| 1 | /* |
| 2 | * Copyright 2016-present Facebook, Inc. |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | /** |
| 18 | * Some arithmetic functions that seem to pop up or get hand-rolled a lot. |
| 19 | * So far they are all focused on integer division. |
| 20 | */ |
| 21 | |
| 22 | #pragma once |
| 23 | |
| 24 | #include <stdint.h> |
| 25 | |
| 26 | #include <limits> |
| 27 | #include <type_traits> |
| 28 | |
| 29 | namespace folly { |
| 30 | |
| 31 | namespace detail { |
| 32 | |
| 33 | template <typename T> |
| 34 | inline constexpr T divFloorBranchless(T num, T denom) { |
| 35 | // floor != trunc when the answer isn't exact and truncation went the |
| 36 | // wrong way (truncation went toward positive infinity). That happens |
| 37 | // when the true answer is negative, which happens when num and denom |
| 38 | // have different signs. The following code compiles branch-free on |
| 39 | // many platforms. |
| 40 | return (num / denom) + |
| 41 | ((num % denom) != 0 ? 1 : 0) * |
| 42 | (std::is_signed<T>::value && (num ^ denom) < 0 ? -1 : 0); |
| 43 | } |
| 44 | |
| 45 | template <typename T> |
| 46 | inline constexpr T divFloorBranchful(T num, T denom) { |
| 47 | // First case handles negative result by preconditioning numerator. |
| 48 | // Preconditioning decreases the magnitude of the numerator, which is |
| 49 | // itself sign-dependent. Second case handles zero or positive rational |
| 50 | // result, where trunc and floor are the same. |
| 51 | return std::is_signed<T>::value && (num ^ denom) < 0 && num != 0 |
| 52 | ? (num + (num > 0 ? -1 : 1)) / denom - 1 |
| 53 | : num / denom; |
| 54 | } |
| 55 | |
| 56 | template <typename T> |
| 57 | inline constexpr T divCeilBranchless(T num, T denom) { |
| 58 | // ceil != trunc when the answer isn't exact (truncation occurred) |
| 59 | // and truncation went away from positive infinity. That happens when |
| 60 | // the true answer is positive, which happens when num and denom have |
| 61 | // the same sign. |
| 62 | return (num / denom) + |
| 63 | ((num % denom) != 0 ? 1 : 0) * |
| 64 | (std::is_signed<T>::value && (num ^ denom) < 0 ? 0 : 1); |
| 65 | } |
| 66 | |
| 67 | template <typename T> |
| 68 | inline constexpr T divCeilBranchful(T num, T denom) { |
| 69 | // First case handles negative or zero rational result, where trunc and ceil |
| 70 | // are the same. |
| 71 | // Second case handles positive result by preconditioning numerator. |
| 72 | // Preconditioning decreases the magnitude of the numerator, which is |
| 73 | // itself sign-dependent. |
| 74 | return (std::is_signed<T>::value && (num ^ denom) < 0) || num == 0 |
| 75 | ? num / denom |
| 76 | : (num + (num > 0 ? -1 : 1)) / denom + 1; |
| 77 | } |
| 78 | |
| 79 | template <typename T> |
| 80 | inline constexpr T divRoundAwayBranchless(T num, T denom) { |
| 81 | // away != trunc whenever truncation actually occurred, which is when |
| 82 | // there is a non-zero remainder. If the unrounded result is negative |
| 83 | // then fixup moves it toward negative infinity. If the unrounded |
| 84 | // result is positive then adjustment makes it larger. |
| 85 | return (num / denom) + |
| 86 | ((num % denom) != 0 ? 1 : 0) * |
| 87 | (std::is_signed<T>::value && (num ^ denom) < 0 ? -1 : 1); |
| 88 | } |
| 89 | |
| 90 | template <typename T> |
| 91 | inline constexpr T divRoundAwayBranchful(T num, T denom) { |
| 92 | // First case of second ternary operator handles negative rational |
| 93 | // result, which is the same as divFloor. Second case of second ternary |
| 94 | // operator handles positive result, which is the same as divCeil. |
| 95 | // Zero case is separated for simplicity. |
| 96 | return num == 0 ? 0 |
| 97 | : (num + (num > 0 ? -1 : 1)) / denom + |
| 98 | (std::is_signed<T>::value && (num ^ denom) < 0 ? -1 : 1); |
| 99 | } |
| 100 | |
| 101 | template <typename N, typename D> |
| 102 | using IdivResultType = typename std::enable_if< |
| 103 | std::is_integral<N>::value && std::is_integral<D>::value && |
| 104 | !std::is_same<N, bool>::value && !std::is_same<D, bool>::value, |
| 105 | decltype(N{1} / D{1})>::type; |
| 106 | } // namespace detail |
| 107 | |
| 108 | #if defined(__arm__) && !FOLLY_AARCH64 |
| 109 | constexpr auto kIntegerDivisionGivesRemainder = false; |
| 110 | #else |
| 111 | constexpr auto kIntegerDivisionGivesRemainder = true; |
| 112 | #endif |
| 113 | |
| 114 | /** |
| 115 | * Returns num/denom, rounded toward negative infinity. Put another way, |
| 116 | * returns the largest integral value that is less than or equal to the |
| 117 | * exact (not rounded) fraction num/denom. |
| 118 | * |
| 119 | * The matching remainder (num - divFloor(num, denom) * denom) can be |
| 120 | * negative only if denom is negative, unlike in truncating division. |
| 121 | * Note that for unsigned types this is the same as the normal integer |
| 122 | * division operator. divFloor is equivalent to python's integral division |
| 123 | * operator //. |
| 124 | * |
| 125 | * This function undergoes the same integer promotion rules as a |
| 126 | * built-in operator, except that we don't allow bool -> int promotion. |
| 127 | * This function is undefined if denom == 0. It is also undefined if the |
| 128 | * result type T is a signed type, num is std::numeric_limits<T>::min(), |
| 129 | * and denom is equal to -1 after conversion to the result type. |
| 130 | */ |
| 131 | template <typename N, typename D> |
| 132 | inline constexpr detail::IdivResultType<N, D> divFloor(N num, D denom) { |
| 133 | using R = decltype(num / denom); |
| 134 | return detail::IdivResultType<N, D>( |
| 135 | kIntegerDivisionGivesRemainder && std::is_signed<R>::value |
| 136 | ? detail::divFloorBranchless<R>(num, denom) |
| 137 | : detail::divFloorBranchful<R>(num, denom)); |
| 138 | } |
| 139 | |
| 140 | /** |
| 141 | * Returns num/denom, rounded toward positive infinity. Put another way, |
| 142 | * returns the smallest integral value that is greater than or equal to |
| 143 | * the exact (not rounded) fraction num/denom. |
| 144 | * |
| 145 | * This function undergoes the same integer promotion rules as a |
| 146 | * built-in operator, except that we don't allow bool -> int promotion. |
| 147 | * This function is undefined if denom == 0. It is also undefined if the |
| 148 | * result type T is a signed type, num is std::numeric_limits<T>::min(), |
| 149 | * and denom is equal to -1 after conversion to the result type. |
| 150 | */ |
| 151 | template <typename N, typename D> |
| 152 | inline constexpr detail::IdivResultType<N, D> divCeil(N num, D denom) { |
| 153 | using R = decltype(num / denom); |
| 154 | return detail::IdivResultType<N, D>( |
| 155 | kIntegerDivisionGivesRemainder && std::is_signed<R>::value |
| 156 | ? detail::divCeilBranchless<R>(num, denom) |
| 157 | : detail::divCeilBranchful<R>(num, denom)); |
| 158 | } |
| 159 | |
| 160 | /** |
| 161 | * Returns num/denom, rounded toward zero. If num and denom are non-zero |
| 162 | * and have different signs (so the unrounded fraction num/denom is |
| 163 | * negative), returns divCeil, otherwise returns divFloor. If T is an |
| 164 | * unsigned type then this is always equal to divFloor. |
| 165 | * |
| 166 | * Note that this is the same as the normal integer division operator, |
| 167 | * at least since C99 (before then the rounding for negative results was |
| 168 | * implementation defined). This function is here for completeness and |
| 169 | * as a place to hang this comment. |
| 170 | * |
| 171 | * This function undergoes the same integer promotion rules as a |
| 172 | * built-in operator, except that we don't allow bool -> int promotion. |
| 173 | * This function is undefined if denom == 0. It is also undefined if the |
| 174 | * result type T is a signed type, num is std::numeric_limits<T>::min(), |
| 175 | * and denom is equal to -1 after conversion to the result type. |
| 176 | */ |
| 177 | template <typename N, typename D> |
| 178 | inline constexpr detail::IdivResultType<N, D> divTrunc(N num, D denom) { |
| 179 | return detail::IdivResultType<N, D>(num / denom); |
| 180 | } |
| 181 | |
| 182 | /** |
| 183 | * Returns num/denom, rounded away from zero. If num and denom are |
| 184 | * non-zero and have different signs (so the unrounded fraction num/denom |
| 185 | * is negative), returns divFloor, otherwise returns divCeil. If T is |
| 186 | * an unsigned type then this is always equal to divCeil. |
| 187 | * |
| 188 | * This function undergoes the same integer promotion rules as a |
| 189 | * built-in operator, except that we don't allow bool -> int promotion. |
| 190 | * This function is undefined if denom == 0. It is also undefined if the |
| 191 | * result type T is a signed type, num is std::numeric_limits<T>::min(), |
| 192 | * and denom is equal to -1 after conversion to the result type. |
| 193 | */ |
| 194 | template <typename N, typename D> |
| 195 | inline constexpr detail::IdivResultType<N, D> divRoundAway(N num, D denom) { |
| 196 | using R = decltype(num / denom); |
| 197 | return detail::IdivResultType<N, D>( |
| 198 | kIntegerDivisionGivesRemainder && std::is_signed<R>::value |
| 199 | ? detail::divRoundAwayBranchless<R>(num, denom) |
| 200 | : detail::divRoundAwayBranchful<R>(num, denom)); |
| 201 | } |
| 202 | |
| 203 | } // namespace folly |
| 204 | |