xref: /third_party/jerryscript/jerry-libm/exp.c (revision 425bb815)
1/* Copyright JS Foundation and other contributors, http://js.foundation
2 *
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at
6 *
7 *     http://www.apache.org/licenses/LICENSE-2.0
8 *
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License.
14 *
15 * This file is based on work under the following copyright and permission
16 * notice:
17 *
18 *     Copyright (C) 2004 by Sun Microsystems, Inc. All rights reserved.
19 *
20 *     Permission to use, copy, modify, and distribute this
21 *     software is freely granted, provided that this notice
22 *     is preserved.
23 *
24 *     @(#)e_exp.c 1.6 04/04/22
25 */
26
27#include "jerry-libm-internal.h"
28
29/* exp(x)
30 * Returns the exponential of x.
31 *
32 * Method:
33 *   1. Argument reduction:
34 *      Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658.
35 *      Given x, find r and integer k such that
36 *
37 *               x = k*ln2 + r,  |r| <= 0.5*ln2.
38 *
39 *      Here r will be represented as r = hi-lo for better
40 *      accuracy.
41 *
42 *   2. Approximation of exp(r) by a special rational function on
43 *      the interval [0,0.34658]:
44 *      Write
45 *          R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ...
46 *      We use a special Remes algorithm on [0,0.34658] to generate
47 *      a polynomial of degree 5 to approximate R. The maximum error
48 *      of this polynomial approximation is bounded by 2**-59. In
49 *      other words,
50 *          R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5
51 *      (where z=r*r, and the values of P1 to P5 are listed below)
52 *      and
53 *          |                  5          |     -59
54 *          | 2.0+P1*z+...+P5*z   -  R(z) | <= 2
55 *          |                             |
56 *      The computation of exp(r) thus becomes
57 *                             2*r
58 *              exp(r) = 1 + -------
59 *                            R - r
60 *                                 r*R1(r)
61 *                     = 1 + r + ----------- (for better accuracy)
62 *                                2 - R1(r)
63 *      where
64 *                               2       4             10
65 *              R1(r) = r - (P1*r  + P2*r  + ... + P5*r   ).
66 *
67 *   3. Scale back to obtain exp(x):
68 *      From step 1, we have
69 *         exp(x) = 2^k * exp(r)
70 *
71 * Special cases:
72 *      exp(INF) is INF, exp(NaN) is NaN;
73 *      exp(-INF) is 0, and
74 *      for finite argument, only exp(0)=1 is exact.
75 *
76 * Accuracy:
77 *      according to an error analysis, the error is always less than
78 *      1 ulp (unit in the last place).
79 *
80 * Misc. info:
81 *      For IEEE double
82 *          if x >  7.09782712893383973096e+02 then exp(x) overflow
83 *          if x < -7.45133219101941108420e+02 then exp(x) underflow
84 *
85 * Constants:
86 * The hexadecimal values are the intended ones for the following
87 * constants. The decimal values may be used, provided that the
88 * compiler will convert from decimal to binary accurately enough
89 * to produce the hexadecimal values shown.
90 */
91
92static const double halF[2] =
93{
94  0.5,
95  -0.5,
96};
97static const double ln2HI[2] =
98{
99  6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
100  -6.93147180369123816490e-01, /* 0xbfe62e42, 0xfee00000 */
101};
102static const double ln2LO[2] =
103{
104  1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */
105  -1.90821492927058770002e-10, /* 0xbdea39ef, 0x35793c76 */
106};
107
108#define one          1.0
109#define huge         1.0e+300
110#define twom1000     9.33263618503218878990e-302 /* 2**-1000=0x01700000,0 */
111#define o_threshold  7.09782712893383973096e+02 /* 0x40862E42, 0xFEFA39EF */
112#define u_threshold -7.45133219101941108420e+02 /* 0xc0874910, 0xD52D3051 */
113#define invln2       1.44269504088896338700e+00 /* 0x3ff71547, 0x652b82fe */
114#define P1           1.66666666666666019037e-01 /* 0x3FC55555, 0x5555553E */
115#define P2          -2.77777777770155933842e-03 /* 0xBF66C16C, 0x16BEBD93 */
116#define P3           6.61375632143793436117e-05 /* 0x3F11566A, 0xAF25DE2C */
117#define P4          -1.65339022054652515390e-06 /* 0xBEBBBD41, 0xC5D26BF1 */
118#define P5           4.13813679705723846039e-08 /* 0x3E663769, 0x72BEA4D0 */
119
120double
121exp (double x) /* default IEEE double exp */
122{
123  double hi, lo, c, t;
124  int k = 0, xsb;
125  unsigned hx;
126
127  hx = __HI (x); /* high word of x */
128  xsb = (hx >> 31) & 1; /* sign bit of x */
129  hx &= 0x7fffffff; /* high word of |x| */
130
131  /* filter out non-finite argument */
132  if (hx >= 0x40862E42) /* if |x| >= 709.78... */
133  {
134    if (hx >= 0x7ff00000)
135    {
136      if (((hx & 0xfffff) | __LO (x)) != 0) /* NaN */
137      {
138        return x + x;
139      }
140      else /* exp(+-inf) = {inf,0} */
141      {
142        return (xsb == 0) ? x : 0.0;
143      }
144    }
145    if (x > o_threshold) /* overflow */
146    {
147      return huge * huge;
148    }
149    if (x < u_threshold) /* underflow */
150    {
151      return twom1000 * twom1000;
152    }
153  }
154
155  /* argument reduction */
156  if (hx > 0x3fd62e42) /* if  |x| > 0.5 ln2 */
157  {
158    if (hx < 0x3FF0A2B2) /* and |x| < 1.5 ln2 */
159    {
160      hi = x - ln2HI[xsb];
161      lo = ln2LO[xsb];
162      k = 1 - xsb - xsb;
163    }
164    else
165    {
166      k = (int) (invln2 * x + halF[xsb]);
167      t = k;
168      hi = x - t * ln2HI[0]; /* t * ln2HI is exact here */
169      lo = t * ln2LO[0];
170    }
171    x = hi - lo;
172  }
173  else if (hx < 0x3e300000) /* when |x| < 2**-28 */
174  {
175    if (huge + x > one) /* trigger inexact */
176    {
177      return one + x;
178    }
179  }
180  else
181  {
182    k = 0;
183  }
184
185  double_accessor ret;
186
187  /* x is now in primary range */
188  t = x * x;
189  c = x - t * (P1 + t * (P2 + t * (P3 + t * (P4 + t * P5))));
190  if (k == 0)
191  {
192    return one - ((x * c) / (c - 2.0) - x);
193  }
194  else
195  {
196    ret.dbl = one - ((lo - (x * c) / (2.0 - c)) - hi);
197  }
198  if (k >= -1021)
199  {
200    ret.as_int.hi += (((unsigned int) k) << 20); /* add k to y's exponent */
201    return ret.dbl;
202  }
203  else
204  {
205    ret.as_int.hi += ((k + 1000) << 20); /* add k to y's exponent */
206    return ret.dbl * twom1000;
207  }
208} /* exp */
209
210#undef one
211#undef huge
212#undef twom1000
213#undef o_threshold
214#undef u_threshold
215#undef invln2
216#undef P1
217#undef P2
218#undef P3
219#undef P4
220#undef P5
221