1 
2 /* @(#)s_log1p.c 5.1 93/09/24 */
3 /*
4  * ====================================================
5  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6  *
7  * Developed at SunPro, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  */
13 
14 /*
15 FUNCTION
16 <<log1p>>, <<log1pf>>---log of <<1 + <[x]>>>
17 
18 INDEX
19 	log1p
20 INDEX
21 	log1pf
22 
23 SYNOPSIS
24 	#include <math.h>
25 	double log1p(double <[x]>);
26 	float log1pf(float <[x]>);
27 
28 DESCRIPTION
29 <<log1p>> calculates
30 @tex
31 $ln(1+x)$,
32 @end tex
33 the natural logarithm of <<1+<[x]>>>.  You can use <<log1p>> rather
34 than `<<log(1+<[x]>)>>' for greater precision when <[x]> is very
35 small.
36 
37 <<log1pf>> calculates the same thing, but accepts and returns
38 <<float>> values rather than <<double>>.
39 
40 RETURNS
41 <<log1p>> returns a <<double>>, the natural log of <<1+<[x]>>>.
42 <<log1pf>> returns a <<float>>, the natural log of <<1+<[x]>>>.
43 
44 PORTABILITY
45 Neither <<log1p>> nor <<log1pf>> is required by ANSI C or by the System V
46 Interface Definition (Issue 2).
47 
48 */
49 
50 /* double log1p(double x)
51  *
52  * Method :
53  *   1. Argument Reduction: find k and f such that
54  *			1+x = 2^k * (1+f),
55  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
56  *
57  *      Note. If k=0, then f=x is exact. However, if k!=0, then f
58  *	may not be representable exactly. In that case, a correction
59  *	term is need. Let u=1+x rounded. Let c = (1+x)-u, then
60  *	log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
61  *	and add back the correction term c/u.
62  *	(Note: when x > 2**53, one can simply return log(x))
63  *
64  *   2. Approximation of log1p(f).
65  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
66  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
67  *	     	 = 2s + s*R
68  *      We use a special Remez algorithm on [0,0.1716] to generate
69  * 	a polynomial of degree 14 to approximate R The maximum error
70  *	of this polynomial approximation is bounded by 2**-58.45. In
71  *	other words,
72  *		        2      4      6      8      10      12      14
73  *	    R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s  +Lp6*s  +Lp7*s
74  *  	(the values of Lp1 to Lp7 are listed in the program)
75  *	and
76  *	    |      2          14          |     -58.45
77  *	    | Lp1*s +...+Lp7*s    -  R(z) | <= 2
78  *	    |                             |
79  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
80  *	In order to guarantee error in log below 1ulp, we compute log
81  *	by
82  *		log1p(f) = f - (hfsq - s*(hfsq+R)).
83  *
84  *	3. Finally, log1p(x) = k*ln2 + log1p(f).
85  *		 	     = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
86  *	   Here ln2 is split into two floating point number:
87  *			ln2_hi + ln2_lo,
88  *	   where n*ln2_hi is always exact for |n| < 2000.
89  *
90  * Special cases:
91  *	log1p(x) is NaN with signal if x < -1 (including -INF) ;
92  *	log1p(+INF) is +INF; log1p(-1) is -INF with signal;
93  *	log1p(NaN) is that NaN with no signal.
94  *
95  * Accuracy:
96  *	according to an error analysis, the error is always less than
97  *	1 ulp (unit in the last place).
98  *
99  * Constants:
100  * The hexadecimal values are the intended ones for the following
101  * constants. The decimal values may be used, provided that the
102  * compiler will convert from decimal to binary accurately enough
103  * to produce the hexadecimal values shown.
104  *
105  * Note: Assuming log() return accurate answer, the following
106  * 	 algorithm can be used to compute log1p(x) to within a few ULP:
107  *
108  *		u = 1+x;
109  *		if(u==1.0) return x ; else
110  *			   return log(u)*(x/(u-1.0));
111  *
112  *	 See HP-15C Advanced Functions Handbook, p.193.
113  */
114 
115 #include "fdlibm.h"
116 #include "math_config.h"
117 
118 #ifdef _NEED_FLOAT64
119 
120 static const __float64
121 ln2_hi  =  _F_64(6.93147180369123816490e-01),	/* 3fe62e42 fee00000 */
122 ln2_lo  =  _F_64(1.90821492927058770002e-10),	/* 3dea39ef 35793c76 */
123 two54   =  _F_64(1.80143985094819840000e+16),  /* 43500000 00000000 */
124 Lp1 = _F_64(6.666666666666735130e-01),  /* 3FE55555 55555593 */
125 Lp2 = _F_64(3.999999999940941908e-01),  /* 3FD99999 9997FA04 */
126 Lp3 = _F_64(2.857142874366239149e-01),  /* 3FD24924 94229359 */
127 Lp4 = _F_64(2.222219843214978396e-01),  /* 3FCC71C5 1D8E78AF */
128 Lp5 = _F_64(1.818357216161805012e-01),  /* 3FC74664 96CB03DE */
129 Lp6 = _F_64(1.531383769920937332e-01),  /* 3FC39A09 D078C69F */
130 Lp7 = _F_64(1.479819860511658591e-01);  /* 3FC2F112 DF3E5244 */
131 
132 static const __float64 zero = _F_64(0.0);
133 
134 __float64
log1p64(__float64 x)135 log1p64(__float64 x)
136 {
137 	__float64 hfsq,f,c=0,s,z,R,u;
138 	__int32_t k,hx,hu,ax;
139 
140 	GET_HIGH_WORD(hx,x);
141 	ax = hx&0x7fffffff;
142 
143 	k = 1;
144 	if (hx < 0x3FDA827A) {			/* x < 0.41422  */
145 	    if(ax>=0x3ff00000) {		/* x <= -1.0 */
146 		if(x==_F_64(-1.0))
147 		    return __math_divzero (1);	/* log1p(-1)=-inf */
148 		else
149 		    return __math_invalid (x);	/* log1p(x<-1)=NaN */
150 	    }
151 	    if(ax<0x3e200000) {			/* |x| < 2**-29 */
152 		if(two54+x>zero			/* raise inexact */
153 	            &&ax<0x3c900000) 		/* |x| < 2**-54 */
154 		    return x;
155 		else
156 		    return x - x*x*_F_64(0.5);
157 	    }
158 	    if(hx>0||hx<=((__int32_t)0xbfd2bec3)) {
159 		k=0;f=x;hu=1;}	/* -0.2929<x<0.41422 */
160 	}
161 	if (hx >= 0x7ff00000) return x+x;
162 	if(k!=0) {
163 	    if(hx<0x43400000) {
164 		u  = _F_64(1.0)+x;
165 		GET_HIGH_WORD(hu,u);
166 	        k  = (hu>>20)-1023;
167 	        c  = (k>0)? _F_64(1.0)-(u-x):x-(u-_F_64(1.0));/* correction term */
168 		c /= u;
169 	    } else {
170 		u  = x;
171 		GET_HIGH_WORD(hu,u);
172 	        k  = (hu>>20)-1023;
173 		c  = 0;
174 	    }
175 	    hu &= 0x000fffff;
176 	    if(hu<0x6a09e) {
177 	        SET_HIGH_WORD(u,hu|0x3ff00000);	/* normalize u */
178 	    } else {
179 	        k += 1;
180 		SET_HIGH_WORD(u,hu|0x3fe00000);	/* normalize u/2 */
181 	        hu = (0x00100000-hu)>>2;
182 	    }
183 	    f = u - _F_64(1.0);
184 	}
185 	hfsq=_F_64(0.5)*f*f;
186 	if(hu==0) {	/* |f| < 2**-20 */
187           if(f==zero) { if(k==0) return zero;
188                       else {c += k*ln2_lo; return k*ln2_hi+c;}}
189           R = hfsq*(_F_64(1.0)-_F_64(0.66666666666666666)*f);
190           if(k==0) return f-R; else
191               return k*ln2_hi-((R-(k*ln2_lo+c))-f);
192 	}
193  	s = f/(_F_64(2.0)+f);
194 	z = s*s;
195 	R = z*(Lp1+z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7))))));
196 	if(k==0) return f-(hfsq-s*(hfsq+R)); else
197 		 return k*ln2_hi-((hfsq-(s*(hfsq+R)+(k*ln2_lo+c)))-f);
198 }
199 
200 _MATH_ALIAS_d_d(log1p)
201 
202 #endif /* _NEED_FLOAT64 */
203