xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/cray/ieee/mul_basecase.c (revision 72c7faa4dbb41dbb0238d6b4a109da0d4b236dd4)
1 /* Cray PVP/IEEE mpn_mul_basecase.
2 
3 Copyright 2000, 2001 Free Software Foundation, Inc.
4 
5 This file is part of the GNU MP Library.
6 
7 The GNU MP Library is free software; you can redistribute it and/or modify
8 it under the terms of either:
9 
10   * the GNU Lesser General Public License as published by the Free
11     Software Foundation; either version 3 of the License, or (at your
12     option) any later version.
13 
14 or
15 
16   * the GNU General Public License as published by the Free Software
17     Foundation; either version 2 of the License, or (at your option) any
18     later version.
19 
20 or both in parallel, as here.
21 
22 The GNU MP Library is distributed in the hope that it will be useful, but
23 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
24 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
25 for more details.
26 
27 You should have received copies of the GNU General Public License and the
28 GNU Lesser General Public License along with the GNU MP Library.  If not,
29 see https://www.gnu.org/licenses/.  */
30 
31 /* The most critical loop of this code runs at about 5 cycles/limb on a T90.
32    That is not perfect, mainly due to vector register shortage.  */
33 
34 #include <intrinsics.h>
35 #include "gmp-impl.h"
36 
37 void
mpn_mul_basecase(mp_ptr rp,mp_srcptr up,mp_size_t un,mp_srcptr vp,mp_size_t vn)38 mpn_mul_basecase (mp_ptr rp,
39 		  mp_srcptr up, mp_size_t un,
40 		  mp_srcptr vp, mp_size_t vn)
41 {
42   mp_limb_t cy[un + vn];
43   mp_limb_t vl;
44   mp_limb_t a, b, r, s0, s1, c0, c1;
45   mp_size_t i, j;
46   int more_carries;
47 
48   for (i = 0; i < un + vn; i++)
49     {
50       rp[i] = 0;
51       cy[i] = 0;
52     }
53 
54 #pragma _CRI novector
55   for (j = 0; j < vn; j++)
56     {
57       vl = vp[j];
58 
59       a = up[0] * vl;
60       r = rp[j];
61       s0 = a + r;
62       rp[j] = s0;
63       c0 = ((a & r) | ((a | r) & ~s0)) >> 63;
64       cy[j] += c0;
65 
66 #pragma _CRI ivdep
67       for (i = 1; i < un; 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[j + i];
74 	  s1 = s0 + r;
75 	  rp[j + i] = s1;
76 	  c1 = ((s0 & r) | ((s0 | r) & ~s1)) >> 63;
77 	  cy[j + i] += c0 + c1;
78 	}
79       rp[j + un] = _int_mult_upper (up[un - 1], vl);
80     }
81 
82   more_carries = 0;
83 #pragma _CRI ivdep
84   for (i = 1; i < un + vn; i++)
85     {
86       r = rp[i];
87       c0 = cy[i - 1];
88       s0 = r + c0;
89       rp[i] = s0;
90       c0 = (r & ~s0) >> 63;
91       more_carries += c0;
92     }
93   /* If that second loop generated carry, handle that in scalar loop.  */
94   if (more_carries)
95     {
96       mp_limb_t cyrec = 0;
97       for (i = 1; i < un + vn; i++)
98 	{
99 	  r = rp[i];
100 	  c0 = (r < cy[i - 1]);
101 	  s0 = r + cyrec;
102 	  rp[i] = s0;
103 	  c1 = (r & ~s0) >> 63;
104 	  cyrec = c0 | c1;
105 	}
106     }
107 }
108