floor, ceil, trunc, truncf, truncl: Defeat GCC optimizations.
[gnulib.git] / lib / expl.c
blob1db6d6da3690c5f9cef04f51c145318ae6357e89
1 /* Exponential function.
2 Copyright (C) 2011-2018 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <https://www.gnu.org/licenses/>. */
17 #include <config.h>
19 /* Specification. */
20 #include <math.h>
22 #if HAVE_SAME_LONG_DOUBLE_AS_DOUBLE
24 long double
25 expl (long double x)
27 return exp (x);
30 #else
32 # include <float.h>
34 /* gl_expl_table[i] = exp((i - 128) * log(2)/256). */
35 extern const long double gl_expl_table[257];
37 /* A value slightly larger than log(2). */
38 #define LOG2_PLUS_EPSILON 0.6931471805599454L
40 /* Best possible approximation of log(2) as a 'long double'. */
41 #define LOG2 0.693147180559945309417232121458176568075L
43 /* Best possible approximation of 1/log(2) as a 'long double'. */
44 #define LOG2_INVERSE 1.44269504088896340735992468100189213743L
46 /* Best possible approximation of log(2)/256 as a 'long double'. */
47 #define LOG2_BY_256 0.00270760617406228636491106297444600221904L
49 /* Best possible approximation of 256/log(2) as a 'long double'. */
50 #define LOG2_BY_256_INVERSE 369.329930467574632284140718336484387181L
52 /* The upper 32 bits of log(2)/256. */
53 #define LOG2_BY_256_HI_PART 0.0027076061733168899081647396087646484375L
54 /* log(2)/256 - LOG2_HI_PART. */
55 #define LOG2_BY_256_LO_PART \
56 0.000000000000745396456746323365681353781544922399845L
58 long double
59 expl (long double x)
61 if (isnanl (x))
62 return x;
64 if (x >= (long double) LDBL_MAX_EXP * LOG2_PLUS_EPSILON)
65 /* x > LDBL_MAX_EXP * log(2)
66 hence exp(x) > 2^LDBL_MAX_EXP, overflows to Infinity. */
67 return HUGE_VALL;
69 if (x <= (long double) (LDBL_MIN_EXP - 1 - LDBL_MANT_DIG) * LOG2_PLUS_EPSILON)
70 /* x < (LDBL_MIN_EXP - 1 - LDBL_MANT_DIG) * log(2)
71 hence exp(x) < 2^(LDBL_MIN_EXP-1-LDBL_MANT_DIG),
72 underflows to zero. */
73 return 0.0L;
75 /* Decompose x into
76 x = n * log(2) + m * log(2)/256 + y
77 where
78 n is an integer,
79 m is an integer, -128 <= m <= 128,
80 y is a number, |y| <= log(2)/512 + epsilon = 0.00135...
81 Then
82 exp(x) = 2^n * exp(m * log(2)/256) * exp(y)
83 The first factor is an ldexpl() call.
84 The second factor is a table lookup.
85 The third factor is computed
86 - either as sinh(y) + cosh(y)
87 where sinh(y) is computed through the power series:
88 sinh(y) = y + y^3/3! + y^5/5! + ...
89 and cosh(y) is computed as hypot(1, sinh(y)),
90 - or as exp(2*z) = (1 + tanh(z)) / (1 - tanh(z))
91 where z = y/2
92 and tanh(z) is computed through its power series:
93 tanh(z) = z
94 - 1/3 * z^3
95 + 2/15 * z^5
96 - 17/315 * z^7
97 + 62/2835 * z^9
98 - 1382/155925 * z^11
99 + 21844/6081075 * z^13
100 - 929569/638512875 * z^15
101 + ...
102 Since |z| <= log(2)/1024 < 0.0007, the relative contribution of the
103 z^13 term is < 0.0007^12 < 2^-120 <= 2^-LDBL_MANT_DIG, therefore we
104 can truncate the series after the z^11 term.
106 Given the usual bounds LDBL_MAX_EXP <= 16384, LDBL_MIN_EXP >= -16381,
107 LDBL_MANT_DIG <= 120, we can estimate x: -11440 <= x <= 11357.
108 This means, when dividing x by log(2), where we want x mod log(2)
109 to be precise to LDBL_MANT_DIG bits, we have to use an approximation
110 to log(2) that has 14+LDBL_MANT_DIG bits. */
113 long double nm = roundl (x * LOG2_BY_256_INVERSE); /* = 256 * n + m */
114 /* n has at most 15 bits, nm therefore has at most 23 bits, therefore
115 n * LOG2_HI_PART is computed exactly, and n * LOG2_LO_PART is computed
116 with an absolute error < 2^15 * 2e-10 * 2^-LDBL_MANT_DIG. */
117 long double y_tmp = x - nm * LOG2_BY_256_HI_PART;
118 long double y = y_tmp - nm * LOG2_BY_256_LO_PART;
119 long double z = 0.5L * y;
121 /* Coefficients of the power series for tanh(z). */
122 #define TANH_COEFF_1 1.0L
123 #define TANH_COEFF_3 -0.333333333333333333333333333333333333334L
124 #define TANH_COEFF_5 0.133333333333333333333333333333333333334L
125 #define TANH_COEFF_7 -0.053968253968253968253968253968253968254L
126 #define TANH_COEFF_9 0.0218694885361552028218694885361552028218L
127 #define TANH_COEFF_11 -0.00886323552990219656886323552990219656886L
128 #define TANH_COEFF_13 0.00359212803657248101692546136990581435026L
129 #define TANH_COEFF_15 -0.00145583438705131826824948518070211191904L
131 long double z2 = z * z;
132 long double tanh_z =
133 (((((TANH_COEFF_11
134 * z2 + TANH_COEFF_9)
135 * z2 + TANH_COEFF_7)
136 * z2 + TANH_COEFF_5)
137 * z2 + TANH_COEFF_3)
138 * z2 + TANH_COEFF_1)
139 * z;
141 long double exp_y = (1.0L + tanh_z) / (1.0L - tanh_z);
143 int n = (int) roundl (nm * (1.0L / 256.0L));
144 int m = (int) nm - 256 * n;
146 return ldexpl (gl_expl_table[128 + m] * exp_y, n);
150 #endif