1 | /* Implementation of gamma function according to ISO C. |
2 | Copyright (C) 1997-2021 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997. |
5 | |
6 | The GNU C Library is free software; you can redistribute it and/or |
7 | modify it under the terms of the GNU Lesser General Public |
8 | License as published by the Free Software Foundation; either |
9 | version 2.1 of the License, or (at your option) any later version. |
10 | |
11 | The GNU C Library is distributed in the hope that it will be useful, |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 | Lesser General Public License for more details. |
15 | |
16 | You should have received a copy of the GNU Lesser General Public |
17 | License along with the GNU C Library; if not, see |
18 | <https://www.gnu.org/licenses/>. */ |
19 | |
20 | #include <math.h> |
21 | #include <math-narrow-eval.h> |
22 | #include <math_private.h> |
23 | #include <fenv_private.h> |
24 | #include <math-underflow.h> |
25 | #include <float.h> |
26 | #include <libm-alias-finite.h> |
27 | |
28 | /* Coefficients B_2k / 2k(2k-1) of x^-(2k-1) inside exp in Stirling's |
29 | approximation to gamma function. */ |
30 | |
31 | static const double gamma_coeff[] = |
32 | { |
33 | 0x1.5555555555555p-4, |
34 | -0xb.60b60b60b60b8p-12, |
35 | 0x3.4034034034034p-12, |
36 | -0x2.7027027027028p-12, |
37 | 0x3.72a3c5631fe46p-12, |
38 | -0x7.daac36664f1f4p-12, |
39 | }; |
40 | |
41 | #define NCOEFF (sizeof (gamma_coeff) / sizeof (gamma_coeff[0])) |
42 | |
43 | /* Return gamma (X), for positive X less than 184, in the form R * |
44 | 2^(*EXP2_ADJ), where R is the return value and *EXP2_ADJ is set to |
45 | avoid overflow or underflow in intermediate calculations. */ |
46 | |
47 | static double |
48 | gamma_positive (double x, int *exp2_adj) |
49 | { |
50 | int local_signgam; |
51 | if (x < 0.5) |
52 | { |
53 | *exp2_adj = 0; |
54 | return __ieee754_exp (__ieee754_lgamma_r (x + 1, &local_signgam)) / x; |
55 | } |
56 | else if (x <= 1.5) |
57 | { |
58 | *exp2_adj = 0; |
59 | return __ieee754_exp (__ieee754_lgamma_r (x, &local_signgam)); |
60 | } |
61 | else if (x < 6.5) |
62 | { |
63 | /* Adjust into the range for using exp (lgamma). */ |
64 | *exp2_adj = 0; |
65 | double n = ceil (x - 1.5); |
66 | double x_adj = x - n; |
67 | double eps; |
68 | double prod = __gamma_product (x_adj, 0, n, &eps); |
69 | return (__ieee754_exp (__ieee754_lgamma_r (x_adj, &local_signgam)) |
70 | * prod * (1.0 + eps)); |
71 | } |
72 | else |
73 | { |
74 | double eps = 0; |
75 | double x_eps = 0; |
76 | double x_adj = x; |
77 | double prod = 1; |
78 | if (x < 12.0) |
79 | { |
80 | /* Adjust into the range for applying Stirling's |
81 | approximation. */ |
82 | double n = ceil (12.0 - x); |
83 | x_adj = math_narrow_eval (x + n); |
84 | x_eps = (x - (x_adj - n)); |
85 | prod = __gamma_product (x_adj - n, x_eps, n, &eps); |
86 | } |
87 | /* The result is now gamma (X_ADJ + X_EPS) / (PROD * (1 + EPS)). |
88 | Compute gamma (X_ADJ + X_EPS) using Stirling's approximation, |
89 | starting by computing pow (X_ADJ, X_ADJ) with a power of 2 |
90 | factored out. */ |
91 | double exp_adj = -eps; |
92 | double x_adj_int = round (x_adj); |
93 | double x_adj_frac = x_adj - x_adj_int; |
94 | int x_adj_log2; |
95 | double x_adj_mant = __frexp (x_adj, &x_adj_log2); |
96 | if (x_adj_mant < M_SQRT1_2) |
97 | { |
98 | x_adj_log2--; |
99 | x_adj_mant *= 2.0; |
100 | } |
101 | *exp2_adj = x_adj_log2 * (int) x_adj_int; |
102 | double ret = (__ieee754_pow (x_adj_mant, x_adj) |
103 | * __ieee754_exp2 (x_adj_log2 * x_adj_frac) |
104 | * __ieee754_exp (-x_adj) |
105 | * sqrt (2 * M_PI / x_adj) |
106 | / prod); |
107 | exp_adj += x_eps * __ieee754_log (x_adj); |
108 | double bsum = gamma_coeff[NCOEFF - 1]; |
109 | double x_adj2 = x_adj * x_adj; |
110 | for (size_t i = 1; i <= NCOEFF - 1; i++) |
111 | bsum = bsum / x_adj2 + gamma_coeff[NCOEFF - 1 - i]; |
112 | exp_adj += bsum / x_adj; |
113 | return ret + ret * __expm1 (exp_adj); |
114 | } |
115 | } |
116 | |
117 | double |
118 | __ieee754_gamma_r (double x, int *signgamp) |
119 | { |
120 | int32_t hx; |
121 | uint32_t lx; |
122 | double ret; |
123 | |
124 | EXTRACT_WORDS (hx, lx, x); |
125 | |
126 | if (__glibc_unlikely (((hx & 0x7fffffff) | lx) == 0)) |
127 | { |
128 | /* Return value for x == 0 is Inf with divide by zero exception. */ |
129 | *signgamp = 0; |
130 | return 1.0 / x; |
131 | } |
132 | if (__builtin_expect (hx < 0, 0) |
133 | && (uint32_t) hx < 0xfff00000 && rint (x) == x) |
134 | { |
135 | /* Return value for integer x < 0 is NaN with invalid exception. */ |
136 | *signgamp = 0; |
137 | return (x - x) / (x - x); |
138 | } |
139 | if (__glibc_unlikely ((unsigned int) hx == 0xfff00000 && lx == 0)) |
140 | { |
141 | /* x == -Inf. According to ISO this is NaN. */ |
142 | *signgamp = 0; |
143 | return x - x; |
144 | } |
145 | if (__glibc_unlikely ((hx & 0x7ff00000) == 0x7ff00000)) |
146 | { |
147 | /* Positive infinity (return positive infinity) or NaN (return |
148 | NaN). */ |
149 | *signgamp = 0; |
150 | return x + x; |
151 | } |
152 | |
153 | if (x >= 172.0) |
154 | { |
155 | /* Overflow. */ |
156 | *signgamp = 0; |
157 | ret = math_narrow_eval (DBL_MAX * DBL_MAX); |
158 | return ret; |
159 | } |
160 | else |
161 | { |
162 | SET_RESTORE_ROUND (FE_TONEAREST); |
163 | if (x > 0.0) |
164 | { |
165 | *signgamp = 0; |
166 | int exp2_adj; |
167 | double tret = gamma_positive (x, &exp2_adj); |
168 | ret = __scalbn (tret, exp2_adj); |
169 | } |
170 | else if (x >= -DBL_EPSILON / 4.0) |
171 | { |
172 | *signgamp = 0; |
173 | ret = 1.0 / x; |
174 | } |
175 | else |
176 | { |
177 | double tx = trunc (x); |
178 | *signgamp = (tx == 2.0 * trunc (tx / 2.0)) ? -1 : 1; |
179 | if (x <= -184.0) |
180 | /* Underflow. */ |
181 | ret = DBL_MIN * DBL_MIN; |
182 | else |
183 | { |
184 | double frac = tx - x; |
185 | if (frac > 0.5) |
186 | frac = 1.0 - frac; |
187 | double sinpix = (frac <= 0.25 |
188 | ? __sin (M_PI * frac) |
189 | : __cos (M_PI * (0.5 - frac))); |
190 | int exp2_adj; |
191 | double tret = M_PI / (-x * sinpix |
192 | * gamma_positive (-x, &exp2_adj)); |
193 | ret = __scalbn (tret, -exp2_adj); |
194 | math_check_force_underflow_nonneg (ret); |
195 | } |
196 | } |
197 | ret = math_narrow_eval (ret); |
198 | } |
199 | if (isinf (ret) && x != 0) |
200 | { |
201 | if (*signgamp < 0) |
202 | { |
203 | ret = math_narrow_eval (-copysign (DBL_MAX, ret) * DBL_MAX); |
204 | ret = -ret; |
205 | } |
206 | else |
207 | ret = math_narrow_eval (copysign (DBL_MAX, ret) * DBL_MAX); |
208 | return ret; |
209 | } |
210 | else if (ret == 0) |
211 | { |
212 | if (*signgamp < 0) |
213 | { |
214 | ret = math_narrow_eval (-copysign (DBL_MIN, ret) * DBL_MIN); |
215 | ret = -ret; |
216 | } |
217 | else |
218 | ret = math_narrow_eval (copysign (DBL_MIN, ret) * DBL_MIN); |
219 | return ret; |
220 | } |
221 | else |
222 | return ret; |
223 | } |
224 | libm_alias_finite (__ieee754_gamma_r, __gamma_r) |
225 | |