| 1 | #pragma once |
| 2 | |
| 3 | #ifdef NNSDK |
| 4 | #include <atomic> |
| 5 | #endif |
| 6 | |
| 7 | namespace sead |
| 8 | { |
| 9 | struct AtomicDirectInitTag |
| 10 | { |
| 11 | }; |
| 12 | |
| 13 | template <class T> |
| 14 | struct AtomicBase |
| 15 | { |
| 16 | public: |
| 17 | AtomicBase(T value = {}); // NOLINT(google-explicit-constructor) |
| 18 | /// Directly initialises the underlying atomic with the specified value. |
| 19 | /// Note that initialisation is not atomic. |
| 20 | AtomicBase(AtomicDirectInitTag, T value); |
| 21 | AtomicBase(const AtomicBase& rhs) { *this = rhs; } |
| 22 | |
| 23 | operator T() const { return load(); } |
| 24 | |
| 25 | AtomicBase& operator=(const AtomicBase& rhs) |
| 26 | { |
| 27 | store(value: rhs.load()); |
| 28 | return *this; |
| 29 | } |
| 30 | |
| 31 | AtomicBase& operator=(T value) |
| 32 | { |
| 33 | store(value); |
| 34 | return *this; |
| 35 | } |
| 36 | |
| 37 | /// Load the current value, as if with memory_order_relaxed. |
| 38 | T load() const; |
| 39 | /// Store a new value, as if with memory_order_relaxed. |
| 40 | void store(T value); |
| 41 | /// Non-atomically store a new value. |
| 42 | void storeNonAtomic(T value); |
| 43 | /// Exchange/swap the current value, as if with memory_order_relaxed. |
| 44 | /// @return the previous value |
| 45 | T exchange(T value); |
| 46 | /// Load the current value and if it is equal to `expected`, store `desired` |
| 47 | /// as if with memory_order_relaxed. |
| 48 | /// Otherwise, this sets `original` to the current value. |
| 49 | /// @param expected The value expected to be found in the atomic object, and to be replaced. |
| 50 | /// @param desired The new value to store in the atomic object if `expected` was found. |
| 51 | /// @param original The value that was found in the atomic object if the comparison fails. May |
| 52 | /// be null. Note that this is only updated when false is returned. |
| 53 | /// @return true if and only if the value was modified |
| 54 | bool compareExchange(T expected, T desired, T* original = nullptr); |
| 55 | |
| 56 | protected: |
| 57 | #ifdef NNSDK |
| 58 | // Nintendo appears to have manually implemented atomics with volatile and platform specific |
| 59 | // intrinsics (e.g. __builtin_arm_ldrex). |
| 60 | // For ease of implementation and portability, we will use std::atomic and cast to volatile |
| 61 | // when necessary. That is formally undefined behavior, but it should be safe because |
| 62 | // sead is built with -fno-strict-aliasing and because of the following static assertions. |
| 63 | std::atomic<T> mValue; |
| 64 | static_assert(sizeof(mValue) == sizeof(T), |
| 65 | "std::atomic<T> and T do not have the same size; unsupported case" ); |
| 66 | static_assert(alignof(decltype(mValue)) == alignof(volatile T), |
| 67 | "std::atomic<T> and T do not have the same alignment; unsupported case" ); |
| 68 | static_assert(std::atomic<T>::is_always_lock_free, |
| 69 | "std::atomic<T>::is_always_lock_free is not true; unsupported case" ); |
| 70 | |
| 71 | const volatile T* getValuePtr() const { return reinterpret_cast<const volatile T*>(&mValue); } |
| 72 | volatile T* getValuePtr() { return reinterpret_cast<volatile T*>(&mValue); } |
| 73 | |
| 74 | #endif |
| 75 | }; |
| 76 | |
| 77 | template <class T> |
| 78 | struct Atomic : AtomicBase<T> |
| 79 | { |
| 80 | using AtomicBase<T>::AtomicBase; |
| 81 | using AtomicBase<T>::operator=; |
| 82 | |
| 83 | T fetchAdd(T x); |
| 84 | T fetchSub(T x); |
| 85 | T fetchAnd(T x); |
| 86 | T fetchOr(T x); |
| 87 | T fetchXor(T x); |
| 88 | T increment() { return fetchAdd(x: 1); } |
| 89 | T decrement() { return fetchSub(x: 1); } |
| 90 | |
| 91 | bool isBitOn(unsigned int bit) const; |
| 92 | /// @return whether the bit was cleared and is now set. |
| 93 | bool setBitOn(unsigned int bit); |
| 94 | /// @return whether the bit was set and is now cleared. |
| 95 | bool setBitOff(unsigned int bit); |
| 96 | |
| 97 | T operator+=(T x) { return fetchAdd(x); } |
| 98 | T operator-=(T x) { return fetchSub(x); } |
| 99 | T operator&=(T x) { return fetchAnd(x); } |
| 100 | T operator|=(T x) { return fetchOr(x); } |
| 101 | T operator^=(T x) { return fetchXor(x); } |
| 102 | T operator++() { return fetchAdd(x: 1) + 1; } |
| 103 | T operator++(int) { return fetchAdd(x: 1); } |
| 104 | T operator--() { return fetchSub(x: 1) - 1; } |
| 105 | T operator--(int) { return fetchSub(x: 1); } |
| 106 | }; |
| 107 | |
| 108 | /// Specialization for pointer types. |
| 109 | template <class T> |
| 110 | struct Atomic<T*> : AtomicBase<T*> |
| 111 | { |
| 112 | using AtomicBase<T*>::AtomicBase; |
| 113 | using AtomicBase<T*>::operator=; |
| 114 | |
| 115 | T& operator*() const { return *this->load(); } |
| 116 | T* operator->() const { return this->load(); } |
| 117 | }; |
| 118 | |
| 119 | // Implementation. |
| 120 | |
| 121 | #ifdef NNSDK |
| 122 | template <class T> |
| 123 | inline AtomicBase<T>::AtomicBase(T value) |
| 124 | { |
| 125 | storeNonAtomic(value); |
| 126 | } |
| 127 | |
| 128 | template <class T> |
| 129 | inline AtomicBase<T>::AtomicBase(AtomicDirectInitTag, T value) : mValue(value) |
| 130 | { |
| 131 | } |
| 132 | |
| 133 | template <class T> |
| 134 | inline T AtomicBase<T>::load() const |
| 135 | { |
| 136 | #ifdef MATCHING_HACK_NX_CLANG |
| 137 | // Using std::atomic<T>::load prevents LLVM from folding ldr+sext into ldrsw. |
| 138 | return *getValuePtr(); |
| 139 | #else |
| 140 | return mValue.load(std::memory_order_relaxed); |
| 141 | #endif |
| 142 | } |
| 143 | |
| 144 | template <class T> |
| 145 | inline void AtomicBase<T>::store(T value) |
| 146 | { |
| 147 | mValue.store(value, std::memory_order_relaxed); |
| 148 | } |
| 149 | |
| 150 | template <class T> |
| 151 | inline void AtomicBase<T>::storeNonAtomic(T value) |
| 152 | { |
| 153 | *getValuePtr() = value; |
| 154 | } |
| 155 | |
| 156 | template <class T> |
| 157 | inline T AtomicBase<T>::exchange(T value) |
| 158 | { |
| 159 | return mValue.exchange(value, std::memory_order_relaxed); |
| 160 | } |
| 161 | |
| 162 | template <class T> |
| 163 | inline bool AtomicBase<T>::compareExchange(T expected, T desired, T* original) |
| 164 | { |
| 165 | #ifdef MATCHING_HACK_NX_CLANG |
| 166 | // Unlike Clang (https://reviews.llvm.org/D13033), Nintendo's implementation does not use clrex. |
| 167 | do |
| 168 | { |
| 169 | T value = __builtin_arm_ldrex(getValuePtr()); |
| 170 | if (value != expected) |
| 171 | { |
| 172 | if (original) |
| 173 | *original = value; |
| 174 | return false; |
| 175 | } |
| 176 | } while (__builtin_arm_strex(desired, getValuePtr())); |
| 177 | return true; |
| 178 | #else |
| 179 | T value = expected; |
| 180 | if (mValue.compare_exchange_strong(value, desired, std::memory_order_relaxed)) |
| 181 | return true; |
| 182 | if (original) |
| 183 | *original = value; |
| 184 | return false; |
| 185 | #endif |
| 186 | } |
| 187 | |
| 188 | #ifdef MATCHING_HACK_NX_CLANG |
| 189 | namespace detail |
| 190 | { |
| 191 | // To match Nintendo's implementation of atomics. |
| 192 | template <typename T, typename F> |
| 193 | inline T atomicReadModifyWrite(volatile T* value_ptr, F op) |
| 194 | { |
| 195 | T value; |
| 196 | do |
| 197 | { |
| 198 | value = __builtin_arm_ldrex(value_ptr); |
| 199 | } while (__builtin_arm_strex(op(value), value_ptr)); |
| 200 | return value; |
| 201 | } |
| 202 | } // namespace detail |
| 203 | #endif |
| 204 | |
| 205 | template <class T> |
| 206 | inline T Atomic<T>::fetchAdd(T x) |
| 207 | { |
| 208 | #ifdef MATCHING_HACK_NX_CLANG |
| 209 | return detail::atomicReadModifyWrite(this->getValuePtr(), [&](T val) { return val + x; }); |
| 210 | #else |
| 211 | return this->mValue.fetch_add(x, std::memory_order_relaxed); |
| 212 | #endif |
| 213 | } |
| 214 | |
| 215 | template <class T> |
| 216 | inline T Atomic<T>::fetchSub(T x) |
| 217 | { |
| 218 | #ifdef MATCHING_HACK_NX_CLANG |
| 219 | return detail::atomicReadModifyWrite(this->getValuePtr(), [&](T val) { return val - x; }); |
| 220 | #else |
| 221 | return this->mValue.fetch_sub(x, std::memory_order_relaxed); |
| 222 | #endif |
| 223 | } |
| 224 | |
| 225 | template <class T> |
| 226 | inline T Atomic<T>::fetchAnd(T x) |
| 227 | { |
| 228 | #ifdef MATCHING_HACK_NX_CLANG |
| 229 | return detail::atomicReadModifyWrite(this->getValuePtr(), [&](T val) { return val & x; }); |
| 230 | #else |
| 231 | return this->mValue.fetch_and(x, std::memory_order_relaxed); |
| 232 | #endif |
| 233 | } |
| 234 | |
| 235 | template <class T> |
| 236 | inline T Atomic<T>::fetchOr(T x) |
| 237 | { |
| 238 | #ifdef MATCHING_HACK_NX_CLANG |
| 239 | return detail::atomicReadModifyWrite(this->getValuePtr(), [&](T val) { return val | x; }); |
| 240 | #else |
| 241 | return this->mValue.fetch_or(x, std::memory_order_relaxed); |
| 242 | #endif |
| 243 | } |
| 244 | |
| 245 | template <class T> |
| 246 | inline T Atomic<T>::fetchXor(T x) |
| 247 | { |
| 248 | #ifdef MATCHING_HACK_NX_CLANG |
| 249 | return detail::atomicReadModifyWrite(this->getValuePtr(), [&](T val) { return val ^ x; }); |
| 250 | #else |
| 251 | return this->mValue.fetch_xor(x, std::memory_order_relaxed); |
| 252 | #endif |
| 253 | } |
| 254 | |
| 255 | template <class T> |
| 256 | bool Atomic<T>::isBitOn(unsigned int bit) const |
| 257 | { |
| 258 | return (this->load() & (1 << bit)) != 0; |
| 259 | } |
| 260 | |
| 261 | template <class T> |
| 262 | bool Atomic<T>::setBitOn(unsigned int bit) |
| 263 | { |
| 264 | #ifdef MATCHING_HACK_NX_CLANG |
| 265 | const auto old = detail::atomicReadModifyWrite(this->getValuePtr(), |
| 266 | [bit](T val) { return val | (1 << bit); }); |
| 267 | #else |
| 268 | const auto old = this->mValue.fetch_or(1 << bit, std::memory_order_relaxed); |
| 269 | #endif |
| 270 | return (old & (1 << bit)) == 0; |
| 271 | } |
| 272 | |
| 273 | template <class T> |
| 274 | bool Atomic<T>::setBitOff(unsigned int bit) |
| 275 | { |
| 276 | #ifdef MATCHING_HACK_NX_CLANG |
| 277 | const auto old = detail::atomicReadModifyWrite(this->getValuePtr(), |
| 278 | [bit](T val) { return val & ~(1 << bit); }); |
| 279 | #else |
| 280 | const auto old = this->mValue.fetch_and(~(1 << bit), std::memory_order_relaxed); |
| 281 | #endif |
| 282 | return (old & (1 << bit)) != 0; |
| 283 | } |
| 284 | #else // NNSDK |
| 285 | #error "Unknown platform" |
| 286 | #endif |
| 287 | } // namespace sead |
| 288 | |