1/* $OpenBSD: whirlpool.c,v 1.3 2024/06/01 07:36:17 tb Exp $ */ 2/** 3 * The Whirlpool hashing function. 4 * 5 * <P> 6 * <b>References</b> 7 * 8 * <P> 9 * The Whirlpool algorithm was developed by 10 * <a href="mailto:pbarreto@scopus.com.br">Paulo S. L. M. Barreto</a> and 11 * <a href="mailto:vincent.rijmen@cryptomathic.com">Vincent Rijmen</a>. 12 * 13 * See 14 * P.S.L.M. Barreto, V. Rijmen, 15 * ``The Whirlpool hashing function,'' 16 * NESSIE submission, 2000 (tweaked version, 2001), 17 * <https://www.cosic.esat.kuleuven.ac.be/nessie/workshop/submissions/whirlpool.zip> 18 * 19 * Based on "@version 3.0 (2003.03.12)" by Paulo S.L.M. Barreto and 20 * Vincent Rijmen. Lookup "reference implementations" on 21 * <http://planeta.terra.com.br/informatica/paulobarreto/> 22 * 23 * ============================================================================= 24 * 25 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ''AS IS'' AND ANY EXPRESS 26 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 27 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 28 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE 29 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 30 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 31 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR 32 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 33 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE 34 * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, 35 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 36 * 37 */ 38 39/* 40 * OpenSSL-specific implementation notes. 41 * 42 * WHIRLPOOL_Update as well as one-stroke WHIRLPOOL both expect 43 * number of *bytes* as input length argument. Bit-oriented routine 44 * as specified by authors is called WHIRLPOOL_BitUpdate[!] and 45 * does not have one-stroke counterpart. 46 * 47 * WHIRLPOOL_BitUpdate implements byte-oriented loop, essentially 48 * to serve WHIRLPOOL_Update. This is done for performance. 49 * 50 * Unlike authors' reference implementation, block processing 51 * routine whirlpool_block is designed to operate on multi-block 52 * input. This is done for performance. 53 */ 54 55#include <endian.h> 56#include <string.h> 57 58#include <openssl/crypto.h> 59#include <openssl/whrlpool.h> 60 61typedef unsigned char u8; 62#if defined(_LP64) 63typedef unsigned long u64; 64#else 65typedef unsigned long long u64; 66#endif 67 68#define ROUNDS 10 69 70#undef SMALL_REGISTER_BANK 71#if defined(__i386) || defined(__i386__) || defined(_M_IX86) 72# define SMALL_REGISTER_BANK 73# if defined(WHIRLPOOL_ASM) 74# ifndef OPENSSL_SMALL_FOOTPRINT 75# define OPENSSL_SMALL_FOOTPRINT /* it appears that for elder non-MMX 76 CPUs this is actually faster! */ 77# endif 78#include "x86_arch.h" 79# define GO_FOR_MMX(ctx,inp,num) \ 80do { \ 81 void whirlpool_block_mmx(void *,const void *,size_t); \ 82 if ((OPENSSL_cpu_caps() & CPUCAP_MASK_MMX) == 0) \ 83 break; \ 84 whirlpool_block_mmx(ctx->H.c,inp,num); \ 85 return; \ 86} while (0) 87# endif 88#elif defined(__arm__) 89# define SMALL_REGISTER_BANK 90#endif 91 92#undef ROTATE 93#if defined(__GNUC__) && __GNUC__>=2 94# if defined(__x86_64) || defined(__x86_64__) 95# define ROTATE(a,n) ({ u64 ret; asm ("rolq %1,%0" \ 96 : "=r"(ret) : "J"(n),"0"(a) : "cc"); ret; }) 97# endif 98#endif 99 100#if defined(OPENSSL_SMALL_FOOTPRINT) 101# if !defined(ROTATE) 102# if BYTE_ORDER == LITTLE_ENDIAN /* little-endians have to rotate left */ 103# define ROTATE(i,n) ((i)<<(n) ^ (i)>>(64-n)) 104# else /* big-endians have to rotate right */ 105# define ROTATE(i,n) ((i)>>(n) ^ (i)<<(64-n)) 106# endif 107# endif 108# if defined(ROTATE) && !defined(__STRICT_ALIGNMENT) 109# define __STRICT_ALIGNMENT /* ensure smallest table size */ 110# endif 111#endif 112 113/* 114 * Table size depends on __STRICT_ALIGNMENT and whether or not endian- 115 * specific ROTATE macro is defined. If __STRICT_ALIGNMENT is not 116 * defined, which is normally the case on x86[_64] CPUs, the table is 117 * 4KB large unconditionally. Otherwise if ROTATE is defined, the 118 * table is 2KB large, and otherwise - 16KB. 2KB table requires a 119 * whole bunch of additional rotations, but I'm willing to "trade," 120 * because 16KB table certainly trashes L1 cache. I wish all CPUs 121 * could handle unaligned load as 4KB table doesn't trash the cache, 122 * nor does it require additional rotations. 123 */ 124/* 125 * Note that every Cn macro expands as two loads: one byte load and 126 * one quadword load. One can argue that that many single-byte loads 127 * is too excessive, as one could load a quadword and "milk" it for 128 * eight 8-bit values instead. Well, yes, but in order to do so *and* 129 * avoid excessive loads you have to accommodate a handful of 64-bit 130 * values in the register bank and issue a bunch of shifts and mask. 131 * It's a tradeoff: loads vs. shift and mask in big register bank[!]. 132 * On most CPUs eight single-byte loads are faster and I let other 133 * ones to depend on smart compiler to fold byte loads if beneficial. 134 * Hand-coded assembler would be another alternative:-) 135 */ 136#ifdef __STRICT_ALIGNMENT 137# if defined(ROTATE) 138# define N 1 139# define LL(c0,c1,c2,c3,c4,c5,c6,c7) c0,c1,c2,c3,c4,c5,c6,c7 140# define C0(K,i) (Cx.q[K.c[(i)*8+0]]) 141# define C1(K,i) ROTATE(Cx.q[K.c[(i)*8+1]],8) 142# define C2(K,i) ROTATE(Cx.q[K.c[(i)*8+2]],16) 143# define C3(K,i) ROTATE(Cx.q[K.c[(i)*8+3]],24) 144# define C4(K,i) ROTATE(Cx.q[K.c[(i)*8+4]],32) 145# define C5(K,i) ROTATE(Cx.q[K.c[(i)*8+5]],40) 146# define C6(K,i) ROTATE(Cx.q[K.c[(i)*8+6]],48) 147# define C7(K,i) ROTATE(Cx.q[K.c[(i)*8+7]],56) 148# else 149# define N 8 150# define LL(c0,c1,c2,c3,c4,c5,c6,c7) c0,c1,c2,c3,c4,c5,c6,c7, \ 151 c7,c0,c1,c2,c3,c4,c5,c6, \ 152 c6,c7,c0,c1,c2,c3,c4,c5, \ 153 c5,c6,c7,c0,c1,c2,c3,c4, \ 154 c4,c5,c6,c7,c0,c1,c2,c3, \ 155 c3,c4,c5,c6,c7,c0,c1,c2, \ 156 c2,c3,c4,c5,c6,c7,c0,c1, \ 157 c1,c2,c3,c4,c5,c6,c7,c0 158# define C0(K,i) (Cx.q[0+8*K.c[(i)*8+0]]) 159# define C1(K,i) (Cx.q[1+8*K.c[(i)*8+1]]) 160# define C2(K,i) (Cx.q[2+8*K.c[(i)*8+2]]) 161# define C3(K,i) (Cx.q[3+8*K.c[(i)*8+3]]) 162# define C4(K,i) (Cx.q[4+8*K.c[(i)*8+4]]) 163# define C5(K,i) (Cx.q[5+8*K.c[(i)*8+5]]) 164# define C6(K,i) (Cx.q[6+8*K.c[(i)*8+6]]) 165# define C7(K,i) (Cx.q[7+8*K.c[(i)*8+7]]) 166# endif 167#else 168# define N 2 169# define LL(c0,c1,c2,c3,c4,c5,c6,c7) c0,c1,c2,c3,c4,c5,c6,c7, \ 170 c0,c1,c2,c3,c4,c5,c6,c7 171# define C0(K,i) (((u64*)(Cx.c+0))[2*K.c[(i)*8+0]]) 172# define C1(K,i) (((u64*)(Cx.c+7))[2*K.c[(i)*8+1]]) 173# define C2(K,i) (((u64*)(Cx.c+6))[2*K.c[(i)*8+2]]) 174# define C3(K,i) (((u64*)(Cx.c+5))[2*K.c[(i)*8+3]]) 175# define C4(K,i) (((u64*)(Cx.c+4))[2*K.c[(i)*8+4]]) 176# define C5(K,i) (((u64*)(Cx.c+3))[2*K.c[(i)*8+5]]) 177# define C6(K,i) (((u64*)(Cx.c+2))[2*K.c[(i)*8+6]]) 178# define C7(K,i) (((u64*)(Cx.c+1))[2*K.c[(i)*8+7]]) 179#endif 180 181static const 182union { 183 u8 c[(256*N+ROUNDS)*sizeof(u64)]; 184 u64 q[(256*N+ROUNDS)]; 185 } Cx = { { 186 /* Note endian-neutral representation:-) */ 187 LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8), 188 LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26), 189 LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8), 190 LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb), 191 LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb), 192 LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11), 193 LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09), 194 LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d), 195 LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b), 196 LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff), 197 LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c), 198 LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e), 199 LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96), 200 LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30), 201 LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d), 202 LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8), 203 LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47), 204 LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35), 205 LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37), 206 LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a), 207 LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2), 208 LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c), 209 LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84), 210 LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80), 211 LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5), 212 LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3), 213 LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21), 214 LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c), 215 LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43), 216 LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29), 217 LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d), 218 LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5), 219 LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd), 220 LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8), 221 LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92), 222 LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e), 223 LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13), 224 LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23), 225 LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20), 226 LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44), 227 LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2), 228 LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf), 229 LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c), 230 LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a), 231 LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50), 232 LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9), 233 LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14), 234 LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9), 235 LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c), 236 LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f), 237 LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90), 238 LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07), 239 LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd), 240 LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3), 241 LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d), 242 LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78), 243 LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97), 244 LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02), 245 LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73), 246 LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7), 247 LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6), 248 LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2), 249 LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49), 250 LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56), 251 LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70), 252 LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd), 253 LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb), 254 LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71), 255 LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b), 256 LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf), 257 LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45), 258 LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a), 259 LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4), 260 LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58), 261 LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e), 262 LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f), 263 LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac), 264 LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0), 265 LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef), 266 LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6), 267 LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c), 268 LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12), 269 LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93), 270 LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde), 271 LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6), 272 LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1), 273 LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b), 274 LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f), 275 LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31), 276 LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8), 277 LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9), 278 LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc), 279 LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e), 280 LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b), 281 LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf), 282 LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59), 283 LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2), 284 LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77), 285 LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33), 286 LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4), 287 LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27), 288 LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb), 289 LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89), 290 LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32), 291 LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54), 292 LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d), 293 LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64), 294 LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d), 295 LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d), 296 LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f), 297 LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca), 298 LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7), 299 LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d), 300 LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce), 301 LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f), 302 LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f), 303 LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63), 304 LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a), 305 LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc), 306 LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82), 307 LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a), 308 LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48), 309 LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95), 310 LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf), 311 LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d), 312 LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0), 313 LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91), 314 LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8), 315 LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b), 316 LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00), 317 LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9), 318 LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e), 319 LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1), 320 LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6), 321 LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28), 322 LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3), 323 LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74), 324 LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe), 325 LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d), 326 LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea), 327 LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57), 328 LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38), 329 LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad), 330 LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4), 331 LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda), 332 LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7), 333 LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb), 334 LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9), 335 LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a), 336 LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03), 337 LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a), 338 LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e), 339 LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60), 340 LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc), 341 LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46), 342 LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f), 343 LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76), 344 LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa), 345 LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36), 346 LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae), 347 LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b), 348 LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85), 349 LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e), 350 LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7), 351 LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55), 352 LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a), 353 LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81), 354 LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52), 355 LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62), 356 LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3), 357 LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10), 358 LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab), 359 LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0), 360 LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5), 361 LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec), 362 LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16), 363 LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94), 364 LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f), 365 LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5), 366 LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98), 367 LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17), 368 LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4), 369 LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1), 370 LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e), 371 LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42), 372 LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34), 373 LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08), 374 LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee), 375 LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61), 376 LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1), 377 LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f), 378 LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24), 379 LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3), 380 LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25), 381 LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22), 382 LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65), 383 LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79), 384 LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69), 385 LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9), 386 LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19), 387 LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe), 388 LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a), 389 LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0), 390 LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99), 391 LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83), 392 LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04), 393 LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66), 394 LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0), 395 LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1), 396 LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd), 397 LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40), 398 LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c), 399 LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18), 400 LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b), 401 LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51), 402 LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05), 403 LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c), 404 LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39), 405 LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa), 406 LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b), 407 LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc), 408 LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e), 409 LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0), 410 LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88), 411 LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67), 412 LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a), 413 LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87), 414 LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1), 415 LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72), 416 LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53), 417 LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01), 418 LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b), 419 LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4), 420 LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3), 421 LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15), 422 LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c), 423 LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5), 424 LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5), 425 LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4), 426 LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba), 427 LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6), 428 LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7), 429 LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06), 430 LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41), 431 LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7), 432 LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f), 433 LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e), 434 LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6), 435 LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2), 436 LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68), 437 LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c), 438 LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed), 439 LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75), 440 LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86), 441 LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b), 442 LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2), 443#define RC (&(Cx.q[256*N])) 444 0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f, /* rc[ROUNDS] */ 445 0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52, 446 0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35, 447 0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57, 448 0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda, 449 0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85, 450 0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67, 451 0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8, 452 0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e, 453 0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33 454 } 455}; 456 457void whirlpool_block(WHIRLPOOL_CTX *ctx,const void *inp,size_t n) 458 { 459 int r; 460 const u8 *p=inp; 461 union { u64 q[8]; u8 c[64]; } S,K,*H=(void *)ctx->H.q; 462 463#ifdef GO_FOR_MMX 464 GO_FOR_MMX(ctx,inp,n); 465#endif 466 do { 467#ifdef OPENSSL_SMALL_FOOTPRINT 468 u64 L[8]; 469 int i; 470 471 for (i=0;i<64;i++) S.c[i] = (K.c[i] = H->c[i]) ^ p[i]; 472 for (r=0;r<ROUNDS;r++) 473 { 474 for (i=0;i<8;i++) 475 { 476 L[i] = i ? 0 : RC[r]; 477 L[i] ^= C0(K,i) ^ C1(K,(i-1)&7) ^ 478 C2(K,(i-2)&7) ^ C3(K,(i-3)&7) ^ 479 C4(K,(i-4)&7) ^ C5(K,(i-5)&7) ^ 480 C6(K,(i-6)&7) ^ C7(K,(i-7)&7); 481 } 482 memcpy (K.q,L,64); 483 for (i=0;i<8;i++) 484 { 485 L[i] ^= C0(S,i) ^ C1(S,(i-1)&7) ^ 486 C2(S,(i-2)&7) ^ C3(S,(i-3)&7) ^ 487 C4(S,(i-4)&7) ^ C5(S,(i-5)&7) ^ 488 C6(S,(i-6)&7) ^ C7(S,(i-7)&7); 489 } 490 memcpy (S.q,L,64); 491 } 492 for (i=0;i<64;i++) H->c[i] ^= S.c[i] ^ p[i]; 493#else 494 u64 L0,L1,L2,L3,L4,L5,L6,L7; 495 496#ifdef __STRICT_ALIGNMENT 497 if ((size_t)p & 7) 498 { 499 memcpy (S.c,p,64); 500 S.q[0] ^= (K.q[0] = H->q[0]); 501 S.q[1] ^= (K.q[1] = H->q[1]); 502 S.q[2] ^= (K.q[2] = H->q[2]); 503 S.q[3] ^= (K.q[3] = H->q[3]); 504 S.q[4] ^= (K.q[4] = H->q[4]); 505 S.q[5] ^= (K.q[5] = H->q[5]); 506 S.q[6] ^= (K.q[6] = H->q[6]); 507 S.q[7] ^= (K.q[7] = H->q[7]); 508 } 509 else 510#endif 511 { 512 const u64 *pa = (const u64*)p; 513 S.q[0] = (K.q[0] = H->q[0]) ^ pa[0]; 514 S.q[1] = (K.q[1] = H->q[1]) ^ pa[1]; 515 S.q[2] = (K.q[2] = H->q[2]) ^ pa[2]; 516 S.q[3] = (K.q[3] = H->q[3]) ^ pa[3]; 517 S.q[4] = (K.q[4] = H->q[4]) ^ pa[4]; 518 S.q[5] = (K.q[5] = H->q[5]) ^ pa[5]; 519 S.q[6] = (K.q[6] = H->q[6]) ^ pa[6]; 520 S.q[7] = (K.q[7] = H->q[7]) ^ pa[7]; 521 } 522 523 for(r=0;r<ROUNDS;r++) 524 { 525#ifdef SMALL_REGISTER_BANK 526 L0 = C0(K,0) ^ C1(K,7) ^ C2(K,6) ^ C3(K,5) ^ 527 C4(K,4) ^ C5(K,3) ^ C6(K,2) ^ C7(K,1) ^ RC[r]; 528 L1 = C0(K,1) ^ C1(K,0) ^ C2(K,7) ^ C3(K,6) ^ 529 C4(K,5) ^ C5(K,4) ^ C6(K,3) ^ C7(K,2); 530 L2 = C0(K,2) ^ C1(K,1) ^ C2(K,0) ^ C3(K,7) ^ 531 C4(K,6) ^ C5(K,5) ^ C6(K,4) ^ C7(K,3); 532 L3 = C0(K,3) ^ C1(K,2) ^ C2(K,1) ^ C3(K,0) ^ 533 C4(K,7) ^ C5(K,6) ^ C6(K,5) ^ C7(K,4); 534 L4 = C0(K,4) ^ C1(K,3) ^ C2(K,2) ^ C3(K,1) ^ 535 C4(K,0) ^ C5(K,7) ^ C6(K,6) ^ C7(K,5); 536 L5 = C0(K,5) ^ C1(K,4) ^ C2(K,3) ^ C3(K,2) ^ 537 C4(K,1) ^ C5(K,0) ^ C6(K,7) ^ C7(K,6); 538 L6 = C0(K,6) ^ C1(K,5) ^ C2(K,4) ^ C3(K,3) ^ 539 C4(K,2) ^ C5(K,1) ^ C6(K,0) ^ C7(K,7); 540 L7 = C0(K,7) ^ C1(K,6) ^ C2(K,5) ^ C3(K,4) ^ 541 C4(K,3) ^ C5(K,2) ^ C6(K,1) ^ C7(K,0); 542 543 K.q[0] = L0; K.q[1] = L1; K.q[2] = L2; K.q[3] = L3; 544 K.q[4] = L4; K.q[5] = L5; K.q[6] = L6; K.q[7] = L7; 545 546 L0 ^= C0(S,0) ^ C1(S,7) ^ C2(S,6) ^ C3(S,5) ^ 547 C4(S,4) ^ C5(S,3) ^ C6(S,2) ^ C7(S,1); 548 L1 ^= C0(S,1) ^ C1(S,0) ^ C2(S,7) ^ C3(S,6) ^ 549 C4(S,5) ^ C5(S,4) ^ C6(S,3) ^ C7(S,2); 550 L2 ^= C0(S,2) ^ C1(S,1) ^ C2(S,0) ^ C3(S,7) ^ 551 C4(S,6) ^ C5(S,5) ^ C6(S,4) ^ C7(S,3); 552 L3 ^= C0(S,3) ^ C1(S,2) ^ C2(S,1) ^ C3(S,0) ^ 553 C4(S,7) ^ C5(S,6) ^ C6(S,5) ^ C7(S,4); 554 L4 ^= C0(S,4) ^ C1(S,3) ^ C2(S,2) ^ C3(S,1) ^ 555 C4(S,0) ^ C5(S,7) ^ C6(S,6) ^ C7(S,5); 556 L5 ^= C0(S,5) ^ C1(S,4) ^ C2(S,3) ^ C3(S,2) ^ 557 C4(S,1) ^ C5(S,0) ^ C6(S,7) ^ C7(S,6); 558 L6 ^= C0(S,6) ^ C1(S,5) ^ C2(S,4) ^ C3(S,3) ^ 559 C4(S,2) ^ C5(S,1) ^ C6(S,0) ^ C7(S,7); 560 L7 ^= C0(S,7) ^ C1(S,6) ^ C2(S,5) ^ C3(S,4) ^ 561 C4(S,3) ^ C5(S,2) ^ C6(S,1) ^ C7(S,0); 562 563 S.q[0] = L0; S.q[1] = L1; S.q[2] = L2; S.q[3] = L3; 564 S.q[4] = L4; S.q[5] = L5; S.q[6] = L6; S.q[7] = L7; 565#else 566 L0 = C0(K,0); L1 = C1(K,0); L2 = C2(K,0); L3 = C3(K,0); 567 L4 = C4(K,0); L5 = C5(K,0); L6 = C6(K,0); L7 = C7(K,0); 568 L0 ^= RC[r]; 569 570 L1 ^= C0(K,1); L2 ^= C1(K,1); L3 ^= C2(K,1); L4 ^= C3(K,1); 571 L5 ^= C4(K,1); L6 ^= C5(K,1); L7 ^= C6(K,1); L0 ^= C7(K,1); 572 573 L2 ^= C0(K,2); L3 ^= C1(K,2); L4 ^= C2(K,2); L5 ^= C3(K,2); 574 L6 ^= C4(K,2); L7 ^= C5(K,2); L0 ^= C6(K,2); L1 ^= C7(K,2); 575 576 L3 ^= C0(K,3); L4 ^= C1(K,3); L5 ^= C2(K,3); L6 ^= C3(K,3); 577 L7 ^= C4(K,3); L0 ^= C5(K,3); L1 ^= C6(K,3); L2 ^= C7(K,3); 578 579 L4 ^= C0(K,4); L5 ^= C1(K,4); L6 ^= C2(K,4); L7 ^= C3(K,4); 580 L0 ^= C4(K,4); L1 ^= C5(K,4); L2 ^= C6(K,4); L3 ^= C7(K,4); 581 582 L5 ^= C0(K,5); L6 ^= C1(K,5); L7 ^= C2(K,5); L0 ^= C3(K,5); 583 L1 ^= C4(K,5); L2 ^= C5(K,5); L3 ^= C6(K,5); L4 ^= C7(K,5); 584 585 L6 ^= C0(K,6); L7 ^= C1(K,6); L0 ^= C2(K,6); L1 ^= C3(K,6); 586 L2 ^= C4(K,6); L3 ^= C5(K,6); L4 ^= C6(K,6); L5 ^= C7(K,6); 587 588 L7 ^= C0(K,7); L0 ^= C1(K,7); L1 ^= C2(K,7); L2 ^= C3(K,7); 589 L3 ^= C4(K,7); L4 ^= C5(K,7); L5 ^= C6(K,7); L6 ^= C7(K,7); 590 591 K.q[0] = L0; K.q[1] = L1; K.q[2] = L2; K.q[3] = L3; 592 K.q[4] = L4; K.q[5] = L5; K.q[6] = L6; K.q[7] = L7; 593 594 L0 ^= C0(S,0); L1 ^= C1(S,0); L2 ^= C2(S,0); L3 ^= C3(S,0); 595 L4 ^= C4(S,0); L5 ^= C5(S,0); L6 ^= C6(S,0); L7 ^= C7(S,0); 596 597 L1 ^= C0(S,1); L2 ^= C1(S,1); L3 ^= C2(S,1); L4 ^= C3(S,1); 598 L5 ^= C4(S,1); L6 ^= C5(S,1); L7 ^= C6(S,1); L0 ^= C7(S,1); 599 600 L2 ^= C0(S,2); L3 ^= C1(S,2); L4 ^= C2(S,2); L5 ^= C3(S,2); 601 L6 ^= C4(S,2); L7 ^= C5(S,2); L0 ^= C6(S,2); L1 ^= C7(S,2); 602 603 L3 ^= C0(S,3); L4 ^= C1(S,3); L5 ^= C2(S,3); L6 ^= C3(S,3); 604 L7 ^= C4(S,3); L0 ^= C5(S,3); L1 ^= C6(S,3); L2 ^= C7(S,3); 605 606 L4 ^= C0(S,4); L5 ^= C1(S,4); L6 ^= C2(S,4); L7 ^= C3(S,4); 607 L0 ^= C4(S,4); L1 ^= C5(S,4); L2 ^= C6(S,4); L3 ^= C7(S,4); 608 609 L5 ^= C0(S,5); L6 ^= C1(S,5); L7 ^= C2(S,5); L0 ^= C3(S,5); 610 L1 ^= C4(S,5); L2 ^= C5(S,5); L3 ^= C6(S,5); L4 ^= C7(S,5); 611 612 L6 ^= C0(S,6); L7 ^= C1(S,6); L0 ^= C2(S,6); L1 ^= C3(S,6); 613 L2 ^= C4(S,6); L3 ^= C5(S,6); L4 ^= C6(S,6); L5 ^= C7(S,6); 614 615 L7 ^= C0(S,7); L0 ^= C1(S,7); L1 ^= C2(S,7); L2 ^= C3(S,7); 616 L3 ^= C4(S,7); L4 ^= C5(S,7); L5 ^= C6(S,7); L6 ^= C7(S,7); 617 618 S.q[0] = L0; S.q[1] = L1; S.q[2] = L2; S.q[3] = L3; 619 S.q[4] = L4; S.q[5] = L5; S.q[6] = L6; S.q[7] = L7; 620#endif 621 } 622 623#ifdef __STRICT_ALIGNMENT 624 if ((size_t)p & 7) 625 { 626 int i; 627 for(i=0;i<64;i++) H->c[i] ^= S.c[i] ^ p[i]; 628 } 629 else 630#endif 631 { 632 const u64 *pa=(const u64 *)p; 633 H->q[0] ^= S.q[0] ^ pa[0]; 634 H->q[1] ^= S.q[1] ^ pa[1]; 635 H->q[2] ^= S.q[2] ^ pa[2]; 636 H->q[3] ^= S.q[3] ^ pa[3]; 637 H->q[4] ^= S.q[4] ^ pa[4]; 638 H->q[5] ^= S.q[5] ^ pa[5]; 639 H->q[6] ^= S.q[6] ^ pa[6]; 640 H->q[7] ^= S.q[7] ^ pa[7]; 641 } 642#endif 643 p += 64; 644 } while(--n); 645 } 646 647int 648WHIRLPOOL_Init(WHIRLPOOL_CTX *c) 649{ 650 memset (c, 0, sizeof(*c)); 651 return (1); 652} 653LCRYPTO_ALIAS(WHIRLPOOL_Init); 654 655int 656WHIRLPOOL_Update(WHIRLPOOL_CTX *c, const void *_inp, size_t bytes) 657{ 658 /* Well, largest suitable chunk size actually is 659 * (1<<(sizeof(size_t)*8-3))-64, but below number 660 * is large enough for not to care about excessive 661 * calls to WHIRLPOOL_BitUpdate... */ 662 size_t chunk = ((size_t)1) << (sizeof(size_t)*8 - 4); 663 const unsigned char *inp = _inp; 664 665 while (bytes >= chunk) { 666 WHIRLPOOL_BitUpdate(c, inp, chunk*8); 667 bytes -= chunk; 668 inp += chunk; 669 } 670 if (bytes) 671 WHIRLPOOL_BitUpdate(c, inp, bytes*8); 672 673 return (1); 674} 675LCRYPTO_ALIAS(WHIRLPOOL_Update); 676 677void 678WHIRLPOOL_BitUpdate(WHIRLPOOL_CTX *c, const void *_inp, size_t bits) 679{ 680 size_t n; 681 unsigned int bitoff = c->bitoff, 682 bitrem = bitoff % 8, 683 inpgap = (8 - (unsigned int)bits % 8)&7; 684 const unsigned char *inp = _inp; 685 686 /* This 256-bit increment procedure relies on the size_t 687 * being natural size of CPU register, so that we don't 688 * have to mask the value in order to detect overflows. */ 689 c->bitlen[0] += bits; 690 if (c->bitlen[0] < bits) /* overflow */ 691 { 692 n = 1; 693 do { 694 c->bitlen[n]++; 695 } while (c->bitlen[n]==0 && 696 ++n < (WHIRLPOOL_COUNTER/sizeof(size_t))); 697 } 698 699#ifndef OPENSSL_SMALL_FOOTPRINT 700reconsider: 701 if (inpgap==0 && bitrem==0) /* byte-oriented loop */ 702 { 703 while (bits) { 704 if (bitoff == 0 && (n = bits/WHIRLPOOL_BBLOCK)) { 705 whirlpool_block(c, inp, n); 706 inp += n*WHIRLPOOL_BBLOCK/8; 707 bits %= WHIRLPOOL_BBLOCK; 708 } else { 709 unsigned int byteoff = bitoff/8; 710 711 bitrem = WHIRLPOOL_BBLOCK - bitoff;/* re-use bitrem */ 712 if (bits >= bitrem) { 713 bits -= bitrem; 714 bitrem /= 8; 715 memcpy(c->data + byteoff, inp, bitrem); 716 inp += bitrem; 717 whirlpool_block(c, c->data, 1); 718 bitoff = 0; 719 } else { 720 memcpy(c->data + byteoff, inp, bits/8); 721 bitoff += (unsigned int)bits; 722 bits = 0; 723 } 724 c->bitoff = bitoff; 725 } 726 } 727 } 728 else /* bit-oriented loop */ 729#endif 730 { 731 /* 732 inp 733 | 734 +-------+-------+------- 735 ||||||||||||||||||||| 736 +-------+-------+------- 737 +-------+-------+-------+-------+------- 738 |||||||||||||| c->data 739 +-------+-------+-------+-------+------- 740 | 741 c->bitoff/8 742 */ 743 while (bits) { 744 unsigned int byteoff = bitoff/8; 745 unsigned char b; 746 747#ifndef OPENSSL_SMALL_FOOTPRINT 748 if (bitrem == inpgap) { 749 c->data[byteoff++] |= inp[0] & (0xff >> inpgap); 750 inpgap = 8 - inpgap; 751 bitoff += inpgap; bitrem = 0; /* bitoff%8 */ 752 bits -= inpgap; inpgap = 0; /* bits%8 */ 753 inp++; 754 if (bitoff == WHIRLPOOL_BBLOCK) { 755 whirlpool_block(c, c->data, 1); 756 bitoff = 0; 757 } 758 c->bitoff = bitoff; 759 goto reconsider; 760 } else 761#endif 762 if (bits >= 8) { 763 b = ((inp[0]<<inpgap) | (inp[1]>>(8 - inpgap))); 764 b &= 0xff; 765 if (bitrem) 766 c->data[byteoff++] |= b >> bitrem; 767 else 768 c->data[byteoff++] = b; 769 bitoff += 8; 770 bits -= 8; 771 inp++; 772 if (bitoff >= WHIRLPOOL_BBLOCK) { 773 whirlpool_block(c, c->data, 1); 774 byteoff = 0; 775 bitoff %= WHIRLPOOL_BBLOCK; 776 } 777 if (bitrem) 778 c->data[byteoff] = b << (8 - bitrem); 779 } 780 else /* remaining less than 8 bits */ 781 { 782 b = (inp[0]<<inpgap)&0xff; 783 if (bitrem) 784 c->data[byteoff++] |= b >> bitrem; 785 else 786 c->data[byteoff++] = b; 787 bitoff += (unsigned int)bits; 788 if (bitoff == WHIRLPOOL_BBLOCK) { 789 whirlpool_block(c, c->data, 1); 790 byteoff = 0; 791 bitoff %= WHIRLPOOL_BBLOCK; 792 } 793 if (bitrem) 794 c->data[byteoff] = b << (8 - bitrem); 795 bits = 0; 796 } 797 c->bitoff = bitoff; 798 } 799 } 800} 801LCRYPTO_ALIAS(WHIRLPOOL_BitUpdate); 802 803int 804WHIRLPOOL_Final(unsigned char *md, WHIRLPOOL_CTX *c) 805{ 806 unsigned int bitoff = c->bitoff, 807 byteoff = bitoff/8; 808 size_t i, j, v; 809 unsigned char *p; 810 811 bitoff %= 8; 812 if (bitoff) 813 c->data[byteoff] |= 0x80 >> bitoff; 814 else 815 c->data[byteoff] = 0x80; 816 byteoff++; 817 818 /* pad with zeros */ 819 if (byteoff > (WHIRLPOOL_BBLOCK/8 - WHIRLPOOL_COUNTER)) { 820 if (byteoff < WHIRLPOOL_BBLOCK/8) 821 memset(&c->data[byteoff], 0, WHIRLPOOL_BBLOCK/8 - byteoff); 822 whirlpool_block(c, c->data, 1); 823 byteoff = 0; 824 } 825 if (byteoff < (WHIRLPOOL_BBLOCK/8 - WHIRLPOOL_COUNTER)) 826 memset(&c->data[byteoff], 0, 827 (WHIRLPOOL_BBLOCK/8 - WHIRLPOOL_COUNTER) - byteoff); 828 /* smash 256-bit c->bitlen in big-endian order */ 829 p = &c->data[WHIRLPOOL_BBLOCK/8-1]; /* last byte in c->data */ 830 for (i = 0; i < WHIRLPOOL_COUNTER/sizeof(size_t); i++) 831 for (v = c->bitlen[i], j = 0; j < sizeof(size_t); j++, v >>= 8) 832 *p-- = (unsigned char)(v&0xff); 833 834 whirlpool_block(c, c->data, 1); 835 836 if (md) { 837 memcpy(md, c->H.c, WHIRLPOOL_DIGEST_LENGTH); 838 memset(c, 0, sizeof(*c)); 839 return (1); 840 } 841 return (0); 842} 843LCRYPTO_ALIAS(WHIRLPOOL_Final); 844 845unsigned char * 846WHIRLPOOL(const void *inp, size_t bytes, unsigned char *md) 847{ 848 WHIRLPOOL_CTX ctx; 849 850 WHIRLPOOL_Init(&ctx); 851 WHIRLPOOL_Update(&ctx, inp, bytes); 852 WHIRLPOOL_Final(md, &ctx); 853 return (md); 854} 855LCRYPTO_ALIAS(WHIRLPOOL); 856