| 1 | /* |
| 2 | * Double-precision log(x) function. |
| 3 | * |
| 4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 5 | * See https://llvm.org/LICENSE.txt for license information. |
| 6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 7 | */ |
| 8 | |
| 9 | #include <float.h> |
| 10 | #include <math.h> |
| 11 | #include <stdint.h> |
| 12 | #include "math_config.h" |
| 13 | |
| 14 | #define T __log_data.tab |
| 15 | #define T2 __log_data.tab2 |
| 16 | #define B __log_data.poly1 |
| 17 | #define A __log_data.poly |
| 18 | #define Ln2hi __log_data.ln2hi |
| 19 | #define Ln2lo __log_data.ln2lo |
| 20 | #define N (1 << LOG_TABLE_BITS) |
| 21 | #define OFF 0x3fe6000000000000 |
| 22 | |
| 23 | /* Top 16 bits of a double. */ |
| 24 | static inline uint32_t |
| 25 | top16 (double x) |
| 26 | { |
| 27 | return asuint64 (f: x) >> 48; |
| 28 | } |
| 29 | |
| 30 | double |
| 31 | log (double x) |
| 32 | { |
| 33 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
| 34 | double_t w, z, r, r2, r3, y, invc, logc, kd, hi, lo; |
| 35 | uint64_t ix, iz, tmp; |
| 36 | uint32_t top; |
| 37 | int k, i; |
| 38 | |
| 39 | ix = asuint64 (f: x); |
| 40 | top = top16 (x); |
| 41 | |
| 42 | #if LOG_POLY1_ORDER == 10 || LOG_POLY1_ORDER == 11 |
| 43 | # define LO asuint64 (1.0 - 0x1p-5) |
| 44 | # define HI asuint64 (1.0 + 0x1.1p-5) |
| 45 | #elif LOG_POLY1_ORDER == 12 |
| 46 | # define LO asuint64 (1.0 - 0x1p-4) |
| 47 | # define HI asuint64 (1.0 + 0x1.09p-4) |
| 48 | #endif |
| 49 | if (unlikely (ix - LO < HI - LO)) |
| 50 | { |
| 51 | /* Handle close to 1.0 inputs separately. */ |
| 52 | /* Fix sign of zero with downward rounding when x==1. */ |
| 53 | if (WANT_ROUNDING && unlikely (ix == asuint64 (1.0))) |
| 54 | return 0; |
| 55 | r = x - 1.0; |
| 56 | r2 = r * r; |
| 57 | r3 = r * r2; |
| 58 | #if LOG_POLY1_ORDER == 10 |
| 59 | /* Worst-case error is around 0.516 ULP. */ |
| 60 | y = r3 * (B[1] + r * B[2] + r2 * B[3] |
| 61 | + r3 * (B[4] + r * B[5] + r2 * B[6] + r3 * (B[7] + r * B[8]))); |
| 62 | w = B[0] * r2; /* B[0] == -0.5. */ |
| 63 | hi = r + w; |
| 64 | y += r - hi + w; |
| 65 | y += hi; |
| 66 | #elif LOG_POLY1_ORDER == 11 |
| 67 | /* Worst-case error is around 0.516 ULP. */ |
| 68 | y = r3 * (B[1] + r * B[2] |
| 69 | + r2 * (B[3] + r * B[4] + r2 * B[5] |
| 70 | + r3 * (B[6] + r * B[7] + r2 * B[8] + r3 * B[9]))); |
| 71 | w = B[0] * r2; /* B[0] == -0.5. */ |
| 72 | hi = r + w; |
| 73 | y += r - hi + w; |
| 74 | y += hi; |
| 75 | #elif LOG_POLY1_ORDER == 12 |
| 76 | y = r3 * (B[1] + r * B[2] + r2 * B[3] |
| 77 | + r3 * (B[4] + r * B[5] + r2 * B[6] |
| 78 | + r3 * (B[7] + r * B[8] + r2 * B[9] + r3 * B[10]))); |
| 79 | # if N <= 64 |
| 80 | /* Worst-case error is around 0.532 ULP. */ |
| 81 | w = B[0] * r2; /* B[0] == -0.5. */ |
| 82 | hi = r + w; |
| 83 | y += r - hi + w; |
| 84 | y += hi; |
| 85 | # else |
| 86 | /* Worst-case error is around 0.507 ULP. */ |
| 87 | w = r * 0x1p27; |
| 88 | double_t rhi = r + w - w; |
| 89 | double_t rlo = r - rhi; |
| 90 | w = rhi * rhi * B[0]; /* B[0] == -0.5. */ |
| 91 | hi = r + w; |
| 92 | lo = r - hi + w; |
| 93 | lo += B[0] * rlo * (rhi + r); |
| 94 | y += lo; |
| 95 | y += hi; |
| 96 | # endif |
| 97 | #endif |
| 98 | return eval_as_double (x: y); |
| 99 | } |
| 100 | if (unlikely (top - 0x0010 >= 0x7ff0 - 0x0010)) |
| 101 | { |
| 102 | /* x < 0x1p-1022 or inf or nan. */ |
| 103 | if (ix * 2 == 0) |
| 104 | return __math_divzero (1); |
| 105 | if (ix == asuint64 (INFINITY)) /* log(inf) == inf. */ |
| 106 | return x; |
| 107 | if ((top & 0x8000) || (top & 0x7ff0) == 0x7ff0) |
| 108 | return __math_invalid (x); |
| 109 | /* x is subnormal, normalize it. */ |
| 110 | ix = asuint64 (f: x * 0x1p52); |
| 111 | ix -= 52ULL << 52; |
| 112 | } |
| 113 | |
| 114 | /* x = 2^k z; where z is in range [OFF,2*OFF) and exact. |
| 115 | The range is split into N subintervals. |
| 116 | The ith subinterval contains z and c is near its center. */ |
| 117 | tmp = ix - OFF; |
| 118 | i = (tmp >> (52 - LOG_TABLE_BITS)) % N; |
| 119 | k = (int64_t) tmp >> 52; /* arithmetic shift */ |
| 120 | iz = ix - (tmp & 0xfffULL << 52); |
| 121 | invc = T[i].invc; |
| 122 | logc = T[i].logc; |
| 123 | z = asdouble (i: iz); |
| 124 | |
| 125 | /* log(x) = log1p(z/c-1) + log(c) + k*Ln2. */ |
| 126 | /* r ~= z/c - 1, |r| < 1/(2*N). */ |
| 127 | #if HAVE_FAST_FMA |
| 128 | /* rounding error: 0x1p-55/N. */ |
| 129 | r = fma (z, invc, -1.0); |
| 130 | #else |
| 131 | /* rounding error: 0x1p-55/N + 0x1p-66. */ |
| 132 | r = (z - T2[i].chi - T2[i].clo) * invc; |
| 133 | #endif |
| 134 | kd = (double_t) k; |
| 135 | |
| 136 | /* hi + lo = r + log(c) + k*Ln2. */ |
| 137 | w = kd * Ln2hi + logc; |
| 138 | hi = w + r; |
| 139 | lo = w - hi + r + kd * Ln2lo; |
| 140 | |
| 141 | /* log(x) = lo + (log1p(r) - r) + hi. */ |
| 142 | r2 = r * r; /* rounding error: 0x1p-54/N^2. */ |
| 143 | /* Worst case error if |y| > 0x1p-5: |
| 144 | 0.5 + 4.13/N + abs-poly-error*2^57 ULP (+ 0.002 ULP without fma) |
| 145 | Worst case error if |y| > 0x1p-4: |
| 146 | 0.5 + 2.06/N + abs-poly-error*2^56 ULP (+ 0.001 ULP without fma). */ |
| 147 | #if LOG_POLY_ORDER == 6 |
| 148 | y = lo + r2 * A[0] + r * r2 * (A[1] + r * A[2] + r2 * (A[3] + r * A[4])) + hi; |
| 149 | #elif LOG_POLY_ORDER == 7 |
| 150 | y = lo |
| 151 | + r2 * (A[0] + r * A[1] + r2 * (A[2] + r * A[3]) |
| 152 | + r2 * r2 * (A[4] + r * A[5])) |
| 153 | + hi; |
| 154 | #endif |
| 155 | return eval_as_double (x: y); |
| 156 | } |
| 157 | #if USE_GLIBC_ABI |
| 158 | strong_alias (log, __log_finite) |
| 159 | hidden_alias (log, __ieee754_log) |
| 160 | # if LDBL_MANT_DIG == 53 |
| 161 | long double logl (long double x) { return log (x); } |
| 162 | # endif |
| 163 | #endif |
| 164 | |