| 1 | /* Double-precision floating point 2^x. |
| 2 | Copyright (C) 1997-2016 Free Software Foundation, Inc. |
| 3 | This file is part of the GNU C Library. |
| 4 | Contributed by Geoffrey Keating <geoffk@ozemail.com.au> |
| 5 | |
| 6 | The GNU C Library is free software; you can redistribute it and/or |
| 7 | modify it under the terms of the GNU Lesser General Public |
| 8 | License as published by the Free Software Foundation; either |
| 9 | version 2.1 of the License, or (at your option) any later version. |
| 10 | |
| 11 | The GNU C Library is distributed in the hope that it will be useful, |
| 12 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 | Lesser General Public License for more details. |
| 15 | |
| 16 | You should have received a copy of the GNU Lesser General Public |
| 17 | License along with the GNU C Library; if not, see |
| 18 | <http://www.gnu.org/licenses/>. */ |
| 19 | |
| 20 | /* The basic design here is from |
| 21 | Shmuel Gal and Boris Bachelis, "An Accurate Elementary Mathematical |
| 22 | Library for the IEEE Floating Point Standard", ACM Trans. Math. Soft., |
| 23 | 17 (1), March 1991, pp. 26-45. |
| 24 | It has been slightly modified to compute 2^x instead of e^x. |
| 25 | */ |
| 26 | #include <stdlib.h> |
| 27 | #include <float.h> |
| 28 | #include <ieee754.h> |
| 29 | #include <math.h> |
| 30 | #include <fenv.h> |
| 31 | #include <inttypes.h> |
| 32 | #include <math_private.h> |
| 33 | |
| 34 | #include "t_exp2.h" |
| 35 | |
| 36 | static const double TWO1023 = 8.988465674311579539e+307; |
| 37 | static const double TWOM1000 = 9.3326361850321887899e-302; |
| 38 | |
| 39 | double |
| 40 | __ieee754_exp2 (double x) |
| 41 | { |
| 42 | static const double himark = (double) DBL_MAX_EXP; |
| 43 | static const double lomark = (double) (DBL_MIN_EXP - DBL_MANT_DIG - 1); |
| 44 | |
| 45 | /* Check for usual case. */ |
| 46 | if (__glibc_likely (isless (x, himark))) |
| 47 | { |
| 48 | /* Exceptional cases: */ |
| 49 | if (__glibc_unlikely (!isgreaterequal (x, lomark))) |
| 50 | { |
| 51 | if (isinf (x)) |
| 52 | /* e^-inf == 0, with no error. */ |
| 53 | return 0; |
| 54 | else |
| 55 | /* Underflow */ |
| 56 | return TWOM1000 * TWOM1000; |
| 57 | } |
| 58 | |
| 59 | static const double THREEp42 = 13194139533312.0; |
| 60 | int tval, unsafe; |
| 61 | double rx, x22, result; |
| 62 | union ieee754_double ex2_u, scale_u; |
| 63 | |
| 64 | if (fabs (x) < DBL_EPSILON / 4.0) |
| 65 | return 1.0 + x; |
| 66 | |
| 67 | { |
| 68 | SET_RESTORE_ROUND_NOEX (FE_TONEAREST); |
| 69 | |
| 70 | /* 1. Argument reduction. |
| 71 | Choose integers ex, -256 <= t < 256, and some real |
| 72 | -1/1024 <= x1 <= 1024 so that |
| 73 | x = ex + t/512 + x1. |
| 74 | |
| 75 | First, calculate rx = ex + t/512. */ |
| 76 | rx = x + THREEp42; |
| 77 | rx -= THREEp42; |
| 78 | x -= rx; /* Compute x=x1. */ |
| 79 | /* Compute tval = (ex*512 + t)+256. |
| 80 | Now, t = (tval mod 512)-256 and ex=tval/512 [that's mod, NOT %; |
| 81 | and /-round-to-nearest not the usual c integer /]. */ |
| 82 | tval = (int) (rx * 512.0 + 256.0); |
| 83 | |
| 84 | /* 2. Adjust for accurate table entry. |
| 85 | Find e so that |
| 86 | x = ex + t/512 + e + x2 |
| 87 | where -1e6 < e < 1e6, and |
| 88 | (double)(2^(t/512+e)) |
| 89 | is accurate to one part in 2^-64. */ |
| 90 | |
| 91 | /* 'tval & 511' is the same as 'tval%512' except that it's always |
| 92 | positive. |
| 93 | Compute x = x2. */ |
| 94 | x -= exp2_deltatable[tval & 511]; |
| 95 | |
| 96 | /* 3. Compute ex2 = 2^(t/512+e+ex). */ |
| 97 | ex2_u.d = exp2_accuratetable[tval & 511]; |
| 98 | tval >>= 9; |
| 99 | /* x2 is an integer multiple of 2^-54; avoid intermediate |
| 100 | underflow from the calculation of x22 * x. */ |
| 101 | unsafe = abs (tval) >= -DBL_MIN_EXP - 56; |
| 102 | ex2_u.ieee.exponent += tval >> unsafe; |
| 103 | scale_u.d = 1.0; |
| 104 | scale_u.ieee.exponent += tval - (tval >> unsafe); |
| 105 | |
| 106 | /* 4. Approximate 2^x2 - 1, using a fourth-degree polynomial, |
| 107 | with maximum error in [-2^-10-2^-30,2^-10+2^-30] |
| 108 | less than 10^-19. */ |
| 109 | |
| 110 | x22 = (((.0096181293647031180 |
| 111 | * x + .055504110254308625) |
| 112 | * x + .240226506959100583) |
| 113 | * x + .69314718055994495) * ex2_u.d; |
| 114 | math_opt_barrier (x22); |
| 115 | } |
| 116 | |
| 117 | /* 5. Return (2^x2-1) * 2^(t/512+e+ex) + 2^(t/512+e+ex). */ |
| 118 | result = x22 * x + ex2_u.d; |
| 119 | |
| 120 | if (!unsafe) |
| 121 | return result; |
| 122 | else |
| 123 | { |
| 124 | result *= scale_u.d; |
| 125 | math_check_force_underflow_nonneg (result); |
| 126 | return result; |
| 127 | } |
| 128 | } |
| 129 | else |
| 130 | /* Return x, if x is a NaN or Inf; or overflow, otherwise. */ |
| 131 | return TWO1023 * x; |
| 132 | } |
| 133 | strong_alias (__ieee754_exp2, __exp2_finite) |
| 134 | |