xref: /illumos-gate/usr/src/lib/libm/common/C/expm1.c (revision 25c28e83beb90e7c80452a7c818c5e6f73a07dc8)
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 expm1 = __expm1
31 
32 /* INDENT OFF */
33 /*
34  * expm1(x)
35  * Returns exp(x)-1, the exponential of x minus 1.
36  *
37  * Method
38  *   1. Arugment reduction:
39  *	Given x, find r and integer k such that
40  *
41  *               x = k*ln2 + r,  |r| <= 0.5*ln2 ~ 0.34658
42  *
43  *      Here a correction term c will be computed to compensate
44  *	the error in r when rounded to a floating-point number.
45  *
46  *   2. Approximating expm1(r) by a special rational function on
47  *	the interval [0,0.34658]:
48  *	Since
49  *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
50  *	we define R1(r*r) by
51  *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 * R1(r*r)
52  *	That is,
53  *	    R1(r**2) = 6/r *((exp(r)+1)/(exp(r)-1) - 2/r)
54  *		     = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
55  *		     = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
56  *      We use a special Reme algorithm on [0,0.347] to generate
57  * 	a polynomial of degree 5 in r*r to approximate R1. The
58  *	maximum error of this polynomial approximation is bounded
59  *	by 2**-61. In other words,
60  *	    R1(z) ~ 1.0 + Q1*z + Q2*z**2 + Q3*z**3 + Q4*z**4 + Q5*z**5
61  *	where 	Q1  =  -1.6666666666666567384E-2,
62  * 		Q2  =   3.9682539681370365873E-4,
63  * 		Q3  =  -9.9206344733435987357E-6,
64  * 		Q4  =   2.5051361420808517002E-7,
65  * 		Q5  =  -6.2843505682382617102E-9;
66  *  	(where z=r*r, and the values of Q1 to Q5 are listed below)
67  *	with error bounded by
68  *	    |                  5           |     -61
69  *	    | 1.0+Q1*z+...+Q5*z   -  R1(z) | <= 2
70  *	    |                              |
71  *
72  *	expm1(r) = exp(r)-1 is then computed by the following
73  * 	specific way which minimize the accumulation rounding error:
74  *			       2     3
75  *			      r     r    [ 3 - (R1 + R1*r/2)  ]
76  *	      expm1(r) = r + --- + --- * [--------------------]
77  *		              2     2    [ 6 - r*(3 - R1*r/2) ]
78  *
79  *	To compensate the error in the argument reduction, we use
80  *		expm1(r+c) = expm1(r) + c + expm1(r)*c
81  *			   ~ expm1(r) + c + r*c
82  *	Thus c+r*c will be added in as the correction terms for
83  *	expm1(r+c). Now rearrange the term to avoid optimization
84  * 	screw up:
85  *		        (      2                                    2 )
86  *		        ({  ( r    [ R1 -  (3 - R1*r/2) ]  )  }    r  )
87  *	 expm1(r+c)~r - ({r*(--- * [--------------------]-c)-c} - --- )
88  *	                ({  ( 2    [ 6 - r*(3 - R1*r/2) ]  )  }    2  )
89  *                      (                                             )
90  *
91  *		   = r - E
92  *   3. Scale back to obtain expm1(x):
93  *	From step 1, we have
94  *	   expm1(x) = either 2^k*[expm1(r)+1] - 1
95  *		    = or     2^k*[expm1(r) + (1-2^-k)]
96  *   4. Implementation notes:
97  *	(A). To save one multiplication, we scale the coefficient Qi
98  *	     to Qi*2^i, and replace z by (x^2)/2.
99  *	(B). To achieve maximum accuracy, we compute expm1(x) by
100  *	  (i)   if x < -56*ln2, return -1.0, (raise inexact if x != inf)
101  *	  (ii)  if k=0, return r-E
102  *	  (iii) if k=-1, return 0.5*(r-E)-0.5
103  *        (iv)	if k=1 if r < -0.25, return 2*((r+0.5)- E)
104  *					else	     return  1.0+2.0*(r-E);
105  *	  (v)   if (k<-2||k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
106  *	  (vi)  if k <= 20, return 2^k((1-2^-k)-(E-r)), else
107  *	  (vii) return 2^k(1-((E+2^-k)-r))
108  *
109  * Special cases:
110  *	expm1(INF) is INF, expm1(NaN) is NaN;
111  *	expm1(-INF) is -1, and
112  *	for finite argument, only expm1(0)=0 is exact.
113  *
114  * Accuracy:
115  *	according to an error analysis, the error is always less than
116  *	1 ulp (unit in the last place).
117  *
118  * Misc. info.
119  *	For IEEE double
120  *	    if x >  7.09782712893383973096e+02 then expm1(x) overflow
121  *
122  * Constants:
123  * The hexadecimal values are the intended ones for the following
124  * constants. The decimal values may be used, provided that the
125  * compiler will convert from decimal to binary accurately enough
126  * to produce the hexadecimal values shown.
127  */
128 /* INDENT ON */
129 
130 #include "libm_synonyms.h"	/* __expm1 */
131 #include "libm_macros.h"
132 #include <math.h>
133 
134 static const double xxx[] = {
135 /* one */		 1.0,
136 /* huge */		 1.0e+300,
137 /* tiny */		 1.0e-300,
138 /* o_threshold */	 7.09782712893383973096e+02,	/* 40862E42 FEFA39EF */
139 /* ln2_hi */		 6.93147180369123816490e-01,	/* 3FE62E42 FEE00000 */
140 /* ln2_lo */		 1.90821492927058770002e-10,	/* 3DEA39EF 35793C76 */
141 /* invln2 */		 1.44269504088896338700e+00,	/* 3FF71547 652B82FE */
142 /* scaled coefficients related to expm1 */
143 /* Q1 */		-3.33333333333331316428e-02,	/* BFA11111 111110F4 */
144 /* Q2 */		 1.58730158725481460165e-03,	/* 3F5A01A0 19FE5585 */
145 /* Q3 */		-7.93650757867487942473e-05,	/* BF14CE19 9EAADBB7 */
146 /* Q4 */		 4.00821782732936239552e-06,	/* 3ED0CFCA 86E65239 */
147 /* Q5 */		-2.01099218183624371326e-07	/* BE8AFDB7 6E09C32D */
148 };
149 #define	one		xxx[0]
150 #define	huge		xxx[1]
151 #define	tiny		xxx[2]
152 #define	o_threshold	xxx[3]
153 #define	ln2_hi		xxx[4]
154 #define	ln2_lo		xxx[5]
155 #define	invln2		xxx[6]
156 #define	Q1		xxx[7]
157 #define	Q2		xxx[8]
158 #define	Q3		xxx[9]
159 #define	Q4		xxx[10]
160 #define	Q5		xxx[11]
161 
162 double
163 expm1(double x) {
164 	double y, hi, lo, c = 0.0L, t, e, hxs, hfx, r1;
165 	int k, xsb;
166 	unsigned hx;
167 
168 	hx = ((unsigned *) &x)[HIWORD];		/* high word of x */
169 	xsb = hx & 0x80000000;			/* sign bit of x */
170 	if (xsb == 0)
171 		y = x;
172 	else
173 		y = -x;				/* y = |x| */
174 	hx &= 0x7fffffff;			/* high word of |x| */
175 
176 	/* filter out huge and non-finite argument */
177 	/* for example exp(38)-1 is approximately 3.1855932e+16 */
178 	if (hx >= 0x4043687A) {
179 		/* if |x|>=56*ln2 (~38.8162...) */
180 		if (hx >= 0x40862E42) {		/* if |x|>=709.78... -> inf */
181 			if (hx >= 0x7ff00000) {
182 				if (((hx & 0xfffff) | ((int *) &x)[LOWORD])
183 					!= 0)
184 					return (x * x);	/* + -> * for Cheetah */
185 				else
186 					/* exp(+-inf)={inf,-1} */
187 					return (xsb == 0 ? x : -1.0);
188 			}
189 			if (x > o_threshold)
190 				return (huge * huge);	/* overflow */
191 		}
192 		if (xsb != 0) {		/* x < -56*ln2, return -1.0 w/inexact */
193 			if (x + tiny < 0.0)		/* raise inexact */
194 				return (tiny - one);	/* return -1 */
195 		}
196 	}
197 
198 	/* argument reduction */
199 	if (hx > 0x3fd62e42) {			/* if  |x| > 0.5 ln2 */
200 		if (hx < 0x3FF0A2B2) {		/* and |x| < 1.5 ln2 */
201 			if (xsb == 0) {		/* positive number */
202 				hi = x - ln2_hi;
203 				lo = ln2_lo;
204 				k = 1;
205 			} else {
206 				/* negative number */
207 				hi = x + ln2_hi;
208 				lo = -ln2_lo;
209 				k = -1;
210 			}
211 		} else {
212 			/* |x| > 1.5 ln2 */
213 			k = (int) (invln2 * x + (xsb == 0 ? 0.5 : -0.5));
214 			t = k;
215 			hi = x - t * ln2_hi;	/* t*ln2_hi is exact here */
216 			lo = t * ln2_lo;
217 		}
218 		x = hi - lo;
219 		c = (hi - x) - lo; /* still at |x| > 0.5 ln2 */
220 	} else if (hx < 0x3c900000) {
221 		/* when |x|<2**-54, return x */
222 		t = huge + x;		/* return x w/inexact when x != 0 */
223 		return (x - (t - (huge + x)));
224 	} else
225 		/* |x| <= 0.5 ln2 */
226 		k = 0;
227 
228 	/* x is now in primary range */
229 	hfx = 0.5 * x;
230 	hxs = x * hfx;
231 	r1 = one + hxs * (Q1 + hxs * (Q2 + hxs * (Q3 + hxs * (Q4 + hxs * Q5))));
232 	t = 3.0 - r1 * hfx;
233 	e = hxs * ((r1 - t) / (6.0 - x * t));
234 	if (k == 0) /* |x| <= 0.5 ln2 */
235 		return (x - (x * e - hxs));
236 	else {		/* |x| > 0.5 ln2 */
237 		e = (x * (e - c) - c);
238 		e -= hxs;
239 		if (k == -1)
240 			return (0.5 * (x - e) - 0.5);
241 		if (k == 1) {
242 			if (x < -0.25)
243 				return (-2.0 * (e - (x + 0.5)));
244 			else
245 				return (one + 2.0 * (x - e));
246 		}
247 		if (k <= -2 || k > 56) {	/* suffice to return exp(x)-1 */
248 			y = one - (e - x);
249 			((int *) &y)[HIWORD] += k << 20;
250 			return (y - one);
251 		}
252 		t = one;
253 		if (k < 20) {
254 			((int *) &t)[HIWORD] = 0x3ff00000 - (0x200000 >> k);
255 							/* t = 1 - 2^-k */
256 			y = t - (e - x);
257 			((int *) &y)[HIWORD] += k << 20;
258 		} else {
259 			((int *) &t)[HIWORD] = (0x3ff - k) << 20; /* 2^-k */
260 			y = x - (e + t);
261 			y += one;
262 			((int *) &y)[HIWORD] += k << 20;
263 		}
264 	}
265 	return (y);
266 }
267