xref: /inferno-os/libmath/fdlibm/e_log.c (revision 37da2899f40661e3e9631e497da8dc59b971cbd0)
1*37da2899SCharles.Forsyth /* derived from /netlib/fdlibm */
2*37da2899SCharles.Forsyth 
3*37da2899SCharles.Forsyth /* @(#)e_log.c 1.3 95/01/18 */
4*37da2899SCharles.Forsyth /*
5*37da2899SCharles.Forsyth  * ====================================================
6*37da2899SCharles.Forsyth  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
7*37da2899SCharles.Forsyth  *
8*37da2899SCharles.Forsyth  * Developed at SunSoft, a Sun Microsystems, Inc. business.
9*37da2899SCharles.Forsyth  * Permission to use, copy, modify, and distribute this
10*37da2899SCharles.Forsyth  * software is freely granted, provided that this notice
11*37da2899SCharles.Forsyth  * is preserved.
12*37da2899SCharles.Forsyth  * ====================================================
13*37da2899SCharles.Forsyth  */
14*37da2899SCharles.Forsyth 
15*37da2899SCharles.Forsyth /* __ieee754_log(x)
16*37da2899SCharles.Forsyth  * Return the logrithm of x
17*37da2899SCharles.Forsyth  *
18*37da2899SCharles.Forsyth  * Method :
19*37da2899SCharles.Forsyth  *   1. Argument Reduction: find k and f such that
20*37da2899SCharles.Forsyth  *			x = 2^k * (1+f),
21*37da2899SCharles.Forsyth  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
22*37da2899SCharles.Forsyth  *
23*37da2899SCharles.Forsyth  *   2. Approximation of log(1+f).
24*37da2899SCharles.Forsyth  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
25*37da2899SCharles.Forsyth  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
26*37da2899SCharles.Forsyth  *	     	 = 2s + s*R
27*37da2899SCharles.Forsyth  *      We use a special Reme algorithm on [0,0.1716] to generate
28*37da2899SCharles.Forsyth  * 	a polynomial of degree 14 to approximate R The maximum error
29*37da2899SCharles.Forsyth  *	of this polynomial approximation is bounded by 2**-58.45. In
30*37da2899SCharles.Forsyth  *	other words,
31*37da2899SCharles.Forsyth  *		        2      4      6      8      10      12      14
32*37da2899SCharles.Forsyth  *	    R(z) ~ Lg1*s +Lg2*s +Lg3*s +Lg4*s +Lg5*s  +Lg6*s  +Lg7*s
33*37da2899SCharles.Forsyth  *  	(the values of Lg1 to Lg7 are listed in the program)
34*37da2899SCharles.Forsyth  *	and
35*37da2899SCharles.Forsyth  *	    |      2          14          |     -58.45
36*37da2899SCharles.Forsyth  *	    | Lg1*s +...+Lg7*s    -  R(z) | <= 2
37*37da2899SCharles.Forsyth  *	    |                             |
38*37da2899SCharles.Forsyth  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
39*37da2899SCharles.Forsyth  *	In order to guarantee error in log below 1ulp, we compute log
40*37da2899SCharles.Forsyth  *	by
41*37da2899SCharles.Forsyth  *		log(1+f) = f - s*(f - R)	(if f is not too large)
42*37da2899SCharles.Forsyth  *		log(1+f) = f - (hfsq - s*(hfsq+R)).	(better accuracy)
43*37da2899SCharles.Forsyth  *
44*37da2899SCharles.Forsyth  *	3. Finally,  log(x) = k*ln2 + log(1+f).
45*37da2899SCharles.Forsyth  *			    = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
46*37da2899SCharles.Forsyth  *	   Here ln2 is split into two floating point number:
47*37da2899SCharles.Forsyth  *			ln2_hi + ln2_lo,
48*37da2899SCharles.Forsyth  *	   where n*ln2_hi is always exact for |n| < 2000.
49*37da2899SCharles.Forsyth  *
50*37da2899SCharles.Forsyth  * Special cases:
51*37da2899SCharles.Forsyth  *	log(x) is NaN with signal if x < 0 (including -INF) ;
52*37da2899SCharles.Forsyth  *	log(+INF) is +INF; log(0) is -INF with signal;
53*37da2899SCharles.Forsyth  *	log(NaN) is that NaN with no signal.
54*37da2899SCharles.Forsyth  *
55*37da2899SCharles.Forsyth  * Accuracy:
56*37da2899SCharles.Forsyth  *	according to an error analysis, the error is always less than
57*37da2899SCharles.Forsyth  *	1 ulp (unit in the last place).
58*37da2899SCharles.Forsyth  *
59*37da2899SCharles.Forsyth  * Constants:
60*37da2899SCharles.Forsyth  * The hexadecimal values are the intended ones for the following
61*37da2899SCharles.Forsyth  * constants. The decimal values may be used, provided that the
62*37da2899SCharles.Forsyth  * compiler will convert from decimal to binary accurately enough
63*37da2899SCharles.Forsyth  * to produce the hexadecimal values shown.
64*37da2899SCharles.Forsyth  */
65*37da2899SCharles.Forsyth 
66*37da2899SCharles.Forsyth #include "fdlibm.h"
67*37da2899SCharles.Forsyth 
68*37da2899SCharles.Forsyth static const double
69*37da2899SCharles.Forsyth ln2_hi  =  6.93147180369123816490e-01,	/* 3fe62e42 fee00000 */
70*37da2899SCharles.Forsyth ln2_lo  =  1.90821492927058770002e-10,	/* 3dea39ef 35793c76 */
71*37da2899SCharles.Forsyth two54   =  1.80143985094819840000e+16,  /* 43500000 00000000 */
72*37da2899SCharles.Forsyth Lg1 = 6.666666666666735130e-01,  /* 3FE55555 55555593 */
73*37da2899SCharles.Forsyth Lg2 = 3.999999999940941908e-01,  /* 3FD99999 9997FA04 */
74*37da2899SCharles.Forsyth Lg3 = 2.857142874366239149e-01,  /* 3FD24924 94229359 */
75*37da2899SCharles.Forsyth Lg4 = 2.222219843214978396e-01,  /* 3FCC71C5 1D8E78AF */
76*37da2899SCharles.Forsyth Lg5 = 1.818357216161805012e-01,  /* 3FC74664 96CB03DE */
77*37da2899SCharles.Forsyth Lg6 = 1.531383769920937332e-01,  /* 3FC39A09 D078C69F */
78*37da2899SCharles.Forsyth Lg7 = 1.479819860511658591e-01;  /* 3FC2F112 DF3E5244 */
79*37da2899SCharles.Forsyth 
80*37da2899SCharles.Forsyth static double zero   =  0.0;
81*37da2899SCharles.Forsyth 
__ieee754_log(double x)82*37da2899SCharles.Forsyth 	double __ieee754_log(double x)
83*37da2899SCharles.Forsyth {
84*37da2899SCharles.Forsyth 	double hfsq,f,s,z,R,w,t1,t2,dk;
85*37da2899SCharles.Forsyth 	int k,hx,i,j;
86*37da2899SCharles.Forsyth 	unsigned lx;
87*37da2899SCharles.Forsyth 
88*37da2899SCharles.Forsyth 	hx = __HI(x);		/* high word of x */
89*37da2899SCharles.Forsyth 	lx = __LO(x);		/* low  word of x */
90*37da2899SCharles.Forsyth 
91*37da2899SCharles.Forsyth 	k=0;
92*37da2899SCharles.Forsyth 	if (hx < 0x00100000) {			/* x < 2**-1022  */
93*37da2899SCharles.Forsyth 	    if (((hx&0x7fffffff)|lx)==0)
94*37da2899SCharles.Forsyth 		return -two54/zero;		/* log(+-0)=-inf */
95*37da2899SCharles.Forsyth 	    if (hx<0) return (x-x)/zero;	/* log(-#) = NaN */
96*37da2899SCharles.Forsyth 	    k -= 54; x *= two54; /* subnormal number, scale up x */
97*37da2899SCharles.Forsyth 	    hx = __HI(x);		/* high word of x */
98*37da2899SCharles.Forsyth 	}
99*37da2899SCharles.Forsyth 	if (hx >= 0x7ff00000) return x+x;
100*37da2899SCharles.Forsyth 	k += (hx>>20)-1023;
101*37da2899SCharles.Forsyth 	hx &= 0x000fffff;
102*37da2899SCharles.Forsyth 	i = (hx+0x95f64)&0x100000;
103*37da2899SCharles.Forsyth 	__HI(x) = hx|(i^0x3ff00000);	/* normalize x or x/2 */
104*37da2899SCharles.Forsyth 	k += (i>>20);
105*37da2899SCharles.Forsyth 	f = x-1.0;
106*37da2899SCharles.Forsyth 	if((0x000fffff&(2+hx))<3) {	/* |f| < 2**-20 */
107*37da2899SCharles.Forsyth 	    if(f==zero) if(k==0) return zero;  else {dk=(double)k;
108*37da2899SCharles.Forsyth 				 return dk*ln2_hi+dk*ln2_lo;}
109*37da2899SCharles.Forsyth 	    R = f*f*(0.5-0.33333333333333333*f);
110*37da2899SCharles.Forsyth 	    if(k==0) return f-R; else {dk=(double)k;
111*37da2899SCharles.Forsyth 	    	     return dk*ln2_hi-((R-dk*ln2_lo)-f);}
112*37da2899SCharles.Forsyth 	}
113*37da2899SCharles.Forsyth  	s = f/(2.0+f);
114*37da2899SCharles.Forsyth 	dk = (double)k;
115*37da2899SCharles.Forsyth 	z = s*s;
116*37da2899SCharles.Forsyth 	i = hx-0x6147a;
117*37da2899SCharles.Forsyth 	w = z*z;
118*37da2899SCharles.Forsyth 	j = 0x6b851-hx;
119*37da2899SCharles.Forsyth 	t1= w*(Lg2+w*(Lg4+w*Lg6));
120*37da2899SCharles.Forsyth 	t2= z*(Lg1+w*(Lg3+w*(Lg5+w*Lg7)));
121*37da2899SCharles.Forsyth 	i |= j;
122*37da2899SCharles.Forsyth 	R = t2+t1;
123*37da2899SCharles.Forsyth 	if(i>0) {
124*37da2899SCharles.Forsyth 	    hfsq=0.5*f*f;
125*37da2899SCharles.Forsyth 	    if(k==0) return f-(hfsq-s*(hfsq+R)); else
126*37da2899SCharles.Forsyth 		     return dk*ln2_hi-((hfsq-(s*(hfsq+R)+dk*ln2_lo))-f);
127*37da2899SCharles.Forsyth 	} else {
128*37da2899SCharles.Forsyth 	    if(k==0) return f-s*(f-R); else
129*37da2899SCharles.Forsyth 		     return dk*ln2_hi-((s*(f-R)-dk*ln2_lo)-f);
130*37da2899SCharles.Forsyth 	}
131*37da2899SCharles.Forsyth }
132