2 * Copyright 2017 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright 2014 Cryptography Research, Inc.
5 * Licensed under the OpenSSL license (the "License"). You may not use
6 * this file except in compliance with the License. You can obtain a copy
7 * in the file LICENSE in the source distribution or at
8 * https://www.openssl.org/source/license.html
10 * Originally written by Mike Hamburg
15 #if (defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) && !I_HATE_UNROLLED_LOOPS) \
16 || defined(DECAF_FORCE_UNROLL)
17 # define REPEAT8(_x) _x _x _x _x _x _x _x _x
18 # define FOR_LIMB(_i,_start,_end,_x) do { _i=_start; REPEAT8( if (_i<_end) { _x; } _i++;) } while (0)
20 # define FOR_LIMB(_i,_start,_end,_x) do { for (_i=_start; _i<_end; _i++) _x; } while (0)
23 void gf_mul(gf_s * __restrict__ cs, const gf as, const gf bs)
25 const uint32_t *a = as->limb, *b = bs->limb;
26 uint32_t *c = cs->limb;
28 uint64_t accum0 = 0, accum1 = 0, accum2 = 0;
29 uint32_t mask = (1 << 28) - 1;
31 uint32_t aa[8], bb[8];
34 for (i = 0; i < 8; i++) {
35 aa[i] = a[i] + a[i + 8];
36 bb[i] = b[i] + b[i + 8];
41 FOR_LIMB(i, 0, j + 1, {
42 accum2 += widemul(a[j - i], b[i]);
43 accum1 += widemul(aa[j - i], bb[i]);
44 accum0 += widemul(a[8 + j - i], b[8 + i]);
46 ); accum1 -= accum2; accum0 += accum2;
48 FOR_LIMB(i, j + 1, 8, {
50 widemul(a[8 + j - i], b[i]);
52 widemul(aa[8 + j - i],
54 accum1 += widemul(a[16 + j - i], b[8 + i]);
59 c[j] = ((uint32_t)(accum0)) & mask;
60 c[j + 8] = ((uint32_t)(accum1)) & mask;
61 accum0 >>= 28; accum1 >>= 28;
67 c[8] = ((uint32_t)(accum0)) & mask;
68 c[0] = ((uint32_t)(accum1)) & mask;
72 c[9] += ((uint32_t)(accum0));
73 c[1] += ((uint32_t)(accum1));
76 void gf_mulw_unsigned(gf_s * __restrict__ cs, const gf as, uint32_t b)
78 const uint32_t *a = as->limb;
79 uint32_t *c = cs->limb;
80 uint64_t accum0 = 0, accum8 = 0;
81 uint32_t mask = (1 << 28) - 1;
87 accum0 += widemul(b, a[i]); accum8 += widemul(b, a[i + 8]);
88 c[i] = accum0 & mask; accum0 >>= 28;
89 c[i + 8] = accum8 & mask; accum8 >>= 28;
92 accum0 += accum8 + c[8];
101 void gf_sqr(gf_s * __restrict__ cs, const gf as)
103 gf_mul(cs, as, as); /* Performs better with a dedicated square */