| 1 | /* Compute a product of 1 + (T/X), 1 + (T/(X+1)), .... |
| 2 | Copyright (C) 2015-2016 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 | <http://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | #include <math.h> |
| 20 | #include <math_private.h> |
| 21 | #include <float.h> |
| 22 | |
| 23 | /* Calculate X * Y exactly and store the result in *HI + *LO. It is |
| 24 | given that the values are small enough that no overflow occurs and |
| 25 | large enough (or zero) that no underflow occurs. */ |
| 26 | |
| 27 | static void |
| 28 | mul_split (long double *hi, long double *lo, long double x, long double y) |
| 29 | { |
| 30 | #ifdef __FP_FAST_FMAL |
| 31 | /* Fast built-in fused multiply-add. */ |
| 32 | *hi = x * y; |
| 33 | *lo = __builtin_fmal (x, y, -*hi); |
| 34 | #elif defined FP_FAST_FMAL |
| 35 | /* Fast library fused multiply-add, compiler before GCC 4.6. */ |
| 36 | *hi = x * y; |
| 37 | *lo = __fmal (x, y, -*hi); |
| 38 | #else |
| 39 | /* Apply Dekker's algorithm. */ |
| 40 | *hi = x * y; |
| 41 | # define C ((1LL << (LDBL_MANT_DIG + 1) / 2) + 1) |
| 42 | long double x1 = x * C; |
| 43 | long double y1 = y * C; |
| 44 | # undef C |
| 45 | x1 = (x - x1) + x1; |
| 46 | y1 = (y - y1) + y1; |
| 47 | long double x2 = x - x1; |
| 48 | long double y2 = y - y1; |
| 49 | *lo = (((x1 * y1 - *hi) + x1 * y2) + x2 * y1) + x2 * y2; |
| 50 | #endif |
| 51 | } |
| 52 | |
| 53 | /* Compute the product of 1 + (T / (X + X_EPS)), 1 + (T / (X + X_EPS + |
| 54 | 1)), ..., 1 + (T / (X + X_EPS + N - 1)), minus 1. X is such that |
| 55 | all the values X + 1, ..., X + N - 1 are exactly representable, and |
| 56 | X_EPS / X is small enough that factors quadratic in it can be |
| 57 | neglected. */ |
| 58 | |
| 59 | long double |
| 60 | __lgamma_productl (long double t, long double x, long double x_eps, int n) |
| 61 | { |
| 62 | long double ret = 0, ret_eps = 0; |
| 63 | for (int i = 0; i < n; i++) |
| 64 | { |
| 65 | long double xi = x + i; |
| 66 | long double quot = t / xi; |
| 67 | long double mhi, mlo; |
| 68 | mul_split (&mhi, &mlo, quot, xi); |
| 69 | long double quot_lo = (t - mhi - mlo) / xi - t * x_eps / (xi * xi); |
| 70 | /* We want (1 + RET + RET_EPS) * (1 + QUOT + QUOT_LO) - 1. */ |
| 71 | long double rhi, rlo; |
| 72 | mul_split (&rhi, &rlo, ret, quot); |
| 73 | long double rpq = ret + quot; |
| 74 | long double rpq_eps = (ret - rpq) + quot; |
| 75 | long double nret = rpq + rhi; |
| 76 | long double nret_eps = (rpq - nret) + rhi; |
| 77 | ret_eps += (rpq_eps + nret_eps + rlo + ret_eps * quot |
| 78 | + quot_lo + quot_lo * (ret + ret_eps)); |
| 79 | ret = nret; |
| 80 | } |
| 81 | return ret + ret_eps; |
| 82 | } |
| 83 | |