1 /*
2  * Copyright (c) 1998, 2001, Oracle and/or its affiliates. All rights reserved.
3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4  *
5  * This code is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 only, as
7  * published by the Free Software Foundation.  Oracle designates this
8  * particular file as subject to the "Classpath" exception as provided
9  * by Oracle in the LICENSE file that accompanied this code.
10  *
11  * This code is distributed in the hope that it will be useful, but WITHOUT
12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
14  * version 2 for more details (a copy is included in the LICENSE file that
15  * accompanied this code).
16  *
17  * You should have received a copy of the GNU General Public License version
18  * 2 along with this work; if not, write to the Free Software Foundation,
19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20  *
21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22  * or visit www.oracle.com if you need additional information or have any
23  * questions.
24  */
25 
26 /* expm1(x)
27  * Returns exp(x)-1, the exponential of x minus 1.
28  *
29  * Method
30  *   1. Argument reduction:
31  *      Given x, find r and integer k such that
32  *
33  *               x = k*ln2 + r,  |r| <= 0.5*ln2 ~ 0.34658
34  *
35  *      Here a correction term c will be computed to compensate
36  *      the error in r when rounded to a floating-point number.
37  *
38  *   2. Approximating expm1(r) by a special rational function on
39  *      the interval [0,0.34658]:
40  *      Since
41  *          r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
42  *      we define R1(r*r) by
43  *          r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 * R1(r*r)
44  *      That is,
45  *          R1(r**2) = 6/r *((exp(r)+1)/(exp(r)-1) - 2/r)
46  *                   = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
47  *                   = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
48  *      We use a special Reme algorithm on [0,0.347] to generate
49  *      a polynomial of degree 5 in r*r to approximate R1. The
50  *      maximum error of this polynomial approximation is bounded
51  *      by 2**-61. In other words,
52  *          R1(z) ~ 1.0 + Q1*z + Q2*z**2 + Q3*z**3 + Q4*z**4 + Q5*z**5
53  *      where   Q1  =  -1.6666666666666567384E-2,
54  *              Q2  =   3.9682539681370365873E-4,
55  *              Q3  =  -9.9206344733435987357E-6,
56  *              Q4  =   2.5051361420808517002E-7,
57  *              Q5  =  -6.2843505682382617102E-9;
58  *      (where z=r*r, and the values of Q1 to Q5 are listed below)
59  *      with error bounded by
60  *          |                  5           |     -61
61  *          | 1.0+Q1*z+...+Q5*z   -  R1(z) | <= 2
62  *          |                              |
63  *
64  *      expm1(r) = exp(r)-1 is then computed by the following
65  *      specific way which minimize the accumulation rounding error:
66  *                             2     3
67  *                            r     r    [ 3 - (R1 + R1*r/2)  ]
68  *            expm1(r) = r + --- + --- * [--------------------]
69  *                            2     2    [ 6 - r*(3 - R1*r/2) ]
70  *
71  *      To compensate the error in the argument reduction, we use
72  *              expm1(r+c) = expm1(r) + c + expm1(r)*c
73  *                         ~ expm1(r) + c + r*c
74  *      Thus c+r*c will be added in as the correction terms for
75  *      expm1(r+c). Now rearrange the term to avoid optimization
76  *      screw up:
77  *                      (      2                                    2 )
78  *                      ({  ( r    [ R1 -  (3 - R1*r/2) ]  )  }    r  )
79  *       expm1(r+c)~r - ({r*(--- * [--------------------]-c)-c} - --- )
80  *                      ({  ( 2    [ 6 - r*(3 - R1*r/2) ]  )  }    2  )
81  *                      (                                             )
82  *
83  *                 = r - E
84  *   3. Scale back to obtain expm1(x):
85  *      From step 1, we have
86  *         expm1(x) = either 2^k*[expm1(r)+1] - 1
87  *                  = or     2^k*[expm1(r) + (1-2^-k)]
88  *   4. Implementation notes:
89  *      (A). To save one multiplication, we scale the coefficient Qi
90  *           to Qi*2^i, and replace z by (x^2)/2.
91  *      (B). To achieve maximum accuracy, we compute expm1(x) by
92  *        (i)   if x < -56*ln2, return -1.0, (raise inexact if x!=inf)
93  *        (ii)  if k=0, return r-E
94  *        (iii) if k=-1, return 0.5*(r-E)-0.5
95  *        (iv)  if k=1 if r < -0.25, return 2*((r+0.5)- E)
96  *                     else          return  1.0+2.0*(r-E);
97  *        (v)   if (k<-2||k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
98  *        (vi)  if k <= 20, return 2^k((1-2^-k)-(E-r)), else
99  *        (vii) return 2^k(1-((E+2^-k)-r))
100  *
101  * Special cases:
102  *      expm1(INF) is INF, expm1(NaN) is NaN;
103  *      expm1(-INF) is -1, and
104  *      for finite argument, only expm1(0)=0 is exact.
105  *
106  * Accuracy:
107  *      according to an error analysis, the error is always less than
108  *      1 ulp (unit in the last place).
109  *
110  * Misc. info.
111  *      For IEEE double
112  *          if x >  7.09782712893383973096e+02 then expm1(x) overflow
113  *
114  * Constants:
115  * The hexadecimal values are the intended ones for the following
116  * constants. The decimal values may be used, provided that the
117  * compiler will convert from decimal to binary accurately enough
118  * to produce the hexadecimal values shown.
119  */
120 
121 #include "fdlibm.h"
122 
123 #ifdef __STDC__
124 static const double
125 #else
126 static double
127 #endif
128 one             = 1.0,
129 huge            = 1.0e+300,
130 tiny            = 1.0e-300,
131 o_threshold     = 7.09782712893383973096e+02,/* 0x40862E42, 0xFEFA39EF */
132 ln2_hi          = 6.93147180369123816490e-01,/* 0x3fe62e42, 0xfee00000 */
133 ln2_lo          = 1.90821492927058770002e-10,/* 0x3dea39ef, 0x35793c76 */
134 invln2          = 1.44269504088896338700e+00,/* 0x3ff71547, 0x652b82fe */
135         /* scaled coefficients related to expm1 */
136 Q1  =  -3.33333333333331316428e-02, /* BFA11111 111110F4 */
137 Q2  =   1.58730158725481460165e-03, /* 3F5A01A0 19FE5585 */
138 Q3  =  -7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
139 Q4  =   4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
140 Q5  =  -2.01099218183624371326e-07; /* BE8AFDB7 6E09C32D */
141 
142 #ifdef __STDC__
expm1(double x)143         double expm1(double x)
144 #else
145         double expm1(x)
146         double x;
147 #endif
148 {
149         double y,hi,lo,c=0,t,e,hxs,hfx,r1;
150         int k,xsb;
151         unsigned hx;
152 
153         hx  = __HI(x);  /* high word of x */
154         xsb = hx&0x80000000;            /* sign bit of x */
155         if(xsb==0) y=x; else y= -x;     /* y = |x| */
156         hx &= 0x7fffffff;               /* high word of |x| */
157 
158     /* filter out huge and non-finite argument */
159         if(hx >= 0x4043687A) {                  /* if |x|>=56*ln2 */
160             if(hx >= 0x40862E42) {              /* if |x|>=709.78... */
161                 if(hx>=0x7ff00000) {
162                     if(((hx&0xfffff)|__LO(x))!=0)
163                          return x+x;     /* NaN */
164                     else return (xsb==0)? x:-1.0;/* exp(+-inf)={inf,-1} */
165                 }
166                 if(x > o_threshold) return huge*huge; /* overflow */
167             }
168             if(xsb!=0) { /* x < -56*ln2, return -1.0 with inexact */
169                 if(x+tiny<0.0)          /* raise inexact */
170                 return tiny-one;        /* return -1 */
171             }
172         }
173 
174     /* argument reduction */
175         if(hx > 0x3fd62e42) {           /* if  |x| > 0.5 ln2 */
176             if(hx < 0x3FF0A2B2) {       /* and |x| < 1.5 ln2 */
177                 if(xsb==0)
178                     {hi = x - ln2_hi; lo =  ln2_lo;  k =  1;}
179                 else
180                     {hi = x + ln2_hi; lo = -ln2_lo;  k = -1;}
181             } else {
182                 k  = invln2*x+((xsb==0)?0.5:-0.5);
183                 t  = k;
184                 hi = x - t*ln2_hi;      /* t*ln2_hi is exact here */
185                 lo = t*ln2_lo;
186             }
187             x  = hi - lo;
188             c  = (hi-x)-lo;
189         }
190         else if(hx < 0x3c900000) {      /* when |x|<2**-54, return x */
191             t = huge+x; /* return x with inexact flags when x!=0 */
192             return x - (t-(huge+x));
193         }
194         else k = 0;
195 
196     /* x is now in primary range */
197         hfx = 0.5*x;
198         hxs = x*hfx;
199         r1 = one+hxs*(Q1+hxs*(Q2+hxs*(Q3+hxs*(Q4+hxs*Q5))));
200         t  = 3.0-r1*hfx;
201         e  = hxs*((r1-t)/(6.0 - x*t));
202         if(k==0) return x - (x*e-hxs);          /* c is 0 */
203         else {
204             e  = (x*(e-c)-c);
205             e -= hxs;
206             if(k== -1) return 0.5*(x-e)-0.5;
207             if(k==1) {
208                 if(x < -0.25) return -2.0*(e-(x+0.5));
209                 else          return  one+2.0*(x-e);
210             }
211             if (k <= -2 || k>56) {   /* suffice to return exp(x)-1 */
212                 y = one-(e-x);
213                 __HI(y) += (k<<20);     /* add k to y's exponent */
214                 return y-one;
215             }
216             t = one;
217             if(k<20) {
218                 __HI(t) = 0x3ff00000 - (0x200000>>k);  /* t=1-2^-k */
219                 y = t-(e-x);
220                 __HI(y) += (k<<20);     /* add k to y's exponent */
221            } else {
222                 __HI(t)  = ((0x3ff-k)<<20);     /* 2^-k */
223                 y = x-(e+t);
224                 y += one;
225                 __HI(y) += (k<<20);     /* add k to y's exponent */
226             }
227         }
228         return y;
229 }
230