granicus.if.org Git - zfs/blob - module/icp/algs/aes/aes_impl.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  23  */
  24
  25 #include <sys/zfs_context.h>
  26 #include <sys/crypto/spi.h>
  27 #include <modes/modes.h>
  28 #include <aes/aes_impl.h>
  29
  30 #ifdef __amd64
  31
  32 #ifdef _KERNEL
  33 /* Workaround for no XMM kernel thread save/restore */
  34 #define KPREEMPT_DISABLE        kpreempt_disable()
  35 #define KPREEMPT_ENABLE         kpreempt_enable()
  36
  37 #else
  38 #define KPREEMPT_DISABLE
  39 #define KPREEMPT_ENABLE
  40 #endif  /* _KERNEL */
  41 #endif  /* __amd64 */
  42
  43
  44 /*
  45  * This file is derived from the file  rijndael-alg-fst.c  taken from the
  46  * "optimized C code v3.0" on the "rijndael home page"
  47  * http://www.iaik.tu-graz.ac.at/research/krypto/AES/old/~rijmen/rijndael/
  48  * pointed by the NIST web-site http://csrc.nist.gov/archive/aes/
  49  *
  50  * The following note is from the original file:
  51  */
  52
  53 /*
  54  * rijndael-alg-fst.c
  55  *
  56  * @version 3.0 (December 2000)
  57  *
  58  * Optimised ANSI C code for the Rijndael cipher (now AES)
  59  *
  60  * @author Vincent Rijmen <vincent.rijmen@esat.kuleuven.ac.be>
  61  * @author Antoon Bosselaers <antoon.bosselaers@esat.kuleuven.ac.be>
  62  * @author Paulo Barreto <paulo.barreto@terra.com.br>
  63  *
  64  * This code is hereby placed in the public domain.
  65  *
  66  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ''AS IS'' AND ANY EXPRESS
  67  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  68  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  69  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE
  70  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  71  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  72  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  73  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  74  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
  75  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
  76  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  77  */
  78
  79 #if defined(__amd64)
  80
  81 /* These functions are used to execute amd64 instructions for AMD or Intel: */
  82 extern int rijndael_key_setup_enc_amd64(uint32_t rk[],
  83         const uint32_t cipherKey[], int keyBits);
  84 extern int rijndael_key_setup_dec_amd64(uint32_t rk[],
  85         const uint32_t cipherKey[], int keyBits);
  86 extern void aes_encrypt_amd64(const uint32_t rk[], int Nr,
  87         const uint32_t pt[4], uint32_t ct[4]);
  88 extern void aes_decrypt_amd64(const uint32_t rk[], int Nr,
  89         const uint32_t ct[4], uint32_t pt[4]);
  90
  91 /* These functions are used to execute Intel-specific AES-NI instructions: */
  92 extern int rijndael_key_setup_enc_intel(uint32_t rk[],
  93         const uint32_t cipherKey[], uint64_t keyBits);
  94 extern int rijndael_key_setup_dec_intel(uint32_t rk[],
  95         const uint32_t cipherKey[], uint64_t keyBits);
  96 extern void aes_encrypt_intel(const uint32_t rk[], int Nr,
  97         const uint32_t pt[4], uint32_t ct[4]);
  98 extern void aes_decrypt_intel(const uint32_t rk[], int Nr,
  99         const uint32_t ct[4], uint32_t pt[4]);
 100
 101 static int intel_aes_instructions_present(void);
 102
 103 #define AES_ENCRYPT_IMPL(a, b, c, d, e) rijndael_encrypt(a, b, c, d, e)
 104 #define AES_DECRYPT_IMPL(a, b, c, d, e) rijndael_decrypt(a, b, c, d, e)
 105
 106 #else /* Generic C implementation */
 107
 108 #define AES_ENCRYPT_IMPL(a, b, c, d, e) rijndael_encrypt(a, b, c, d)
 109 #define AES_DECRYPT_IMPL(a, b, c, d, e) rijndael_decrypt(a, b, c, d)
 110 #define rijndael_key_setup_enc_raw      rijndael_key_setup_enc
 111 #endif  /* __amd64 */
 112
 113 #if defined(_LITTLE_ENDIAN) && !defined(__amd64)
 114 #define AES_BYTE_SWAP
 115 #endif
 116
 117
 118 #if !defined(__amd64)
 119 /*
 120  *  Constant tables
 121  */
 122
 123 /*
 124  * Te0[x] = S [x].[02, 01, 01, 03];
 125  * Te1[x] = S [x].[03, 02, 01, 01];
 126  * Te2[x] = S [x].[01, 03, 02, 01];
 127  * Te3[x] = S [x].[01, 01, 03, 02];
 128  * Te4[x] = S [x].[01, 01, 01, 01];
 129  *
 130  * Td0[x] = Si[x].[0e, 09, 0d, 0b];
 131  * Td1[x] = Si[x].[0b, 0e, 09, 0d];
 132  * Td2[x] = Si[x].[0d, 0b, 0e, 09];
 133  * Td3[x] = Si[x].[09, 0d, 0b, 0e];
 134  * Td4[x] = Si[x].[01, 01, 01, 01];
 135  */
 136
 137 /* Encrypt Sbox constants (for the substitute bytes operation) */
 138
 139 static const uint32_t Te0[256] =
 140 {
 141         0xc66363a5U, 0xf87c7c84U, 0xee777799U, 0xf67b7b8dU,
 142         0xfff2f20dU, 0xd66b6bbdU, 0xde6f6fb1U, 0x91c5c554U,
 143         0x60303050U, 0x02010103U, 0xce6767a9U, 0x562b2b7dU,
 144         0xe7fefe19U, 0xb5d7d762U, 0x4dababe6U, 0xec76769aU,
 145         0x8fcaca45U, 0x1f82829dU, 0x89c9c940U, 0xfa7d7d87U,
 146         0xeffafa15U, 0xb25959ebU, 0x8e4747c9U, 0xfbf0f00bU,
 147         0x41adadecU, 0xb3d4d467U, 0x5fa2a2fdU, 0x45afafeaU,
 148         0x239c9cbfU, 0x53a4a4f7U, 0xe4727296U, 0x9bc0c05bU,
 149         0x75b7b7c2U, 0xe1fdfd1cU, 0x3d9393aeU, 0x4c26266aU,
 150         0x6c36365aU, 0x7e3f3f41U, 0xf5f7f702U, 0x83cccc4fU,
 151         0x6834345cU, 0x51a5a5f4U, 0xd1e5e534U, 0xf9f1f108U,
 152         0xe2717193U, 0xabd8d873U, 0x62313153U, 0x2a15153fU,
 153         0x0804040cU, 0x95c7c752U, 0x46232365U, 0x9dc3c35eU,
 154         0x30181828U, 0x379696a1U, 0x0a05050fU, 0x2f9a9ab5U,
 155         0x0e070709U, 0x24121236U, 0x1b80809bU, 0xdfe2e23dU,
 156         0xcdebeb26U, 0x4e272769U, 0x7fb2b2cdU, 0xea75759fU,
 157         0x1209091bU, 0x1d83839eU, 0x582c2c74U, 0x341a1a2eU,
 158         0x361b1b2dU, 0xdc6e6eb2U, 0xb45a5aeeU, 0x5ba0a0fbU,
 159         0xa45252f6U, 0x763b3b4dU, 0xb7d6d661U, 0x7db3b3ceU,
 160         0x5229297bU, 0xdde3e33eU, 0x5e2f2f71U, 0x13848497U,
 161         0xa65353f5U, 0xb9d1d168U, 0x00000000U, 0xc1eded2cU,
 162         0x40202060U, 0xe3fcfc1fU, 0x79b1b1c8U, 0xb65b5bedU,
 163         0xd46a6abeU, 0x8dcbcb46U, 0x67bebed9U, 0x7239394bU,
 164         0x944a4adeU, 0x984c4cd4U, 0xb05858e8U, 0x85cfcf4aU,
 165         0xbbd0d06bU, 0xc5efef2aU, 0x4faaaae5U, 0xedfbfb16U,
 166         0x864343c5U, 0x9a4d4dd7U, 0x66333355U, 0x11858594U,
 167         0x8a4545cfU, 0xe9f9f910U, 0x04020206U, 0xfe7f7f81U,
 168         0xa05050f0U, 0x783c3c44U, 0x259f9fbaU, 0x4ba8a8e3U,
 169         0xa25151f3U, 0x5da3a3feU, 0x804040c0U, 0x058f8f8aU,
 170         0x3f9292adU, 0x219d9dbcU, 0x70383848U, 0xf1f5f504U,
 171         0x63bcbcdfU, 0x77b6b6c1U, 0xafdada75U, 0x42212163U,
 172         0x20101030U, 0xe5ffff1aU, 0xfdf3f30eU, 0xbfd2d26dU,
 173         0x81cdcd4cU, 0x180c0c14U, 0x26131335U, 0xc3ecec2fU,
 174         0xbe5f5fe1U, 0x359797a2U, 0x884444ccU, 0x2e171739U,
 175         0x93c4c457U, 0x55a7a7f2U, 0xfc7e7e82U, 0x7a3d3d47U,
 176         0xc86464acU, 0xba5d5de7U, 0x3219192bU, 0xe6737395U,
 177         0xc06060a0U, 0x19818198U, 0x9e4f4fd1U, 0xa3dcdc7fU,
 178         0x44222266U, 0x542a2a7eU, 0x3b9090abU, 0x0b888883U,
 179         0x8c4646caU, 0xc7eeee29U, 0x6bb8b8d3U, 0x2814143cU,
 180         0xa7dede79U, 0xbc5e5ee2U, 0x160b0b1dU, 0xaddbdb76U,
 181         0xdbe0e03bU, 0x64323256U, 0x743a3a4eU, 0x140a0a1eU,
 182         0x924949dbU, 0x0c06060aU, 0x4824246cU, 0xb85c5ce4U,
 183         0x9fc2c25dU, 0xbdd3d36eU, 0x43acacefU, 0xc46262a6U,
 184         0x399191a8U, 0x319595a4U, 0xd3e4e437U, 0xf279798bU,
 185         0xd5e7e732U, 0x8bc8c843U, 0x6e373759U, 0xda6d6db7U,
 186         0x018d8d8cU, 0xb1d5d564U, 0x9c4e4ed2U, 0x49a9a9e0U,
 187         0xd86c6cb4U, 0xac5656faU, 0xf3f4f407U, 0xcfeaea25U,
 188         0xca6565afU, 0xf47a7a8eU, 0x47aeaee9U, 0x10080818U,
 189         0x6fbabad5U, 0xf0787888U, 0x4a25256fU, 0x5c2e2e72U,
 190         0x381c1c24U, 0x57a6a6f1U, 0x73b4b4c7U, 0x97c6c651U,
 191         0xcbe8e823U, 0xa1dddd7cU, 0xe874749cU, 0x3e1f1f21U,
 192         0x964b4bddU, 0x61bdbddcU, 0x0d8b8b86U, 0x0f8a8a85U,
 193         0xe0707090U, 0x7c3e3e42U, 0x71b5b5c4U, 0xcc6666aaU,
 194         0x904848d8U, 0x06030305U, 0xf7f6f601U, 0x1c0e0e12U,
 195         0xc26161a3U, 0x6a35355fU, 0xae5757f9U, 0x69b9b9d0U,
 196         0x17868691U, 0x99c1c158U, 0x3a1d1d27U, 0x279e9eb9U,
 197         0xd9e1e138U, 0xebf8f813U, 0x2b9898b3U, 0x22111133U,
 198         0xd26969bbU, 0xa9d9d970U, 0x078e8e89U, 0x339494a7U,
 199         0x2d9b9bb6U, 0x3c1e1e22U, 0x15878792U, 0xc9e9e920U,
 200         0x87cece49U, 0xaa5555ffU, 0x50282878U, 0xa5dfdf7aU,
 201         0x038c8c8fU, 0x59a1a1f8U, 0x09898980U, 0x1a0d0d17U,
 202         0x65bfbfdaU, 0xd7e6e631U, 0x844242c6U, 0xd06868b8U,
 203         0x824141c3U, 0x299999b0U, 0x5a2d2d77U, 0x1e0f0f11U,
 204         0x7bb0b0cbU, 0xa85454fcU, 0x6dbbbbd6U, 0x2c16163aU
 205 };
 206
 207
 208 static const uint32_t Te1[256] =
 209 {
 210         0xa5c66363U, 0x84f87c7cU, 0x99ee7777U, 0x8df67b7bU,
 211         0x0dfff2f2U, 0xbdd66b6bU, 0xb1de6f6fU, 0x5491c5c5U,
 212         0x50603030U, 0x03020101U, 0xa9ce6767U, 0x7d562b2bU,
 213         0x19e7fefeU, 0x62b5d7d7U, 0xe64dababU, 0x9aec7676U,
 214         0x458fcacaU, 0x9d1f8282U, 0x4089c9c9U, 0x87fa7d7dU,
 215         0x15effafaU, 0xebb25959U, 0xc98e4747U, 0x0bfbf0f0U,
 216         0xec41adadU, 0x67b3d4d4U, 0xfd5fa2a2U, 0xea45afafU,
 217         0xbf239c9cU, 0xf753a4a4U, 0x96e47272U, 0x5b9bc0c0U,
 218         0xc275b7b7U, 0x1ce1fdfdU, 0xae3d9393U, 0x6a4c2626U,
 219         0x5a6c3636U, 0x417e3f3fU, 0x02f5f7f7U, 0x4f83ccccU,
 220         0x5c683434U, 0xf451a5a5U, 0x34d1e5e5U, 0x08f9f1f1U,
 221         0x93e27171U, 0x73abd8d8U, 0x53623131U, 0x3f2a1515U,
 222         0x0c080404U, 0x5295c7c7U, 0x65462323U, 0x5e9dc3c3U,
 223         0x28301818U, 0xa1379696U, 0x0f0a0505U, 0xb52f9a9aU,
 224         0x090e0707U, 0x36241212U, 0x9b1b8080U, 0x3ddfe2e2U,
 225         0x26cdebebU, 0x694e2727U, 0xcd7fb2b2U, 0x9fea7575U,
 226         0x1b120909U, 0x9e1d8383U, 0x74582c2cU, 0x2e341a1aU,
 227         0x2d361b1bU, 0xb2dc6e6eU, 0xeeb45a5aU, 0xfb5ba0a0U,
 228         0xf6a45252U, 0x4d763b3bU, 0x61b7d6d6U, 0xce7db3b3U,
 229         0x7b522929U, 0x3edde3e3U, 0x715e2f2fU, 0x97138484U,
 230         0xf5a65353U, 0x68b9d1d1U, 0x00000000U, 0x2cc1ededU,
 231         0x60402020U, 0x1fe3fcfcU, 0xc879b1b1U, 0xedb65b5bU,
 232         0xbed46a6aU, 0x468dcbcbU, 0xd967bebeU, 0x4b723939U,
 233         0xde944a4aU, 0xd4984c4cU, 0xe8b05858U, 0x4a85cfcfU,
 234         0x6bbbd0d0U, 0x2ac5efefU, 0xe54faaaaU, 0x16edfbfbU,
 235         0xc5864343U, 0xd79a4d4dU, 0x55663333U, 0x94118585U,
 236         0xcf8a4545U, 0x10e9f9f9U, 0x06040202U, 0x81fe7f7fU,
 237         0xf0a05050U, 0x44783c3cU, 0xba259f9fU, 0xe34ba8a8U,
 238         0xf3a25151U, 0xfe5da3a3U, 0xc0804040U, 0x8a058f8fU,
 239         0xad3f9292U, 0xbc219d9dU, 0x48703838U, 0x04f1f5f5U,
 240         0xdf63bcbcU, 0xc177b6b6U, 0x75afdadaU, 0x63422121U,
 241         0x30201010U, 0x1ae5ffffU, 0x0efdf3f3U, 0x6dbfd2d2U,
 242         0x4c81cdcdU, 0x14180c0cU, 0x35261313U, 0x2fc3ececU,
 243         0xe1be5f5fU, 0xa2359797U, 0xcc884444U, 0x392e1717U,
 244         0x5793c4c4U, 0xf255a7a7U, 0x82fc7e7eU, 0x477a3d3dU,
 245         0xacc86464U, 0xe7ba5d5dU, 0x2b321919U, 0x95e67373U,
 246         0xa0c06060U, 0x98198181U, 0xd19e4f4fU, 0x7fa3dcdcU,
 247         0x66442222U, 0x7e542a2aU, 0xab3b9090U, 0x830b8888U,
 248         0xca8c4646U, 0x29c7eeeeU, 0xd36bb8b8U, 0x3c281414U,
 249         0x79a7dedeU, 0xe2bc5e5eU, 0x1d160b0bU, 0x76addbdbU,
 250         0x3bdbe0e0U, 0x56643232U, 0x4e743a3aU, 0x1e140a0aU,
 251         0xdb924949U, 0x0a0c0606U, 0x6c482424U, 0xe4b85c5cU,
 252         0x5d9fc2c2U, 0x6ebdd3d3U, 0xef43acacU, 0xa6c46262U,
 253         0xa8399191U, 0xa4319595U, 0x37d3e4e4U, 0x8bf27979U,
 254         0x32d5e7e7U, 0x438bc8c8U, 0x596e3737U, 0xb7da6d6dU,
 255         0x8c018d8dU, 0x64b1d5d5U, 0xd29c4e4eU, 0xe049a9a9U,
 256         0xb4d86c6cU, 0xfaac5656U, 0x07f3f4f4U, 0x25cfeaeaU,
 257         0xafca6565U, 0x8ef47a7aU, 0xe947aeaeU, 0x18100808U,
 258         0xd56fbabaU, 0x88f07878U, 0x6f4a2525U, 0x725c2e2eU,
 259         0x24381c1cU, 0xf157a6a6U, 0xc773b4b4U, 0x5197c6c6U,
 260         0x23cbe8e8U, 0x7ca1ddddU, 0x9ce87474U, 0x213e1f1fU,
 261         0xdd964b4bU, 0xdc61bdbdU, 0x860d8b8bU, 0x850f8a8aU,
 262         0x90e07070U, 0x427c3e3eU, 0xc471b5b5U, 0xaacc6666U,
 263         0xd8904848U, 0x05060303U, 0x01f7f6f6U, 0x121c0e0eU,
 264         0xa3c26161U, 0x5f6a3535U, 0xf9ae5757U, 0xd069b9b9U,
 265         0x91178686U, 0x5899c1c1U, 0x273a1d1dU, 0xb9279e9eU,
 266         0x38d9e1e1U, 0x13ebf8f8U, 0xb32b9898U, 0x33221111U,
 267         0xbbd26969U, 0x70a9d9d9U, 0x89078e8eU, 0xa7339494U,
 268         0xb62d9b9bU, 0x223c1e1eU, 0x92158787U, 0x20c9e9e9U,
 269         0x4987ceceU, 0xffaa5555U, 0x78502828U, 0x7aa5dfdfU,
 270         0x8f038c8cU, 0xf859a1a1U, 0x80098989U, 0x171a0d0dU,
 271         0xda65bfbfU, 0x31d7e6e6U, 0xc6844242U, 0xb8d06868U,
 272         0xc3824141U, 0xb0299999U, 0x775a2d2dU, 0x111e0f0fU,
 273         0xcb7bb0b0U, 0xfca85454U, 0xd66dbbbbU, 0x3a2c1616U
 274 };
 275
 276
 277 static const uint32_t Te2[256] =
 278 {
 279         0x63a5c663U, 0x7c84f87cU, 0x7799ee77U, 0x7b8df67bU,
 280         0xf20dfff2U, 0x6bbdd66bU, 0x6fb1de6fU, 0xc55491c5U,
 281         0x30506030U, 0x01030201U, 0x67a9ce67U, 0x2b7d562bU,
 282         0xfe19e7feU, 0xd762b5d7U, 0xabe64dabU, 0x769aec76U,
 283         0xca458fcaU, 0x829d1f82U, 0xc94089c9U, 0x7d87fa7dU,
 284         0xfa15effaU, 0x59ebb259U, 0x47c98e47U, 0xf00bfbf0U,
 285         0xadec41adU, 0xd467b3d4U, 0xa2fd5fa2U, 0xafea45afU,
 286         0x9cbf239cU, 0xa4f753a4U, 0x7296e472U, 0xc05b9bc0U,
 287         0xb7c275b7U, 0xfd1ce1fdU, 0x93ae3d93U, 0x266a4c26U,
 288         0x365a6c36U, 0x3f417e3fU, 0xf702f5f7U, 0xcc4f83ccU,
 289         0x345c6834U, 0xa5f451a5U, 0xe534d1e5U, 0xf108f9f1U,
 290         0x7193e271U, 0xd873abd8U, 0x31536231U, 0x153f2a15U,
 291         0x040c0804U, 0xc75295c7U, 0x23654623U, 0xc35e9dc3U,
 292         0x18283018U, 0x96a13796U, 0x050f0a05U, 0x9ab52f9aU,
 293         0x07090e07U, 0x12362412U, 0x809b1b80U, 0xe23ddfe2U,
 294         0xeb26cdebU, 0x27694e27U, 0xb2cd7fb2U, 0x759fea75U,
 295         0x091b1209U, 0x839e1d83U, 0x2c74582cU, 0x1a2e341aU,
 296         0x1b2d361bU, 0x6eb2dc6eU, 0x5aeeb45aU, 0xa0fb5ba0U,
 297         0x52f6a452U, 0x3b4d763bU, 0xd661b7d6U, 0xb3ce7db3U,
 298         0x297b5229U, 0xe33edde3U, 0x2f715e2fU, 0x84971384U,
 299         0x53f5a653U, 0xd168b9d1U, 0x00000000U, 0xed2cc1edU,
 300         0x20604020U, 0xfc1fe3fcU, 0xb1c879b1U, 0x5bedb65bU,
 301         0x6abed46aU, 0xcb468dcbU, 0xbed967beU, 0x394b7239U,
 302         0x4ade944aU, 0x4cd4984cU, 0x58e8b058U, 0xcf4a85cfU,
 303         0xd06bbbd0U, 0xef2ac5efU, 0xaae54faaU, 0xfb16edfbU,
 304         0x43c58643U, 0x4dd79a4dU, 0x33556633U, 0x85941185U,
 305         0x45cf8a45U, 0xf910e9f9U, 0x02060402U, 0x7f81fe7fU,
 306         0x50f0a050U, 0x3c44783cU, 0x9fba259fU, 0xa8e34ba8U,
 307         0x51f3a251U, 0xa3fe5da3U, 0x40c08040U, 0x8f8a058fU,
 308         0x92ad3f92U, 0x9dbc219dU, 0x38487038U, 0xf504f1f5U,
 309         0xbcdf63bcU, 0xb6c177b6U, 0xda75afdaU, 0x21634221U,
 310         0x10302010U, 0xff1ae5ffU, 0xf30efdf3U, 0xd26dbfd2U,
 311         0xcd4c81cdU, 0x0c14180cU, 0x13352613U, 0xec2fc3ecU,
 312         0x5fe1be5fU, 0x97a23597U, 0x44cc8844U, 0x17392e17U,
 313         0xc45793c4U, 0xa7f255a7U, 0x7e82fc7eU, 0x3d477a3dU,
 314         0x64acc864U, 0x5de7ba5dU, 0x192b3219U, 0x7395e673U,
 315         0x60a0c060U, 0x81981981U, 0x4fd19e4fU, 0xdc7fa3dcU,
 316         0x22664422U, 0x2a7e542aU, 0x90ab3b90U, 0x88830b88U,
 317         0x46ca8c46U, 0xee29c7eeU, 0xb8d36bb8U, 0x143c2814U,
 318         0xde79a7deU, 0x5ee2bc5eU, 0x0b1d160bU, 0xdb76addbU,
 319         0xe03bdbe0U, 0x32566432U, 0x3a4e743aU, 0x0a1e140aU,
 320         0x49db9249U, 0x060a0c06U, 0x246c4824U, 0x5ce4b85cU,
 321         0xc25d9fc2U, 0xd36ebdd3U, 0xacef43acU, 0x62a6c462U,
 322         0x91a83991U, 0x95a43195U, 0xe437d3e4U, 0x798bf279U,
 323         0xe732d5e7U, 0xc8438bc8U, 0x37596e37U, 0x6db7da6dU,
 324         0x8d8c018dU, 0xd564b1d5U, 0x4ed29c4eU, 0xa9e049a9U,
 325         0x6cb4d86cU, 0x56faac56U, 0xf407f3f4U, 0xea25cfeaU,
 326         0x65afca65U, 0x7a8ef47aU, 0xaee947aeU, 0x08181008U,
 327         0xbad56fbaU, 0x7888f078U, 0x256f4a25U, 0x2e725c2eU,
 328         0x1c24381cU, 0xa6f157a6U, 0xb4c773b4U, 0xc65197c6U,
 329         0xe823cbe8U, 0xdd7ca1ddU, 0x749ce874U, 0x1f213e1fU,
 330         0x4bdd964bU, 0xbddc61bdU, 0x8b860d8bU, 0x8a850f8aU,
 331         0x7090e070U, 0x3e427c3eU, 0xb5c471b5U, 0x66aacc66U,
 332         0x48d89048U, 0x03050603U, 0xf601f7f6U, 0x0e121c0eU,
 333         0x61a3c261U, 0x355f6a35U, 0x57f9ae57U, 0xb9d069b9U,
 334         0x86911786U, 0xc15899c1U, 0x1d273a1dU, 0x9eb9279eU,
 335         0xe138d9e1U, 0xf813ebf8U, 0x98b32b98U, 0x11332211U,
 336         0x69bbd269U, 0xd970a9d9U, 0x8e89078eU, 0x94a73394U,
 337         0x9bb62d9bU, 0x1e223c1eU, 0x87921587U, 0xe920c9e9U,
 338         0xce4987ceU, 0x55ffaa55U, 0x28785028U, 0xdf7aa5dfU,
 339         0x8c8f038cU, 0xa1f859a1U, 0x89800989U, 0x0d171a0dU,
 340         0xbfda65bfU, 0xe631d7e6U, 0x42c68442U, 0x68b8d068U,
 341         0x41c38241U, 0x99b02999U, 0x2d775a2dU, 0x0f111e0fU,
 342         0xb0cb7bb0U, 0x54fca854U, 0xbbd66dbbU, 0x163a2c16U
 343 };
 344
 345
 346 static const uint32_t Te3[256] =
 347 {
 348         0x6363a5c6U, 0x7c7c84f8U, 0x777799eeU, 0x7b7b8df6U,
 349         0xf2f20dffU, 0x6b6bbdd6U, 0x6f6fb1deU, 0xc5c55491U,
 350         0x30305060U, 0x01010302U, 0x6767a9ceU, 0x2b2b7d56U,
 351         0xfefe19e7U, 0xd7d762b5U, 0xababe64dU, 0x76769aecU,
 352         0xcaca458fU, 0x82829d1fU, 0xc9c94089U, 0x7d7d87faU,
 353         0xfafa15efU, 0x5959ebb2U, 0x4747c98eU, 0xf0f00bfbU,
 354         0xadadec41U, 0xd4d467b3U, 0xa2a2fd5fU, 0xafafea45U,
 355         0x9c9cbf23U, 0xa4a4f753U, 0x727296e4U, 0xc0c05b9bU,
 356         0xb7b7c275U, 0xfdfd1ce1U, 0x9393ae3dU, 0x26266a4cU,
 357         0x36365a6cU, 0x3f3f417eU, 0xf7f702f5U, 0xcccc4f83U,
 358         0x34345c68U, 0xa5a5f451U, 0xe5e534d1U, 0xf1f108f9U,
 359         0x717193e2U, 0xd8d873abU, 0x31315362U, 0x15153f2aU,
 360         0x04040c08U, 0xc7c75295U, 0x23236546U, 0xc3c35e9dU,
 361         0x18182830U, 0x9696a137U, 0x05050f0aU, 0x9a9ab52fU,
 362         0x0707090eU, 0x12123624U, 0x80809b1bU, 0xe2e23ddfU,
 363         0xebeb26cdU, 0x2727694eU, 0xb2b2cd7fU, 0x75759feaU,
 364         0x09091b12U, 0x83839e1dU, 0x2c2c7458U, 0x1a1a2e34U,
 365         0x1b1b2d36U, 0x6e6eb2dcU, 0x5a5aeeb4U, 0xa0a0fb5bU,
 366         0x5252f6a4U, 0x3b3b4d76U, 0xd6d661b7U, 0xb3b3ce7dU,
 367         0x29297b52U, 0xe3e33eddU, 0x2f2f715eU, 0x84849713U,
 368         0x5353f5a6U, 0xd1d168b9U, 0x00000000U, 0xeded2cc1U,
 369         0x20206040U, 0xfcfc1fe3U, 0xb1b1c879U, 0x5b5bedb6U,
 370         0x6a6abed4U, 0xcbcb468dU, 0xbebed967U, 0x39394b72U,
 371         0x4a4ade94U, 0x4c4cd498U, 0x5858e8b0U, 0xcfcf4a85U,
 372         0xd0d06bbbU, 0xefef2ac5U, 0xaaaae54fU, 0xfbfb16edU,
 373         0x4343c586U, 0x4d4dd79aU, 0x33335566U, 0x85859411U,
 374         0x4545cf8aU, 0xf9f910e9U, 0x02020604U, 0x7f7f81feU,
 375         0x5050f0a0U, 0x3c3c4478U, 0x9f9fba25U, 0xa8a8e34bU,
 376         0x5151f3a2U, 0xa3a3fe5dU, 0x4040c080U, 0x8f8f8a05U,
 377         0x9292ad3fU, 0x9d9dbc21U, 0x38384870U, 0xf5f504f1U,
 378         0xbcbcdf63U, 0xb6b6c177U, 0xdada75afU, 0x21216342U,
 379         0x10103020U, 0xffff1ae5U, 0xf3f30efdU, 0xd2d26dbfU,
 380         0xcdcd4c81U, 0x0c0c1418U, 0x13133526U, 0xecec2fc3U,
 381         0x5f5fe1beU, 0x9797a235U, 0x4444cc88U, 0x1717392eU,
 382         0xc4c45793U, 0xa7a7f255U, 0x7e7e82fcU, 0x3d3d477aU,
 383         0x6464acc8U, 0x5d5de7baU, 0x19192b32U, 0x737395e6U,
 384         0x6060a0c0U, 0x81819819U, 0x4f4fd19eU, 0xdcdc7fa3U,
 385         0x22226644U, 0x2a2a7e54U, 0x9090ab3bU, 0x8888830bU,
 386         0x4646ca8cU, 0xeeee29c7U, 0xb8b8d36bU, 0x14143c28U,
 387         0xdede79a7U, 0x5e5ee2bcU, 0x0b0b1d16U, 0xdbdb76adU,
 388         0xe0e03bdbU, 0x32325664U, 0x3a3a4e74U, 0x0a0a1e14U,
 389         0x4949db92U, 0x06060a0cU, 0x24246c48U, 0x5c5ce4b8U,
 390         0xc2c25d9fU, 0xd3d36ebdU, 0xacacef43U, 0x6262a6c4U,
 391         0x9191a839U, 0x9595a431U, 0xe4e437d3U, 0x79798bf2U,
 392         0xe7e732d5U, 0xc8c8438bU, 0x3737596eU, 0x6d6db7daU,
 393         0x8d8d8c01U, 0xd5d564b1U, 0x4e4ed29cU, 0xa9a9e049U,
 394         0x6c6cb4d8U, 0x5656faacU, 0xf4f407f3U, 0xeaea25cfU,
 395         0x6565afcaU, 0x7a7a8ef4U, 0xaeaee947U, 0x08081810U,
 396         0xbabad56fU, 0x787888f0U, 0x25256f4aU, 0x2e2e725cU,
 397         0x1c1c2438U, 0xa6a6f157U, 0xb4b4c773U, 0xc6c65197U,
 398         0xe8e823cbU, 0xdddd7ca1U, 0x74749ce8U, 0x1f1f213eU,
 399         0x4b4bdd96U, 0xbdbddc61U, 0x8b8b860dU, 0x8a8a850fU,
 400         0x707090e0U, 0x3e3e427cU, 0xb5b5c471U, 0x6666aaccU,
 401         0x4848d890U, 0x03030506U, 0xf6f601f7U, 0x0e0e121cU,
 402         0x6161a3c2U, 0x35355f6aU, 0x5757f9aeU, 0xb9b9d069U,
 403         0x86869117U, 0xc1c15899U, 0x1d1d273aU, 0x9e9eb927U,
 404         0xe1e138d9U, 0xf8f813ebU, 0x9898b32bU, 0x11113322U,
 405         0x6969bbd2U, 0xd9d970a9U, 0x8e8e8907U, 0x9494a733U,
 406         0x9b9bb62dU, 0x1e1e223cU, 0x87879215U, 0xe9e920c9U,
 407         0xcece4987U, 0x5555ffaaU, 0x28287850U, 0xdfdf7aa5U,
 408         0x8c8c8f03U, 0xa1a1f859U, 0x89898009U, 0x0d0d171aU,
 409         0xbfbfda65U, 0xe6e631d7U, 0x4242c684U, 0x6868b8d0U,
 410         0x4141c382U, 0x9999b029U, 0x2d2d775aU, 0x0f0f111eU,
 411         0xb0b0cb7bU, 0x5454fca8U, 0xbbbbd66dU, 0x16163a2cU
 412 };
 413
 414 static const uint32_t Te4[256] =
 415 {
 416         0x63636363U, 0x7c7c7c7cU, 0x77777777U, 0x7b7b7b7bU,
 417         0xf2f2f2f2U, 0x6b6b6b6bU, 0x6f6f6f6fU, 0xc5c5c5c5U,
 418         0x30303030U, 0x01010101U, 0x67676767U, 0x2b2b2b2bU,
 419         0xfefefefeU, 0xd7d7d7d7U, 0xababababU, 0x76767676U,
 420         0xcacacacaU, 0x82828282U, 0xc9c9c9c9U, 0x7d7d7d7dU,
 421         0xfafafafaU, 0x59595959U, 0x47474747U, 0xf0f0f0f0U,
 422         0xadadadadU, 0xd4d4d4d4U, 0xa2a2a2a2U, 0xafafafafU,
 423         0x9c9c9c9cU, 0xa4a4a4a4U, 0x72727272U, 0xc0c0c0c0U,
 424         0xb7b7b7b7U, 0xfdfdfdfdU, 0x93939393U, 0x26262626U,
 425         0x36363636U, 0x3f3f3f3fU, 0xf7f7f7f7U, 0xccccccccU,
 426         0x34343434U, 0xa5a5a5a5U, 0xe5e5e5e5U, 0xf1f1f1f1U,
 427         0x71717171U, 0xd8d8d8d8U, 0x31313131U, 0x15151515U,
 428         0x04040404U, 0xc7c7c7c7U, 0x23232323U, 0xc3c3c3c3U,
 429         0x18181818U, 0x96969696U, 0x05050505U, 0x9a9a9a9aU,
 430         0x07070707U, 0x12121212U, 0x80808080U, 0xe2e2e2e2U,
 431         0xebebebebU, 0x27272727U, 0xb2b2b2b2U, 0x75757575U,
 432         0x09090909U, 0x83838383U, 0x2c2c2c2cU, 0x1a1a1a1aU,
 433         0x1b1b1b1bU, 0x6e6e6e6eU, 0x5a5a5a5aU, 0xa0a0a0a0U,
 434         0x52525252U, 0x3b3b3b3bU, 0xd6d6d6d6U, 0xb3b3b3b3U,
 435         0x29292929U, 0xe3e3e3e3U, 0x2f2f2f2fU, 0x84848484U,
 436         0x53535353U, 0xd1d1d1d1U, 0x00000000U, 0xededededU,
 437         0x20202020U, 0xfcfcfcfcU, 0xb1b1b1b1U, 0x5b5b5b5bU,
 438         0x6a6a6a6aU, 0xcbcbcbcbU, 0xbebebebeU, 0x39393939U,
 439         0x4a4a4a4aU, 0x4c4c4c4cU, 0x58585858U, 0xcfcfcfcfU,
 440         0xd0d0d0d0U, 0xefefefefU, 0xaaaaaaaaU, 0xfbfbfbfbU,
 441         0x43434343U, 0x4d4d4d4dU, 0x33333333U, 0x85858585U,
 442         0x45454545U, 0xf9f9f9f9U, 0x02020202U, 0x7f7f7f7fU,
 443         0x50505050U, 0x3c3c3c3cU, 0x9f9f9f9fU, 0xa8a8a8a8U,
 444         0x51515151U, 0xa3a3a3a3U, 0x40404040U, 0x8f8f8f8fU,
 445         0x92929292U, 0x9d9d9d9dU, 0x38383838U, 0xf5f5f5f5U,
 446         0xbcbcbcbcU, 0xb6b6b6b6U, 0xdadadadaU, 0x21212121U,
 447         0x10101010U, 0xffffffffU, 0xf3f3f3f3U, 0xd2d2d2d2U,
 448         0xcdcdcdcdU, 0x0c0c0c0cU, 0x13131313U, 0xececececU,
 449         0x5f5f5f5fU, 0x97979797U, 0x44444444U, 0x17171717U,
 450         0xc4c4c4c4U, 0xa7a7a7a7U, 0x7e7e7e7eU, 0x3d3d3d3dU,
 451         0x64646464U, 0x5d5d5d5dU, 0x19191919U, 0x73737373U,
 452         0x60606060U, 0x81818181U, 0x4f4f4f4fU, 0xdcdcdcdcU,
 453         0x22222222U, 0x2a2a2a2aU, 0x90909090U, 0x88888888U,
 454         0x46464646U, 0xeeeeeeeeU, 0xb8b8b8b8U, 0x14141414U,
 455         0xdedededeU, 0x5e5e5e5eU, 0x0b0b0b0bU, 0xdbdbdbdbU,
 456         0xe0e0e0e0U, 0x32323232U, 0x3a3a3a3aU, 0x0a0a0a0aU,
 457         0x49494949U, 0x06060606U, 0x24242424U, 0x5c5c5c5cU,
 458         0xc2c2c2c2U, 0xd3d3d3d3U, 0xacacacacU, 0x62626262U,
 459         0x91919191U, 0x95959595U, 0xe4e4e4e4U, 0x79797979U,
 460         0xe7e7e7e7U, 0xc8c8c8c8U, 0x37373737U, 0x6d6d6d6dU,
 461         0x8d8d8d8dU, 0xd5d5d5d5U, 0x4e4e4e4eU, 0xa9a9a9a9U,
 462         0x6c6c6c6cU, 0x56565656U, 0xf4f4f4f4U, 0xeaeaeaeaU,
 463         0x65656565U, 0x7a7a7a7aU, 0xaeaeaeaeU, 0x08080808U,
 464         0xbabababaU, 0x78787878U, 0x25252525U, 0x2e2e2e2eU,
 465         0x1c1c1c1cU, 0xa6a6a6a6U, 0xb4b4b4b4U, 0xc6c6c6c6U,
 466         0xe8e8e8e8U, 0xddddddddU, 0x74747474U, 0x1f1f1f1fU,
 467         0x4b4b4b4bU, 0xbdbdbdbdU, 0x8b8b8b8bU, 0x8a8a8a8aU,
 468         0x70707070U, 0x3e3e3e3eU, 0xb5b5b5b5U, 0x66666666U,
 469         0x48484848U, 0x03030303U, 0xf6f6f6f6U, 0x0e0e0e0eU,
 470         0x61616161U, 0x35353535U, 0x57575757U, 0xb9b9b9b9U,
 471         0x86868686U, 0xc1c1c1c1U, 0x1d1d1d1dU, 0x9e9e9e9eU,
 472         0xe1e1e1e1U, 0xf8f8f8f8U, 0x98989898U, 0x11111111U,
 473         0x69696969U, 0xd9d9d9d9U, 0x8e8e8e8eU, 0x94949494U,
 474         0x9b9b9b9bU, 0x1e1e1e1eU, 0x87878787U, 0xe9e9e9e9U,
 475         0xcecececeU, 0x55555555U, 0x28282828U, 0xdfdfdfdfU,
 476         0x8c8c8c8cU, 0xa1a1a1a1U, 0x89898989U, 0x0d0d0d0dU,
 477         0xbfbfbfbfU, 0xe6e6e6e6U, 0x42424242U, 0x68686868U,
 478         0x41414141U, 0x99999999U, 0x2d2d2d2dU, 0x0f0f0f0fU,
 479         0xb0b0b0b0U, 0x54545454U, 0xbbbbbbbbU, 0x16161616U
 480 };
 481
 482 /* Decrypt Sbox constants (for the substitute bytes operation) */
 483
 484 static const uint32_t Td0[256] =
 485 {
 486         0x51f4a750U, 0x7e416553U, 0x1a17a4c3U, 0x3a275e96U,
 487         0x3bab6bcbU, 0x1f9d45f1U, 0xacfa58abU, 0x4be30393U,
 488         0x2030fa55U, 0xad766df6U, 0x88cc7691U, 0xf5024c25U,
 489         0x4fe5d7fcU, 0xc52acbd7U, 0x26354480U, 0xb562a38fU,
 490         0xdeb15a49U, 0x25ba1b67U, 0x45ea0e98U, 0x5dfec0e1U,
 491         0xc32f7502U, 0x814cf012U, 0x8d4697a3U, 0x6bd3f9c6U,
 492         0x038f5fe7U, 0x15929c95U, 0xbf6d7aebU, 0x955259daU,
 493         0xd4be832dU, 0x587421d3U, 0x49e06929U, 0x8ec9c844U,
 494         0x75c2896aU, 0xf48e7978U, 0x99583e6bU, 0x27b971ddU,
 495         0xbee14fb6U, 0xf088ad17U, 0xc920ac66U, 0x7dce3ab4U,
 496         0x63df4a18U, 0xe51a3182U, 0x97513360U, 0x62537f45U,
 497         0xb16477e0U, 0xbb6bae84U, 0xfe81a01cU, 0xf9082b94U,
 498         0x70486858U, 0x8f45fd19U, 0x94de6c87U, 0x527bf8b7U,
 499         0xab73d323U, 0x724b02e2U, 0xe31f8f57U, 0x6655ab2aU,
 500         0xb2eb2807U, 0x2fb5c203U, 0x86c57b9aU, 0xd33708a5U,
 501         0x302887f2U, 0x23bfa5b2U, 0x02036abaU, 0xed16825cU,
 502         0x8acf1c2bU, 0xa779b492U, 0xf307f2f0U, 0x4e69e2a1U,
 503         0x65daf4cdU, 0x0605bed5U, 0xd134621fU, 0xc4a6fe8aU,
 504         0x342e539dU, 0xa2f355a0U, 0x058ae132U, 0xa4f6eb75U,
 505         0x0b83ec39U, 0x4060efaaU, 0x5e719f06U, 0xbd6e1051U,
 506         0x3e218af9U, 0x96dd063dU, 0xdd3e05aeU, 0x4de6bd46U,
 507         0x91548db5U, 0x71c45d05U, 0x0406d46fU, 0x605015ffU,
 508         0x1998fb24U, 0xd6bde997U, 0x894043ccU, 0x67d99e77U,
 509         0xb0e842bdU, 0x07898b88U, 0xe7195b38U, 0x79c8eedbU,
 510         0xa17c0a47U, 0x7c420fe9U, 0xf8841ec9U, 0x00000000U,
 511         0x09808683U, 0x322bed48U, 0x1e1170acU, 0x6c5a724eU,
 512         0xfd0efffbU, 0x0f853856U, 0x3daed51eU, 0x362d3927U,
 513         0x0a0fd964U, 0x685ca621U, 0x9b5b54d1U, 0x24362e3aU,
 514         0x0c0a67b1U, 0x9357e70fU, 0xb4ee96d2U, 0x1b9b919eU,
 515         0x80c0c54fU, 0x61dc20a2U, 0x5a774b69U, 0x1c121a16U,
 516         0xe293ba0aU, 0xc0a02ae5U, 0x3c22e043U, 0x121b171dU,
 517         0x0e090d0bU, 0xf28bc7adU, 0x2db6a8b9U, 0x141ea9c8U,
 518         0x57f11985U, 0xaf75074cU, 0xee99ddbbU, 0xa37f60fdU,
 519         0xf701269fU, 0x5c72f5bcU, 0x44663bc5U, 0x5bfb7e34U,
 520         0x8b432976U, 0xcb23c6dcU, 0xb6edfc68U, 0xb8e4f163U,
 521         0xd731dccaU, 0x42638510U, 0x13972240U, 0x84c61120U,
 522         0x854a247dU, 0xd2bb3df8U, 0xaef93211U, 0xc729a16dU,
 523         0x1d9e2f4bU, 0xdcb230f3U, 0x0d8652ecU, 0x77c1e3d0U,
 524         0x2bb3166cU, 0xa970b999U, 0x119448faU, 0x47e96422U,
 525         0xa8fc8cc4U, 0xa0f03f1aU, 0x567d2cd8U, 0x223390efU,
 526         0x87494ec7U, 0xd938d1c1U, 0x8ccaa2feU, 0x98d40b36U,
 527         0xa6f581cfU, 0xa57ade28U, 0xdab78e26U, 0x3fadbfa4U,
 528         0x2c3a9de4U, 0x5078920dU, 0x6a5fcc9bU, 0x547e4662U,
 529         0xf68d13c2U, 0x90d8b8e8U, 0x2e39f75eU, 0x82c3aff5U,
 530         0x9f5d80beU, 0x69d0937cU, 0x6fd52da9U, 0xcf2512b3U,
 531         0xc8ac993bU, 0x10187da7U, 0xe89c636eU, 0xdb3bbb7bU,
 532         0xcd267809U, 0x6e5918f4U, 0xec9ab701U, 0x834f9aa8U,
 533         0xe6956e65U, 0xaaffe67eU, 0x21bccf08U, 0xef15e8e6U,
 534         0xbae79bd9U, 0x4a6f36ceU, 0xea9f09d4U, 0x29b07cd6U,
 535         0x31a4b2afU, 0x2a3f2331U, 0xc6a59430U, 0x35a266c0U,
 536         0x744ebc37U, 0xfc82caa6U, 0xe090d0b0U, 0x33a7d815U,
 537         0xf104984aU, 0x41ecdaf7U, 0x7fcd500eU, 0x1791f62fU,
 538         0x764dd68dU, 0x43efb04dU, 0xccaa4d54U, 0xe49604dfU,
 539         0x9ed1b5e3U, 0x4c6a881bU, 0xc12c1fb8U, 0x4665517fU,
 540         0x9d5eea04U, 0x018c355dU, 0xfa877473U, 0xfb0b412eU,
 541         0xb3671d5aU, 0x92dbd252U, 0xe9105633U, 0x6dd64713U,
 542         0x9ad7618cU, 0x37a10c7aU, 0x59f8148eU, 0xeb133c89U,
 543         0xcea927eeU, 0xb761c935U, 0xe11ce5edU, 0x7a47b13cU,
 544         0x9cd2df59U, 0x55f2733fU, 0x1814ce79U, 0x73c737bfU,
 545         0x53f7cdeaU, 0x5ffdaa5bU, 0xdf3d6f14U, 0x7844db86U,
 546         0xcaaff381U, 0xb968c43eU, 0x3824342cU, 0xc2a3405fU,
 547         0x161dc372U, 0xbce2250cU, 0x283c498bU, 0xff0d9541U,
 548         0x39a80171U, 0x080cb3deU, 0xd8b4e49cU, 0x6456c190U,
 549         0x7bcb8461U, 0xd532b670U, 0x486c5c74U, 0xd0b85742U
 550 };
 551
 552 static const uint32_t Td1[256] =
 553 {
 554         0x5051f4a7U, 0x537e4165U, 0xc31a17a4U, 0x963a275eU,
 555         0xcb3bab6bU, 0xf11f9d45U, 0xabacfa58U, 0x934be303U,
 556         0x552030faU, 0xf6ad766dU, 0x9188cc76U, 0x25f5024cU,
 557         0xfc4fe5d7U, 0xd7c52acbU, 0x80263544U, 0x8fb562a3U,
 558         0x49deb15aU, 0x6725ba1bU, 0x9845ea0eU, 0xe15dfec0U,
 559         0x02c32f75U, 0x12814cf0U, 0xa38d4697U, 0xc66bd3f9U,
 560         0xe7038f5fU, 0x9515929cU, 0xebbf6d7aU, 0xda955259U,
 561         0x2dd4be83U, 0xd3587421U, 0x2949e069U, 0x448ec9c8U,
 562         0x6a75c289U, 0x78f48e79U, 0x6b99583eU, 0xdd27b971U,
 563         0xb6bee14fU, 0x17f088adU, 0x66c920acU, 0xb47dce3aU,
 564         0x1863df4aU, 0x82e51a31U, 0x60975133U, 0x4562537fU,
 565         0xe0b16477U, 0x84bb6baeU, 0x1cfe81a0U, 0x94f9082bU,
 566         0x58704868U, 0x198f45fdU, 0x8794de6cU, 0xb7527bf8U,
 567         0x23ab73d3U, 0xe2724b02U, 0x57e31f8fU, 0x2a6655abU,
 568         0x07b2eb28U, 0x032fb5c2U, 0x9a86c57bU, 0xa5d33708U,
 569         0xf2302887U, 0xb223bfa5U, 0xba02036aU, 0x5ced1682U,
 570         0x2b8acf1cU, 0x92a779b4U, 0xf0f307f2U, 0xa14e69e2U,
 571         0xcd65daf4U, 0xd50605beU, 0x1fd13462U, 0x8ac4a6feU,
 572         0x9d342e53U, 0xa0a2f355U, 0x32058ae1U, 0x75a4f6ebU,
 573         0x390b83ecU, 0xaa4060efU, 0x065e719fU, 0x51bd6e10U,
 574         0xf93e218aU, 0x3d96dd06U, 0xaedd3e05U, 0x464de6bdU,
 575         0xb591548dU, 0x0571c45dU, 0x6f0406d4U, 0xff605015U,
 576         0x241998fbU, 0x97d6bde9U, 0xcc894043U, 0x7767d99eU,
 577         0xbdb0e842U, 0x8807898bU, 0x38e7195bU, 0xdb79c8eeU,
 578         0x47a17c0aU, 0xe97c420fU, 0xc9f8841eU, 0x00000000U,
 579         0x83098086U, 0x48322bedU, 0xac1e1170U, 0x4e6c5a72U,
 580         0xfbfd0effU, 0x560f8538U, 0x1e3daed5U, 0x27362d39U,
 581         0x640a0fd9U, 0x21685ca6U, 0xd19b5b54U, 0x3a24362eU,
 582         0xb10c0a67U, 0x0f9357e7U, 0xd2b4ee96U, 0x9e1b9b91U,
 583         0x4f80c0c5U, 0xa261dc20U, 0x695a774bU, 0x161c121aU,
 584         0x0ae293baU, 0xe5c0a02aU, 0x433c22e0U, 0x1d121b17U,
 585         0x0b0e090dU, 0xadf28bc7U, 0xb92db6a8U, 0xc8141ea9U,
 586         0x8557f119U, 0x4caf7507U, 0xbbee99ddU, 0xfda37f60U,
 587         0x9ff70126U, 0xbc5c72f5U, 0xc544663bU, 0x345bfb7eU,
 588         0x768b4329U, 0xdccb23c6U, 0x68b6edfcU, 0x63b8e4f1U,
 589         0xcad731dcU, 0x10426385U, 0x40139722U, 0x2084c611U,
 590         0x7d854a24U, 0xf8d2bb3dU, 0x11aef932U, 0x6dc729a1U,
 591         0x4b1d9e2fU, 0xf3dcb230U, 0xec0d8652U, 0xd077c1e3U,
 592         0x6c2bb316U, 0x99a970b9U, 0xfa119448U, 0x2247e964U,
 593         0xc4a8fc8cU, 0x1aa0f03fU, 0xd8567d2cU, 0xef223390U,
 594         0xc787494eU, 0xc1d938d1U, 0xfe8ccaa2U, 0x3698d40bU,
 595         0xcfa6f581U, 0x28a57adeU, 0x26dab78eU, 0xa43fadbfU,
 596         0xe42c3a9dU, 0x0d507892U, 0x9b6a5fccU, 0x62547e46U,
 597         0xc2f68d13U, 0xe890d8b8U, 0x5e2e39f7U, 0xf582c3afU,
 598         0xbe9f5d80U, 0x7c69d093U, 0xa96fd52dU, 0xb3cf2512U,
 599         0x3bc8ac99U, 0xa710187dU, 0x6ee89c63U, 0x7bdb3bbbU,
 600         0x09cd2678U, 0xf46e5918U, 0x01ec9ab7U, 0xa8834f9aU,
 601         0x65e6956eU, 0x7eaaffe6U, 0x0821bccfU, 0xe6ef15e8U,
 602         0xd9bae79bU, 0xce4a6f36U, 0xd4ea9f09U, 0xd629b07cU,
 603         0xaf31a4b2U, 0x312a3f23U, 0x30c6a594U, 0xc035a266U,
 604         0x37744ebcU, 0xa6fc82caU, 0xb0e090d0U, 0x1533a7d8U,
 605         0x4af10498U, 0xf741ecdaU, 0x0e7fcd50U, 0x2f1791f6U,
 606         0x8d764dd6U, 0x4d43efb0U, 0x54ccaa4dU, 0xdfe49604U,
 607         0xe39ed1b5U, 0x1b4c6a88U, 0xb8c12c1fU, 0x7f466551U,
 608         0x049d5eeaU, 0x5d018c35U, 0x73fa8774U, 0x2efb0b41U,
 609         0x5ab3671dU, 0x5292dbd2U, 0x33e91056U, 0x136dd647U,
 610         0x8c9ad761U, 0x7a37a10cU, 0x8e59f814U, 0x89eb133cU,
 611         0xeecea927U, 0x35b761c9U, 0xede11ce5U, 0x3c7a47b1U,
 612         0x599cd2dfU, 0x3f55f273U, 0x791814ceU, 0xbf73c737U,
 613         0xea53f7cdU, 0x5b5ffdaaU, 0x14df3d6fU, 0x867844dbU,
 614         0x81caaff3U, 0x3eb968c4U, 0x2c382434U, 0x5fc2a340U,
 615         0x72161dc3U, 0x0cbce225U, 0x8b283c49U, 0x41ff0d95U,
 616         0x7139a801U, 0xde080cb3U, 0x9cd8b4e4U, 0x906456c1U,
 617         0x617bcb84U, 0x70d532b6U, 0x74486c5cU, 0x42d0b857U
 618 };
 619
 620 static const uint32_t Td2[256] =
 621 {
 622         0xa75051f4U, 0x65537e41U, 0xa4c31a17U, 0x5e963a27U,
 623         0x6bcb3babU, 0x45f11f9dU, 0x58abacfaU, 0x03934be3U,
 624         0xfa552030U, 0x6df6ad76U, 0x769188ccU, 0x4c25f502U,
 625         0xd7fc4fe5U, 0xcbd7c52aU, 0x44802635U, 0xa38fb562U,
 626         0x5a49deb1U, 0x1b6725baU, 0x0e9845eaU, 0xc0e15dfeU,
 627         0x7502c32fU, 0xf012814cU, 0x97a38d46U, 0xf9c66bd3U,
 628         0x5fe7038fU, 0x9c951592U, 0x7aebbf6dU, 0x59da9552U,
 629         0x832dd4beU, 0x21d35874U, 0x692949e0U, 0xc8448ec9U,
 630         0x896a75c2U, 0x7978f48eU, 0x3e6b9958U, 0x71dd27b9U,
 631         0x4fb6bee1U, 0xad17f088U, 0xac66c920U, 0x3ab47dceU,
 632         0x4a1863dfU, 0x3182e51aU, 0x33609751U, 0x7f456253U,
 633         0x77e0b164U, 0xae84bb6bU, 0xa01cfe81U, 0x2b94f908U,
 634         0x68587048U, 0xfd198f45U, 0x6c8794deU, 0xf8b7527bU,
 635         0xd323ab73U, 0x02e2724bU, 0x8f57e31fU, 0xab2a6655U,
 636         0x2807b2ebU, 0xc2032fb5U, 0x7b9a86c5U, 0x08a5d337U,
 637         0x87f23028U, 0xa5b223bfU, 0x6aba0203U, 0x825ced16U,
 638         0x1c2b8acfU, 0xb492a779U, 0xf2f0f307U, 0xe2a14e69U,
 639         0xf4cd65daU, 0xbed50605U, 0x621fd134U, 0xfe8ac4a6U,
 640         0x539d342eU, 0x55a0a2f3U, 0xe132058aU, 0xeb75a4f6U,
 641         0xec390b83U, 0xefaa4060U, 0x9f065e71U, 0x1051bd6eU,
 642         0x8af93e21U, 0x063d96ddU, 0x05aedd3eU, 0xbd464de6U,
 643         0x8db59154U, 0x5d0571c4U, 0xd46f0406U, 0x15ff6050U,
 644         0xfb241998U, 0xe997d6bdU, 0x43cc8940U, 0x9e7767d9U,
 645         0x42bdb0e8U, 0x8b880789U, 0x5b38e719U, 0xeedb79c8U,
 646         0x0a47a17cU, 0x0fe97c42U, 0x1ec9f884U, 0x00000000U,
 647         0x86830980U, 0xed48322bU, 0x70ac1e11U, 0x724e6c5aU,
 648         0xfffbfd0eU, 0x38560f85U, 0xd51e3daeU, 0x3927362dU,
 649         0xd9640a0fU, 0xa621685cU, 0x54d19b5bU, 0x2e3a2436U,
 650         0x67b10c0aU, 0xe70f9357U, 0x96d2b4eeU, 0x919e1b9bU,
 651         0xc54f80c0U, 0x20a261dcU, 0x4b695a77U, 0x1a161c12U,
 652         0xba0ae293U, 0x2ae5c0a0U, 0xe0433c22U, 0x171d121bU,
 653         0x0d0b0e09U, 0xc7adf28bU, 0xa8b92db6U, 0xa9c8141eU,
 654         0x198557f1U, 0x074caf75U, 0xddbbee99U, 0x60fda37fU,
 655         0x269ff701U, 0xf5bc5c72U, 0x3bc54466U, 0x7e345bfbU,
 656         0x29768b43U, 0xc6dccb23U, 0xfc68b6edU, 0xf163b8e4U,
 657         0xdccad731U, 0x85104263U, 0x22401397U, 0x112084c6U,
 658         0x247d854aU, 0x3df8d2bbU, 0x3211aef9U, 0xa16dc729U,
 659         0x2f4b1d9eU, 0x30f3dcb2U, 0x52ec0d86U, 0xe3d077c1U,
 660         0x166c2bb3U, 0xb999a970U, 0x48fa1194U, 0x642247e9U,
 661         0x8cc4a8fcU, 0x3f1aa0f0U, 0x2cd8567dU, 0x90ef2233U,
 662         0x4ec78749U, 0xd1c1d938U, 0xa2fe8ccaU, 0x0b3698d4U,
 663         0x81cfa6f5U, 0xde28a57aU, 0x8e26dab7U, 0xbfa43fadU,
 664         0x9de42c3aU, 0x920d5078U, 0xcc9b6a5fU, 0x4662547eU,
 665         0x13c2f68dU, 0xb8e890d8U, 0xf75e2e39U, 0xaff582c3U,
 666         0x80be9f5dU, 0x937c69d0U, 0x2da96fd5U, 0x12b3cf25U,
 667         0x993bc8acU, 0x7da71018U, 0x636ee89cU, 0xbb7bdb3bU,
 668         0x7809cd26U, 0x18f46e59U, 0xb701ec9aU, 0x9aa8834fU,
 669         0x6e65e695U, 0xe67eaaffU, 0xcf0821bcU, 0xe8e6ef15U,
 670         0x9bd9bae7U, 0x36ce4a6fU, 0x09d4ea9fU, 0x7cd629b0U,
 671         0xb2af31a4U, 0x23312a3fU, 0x9430c6a5U, 0x66c035a2U,
 672         0xbc37744eU, 0xcaa6fc82U, 0xd0b0e090U, 0xd81533a7U,
 673         0x984af104U, 0xdaf741ecU, 0x500e7fcdU, 0xf62f1791U,
 674         0xd68d764dU, 0xb04d43efU, 0x4d54ccaaU, 0x04dfe496U,
 675         0xb5e39ed1U, 0x881b4c6aU, 0x1fb8c12cU, 0x517f4665U,
 676         0xea049d5eU, 0x355d018cU, 0x7473fa87U, 0x412efb0bU,
 677         0x1d5ab367U, 0xd25292dbU, 0x5633e910U, 0x47136dd6U,
 678         0x618c9ad7U, 0x0c7a37a1U, 0x148e59f8U, 0x3c89eb13U,
 679         0x27eecea9U, 0xc935b761U, 0xe5ede11cU, 0xb13c7a47U,
 680         0xdf599cd2U, 0x733f55f2U, 0xce791814U, 0x37bf73c7U,
 681         0xcdea53f7U, 0xaa5b5ffdU, 0x6f14df3dU, 0xdb867844U,
 682         0xf381caafU, 0xc43eb968U, 0x342c3824U, 0x405fc2a3U,
 683         0xc372161dU, 0x250cbce2U, 0x498b283cU, 0x9541ff0dU,
 684         0x017139a8U, 0xb3de080cU, 0xe49cd8b4U, 0xc1906456U,
 685         0x84617bcbU, 0xb670d532U, 0x5c74486cU, 0x5742d0b8U
 686 };
 687
 688 static const uint32_t Td3[256] =
 689 {
 690         0xf4a75051U, 0x4165537eU, 0x17a4c31aU, 0x275e963aU,
 691         0xab6bcb3bU, 0x9d45f11fU, 0xfa58abacU, 0xe303934bU,
 692         0x30fa5520U, 0x766df6adU, 0xcc769188U, 0x024c25f5U,
 693         0xe5d7fc4fU, 0x2acbd7c5U, 0x35448026U, 0x62a38fb5U,
 694         0xb15a49deU, 0xba1b6725U, 0xea0e9845U, 0xfec0e15dU,
 695         0x2f7502c3U, 0x4cf01281U, 0x4697a38dU, 0xd3f9c66bU,
 696         0x8f5fe703U, 0x929c9515U, 0x6d7aebbfU, 0x5259da95U,
 697         0xbe832dd4U, 0x7421d358U, 0xe0692949U, 0xc9c8448eU,
 698         0xc2896a75U, 0x8e7978f4U, 0x583e6b99U, 0xb971dd27U,
 699         0xe14fb6beU, 0x88ad17f0U, 0x20ac66c9U, 0xce3ab47dU,
 700         0xdf4a1863U, 0x1a3182e5U, 0x51336097U, 0x537f4562U,
 701         0x6477e0b1U, 0x6bae84bbU, 0x81a01cfeU, 0x082b94f9U,
 702         0x48685870U, 0x45fd198fU, 0xde6c8794U, 0x7bf8b752U,
 703         0x73d323abU, 0x4b02e272U, 0x1f8f57e3U, 0x55ab2a66U,
 704         0xeb2807b2U, 0xb5c2032fU, 0xc57b9a86U, 0x3708a5d3U,
 705         0x2887f230U, 0xbfa5b223U, 0x036aba02U, 0x16825cedU,
 706         0xcf1c2b8aU, 0x79b492a7U, 0x07f2f0f3U, 0x69e2a14eU,
 707         0xdaf4cd65U, 0x05bed506U, 0x34621fd1U, 0xa6fe8ac4U,
 708         0x2e539d34U, 0xf355a0a2U, 0x8ae13205U, 0xf6eb75a4U,
 709         0x83ec390bU, 0x60efaa40U, 0x719f065eU, 0x6e1051bdU,
 710         0x218af93eU, 0xdd063d96U, 0x3e05aeddU, 0xe6bd464dU,
 711         0x548db591U, 0xc45d0571U, 0x06d46f04U, 0x5015ff60U,
 712         0x98fb2419U, 0xbde997d6U, 0x4043cc89U, 0xd99e7767U,
 713         0xe842bdb0U, 0x898b8807U, 0x195b38e7U, 0xc8eedb79U,
 714         0x7c0a47a1U, 0x420fe97cU, 0x841ec9f8U, 0x00000000U,
 715         0x80868309U, 0x2bed4832U, 0x1170ac1eU, 0x5a724e6cU,
 716         0x0efffbfdU, 0x8538560fU, 0xaed51e3dU, 0x2d392736U,
 717         0x0fd9640aU, 0x5ca62168U, 0x5b54d19bU, 0x362e3a24U,
 718         0x0a67b10cU, 0x57e70f93U, 0xee96d2b4U, 0x9b919e1bU,
 719         0xc0c54f80U, 0xdc20a261U, 0x774b695aU, 0x121a161cU,
 720         0x93ba0ae2U, 0xa02ae5c0U, 0x22e0433cU, 0x1b171d12U,
 721         0x090d0b0eU, 0x8bc7adf2U, 0xb6a8b92dU, 0x1ea9c814U,
 722         0xf1198557U, 0x75074cafU, 0x99ddbbeeU, 0x7f60fda3U,
 723         0x01269ff7U, 0x72f5bc5cU, 0x663bc544U, 0xfb7e345bU,
 724         0x4329768bU, 0x23c6dccbU, 0xedfc68b6U, 0xe4f163b8U,
 725         0x31dccad7U, 0x63851042U, 0x97224013U, 0xc6112084U,
 726         0x4a247d85U, 0xbb3df8d2U, 0xf93211aeU, 0x29a16dc7U,
 727         0x9e2f4b1dU, 0xb230f3dcU, 0x8652ec0dU, 0xc1e3d077U,
 728         0xb3166c2bU, 0x70b999a9U, 0x9448fa11U, 0xe9642247U,
 729         0xfc8cc4a8U, 0xf03f1aa0U, 0x7d2cd856U, 0x3390ef22U,
 730         0x494ec787U, 0x38d1c1d9U, 0xcaa2fe8cU, 0xd40b3698U,
 731         0xf581cfa6U, 0x7ade28a5U, 0xb78e26daU, 0xadbfa43fU,
 732         0x3a9de42cU, 0x78920d50U, 0x5fcc9b6aU, 0x7e466254U,
 733         0x8d13c2f6U, 0xd8b8e890U, 0x39f75e2eU, 0xc3aff582U,
 734         0x5d80be9fU, 0xd0937c69U, 0xd52da96fU, 0x2512b3cfU,
 735         0xac993bc8U, 0x187da710U, 0x9c636ee8U, 0x3bbb7bdbU,
 736         0x267809cdU, 0x5918f46eU, 0x9ab701ecU, 0x4f9aa883U,
 737         0x956e65e6U, 0xffe67eaaU, 0xbccf0821U, 0x15e8e6efU,
 738         0xe79bd9baU, 0x6f36ce4aU, 0x9f09d4eaU, 0xb07cd629U,
 739         0xa4b2af31U, 0x3f23312aU, 0xa59430c6U, 0xa266c035U,
 740         0x4ebc3774U, 0x82caa6fcU, 0x90d0b0e0U, 0xa7d81533U,
 741         0x04984af1U, 0xecdaf741U, 0xcd500e7fU, 0x91f62f17U,
 742         0x4dd68d76U, 0xefb04d43U, 0xaa4d54ccU, 0x9604dfe4U,
 743         0xd1b5e39eU, 0x6a881b4cU, 0x2c1fb8c1U, 0x65517f46U,
 744         0x5eea049dU, 0x8c355d01U, 0x877473faU, 0x0b412efbU,
 745         0x671d5ab3U, 0xdbd25292U, 0x105633e9U, 0xd647136dU,
 746         0xd7618c9aU, 0xa10c7a37U, 0xf8148e59U, 0x133c89ebU,
 747         0xa927eeceU, 0x61c935b7U, 0x1ce5ede1U, 0x47b13c7aU,
 748         0xd2df599cU, 0xf2733f55U, 0x14ce7918U, 0xc737bf73U,
 749         0xf7cdea53U, 0xfdaa5b5fU, 0x3d6f14dfU, 0x44db8678U,
 750         0xaff381caU, 0x68c43eb9U, 0x24342c38U, 0xa3405fc2U,
 751         0x1dc37216U, 0xe2250cbcU, 0x3c498b28U, 0x0d9541ffU,
 752         0xa8017139U, 0x0cb3de08U, 0xb4e49cd8U, 0x56c19064U,
 753         0xcb84617bU, 0x32b670d5U, 0x6c5c7448U, 0xb85742d0U
 754 };
 755
 756 static const uint32_t Td4[256] =
 757 {
 758         0x52525252U, 0x09090909U, 0x6a6a6a6aU, 0xd5d5d5d5U,
 759         0x30303030U, 0x36363636U, 0xa5a5a5a5U, 0x38383838U,
 760         0xbfbfbfbfU, 0x40404040U, 0xa3a3a3a3U, 0x9e9e9e9eU,
 761         0x81818181U, 0xf3f3f3f3U, 0xd7d7d7d7U, 0xfbfbfbfbU,
 762         0x7c7c7c7cU, 0xe3e3e3e3U, 0x39393939U, 0x82828282U,
 763         0x9b9b9b9bU, 0x2f2f2f2fU, 0xffffffffU, 0x87878787U,
 764         0x34343434U, 0x8e8e8e8eU, 0x43434343U, 0x44444444U,
 765         0xc4c4c4c4U, 0xdedededeU, 0xe9e9e9e9U, 0xcbcbcbcbU,
 766         0x54545454U, 0x7b7b7b7bU, 0x94949494U, 0x32323232U,
 767         0xa6a6a6a6U, 0xc2c2c2c2U, 0x23232323U, 0x3d3d3d3dU,
 768         0xeeeeeeeeU, 0x4c4c4c4cU, 0x95959595U, 0x0b0b0b0bU,
 769         0x42424242U, 0xfafafafaU, 0xc3c3c3c3U, 0x4e4e4e4eU,
 770         0x08080808U, 0x2e2e2e2eU, 0xa1a1a1a1U, 0x66666666U,
 771         0x28282828U, 0xd9d9d9d9U, 0x24242424U, 0xb2b2b2b2U,
 772         0x76767676U, 0x5b5b5b5bU, 0xa2a2a2a2U, 0x49494949U,
 773         0x6d6d6d6dU, 0x8b8b8b8bU, 0xd1d1d1d1U, 0x25252525U,
 774         0x72727272U, 0xf8f8f8f8U, 0xf6f6f6f6U, 0x64646464U,
 775         0x86868686U, 0x68686868U, 0x98989898U, 0x16161616U,
 776         0xd4d4d4d4U, 0xa4a4a4a4U, 0x5c5c5c5cU, 0xccccccccU,
 777         0x5d5d5d5dU, 0x65656565U, 0xb6b6b6b6U, 0x92929292U,
 778         0x6c6c6c6cU, 0x70707070U, 0x48484848U, 0x50505050U,
 779         0xfdfdfdfdU, 0xededededU, 0xb9b9b9b9U, 0xdadadadaU,
 780         0x5e5e5e5eU, 0x15151515U, 0x46464646U, 0x57575757U,
 781         0xa7a7a7a7U, 0x8d8d8d8dU, 0x9d9d9d9dU, 0x84848484U,
 782         0x90909090U, 0xd8d8d8d8U, 0xababababU, 0x00000000U,
 783         0x8c8c8c8cU, 0xbcbcbcbcU, 0xd3d3d3d3U, 0x0a0a0a0aU,
 784         0xf7f7f7f7U, 0xe4e4e4e4U, 0x58585858U, 0x05050505U,
 785         0xb8b8b8b8U, 0xb3b3b3b3U, 0x45454545U, 0x06060606U,
 786         0xd0d0d0d0U, 0x2c2c2c2cU, 0x1e1e1e1eU, 0x8f8f8f8fU,
 787         0xcacacacaU, 0x3f3f3f3fU, 0x0f0f0f0fU, 0x02020202U,
 788         0xc1c1c1c1U, 0xafafafafU, 0xbdbdbdbdU, 0x03030303U,
 789         0x01010101U, 0x13131313U, 0x8a8a8a8aU, 0x6b6b6b6bU,
 790         0x3a3a3a3aU, 0x91919191U, 0x11111111U, 0x41414141U,
 791         0x4f4f4f4fU, 0x67676767U, 0xdcdcdcdcU, 0xeaeaeaeaU,
 792         0x97979797U, 0xf2f2f2f2U, 0xcfcfcfcfU, 0xcecececeU,
 793         0xf0f0f0f0U, 0xb4b4b4b4U, 0xe6e6e6e6U, 0x73737373U,
 794         0x96969696U, 0xacacacacU, 0x74747474U, 0x22222222U,
 795         0xe7e7e7e7U, 0xadadadadU, 0x35353535U, 0x85858585U,
 796         0xe2e2e2e2U, 0xf9f9f9f9U, 0x37373737U, 0xe8e8e8e8U,
 797         0x1c1c1c1cU, 0x75757575U, 0xdfdfdfdfU, 0x6e6e6e6eU,
 798         0x47474747U, 0xf1f1f1f1U, 0x1a1a1a1aU, 0x71717171U,
 799         0x1d1d1d1dU, 0x29292929U, 0xc5c5c5c5U, 0x89898989U,
 800         0x6f6f6f6fU, 0xb7b7b7b7U, 0x62626262U, 0x0e0e0e0eU,
 801         0xaaaaaaaaU, 0x18181818U, 0xbebebebeU, 0x1b1b1b1bU,
 802         0xfcfcfcfcU, 0x56565656U, 0x3e3e3e3eU, 0x4b4b4b4bU,
 803         0xc6c6c6c6U, 0xd2d2d2d2U, 0x79797979U, 0x20202020U,
 804         0x9a9a9a9aU, 0xdbdbdbdbU, 0xc0c0c0c0U, 0xfefefefeU,
 805         0x78787878U, 0xcdcdcdcdU, 0x5a5a5a5aU, 0xf4f4f4f4U,
 806         0x1f1f1f1fU, 0xddddddddU, 0xa8a8a8a8U, 0x33333333U,
 807         0x88888888U, 0x07070707U, 0xc7c7c7c7U, 0x31313131U,
 808         0xb1b1b1b1U, 0x12121212U, 0x10101010U, 0x59595959U,
 809         0x27272727U, 0x80808080U, 0xececececU, 0x5f5f5f5fU,
 810         0x60606060U, 0x51515151U, 0x7f7f7f7fU, 0xa9a9a9a9U,
 811         0x19191919U, 0xb5b5b5b5U, 0x4a4a4a4aU, 0x0d0d0d0dU,
 812         0x2d2d2d2dU, 0xe5e5e5e5U, 0x7a7a7a7aU, 0x9f9f9f9fU,
 813         0x93939393U, 0xc9c9c9c9U, 0x9c9c9c9cU, 0xefefefefU,
 814         0xa0a0a0a0U, 0xe0e0e0e0U, 0x3b3b3b3bU, 0x4d4d4d4dU,
 815         0xaeaeaeaeU, 0x2a2a2a2aU, 0xf5f5f5f5U, 0xb0b0b0b0U,
 816         0xc8c8c8c8U, 0xebebebebU, 0xbbbbbbbbU, 0x3c3c3c3cU,
 817         0x83838383U, 0x53535353U, 0x99999999U, 0x61616161U,
 818         0x17171717U, 0x2b2b2b2bU, 0x04040404U, 0x7e7e7e7eU,
 819         0xbabababaU, 0x77777777U, 0xd6d6d6d6U, 0x26262626U,
 820         0xe1e1e1e1U, 0x69696969U, 0x14141414U, 0x63636363U,
 821         0x55555555U, 0x21212121U, 0x0c0c0c0cU, 0x7d7d7d7dU
 822 };
 823
 824 /* Rcon is Round Constant; used for encryption key expansion */
 825 static const uint32_t rcon[RC_LENGTH] =
 826 {
 827         /* for 128-bit blocks, Rijndael never uses more than 10 rcon values */
 828         0x01000000, 0x02000000, 0x04000000, 0x08000000,
 829         0x10000000, 0x20000000, 0x40000000, 0x80000000,
 830         0x1B000000, 0x36000000
 831 };
 832
 833
 834 /*
 835  * Expand the cipher key into the encryption key schedule.
 836  *
 837  * Return the number of rounds for the given cipher key size.
 838  * The size of the key schedule depends on the number of rounds
 839  * (which can be computed from the size of the key), i.e. 4*(Nr + 1).
 840  *
 841  * Parameters:
 842  * rk           AES key schedule 32-bit array to be initialized
 843  * cipherKey    User key
 844  * keyBits      AES key size (128, 192, or 256 bits)
 845  */
 846 static int
 847 rijndael_key_setup_enc_raw(uint32_t rk[], const uint32_t cipherKey[],
 848     int keyBits)
 849 {
 850         int             i = 0;
 851         uint32_t        temp;
 852
 853         rk[0] = cipherKey[0];
 854         rk[1] = cipherKey[1];
 855         rk[2] = cipherKey[2];
 856         rk[3] = cipherKey[3];
 857
 858         if (keyBits == 128) {
 859                 for (;;) {
 860                         temp  = rk[3];
 861                         rk[4] = rk[0] ^
 862                             (Te4[(temp >> 16) & 0xff] & 0xff000000) ^
 863                             (Te4[(temp >> 8) & 0xff] & 0x00ff0000) ^
 864                             (Te4[temp & 0xff] & 0x0000ff00) ^
 865                             (Te4[temp >> 24] & 0x000000ff) ^
 866                             rcon[i];
 867                         rk[5] = rk[1] ^ rk[4];
 868                         rk[6] = rk[2] ^ rk[5];
 869                         rk[7] = rk[3] ^ rk[6];
 870
 871                         if (++i == 10) {
 872                                 return (10);
 873                         }
 874                         rk += 4;
 875                 }
 876         }
 877
 878         rk[4] = cipherKey[4];
 879         rk[5] = cipherKey[5];
 880
 881         if (keyBits == 192) {
 882                 for (;;) {
 883                         temp = rk[5];
 884                         rk[6] = rk[0] ^
 885                             (Te4[(temp >> 16) & 0xff] & 0xff000000) ^
 886                             (Te4[(temp >> 8) & 0xff] & 0x00ff0000) ^
 887                             (Te4[temp & 0xff] & 0x0000ff00) ^
 888                             (Te4[temp >> 24] & 0x000000ff) ^
 889                             rcon[i];
 890                         rk[7] = rk[1] ^ rk[6];
 891                         rk[8] = rk[2] ^ rk[7];
 892                         rk[9] = rk[3] ^ rk[8];
 893
 894                         if (++i == 8) {
 895                                 return (12);
 896                         }
 897
 898                         rk[10] = rk[4] ^ rk[9];
 899                         rk[11] = rk[5] ^ rk[10];
 900                         rk += 6;
 901                 }
 902         }
 903
 904         rk[6] = cipherKey[6];
 905         rk[7] = cipherKey[7];
 906
 907         if (keyBits == 256) {
 908                 for (;;) {
 909                         temp = rk[7];
 910                         rk[8] = rk[0] ^
 911                             (Te4[(temp >> 16) & 0xff] & 0xff000000) ^
 912                             (Te4[(temp >> 8) & 0xff] & 0x00ff0000) ^
 913                             (Te4[temp & 0xff] & 0x0000ff00) ^
 914                             (Te4[temp >> 24] & 0x000000ff) ^
 915                             rcon[i];
 916                         rk[9] = rk[1] ^ rk[8];
 917                         rk[10] = rk[2] ^ rk[9];
 918                         rk[11] = rk[3] ^ rk[10];
 919
 920                         if (++i == 7) {
 921                                 return (14);
 922                         }
 923                         temp = rk[11];
 924                         rk[12] = rk[4] ^
 925                             (Te4[temp >> 24] & 0xff000000) ^
 926                             (Te4[(temp >> 16) & 0xff] & 0x00ff0000) ^
 927                             (Te4[(temp >> 8) & 0xff] & 0x0000ff00) ^
 928                             (Te4[temp & 0xff] & 0x000000ff);
 929                         rk[13] = rk[5] ^ rk[12];
 930                         rk[14] = rk[6] ^ rk[13];
 931                         rk[15] = rk[7] ^ rk[14];
 932
 933                         rk += 8;
 934                 }
 935         }
 936
 937         return (0);
 938 }
 939 #endif  /* !__amd64 */
 940
 941 #if defined(__amd64)
 942
 943 /*
 944  * Expand the 32-bit AES cipher key array into the encryption and decryption
 945  * key schedules.
 946  *
 947  * Parameters:
 948  * key          AES key schedule to be initialized
 949  * keyarr32     User key
 950  * keyBits      AES key size (128, 192, or 256 bits)
 951  */
 952 static void
 953 aes_setupkeys(aes_key_t *key, const uint32_t *keyarr32, int keybits)
 954 {
 955         if (intel_aes_instructions_present()) {
 956                 key->flags = INTEL_AES_NI_CAPABLE;
 957                 KPREEMPT_DISABLE;
 958                 key->nr = rijndael_key_setup_enc_intel(&(key->encr_ks.ks32[0]),
 959                     keyarr32, keybits);
 960                 key->nr = rijndael_key_setup_dec_intel(&(key->decr_ks.ks32[0]),
 961                     keyarr32, keybits);
 962                 KPREEMPT_ENABLE;
 963         } else {
 964                 key->flags = 0;
 965                 key->nr = rijndael_key_setup_enc_amd64(&(key->encr_ks.ks32[0]),
 966                     keyarr32, keybits);
 967                 key->nr = rijndael_key_setup_dec_amd64(&(key->decr_ks.ks32[0]),
 968                     keyarr32, keybits);
 969         }
 970
 971         key->type = AES_32BIT_KS;
 972 }
 973
 974 /*
 975  * Encrypt one block of data. The block is assumed to be an array
 976  * of four uint32_t values, so copy for alignment (and byte-order
 977  * reversal for little endian systems might be necessary on the
 978  * input and output byte streams.
 979  * The size of the key schedule depends on the number of rounds
 980  * (which can be computed from the size of the key), i.e. 4*(Nr + 1).
 981  *
 982  * Parameters:
 983  * rk           Key schedule, of aes_ks_t (60 32-bit integers)
 984  * Nr           Number of rounds
 985  * pt           Input block (plain text)
 986  * ct           Output block (crypto text).  Can overlap with pt
 987  * flags        Indicates whether we're on Intel AES-NI-capable hardware
 988  */
 989 static void
 990 rijndael_encrypt(const uint32_t rk[], int Nr, const uint32_t pt[4],
 991     uint32_t ct[4], int flags) {
 992         if (flags & INTEL_AES_NI_CAPABLE) {
 993                 KPREEMPT_DISABLE;
 994                 aes_encrypt_intel(rk, Nr, pt, ct);
 995                 KPREEMPT_ENABLE;
 996         } else {
 997                 aes_encrypt_amd64(rk, Nr, pt, ct);
 998         }
 999 }
1000
1001 /*
1002  * Decrypt one block of data. The block is assumed to be an array
1003  * of four uint32_t values, so copy for alignment (and byte-order
1004  * reversal for little endian systems might be necessary on the
1005  * input and output byte streams.
1006  * The size of the key schedule depends on the number of rounds
1007  * (which can be computed from the size of the key), i.e. 4*(Nr + 1).
1008  *
1009  * Parameters:
1010  * rk           Key schedule, of aes_ks_t (60 32-bit integers)
1011  * Nr           Number of rounds
1012  * ct           Input block (crypto text)
1013  * pt           Output block (plain text). Can overlap with pt
1014  * flags        Indicates whether we're on Intel AES-NI-capable hardware
1015  */
1016 static void
1017 rijndael_decrypt(const uint32_t rk[], int Nr, const uint32_t ct[4],
1018     uint32_t pt[4], int flags) {
1019         if (flags & INTEL_AES_NI_CAPABLE) {
1020                 KPREEMPT_DISABLE;
1021                 aes_decrypt_intel(rk, Nr, ct, pt);
1022                 KPREEMPT_ENABLE;
1023         } else {
1024                 aes_decrypt_amd64(rk, Nr, ct, pt);
1025         }
1026 }
1027
1028
1029 #else /* generic C implementation */
1030
1031 /*
1032  *  Expand the cipher key into the decryption key schedule.
1033  *  Return the number of rounds for the given cipher key size.
1034  *  The size of the key schedule depends on the number of rounds
1035  *  (which can be computed from the size of the key), i.e. 4*(Nr + 1).
1036  *
1037  * Parameters:
1038  * rk           AES key schedule 32-bit array to be initialized
1039  * cipherKey    User key
1040  * keyBits      AES key size (128, 192, or 256 bits)
1041  */
1042 static int
1043 rijndael_key_setup_dec(uint32_t rk[], const uint32_t cipherKey[], int keyBits)
1044 {
1045         int      Nr, i, j;
1046         uint32_t temp;
1047
1048         /* expand the cipher key: */
1049         Nr = rijndael_key_setup_enc_raw(rk, cipherKey, keyBits);
1050
1051         /* invert the order of the round keys: */
1052         for (i = 0, j = 4 * Nr; i < j; i += 4, j -= 4) {
1053                 temp = rk[i];
1054                 rk[i] = rk[j];
1055                 rk[j] = temp;
1056                 temp = rk[i + 1];
1057                 rk[i + 1] = rk[j + 1];
1058                 rk[j + 1] = temp;
1059                 temp = rk[i + 2];
1060                 rk[i + 2] = rk[j + 2];
1061                 rk[j + 2] = temp;
1062                 temp = rk[i + 3];
1063                 rk[i + 3] = rk[j + 3];
1064                 rk[j + 3] = temp;
1065         }
1066
1067         /*
1068          * apply the inverse MixColumn transform to all
1069          * round keys but the first and the last:
1070          */
1071         for (i = 1; i < Nr; i++) {
1072                 rk += 4;
1073                 rk[0] = Td0[Te4[rk[0] >> 24] & 0xff] ^
1074                     Td1[Te4[(rk[0] >> 16) & 0xff] & 0xff] ^
1075                     Td2[Te4[(rk[0] >> 8) & 0xff] & 0xff] ^
1076                     Td3[Te4[rk[0] & 0xff] & 0xff];
1077                 rk[1] = Td0[Te4[rk[1] >> 24] & 0xff] ^
1078                     Td1[Te4[(rk[1] >> 16) & 0xff] & 0xff] ^
1079                     Td2[Te4[(rk[1] >> 8) & 0xff] & 0xff] ^
1080                     Td3[Te4[rk[1] & 0xff] & 0xff];
1081                 rk[2] = Td0[Te4[rk[2] >> 24] & 0xff] ^
1082                     Td1[Te4[(rk[2] >> 16) & 0xff] & 0xff] ^
1083                     Td2[Te4[(rk[2] >> 8) & 0xff] & 0xff] ^
1084                     Td3[Te4[rk[2] & 0xff] & 0xff];
1085                 rk[3] = Td0[Te4[rk[3] >> 24] & 0xff] ^
1086                     Td1[Te4[(rk[3] >> 16) & 0xff] & 0xff] ^
1087                     Td2[Te4[(rk[3] >> 8) & 0xff] & 0xff] ^
1088                     Td3[Te4[rk[3] & 0xff] & 0xff];
1089         }
1090
1091         return (Nr);
1092 }
1093
1094
1095 /*
1096  * Expand the 32-bit AES cipher key array into the encryption and decryption
1097  * key schedules.
1098  *
1099  * Parameters:
1100  * key          AES key schedule to be initialized
1101  * keyarr32     User key
1102  * keyBits      AES key size (128, 192, or 256 bits)
1103  */
1104 static void
1105 aes_setupkeys(aes_key_t *key, const uint32_t *keyarr32, int keybits)
1106 {
1107         key->nr = rijndael_key_setup_enc(&(key->encr_ks.ks32[0]), keyarr32,
1108             keybits);
1109         key->nr = rijndael_key_setup_dec(&(key->decr_ks.ks32[0]), keyarr32,
1110             keybits);
1111         key->type = AES_32BIT_KS;
1112 }
1113
1114
1115 /*
1116  * Encrypt one block of data. The block is assumed to be an array
1117  * of four uint32_t values, so copy for alignment (and byte-order
1118  * reversal for little endian systems might be necessary on the
1119  * input and output byte streams.
1120  * The size of the key schedule depends on the number of rounds
1121  * (which can be computed from the size of the key), i.e. 4*(Nr + 1).
1122  *
1123  * Parameters:
1124  * rk   Key schedule, of aes_ks_t (60 32-bit integers)
1125  * Nr   Number of rounds
1126  * pt   Input block (plain text)
1127  * ct   Output block (crypto text).  Can overlap with pt
1128  */
1129 static void
1130 rijndael_encrypt(const uint32_t rk[], int Nr, const uint32_t pt[4],
1131     uint32_t ct[4])
1132 {
1133         uint32_t        s0, s1, s2, s3, t0, t1, t2, t3;
1134         int             r;
1135
1136         /*
1137          * map byte array block to cipher state
1138          * and add initial round key:
1139          */
1140
1141         s0 = pt[0] ^ rk[0];
1142         s1 = pt[1] ^ rk[1];
1143         s2 = pt[2] ^ rk[2];
1144         s3 = pt[3] ^ rk[3];
1145
1146         /*
1147          * Nr - 1 full rounds:
1148          */
1149
1150         r = Nr >> 1;
1151
1152         for (;;) {
1153                 t0 = Te0[s0 >> 24] ^
1154                     Te1[(s1 >> 16) & 0xff] ^
1155                     Te2[(s2 >>  8) & 0xff] ^
1156                     Te3[s3 & 0xff] ^
1157                     rk[4];
1158
1159                 t1 = Te0[s1 >> 24] ^
1160                     Te1[(s2 >> 16) & 0xff] ^
1161                     Te2[(s3 >>  8) & 0xff] ^
1162                     Te3[s0 & 0xff] ^
1163                     rk[5];
1164
1165                 t2 = Te0[s2 >> 24] ^
1166                     Te1[(s3 >> 16) & 0xff] ^
1167                     Te2[(s0 >>  8) & 0xff] ^
1168                     Te3[s1 & 0xff] ^
1169                     rk[6];
1170
1171                 t3 = Te0[s3 >> 24] ^
1172                     Te1[(s0 >> 16) & 0xff] ^
1173                     Te2[(s1 >>  8) & 0xff] ^
1174                     Te3[s2 & 0xff] ^
1175                     rk[7];
1176
1177                 rk += 8;
1178
1179                 if (--r == 0) {
1180                         break;
1181                 }
1182
1183                 s0 = Te0[t0 >> 24] ^
1184                     Te1[(t1 >> 16) & 0xff] ^
1185                     Te2[(t2 >>  8) & 0xff] ^
1186                     Te3[t3 & 0xff] ^
1187                     rk[0];
1188
1189                 s1 = Te0[t1 >> 24] ^
1190                     Te1[(t2 >> 16) & 0xff] ^
1191                     Te2[(t3 >>  8) & 0xff] ^
1192                     Te3[t0 & 0xff] ^
1193                     rk[1];
1194
1195                 s2 = Te0[t2 >> 24] ^
1196                     Te1[(t3 >> 16) & 0xff] ^
1197                     Te2[(t0 >>  8) & 0xff] ^
1198                     Te3[t1 & 0xff] ^
1199                     rk[2];
1200
1201                 s3 = Te0[t3 >> 24] ^
1202                     Te1[(t0 >> 16) & 0xff] ^
1203                     Te2[(t1 >>  8) & 0xff] ^
1204                     Te3[t2 & 0xff] ^
1205                     rk[3];
1206         }
1207
1208         /*
1209          * apply last round and
1210          * map cipher state to byte array block:
1211          */
1212
1213         s0 = (Te4[(t0 >> 24)] & 0xff000000) ^
1214             (Te4[(t1 >> 16) & 0xff] & 0x00ff0000) ^
1215             (Te4[(t2 >>  8) & 0xff] & 0x0000ff00) ^
1216             (Te4[t3 & 0xff] & 0x000000ff) ^
1217             rk[0];
1218         ct[0] = s0;
1219
1220         s1 = (Te4[(t1 >> 24)] & 0xff000000) ^
1221             (Te4[(t2 >> 16) & 0xff] & 0x00ff0000) ^
1222             (Te4[(t3 >>  8) & 0xff] & 0x0000ff00) ^
1223             (Te4[t0 & 0xff] & 0x000000ff) ^
1224             rk[1];
1225         ct[1] = s1;
1226
1227         s2 = (Te4[(t2 >> 24)] & 0xff000000) ^
1228             (Te4[(t3 >> 16) & 0xff] & 0x00ff0000) ^
1229             (Te4[(t0 >>  8) & 0xff] & 0x0000ff00) ^
1230             (Te4[t1 & 0xff] & 0x000000ff) ^
1231             rk[2];
1232         ct[2] = s2;
1233
1234         s3 = (Te4[(t3 >> 24)] & 0xff000000) ^
1235             (Te4[(t0 >> 16) & 0xff] & 0x00ff0000) ^
1236             (Te4[(t1 >>  8) & 0xff] & 0x0000ff00) ^
1237             (Te4[t2 & 0xff] & 0x000000ff) ^
1238             rk[3];
1239         ct[3] = s3;
1240 }
1241
1242
1243 /*
1244  * Decrypt one block of data. The block is assumed to be an array
1245  * of four uint32_t values, so copy for alignment (and byte-order
1246  * reversal for little endian systems might be necessary on the
1247  * input and output byte streams.
1248  * The size of the key schedule depends on the number of rounds
1249  * (which can be computed from the size of the key), i.e. 4*(Nr + 1).
1250  *
1251  * Parameters:
1252  * rk   Key schedule, of aes_ks_t (60 32-bit integers)
1253  * Nr   Number of rounds
1254  * ct   Input block (crypto text)
1255  * pt   Output block (plain text). Can overlap with pt
1256  */
1257 static void
1258 rijndael_decrypt(const uint32_t rk[], int Nr, const uint32_t ct[4],
1259     uint32_t pt[4])
1260 {
1261         uint32_t s0, s1, s2, s3, t0, t1, t2, t3;
1262         int      r;
1263
1264         /*
1265          * map byte array block to cipher state
1266          * and add initial round key:
1267          */
1268         s0 = ct[0] ^ rk[0];
1269         s1 = ct[1] ^ rk[1];
1270         s2 = ct[2] ^ rk[2];
1271         s3 = ct[3] ^ rk[3];
1272
1273         /*
1274          * Nr - 1 full rounds:
1275          */
1276
1277         r = Nr >> 1;
1278
1279         for (;;) {
1280                 t0 = Td0[s0 >> 24] ^
1281                     Td1[(s3 >> 16) & 0xff] ^
1282                     Td2[(s2 >> 8) & 0xff] ^
1283                     Td3[s1 & 0xff] ^
1284                     rk[4];
1285
1286                 t1 = Td0[s1 >> 24] ^
1287                     Td1[(s0 >> 16) & 0xff] ^
1288                     Td2[(s3 >>  8) & 0xff] ^
1289                     Td3[s2 & 0xff] ^
1290                     rk[5];
1291
1292                 t2 = Td0[s2 >> 24] ^
1293                     Td1[(s1 >> 16) & 0xff] ^
1294                     Td2[(s0 >>  8) & 0xff] ^
1295                     Td3[s3 & 0xff] ^
1296                     rk[6];
1297
1298                 t3 = Td0[s3 >> 24] ^
1299                     Td1[(s2 >> 16) & 0xff] ^
1300                     Td2[(s1 >> 8) & 0xff] ^
1301                     Td3[s0 & 0xff] ^
1302                     rk[7];
1303
1304                 rk += 8;
1305
1306                 if (--r == 0) {
1307                         break;
1308                 }
1309
1310                 s0 = Td0[t0 >> 24] ^
1311                     Td1[(t3 >> 16) & 0xff] ^
1312                     Td2[(t2 >> 8) & 0xff] ^
1313                     Td3[t1 & 0xff] ^
1314                     rk[0];
1315
1316                 s1 = Td0[t1 >> 24] ^
1317                     Td1[(t0 >> 16) & 0xff] ^
1318                     Td2[(t3 >> 8) & 0xff] ^
1319                     Td3[t2 & 0xff] ^
1320                     rk[1];
1321
1322                 s2 = Td0[t2 >> 24] ^
1323                     Td1[(t1 >> 16) & 0xff] ^
1324                     Td2[(t0 >> 8) & 0xff] ^
1325                     Td3[t3 & 0xff] ^
1326                     rk[2];
1327
1328                 s3 = Td0[t3 >> 24] ^
1329                     Td1[(t2 >> 16) & 0xff] ^
1330                     Td2[(t1 >> 8) & 0xff] ^
1331                     Td3[t0 & 0xff] ^
1332                     rk[3];
1333         }
1334
1335         /*
1336          * apply last round and
1337          * map cipher state to byte array block:
1338          */
1339
1340         s0 = (Td4[t0 >> 24] & 0xff000000) ^
1341             (Td4[(t3 >> 16) & 0xff] & 0x00ff0000) ^
1342             (Td4[(t2 >> 8) & 0xff] & 0x0000ff00) ^
1343             (Td4[t1 & 0xff] & 0x000000ff) ^
1344             rk[0];
1345         pt[0] = s0;
1346
1347         s1 = (Td4[t1 >> 24] & 0xff000000) ^
1348             (Td4[(t0 >> 16) & 0xff] & 0x00ff0000) ^
1349             (Td4[(t3 >>  8) & 0xff] & 0x0000ff00) ^
1350             (Td4[t2 & 0xff] & 0x000000ff) ^
1351             rk[1];
1352         pt[1] = s1;
1353
1354         s2 = (Td4[t2 >> 24] & 0xff000000) ^
1355             (Td4[(t1 >> 16) & 0xff] & 0x00ff0000) ^
1356             (Td4[(t0 >> 8) & 0xff] & 0x0000ff00) ^
1357             (Td4[t3 & 0xff] & 0x000000ff) ^
1358             rk[2];
1359         pt[2] = s2;
1360
1361         s3 = (Td4[t3 >> 24] & 0xff000000) ^
1362             (Td4[(t2 >> 16) & 0xff] & 0x00ff0000) ^
1363             (Td4[(t1 >>  8) & 0xff] & 0x0000ff00) ^
1364             (Td4[t0 & 0xff] & 0x000000ff) ^
1365             rk[3];
1366         pt[3] = s3;
1367 }
1368 #endif  /* __amd64 */
1369
1370
1371 /*
1372  * Initialize AES encryption and decryption key schedules.
1373  *
1374  * Parameters:
1375  * cipherKey    User key
1376  * keyBits      AES key size (128, 192, or 256 bits)
1377  * keysched     AES key schedule to be initialized, of type aes_key_t.
1378  *              Allocated by aes_alloc_keysched().
1379  */
1380 void
1381 aes_init_keysched(const uint8_t *cipherKey, uint_t keyBits, void *keysched)
1382 {
1383         aes_key_t       *newbie = keysched;
1384         uint_t          keysize, i, j;
1385         union {
1386                 uint64_t        ka64[4];
1387                 uint32_t        ka32[8];
1388                 } keyarr;
1389
1390         switch (keyBits) {
1391         case 128:
1392                 newbie->nr = 10;
1393                 break;
1394
1395         case 192:
1396                 newbie->nr = 12;
1397                 break;
1398
1399         case 256:
1400                 newbie->nr = 14;
1401                 break;
1402
1403         default:
1404                 /* should never get here */
1405                 return;
1406         }
1407         keysize = CRYPTO_BITS2BYTES(keyBits);
1408
1409         /*
1410          * For _LITTLE_ENDIAN machines (except AMD64), reverse every
1411          * 4 bytes in the key.  On _BIG_ENDIAN and AMD64, copy the key
1412          * without reversing bytes.
1413          * For AMD64, do not byte swap for aes_setupkeys().
1414          *
1415          * SPARCv8/v9 uses a key schedule array with 64-bit elements.
1416          * X86/AMD64  uses a key schedule array with 32-bit elements.
1417          */
1418 #ifndef AES_BYTE_SWAP
1419         if (IS_P2ALIGNED(cipherKey, sizeof (uint64_t))) {
1420                 for (i = 0, j = 0; j < keysize; i++, j += 8) {
1421                         /* LINTED: pointer alignment */
1422                         keyarr.ka64[i] = *((uint64_t *)&cipherKey[j]);
1423                 }
1424         } else {
1425                 bcopy(cipherKey, keyarr.ka32, keysize);
1426         }
1427
1428 #else   /* byte swap */
1429         for (i = 0, j = 0; j < keysize; i++, j += 4) {
1430                 keyarr.ka32[i] = htonl(*(uint32_t *)(void *)&cipherKey[j]);
1431         }
1432 #endif
1433
1434         aes_setupkeys(newbie, keyarr.ka32, keyBits);
1435 }
1436
1437
1438 /*
1439  * Encrypt one block using AES.
1440  * Align if needed and (for x86 32-bit only) byte-swap.
1441  *
1442  * Parameters:
1443  * ks   Key schedule, of type aes_key_t
1444  * pt   Input block (plain text)
1445  * ct   Output block (crypto text).  Can overlap with pt
1446  */
1447 int
1448 aes_encrypt_block(const void *ks, const uint8_t *pt, uint8_t *ct)
1449 {
1450         aes_key_t       *ksch = (aes_key_t *)ks;
1451
1452 #ifndef AES_BYTE_SWAP
1453         if (IS_P2ALIGNED2(pt, ct, sizeof (uint32_t))) {
1454                 /* LINTED:  pointer alignment */
1455                 AES_ENCRYPT_IMPL(&ksch->encr_ks.ks32[0], ksch->nr,
1456                     /* LINTED:  pointer alignment */
1457                     (uint32_t *)pt, (uint32_t *)ct, ksch->flags);
1458         } else {
1459 #endif
1460                 uint32_t buffer[AES_BLOCK_LEN / sizeof (uint32_t)];
1461
1462                 /* Copy input block into buffer */
1463 #ifndef AES_BYTE_SWAP
1464                 bcopy(pt, &buffer, AES_BLOCK_LEN);
1465
1466 #else   /* byte swap */
1467                 buffer[0] = htonl(*(uint32_t *)(void *)&pt[0]);
1468                 buffer[1] = htonl(*(uint32_t *)(void *)&pt[4]);
1469                 buffer[2] = htonl(*(uint32_t *)(void *)&pt[8]);
1470                 buffer[3] = htonl(*(uint32_t *)(void *)&pt[12]);
1471 #endif
1472
1473                 AES_ENCRYPT_IMPL(&ksch->encr_ks.ks32[0], ksch->nr,
1474                     buffer, buffer, ksch->flags);
1475
1476                 /* Copy result from buffer to output block */
1477 #ifndef AES_BYTE_SWAP
1478                 bcopy(&buffer, ct, AES_BLOCK_LEN);
1479         }
1480
1481 #else   /* byte swap */
1482                 *(uint32_t *)(void *)&ct[0] = htonl(buffer[0]);
1483                 *(uint32_t *)(void *)&ct[4] = htonl(buffer[1]);
1484                 *(uint32_t *)(void *)&ct[8] = htonl(buffer[2]);
1485                 *(uint32_t *)(void *)&ct[12] = htonl(buffer[3]);
1486 #endif
1487         return (CRYPTO_SUCCESS);
1488 }
1489
1490
1491 /*
1492  * Decrypt one block using AES.
1493  * Align and byte-swap if needed.
1494  *
1495  * Parameters:
1496  * ks   Key schedule, of type aes_key_t
1497  * ct   Input block (crypto text)
1498  * pt   Output block (plain text). Can overlap with pt
1499  */
1500 int
1501 aes_decrypt_block(const void *ks, const uint8_t *ct, uint8_t *pt)
1502 {
1503         aes_key_t       *ksch = (aes_key_t *)ks;
1504
1505 #ifndef AES_BYTE_SWAP
1506         if (IS_P2ALIGNED2(ct, pt, sizeof (uint32_t))) {
1507                 /* LINTED:  pointer alignment */
1508                 AES_DECRYPT_IMPL(&ksch->decr_ks.ks32[0], ksch->nr,
1509                     /* LINTED:  pointer alignment */
1510                     (uint32_t *)ct, (uint32_t *)pt, ksch->flags);
1511         } else {
1512 #endif
1513                 uint32_t buffer[AES_BLOCK_LEN / sizeof (uint32_t)];
1514
1515                 /* Copy input block into buffer */
1516 #ifndef AES_BYTE_SWAP
1517                 bcopy(ct, &buffer, AES_BLOCK_LEN);
1518
1519 #else   /* byte swap */
1520                 buffer[0] = htonl(*(uint32_t *)(void *)&ct[0]);
1521                 buffer[1] = htonl(*(uint32_t *)(void *)&ct[4]);
1522                 buffer[2] = htonl(*(uint32_t *)(void *)&ct[8]);
1523                 buffer[3] = htonl(*(uint32_t *)(void *)&ct[12]);
1524 #endif
1525
1526                 AES_DECRYPT_IMPL(&ksch->decr_ks.ks32[0], ksch->nr,
1527                     buffer, buffer, ksch->flags);
1528
1529                 /* Copy result from buffer to output block */
1530 #ifndef AES_BYTE_SWAP
1531                 bcopy(&buffer, pt, AES_BLOCK_LEN);
1532         }
1533
1534 #else   /* byte swap */
1535         *(uint32_t *)(void *)&pt[0] = htonl(buffer[0]);
1536         *(uint32_t *)(void *)&pt[4] = htonl(buffer[1]);
1537         *(uint32_t *)(void *)&pt[8] = htonl(buffer[2]);
1538         *(uint32_t *)(void *)&pt[12] = htonl(buffer[3]);
1539 #endif
1540
1541         return (CRYPTO_SUCCESS);
1542 }
1543
1544
1545 /*
1546  * Allocate key schedule for AES.
1547  *
1548  * Return the pointer and set size to the number of bytes allocated.
1549  * Memory allocated must be freed by the caller when done.
1550  *
1551  * Parameters:
1552  * size         Size of key schedule allocated, in bytes
1553  * kmflag       Flag passed to kmem_alloc(9F); ignored in userland.
1554  */
1555 /* ARGSUSED */
1556 void *
1557 aes_alloc_keysched(size_t *size, int kmflag)
1558 {
1559         aes_key_t *keysched;
1560
1561         keysched = (aes_key_t *)kmem_alloc(sizeof (aes_key_t), kmflag);
1562         if (keysched != NULL) {
1563                 *size = sizeof (aes_key_t);
1564                 return (keysched);
1565         }
1566         return (NULL);
1567 }
1568
1569
1570 #ifdef __amd64
1571
1572 #define INTEL_AESNI_FLAG (1 << 25)
1573
1574 /*
1575  * Return 1 if executing on Intel with AES-NI instructions,
1576  * otherwise 0 (i.e., Intel without AES-NI or AMD64).
1577  * Cache the result, as the CPU can't change.
1578  */
1579 static int
1580 intel_aes_instructions_present(void)
1581 {
1582         static int cached_result = -1;
1583         unsigned eax, ebx, ecx, edx;
1584         unsigned func, subfunc;
1585
1586         if (cached_result == -1) { /* first time */
1587                 /* check for an intel cpu */
1588                 func = 0;
1589                 subfunc = 0;
1590
1591                 __asm__ __volatile__(
1592                     "cpuid"
1593                     : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx)
1594                     : "a"(func), "c"(subfunc));
1595
1596                 if (memcmp((char *) (&ebx), "Genu", 4) == 0 &&
1597                     memcmp((char *) (&edx), "ineI", 4) == 0 &&
1598                         memcmp((char *) (&ecx), "ntel", 4) == 0) {
1599
1600                         func = 1;
1601                         subfunc = 0;
1602
1603                         /* check for aes-ni instruction set */
1604                         __asm__ __volatile__(
1605                                 "cpuid"
1606                                 : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx)
1607                                 : "a"(func), "c"(subfunc));
1608
1609                         cached_result = !!(ecx & INTEL_AESNI_FLAG);
1610                 } else {
1611                         cached_result = 0;
1612                 }
1613         }
1614
1615         return (cached_result);
1616 }
1617
1618 #endif  /* __amd64 */