xref: /illumos-gate/usr/src/lib/libm/common/Q/expm1l.c (revision d8e10381a0083d7717710b0db7e64707bc0f3ff8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __expm1l = expm1l
31 
32 #if !defined(__sparc)
33 #error Unsupported architecture
34 #endif
35 
36 /*
37  * expm1l(x)
38  *
39  * Table driven method
40  * Written by K.C. Ng, June 1995.
41  * Algorithm :
42  *	1. expm1(x) = x if x<2**-114
43  *	2. if |x| <= 0.0625 = 1/16, use approximation
44  *		expm1(x) = x + x*P/(2-P)
45  * where
46  * 	P = x - z*(P1+z*(P2+z*(P3+z*(P4+z*(P5+z*P6+z*P7))))), z = x*x;
47  * (this formula is derived from
48  *	2-P+x = R = x*(exp(x)+1)/(exp(x)-1) ~ 2 + x*x/6 - x^4/360 + ...)
49  *
50  * P1 =   1.66666666666666666666666666666638500528074603030e-0001
51  * P2 =  -2.77777777777777777777777759668391122822266551158e-0003
52  * P3 =   6.61375661375661375657437408890138814721051293054e-0005
53  * P4 =  -1.65343915343915303310185228411892601606669528828e-0006
54  * P5 =   4.17535139755122945763580609663414647067443411178e-0008
55  * P6 =  -1.05683795988668526689182102605260986731620026832e-0009
56  * P7 =   2.67544168821852702827123344217198187229611470514e-0011
57  *
58  * Accuracy: |R-x*(exp(x)+1)/(exp(x)-1)|<=2**-119.13
59  *
60  *	3. For 1/16 < |x| < 1.125, choose x(+-i) ~ +-(i+4.5)/64, i=0,..,67
61  *	   since
62  *		exp(x) = exp(xi+(x-xi))= exp(xi)*exp((x-xi))
63  *	   we have
64  *		expm1(x) = expm1(xi)+(exp(xi))*(expm1(x-xi))
65  *	   where
66  *		|s=x-xi| <= 1/128
67  *	   and
68  *	expm1(s)=2s/(2-R), R= s-s^2*(T1+s^2*(T2+s^2*(T3+s^2*(T4+s^2*T5))))
69  *
70  * T1 =   1.666666666666666666666666666660876387437e-1L,
71  * T2 =  -2.777777777777777777777707812093173478756e-3L,
72  * T3 =   6.613756613756613482074280932874221202424e-5L,
73  * T4 =  -1.653439153392139954169609822742235851120e-6L,
74  * T5 =   4.175314851769539751387852116610973796053e-8L;
75  *
76  *	4. For |x| >= 1.125, return exp(x)-1.
77  *	    (see algorithm for exp)
78  *
79  * Special cases:
80  *	expm1l(INF) is INF, expm1l(NaN) is NaN;
81  *	expm1l(-INF)= -1;
82  *	for finite argument, only expm1l(0)=0 is exact.
83  *
84  * Accuracy:
85  *	according to an error analysis, the error is always less than
86  *	2 ulp (unit in the last place).
87  *
88  * Misc. info.
89  *	For 113 bit long double
90  *		if x >  1.135652340629414394949193107797076342845e+4
91  *      then expm1l(x) overflow;
92  *
93  * Constants:
94  * Only decimal values are given. We assume that the compiler will convert
95  * from decimal to binary accurately enough to produce the correct
96  * hexadecimal values.
97  */
98 
99 #include "libm.h"
100 
101 extern const long double _TBL_expl_hi[], _TBL_expl_lo[];
102 extern const long double _TBL_expm1lx[], _TBL_expm1l[];
103 
104 static const long double
105 	zero		= +0.0L,
106 	one		= +1.0L,
107 	two		= +2.0L,
108 	ln2_64		= +1.083042469624914545964425189778400898568e-2L,
109 	ovflthreshold	= +1.135652340629414394949193107797076342845e+4L,
110 	invln2_32	= +4.616624130844682903551758979206054839765e+1L,
111 	ln2_32hi	= +2.166084939249829091928849858592451515688e-2L,
112 	ln2_32lo	= +5.209643502595475652782654157501186731779e-27L,
113 	huge		= +1.0e4000L,
114 	tiny		= +1.0e-4000L,
115 	P1 = +1.66666666666666666666666666666638500528074603030e-0001L,
116 	P2 = -2.77777777777777777777777759668391122822266551158e-0003L,
117 	P3 = +6.61375661375661375657437408890138814721051293054e-0005L,
118 	P4 = -1.65343915343915303310185228411892601606669528828e-0006L,
119 	P5 = +4.17535139755122945763580609663414647067443411178e-0008L,
120 	P6 = -1.05683795988668526689182102605260986731620026832e-0009L,
121 	P7 = +2.67544168821852702827123344217198187229611470514e-0011L,
122 /* rational approximation coeffs for [-(ln2)/64,(ln2)/64] */
123 	T1 = +1.666666666666666666666666666660876387437e-1L,
124 	T2 = -2.777777777777777777777707812093173478756e-3L,
125 	T3 = +6.613756613756613482074280932874221202424e-5L,
126 	T4 = -1.653439153392139954169609822742235851120e-6L,
127 	T5 = +4.175314851769539751387852116610973796053e-8L;
128 
129 long double
130 expm1l(long double x) {
131 	int hx, ix, j, k, m;
132 	long double t, r, s, w;
133 
134 	hx = ((int *) &x)[HIXWORD];
135 	ix = hx & ~0x80000000;
136 	if (ix >= 0x7fff0000) {
137 		if (x != x)
138 			return (x + x);	/* NaN */
139 		if (x < zero)
140 			return (-one);	/* -inf */
141 		return (x);	/* +inf */
142 	}
143 	if (ix < 0x3fff4000) {	/* |x| < 1.25 */
144 		if (ix < 0x3ffb0000) {	/* |x| < 0.0625 */
145 			if (ix < 0x3f8d0000) {
146 				if ((int) x == 0)
147 					return (x);	/* |x|<2^-114 */
148 			}
149 			t = x * x;
150 			r = (x - t * (P1 + t * (P2 + t * (P3 + t * (P4 + t *
151 				(P5 + t * (P6 + t * P7)))))));
152 			return (x + (x * r) / (two - r));
153 		}
154 		/* compute i = [64*x] */
155 		m = 0x4009 - (ix >> 16);
156 		j = ((ix & 0x0000ffff) | 0x10000) >> m;	/* j=4,...,67 */
157 		if (hx < 0)
158 			j += 82;			/* negative */
159 		s = x - _TBL_expm1lx[j];
160 		t = s * s;
161 		r = s - t * (T1 + t * (T2 + t * (T3 + t * (T4 + t * T5))));
162 		r = (s + s) / (two - r);
163 		w = _TBL_expm1l[j];
164 		return (w + (w + one) * r);
165 	}
166 	if (hx > 0) {
167 		if (x > ovflthreshold)
168 			return (huge * huge);
169 		k = (int) (invln2_32 * (x + ln2_64));
170 	} else {
171 		if (x < -80.0)
172 			return (tiny - x / x);
173 		k = (int) (invln2_32 * (x - ln2_64));
174 	}
175 	j = k & 0x1f;
176 	m = k >> 5;
177 	t = (long double) k;
178 	x = (x - t * ln2_32hi) - t * ln2_32lo;
179 	t = x * x;
180 	r = (x - t * (T1 + t * (T2 + t * (T3 + t * (T4 + t * T5))))) - two;
181 	x = _TBL_expl_hi[j] - ((_TBL_expl_hi[j] * (x + x)) / r -
182 		_TBL_expl_lo[j]);
183 	return (scalbnl(x, m) - one);
184 }
185