1/*	$OpenBSD: e_expl.c,v 1.4 2016/09/12 19:47:02 guenther Exp $	*/
2
3/*
4 * Copyright (c) 2008 Stephen L. Moshier <steve@moshier.net>
5 *
6 * Permission to use, copy, modify, and distribute this software for any
7 * purpose with or without fee is hereby granted, provided that the above
8 * copyright notice and this permission notice appear in all copies.
9 *
10 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
11 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
12 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
13 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
14 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
15 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
16 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
17 */
18
19/*							expl.c
20 *
21 *	Exponential function, 128-bit long double precision
22 *
23 *
24 *
25 * SYNOPSIS:
26 *
27 * long double x, y, expl();
28 *
29 * y = expl( x );
30 *
31 *
32 *
33 * DESCRIPTION:
34 *
35 * Returns e (2.71828...) raised to the x power.
36 *
37 * Range reduction is accomplished by separating the argument
38 * into an integer k and fraction f such that
39 *
40 *     x    k  f
41 *    e  = 2  e.
42 *
43 * A Pade' form of degree 2/3 is used to approximate exp(f) - 1
44 * in the basic range [-0.5 ln 2, 0.5 ln 2].
45 *
46 *
47 * ACCURACY:
48 *
49 *                      Relative error:
50 * arithmetic   domain     # trials      peak         rms
51 *    IEEE      +-MAXLOG    100,000     2.6e-34     8.6e-35
52 *
53 *
54 * Error amplification in the exponential function can be
55 * a serious matter.  The error propagation involves
56 * exp( X(1+delta) ) = exp(X) ( 1 + X*delta + ... ),
57 * which shows that a 1 lsb error in representing X produces
58 * a relative error of X times 1 lsb in the function.
59 * While the routine gives an accurate result for arguments
60 * that are exactly represented by a long double precision
61 * computer number, the result contains amplified roundoff
62 * error for large arguments not exactly represented.
63 *
64 *
65 * ERROR MESSAGES:
66 *
67 *   message         condition      value returned
68 * exp underflow    x < MINLOG         0.0
69 * exp overflow     x > MAXLOG         MAXNUM
70 *
71 */
72
73/*	Exponential function	*/
74
75#include <float.h>
76#include <math.h>
77
78#include "math_private.h"
79
80/* Pade' coefficients for exp(x) - 1
81   Theoretical peak relative error = 2.2e-37,
82   relative peak error spread = 9.2e-38
83 */
84static long double P[5] = {
85 3.279723985560247033712687707263393506266E-10L,
86 6.141506007208645008909088812338454698548E-7L,
87 2.708775201978218837374512615596512792224E-4L,
88 3.508710990737834361215404761139478627390E-2L,
89 9.999999999999999999999999999999999998502E-1L
90};
91static long double Q[6] = {
92 2.980756652081995192255342779918052538681E-12L,
93 1.771372078166251484503904874657985291164E-8L,
94 1.504792651814944826817779302637284053660E-5L,
95 3.611828913847589925056132680618007270344E-3L,
96 2.368408864814233538909747618894558968880E-1L,
97 2.000000000000000000000000000000000000150E0L
98};
99/* C1 + C2 = ln 2 */
100static const long double C1 = -6.93145751953125E-1L;
101static const long double C2 = -1.428606820309417232121458176568075500134E-6L;
102
103static const long double LOG2EL = 1.442695040888963407359924681001892137426646L;
104static const long double MAXLOGL = 1.1356523406294143949491931077970764891253E4L;
105static const long double MINLOGL = -1.143276959615573793352782661133116431383730e4L;
106static const long double huge = 0x1p10000L;
107#if 0 /* XXX Prevent gcc from erroneously constant folding this. */
108static const long double twom10000 = 0x1p-10000L;
109#else
110static volatile long double twom10000 = 0x1p-10000L;
111#endif
112
113long double
114expl(long double x)
115{
116long double px, xx;
117int n;
118
119if( x > MAXLOGL)
120	return (huge*huge);		/* overflow */
121
122if( x < MINLOGL )
123	return (twom10000*twom10000);	/* underflow */
124
125/* Express e**x = e**g 2**n
126 *   = e**g e**( n loge(2) )
127 *   = e**( g + n loge(2) )
128 */
129px = floorl( LOG2EL * x + 0.5L ); /* floor() truncates toward -infinity. */
130n = px;
131x += px * C1;
132x += px * C2;
133/* rational approximation for exponential
134 * of the fractional part:
135 * e**x =  1 + 2x P(x**2)/( Q(x**2) - P(x**2) )
136 */
137xx = x * x;
138px = x * __polevll( xx, P, 4 );
139xx = __polevll( xx, Q, 5 );
140x =  px/( xx - px );
141x = 1.0L + x + x;
142
143x = ldexpl( x, n );
144return(x);
145}
146DEF_STD(expl);
147