-16+3`,21,28 10301f13597dSJung-uk Kim lwzx $acc00,$Tbl0,$acc00 10311f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8+3`,21,28 10321f13597dSJung-uk Kim lwzx $acc01,$Tbl0,$acc01 10331f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8+3`,21,28 10341f13597dSJung-uk Kim lwzx $acc02,$Tbl0,$acc02 10351f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8+3`,21,28 10361f13597dSJung-uk Kim lwzx $acc03,$Tbl0,$acc03 10371f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8+3`,21,28 10381f13597dSJung-uk Kim lwzx $acc04,$Tbl1,$acc04 10391f13597dSJung-uk Kim rlwinm $acc12,$s1,`0+3`,21,28 10401f13597dSJung-uk Kim lwzx $acc05,$Tbl1,$acc05 10411f13597dSJung-uk Kim rlwinm $acc13,$s2,`0+3`,21,28 10421f13597dSJung-uk Kim lwzx $acc06,$Tbl1,$acc06 10431f13597dSJung-uk Kim rlwinm $acc14,$s3,`0+3`,21,28 10441f13597dSJung-uk Kim lwzx $acc07,$Tbl1,$acc07 10451f13597dSJung-uk Kim rlwinm $acc15,$s0,`0+3`,21,28 10461f13597dSJung-uk Kim lwzx $acc08,$Tbl2,$acc08 10471f13597dSJung-uk Kim xor $t0,$t0,$acc00 10481f13597dSJung-uk Kim lwzx $acc09,$Tbl2,$acc09 10491f13597dSJung-uk Kim xor $t1,$t1,$acc01 10501f13597dSJung-uk Kim lwzx $acc10,$Tbl2,$acc10 10511f13597dSJung-uk Kim xor $t2,$t2,$acc02 10521f13597dSJung-uk Kim lwzx $acc11,$Tbl2,$acc11 10531f13597dSJung-uk Kim xor $t3,$t3,$acc03 10541f13597dSJung-uk Kim lwzx $acc12,$Tbl3,$acc12 10551f13597dSJung-uk Kim xor $t0,$t0,$acc04 10561f13597dSJung-uk Kim lwzx $acc13,$Tbl3,$acc13 10571f13597dSJung-uk Kim xor $t1,$t1,$acc05 10581f13597dSJung-uk Kim lwzx $acc14,$Tbl3,$acc14 10591f13597dSJung-uk Kim xor $t2,$t2,$acc06 10601f13597dSJung-uk Kim lwzx $acc15,$Tbl3,$acc15 10611f13597dSJung-uk Kim xor $t3,$t3,$acc07 10621f13597dSJung-uk Kim xor $t0,$t0,$acc08 10631f13597dSJung-uk Kim xor $t1,$t1,$acc09 10641f13597dSJung-uk Kim xor $t2,$t2,$acc10 10651f13597dSJung-uk Kim xor $t3,$t3,$acc11 10661f13597dSJung-uk Kim xor $s0,$t0,$acc12 10671f13597dSJung-uk Kim xor $s1,$t1,$acc13 10681f13597dSJung-uk Kim xor $s2,$t2,$acc14 10691f13597dSJung-uk Kim xor $s3,$t3,$acc15 10701f13597dSJung-uk Kim addi $key,$key,16 1071*b8721c16SJung-uk Kim bdnz Ldec_loop 10721f13597dSJung-uk Kim 10731f13597dSJung-uk Kim addi $Tbl2,$Tbl0,2048 10741f13597dSJung-uk Kim nop 10751f13597dSJung-uk Kim lwz $t0,0($key) 10761f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24`,24,31 10771f13597dSJung-uk Kim lwz $t1,4($key) 10781f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24`,24,31 10791f13597dSJung-uk Kim lwz $t2,8($key) 10801f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24`,24,31 10811f13597dSJung-uk Kim lwz $t3,12($key) 10821f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24`,24,31 10831f13597dSJung-uk Kim lwz $acc08,`2048+0`($Tbl0) ! prefetch Td4 10841f13597dSJung-uk Kim rlwinm $acc04,$s3,`32-16`,24,31 10851f13597dSJung-uk Kim lwz $acc09,`2048+32`($Tbl0) 10861f13597dSJung-uk Kim rlwinm $acc05,$s0,`32-16`,24,31 10871f13597dSJung-uk Kim lwz $acc10,`2048+64`($Tbl0) 10881f13597dSJung-uk Kim lbzx $acc00,$Tbl2,$acc00 10891f13597dSJung-uk Kim lwz $acc11,`2048+96`($Tbl0) 10901f13597dSJung-uk Kim lbzx $acc01,$Tbl2,$acc01 10911f13597dSJung-uk Kim lwz $acc12,`2048+128`($Tbl0) 10921f13597dSJung-uk Kim rlwinm $acc06,$s1,`32-16`,24,31 10931f13597dSJung-uk Kim lwz $acc13,`2048+160`($Tbl0) 10941f13597dSJung-uk Kim rlwinm $acc07,$s2,`32-16`,24,31 10951f13597dSJung-uk Kim lwz $acc14,`2048+192`($Tbl0) 10961f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8`,24,31 10971f13597dSJung-uk Kim lwz $acc15,`2048+224`($Tbl0) 10981f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8`,24,31 10991f13597dSJung-uk Kim lbzx $acc02,$Tbl2,$acc02 11001f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8`,24,31 11011f13597dSJung-uk Kim lbzx $acc03,$Tbl2,$acc03 11021f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8`,24,31 11031f13597dSJung-uk Kim lbzx $acc04,$Tbl2,$acc04 11041f13597dSJung-uk Kim rlwinm $acc12,$s1,`0`,24,31 11051f13597dSJung-uk Kim lbzx $acc05,$Tbl2,$acc05 11061f13597dSJung-uk Kim rlwinm $acc13,$s2,`0`,24,31 11071f13597dSJung-uk Kim lbzx $acc06,$Tbl2,$acc06 11081f13597dSJung-uk Kim rlwinm $acc14,$s3,`0`,24,31 11091f13597dSJung-uk Kim lbzx $acc07,$Tbl2,$acc07 11101f13597dSJung-uk Kim rlwinm $acc15,$s0,`0`,24,31 11111f13597dSJung-uk Kim lbzx $acc08,$Tbl2,$acc08 11121f13597dSJung-uk Kim rlwinm $s0,$acc00,24,0,7 11131f13597dSJung-uk Kim lbzx $acc09,$Tbl2,$acc09 11141f13597dSJung-uk Kim rlwinm $s1,$acc01,24,0,7 11151f13597dSJung-uk Kim lbzx $acc10,$Tbl2,$acc10 11161f13597dSJung-uk Kim rlwinm $s2,$acc02,24,0,7 11171f13597dSJung-uk Kim lbzx $acc11,$Tbl2,$acc11 11181f13597dSJung-uk Kim rlwinm $s3,$acc03,24,0,7 11191f13597dSJung-uk Kim lbzx $acc12,$Tbl2,$acc12 11201f13597dSJung-uk Kim rlwimi $s0,$acc04,16,8,15 11211f13597dSJung-uk Kim lbzx $acc13,$Tbl2,$acc13 11221f13597dSJung-uk Kim rlwimi $s1,$acc05,16,8,15 11231f13597dSJung-uk Kim lbzx $acc14,$Tbl2,$acc14 11241f13597dSJung-uk Kim rlwimi $s2,$acc06,16,8,15 11251f13597dSJung-uk Kim lbzx $acc15,$Tbl2,$acc15 11261f13597dSJung-uk Kim rlwimi $s3,$acc07,16,8,15 11271f13597dSJung-uk Kim rlwimi $s0,$acc08,8,16,23 11281f13597dSJung-uk Kim rlwimi $s1,$acc09,8,16,23 11291f13597dSJung-uk Kim rlwimi $s2,$acc10,8,16,23 11301f13597dSJung-uk Kim rlwimi $s3,$acc11,8,16,23 11311f13597dSJung-uk Kim or $s0,$s0,$acc12 11321f13597dSJung-uk Kim or $s1,$s1,$acc13 11331f13597dSJung-uk Kim or $s2,$s2,$acc14 11341f13597dSJung-uk Kim or $s3,$s3,$acc15 11351f13597dSJung-uk Kim xor $s0,$s0,$t0 11361f13597dSJung-uk Kim xor $s1,$s1,$t1 11371f13597dSJung-uk Kim xor $s2,$s2,$t2 11381f13597dSJung-uk Kim xor $s3,$s3,$t3 11391f13597dSJung-uk Kim blr 11401f13597dSJung-uk Kim .long 0 11411f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 11421f13597dSJung-uk Kim 11431f13597dSJung-uk Kim.align 4 11441f13597dSJung-uk KimLppc_AES_decrypt_compact: 11451f13597dSJung-uk Kim lwz $acc00,240($key) 11461f13597dSJung-uk Kim addi $Tbl1,$Tbl0,2048 11471f13597dSJung-uk Kim lwz $t0,0($key) 11481f13597dSJung-uk Kim lis $mask80,0x8080 11491f13597dSJung-uk Kim lwz $t1,4($key) 11501f13597dSJung-uk Kim lis $mask1b,0x1b1b 11511f13597dSJung-uk Kim lwz $t2,8($key) 11521f13597dSJung-uk Kim ori $mask80,$mask80,0x8080 11531f13597dSJung-uk Kim lwz $t3,12($key) 11541f13597dSJung-uk Kim ori $mask1b,$mask1b,0x1b1b 11551f13597dSJung-uk Kim addi $key,$key,16 11561f13597dSJung-uk Kim___ 11571f13597dSJung-uk Kim$code.=<<___ if ($SIZE_T==8); 11581f13597dSJung-uk Kim insrdi $mask80,$mask80,32,0 11591f13597dSJung-uk Kim insrdi $mask1b,$mask1b,32,0 11601f13597dSJung-uk Kim___ 11611f13597dSJung-uk Kim$code.=<<___; 11621f13597dSJung-uk Kim mtctr $acc00 11631f13597dSJung-uk Kim.align 4 11641f13597dSJung-uk KimLdec_compact_loop: 11651f13597dSJung-uk Kim xor $s0,$s0,$t0 11661f13597dSJung-uk Kim xor $s1,$s1,$t1 11671f13597dSJung-uk Kim rlwinm $acc00,$s0,`32-24`,24,31 11681f13597dSJung-uk Kim xor $s2,$s2,$t2 11691f13597dSJung-uk Kim rlwinm $acc01,$s1,`32-24`,24,31 11701f13597dSJung-uk Kim xor $s3,$s3,$t3 11711f13597dSJung-uk Kim rlwinm $acc02,$s2,`32-24`,24,31 11721f13597dSJung-uk Kim rlwinm $acc03,$s3,`32-24`,24,31 11731f13597dSJung-uk Kim rlwinm $acc04,$s3,`32-16`,24,31 11741f13597dSJung-uk Kim rlwinm $acc05,$s0,`32-16`,24,31 11751f13597dSJung-uk Kim rlwinm $acc06,$s1,`32-16`,24,31 11761f13597dSJung-uk Kim rlwinm $acc07,$s2,`32-16`,24,31 11771f13597dSJung-uk Kim lbzx $acc00,$Tbl1,$acc00 11781f13597dSJung-uk Kim rlwinm $acc08,$s2,`32-8`,24,31 11791f13597dSJung-uk Kim lbzx $acc01,$Tbl1,$acc01 11801f13597dSJung-uk Kim rlwinm $acc09,$s3,`32-8`,24,31 11811f13597dSJung-uk Kim lbzx $acc02,$Tbl1,$acc02 11821f13597dSJung-uk Kim rlwinm $acc10,$s0,`32-8`,24,31 11831f13597dSJung-uk Kim lbzx $acc03,$Tbl1,$acc03 11841f13597dSJung-uk Kim rlwinm $acc11,$s1,`32-8`,24,31 11851f13597dSJung-uk Kim lbzx $acc04,$Tbl1,$acc04 11861f13597dSJung-uk Kim rlwinm $acc12,$s1,`0`,24,31 11871f13597dSJung-uk Kim lbzx $acc05,$Tbl1,$acc05 11881f13597dSJung-uk Kim rlwinm $acc13,$s2,`0`,24,31 11891f13597dSJung-uk Kim lbzx $acc06,$Tbl1,$acc06 11901f13597dSJung-uk Kim rlwinm $acc14,$s3,`0`,24,31 11911f13597dSJung-uk Kim lbzx $acc07,$Tbl1,$acc07 11921f13597dSJung-uk Kim rlwinm $acc15,$s0,`0`,24,31 11931f13597dSJung-uk Kim lbzx $acc08,$Tbl1,$acc08 11941f13597dSJung-uk Kim rlwinm $s0,$acc00,24,0,7 11951f13597dSJung-uk Kim lbzx $acc09,$Tbl1,$acc09 11961f13597dSJung-uk Kim rlwinm $s1,$acc01,24,0,7 11971f13597dSJung-uk Kim lbzx $acc10,$Tbl1,$acc10 11981f13597dSJung-uk Kim rlwinm $s2,$acc02,24,0,7 11991f13597dSJung-uk Kim lbzx $acc11,$Tbl1,$acc11 12001f13597dSJung-uk Kim rlwinm $s3,$acc03,24,0,7 12011f13597dSJung-uk Kim lbzx $acc12,$Tbl1,$acc12 12021f13597dSJung-uk Kim rlwimi $s0,$acc04,16,8,15 12031f13597dSJung-uk Kim lbzx $acc13,$Tbl1,$acc13 12041f13597dSJung-uk Kim rlwimi $s1,$acc05,16,8,15 12051f13597dSJung-uk Kim lbzx $acc14,$Tbl1,$acc14 12061f13597dSJung-uk Kim rlwimi $s2,$acc06,16,8,15 12071f13597dSJung-uk Kim lbzx $acc15,$Tbl1,$acc15 12081f13597dSJung-uk Kim rlwimi $s3,$acc07,16,8,15 12091f13597dSJung-uk Kim rlwimi $s0,$acc08,8,16,23 12101f13597dSJung-uk Kim rlwimi $s1,$acc09,8,16,23 12111f13597dSJung-uk Kim rlwimi $s2,$acc10,8,16,23 12121f13597dSJung-uk Kim rlwimi $s3,$acc11,8,16,23 12131f13597dSJung-uk Kim lwz $t0,0($key) 12141f13597dSJung-uk Kim or $s0,$s0,$acc12 12151f13597dSJung-uk Kim lwz $t1,4($key) 12161f13597dSJung-uk Kim or $s1,$s1,$acc13 12171f13597dSJung-uk Kim lwz $t2,8($key) 12181f13597dSJung-uk Kim or $s2,$s2,$acc14 12191f13597dSJung-uk Kim lwz $t3,12($key) 12201f13597dSJung-uk Kim or $s3,$s3,$acc15 12211f13597dSJung-uk Kim 12221f13597dSJung-uk Kim addi $key,$key,16 12231f13597dSJung-uk Kim bdz Ldec_compact_done 12241f13597dSJung-uk Kim___ 12251f13597dSJung-uk Kim$code.=<<___ if ($SIZE_T==8); 12261f13597dSJung-uk Kim # vectorized permutation improves decrypt performance by 10% 12271f13597dSJung-uk Kim insrdi $s0,$s1,32,0 12281f13597dSJung-uk Kim insrdi $s2,$s3,32,0 12291f13597dSJung-uk Kim 12301f13597dSJung-uk Kim and $acc00,$s0,$mask80 # r1=r0&0x80808080 12311f13597dSJung-uk Kim and $acc02,$s2,$mask80 12321f13597dSJung-uk Kim srdi $acc04,$acc00,7 # r1>>7 12331f13597dSJung-uk Kim srdi $acc06,$acc02,7 12341f13597dSJung-uk Kim andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f 12351f13597dSJung-uk Kim andc $acc10,$s2,$mask80 12361f13597dSJung-uk Kim sub $acc00,$acc00,$acc04 # r1-(r1>>7) 12371f13597dSJung-uk Kim sub $acc02,$acc02,$acc06 12381f13597dSJung-uk Kim add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 12391f13597dSJung-uk Kim add $acc10,$acc10,$acc10 12401f13597dSJung-uk Kim and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 12411f13597dSJung-uk Kim and $acc02,$acc02,$mask1b 12421f13597dSJung-uk Kim xor $acc00,$acc00,$acc08 # r2 12431f13597dSJung-uk Kim xor $acc02,$acc02,$acc10 12441f13597dSJung-uk Kim 12451f13597dSJung-uk Kim and $acc04,$acc00,$mask80 # r1=r2&0x80808080 12461f13597dSJung-uk Kim and $acc06,$acc02,$mask80 12471f13597dSJung-uk Kim srdi $acc08,$acc04,7 # r1>>7 12481f13597dSJung-uk Kim srdi $acc10,$acc06,7 12491f13597dSJung-uk Kim andc $acc12,$acc00,$mask80 # r2&0x7f7f7f7f 12501f13597dSJung-uk Kim andc $acc14,$acc02,$mask80 12511f13597dSJung-uk Kim sub $acc04,$acc04,$acc08 # r1-(r1>>7) 12521f13597dSJung-uk Kim sub $acc06,$acc06,$acc10 12531f13597dSJung-uk Kim add $acc12,$acc12,$acc12 # (r2&0x7f7f7f7f)<<1 12541f13597dSJung-uk Kim add $acc14,$acc14,$acc14 12551f13597dSJung-uk Kim and $acc04,$acc04,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 12561f13597dSJung-uk Kim and $acc06,$acc06,$mask1b 12571f13597dSJung-uk Kim xor $acc04,$acc04,$acc12 # r4 12581f13597dSJung-uk Kim xor $acc06,$acc06,$acc14 12591f13597dSJung-uk Kim 12601f13597dSJung-uk Kim and $acc08,$acc04,$mask80 # r1=r4&0x80808080 12611f13597dSJung-uk Kim and $acc10,$acc06,$mask80 12621f13597dSJung-uk Kim srdi $acc12,$acc08,7 # r1>>7 12631f13597dSJung-uk Kim srdi $acc14,$acc10,7 12641f13597dSJung-uk Kim sub $acc08,$acc08,$acc12 # r1-(r1>>7) 12651f13597dSJung-uk Kim sub $acc10,$acc10,$acc14 12661f13597dSJung-uk Kim andc $acc12,$acc04,$mask80 # r4&0x7f7f7f7f 12671f13597dSJung-uk Kim andc $acc14,$acc06,$mask80 12681f13597dSJung-uk Kim add $acc12,$acc12,$acc12 # (r4&0x7f7f7f7f)<<1 12691f13597dSJung-uk Kim add $acc14,$acc14,$acc14 12701f13597dSJung-uk Kim and $acc08,$acc08,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 12711f13597dSJung-uk Kim and $acc10,$acc10,$mask1b 12721f13597dSJung-uk Kim xor $acc08,$acc08,$acc12 # r8 12731f13597dSJung-uk Kim xor $acc10,$acc10,$acc14 12741f13597dSJung-uk Kim 12751f13597dSJung-uk Kim xor $acc00,$acc00,$s0 # r2^r0 12761f13597dSJung-uk Kim xor $acc02,$acc02,$s2 12771f13597dSJung-uk Kim xor $acc04,$acc04,$s0 # r4^r0 12781f13597dSJung-uk Kim xor $acc06,$acc06,$s2 12791f13597dSJung-uk Kim 12801f13597dSJung-uk Kim extrdi $acc01,$acc00,32,0 12811f13597dSJung-uk Kim extrdi $acc03,$acc02,32,0 12821f13597dSJung-uk Kim extrdi $acc05,$acc04,32,0 12831f13597dSJung-uk Kim extrdi $acc07,$acc06,32,0 12841f13597dSJung-uk Kim extrdi $acc09,$acc08,32,0 12851f13597dSJung-uk Kim extrdi $acc11,$acc10,32,0 12861f13597dSJung-uk Kim___ 12871f13597dSJung-uk Kim$code.=<<___ if ($SIZE_T==4); 12881f13597dSJung-uk Kim and $acc00,$s0,$mask80 # r1=r0&0x80808080 12891f13597dSJung-uk Kim and $acc01,$s1,$mask80 12901f13597dSJung-uk Kim and $acc02,$s2,$mask80 12911f13597dSJung-uk Kim and $acc03,$s3,$mask80 12921f13597dSJung-uk Kim srwi $acc04,$acc00,7 # r1>>7 12931f13597dSJung-uk Kim andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f 12941f13597dSJung-uk Kim srwi $acc05,$acc01,7 12951f13597dSJung-uk Kim andc $acc09,$s1,$mask80 12961f13597dSJung-uk Kim srwi $acc06,$acc02,7 12971f13597dSJung-uk Kim andc $acc10,$s2,$mask80 12981f13597dSJung-uk Kim srwi $acc07,$acc03,7 12991f13597dSJung-uk Kim andc $acc11,$s3,$mask80 13001f13597dSJung-uk Kim sub $acc00,$acc00,$acc04 # r1-(r1>>7) 13011f13597dSJung-uk Kim sub $acc01,$acc01,$acc05 13021f13597dSJung-uk Kim sub $acc02,$acc02,$acc06 13031f13597dSJung-uk Kim sub $acc03,$acc03,$acc07 13041f13597dSJung-uk Kim add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 13051f13597dSJung-uk Kim add $acc09,$acc09,$acc09 13061f13597dSJung-uk Kim add $acc10,$acc10,$acc10 13071f13597dSJung-uk Kim add $acc11,$acc11,$acc11 13081f13597dSJung-uk Kim and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 13091f13597dSJung-uk Kim and $acc01,$acc01,$mask1b 13101f13597dSJung-uk Kim and $acc02,$acc02,$mask1b 13111f13597dSJung-uk Kim and $acc03,$acc03,$mask1b 13121f13597dSJung-uk Kim xor $acc00,$acc00,$acc08 # r2 13131f13597dSJung-uk Kim xor $acc01,$acc01,$acc09 13141f13597dSJung-uk Kim xor $acc02,$acc02,$acc10 13151f13597dSJung-uk Kim xor $acc03,$acc03,$acc11 13161f13597dSJung-uk Kim 13171f13597dSJung-uk Kim and $acc04,$acc00,$mask80 # r1=r2&0x80808080 13181f13597dSJung-uk Kim and $acc05,$acc01,$mask80 13191f13597dSJung-uk Kim and $acc06,$acc02,$mask80 13201f13597dSJung-uk Kim and $acc07,$acc03,$mask80 13211f13597dSJung-uk Kim srwi $acc08,$acc04,7 # r1>>7 13221f13597dSJung-uk Kim andc $acc12,$acc00,$mask80 # r2&0x7f7f7f7f 13231f13597dSJung-uk Kim srwi $acc09,$acc05,7 13241f13597dSJung-uk Kim andc $acc13,$acc01,$mask80 13251f13597dSJung-uk Kim srwi $acc10,$acc06,7 13261f13597dSJung-uk Kim andc $acc14,$acc02,$mask80 13271f13597dSJung-uk Kim srwi $acc11,$acc07,7 13281f13597dSJung-uk Kim andc $acc15,$acc03,$mask80 13291f13597dSJung-uk Kim sub $acc04,$acc04,$acc08 # r1-(r1>>7) 13301f13597dSJung-uk Kim sub $acc05,$acc05,$acc09 13311f13597dSJung-uk Kim sub $acc06,$acc06,$acc10 13321f13597dSJung-uk Kim sub $acc07,$acc07,$acc11 13331f13597dSJung-uk Kim add $acc12,$acc12,$acc12 # (r2&0x7f7f7f7f)<<1 13341f13597dSJung-uk Kim add $acc13,$acc13,$acc13 13351f13597dSJung-uk Kim add $acc14,$acc14,$acc14 13361f13597dSJung-uk Kim add $acc15,$acc15,$acc15 13371f13597dSJung-uk Kim and $acc04,$acc04,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 13381f13597dSJung-uk Kim and $acc05,$acc05,$mask1b 13391f13597dSJung-uk Kim and $acc06,$acc06,$mask1b 13401f13597dSJung-uk Kim and $acc07,$acc07,$mask1b 13411f13597dSJung-uk Kim xor $acc04,$acc04,$acc12 # r4 13421f13597dSJung-uk Kim xor $acc05,$acc05,$acc13 13431f13597dSJung-uk Kim xor $acc06,$acc06,$acc14 13441f13597dSJung-uk Kim xor $acc07,$acc07,$acc15 13451f13597dSJung-uk Kim 13461f13597dSJung-uk Kim and $acc08,$acc04,$mask80 # r1=r4&0x80808080 13471f13597dSJung-uk Kim and $acc09,$acc05,$mask80 13481f13597dSJung-uk Kim srwi $acc12,$acc08,7 # r1>>7 13491f13597dSJung-uk Kim and $acc10,$acc06,$mask80 13501f13597dSJung-uk Kim srwi $acc13,$acc09,7 13511f13597dSJung-uk Kim and $acc11,$acc07,$mask80 13521f13597dSJung-uk Kim srwi $acc14,$acc10,7 13531f13597dSJung-uk Kim sub $acc08,$acc08,$acc12 # r1-(r1>>7) 13541f13597dSJung-uk Kim srwi $acc15,$acc11,7 13551f13597dSJung-uk Kim sub $acc09,$acc09,$acc13 13561f13597dSJung-uk Kim sub $acc10,$acc10,$acc14 13571f13597dSJung-uk Kim sub $acc11,$acc11,$acc15 13581f13597dSJung-uk Kim andc $acc12,$acc04,$mask80 # r4&0x7f7f7f7f 13591f13597dSJung-uk Kim andc $acc13,$acc05,$mask80 13601f13597dSJung-uk Kim andc $acc14,$acc06,$mask80 13611f13597dSJung-uk Kim andc $acc15,$acc07,$mask80 13621f13597dSJung-uk Kim add $acc12,$acc12,$acc12 # (r4&0x7f7f7f7f)<<1 13631f13597dSJung-uk Kim add $acc13,$acc13,$acc13 13641f13597dSJung-uk Kim add $acc14,$acc14,$acc14 13651f13597dSJung-uk Kim add $acc15,$acc15,$acc15 13661f13597dSJung-uk Kim and $acc08,$acc08,$mask1b # (r1-(r1>>7))&0x1b1b1b1b 13671f13597dSJung-uk Kim and $acc09,$acc09,$mask1b 13681f13597dSJung-uk Kim and $acc10,$acc10,$mask1b 13691f13597dSJung-uk Kim and $acc11,$acc11,$mask1b 13701f13597dSJung-uk Kim xor $acc08,$acc08,$acc12 # r8 13711f13597dSJung-uk Kim xor $acc09,$acc09,$acc13 13721f13597dSJung-uk Kim xor $acc10,$acc10,$acc14 13731f13597dSJung-uk Kim xor $acc11,$acc11,$acc15 13741f13597dSJung-uk Kim 13751f13597dSJung-uk Kim xor $acc00,$acc00,$s0 # r2^r0 13761f13597dSJung-uk Kim xor $acc01,$acc01,$s1 13771f13597dSJung-uk Kim xor $acc02,$acc02,$s2 13781f13597dSJung-uk Kim xor $acc03,$acc03,$s3 13791f13597dSJung-uk Kim xor $acc04,$acc04,$s0 # r4^r0 13801f13597dSJung-uk Kim xor $acc05,$acc05,$s1 13811f13597dSJung-uk Kim xor $acc06,$acc06,$s2 13821f13597dSJung-uk Kim xor $acc07,$acc07,$s3 13831f13597dSJung-uk Kim___ 13841f13597dSJung-uk Kim$code.=<<___; 13851f13597dSJung-uk Kim rotrwi $s0,$s0,8 # = ROTATE(r0,8) 13861f13597dSJung-uk Kim rotrwi $s1,$s1,8 13871f13597dSJung-uk Kim xor $s0,$s0,$acc00 # ^= r2^r0 13881f13597dSJung-uk Kim rotrwi $s2,$s2,8 13891f13597dSJung-uk Kim xor $s1,$s1,$acc01 13901f13597dSJung-uk Kim rotrwi $s3,$s3,8 13911f13597dSJung-uk Kim xor $s2,$s2,$acc02 13921f13597dSJung-uk Kim xor $s3,$s3,$acc03 13931f13597dSJung-uk Kim xor $acc00,$acc00,$acc08 13941f13597dSJung-uk Kim xor $acc01,$acc01,$acc09 13951f13597dSJung-uk Kim xor $acc02,$acc02,$acc10 13961f13597dSJung-uk Kim xor $acc03,$acc03,$acc11 13971f13597dSJung-uk Kim xor $s0,$s0,$acc04 # ^= r4^r0 13981f13597dSJung-uk Kim rotrwi $acc00,$acc00,24 13991f13597dSJung-uk Kim xor $s1,$s1,$acc05 14001f13597dSJung-uk Kim rotrwi $acc01,$acc01,24 14011f13597dSJung-uk Kim xor $s2,$s2,$acc06 14021f13597dSJung-uk Kim rotrwi $acc02,$acc02,24 14031f13597dSJung-uk Kim xor $s3,$s3,$acc07 14041f13597dSJung-uk Kim rotrwi $acc03,$acc03,24 14051f13597dSJung-uk Kim xor $acc04,$acc04,$acc08 14061f13597dSJung-uk Kim xor $acc05,$acc05,$acc09 14071f13597dSJung-uk Kim xor $acc06,$acc06,$acc10 14081f13597dSJung-uk Kim xor $acc07,$acc07,$acc11 14091f13597dSJung-uk Kim xor $s0,$s0,$acc08 # ^= r8 [^((r4^r0)^(r2^r0)=r4^r2)] 14101f13597dSJung-uk Kim rotrwi $acc04,$acc04,16 14111f13597dSJung-uk Kim xor $s1,$s1,$acc09 14121f13597dSJung-uk Kim rotrwi $acc05,$acc05,16 14131f13597dSJung-uk Kim xor $s2,$s2,$acc10 14141f13597dSJung-uk Kim rotrwi $acc06,$acc06,16 14151f13597dSJung-uk Kim xor $s3,$s3,$acc11 14161f13597dSJung-uk Kim rotrwi $acc07,$acc07,16 14171f13597dSJung-uk Kim xor $s0,$s0,$acc00 # ^= ROTATE(r8^r2^r0,24) 14181f13597dSJung-uk Kim rotrwi $acc08,$acc08,8 14191f13597dSJung-uk Kim xor $s1,$s1,$acc01 14201f13597dSJung-uk Kim rotrwi $acc09,$acc09,8 14211f13597dSJung-uk Kim xor $s2,$s2,$acc02 14221f13597dSJung-uk Kim rotrwi $acc10,$acc10,8 14231f13597dSJung-uk Kim xor $s3,$s3,$acc03 14241f13597dSJung-uk Kim rotrwi $acc11,$acc11,8 14251f13597dSJung-uk Kim xor $s0,$s0,$acc04 # ^= ROTATE(r8^r4^r0,16) 14261f13597dSJung-uk Kim xor $s1,$s1,$acc05 14271f13597dSJung-uk Kim xor $s2,$s2,$acc06 14281f13597dSJung-uk Kim xor $s3,$s3,$acc07 14291f13597dSJung-uk Kim xor $s0,$s0,$acc08 # ^= ROTATE(r8,8) 14301f13597dSJung-uk Kim xor $s1,$s1,$acc09 14311f13597dSJung-uk Kim xor $s2,$s2,$acc10 14321f13597dSJung-uk Kim xor $s3,$s3,$acc11 14331f13597dSJung-uk Kim 14341f13597dSJung-uk Kim b Ldec_compact_loop 14351f13597dSJung-uk Kim.align 4 14361f13597dSJung-uk KimLdec_compact_done: 14371f13597dSJung-uk Kim xor $s0,$s0,$t0 14381f13597dSJung-uk Kim xor $s1,$s1,$t1 14391f13597dSJung-uk Kim xor $s2,$s2,$t2 14401f13597dSJung-uk Kim xor $s3,$s3,$t3 14411f13597dSJung-uk Kim blr 14421f13597dSJung-uk Kim .long 0 14431f13597dSJung-uk Kim .byte 0,12,0x14,0,0,0,0,0 14447bded2dbSJung-uk Kim.size .AES_decrypt,.-.AES_decrypt 14451f13597dSJung-uk Kim 14461f13597dSJung-uk Kim.asciz "AES for PPC, CRYPTOGAMS by <appro\@openssl.org>" 14471f13597dSJung-uk Kim.align 7 14481f13597dSJung-uk Kim___ 14491f13597dSJung-uk Kim 14501f13597dSJung-uk Kim$code =~ s/\`([^\`]*)\`/eval $1/gem; 14511f13597dSJung-uk Kimprint $code; 14521f13597dSJung-uk Kimclose STDOUT; 1453