mul_2exp.c revision 1.1.1.1
1/* mpf_mul_2exp -- Multiply a float by 2^n. 2 3Copyright 1993, 1994, 1996, 2000, 2001, 2002, 2004 Free Software Foundation, 4Inc. 5 6This file is part of the GNU MP Library. 7 8The GNU MP Library is free software; you can redistribute it and/or modify 9it under the terms of the GNU Lesser General Public License as published by 10the Free Software Foundation; either version 3 of the License, or (at your 11option) any later version. 12 13The GNU MP Library is distributed in the hope that it will be useful, but 14WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 15or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 16License for more details. 17 18You should have received a copy of the GNU Lesser General Public License 19along 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 added to EXP(u) 26 to set EXP(r). The remainder exp%GMP_NUMB_BITS is then a left shift for 27 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_mul_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_lshift (PTR(r), PTR(u)+k, size, ...), where k is the number of 36 low limbs dropped from u, and the carry out is stored to PTR(r)[size]. 37 38 It may be noted that the low limb PTR(r)[0] doesn't incorporate bits from 39 PTR(u)[k-1] (when k>=1 makes that limb available). Taking just prec 40 limbs from the input (with the high non-zero) is enough bits for the 41 application requested precision, there's no need for extra work. 42 43 If r==u the shift will have overlapping operands. When k==0 (ie. when 44 usize <= prec), the overlap is supported by lshift (ie. dst == src). 45 46 But when r==u and k>=1 (ie. usize > prec), we would have an invalid 47 overlap (ie. mpn_lshift (rp, rp+k, ...)). In this case we must instead 48 use mpn_rshift (PTR(r)+1, PTR(u)+k, size, NUMB-shift) with the carry out 49 stored to PTR(r)[0]. An rshift by NUMB-shift bits like this gives 50 identical data, it's just its overlap restrictions which differ. 51 52 Enhancements: 53 54 The way mpn_lshift is used means successive mpf_mul_2exp calls on the 55 same operand will accumulate low zero limbs, until prec+1 limbs is 56 reached. This is wasteful for subsequent operations. When abs_usize <= 57 prec, we should test the low exp%GMP_NUMB_BITS many bits of PTR(u)[0], 58 ie. those which would be shifted out by an mpn_rshift. If they're zero 59 then use that mpn_rshift. */ 60 61void 62mpf_mul_2exp (mpf_ptr r, mpf_srcptr u, mp_bitcnt_t exp) 63{ 64 mp_srcptr up; 65 mp_ptr rp = r->_mp_d; 66 mp_size_t usize; 67 mp_size_t abs_usize; 68 mp_size_t prec = r->_mp_prec; 69 mp_exp_t uexp = u->_mp_exp; 70 71 usize = u->_mp_size; 72 73 if (UNLIKELY (usize == 0)) 74 { 75 r->_mp_size = 0; 76 r->_mp_exp = 0; 77 return; 78 } 79 80 abs_usize = ABS (usize); 81 up = u->_mp_d; 82 83 if (exp % GMP_NUMB_BITS == 0) 84 { 85 prec++; /* retain more precision here as we don't need 86 to account for carry-out here */ 87 if (abs_usize > prec) 88 { 89 up += abs_usize - prec; 90 abs_usize = prec; 91 } 92 if (rp != up) 93 MPN_COPY_INCR (rp, up, abs_usize); 94 r->_mp_exp = uexp + exp / GMP_NUMB_BITS; 95 } 96 else 97 { 98 mp_limb_t cy_limb; 99 mp_size_t adj; 100 if (abs_usize > prec) 101 { 102 up += abs_usize - prec; 103 abs_usize = prec; 104 /* Use mpn_rshift since mpn_lshift operates downwards, and we 105 therefore would clobber part of U before using that part, in case 106 R is the same variable as U. */ 107 cy_limb = mpn_rshift (rp + 1, up, abs_usize, 108 GMP_NUMB_BITS - exp % GMP_NUMB_BITS); 109 rp[0] = cy_limb; 110 adj = rp[abs_usize] != 0; 111 } 112 else 113 { 114 cy_limb = mpn_lshift (rp, up, abs_usize, exp % GMP_NUMB_BITS); 115 rp[abs_usize] = cy_limb; 116 adj = cy_limb != 0; 117 } 118 119 abs_usize += adj; 120 r->_mp_exp = uexp + exp / GMP_NUMB_BITS + adj; 121 } 122 r->_mp_size = usize >= 0 ? abs_usize : -abs_usize; 123} 124