Team Fortress 2 Source Code as on 22/4/2020
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

717 lines
24 KiB

  1. // sosemanuk.cpp - written and placed in the public domain by Wei Dai
  2. // use "cl /EP /P /DCRYPTOPP_GENERATE_X64_MASM sosemanuk.cpp" to generate MASM code
  3. #include "pch.h"
  4. #include "config.h"
  5. #if CRYPTOPP_MSC_VERSION
  6. # pragma warning(disable: 4702 4731)
  7. #endif
  8. #ifndef CRYPTOPP_GENERATE_X64_MASM
  9. #include "sosemanuk.h"
  10. #include "serpentp.h"
  11. #include "secblock.h"
  12. #include "misc.h"
  13. #include "cpu.h"
  14. NAMESPACE_BEGIN(CryptoPP)
  15. void SosemanukPolicy::CipherSetKey(const NameValuePairs &params, const byte *userKey, size_t keylen)
  16. {
  17. CRYPTOPP_UNUSED(params);
  18. Serpent_KeySchedule(m_key, 24, userKey, keylen);
  19. }
  20. void SosemanukPolicy::CipherResynchronize(byte *keystreamBuffer, const byte *iv, size_t length)
  21. {
  22. CRYPTOPP_UNUSED(keystreamBuffer), CRYPTOPP_UNUSED(iv), CRYPTOPP_UNUSED(length);
  23. assert(length==16);
  24. word32 a, b, c, d, e;
  25. typedef BlockGetAndPut<word32, LittleEndian> Block;
  26. Block::Get(iv)(a)(b)(c)(d);
  27. const word32 *k = m_key;
  28. unsigned int i=1;
  29. do
  30. {
  31. beforeS0(KX); beforeS0(S0); afterS0(LT);
  32. afterS0(KX); afterS0(S1); afterS1(LT);
  33. if (i == 3) // after 18th round
  34. {
  35. m_state[4] = b;
  36. m_state[5] = e;
  37. m_state[10] = c;
  38. m_state[11] = a;
  39. }
  40. afterS1(KX); afterS1(S2); afterS2(LT);
  41. afterS2(KX); afterS2(S3); afterS3(LT);
  42. if (i == 2) // after 12th round
  43. {
  44. m_state[6] = c;
  45. m_state[7] = d;
  46. m_state[8] = b;
  47. m_state[9] = e;
  48. }
  49. afterS3(KX); afterS3(S4); afterS4(LT);
  50. afterS4(KX); afterS4(S5); afterS5(LT);
  51. afterS5(KX); afterS5(S6); afterS6(LT);
  52. afterS6(KX); afterS6(S7); afterS7(LT);
  53. if (i == 3)
  54. break;
  55. ++i;
  56. c = b;
  57. b = e;
  58. e = d;
  59. d = a;
  60. a = e;
  61. k += 32;
  62. }
  63. while (true);
  64. afterS7(KX);
  65. m_state[0] = a;
  66. m_state[1] = b;
  67. m_state[2] = e;
  68. m_state[3] = d;
  69. #define XMUX(c, x, y) (x ^ (y & (0 - (c & 1))))
  70. m_state[11] += XMUX(m_state[10], m_state[1], m_state[8]);
  71. m_state[10] = rotlFixed(m_state[10] * 0x54655307, 7);
  72. }
  73. extern "C" {
  74. word32 s_sosemanukMulTables[512] = {
  75. #if (CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X64) && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  76. 0x00000000, 0xE19FCF12, 0x6B973724, 0x8A08F836,
  77. 0xD6876E48, 0x3718A15A, 0xBD10596C, 0x5C8F967E,
  78. 0x05A7DC90, 0xE4381382, 0x6E30EBB4, 0x8FAF24A6,
  79. 0xD320B2D8, 0x32BF7DCA, 0xB8B785FC, 0x59284AEE,
  80. 0x0AE71189, 0xEB78DE9B, 0x617026AD, 0x80EFE9BF,
  81. 0xDC607FC1, 0x3DFFB0D3, 0xB7F748E5, 0x566887F7,
  82. 0x0F40CD19, 0xEEDF020B, 0x64D7FA3D, 0x8548352F,
  83. 0xD9C7A351, 0x38586C43, 0xB2509475, 0x53CF5B67,
  84. 0x146722BB, 0xF5F8EDA9, 0x7FF0159F, 0x9E6FDA8D,
  85. 0xC2E04CF3, 0x237F83E1, 0xA9777BD7, 0x48E8B4C5,
  86. 0x11C0FE2B, 0xF05F3139, 0x7A57C90F, 0x9BC8061D,
  87. 0xC7479063, 0x26D85F71, 0xACD0A747, 0x4D4F6855,
  88. 0x1E803332, 0xFF1FFC20, 0x75170416, 0x9488CB04,
  89. 0xC8075D7A, 0x29989268, 0xA3906A5E, 0x420FA54C,
  90. 0x1B27EFA2, 0xFAB820B0, 0x70B0D886, 0x912F1794,
  91. 0xCDA081EA, 0x2C3F4EF8, 0xA637B6CE, 0x47A879DC,
  92. 0x28CE44DF, 0xC9518BCD, 0x435973FB, 0xA2C6BCE9,
  93. 0xFE492A97, 0x1FD6E585, 0x95DE1DB3, 0x7441D2A1,
  94. 0x2D69984F, 0xCCF6575D, 0x46FEAF6B, 0xA7616079,
  95. 0xFBEEF607, 0x1A713915, 0x9079C123, 0x71E60E31,
  96. 0x22295556, 0xC3B69A44, 0x49BE6272, 0xA821AD60,
  97. 0xF4AE3B1E, 0x1531F40C, 0x9F390C3A, 0x7EA6C328,
  98. 0x278E89C6, 0xC61146D4, 0x4C19BEE2, 0xAD8671F0,
  99. 0xF109E78E, 0x1096289C, 0x9A9ED0AA, 0x7B011FB8,
  100. 0x3CA96664, 0xDD36A976, 0x573E5140, 0xB6A19E52,
  101. 0xEA2E082C, 0x0BB1C73E, 0x81B93F08, 0x6026F01A,
  102. 0x390EBAF4, 0xD89175E6, 0x52998DD0, 0xB30642C2,
  103. 0xEF89D4BC, 0x0E161BAE, 0x841EE398, 0x65812C8A,
  104. 0x364E77ED, 0xD7D1B8FF, 0x5DD940C9, 0xBC468FDB,
  105. 0xE0C919A5, 0x0156D6B7, 0x8B5E2E81, 0x6AC1E193,
  106. 0x33E9AB7D, 0xD276646F, 0x587E9C59, 0xB9E1534B,
  107. 0xE56EC535, 0x04F10A27, 0x8EF9F211, 0x6F663D03,
  108. 0x50358817, 0xB1AA4705, 0x3BA2BF33, 0xDA3D7021,
  109. 0x86B2E65F, 0x672D294D, 0xED25D17B, 0x0CBA1E69,
  110. 0x55925487, 0xB40D9B95, 0x3E0563A3, 0xDF9AACB1,
  111. 0x83153ACF, 0x628AF5DD, 0xE8820DEB, 0x091DC2F9,
  112. 0x5AD2999E, 0xBB4D568C, 0x3145AEBA, 0xD0DA61A8,
  113. 0x8C55F7D6, 0x6DCA38C4, 0xE7C2C0F2, 0x065D0FE0,
  114. 0x5F75450E, 0xBEEA8A1C, 0x34E2722A, 0xD57DBD38,
  115. 0x89F22B46, 0x686DE454, 0xE2651C62, 0x03FAD370,
  116. 0x4452AAAC, 0xA5CD65BE, 0x2FC59D88, 0xCE5A529A,
  117. 0x92D5C4E4, 0x734A0BF6, 0xF942F3C0, 0x18DD3CD2,
  118. 0x41F5763C, 0xA06AB92E, 0x2A624118, 0xCBFD8E0A,
  119. 0x97721874, 0x76EDD766, 0xFCE52F50, 0x1D7AE042,
  120. 0x4EB5BB25, 0xAF2A7437, 0x25228C01, 0xC4BD4313,
  121. 0x9832D56D, 0x79AD1A7F, 0xF3A5E249, 0x123A2D5B,
  122. 0x4B1267B5, 0xAA8DA8A7, 0x20855091, 0xC11A9F83,
  123. 0x9D9509FD, 0x7C0AC6EF, 0xF6023ED9, 0x179DF1CB,
  124. 0x78FBCCC8, 0x996403DA, 0x136CFBEC, 0xF2F334FE,
  125. 0xAE7CA280, 0x4FE36D92, 0xC5EB95A4, 0x24745AB6,
  126. 0x7D5C1058, 0x9CC3DF4A, 0x16CB277C, 0xF754E86E,
  127. 0xABDB7E10, 0x4A44B102, 0xC04C4934, 0x21D38626,
  128. 0x721CDD41, 0x93831253, 0x198BEA65, 0xF8142577,
  129. 0xA49BB309, 0x45047C1B, 0xCF0C842D, 0x2E934B3F,
  130. 0x77BB01D1, 0x9624CEC3, 0x1C2C36F5, 0xFDB3F9E7,
  131. 0xA13C6F99, 0x40A3A08B, 0xCAAB58BD, 0x2B3497AF,
  132. 0x6C9CEE73, 0x8D032161, 0x070BD957, 0xE6941645,
  133. 0xBA1B803B, 0x5B844F29, 0xD18CB71F, 0x3013780D,
  134. 0x693B32E3, 0x88A4FDF1, 0x02AC05C7, 0xE333CAD5,
  135. 0xBFBC5CAB, 0x5E2393B9, 0xD42B6B8F, 0x35B4A49D,
  136. 0x667BFFFA, 0x87E430E8, 0x0DECC8DE, 0xEC7307CC,
  137. 0xB0FC91B2, 0x51635EA0, 0xDB6BA696, 0x3AF46984,
  138. 0x63DC236A, 0x8243EC78, 0x084B144E, 0xE9D4DB5C,
  139. 0xB55B4D22, 0x54C48230, 0xDECC7A06, 0x3F53B514,
  140. #else
  141. 0x00000000, 0xE19FCF13, 0x6B973726, 0x8A08F835,
  142. 0xD6876E4C, 0x3718A15F, 0xBD10596A, 0x5C8F9679,
  143. 0x05A7DC98, 0xE438138B, 0x6E30EBBE, 0x8FAF24AD,
  144. 0xD320B2D4, 0x32BF7DC7, 0xB8B785F2, 0x59284AE1,
  145. 0x0AE71199, 0xEB78DE8A, 0x617026BF, 0x80EFE9AC,
  146. 0xDC607FD5, 0x3DFFB0C6, 0xB7F748F3, 0x566887E0,
  147. 0x0F40CD01, 0xEEDF0212, 0x64D7FA27, 0x85483534,
  148. 0xD9C7A34D, 0x38586C5E, 0xB250946B, 0x53CF5B78,
  149. 0x1467229B, 0xF5F8ED88, 0x7FF015BD, 0x9E6FDAAE,
  150. 0xC2E04CD7, 0x237F83C4, 0xA9777BF1, 0x48E8B4E2,
  151. 0x11C0FE03, 0xF05F3110, 0x7A57C925, 0x9BC80636,
  152. 0xC747904F, 0x26D85F5C, 0xACD0A769, 0x4D4F687A,
  153. 0x1E803302, 0xFF1FFC11, 0x75170424, 0x9488CB37,
  154. 0xC8075D4E, 0x2998925D, 0xA3906A68, 0x420FA57B,
  155. 0x1B27EF9A, 0xFAB82089, 0x70B0D8BC, 0x912F17AF,
  156. 0xCDA081D6, 0x2C3F4EC5, 0xA637B6F0, 0x47A879E3,
  157. 0x28CE449F, 0xC9518B8C, 0x435973B9, 0xA2C6BCAA,
  158. 0xFE492AD3, 0x1FD6E5C0, 0x95DE1DF5, 0x7441D2E6,
  159. 0x2D699807, 0xCCF65714, 0x46FEAF21, 0xA7616032,
  160. 0xFBEEF64B, 0x1A713958, 0x9079C16D, 0x71E60E7E,
  161. 0x22295506, 0xC3B69A15, 0x49BE6220, 0xA821AD33,
  162. 0xF4AE3B4A, 0x1531F459, 0x9F390C6C, 0x7EA6C37F,
  163. 0x278E899E, 0xC611468D, 0x4C19BEB8, 0xAD8671AB,
  164. 0xF109E7D2, 0x109628C1, 0x9A9ED0F4, 0x7B011FE7,
  165. 0x3CA96604, 0xDD36A917, 0x573E5122, 0xB6A19E31,
  166. 0xEA2E0848, 0x0BB1C75B, 0x81B93F6E, 0x6026F07D,
  167. 0x390EBA9C, 0xD891758F, 0x52998DBA, 0xB30642A9,
  168. 0xEF89D4D0, 0x0E161BC3, 0x841EE3F6, 0x65812CE5,
  169. 0x364E779D, 0xD7D1B88E, 0x5DD940BB, 0xBC468FA8,
  170. 0xE0C919D1, 0x0156D6C2, 0x8B5E2EF7, 0x6AC1E1E4,
  171. 0x33E9AB05, 0xD2766416, 0x587E9C23, 0xB9E15330,
  172. 0xE56EC549, 0x04F10A5A, 0x8EF9F26F, 0x6F663D7C,
  173. 0x50358897, 0xB1AA4784, 0x3BA2BFB1, 0xDA3D70A2,
  174. 0x86B2E6DB, 0x672D29C8, 0xED25D1FD, 0x0CBA1EEE,
  175. 0x5592540F, 0xB40D9B1C, 0x3E056329, 0xDF9AAC3A,
  176. 0x83153A43, 0x628AF550, 0xE8820D65, 0x091DC276,
  177. 0x5AD2990E, 0xBB4D561D, 0x3145AE28, 0xD0DA613B,
  178. 0x8C55F742, 0x6DCA3851, 0xE7C2C064, 0x065D0F77,
  179. 0x5F754596, 0xBEEA8A85, 0x34E272B0, 0xD57DBDA3,
  180. 0x89F22BDA, 0x686DE4C9, 0xE2651CFC, 0x03FAD3EF,
  181. 0x4452AA0C, 0xA5CD651F, 0x2FC59D2A, 0xCE5A5239,
  182. 0x92D5C440, 0x734A0B53, 0xF942F366, 0x18DD3C75,
  183. 0x41F57694, 0xA06AB987, 0x2A6241B2, 0xCBFD8EA1,
  184. 0x977218D8, 0x76EDD7CB, 0xFCE52FFE, 0x1D7AE0ED,
  185. 0x4EB5BB95, 0xAF2A7486, 0x25228CB3, 0xC4BD43A0,
  186. 0x9832D5D9, 0x79AD1ACA, 0xF3A5E2FF, 0x123A2DEC,
  187. 0x4B12670D, 0xAA8DA81E, 0x2085502B, 0xC11A9F38,
  188. 0x9D950941, 0x7C0AC652, 0xF6023E67, 0x179DF174,
  189. 0x78FBCC08, 0x9964031B, 0x136CFB2E, 0xF2F3343D,
  190. 0xAE7CA244, 0x4FE36D57, 0xC5EB9562, 0x24745A71,
  191. 0x7D5C1090, 0x9CC3DF83, 0x16CB27B6, 0xF754E8A5,
  192. 0xABDB7EDC, 0x4A44B1CF, 0xC04C49FA, 0x21D386E9,
  193. 0x721CDD91, 0x93831282, 0x198BEAB7, 0xF81425A4,
  194. 0xA49BB3DD, 0x45047CCE, 0xCF0C84FB, 0x2E934BE8,
  195. 0x77BB0109, 0x9624CE1A, 0x1C2C362F, 0xFDB3F93C,
  196. 0xA13C6F45, 0x40A3A056, 0xCAAB5863, 0x2B349770,
  197. 0x6C9CEE93, 0x8D032180, 0x070BD9B5, 0xE69416A6,
  198. 0xBA1B80DF, 0x5B844FCC, 0xD18CB7F9, 0x301378EA,
  199. 0x693B320B, 0x88A4FD18, 0x02AC052D, 0xE333CA3E,
  200. 0xBFBC5C47, 0x5E239354, 0xD42B6B61, 0x35B4A472,
  201. 0x667BFF0A, 0x87E43019, 0x0DECC82C, 0xEC73073F,
  202. 0xB0FC9146, 0x51635E55, 0xDB6BA660, 0x3AF46973,
  203. 0x63DC2392, 0x8243EC81, 0x084B14B4, 0xE9D4DBA7,
  204. 0xB55B4DDE, 0x54C482CD, 0xDECC7AF8, 0x3F53B5EB,
  205. #endif
  206. 0x00000000, 0x180F40CD, 0x301E8033, 0x2811C0FE,
  207. 0x603CA966, 0x7833E9AB, 0x50222955, 0x482D6998,
  208. 0xC078FBCC, 0xD877BB01, 0xF0667BFF, 0xE8693B32,
  209. 0xA04452AA, 0xB84B1267, 0x905AD299, 0x88559254,
  210. 0x29F05F31, 0x31FF1FFC, 0x19EEDF02, 0x01E19FCF,
  211. 0x49CCF657, 0x51C3B69A, 0x79D27664, 0x61DD36A9,
  212. 0xE988A4FD, 0xF187E430, 0xD99624CE, 0xC1996403,
  213. 0x89B40D9B, 0x91BB4D56, 0xB9AA8DA8, 0xA1A5CD65,
  214. 0x5249BE62, 0x4A46FEAF, 0x62573E51, 0x7A587E9C,
  215. 0x32751704, 0x2A7A57C9, 0x026B9737, 0x1A64D7FA,
  216. 0x923145AE, 0x8A3E0563, 0xA22FC59D, 0xBA208550,
  217. 0xF20DECC8, 0xEA02AC05, 0xC2136CFB, 0xDA1C2C36,
  218. 0x7BB9E153, 0x63B6A19E, 0x4BA76160, 0x53A821AD,
  219. 0x1B854835, 0x038A08F8, 0x2B9BC806, 0x339488CB,
  220. 0xBBC11A9F, 0xA3CE5A52, 0x8BDF9AAC, 0x93D0DA61,
  221. 0xDBFDB3F9, 0xC3F2F334, 0xEBE333CA, 0xF3EC7307,
  222. 0xA492D5C4, 0xBC9D9509, 0x948C55F7, 0x8C83153A,
  223. 0xC4AE7CA2, 0xDCA13C6F, 0xF4B0FC91, 0xECBFBC5C,
  224. 0x64EA2E08, 0x7CE56EC5, 0x54F4AE3B, 0x4CFBEEF6,
  225. 0x04D6876E, 0x1CD9C7A3, 0x34C8075D, 0x2CC74790,
  226. 0x8D628AF5, 0x956DCA38, 0xBD7C0AC6, 0xA5734A0B,
  227. 0xED5E2393, 0xF551635E, 0xDD40A3A0, 0xC54FE36D,
  228. 0x4D1A7139, 0x551531F4, 0x7D04F10A, 0x650BB1C7,
  229. 0x2D26D85F, 0x35299892, 0x1D38586C, 0x053718A1,
  230. 0xF6DB6BA6, 0xEED42B6B, 0xC6C5EB95, 0xDECAAB58,
  231. 0x96E7C2C0, 0x8EE8820D, 0xA6F942F3, 0xBEF6023E,
  232. 0x36A3906A, 0x2EACD0A7, 0x06BD1059, 0x1EB25094,
  233. 0x569F390C, 0x4E9079C1, 0x6681B93F, 0x7E8EF9F2,
  234. 0xDF2B3497, 0xC724745A, 0xEF35B4A4, 0xF73AF469,
  235. 0xBF179DF1, 0xA718DD3C, 0x8F091DC2, 0x97065D0F,
  236. 0x1F53CF5B, 0x075C8F96, 0x2F4D4F68, 0x37420FA5,
  237. 0x7F6F663D, 0x676026F0, 0x4F71E60E, 0x577EA6C3,
  238. 0xE18D0321, 0xF98243EC, 0xD1938312, 0xC99CC3DF,
  239. 0x81B1AA47, 0x99BEEA8A, 0xB1AF2A74, 0xA9A06AB9,
  240. 0x21F5F8ED, 0x39FAB820, 0x11EB78DE, 0x09E43813,
  241. 0x41C9518B, 0x59C61146, 0x71D7D1B8, 0x69D89175,
  242. 0xC87D5C10, 0xD0721CDD, 0xF863DC23, 0xE06C9CEE,
  243. 0xA841F576, 0xB04EB5BB, 0x985F7545, 0x80503588,
  244. 0x0805A7DC, 0x100AE711, 0x381B27EF, 0x20146722,
  245. 0x68390EBA, 0x70364E77, 0x58278E89, 0x4028CE44,
  246. 0xB3C4BD43, 0xABCBFD8E, 0x83DA3D70, 0x9BD57DBD,
  247. 0xD3F81425, 0xCBF754E8, 0xE3E69416, 0xFBE9D4DB,
  248. 0x73BC468F, 0x6BB30642, 0x43A2C6BC, 0x5BAD8671,
  249. 0x1380EFE9, 0x0B8FAF24, 0x239E6FDA, 0x3B912F17,
  250. 0x9A34E272, 0x823BA2BF, 0xAA2A6241, 0xB225228C,
  251. 0xFA084B14, 0xE2070BD9, 0xCA16CB27, 0xD2198BEA,
  252. 0x5A4C19BE, 0x42435973, 0x6A52998D, 0x725DD940,
  253. 0x3A70B0D8, 0x227FF015, 0x0A6E30EB, 0x12617026,
  254. 0x451FD6E5, 0x5D109628, 0x750156D6, 0x6D0E161B,
  255. 0x25237F83, 0x3D2C3F4E, 0x153DFFB0, 0x0D32BF7D,
  256. 0x85672D29, 0x9D686DE4, 0xB579AD1A, 0xAD76EDD7,
  257. 0xE55B844F, 0xFD54C482, 0xD545047C, 0xCD4A44B1,
  258. 0x6CEF89D4, 0x74E0C919, 0x5CF109E7, 0x44FE492A,
  259. 0x0CD320B2, 0x14DC607F, 0x3CCDA081, 0x24C2E04C,
  260. 0xAC977218, 0xB49832D5, 0x9C89F22B, 0x8486B2E6,
  261. 0xCCABDB7E, 0xD4A49BB3, 0xFCB55B4D, 0xE4BA1B80,
  262. 0x17566887, 0x0F59284A, 0x2748E8B4, 0x3F47A879,
  263. 0x776AC1E1, 0x6F65812C, 0x477441D2, 0x5F7B011F,
  264. 0xD72E934B, 0xCF21D386, 0xE7301378, 0xFF3F53B5,
  265. 0xB7123A2D, 0xAF1D7AE0, 0x870CBA1E, 0x9F03FAD3,
  266. 0x3EA637B6, 0x26A9777B, 0x0EB8B785, 0x16B7F748,
  267. 0x5E9A9ED0, 0x4695DE1D, 0x6E841EE3, 0x768B5E2E,
  268. 0xFEDECC7A, 0xE6D18CB7, 0xCEC04C49, 0xD6CF0C84,
  269. 0x9EE2651C, 0x86ED25D1, 0xAEFCE52F, 0xB6F3A5E2
  270. };
  271. }
  272. #if (CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X64) && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  273. unsigned int SosemanukPolicy::GetAlignment() const
  274. {
  275. #if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  276. #ifdef __INTEL_COMPILER
  277. if (HasSSE2() && !IsP4()) // Intel compiler produces faster code for this algorithm on the P4
  278. #else
  279. if (HasSSE2())
  280. #endif
  281. return 16;
  282. else
  283. #endif
  284. return GetAlignmentOf<word32>();
  285. }
  286. unsigned int SosemanukPolicy::GetOptimalBlockSize() const
  287. {
  288. #if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  289. #ifdef __INTEL_COMPILER
  290. if (HasSSE2() && !IsP4()) // Intel compiler produces faster code for this algorithm on the P4
  291. #else
  292. if (HasSSE2())
  293. #endif
  294. return 4*BYTES_PER_ITERATION;
  295. else
  296. #endif
  297. return BYTES_PER_ITERATION;
  298. }
  299. #endif
  300. #ifdef CRYPTOPP_X64_MASM_AVAILABLE
  301. extern "C" {
  302. void Sosemanuk_OperateKeystream(size_t iterationCount, const byte *input, byte *output, word32 *state);
  303. }
  304. #endif
  305. void SosemanukPolicy::OperateKeystream(KeystreamOperation operation, byte *output, const byte *input, size_t iterationCount)
  306. {
  307. #endif // #ifdef CRYPTOPP_GENERATE_X64_MASM
  308. #ifdef CRYPTOPP_X64_MASM_AVAILABLE
  309. Sosemanuk_OperateKeystream(iterationCount, input, output, m_state.data());
  310. return;
  311. #endif
  312. #if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  313. #ifdef CRYPTOPP_GENERATE_X64_MASM
  314. ALIGN 8
  315. Sosemanuk_OperateKeystream PROC FRAME
  316. rex_push_reg rsi
  317. push_reg rdi
  318. alloc_stack(80*4*2+12*4+8*WORD_SZ + 2*16+8)
  319. save_xmm128 xmm6, 02f0h
  320. save_xmm128 xmm7, 0300h
  321. .endprolog
  322. mov rdi, r8
  323. mov rax, r9
  324. #else
  325. #ifdef __INTEL_COMPILER
  326. if (HasSSE2() && !IsP4()) // Intel compiler produces faster code for this algorithm on the P4
  327. #else
  328. if (HasSSE2())
  329. #endif
  330. {
  331. #ifdef __GNUC__
  332. #if CRYPTOPP_BOOL_X64
  333. FixedSizeAlignedSecBlock<byte, 80*4*2+12*4+8*WORD_SZ> workspace;
  334. #endif
  335. __asm__ __volatile__
  336. (
  337. INTEL_NOPREFIX
  338. AS_PUSH_IF86( bx)
  339. #else
  340. word32 *state = m_state;
  341. AS2( mov WORD_REG(ax), state)
  342. AS2( mov WORD_REG(di), output)
  343. AS2( mov WORD_REG(dx), input)
  344. AS2( mov WORD_REG(cx), iterationCount)
  345. #endif
  346. #endif // #ifdef CRYPTOPP_GENERATE_X64_MASM
  347. #if defined(__GNUC__) && CRYPTOPP_BOOL_X64
  348. #define SSE2_workspace %5
  349. #else
  350. #define SSE2_workspace WORD_REG(sp)
  351. #endif
  352. #define SSE2_output WORD_PTR [SSE2_workspace+1*WORD_SZ]
  353. #define SSE2_input WORD_PTR [SSE2_workspace+2*WORD_SZ]
  354. #define SSE2_wordsLeft WORD_PTR [SSE2_workspace+3*WORD_SZ]
  355. #define SSE2_diEnd WORD_PTR [SSE2_workspace+4*WORD_SZ]
  356. #define SSE2_pMulTables WORD_PTR [SSE2_workspace+5*WORD_SZ]
  357. #define SSE2_state WORD_PTR [SSE2_workspace+6*WORD_SZ]
  358. #define SSE2_wordsLeft2 WORD_PTR [SSE2_workspace+7*WORD_SZ]
  359. #define SSE2_stateCopy SSE2_workspace + 8*WORD_SZ
  360. #define SSE2_uvStart SSE2_stateCopy + 12*4
  361. #if (CRYPTOPP_BOOL_X86) && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  362. AS_PUSH_IF86( bp)
  363. AS2( mov AS_REG_6, esp)
  364. AS2( and esp, -16)
  365. AS2( sub esp, 80*4*2+12*4+8*WORD_SZ) // 80 v's, 80 u's, 12 state, 8 locals
  366. AS2( mov [esp], AS_REG_6)
  367. #endif
  368. AS2( mov SSE2_output, WORD_REG(di))
  369. AS2( mov SSE2_input, WORD_REG(dx))
  370. AS2( mov SSE2_state, WORD_REG(ax))
  371. #ifndef _MSC_VER
  372. AS2( mov SSE2_pMulTables, WORD_REG(si))
  373. #endif
  374. AS2( lea WORD_REG(cx), [4*WORD_REG(cx)+WORD_REG(cx)])
  375. AS2( lea WORD_REG(si), [4*WORD_REG(cx)])
  376. AS2( mov SSE2_wordsLeft, WORD_REG(si))
  377. AS2( movdqa xmm0, [WORD_REG(ax)+0*16]) // copy state to stack to save a register
  378. AS2( movdqa [SSE2_stateCopy+0*16], xmm0)
  379. AS2( movdqa xmm0, [WORD_REG(ax)+1*16])
  380. AS2( movdqa [SSE2_stateCopy+1*16], xmm0)
  381. AS2( movq xmm0, QWORD PTR [WORD_REG(ax)+2*16])
  382. AS2( movq QWORD PTR [SSE2_stateCopy+2*16], xmm0)
  383. AS2( psrlq xmm0, 32)
  384. AS2( movd AS_REG_6d, xmm0) // s(9)
  385. AS2( mov ecx, [WORD_REG(ax)+10*4])
  386. AS2( mov edx, [WORD_REG(ax)+11*4])
  387. AS2( pcmpeqb xmm7, xmm7) // all ones
  388. #define s(i) SSE2_stateCopy + ASM_MOD(i,10)*4
  389. #define u(j) WORD_REG(di) + (ASM_MOD(j,4)*20 + (j/4)) * 4
  390. #define v(j) WORD_REG(di) + (ASM_MOD(j,4)*20 + (j/4)) * 4 + 80*4
  391. #define R10 ecx
  392. #define R11 edx
  393. #define R20 edx
  394. #define R21 ecx
  395. // workaround bug in GAS 2.15
  396. #define R20r WORD_REG(dx)
  397. #define R21r WORD_REG(cx)
  398. #define SSE2_STEP(i, j) \
  399. AS2( mov eax, [s(i+0)])\
  400. AS2( mov [v(i)], eax)\
  401. AS2( rol eax, 8)\
  402. AS2( lea AS_REG_7, [AS_REG_6 + R2##j##r])\
  403. AS2( xor AS_REG_7d, R1##j)\
  404. AS2( mov [u(i)], AS_REG_7d)\
  405. AS2( mov AS_REG_7d, 1)\
  406. AS2( and AS_REG_7d, R2##j)\
  407. AS1( neg AS_REG_7d)\
  408. AS2( and AS_REG_7d, AS_REG_6d)\
  409. AS2( xor AS_REG_6d, eax)\
  410. AS2( movzx eax, al)\
  411. AS2( xor AS_REG_6d, [WORD_REG(si)+WORD_REG(ax)*4])\
  412. AS2( mov eax, [s(i+3)])\
  413. AS2( xor AS_REG_7d, [s(i+2)])\
  414. AS2( add R1##j, AS_REG_7d)\
  415. AS2( movzx AS_REG_7d, al)\
  416. AS2( shr eax, 8)\
  417. AS2( xor AS_REG_6d, [WORD_REG(si)+1024+AS_REG_7*4])\
  418. AS2( xor AS_REG_6d, eax)\
  419. AS2( imul R2##j, AS_HEX(54655307))\
  420. AS2( rol R2##j, 7)\
  421. AS2( mov [s(i+0)], AS_REG_6d)\
  422. ASL(2) // outer loop, each iteration of this processes 80 words
  423. AS2( lea WORD_REG(di), [SSE2_uvStart]) // start of v and u
  424. AS2( mov WORD_REG(ax), 80)
  425. AS2( cmp WORD_REG(si), 80)
  426. AS2( cmovg WORD_REG(si), WORD_REG(ax))
  427. AS2( mov SSE2_wordsLeft2, WORD_REG(si))
  428. AS2( lea WORD_REG(si), [WORD_REG(di)+WORD_REG(si)]) // use to end first inner loop
  429. AS2( mov SSE2_diEnd, WORD_REG(si))
  430. #ifdef _MSC_VER
  431. AS2( lea WORD_REG(si), s_sosemanukMulTables)
  432. #else
  433. AS2( mov WORD_REG(si), SSE2_pMulTables)
  434. #endif
  435. ASL(0) // first inner loop, 20 words each, 4 iterations
  436. SSE2_STEP(0, 0)
  437. SSE2_STEP(1, 1)
  438. SSE2_STEP(2, 0)
  439. SSE2_STEP(3, 1)
  440. SSE2_STEP(4, 0)
  441. SSE2_STEP(5, 1)
  442. SSE2_STEP(6, 0)
  443. SSE2_STEP(7, 1)
  444. SSE2_STEP(8, 0)
  445. SSE2_STEP(9, 1)
  446. SSE2_STEP(10, 0)
  447. SSE2_STEP(11, 1)
  448. SSE2_STEP(12, 0)
  449. SSE2_STEP(13, 1)
  450. SSE2_STEP(14, 0)
  451. SSE2_STEP(15, 1)
  452. SSE2_STEP(16, 0)
  453. SSE2_STEP(17, 1)
  454. SSE2_STEP(18, 0)
  455. SSE2_STEP(19, 1)
  456. // loop
  457. AS2( add WORD_REG(di), 5*4)
  458. AS2( cmp WORD_REG(di), SSE2_diEnd)
  459. ASJ( jne, 0, b)
  460. AS2( mov WORD_REG(ax), SSE2_input)
  461. AS2( mov AS_REG_7, SSE2_output)
  462. AS2( lea WORD_REG(di), [SSE2_uvStart]) // start of v and u
  463. AS2( mov WORD_REG(si), SSE2_wordsLeft2)
  464. ASL(1) // second inner loop, 16 words each, 5 iterations
  465. AS2( movdqa xmm0, [WORD_REG(di)+0*20*4])
  466. AS2( movdqa xmm2, [WORD_REG(di)+2*20*4])
  467. AS2( movdqa xmm3, [WORD_REG(di)+3*20*4])
  468. AS2( movdqa xmm1, [WORD_REG(di)+1*20*4])
  469. // S2
  470. AS2( movdqa xmm4, xmm0)
  471. AS2( pand xmm0, xmm2)
  472. AS2( pxor xmm0, xmm3)
  473. AS2( pxor xmm2, xmm1)
  474. AS2( pxor xmm2, xmm0)
  475. AS2( por xmm3, xmm4)
  476. AS2( pxor xmm3, xmm1)
  477. AS2( pxor xmm4, xmm2)
  478. AS2( movdqa xmm1, xmm3)
  479. AS2( por xmm3, xmm4)
  480. AS2( pxor xmm3, xmm0)
  481. AS2( pand xmm0, xmm1)
  482. AS2( pxor xmm4, xmm0)
  483. AS2( pxor xmm1, xmm3)
  484. AS2( pxor xmm1, xmm4)
  485. AS2( pxor xmm4, xmm7)
  486. // xor with v
  487. AS2( pxor xmm2, [WORD_REG(di)+80*4])
  488. AS2( pxor xmm3, [WORD_REG(di)+80*5])
  489. AS2( pxor xmm1, [WORD_REG(di)+80*6])
  490. AS2( pxor xmm4, [WORD_REG(di)+80*7])
  491. // exit loop early if less than 16 words left to output
  492. // this is necessary because block size is 20 words, and we output 16 words in each iteration of this loop
  493. AS2( cmp WORD_REG(si), 16)
  494. ASJ( jl, 4, f)
  495. // unpack
  496. AS2( movdqa xmm6, xmm2)
  497. AS2( punpckldq xmm2, xmm3)
  498. AS2( movdqa xmm5, xmm1)
  499. AS2( punpckldq xmm1, xmm4)
  500. AS2( movdqa xmm0, xmm2)
  501. AS2( punpcklqdq xmm2, xmm1)
  502. AS2( punpckhqdq xmm0, xmm1)
  503. AS2( punpckhdq xmm6, xmm3)
  504. AS2( punpckhdq xmm5, xmm4)
  505. AS2( movdqa xmm3, xmm6)
  506. AS2( punpcklqdq xmm6, xmm5)
  507. AS2( punpckhqdq xmm3, xmm5)
  508. // output keystream
  509. AS_XMM_OUTPUT4(SSE2_Sosemanuk_Output, WORD_REG(ax), AS_REG_7, 2,0,6,3, 1, 0,1,2,3, 4)
  510. // loop
  511. AS2( add WORD_REG(di), 4*4)
  512. AS2( sub WORD_REG(si), 16)
  513. ASJ( jnz, 1, b)
  514. // outer loop
  515. AS2( mov WORD_REG(si), SSE2_wordsLeft)
  516. AS2( sub WORD_REG(si), 80)
  517. ASJ( jz, 6, f)
  518. AS2( mov SSE2_wordsLeft, WORD_REG(si))
  519. AS2( mov SSE2_input, WORD_REG(ax))
  520. AS2( mov SSE2_output, AS_REG_7)
  521. ASJ( jmp, 2, b)
  522. ASL(4) // final output of less than 16 words
  523. AS2( test WORD_REG(ax), WORD_REG(ax))
  524. ASJ( jz, 5, f)
  525. AS2( movd xmm0, dword ptr [WORD_REG(ax)+0*4])
  526. AS2( pxor xmm2, xmm0)
  527. AS2( movd xmm0, dword ptr [WORD_REG(ax)+1*4])
  528. AS2( pxor xmm3, xmm0)
  529. AS2( movd xmm0, dword ptr [WORD_REG(ax)+2*4])
  530. AS2( pxor xmm1, xmm0)
  531. AS2( movd xmm0, dword ptr [WORD_REG(ax)+3*4])
  532. AS2( pxor xmm4, xmm0)
  533. AS2( add WORD_REG(ax), 16)
  534. ASL(5)
  535. AS2( movd dword ptr [AS_REG_7+0*4], xmm2)
  536. AS2( movd dword ptr [AS_REG_7+1*4], xmm3)
  537. AS2( movd dword ptr [AS_REG_7+2*4], xmm1)
  538. AS2( movd dword ptr [AS_REG_7+3*4], xmm4)
  539. AS2( sub WORD_REG(si), 4)
  540. ASJ( jz, 6, f)
  541. AS2( add AS_REG_7, 16)
  542. AS2( psrldq xmm2, 4)
  543. AS2( psrldq xmm3, 4)
  544. AS2( psrldq xmm1, 4)
  545. AS2( psrldq xmm4, 4)
  546. ASJ( jmp, 4, b)
  547. ASL(6) // save state
  548. AS2( mov AS_REG_6, SSE2_state)
  549. AS2( movdqa xmm0, [SSE2_stateCopy+0*16])
  550. AS2( movdqa [AS_REG_6+0*16], xmm0)
  551. AS2( movdqa xmm0, [SSE2_stateCopy+1*16])
  552. AS2( movdqa [AS_REG_6+1*16], xmm0)
  553. AS2( movq xmm0, QWORD PTR [SSE2_stateCopy+2*16])
  554. AS2( movq QWORD PTR [AS_REG_6+2*16], xmm0)
  555. AS2( mov [AS_REG_6+10*4], ecx)
  556. AS2( mov [AS_REG_6+11*4], edx)
  557. AS_POP_IF86( sp)
  558. AS_POP_IF86( bp)
  559. #ifdef __GNUC__
  560. AS_POP_IF86( bx)
  561. ATT_PREFIX
  562. :
  563. : "a" (m_state.m_ptr), "c" (iterationCount), "S" (s_sosemanukMulTables), "D" (output), "d" (input)
  564. #if CRYPTOPP_BOOL_X64
  565. , "r" (workspace.m_ptr)
  566. : "memory", "cc", "%r9", "%r10", "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7"
  567. #else
  568. : "memory", "cc"
  569. #endif
  570. );
  571. #endif
  572. #ifdef CRYPTOPP_GENERATE_X64_MASM
  573. movdqa xmm6, [rsp + 02f0h]
  574. movdqa xmm7, [rsp + 0300h]
  575. add rsp, 80*4*2+12*4+8*WORD_SZ + 2*16+8
  576. pop rdi
  577. pop rsi
  578. ret
  579. Sosemanuk_OperateKeystream ENDP
  580. #else
  581. }
  582. else
  583. #endif
  584. #endif
  585. #ifndef CRYPTOPP_GENERATE_X64_MASM
  586. {
  587. #if (CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X64) && !defined(CRYPTOPP_DISABLE_SOSEMANUK_ASM)
  588. #define MUL_A(x) (x = rotlFixed(x, 8), x ^ s_sosemanukMulTables[byte(x)])
  589. #else
  590. #define MUL_A(x) (((x) << 8) ^ s_sosemanukMulTables[(x) >> 24])
  591. #endif
  592. #define DIV_A(x) (((x) >> 8) ^ s_sosemanukMulTables[256 + byte(x)])
  593. #define r1(i) ((i%2) ? reg2 : reg1)
  594. #define r2(i) ((i%2) ? reg1 : reg2)
  595. #define STEP(x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, v, u) \
  596. u = (s##x9 + r2(x0)) ^ r1(x0);\
  597. v = s##x0;\
  598. s##x0 = MUL_A(s##x0) ^ DIV_A(s##x3) ^ s##x9;\
  599. r1(x0) += XMUX(r2(x0), s##x2, s##x9);\
  600. r2(x0) = rotlFixed(r2(x0) * 0x54655307, 7);\
  601. #define SOSEMANUK_OUTPUT(x) \
  602. CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 0, u2 ^ v0);\
  603. CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 1, u3 ^ v1);\
  604. CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 2, u1 ^ v2);\
  605. CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 3, u4 ^ v3);
  606. #define OUTPUT4 \
  607. S2(0, u0, u1, u2, u3, u4);\
  608. CRYPTOPP_KEYSTREAM_OUTPUT_SWITCH(SOSEMANUK_OUTPUT, 4*4);
  609. word32 s0 = m_state[0];
  610. word32 s1 = m_state[1];
  611. word32 s2 = m_state[2];
  612. word32 s3 = m_state[3];
  613. word32 s4 = m_state[4];
  614. word32 s5 = m_state[5];
  615. word32 s6 = m_state[6];
  616. word32 s7 = m_state[7];
  617. word32 s8 = m_state[8];
  618. word32 s9 = m_state[9];
  619. word32 reg1 = m_state[10];
  620. word32 reg2 = m_state[11];
  621. word32 u0, u1, u2, u3, u4, v0, v1, v2, v3;
  622. do
  623. {
  624. STEP(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, v0, u0)
  625. STEP(1, 2, 3, 4, 5, 6, 7, 8, 9, 0, v1, u1)
  626. STEP(2, 3, 4, 5, 6, 7, 8, 9, 0, 1, v2, u2)
  627. STEP(3, 4, 5, 6, 7, 8, 9, 0, 1, 2, v3, u3)
  628. OUTPUT4
  629. STEP(4, 5, 6, 7, 8, 9, 0, 1, 2, 3, v0, u0)
  630. STEP(5, 6, 7, 8, 9, 0, 1, 2, 3, 4, v1, u1)
  631. STEP(6, 7, 8, 9, 0, 1, 2, 3, 4, 5, v2, u2)
  632. STEP(7, 8, 9, 0, 1, 2, 3, 4, 5, 6, v3, u3)
  633. OUTPUT4
  634. STEP(8, 9, 0, 1, 2, 3, 4, 5, 6, 7, v0, u0)
  635. STEP(9, 0, 1, 2, 3, 4, 5, 6, 7, 8, v1, u1)
  636. STEP(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, v2, u2)
  637. STEP(1, 2, 3, 4, 5, 6, 7, 8, 9, 0, v3, u3)
  638. OUTPUT4
  639. STEP(2, 3, 4, 5, 6, 7, 8, 9, 0, 1, v0, u0)
  640. STEP(3, 4, 5, 6, 7, 8, 9, 0, 1, 2, v1, u1)
  641. STEP(4, 5, 6, 7, 8, 9, 0, 1, 2, 3, v2, u2)
  642. STEP(5, 6, 7, 8, 9, 0, 1, 2, 3, 4, v3, u3)
  643. OUTPUT4
  644. STEP(6, 7, 8, 9, 0, 1, 2, 3, 4, 5, v0, u0)
  645. STEP(7, 8, 9, 0, 1, 2, 3, 4, 5, 6, v1, u1)
  646. STEP(8, 9, 0, 1, 2, 3, 4, 5, 6, 7, v2, u2)
  647. STEP(9, 0, 1, 2, 3, 4, 5, 6, 7, 8, v3, u3)
  648. OUTPUT4
  649. }
  650. while (--iterationCount);
  651. m_state[0] = s0;
  652. m_state[1] = s1;
  653. m_state[2] = s2;
  654. m_state[3] = s3;
  655. m_state[4] = s4;
  656. m_state[5] = s5;
  657. m_state[6] = s6;
  658. m_state[7] = s7;
  659. m_state[8] = s8;
  660. m_state[9] = s9;
  661. m_state[10] = reg1;
  662. m_state[11] = reg2;
  663. }
  664. }
  665. NAMESPACE_END
  666. #endif // #ifndef CRYPTOPP_GENERATE_X64_MASM