Mercurial > dropbear
annotate libtommath/bn_s_mp_mul_digs_fast.c @ 1714:c0f12eaf95c9
Disable by default 3des, cbc, hmac-sha1-96, x11 forwarding
author | Matt Johnston <matt@ucc.asn.au> |
---|---|
date | Thu, 11 Jun 2020 00:09:15 +0800 |
parents | 1051e4eea25a |
children |
rev | line source |
---|---|
1692
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
1 #include "tommath_private.h" |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
2 #ifdef BN_S_MP_MUL_DIGS_FAST_C |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
3 /* LibTomMath, multiple-precision integer library -- Tom St Denis */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
4 /* SPDX-License-Identifier: Unlicense */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
5 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
6 /* Fast (comba) multiplier |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
7 * |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
8 * This is the fast column-array [comba] multiplier. It is |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
9 * designed to compute the columns of the product first |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
10 * then handle the carries afterwards. This has the effect |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
11 * of making the nested loops that compute the columns very |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
12 * simple and schedulable on super-scalar processors. |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
13 * |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
14 * This has been modified to produce a variable number of |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
15 * digits of output so if say only a half-product is required |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
16 * you don't have to compute the upper half (a feature |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
17 * required for fast Barrett reduction). |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
18 * |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
19 * Based on Algorithm 14.12 on pp.595 of HAC. |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
20 * |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
21 */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
22 mp_err s_mp_mul_digs_fast(const mp_int *a, const mp_int *b, mp_int *c, int digs) |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
23 { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
24 int olduse, pa, ix, iz; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
25 mp_err err; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
26 mp_digit W[MP_WARRAY]; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
27 mp_word _W; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
28 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
29 /* grow the destination as required */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
30 if (c->alloc < digs) { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
31 if ((err = mp_grow(c, digs)) != MP_OKAY) { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
32 return err; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
33 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
34 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
35 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
36 /* number of output digits to produce */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
37 pa = MP_MIN(digs, a->used + b->used); |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
38 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
39 /* clear the carry */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
40 _W = 0; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
41 for (ix = 0; ix < pa; ix++) { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
42 int tx, ty; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
43 int iy; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
44 mp_digit *tmpx, *tmpy; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
45 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
46 /* get offsets into the two bignums */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
47 ty = MP_MIN(b->used-1, ix); |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
48 tx = ix - ty; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
49 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
50 /* setup temp aliases */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
51 tmpx = a->dp + tx; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
52 tmpy = b->dp + ty; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
53 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
54 /* this is the number of times the loop will iterrate, essentially |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
55 while (tx++ < a->used && ty-- >= 0) { ... } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
56 */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
57 iy = MP_MIN(a->used-tx, ty+1); |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
58 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
59 /* execute loop */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
60 for (iz = 0; iz < iy; ++iz) { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
61 _W += (mp_word)*tmpx++ * (mp_word)*tmpy--; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
62 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
63 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
64 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
65 /* store term */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
66 W[ix] = (mp_digit)_W & MP_MASK; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
67 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
68 /* make next carry */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
69 _W = _W >> (mp_word)MP_DIGIT_BIT; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
70 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
71 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
72 /* setup dest */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
73 olduse = c->used; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
74 c->used = pa; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
75 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
76 { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
77 mp_digit *tmpc; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
78 tmpc = c->dp; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
79 for (ix = 0; ix < pa; ix++) { |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
80 /* now extract the previous digit [below the carry] */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
81 *tmpc++ = W[ix]; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
82 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
83 |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
84 /* clear unused digits [that existed in the old copy of c] */ |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
85 MP_ZERO_DIGITS(tmpc, olduse - ix); |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
86 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
87 mp_clamp(c); |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
88 return MP_OKAY; |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
89 } |
1051e4eea25a
Update LibTomMath to 1.2.0 (#84)
Steffen Jaeckel <s@jaeckel.eu>
parents:
diff
changeset
|
90 #endif |