1 /* Definitions of some C99 math library functions, for those platforms
2    that don't implement these functions already. */
3 
4 #include "Python.h"
5 #include <float.h>
6 #include "_math.h"
7 
8 /* The following copyright notice applies to the original
9    implementations of acosh, asinh and atanh. */
10 
11 /*
12  * ====================================================
13  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
14  *
15  * Developed at SunPro, a Sun Microsystems, Inc. business.
16  * Permission to use, copy, modify, and distribute this
17  * software is freely granted, provided that this notice
18  * is preserved.
19  * ====================================================
20  */
21 
22 #if !defined(HAVE_ACOSH) || !defined(HAVE_ASINH)
23 static const double ln2 = 6.93147180559945286227E-01;
24 static const double two_pow_p28 = 268435456.0; /* 2**28 */
25 #endif
26 #if !defined(HAVE_ASINH) || !defined(HAVE_ATANH)
27 static const double two_pow_m28 = 3.7252902984619141E-09; /* 2**-28 */
28 #endif
29 #if !defined(HAVE_ATANH) && !defined(Py_NAN)
30 static const double zero = 0.0;
31 #endif
32 
33 
34 #ifndef HAVE_ACOSH
35 /* acosh(x)
36  * Method :
37  *      Based on
38  *            acosh(x) = log [ x + sqrt(x*x-1) ]
39  *      we have
40  *            acosh(x) := log(x)+ln2, if x is large; else
41  *            acosh(x) := log(2x-1/(sqrt(x*x-1)+x)) if x>2; else
42  *            acosh(x) := log1p(t+sqrt(2.0*t+t*t)); where t=x-1.
43  *
44  * Special cases:
45  *      acosh(x) is NaN with signal if x<1.
46  *      acosh(NaN) is NaN without signal.
47  */
48 
49 double
_Py_acosh(double x)50 _Py_acosh(double x)
51 {
52     if (Py_IS_NAN(x)) {
53         return x+x;
54     }
55     if (x < 1.) {                       /* x < 1;  return a signaling NaN */
56         errno = EDOM;
57 #ifdef Py_NAN
58         return Py_NAN;
59 #else
60         return (x-x)/(x-x);
61 #endif
62     }
63     else if (x >= two_pow_p28) {        /* x > 2**28 */
64         if (Py_IS_INFINITY(x)) {
65             return x+x;
66         }
67         else {
68             return log(x) + ln2;          /* acosh(huge)=log(2x) */
69         }
70     }
71     else if (x == 1.) {
72         return 0.0;                     /* acosh(1) = 0 */
73     }
74     else if (x > 2.) {                  /* 2 < x < 2**28 */
75         double t = x * x;
76         return log(2.0 * x - 1.0 / (x + sqrt(t - 1.0)));
77     }
78     else {                              /* 1 < x <= 2 */
79         double t = x - 1.0;
80         return m_log1p(t + sqrt(2.0 * t + t * t));
81     }
82 }
83 #endif   /* HAVE_ACOSH */
84 
85 
86 #ifndef HAVE_ASINH
87 /* asinh(x)
88  * Method :
89  *      Based on
90  *              asinh(x) = sign(x) * log [ |x| + sqrt(x*x+1) ]
91  *      we have
92  *      asinh(x) := x  if  1+x*x=1,
93  *               := sign(x)*(log(x)+ln2)) for large |x|, else
94  *               := sign(x)*log(2|x|+1/(|x|+sqrt(x*x+1))) if|x|>2, else
95  *               := sign(x)*log1p(|x| + x^2/(1 + sqrt(1+x^2)))
96  */
97 
98 double
_Py_asinh(double x)99 _Py_asinh(double x)
100 {
101     double w;
102     double absx = fabs(x);
103 
104     if (Py_IS_NAN(x) || Py_IS_INFINITY(x)) {
105         return x+x;
106     }
107     if (absx < two_pow_m28) {           /* |x| < 2**-28 */
108         return x;                       /* return x inexact except 0 */
109     }
110     if (absx > two_pow_p28) {           /* |x| > 2**28 */
111         w = log(absx) + ln2;
112     }
113     else if (absx > 2.0) {              /* 2 < |x| < 2**28 */
114         w = log(2.0 * absx + 1.0 / (sqrt(x * x + 1.0) + absx));
115     }
116     else {                              /* 2**-28 <= |x| < 2= */
117         double t = x*x;
118         w = m_log1p(absx + t / (1.0 + sqrt(1.0 + t)));
119     }
120     return copysign(w, x);
121 
122 }
123 #endif   /* HAVE_ASINH */
124 
125 
126 #ifndef HAVE_ATANH
127 /* atanh(x)
128  * Method :
129  *    1.Reduced x to positive by atanh(-x) = -atanh(x)
130  *    2.For x>=0.5
131  *                  1              2x                          x
132  *      atanh(x) = --- * log(1 + -------) = 0.5 * log1p(2 * -------)
133  *                  2             1 - x                      1 - x
134  *
135  *      For x<0.5
136  *      atanh(x) = 0.5*log1p(2x+2x*x/(1-x))
137  *
138  * Special cases:
139  *      atanh(x) is NaN if |x| >= 1 with signal;
140  *      atanh(NaN) is that NaN with no signal;
141  *
142  */
143 
144 double
_Py_atanh(double x)145 _Py_atanh(double x)
146 {
147     double absx;
148     double t;
149 
150     if (Py_IS_NAN(x)) {
151         return x+x;
152     }
153     absx = fabs(x);
154     if (absx >= 1.) {                   /* |x| >= 1 */
155         errno = EDOM;
156 #ifdef Py_NAN
157         return Py_NAN;
158 #else
159         return x / zero;
160 #endif
161     }
162     if (absx < two_pow_m28) {           /* |x| < 2**-28 */
163         return x;
164     }
165     if (absx < 0.5) {                   /* |x| < 0.5 */
166         t = absx+absx;
167         t = 0.5 * m_log1p(t + t*absx / (1.0 - absx));
168     }
169     else {                              /* 0.5 <= |x| <= 1.0 */
170         t = 0.5 * m_log1p((absx + absx) / (1.0 - absx));
171     }
172     return copysign(t, x);
173 }
174 #endif   /* HAVE_ATANH */
175 
176 
177 #ifndef HAVE_EXPM1
178 /* Mathematically, expm1(x) = exp(x) - 1.  The expm1 function is designed
179    to avoid the significant loss of precision that arises from direct
180    evaluation of the expression exp(x) - 1, for x near 0. */
181 
182 double
_Py_expm1(double x)183 _Py_expm1(double x)
184 {
185     /* For abs(x) >= log(2), it's safe to evaluate exp(x) - 1 directly; this
186        also works fine for infinities and nans.
187 
188        For smaller x, we can use a method due to Kahan that achieves close to
189        full accuracy.
190     */
191 
192     if (fabs(x) < 0.7) {
193         double u;
194         u = exp(x);
195         if (u == 1.0)
196             return x;
197         else
198             return (u - 1.0) * x / log(u);
199     }
200     else
201         return exp(x) - 1.0;
202 }
203 #endif   /* HAVE_EXPM1 */
204 
205 
206 /* log1p(x) = log(1+x).  The log1p function is designed to avoid the
207    significant loss of precision that arises from direct evaluation when x is
208    small. */
209 
210 double
_Py_log1p(double x)211 _Py_log1p(double x)
212 {
213 #ifdef HAVE_LOG1P
214     /* Some platforms supply a log1p function but don't respect the sign of
215        zero:  log1p(-0.0) gives 0.0 instead of the correct result of -0.0.
216 
217        To save fiddling with configure tests and platform checks, we handle the
218        special case of zero input directly on all platforms.
219     */
220     if (x == 0.0) {
221         return x;
222     }
223     else {
224         return log1p(x);
225     }
226 #else
227     /* For x small, we use the following approach.  Let y be the nearest float
228        to 1+x, then
229 
230          1+x = y * (1 - (y-1-x)/y)
231 
232        so log(1+x) = log(y) + log(1-(y-1-x)/y).  Since (y-1-x)/y is tiny, the
233        second term is well approximated by (y-1-x)/y.  If abs(x) >=
234        DBL_EPSILON/2 or the rounding-mode is some form of round-to-nearest
235        then y-1-x will be exactly representable, and is computed exactly by
236        (y-1)-x.
237 
238        If abs(x) < DBL_EPSILON/2 and the rounding mode is not known to be
239        round-to-nearest then this method is slightly dangerous: 1+x could be
240        rounded up to 1+DBL_EPSILON instead of down to 1, and in that case
241        y-1-x will not be exactly representable any more and the result can be
242        off by many ulps.  But this is easily fixed: for a floating-point
243        number |x| < DBL_EPSILON/2., the closest floating-point number to
244        log(1+x) is exactly x.
245     */
246 
247     double y;
248     if (fabs(x) < DBL_EPSILON / 2.) {
249         return x;
250     }
251     else if (-0.5 <= x && x <= 1.) {
252         /* WARNING: it's possible that an overeager compiler
253            will incorrectly optimize the following two lines
254            to the equivalent of "return log(1.+x)". If this
255            happens, then results from log1p will be inaccurate
256            for small x. */
257         y = 1.+x;
258         return log(y) - ((y - 1.) - x) / y;
259     }
260     else {
261         /* NaNs and infinities should end up here */
262         return log(1.+x);
263     }
264 #endif /* ifdef HAVE_LOG1P */
265 }
266 
267