xref: /openbsd-src/lib/libm/src/e_exp.c (revision d13be5d47e4149db2549a9828e244d59dbc43f15)
1 /* @(#)e_exp.c 5.1 93/09/24 */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunPro, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 /* LINTLIBRARY */
14 
15 /* exp(x)
16  * Returns the exponential of x.
17  *
18  * Method
19  *   1. Argument reduction:
20  *      Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658.
21  *	Given x, find r and integer k such that
22  *
23  *               x = k*ln2 + r,  |r| <= 0.5*ln2.
24  *
25  *      Here r will be represented as r = hi-lo for better
26  *	accuracy.
27  *
28  *   2. Approximation of exp(r) by a special rational function on
29  *	the interval [0,0.34658]:
30  *	Write
31  *	    R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ...
32  *      We use a special Remes algorithm on [0,0.34658] to generate
33  * 	a polynomial of degree 5 to approximate R. The maximum error
34  *	of this polynomial approximation is bounded by 2**-59. In
35  *	other words,
36  *	    R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5
37  *  	(where z=r*r, and the values of P1 to P5 are listed below)
38  *	and
39  *	    |                  5          |     -59
40  *	    | 2.0+P1*z+...+P5*z   -  R(z) | <= 2
41  *	    |                             |
42  *	The computation of exp(r) thus becomes
43  *                             2*r
44  *		exp(r) = 1 + -------
45  *		              R - r
46  *                                 r*R1(r)
47  *		       = 1 + r + ----------- (for better accuracy)
48  *		                  2 - R1(r)
49  *	where
50  *			         2       4             10
51  *		R1(r) = r - (P1*r  + P2*r  + ... + P5*r   ).
52  *
53  *   3. Scale back to obtain exp(x):
54  *	From step 1, we have
55  *	   exp(x) = 2^k * exp(r)
56  *
57  * Special cases:
58  *	exp(INF) is INF, exp(NaN) is NaN;
59  *	exp(-INF) is 0, and
60  *	for finite argument, only exp(0)=1 is exact.
61  *
62  * Accuracy:
63  *	according to an error analysis, the error is always less than
64  *	1 ulp (unit in the last place).
65  *
66  * Misc. info.
67  *	For IEEE double
68  *	    if x >  7.09782712893383973096e+02 then exp(x) overflow
69  *	    if x < -7.45133219101941108420e+02 then exp(x) underflow
70  *
71  * Constants:
72  * The hexadecimal values are the intended ones for the following
73  * constants. The decimal values may be used, provided that the
74  * compiler will convert from decimal to binary accurately enough
75  * to produce the hexadecimal values shown.
76  */
77 
78 #include <sys/cdefs.h>
79 #include <float.h>
80 #include <math.h>
81 
82 #include "math_private.h"
83 
84 static const double
85 one	= 1.0,
86 halF[2]	= {0.5,-0.5,},
87 huge	= 1.0e+300,
88 twom1000= 9.33263618503218878990e-302,     /* 2**-1000=0x01700000,0*/
89 o_threshold=  7.09782712893383973096e+02,  /* 0x40862E42, 0xFEFA39EF */
90 u_threshold= -7.45133219101941108420e+02,  /* 0xc0874910, 0xD52D3051 */
91 ln2HI[2]   ={ 6.93147180369123816490e-01,  /* 0x3fe62e42, 0xfee00000 */
92 	     -6.93147180369123816490e-01,},/* 0xbfe62e42, 0xfee00000 */
93 ln2LO[2]   ={ 1.90821492927058770002e-10,  /* 0x3dea39ef, 0x35793c76 */
94 	     -1.90821492927058770002e-10,},/* 0xbdea39ef, 0x35793c76 */
95 invln2 =  1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
96 P1   =  1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */
97 P2   = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */
98 P3   =  6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */
99 P4   = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
100 P5   =  4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */
101 
102 
103 double
104 exp(double x)	/* default IEEE double exp */
105 {
106 	double y,hi,lo,c,t;
107 	int32_t k,xsb;
108 	u_int32_t hx;
109 
110 	GET_HIGH_WORD(hx,x);
111 	xsb = (hx>>31)&1;		/* sign bit of x */
112 	hx &= 0x7fffffff;		/* high word of |x| */
113 
114     /* filter out non-finite argument */
115 	if(hx >= 0x40862E42) {			/* if |x|>=709.78... */
116             if(hx>=0x7ff00000) {
117 	        u_int32_t lx;
118 		GET_LOW_WORD(lx,x);
119 		if(((hx&0xfffff)|lx)!=0)
120 		     return x+x; 		/* NaN */
121 		else return (xsb==0)? x:0.0;	/* exp(+-inf)={inf,0} */
122 	    }
123 	    if(x > o_threshold) return huge*huge; /* overflow */
124 	    if(x < u_threshold) return twom1000*twom1000; /* underflow */
125 	}
126 
127     /* argument reduction */
128 	if(hx > 0x3fd62e42) {		/* if  |x| > 0.5 ln2 */
129 	    if(hx < 0x3FF0A2B2) {	/* and |x| < 1.5 ln2 */
130 		hi = x-ln2HI[xsb]; lo=ln2LO[xsb]; k = 1-xsb-xsb;
131 	    } else {
132 		k  = invln2*x+halF[xsb];
133 		t  = k;
134 		hi = x - t*ln2HI[0];	/* t*ln2HI is exact here */
135 		lo = t*ln2LO[0];
136 	    }
137 	    x  = hi - lo;
138 	}
139 	else if(hx < 0x3e300000)  {	/* when |x|<2**-28 */
140 	    if(huge+x>one) return one+x;/* trigger inexact */
141 	}
142 	else k = 0;
143 
144     /* x is now in primary range */
145 	t  = x*x;
146 	c  = x - t*(P1+t*(P2+t*(P3+t*(P4+t*P5))));
147 	if(k==0) 	return one-((x*c)/(c-2.0)-x);
148 	else 		y = one-((lo-(x*c)/(2.0-c))-hi);
149 	if(k >= -1021) {
150 	    u_int32_t hy;
151 	    GET_HIGH_WORD(hy,y);
152 	    SET_HIGH_WORD(y,hy+(k<<20));	/* add k to y's exponent */
153 	    return y;
154 	} else {
155 	    u_int32_t hy;
156 	    GET_HIGH_WORD(hy,y);
157 	    SET_HIGH_WORD(y,hy+((k+1000)<<20));	/* add k to y's exponent */
158 	    return y*twom1000;
159 	}
160 }
161 
162 #if	LDBL_MANT_DIG == 53
163 #ifdef	lint
164 /* PROTOLIB1 */
165 long double expl(long double);
166 #else	/* lint */
167 __weak_alias(expl, exp);
168 #endif	/* lint */
169 #endif	/* LDBL_MANT_DIG == 53 */
170