1  /* Implementation of gamma function according to ISO C.
       2     Copyright (C) 1997-2023 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 long double gamma_coeff[] =
      30    {
      31      0x1.5555555555555556p-4L,
      32      -0xb.60b60b60b60b60bp-12L,
      33      0x3.4034034034034034p-12L,
      34      -0x2.7027027027027028p-12L,
      35      0x3.72a3c5631fe46aep-12L,
      36      -0x7.daac36664f1f208p-12L,
      37      0x1.a41a41a41a41a41ap-8L,
      38      -0x7.90a1b2c3d4e5f708p-8L,
      39    };
      40  
      41  #define NCOEFF (sizeof (gamma_coeff) / sizeof (gamma_coeff[0]))
      42  
      43  /* Return gamma (X), for positive X less than 1766, 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 long double
      48  gammal_positive (long double x, int *exp2_adj)
      49  {
      50    int local_signgam;
      51    if (x < 0.5L)
      52      {
      53        *exp2_adj = 0;
      54        return __ieee754_expl (__ieee754_lgammal_r (x + 1, &local_signgam)) / x;
      55      }
      56    else if (x <= 1.5L)
      57      {
      58        *exp2_adj = 0;
      59        return __ieee754_expl (__ieee754_lgammal_r (x, &local_signgam));
      60      }
      61    else if (x < 7.5L)
      62      {
      63        /* Adjust into the range for using exp (lgamma).  */
      64        *exp2_adj = 0;
      65        long double n = ceill (x - 1.5L);
      66        long double x_adj = x - n;
      67        long double eps;
      68        long double prod = __gamma_productl (x_adj, 0, n, &eps);
      69        return (__ieee754_expl (__ieee754_lgammal_r (x_adj, &local_signgam))
      70  	      * prod * (1.0L + eps));
      71      }
      72    else
      73      {
      74        long double eps = 0;
      75        long double x_eps = 0;
      76        long double x_adj = x;
      77        long double prod = 1;
      78        if (x < 13.0L)
      79  	{
      80  	  /* Adjust into the range for applying Stirling's
      81  	     approximation.  */
      82  	  long double n = ceill (13.0L - x);
      83  	  x_adj = x + n;
      84  	  x_eps = (x - (x_adj - n));
      85  	  prod = __gamma_productl (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        long double exp_adj = -eps;
      92        long double x_adj_int = roundl (x_adj);
      93        long double x_adj_frac = x_adj - x_adj_int;
      94        int x_adj_log2;
      95        long double x_adj_mant = __frexpl (x_adj, &x_adj_log2);
      96        if (x_adj_mant < M_SQRT1_2l)
      97  	{
      98  	  x_adj_log2--;
      99  	  x_adj_mant *= 2.0L;
     100  	}
     101        *exp2_adj = x_adj_log2 * (int) x_adj_int;
     102        long double ret = (__ieee754_powl (x_adj_mant, x_adj)
     103  			 * __ieee754_exp2l (x_adj_log2 * x_adj_frac)
     104  			 * __ieee754_expl (-x_adj)
     105  			 * sqrtl (2 * M_PIl / x_adj)
     106  			 / prod);
     107        exp_adj += x_eps * __ieee754_logl (x_adj);
     108        long double bsum = gamma_coeff[NCOEFF - 1];
     109        long 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 * __expm1l (exp_adj);
     114      }
     115  }
     116  
     117  long double
     118  __ieee754_gammal_r (long double x, int *signgamp)
     119  {
     120    uint32_t es, hx, lx;
     121    long double ret;
     122  
     123    GET_LDOUBLE_WORDS (es, hx, lx, x);
     124  
     125    if (__glibc_unlikely (((es & 0x7fff) | hx | lx) == 0))
     126      {
     127        /* Return value for x == 0 is Inf with divide by zero exception.  */
     128        *signgamp = 0;
     129        return 1.0 / x;
     130      }
     131    if (__glibc_unlikely (es == 0xffffffff && ((hx & 0x7fffffff) | lx) == 0))
     132      {
     133        /* x == -Inf.  According to ISO this is NaN.  */
     134        *signgamp = 0;
     135        return x - x;
     136      }
     137    if (__glibc_unlikely ((es & 0x7fff) == 0x7fff))
     138      {
     139        /* Positive infinity (return positive infinity) or NaN (return
     140  	 NaN).  */
     141        *signgamp = 0;
     142        return x + x;
     143      }
     144    if (__builtin_expect ((es & 0x8000) != 0, 0) && rintl (x) == x)
     145      {
     146        /* Return value for integer x < 0 is NaN with invalid exception.  */
     147        *signgamp = 0;
     148        return (x - x) / (x - x);
     149      }
     150  
     151    if (x >= 1756.0L)
     152      {
     153        /* Overflow.  */
     154        *signgamp = 0;
     155        return LDBL_MAX * LDBL_MAX;
     156      }
     157    else
     158      {
     159        SET_RESTORE_ROUNDL (FE_TONEAREST);
     160        if (x > 0.0L)
     161  	{
     162  	  *signgamp = 0;
     163  	  int exp2_adj;
     164  	  ret = gammal_positive (x, &exp2_adj);
     165  	  ret = __scalbnl (ret, exp2_adj);
     166  	}
     167        else if (x >= -LDBL_EPSILON / 4.0L)
     168  	{
     169  	  *signgamp = 0;
     170  	  ret = 1.0L / x;
     171  	}
     172        else
     173  	{
     174  	  long double tx = truncl (x);
     175  	  *signgamp = (tx == 2.0L * truncl (tx / 2.0L)) ? -1 : 1;
     176  	  if (x <= -1766.0L)
     177  	    /* Underflow.  */
     178  	    ret = LDBL_MIN * LDBL_MIN;
     179  	  else
     180  	    {
     181  	      long double frac = tx - x;
     182  	      if (frac > 0.5L)
     183  		frac = 1.0L - frac;
     184  	      long double sinpix = (frac <= 0.25L
     185  				    ? __sinl (M_PIl * frac)
     186  				    : __cosl (M_PIl * (0.5L - frac)));
     187  	      int exp2_adj;
     188  	      ret = M_PIl / (-x * sinpix
     189  			     * gammal_positive (-x, &exp2_adj));
     190  	      ret = __scalbnl (ret, -exp2_adj);
     191  	      math_check_force_underflow_nonneg (ret);
     192  	    }
     193  	}
     194      }
     195    if (isinf (ret) && x != 0)
     196      {
     197        if (*signgamp < 0)
     198  	return -(-copysignl (LDBL_MAX, ret) * LDBL_MAX);
     199        else
     200  	return copysignl (LDBL_MAX, ret) * LDBL_MAX;
     201      }
     202    else if (ret == 0)
     203      {
     204        if (*signgamp < 0)
     205  	return -(-copysignl (LDBL_MIN, ret) * LDBL_MIN);
     206        else
     207  	return copysignl (LDBL_MIN, ret) * LDBL_MIN;
     208      }
     209    else
     210      return ret;
     211  }
     212  libm_alias_finite (__ieee754_gammal_r, __gammal_r)