From 08c1e93fa36a49f49325a07fe91ff92c964c2b6c Mon Sep 17 00:00:00 2001 From: Chanho Park Date: Thu, 11 Dec 2014 18:55:56 +0900 Subject: Imported Upstream version 1.57.0 --- boost/atomic/detail/atomic_flag.hpp | 70 ++ boost/atomic/detail/atomic_template.hpp | 774 +++++++++++++++++++ boost/atomic/detail/caps_gcc_alpha.hpp | 34 + boost/atomic/detail/caps_gcc_arm.hpp | 56 ++ boost/atomic/detail/caps_gcc_atomic.hpp | 134 ++++ boost/atomic/detail/caps_gcc_ppc.hpp | 36 + boost/atomic/detail/caps_gcc_sparc.hpp | 34 + boost/atomic/detail/caps_gcc_sync.hpp | 62 ++ boost/atomic/detail/caps_gcc_x86.hpp | 52 ++ boost/atomic/detail/caps_linux_arm.hpp | 35 + boost/atomic/detail/caps_msvc_arm.hpp | 34 + boost/atomic/detail/caps_msvc_x86.hpp | 50 ++ boost/atomic/detail/caps_windows.hpp | 33 + boost/atomic/detail/casts.hpp | 64 ++ boost/atomic/detail/config.hpp | 24 + boost/atomic/detail/int_sizes.hpp | 140 ++++ boost/atomic/detail/interlocked.hpp | 481 ++++++++++++ boost/atomic/detail/link.hpp | 58 ++ boost/atomic/detail/lockpool.hpp | 51 ++ boost/atomic/detail/operations.hpp | 24 + boost/atomic/detail/operations_fwd.hpp | 34 + boost/atomic/detail/operations_lockfree.hpp | 30 + boost/atomic/detail/ops_cas_based.hpp | 91 +++ boost/atomic/detail/ops_emulated.hpp | 149 ++++ boost/atomic/detail/ops_extending_cas_based.hpp | 65 ++ boost/atomic/detail/ops_gcc_alpha.hpp | 874 +++++++++++++++++++++ boost/atomic/detail/ops_gcc_arm.hpp | 971 ++++++++++++++++++++++++ boost/atomic/detail/ops_gcc_atomic.hpp | 380 ++++++++++ boost/atomic/detail/ops_gcc_ppc.hpp | 775 +++++++++++++++++++ boost/atomic/detail/ops_gcc_sparc.hpp | 245 ++++++ boost/atomic/detail/ops_gcc_sync.hpp | 237 ++++++ boost/atomic/detail/ops_gcc_x86.hpp | 510 +++++++++++++ boost/atomic/detail/ops_gcc_x86_dcas.hpp | 308 ++++++++ boost/atomic/detail/ops_linux_arm.hpp | 177 +++++ boost/atomic/detail/ops_msvc_arm.hpp | 820 ++++++++++++++++++++ boost/atomic/detail/ops_msvc_common.hpp | 38 + boost/atomic/detail/ops_msvc_x86.hpp | 879 +++++++++++++++++++++ boost/atomic/detail/ops_windows.hpp | 215 ++++++ boost/atomic/detail/pause.hpp | 43 ++ boost/atomic/detail/platform.hpp | 115 +++ boost/atomic/detail/storage_type.hpp | 168 ++++ 41 files changed, 9370 insertions(+) create mode 100644 boost/atomic/detail/atomic_flag.hpp create mode 100644 boost/atomic/detail/atomic_template.hpp create mode 100644 boost/atomic/detail/caps_gcc_alpha.hpp create mode 100644 boost/atomic/detail/caps_gcc_arm.hpp create mode 100644 boost/atomic/detail/caps_gcc_atomic.hpp create mode 100644 boost/atomic/detail/caps_gcc_ppc.hpp create mode 100644 boost/atomic/detail/caps_gcc_sparc.hpp create mode 100644 boost/atomic/detail/caps_gcc_sync.hpp create mode 100644 boost/atomic/detail/caps_gcc_x86.hpp create mode 100644 boost/atomic/detail/caps_linux_arm.hpp create mode 100644 boost/atomic/detail/caps_msvc_arm.hpp create mode 100644 boost/atomic/detail/caps_msvc_x86.hpp create mode 100644 boost/atomic/detail/caps_windows.hpp create mode 100644 boost/atomic/detail/casts.hpp create mode 100644 boost/atomic/detail/config.hpp create mode 100644 boost/atomic/detail/int_sizes.hpp create mode 100644 boost/atomic/detail/interlocked.hpp create mode 100644 boost/atomic/detail/link.hpp create mode 100644 boost/atomic/detail/lockpool.hpp create mode 100644 boost/atomic/detail/operations.hpp create mode 100644 boost/atomic/detail/operations_fwd.hpp create mode 100644 boost/atomic/detail/operations_lockfree.hpp create mode 100644 boost/atomic/detail/ops_cas_based.hpp create mode 100644 boost/atomic/detail/ops_emulated.hpp create mode 100644 boost/atomic/detail/ops_extending_cas_based.hpp create mode 100644 boost/atomic/detail/ops_gcc_alpha.hpp create mode 100644 boost/atomic/detail/ops_gcc_arm.hpp create mode 100644 boost/atomic/detail/ops_gcc_atomic.hpp create mode 100644 boost/atomic/detail/ops_gcc_ppc.hpp create mode 100644 boost/atomic/detail/ops_gcc_sparc.hpp create mode 100644 boost/atomic/detail/ops_gcc_sync.hpp create mode 100644 boost/atomic/detail/ops_gcc_x86.hpp create mode 100644 boost/atomic/detail/ops_gcc_x86_dcas.hpp create mode 100644 boost/atomic/detail/ops_linux_arm.hpp create mode 100644 boost/atomic/detail/ops_msvc_arm.hpp create mode 100644 boost/atomic/detail/ops_msvc_common.hpp create mode 100644 boost/atomic/detail/ops_msvc_x86.hpp create mode 100644 boost/atomic/detail/ops_windows.hpp create mode 100644 boost/atomic/detail/pause.hpp create mode 100644 boost/atomic/detail/platform.hpp create mode 100644 boost/atomic/detail/storage_type.hpp (limited to 'boost/atomic/detail') diff --git a/boost/atomic/detail/atomic_flag.hpp b/boost/atomic/detail/atomic_flag.hpp new file mode 100644 index 0000000000..6a6667d8eb --- /dev/null +++ b/boost/atomic/detail/atomic_flag.hpp @@ -0,0 +1,70 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/atomic_flag.hpp + * + * This header contains interface definition of \c atomic_flag. + */ + +#ifndef BOOST_ATOMIC_DETAIL_ATOMIC_FLAG_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_ATOMIC_FLAG_HPP_INCLUDED_ + +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +/* + * IMPLEMENTATION NOTE: All interface functions MUST be declared with BOOST_FORCEINLINE, + * see comment for convert_memory_order_to_gcc in ops_gcc_atomic.hpp. + */ + +namespace boost { +namespace atomics { + +#if defined(BOOST_NO_CXX11_CONSTEXPR) || defined(BOOST_NO_CXX11_UNIFIED_INITIALIZATION_SYNTAX) +#define BOOST_ATOMIC_NO_ATOMIC_FLAG_INIT +#else +#define BOOST_ATOMIC_FLAG_INIT {} +#endif + +struct atomic_flag +{ + typedef atomics::detail::operations< 1u, false > operations; + typedef operations::storage_type storage_type; + + storage_type m_storage; + + BOOST_FORCEINLINE BOOST_CONSTEXPR atomic_flag() BOOST_NOEXCEPT : m_storage(0) + { + } + + BOOST_FORCEINLINE bool test_and_set(memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return operations::test_and_set(m_storage, order); + } + + BOOST_FORCEINLINE void clear(memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + operations::clear(m_storage, order); + } + + BOOST_DELETED_FUNCTION(atomic_flag(atomic_flag const&)) + BOOST_DELETED_FUNCTION(atomic_flag& operator= (atomic_flag const&)) +}; + +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_ATOMIC_FLAG_HPP_INCLUDED_ diff --git a/boost/atomic/detail/atomic_template.hpp b/boost/atomic/detail/atomic_template.hpp new file mode 100644 index 0000000000..bc3922a711 --- /dev/null +++ b/boost/atomic/detail/atomic_template.hpp @@ -0,0 +1,774 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2011 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/atomic_template.hpp + * + * This header contains interface definition of \c atomic template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_ATOMIC_TEMPLATE_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_ATOMIC_TEMPLATE_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(BOOST_MSVC) +#pragma warning(push) +// 'boost::atomics::atomic' : multiple assignment operators specified +#pragma warning(disable: 4522) +#endif + +/* + * IMPLEMENTATION NOTE: All interface functions MUST be declared with BOOST_FORCEINLINE, + * see comment for convert_memory_order_to_gcc in ops_gcc_atomic.hpp. + */ + +namespace boost { +namespace atomics { +namespace detail { + +BOOST_FORCEINLINE BOOST_CONSTEXPR memory_order deduce_failure_order(memory_order order) BOOST_NOEXCEPT +{ + return order == memory_order_acq_rel ? memory_order_acquire : (order == memory_order_release ? memory_order_relaxed : order); +} + +BOOST_FORCEINLINE BOOST_CONSTEXPR bool cas_failure_order_must_not_be_stronger_than_success_order(memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT +{ + // 15 == (memory_order_seq_cst | memory_order_consume), see memory_order.hpp + // Given the enum values we can test the strength of memory order requirements with this single condition. + return (failure_order & 15u) <= (success_order & 15u); +} + +template< typename T, bool IsInt = boost::is_integral< T >::value > +struct classify +{ + typedef void type; +}; + +template< typename T > +struct classify< T, true > { typedef int type; }; + +template< typename T > +struct classify< T*, false > { typedef void* type; }; + +template< typename T, typename Kind > +class base_atomic; + +//! Implementation for integers +template< typename T > +class base_atomic< T, int > +{ +private: + typedef T value_type; + typedef T difference_type; + typedef atomics::detail::operations< storage_size_of< value_type >::value, boost::is_signed< T >::value > operations; + +protected: + typedef value_type value_arg_type; + +public: + typedef typename operations::storage_type storage_type; + +protected: + storage_type m_storage; + +public: + BOOST_DEFAULTED_FUNCTION(base_atomic(), {}) + BOOST_CONSTEXPR explicit base_atomic(value_type v) BOOST_NOEXCEPT : m_storage(v) {} + + BOOST_FORCEINLINE void store(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_consume); + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + + operations::store(m_storage, static_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE value_type load(memory_order order = memory_order_seq_cst) const volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_release); + BOOST_ASSERT(order != memory_order_acq_rel); + + return static_cast< value_type >(operations::load(m_storage, order)); + } + + BOOST_FORCEINLINE value_type fetch_add(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::fetch_add(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type fetch_sub(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::fetch_sub(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type exchange(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::exchange(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = static_cast< storage_type >(expected); + const bool res = operations::compare_exchange_strong(m_storage, old_value, static_cast< storage_type >(desired), success_order, failure_order); + expected = static_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_strong(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = static_cast< storage_type >(expected); + const bool res = operations::compare_exchange_weak(m_storage, old_value, static_cast< storage_type >(desired), success_order, failure_order); + expected = static_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_weak(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE value_type fetch_and(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::fetch_and(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type fetch_or(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::fetch_or(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type fetch_xor(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return static_cast< value_type >(operations::fetch_xor(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE bool is_lock_free() const volatile BOOST_NOEXCEPT + { + return operations::is_lock_free(m_storage); + } + + BOOST_FORCEINLINE value_type operator++(int) volatile BOOST_NOEXCEPT + { + return fetch_add(1); + } + + BOOST_FORCEINLINE value_type operator++() volatile BOOST_NOEXCEPT + { + return fetch_add(1) + 1; + } + + BOOST_FORCEINLINE value_type operator--(int) volatile BOOST_NOEXCEPT + { + return fetch_sub(1); + } + + BOOST_FORCEINLINE value_type operator--() volatile BOOST_NOEXCEPT + { + return fetch_sub(1) - 1; + } + + BOOST_FORCEINLINE value_type operator+=(difference_type v) volatile BOOST_NOEXCEPT + { + return fetch_add(v) + v; + } + + BOOST_FORCEINLINE value_type operator-=(difference_type v) volatile BOOST_NOEXCEPT + { + return fetch_sub(v) - v; + } + + BOOST_FORCEINLINE value_type operator&=(value_type v) volatile BOOST_NOEXCEPT + { + return fetch_and(v) & v; + } + + BOOST_FORCEINLINE value_type operator|=(value_type v) volatile BOOST_NOEXCEPT + { + return fetch_or(v) | v; + } + + BOOST_FORCEINLINE value_type operator^=(value_type v) volatile BOOST_NOEXCEPT + { + return fetch_xor(v) ^ v; + } + + BOOST_DELETED_FUNCTION(base_atomic(base_atomic const&)) + BOOST_DELETED_FUNCTION(base_atomic& operator=(base_atomic const&)) +}; + +//! Implementation for bool +template< > +class base_atomic< bool, int > +{ +private: + typedef bool value_type; + typedef atomics::detail::operations< 1u, false > operations; + +protected: + typedef value_type value_arg_type; + +public: + typedef operations::storage_type storage_type; + +protected: + storage_type m_storage; + +public: + BOOST_DEFAULTED_FUNCTION(base_atomic(), {}) + BOOST_CONSTEXPR explicit base_atomic(value_type v) BOOST_NOEXCEPT : m_storage(v) {} + + BOOST_FORCEINLINE void store(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_consume); + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + + operations::store(m_storage, static_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE value_type load(memory_order order = memory_order_seq_cst) const volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_release); + BOOST_ASSERT(order != memory_order_acq_rel); + + return !!operations::load(m_storage, order); + } + + BOOST_FORCEINLINE value_type exchange(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return !!operations::exchange(m_storage, static_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = static_cast< storage_type >(expected); + const bool res = operations::compare_exchange_strong(m_storage, old_value, static_cast< storage_type >(desired), success_order, failure_order); + expected = !!old_value; + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_strong(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = static_cast< storage_type >(expected); + const bool res = operations::compare_exchange_weak(m_storage, old_value, static_cast< storage_type >(desired), success_order, failure_order); + expected = !!old_value; + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_weak(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool is_lock_free() const volatile BOOST_NOEXCEPT + { + return operations::is_lock_free(m_storage); + } + + BOOST_DELETED_FUNCTION(base_atomic(base_atomic const&)) + BOOST_DELETED_FUNCTION(base_atomic& operator=(base_atomic const&)) +}; + + +//! Implementation for user-defined types, such as structs and enums +template< typename T > +class base_atomic< T, void > +{ +private: + typedef T value_type; + typedef atomics::detail::operations< storage_size_of< value_type >::value, false > operations; + +protected: + typedef value_type const& value_arg_type; + +public: + typedef typename operations::storage_type storage_type; + +protected: + storage_type m_storage; + +public: + BOOST_FORCEINLINE explicit base_atomic(value_type const& v = value_type()) BOOST_NOEXCEPT : m_storage(atomics::detail::memcpy_cast< storage_type >(v)) + { + } + + BOOST_FORCEINLINE void store(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_consume); + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + + operations::store(m_storage, atomics::detail::memcpy_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE value_type load(memory_order order = memory_order_seq_cst) const volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_release); + BOOST_ASSERT(order != memory_order_acq_rel); + + return atomics::detail::memcpy_cast< value_type >(operations::load(m_storage, order)); + } + + BOOST_FORCEINLINE value_type exchange(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::memcpy_cast< value_type >(operations::exchange(m_storage, atomics::detail::memcpy_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::memcpy_cast< storage_type >(expected); + const bool res = operations::compare_exchange_strong(m_storage, old_value, atomics::detail::memcpy_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::memcpy_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_strong(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::memcpy_cast< storage_type >(expected); + const bool res = operations::compare_exchange_weak(m_storage, old_value, atomics::detail::memcpy_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::memcpy_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_weak(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool is_lock_free() const volatile BOOST_NOEXCEPT + { + return operations::is_lock_free(m_storage); + } + + BOOST_DELETED_FUNCTION(base_atomic(base_atomic const&)) + BOOST_DELETED_FUNCTION(base_atomic& operator=(base_atomic const&)) +}; + + +//! Implementation for pointers +template< typename T > +class base_atomic< T*, void* > +{ +private: + typedef T* value_type; + typedef std::ptrdiff_t difference_type; + typedef atomics::detail::operations< storage_size_of< value_type >::value, false > operations; + +protected: + typedef value_type value_arg_type; + +public: + typedef typename operations::storage_type storage_type; + +protected: + storage_type m_storage; + +public: + BOOST_DEFAULTED_FUNCTION(base_atomic(), {}) + BOOST_FORCEINLINE explicit base_atomic(value_type const& v) BOOST_NOEXCEPT : m_storage(atomics::detail::union_cast< storage_type >(v)) + { + } + + BOOST_FORCEINLINE void store(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_consume); + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + + operations::store(m_storage, atomics::detail::union_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE value_type load(memory_order order = memory_order_seq_cst) const volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_release); + BOOST_ASSERT(order != memory_order_acq_rel); + + return atomics::detail::union_cast< value_type >(operations::load(m_storage, order)); + } + + BOOST_FORCEINLINE value_type fetch_add(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::fetch_add(m_storage, static_cast< storage_type >(v * sizeof(T)), order)); + } + + BOOST_FORCEINLINE value_type fetch_sub(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::fetch_sub(m_storage, static_cast< storage_type >(v * sizeof(T)), order)); + } + + BOOST_FORCEINLINE value_type exchange(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::exchange(m_storage, atomics::detail::union_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::union_cast< storage_type >(expected); + const bool res = operations::compare_exchange_strong(m_storage, old_value, atomics::detail::union_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::union_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_strong(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::union_cast< storage_type >(expected); + const bool res = operations::compare_exchange_weak(m_storage, old_value, atomics::detail::union_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::union_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_weak(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool is_lock_free() const volatile BOOST_NOEXCEPT + { + return operations::is_lock_free(m_storage); + } + + BOOST_FORCEINLINE value_type operator++(int) volatile BOOST_NOEXCEPT + { + return fetch_add(1); + } + + BOOST_FORCEINLINE value_type operator++() volatile BOOST_NOEXCEPT + { + return fetch_add(1) + 1; + } + + BOOST_FORCEINLINE value_type operator--(int) volatile BOOST_NOEXCEPT + { + return fetch_sub(1); + } + + BOOST_FORCEINLINE value_type operator--() volatile BOOST_NOEXCEPT + { + return fetch_sub(1) - 1; + } + + BOOST_FORCEINLINE value_type operator+=(difference_type v) volatile BOOST_NOEXCEPT + { + return fetch_add(v) + v; + } + + BOOST_FORCEINLINE value_type operator-=(difference_type v) volatile BOOST_NOEXCEPT + { + return fetch_sub(v) - v; + } + + BOOST_DELETED_FUNCTION(base_atomic(base_atomic const&)) + BOOST_DELETED_FUNCTION(base_atomic& operator=(base_atomic const&)) +}; + + +//! Implementation for void pointers +template< > +class base_atomic< void*, void* > +{ +private: + typedef void* value_type; + typedef std::ptrdiff_t difference_type; + typedef atomics::detail::operations< storage_size_of< value_type >::value, false > operations; + +protected: + typedef value_type value_arg_type; + +public: + typedef operations::storage_type storage_type; + +protected: + storage_type m_storage; + +public: + BOOST_DEFAULTED_FUNCTION(base_atomic(), {}) + BOOST_FORCEINLINE explicit base_atomic(value_type const& v) BOOST_NOEXCEPT : m_storage(atomics::detail::union_cast< storage_type >(v)) + { + } + + BOOST_FORCEINLINE void store(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_consume); + BOOST_ASSERT(order != memory_order_acquire); + BOOST_ASSERT(order != memory_order_acq_rel); + + operations::store(m_storage, atomics::detail::union_cast< storage_type >(v), order); + } + + BOOST_FORCEINLINE value_type load(memory_order order = memory_order_seq_cst) const volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(order != memory_order_release); + BOOST_ASSERT(order != memory_order_acq_rel); + + return atomics::detail::union_cast< value_type >(operations::load(m_storage, order)); + } + + BOOST_FORCEINLINE value_type fetch_add(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::fetch_add(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type fetch_sub(difference_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::fetch_sub(m_storage, static_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE value_type exchange(value_type v, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return atomics::detail::union_cast< value_type >(operations::exchange(m_storage, atomics::detail::union_cast< storage_type >(v), order)); + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::union_cast< storage_type >(expected); + const bool res = operations::compare_exchange_strong(m_storage, old_value, atomics::detail::union_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::union_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_strong(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_strong(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order success_order, memory_order failure_order) volatile BOOST_NOEXCEPT + { + BOOST_ASSERT(failure_order != memory_order_release); + BOOST_ASSERT(failure_order != memory_order_acq_rel); + BOOST_ASSERT(cas_failure_order_must_not_be_stronger_than_success_order(success_order, failure_order)); + + storage_type old_value = atomics::detail::union_cast< storage_type >(expected); + const bool res = operations::compare_exchange_weak(m_storage, old_value, atomics::detail::union_cast< storage_type >(desired), success_order, failure_order); + expected = atomics::detail::union_cast< value_type >(old_value); + return res; + } + + BOOST_FORCEINLINE bool compare_exchange_weak(value_type& expected, value_type desired, memory_order order = memory_order_seq_cst) volatile BOOST_NOEXCEPT + { + return compare_exchange_weak(expected, desired, order, atomics::detail::deduce_failure_order(order)); + } + + BOOST_FORCEINLINE bool is_lock_free() const volatile BOOST_NOEXCEPT + { + return operations::is_lock_free(m_storage); + } + + BOOST_FORCEINLINE value_type operator++(int) volatile BOOST_NOEXCEPT + { + return fetch_add(1); + } + + BOOST_FORCEINLINE value_type operator++() volatile BOOST_NOEXCEPT + { + return (char*)fetch_add(1) + 1; + } + + BOOST_FORCEINLINE value_type operator--(int) volatile BOOST_NOEXCEPT + { + return fetch_sub(1); + } + + BOOST_FORCEINLINE value_type operator--() volatile BOOST_NOEXCEPT + { + return (char*)fetch_sub(1) - 1; + } + + BOOST_FORCEINLINE value_type operator+=(difference_type v) volatile BOOST_NOEXCEPT + { + return (char*)fetch_add(v) + v; + } + + BOOST_FORCEINLINE value_type operator-=(difference_type v) volatile BOOST_NOEXCEPT + { + return (char*)fetch_sub(v) - v; + } + + BOOST_DELETED_FUNCTION(base_atomic(base_atomic const&)) + BOOST_DELETED_FUNCTION(base_atomic& operator=(base_atomic const&)) +}; + +} // namespace detail + +template< typename T > +class atomic : + public atomics::detail::base_atomic< T, typename atomics::detail::classify< T >::type > +{ +private: + typedef T value_type; + typedef atomics::detail::base_atomic< T, typename atomics::detail::classify< T >::type > base_type; + typedef typename base_type::value_arg_type value_arg_type; + +public: + typedef typename base_type::storage_type storage_type; + +public: + BOOST_DEFAULTED_FUNCTION(atomic(), BOOST_NOEXCEPT {}) + + // NOTE: The constructor is made explicit because gcc 4.7 complains that + // operator=(value_arg_type) is considered ambiguous with operator=(atomic const&) + // in assignment expressions, even though conversion to atomic<> is less preferred + // than conversion to value_arg_type. + BOOST_FORCEINLINE explicit BOOST_CONSTEXPR atomic(value_arg_type v) BOOST_NOEXCEPT : base_type(v) {} + + BOOST_FORCEINLINE value_type operator= (value_arg_type v) volatile BOOST_NOEXCEPT + { + this->store(v); + return v; + } + + BOOST_FORCEINLINE operator value_type() volatile const BOOST_NOEXCEPT + { + return this->load(); + } + + BOOST_FORCEINLINE storage_type& storage() BOOST_NOEXCEPT { return this->m_storage; } + BOOST_FORCEINLINE storage_type volatile& storage() volatile BOOST_NOEXCEPT { return this->m_storage; } + BOOST_FORCEINLINE storage_type const& storage() const BOOST_NOEXCEPT { return this->m_storage; } + BOOST_FORCEINLINE storage_type const volatile& storage() const volatile BOOST_NOEXCEPT { return this->m_storage; } + + BOOST_DELETED_FUNCTION(atomic(atomic const&)) + BOOST_DELETED_FUNCTION(atomic& operator= (atomic const&)) + BOOST_DELETED_FUNCTION(atomic& operator= (atomic const&) volatile) +}; + +typedef atomic< char > atomic_char; +typedef atomic< unsigned char > atomic_uchar; +typedef atomic< signed char > atomic_schar; +typedef atomic< uint8_t > atomic_uint8_t; +typedef atomic< int8_t > atomic_int8_t; +typedef atomic< unsigned short > atomic_ushort; +typedef atomic< short > atomic_short; +typedef atomic< uint16_t > atomic_uint16_t; +typedef atomic< int16_t > atomic_int16_t; +typedef atomic< unsigned int > atomic_uint; +typedef atomic< int > atomic_int; +typedef atomic< uint32_t > atomic_uint32_t; +typedef atomic< int32_t > atomic_int32_t; +typedef atomic< unsigned long > atomic_ulong; +typedef atomic< long > atomic_long; +typedef atomic< uint64_t > atomic_uint64_t; +typedef atomic< int64_t > atomic_int64_t; +#ifdef BOOST_HAS_LONG_LONG +typedef atomic< boost::ulong_long_type > atomic_ullong; +typedef atomic< boost::long_long_type > atomic_llong; +#endif +typedef atomic< void* > atomic_address; +typedef atomic< bool > atomic_bool; +typedef atomic< wchar_t > atomic_wchar_t; +#if !defined(BOOST_NO_CXX11_CHAR16_T) +typedef atomic< char16_t > atomic_char16_t; +#endif +#if !defined(BOOST_NO_CXX11_CHAR32_T) +typedef atomic< char32_t > atomic_char32_t; +#endif + +typedef atomic< int_least8_t > atomic_int_least8_t; +typedef atomic< uint_least8_t > atomic_uint_least8_t; +typedef atomic< int_least16_t > atomic_int_least16_t; +typedef atomic< uint_least16_t > atomic_uint_least16_t; +typedef atomic< int_least32_t > atomic_int_least32_t; +typedef atomic< uint_least32_t > atomic_uint_least32_t; +typedef atomic< int_least64_t > atomic_int_least64_t; +typedef atomic< uint_least64_t > atomic_uint_least64_t; +typedef atomic< int_fast8_t > atomic_int_fast8_t; +typedef atomic< uint_fast8_t > atomic_uint_fast8_t; +typedef atomic< int_fast16_t > atomic_int_fast16_t; +typedef atomic< uint_fast16_t > atomic_uint_fast16_t; +typedef atomic< int_fast32_t > atomic_int_fast32_t; +typedef atomic< uint_fast32_t > atomic_uint_fast32_t; +typedef atomic< int_fast64_t > atomic_int_fast64_t; +typedef atomic< uint_fast64_t > atomic_uint_fast64_t; +typedef atomic< intmax_t > atomic_intmax_t; +typedef atomic< uintmax_t > atomic_uintmax_t; + +typedef atomic< std::size_t > atomic_size_t; +typedef atomic< std::ptrdiff_t > atomic_ptrdiff_t; + +#if defined(BOOST_HAS_INTPTR_T) +typedef atomic< intptr_t > atomic_intptr_t; +typedef atomic< uintptr_t > atomic_uintptr_t; +#endif + +} // namespace atomics +} // namespace boost + +#if defined(BOOST_MSVC) +#pragma warning(pop) +#endif + +#endif // BOOST_ATOMIC_DETAIL_ATOMIC_TEMPLATE_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_alpha.hpp b/boost/atomic/detail/caps_gcc_alpha.hpp new file mode 100644 index 0000000000..861432f58a --- /dev/null +++ b/boost/atomic/detail/caps_gcc_alpha.hpp @@ -0,0 +1,34 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_alpha.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_ALPHA_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_ALPHA_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_ALPHA_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_arm.hpp b/boost/atomic/detail/caps_gcc_arm.hpp new file mode 100644 index 0000000000..b827c648d7 --- /dev/null +++ b/boost/atomic/detail/caps_gcc_arm.hpp @@ -0,0 +1,56 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2009 Phil Endecott + * Copyright (c) 2013 Tim Blechmann + * ARM Code by Phil Endecott, based on other architectures. + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_arm.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_ARM_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if !(defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6ZK__)) +// ARMv7 and later have dmb instruction +#define BOOST_ATOMIC_DETAIL_ARM_HAS_DMB 1 +#endif + +#if !(defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6Z__)) +// ARMv6k and ARMv7 have 8 and 16 ldrex/strex variants +#define BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXB_STREXB 1 +#define BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXH_STREXH 1 +#if !(((defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6ZK__)) && defined(__thumb__)) || defined(__ARM_ARCH_7M__)) +// ARMv6k and ARMv7 except ARMv7-M have 64-bit ldrex/strex variants. +// Unfortunately, GCC (at least 4.7.3 on Ubuntu) does not allocate register pairs properly when targeting ARMv6k Thumb, +// which is required for ldrexd/strexd instructions, so we disable 64-bit support. When targeting ARMv6k ARM +// or ARMv7 (both ARM and Thumb 2) it works as expected. +#define BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD 1 +#endif +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#if defined(BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD) +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#endif +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_atomic.hpp b/boost/atomic/detail/caps_gcc_atomic.hpp new file mode 100644 index 0000000000..f4e7a7023e --- /dev/null +++ b/boost/atomic/detail/caps_gcc_atomic.hpp @@ -0,0 +1,134 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_atomic.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_ATOMIC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_ATOMIC_HPP_INCLUDED_ + +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(__i386__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B 1 +#endif + +#if defined(__x86_64__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B 1 +#endif + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) && (defined(BOOST_HAS_INT128) || !defined(BOOST_NO_ALIGNMENT)) +#define BOOST_ATOMIC_INT128_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_INT128_LOCK_FREE 0 +#endif + +#if __GCC_ATOMIC_LLONG_LOCK_FREE == 2 +#define BOOST_ATOMIC_LLONG_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_LLONG_LOCK_FREE BOOST_ATOMIC_INT128_LOCK_FREE +#endif + +#if __GCC_ATOMIC_LONG_LOCK_FREE == 2 +#define BOOST_ATOMIC_LONG_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_LONG_LOCK_FREE BOOST_ATOMIC_LLONG_LOCK_FREE +#endif + +#if __GCC_ATOMIC_INT_LOCK_FREE == 2 +#define BOOST_ATOMIC_INT_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_INT_LOCK_FREE BOOST_ATOMIC_LONG_LOCK_FREE +#endif + +#if __GCC_ATOMIC_SHORT_LOCK_FREE == 2 +#define BOOST_ATOMIC_SHORT_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_SHORT_LOCK_FREE BOOST_ATOMIC_INT_LOCK_FREE +#endif + +#if __GCC_ATOMIC_CHAR_LOCK_FREE == 2 +#define BOOST_ATOMIC_CHAR_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_CHAR_LOCK_FREE BOOST_ATOMIC_SHORT_LOCK_FREE +#endif + +#if __GCC_ATOMIC_POINTER_LOCK_FREE == 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 +#else +#define BOOST_ATOMIC_POINTER_LOCK_FREE 0 +#endif + + +#define BOOST_ATOMIC_INT8_LOCK_FREE BOOST_ATOMIC_CHAR_LOCK_FREE + +#if BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE BOOST_ATOMIC_SHORT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_INT == 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE BOOST_ATOMIC_INT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE BOOST_ATOMIC_LONG_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE BOOST_ATOMIC_LLONG_LOCK_FREE +#else +#define BOOST_ATOMIC_INT16_LOCK_FREE 0 +#endif + +#if BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 4 +#define BOOST_ATOMIC_INT32_LOCK_FREE BOOST_ATOMIC_SHORT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_INT == 4 +#define BOOST_ATOMIC_INT32_LOCK_FREE BOOST_ATOMIC_INT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 4 +#define BOOST_ATOMIC_INT32_LOCK_FREE BOOST_ATOMIC_LONG_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 4 +#define BOOST_ATOMIC_INT32_LOCK_FREE BOOST_ATOMIC_LLONG_LOCK_FREE +#else +#define BOOST_ATOMIC_INT32_LOCK_FREE 0 +#endif + +#if BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 8 +#define BOOST_ATOMIC_INT64_LOCK_FREE BOOST_ATOMIC_SHORT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_INT == 8 +#define BOOST_ATOMIC_INT64_LOCK_FREE BOOST_ATOMIC_INT_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 8 +#define BOOST_ATOMIC_INT64_LOCK_FREE BOOST_ATOMIC_LONG_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 8 +#define BOOST_ATOMIC_INT64_LOCK_FREE BOOST_ATOMIC_LLONG_LOCK_FREE +#else +#define BOOST_ATOMIC_INT64_LOCK_FREE 0 +#endif + + +#if __GCC_ATOMIC_WCHAR_T_LOCK_FREE == 2 +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE 2 +#elif BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 8 +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE BOOST_ATOMIC_INT64_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 4 +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE BOOST_ATOMIC_INT32_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 2 +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE BOOST_ATOMIC_INT16_LOCK_FREE +#elif BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 1 +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE BOOST_ATOMIC_INT8_LOCK_FREE +#else +#define BOOST_ATOMIC_WCHAR_T_LOCK_FREE 0 +#endif + +#define BOOST_ATOMIC_CHAR32_T_LOCK_FREE BOOST_ATOMIC_INT32_LOCK_FREE +#define BOOST_ATOMIC_CHAR16_T_LOCK_FREE BOOST_ATOMIC_INT16_LOCK_FREE + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_ATOMIC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_ppc.hpp b/boost/atomic/detail/caps_gcc_ppc.hpp new file mode 100644 index 0000000000..6dbdde826d --- /dev/null +++ b/boost/atomic/detail/caps_gcc_ppc.hpp @@ -0,0 +1,36 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_ppc.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_PPC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_PPC_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#if defined(__powerpc64__) +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#endif +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_PPC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_sparc.hpp b/boost/atomic/detail/caps_gcc_sparc.hpp new file mode 100644 index 0000000000..5806684926 --- /dev/null +++ b/boost/atomic/detail/caps_gcc_sparc.hpp @@ -0,0 +1,34 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2010 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_sparc.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_SPARC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_SPARC_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_SPARC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_sync.hpp b/boost/atomic/detail/caps_gcc_sync.hpp new file mode 100644 index 0000000000..7fac07a130 --- /dev/null +++ b/boost/atomic/detail/caps_gcc_sync.hpp @@ -0,0 +1,62 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2011 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_sync.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_SYNC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_SYNC_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(__i386__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B 1 +#endif + +#if defined(__x86_64__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B 1 +#endif + +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#endif +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#endif +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#endif +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)\ + || defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#endif +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_INT128_LOCK_FREE 2 +#endif + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_SYNC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_gcc_x86.hpp b/boost/atomic/detail/caps_gcc_x86.hpp new file mode 100644 index 0000000000..0696bf1d2f --- /dev/null +++ b/boost/atomic/detail/caps_gcc_x86.hpp @@ -0,0 +1,52 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2013 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_gcc_x86.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_GCC_X86_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_GCC_X86_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(__i386__) &&\ + (\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) ||\ + defined(__i586__) || defined(__i686__) || defined(__pentium4__) || defined(__nocona__) || defined(__core2__) || defined(__corei7__) ||\ + defined(__k6__) || defined(__athlon__) || defined(__k8__) || defined(__amdfam10__) || defined(__bdver1__) || defined(__bdver2__) || defined(__bdver3__) || defined(__btver1__) || defined(__btver2__)\ + ) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B 1 +#endif + +#if defined(__x86_64__) && defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B 1 +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#if defined(__x86_64__) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#endif +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) && (defined(BOOST_HAS_INT128) || !defined(BOOST_NO_ALIGNMENT)) +#define BOOST_ATOMIC_INT128_LOCK_FREE 2 +#endif +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_GCC_X86_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_linux_arm.hpp b/boost/atomic/detail/caps_linux_arm.hpp new file mode 100644 index 0000000000..abe6fb81af --- /dev/null +++ b/boost/atomic/detail/caps_linux_arm.hpp @@ -0,0 +1,35 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009, 2011 Helge Bahmann + * Copyright (c) 2009 Phil Endecott + * Copyright (c) 2013 Tim Blechmann + * Linux-specific code by Phil Endecott + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_linux_arm.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_LINUX_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_LINUX_ARM_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_LINUX_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_msvc_arm.hpp b/boost/atomic/detail/caps_msvc_arm.hpp new file mode 100644 index 0000000000..6b3c61fb3e --- /dev/null +++ b/boost/atomic/detail/caps_msvc_arm.hpp @@ -0,0 +1,34 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2012 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_msvc_arm.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_MSVC_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_MSVC_ARM_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_MSVC_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_msvc_x86.hpp b/boost/atomic/detail/caps_msvc_x86.hpp new file mode 100644 index 0000000000..5661a5b7ab --- /dev/null +++ b/boost/atomic/detail/caps_msvc_x86.hpp @@ -0,0 +1,50 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2012 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_msvc_x86.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_MSVC_X86_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_MSVC_X86_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(_M_IX86) && _M_IX86 >= 500 +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B 1 +#endif + +#if _MSC_VER >= 1500 && defined(_M_AMD64) && !defined(BOOST_ATOMIC_NO_CMPXCHG16B) +#define BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B 1 +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 + +#if defined(_M_AMD64) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) +#define BOOST_ATOMIC_INT64_LOCK_FREE 2 +#endif + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) && (defined(BOOST_HAS_INT128) || !defined(BOOST_NO_ALIGNMENT)) +#define BOOST_ATOMIC_INT128_LOCK_FREE 2 +#endif + +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_MSVC_X86_HPP_INCLUDED_ diff --git a/boost/atomic/detail/caps_windows.hpp b/boost/atomic/detail/caps_windows.hpp new file mode 100644 index 0000000000..1cc0ded833 --- /dev/null +++ b/boost/atomic/detail/caps_windows.hpp @@ -0,0 +1,33 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2012 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/caps_windows.hpp + * + * This header defines feature capabilities macros + */ + +#ifndef BOOST_ATOMIC_DETAIL_CAPS_WINDOWS_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CAPS_WINDOWS_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_INT8_LOCK_FREE 2 +#define BOOST_ATOMIC_INT16_LOCK_FREE 2 +#define BOOST_ATOMIC_INT32_LOCK_FREE 2 +#define BOOST_ATOMIC_POINTER_LOCK_FREE 2 + +#define BOOST_ATOMIC_THREAD_FENCE 2 +#define BOOST_ATOMIC_SIGNAL_FENCE 2 + +#endif // BOOST_ATOMIC_DETAIL_CAPS_WINDOWS_HPP_INCLUDED_ diff --git a/boost/atomic/detail/casts.hpp b/boost/atomic/detail/casts.hpp new file mode 100644 index 0000000000..db28bc25ff --- /dev/null +++ b/boost/atomic/detail/casts.hpp @@ -0,0 +1,64 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2013 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/casts.hpp + * + * This header defines \c union_cast and \c memcpy_cast used to convert between storage and value types + */ + +#ifndef BOOST_ATOMIC_DETAIL_CASTS_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CASTS_HPP_INCLUDED_ + +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< typename To, typename From > +BOOST_FORCEINLINE To union_cast(From const& from) BOOST_NOEXCEPT +{ + union + { + To as_to; + From as_from; + } + caster = {}; + caster.as_from = from; + return caster.as_to; +} + +template< typename To, typename From > +BOOST_FORCEINLINE To memcpy_cast(From const& from) BOOST_NOEXCEPT +{ + struct + { + To to; + } + value = {}; + std::memcpy + ( + &reinterpret_cast< char& >(value.to), + &reinterpret_cast< const char& >(from), + (sizeof(From) < sizeof(To) ? sizeof(From) : sizeof(To)) + ); + return value.to; +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_CASTS_HPP_INCLUDED_ diff --git a/boost/atomic/detail/config.hpp b/boost/atomic/detail/config.hpp new file mode 100644 index 0000000000..d03ec6aec6 --- /dev/null +++ b/boost/atomic/detail/config.hpp @@ -0,0 +1,24 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2012 Hartmut Kaiser + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/config.hpp + * + * This header defines configuraion macros for Boost.Atomic + */ + +#ifndef BOOST_ATOMIC_DETAIL_CONFIG_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_CONFIG_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#endif // BOOST_ATOMIC_DETAIL_CONFIG_HPP_INCLUDED_ diff --git a/boost/atomic/detail/int_sizes.hpp b/boost/atomic/detail/int_sizes.hpp new file mode 100644 index 0000000000..d06ed42a11 --- /dev/null +++ b/boost/atomic/detail/int_sizes.hpp @@ -0,0 +1,140 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/int_sizes.hpp + * + * This header defines macros for testing buitin integer type sizes + */ + +#ifndef BOOST_ATOMIC_DETAIL_INT_SIZES_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_INT_SIZES_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +// GCC and compatible compilers define internal macros with builtin type traits +#if defined(__SIZEOF_SHORT__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_SHORT __SIZEOF_SHORT__ +#endif +#if defined(__SIZEOF_INT__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_INT __SIZEOF_INT__ +#endif +#if defined(__SIZEOF_LONG__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_LONG __SIZEOF_LONG__ +#endif +#if defined(__SIZEOF_LONG_LONG__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG __SIZEOF_LONG_LONG__ +#endif +#if defined(__SIZEOF_WCHAR_T__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T __SIZEOF_WCHAR_T__ +#endif +#if defined(__SIZEOF_POINTER__) +#define BOOST_ATOMIC_DETAIL_SIZEOF_POINTER __SIZEOF_POINTER__ +#elif defined(_MSC_VER) +#if defined(_M_AMD64) || defined(_M_IA64) +#define BOOST_ATOMIC_DETAIL_SIZEOF_POINTER 8 +#else +#define BOOST_ATOMIC_DETAIL_SIZEOF_POINTER 4 +#endif +#endif + +#if !defined(BOOST_ATOMIC_DETAIL_SIZEOF_SHORT) || !defined(BOOST_ATOMIC_DETAIL_SIZEOF_INT) ||\ + !defined(BOOST_ATOMIC_DETAIL_SIZEOF_LONG) || !defined(BOOST_ATOMIC_DETAIL_SIZEOF_LLONG) + +// Try to deduce sizes from limits +#include +#include + +#if (USHRT_MAX + 0) == 0xff +#define BOOST_ATOMIC_DETAIL_SIZEOF_SHORT 1 +#elif (USHRT_MAX + 0) == 0xffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_SHORT 2 +#elif (USHRT_MAX + 0) == 0xffffffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_SHORT 4 +#elif (USHRT_MAX + 0) == UINT64_C(0xffffffffffffffff) +#define BOOST_ATOMIC_DETAIL_SIZEOF_SHORT 8 +#endif + +#if (UINT_MAX + 0) == 0xff +#define BOOST_ATOMIC_DETAIL_SIZEOF_INT 1 +#elif (UINT_MAX + 0) == 0xffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_INT 2 +#elif (UINT_MAX + 0) == 0xffffffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_INT 4 +#elif (UINT_MAX + 0) == UINT64_C(0xffffffffffffffff) +#define BOOST_ATOMIC_DETAIL_SIZEOF_INT 8 +#endif + +#if (ULONG_MAX + 0) == 0xff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LONG 1 +#elif (ULONG_MAX + 0) == 0xffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LONG 2 +#elif (ULONG_MAX + 0) == 0xffffffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LONG 4 +#elif (ULONG_MAX + 0) == UINT64_C(0xffffffffffffffff) +#define BOOST_ATOMIC_DETAIL_SIZEOF_LONG 8 +#endif + +#if defined(__hpux) // HP-UX's value of ULONG_LONG_MAX is unusable in preprocessor expressions +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG 8 +#else + +// The list of the non-standard macros (the ones except ULLONG_MAX) is taken from cstdint.hpp +#if defined(ULLONG_MAX) +#define BOOST_ATOMIC_DETAIL_ULLONG_MAX ULLONG_MAX +#elif defined(ULONG_LONG_MAX) +#define BOOST_ATOMIC_DETAIL_ULLONG_MAX ULONG_LONG_MAX +#elif defined(ULONGLONG_MAX) +#define BOOST_ATOMIC_DETAIL_ULLONG_MAX ULONGLONG_MAX +#elif defined(_LLONG_MAX) // strangely enough, this one seems to be holding the limit for the unsigned integer +#define BOOST_ATOMIC_DETAIL_ULLONG_MAX _LLONG_MAX +#endif + +#if (BOOST_ATOMIC_DETAIL_ULLONG_MAX + 0) == 0xff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG 1 +#elif (BOOST_ATOMIC_DETAIL_ULLONG_MAX + 0) == 0xffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG 2 +#elif (BOOST_ATOMIC_DETAIL_ULLONG_MAX + 0) == 0xffffffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG 4 +#elif (BOOST_ATOMIC_DETAIL_ULLONG_MAX + 0) == UINT64_C(0xffffffffffffffff) +#define BOOST_ATOMIC_DETAIL_SIZEOF_LLONG 8 +#endif + +#endif // defined(__hpux) + +#endif + +#if !defined(BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T) + +#include +#include + +#if defined(_MSC_VER) && _MSC_VER <= 1310 +// MSVC 7.1 defines WCHAR_MAX to a value not suitable for constant expressions +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T 2 +#elif (WCHAR_MAX + 0) == 0xff || (WCHAR_MAX + 0) == 0x7f +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T 1 +#elif (WCHAR_MAX + 0) == 0xffff || (WCHAR_MAX + 0) == 0x7fff +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T 2 +#elif (WCHAR_MAX + 0) == 0xffffffff || (WCHAR_MAX + 0) == 0x7fffffff +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T 4 +#elif (WCHAR_MAX + 0) == UINT64_C(0xffffffffffffffff) || (WCHAR_MAX + 0) == INT64_C(0x7fffffffffffffff) +#define BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T 8 +#endif +#endif + +#if !defined(BOOST_ATOMIC_DETAIL_SIZEOF_SHORT) || !defined(BOOST_ATOMIC_DETAIL_SIZEOF_INT) ||\ + !defined(BOOST_ATOMIC_DETAIL_SIZEOF_LONG) || !defined(BOOST_ATOMIC_DETAIL_SIZEOF_LLONG) ||\ + !defined(BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T) +#error Boost.Atomic: Failed to determine builtin integer sizes, the target platform is not supported. Please, report to the developers. +#endif + +#endif // BOOST_ATOMIC_DETAIL_INT_SIZES_HPP_INCLUDED_ diff --git a/boost/atomic/detail/interlocked.hpp b/boost/atomic/detail/interlocked.hpp new file mode 100644 index 0000000000..1c62396bf5 --- /dev/null +++ b/boost/atomic/detail/interlocked.hpp @@ -0,0 +1,481 @@ +#ifndef BOOST_ATOMIC_DETAIL_INTERLOCKED_HPP +#define BOOST_ATOMIC_DETAIL_INTERLOCKED_HPP + +// Copyright (c) 2009 Helge Bahmann +// Copyright (c) 2012 - 2014 Andrey Semashev +// +// Distributed under the Boost Software License, Version 1.0. +// See accompanying file LICENSE_1_0.txt or copy at +// http://www.boost.org/LICENSE_1_0.txt) + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(_WIN32_WCE) + +#if _WIN32_WCE >= 0x600 + +extern "C" long __cdecl _InterlockedCompareExchange( long volatile *, long, long ); +extern "C" long __cdecl _InterlockedExchangeAdd( long volatile *, long ); +extern "C" long __cdecl _InterlockedExchange( long volatile *, long ); + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) _InterlockedCompareExchange((long*)(dest), exchange, compare) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) _InterlockedExchangeAdd((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) _InterlockedExchange((long*)(dest), (long)(newval)) + +#else // _WIN32_WCE >= 0x600 + +extern "C" long __cdecl InterlockedCompareExchange( long*, long, long ); +extern "C" long __cdecl InterlockedExchangeAdd( long*, long ); +extern "C" long __cdecl InterlockedExchange( long*, long ); + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) InterlockedCompareExchange((long*)(dest), exchange, compare) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) InterlockedExchangeAdd((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) InterlockedExchange((long*)(dest), (long)(newval)) + +#endif // _WIN32_WCE >= 0x600 + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) ((void*)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE((long*)(dest), (long)(exchange), (long)(compare))) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, exchange) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE((long*)(dest), (long)(exchange))) + +#elif defined(_MSC_VER) && _MSC_VER >= 1310 + +#if _MSC_VER < 1400 + +extern "C" long __cdecl _InterlockedCompareExchange( long volatile *, long, long ); +extern "C" long __cdecl _InterlockedExchangeAdd( long volatile *, long ); +extern "C" long __cdecl _InterlockedExchange( long volatile *, long ); + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange) +#pragma intrinsic(_InterlockedExchangeAdd) +#pragma intrinsic(_InterlockedExchange) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) _InterlockedCompareExchange((long*)(dest), exchange, compare) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) _InterlockedExchangeAdd((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) _InterlockedExchange((long*)(dest), (long)(newval)) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) ((void*)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE((long*)(dest), (long)(exchange), (long)(compare))) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, exchange) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE((long*)(dest), (long)(exchange))) + +#else // _MSC_VER < 1400 + +#include + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange) +#pragma intrinsic(_InterlockedExchangeAdd) +#pragma intrinsic(_InterlockedExchange) +#pragma intrinsic(_InterlockedAnd) +#pragma intrinsic(_InterlockedOr) +#pragma intrinsic(_InterlockedXor) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) _InterlockedCompareExchange((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) _InterlockedExchangeAdd((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) _InterlockedExchange((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_AND(dest, arg) _InterlockedAnd((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR(dest, arg) _InterlockedOr((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR(dest, arg) _InterlockedXor((long*)(dest), (long)(arg)) + +#if (defined(_M_IX86) && _M_IX86 >= 500) || defined(_M_AMD64) || defined(_M_IA64) +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange64) +#endif +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(dest, exchange, compare) _InterlockedCompareExchange64((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#endif + +#if _MSC_VER >= 1500 && defined(_M_AMD64) +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange128) +#endif +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(dest, exchange, compare) _InterlockedCompareExchange128((__int64*)(dest), ((const __int64*)(&exchange))[1], ((const __int64*)(&exchange))[0], (__int64*)(compare)) +#endif + +#if _MSC_VER >= 1600 + +// MSVC 2010 and later provide intrinsics for 8 and 16 bit integers. +// Note that for each bit count these macros must be either all defined or all not defined. +// Otherwise atomic<> operations will be implemented inconsistently. + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange8) +#pragma intrinsic(_InterlockedExchangeAdd8) +#pragma intrinsic(_InterlockedExchange8) +#pragma intrinsic(_InterlockedAnd8) +#pragma intrinsic(_InterlockedOr8) +#pragma intrinsic(_InterlockedXor8) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8(dest, exchange, compare) _InterlockedCompareExchange8((char*)(dest), (char)(exchange), (char)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8(dest, addend) _InterlockedExchangeAdd8((char*)(dest), (char)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE8(dest, newval) _InterlockedExchange8((char*)(dest), (char)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_AND8(dest, arg) _InterlockedAnd8((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR8(dest, arg) _InterlockedOr8((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR8(dest, arg) _InterlockedXor8((char*)(dest), (char)(arg)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange16) +#pragma intrinsic(_InterlockedExchangeAdd16) +#pragma intrinsic(_InterlockedExchange16) +#pragma intrinsic(_InterlockedAnd16) +#pragma intrinsic(_InterlockedOr16) +#pragma intrinsic(_InterlockedXor16) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16(dest, exchange, compare) _InterlockedCompareExchange16((short*)(dest), (short)(exchange), (short)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16(dest, addend) _InterlockedExchangeAdd16((short*)(dest), (short)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE16(dest, newval) _InterlockedExchange16((short*)(dest), (short)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_AND16(dest, arg) _InterlockedAnd16((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR16(dest, arg) _InterlockedOr16((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR16(dest, arg) _InterlockedXor16((short*)(dest), (short)(arg)) + +#endif // _MSC_VER >= 1600 + +#if defined(_M_AMD64) || defined(_M_IA64) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedExchangeAdd64) +#pragma intrinsic(_InterlockedExchange64) +#pragma intrinsic(_InterlockedAnd64) +#pragma intrinsic(_InterlockedOr64) +#pragma intrinsic(_InterlockedXor64) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, addend) _InterlockedExchangeAdd64((__int64*)(dest), (__int64)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(dest, newval) _InterlockedExchange64((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_AND64(dest, arg) _InterlockedAnd64((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR64(dest, arg) _InterlockedOr64((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR64(dest, arg) _InterlockedXor64((__int64*)(dest), (__int64)(arg)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchangePointer) +#pragma intrinsic(_InterlockedExchangePointer) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) _InterlockedCompareExchangePointer((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) _InterlockedExchangePointer((void**)(dest), (void*)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64((long*)(dest), byte_offset)) + +#elif defined(_M_IX86) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) ((void*)_InterlockedCompareExchange((long*)(dest), (long)(exchange), (long)(compare))) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) ((void*)_InterlockedExchange((long*)(dest), (long)(newval))) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD((long*)(dest), byte_offset)) + +#endif + +#if _MSC_VER >= 1700 && defined(_M_ARM) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedExchangeAdd64) +#pragma intrinsic(_InterlockedExchange64) +#pragma intrinsic(_InterlockedAnd64) +#pragma intrinsic(_InterlockedOr64) +#pragma intrinsic(_InterlockedXor64) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, addend) _InterlockedExchangeAdd64((__int64*)(dest), (__int64)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(dest, newval) _InterlockedExchange64((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_AND64(dest, arg) _InterlockedAnd64((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR64(dest, arg) _InterlockedOr64((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR64(dest, arg) _InterlockedXor64((__int64*)(dest), (__int64)(arg)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedCompareExchange8_nf) +#pragma intrinsic(_InterlockedCompareExchange8_acq) +#pragma intrinsic(_InterlockedCompareExchange8_rel) +#pragma intrinsic(_InterlockedCompareExchange16_nf) +#pragma intrinsic(_InterlockedCompareExchange16_acq) +#pragma intrinsic(_InterlockedCompareExchange16_rel) +#pragma intrinsic(_InterlockedCompareExchange_nf) +#pragma intrinsic(_InterlockedCompareExchange_acq) +#pragma intrinsic(_InterlockedCompareExchange_rel) +#pragma intrinsic(_InterlockedCompareExchange64) +#pragma intrinsic(_InterlockedCompareExchange64_nf) +#pragma intrinsic(_InterlockedCompareExchange64_acq) +#pragma intrinsic(_InterlockedCompareExchange64_rel) +#pragma intrinsic(_InterlockedCompareExchangePointer) +#pragma intrinsic(_InterlockedCompareExchangePointer_nf) +#pragma intrinsic(_InterlockedCompareExchangePointer_acq) +#pragma intrinsic(_InterlockedCompareExchangePointer_rel) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_RELAXED(dest, exchange, compare) _InterlockedCompareExchange8_nf((char*)(dest), (char)(exchange), (char)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_ACQUIRE(dest, exchange, compare) _InterlockedCompareExchange8_acq((char*)(dest), (char)(exchange), (char)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_RELEASE(dest, exchange, compare) _InterlockedCompareExchange8_rel((char*)(dest), (char)(exchange), (char)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_RELAXED(dest, exchange, compare) _InterlockedCompareExchange16_nf((short*)(dest), (short)(exchange), (short)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_ACQUIRE(dest, exchange, compare) _InterlockedCompareExchange16_acq((short*)(dest), (short)(exchange), (short)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_RELEASE(dest, exchange, compare) _InterlockedCompareExchange16_rel((short*)(dest), (short)(exchange), (short)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_RELAXED(dest, exchange, compare) _InterlockedCompareExchange_nf((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_ACQUIRE(dest, exchange, compare) _InterlockedCompareExchange_acq((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_RELEASE(dest, exchange, compare) _InterlockedCompareExchange_rel((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(dest, exchange, compare) _InterlockedCompareExchange64((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_RELAXED(dest, exchange, compare) _InterlockedCompareExchange64_nf((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_ACQUIRE(dest, exchange, compare) _InterlockedCompareExchange64_acq((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_RELEASE(dest, exchange, compare) _InterlockedCompareExchange64_rel((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) _InterlockedCompareExchangePointer((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER_RELAXED(dest, exchange, compare) _InterlockedCompareExchangePointer_nf((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER_ACQUIRE(dest, exchange, compare) _InterlockedCompareExchangePointer_acq((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER_RELEASE(dest, exchange, compare) _InterlockedCompareExchangePointer_rel((void**)(dest), (void*)(exchange), (void*)(compare)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedExchangeAdd8_nf) +#pragma intrinsic(_InterlockedExchangeAdd8_acq) +#pragma intrinsic(_InterlockedExchangeAdd8_rel) +#pragma intrinsic(_InterlockedExchangeAdd16_nf) +#pragma intrinsic(_InterlockedExchangeAdd16_acq) +#pragma intrinsic(_InterlockedExchangeAdd16_rel) +#pragma intrinsic(_InterlockedExchangeAdd_nf) +#pragma intrinsic(_InterlockedExchangeAdd_acq) +#pragma intrinsic(_InterlockedExchangeAdd_rel) +#pragma intrinsic(_InterlockedExchangeAdd64_nf) +#pragma intrinsic(_InterlockedExchangeAdd64_acq) +#pragma intrinsic(_InterlockedExchangeAdd64_rel) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_RELAXED(dest, addend) _InterlockedExchangeAdd8_nf((char*)(dest), (char)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_ACQUIRE(dest, addend) _InterlockedExchangeAdd8_acq((char*)(dest), (char)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_RELEASE(dest, addend) _InterlockedExchangeAdd8_rel((char*)(dest), (char)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_RELAXED(dest, addend) _InterlockedExchangeAdd16_nf((short*)(dest), (short)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_ACQUIRE(dest, addend) _InterlockedExchangeAdd16_acq((short*)(dest), (short)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_RELEASE(dest, addend) _InterlockedExchangeAdd16_rel((short*)(dest), (short)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELAXED(dest, addend) _InterlockedExchangeAdd_nf((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_ACQUIRE(dest, addend) _InterlockedExchangeAdd_acq((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELEASE(dest, addend) _InterlockedExchangeAdd_rel((long*)(dest), (long)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_RELAXED(dest, addend) _InterlockedExchangeAdd64_nf((__int64*)(dest), (__int64)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_ACQUIRE(dest, addend) _InterlockedExchangeAdd64_acq((__int64*)(dest), (__int64)(addend)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_RELEASE(dest, addend) _InterlockedExchangeAdd64_rel((__int64*)(dest), (__int64)(addend)) + +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD((long*)(dest), byte_offset)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER_RELAXED(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELAXED((long*)(dest), byte_offset)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER_ACQUIRE(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_ACQUIRE((long*)(dest), byte_offset)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER_RELEASE(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELEASE((long*)(dest), byte_offset)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedExchange8_nf) +#pragma intrinsic(_InterlockedExchange8_acq) +#pragma intrinsic(_InterlockedExchange16_nf) +#pragma intrinsic(_InterlockedExchange16_acq) +#pragma intrinsic(_InterlockedExchange_nf) +#pragma intrinsic(_InterlockedExchange_acq) +#pragma intrinsic(_InterlockedExchange64_nf) +#pragma intrinsic(_InterlockedExchange64_acq) +#pragma intrinsic(_InterlockedExchangePointer) +#pragma intrinsic(_InterlockedExchangePointer_nf) +#pragma intrinsic(_InterlockedExchangePointer_acq) +#if _MSC_VER >= 1800 +#pragma intrinsic(_InterlockedExchange8_rel) +#pragma intrinsic(_InterlockedExchange16_rel) +#pragma intrinsic(_InterlockedExchange_rel) +#pragma intrinsic(_InterlockedExchange64_rel) +#pragma intrinsic(_InterlockedExchangePointer_rel) +#endif +#endif + +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_RELAXED(dest, newval) _InterlockedExchange8_nf((char*)(dest), (char)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_ACQUIRE(dest, newval) _InterlockedExchange8_acq((char*)(dest), (char)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_RELAXED(dest, newval) _InterlockedExchange16_nf((short*)(dest), (short)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_ACQUIRE(dest, newval) _InterlockedExchange16_acq((short*)(dest), (short)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_RELAXED(dest, newval) _InterlockedExchange_nf((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ACQUIRE(dest, newval) _InterlockedExchange_acq((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_RELAXED(dest, newval) _InterlockedExchange64_nf((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_ACQUIRE(dest, newval) _InterlockedExchange64_acq((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) _InterlockedExchangePointer((void**)(dest), (void*)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER_RELAXED(dest, newval) _InterlockedExchangePointer_nf((void**)(dest), (void*)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER_ACQUIRE(dest, newval) _InterlockedExchangePointer_acq((void**)(dest), (void*)(newval)) + +#if _MSC_VER >= 1800 +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_RELEASE(dest, newval) _InterlockedExchange8_rel((char*)(dest), (char)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_RELEASE(dest, newval) _InterlockedExchange16_rel((short*)(dest), (short)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_RELEASE(dest, newval) _InterlockedExchange_rel((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_RELEASE(dest, newval) _InterlockedExchange64_rel((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER_RELEASE(dest, newval) _InterlockedExchangePointer_rel((void**)(dest), (void*)(newval)) +#else +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_RELEASE(dest, newval) BOOST_ATOMIC_INTERLOCKED_EXCHANGE8(dest, newval) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_RELEASE(dest, newval) BOOST_ATOMIC_INTERLOCKED_EXCHANGE16(dest, newval) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_RELEASE(dest, newval) BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_RELEASE(dest, newval) BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(dest, newval) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER_RELEASE(dest, newval) BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) +#endif + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedAnd8_nf) +#pragma intrinsic(_InterlockedAnd8_acq) +#pragma intrinsic(_InterlockedAnd8_rel) +#pragma intrinsic(_InterlockedAnd16_nf) +#pragma intrinsic(_InterlockedAnd16_acq) +#pragma intrinsic(_InterlockedAnd16_rel) +#pragma intrinsic(_InterlockedAnd_nf) +#pragma intrinsic(_InterlockedAnd_acq) +#pragma intrinsic(_InterlockedAnd_rel) +#pragma intrinsic(_InterlockedAnd64_nf) +#pragma intrinsic(_InterlockedAnd64_acq) +#pragma intrinsic(_InterlockedAnd64_rel) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_AND8_RELAXED(dest, arg) _InterlockedAnd8_nf((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND8_ACQUIRE(dest, arg) _InterlockedAnd8_acq((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND8_RELEASE(dest, arg) _InterlockedAnd8_rel((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND16_RELAXED(dest, arg) _InterlockedAnd16_nf((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND16_ACQUIRE(dest, arg) _InterlockedAnd16_acq((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND16_RELEASE(dest, arg) _InterlockedAnd16_rel((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND_RELAXED(dest, arg) _InterlockedAnd_nf((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND_ACQUIRE(dest, arg) _InterlockedAnd_acq((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND_RELEASE(dest, arg) _InterlockedAnd_rel((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND64_RELAXED(dest, arg) _InterlockedAnd64_nf((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND64_ACQUIRE(dest, arg) _InterlockedAnd64_acq((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_AND64_RELEASE(dest, arg) _InterlockedAnd64_rel((__int64*)(dest), (__int64)(arg)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedOr8_nf) +#pragma intrinsic(_InterlockedOr8_acq) +#pragma intrinsic(_InterlockedOr8_rel) +#pragma intrinsic(_InterlockedOr16_nf) +#pragma intrinsic(_InterlockedOr16_acq) +#pragma intrinsic(_InterlockedOr16_rel) +#pragma intrinsic(_InterlockedOr_nf) +#pragma intrinsic(_InterlockedOr_acq) +#pragma intrinsic(_InterlockedOr_rel) +#pragma intrinsic(_InterlockedOr64_nf) +#pragma intrinsic(_InterlockedOr64_acq) +#pragma intrinsic(_InterlockedOr64_rel) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_OR8_RELAXED(dest, arg) _InterlockedOr8_nf((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR8_ACQUIRE(dest, arg) _InterlockedOr8_acq((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR8_RELEASE(dest, arg) _InterlockedOr8_rel((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR16_RELAXED(dest, arg) _InterlockedOr16_nf((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR16_ACQUIRE(dest, arg) _InterlockedOr16_acq((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR16_RELEASE(dest, arg) _InterlockedOr16_rel((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR_RELAXED(dest, arg) _InterlockedOr_nf((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR_ACQUIRE(dest, arg) _InterlockedOr_acq((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR_RELEASE(dest, arg) _InterlockedOr_rel((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR64_RELAXED(dest, arg) _InterlockedOr64_nf((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR64_ACQUIRE(dest, arg) _InterlockedOr64_acq((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_OR64_RELEASE(dest, arg) _InterlockedOr64_rel((__int64*)(dest), (__int64)(arg)) + +#if defined(BOOST_MSVC) +#pragma intrinsic(_InterlockedXor8_nf) +#pragma intrinsic(_InterlockedXor8_acq) +#pragma intrinsic(_InterlockedXor8_rel) +#pragma intrinsic(_InterlockedXor16_nf) +#pragma intrinsic(_InterlockedXor16_acq) +#pragma intrinsic(_InterlockedXor16_rel) +#pragma intrinsic(_InterlockedXor_nf) +#pragma intrinsic(_InterlockedXor_acq) +#pragma intrinsic(_InterlockedXor_rel) +#pragma intrinsic(_InterlockedXor64_nf) +#pragma intrinsic(_InterlockedXor64_acq) +#pragma intrinsic(_InterlockedXor64_rel) +#endif + +#define BOOST_ATOMIC_INTERLOCKED_XOR8_RELAXED(dest, arg) _InterlockedXor8_nf((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR8_ACQUIRE(dest, arg) _InterlockedXor8_acq((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR8_RELEASE(dest, arg) _InterlockedXor8_rel((char*)(dest), (char)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR16_RELAXED(dest, arg) _InterlockedXor16_nf((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR16_ACQUIRE(dest, arg) _InterlockedXor16_acq((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR16_RELEASE(dest, arg) _InterlockedXor16_rel((short*)(dest), (short)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR_RELAXED(dest, arg) _InterlockedXor_nf((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR_ACQUIRE(dest, arg) _InterlockedXor_acq((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR_RELEASE(dest, arg) _InterlockedXor_rel((long*)(dest), (long)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR64_RELAXED(dest, arg) _InterlockedXor64_nf((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR64_ACQUIRE(dest, arg) _InterlockedXor64_acq((__int64*)(dest), (__int64)(arg)) +#define BOOST_ATOMIC_INTERLOCKED_XOR64_RELEASE(dest, arg) _InterlockedXor64_rel((__int64*)(dest), (__int64)(arg)) + +#endif // _MSC_VER >= 1700 && defined(_M_ARM) + +#endif // _MSC_VER < 1400 + +#else // defined(_MSC_VER) && _MSC_VER >= 1310 + +#if defined(BOOST_USE_WINDOWS_H) + +#include + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) InterlockedCompareExchange((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) InterlockedExchange((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) InterlockedExchangeAdd((long*)(dest), (long)(addend)) + +#if defined(_WIN64) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(dest, exchange, compare) InterlockedCompareExchange64((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(dest, newval) InterlockedExchange64((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, addend) InterlockedExchangeAdd64((__int64*)(dest), (__int64)(addend)) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) InterlockedCompareExchangePointer((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) InterlockedExchangePointer((void**)(dest), (void*)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, byte_offset)) + +#else // defined(_WIN64) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) ((void*)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, byte_offset)) + +#endif // defined(_WIN64) + +#else // defined(BOOST_USE_WINDOWS_H) + +#if defined(__MINGW64__) +#define BOOST_ATOMIC_INTERLOCKED_IMPORT +#else +#define BOOST_ATOMIC_INTERLOCKED_IMPORT __declspec(dllimport) +#endif + +namespace boost { +namespace atomics { +namespace detail { + +extern "C" { + +BOOST_ATOMIC_INTERLOCKED_IMPORT long __stdcall InterlockedCompareExchange(long volatile*, long, long); +BOOST_ATOMIC_INTERLOCKED_IMPORT long __stdcall InterlockedExchange(long volatile*, long); +BOOST_ATOMIC_INTERLOCKED_IMPORT long __stdcall InterlockedExchangeAdd(long volatile*, long); + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare) boost::atomics::detail::InterlockedCompareExchange((long*)(dest), (long)(exchange), (long)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval) boost::atomics::detail::InterlockedExchange((long*)(dest), (long)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, addend) boost::atomics::detail::InterlockedExchangeAdd((long*)(dest), (long)(addend)) + +#if defined(_WIN64) + +BOOST_ATOMIC_INTERLOCKED_IMPORT __int64 __stdcall InterlockedCompareExchange64(__int64 volatile*, __int64, __int64); +BOOST_ATOMIC_INTERLOCKED_IMPORT __int64 __stdcall InterlockedExchange64(__int64 volatile*, __int64); +BOOST_ATOMIC_INTERLOCKED_IMPORT __int64 __stdcall InterlockedExchangeAdd64(__int64 volatile*, __int64); + +BOOST_ATOMIC_INTERLOCKED_IMPORT void* __stdcall InterlockedCompareExchangePointer(void* volatile *, void*, void*); +BOOST_ATOMIC_INTERLOCKED_IMPORT void* __stdcall InterlockedExchangePointer(void* volatile *, void*); + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(dest, exchange, compare) boost::atomics::detail::InterlockedCompareExchange64((__int64*)(dest), (__int64)(exchange), (__int64)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(dest, newval) boost::atomics::detail::InterlockedExchange64((__int64*)(dest), (__int64)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, addend) boost::atomics::detail::InterlockedExchangeAdd64((__int64*)(dest), (__int64)(addend)) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) boost::atomics::detail::InterlockedCompareExchangePointer((void**)(dest), (void*)(exchange), (void*)(compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) boost::atomics::detail::InterlockedExchangePointer((void**)(dest), (void*)(newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(dest, byte_offset)) + +#else // defined(_WIN64) + +#define BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_POINTER(dest, exchange, compare) ((void*)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(dest, exchange, compare)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_POINTER(dest, newval) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE(dest, newval)) +#define BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_POINTER(dest, byte_offset) ((void*)BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(dest, byte_offset)) + +#endif // defined(_WIN64) + +} // extern "C" + +} // namespace detail +} // namespace atomics +} // namespace boost + +#undef BOOST_ATOMIC_INTERLOCKED_IMPORT + +#endif // defined(BOOST_USE_WINDOWS_H) + +#endif // defined(_MSC_VER) + +#endif diff --git a/boost/atomic/detail/link.hpp b/boost/atomic/detail/link.hpp new file mode 100644 index 0000000000..4f522acbcf --- /dev/null +++ b/boost/atomic/detail/link.hpp @@ -0,0 +1,58 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2012 Hartmut Kaiser + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/config.hpp + * + * This header defines macros for linking with compiled library of Boost.Atomic + */ + +#ifndef BOOST_ATOMIC_DETAIL_LINK_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_LINK_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +/////////////////////////////////////////////////////////////////////////////// +// Set up dll import/export options +#if (defined(BOOST_ATOMIC_DYN_LINK) || defined(BOOST_ALL_DYN_LINK)) && \ + !defined(BOOST_ATOMIC_STATIC_LINK) + +#if defined(BOOST_ATOMIC_SOURCE) +#define BOOST_ATOMIC_DECL BOOST_SYMBOL_EXPORT +#define BOOST_ATOMIC_BUILD_DLL +#else +#define BOOST_ATOMIC_DECL BOOST_SYMBOL_IMPORT +#endif + +#endif // building a shared library + +#ifndef BOOST_ATOMIC_DECL +#define BOOST_ATOMIC_DECL +#endif + +/////////////////////////////////////////////////////////////////////////////// +// Auto library naming +#if !defined(BOOST_ATOMIC_SOURCE) && !defined(BOOST_ALL_NO_LIB) && \ + !defined(BOOST_ATOMIC_NO_LIB) + +#define BOOST_LIB_NAME boost_atomic + +// tell the auto-link code to select a dll when required: +#if defined(BOOST_ALL_DYN_LINK) || defined(BOOST_ATOMIC_DYN_LINK) +#define BOOST_DYN_LINK +#endif + +#include + +#endif // auto-linking disabled + +#endif diff --git a/boost/atomic/detail/lockpool.hpp b/boost/atomic/detail/lockpool.hpp new file mode 100644 index 0000000000..4e249aa048 --- /dev/null +++ b/boost/atomic/detail/lockpool.hpp @@ -0,0 +1,51 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2011 Helge Bahmann + * Copyright (c) 2013-2014 Andrey Semashev + */ +/*! + * \file atomic/detail/lockpool.hpp + * + * This header contains declaration of the lockpool used to emulate atomic ops. + */ + +#ifndef BOOST_ATOMIC_DETAIL_LOCKPOOL_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_LOCKPOOL_HPP_INCLUDED_ + +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +struct lockpool +{ + class scoped_lock + { + void* m_lock; + + public: + explicit BOOST_ATOMIC_DECL scoped_lock(const volatile void* addr) BOOST_NOEXCEPT; + BOOST_ATOMIC_DECL ~scoped_lock() BOOST_NOEXCEPT; + + BOOST_DELETED_FUNCTION(scoped_lock(scoped_lock const&)) + BOOST_DELETED_FUNCTION(scoped_lock& operator=(scoped_lock const&)) + }; + + static BOOST_ATOMIC_DECL void thread_fence() BOOST_NOEXCEPT; + static BOOST_ATOMIC_DECL void signal_fence() BOOST_NOEXCEPT; +}; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_LOCKPOOL_HPP_INCLUDED_ diff --git a/boost/atomic/detail/operations.hpp b/boost/atomic/detail/operations.hpp new file mode 100644 index 0000000000..d81399a8e3 --- /dev/null +++ b/boost/atomic/detail/operations.hpp @@ -0,0 +1,24 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/operations.hpp + * + * This header defines atomic operations, including the emulated version. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPERATIONS_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPERATIONS_HPP_INCLUDED_ + +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#endif // BOOST_ATOMIC_DETAIL_OPERATIONS_HPP_INCLUDED_ diff --git a/boost/atomic/detail/operations_fwd.hpp b/boost/atomic/detail/operations_fwd.hpp new file mode 100644 index 0000000000..69049e4630 --- /dev/null +++ b/boost/atomic/detail/operations_fwd.hpp @@ -0,0 +1,34 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/operations_fwd.hpp + * + * This header contains forward declaration of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPERATIONS_FWD_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPERATIONS_FWD_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< unsigned int Size, bool Signed > +struct operations; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPERATIONS_FWD_HPP_INCLUDED_ diff --git a/boost/atomic/detail/operations_lockfree.hpp b/boost/atomic/detail/operations_lockfree.hpp new file mode 100644 index 0000000000..b465403a6d --- /dev/null +++ b/boost/atomic/detail/operations_lockfree.hpp @@ -0,0 +1,30 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/operations_lockfree.hpp + * + * This header defines lockfree atomic operations. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPERATIONS_LOCKFREE_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPERATIONS_LOCKFREE_HPP_INCLUDED_ + +#include +#include + +#if !defined(BOOST_ATOMIC_EMULATED) +#include BOOST_ATOMIC_DETAIL_HEADER(boost/atomic/detail/ops_) +#else +#include +#endif + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#endif // BOOST_ATOMIC_DETAIL_OPERATIONS_LOCKFREE_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_cas_based.hpp b/boost/atomic/detail/ops_cas_based.hpp new file mode 100644 index 0000000000..7f8d288f7f --- /dev/null +++ b/boost/atomic/detail/ops_cas_based.hpp @@ -0,0 +1,91 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_cas_based.hpp + * + * This header contains CAS-based implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_CAS_BASED_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_CAS_BASED_HPP_INCLUDED_ + +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< typename Base > +struct cas_based_operations : + public Base +{ + typedef typename Base::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, old_val + v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, old_val - v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, old_val & v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, old_val | v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + while (!Base::compare_exchange_weak(storage, old_val, old_val ^ v, order, memory_order_relaxed)) {} + return old_val; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + Base::store(storage, (storage_type)0, order); + } +}; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_CAS_BASED_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_emulated.hpp b/boost/atomic/detail/ops_emulated.hpp new file mode 100644 index 0000000000..597490f2d7 --- /dev/null +++ b/boost/atomic/detail/ops_emulated.hpp @@ -0,0 +1,149 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_emulated.hpp + * + * This header contains lockpool-based implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_EMULATED_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_EMULATED_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< typename T > +struct emulated_operations +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + lockpool::scoped_lock lock(&storage); + const_cast< storage_type& >(storage) = v; + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT + { + lockpool::scoped_lock lock(&storage); + return const_cast< storage_type const& >(storage); + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s += v; + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s -= v; + return old_val; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s = v; + return old_val; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + const bool res = old_val == expected; + if (res) + s = desired; + expected = old_val; + + return res; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s &= v; + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s |= v; + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type& s = const_cast< storage_type& >(storage); + lockpool::scoped_lock lock(&storage); + storage_type old_val = s; + s ^= v; + return old_val; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, (storage_type)0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return false; + } +}; + +template< unsigned int Size, bool Signed > +struct operations : + public emulated_operations< typename make_storage_type< Size, Signed >::type > +{ +}; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_EMULATED_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_extending_cas_based.hpp b/boost/atomic/detail/ops_extending_cas_based.hpp new file mode 100644 index 0000000000..d7f3c5f5de --- /dev/null +++ b/boost/atomic/detail/ops_extending_cas_based.hpp @@ -0,0 +1,65 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_extending_cas_based.hpp + * + * This header contains a boilerplate of the \c operations template implementation that requires sign/zero extension in arithmetic operations. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_EXTENDING_CAS_BASED_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_EXTENDING_CAS_BASED_HPP_INCLUDED_ + +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< typename Base, unsigned int Size, bool Signed > +struct extending_cas_based_operations : + public Base +{ + typedef typename Base::storage_type storage_type; + typedef typename make_storage_type< Size, Signed >::type emulated_storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + emulated_storage_type new_val; + do + { + new_val = static_cast< emulated_storage_type >(old_val) + static_cast< emulated_storage_type >(v); + } + while (!Base::compare_exchange_weak(storage, old_val, static_cast< storage_type >(new_val), order, memory_order_relaxed)); + return old_val; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type old_val = Base::load(storage, memory_order_relaxed); + emulated_storage_type new_val; + do + { + new_val = static_cast< emulated_storage_type >(old_val) - static_cast< emulated_storage_type >(v); + } + while (!Base::compare_exchange_weak(storage, old_val, static_cast< storage_type >(new_val), order, memory_order_relaxed)); + return old_val; + } +}; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_EXTENDING_CAS_BASED_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_alpha.hpp b/boost/atomic/detail/ops_gcc_alpha.hpp new file mode 100644 index 0000000000..d17c61d7a5 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_alpha.hpp @@ -0,0 +1,874 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_alpha.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_ALPHA_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_ALPHA_HPP_INCLUDED_ + +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +/* + Refer to http://h71000.www7.hp.com/doc/82final/5601/5601pro_004.html + (HP OpenVMS systems documentation) and the Alpha Architecture Reference Manual. + */ + +/* + NB: The most natural thing would be to write the increment/decrement + operators along the following lines: + + __asm__ __volatile__ + ( + "1: ldl_l %0,%1 \n" + "addl %0,1,%0 \n" + "stl_c %0,%1 \n" + "beq %0,1b\n" + : "=&b" (tmp) + : "m" (value) + : "cc" + ); + + However according to the comments on the HP website and matching + comments in the Linux kernel sources this defies branch prediction, + as the cpu assumes that backward branches are always taken; so + instead copy the trick from the Linux kernel, introduce a forward + branch and back again. + + I have, however, had a hard time measuring the difference between + the two versions in microbenchmarks -- I am leaving it in nevertheless + as it apparently does not hurt either. +*/ + +struct gcc_alpha_operations_base +{ + static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + __asm__ __volatile__ ("mb" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT + { + if ((order & (memory_order_consume | memory_order_acquire)) != 0) + __asm__ __volatile__ ("mb" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("mb" ::: "memory"); + } +}; + + +template< bool Signed > +struct operations< 4u, Signed > : + public gcc_alpha_operations_base +{ + typedef typename make_storage_type< 4u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "mov %3, %1\n" + "ldl_l %0, %2\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (tmp) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + int success; + storage_type current; + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %2, %4\n" // current = *(&storage) + "cmpeq %2, %0, %3\n" // success = current == expected + "mov %2, %0\n" // expected = current + "beq %3, 2f\n" // if (success == 0) goto end + "stl_c %1, %4\n" // storage = desired; desired = store succeeded + "mov %1, %3\n" // success = desired + "2:\n" + : "+&r" (expected), // %0 + "+&r" (desired), // %1 + "=&r" (current), // %2 + "=&r" (success) // %3 + : "m" (storage) // %4 + : + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + storage_type current, tmp; + fence_before(success_order); + __asm__ __volatile__ + ( + "1:\n" + "mov %5, %1\n" // tmp = desired + "ldl_l %2, %4\n" // current = *(&storage) + "cmpeq %2, %0, %3\n" // success = current == expected + "mov %2, %0\n" // expected = current + "beq %3, 2f\n" // if (success == 0) goto end + "stl_c %1, %4\n" // storage = tmp; tmp = store succeeded + "beq %1, 3f\n" // if (tmp == 0) goto retry + "mov %1, %3\n" // success = tmp + "2:\n" + + ".subsection 2\n" + "3: br 1b\n" + ".previous\n" + + : "+&r" (expected), // %0 + "=&r" (tmp), // %1 + "=&r" (current), // %2 + "=&r" (success) // %3 + : "m" (storage), // %4 + "r" (desired) // %5 + : + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "addl %0, %3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "subl %0, %3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "and %0, %3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "bis %0, %3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "xor %0, %3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + + +template< > +struct operations< 1u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "addl %0, %3, %1\n" + "zapnot %1, #1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "subl %0, %3, %1\n" + "zapnot %1, #1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 1u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "addl %0, %3, %1\n" + "sextb %1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "subl %0, %3, %1\n" + "sextb %1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } +}; + + +template< > +struct operations< 2u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "addl %0, %3, %1\n" + "zapnot %1, #3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "subl %0, %3, %1\n" + "zapnot %1, #3, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 2u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "addl %0, %3, %1\n" + "sextw %1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldl_l %0, %2\n" + "subl %0, %3, %1\n" + "sextw %1, %1\n" + "stl_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } +}; + + +template< bool Signed > +struct operations< 8u, Signed > : + public gcc_alpha_operations_base +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "mov %3, %1\n" + "ldq_l %0, %2\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (tmp) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + int success; + storage_type current; + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %2, %4\n" // current = *(&storage) + "cmpeq %2, %0, %3\n" // success = current == expected + "mov %2, %0\n" // expected = current + "beq %3, 2f\n" // if (success == 0) goto end + "stq_c %1, %4\n" // storage = desired; desired = store succeeded + "mov %1, %3\n" // success = desired + "2:\n" + : "+&r" (expected), // %0 + "+&r" (desired), // %1 + "=&r" (current), // %2 + "=&r" (success) // %3 + : "m" (storage) // %4 + : + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + storage_type current, tmp; + fence_before(success_order); + __asm__ __volatile__ + ( + "1:\n" + "mov %5, %1\n" // tmp = desired + "ldq_l %2, %4\n" // current = *(&storage) + "cmpeq %2, %0, %3\n" // success = current == expected + "mov %2, %0\n" // expected = current + "beq %3, 2f\n" // if (success == 0) goto end + "stq_c %1, %4\n" // storage = tmp; tmp = store succeeded + "beq %1, 3f\n" // if (tmp == 0) goto retry + "mov %1, %3\n" // success = tmp + "2:\n" + + ".subsection 2\n" + "3: br 1b\n" + ".previous\n" + + : "+&r" (expected), // %0 + "=&r" (tmp), // %1 + "=&r" (current), // %2 + "=&r" (success) // %3 + : "m" (storage), // %4 + "r" (desired) // %5 + : + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %0, %2\n" + "addq %0, %3, %1\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %0, %2\n" + "subq %0, %3, %1\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %0, %2\n" + "and %0, %3, %1\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %0, %2\n" + "bis %0, %3, %1\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, modified; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldq_l %0, %2\n" + "xor %0, %3, %1\n" + "stq_c %1, %2\n" + "beq %1, 2f\n" + + ".subsection 2\n" + "2: br 1b\n" + ".previous\n" + + : "=&r" (original), // %0 + "=&r" (modified) // %1 + : "m" (storage), // %2 + "r" (v) // %3 + : + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("mb" ::: "memory"); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_ALPHA_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_arm.hpp b/boost/atomic/detail/ops_gcc_arm.hpp new file mode 100644 index 0000000000..29e1e5a319 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_arm.hpp @@ -0,0 +1,971 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_arm.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +// From the ARM Architecture Reference Manual for architecture v6: +// +// LDREX{} , [] +// Specifies the destination register for the memory word addressed by +// Specifies the register containing the address. +// +// STREX{} , , [] +// Specifies the destination register for the returned status value. +// 0 if the operation updates memory +// 1 if the operation fails to update memory +// Specifies the register containing the word to be stored to memory. +// Specifies the register containing the address. +// Rd must not be the same register as Rm or Rn. +// +// ARM v7 is like ARM v6 plus: +// There are half-word and byte versions of the LDREX and STREX instructions, +// LDREXH, LDREXB, STREXH and STREXB. +// There are also double-word versions, LDREXD and STREXD. +// (Actually it looks like these are available from version 6k onwards.) +// FIXME these are not yet used; should be mostly a matter of copy-and-paste. +// I think you can supply an immediate offset to the address. +// +// A memory barrier is effected using a "co-processor 15" instruction, +// though a separate assembler mnemonic is available for it in v7. +// +// "Thumb 1" is a subset of the ARM instruction set that uses a 16-bit encoding. It +// doesn't include all instructions and in particular it doesn't include the co-processor +// instruction used for the memory barrier or the load-locked/store-conditional +// instructions. So, if we're compiling in "Thumb 1" mode, we need to wrap all of our +// asm blocks with code to temporarily change to ARM mode. +// +// You can only change between ARM and Thumb modes when branching using the bx instruction. +// bx takes an address specified in a register. The least significant bit of the address +// indicates the mode, so 1 is added to indicate that the destination code is Thumb. +// A temporary register is needed for the address and is passed as an argument to these +// macros. It must be one of the "low" registers accessible to Thumb code, specified +// using the "l" attribute in the asm statement. +// +// Architecture v7 introduces "Thumb 2", which does include (almost?) all of the ARM +// instruction set. (Actually, there was an extension of v6 called v6T2 which supported +// "Thumb 2" mode, but its architecture manual is no longer available, referring to v7.) +// So in v7 we don't need to change to ARM mode; we can write "universal +// assembler" which will assemble to Thumb 2 or ARM code as appropriate. The only thing +// we need to do to make this "universal" assembler mode work is to insert "IT" instructions +// to annotate the conditional instructions. These are ignored in other modes (e.g. v6), +// so they can always be present. + +// A note about memory_order_consume. Technically, this architecture allows to avoid +// unnecessary memory barrier after consume load since it supports data dependency ordering. +// However, some compiler optimizations may break a seemingly valid code relying on data +// dependency tracking by injecting bogus branches to aid out of order execution. +// This may happen not only in Boost.Atomic code but also in user's code, which we have no +// control of. See this thread: http://lists.boost.org/Archives/boost/2014/06/213890.php. +// For this reason we promote memory_order_consume to memory_order_acquire. + +#if defined(__thumb__) && !defined(__thumb2__) +#define BOOST_ATOMIC_DETAIL_ARM_ASM_START(TMPREG) "adr " #TMPREG ", 8f\n" "bx " #TMPREG "\n" ".arm\n" ".align 4\n" "8:\n" +#define BOOST_ATOMIC_DETAIL_ARM_ASM_END(TMPREG) "adr " #TMPREG ", 9f + 1\n" "bx " #TMPREG "\n" ".thumb\n" ".align 2\n" "9:\n" +#define BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(var) "=&l" (var) +#else +// The tmpreg may be wasted in this case, which is non-optimal. +#define BOOST_ATOMIC_DETAIL_ARM_ASM_START(TMPREG) +#define BOOST_ATOMIC_DETAIL_ARM_ASM_END(TMPREG) +#define BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(var) "=&r" (var) +#endif + +struct gcc_arm_operations_base +{ + static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT + { + if ((order & (memory_order_consume | memory_order_acquire)) != 0) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT + { +#if defined(BOOST_ATOMIC_DETAIL_ARM_HAS_DMB) + // Older binutils (supposedly, older than 2.21.1) didn't support symbolic or numeric arguments of the "dmb" instruction such as "ish" or "#11". + // As a workaround we have to inject encoded bytes of the instruction. There are two encodings for the instruction: ARM and Thumb. See ARM Architecture Reference Manual, A8.8.43. + // Since we cannot detect binutils version at compile time, we'll have to always use this hack. + __asm__ __volatile__ + ( +#if defined(__thumb2__) + ".short 0xF3BF, 0x8F5B\n" // dmb ish +#else + ".word 0xF57FF05B\n" // dmb ish +#endif + : + : + : "memory" + ); +#else + int tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "mcr\tp15, 0, r0, c7, c10, 5\n" + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : "=&l" (tmp) + : + : "memory" + ); +#endif + } +}; + + +template< bool Signed > +struct operations< 4u, Signed > : + public gcc_arm_operations_base +{ + typedef typename make_storage_type< 4u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original; + fence_before(order); + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // load the original value + "strex %[tmp], %[value], %[storage]\n" // store the replacement, tmp = store failed + "teq %[tmp], #0\n" // check if store succeeded + "bne 1b\n" + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [tmp] "=&l" (tmp), [original] "=&r" (original), [storage] "+Q" (storage) + : [value] "r" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + uint32_t success; + uint32_t tmp; + storage_type original; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "mov %[success], #0\n" // success = 0 + "ldrex %[original], %[storage]\n" // original = *(&storage) + "cmp %[original], %[expected]\n" // flags = original==expected + "itt eq\n" // [hint that the following 2 instructions are conditional on flags.equal] + "strexeq %[success], %[desired], %[storage]\n" // if (flags.equal) *(&storage) = desired, success = store failed + "eoreq %[success], %[success], #1\n" // if (flags.equal) success ^= 1 (i.e. make it 1 if store succeeded) + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [success] "=&r" (success), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [expected] "r" (expected), // %4 + [desired] "r" (desired) // %5 + : "cc" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = original; + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + uint32_t success; + uint32_t tmp; + storage_type original; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "mov %[success], #0\n" // success = 0 + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "cmp %[original], %[expected]\n" // flags = original==expected + "bne 2f\n" // if (!flags.equal) goto end + "strex %[success], %[desired], %[storage]\n" // *(&storage) = desired, success = store failed + "eors %[success], %[success], #1\n" // success ^= 1 (i.e. make it 1 if store succeeded); flags.equal = success == 0 + "beq 1b\n" // if (flags.equal) goto retry + "2:\n" + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [success] "=&r" (success), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [expected] "r" (expected), // %4 + [desired] "r" (desired) // %5 + : "cc" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = original; + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "add %[result], %[original], %[value]\n" // result = original + value + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "sub %[result], %[original], %[value]\n" // result = original - value + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "and %[result], %[original], %[value]\n" // result = original & value + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "orr %[result], %[original], %[value]\n" // result = original | value + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "eor %[result], %[original], %[value]\n" // result = original ^ value + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + + +template< > +struct operations< 1u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "add %[result], %[original], %[value]\n" // result = original + value + "uxtb %[result], %[result]\n" // zero extend result from 8 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "sub %[result], %[original], %[value]\n" // result = original - value + "uxtb %[result], %[result]\n" // zero extend result from 8 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 1u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "add %[result], %[original], %[value]\n" // result = original + value + "sxtb %[result], %[result]\n" // sign extend result from 8 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "sub %[result], %[original], %[value]\n" // result = original - value + "sxtb %[result], %[result]\n" // sign extend result from 8 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } +}; + + +template< > +struct operations< 2u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "add %[result], %[original], %[value]\n" // result = original + value + "uxth %[result], %[result]\n" // zero extend result from 16 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "sub %[result], %[original], %[value]\n" // result = original - value + "uxth %[result], %[result]\n" // zero extend result from 16 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 2u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "add %[result], %[original], %[value]\n" // result = original + value + "sxth %[result], %[result]\n" // sign extend result from 16 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + uint32_t tmp; + storage_type original, result; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp]) + "1:\n" + "ldrex %[original], %[storage]\n" // original = *(&storage) + "sub %[result], %[original], %[value]\n" // result = original - value + "sxth %[result], %[result]\n" // sign extend result from 16 to 32 bits + "strex %[tmp], %[result], %[storage]\n" // *(&storage) = result, tmp = store failed + "teq %[tmp], #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp]) + : [original] "=&r" (original), // %0 + [result] "=&r" (result), // %1 + [tmp] "=&l" (tmp), // %2 + [storage] "+Q" (storage) // %3 + : [value] "r" (v) // %4 + : "cc" + ); + fence_after(order); + return original; + } +}; + + +#if defined(BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD) + +// Unlike 32-bit operations, for 64-bit loads and stores we must use ldrexd/strexd. +// Any other instructions result in a non-atomic sequence of 32-bit accesses. +// See "ARM Architecture Reference Manual ARMv7-A and ARMv7-R edition", +// Section A3.5.3 "Atomicity in the ARM architecture". + +// In the asm blocks below we have to use 32-bit register pairs to compose 64-bit values. +// In order to pass the 64-bit operands to/from asm blocks, we use undocumented gcc feature: +// the lower half (Rt) of the operand is accessible normally, via the numbered placeholder (e.g. %0), +// and the upper half (Rt2) - via the same placeholder with an 'H' after the '%' sign (e.g. %H0). +// See: http://hardwarebug.org/2010/07/06/arm-inline-asm-secrets/ + +template< bool Signed > +struct operations< 8u, Signed > : + public gcc_arm_operations_base +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + exchange(storage, v, order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type original; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "ldrexd %1, %H1, [%2]\n" + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original) // %1 + : "r" (&storage) // %2 + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original; + fence_before(order); + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // load the original value + "strexd %0, %2, %H2, [%3]\n" // store the replacement, tmp = store failed + "teq %0, #0\n" // check if store succeeded + "bne 1b\n" + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original) // %1 + : "r" (v), // %2 + "r" (&storage) // %3 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + uint32_t tmp; + storage_type original, old_val = expected; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "cmp %1, %2\n" // flags = original.lo==old_val.lo + "ittt eq\n" // [hint that the following 3 instructions are conditional on flags.equal] + "cmpeq %H1, %H2\n" // if (flags.equal) flags = original.hi==old_val.hi + "strexdeq %0, %4, %H4, [%3]\n" // if (flags.equal) *(&storage) = desired, tmp = store failed + "teqeq %0, #0\n" // if (flags.equal) flags = tmp==0 + "ite eq\n" // [hint that the following 2 instructions are conditional on flags.equal] + "moveq %2, #1\n" // if (flags.equal) old_val.lo = 1 + "movne %2, #0\n" // if (!flags.equal) old_val.lo = 0 + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "+r" (old_val) // %2 + : "r" (&storage), // %3 + "r" (desired) // %4 + : "cc", "memory" + ); + const uint32_t success = (uint32_t)old_val; + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = original; + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + uint32_t tmp; + storage_type original, old_val = expected; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "cmp %1, %2\n" // flags = original.lo==old_val.lo + "it eq\n" // [hint that the following instruction is conditional on flags.equal] + "cmpeq %H1, %H2\n" // if (flags.equal) flags = original.hi==old_val.hi + "bne 2f\n" // if (!flags.equal) goto end + "strexd %0, %4, %H4, [%3]\n" // *(&storage) = desired, tmp = store failed + "teq %0, #0\n" // flags.equal = tmp == 0 + "bne 1b\n" // if (flags.equal) goto retry + "2:\n" + "ite eq\n" // [hint that the following 2 instructions are conditional on flags.equal] + "moveq %2, #1\n" // if (flags.equal) old_val.lo = 1 + "movne %2, #0\n" // if (!flags.equal) old_val.lo = 0 + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "+r" (old_val) // %2 + : "r" (&storage), // %3 + "r" (desired) // %4 + : "cc", "memory" + ); + const uint32_t success = (uint32_t)old_val; + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = original; + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage_type original, result; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "adds %2, %1, %4\n" // result = original + value + "adc %H2, %H1, %H4\n" + "strexd %0, %2, %H2, [%3]\n" // *(&storage) = result, tmp = store failed + "teq %0, #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "=&r" (result) // %2 + : "r" (&storage), // %3 + "r" (v) // %4 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage_type original, result; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "subs %2, %1, %4\n" // result = original - value + "sbc %H2, %H1, %H4\n" + "strexd %0, %2, %H2, [%3]\n" // *(&storage) = result, tmp = store failed + "teq %0, #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "=&r" (result) // %2 + : "r" (&storage), // %3 + "r" (v) // %4 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage_type original, result; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "and %2, %1, %4\n" // result = original & value + "and %H2, %H1, %H4\n" + "strexd %0, %2, %H2, [%3]\n" // *(&storage) = result, tmp = store failed + "teq %0, #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "=&r" (result) // %2 + : "r" (&storage), // %3 + "r" (v) // %4 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage_type original, result; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "orr %2, %1, %4\n" // result = original | value + "orr %H2, %H1, %H4\n" + "strexd %0, %2, %H2, [%3]\n" // *(&storage) = result, tmp = store failed + "teq %0, #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "=&r" (result) // %2 + : "r" (&storage), // %3 + "r" (v) // %4 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + storage_type original, result; + uint32_t tmp; + __asm__ __volatile__ + ( + BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0) + "1:\n" + "ldrexd %1, %H1, [%3]\n" // original = *(&storage) + "eor %2, %1, %4\n" // result = original ^ value + "eor %H2, %H1, %H4\n" + "strexd %0, %2, %H2, [%3]\n" // *(&storage) = result, tmp = store failed + "teq %0, #0\n" // flags = tmp==0 + "bne 1b\n" // if (!flags.equal) goto retry + BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0) + : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0 + "=&r" (original), // %1 + "=&r" (result) // %2 + : "r" (&storage), // %3 + "r" (v) // %4 + : "cc", "memory" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +#endif // defined(BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD) + + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + gcc_arm_operations_base::hardware_full_fence(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_atomic.hpp b/boost/atomic/detail/ops_gcc_atomic.hpp new file mode 100644 index 0000000000..2e4c37bec7 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_atomic.hpp @@ -0,0 +1,380 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_atomic.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_ATOMIC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_ATOMIC_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#if defined(__clang__) && (defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)) +#include +#include +#endif + +#if __GCC_ATOMIC_LLONG_LOCK_FREE != BOOST_ATOMIC_LLONG_LOCK_FREE || __GCC_ATOMIC_LONG_LOCK_FREE != BOOST_ATOMIC_LONG_LOCK_FREE ||\ + __GCC_ATOMIC_INT_LOCK_FREE != BOOST_ATOMIC_INT_LOCK_FREE || __GCC_ATOMIC_SHORT_LOCK_FREE != BOOST_ATOMIC_SHORT_LOCK_FREE ||\ + __GCC_ATOMIC_CHAR_LOCK_FREE != BOOST_ATOMIC_CHAR_LOCK_FREE || __GCC_ATOMIC_BOOL_LOCK_FREE != BOOST_ATOMIC_BOOL_LOCK_FREE ||\ + __GCC_ATOMIC_WCHAR_T_LOCK_FREE != BOOST_ATOMIC_WCHAR_T_LOCK_FREE +// There are platforms where we need to use larger storage types +#include +#include +#endif + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(__INTEL_COMPILER) +// This is used to suppress warning #32013 described below for Intel Compiler. +// In debug builds the compiler does not inline any functions, so basically +// every atomic function call results in this warning. I don't know any other +// way to selectively disable just this one warning. +#pragma system_header +#endif + +namespace boost { +namespace atomics { +namespace detail { + +/*! + * The function converts \c boost::memory_order values to the compiler-specific constants. + * + * NOTE: The intention is that the function is optimized away by the compiler, and the + * compiler-specific constants are passed to the intrinsics. I know constexpr doesn't + * work in this case because the standard atomics interface require memory ordering + * constants to be passed as function arguments, at which point they stop being constexpr. + * However it is crucial that the compiler sees constants and not runtime values, + * because otherwise it just ignores the ordering value and always uses seq_cst. + * This is the case with Intel C++ Compiler 14.0.3 (Composer XE 2013 SP1, update 3) and + * gcc 4.8.2. Intel Compiler issues a warning in this case: + * + * warning #32013: Invalid memory order specified. Defaulting to seq_cst memory order. + * + * while gcc acts silently. + * + * To mitigate the problem ALL functions, including the atomic<> members must be + * declared with BOOST_FORCEINLINE. In this case the compilers are able to see that + * all functions are called with constant orderings and call intrinstcts properly. + * + * Unfortunately, this still doesn't work in debug mode as the compiler doesn't + * inline functions even when marked with BOOST_FORCEINLINE. In this case all atomic + * operaions will be executed with seq_cst semantics. + */ +BOOST_FORCEINLINE BOOST_CONSTEXPR int convert_memory_order_to_gcc(memory_order order) BOOST_NOEXCEPT +{ + return (order == memory_order_relaxed ? __ATOMIC_RELAXED : (order == memory_order_consume ? __ATOMIC_CONSUME : + (order == memory_order_acquire ? __ATOMIC_ACQUIRE : (order == memory_order_release ? __ATOMIC_RELEASE : + (order == memory_order_acq_rel ? __ATOMIC_ACQ_REL : __ATOMIC_SEQ_CST))))); +} + +template< typename T > +struct gcc_atomic_operations +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + __atomic_store_n(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return __atomic_load_n(&storage, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_fetch_add(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_fetch_sub(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_exchange_n(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return __atomic_compare_exchange_n + ( + &storage, &expected, desired, false, + atomics::detail::convert_memory_order_to_gcc(success_order), + atomics::detail::convert_memory_order_to_gcc(failure_order) + ); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return __atomic_compare_exchange_n + ( + &storage, &expected, desired, true, + atomics::detail::convert_memory_order_to_gcc(success_order), + atomics::detail::convert_memory_order_to_gcc(failure_order) + ); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_fetch_and(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_fetch_or(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return __atomic_fetch_xor(&storage, v, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return __atomic_test_and_set(&storage, atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + __atomic_clear(const_cast< storage_type* >(&storage), atomics::detail::convert_memory_order_to_gcc(order)); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile& storage) BOOST_NOEXCEPT + { + return __atomic_is_lock_free(sizeof(storage_type), &storage); + } +}; + +#if BOOST_ATOMIC_INT128_LOCK_FREE > 0 +#if defined(__clang__) && defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +// Workaround for clang bug: http://llvm.org/bugs/show_bug.cgi?id=19149 +// Clang 3.4 does not implement 128-bit __atomic* intrinsics even though it defines __GCC_HAVE_SYNC_COMPARE_AND_SWAP_16 +template< bool Signed > +struct operations< 16u, Signed > : + public cas_based_operations< gcc_dcas_x86_64< Signed > > +{ +}; + +#else + +template< bool Signed > +struct operations< 16u, Signed > : + public gcc_atomic_operations< typename make_storage_type< 16u, Signed >::type > +{ +}; + +#endif +#endif + + +#if BOOST_ATOMIC_INT64_LOCK_FREE > 0 +#if defined(__clang__) && defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) + +// Workaround for clang bug http://llvm.org/bugs/show_bug.cgi?id=19355 +template< bool Signed > +struct operations< 8u, Signed > : + public cas_based_operations< gcc_dcas_x86< Signed > > +{ +}; + +#elif (BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 8 && __GCC_ATOMIC_LLONG_LOCK_FREE != BOOST_ATOMIC_LLONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 8 && __GCC_ATOMIC_LONG_LOCK_FREE != BOOST_ATOMIC_LONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_INT == 8 && __GCC_ATOMIC_INT_LOCK_FREE != BOOST_ATOMIC_INT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 8 && __GCC_ATOMIC_SHORT_LOCK_FREE != BOOST_ATOMIC_SHORT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 8 && __GCC_ATOMIC_WCHAR_T_LOCK_FREE != BOOST_ATOMIC_WCHAR_T_LOCK_FREE) + +#define BOOST_ATOMIC_DETAIL_INT64_EXTENDED + +template< bool Signed > +struct operations< 8u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 16u, Signed >::type >, 8u, Signed > +{ +}; + +#else + +template< bool Signed > +struct operations< 8u, Signed > : + public gcc_atomic_operations< typename make_storage_type< 8u, Signed >::type > +{ +}; + +#endif +#endif + +#if BOOST_ATOMIC_INT32_LOCK_FREE > 0 +#if (BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 4 && __GCC_ATOMIC_LLONG_LOCK_FREE != BOOST_ATOMIC_LLONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 4 && __GCC_ATOMIC_LONG_LOCK_FREE != BOOST_ATOMIC_LONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_INT == 4 && __GCC_ATOMIC_INT_LOCK_FREE != BOOST_ATOMIC_INT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 4 && __GCC_ATOMIC_SHORT_LOCK_FREE != BOOST_ATOMIC_SHORT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 4 && __GCC_ATOMIC_WCHAR_T_LOCK_FREE != BOOST_ATOMIC_WCHAR_T_LOCK_FREE) + +#define BOOST_ATOMIC_DETAIL_INT32_EXTENDED + +#if !defined(BOOST_ATOMIC_DETAIL_INT64_EXTENDED) + +template< bool Signed > +struct operations< 4u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 8u, Signed >::type >, 4u, Signed > +{ +}; + +#else // !defined(BOOST_ATOMIC_DETAIL_INT64_EXTENDED) + +template< bool Signed > +struct operations< 4u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 16u, Signed >::type >, 4u, Signed > +{ +}; + +#endif // !defined(BOOST_ATOMIC_DETAIL_INT64_EXTENDED) + +#else + +template< bool Signed > +struct operations< 4u, Signed > : + public gcc_atomic_operations< typename make_storage_type< 4u, Signed >::type > +{ +}; + +#endif +#endif + +#if BOOST_ATOMIC_INT16_LOCK_FREE > 0 +#if (BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 2 && __GCC_ATOMIC_LLONG_LOCK_FREE != BOOST_ATOMIC_LLONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 2 && __GCC_ATOMIC_LONG_LOCK_FREE != BOOST_ATOMIC_LONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_INT == 2 && __GCC_ATOMIC_INT_LOCK_FREE != BOOST_ATOMIC_INT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 2 && __GCC_ATOMIC_SHORT_LOCK_FREE != BOOST_ATOMIC_SHORT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 2 && __GCC_ATOMIC_WCHAR_T_LOCK_FREE != BOOST_ATOMIC_WCHAR_T_LOCK_FREE) + +#define BOOST_ATOMIC_DETAIL_INT16_EXTENDED + +#if !defined(BOOST_ATOMIC_DETAIL_INT32_EXTENDED) + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 4u, Signed >::type >, 2u, Signed > +{ +}; + +#elif !defined(BOOST_ATOMIC_DETAIL_INT64_EXTENDED) + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 8u, Signed >::type >, 2u, Signed > +{ +}; + +#else + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 16u, Signed >::type >, 2u, Signed > +{ +}; + +#endif + +#else + +template< bool Signed > +struct operations< 2u, Signed > : + public gcc_atomic_operations< typename make_storage_type< 2u, Signed >::type > +{ +}; + +#endif +#endif + +#if BOOST_ATOMIC_INT8_LOCK_FREE > 0 +#if (BOOST_ATOMIC_DETAIL_SIZEOF_LLONG == 1 && __GCC_ATOMIC_LLONG_LOCK_FREE != BOOST_ATOMIC_LLONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_LONG == 1 && __GCC_ATOMIC_LONG_LOCK_FREE != BOOST_ATOMIC_LONG_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_INT == 1 && __GCC_ATOMIC_INT_LOCK_FREE != BOOST_ATOMIC_INT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_SHORT == 1 && __GCC_ATOMIC_SHORT_LOCK_FREE != BOOST_ATOMIC_SHORT_LOCK_FREE) ||\ + (BOOST_ATOMIC_DETAIL_SIZEOF_WCHAR_T == 1 && __GCC_ATOMIC_WCHAR_T_LOCK_FREE != BOOST_ATOMIC_WCHAR_T_LOCK_FREE) ||\ + (__GCC_ATOMIC_CHAR_LOCK_FREE != BOOST_ATOMIC_CHAR_LOCK_FREE) ||\ + (__GCC_ATOMIC_BOOL_LOCK_FREE != BOOST_ATOMIC_BOOL_LOCK_FREE) + +#if !defined(BOOST_ATOMIC_DETAIL_INT16_EXTENDED) + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 2u, Signed >::type >, 1u, Signed > +{ +}; + +#elif !defined(BOOST_ATOMIC_DETAIL_INT32_EXTENDED) + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 4u, Signed >::type >, 1u, Signed > +{ +}; + +#elif !defined(BOOST_ATOMIC_DETAIL_INT64_EXTENDED) + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 8u, Signed >::type >, 1u, Signed > +{ +}; + +#else + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< gcc_atomic_operations< typename make_storage_type< 16u, Signed >::type >, 1u, Signed > +{ +}; + +#endif + +#else + +template< bool Signed > +struct operations< 1u, Signed > : + public gcc_atomic_operations< typename make_storage_type< 1u, Signed >::type > +{ +}; + +#endif +#endif + +#undef BOOST_ATOMIC_DETAIL_INT16_EXTENDED +#undef BOOST_ATOMIC_DETAIL_INT32_EXTENDED +#undef BOOST_ATOMIC_DETAIL_INT64_EXTENDED + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + __atomic_thread_fence(atomics::detail::convert_memory_order_to_gcc(order)); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + __atomic_signal_fence(atomics::detail::convert_memory_order_to_gcc(order)); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_ATOMIC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_ppc.hpp b/boost/atomic/detail/ops_gcc_ppc.hpp new file mode 100644 index 0000000000..1a1fbb7f94 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_ppc.hpp @@ -0,0 +1,775 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_ppc.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_ + +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +/* + Refer to: Motorola: "Programming Environments Manual for 32-Bit + Implementations of the PowerPC Architecture", Appendix E: + "Synchronization Programming Examples" for an explanation of what is + going on here (can be found on the web at various places by the + name "MPCFPE32B.pdf", Google is your friend...) + + Most of the atomic operations map to instructions in a relatively + straight-forward fashion, but "load"s may at first glance appear + a bit strange as they map to: + + lwz %rX, addr + cmpw %rX, %rX + bne- 1f + 1: + + That is, the CPU is forced to perform a branch that "formally" depends + on the value retrieved from memory. This scheme has an overhead of + about 1-2 clock cycles per load, but it allows to map "acquire" to + the "isync" instruction instead of "sync" uniformly and for all type + of atomic operations. Since "isync" has a cost of about 15 clock + cycles, while "sync" hast a cost of about 50 clock cycles, the small + penalty to atomic loads more than compensates for this. + + Byte- and halfword-sized atomic values are realized by encoding the + value to be represented into a word, performing sign/zero extension + as appropriate. This means that after add/sub operations the value + needs fixing up to accurately preserve the wrap-around semantic of + the smaller type. (Nothing special needs to be done for the bit-wise + and the "exchange type" operators as the compiler already sees to + it that values carried in registers are extended appropriately and + everything falls into place naturally). + + The register constraint "b" instructs gcc to use any register + except r0; this is sometimes required because the encoding for + r0 is used to signify "constant zero" in a number of instructions, + making r0 unusable in this place. For simplicity this constraint + is used everywhere since I am to lazy to look this up on a + per-instruction basis, and ppc has enough registers for this not + to pose a problem. +*/ + +// A note about memory_order_consume. Technically, this architecture allows to avoid +// unnecessary memory barrier after consume load since it supports data dependency ordering. +// However, some compiler optimizations may break a seemingly valid code relying on data +// dependency tracking by injecting bogus branches to aid out of order execution. +// This may happen not only in Boost.Atomic code but also in user's code, which we have no +// control of. See this thread: http://lists.boost.org/Archives/boost/2014/06/213890.php. +// For this reason we promote memory_order_consume to memory_order_acquire. + +struct gcc_ppc_operations_base +{ + static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT + { +#if defined(__powerpc64__) + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("sync" ::: "memory"); + else if ((order & memory_order_release) != 0) + __asm__ __volatile__ ("lwsync" ::: "memory"); +#else + if ((order & memory_order_release) != 0) + __asm__ __volatile__ ("sync" ::: "memory"); +#endif + } + + static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT + { + if ((order & (memory_order_consume | memory_order_acquire)) != 0) + __asm__ __volatile__ ("isync" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("sync" ::: "memory"); + } +}; + + +template< bool Signed > +struct operations< 4u, Signed > : + public gcc_ppc_operations_base +{ + typedef typename make_storage_type< 4u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + __asm__ __volatile__ + ( + "stw %1, %0\n" + : "+m" (storage) + : "r" (v) + ); + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v; + __asm__ __volatile__ + ( + "lwz %0, %1\n" + "cmpw %0, %0\n" + "bne- 1f\n" + "1:\n" + : "=&r" (v) + : "m" (storage) + : "cr0" + ); + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y1\n" + "stwcx. %2,%y1\n" + "bne- 1b\n" + : "=&b" (original), "+Z" (storage) + : "b" (v) + : "cr0" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + fence_before(success_order); + __asm__ __volatile__ + ( + "li %1, 0\n" + "lwarx %0,%y2\n" + "cmpw %0, %3\n" + "bne- 1f\n" + "stwcx. %4,%y2\n" + "bne- 1f\n" + "li %1, 1\n" + "1:" + : "=&b" (expected), "=&b" (success), "+Z" (storage) + : "b" (expected), "b" (desired) + : "cr0" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + fence_before(success_order); + __asm__ __volatile__ + ( + "li %1, 0\n" + "0: lwarx %0,%y2\n" + "cmpw %0, %3\n" + "bne- 1f\n" + "stwcx. %4,%y2\n" + "bne- 0b\n" + "li %1, 1\n" + "1:" + : "=&b" (expected), "=&b" (success), "+Z" (storage) + : "b" (expected), "b" (desired) + : "cr0" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "add %1,%0,%3\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "sub %1,%0,%3\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "and %1,%0,%3\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "or %1,%0,%3\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "xor %1,%0,%3\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + + +template< > +struct operations< 1u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "add %1,%0,%3\n" + "rlwinm %1, %1, 0, 0xff\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "sub %1,%0,%3\n" + "rlwinm %1, %1, 0, 0xff\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 1u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "add %1,%0,%3\n" + "extsb %1, %1\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "sub %1,%0,%3\n" + "extsb %1, %1\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } +}; + + +template< > +struct operations< 2u, false > : + public operations< 4u, false > +{ + typedef operations< 4u, false > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "add %1,%0,%3\n" + "rlwinm %1, %1, 0, 0xffff\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "sub %1,%0,%3\n" + "rlwinm %1, %1, 0, 0xffff\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } +}; + +template< > +struct operations< 2u, true > : + public operations< 4u, true > +{ + typedef operations< 4u, true > base_type; + typedef base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "add %1,%0,%3\n" + "extsh %1, %1\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "lwarx %0,%y2\n" + "sub %1,%0,%3\n" + "extsh %1, %1\n" + "stwcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } +}; + + +#if defined(__powerpc64__) + +template< bool Signed > +struct operations< 8u, Signed > : + public gcc_ppc_operations_base +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before(order); + __asm__ __volatile__ + ( + "std %1, %0\n" + : "+m" (storage) + : "r" (v) + ); + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v; + __asm__ __volatile__ + ( + "ld %0, %1\n" + "cmpd %0, %0\n" + "bne- 1f\n" + "1:\n" + : "=&b" (v) + : "m" (storage) + : "cr0" + ); + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y1\n" + "stdcx. %2,%y1\n" + "bne- 1b\n" + : "=&b" (original), "+Z" (storage) + : "b" (v) + : "cr0" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + fence_before(success_order); + __asm__ __volatile__ + ( + "li %1, 0\n" + "ldarx %0,%y2\n" + "cmpd %0, %3\n" + "bne- 1f\n" + "stdcx. %4,%y2\n" + "bne- 1f\n" + "li %1, 1\n" + "1:" + : "=&b" (expected), "=&b" (success), "+Z" (storage) + : "b" (expected), "b" (desired) + : "cr0" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + int success; + fence_before(success_order); + __asm__ __volatile__ + ( + "li %1, 0\n" + "0: ldarx %0,%y2\n" + "cmpd %0, %3\n" + "bne- 1f\n" + "stdcx. %4,%y2\n" + "bne- 0b\n" + "li %1, 1\n" + "1:" + : "=&b" (expected), "=&b" (success), "+Z" (storage) + : "b" (expected), "b" (desired) + : "cr0" + ); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + return !!success; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y2\n" + "add %1,%0,%3\n" + "stdcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y2\n" + "sub %1,%0,%3\n" + "stdcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y2\n" + "and %1,%0,%3\n" + "stdcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y2\n" + "or %1,%0,%3\n" + "stdcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type original, tmp; + fence_before(order); + __asm__ __volatile__ + ( + "1:\n" + "ldarx %0,%y2\n" + "xor %1,%0,%3\n" + "stdcx. %1,%y2\n" + "bne- 1b\n" + : "=&b" (original), "=&b" (tmp), "+Z" (storage) + : "b" (v) + : "cc" + ); + fence_after(order); + return original; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, 0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +#endif // defined(__powerpc64__) + + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + switch (order) + { + case memory_order_consume: + case memory_order_acquire: + __asm__ __volatile__ ("isync" ::: "memory"); + break; + case memory_order_release: +#if defined(__powerpc64__) + __asm__ __volatile__ ("lwsync" ::: "memory"); + break; +#endif + case memory_order_acq_rel: + case memory_order_seq_cst: + __asm__ __volatile__ ("sync" ::: "memory"); + break; + default:; + } +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_sparc.hpp b/boost/atomic/detail/ops_gcc_sparc.hpp new file mode 100644 index 0000000000..ea6df91dc6 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_sparc.hpp @@ -0,0 +1,245 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2010 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_sparc.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_SPARC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_SPARC_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +struct gcc_sparc_cas_base +{ + static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("membar #Sync" ::: "memory"); + else if ((order & memory_order_release) != 0) + __asm__ __volatile__ ("membar #StoreStore | #LoadStore" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("membar #Sync" ::: "memory"); + else if ((order & (memory_order_consume | memory_order_acquire)) != 0) + __asm__ __volatile__ ("membar #StoreStore | #LoadStore" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __asm__ __volatile__ ("membar #Sync" ::: "memory"); + } +}; + +template< bool Signed > +struct gcc_sparc_cas32 : + public gcc_sparc_cas_base +{ + typedef typename make_storage_type< 4u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before_store(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + storage_type previous = expected; + __asm__ __volatile__ + ( + "cas [%1], %2, %0" + : "+r" (desired) + : "r" (&storage), "r" (previous) + : "memory" + ); + const bool success = (desired == previous); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = desired; + return success; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public cas_based_operations< gcc_sparc_cas32< Signed > > +{ + typedef cas_based_operations< gcc_sparc_cas32< Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + __asm__ __volatile__ + ( + "swap [%1], %0" + : "+r" (v) + : "r" (&storage) + : "memory" + ); + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!exchange(storage, (storage_type)1, order); + } +}; + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 1u, Signed > +{ +}; + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 2u, Signed > +{ +}; + +template< bool Signed > +struct gcc_sparc_cas64 : + public gcc_sparc_cas_base +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before_store(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + fence_before(success_order); + storage_type previous = expected; + __asm__ __volatile__ + ( + "casx [%1], %2, %0" + : "+r" (desired) + : "r" (&storage), "r" (previous) + : "memory" + ); + const bool success = (desired == previous); + if (success) + fence_after(success_order); + else + fence_after(failure_order); + expected = desired; + return success; + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 8u, Signed > : + public cas_based_operations< gcc_sparc_cas64< Signed > > +{ +}; + + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + switch (order) + { + case memory_order_release: + __asm__ __volatile__ ("membar #StoreStore | #LoadStore" ::: "memory"); + break; + case memory_order_consume: + case memory_order_acquire: + __asm__ __volatile__ ("membar #LoadLoad | #LoadStore" ::: "memory"); + break; + case memory_order_acq_rel: + __asm__ __volatile__ ("membar #LoadLoad | #LoadStore | #StoreStore" ::: "memory"); + break; + case memory_order_seq_cst: + __asm__ __volatile__ ("membar #Sync" ::: "memory"); + break; + case memory_order_relaxed: + default: + break; + } +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_SPARC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_sync.hpp b/boost/atomic/detail/ops_gcc_sync.hpp new file mode 100644 index 0000000000..f4fc3331c6 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_sync.hpp @@ -0,0 +1,237 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2011 Helge Bahmann + * Copyright (c) 2013 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_sync.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_SYNC_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_SYNC_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +struct gcc_sync_operations_base +{ + static BOOST_FORCEINLINE void fence_before_store(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + __sync_synchronize(); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + __sync_synchronize(); + } + + static BOOST_FORCEINLINE void fence_after_load(memory_order order) BOOST_NOEXCEPT + { + if ((order & (memory_order_acquire | memory_order_consume)) != 0) + __sync_synchronize(); + } +}; + +template< typename T > +struct gcc_sync_operations : + public gcc_sync_operations_base +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before_store(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return __sync_fetch_and_add(&storage, v); + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return __sync_fetch_and_sub(&storage, v); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + // GCC docs mention that not all architectures may support full exchange semantics for this intrinsic. However, GCC's implementation of + // std::atomic<> uses this intrinsic unconditionally. We do so as well. In case if some architectures actually don't support this, we can always + // add a check here and fall back to a CAS loop. + if ((order & memory_order_release) != 0) + __sync_synchronize(); + return __sync_lock_test_and_set(&storage, v); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type expected2 = expected; + storage_type old_val = __sync_val_compare_and_swap(&storage, expected2, desired); + + if (old_val == expected2) + { + return true; + } + else + { + expected = old_val; + return false; + } + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return __sync_fetch_and_and(&storage, v); + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return __sync_fetch_and_or(&storage, v); + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return __sync_fetch_and_xor(&storage, v); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + __sync_synchronize(); + return !!__sync_lock_test_and_set(&storage, 1); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + __sync_lock_release(&storage); + if (order == memory_order_seq_cst) + __sync_synchronize(); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +#if BOOST_ATOMIC_INT8_LOCK_FREE > 0 +template< bool Signed > +struct operations< 1u, Signed > : +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1) + public gcc_sync_operations< typename make_storage_type< 1u, Signed >::type > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 2u, Signed >::type >, 1u, Signed > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 4u, Signed >::type >, 1u, Signed > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 8u, Signed >::type >, 1u, Signed > +#else + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 16u, Signed >::type >, 1u, Signed > +#endif +{ +}; +#endif + +#if BOOST_ATOMIC_INT16_LOCK_FREE > 0 +template< bool Signed > +struct operations< 2u, Signed > : +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2) + public gcc_sync_operations< typename make_storage_type< 2u, Signed >::type > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 4u, Signed >::type >, 2u, Signed > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 8u, Signed >::type >, 2u, Signed > +#else + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 16u, Signed >::type >, 2u, Signed > +#endif +{ +}; +#endif + +#if BOOST_ATOMIC_INT32_LOCK_FREE > 0 +template< bool Signed > +struct operations< 4u, Signed > : +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4) + public gcc_sync_operations< typename make_storage_type< 4u, Signed >::type > +#elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 8u, Signed >::type >, 4u, Signed > +#else + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 16u, Signed >::type >, 4u, Signed > +#endif +{ +}; +#endif + +#if BOOST_ATOMIC_INT64_LOCK_FREE > 0 +template< bool Signed > +struct operations< 8u, Signed > : +#if defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) + public gcc_sync_operations< typename make_storage_type< 8u, Signed >::type > +#else + public extending_cas_based_operations< gcc_sync_operations< typename make_storage_type< 16u, Signed >::type >, 8u, Signed > +#endif +{ +}; +#endif + +#if BOOST_ATOMIC_INT128_LOCK_FREE > 0 +template< bool Signed > +struct operations< 16u, Signed > : + public gcc_sync_operations< typename make_storage_type< 16u, Signed >::type > +{ +}; +#endif + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __sync_synchronize(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_SYNC_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_x86.hpp b/boost/atomic/detail/ops_gcc_x86.hpp new file mode 100644 index 0000000000..f18227f8a9 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_x86.hpp @@ -0,0 +1,510 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_x86.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_X86_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_X86_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) +#include +#include +#endif + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(__x86_64__) +#define BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "rdx" +#else +#define BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "edx" +#endif + +namespace boost { +namespace atomics { +namespace detail { + +struct gcc_x86_operations_base +{ + static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + __asm__ __volatile__ ("" ::: "memory"); + } + + static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_acquire) != 0) + __asm__ __volatile__ ("" ::: "memory"); + } +}; + +template< typename T, typename Derived > +struct gcc_x86_operations : + public gcc_x86_operations_base +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + if (order != memory_order_seq_cst) + { + fence_before(order); + storage = v; + fence_after(order); + } + else + { + Derived::exchange(storage, v, order); + } + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + return Derived::fetch_add(storage, -v, order); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return Derived::compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!Derived::exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, (storage_type)0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 1u, Signed > : + public gcc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > +{ + typedef gcc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "lock; xaddb %0, %1" + : "+q" (v), "+m" (storage) + : + : "cc", "memory" + ); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "xchgb %0, %1" + : "+q" (v), "+m" (storage) + : + : "memory" + ); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchgb %3, %1\n\t" + "sete %2" + : "+a" (previous), "+m" (storage), "=q" (success) + : "q" (desired) + : "cc", "memory" + ); + expected = previous; + return success; + } + +#define BOOST_ATOMIC_DETAIL_CAS_LOOP(op, argument, result)\ + __asm__ __volatile__\ + (\ + "xor %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER ", %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "\n\t"\ + ".align 16\n\t"\ + "1: movb %[arg], %%dl\n\t"\ + op " %%al, %%dl\n\t"\ + "lock; cmpxchgb %%dl, %[storage]\n\t"\ + "jne 1b"\ + : [res] "+a" (result), [storage] "+m" (storage)\ + : [arg] "q" (argument)\ + : "cc", BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER, "memory"\ + ) + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("andb", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("orb", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("xorb", v, res); + return res; + } + +#undef BOOST_ATOMIC_DETAIL_CAS_LOOP +}; + +template< bool Signed > +struct operations< 2u, Signed > : + public gcc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > +{ + typedef gcc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "lock; xaddw %0, %1" + : "+q" (v), "+m" (storage) + : + : "cc", "memory" + ); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "xchgw %0, %1" + : "+q" (v), "+m" (storage) + : + : "memory" + ); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchgw %3, %1\n\t" + "sete %2" + : "+a" (previous), "+m" (storage), "=q" (success) + : "q" (desired) + : "cc", "memory" + ); + expected = previous; + return success; + } + +#define BOOST_ATOMIC_DETAIL_CAS_LOOP(op, argument, result)\ + __asm__ __volatile__\ + (\ + "xor %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER ", %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "\n\t"\ + ".align 16\n\t"\ + "1: movw %[arg], %%dx\n\t"\ + op " %%ax, %%dx\n\t"\ + "lock; cmpxchgw %%dx, %[storage]\n\t"\ + "jne 1b"\ + : [res] "+a" (result), [storage] "+m" (storage)\ + : [arg] "q" (argument)\ + : "cc", BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER, "memory"\ + ) + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("andw", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("orw", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("xorw", v, res); + return res; + } + +#undef BOOST_ATOMIC_DETAIL_CAS_LOOP +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public gcc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > +{ + typedef gcc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "lock; xaddl %0, %1" + : "+r" (v), "+m" (storage) + : + : "cc", "memory" + ); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "xchgl %0, %1" + : "+r" (v), "+m" (storage) + : + : "memory" + ); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchgl %3, %1\n\t" + "sete %2" + : "+a" (previous), "+m" (storage), "=q" (success) + : "r" (desired) + : "cc", "memory" + ); + expected = previous; + return success; + } + +#define BOOST_ATOMIC_DETAIL_CAS_LOOP(op, argument, result)\ + __asm__ __volatile__\ + (\ + "xor %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER ", %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "\n\t"\ + ".align 16\n\t"\ + "1: movl %[arg], %%edx\n\t"\ + op " %%eax, %%edx\n\t"\ + "lock; cmpxchgl %%edx, %[storage]\n\t"\ + "jne 1b"\ + : [res] "+a" (result), [storage] "+m" (storage)\ + : [arg] "r" (argument)\ + : "cc", BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER, "memory"\ + ) + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("andl", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("orl", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("xorl", v, res); + return res; + } + +#undef BOOST_ATOMIC_DETAIL_CAS_LOOP +}; + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) + +template< bool Signed > +struct operations< 8u, Signed > : + public cas_based_operations< gcc_dcas_x86< Signed > > +{ +}; + +#elif defined(__x86_64__) + +template< bool Signed > +struct operations< 8u, Signed > : + public gcc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > +{ + typedef gcc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "lock; xaddq %0, %1" + : "+r" (v), "+m" (storage) + : + : "cc", "memory" + ); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + __asm__ __volatile__ + ( + "xchgq %0, %1" + : "+r" (v), "+m" (storage) + : + : "memory" + ); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchgq %3, %1\n\t" + "sete %2" + : "+a" (previous), "+m" (storage), "=q" (success) + : "r" (desired) + : "cc", "memory" + ); + expected = previous; + return success; + } + +#define BOOST_ATOMIC_DETAIL_CAS_LOOP(op, argument, result)\ + __asm__ __volatile__\ + (\ + "xor %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER ", %%" BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER "\n\t"\ + ".align 16\n\t"\ + "1: movq %[arg], %%rdx\n\t"\ + op " %%rax, %%rdx\n\t"\ + "lock; cmpxchgq %%rdx, %[storage]\n\t"\ + "jne 1b"\ + : [res] "+a" (result), [storage] "+m" (storage)\ + : [arg] "r" (argument)\ + : "cc", BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER, "memory"\ + ) + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("andq", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("orq", v, res); + return res; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type res = storage; + BOOST_ATOMIC_DETAIL_CAS_LOOP("xorq", v, res); + return res; + } + +#undef BOOST_ATOMIC_DETAIL_CAS_LOOP +}; + +#endif + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +template< bool Signed > +struct operations< 16u, Signed > : + public cas_based_operations< gcc_dcas_x86_64< Signed > > +{ +}; + +#endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order == memory_order_seq_cst) + { + __asm__ __volatile__ + ( +#if defined(__x86_64__) || defined(__SSE2__) + "mfence\n" +#else + "lock; addl $0, (%%esp)\n" +#endif + ::: "memory" + ); + } + else if ((order & (memory_order_acquire | memory_order_release)) != 0) + { + __asm__ __volatile__ ("" ::: "memory"); + } +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#undef BOOST_ATOMIC_DETAIL_TEMP_CAS_REGISTER + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_X86_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_gcc_x86_dcas.hpp b/boost/atomic/detail/ops_gcc_x86_dcas.hpp new file mode 100644 index 0000000000..5e005359b5 --- /dev/null +++ b/boost/atomic/detail/ops_gcc_x86_dcas.hpp @@ -0,0 +1,308 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_gcc_x86_dcas.hpp + * + * This header contains implementation of the double-width CAS primitive for x86. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_ + +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) + +template< bool Signed > +struct gcc_dcas_x86 +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + if ((((uint32_t)&storage) & 0x00000007) == 0) + { +#if defined(__SSE2__) + __asm__ __volatile__ + ( +#if defined(__AVX__) + "vmovq %1, %%xmm4\n\t" + "vmovq %%xmm4, %0\n\t" +#else + "movq %1, %%xmm4\n\t" + "movq %%xmm4, %0\n\t" +#endif + : "=m" (storage) + : "m" (v) + : "memory", "xmm4" + ); +#else + __asm__ __volatile__ + ( + "fildll %1\n\t" + "fistpll %0\n\t" + : "=m" (storage) + : "m" (v) + : "memory" + ); +#endif + } + else + { +#if defined(__PIC__) + uint32_t scratch; + __asm__ __volatile__ + ( + "movl %%ebx, %[scratch]\n\t" + "movl %[value_lo], %%ebx\n\t" + "movl 0(%[dest]), %%eax\n\t" + "movl 4(%[dest]), %%edx\n\t" + ".align 16\n\t" + "1: lock; cmpxchg8b 0(%[dest])\n\t" + "jne 1b\n\t" + "movl %[scratch], %%ebx" + : [scratch] "=m,m" (scratch) + : [value_lo] "a,a" ((uint32_t)v), "c,c" ((uint32_t)(v >> 32)), [dest] "D,S" (&storage) + : "cc", "edx", "memory" + ); +#else + __asm__ __volatile__ + ( + "movl 0(%[dest]), %%eax\n\t" + "movl 4(%[dest]), %%edx\n\t" + ".align 16\n\t" + "1: lock; cmpxchg8b 0(%[dest])\n\t" + "jne 1b\n\t" + : + : [value_lo] "b,b" ((uint32_t)v), "c,c" ((uint32_t)(v >> 32)), [dest] "D,S" (&storage) + : "cc", "eax", "edx", "memory" + ); +#endif + } + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT + { + storage_type value; + + if ((((uint32_t)&storage) & 0x00000007) == 0) + { +#if defined(__SSE2__) + __asm__ __volatile__ + ( +#if defined(__AVX__) + "vmovq %1, %%xmm4\n\t" + "vmovq %%xmm4, %0\n\t" +#else + "movq %1, %%xmm4\n\t" + "movq %%xmm4, %0\n\t" +#endif + : "=m" (value) + : "m" (storage) + : "memory", "xmm4" + ); +#else + __asm__ __volatile__ + ( + "fildll %1\n\t" + "fistpll %0\n\t" + : "=m" (value) + : "m" (storage) + : "memory" + ); +#endif + } + else + { +#if defined(__clang__) + // Clang cannot allocate eax:edx register pairs but it has sync intrinsics + value = __sync_val_compare_and_swap(&storage, (storage_type)0, (storage_type)0); +#else + // We don't care for comparison result here; the previous value will be stored into value anyway. + // Also we don't care for ebx and ecx values, they just have to be equal to eax and edx before cmpxchg8b. + __asm__ __volatile__ + ( + "movl %%ebx, %%eax\n\t" + "movl %%ecx, %%edx\n\t" + "lock; cmpxchg8b %[storage]" + : "=&A" (value) + : [storage] "m" (storage) + : "cc", "memory" + ); +#endif + } + + return value; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { +#if defined(__clang__) + // Clang cannot allocate eax:edx register pairs but it has sync intrinsics + storage_type old_expected = expected; + expected = __sync_val_compare_and_swap(&storage, old_expected, desired); + return expected == old_expected; +#elif defined(__PIC__) + // Make sure ebx is saved and restored properly in case + // of position independent code. To make this work + // setup register constraints such that ebx can not be + // used by accident e.g. as base address for the variable + // to be modified. Accessing "scratch" should always be okay, + // as it can only be placed on the stack (and therefore + // accessed through ebp or esp only). + // + // In theory, could push/pop ebx onto/off the stack, but movs + // to a prepared stack slot turn out to be faster. + + uint32_t scratch; + bool success; + __asm__ __volatile__ + ( + "movl %%ebx, %[scratch]\n\t" + "movl %[desired_lo], %%ebx\n\t" + "lock; cmpxchg8b %[dest]\n\t" + "movl %[scratch], %%ebx\n\t" + "sete %[success]" + : "+A,A,A,A,A,A" (expected), [dest] "+m,m,m,m,m,m" (storage), [scratch] "=m,m,m,m,m,m" (scratch), [success] "=q,m,q,m,q,m" (success) + : [desired_lo] "S,S,D,D,m,m" ((uint32_t)desired), "c,c,c,c,c,c" ((uint32_t)(desired >> 32)) + : "cc", "memory" + ); + return success; +#else + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchg8b %[dest]\n\t" + "sete %[success]" + : "+A,A" (expected), [dest] "+m,m" (storage), [success] "=q,m" (success) + : "b,b" ((uint32_t)desired), "c,c" ((uint32_t)(desired >> 32)) + : "cc", "memory" + ); + return success; +#endif + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +#endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +template< bool Signed > +struct gcc_dcas_x86_64 +{ + typedef typename make_storage_type< 16u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + uint64_t const* p_value = (uint64_t const*)&v; + __asm__ __volatile__ + ( + "movq 0(%[dest]), %%rax\n\t" + "movq 8(%[dest]), %%rdx\n\t" + ".align 16\n\t" + "1: lock; cmpxchg16b 0(%[dest])\n\t" + "jne 1b" + : + : "b" (p_value[0]), "c" (p_value[1]), [dest] "r" (&storage) + : "cc", "rax", "rdx", "memory" + ); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT + { +#if defined(__clang__) + // Clang cannot allocate rax:rdx register pairs but it has sync intrinsics + storage_type value = storage_type(); + return __sync_val_compare_and_swap(&storage, value, value); +#else + storage_type value; + + // We don't care for comparison result here; the previous value will be stored into value anyway. + // Also we don't care for rbx and rcx values, they just have to be equal to rax and rdx before cmpxchg16b. + __asm__ __volatile__ + ( + "movq %%rbx, %%rax\n\t" + "movq %%rcx, %%rdx\n\t" + "lock; cmpxchg16b %[storage]" + : "=&A" (value) + : [storage] "m" (storage) + : "cc", "memory" + ); + + return value; +#endif + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { +#if defined(__clang__) + // Clang cannot allocate rax:rdx register pairs but it has sync intrinsics + storage_type old_expected = expected; + expected = __sync_val_compare_and_swap(&storage, old_expected, desired); + return expected == old_expected; +#else + uint64_t const* p_desired = (uint64_t const*)&desired; + bool success; + __asm__ __volatile__ + ( + "lock; cmpxchg16b %[dest]\n\t" + "sete %[success]" + : "+A,A" (expected), [dest] "+m,m" (storage), [success] "=q,m" (success) + : "b,b" (p_desired[0]), "c,c" (p_desired[1]) + : "cc", "memory" + ); + return success; +#endif + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +#endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_linux_arm.hpp b/boost/atomic/detail/ops_linux_arm.hpp new file mode 100644 index 0000000000..25167b1974 --- /dev/null +++ b/boost/atomic/detail/ops_linux_arm.hpp @@ -0,0 +1,177 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009, 2011 Helge Bahmann + * Copyright (c) 2009 Phil Endecott + * Copyright (c) 2013 Tim Blechmann + * Linux-specific code by Phil Endecott + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_linux_arm.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_LINUX_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_LINUX_ARM_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +// Different ARM processors have different atomic instructions. In particular, +// architecture versions before v6 (which are still in widespread use, e.g. the +// Intel/Marvell XScale chips like the one in the NSLU2) have only atomic swap. +// On Linux the kernel provides some support that lets us abstract away from +// these differences: it provides emulated CAS and barrier functions at special +// addresses that are guaranteed not to be interrupted by the kernel. Using +// this facility is slightly slower than inline assembler would be, but much +// faster than a system call. +// +// While this emulated CAS is "strong" in the sense that it does not fail +// "spuriously" (i.e.: it never fails to perform the exchange when the value +// found equals the value expected), it does not return the found value on +// failure. To satisfy the atomic API, compare_exchange_{weak|strong} must +// return the found value on failure, and we have to manually load this value +// after the emulated CAS reports failure. This in turn introduces a race +// between the CAS failing (due to the "wrong" value being found) and subsequently +// loading (which might turn up the "right" value). From an application's +// point of view this looks like "spurious failure", and therefore the +// emulated CAS is only good enough to provide compare_exchange_weak +// semantics. + +struct linux_arm_cas_base +{ + static BOOST_FORCEINLINE void fence_before_store(memory_order order) BOOST_NOEXCEPT + { + if ((order & memory_order_release) != 0) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + if (order == memory_order_seq_cst) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void fence_after_load(memory_order order) BOOST_NOEXCEPT + { + if ((order & (memory_order_consume | memory_order_acquire)) != 0) + hardware_full_fence(); + } + + static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT + { + typedef void (*kernel_dmb_t)(void); + ((kernel_dmb_t)0xffff0fa0)(); + } +}; + +template< bool Signed > +struct linux_arm_cas : + public linux_arm_cas_base +{ + typedef typename make_storage_type< 4u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + fence_before_store(order); + storage = v; + fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + while (true) + { + storage_type tmp = expected; + if (compare_exchange_weak(storage, tmp, desired, success_order, failure_order)) + return true; + if (tmp != expected) + { + expected = tmp; + return false; + } + } + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + typedef storage_type (*kernel_cmpxchg32_t)(storage_type oldval, storage_type newval, volatile storage_type* ptr); + + if (((kernel_cmpxchg32_t)0xffff0fc0)(expected, desired, &storage) == 0) + { + return true; + } + else + { + expected = storage; + return false; + } + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< cas_based_operations< linux_arm_cas< Signed > >, 1u, Signed > +{ +}; + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< cas_based_operations< linux_arm_cas< Signed > >, 2u, Signed > +{ +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public cas_based_operations< linux_arm_cas< Signed > > +{ +}; + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + linux_arm_cas_base::hardware_full_fence(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + __asm__ __volatile__ ("" ::: "memory"); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_LINUX_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_msvc_arm.hpp b/boost/atomic/detail/ops_msvc_arm.hpp new file mode 100644 index 0000000000..349f7a5ae8 --- /dev/null +++ b/boost/atomic/detail/ops_msvc_arm.hpp @@ -0,0 +1,820 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_msvc_arm.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_MSVC_ARM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_MSVC_ARM_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#define BOOST_ATOMIC_DETAIL_ARM_LOAD8(p) __iso_volatile_load8((const volatile __int8*)(p)) +#define BOOST_ATOMIC_DETAIL_ARM_LOAD16(p) __iso_volatile_load16((const volatile __int16*)(p)) +#define BOOST_ATOMIC_DETAIL_ARM_LOAD32(p) __iso_volatile_load32((const volatile __int32*)(p)) +#define BOOST_ATOMIC_DETAIL_ARM_LOAD64(p) __iso_volatile_load64((const volatile __int64*)(p)) +#define BOOST_ATOMIC_DETAIL_ARM_STORE8(p, v) __iso_volatile_store8((volatile __int8*)(p), (__int8)(v)) +#define BOOST_ATOMIC_DETAIL_ARM_STORE16(p, v) __iso_volatile_store16((volatile __int16*)(p), (__int16)(v)) +#define BOOST_ATOMIC_DETAIL_ARM_STORE32(p, v) __iso_volatile_store32((volatile __int32*)(p), (__int32)(v)) +#define BOOST_ATOMIC_DETAIL_ARM_STORE64(p, v) __iso_volatile_store64((volatile __int64*)(p), (__int64)(v)) + +namespace boost { +namespace atomics { +namespace detail { + +// A note about memory_order_consume. Technically, this architecture allows to avoid +// unnecessary memory barrier after consume load since it supports data dependency ordering. +// However, some compiler optimizations may break a seemingly valid code relying on data +// dependency tracking by injecting bogus branches to aid out of order execution. +// This may happen not only in Boost.Atomic code but also in user's code, which we have no +// control of. See this thread: http://lists.boost.org/Archives/boost/2014/06/213890.php. +// For this reason we promote memory_order_consume to memory_order_acquire. + +struct msvc_arm_operations_base +{ + static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT + { + __dmb(0xB); // _ARM_BARRIER_ISH, see armintr.h from MSVC 11 and later + } + + static BOOST_FORCEINLINE void fence_before_store(memory_order order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + + if ((order & memory_order_release) != 0) + hardware_full_fence(); + + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + + if (order == memory_order_seq_cst) + hardware_full_fence(); + + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE void fence_after_load(memory_order order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + + if ((order & (memory_order_consume | memory_order_acquire)) != 0) + hardware_full_fence(); + + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE BOOST_CONSTEXPR memory_order cas_common_order(memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + // Combine order flags together and promote memory_order_consume to memory_order_acquire + return static_cast< memory_order >(((failure_order | success_order) & ~memory_order_consume) | (((failure_order | success_order) & memory_order_consume) << 1u)); + } +}; + +template< typename T, typename Derived > +struct msvc_arm_operations : + public msvc_arm_operations_base +{ + typedef T storage_type; + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + typedef typename make_signed< storage_type >::type signed_storage_type; + return Derived::fetch_add(storage, static_cast< storage_type >(-static_cast< signed_storage_type >(v)), order); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return Derived::compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!Derived::exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + Derived::store(storage, (storage_type)0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 1u, Signed > : + public msvc_arm_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > +{ + typedef msvc_arm_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before_store(order); + BOOST_ATOMIC_DETAIL_ARM_STORE8(&storage, v); + base_type::fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = BOOST_ATOMIC_DETAIL_ARM_LOAD8(&storage); + base_type::fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + storage_type previous = expected, old_val; + + switch (cas_common_order(success_order, failure_order)) + { + case memory_order_relaxed: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_RELAXED(&storage, desired, previous)); + break; + case memory_order_consume: + case memory_order_acquire: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_ACQUIRE(&storage, desired, previous)); + break; + case memory_order_release: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8_RELEASE(&storage, desired, previous)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8(&storage, desired, previous)); + break; + } + expected = old_val; + + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8(&storage, v)); + break; + } + return v; + } +}; + +template< bool Signed > +struct operations< 2u, Signed > : + public msvc_arm_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > +{ + typedef msvc_arm_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before_store(order); + BOOST_ATOMIC_DETAIL_ARM_STORE16(&storage, v); + base_type::fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = BOOST_ATOMIC_DETAIL_ARM_LOAD16(&storage); + base_type::fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + storage_type previous = expected, old_val; + + switch (cas_common_order(success_order, failure_order)) + { + case memory_order_relaxed: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_RELAXED(&storage, desired, previous)); + break; + case memory_order_consume: + case memory_order_acquire: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_ACQUIRE(&storage, desired, previous)); + break; + case memory_order_release: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16_RELEASE(&storage, desired, previous)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16(&storage, desired, previous)); + break; + } + expected = old_val; + + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16(&storage, v)); + break; + } + return v; + } +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public msvc_arm_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > +{ + typedef msvc_arm_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before_store(order); + BOOST_ATOMIC_DETAIL_ARM_STORE32(&storage, v); + base_type::fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = BOOST_ATOMIC_DETAIL_ARM_LOAD32(&storage); + base_type::fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + storage_type previous = expected, old_val; + + switch (cas_common_order(success_order, failure_order)) + { + case memory_order_relaxed: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_RELAXED(&storage, desired, previous)); + break; + case memory_order_consume: + case memory_order_acquire: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_ACQUIRE(&storage, desired, previous)); + break; + case memory_order_release: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE_RELEASE(&storage, desired, previous)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(&storage, desired, previous)); + break; + } + expected = old_val; + + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR(&storage, v)); + break; + } + return v; + } +}; + +template< bool Signed > +struct operations< 8u, Signed > : + public msvc_arm_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > +{ + typedef msvc_arm_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before_store(order); + BOOST_ATOMIC_DETAIL_ARM_STORE64(&storage, v); + base_type::fence_after_store(order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = BOOST_ATOMIC_DETAIL_ARM_LOAD64(&storage); + base_type::fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + storage_type previous = expected, old_val; + + switch (cas_common_order(success_order, failure_order)) + { + case memory_order_relaxed: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_RELAXED(&storage, desired, previous)); + break; + case memory_order_consume: + case memory_order_acquire: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_ACQUIRE(&storage, desired, previous)); + break; + case memory_order_release: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64_RELEASE(&storage, desired, previous)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(&storage, desired, previous)); + break; + } + expected = old_val; + + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64(&storage, v)); + break; + } + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + switch (order) + { + case memory_order_relaxed: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64_RELAXED(&storage, v)); + break; + case memory_order_consume: + case memory_order_acquire: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64_ACQUIRE(&storage, v)); + break; + case memory_order_release: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64_RELEASE(&storage, v)); + break; + case memory_order_acq_rel: + case memory_order_seq_cst: + default: + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64(&storage, v)); + break; + } + return v; + } +}; + + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + if (order != memory_order_relaxed) + msvc_arm_operations_base::hardware_full_fence(); + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#undef BOOST_ATOMIC_DETAIL_ARM_LOAD8 +#undef BOOST_ATOMIC_DETAIL_ARM_LOAD16 +#undef BOOST_ATOMIC_DETAIL_ARM_LOAD32 +#undef BOOST_ATOMIC_DETAIL_ARM_LOAD64 +#undef BOOST_ATOMIC_DETAIL_ARM_STORE8 +#undef BOOST_ATOMIC_DETAIL_ARM_STORE16 +#undef BOOST_ATOMIC_DETAIL_ARM_STORE32 +#undef BOOST_ATOMIC_DETAIL_ARM_STORE64 + +#endif // BOOST_ATOMIC_DETAIL_OPS_MSVC_ARM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_msvc_common.hpp b/boost/atomic/detail/ops_msvc_common.hpp new file mode 100644 index 0000000000..53628f3600 --- /dev/null +++ b/boost/atomic/detail/ops_msvc_common.hpp @@ -0,0 +1,38 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_msvc_common.hpp + * + * This header contains common tools for MSVC implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_MSVC_COMMON_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_MSVC_COMMON_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +// Define compiler barriers +#if defined(__INTEL_COMPILER) +#define BOOST_ATOMIC_DETAIL_COMPILER_BARRIER() __memory_barrier() +#elif defined(_MSC_VER) && !defined(_WIN32_WCE) +extern "C" void _ReadWriteBarrier(void); +#pragma intrinsic(_ReadWriteBarrier) +#define BOOST_ATOMIC_DETAIL_COMPILER_BARRIER() _ReadWriteBarrier() +#endif + +#ifndef BOOST_ATOMIC_DETAIL_COMPILER_BARRIER +#define BOOST_ATOMIC_DETAIL_COMPILER_BARRIER() +#endif + +#endif // BOOST_ATOMIC_DETAIL_OPS_MSVC_COMMON_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_msvc_x86.hpp b/boost/atomic/detail/ops_msvc_x86.hpp new file mode 100644 index 0000000000..501d9c622d --- /dev/null +++ b/boost/atomic/detail/ops_msvc_x86.hpp @@ -0,0 +1,879 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_msvc_x86.hpp + * + * This header contains implementation of the \c operations template. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) +#include +#include +#endif +#include +#if !defined(_M_IX86) && !(defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8) && defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16)) +#include +#endif + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(BOOST_MSVC) +#pragma warning(push) +// frame pointer register 'ebx' modified by inline assembly code. See the note below. +#pragma warning(disable: 4731) +#endif + +#if defined(_MSC_VER) && (defined(_M_AMD64) || (defined(_M_IX86) && defined(_M_IX86_FP) && _M_IX86_FP >= 2)) +extern "C" void _mm_mfence(void); +#if defined(BOOST_MSVC) +#pragma intrinsic(_mm_mfence) +#endif +#endif + +namespace boost { +namespace atomics { +namespace detail { + +/* + * Implementation note for asm blocks. + * + * http://msdn.microsoft.com/en-us/data/k1a8ss06%28v=vs.105%29 + * + * Some SSE types require eight-byte stack alignment, forcing the compiler to emit dynamic stack-alignment code. + * To be able to access both the local variables and the function parameters after the alignment, the compiler + * maintains two frame pointers. If the compiler performs frame pointer omission (FPO), it will use EBP and ESP. + * If the compiler does not perform FPO, it will use EBX and EBP. To ensure code runs correctly, do not modify EBX + * in asm code if the function requires dynamic stack alignment as it could modify the frame pointer. + * Either move the eight-byte aligned types out of the function, or avoid using EBX. + * + * Since we have no way of knowing that the compiler uses FPO, we have to always save and restore ebx + * whenever we have to clobber it. Additionally, we disable warning C4731 above so that the compiler + * doesn't spam about ebx use. + */ + +struct msvc_x86_operations_base +{ + static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT + { +#if defined(_MSC_VER) && (defined(_M_AMD64) || (defined(_M_IX86) && defined(_M_IX86_FP) && _M_IX86_FP >= 2)) + // Use mfence only if SSE2 is available + _mm_mfence(); +#else + long tmp; + BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&tmp, 0); +#endif + } + + static BOOST_FORCEINLINE void fence_before(memory_order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE void fence_after(memory_order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE void fence_after_load(memory_order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + + // On x86 and x86_64 there is no need for a hardware barrier, + // even if seq_cst memory order is requested, because all + // seq_cst writes are implemented with lock-prefixed operations + // or xchg which has implied lock prefix. Therefore normal loads + // are already ordered with seq_cst stores on these architectures. + } +}; + +template< typename T, typename Derived > +struct msvc_x86_operations : + public msvc_x86_operations_base +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + if (order != memory_order_seq_cst) + { + fence_before(order); + storage = v; + fence_after(order); + } + else + { + Derived::exchange(storage, v, order); + } + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + storage_type v = storage; + fence_after_load(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + typedef typename make_signed< storage_type >::type signed_storage_type; + return Derived::fetch_add(storage, static_cast< storage_type >(-static_cast< signed_storage_type >(v)), order); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return Derived::compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!Derived::exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, (storage_type)0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public msvc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&storage, v)); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(&storage, desired, previous)); + expected = old_val; + return (previous == old_val); + } + +#if defined(BOOST_ATOMIC_INTERLOCKED_AND) + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND(&storage, v)); + } +#else + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res & v, order, memory_order_relaxed)) {} + return res; + } +#endif + +#if defined(BOOST_ATOMIC_INTERLOCKED_OR) + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR(&storage, v)); + } +#else + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res | v, order, memory_order_relaxed)) {} + return res; + } +#endif + +#if defined(BOOST_ATOMIC_INTERLOCKED_XOR) + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR(&storage, v)); + } +#else + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res ^ v, order, memory_order_relaxed)) {} + return res; + } +#endif +}; + +#if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8) + +template< bool Signed > +struct operations< 1u, Signed > : + public msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8(&storage, v)); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8(&storage, desired, previous)); + expected = old_val; + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8(&storage, v)); + } +}; + +#elif defined(_M_IX86) + +template< bool Signed > +struct operations< 1u, Signed > : + public msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + __asm + { + mov edx, storage + movzx eax, v + lock xadd byte ptr [edx], al + mov v, al + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + __asm + { + mov edx, storage + movzx eax, v + xchg byte ptr [edx], al + mov v, al + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order) BOOST_NOEXCEPT + { + base_type::fence_before(success_order); + bool success; + __asm + { + mov esi, expected + mov edi, storage + movzx eax, byte ptr [esi] + movzx edx, desired + lock cmpxchg byte ptr [edi], dl + mov byte ptr [esi], al + sete success + }; + // The success and failure fences are equivalent anyway + base_type::fence_after(success_order); + return success; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, byte ptr [edi] + align 16 + again: + mov dl, al + and dl, bl + lock cmpxchg byte ptr [edi], dl + jne again + mov v, al + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, byte ptr [edi] + align 16 + again: + mov dl, al + or dl, bl + lock cmpxchg byte ptr [edi], dl + jne again + mov v, al + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, byte ptr [edi] + align 16 + again: + mov dl, al + xor dl, bl + lock cmpxchg byte ptr [edi], dl + jne again + mov v, al + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } +}; + +#else + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 1u, Signed > +{ +}; + +#endif + +#if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16) + +template< bool Signed > +struct operations< 2u, Signed > : + public msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16(&storage, v)); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16(&storage, desired, previous)); + expected = old_val; + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16(&storage, v)); + } +}; + +#elif defined(_M_IX86) + +template< bool Signed > +struct operations< 2u, Signed > : + public msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + __asm + { + mov edx, storage + movzx eax, v + lock xadd word ptr [edx], ax + mov v, ax + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + __asm + { + mov edx, storage + movzx eax, v + xchg word ptr [edx], ax + mov v, ax + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order) BOOST_NOEXCEPT + { + base_type::fence_before(success_order); + bool success; + __asm + { + mov esi, expected + mov edi, storage + movzx eax, word ptr [esi] + movzx edx, desired + lock cmpxchg word ptr [edi], dx + mov word ptr [esi], ax + sete success + }; + // The success and failure fences are equivalent anyway + base_type::fence_after(success_order); + return success; + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, word ptr [edi] + align 16 + again: + mov dx, ax + and dx, bx + lock cmpxchg word ptr [edi], dx + jne again + mov v, ax + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, word ptr [edi] + align 16 + again: + mov dx, ax + or dx, bx + lock cmpxchg word ptr [edi], dx + jne again + mov v, ax + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + int backup; + __asm + { + mov backup, ebx + xor edx, edx + mov edi, storage + movzx ebx, v + movzx eax, word ptr [edi] + align 16 + again: + mov dx, ax + xor dx, bx + lock cmpxchg word ptr [edi], dx + jne again + mov v, ax + mov ebx, backup + }; + base_type::fence_after(order); + return v; + } +}; + +#else + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 2u, Signed > +{ +}; + +#endif + + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) + +template< bool Signed > +struct msvc_dcas_x86 +{ + typedef typename make_storage_type< 8u, Signed >::type storage_type; + + // Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 3A, 8.1.1. Guaranteed Atomic Operations: + // + // The Pentium processor (and newer processors since) guarantees that the following additional memory operations will always be carried out atomically: + // * Reading or writing a quadword aligned on a 64-bit boundary + // + // Luckily, the memory is almost always 8-byte aligned in our case because atomic<> uses 64 bit native types for storage and dynamic memory allocations + // have at least 8 byte alignment. The only unfortunate case is when atomic is placeod on the stack and it is not 8-byte aligned (like on 32 bit Windows). + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type volatile* p = &storage; + if (((uint32_t)p & 0x00000007) == 0) + { +#if defined(_M_IX86_FP) && _M_IX86_FP >= 2 +#if defined(__AVX__) + __asm + { + mov edx, p + vmovq xmm4, v + vmovq qword ptr [edx], xmm4 + }; +#else + __asm + { + mov edx, p + movq xmm4, v + movq qword ptr [edx], xmm4 + }; +#endif +#else + __asm + { + mov edx, p + fild v + fistp qword ptr [edx] + }; +#endif + } + else + { + int backup; + __asm + { + mov backup, ebx + mov edi, p + mov ebx, dword ptr [v] + mov ecx, dword ptr [v + 4] + mov eax, dword ptr [edi] + mov edx, dword ptr [edi + 4] + align 16 + again: + lock cmpxchg8b qword ptr [edi] + jne again + mov ebx, backup + }; + } + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT + { + storage_type const volatile* p = &storage; + storage_type value; + + if (((uint32_t)p & 0x00000007) == 0) + { +#if defined(_M_IX86_FP) && _M_IX86_FP >= 2 +#if defined(__AVX__) + __asm + { + mov edx, p + vmovq xmm4, qword ptr [edx] + vmovq value, xmm4 + }; +#else + __asm + { + mov edx, p + movq xmm4, qword ptr [edx] + movq value, xmm4 + }; +#endif +#else + __asm + { + mov edx, p + fild qword ptr [edx] + fistp value + }; +#endif + } + else + { + // We don't care for comparison result here; the previous value will be stored into value anyway. + // Also we don't care for ebx and ecx values, they just have to be equal to eax and edx before cmpxchg8b. + __asm + { + mov edi, p + mov eax, ebx + mov edx, ecx + lock cmpxchg8b qword ptr [edi] + mov dword ptr [value], eax + mov dword ptr [value + 4], edx + }; + } + + return value; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type volatile* p = &storage; +#if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64) + const storage_type old_val = (storage_type)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(p, desired, expected); + const bool result = (old_val == expected); + expected = old_val; + return result; +#else + bool result; + int backup; + __asm + { + mov backup, ebx + mov edi, p + mov esi, expected + mov ebx, dword ptr [desired] + mov ecx, dword ptr [desired + 4] + mov eax, dword ptr [esi] + mov edx, dword ptr [esi + 4] + lock cmpxchg8b qword ptr [edi] + mov dword ptr [esi], eax + mov dword ptr [esi + 4], edx + mov ebx, backup + sete result + }; + return result; +#endif + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 8u, Signed > : + public cas_based_operations< msvc_dcas_x86< Signed > > +{ +}; + +#elif defined(_M_AMD64) + +template< bool Signed > +struct operations< 8u, Signed > : + public msvc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > +{ + typedef msvc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(&storage, v)); + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(&storage, desired, previous)); + expected = old_val; + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64(&storage, v)); + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64(&storage, v)); + } +}; + +#endif + +#if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +template< bool Signed > +struct msvc_dcas_x86_64 +{ + typedef typename make_storage_type< 16u, Signed >::type storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT + { + storage_type value = const_cast< storage_type& >(storage); + while (!BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, v, &value)) {} + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT + { + storage_type value = storage_type(); + BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, value, &value); + return value; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT + { + return !!BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, desired, &expected); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 16u, Signed > : + public cas_based_operations< msvc_dcas_x86_64< Signed > > +{ +}; + +#endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B) + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + if (order == memory_order_seq_cst) + msvc_x86_operations_base::hardware_full_fence(); + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#if defined(BOOST_MSVC) +#pragma warning(pop) +#endif + +#endif // BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_ diff --git a/boost/atomic/detail/ops_windows.hpp b/boost/atomic/detail/ops_windows.hpp new file mode 100644 index 0000000000..1b4b04c8bc --- /dev/null +++ b/boost/atomic/detail/ops_windows.hpp @@ -0,0 +1,215 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/ops_windows.hpp + * + * This header contains implementation of the \c operations template. + * + * This implementation is the most basic version for Windows. It should + * work for any non-MSVC-like compilers as long as there are Interlocked WinAPI + * functions available. This version is also used for WinCE. + * + * Notably, this implementation is not as efficient as other + * versions based on compiler intrinsics. + */ + +#ifndef BOOST_ATOMIC_DETAIL_OPS_WINDOWS_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_OPS_WINDOWS_HPP_INCLUDED_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +struct windows_operations_base +{ + static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT + { + long tmp; + BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&tmp, 0); + } + + static BOOST_FORCEINLINE void fence_before(memory_order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } + + static BOOST_FORCEINLINE void fence_after(memory_order) BOOST_NOEXCEPT + { + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + } +}; + +template< typename T, typename Derived > +struct windows_operations : + public windows_operations_base +{ + typedef T storage_type; + + static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + Derived::exchange(storage, v, order); + } + + static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return Derived::fetch_add(const_cast< storage_type volatile& >(storage), (storage_type)0, order); + } + + static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + typedef typename make_signed< storage_type >::type signed_storage_type; + return Derived::fetch_add(storage, static_cast< storage_type >(-static_cast< signed_storage_type >(v)), order); + } + + static BOOST_FORCEINLINE bool compare_exchange_weak( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + return Derived::compare_exchange_strong(storage, expected, desired, success_order, failure_order); + } + + static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + return !!Derived::exchange(storage, (storage_type)1, order); + } + + static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT + { + store(storage, (storage_type)0, order); + } + + static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT + { + return true; + } +}; + +template< bool Signed > +struct operations< 4u, Signed > : + public windows_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > +{ + typedef windows_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > base_type; + typedef typename base_type::storage_type storage_type; + + static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(&storage, v)); + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { + base_type::fence_before(order); + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&storage, v)); + base_type::fence_after(order); + return v; + } + + static BOOST_FORCEINLINE bool compare_exchange_strong( + storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT + { + storage_type previous = expected; + base_type::fence_before(success_order); + storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(&storage, desired, previous)); + expected = old_val; + // The success and failure fences are the same anyway + base_type::fence_after(success_order); + return (previous == old_val); + } + + static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { +#if defined(BOOST_ATOMIC_INTERLOCKED_AND) + base_type::fence_before(order); + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND(&storage, v)); + base_type::fence_after(order); + return v; +#else + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res & v, order, memory_order_relaxed)) {} + return res; +#endif + } + + static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { +#if defined(BOOST_ATOMIC_INTERLOCKED_OR) + base_type::fence_before(order); + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR(&storage, v)); + base_type::fence_after(order); + return v; +#else + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res | v, order, memory_order_relaxed)) {} + return res; +#endif + } + + static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT + { +#if defined(BOOST_ATOMIC_INTERLOCKED_XOR) + base_type::fence_before(order); + v = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR(&storage, v)); + base_type::fence_after(order); + return v; +#else + storage_type res = storage; + while (!compare_exchange_strong(storage, res, res ^ v, order, memory_order_relaxed)) {} + return res; +#endif + } +}; + +template< bool Signed > +struct operations< 1u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 1u, Signed > +{ +}; + +template< bool Signed > +struct operations< 2u, Signed > : + public extending_cas_based_operations< operations< 4u, Signed >, 2u, Signed > +{ +}; + +BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT +{ + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); + if (order == memory_order_seq_cst) + windows_operations_base::hardware_full_fence(); + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT +{ + if (order != memory_order_relaxed) + BOOST_ATOMIC_DETAIL_COMPILER_BARRIER(); +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_OPS_WINDOWS_HPP_INCLUDED_ diff --git a/boost/atomic/detail/pause.hpp b/boost/atomic/detail/pause.hpp new file mode 100644 index 0000000000..37aa5ca84e --- /dev/null +++ b/boost/atomic/detail/pause.hpp @@ -0,0 +1,43 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * (C) Copyright 2013 Tim Blechmann + * (C) Copyright 2013 Andrey Semashev + */ + +#ifndef BOOST_ATOMIC_DETAIL_PAUSE_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_PAUSE_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if defined(_MSC_VER) && (defined(_M_AMD64) || defined(_M_IX86)) +extern "C" void _mm_pause(void); +#if defined(BOOST_MSVC) +#pragma intrinsic(_mm_pause) +#endif +#endif + +namespace boost { +namespace atomics { +namespace detail { + +BOOST_FORCEINLINE void pause() BOOST_NOEXCEPT +{ +#if defined(_MSC_VER) && (defined(_M_AMD64) || defined(_M_IX86)) + _mm_pause(); +#elif defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)) + __asm__ __volatile__("pause;"); +#endif +} + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_PAUSE_HPP_INCLUDED_ diff --git a/boost/atomic/detail/platform.hpp b/boost/atomic/detail/platform.hpp new file mode 100644 index 0000000000..76ad4ebb8f --- /dev/null +++ b/boost/atomic/detail/platform.hpp @@ -0,0 +1,115 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/platform.hpp + * + * This header defines macros for the target platform detection + */ + +#ifndef BOOST_ATOMIC_DETAIL_PLATFORM_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_PLATFORM_HPP_INCLUDED_ + +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +#if !defined(BOOST_ATOMIC_FORCE_FALLBACK) + +// Compiler-based backends +#if ((defined(__GNUC__) && ((__GNUC__ * 100 + __GNUC_MINOR__) >= 407)) ||\ + (defined(BOOST_CLANG) && ((__clang_major__ * 100 + __clang_minor__) >= 302))) &&\ + (\ + (__GCC_ATOMIC_BOOL_LOCK_FREE + 0) == 2 ||\ + (__GCC_ATOMIC_CHAR_LOCK_FREE + 0) == 2 ||\ + (__GCC_ATOMIC_SHORT_LOCK_FREE + 0) == 2 ||\ + (__GCC_ATOMIC_INT_LOCK_FREE + 0) == 2 ||\ + (__GCC_ATOMIC_LONG_LOCK_FREE + 0) == 2 ||\ + (__GCC_ATOMIC_LLONG_LOCK_FREE + 0) == 2\ + ) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_atomic + +#elif defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_x86 + +#elif defined(__GNUC__) && (defined(__POWERPC__) || defined(__PPC__)) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_ppc + +// This list of ARM architecture versions comes from Apple's arm/arch.h header. +// I don't know how complete it is. +#elif defined(__GNUC__) &&\ + (\ + defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) ||\ + defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) ||\ + defined(__ARM_ARCH_6ZK__) ||\ + defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) ||\ + defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) ||\ + defined(__ARM_ARCH_7EM__) || defined(__ARM_ARCH_7S__)\ + ) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_arm + +#elif defined(__GNUC__) && defined(__sparc_v9__) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_sparc + +#elif defined(__GNUC__) && defined(__alpha__) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_alpha + +#elif defined(__GNUC__) && ((__GNUC__ * 100 + __GNUC_MINOR__) >= 401) &&\ + (\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1) ||\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2) ||\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4) ||\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8) ||\ + defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16)\ + ) + +#define BOOST_ATOMIC_DETAIL_PLATFORM gcc_sync + +#elif defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) + +#define BOOST_ATOMIC_DETAIL_PLATFORM msvc_x86 + +#elif defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_M_ARM) + +#define BOOST_ATOMIC_DETAIL_PLATFORM msvc_arm + +#endif + +// OS-based backends +#if !defined(BOOST_ATOMIC_DETAIL_PLATFORM) + +#if defined(__linux__) && defined(__arm__) + +#define BOOST_ATOMIC_DETAIL_PLATFORM linux_arm + +#elif defined(BOOST_WINDOWS) || defined(_WIN32_CE) + +#define BOOST_ATOMIC_DETAIL_PLATFORM windows + +#endif + +#endif // !defined(BOOST_ATOMIC_DETAIL_PLATFORM) + +#endif // !defined(BOOST_ATOMIC_FORCE_FALLBACK) + +#if !defined(BOOST_ATOMIC_DETAIL_PLATFORM) +#define BOOST_ATOMIC_DETAIL_PLATFORM emulated +#define BOOST_ATOMIC_EMULATED +#endif + +#define BOOST_ATOMIC_DETAIL_HEADER(prefix) + +#endif // BOOST_ATOMIC_DETAIL_PLATFORM_HPP_INCLUDED_ diff --git a/boost/atomic/detail/storage_type.hpp b/boost/atomic/detail/storage_type.hpp new file mode 100644 index 0000000000..a024f1d327 --- /dev/null +++ b/boost/atomic/detail/storage_type.hpp @@ -0,0 +1,168 @@ +/* + * Distributed under the Boost Software License, Version 1.0. + * (See accompanying file LICENSE_1_0.txt or copy at + * http://www.boost.org/LICENSE_1_0.txt) + * + * Copyright (c) 2009 Helge Bahmann + * Copyright (c) 2012 Tim Blechmann + * Copyright (c) 2013 - 2014 Andrey Semashev + */ +/*! + * \file atomic/detail/storage_type.hpp + * + * This header defines underlying types used as storage + */ + +#ifndef BOOST_ATOMIC_DETAIL_STORAGE_TYPE_HPP_INCLUDED_ +#define BOOST_ATOMIC_DETAIL_STORAGE_TYPE_HPP_INCLUDED_ + +#include +#include +#include + +#ifdef BOOST_HAS_PRAGMA_ONCE +#pragma once +#endif + +namespace boost { +namespace atomics { +namespace detail { + +template< unsigned int Size > +struct buffer_storage +{ + unsigned char data[Size]; + + BOOST_FORCEINLINE bool operator! () const BOOST_NOEXCEPT + { + bool result = true; + for (unsigned int i = 0; i < Size && result; ++i) + { + result &= data[i] == 0; + } + return result; + } + + BOOST_FORCEINLINE bool operator== (buffer_storage const& that) const BOOST_NOEXCEPT + { + return std::memcmp(data, that.data, Size) == 0; + } + + BOOST_FORCEINLINE bool operator!= (buffer_storage const& that) const BOOST_NOEXCEPT + { + return std::memcmp(data, that.data, Size) != 0; + } +}; + +template< unsigned int Size, bool Signed > +struct make_storage_type +{ + typedef buffer_storage< Size > type; +}; + +template< > +struct make_storage_type< 1u, false > +{ + typedef boost::uint8_t type; +}; + +template< > +struct make_storage_type< 1u, true > +{ + typedef boost::int8_t type; +}; + +template< > +struct make_storage_type< 2u, false > +{ + typedef boost::uint16_t type; +}; + +template< > +struct make_storage_type< 2u, true > +{ + typedef boost::int16_t type; +}; + +template< > +struct make_storage_type< 4u, false > +{ + typedef boost::uint32_t type; +}; + +template< > +struct make_storage_type< 4u, true > +{ + typedef boost::int32_t type; +}; + +template< > +struct make_storage_type< 8u, false > +{ + typedef boost::uint64_t type; +}; + +template< > +struct make_storage_type< 8u, true > +{ + typedef boost::int64_t type; +}; + +#if defined(BOOST_HAS_INT128) + +template< > +struct make_storage_type< 16u, false > +{ + typedef boost::uint128_type type; +}; + +template< > +struct make_storage_type< 16u, true > +{ + typedef boost::int128_type type; +}; + +#elif !defined(BOOST_NO_ALIGNMENT) + +struct BOOST_ALIGNMENT(16) storage128_t +{ + boost::uint64_t data[2]; + + BOOST_FORCEINLINE bool operator! () const BOOST_NOEXCEPT + { + return data[0] == 0 && data[1] == 0; + } +}; + +BOOST_FORCEINLINE bool operator== (storage128_t const& left, storage128_t const& right) BOOST_NOEXCEPT +{ + return left.data[0] == right.data[0] && left.data[1] == right.data[1]; +} +BOOST_FORCEINLINE bool operator!= (storage128_t const& left, storage128_t const& right) BOOST_NOEXCEPT +{ + return !(left == right); +} + +template< bool Signed > +struct make_storage_type< 16u, Signed > +{ + typedef storage128_t type; +}; + +#endif + +template< typename T > +struct storage_size_of +{ + enum _ + { + size = sizeof(T), + value = (size == 3 ? 4 : (size >= 5 && size <= 7 ? 8 : (size >= 9 && size <= 15 ? 16 : size))) + }; +}; + +} // namespace detail +} // namespace atomics +} // namespace boost + +#endif // BOOST_ATOMIC_DETAIL_STORAGE_TYPE_HPP_INCLUDED_ -- cgit v1.2.3