atomic/detail/core_ops_gcc_sync.hpp

0001 /*
0002  * Distributed under the Boost Software License, Version 1.0.
0003  * (See accompanying file LICENSE_1_0.txt or copy at
0004  * http://www.boost.org/LICENSE_1_0.txt)
0005  *
0006  * Copyright (c) 2011 Helge Bahmann
0007  * Copyright (c) 2013 Tim Blechmann
0008  * Copyright (c) 2014 Andrey Semashev
0009  */
0010 /*!
0011  * \file   atomic/detail/core_ops_gcc_sync.hpp
0012  *
0013  * This header contains implementation of the \c core_operations template.
0014  */
0015
0016 #ifndef BOOST_ATOMIC_DETAIL_CORE_OPS_GCC_SYNC_HPP_INCLUDED_
0017 #define BOOST_ATOMIC_DETAIL_CORE_OPS_GCC_SYNC_HPP_INCLUDED_
0018
0019 #include <cstddef>
0020 #include <boost/memory_order.hpp>
0021 #include <boost/atomic/detail/config.hpp>
0022 #include <boost/atomic/detail/storage_traits.hpp>
0023 #include <boost/atomic/detail/core_operations_fwd.hpp>
0024 #include <boost/atomic/detail/extending_cas_based_arithmetic.hpp>
0025 #include <boost/atomic/detail/type_traits/integral_constant.hpp>
0026 #include <boost/atomic/detail/capabilities.hpp>
0027 #include <boost/atomic/detail/header.hpp>
0028
0029 #ifdef BOOST_HAS_PRAGMA_ONCE
0030 #pragma once
0031 #endif
0032
0033 namespace boost {
0034 namespace atomics {
0035 namespace detail {
0036
0037 struct core_operations_gcc_sync_base
0038 {
0039     static BOOST_CONSTEXPR_OR_CONST bool full_cas_based = false;
0040     static BOOST_CONSTEXPR_OR_CONST bool is_always_lock_free = true;
0041
0042     static BOOST_FORCEINLINE void fence_before_store(memory_order order) BOOST_NOEXCEPT
0043     {
0044         if ((static_cast< unsigned int >(order) & static_cast< unsigned int >(memory_order_release)) != 0u)
0045             __sync_synchronize();
0046     }
0047
0048     static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT
0049     {
0050         if (order == memory_order_seq_cst)
0051             __sync_synchronize();
0052     }
0053
0054     static BOOST_FORCEINLINE void fence_after_load(memory_order order) BOOST_NOEXCEPT
0055     {
0056         if ((static_cast< unsigned int >(order) & (static_cast< unsigned int >(memory_order_acquire) | static_cast< unsigned int >(memory_order_consume))) != 0u)
0057             __sync_synchronize();
0058     }
0059 };
0060
0061 template< std::size_t Size, bool Signed, bool Interprocess >
0062 struct core_operations_gcc_sync :
0063     public core_operations_gcc_sync_base
0064 {
0065     typedef typename storage_traits< Size >::type storage_type;
0066
0067     static BOOST_CONSTEXPR_OR_CONST std::size_t storage_size = Size;
0068     static BOOST_CONSTEXPR_OR_CONST std::size_t storage_alignment = storage_traits< storage_size >::alignment;
0069     static BOOST_CONSTEXPR_OR_CONST bool is_signed = Signed;
0070     static BOOST_CONSTEXPR_OR_CONST bool is_interprocess = Interprocess;
0071
0072     // In general, we cannot guarantee atomicity of plain loads and stores of anything larger than a single byte on
0073     // an arbitrary CPU architecture. However, all modern architectures seem to guarantee atomic loads and stores of
0074     // suitably aligned objects of up to a pointer size. For larger objects we should probably use intrinsics to guarantee
0075     // atomicity. If there appears an architecture where this doesn't hold, this threshold needs to be updated (patches are welcome).
0076     typedef atomics::detail::integral_constant< bool, storage_size <= sizeof(void*) > plain_stores_loads_are_atomic;
0077
0078     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
0079     {
0080         store(storage, v, order, plain_stores_loads_are_atomic());
0081     }
0082
0083     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order, atomics::detail::true_type) BOOST_NOEXCEPT
0084     {
0085         fence_before_store(order);
0086         storage = v;
0087         fence_after_store(order);
0088     }
0089
0090     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order, atomics::detail::false_type) BOOST_NOEXCEPT
0091     {
0092         exchange(storage, v, order);
0093     }
0094
0095     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
0096     {
0097         return load(storage, order, plain_stores_loads_are_atomic());
0098     }
0099
0100     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order, atomics::detail::true_type) BOOST_NOEXCEPT
0101     {
0102         storage_type v = storage;
0103         fence_after_load(order);
0104         return v;
0105     }
0106
0107     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order, atomics::detail::false_type) BOOST_NOEXCEPT
0108     {
0109         // Note: don't use fetch_add or other arithmetics here since storage_type may not be an arithmetic type.
0110         storage_type expected = storage_type();
0111         storage_type desired = expected;
0112         // We don't care if CAS succeeds or not. If it does, it will just write the same value there was before.
0113         return __sync_val_compare_and_swap(const_cast< storage_type volatile* >(&storage), expected, desired);
0114     }
0115
0116     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
0117     {
0118         return __sync_fetch_and_add(&storage, v);
0119     }
0120
0121     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
0122     {
0123         return __sync_fetch_and_sub(&storage, v);
0124     }
0125
0126     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
0127     {
0128         // GCC docs mention that not all architectures may support full exchange semantics for this intrinsic. However, GCC's implementation of
0129         // std::atomic<> uses this intrinsic unconditionally. We do so as well. In case if some architectures actually don't support this, we can always
0130         // add a check here and fall back to a CAS loop.
0131         if ((static_cast< unsigned int >(order) & static_cast< unsigned int >(memory_order_release)) != 0u)
0132             __sync_synchronize();
0133         return __sync_lock_test_and_set(&storage, v);
0134     }
0135
0136     static BOOST_FORCEINLINE bool compare_exchange_strong(
0137         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
0138     {
0139         storage_type expected2 = expected;
0140         storage_type old_val = __sync_val_compare_and_swap(&storage, expected2, desired);
0141
0142         if (old_val == expected2)
0143         {
0144             return true;
0145         }
0146         else
0147         {
0148             expected = old_val;
0149             return false;
0150         }
0151     }
0152
0153     static BOOST_FORCEINLINE bool compare_exchange_weak(
0154         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
0155     {
0156         return compare_exchange_strong(storage, expected, desired, success_order, failure_order);
0157     }
0158
0159     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
0160     {
0161         return __sync_fetch_and_and(&storage, v);
0162     }
0163
0164     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
0165     {
0166         return __sync_fetch_and_or(&storage, v);
0167     }
0168
0169     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
0170     {
0171         return __sync_fetch_and_xor(&storage, v);
0172     }
0173
0174     static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
0175     {
0176         if ((static_cast< unsigned int >(order) & static_cast< unsigned int >(memory_order_release)) != 0u)
0177             __sync_synchronize();
0178         return !!__sync_lock_test_and_set(&storage, 1);
0179     }
0180
0181     static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
0182     {
0183         __sync_lock_release(&storage);
0184         if (order == memory_order_seq_cst)
0185             __sync_synchronize();
0186     }
0187 };
0188
0189 #if BOOST_ATOMIC_INT8_LOCK_FREE > 0
0190 template< bool Signed, bool Interprocess >
0191 struct core_operations< 1u, Signed, Interprocess > :
0192 #if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1)
0193     public core_operations_gcc_sync< 1u, Signed, Interprocess >
0194 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2)
0195     public extending_cas_based_arithmetic< core_operations_gcc_sync< 2u, Signed, Interprocess >, 1u, Signed >
0196 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
0197     public extending_cas_based_arithmetic< core_operations_gcc_sync< 4u, Signed, Interprocess >, 1u, Signed >
0198 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)
0199     public extending_cas_based_arithmetic< core_operations_gcc_sync< 8u, Signed, Interprocess >, 1u, Signed >
0200 #else
0201     public extending_cas_based_arithmetic< core_operations_gcc_sync< 16u, Signed, Interprocess >, 1u, Signed >
0202 #endif
0203 {
0204 };
0205 #endif
0206
0207 #if BOOST_ATOMIC_INT16_LOCK_FREE > 0
0208 template< bool Signed, bool Interprocess >
0209 struct core_operations< 2u, Signed, Interprocess > :
0210 #if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2)
0211     public core_operations_gcc_sync< 2u, Signed, Interprocess >
0212 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
0213     public extending_cas_based_arithmetic< core_operations_gcc_sync< 4u, Signed, Interprocess >, 2u, Signed >
0214 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)
0215     public extending_cas_based_arithmetic< core_operations_gcc_sync< 8u, Signed, Interprocess >, 2u, Signed >
0216 #else
0217     public extending_cas_based_arithmetic< core_operations_gcc_sync< 16u, Signed, Interprocess >, 2u, Signed >
0218 #endif
0219 {
0220 };
0221 #endif
0222
0223 #if BOOST_ATOMIC_INT32_LOCK_FREE > 0
0224 template< bool Signed, bool Interprocess >
0225 struct core_operations< 4u, Signed, Interprocess > :
0226 #if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
0227     public core_operations_gcc_sync< 4u, Signed, Interprocess >
0228 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)
0229     public extending_cas_based_arithmetic< core_operations_gcc_sync< 8u, Signed, Interprocess >, 4u, Signed >
0230 #else
0231     public extending_cas_based_arithmetic< core_operations_gcc_sync< 16u, Signed, Interprocess >, 4u, Signed >
0232 #endif
0233 {
0234 };
0235 #endif
0236
0237 #if BOOST_ATOMIC_INT64_LOCK_FREE > 0
0238 template< bool Signed, bool Interprocess >
0239 struct core_operations< 8u, Signed, Interprocess > :
0240 #if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)
0241     public core_operations_gcc_sync< 8u, Signed, Interprocess >
0242 #else
0243     public extending_cas_based_arithmetic< core_operations_gcc_sync< 16u, Signed, Interprocess >, 8u, Signed >
0244 #endif
0245 {
0246 };
0247 #endif
0248
0249 #if BOOST_ATOMIC_INT128_LOCK_FREE > 0
0250 template< bool Signed, bool Interprocess >
0251 struct core_operations< 16u, Signed, Interprocess > :
0252     public core_operations_gcc_sync< 16u, Signed, Interprocess >
0253 {
0254 };
0255 #endif
0256
0257 } // namespace detail
0258 } // namespace atomics
0259 } // namespace boost
0260
0261 #include <boost/atomic/detail/footer.hpp>
0262
0263 #endif // BOOST_ATOMIC_DETAIL_CORE_OPS_GCC_SYNC_HPP_INCLUDED_