2 * Copyright (C) 2008 The Android Open Source Project
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
8 * http://www.apache.org/licenses/LICENSE-2.0
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.
17 #ifndef ART_RUNTIME_ATOMIC_H_
18 #define ART_RUNTIME_ATOMIC_H_
25 #include "arch/instruction_set.h"
26 #include "base/logging.h"
27 #include "base/macros.h"
33 // QuasiAtomic encapsulates two separate facilities that we are
34 // trying to move away from: "quasiatomic" 64 bit operations
35 // and custom memory fences. For the time being, they remain
36 // exposed. Clients should be converted to use either class Atomic
37 // below whenever possible, and should eventually use C++11 atomics.
38 // The two facilities that do not have a good C++11 analog are
39 // ThreadFenceForConstructor and Atomic::*JavaData.
41 // NOTE: Two "quasiatomic" operations on the exact same memory address
42 // are guaranteed to operate atomically with respect to each other,
43 // but no guarantees are made about quasiatomic operations mixed with
44 // non-quasiatomic operations on the same address, nor about
45 // quasiatomic operations that are performed on partially-overlapping
48 static constexpr bool NeedSwapMutexes(InstructionSet isa) {
49 // TODO - mips64 still need this for Cas64 ???
50 return (isa == kMips) || (isa == kMips64);
54 static void Startup();
56 static void Shutdown();
58 // Reads the 64-bit value at "addr" without tearing.
59 static int64_t Read64(volatile const int64_t* addr) {
60 if (!NeedSwapMutexes(kRuntimeISA)) {
66 #if defined(__ARM_FEATURE_LPAE)
67 // With LPAE support (such as Cortex-A15) then ldrd is defined not to tear.
68 __asm__ __volatile__("@ QuasiAtomic::Read64\n"
73 // Exclusive loads are defined not to tear, clearing the exclusive state isn't necessary.
74 __asm__ __volatile__("@ QuasiAtomic::Read64\n"
79 #elif defined(__i386__)
85 LOG(FATAL) << "Unsupported architecture";
87 #endif // defined(__LP64__)
90 return SwapMutexRead64(addr);
94 // Writes to the 64-bit value at "addr" without tearing.
95 static void Write64(volatile int64_t* addr, int64_t value) {
96 if (!NeedSwapMutexes(kRuntimeISA)) {
101 #if defined(__ARM_FEATURE_LPAE)
102 // If we know that ARM architecture has LPAE (such as Cortex-A15) strd is defined not to tear.
103 __asm__ __volatile__("@ QuasiAtomic::Write64\n"
108 // The write is done as a swap so that the cache-line is in the exclusive state for the store.
112 __asm__ __volatile__("@ QuasiAtomic::Write64\n"
113 "ldrexd %0, %H0, %2\n"
114 "strexd %1, %3, %H3, %2"
115 : "=&r" (prev), "=&r" (status), "+Q"(*addr)
118 } while (UNLIKELY(status != 0));
120 #elif defined(__i386__)
121 __asm__ __volatile__(
126 LOG(FATAL) << "Unsupported architecture";
128 #endif // defined(__LP64__)
130 SwapMutexWrite64(addr, value);
134 // Atomically compare the value at "addr" to "old_value", if equal replace it with "new_value"
135 // and return true. Otherwise, don't swap, and return false.
136 // This is fully ordered, i.e. it has C++11 memory_order_seq_cst
137 // semantics (assuming all other accesses use a mutex if this one does).
138 // This has "strong" semantics; if it fails then it is guaranteed that
139 // at some point during the execution of Cas64, *addr was not equal to
141 static bool Cas64(int64_t old_value, int64_t new_value, volatile int64_t* addr) {
142 if (!NeedSwapMutexes(kRuntimeISA)) {
143 return __sync_bool_compare_and_swap(addr, old_value, new_value);
145 return SwapMutexCas64(old_value, new_value, addr);
149 // Does the architecture provide reasonable atomic long operations or do we fall back on mutexes?
150 static bool LongAtomicsUseMutexes(InstructionSet isa) {
151 return NeedSwapMutexes(isa);
154 static void ThreadFenceAcquire() {
155 std::atomic_thread_fence(std::memory_order_acquire);
158 static void ThreadFenceRelease() {
159 std::atomic_thread_fence(std::memory_order_release);
162 static void ThreadFenceForConstructor() {
163 #if defined(__aarch64__)
164 __asm__ __volatile__("dmb ishst" : : : "memory");
166 std::atomic_thread_fence(std::memory_order_release);
170 static void ThreadFenceSequentiallyConsistent() {
171 std::atomic_thread_fence(std::memory_order_seq_cst);
175 static Mutex* GetSwapMutex(const volatile int64_t* addr);
176 static int64_t SwapMutexRead64(volatile const int64_t* addr);
177 static void SwapMutexWrite64(volatile int64_t* addr, int64_t val);
178 static bool SwapMutexCas64(int64_t old_value, int64_t new_value, volatile int64_t* addr);
180 // We stripe across a bunch of different mutexes to reduce contention.
181 static constexpr size_t kSwapMutexCount = 32;
182 static std::vector<Mutex*>* gSwapMutexes;
184 DISALLOW_COPY_AND_ASSIGN(QuasiAtomic);
188 class PACKED(sizeof(T)) Atomic : public std::atomic<T> {
190 Atomic<T>() : std::atomic<T>(0) { }
192 explicit Atomic<T>(T value) : std::atomic<T>(value) { }
194 // Load from memory without ordering or synchronization constraints.
195 T LoadRelaxed() const {
196 return this->load(std::memory_order_relaxed);
199 // Load from memory with acquire ordering.
200 T LoadAcquire() const {
201 return this->load(std::memory_order_acquire);
204 // Word tearing allowed, but may race.
206 // There has been some discussion of eventually disallowing word
207 // tearing for Java data loads.
208 T LoadJavaData() const {
209 return this->load(std::memory_order_relaxed);
212 // Load from memory with a total ordering.
213 // Corresponds exactly to a Java volatile load.
214 T LoadSequentiallyConsistent() const {
215 return this->load(std::memory_order_seq_cst);
218 // Store to memory without ordering or synchronization constraints.
219 void StoreRelaxed(T desired) {
220 this->store(desired, std::memory_order_relaxed);
223 // Word tearing allowed, but may race.
224 void StoreJavaData(T desired) {
225 this->store(desired, std::memory_order_relaxed);
228 // Store to memory with release ordering.
229 void StoreRelease(T desired) {
230 this->store(desired, std::memory_order_release);
233 // Store to memory with a total ordering.
234 void StoreSequentiallyConsistent(T desired) {
235 this->store(desired, std::memory_order_seq_cst);
238 // Atomically replace the value with desired value.
239 T ExchangeRelaxed(T desired_value) {
240 return this->exchange(desired_value, std::memory_order_relaxed);
243 // Atomically replace the value with desired value if it matches the expected value.
244 // Participates in total ordering of atomic operations.
245 bool CompareExchangeStrongSequentiallyConsistent(T expected_value, T desired_value) {
246 return this->compare_exchange_strong(expected_value, desired_value, std::memory_order_seq_cst);
249 // The same, except it may fail spuriously.
250 bool CompareExchangeWeakSequentiallyConsistent(T expected_value, T desired_value) {
251 return this->compare_exchange_weak(expected_value, desired_value, std::memory_order_seq_cst);
254 // Atomically replace the value with desired value if it matches the expected value. Doesn't
255 // imply ordering or synchronization constraints.
256 bool CompareExchangeStrongRelaxed(T expected_value, T desired_value) {
257 return this->compare_exchange_strong(expected_value, desired_value, std::memory_order_relaxed);
260 // Atomically replace the value with desired value if it matches the expected value. Prior writes
261 // to other memory locations become visible to the threads that do a consume or an acquire on the
263 bool CompareExchangeStrongRelease(T expected_value, T desired_value) {
264 return this->compare_exchange_strong(expected_value, desired_value, std::memory_order_release);
267 // The same, except it may fail spuriously.
268 bool CompareExchangeWeakRelaxed(T expected_value, T desired_value) {
269 return this->compare_exchange_weak(expected_value, desired_value, std::memory_order_relaxed);
272 // Atomically replace the value with desired value if it matches the expected value. Prior writes
273 // made to other memory locations by the thread that did the release become visible in this
275 bool CompareExchangeWeakAcquire(T expected_value, T desired_value) {
276 return this->compare_exchange_weak(expected_value, desired_value, std::memory_order_acquire);
279 // Atomically replace the value with desired value if it matches the expected value. prior writes
280 // to other memory locations become visible to the threads that do a consume or an acquire on the
282 bool CompareExchangeWeakRelease(T expected_value, T desired_value) {
283 return this->compare_exchange_weak(expected_value, desired_value, std::memory_order_release);
286 T FetchAndAddSequentiallyConsistent(const T value) {
287 return this->fetch_add(value, std::memory_order_seq_cst); // Return old_value.
290 T FetchAndAddRelaxed(const T value) {
291 return this->fetch_add(value, std::memory_order_relaxed); // Return old_value.
294 T FetchAndSubSequentiallyConsistent(const T value) {
295 return this->fetch_sub(value, std::memory_order_seq_cst); // Return old value.
298 T FetchAndSubRelaxed(const T value) {
299 return this->fetch_sub(value, std::memory_order_relaxed); // Return old value.
302 T FetchAndOrSequentiallyConsistent(const T value) {
303 return this->fetch_or(value, std::memory_order_seq_cst); // Return old_value.
306 T FetchAndAndSequentiallyConsistent(const T value) {
307 return this->fetch_and(value, std::memory_order_seq_cst); // Return old_value.
310 volatile T* Address() {
311 return reinterpret_cast<T*>(this);
314 static T MaxValue() {
315 return std::numeric_limits<T>::max();
319 typedef Atomic<int32_t> AtomicInteger;
321 static_assert(sizeof(AtomicInteger) == sizeof(int32_t), "Weird AtomicInteger size");
322 static_assert(alignof(AtomicInteger) == alignof(int32_t),
323 "AtomicInteger alignment differs from that of underlyingtype");
324 static_assert(sizeof(Atomic<int64_t>) == sizeof(int64_t), "Weird Atomic<int64> size");
326 // Assert the alignment of 64-bit integers is 64-bit. This isn't true on certain 32-bit
327 // architectures (e.g. x86-32) but we know that 64-bit integers here are arranged to be 8-byte
329 #if defined(__LP64__)
330 static_assert(alignof(Atomic<int64_t>) == alignof(int64_t),
331 "Atomic<int64> alignment differs from that of underlying type");
336 #endif // ART_RUNTIME_ATOMIC_H_