xref: /dragonfly/contrib/gmp/mpf/div_2exp.c (revision cfd1aba3)
1 /* mpf_div_2exp -- Divide a float by 2^n.
2 
3 Copyright 1993, 1994, 1996, 2000, 2001, 2002, 2004 Free Software Foundation,
4 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 the GNU Lesser General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or (at your
11 option) any later version.
12 
13 The GNU MP Library is distributed in the hope that it will be useful, but
14 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
15 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
16 License for more details.
17 
18 You should have received a copy of the GNU Lesser General Public License
19 along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.  */
20 
21 #include "gmp.h"
22 #include "gmp-impl.h"
23 
24 
25 /* Multiples of GMP_NUMB_BITS in exp simply mean an amount subtracted from
26    EXP(u) to set EXP(r).  The remainder exp%GMP_NUMB_BITS is then a right
27    shift for the limb data.
28 
29    If exp%GMP_NUMB_BITS == 0 then there's no shifting, we effectively just
30    do an mpz_set with changed EXP(r).  Like mpz_set we take prec+1 limbs in
31    this case.  Although just prec would suffice, it's nice to have
32    mpf_div_2exp with exp==0 come out the same as mpz_set.
33 
34    When shifting we take up to prec many limbs from the input.  Our shift is
35    cy = mpn_rshift (PTR(r)+1, PTR(u)+k, ...), where k is the number of low
36    limbs dropped from u, and the carry out is stored to PTR(r)[0].  We don't
37    try to work extra bits from PTR(u)[k-1] (when k>=1 makes it available)
38    into that low carry limb.  Just prec limbs (with the high non-zero) from
39    the input is enough bits for the application requested precision, no need
40    to do extra work.
41 
42    If r==u the shift will have overlapping operands.  When k>=1 (ie. when
43    usize > prec), the overlap is in the style supported by rshift (ie. dst
44    <= src).
45 
46    But when r==u and k==0 (ie. usize <= prec), we would have an invalid
47    overlap (mpn_rshift (rp+1, rp, ...)).  In this case we must instead use
48    mpn_lshift (PTR(r), PTR(u), size, NUMB-shift).  An lshift by NUMB-shift
49    bits gives identical data of course, it's just its overlap restrictions
50    which differ.
51 
52    In both shift cases, the resulting data is abs_usize+1 limbs.  "adj" is
53    used to add +1 to that size if the high is non-zero (it may of course
54    have become zero by the shifting).  EXP(u) is the exponent just above
55    those abs_usize+1 limbs, so it gets -1+adj, which means -1 if the high is
56    zero, or no change if the high is non-zero.
57 
58    Enhancements:
59 
60    The way mpn_lshift is used means successive mpf_div_2exp calls on the
61    same operand will accumulate low zero limbs, until prec+1 limbs is
62    reached.  This is wasteful for subsequent operations.  When abs_usize <=
63    prec, we should test the low exp%GMP_NUMB_BITS many bits of PTR(u)[0],
64    ie. those which would be shifted out by an mpn_rshift.  If they're zero
65    then use that mpn_rshift.  */
66 
67 void
68 mpf_div_2exp (mpf_ptr r, mpf_srcptr u, mp_bitcnt_t exp)
69 {
70   mp_srcptr up;
71   mp_ptr rp = r->_mp_d;
72   mp_size_t usize;
73   mp_size_t abs_usize;
74   mp_size_t prec = r->_mp_prec;
75   mp_exp_t uexp = u->_mp_exp;
76 
77   usize = u->_mp_size;
78 
79   if (UNLIKELY (usize == 0))
80     {
81       r->_mp_size = 0;
82       r->_mp_exp = 0;
83       return;
84     }
85 
86   abs_usize = ABS (usize);
87   up = u->_mp_d;
88 
89   if (exp % GMP_NUMB_BITS == 0)
90     {
91       prec++;			/* retain more precision here as we don't need
92 				   to account for carry-out here */
93       if (abs_usize > prec)
94 	{
95 	  up += abs_usize - prec;
96 	  abs_usize = prec;
97 	}
98       if (rp != up)
99 	MPN_COPY_INCR (rp, up, abs_usize);
100       r->_mp_exp = uexp - exp / GMP_NUMB_BITS;
101     }
102   else
103     {
104       mp_limb_t cy_limb;
105       mp_size_t adj;
106       if (abs_usize > prec)
107 	{
108 	  up += abs_usize - prec;
109 	  abs_usize = prec;
110 	  /* Use mpn_rshift since mpn_lshift operates downwards, and we
111 	     therefore would clobber part of U before using that part, in case
112 	     R is the same variable as U.  */
113 	  cy_limb = mpn_rshift (rp + 1, up, abs_usize, exp % GMP_NUMB_BITS);
114 	  rp[0] = cy_limb;
115 	  adj = rp[abs_usize] != 0;
116 	}
117       else
118 	{
119 	  cy_limb = mpn_lshift (rp, up, abs_usize,
120 				GMP_NUMB_BITS - exp % GMP_NUMB_BITS);
121 	  rp[abs_usize] = cy_limb;
122 	  adj = cy_limb != 0;
123 	}
124 
125       abs_usize += adj;
126       r->_mp_exp = uexp - exp / GMP_NUMB_BITS - 1 + adj;
127     }
128   r->_mp_size = usize >= 0 ? abs_usize : -abs_usize;
129 }
130