Austin Schuh | dace2a6 | 2020-08-18 10:56:48 -0700 | [diff] [blame^] | 1 | /* mpf_mul_2exp -- Multiply a float by 2^n. |
| 2 | |
| 3 | Copyright 1993, 1994, 1996, 2000-2002, 2004 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 | #include "gmp-impl.h" |
| 32 | |
| 33 | |
| 34 | /* Multiples of GMP_NUMB_BITS in exp simply mean an amount added to EXP(u) |
| 35 | to set EXP(r). The remainder exp%GMP_NUMB_BITS is then a left shift for |
| 36 | the limb data. |
| 37 | |
| 38 | If exp%GMP_NUMB_BITS == 0 then there's no shifting, we effectively just |
| 39 | do an mpz_set with changed EXP(r). Like mpz_set we take prec+1 limbs in |
| 40 | this case. Although just prec would suffice, it's nice to have |
| 41 | mpf_mul_2exp with exp==0 come out the same as mpz_set. |
| 42 | |
| 43 | When shifting we take up to prec many limbs from the input. Our shift is |
| 44 | cy = mpn_lshift (PTR(r), PTR(u)+k, size, ...), where k is the number of |
| 45 | low limbs dropped from u, and the carry out is stored to PTR(r)[size]. |
| 46 | |
| 47 | It may be noted that the low limb PTR(r)[0] doesn't incorporate bits from |
| 48 | PTR(u)[k-1] (when k>=1 makes that limb available). Taking just prec |
| 49 | limbs from the input (with the high non-zero) is enough bits for the |
| 50 | application requested precision, there's no need for extra work. |
| 51 | |
| 52 | If r==u the shift will have overlapping operands. When k==0 (ie. when |
| 53 | usize <= prec), the overlap is supported by lshift (ie. dst == src). |
| 54 | |
| 55 | But when r==u and k>=1 (ie. usize > prec), we would have an invalid |
| 56 | overlap (ie. mpn_lshift (rp, rp+k, ...)). In this case we must instead |
| 57 | use mpn_rshift (PTR(r)+1, PTR(u)+k, size, NUMB-shift) with the carry out |
| 58 | stored to PTR(r)[0]. An rshift by NUMB-shift bits like this gives |
| 59 | identical data, it's just its overlap restrictions which differ. |
| 60 | |
| 61 | Enhancements: |
| 62 | |
| 63 | The way mpn_lshift is used means successive mpf_mul_2exp calls on the |
| 64 | same operand will accumulate low zero limbs, until prec+1 limbs is |
| 65 | reached. This is wasteful for subsequent operations. When abs_usize <= |
| 66 | prec, we should test the low exp%GMP_NUMB_BITS many bits of PTR(u)[0], |
| 67 | ie. those which would be shifted out by an mpn_rshift. If they're zero |
| 68 | then use that mpn_rshift. */ |
| 69 | |
| 70 | void |
| 71 | mpf_mul_2exp (mpf_ptr r, mpf_srcptr u, mp_bitcnt_t exp) |
| 72 | { |
| 73 | mp_srcptr up; |
| 74 | mp_ptr rp = r->_mp_d; |
| 75 | mp_size_t usize; |
| 76 | mp_size_t abs_usize; |
| 77 | mp_size_t prec = r->_mp_prec; |
| 78 | mp_exp_t uexp = u->_mp_exp; |
| 79 | |
| 80 | usize = u->_mp_size; |
| 81 | |
| 82 | if (UNLIKELY (usize == 0)) |
| 83 | { |
| 84 | r->_mp_size = 0; |
| 85 | r->_mp_exp = 0; |
| 86 | return; |
| 87 | } |
| 88 | |
| 89 | abs_usize = ABS (usize); |
| 90 | up = u->_mp_d; |
| 91 | |
| 92 | if (exp % GMP_NUMB_BITS == 0) |
| 93 | { |
| 94 | prec++; /* retain more precision here as we don't need |
| 95 | to account for carry-out here */ |
| 96 | if (abs_usize > prec) |
| 97 | { |
| 98 | up += abs_usize - prec; |
| 99 | abs_usize = prec; |
| 100 | } |
| 101 | if (rp != up) |
| 102 | MPN_COPY_INCR (rp, up, abs_usize); |
| 103 | r->_mp_exp = uexp + exp / GMP_NUMB_BITS; |
| 104 | } |
| 105 | else |
| 106 | { |
| 107 | mp_limb_t cy_limb; |
| 108 | mp_size_t adj; |
| 109 | if (abs_usize > prec) |
| 110 | { |
| 111 | up += abs_usize - prec; |
| 112 | abs_usize = prec; |
| 113 | /* Use mpn_rshift since mpn_lshift operates downwards, and we |
| 114 | therefore would clobber part of U before using that part, in case |
| 115 | R is the same variable as U. */ |
| 116 | cy_limb = mpn_rshift (rp + 1, up, abs_usize, |
| 117 | GMP_NUMB_BITS - exp % GMP_NUMB_BITS); |
| 118 | rp[0] = cy_limb; |
| 119 | adj = rp[abs_usize] != 0; |
| 120 | } |
| 121 | else |
| 122 | { |
| 123 | cy_limb = mpn_lshift (rp, up, abs_usize, exp % GMP_NUMB_BITS); |
| 124 | rp[abs_usize] = cy_limb; |
| 125 | adj = cy_limb != 0; |
| 126 | } |
| 127 | |
| 128 | abs_usize += adj; |
| 129 | r->_mp_exp = uexp + exp / GMP_NUMB_BITS + adj; |
| 130 | } |
| 131 | r->_mp_size = usize >= 0 ? abs_usize : -abs_usize; |
| 132 | } |