| 1 | //===-- Add and subtract IEEE 754 floating-point numbers --------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_ADD_SUB_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_ADD_SUB_H |
| 11 | |
| 12 | #include "hdr/fenv_macros.h" |
| 13 | #include "src/__support/CPP/algorithm.h" |
| 14 | #include "src/__support/CPP/bit.h" |
| 15 | #include "src/__support/CPP/type_traits.h" |
| 16 | #include "src/__support/FPUtil/BasicOperations.h" |
| 17 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 18 | #include "src/__support/FPUtil/FPBits.h" |
| 19 | #include "src/__support/FPUtil/cast.h" |
| 20 | #include "src/__support/FPUtil/dyadic_float.h" |
| 21 | #include "src/__support/FPUtil/rounding_mode.h" |
| 22 | #include "src/__support/macros/attributes.h" |
| 23 | #include "src/__support/macros/config.h" |
| 24 | #include "src/__support/macros/optimization.h" |
| 25 | |
| 26 | namespace LIBC_NAMESPACE_DECL { |
| 27 | namespace fputil::generic { |
| 28 | |
| 29 | template <bool IsSub, typename OutType, typename InType> |
| 30 | LIBC_INLINE constexpr cpp::enable_if_t<cpp::is_floating_point_v<OutType> && |
| 31 | cpp::is_floating_point_v<InType> && |
| 32 | sizeof(OutType) <= sizeof(InType), |
| 33 | OutType> |
| 34 | add_or_sub(InType x, InType y) { |
| 35 | using OutFPBits = FPBits<OutType>; |
| 36 | using OutStorageType = typename OutFPBits::StorageType; |
| 37 | using InFPBits = FPBits<InType>; |
| 38 | using InStorageType = typename InFPBits::StorageType; |
| 39 | |
| 40 | constexpr int GUARD_BITS_LEN = 3; |
| 41 | constexpr int RESULT_FRACTION_LEN = InFPBits::FRACTION_LEN + GUARD_BITS_LEN; |
| 42 | constexpr int RESULT_MANTISSA_LEN = RESULT_FRACTION_LEN + 1; |
| 43 | |
| 44 | using DyadicFloat = |
| 45 | DyadicFloat<cpp::bit_ceil(value: static_cast<size_t>(RESULT_MANTISSA_LEN))>; |
| 46 | |
| 47 | InFPBits x_bits(x); |
| 48 | InFPBits y_bits(y); |
| 49 | |
| 50 | bool is_effectively_add = (x_bits.sign() == y_bits.sign()) != IsSub; |
| 51 | |
| 52 | if (LIBC_UNLIKELY(x_bits.is_inf_or_nan() || y_bits.is_inf_or_nan() || |
| 53 | x_bits.is_zero() || y_bits.is_zero())) { |
| 54 | if (x_bits.is_nan() || y_bits.is_nan()) { |
| 55 | if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) |
| 56 | raise_except_if_required(FE_INVALID); |
| 57 | |
| 58 | if (x_bits.is_quiet_nan()) { |
| 59 | InStorageType x_payload = x_bits.get_mantissa(); |
| 60 | x_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN; |
| 61 | return OutFPBits::quiet_nan(x_bits.sign(), |
| 62 | static_cast<OutStorageType>(x_payload)) |
| 63 | .get_val(); |
| 64 | } |
| 65 | |
| 66 | if (y_bits.is_quiet_nan()) { |
| 67 | InStorageType y_payload = y_bits.get_mantissa(); |
| 68 | y_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN; |
| 69 | return OutFPBits::quiet_nan(y_bits.sign(), |
| 70 | static_cast<OutStorageType>(y_payload)) |
| 71 | .get_val(); |
| 72 | } |
| 73 | |
| 74 | return OutFPBits::quiet_nan().get_val(); |
| 75 | } |
| 76 | |
| 77 | if (x_bits.is_inf()) { |
| 78 | if (y_bits.is_inf()) { |
| 79 | if (!is_effectively_add) { |
| 80 | raise_except_if_required(FE_INVALID); |
| 81 | return OutFPBits::quiet_nan().get_val(); |
| 82 | } |
| 83 | |
| 84 | return OutFPBits::inf(x_bits.sign()).get_val(); |
| 85 | } |
| 86 | |
| 87 | return OutFPBits::inf(x_bits.sign()).get_val(); |
| 88 | } |
| 89 | |
| 90 | if (y_bits.is_inf()) { |
| 91 | if constexpr (IsSub) |
| 92 | return OutFPBits::inf(y_bits.sign().negate()).get_val(); |
| 93 | else |
| 94 | return OutFPBits::inf(y_bits.sign()).get_val(); |
| 95 | } |
| 96 | |
| 97 | if (x_bits.is_zero()) { |
| 98 | if (y_bits.is_zero()) { |
| 99 | if (is_effectively_add) |
| 100 | return OutFPBits::zero(x_bits.sign()).get_val(); |
| 101 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 102 | return OutFPBits::zero(Sign::POS).get_val(); |
| 103 | #else |
| 104 | switch (fputil::quick_get_round()) { |
| 105 | case FE_DOWNWARD: |
| 106 | return OutFPBits::zero(Sign::NEG).get_val(); |
| 107 | default: |
| 108 | return OutFPBits::zero(Sign::POS).get_val(); |
| 109 | } |
| 110 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 111 | } |
| 112 | |
| 113 | if constexpr (cpp::is_same_v<InType, bfloat16> && |
| 114 | cpp::is_same_v<OutType, bfloat16>) { |
| 115 | OutFPBits out_y_bits(y); |
| 116 | if constexpr (IsSub) |
| 117 | out_y_bits.set_sign(out_y_bits.sign().negate()); |
| 118 | return out_y_bits.get_val(); |
| 119 | } else { |
| 120 | |
| 121 | #ifdef LIBC_USE_CONSTEXPR |
| 122 | InType tmp = y; |
| 123 | #else |
| 124 | // volatile prevents Clang from converting tmp to OutType and then |
| 125 | // immediately back to InType before negating it, resulting in double |
| 126 | // rounding. |
| 127 | volatile InType tmp = y; |
| 128 | #endif // LIBC_USE_CONSTEXPR |
| 129 | if constexpr (IsSub) |
| 130 | tmp = -tmp; |
| 131 | return cast<OutType>(tmp); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | if (y_bits.is_zero()) |
| 136 | return cast<OutType>(x); |
| 137 | } |
| 138 | |
| 139 | InType x_abs = x_bits.abs().get_val(); |
| 140 | InType y_abs = y_bits.abs().get_val(); |
| 141 | |
| 142 | if (x_abs == y_abs && !is_effectively_add) { |
| 143 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 144 | return OutFPBits::zero(Sign::POS).get_val(); |
| 145 | #else |
| 146 | switch (fputil::quick_get_round()) { |
| 147 | case FE_DOWNWARD: |
| 148 | return OutFPBits::zero(Sign::NEG).get_val(); |
| 149 | default: |
| 150 | return OutFPBits::zero(Sign::POS).get_val(); |
| 151 | } |
| 152 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 153 | } |
| 154 | |
| 155 | Sign result_sign = Sign::POS; |
| 156 | |
| 157 | if (x_abs > y_abs) { |
| 158 | result_sign = x_bits.sign(); |
| 159 | } else if (x_abs < y_abs) { |
| 160 | result_sign = y_bits.sign(); |
| 161 | if constexpr (IsSub) |
| 162 | result_sign = result_sign.negate(); |
| 163 | } else if (is_effectively_add) { |
| 164 | result_sign = x_bits.sign(); |
| 165 | } |
| 166 | |
| 167 | InFPBits max_bits(cpp::max(x_abs, y_abs)); |
| 168 | InFPBits min_bits(cpp::min(x_abs, y_abs)); |
| 169 | |
| 170 | InStorageType result_mant{}; |
| 171 | |
| 172 | if (max_bits.is_subnormal()) { |
| 173 | // min_bits must be subnormal too. |
| 174 | |
| 175 | if (is_effectively_add) |
| 176 | result_mant = max_bits.get_mantissa() + min_bits.get_mantissa(); |
| 177 | else |
| 178 | result_mant = max_bits.get_mantissa() - min_bits.get_mantissa(); |
| 179 | |
| 180 | result_mant <<= GUARD_BITS_LEN; |
| 181 | } else { |
| 182 | InStorageType max_mant = static_cast<InStorageType>( |
| 183 | max_bits.get_explicit_mantissa() << GUARD_BITS_LEN); |
| 184 | InStorageType min_mant = static_cast<InStorageType>( |
| 185 | min_bits.get_explicit_mantissa() << GUARD_BITS_LEN); |
| 186 | |
| 187 | int alignment = (max_bits.get_biased_exponent() - max_bits.is_normal()) - |
| 188 | (min_bits.get_biased_exponent() - min_bits.is_normal()); |
| 189 | |
| 190 | InStorageType aligned_min_mant = static_cast<InStorageType>( |
| 191 | min_mant >> cpp::min(a: alignment, b: RESULT_MANTISSA_LEN)); |
| 192 | bool aligned_min_mant_sticky{}; |
| 193 | |
| 194 | if (alignment <= GUARD_BITS_LEN) |
| 195 | aligned_min_mant_sticky = false; |
| 196 | else if (alignment > InFPBits::FRACTION_LEN + GUARD_BITS_LEN) |
| 197 | aligned_min_mant_sticky = true; |
| 198 | else |
| 199 | aligned_min_mant_sticky = |
| 200 | (static_cast<InStorageType>( |
| 201 | min_mant << (InFPBits::STORAGE_LEN - alignment))) != 0; |
| 202 | |
| 203 | InStorageType min_mant_sticky = |
| 204 | static_cast<InStorageType>(static_cast<int>(aligned_min_mant_sticky)); |
| 205 | |
| 206 | if (is_effectively_add) |
| 207 | result_mant = max_mant + (aligned_min_mant | min_mant_sticky); |
| 208 | else |
| 209 | result_mant = max_mant - (aligned_min_mant | min_mant_sticky); |
| 210 | } |
| 211 | |
| 212 | int result_exp = max_bits.get_explicit_exponent() - RESULT_FRACTION_LEN; |
| 213 | DyadicFloat result(result_sign, result_exp, result_mant); |
| 214 | return result.template as<OutType, /*ShouldSignalExceptions=*/true>(); |
| 215 | } |
| 216 | |
| 217 | template <typename OutType, typename InType> |
| 218 | LIBC_INLINE constexpr cpp::enable_if_t<cpp::is_floating_point_v<OutType> && |
| 219 | cpp::is_floating_point_v<InType> && |
| 220 | sizeof(OutType) <= sizeof(InType), |
| 221 | OutType> |
| 222 | add(InType x, InType y) { |
| 223 | return add_or_sub</*IsSub=*/false, OutType>(x, y); |
| 224 | } |
| 225 | |
| 226 | template <typename OutType, typename InType> |
| 227 | LIBC_INLINE constexpr cpp::enable_if_t<cpp::is_floating_point_v<OutType> && |
| 228 | cpp::is_floating_point_v<InType> && |
| 229 | sizeof(OutType) <= sizeof(InType), |
| 230 | OutType> |
| 231 | sub(InType x, InType y) { |
| 232 | return add_or_sub</*IsSub=*/true, OutType>(x, y); |
| 233 | } |
| 234 | |
| 235 | } // namespace fputil::generic |
| 236 | } // namespace LIBC_NAMESPACE_DECL |
| 237 | |
| 238 | #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_ADD_SUB_H |
| 239 | |