1 | /* Implementation of gamma function according to ISO C. |
2 | Copyright (C) 1997-2017 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997 and |
5 | Jakub Jelinek <jj@ultra.linux.cz, 1999. |
6 | |
7 | The GNU C Library is free software; you can redistribute it and/or |
8 | modify it under the terms of the GNU Lesser General Public |
9 | License as published by the Free Software Foundation; either |
10 | version 2.1 of the License, or (at your option) any later version. |
11 | |
12 | The GNU C Library is distributed in the hope that it will be useful, |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
15 | Lesser General Public License for more details. |
16 | |
17 | You should have received a copy of the GNU Lesser General Public |
18 | License along with the GNU C Library; if not, see |
19 | <http://www.gnu.org/licenses/>. */ |
20 | |
21 | #include <math.h> |
22 | #include <math_private.h> |
23 | #include <float.h> |
24 | |
25 | /* Coefficients B_2k / 2k(2k-1) of x^-(2k-1) inside exp in Stirling's |
26 | approximation to gamma function. */ |
27 | |
28 | static const _Float128 gamma_coeff[] = |
29 | { |
30 | L(0x1.5555555555555555555555555555p-4), |
31 | L(-0xb.60b60b60b60b60b60b60b60b60b8p-12), |
32 | L(0x3.4034034034034034034034034034p-12), |
33 | L(-0x2.7027027027027027027027027028p-12), |
34 | L(0x3.72a3c5631fe46ae1d4e700dca8f2p-12), |
35 | L(-0x7.daac36664f1f207daac36664f1f4p-12), |
36 | L(0x1.a41a41a41a41a41a41a41a41a41ap-8), |
37 | L(-0x7.90a1b2c3d4e5f708192a3b4c5d7p-8), |
38 | L(0x2.dfd2c703c0cfff430edfd2c703cp-4), |
39 | L(-0x1.6476701181f39edbdb9ce625987dp+0), |
40 | L(0xd.672219167002d3a7a9c886459cp+0), |
41 | L(-0x9.cd9292e6660d55b3f712eb9e07c8p+4), |
42 | L(0x8.911a740da740da740da740da741p+8), |
43 | L(-0x8.d0cc570e255bf59ff6eec24b49p+12), |
44 | }; |
45 | |
46 | #define NCOEFF (sizeof (gamma_coeff) / sizeof (gamma_coeff[0])) |
47 | |
48 | /* Return gamma (X), for positive X less than 1775, in the form R * |
49 | 2^(*EXP2_ADJ), where R is the return value and *EXP2_ADJ is set to |
50 | avoid overflow or underflow in intermediate calculations. */ |
51 | |
52 | static _Float128 |
53 | gammal_positive (_Float128 x, int *exp2_adj) |
54 | { |
55 | int local_signgam; |
56 | if (x < L(0.5)) |
57 | { |
58 | *exp2_adj = 0; |
59 | return __ieee754_expl (__ieee754_lgammal_r (x + 1, &local_signgam)) / x; |
60 | } |
61 | else if (x <= L(1.5)) |
62 | { |
63 | *exp2_adj = 0; |
64 | return __ieee754_expl (__ieee754_lgammal_r (x, &local_signgam)); |
65 | } |
66 | else if (x < L(12.5)) |
67 | { |
68 | /* Adjust into the range for using exp (lgamma). */ |
69 | *exp2_adj = 0; |
70 | _Float128 n = __ceill (x - L(1.5)); |
71 | _Float128 x_adj = x - n; |
72 | _Float128 eps; |
73 | _Float128 prod = __gamma_productl (x_adj, 0, n, &eps); |
74 | return (__ieee754_expl (__ieee754_lgammal_r (x_adj, &local_signgam)) |
75 | * prod * (1 + eps)); |
76 | } |
77 | else |
78 | { |
79 | _Float128 eps = 0; |
80 | _Float128 x_eps = 0; |
81 | _Float128 x_adj = x; |
82 | _Float128 prod = 1; |
83 | if (x < 24) |
84 | { |
85 | /* Adjust into the range for applying Stirling's |
86 | approximation. */ |
87 | _Float128 n = __ceill (24 - x); |
88 | x_adj = x + n; |
89 | x_eps = (x - (x_adj - n)); |
90 | prod = __gamma_productl (x_adj - n, x_eps, n, &eps); |
91 | } |
92 | /* The result is now gamma (X_ADJ + X_EPS) / (PROD * (1 + EPS)). |
93 | Compute gamma (X_ADJ + X_EPS) using Stirling's approximation, |
94 | starting by computing pow (X_ADJ, X_ADJ) with a power of 2 |
95 | factored out. */ |
96 | _Float128 exp_adj = -eps; |
97 | _Float128 x_adj_int = __roundl (x_adj); |
98 | _Float128 x_adj_frac = x_adj - x_adj_int; |
99 | int x_adj_log2; |
100 | _Float128 x_adj_mant = __frexpl (x_adj, &x_adj_log2); |
101 | if (x_adj_mant < M_SQRT1_2l) |
102 | { |
103 | x_adj_log2--; |
104 | x_adj_mant *= 2; |
105 | } |
106 | *exp2_adj = x_adj_log2 * (int) x_adj_int; |
107 | _Float128 ret = (__ieee754_powl (x_adj_mant, x_adj) |
108 | * __ieee754_exp2l (x_adj_log2 * x_adj_frac) |
109 | * __ieee754_expl (-x_adj) |
110 | * __ieee754_sqrtl (2 * M_PIl / x_adj) |
111 | / prod); |
112 | exp_adj += x_eps * __ieee754_logl (x_adj); |
113 | _Float128 bsum = gamma_coeff[NCOEFF - 1]; |
114 | _Float128 x_adj2 = x_adj * x_adj; |
115 | for (size_t i = 1; i <= NCOEFF - 1; i++) |
116 | bsum = bsum / x_adj2 + gamma_coeff[NCOEFF - 1 - i]; |
117 | exp_adj += bsum / x_adj; |
118 | return ret + ret * __expm1l (exp_adj); |
119 | } |
120 | } |
121 | |
122 | _Float128 |
123 | __ieee754_gammal_r (_Float128 x, int *signgamp) |
124 | { |
125 | int64_t hx; |
126 | u_int64_t lx; |
127 | _Float128 ret; |
128 | |
129 | GET_LDOUBLE_WORDS64 (hx, lx, x); |
130 | |
131 | if (((hx & 0x7fffffffffffffffLL) | lx) == 0) |
132 | { |
133 | /* Return value for x == 0 is Inf with divide by zero exception. */ |
134 | *signgamp = 0; |
135 | return 1.0 / x; |
136 | } |
137 | if (hx < 0 && (u_int64_t) hx < 0xffff000000000000ULL && __rintl (x) == x) |
138 | { |
139 | /* Return value for integer x < 0 is NaN with invalid exception. */ |
140 | *signgamp = 0; |
141 | return (x - x) / (x - x); |
142 | } |
143 | if (hx == 0xffff000000000000ULL && lx == 0) |
144 | { |
145 | /* x == -Inf. According to ISO this is NaN. */ |
146 | *signgamp = 0; |
147 | return x - x; |
148 | } |
149 | if ((hx & 0x7fff000000000000ULL) == 0x7fff000000000000ULL) |
150 | { |
151 | /* Positive infinity (return positive infinity) or NaN (return |
152 | NaN). */ |
153 | *signgamp = 0; |
154 | return x + x; |
155 | } |
156 | |
157 | if (x >= 1756) |
158 | { |
159 | /* Overflow. */ |
160 | *signgamp = 0; |
161 | return LDBL_MAX * LDBL_MAX; |
162 | } |
163 | else |
164 | { |
165 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
166 | if (x > 0) |
167 | { |
168 | *signgamp = 0; |
169 | int exp2_adj; |
170 | ret = gammal_positive (x, &exp2_adj); |
171 | ret = __scalbnl (ret, exp2_adj); |
172 | } |
173 | else if (x >= -LDBL_EPSILON / 4) |
174 | { |
175 | *signgamp = 0; |
176 | ret = 1 / x; |
177 | } |
178 | else |
179 | { |
180 | _Float128 tx = __truncl (x); |
181 | *signgamp = (tx == 2 * __truncl (tx / 2)) ? -1 : 1; |
182 | if (x <= -1775) |
183 | /* Underflow. */ |
184 | ret = LDBL_MIN * LDBL_MIN; |
185 | else |
186 | { |
187 | _Float128 frac = tx - x; |
188 | if (frac > L(0.5)) |
189 | frac = 1 - frac; |
190 | _Float128 sinpix = (frac <= L(0.25) |
191 | ? __sinl (M_PIl * frac) |
192 | : __cosl (M_PIl * (L(0.5) - frac))); |
193 | int exp2_adj; |
194 | ret = M_PIl / (-x * sinpix |
195 | * gammal_positive (-x, &exp2_adj)); |
196 | ret = __scalbnl (ret, -exp2_adj); |
197 | math_check_force_underflow_nonneg (ret); |
198 | } |
199 | } |
200 | } |
201 | if (isinf (ret) && x != 0) |
202 | { |
203 | if (*signgamp < 0) |
204 | return -(-__copysignl (LDBL_MAX, ret) * LDBL_MAX); |
205 | else |
206 | return __copysignl (LDBL_MAX, ret) * LDBL_MAX; |
207 | } |
208 | else if (ret == 0) |
209 | { |
210 | if (*signgamp < 0) |
211 | return -(-__copysignl (LDBL_MIN, ret) * LDBL_MIN); |
212 | else |
213 | return __copysignl (LDBL_MIN, ret) * LDBL_MIN; |
214 | } |
215 | else |
216 | return ret; |
217 | } |
218 | strong_alias (__ieee754_gammal_r, __gammal_r_finite) |
219 | |