1 | /* Implementation of gamma function according to ISO C. |
2 | Copyright (C) 1997-2018 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 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include <math.h> |
21 | #include <math-narrow-eval.h> |
22 | #include <math_private.h> |
23 | #include <math-underflow.h> |
24 | #include <float.h> |
25 | |
26 | /* Coefficients B_2k / 2k(2k-1) of x^-(2k-1) inside exp in Stirling's |
27 | approximation to gamma function. */ |
28 | |
29 | static const float gamma_coeff[] = |
30 | { |
31 | 0x1.555556p-4f, |
32 | -0xb.60b61p-12f, |
33 | 0x3.403404p-12f, |
34 | }; |
35 | |
36 | #define NCOEFF (sizeof (gamma_coeff) / sizeof (gamma_coeff[0])) |
37 | |
38 | /* Return gamma (X), for positive X less than 42, in the form R * |
39 | 2^(*EXP2_ADJ), where R is the return value and *EXP2_ADJ is set to |
40 | avoid overflow or underflow in intermediate calculations. */ |
41 | |
42 | static float |
43 | gammaf_positive (float x, int *exp2_adj) |
44 | { |
45 | int local_signgam; |
46 | if (x < 0.5f) |
47 | { |
48 | *exp2_adj = 0; |
49 | return __ieee754_expf (__ieee754_lgammaf_r (x + 1, &local_signgam)) / x; |
50 | } |
51 | else if (x <= 1.5f) |
52 | { |
53 | *exp2_adj = 0; |
54 | return __ieee754_expf (__ieee754_lgammaf_r (x, &local_signgam)); |
55 | } |
56 | else if (x < 2.5f) |
57 | { |
58 | *exp2_adj = 0; |
59 | float x_adj = x - 1; |
60 | return (__ieee754_expf (__ieee754_lgammaf_r (x_adj, &local_signgam)) |
61 | * x_adj); |
62 | } |
63 | else |
64 | { |
65 | float eps = 0; |
66 | float x_eps = 0; |
67 | float x_adj = x; |
68 | float prod = 1; |
69 | if (x < 4.0f) |
70 | { |
71 | /* Adjust into the range for applying Stirling's |
72 | approximation. */ |
73 | float n = __ceilf (4.0f - x); |
74 | x_adj = math_narrow_eval (x + n); |
75 | x_eps = (x - (x_adj - n)); |
76 | prod = __gamma_productf (x_adj - n, x_eps, n, &eps); |
77 | } |
78 | /* The result is now gamma (X_ADJ + X_EPS) / (PROD * (1 + EPS)). |
79 | Compute gamma (X_ADJ + X_EPS) using Stirling's approximation, |
80 | starting by computing pow (X_ADJ, X_ADJ) with a power of 2 |
81 | factored out. */ |
82 | float exp_adj = -eps; |
83 | float x_adj_int = __roundf (x_adj); |
84 | float x_adj_frac = x_adj - x_adj_int; |
85 | int x_adj_log2; |
86 | float x_adj_mant = __frexpf (x_adj, &x_adj_log2); |
87 | if (x_adj_mant < (float) M_SQRT1_2) |
88 | { |
89 | x_adj_log2--; |
90 | x_adj_mant *= 2.0f; |
91 | } |
92 | *exp2_adj = x_adj_log2 * (int) x_adj_int; |
93 | float ret = (__ieee754_powf (x_adj_mant, x_adj) |
94 | * __ieee754_exp2f (x_adj_log2 * x_adj_frac) |
95 | * __ieee754_expf (-x_adj) |
96 | * sqrtf (2 * (float) M_PI / x_adj) |
97 | / prod); |
98 | exp_adj += x_eps * __ieee754_logf (x_adj); |
99 | float bsum = gamma_coeff[NCOEFF - 1]; |
100 | float x_adj2 = x_adj * x_adj; |
101 | for (size_t i = 1; i <= NCOEFF - 1; i++) |
102 | bsum = bsum / x_adj2 + gamma_coeff[NCOEFF - 1 - i]; |
103 | exp_adj += bsum / x_adj; |
104 | return ret + ret * __expm1f (exp_adj); |
105 | } |
106 | } |
107 | |
108 | float |
109 | __ieee754_gammaf_r (float x, int *signgamp) |
110 | { |
111 | int32_t hx; |
112 | float ret; |
113 | |
114 | GET_FLOAT_WORD (hx, x); |
115 | |
116 | if (__glibc_unlikely ((hx & 0x7fffffff) == 0)) |
117 | { |
118 | /* Return value for x == 0 is Inf with divide by zero exception. */ |
119 | *signgamp = 0; |
120 | return 1.0 / x; |
121 | } |
122 | if (__builtin_expect (hx < 0, 0) |
123 | && (uint32_t) hx < 0xff800000 && __rintf (x) == x) |
124 | { |
125 | /* Return value for integer x < 0 is NaN with invalid exception. */ |
126 | *signgamp = 0; |
127 | return (x - x) / (x - x); |
128 | } |
129 | if (__glibc_unlikely (hx == 0xff800000)) |
130 | { |
131 | /* x == -Inf. According to ISO this is NaN. */ |
132 | *signgamp = 0; |
133 | return x - x; |
134 | } |
135 | if (__glibc_unlikely ((hx & 0x7f800000) == 0x7f800000)) |
136 | { |
137 | /* Positive infinity (return positive infinity) or NaN (return |
138 | NaN). */ |
139 | *signgamp = 0; |
140 | return x + x; |
141 | } |
142 | |
143 | if (x >= 36.0f) |
144 | { |
145 | /* Overflow. */ |
146 | *signgamp = 0; |
147 | ret = math_narrow_eval (FLT_MAX * FLT_MAX); |
148 | return ret; |
149 | } |
150 | else |
151 | { |
152 | SET_RESTORE_ROUNDF (FE_TONEAREST); |
153 | if (x > 0.0f) |
154 | { |
155 | *signgamp = 0; |
156 | int exp2_adj; |
157 | float tret = gammaf_positive (x, &exp2_adj); |
158 | ret = __scalbnf (tret, exp2_adj); |
159 | } |
160 | else if (x >= -FLT_EPSILON / 4.0f) |
161 | { |
162 | *signgamp = 0; |
163 | ret = 1.0f / x; |
164 | } |
165 | else |
166 | { |
167 | float tx = __truncf (x); |
168 | *signgamp = (tx == 2.0f * __truncf (tx / 2.0f)) ? -1 : 1; |
169 | if (x <= -42.0f) |
170 | /* Underflow. */ |
171 | ret = FLT_MIN * FLT_MIN; |
172 | else |
173 | { |
174 | float frac = tx - x; |
175 | if (frac > 0.5f) |
176 | frac = 1.0f - frac; |
177 | float sinpix = (frac <= 0.25f |
178 | ? __sinf ((float) M_PI * frac) |
179 | : __cosf ((float) M_PI * (0.5f - frac))); |
180 | int exp2_adj; |
181 | float tret = (float) M_PI / (-x * sinpix |
182 | * gammaf_positive (-x, &exp2_adj)); |
183 | ret = __scalbnf (tret, -exp2_adj); |
184 | math_check_force_underflow_nonneg (ret); |
185 | } |
186 | } |
187 | ret = math_narrow_eval (ret); |
188 | } |
189 | if (isinf (ret) && x != 0) |
190 | { |
191 | if (*signgamp < 0) |
192 | { |
193 | ret = math_narrow_eval (-__copysignf (FLT_MAX, ret) * FLT_MAX); |
194 | ret = -ret; |
195 | } |
196 | else |
197 | ret = math_narrow_eval (__copysignf (FLT_MAX, ret) * FLT_MAX); |
198 | return ret; |
199 | } |
200 | else if (ret == 0) |
201 | { |
202 | if (*signgamp < 0) |
203 | { |
204 | ret = math_narrow_eval (-__copysignf (FLT_MIN, ret) * FLT_MIN); |
205 | ret = -ret; |
206 | } |
207 | else |
208 | ret = math_narrow_eval (__copysignf (FLT_MIN, ret) * FLT_MIN); |
209 | return ret; |
210 | } |
211 | else |
212 | return ret; |
213 | } |
214 | strong_alias (__ieee754_gammaf_r, __gammaf_r_finite) |
215 | |