| 1 | //===----------------------------------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | /// |
| 9 | /// \file |
| 10 | /// Common utilities and tables for double-precision pow(x, y). |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_POW_UTILS_H |
| 15 | #define LLVM_LIBC_SRC___SUPPORT_MATH_POW_UTILS_H |
| 16 | |
| 17 | #include "hdr/errno_macros.h" |
| 18 | #include "hdr/fenv_macros.h" |
| 19 | #include "src/__support/CPP/bit.h" |
| 20 | #include "src/__support/CPP/optional.h" |
| 21 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 22 | #include "src/__support/FPUtil/FPBits.h" |
| 23 | #include "src/__support/FPUtil/double_double.h" |
| 24 | #include "src/__support/FPUtil/nearest_integer.h" |
| 25 | #include "src/__support/FPUtil/rounding_mode.h" |
| 26 | #include "src/__support/FPUtil/sqrt.h" |
| 27 | #include "src/__support/common.h" |
| 28 | #include "src/__support/macros/attributes.h" |
| 29 | #include "src/__support/macros/config.h" |
| 30 | #include "src/__support/macros/optimization.h" |
| 31 | #include "src/__support/macros/properties/cpu_features.h" |
| 32 | #include "src/__support/math/common_constants.h" |
| 33 | #include "src/__support/math/exp10.h" |
| 34 | #include "src/__support/math/exp2.h" |
| 35 | |
| 36 | namespace LIBC_NAMESPACE_DECL { |
| 37 | namespace math { |
| 38 | namespace pow_internal { |
| 39 | |
| 40 | using fputil::DoubleDouble; |
| 41 | using namespace common_constants_internal; |
| 42 | |
| 43 | // Constants for log2(x) range reduction, generated by Sollya with: |
| 44 | // > for i from 0 to 127 do { |
| 45 | // r = 2^-8 * ceil( 2^8 * (1 - 2^(-8)) / (1 + i*2^-7) ); |
| 46 | // b = nearestint(log2(r) * 2^41) * 2^-41; |
| 47 | // c = round(log2(r) - b, D, RN); |
| 48 | // print("{ ", -c, ", ", -b, " }, "); |
| 49 | // }; |
| 50 | // This is the same as -log2(RD[i]), with the least significant bits of the |
| 51 | // high part set to be 2^-41, so that the sum of high parts + e_x is exact in |
| 52 | // double precision. |
| 53 | // We also replace the first and the last ones to be 0. |
| 54 | LIBC_INLINE_VAR constexpr DoubleDouble LOG2_R_DD[128] = { |
| 55 | {.lo: 0.0, .hi: 0.0}, |
| 56 | {.lo: -0x1.19b14945cf6bap-44, .hi: 0x1.72c7ba21p-7}, |
| 57 | {.lo: -0x1.95539356f93dcp-43, .hi: 0x1.743ee862p-6}, |
| 58 | {.lo: 0x1.abe0a48f83604p-43, .hi: 0x1.184b8e4c5p-5}, |
| 59 | {.lo: 0x1.635577970e04p-43, .hi: 0x1.77394c9d9p-5}, |
| 60 | {.lo: -0x1.401fbaaa67e3cp-45, .hi: 0x1.d6ebd1f2p-5}, |
| 61 | {.lo: -0x1.5b1799ceaeb51p-43, .hi: 0x1.1bb32a6008p-4}, |
| 62 | {.lo: 0x1.7c407050799bfp-43, .hi: 0x1.4c560fe688p-4}, |
| 63 | {.lo: 0x1.da6339da288fcp-43, .hi: 0x1.7d60496cf8p-4}, |
| 64 | {.lo: 0x1.be4f6f22dbbadp-43, .hi: 0x1.960caf9ab8p-4}, |
| 65 | {.lo: -0x1.c760bc9b188c4p-45, .hi: 0x1.c7b528b71p-4}, |
| 66 | {.lo: 0x1.164e932b2d51cp-44, .hi: 0x1.f9c95dc1dp-4}, |
| 67 | {.lo: 0x1.924ae921f7ecap-45, .hi: 0x1.097e38ce6p-3}, |
| 68 | {.lo: -0x1.6d25a5b8a19b2p-44, .hi: 0x1.22dadc2ab4p-3}, |
| 69 | {.lo: 0x1.e50a1644ac794p-43, .hi: 0x1.3c6fb650ccp-3}, |
| 70 | {.lo: 0x1.f34baa74a7942p-43, .hi: 0x1.494f863b8cp-3}, |
| 71 | {.lo: -0x1.8f7aac147fdc1p-46, .hi: 0x1.633a8bf438p-3}, |
| 72 | {.lo: 0x1.f84be19cb9578p-43, .hi: 0x1.7046031c78p-3}, |
| 73 | {.lo: -0x1.66cccab240e9p-46, .hi: 0x1.8a8980abfcp-3}, |
| 74 | {.lo: -0x1.3f7a55cd2af4cp-47, .hi: 0x1.97c1cb13c8p-3}, |
| 75 | {.lo: 0x1.3458cde69308cp-43, .hi: 0x1.b2602497d4p-3}, |
| 76 | {.lo: -0x1.667f21fa8423fp-44, .hi: 0x1.bfc67a8p-3}, |
| 77 | {.lo: 0x1.d2fe4574e09b9p-47, .hi: 0x1.dac22d3e44p-3}, |
| 78 | {.lo: 0x1.367bde40c5e6dp-43, .hi: 0x1.e857d3d36p-3}, |
| 79 | {.lo: 0x1.d45da26510033p-46, .hi: 0x1.01d9bbcfa6p-2}, |
| 80 | {.lo: -0x1.7204f55bbf90dp-44, .hi: 0x1.08bce0d96p-2}, |
| 81 | {.lo: -0x1.d4f1b95e0ff45p-43, .hi: 0x1.169c05364p-2}, |
| 82 | {.lo: 0x1.c20d74c0211bfp-44, .hi: 0x1.1d982c9d52p-2}, |
| 83 | {.lo: 0x1.ad89a083e072ap-43, .hi: 0x1.249cd2b13cp-2}, |
| 84 | {.lo: 0x1.cd0cb4492f1bcp-43, .hi: 0x1.32bfee370ep-2}, |
| 85 | {.lo: -0x1.2101a9685c779p-47, .hi: 0x1.39de8e155ap-2}, |
| 86 | {.lo: 0x1.9451cd394fe8dp-43, .hi: 0x1.4106017c3ep-2}, |
| 87 | {.lo: 0x1.661e393a16b95p-44, .hi: 0x1.4f6fbb2cecp-2}, |
| 88 | {.lo: -0x1.c6d8d86531d56p-44, .hi: 0x1.56b22e6b58p-2}, |
| 89 | {.lo: 0x1.c1c885adb21d3p-43, .hi: 0x1.5dfdcf1eeap-2}, |
| 90 | {.lo: 0x1.3bb5921006679p-45, .hi: 0x1.6552b49986p-2}, |
| 91 | {.lo: 0x1.1d406db502403p-43, .hi: 0x1.6cb0f6865cp-2}, |
| 92 | {.lo: 0x1.55a63e278bad5p-43, .hi: 0x1.7b89f02cf2p-2}, |
| 93 | {.lo: -0x1.66ae2a7ada553p-49, .hi: 0x1.8304d90c12p-2}, |
| 94 | {.lo: -0x1.66cccab240e9p-45, .hi: 0x1.8a8980abfcp-2}, |
| 95 | {.lo: -0x1.62404772a151dp-45, .hi: 0x1.921800924ep-2}, |
| 96 | {.lo: 0x1.ac9bca36fd02ep-44, .hi: 0x1.99b072a96cp-2}, |
| 97 | {.lo: 0x1.4bc302ffa76fbp-43, .hi: 0x1.a8ff97181p-2}, |
| 98 | {.lo: 0x1.01fea1ec47c71p-43, .hi: 0x1.b0b67f4f46p-2}, |
| 99 | {.lo: -0x1.f20203b3186a6p-43, .hi: 0x1.b877c57b1cp-2}, |
| 100 | {.lo: -0x1.2642415d47384p-45, .hi: 0x1.c043859e3p-2}, |
| 101 | {.lo: -0x1.bc76a2753b99bp-50, .hi: 0x1.c819dc2d46p-2}, |
| 102 | {.lo: -0x1.da93ae3a5f451p-43, .hi: 0x1.cffae611aep-2}, |
| 103 | {.lo: -0x1.50e785694a8c6p-43, .hi: 0x1.d7e6c0abc4p-2}, |
| 104 | {.lo: 0x1.c56138c894641p-43, .hi: 0x1.dfdd89d586p-2}, |
| 105 | {.lo: 0x1.5669df6a2b592p-43, .hi: 0x1.e7df5fe538p-2}, |
| 106 | {.lo: -0x1.ea92d9e0e8ac2p-48, .hi: 0x1.efec61b012p-2}, |
| 107 | {.lo: 0x1.a0331af2e6feap-43, .hi: 0x1.f804ae8d0cp-2}, |
| 108 | {.lo: 0x1.9518ce032f41dp-48, .hi: 0x1.0014332bep-1}, |
| 109 | {.lo: -0x1.b3b3864c60011p-44, .hi: 0x1.042bd4b9a8p-1}, |
| 110 | {.lo: -0x1.103e8f00d41c8p-45, .hi: 0x1.08494c66b9p-1}, |
| 111 | {.lo: 0x1.65be75cc3da17p-43, .hi: 0x1.0c6caaf0c5p-1}, |
| 112 | {.lo: 0x1.3676289cd3dd4p-43, .hi: 0x1.1096015deep-1}, |
| 113 | {.lo: -0x1.41dfc7d7c3321p-43, .hi: 0x1.14c560fe69p-1}, |
| 114 | {.lo: 0x1.e0cda8bd74461p-44, .hi: 0x1.18fadb6e2dp-1}, |
| 115 | {.lo: 0x1.2a606046ad444p-44, .hi: 0x1.1d368296b5p-1}, |
| 116 | {.lo: 0x1.f9ea977a639cp-43, .hi: 0x1.217868b0c3p-1}, |
| 117 | {.lo: -0x1.50520a377c7ecp-45, .hi: 0x1.25c0a0463cp-1}, |
| 118 | {.lo: 0x1.6e3cb71b554e7p-47, .hi: 0x1.2a0f3c3407p-1}, |
| 119 | {.lo: -0x1.4275f1035e5e8p-48, .hi: 0x1.2e644fac05p-1}, |
| 120 | {.lo: -0x1.4275f1035e5e8p-48, .hi: 0x1.2e644fac05p-1}, |
| 121 | {.lo: -0x1.979a5db68721dp-45, .hi: 0x1.32bfee370fp-1}, |
| 122 | {.lo: 0x1.1ee969a95f529p-43, .hi: 0x1.37222bb707p-1}, |
| 123 | {.lo: 0x1.bb4b69336b66ep-43, .hi: 0x1.3b8b1c68fap-1}, |
| 124 | {.lo: 0x1.d5e6a8a4fb059p-45, .hi: 0x1.3ffad4e74fp-1}, |
| 125 | {.lo: 0x1.3106e404cabb7p-44, .hi: 0x1.44716a2c08p-1}, |
| 126 | {.lo: 0x1.3106e404cabb7p-44, .hi: 0x1.44716a2c08p-1}, |
| 127 | {.lo: -0x1.9bcaf1aa4168ap-43, .hi: 0x1.48eef19318p-1}, |
| 128 | {.lo: 0x1.1646b761c48dep-44, .hi: 0x1.4d7380dcc4p-1}, |
| 129 | {.lo: 0x1.2f0c0bfe9dbecp-43, .hi: 0x1.51ff2e3021p-1}, |
| 130 | {.lo: 0x1.29904613e33cp-43, .hi: 0x1.5692101d9bp-1}, |
| 131 | {.lo: 0x1.1d406db502403p-44, .hi: 0x1.5b2c3da197p-1}, |
| 132 | {.lo: 0x1.1d406db502403p-44, .hi: 0x1.5b2c3da197p-1}, |
| 133 | {.lo: -0x1.125d6cbcd1095p-44, .hi: 0x1.5fcdce2728p-1}, |
| 134 | {.lo: -0x1.bd9b32266d92cp-43, .hi: 0x1.6476d98adap-1}, |
| 135 | {.lo: 0x1.54243b21709cep-44, .hi: 0x1.6927781d93p-1}, |
| 136 | {.lo: 0x1.54243b21709cep-44, .hi: 0x1.6927781d93p-1}, |
| 137 | {.lo: -0x1.ce60916e52e91p-44, .hi: 0x1.6ddfc2a79p-1}, |
| 138 | {.lo: 0x1.f1f5ae718f241p-43, .hi: 0x1.729fd26b7p-1}, |
| 139 | {.lo: -0x1.6eb9612e0b4f3p-43, .hi: 0x1.7767c12968p-1}, |
| 140 | {.lo: -0x1.6eb9612e0b4f3p-43, .hi: 0x1.7767c12968p-1}, |
| 141 | {.lo: 0x1.fed21f9cb2cc5p-43, .hi: 0x1.7c37a9227ep-1}, |
| 142 | {.lo: 0x1.7f5dc57266758p-43, .hi: 0x1.810fa51bf6p-1}, |
| 143 | {.lo: 0x1.7f5dc57266758p-43, .hi: 0x1.810fa51bf6p-1}, |
| 144 | {.lo: 0x1.5b338360c2ae2p-43, .hi: 0x1.85efd062c6p-1}, |
| 145 | {.lo: -0x1.96fc8f4b56502p-43, .hi: 0x1.8ad846cf37p-1}, |
| 146 | {.lo: -0x1.96fc8f4b56502p-43, .hi: 0x1.8ad846cf37p-1}, |
| 147 | {.lo: -0x1.bdc81c4db3134p-44, .hi: 0x1.8fc924c89bp-1}, |
| 148 | {.lo: 0x1.36c101ee1344p-43, .hi: 0x1.94c287492cp-1}, |
| 149 | {.lo: 0x1.36c101ee1344p-43, .hi: 0x1.94c287492cp-1}, |
| 150 | {.lo: 0x1.e41fa0a62e6aep-44, .hi: 0x1.99c48be206p-1}, |
| 151 | {.lo: -0x1.d97ee9124773bp-46, .hi: 0x1.9ecf50bf44p-1}, |
| 152 | {.lo: -0x1.d97ee9124773bp-46, .hi: 0x1.9ecf50bf44p-1}, |
| 153 | {.lo: -0x1.3f94e00e7d6bcp-46, .hi: 0x1.a3e2f4ac44p-1}, |
| 154 | {.lo: -0x1.6879fa00b120ap-43, .hi: 0x1.a8ff971811p-1}, |
| 155 | {.lo: -0x1.6879fa00b120ap-43, .hi: 0x1.a8ff971811p-1}, |
| 156 | {.lo: 0x1.1659d8e2d7d38p-44, .hi: 0x1.ae255819fp-1}, |
| 157 | {.lo: 0x1.1e5e0ae0d3f8ap-43, .hi: 0x1.b35458761dp-1}, |
| 158 | {.lo: 0x1.1e5e0ae0d3f8ap-43, .hi: 0x1.b35458761dp-1}, |
| 159 | {.lo: 0x1.484a15babcf88p-43, .hi: 0x1.b88cb9a2abp-1}, |
| 160 | {.lo: 0x1.484a15babcf88p-43, .hi: 0x1.b88cb9a2abp-1}, |
| 161 | {.lo: 0x1.871a7610e40bdp-45, .hi: 0x1.bdce9dcc96p-1}, |
| 162 | {.lo: -0x1.2d90e5edaeceep-43, .hi: 0x1.c31a27dd01p-1}, |
| 163 | {.lo: -0x1.2d90e5edaeceep-43, .hi: 0x1.c31a27dd01p-1}, |
| 164 | {.lo: -0x1.5dd31d962d373p-43, .hi: 0x1.c86f7b7ea5p-1}, |
| 165 | {.lo: -0x1.5dd31d962d373p-43, .hi: 0x1.c86f7b7ea5p-1}, |
| 166 | {.lo: -0x1.9ad57391924a7p-43, .hi: 0x1.cdcebd2374p-1}, |
| 167 | {.lo: -0x1.3167ccc538261p-44, .hi: 0x1.d338120a6ep-1}, |
| 168 | {.lo: -0x1.3167ccc538261p-44, .hi: 0x1.d338120a6ep-1}, |
| 169 | {.lo: 0x1.c7a4ff65ddbc9p-45, .hi: 0x1.d8aba045bp-1}, |
| 170 | {.lo: 0x1.c7a4ff65ddbc9p-45, .hi: 0x1.d8aba045bp-1}, |
| 171 | {.lo: -0x1.f9ab3cf74babap-44, .hi: 0x1.de298ec0bbp-1}, |
| 172 | {.lo: -0x1.f9ab3cf74babap-44, .hi: 0x1.de298ec0bbp-1}, |
| 173 | {.lo: 0x1.52842c1c1e586p-43, .hi: 0x1.e3b20546f5p-1}, |
| 174 | {.lo: 0x1.52842c1c1e586p-43, .hi: 0x1.e3b20546f5p-1}, |
| 175 | {.lo: 0x1.3c6764fc87b4ap-48, .hi: 0x1.e9452c8a71p-1}, |
| 176 | {.lo: 0x1.3c6764fc87b4ap-48, .hi: 0x1.e9452c8a71p-1}, |
| 177 | {.lo: -0x1.a0976c0a2827dp-44, .hi: 0x1.eee32e2aedp-1}, |
| 178 | {.lo: -0x1.a0976c0a2827dp-44, .hi: 0x1.eee32e2aedp-1}, |
| 179 | {.lo: -0x1.a45314dc4fc42p-43, .hi: 0x1.f48c34bd1fp-1}, |
| 180 | {.lo: -0x1.a45314dc4fc42p-43, .hi: 0x1.f48c34bd1fp-1}, |
| 181 | {.lo: 0x1.ef5d00e390ap-44, .hi: 0x1.fa406bd244p-1}, |
| 182 | {.lo: 0.0, .hi: 1.0}, |
| 183 | }; |
| 184 | |
| 185 | // Check if x is an odd integer: the lowest set bit must be at the unit |
| 186 | // position: |
| 187 | // x_e + lsb == UNIT_EXPONENT. |
| 188 | LIBC_INLINE bool is_odd_integer(double x) { |
| 189 | using FPBits = fputil::FPBits<double>; |
| 190 | FPBits xbits(x); |
| 191 | uint64_t x_u = xbits.uintval(); |
| 192 | unsigned x_e = static_cast<unsigned>(xbits.get_biased_exponent()); |
| 193 | unsigned lsb = |
| 194 | static_cast<unsigned>(cpp::countr_zero(value: x_u | FPBits::EXP_MASK)); |
| 195 | constexpr unsigned UNIT_EXPONENT = |
| 196 | static_cast<unsigned>(FPBits::EXP_BIAS + FPBits::FRACTION_LEN); |
| 197 | return (x_e + lsb == UNIT_EXPONENT); |
| 198 | } |
| 199 | |
| 200 | // Check if x is an integer: the lowest set bit must be at or above the unit |
| 201 | // position: |
| 202 | // x_e + lsb >= UNIT_EXPONENT. |
| 203 | LIBC_INLINE bool is_integer(double x) { |
| 204 | if (x == 0.0) |
| 205 | return true; |
| 206 | using FPBits = fputil::FPBits<double>; |
| 207 | FPBits xbits(x); |
| 208 | uint64_t x_u = xbits.uintval(); |
| 209 | unsigned x_e = static_cast<unsigned>(xbits.get_biased_exponent()); |
| 210 | unsigned lsb = |
| 211 | static_cast<unsigned>(cpp::countr_zero(value: x_u | FPBits::EXP_MASK)); |
| 212 | constexpr unsigned UNIT_EXPONENT = |
| 213 | static_cast<unsigned>(FPBits::EXP_BIAS + FPBits::FRACTION_LEN); |
| 214 | return (x_e + lsb >= UNIT_EXPONENT); |
| 215 | } |
| 216 | |
| 217 | LIBC_INLINE double set_overflow(bool is_neg) { |
| 218 | fputil::set_errno_if_required(ERANGE); |
| 219 | fputil::raise_overflow_except_if_required<double>(); |
| 220 | using FPBits = fputil::FPBits<double>; |
| 221 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 222 | int rounding = fputil::quick_get_round(); |
| 223 | if (rounding == FE_TOWARDZERO) |
| 224 | return is_neg ? -FPBits::max_normal().get_val() |
| 225 | : FPBits::max_normal().get_val(); |
| 226 | if (rounding == FE_DOWNWARD) |
| 227 | return is_neg ? -FPBits::inf().get_val() : FPBits::max_normal().get_val(); |
| 228 | if (rounding == FE_UPWARD) |
| 229 | return is_neg ? -FPBits::max_normal().get_val() : FPBits::inf().get_val(); |
| 230 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 231 | return is_neg ? -FPBits::inf().get_val() : FPBits::inf().get_val(); |
| 232 | } |
| 233 | |
| 234 | LIBC_INLINE double set_underflow(bool is_neg) { |
| 235 | fputil::set_errno_if_required(ERANGE); |
| 236 | fputil::raise_underflow_except_if_required<double>(); |
| 237 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 238 | using FPBits = fputil::FPBits<double>; |
| 239 | int rounding = fputil::quick_get_round(); |
| 240 | if (rounding == FE_UPWARD && !is_neg) |
| 241 | return FPBits::min_subnormal().get_val(); |
| 242 | if (rounding == FE_DOWNWARD && is_neg) |
| 243 | return -FPBits::min_subnormal().get_val(); |
| 244 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 245 | return is_neg ? -0.0 : 0.0; |
| 246 | } |
| 247 | |
| 248 | // Rounding tests for 2^hi * (mid + lo) when the output might be denormal. We |
| 249 | // assume further that |
| 250 | // 1 <= mid < 2, |
| 251 | // mid + lo < 2, |
| 252 | // and |lo| << mid. |
| 253 | // |
| 254 | // Notice that, if 0 < x < 2^-1022: |
| 255 | // double( 2^-1022 + x ) - 2^-1022 = double(x), |
| 256 | // and if -2^-1022 < x < 0: |
| 257 | // double(-2^-1022 + x ) - -2^-1022 = double(x). |
| 258 | // |
| 259 | // So if we scale |x| up by 2^1022, we can use the "+1 trick" (or "-1 trick" |
| 260 | // when x < 0): |
| 261 | // double( 1.0 + 2^1022 * x) - 1.0, for x > 0, and |
| 262 | // double(-1.0 + 2^1022 * x) - -1.0, for x < 0, |
| 263 | // to test how x is rounded in the denormal range. By adding +-1.0, the |
| 264 | // results will have similar rounding points as denormal outputs, aligning the |
| 265 | // least significant bit of the mantissa with 2^-52 * 2^-1022 = 2^-1074. |
| 266 | // |
| 267 | // Finally, subtracting 1023 in the exponent field converts the rounded normal |
| 268 | // value into the subnormal value. |
| 269 | template <bool SKIP_ZIV_TEST = false> |
| 270 | LIBC_INLINE constexpr cpp::optional<double> |
| 271 | ziv_test_denorm(int hi, double mid, double lo, double err, bool is_neg) { |
| 272 | using FPBits = typename fputil::FPBits<double>; |
| 273 | |
| 274 | uint64_t scale_u = static_cast<uint64_t>(hi + 2045) << FPBits::FRACTION_LEN; |
| 275 | double exp_scale = cpp::bit_cast<double>(from: scale_u); |
| 276 | |
| 277 | double mid_hi = mid * exp_scale; |
| 278 | double lo_scaled = lo * exp_scale; |
| 279 | |
| 280 | double = 0.0; |
| 281 | uint64_t scale_down = 0x3FE0'0000'0000'0000ULL; // 1022 in the exponent field. |
| 282 | |
| 283 | // Result is denormal if |mid_hi + lo_scaled| < 1.0. |
| 284 | // In that case, add +-1.0 to mimic denormal rounding points. |
| 285 | if (is_neg) { |
| 286 | if ((-1.0 - mid_hi) < lo_scaled) { |
| 287 | // Extra rounding step is needed, which adds more rounding errors. |
| 288 | extra_factor = -1.0; |
| 289 | scale_down = 0x3FF0'0000'0000'0000ULL; // 1023 in the exponent field. |
| 290 | err += 0x1.0p-52; |
| 291 | } |
| 292 | } else { |
| 293 | if ((1.0 - mid_hi) > lo_scaled) { |
| 294 | // Extra rounding step is needed, which adds more rounding errors. |
| 295 | extra_factor = 1.0; |
| 296 | scale_down = 0x3FF0'0000'0000'0000ULL; // 1023 in the exponent field. |
| 297 | err += 0x1.0p-52; |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | // By adding +-1.0, the hardware adder performs rounding at the exact same |
| 302 | // bit position as denormal rounding. |
| 303 | if constexpr (SKIP_ZIV_TEST) { |
| 304 | double r = extra_factor + (mid_hi + lo_scaled); |
| 305 | return cpp::bit_cast<double>(from: cpp::bit_cast<uint64_t>(from: r) - scale_down); |
| 306 | } else { |
| 307 | double err_scaled = err * exp_scale; |
| 308 | |
| 309 | double lo_u = lo_scaled + err_scaled; |
| 310 | double lo_l = lo_scaled - err_scaled; |
| 311 | |
| 312 | double upper = extra_factor + (mid_hi + lo_u); |
| 313 | double lower = extra_factor + (mid_hi + lo_l); |
| 314 | |
| 315 | if (LIBC_LIKELY(upper == lower)) { |
| 316 | return cpp::bit_cast<double>(from: cpp::bit_cast<uint64_t>(from: upper) - scale_down); |
| 317 | } |
| 318 | |
| 319 | return cpp::nullopt; |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | // Upper bound for y = 1075 / |log2(1 - 2^(-53)|, generated by Sollya: |
| 324 | // > y = round(-1075 / log2(1 - 2^(-53)), D, RU); |
| 325 | // > y; |
| 326 | // 0x1.74910d52d3052p62 |
| 327 | // > printdouble(y); |
| 328 | // 0x43d74910d52d3052 |
| 329 | constexpr uint64_t Y_UPPER_BOUND = 0x43d7'4910'd52d'3052; |
| 330 | // Lower bound for y = 2^(-54) / 1074), generated by Sollya: |
| 331 | // > y = round(2^(-54) / 1074), D, RD); |
| 332 | // > y; |
| 333 | // 0x1.e829f39aef509p-65 |
| 334 | // > printdouble(y); |
| 335 | // 0x3bee829f39ae'f509 |
| 336 | constexpr uint64_t Y_LOWER_BOUND = 0x3bee'829f'39ae'f509; |
| 337 | |
| 338 | // Fast checks for special inputs: |
| 339 | // x = 0, +-1, 2^k, 10, +- inf |
| 340 | // y = 0, +-1, 2, 0.5 |
| 341 | LIBC_ALWAYS_INLINE cpp::optional<double> check_special_inputs(double x, |
| 342 | double y) { |
| 343 | using FPBits = fputil::FPBits<double>; |
| 344 | FPBits xbits(x), ybits(y); |
| 345 | |
| 346 | bool x_sign = xbits.sign() == Sign::NEG; |
| 347 | bool y_sign = ybits.sign() == Sign::NEG; |
| 348 | |
| 349 | FPBits x_abs = xbits.abs(); |
| 350 | FPBits y_abs = ybits.abs(); |
| 351 | |
| 352 | uint64_t x_u = xbits.uintval(); |
| 353 | uint64_t y_u = ybits.uintval(); |
| 354 | uint64_t y_a = y_abs.uintval(); |
| 355 | |
| 356 | if (LIBC_UNLIKELY((x_u & 0x0003'FFFF'FFFF'FFFF) == 0) || |
| 357 | ((y_u & 0x000F'FFFF'FFFF'FFFF) == 0)) { |
| 358 | // If x or y is signaling NaN |
| 359 | if (x_abs.is_signaling_nan() || y_abs.is_signaling_nan()) { |
| 360 | fputil::raise_except_if_required(FE_INVALID); |
| 361 | return FPBits::quiet_nan().get_val(); |
| 362 | } |
| 363 | |
| 364 | if (x == 1.0 || y == 0.0) |
| 365 | return 1.0; |
| 366 | |
| 367 | if (x == 0.0) { |
| 368 | if (y_abs.is_nan()) |
| 369 | return y; |
| 370 | if (y_abs.is_inf()) |
| 371 | return y_sign ? FPBits::inf().get_val() : 0.0; |
| 372 | bool out_is_neg = x_sign && is_odd_integer(x: y); |
| 373 | if (y_sign) { |
| 374 | // pow(0, negative number) = inf |
| 375 | fputil::set_errno_if_required(EDOM); |
| 376 | fputil::raise_except_if_required(FE_DIVBYZERO); |
| 377 | return FPBits::inf(sign: out_is_neg ? Sign::NEG : Sign::POS).get_val(); |
| 378 | } |
| 379 | // pow(0, positive number) = 0 |
| 380 | return out_is_neg ? -0.0 : 0.0; |
| 381 | } |
| 382 | |
| 383 | if (y == 1.0) |
| 384 | return x; |
| 385 | |
| 386 | if (y == 2.0) |
| 387 | return x * x; |
| 388 | |
| 389 | if (y == 0.5 && !x_sign) |
| 390 | return fputil::sqrt<double>(x); |
| 391 | |
| 392 | // TODO: Add special case y = -0.5 when rsqrt is available. |
| 393 | |
| 394 | if (x == 2.0) |
| 395 | return math::exp2(x: y); |
| 396 | |
| 397 | if (x == 10.0) |
| 398 | return math::exp10(x: y); |
| 399 | |
| 400 | // For x = 2^(+- 2^n): |
| 401 | // x^y = 2^(+- 2^n * y). |
| 402 | if (!x_sign && x_abs.is_normal() && xbits.get_mantissa() == 0 && |
| 403 | y_a > Y_LOWER_BOUND && y_a < Y_UPPER_BOUND) { |
| 404 | int e_x = xbits.get_exponent(); |
| 405 | uint32_t abs_e = static_cast<uint32_t>(e_x > 0 ? e_x : -e_x); |
| 406 | if (cpp::has_single_bit(value: abs_e)) { |
| 407 | double hi = static_cast<double>(e_x) * y; |
| 408 | if (hi >= 1024.0) |
| 409 | return set_overflow(false); |
| 410 | if (hi < -1075.0) |
| 411 | return set_underflow(false); |
| 412 | if (hi >= -1022.0) |
| 413 | return math::exp2(x: hi); |
| 414 | } |
| 415 | } |
| 416 | } |
| 417 | |
| 418 | return cpp::nullopt; |
| 419 | } |
| 420 | |
| 421 | // Filters out extreme input ranges, infinities, NaNs, normalizes denormal |
| 422 | // inputs, and handles negative bases. Returns a value if the result is |
| 423 | // determined, or nullopt if regular evaluation should proceed (in which case x, |
| 424 | // y, e_x, x_mant, is_neg, sign_d may be updated). |
| 425 | LIBC_ALWAYS_INLINE cpp::optional<double> |
| 426 | check_exceptional_cases(double &x, double &y, double &e_x, uint64_t &x_mant, |
| 427 | bool &is_neg, double &sign_d) { |
| 428 | using FPBits = fputil::FPBits<double>; |
| 429 | |
| 430 | FPBits xbits(x), ybits(y); |
| 431 | bool x_sign = xbits.sign() == Sign::NEG; |
| 432 | bool y_sign = ybits.sign() == Sign::NEG; |
| 433 | |
| 434 | FPBits x_abs = xbits.abs(); |
| 435 | FPBits y_abs = ybits.abs(); |
| 436 | |
| 437 | uint64_t x_u = xbits.uintval(); |
| 438 | uint64_t y_mant = ybits.get_mantissa(); |
| 439 | uint64_t x_a = x_abs.uintval(); |
| 440 | uint64_t y_a = y_abs.uintval(); |
| 441 | |
| 442 | if (LIBC_UNLIKELY(y_a <= Y_LOWER_BOUND || y_a >= Y_UPPER_BOUND || |
| 443 | x_u >= FPBits::inf().uintval() || |
| 444 | x_u < FPBits::min_normal().uintval())) { |
| 445 | // If x or y is signaling NaN |
| 446 | if (x_abs.is_signaling_nan() || y_abs.is_signaling_nan()) { |
| 447 | fputil::raise_except_if_required(FE_INVALID); |
| 448 | return FPBits::quiet_nan().get_val(); |
| 449 | } |
| 450 | |
| 451 | // |y| > |1075 / log2(1 - 2^-53)|. |
| 452 | if (y_a > 0x43d7'4910'd52d'3052) { |
| 453 | if (y_a >= 0x7ff0'0000'0000'0000) { |
| 454 | // y is inf or nan |
| 455 | if (y_mant != 0) { |
| 456 | // y is NaN |
| 457 | // pow(1, NaN) = 1 is already dealt with. |
| 458 | // pow(x, NaN) = NaN |
| 459 | return y; |
| 460 | } |
| 461 | |
| 462 | // Now y is +-Inf |
| 463 | if (x_abs.is_nan()) { |
| 464 | // pow(NaN, +-Inf) = NaN |
| 465 | return x; |
| 466 | } |
| 467 | |
| 468 | if (x_a == 0x3ff0'0000'0000'0000) { |
| 469 | // pow(+-1, +-Inf) = 1.0 |
| 470 | return 1.0; |
| 471 | } |
| 472 | |
| 473 | // pow (|x| < 1, -inf) = +inf |
| 474 | // pow (|x| < 1, +inf) = 0.0 |
| 475 | // pow (|x| > 1, -inf) = 0.0 |
| 476 | // pow (|x| > 1, +inf) = +inf |
| 477 | return ((x_a < FPBits::one().uintval()) == y_sign) |
| 478 | ? FPBits::inf().get_val() |
| 479 | : 0.0; |
| 480 | } |
| 481 | // x^y will overflow / underflow in double precision. Set y to a |
| 482 | // large enough exponent but not too large, so that the computations |
| 483 | // won't overflow in double precision. |
| 484 | y = y_sign ? -0x1.0p100 : 0x1.0p100; |
| 485 | } |
| 486 | |
| 487 | // y is finite and non-zero. |
| 488 | |
| 489 | if (x_a == FPBits::inf().uintval()) { |
| 490 | bool out_is_neg = x_sign && is_odd_integer(x: y); |
| 491 | Sign out_sign = out_is_neg ? Sign::NEG : Sign::POS; |
| 492 | return y_sign ? FPBits::zero(sign: out_sign).get_val() |
| 493 | : FPBits::inf(sign: out_sign).get_val(); |
| 494 | } |
| 495 | |
| 496 | if (x_a > FPBits::inf().uintval()) { |
| 497 | // x is NaN. |
| 498 | // pow (aNaN, 0) is already taken care above. |
| 499 | return x; |
| 500 | } |
| 501 | |
| 502 | // Normalize denormal inputs. |
| 503 | if (x_a < FPBits::min_normal().uintval()) { |
| 504 | FPBits x_norm(x * 0x1.0p64); |
| 505 | e_x = static_cast<double>(x_norm.get_exponent()) - 64.0; |
| 506 | x_mant = x_norm.get_mantissa(); |
| 507 | } |
| 508 | |
| 509 | // x is finite and negative, and y is a finite integer. |
| 510 | if (x_sign) { |
| 511 | if (is_integer(x: y)) { |
| 512 | x = -x; |
| 513 | if (is_odd_integer(x: y)) { |
| 514 | is_neg = true; |
| 515 | sign_d = -1.0; |
| 516 | } |
| 517 | } else { |
| 518 | // pow( negative, non-integer ) = NaN |
| 519 | fputil::set_errno_if_required(EDOM); |
| 520 | fputil::raise_except_if_required(FE_INVALID); |
| 521 | return FPBits::quiet_nan().get_val(); |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | if (y_a <= Y_LOWER_BOUND) { |
| 526 | volatile double one = 1.0; |
| 527 | volatile double eps = ((x_a < FPBits::one().uintval()) == !y_sign) |
| 528 | ? -0x1.0p-100 |
| 529 | : 0x1.0p-100; |
| 530 | return one + eps; |
| 531 | } |
| 532 | } |
| 533 | |
| 534 | return cpp::nullopt; |
| 535 | } |
| 536 | |
| 537 | // Check if x^y is an exact rounding boundary case (a 54-bit dyadic float, |
| 538 | // which is either an exact 53-bit float or an exact halfway midpoint). |
| 539 | // |
| 540 | // Reference: |
| 541 | // Lauter, C. and Lefevre, V., "Rounding Boundary Cases for Values of the |
| 542 | // Exponential and Power Functions," IEEE Trans. Comput. 58(8):1063-1074. |
| 543 | // |
| 544 | // For a normalized 128-bit dyadic float with MSB at bit 127, the top 54 bits |
| 545 | // occupy bits [127:74]. For any 54-bit number, bits [73:0] are zero. |
| 546 | // |
| 547 | // If x^y is not in F_54: |
| 548 | // |x^y - o_54(x^y)| / x^y >= 2^-114. |
| 549 | // With an evaluation error < 2^-115, the distance from bits [73:0] to the |
| 550 | // nearest 54-bit boundary must satisfy: |
| 551 | // dist >= (2^-114 - 2^-115) * 2^127 = 2^12. |
| 552 | // We use a threshold of 2^-114 (dist < 2^13) to capture all boundary |
| 553 | // candidates: |
| 554 | // |
| 555 | // 1. x = 2^e: x^y = 2^(e * y) is in F_54 iff e * y is an integer. |
| 556 | // 2. x = m * 2^e, y = n * 2^f with m, n odd integers: |
| 557 | // x^y in F_54 implies: |
| 558 | // - 0 <= y <= 35, n <= 35, and f >= -5. |
| 559 | // - e * y is an integer. |
| 560 | // - When f < 0 (i.e. y = n / 2^(-f)): m^(2^f) is an integer, verified |
| 561 | // by taking repeated square roots -f times. |
| 562 | // - The resulting significand (m^(2^f))^n <= 2^54. |
| 563 | // |
| 564 | // Returns true and sets exact_m and exact_exp if x^y is an exact boundary. |
| 565 | LIBC_INLINE bool is_exact_rounding_boundary(double x, double y, |
| 566 | uint64_t &exact_m, int &exact_exp) { |
| 567 | using FPBits = fputil::FPBits<double>; |
| 568 | |
| 569 | FPBits xbits(x); |
| 570 | int x_e = 0; |
| 571 | uint64_t x_mant = xbits.get_mantissa(); |
| 572 | if (LIBC_UNLIKELY(xbits.get_biased_exponent() == 0)) { |
| 573 | if (x_mant != 0) { |
| 574 | int shift = cpp::countl_zero(value: x_mant) - 11; |
| 575 | x_mant = (x_mant << shift) & FPBits::FRACTION_MASK; |
| 576 | x_e = -1022 - shift; |
| 577 | } |
| 578 | } else { |
| 579 | x_e = xbits.get_exponent(); |
| 580 | } |
| 581 | |
| 582 | // Case 1: x = 2^e. |
| 583 | if (x_mant == 0) { |
| 584 | double e = static_cast<double>(x_e); |
| 585 | double ey = e * y; |
| 586 | if (is_integer(x: ey)) { |
| 587 | exact_m = 1; |
| 588 | exact_exp = static_cast<int>(ey); |
| 589 | return true; |
| 590 | } |
| 591 | return false; |
| 592 | } |
| 593 | |
| 594 | // Case 2: x is not a power of 2. |
| 595 | if (y < 0.0 || y > 35.0) |
| 596 | return false; |
| 597 | |
| 598 | // Decompose y = n * 2^f with n odd. |
| 599 | FPBits ybits(y); |
| 600 | uint64_t y_mant = ybits.get_mantissa() | (1ULL << FPBits::FRACTION_LEN); |
| 601 | int y_exp = ybits.get_exponent() - static_cast<int>(FPBits::FRACTION_LEN); |
| 602 | int tz_y = cpp::countr_zero(value: y_mant); |
| 603 | uint64_t n = y_mant >> tz_y; |
| 604 | int f = y_exp + tz_y; |
| 605 | |
| 606 | if (n > 35 || f < -5) |
| 607 | return false; |
| 608 | |
| 609 | // Decompose x = m * 2^e with m odd. |
| 610 | uint64_t full_x_mant = x_mant | (1ULL << FPBits::FRACTION_LEN); |
| 611 | int tz_x = cpp::countr_zero(value: full_x_mant); |
| 612 | uint64_t m = full_x_mant >> tz_x; |
| 613 | int e = x_e - static_cast<int>(FPBits::FRACTION_LEN) + tz_x; |
| 614 | |
| 615 | if (f < 0) { |
| 616 | // Non-integer exponent: y = n / 2^(-f) with -f in {1, ..., 5}. |
| 617 | double ey = static_cast<double>(e) * y; |
| 618 | if (!is_integer(x: ey)) |
| 619 | return false; |
| 620 | |
| 621 | // Check if m^(2^f) is an integer by taking repeated square roots -f times. |
| 622 | int count = -f; |
| 623 | uint64_t cur = m; |
| 624 | for (int i = 0; i < count; ++i) { |
| 625 | uint64_t s = |
| 626 | static_cast<uint64_t>(fputil::sqrt<double>(x: static_cast<double>(cur))); |
| 627 | if (s * s != cur) |
| 628 | return false; |
| 629 | cur = s; |
| 630 | } |
| 631 | |
| 632 | // Compute res = cur^n and ensure res <= 2^54. |
| 633 | uint64_t res = 1; |
| 634 | for (uint64_t i = 0; i < n; ++i) { |
| 635 | if (res > (1ULL << 54) / cur) |
| 636 | return false; |
| 637 | res *= cur; |
| 638 | } |
| 639 | exact_m = res; |
| 640 | exact_exp = static_cast<int>(ey); |
| 641 | return true; |
| 642 | } else { |
| 643 | // Integer exponent: f >= 0, so y is an integer. |
| 644 | int y_int = static_cast<int>(y); |
| 645 | uint64_t res = 1; |
| 646 | for (int i = 0; i < y_int; ++i) { |
| 647 | if (res > (1ULL << 54) / m) |
| 648 | return false; |
| 649 | res *= m; |
| 650 | } |
| 651 | exact_m = res; |
| 652 | exact_exp = e * y_int; |
| 653 | return true; |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | } // namespace pow_internal |
| 658 | } // namespace math |
| 659 | } // namespace LIBC_NAMESPACE_DECL |
| 660 | |
| 661 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_POW_UTILS_H |
| 662 | |