1 /* Cray PVP/IEEE mpn_addmul_1 -- multiply a limb vector with a limb and add the
2 result to a second limb vector.
3
4 Copyright 2000-2002 Free Software Foundation, Inc.
5
6 This file is part of the GNU MP Library.
7
8 The GNU MP Library is free software; you can redistribute it and/or modify
9 it under the terms of either:
10
11 * the GNU Lesser General Public License as published by the Free
12 Software Foundation; either version 3 of the License, or (at your
13 option) any later version.
14
15 or
16
17 * the GNU General Public License as published by the Free Software
18 Foundation; either version 2 of the License, or (at your option) any
19 later version.
20
21 or both in parallel, as here.
22
23 The GNU MP Library is distributed in the hope that it will be useful, but
24 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
25 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
26 for more details.
27
28 You should have received copies of the GNU General Public License and the
29 GNU Lesser General Public License along with the GNU MP Library. If not,
30 see https://www.gnu.org/licenses/. */
31
32 /* This code runs at just under 9 cycles/limb on a T90. That is not perfect,
33 mainly due to vector register shortage in the main loop. Assembly code
34 should bring it down to perhaps 7 cycles/limb. */
35
36 #include <intrinsics.h>
37 #include "gmp.h"
38 #include "gmp-impl.h"
39
40 mp_limb_t
mpn_addmul_1(mp_ptr rp,mp_srcptr up,mp_size_t n,mp_limb_t vl)41 mpn_addmul_1 (mp_ptr rp, mp_srcptr up, mp_size_t n, mp_limb_t vl)
42 {
43 mp_limb_t cy[n];
44 mp_limb_t a, b, r, s0, s1, c0, c1;
45 mp_size_t i;
46 int more_carries;
47
48 if (up == rp)
49 {
50 /* The algorithm used below cannot handle overlap. Handle it here by
51 making a temporary copy of the source vector, then call ourselves. */
52 mp_limb_t xp[n];
53 MPN_COPY (xp, up, n);
54 return mpn_addmul_1 (rp, xp, n, vl);
55 }
56
57 a = up[0] * vl;
58 r = rp[0];
59 s0 = a + r;
60 rp[0] = s0;
61 c0 = ((a & r) | ((a | r) & ~s0)) >> 63;
62 cy[0] = c0;
63
64 /* Main multiply loop. Generate a raw accumulated output product in rp[]
65 and a carry vector in cy[]. */
66 #pragma _CRI ivdep
67 for (i = 1; i < n; i++)
68 {
69 a = up[i] * vl;
70 b = _int_mult_upper (up[i - 1], vl);
71 s0 = a + b;
72 c0 = ((a & b) | ((a | b) & ~s0)) >> 63;
73 r = rp[i];
74 s1 = s0 + r;
75 rp[i] = s1;
76 c1 = ((s0 & r) | ((s0 | r) & ~s1)) >> 63;
77 cy[i] = c0 + c1;
78 }
79 /* Carry add loop. Add the carry vector cy[] to the raw result rp[] and
80 store the new result back to rp[]. */
81 more_carries = 0;
82 #pragma _CRI ivdep
83 for (i = 1; i < n; i++)
84 {
85 r = rp[i];
86 c0 = cy[i - 1];
87 s0 = r + c0;
88 rp[i] = s0;
89 c0 = (r & ~s0) >> 63;
90 more_carries += c0;
91 }
92 /* If that second loop generated carry, handle that in scalar loop. */
93 if (more_carries)
94 {
95 mp_limb_t cyrec = 0;
96 /* Look for places where rp[k] == 0 and cy[k-1] == 1 or
97 rp[k] == 1 and cy[k-1] == 2.
98 These are where we got a recurrency carry. */
99 for (i = 1; i < n; i++)
100 {
101 r = rp[i];
102 c0 = r < cy[i - 1];
103 s0 = r + cyrec;
104 rp[i] = s0;
105 c1 = (r & ~s0) >> 63;
106 cyrec = c0 | c1;
107 }
108 return _int_mult_upper (up[n - 1], vl) + cyrec + cy[n - 1];
109 }
110
111 return _int_mult_upper (up[n - 1], vl) + cy[n - 1];
112 }
113