1 /* Implementation of gamma function according to ISO C.
2    Copyright (C) 1997-2022 Free Software Foundation, Inc.
3    This file is part of the GNU C Library.
4 
5    The GNU C Library is free software; you can redistribute it and/or
6    modify it under the terms of the GNU Lesser General Public
7    License as published by the Free Software Foundation; either
8    version 2.1 of the License, or (at your option) any later version.
9 
10    The GNU C Library is distributed in the hope that it will be useful,
11    but WITHOUT ANY WARRANTY; without even the implied warranty of
12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13    Lesser General Public License for more details.
14 
15    You should have received a copy of the GNU Lesser General Public
16    License along with the GNU C Library; if not, see
17    <https://www.gnu.org/licenses/>.  */
18 
19 #include <math.h>
20 #include <math_private.h>
21 #include <fenv_private.h>
22 #include <math-underflow.h>
23 #include <float.h>
24 #include <libm-alias-finite.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 _Float128 gamma_coeff[] =
30   {
31     L(0x1.5555555555555555555555555555p-4),
32     L(-0xb.60b60b60b60b60b60b60b60b60b8p-12),
33     L(0x3.4034034034034034034034034034p-12),
34     L(-0x2.7027027027027027027027027028p-12),
35     L(0x3.72a3c5631fe46ae1d4e700dca8f2p-12),
36     L(-0x7.daac36664f1f207daac36664f1f4p-12),
37     L(0x1.a41a41a41a41a41a41a41a41a41ap-8),
38     L(-0x7.90a1b2c3d4e5f708192a3b4c5d7p-8),
39     L(0x2.dfd2c703c0cfff430edfd2c703cp-4),
40     L(-0x1.6476701181f39edbdb9ce625987dp+0),
41     L(0xd.672219167002d3a7a9c886459cp+0),
42     L(-0x9.cd9292e6660d55b3f712eb9e07c8p+4),
43     L(0x8.911a740da740da740da740da741p+8),
44     L(-0x8.d0cc570e255bf59ff6eec24b49p+12),
45   };
46 
47 #define NCOEFF (sizeof (gamma_coeff) / sizeof (gamma_coeff[0]))
48 
49 /* Return gamma (X), for positive X less than 1775, in the form R *
50    2^(*EXP2_ADJ), where R is the return value and *EXP2_ADJ is set to
51    avoid overflow or underflow in intermediate calculations.  */
52 
53 static _Float128
gammal_positive(_Float128 x,int * exp2_adj)54 gammal_positive (_Float128 x, int *exp2_adj)
55 {
56   int local_signgam;
57   if (x < L(0.5))
58     {
59       *exp2_adj = 0;
60       return __ieee754_expl (__ieee754_lgammal_r (x + 1, &local_signgam)) / x;
61     }
62   else if (x <= L(1.5))
63     {
64       *exp2_adj = 0;
65       return __ieee754_expl (__ieee754_lgammal_r (x, &local_signgam));
66     }
67   else if (x < L(12.5))
68     {
69       /* Adjust into the range for using exp (lgamma).  */
70       *exp2_adj = 0;
71       _Float128 n = ceill (x - L(1.5));
72       _Float128 x_adj = x - n;
73       _Float128 eps;
74       _Float128 prod = __gamma_productl (x_adj, 0, n, &eps);
75       return (__ieee754_expl (__ieee754_lgammal_r (x_adj, &local_signgam))
76 	      * prod * (1 + eps));
77     }
78   else
79     {
80       _Float128 eps = 0;
81       _Float128 x_eps = 0;
82       _Float128 x_adj = x;
83       _Float128 prod = 1;
84       if (x < 24)
85 	{
86 	  /* Adjust into the range for applying Stirling's
87 	     approximation.  */
88 	  _Float128 n = ceill (24 - x);
89 	  x_adj = x + n;
90 	  x_eps = (x - (x_adj - n));
91 	  prod = __gamma_productl (x_adj - n, x_eps, n, &eps);
92 	}
93       /* The result is now gamma (X_ADJ + X_EPS) / (PROD * (1 + EPS)).
94 	 Compute gamma (X_ADJ + X_EPS) using Stirling's approximation,
95 	 starting by computing pow (X_ADJ, X_ADJ) with a power of 2
96 	 factored out.  */
97       _Float128 exp_adj = -eps;
98       _Float128 x_adj_int = roundl (x_adj);
99       _Float128 x_adj_frac = x_adj - x_adj_int;
100       int x_adj_log2;
101       _Float128 x_adj_mant = __frexpl (x_adj, &x_adj_log2);
102       if (x_adj_mant < M_SQRT1_2l)
103 	{
104 	  x_adj_log2--;
105 	  x_adj_mant *= 2;
106 	}
107       *exp2_adj = x_adj_log2 * (int) x_adj_int;
108       _Float128 ret = (__ieee754_powl (x_adj_mant, x_adj)
109 		       * __ieee754_exp2l (x_adj_log2 * x_adj_frac)
110 		       * __ieee754_expl (-x_adj)
111 		       * sqrtl (2 * M_PIl / x_adj)
112 		       / prod);
113       exp_adj += x_eps * __ieee754_logl (x_adj);
114       _Float128 bsum = gamma_coeff[NCOEFF - 1];
115       _Float128 x_adj2 = x_adj * x_adj;
116       for (size_t i = 1; i <= NCOEFF - 1; i++)
117 	bsum = bsum / x_adj2 + gamma_coeff[NCOEFF - 1 - i];
118       exp_adj += bsum / x_adj;
119       return ret + ret * __expm1l (exp_adj);
120     }
121 }
122 
123 _Float128
__ieee754_gammal_r(_Float128 x,int * signgamp)124 __ieee754_gammal_r (_Float128 x, int *signgamp)
125 {
126   int64_t hx;
127   uint64_t lx;
128   _Float128 ret;
129 
130   GET_LDOUBLE_WORDS64 (hx, lx, x);
131 
132   if (((hx & 0x7fffffffffffffffLL) | lx) == 0)
133     {
134       /* Return value for x == 0 is Inf with divide by zero exception.  */
135       *signgamp = 0;
136       return 1.0 / x;
137     }
138   if (hx < 0 && (uint64_t) hx < 0xffff000000000000ULL && rintl (x) == x)
139     {
140       /* Return value for integer x < 0 is NaN with invalid exception.  */
141       *signgamp = 0;
142       return (x - x) / (x - x);
143     }
144   if (hx == 0xffff000000000000ULL && lx == 0)
145     {
146       /* x == -Inf.  According to ISO this is NaN.  */
147       *signgamp = 0;
148       return x - x;
149     }
150   if ((hx & 0x7fff000000000000ULL) == 0x7fff000000000000ULL)
151     {
152       /* Positive infinity (return positive infinity) or NaN (return
153 	 NaN).  */
154       *signgamp = 0;
155       return x + x;
156     }
157 
158   if (x >= 1756)
159     {
160       /* Overflow.  */
161       *signgamp = 0;
162       return LDBL_MAX * LDBL_MAX;
163     }
164   else
165     {
166       SET_RESTORE_ROUNDL (FE_TONEAREST);
167       if (x > 0)
168 	{
169 	  *signgamp = 0;
170 	  int exp2_adj;
171 	  ret = gammal_positive (x, &exp2_adj);
172 	  ret = __scalbnl (ret, exp2_adj);
173 	}
174       else if (x >= -LDBL_EPSILON / 4)
175 	{
176 	  *signgamp = 0;
177 	  ret = 1 / x;
178 	}
179       else
180 	{
181 	  _Float128 tx = truncl (x);
182 	  *signgamp = (tx == 2 * truncl (tx / 2)) ? -1 : 1;
183 	  if (x <= -1775)
184 	    /* Underflow.  */
185 	    ret = LDBL_MIN * LDBL_MIN;
186 	  else
187 	    {
188 	      _Float128 frac = tx - x;
189 	      if (frac > L(0.5))
190 		frac = 1 - frac;
191 	      _Float128 sinpix = (frac <= L(0.25)
192 				  ? __sinl (M_PIl * frac)
193 				  : __cosl (M_PIl * (L(0.5) - frac)));
194 	      int exp2_adj;
195 	      ret = M_PIl / (-x * sinpix
196 			     * gammal_positive (-x, &exp2_adj));
197 	      ret = __scalbnl (ret, -exp2_adj);
198 	      math_check_force_underflow_nonneg (ret);
199 	    }
200 	}
201     }
202   if (isinf (ret) && x != 0)
203     {
204       if (*signgamp < 0)
205 	return -(-copysignl (LDBL_MAX, ret) * LDBL_MAX);
206       else
207 	return copysignl (LDBL_MAX, ret) * LDBL_MAX;
208     }
209   else if (ret == 0)
210     {
211       if (*signgamp < 0)
212 	return -(-copysignl (LDBL_MIN, ret) * LDBL_MIN);
213       else
214 	return copysignl (LDBL_MIN, ret) * LDBL_MIN;
215     }
216   else
217     return ret;
218 }
219 libm_alias_finite (__ieee754_gammal_r, __gammal_r)
220