1 /* mpn_divexact_1 -- mpn by limb exact division.
2
3 THE FUNCTIONS IN THIS FILE ARE FOR INTERNAL USE ONLY. THEY'RE ALMOST
4 CERTAIN TO BE SUBJECT TO INCOMPATIBLE CHANGES OR DISAPPEAR COMPLETELY IN
5 FUTURE GNU MP RELEASES.
6
7 Copyright 2000, 2001, 2002, 2003, 2005 Free Software Foundation, Inc.
8
9 This file is part of the GNU MP Library.
10
11 The GNU MP Library is free software; you can redistribute it and/or modify
12 it under the terms of the GNU Lesser General Public License as published by
13 the Free Software Foundation; either version 3 of the License, or (at your
14 option) any later version.
15
16 The GNU MP Library is distributed in the hope that it will be useful, but
17 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
18 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
19 License for more details.
20
21 You should have received a copy of the GNU Lesser General Public License
22 along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. */
23
24 #include "gmp.h"
25 #include "gmp-impl.h"
26 #include "longlong.h"
27
28
29
30 /* Divide a={src,size} by d=divisor and store the quotient in q={dst,size}.
31 q will only be correct if d divides a exactly.
32
33 A separate loop is used for shift==0 because n<<GMP_LIMB_BITS doesn't
34 give zero on all CPUs (for instance it doesn't on the x86s). This
35 separate loop might run faster too, helping odd divisors.
36
37 Possibilities:
38
39 mpn_divexact_1c could be created, accepting and returning c. This would
40 let a long calculation be done piece by piece. Currently there's no
41 particular need for that, and not returning c means that a final umul can
42 be skipped.
43
44 Another use for returning c would be letting the caller know whether the
45 division was in fact exact. It would work just to return the carry bit
46 "c=(l>s)" and let the caller do a final umul if interested.
47
48 When the divisor is even, the factors of two could be handled with a
49 separate mpn_rshift, instead of shifting on the fly. That might be
50 faster on some CPUs and would mean just the shift==0 style loop would be
51 needed.
52
53 If n<<GMP_LIMB_BITS gives zero on a particular CPU then the separate
54 shift==0 loop is unnecessary, and could be eliminated if there's no great
55 speed difference.
56
57 It's not clear whether "/" is the best way to handle size==1. Alpha gcc
58 2.95 for instance has a poor "/" and might prefer the modular method.
59 Perhaps a tuned parameter should control this.
60
61 If src[size-1] < divisor then dst[size-1] will be zero, and one divide
62 step could be skipped. A test at last step for s<divisor (or ls in the
63 even case) might be a good way to do that. But if this code is often
64 used with small divisors then it might not be worth bothering */
65
66 void
mpn_divexact_1(mp_ptr dst,mp_srcptr src,mp_size_t size,mp_limb_t divisor)67 mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor)
68 {
69 mp_size_t i;
70 mp_limb_t c, h, l, ls, s, s_next, inverse, dummy;
71 unsigned shift;
72
73 ASSERT (size >= 1);
74 ASSERT (divisor != 0);
75 ASSERT (MPN_SAME_OR_SEPARATE_P (dst, src, size));
76 ASSERT_MPN (src, size);
77 ASSERT_LIMB (divisor);
78
79 s = src[0];
80
81 if (size == 1)
82 {
83 dst[0] = s / divisor;
84 return;
85 }
86
87 if ((divisor & 1) == 0)
88 {
89 count_trailing_zeros (shift, divisor);
90 divisor >>= shift;
91 }
92 else
93 shift = 0;
94
95 binvert_limb (inverse, divisor);
96 divisor <<= GMP_NAIL_BITS;
97
98 if (shift != 0)
99 {
100 c = 0;
101 i = 0;
102 size--;
103
104 do
105 {
106 s_next = src[i+1];
107 ls = ((s >> shift) | (s_next << (GMP_NUMB_BITS-shift))) & GMP_NUMB_MASK;
108 s = s_next;
109
110 SUBC_LIMB (c, l, ls, c);
111
112 l = (l * inverse) & GMP_NUMB_MASK;
113 dst[i] = l;
114
115 umul_ppmm (h, dummy, l, divisor);
116 c += h;
117
118 i++;
119 }
120 while (i < size);
121
122 ls = s >> shift;
123 l = ls - c;
124 l = (l * inverse) & GMP_NUMB_MASK;
125 dst[i] = l;
126 }
127 else
128 {
129 l = (s * inverse) & GMP_NUMB_MASK;
130 dst[0] = l;
131 i = 1;
132 c = 0;
133
134 do
135 {
136 umul_ppmm (h, dummy, l, divisor);
137 c += h;
138
139 s = src[i];
140 SUBC_LIMB (c, l, s, c);
141
142 l = (l * inverse) & GMP_NUMB_MASK;
143 dst[i] = l;
144 i++;
145 }
146 while (i < size);
147 }
148 }
149