Chris@102: /*
Chris@102:  * Distributed under the Boost Software License, Version 1.0.
Chris@102:  * (See accompanying file LICENSE_1_0.txt or copy at
Chris@102:  * http://www.boost.org/LICENSE_1_0.txt)
Chris@102:  *
Chris@102:  * Copyright (c) 2009 Helge Bahmann
Chris@102:  * Copyright (c) 2012 Tim Blechmann
Chris@102:  * Copyright (c) 2014 Andrey Semashev
Chris@102:  */
Chris@102: /*!
Chris@102:  * \file   atomic/detail/ops_msvc_x86.hpp
Chris@102:  *
Chris@102:  * This header contains implementation of the \c operations template.
Chris@102:  */
Chris@102: 
Chris@102: #ifndef BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_
Chris@102: #define BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_
Chris@102: 
Chris@102: #include <boost/memory_order.hpp>
Chris@102: #include <boost/type_traits/make_signed.hpp>
Chris@102: #include <boost/atomic/detail/config.hpp>
Chris@102: #include <boost/atomic/detail/interlocked.hpp>
Chris@102: #include <boost/atomic/detail/storage_type.hpp>
Chris@102: #include <boost/atomic/detail/operations_fwd.hpp>
Chris@102: #include <boost/atomic/capabilities.hpp>
Chris@102: #if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B) || defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)
Chris@102: #include <boost/cstdint.hpp>
Chris@102: #include <boost/atomic/detail/ops_cas_based.hpp>
Chris@102: #endif
Chris@102: #include <boost/atomic/detail/ops_msvc_common.hpp>
Chris@102: #if !defined(_M_IX86) && !(defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8) && defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16))
Chris@102: #include <boost/atomic/detail/ops_extending_cas_based.hpp>
Chris@102: #endif
Chris@102: 
Chris@102: #ifdef BOOST_HAS_PRAGMA_ONCE
Chris@102: #pragma once
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(BOOST_MSVC)
Chris@102: #pragma warning(push)
Chris@102: // frame pointer register 'ebx' modified by inline assembly code. See the note below.
Chris@102: #pragma warning(disable: 4731)
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(_MSC_VER) && (defined(_M_AMD64) || (defined(_M_IX86) && defined(_M_IX86_FP) && _M_IX86_FP >= 2))
Chris@102: extern "C" void _mm_mfence(void);
Chris@102: #if defined(BOOST_MSVC)
Chris@102: #pragma intrinsic(_mm_mfence)
Chris@102: #endif
Chris@102: #endif
Chris@102: 
Chris@102: namespace boost {
Chris@102: namespace atomics {
Chris@102: namespace detail {
Chris@102: 
Chris@102: /*
Chris@102:  * Implementation note for asm blocks.
Chris@102:  *
Chris@102:  * http://msdn.microsoft.com/en-us/data/k1a8ss06%28v=vs.105%29
Chris@102:  *
Chris@102:  * Some SSE types require eight-byte stack alignment, forcing the compiler to emit dynamic stack-alignment code.
Chris@102:  * To be able to access both the local variables and the function parameters after the alignment, the compiler
Chris@102:  * maintains two frame pointers. If the compiler performs frame pointer omission (FPO), it will use EBP and ESP.
Chris@102:  * If the compiler does not perform FPO, it will use EBX and EBP. To ensure code runs correctly, do not modify EBX
Chris@102:  * in asm code if the function requires dynamic stack alignment as it could modify the frame pointer.
Chris@102:  * Either move the eight-byte aligned types out of the function, or avoid using EBX.
Chris@102:  *
Chris@102:  * Since we have no way of knowing that the compiler uses FPO, we have to always save and restore ebx
Chris@102:  * whenever we have to clobber it. Additionally, we disable warning C4731 above so that the compiler
Chris@102:  * doesn't spam about ebx use.
Chris@102:  */
Chris@102: 
Chris@102: struct msvc_x86_operations_base
Chris@102: {
Chris@102:     static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT
Chris@102:     {
Chris@102: #if defined(_MSC_VER) && (defined(_M_AMD64) || (defined(_M_IX86) && defined(_M_IX86_FP) && _M_IX86_FP >= 2))
Chris@102:         // Use mfence only if SSE2 is available
Chris@102:         _mm_mfence();
Chris@102: #else
Chris@102:         long tmp;
Chris@102:         BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&tmp, 0);
Chris@102: #endif
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void fence_before(memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void fence_after(memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void fence_after_load(memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102: 
Chris@102:         // On x86 and x86_64 there is no need for a hardware barrier,
Chris@102:         // even if seq_cst memory order is requested, because all
Chris@102:         // seq_cst writes are implemented with lock-prefixed operations
Chris@102:         // or xchg which has implied lock prefix. Therefore normal loads
Chris@102:         // are already ordered with seq_cst stores on these architectures.
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< typename T, typename Derived >
Chris@102: struct msvc_x86_operations :
Chris@102:     public msvc_x86_operations_base
Chris@102: {
Chris@102:     typedef T storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         if (order != memory_order_seq_cst)
Chris@102:         {
Chris@102:             fence_before(order);
Chris@102:             storage = v;
Chris@102:             fence_after(order);
Chris@102:         }
Chris@102:         else
Chris@102:         {
Chris@102:             Derived::exchange(storage, v, order);
Chris@102:         }
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type v = storage;
Chris@102:         fence_after_load(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         typedef typename make_signed< storage_type >::type signed_storage_type;
Chris@102:         return Derived::fetch_add(storage, static_cast< storage_type >(-static_cast< signed_storage_type >(v)), order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_weak(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return Derived::compare_exchange_strong(storage, expected, desired, success_order, failure_order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return !!Derived::exchange(storage, (storage_type)1, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         store(storage, (storage_type)0, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return true;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 4u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 4u, Signed >::type, operations< 4u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type previous = expected;
Chris@102:         storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE(&storage, desired, previous));
Chris@102:         expected = old_val;
Chris@102:         return (previous == old_val);
Chris@102:     }
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_AND)
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND(&storage, v));
Chris@102:     }
Chris@102: #else
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type res = storage;
Chris@102:         while (!compare_exchange_strong(storage, res, res & v, order, memory_order_relaxed)) {}
Chris@102:         return res;
Chris@102:     }
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_OR)
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR(&storage, v));
Chris@102:     }
Chris@102: #else
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type res = storage;
Chris@102:         while (!compare_exchange_strong(storage, res, res | v, order, memory_order_relaxed)) {}
Chris@102:         return res;
Chris@102:     }
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_XOR)
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR(&storage, v));
Chris@102:     }
Chris@102: #else
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type res = storage;
Chris@102:         while (!compare_exchange_strong(storage, res, res ^ v, order, memory_order_relaxed)) {}
Chris@102:         return res;
Chris@102:     }
Chris@102: #endif
Chris@102: };
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 1u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD8(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE8(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type previous = expected;
Chris@102:         storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE8(&storage, desired, previous));
Chris@102:         expected = old_val;
Chris@102:         return (previous == old_val);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND8(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR8(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR8(&storage, v));
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #elif defined(_M_IX86)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 1u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 1u, Signed >::type, operations< 1u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov edx, storage
Chris@102:             movzx eax, v
Chris@102:             lock xadd byte ptr [edx], al
Chris@102:             mov v, al
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov edx, storage
Chris@102:             movzx eax, v
Chris@102:             xchg byte ptr [edx], al
Chris@102:             mov v, al
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(success_order);
Chris@102:         bool success;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov esi, expected
Chris@102:             mov edi, storage
Chris@102:             movzx eax, byte ptr [esi]
Chris@102:             movzx edx, desired
Chris@102:             lock cmpxchg byte ptr [edi], dl
Chris@102:             mov byte ptr [esi], al
Chris@102:             sete success
Chris@102:         };
Chris@102:         // The success and failure fences are equivalent anyway
Chris@102:         base_type::fence_after(success_order);
Chris@102:         return success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, byte ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dl, al
Chris@102:             and dl, bl
Chris@102:             lock cmpxchg byte ptr [edi], dl
Chris@102:             jne again
Chris@102:             mov v, al
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, byte ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dl, al
Chris@102:             or dl, bl
Chris@102:             lock cmpxchg byte ptr [edi], dl
Chris@102:             jne again
Chris@102:             mov v, al
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, byte ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dl, al
Chris@102:             xor dl, bl
Chris@102:             lock cmpxchg byte ptr [edi], dl
Chris@102:             jne again
Chris@102:             mov v, al
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #else
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 1u, Signed > :
Chris@102:     public extending_cas_based_operations< operations< 4u, Signed >, 1u, Signed >
Chris@102: {
Chris@102: };
Chris@102: 
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 2u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD16(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE16(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type previous = expected;
Chris@102:         storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE16(&storage, desired, previous));
Chris@102:         expected = old_val;
Chris@102:         return (previous == old_val);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND16(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR16(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR16(&storage, v));
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #elif defined(_M_IX86)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 2u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 2u, Signed >::type, operations< 2u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov edx, storage
Chris@102:             movzx eax, v
Chris@102:             lock xadd word ptr [edx], ax
Chris@102:             mov v, ax
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov edx, storage
Chris@102:             movzx eax, v
Chris@102:             xchg word ptr [edx], ax
Chris@102:             mov v, ax
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(success_order);
Chris@102:         bool success;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov esi, expected
Chris@102:             mov edi, storage
Chris@102:             movzx eax, word ptr [esi]
Chris@102:             movzx edx, desired
Chris@102:             lock cmpxchg word ptr [edi], dx
Chris@102:             mov word ptr [esi], ax
Chris@102:             sete success
Chris@102:         };
Chris@102:         // The success and failure fences are equivalent anyway
Chris@102:         base_type::fence_after(success_order);
Chris@102:         return success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, word ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dx, ax
Chris@102:             and dx, bx
Chris@102:             lock cmpxchg word ptr [edi], dx
Chris@102:             jne again
Chris@102:             mov v, ax
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, word ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dx, ax
Chris@102:             or dx, bx
Chris@102:             lock cmpxchg word ptr [edi], dx
Chris@102:             jne again
Chris@102:             mov v, ax
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         base_type::fence_before(order);
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             xor edx, edx
Chris@102:             mov edi, storage
Chris@102:             movzx ebx, v
Chris@102:             movzx eax, word ptr [edi]
Chris@102:             align 16
Chris@102:         again:
Chris@102:             mov dx, ax
Chris@102:             xor dx, bx
Chris@102:             lock cmpxchg word ptr [edi], dx
Chris@102:             jne again
Chris@102:             mov v, ax
Chris@102:             mov ebx, backup
Chris@102:         };
Chris@102:         base_type::fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #else
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 2u, Signed > :
Chris@102:     public extending_cas_based_operations< operations< 4u, Signed >, 2u, Signed >
Chris@102: {
Chris@102: };
Chris@102: 
Chris@102: #endif
Chris@102: 
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct msvc_dcas_x86
Chris@102: {
Chris@102:     typedef typename make_storage_type< 8u, Signed >::type storage_type;
Chris@102: 
Chris@102:     // Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 3A, 8.1.1. Guaranteed Atomic Operations:
Chris@102:     //
Chris@102:     // The Pentium processor (and newer processors since) guarantees that the following additional memory operations will always be carried out atomically:
Chris@102:     // * Reading or writing a quadword aligned on a 64-bit boundary
Chris@102:     //
Chris@102:     // 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
Chris@102:     // 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).
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type volatile* p = &storage;
Chris@102:         if (((uint32_t)p & 0x00000007) == 0)
Chris@102:         {
Chris@102: #if defined(_M_IX86_FP) && _M_IX86_FP >= 2
Chris@102: #if defined(__AVX__)
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 vmovq xmm4, v
Chris@102:                 vmovq qword ptr [edx], xmm4
Chris@102:             };
Chris@102: #else
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 movq xmm4, v
Chris@102:                 movq qword ptr [edx], xmm4
Chris@102:             };
Chris@102: #endif
Chris@102: #else
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 fild v
Chris@102:                 fistp qword ptr [edx]
Chris@102:             };
Chris@102: #endif
Chris@102:         }
Chris@102:         else
Chris@102:         {
Chris@102:             int backup;
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov backup, ebx
Chris@102:                 mov edi, p
Chris@102:                 mov ebx, dword ptr [v]
Chris@102:                 mov ecx, dword ptr [v + 4]
Chris@102:                 mov eax, dword ptr [edi]
Chris@102:                 mov edx, dword ptr [edi + 4]
Chris@102:                 align 16
Chris@102:             again:
Chris@102:                 lock cmpxchg8b qword ptr [edi]
Chris@102:                 jne again
Chris@102:                 mov ebx, backup
Chris@102:             };
Chris@102:         }
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type const volatile* p = &storage;
Chris@102:         storage_type value;
Chris@102: 
Chris@102:         if (((uint32_t)p & 0x00000007) == 0)
Chris@102:         {
Chris@102: #if defined(_M_IX86_FP) && _M_IX86_FP >= 2
Chris@102: #if defined(__AVX__)
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 vmovq xmm4, qword ptr [edx]
Chris@102:                 vmovq value, xmm4
Chris@102:             };
Chris@102: #else
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 movq xmm4, qword ptr [edx]
Chris@102:                 movq value, xmm4
Chris@102:             };
Chris@102: #endif
Chris@102: #else
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edx, p
Chris@102:                 fild qword ptr [edx]
Chris@102:                 fistp value
Chris@102:             };
Chris@102: #endif
Chris@102:         }
Chris@102:         else
Chris@102:         {
Chris@102:             // We don't care for comparison result here; the previous value will be stored into value anyway.
Chris@102:             // Also we don't care for ebx and ecx values, they just have to be equal to eax and edx before cmpxchg8b.
Chris@102:             __asm
Chris@102:             {
Chris@102:                 mov edi, p
Chris@102:                 mov eax, ebx
Chris@102:                 mov edx, ecx
Chris@102:                 lock cmpxchg8b qword ptr [edi]
Chris@102:                 mov dword ptr [value], eax
Chris@102:                 mov dword ptr [value + 4], edx
Chris@102:             };
Chris@102:         }
Chris@102: 
Chris@102:         return value;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type volatile* p = &storage;
Chris@102: #if defined(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64)
Chris@102:         const storage_type old_val = (storage_type)BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(p, desired, expected);
Chris@102:         const bool result = (old_val == expected);
Chris@102:         expected = old_val;
Chris@102:         return result;
Chris@102: #else
Chris@102:         bool result;
Chris@102:         int backup;
Chris@102:         __asm
Chris@102:         {
Chris@102:             mov backup, ebx
Chris@102:             mov edi, p
Chris@102:             mov esi, expected
Chris@102:             mov ebx, dword ptr [desired]
Chris@102:             mov ecx, dword ptr [desired + 4]
Chris@102:             mov eax, dword ptr [esi]
Chris@102:             mov edx, dword ptr [esi + 4]
Chris@102:             lock cmpxchg8b qword ptr [edi]
Chris@102:             mov dword ptr [esi], eax
Chris@102:             mov dword ptr [esi + 4], edx
Chris@102:             mov ebx, backup
Chris@102:             sete result
Chris@102:         };
Chris@102:         return result;
Chris@102: #endif
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_weak(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return compare_exchange_strong(storage, expected, desired, success_order, failure_order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return true;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 8u, Signed > :
Chris@102:     public cas_based_operations< msvc_dcas_x86< Signed > >
Chris@102: {
Chris@102: };
Chris@102: 
Chris@102: #elif defined(_M_AMD64)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 8u, Signed > :
Chris@102:     public msvc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > >
Chris@102: {
Chris@102:     typedef msvc_x86_operations< typename make_storage_type< 8u, Signed >::type, operations< 8u, Signed > > base_type;
Chris@102:     typedef typename base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE_ADD64(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_EXCHANGE64(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type previous = expected;
Chris@102:         storage_type old_val = static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE64(&storage, desired, previous));
Chris@102:         expected = old_val;
Chris@102:         return (previous == old_val);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_AND64(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_OR64(&storage, v));
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return static_cast< storage_type >(BOOST_ATOMIC_INTERLOCKED_XOR64(&storage, v));
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #endif
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct msvc_dcas_x86_64
Chris@102: {
Chris@102:     typedef typename make_storage_type< 16u, Signed >::type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type value = const_cast< storage_type& >(storage);
Chris@102:         while (!BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, v, &value)) {}
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type value = storage_type();
Chris@102:         BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, value, &value);
Chris@102:         return value;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return !!BOOST_ATOMIC_INTERLOCKED_COMPARE_EXCHANGE128(&storage, desired, &expected);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_weak(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return compare_exchange_strong(storage, expected, desired, success_order, failure_order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return true;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 16u, Signed > :
Chris@102:     public cas_based_operations< msvc_dcas_x86_64< Signed > >
Chris@102: {
Chris@102: };
Chris@102: 
Chris@102: #endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)
Chris@102: 
Chris@102: BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102:     if (order == memory_order_seq_cst)
Chris@102:         msvc_x86_operations_base::hardware_full_fence();
Chris@102:     BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102: }
Chris@102: 
Chris@102: BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     if (order != memory_order_relaxed)
Chris@102:         BOOST_ATOMIC_DETAIL_COMPILER_BARRIER();
Chris@102: }
Chris@102: 
Chris@102: } // namespace detail
Chris@102: } // namespace atomics
Chris@102: } // namespace boost
Chris@102: 
Chris@102: #if defined(BOOST_MSVC)
Chris@102: #pragma warning(pop)
Chris@102: #endif
Chris@102: 
Chris@102: #endif // BOOST_ATOMIC_DETAIL_OPS_MSVC_X86_HPP_INCLUDED_