1/* @(#)s_expm1.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#include <sys/cdefs.h>
14#if defined(LIBM_SCCS) && !defined(lint)
15__RCSID("$NetBSD: s_expm1.c,v 1.13 2017/02/09 22:11:09 maya Exp $");
16#endif
17
18/* expm1(x)
19 * Returns exp(x)-1, the exponential of x minus 1.
20 *
21 * Method
22 *   1. Argument reduction:
23 *	Given x, find r and integer k such that
24 *
25 *               x = k*ln2 + r,  |r| <= 0.5*ln2 ~ 0.34658
26 *
27 *      Here a correction term c will be computed to compensate
28 *	the error in r when rounded to a floating-point number.
29 *
30 *   2. Approximating expm1(r) by a special rational function on
31 *	the interval [0,0.34658]:
32 *	Since
33 *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
34 *	we define R1(r*r) by
35 *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 * R1(r*r)
36 *	That is,
37 *	    R1(r**2) = 6/r *((exp(r)+1)/(exp(r)-1) - 2/r)
38 *		     = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
39 *		     = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
40 *      We use a special Reme algorithm on [0,0.347] to generate
41 * 	a polynomial of degree 5 in r*r to approximate R1. The
42 *	maximum error of this polynomial approximation is bounded
43 *	by 2**-61. In other words,
44 *	    R1(z) ~ 1.0 + Q1*z + Q2*z**2 + Q3*z**3 + Q4*z**4 + Q5*z**5
45 *	where 	Q1  =  -1.6666666666666567384E-2,
46 * 		Q2  =   3.9682539681370365873E-4,
47 * 		Q3  =  -9.9206344733435987357E-6,
48 * 		Q4  =   2.5051361420808517002E-7,
49 * 		Q5  =  -6.2843505682382617102E-9;
50 *  	(where z=r*r, and the values of Q1 to Q5 are listed below)
51 *	with error bounded by
52 *	    |                  5           |     -61
53 *	    | 1.0+Q1*z+...+Q5*z   -  R1(z) | <= 2
54 *	    |                              |
55 *
56 *	expm1(r) = exp(r)-1 is then computed by the following
57 * 	specific way which minimize the accumulation rounding error:
58 *			       2     3
59 *			      r     r    [ 3 - (R1 + R1*r/2)  ]
60 *	      expm1(r) = r + --- + --- * [--------------------]
61 *		              2     2    [ 6 - r*(3 - R1*r/2) ]
62 *
63 *	To compensate the error in the argument reduction, we use
64 *		expm1(r+c) = expm1(r) + c + expm1(r)*c
65 *			   ~ expm1(r) + c + r*c
66 *	Thus c+r*c will be added in as the correction terms for
67 *	expm1(r+c). Now rearrange the term to avoid optimization
68 * 	screw up:
69 *		        (      2                                    2 )
70 *		        ({  ( r    [ R1 -  (3 - R1*r/2) ]  )  }    r  )
71 *	 expm1(r+c)~r - ({r*(--- * [--------------------]-c)-c} - --- )
72 *	                ({  ( 2    [ 6 - r*(3 - R1*r/2) ]  )  }    2  )
73 *                      (                                             )
74 *
75 *		   = r - E
76 *   3. Scale back to obtain expm1(x):
77 *	From step 1, we have
78 *	   expm1(x) = either 2^k*[expm1(r)+1] - 1
79 *		    = or     2^k*[expm1(r) + (1-2^-k)]
80 *   4. Implementation notes:
81 *	(A). To save one multiplication, we scale the coefficient Qi
82 *	     to Qi*2^i, and replace z by (x^2)/2.
83 *	(B). To achieve maximum accuracy, we compute expm1(x) by
84 *	  (i)   if x < -56*ln2, return -1.0, (raise inexact if x!=inf)
85 *	  (ii)  if k=0, return r-E
86 *	  (iii) if k=-1, return 0.5*(r-E)-0.5
87 *        (iv)	if k=1 if r < -0.25, return 2*((r+0.5)- E)
88 *	       	       else	     return  1.0+2.0*(r-E);
89 *	  (v)   if (k<-2||k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
90 *	  (vi)  if k <= 20, return 2^k((1-2^-k)-(E-r)), else
91 *	  (vii) return 2^k(1-((E+2^-k)-r))
92 *
93 * Special cases:
94 *	expm1(INF) is INF, expm1(NaN) is NaN;
95 *	expm1(-INF) is -1, and
96 *	for finite argument, only expm1(0)=0 is exact.
97 *
98 * Accuracy:
99 *	according to an error analysis, the error is always less than
100 *	1 ulp (unit in the last place).
101 *
102 * Misc. info.
103 *	For IEEE double
104 *	    if x >  7.09782712893383973096e+02 then expm1(x) overflow
105 *
106 * Constants:
107 * The hexadecimal values are the intended ones for the following
108 * constants. The decimal values may be used, provided that the
109 * compiler will convert from decimal to binary accurately enough
110 * to produce the hexadecimal values shown.
111 */
112
113#include "math.h"
114#include "math_private.h"
115
116static const double
117one		= 1.0,
118huge		= 1.0e+300,
119tiny		= 1.0e-300,
120o_threshold	= 7.09782712893383973096e+02,/* 0x40862E42, 0xFEFA39EF */
121ln2_hi		= 6.93147180369123816490e-01,/* 0x3fe62e42, 0xfee00000 */
122ln2_lo		= 1.90821492927058770002e-10,/* 0x3dea39ef, 0x35793c76 */
123invln2		= 1.44269504088896338700e+00,/* 0x3ff71547, 0x652b82fe */
124	/* scaled coefficients related to expm1 */
125Q1  =  -3.33333333333331316428e-02, /* BFA11111 111110F4 */
126Q2  =   1.58730158725481460165e-03, /* 3F5A01A0 19FE5585 */
127Q3  =  -7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
128Q4  =   4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
129Q5  =  -2.01099218183624371326e-07; /* BE8AFDB7 6E09C32D */
130
131double
132expm1(double x)
133{
134	double y,hi,lo,c,t,e,hxs,hfx,r1;
135	int32_t k,xsb;
136	u_int32_t hx;
137
138	c = 0;
139	GET_HIGH_WORD(hx,x);
140	xsb = hx&0x80000000;		/* sign bit of x */
141#ifdef DEAD_CODE
142	if(xsb==0) y=x; else y= -x;	/* y = |x| */
143#endif
144	hx &= 0x7fffffff;		/* high word of |x| */
145
146    /* filter out huge and non-finite argument */
147	if(hx >= 0x4043687A) {			/* if |x|>=56*ln2 */
148	    if(hx >= 0x40862E42) {		/* if |x|>=709.78... */
149                if(hx>=0x7ff00000) {
150		    u_int32_t low;
151		    GET_LOW_WORD(low,x);
152		    if(((hx&0xfffff)|low)!=0)
153		         return x+x; 	 /* NaN */
154		    else return (xsb==0)? x:-1.0;/* exp(+-inf)={inf,-1} */
155	        }
156	        if(x > o_threshold) return huge*huge; /* overflow */
157	    }
158	    if(xsb!=0) { /* x < -56*ln2, return -1.0 with inexact */
159		if(x+tiny<0.0)		/* raise inexact */
160		return tiny-one;	/* return -1 */
161	    }
162	}
163
164    /* argument reduction */
165	if(hx > 0x3fd62e42) {		/* if  |x| > 0.5 ln2 */
166	    if(hx < 0x3FF0A2B2) {	/* and |x| < 1.5 ln2 */
167		if(xsb==0)
168		    {hi = x - ln2_hi; lo =  ln2_lo;  k =  1;}
169		else
170		    {hi = x + ln2_hi; lo = -ln2_lo;  k = -1;}
171	    } else {
172		k  = invln2*x+((xsb==0)?0.5:-0.5);
173		t  = k;
174		hi = x - t*ln2_hi;	/* t*ln2_hi is exact here */
175		lo = t*ln2_lo;
176	    }
177	    x  = hi - lo;
178	    c  = (hi-x)-lo;
179	}
180	else if(hx < 0x3c900000) {  	/* when |x|<2**-54, return x */
181	    t = huge+x;	/* return x with inexact flags when x!=0 */
182	    return x - (t-(huge+x));
183	}
184	else k = 0;
185
186    /* x is now in primary range */
187	hfx = 0.5*x;
188	hxs = x*hfx;
189	r1 = one+hxs*(Q1+hxs*(Q2+hxs*(Q3+hxs*(Q4+hxs*Q5))));
190	t  = 3.0-r1*hfx;
191	e  = hxs*((r1-t)/(6.0 - x*t));
192	if(k==0) return x - (x*e-hxs);		/* c is 0 */
193	else {
194	    e  = (x*(e-c)-c);
195	    e -= hxs;
196	    if(k== -1) return 0.5*(x-e)-0.5;
197	    if(k==1)  {
198	       	if(x < -0.25) return -2.0*(e-(x+0.5));
199	       	else 	      return  one+2.0*(x-e);
200	    }
201	    if (k <= -2 || k>56) {   /* suffice to return exp(x)-1 */
202	        u_int32_t high;
203	        y = one-(e-x);
204		GET_HIGH_WORD(high,y);
205		SET_HIGH_WORD(y,high+(k<<20));	/* add k to y's exponent */
206	        return y-one;
207	    }
208	    t = one;
209	    if(k<20) {
210	        u_int32_t high;
211	        SET_HIGH_WORD(t,0x3ff00000 - (0x200000>>k));  /* t=1-2^-k */
212	       	y = t-(e-x);
213		GET_HIGH_WORD(high,y);
214		SET_HIGH_WORD(y,high+(k<<20));	/* add k to y's exponent */
215	   } else {
216	        u_int32_t high;
217		SET_HIGH_WORD(t,((0x3ff-k)<<20));	/* 2^-k */
218	       	y = x-(e+t);
219	       	y += one;
220		GET_HIGH_WORD(high,y);
221		SET_HIGH_WORD(y,high+(k<<20));	/* add k to y's exponent */
222	    }
223	}
224	return y;
225}
226