1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/mfbt/SHA1.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,327 @@
1.4 +/* This Source Code Form is subject to the terms of the Mozilla Public
1.5 + * License, v. 2.0. If a copy of the MPL was not distributed with this
1.6 + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
1.7 +
1.8 +#include "mozilla/Assertions.h"
1.9 +#include "mozilla/Endian.h"
1.10 +#include "mozilla/SHA1.h"
1.11 +
1.12 +#include <string.h>
1.13 +
1.14 +using mozilla::NativeEndian;
1.15 +using mozilla::SHA1Sum;
1.16 +
1.17 +static inline uint32_t
1.18 +SHA_ROTL(uint32_t t, uint32_t n)
1.19 +{
1.20 + MOZ_ASSERT(n < 32);
1.21 + return (t << n) | (t >> (32 - n));
1.22 +}
1.23 +
1.24 +static void
1.25 +shaCompress(volatile unsigned* X, const uint32_t* datain);
1.26 +
1.27 +#define SHA_F1(X, Y, Z) ((((Y) ^ (Z)) & (X)) ^ (Z))
1.28 +#define SHA_F2(X, Y, Z) ((X) ^ (Y) ^ (Z))
1.29 +#define SHA_F3(X, Y, Z) (((X) & (Y)) | ((Z) & ((X) | (Y))))
1.30 +#define SHA_F4(X, Y, Z) ((X) ^ (Y) ^ (Z))
1.31 +
1.32 +#define SHA_MIX(n, a, b, c) XW(n) = SHA_ROTL(XW(a) ^ XW(b) ^ XW(c) ^XW(n), 1)
1.33 +
1.34 +SHA1Sum::SHA1Sum()
1.35 + : size(0), mDone(false)
1.36 +{
1.37 + // Initialize H with constants from FIPS180-1.
1.38 + H[0] = 0x67452301L;
1.39 + H[1] = 0xefcdab89L;
1.40 + H[2] = 0x98badcfeL;
1.41 + H[3] = 0x10325476L;
1.42 + H[4] = 0xc3d2e1f0L;
1.43 +}
1.44 +
1.45 +/*
1.46 + * Explanation of H array and index values:
1.47 + *
1.48 + * The context's H array is actually the concatenation of two arrays
1.49 + * defined by SHA1, the H array of state variables (5 elements),
1.50 + * and the W array of intermediate values, of which there are 16 elements.
1.51 + * The W array starts at H[5], that is W[0] is H[5].
1.52 + * Although these values are defined as 32-bit values, we use 64-bit
1.53 + * variables to hold them because the AMD64 stores 64 bit values in
1.54 + * memory MUCH faster than it stores any smaller values.
1.55 + *
1.56 + * Rather than passing the context structure to shaCompress, we pass
1.57 + * this combined array of H and W values. We do not pass the address
1.58 + * of the first element of this array, but rather pass the address of an
1.59 + * element in the middle of the array, element X. Presently X[0] is H[11].
1.60 + * So we pass the address of H[11] as the address of array X to shaCompress.
1.61 + * Then shaCompress accesses the members of the array using positive AND
1.62 + * negative indexes.
1.63 + *
1.64 + * Pictorially: (each element is 8 bytes)
1.65 + * H | H0 H1 H2 H3 H4 W0 W1 W2 W3 W4 W5 W6 W7 W8 W9 Wa Wb Wc Wd We Wf |
1.66 + * X |-11-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 |
1.67 + *
1.68 + * The byte offset from X[0] to any member of H and W is always
1.69 + * representable in a signed 8-bit value, which will be encoded
1.70 + * as a single byte offset in the X86-64 instruction set.
1.71 + * If we didn't pass the address of H[11], and instead passed the
1.72 + * address of H[0], the offsets to elements H[16] and above would be
1.73 + * greater than 127, not representable in a signed 8-bit value, and the
1.74 + * x86-64 instruction set would encode every such offset as a 32-bit
1.75 + * signed number in each instruction that accessed element H[16] or
1.76 + * higher. This results in much bigger and slower code.
1.77 + */
1.78 +#define H2X 11 /* X[0] is H[11], and H[0] is X[-11] */
1.79 +#define W2X 6 /* X[0] is W[6], and W[0] is X[-6] */
1.80 +
1.81 +/*
1.82 + * SHA: Add data to context.
1.83 + */
1.84 +void
1.85 +SHA1Sum::update(const void* dataIn, uint32_t len)
1.86 +{
1.87 + MOZ_ASSERT(!mDone, "SHA1Sum can only be used to compute a single hash.");
1.88 +
1.89 + const uint8_t* data = static_cast<const uint8_t*>(dataIn);
1.90 +
1.91 + if (len == 0)
1.92 + return;
1.93 +
1.94 + /* Accumulate the byte count. */
1.95 + unsigned int lenB = static_cast<unsigned int>(size) & 63U;
1.96 +
1.97 + size += len;
1.98 +
1.99 + /* Read the data into W and process blocks as they get full. */
1.100 + unsigned int togo;
1.101 + if (lenB > 0) {
1.102 + togo = 64U - lenB;
1.103 + if (len < togo)
1.104 + togo = len;
1.105 + memcpy(u.b + lenB, data, togo);
1.106 + len -= togo;
1.107 + data += togo;
1.108 + lenB = (lenB + togo) & 63U;
1.109 + if (!lenB)
1.110 + shaCompress(&H[H2X], u.w);
1.111 + }
1.112 +
1.113 + while (len >= 64U) {
1.114 + len -= 64U;
1.115 + shaCompress(&H[H2X], reinterpret_cast<const uint32_t*>(data));
1.116 + data += 64U;
1.117 + }
1.118 +
1.119 + if (len > 0)
1.120 + memcpy(u.b, data, len);
1.121 +}
1.122 +
1.123 +
1.124 +/*
1.125 + * SHA: Generate hash value
1.126 + */
1.127 +void
1.128 +SHA1Sum::finish(SHA1Sum::Hash& hashOut)
1.129 +{
1.130 + MOZ_ASSERT(!mDone, "SHA1Sum can only be used to compute a single hash.");
1.131 +
1.132 + uint64_t size2 = size;
1.133 + uint32_t lenB = uint32_t(size2) & 63;
1.134 +
1.135 + static const uint8_t bulk_pad[64] =
1.136 + { 0x80,0,0,0,0,0,0,0,0,0,
1.137 + 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
1.138 + 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
1.139 +
1.140 + /* Pad with a binary 1 (e.g. 0x80), then zeroes, then length in bits. */
1.141 + update(bulk_pad, (((55 + 64) - lenB) & 63) + 1);
1.142 + MOZ_ASSERT((uint32_t(size) & 63) == 56);
1.143 +
1.144 + /* Convert size from bytes to bits. */
1.145 + size2 <<= 3;
1.146 + u.w[14] = NativeEndian::swapToBigEndian(uint32_t(size2 >> 32));
1.147 + u.w[15] = NativeEndian::swapToBigEndian(uint32_t(size2));
1.148 + shaCompress(&H[H2X], u.w);
1.149 +
1.150 + /* Output hash. */
1.151 + u.w[0] = NativeEndian::swapToBigEndian(H[0]);
1.152 + u.w[1] = NativeEndian::swapToBigEndian(H[1]);
1.153 + u.w[2] = NativeEndian::swapToBigEndian(H[2]);
1.154 + u.w[3] = NativeEndian::swapToBigEndian(H[3]);
1.155 + u.w[4] = NativeEndian::swapToBigEndian(H[4]);
1.156 + memcpy(hashOut, u.w, 20);
1.157 + mDone = true;
1.158 +}
1.159 +
1.160 +/*
1.161 + * SHA: Compression function, unrolled.
1.162 + *
1.163 + * Some operations in shaCompress are done as 5 groups of 16 operations.
1.164 + * Others are done as 4 groups of 20 operations.
1.165 + * The code below shows that structure.
1.166 + *
1.167 + * The functions that compute the new values of the 5 state variables
1.168 + * A-E are done in 4 groups of 20 operations (or you may also think
1.169 + * of them as being done in 16 groups of 5 operations). They are
1.170 + * done by the SHA_RNDx macros below, in the right column.
1.171 + *
1.172 + * The functions that set the 16 values of the W array are done in
1.173 + * 5 groups of 16 operations. The first group is done by the
1.174 + * LOAD macros below, the latter 4 groups are done by SHA_MIX below,
1.175 + * in the left column.
1.176 + *
1.177 + * gcc's optimizer observes that each member of the W array is assigned
1.178 + * a value 5 times in this code. It reduces the number of store
1.179 + * operations done to the W array in the context (that is, in the X array)
1.180 + * by creating a W array on the stack, and storing the W values there for
1.181 + * the first 4 groups of operations on W, and storing the values in the
1.182 + * context's W array only in the fifth group. This is undesirable.
1.183 + * It is MUCH bigger code than simply using the context's W array, because
1.184 + * all the offsets to the W array in the stack are 32-bit signed offsets,
1.185 + * and it is no faster than storing the values in the context's W array.
1.186 + *
1.187 + * The original code for sha_fast.c prevented this creation of a separate
1.188 + * W array in the stack by creating a W array of 80 members, each of
1.189 + * whose elements is assigned only once. It also separated the computations
1.190 + * of the W array values and the computations of the values for the 5
1.191 + * state variables into two separate passes, W's, then A-E's so that the
1.192 + * second pass could be done all in registers (except for accessing the W
1.193 + * array) on machines with fewer registers. The method is suboptimal
1.194 + * for machines with enough registers to do it all in one pass, and it
1.195 + * necessitates using many instructions with 32-bit offsets.
1.196 + *
1.197 + * This code eliminates the separate W array on the stack by a completely
1.198 + * different means: by declaring the X array volatile. This prevents
1.199 + * the optimizer from trying to reduce the use of the X array by the
1.200 + * creation of a MORE expensive W array on the stack. The result is
1.201 + * that all instructions use signed 8-bit offsets and not 32-bit offsets.
1.202 + *
1.203 + * The combination of this code and the -O3 optimizer flag on GCC 3.4.3
1.204 + * results in code that is 3 times faster than the previous NSS sha_fast
1.205 + * code on AMD64.
1.206 + */
1.207 +static void
1.208 +shaCompress(volatile unsigned *X, const uint32_t *inbuf)
1.209 +{
1.210 + unsigned A, B, C, D, E;
1.211 +
1.212 +#define XH(n) X[n - H2X]
1.213 +#define XW(n) X[n - W2X]
1.214 +
1.215 +#define K0 0x5a827999L
1.216 +#define K1 0x6ed9eba1L
1.217 +#define K2 0x8f1bbcdcL
1.218 +#define K3 0xca62c1d6L
1.219 +
1.220 +#define SHA_RND1(a, b, c, d, e, n) \
1.221 + a = SHA_ROTL(b, 5) + SHA_F1(c, d, e) + a + XW(n) + K0; c = SHA_ROTL(c, 30)
1.222 +#define SHA_RND2(a, b, c, d, e, n) \
1.223 + a = SHA_ROTL(b, 5) + SHA_F2(c, d, e) + a + XW(n) + K1; c = SHA_ROTL(c, 30)
1.224 +#define SHA_RND3(a, b, c, d, e, n) \
1.225 + a = SHA_ROTL(b, 5) + SHA_F3(c, d, e) + a + XW(n) + K2; c = SHA_ROTL(c, 30)
1.226 +#define SHA_RND4(a, b, c, d, e, n) \
1.227 + a = SHA_ROTL(b ,5) + SHA_F4(c, d, e) + a + XW(n) + K3; c = SHA_ROTL(c, 30)
1.228 +
1.229 +#define LOAD(n) XW(n) = NativeEndian::swapToBigEndian(inbuf[n])
1.230 +
1.231 + A = XH(0);
1.232 + B = XH(1);
1.233 + C = XH(2);
1.234 + D = XH(3);
1.235 + E = XH(4);
1.236 +
1.237 + LOAD(0); SHA_RND1(E,A,B,C,D, 0);
1.238 + LOAD(1); SHA_RND1(D,E,A,B,C, 1);
1.239 + LOAD(2); SHA_RND1(C,D,E,A,B, 2);
1.240 + LOAD(3); SHA_RND1(B,C,D,E,A, 3);
1.241 + LOAD(4); SHA_RND1(A,B,C,D,E, 4);
1.242 + LOAD(5); SHA_RND1(E,A,B,C,D, 5);
1.243 + LOAD(6); SHA_RND1(D,E,A,B,C, 6);
1.244 + LOAD(7); SHA_RND1(C,D,E,A,B, 7);
1.245 + LOAD(8); SHA_RND1(B,C,D,E,A, 8);
1.246 + LOAD(9); SHA_RND1(A,B,C,D,E, 9);
1.247 + LOAD(10); SHA_RND1(E,A,B,C,D,10);
1.248 + LOAD(11); SHA_RND1(D,E,A,B,C,11);
1.249 + LOAD(12); SHA_RND1(C,D,E,A,B,12);
1.250 + LOAD(13); SHA_RND1(B,C,D,E,A,13);
1.251 + LOAD(14); SHA_RND1(A,B,C,D,E,14);
1.252 + LOAD(15); SHA_RND1(E,A,B,C,D,15);
1.253 +
1.254 + SHA_MIX( 0, 13, 8, 2); SHA_RND1(D,E,A,B,C, 0);
1.255 + SHA_MIX( 1, 14, 9, 3); SHA_RND1(C,D,E,A,B, 1);
1.256 + SHA_MIX( 2, 15, 10, 4); SHA_RND1(B,C,D,E,A, 2);
1.257 + SHA_MIX( 3, 0, 11, 5); SHA_RND1(A,B,C,D,E, 3);
1.258 +
1.259 + SHA_MIX( 4, 1, 12, 6); SHA_RND2(E,A,B,C,D, 4);
1.260 + SHA_MIX( 5, 2, 13, 7); SHA_RND2(D,E,A,B,C, 5);
1.261 + SHA_MIX( 6, 3, 14, 8); SHA_RND2(C,D,E,A,B, 6);
1.262 + SHA_MIX( 7, 4, 15, 9); SHA_RND2(B,C,D,E,A, 7);
1.263 + SHA_MIX( 8, 5, 0, 10); SHA_RND2(A,B,C,D,E, 8);
1.264 + SHA_MIX( 9, 6, 1, 11); SHA_RND2(E,A,B,C,D, 9);
1.265 + SHA_MIX(10, 7, 2, 12); SHA_RND2(D,E,A,B,C,10);
1.266 + SHA_MIX(11, 8, 3, 13); SHA_RND2(C,D,E,A,B,11);
1.267 + SHA_MIX(12, 9, 4, 14); SHA_RND2(B,C,D,E,A,12);
1.268 + SHA_MIX(13, 10, 5, 15); SHA_RND2(A,B,C,D,E,13);
1.269 + SHA_MIX(14, 11, 6, 0); SHA_RND2(E,A,B,C,D,14);
1.270 + SHA_MIX(15, 12, 7, 1); SHA_RND2(D,E,A,B,C,15);
1.271 +
1.272 + SHA_MIX( 0, 13, 8, 2); SHA_RND2(C,D,E,A,B, 0);
1.273 + SHA_MIX( 1, 14, 9, 3); SHA_RND2(B,C,D,E,A, 1);
1.274 + SHA_MIX( 2, 15, 10, 4); SHA_RND2(A,B,C,D,E, 2);
1.275 + SHA_MIX( 3, 0, 11, 5); SHA_RND2(E,A,B,C,D, 3);
1.276 + SHA_MIX( 4, 1, 12, 6); SHA_RND2(D,E,A,B,C, 4);
1.277 + SHA_MIX( 5, 2, 13, 7); SHA_RND2(C,D,E,A,B, 5);
1.278 + SHA_MIX( 6, 3, 14, 8); SHA_RND2(B,C,D,E,A, 6);
1.279 + SHA_MIX( 7, 4, 15, 9); SHA_RND2(A,B,C,D,E, 7);
1.280 +
1.281 + SHA_MIX( 8, 5, 0, 10); SHA_RND3(E,A,B,C,D, 8);
1.282 + SHA_MIX( 9, 6, 1, 11); SHA_RND3(D,E,A,B,C, 9);
1.283 + SHA_MIX(10, 7, 2, 12); SHA_RND3(C,D,E,A,B,10);
1.284 + SHA_MIX(11, 8, 3, 13); SHA_RND3(B,C,D,E,A,11);
1.285 + SHA_MIX(12, 9, 4, 14); SHA_RND3(A,B,C,D,E,12);
1.286 + SHA_MIX(13, 10, 5, 15); SHA_RND3(E,A,B,C,D,13);
1.287 + SHA_MIX(14, 11, 6, 0); SHA_RND3(D,E,A,B,C,14);
1.288 + SHA_MIX(15, 12, 7, 1); SHA_RND3(C,D,E,A,B,15);
1.289 +
1.290 + SHA_MIX( 0, 13, 8, 2); SHA_RND3(B,C,D,E,A, 0);
1.291 + SHA_MIX( 1, 14, 9, 3); SHA_RND3(A,B,C,D,E, 1);
1.292 + SHA_MIX( 2, 15, 10, 4); SHA_RND3(E,A,B,C,D, 2);
1.293 + SHA_MIX( 3, 0, 11, 5); SHA_RND3(D,E,A,B,C, 3);
1.294 + SHA_MIX( 4, 1, 12, 6); SHA_RND3(C,D,E,A,B, 4);
1.295 + SHA_MIX( 5, 2, 13, 7); SHA_RND3(B,C,D,E,A, 5);
1.296 + SHA_MIX( 6, 3, 14, 8); SHA_RND3(A,B,C,D,E, 6);
1.297 + SHA_MIX( 7, 4, 15, 9); SHA_RND3(E,A,B,C,D, 7);
1.298 + SHA_MIX( 8, 5, 0, 10); SHA_RND3(D,E,A,B,C, 8);
1.299 + SHA_MIX( 9, 6, 1, 11); SHA_RND3(C,D,E,A,B, 9);
1.300 + SHA_MIX(10, 7, 2, 12); SHA_RND3(B,C,D,E,A,10);
1.301 + SHA_MIX(11, 8, 3, 13); SHA_RND3(A,B,C,D,E,11);
1.302 +
1.303 + SHA_MIX(12, 9, 4, 14); SHA_RND4(E,A,B,C,D,12);
1.304 + SHA_MIX(13, 10, 5, 15); SHA_RND4(D,E,A,B,C,13);
1.305 + SHA_MIX(14, 11, 6, 0); SHA_RND4(C,D,E,A,B,14);
1.306 + SHA_MIX(15, 12, 7, 1); SHA_RND4(B,C,D,E,A,15);
1.307 +
1.308 + SHA_MIX( 0, 13, 8, 2); SHA_RND4(A,B,C,D,E, 0);
1.309 + SHA_MIX( 1, 14, 9, 3); SHA_RND4(E,A,B,C,D, 1);
1.310 + SHA_MIX( 2, 15, 10, 4); SHA_RND4(D,E,A,B,C, 2);
1.311 + SHA_MIX( 3, 0, 11, 5); SHA_RND4(C,D,E,A,B, 3);
1.312 + SHA_MIX( 4, 1, 12, 6); SHA_RND4(B,C,D,E,A, 4);
1.313 + SHA_MIX( 5, 2, 13, 7); SHA_RND4(A,B,C,D,E, 5);
1.314 + SHA_MIX( 6, 3, 14, 8); SHA_RND4(E,A,B,C,D, 6);
1.315 + SHA_MIX( 7, 4, 15, 9); SHA_RND4(D,E,A,B,C, 7);
1.316 + SHA_MIX( 8, 5, 0, 10); SHA_RND4(C,D,E,A,B, 8);
1.317 + SHA_MIX( 9, 6, 1, 11); SHA_RND4(B,C,D,E,A, 9);
1.318 + SHA_MIX(10, 7, 2, 12); SHA_RND4(A,B,C,D,E,10);
1.319 + SHA_MIX(11, 8, 3, 13); SHA_RND4(E,A,B,C,D,11);
1.320 + SHA_MIX(12, 9, 4, 14); SHA_RND4(D,E,A,B,C,12);
1.321 + SHA_MIX(13, 10, 5, 15); SHA_RND4(C,D,E,A,B,13);
1.322 + SHA_MIX(14, 11, 6, 0); SHA_RND4(B,C,D,E,A,14);
1.323 + SHA_MIX(15, 12, 7, 1); SHA_RND4(A,B,C,D,E,15);
1.324 +
1.325 + XH(0) += A;
1.326 + XH(1) += B;
1.327 + XH(2) += C;
1.328 + XH(3) += D;
1.329 + XH(4) += E;
1.330 +}
