| 1 | /* Euclidean distance function. Long Double/Binary96 version. |
| 2 | Copyright (C) 2021-2023 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 | /* This implementation is based on 'An Improved Algorithm for hypot(a,b)' by |
| 20 | Carlos F. Borges [1] using the MyHypot3 with the following changes: |
| 21 | |
| 22 | - Handle qNaN and sNaN. |
| 23 | - Tune the 'widely varying operands' to avoid spurious underflow |
| 24 | due the multiplication and fix the return value for upwards |
| 25 | rounding mode. |
| 26 | - Handle required underflow exception for subnormal results. |
| 27 | |
| 28 | [1] https://arxiv.org/pdf/1904.09481.pdf */ |
| 29 | |
| 30 | #include <math.h> |
| 31 | #include <math_private.h> |
| 32 | #include <math-underflow.h> |
| 33 | #include <libm-alias-finite.h> |
| 34 | |
| 35 | #define SCALE 0x8p-8257L |
| 36 | #define LARGE_VAL 0xb.504f333f9de6484p+8188L |
| 37 | #define TINY_VAL 0x8p-8194L |
| 38 | #define EPS 0x8p-68L |
| 39 | |
| 40 | /* Hypot kernel. The inputs must be adjusted so that ax >= ay >= 0 |
| 41 | and squaring ax, ay and (ax - ay) does not overflow or underflow. */ |
| 42 | static inline long double |
| 43 | kernel (long double ax, long double ay) |
| 44 | { |
| 45 | long double t1, t2; |
| 46 | long double h = sqrtl (ax * ax + ay * ay); |
| 47 | if (h <= 2.0L * ay) |
| 48 | { |
| 49 | long double delta = h - ay; |
| 50 | t1 = ax * (2.0L * delta - ax); |
| 51 | t2 = (delta - 2.0L * (ax - ay)) * delta; |
| 52 | } |
| 53 | else |
| 54 | { |
| 55 | long double delta = h - ax; |
| 56 | t1 = 2.0L * delta * (ax - 2.0L * ay); |
| 57 | t2 = (4.0L * delta - ay) * ay + delta * delta; |
| 58 | } |
| 59 | |
| 60 | h -= (t1 + t2) / (2.0L * h); |
| 61 | return h; |
| 62 | } |
| 63 | |
| 64 | long double |
| 65 | __ieee754_hypotl (long double x, long double y) |
| 66 | { |
| 67 | if (!isfinite(x) || !isfinite(y)) |
| 68 | { |
| 69 | if ((isinf (x) || isinf (y)) |
| 70 | && !issignaling (x) && !issignaling (y)) |
| 71 | return INFINITY; |
| 72 | return x + y; |
| 73 | } |
| 74 | |
| 75 | x = fabsl (x); |
| 76 | y = fabsl (y); |
| 77 | |
| 78 | long double ax = x < y ? y : x; |
| 79 | long double ay = x < y ? x : y; |
| 80 | |
| 81 | /* If ax is huge, scale both inputs down. */ |
| 82 | if (__glibc_unlikely (ax > LARGE_VAL)) |
| 83 | { |
| 84 | if (__glibc_unlikely (ay <= ax * EPS)) |
| 85 | return ax + ay; |
| 86 | |
| 87 | return kernel (ax * SCALE, ay * SCALE) / SCALE; |
| 88 | } |
| 89 | |
| 90 | /* If ay is tiny, scale both inputs up. */ |
| 91 | if (__glibc_unlikely (ay < TINY_VAL)) |
| 92 | { |
| 93 | if (__glibc_unlikely (ax >= ay / EPS)) |
| 94 | return ax + ay; |
| 95 | |
| 96 | ax = kernel (ax / SCALE, ay / SCALE) * SCALE; |
| 97 | math_check_force_underflow_nonneg (ax); |
| 98 | return ax; |
| 99 | } |
| 100 | |
| 101 | /* Common case: ax is not huge and ay is not tiny. */ |
| 102 | if (__glibc_unlikely (ay <= ax * EPS)) |
| 103 | return ax + ay; |
| 104 | |
| 105 | return kernel (ax, ay); |
| 106 | } |
| 107 | libm_alias_finite (__ieee754_hypotl, __hypotl) |
| 108 | |