Revision History

Added an io_uring backend. This backend may be used for all I/O objects, including sockets, timers, and posix descriptors.

The backend is disabled by default, and must be enabled by defining both ASIO_HAS_IO_URING and ASIO_DISABLE_EPOLL.

Simply defining ASIO_HAS_IO_URING alone will enable the backend without using it for the existing I/O objects. This allows it to be used for I/O objects that require io_uring support, such as files.

This support depends on the liburing library at both compile and link time. Add -luring to your list of libraries for linking.

Added support for files. This introduces new classes for stream-oriented and random-access files. For example, to write to a newly created stream-oriented file:

asio::stream_file file(
    my_io_context, "/path/to/file",
    asio::stream_file::write_only
      | asio::stream_file::create
      | asio::stream_file::truncate);

file.async_write_some(my_buffer,
    [](error_code e, size_t n)
    {
      // ...
    });

or to read from a random-access file:

asio::random_access_file file(
    my_io_context, "/path/to/file",
    asio::random_access_file::read_only);

file.async_read_some_at(1234, my_buffer,
    [](error_code e, size_t n)
    {
      // ...
    });

This feature currently supports I/O completion ports on Windows, and io_uring on Linux (define ASIO_HAS_IO_URING to enable).

Added support for portable pipes. This change add supports for pipes on POSIX and Windows (when I/O completion ports are available). For example, to create and use a connected pair of pipe objects:

asio::readable_pipe read_end;
asio::writable_pipe write_end;
asio::connect_pipe(read_end, write_end);

write_end.async_write_some(my_write_buffer,
    [](error_code e, size_t n)
    {
      // ...
    });

read_end.async_read_some(my_read_buffer,
    [](error_code e, size_t n)
    {
      // ...
    });

Added support for registered buffers. The mutable_registered_buffer and const_registered_buffer classes are buffer sequence types that represent registered buffers. These buffers are obtained by first performing a buffer registration:
```
auto my_registration =
  asio::register_buffers(
      my_execution_context,
      my_buffer_sequence);
```
The registration object must be maintained for as long as the buffer registration is required. The supplied buffer sequence represents the memory location or locations that will be registered, and the caller must ensure they remain valid for as long as they are registered. The registration is automatically removed when the registration object is destroyed. There can be at most one active registration per execution context.

The registration object is a container of registered buffers. Buffers may be obtained from it by iterating over the container, or via direct index access:
```
asio::mutable_registered_buffer my_buffer
  = my_registration[i];
```
The registered buffers may then be passed directly to operations:
```
asio::async_read(my_socket, my_buffer,
    [](error_code ec, size_t n)
    {
      // ...
    });
```
Buffer registration supports the io_uring backend when used with read and write operations on descriptors, files, pipes, and sockets. For portability, the facility may be used on other platforms, but the registered buffers will behave as normal buffers.

Added experimental support for channels. This adds templates experimental::basic_channel and experimental::basic_concurrent_channel, with aliases experimental::channel and experimental::concurrent_channel. Channels may be used to send completions as messages. For example:

// Create a channel with no buffer space.
channel<void(error_code, size_t)> ch(ctx);

// The call to try_send fails as there is no buffer
// space and no waiting receive operations.
bool ok = ch.try_send(asio::error::eof, 123);
assert(!ok);

// The async_send operation is outstanding until
// a receive operation consumes the message.
ch.async_send(asio::error::eof, 123,
    [](error_code ec)
    {
      // ...
    });

// The async_receive consumes the message. Both the
// async_send and async_receive operations complete
// immediately.
ch.async_receive(
    [](error_code ec, size_t n)
    {
      // ...
    });

Implemented improvements to experimental::coro.
- Added overload so member functions can provide an explicit executor.
- Added co_spawn for coro tasks.
- Added reference and overview documentation.
- Adopted awaitable cancellation model.
- Refactored implementation.
Disabled aligned_alloc on clang when using an MSVC runtime.
Changed to use a faster implementation for ip::network_v4::canonical().
Added template specialisations for common uses to improve compile time.
Reduced the size of io_context executors to a single pointer.
Increased the small object buffer size for execution::any_executor and any_io_executor.
Fixed multi-signature handling when variadic templates are disabled.
Fixed compatibility with new versions of gcc and clang.
Fixed compilation on Solaris.
Fix defence against Qt-defined macros when building with MSVC.
Fixed various warnings.

Fixed experimental::coro compatibility with gcc.
Fixed experimental::promise compatibility with gcc.
Added documentation for per-operation cancellation.
Added documentation for parallel_group.
Added overview documentation for experimental::coro.
Added some missing C++14 examples.
Updated C++20 coroutines overview documentation.

Cleaned up memory recycling to ensure the cache size is used consistently.
Changed parallel_group to use the recycling allocator for shared state.
Changed awaitable memory recycling to use the configurable cache size.
Ensured parallel_group respects operations' associated executors.
Added defence against macros defined by Qt.
Added missing boost_ prefix to namespaces during boostification.
Added missing lvalue-qualified overloads for deferred operator().
Added a tag to disambiguate deferred constructors.
Added missing push/pop_options includes.
Fixed argument evaluation order issue with a potentially moved-from variable in awaitable implementation.
Enabled "expression SFINAE" for recent MSVC using /std:c++latest.
Fixed compilation errors when dev_poll_reactor backend is used.
Fixed handler type requirements checking to reflect rvalue completion handler invocation.
Fixed posix::basic_stream_descriptor move operations to work with non-defaulted executors.
Added missing move of executor in strand implementation.

Fixed handling of move-only results with awaitable operators && and ||.
Fixed strand<> to avoid using a potentially moved-from executor.
Fixed order of _aligned_malloc arguments for MSVC.
Ensured handler alignment requirements are respected by cancellation_signal.
Ensured gcc tests are not used for clang when detecting compiler features.
Disabled coroutines support for the clang shipped with MSVC.
Fixed compatibility with recent LibreSSL when OPENSSL_NO_SSL_INTERN is defined.
Documented supported cancellation types for individual asynchronous operations.
Ensured un-cancelled ops are correctly placed back in the queue.

Added support for cancellation of individual asynchronous operations. Cancellation is implemented by having each completion handler carry an associated CancellationSlot, a lightweight cancellation channel that is specified through the new associated_cancellation_slot associator. A concrete CancellationSlot implementation is provided in the form of the cancellation_signal and cancellation_slot classes. In conjunction with the bind_cancellation_slot helper function, these may be used to hook cancellation into asynchronous operations. However, it should be noted that these classes are the low-level building blocks of cancellation, and most use cases should use a higher level abstraction for cancellation, such as experimental::parallel_group or the new logical operators for awaitable (see below). The ability to cancel individual operations, or composed operations, is currently supported by:
- timers
- sockets on POSIX and Windows
- POSIX descriptors
- Windows HANDLEs
- SSL streams
- all Asio-provided composed operations such as async_read and async_write
- compositions based on async_compose
- C++20 coroutines that use awaitable
- C++20 coroutines that use the new experimental::coro (see below)
- the new experimental::parallel_group operation (see below)
- the new experimental::promise class (see below)

Added the associator trait. The associator trait is used to generically forward associators, such as associated_executor and associated_allocator, through intermediate completion handlers. For example:

template <typename Handler>
struct intermediate_handler
{
  Handler handler_;

  template <typename... Args>
  void operator()(Args&... args)
  {
    // ...
  }
};

namespace asio {
  template <
      template <typename, typename> class Associator,
      typename Handler,
      typename DefaultCandidate>
  struct associator<
      Associator,
      intermediate_handler<Handler>,
      DefaultCandidate>
  {
    using type =
      typename Associator<Handler, DefaultCandidate>::type;

    static type get(
        const intermediate_handler<Handler>& h,
        const DefaultCandidate& c = DefaultCandidate()) noexcept
    {
      return Associator<Handler, DefaultCandidate>::get(
          h.handler_, c);
    }
  };
} // namespace asio

Changed all asynchronous operations implemented in Asio to invoke their completion handlers as rvalue references. This allows an rvalue reference qualifier to be added to the function call operator.
Added support for asynchronous operations with multiple completion signatures. Completion signatures may also specify that they are noexcept, and whether they are lvalue-invocable (and thus do not "consume" the completion handler) or rvalue-invocable (and thus do "consume" the handler, indicating an end to the asynchronous operation). For example:
```
auto my_async_operation(...,
    asio::completion_token_for<
      void(intermediate_result_type) & noexcept,
      void(final_result_type) &&
    > auto&& token)
{
  // ...
}
```
Added operator&& and operator|| for awaitable<>. The logical operators || and && have been overloaded for awaitable<>, to allow coroutines to be trivially awaited in parallel.

When awaited using &&, the co_await expression waits until both operations have completed successfully. As a "short-circuit" evaluation, if one operation fails with an exception, the other is immediately cancelled. For example:
```
std::tuple<std::size_t, std::size_t> results =
  co_await (
    async_read(socket, input_buffer, use_awaitable)
      && async_write(socket, output_buffer, use_awaitable)
  );
```
When awaited using ||, the co_await expression waits until either operation succeeds. As a "short-circuit" evaluation, if one operation succeeds without throwing an exception, the other is immediately cancelled. For example:
```
std::variant<std::size_t, std::monostate> results =
  co_await (
    async_read(socket, input_buffer, use_awaitable)
      || timer.async_wait(use_awaitable)
  );
```
The operators may be enabled by adding the #include:
```
#include <asio/experimental/awaitable_operators.hpp>
```
and then bringing the contents of the experimental::awaitable_operators namespace into scope:
```
using namespace asio::experimental::awaitable_operators;
```
Added the experimental::as_tuple completion token adapter. The as_tuple completion token adapter can be used to specify that the completion handler arguments should be combined into a single tuple argument. The as_tuple adapter may be used in conjunction with use_awaitable and structured bindings as follows:
```
auto [e, n] = co_await socket.async_read_some(
    asio::buffer(data), as_tuple(use_awaitable));
```
Alternatively, it may be used as a default completion token like so:
```
using default_token = as_tuple_t<use_awaitable_t<>>;
using tcp_socket = default_token::as_default_on_t<tcp::socket>;
// ...
awaitable<void> do_read(tcp_socket socket)
{
  // ...
  auto [e, n] = co_await socket.async_read_some(asio::buffer(data));
  // ...
}
```

Added the experimental::append completion token adapter. The append completion token adapter can be used to pass additional completion handler arguments at the end of the completion signature. For example:

timer.async_wait(
    asio::experimental::append(
      [](std::error_code ec, int i)
      {
        // ...
      },
    42)
  );

std::future<int> f = timer.async_wait(
    asio::experimental::append(
      asio::use_future,
      42
    )
  );

Added the experimental::prepend completion token adapter. The prepend completion token adapter can be used to pass additional arguments before the existing completion handler arguments. For example:

timer.async_wait(
    asio::experimental::prepend(
      [](int i, std::error_code ec)
      {
        // ...
      },
    42)
  );

std::future<std::tuple<int, std::error_code>> f = timer.async_wait(
    asio::experimental::prepend(
      asio::use_future,
      42
    )
  );

Added the experimental::deferred completion token. The deferred completion token takes a call to an asynchronous operation's initiating function and turns it into a function object that accepts a completion token. For example:

auto deferred_op =
  timer.async_wait(
    asio::experimental::deferred);
...
std::move(deferred_op)(
    [](std::error_code ec){ ... });

or:

auto deferred_op =
  timer.async_wait(
    asio::experimental::deferred);
...
std::future<void> =
  std::move(deferred_op)(
    asio::use_future);

The deferred token also supports chaining, to create simple compositions:

auto deferred_op =
  timer.async_wait(
    asio::experimental::deferred(
      [&](std::error_code ec)
      {
        timer.expires_after(
            std::chrono::seconds(1));

        return timer.async_wait(
            asio::experimental::deferred);
      });
...
std::future<void> = std::move(deferred_op)(asio::use_future);

Added the experimental::parallel_group class and experimental::make_parallel_group function. This utility may be used to launch work that is performed in parallel, and wait for one or all of the operations to complete. A parallel_group implements automatic cancellation of incomplete operations. For example:
```
experimental::make_parallel_group(
    [&](auto token)
    {
      return stream.async_read_some(asio::buffer(data), token);
    },
    [&](auto token)
    {
      return timer.async_wait(token);
    }
  ).async_wait(
    experimental::wait_for_one(),
    [](
        std::array<std::size_t, 2> completion_order,
        std::error_code ec1, std::size_t n1,
        std::error_code ec2
    )
    {
      // ...
    }
  );
```
The conditions for completion of the group may be specified using one of the four provided function objects wait_for_all, wait_for_one, wait_for_one_success, and wait_for_one_error, or with a custom function. The parallel_group class can also be combined with deferred as follows:
```
experimental::make_parallel_group(
    stream.async_read_some(asio::buffer(data), experimental::deferred),
    timer.async_wait(experimental::deferred)
  ).async_wait(
    // ...
  );
```
Note: for maximum flexibility, parallel_group does not propagate the executor automatically to the operations within the group.

Added experimental::promise. The promise type allows eager execution and synchronisation of async operations. For example:

auto promise = async_read(
    stream, asio::buffer(my_buffer),
    asio::experimental::use_promise);

... do other stuff while the read is going on ...

promise.async_wait( // completion the operation
    [](error_code ec, std::size_t bytes_read)
    {
      ...
    });

Promises can be safely disregarded if the result is no longer required. Different operations can be combined to either wait for all to complete or for one to complete (and cancel the rest). For example, to wait for one to complete:

auto timeout_promise =
  timer.async_wait(
    asio::experimental::use_promise);

auto read_promise = async_read(
    stream, asio::buffer(my_buffer),
    asio::experimental::use_promise);

auto promise =
  asio::experimental::promise<>::race(
    timeout_promise, read_promise);

promise.async_wait(
    [](std::variant<error_code, std::tuple<error_code, std::size_t>> v)
    {
      if (v.index() == 0) {} //timed out
      else if (v.index() == 1) // completed in time
    });

or to wait for all to complete:

auto write_promise = async_write(
    stream, asio::buffer(my_write_buffer),
    asio::experimental::use_promise);

auto read_promise = async_read(
    stream, asio::buffer(my_buffer),
    asio::experimental::use_promise);

auto promise =
  asio::experimental::promise<>::all(
    write_promise, read_promise);

promise.async_wait(
    [](std::tuple<error_code, std::size_t> write_result,
      std::tuple<error_code, std::size_t> read_result)
    {
    });

Thanks go to Klemens Morgenstern for contributing this feature.

Added the experimental::coro class template. The coro type is a C++20 coroutine primitive for resumable functions, with the ability to combine both asynchronous waiting (co_await) and yielding (co_yield) into a single, stateful control flow. For example:

#include <asio.hpp>
#include <asio/experimental/coro.hpp>

using asio::ip::tcp;

asio::experimental::coro<std::string> reader(tcp::socket& sock)
{
  std::string buf;
  while (sock.is_open())
  {
    std::size_t n = co_await asio::async_read_until(
        sock, asio::dynamic_buffer(buf), '\n',
        asio::experimental::use_coro);
    co_yield buf.substr(0, n);
    buf.erase(0, n);
  }
}

asio::awaitable<void> consumer(tcp::socket sock)
{
  auto r = reader(sock);
  auto msg1 = co_await r.async_resume(asio::use_awaitable);
  std::cout << "Message 1: " << msg1.value_or("\n");
  auto msg2 = co_await r.async_resume(asio::use_awaitable);
  std::cout << "Message 2: " << msg2.value_or("\n");
}

asio::awaitable<void> listen(tcp::acceptor& acceptor)
{
  for (;;)
  {
    co_spawn(
        acceptor.get_executor(),
        consumer(co_await acceptor.async_accept(asio::use_awaitable)),
        asio::detached);
  }
}

int main()
{
  asio::io_context ctx;
  tcp::acceptor acceptor(ctx, {tcp::v4(), 54321});
  co_spawn(ctx, listen(acceptor), asio::detached);
  ctx.run();
}

Thanks go to Klemens Morgenstern for contributing this feature.

Added move assignment to ssl::stream<>.
Changed co_spawn to dispatch the coroutine's initial step to the executor, and to only post the completion handler if the coroutine did not otherwise perform a context switch (i.e. a co_await on an asynchronous operation).
Enabled additional optimisations for any_executor and any_io_executor when used with asynchronous operations.
Added the nodiscard attribute to awaitable<>.
Increased the number of cached slots in the default recycling allocator. This number defaults to 2, but may also be specified by defining the ASIO_RECYCLING_ALLOCATOR_CACHE_SIZE macro.
Disabled the std::system_error message workaround for recent MSVC.
Changed the default allocator behaviour to respect alignment requirements, to support over-aligned types. Requires C++17 or later, or Boost.
Ensured the result strings are always initialised in reverse name resolution.
Fixed recursive template instantiation issues in use_awaitable_t::executor_with_default.
Fixed the any_io_executor equality operators to correctly return a result based on the target executor.

Added ip::scope_id_type type alias.
Added ip::port_type type alias.
Added std::hash specialisations for IP addresses.
Added std::hash specialisations for ip::basic_endpoint<>.
Refactored SFINAE usage to improve compile times.
Added friendship support to customisation points, and made most customisations private.
Changed any_io_executor to a "strong typedef"-style class.
Fixed experimental::as_single to work with handler hook deprecation.
Ensured pthread condition variable attributes are cleaned up on all platforms.
Clarified thread safety notes on sockets and descriptors.
Ensured errno is not overwritten if socket() fails on macOS/FreeBSD.
Fixed work tracking for io_context and thread_pool executors when move-assigned.
Ensured internal call_stack objects are accessed only from implementation files.
Fixed I/O object move-assignment to ensure the executor is left in a valid state.
Fixed detection of compiler support for defaulted template argument on functions with MSVC.
Prevented the blocking.always property from being used with strand<>, as it did not produce the correct semantics.
Removed deprecated file asio/impl/src.cpp.

Enabled support for UNIX domain sockets on Windows. From Windows 10, UNIX domain sockets (a.k.a "local" sockets) are supported on Windows, with the exception of the connect_pair function (which will fail with an operation_not_supported error).
Added executor-converting construction and assignment to ip::basic_resolver.
Added compatibility between polymorphic executors and the (deprecated) handler invocation hook.
Added the experimental::as_single completion token adapter. The as_single completion token adapter can be used to specify that the completion handler arguments should be combined into a single argument. For completion signatures with a single parameter, the argument is passed through as-is. For signatures with two or more parameters, the arguments are combined into a tuple. The as_single adapter may be used in conjunction with use_awaitable and structured bindings as follows:
```
auto [e, n] = co_await socket.async_read_some(
    asio::buffer(data), as_single(use_awaitable));
```
Alternatively, it may be used as a default completion token like so:
```
using default_token = as_single_t<use_awaitable_t<>>;
using tcp_socket = default_token::as_default_on_t<tcp::socket>;
// ...
awaitable<void> do_read(tcp_socket socket)
{
  // ...
  auto [e, n] = co_await socket.async_read_some(asio::buffer(data));
  // ...
}
```
Added support for MSG_NOSIGNAL on more platforms by using _POSIX_VERSION to detect whether it is supported.
Added the ability to compile using libpthread on Windows.
Added workarounds for the Intel C++ compiler.
Added more support for detecting and optimising for handlers that have no custom executor.
Reduced lock contention for timer cancellation on Windows.
Reinstated a previously removed null-pointer check, as it had a measurable impact on performance.
Fixed the executor concept to test for a const-qualified execute().
Fixed any_executor support for builds without RTTI support.
Fixed the thread_pool unit test to work without RTTI support.
Fixed C++20 coroutines compatibility with clang on Windows.
Fixed some compatibility issues with Windows Runtime.
Fixed shadow name warnings caused by addition of asio::query.
Fixed a "logical ‘or’ of equal expressions" warning on linux.
Fixed a benign switch fallthrough warning.
Added missing push/pop_options.hpp includes.
Suppressed zero-as-null-pointer-constant warnings.
Fixed a comma-operator warning.
Updated the documentation to clarify when the select reactor is used on Windows.
Fixed potential ambiguity caused by any_executor comparisons and conversion.
Added detection of non-experimental C++20 coroutines on MSVC 19.8.
Fixed compatibility with uClibc.
Fixed strand<> adaptation of Networking TS executors when targeting older C++ versions or less conformant compilers.

Marked the basic_socket_acceptor move constructor as noexcept.
Added workarounds for various issues in gcc 10's coroutine support.
Added standard executor support to windows::overlapped_ptr.
Added missing compatibility macros for the execution::receiver_of concept.
Added short-circuited evaluation to execution::executor, execution::operation_state,execution::sender, execution::receiver, execution::receiver_of, and execution::scheduler concepts and the corresponding traits. These now test first for well-formed CPO expressions (or, in the case of senders, a specialised sender_traits template) and, if not valid, short-circuit the remainder of the evaluation. This helps prevent recursive template instantiations that can occur in some contexts.
Added constraints to the strand template's constructor, to prevent template instantiation recursion.
Changed Asio's internal executor adapters to prevent template instantiation recursion.
Added constraints to the execution::any_executor template's converting constructors, as per the specification.
Added missing execution::sender_traits specialisation and connect() member function to the thread_pool executor, as per the specification.
Changed execution::blocking_t::always_t::is_preferable to be false as per the specification.
Added shape_type and index_type to thread_pool executors, as per the specification.
Ensured that the standard executor concept-related traits (such as execution::is_executor) work with void.
Fixed async_compose support for standard executors.
Fixed the forward declaration of any_io_executor in asio/ts/netfwd.hpp.
Fixed use_future compatibility with older compilers.
Fixed compatibility with MinGW.
Improved compatibility with older versions of gcc.
Added compatiblity with clang-cl on Windows.
Fixed various compiler warnings.
Added basic overview documentation for the standard executors support.
Added detailed descriptions to the standard executor-related member functions of io_context executors, thread_pool executors, system_executor, and strand<>.
Changed introductory documentation to reflect that there our now multiple types of I/O execution context.
Marked constructors, destructors, and static data members in class synopses in the reference documentation.
Fixed various minor documentation errors.

Added an implementation of the proposed standard executors (P0443r13, P1348r0, and P1393r0).
Added support for the proposed standard executors to Asio's I/O facilities:
- The io_context::executor_type, thread_pool::executor_type, system_executor, and strand executors now meet the requirements for the proposed standard executors. These classes also continue to meet the existing requirements for the Networking TS model of executors.
- All I/O objects, asynchronous operations, and utilities including dispatch, post, defer, get_associated_executor, bind_executor, make_work_guard, spawn, co_spawn, async_compose, use_future, etc., can interoperate with both new proposed standard executors, and with existing Networking TS executors. The implementation determines at compile time which model a particular executor meets; the proposed standard executor model is used in preference if both are detected.
- The any_io_executor type alias has been introduced as the new default runtime-polymorphic executor for all I/O objects. This type alias points to the execution::any_executor<> template with a set of supportable properties specified for use with I/O. This change may break existing code that directly uses the old polymorphic wrapper, executor. If required for backward compatibility, ASIO_USE_TS_EXECUTOR_AS_DEFAULT can be defined, which changes the any_io_executor type alias to instead point to the executor polymorphic wrapper.
- Support for the existing Networking TS model of executors can be disabled by defining ASIO_NO_TS_EXECUTORS.

Added converting move construction and assignment to basic_waitable_timer. This enables move construction and assignment between different timer types, provided the executor types are convertible. For example:

basic_waitable_timer<
    clock_type,
    traits_type,
    io_context::executor_type
  > timer1(my_io_context);

basic_waitable_timer<
    clock_type,
    traits_type,
    any_io_executor // polymorphic wrapper
  > timer2(std::move(timer1));

Enabled C++20 coroutine support when using gcc 10.

Added overloads of co_spawn that launch an awaitable. This change allows us to write:

co_spawn(executor,
    echo(std::move(socket)),
    detached);

instead of:

co_spawn(executor,
    [socket = std::move(socket)]() mutable
    {
      return echo(std::move(socket));
    },
    detached);

Added a new constructor overload to use_awaitable_t's default executor adapter, to enable conversion between executor types.
Added support for using detached_t as a default completion token, by adding members as_default_on() and as_default_on_t<>.
Added a move constructor to ssl::stream<>.
Changed ssl::stream<> write operations to linearise gather-write buffer sequences.
Added compile-time detection of the deprecated asio_handler_invoke hook. This hook was deprecated with the introduction of the Networking TS trait associated_executor and function get_associated_executor(). Compiling an application with ASIO_NO_DEPRECATED will now trigger a compile error if any handler implements the asio_handler_invoke hook.
Added compile-time detection of the deprecated asio_handler_allocate and asio_handle_deallocate hooks. These hooks were deprecated with the introduction of the Networking TS trait associated_allocator and function get_associated_allocator(). Compiling an application with ASIO_NO_DEPRECATED will now trigger a compile error if any handler implements the asio_handler_allocate or asio_handler_deallocate hooks.
Implemented a number of performance optimisations, including:
- Specialising single-buffer operations to use recv rather than recvmsg, send rather than sendmsg, read rather than readv, and write rather than writev.
- Lightening the reference counting overhead of the polymorphic wrapper executor.
- Returning from system call operation wrappers as early as possible, and only accessing errno and error codes when on an error path.
- Applying additional optimisations if a "native" I/O executor (such as io_context::exeutor_type) is detected.
Added source location support to handler tracking. The new ASIO_HANDLER_LOCATION((file_name, line, function_name)) macro may be used to inform the handler tracking mechanism of a source location. This macro declares an object that is placed on the stack. Then, when an asynchronous operation is launched with location information, it outputs lines using the <action> n^m, prior to the n*m line that signifies the beginning of the asynchronous operation. For example:
```
@asio|1589423304.861944|>7|ec=system:0,bytes_transferred=5
@asio|1589423304.861952|7^8|in 'async_write' (./../../../include/asio/impl/write.hpp:330)
@asio|1589423304.861952|7^8|called from 'do_write' (handler_tracking/async_tcp_echo_server.cpp:62)
@asio|1589423304.861952|7^8|called from 'operator()' (handler_tracking/async_tcp_echo_server.cpp:51)
@asio|1589423304.861952|7*8|socket@0x7ff61c008230.async_send
@asio|1589423304.861975|.8|non_blocking_send,ec=system:0,bytes_transferred=5
@asio|1589423304.861980|<7|
```
If std::source_location or std::experimental::source_location are available, the use_awaitable_t token (when default-constructed or used as a default completion token) will also cause handler tracking to output a source location for each newly created asynchronous operation. A use_awaitable_t object may also be explicitly constructed with location information.
Implemented various improvements to the handlerviz.pl tool.
- Add nodes for pending handlers at bottom of graph, outlined in red.
- Display source location in a tooltip on the edge label (for SVG).
- Use invisible nodes to enforce order to keep related control flow vertical.
Added the handlerlive.pl tool, which processes handler tracking output to produce a list of "live" handlers. Live handlers are those that are associated with pending asynchronous operations, as well as handlers that are currently executing. For example:
```
cat output.txt | perl handlerlive.pl
```
or:
```
perl handerlive.pl < output.txt
```
or:
```
perl handlerlive.pl output.txt
```
Added the handlertree.pl tool, which filters handler tracking output to include only those events in the tree that produced the nominated handlers. For example, to filter the output to include only the events associated with handlers 123, 456, and their predecessors:
```
cat output.txt | perl handlertree.pl 123 456
```
or:
```
perl handlertree.pl 123 456 < output.txt
```
This script may be combined with handerlive.pl and handlerviz.pl to produce a graph of the "live" asynchronous operation chains. For example:
```
cat output.txt | \
  perl handlertree.pl `perl handlerlive.pl output.txt` | \
  perl handlerviz.pl | \
  dot -Tsvg > output.svg
```
Added changes for clang-based Embarcadero C++ compilers.
Fixed a deadlock that can occur when multiple threads concurrently initialise the Windows I/O completion port backend.
Fixed async_compose to work with copyable handlers when passed by lvalue.
Fixed completion signature deduction in co_spawn.
Removed a spurious Executor base class from the executor_binder implementation.
Various fixes and improvements in the documentation and examples.

Fixed compatibility with C++20 concept syntax.
Marked the POSIX descriptor classes' move constructors as noexcept.
Added the ssl::host_name_verification class, which is a drop-in replacement for ssl::rfc2818_verification. The ssl::rfc2818_verification class has been marked as deprecated. As a consequence of this change, SSL support now depends on functions that were introduced in OpenSSL 1.0.2.
Added an ssl::context constructor to take ownership of a native handle.
Changed C++ language version detection with gcc to use __cplusplus macro.
Fixed a work counting issue in the asynchronous resolve operation for endpoints.
Fixed the strand<> converting constructors and assignment operators.
Ensured that resolvers are restarted correctly after a fork.
Fixed compatibility with the current NetBSD release.
Removed spurious handler requirement checks in some async_read overloads.
Changed the ssl::context class to propagate non-EOF errors from the add_certificate_authority function.
Fixed a Windows-specific thread_pool destructor hang that occurred when the pool had an associated I/O object.
Changed the select reactor to recreate the "self pipe trick" sockets on error. This addresses an issue on some versions of Windows, where these sockets are discconected after a system sleep.
Fixed a compile error in the buffered streams due to the lack of reference collapsing in C++98.
Changed the priority_scheduler example to demonstrate calls to shutdown() and destroy().
Removed some unnecessary null pointer checks.
Changed Windows platform detection to recognise TV titles as Windows apps.
Added some emscripten compatibility patches.
Fixed a compile error in the use_awaitable_t::as_default_on function.
Changed all uses of the boost.bind placeholders to use the boost::placeholders namespace.
Fixed a potential compile error in the async_compose implementation due to incorrect overload selection.
Suppressed some non-virtual destructor warnings.
Various documentation fixes and improvements.

Changed the async_initiate helper function to automatically deduce its return type. This is enabled for C++11 or later.
Changed all asynchronous operations to use automatically deduced return types. This allows completion token implementations to incorporate the asynchronous operation initiation into the initiating function's return type, without type erasure. Note that C++14 or later is required to support completion tokens that use per-operation return type deduction. For C++11 or earlier, a completion token's async_result specialisation must still provide the nested typedef return_type.
Introduced three new concepts to support async_initiate.
- completion_signature<T>: Checks if T is a signature of the form R(Args...).
- completion_handler_for<T, Signature>: Checks if T is usable as a completion handler with the specified signature.
- completion_token_for<T, Signature>: Checks if T is a completion token that can be used with async_initiate and the specified signature.
- For backward compatibility with pre-concepts C++, the macros (BOOST_)ASIO_COMPLETION_SIGNATURE, (BOOST_)ASIO_COMPLETION_HANDLER_FOR, and (BOOST_)ASIO_COMPLETION_TOKEN_FOR are provided. These macros expand to typename when concepts are unsupported.
Added the nested template type rebind_executor to all I/O object types, as a way to generically rebind them to use alternative I/O executors. For example:
```
using my_socket_type = tcp::socket::rebind_executor<my_executor_type>::other;
```
Changed the asynchronous operations' initiation function objects to report their associated I/O executor via the nested type executor_type and member function get_executor(). Note that the presence of executor_type and get_executor() should be treated as optional, and consequently it may be preferable to access them via the associated_executor trait and the get_associated_executor() helper function.
Added the default_completion_token trait, so that every I/O executor type now has an associated default completion token type. This trait may be used in asynchronous operation declarations as follows:
```
template <
    typename IoObject,
    typename CompletionToken =
      typename default_completion_token<
        typename IoObject::executor_type
      >::type
  >
auto async_xyz(
    IoObject& io_object,
    CompletionToken&& token =
      typename default_completion_token<
        typename IoObject::executor_type
      >::type{}
  );
```
If not specialised, this trait type is void, meaning no default completion token type is available for the given I/O executor.
Specialised the default_completion_token trait for the use_awaitable completion token, so that it may be used as shown in the following example:
```
auto socket = use_awaitable.as_default_on(tcp::socket(my_context));
// ...
co_await socket.async_connect(my_endpoint); // Defaults to use_awaitable.
```
In this example, the type of the socket object is transformed from tcp::socket to have an I/O executor with the default completion token set to use_awaitable. Alternatively, the socket type may be computed directly:
```
using tcp_socket = use_awaitable_t<>::as_default_on_t<tcp::socket>;
tcp_socket socket(my_context);
// ...
co_await socket.async_connect(my_endpoint); // Defaults to use_awaitable.
```
Added missing async_initiate to the Windows-specific I/O objects' asynchronous operations.
Ensured that the executor type is propagated to newly accepted sockets. When synchronously or asynchronously accepting a new connection, but without specifying an executor or execution context, the accept operation will now correctly propagate the executor type from the acceptor to the socket. For example, if your acceptor type is:
```
basic_socket_acceptor<ip::tcp, my_executor_type>
```
then your accepted socket type will be:
```
basic_stream_socket<ip::tcp, my_executor_type>
```
Changed to require that Protocol copy and move operations never throw.
Changed to require that Endpoint default constructor and move operations never throw.
Added the noexcept qualifier to protocol accessors.
Added the noexcept qualifier to socket move constructors.
Fixed issues associated with opening serial ports on Windows:
- Use the correct constant to initialise the RTS control flag.
- Specify a default baud rate (9600).
Fixed a lost "outstanding work count" that can occur when an asynchronous accept operation is automatically restarted.

Improved performance slightly by eliminating a redundant move construction when completed handlers are dispatched.
Eliminated a compiler warning by annotating a case fall-through in the free function connect() implementation.
Fixed the is_*_buffer_sequence detection traits for user-defined sequence types.
Fixed some Windows-specific warnings about an incompatible pointer cast when obtaining the CancelIoEx entry point.
Changed to automatically set the defaults when opening a serial port on Windows.
Changed the serial port get_option() member function to be const.
Fixed a name hiding issue with the WinRT stream-oriented socket backend's shutdown function.
Applied a minor fix to the documentation for is_dynamic_buffer.
Added some support for Haiku OS.
Added wolfSSL compatability.
Changed to require C++17 or later for coroutines TS support with clang.
Fixed a doxygen generation problem in the tutorial.

Improved I/O object performance by adding runtime detection of native I/O executors when using the polymorphic executor wrapper.
Changed I/O object move constructors so that the executor is copied, not moved. This ensures that the moved-from I/O object is left in the same state as if constructed with a valid executor but without a resource.
On Windows, fixed an issue where global object destructors were not being run.
Fixed move-based async_accept between sockets with different executor types.

Added custom I/O executor support to I/O objects.
- All I/O objects now have an additional Executor template parameter. This template parameter defaults to the asio::executor type (the polymorphic executor wrapper) but can be used to specify a user-defined executor type.
- I/O objects' constructors and functions that previously took an asio::io_context& now accept either an Executor or a reference to a concrete ExecutionContext (such as asio::io_context or asio::thread_pool).
- Note: One potential source of breakage in existing user code is when reusing an I/O object's io_context for constructing another I/O object, as in:
```
asio::steady_timer my_timer(my_socket.get_executor().context());
```
  To fix this, either construct the second I/O object using the first I/O object's executor:
```
asio::steady_timer my_timer(my_socket.get_executor());
```
  or otherwise explicitly pass the io_context:
```
asio::steady_timer my_timer(my_io_context);
```
- The previously deprecated get_io_context and get_io_service member functions have now been removed.
- The previously deprecated service template parameters, and the corresponding classes, have now been removed.
Added a new async_result form with an initiate static member function.
- The async_result template now supports a new form:
```
template <typename CompletionToken, typename Signature>
struct async_result
{
  typedef /* ... */ return_type;

  template <typename Initiation,
      typename RawCompletionToken,
      typename... Args>
  static return_type initiate(
      Initiation&& initiation,
      RawCompletionToken&& token,
      Args&&... args);
};
```
- The initiate member function must: (a) transform the token into a completion handler object handler; (b) cause the invocation of the function object initiation as if by calling std::forward<Initiation>(initiation)(std::move(handler), std::forward<Args>(args)...). Note that the invocation of initiation may be deferred (e.g. lazily evaluated), in which case initiation and args must be decay-copied and moved as required.
- A helper function template async_initiate has also been added as a wrapper for the invocation of async_result<>::initiate. For backward compatibility, this function supports both the old and new async_result forms.
- The composed operations examples have been updated to use async_initiate.
- The previously deprecated handler_type trait and single-argument form of async_result have now been removed.
Updated the Coroutines TS support and promoted it to the asio namespace.
- The awaitable<>, co_spawn, this_coro, detached, and redirect_error facilities have been moved from the asio::experimental namespace to namespace asio. As part of this change, the this_coro::token() awaitable has been superseded by the asio::use_awaitable completion token.
- Please note that the use_awaitable and redirect_error completion tokens work only with asynchronous operations that use the new form of async_result with member function initiate. Furthermore, when using use_awaitable, please be aware that the asynchronous operation is not initiated until co_await is applied to the awaitable<>.
Added a new DynamicBuffer_v2 concept which is CopyConstructible.
- This change adds a new set of type requirements for dynamic buffers, DynamicBuffer_v2, which supports copy construction. These new type requirements enable dynamic buffers to be used as arguments to user-defined composed operations, where the same dynamic buffer object is used repeatedly for multiple underlying operations. For example:
```
template <typename DynamicBuffer>
void echo_line(tcp::socket& sock, DynamicBuffer buf)
{
  n = asio::read_until(sock, buf, '\n');
  asio::write(sock, buf, asio::transfer_exactly(n));
}
```
- The original DynamicBuffer type requirements have been renamed to DynamicBuffer_v1. These requirements continue to be compatible with the Networking TS.
- New type traits is_dynamic_buffer_v1 and is_dynamic_buffer_v2 have been added to test for conformance to DynamicBuffer_v1 and DynamicBuffer_v2 respectively. The existing is_dynamic_buffer trait has been retained and delegates to is_dynamic_buffer_v1 (unless ASIO_NO_DYNAMIC_BUFFER_V1 is explicitly defined, in which case it delegates to is_dynamic_buffer_v2).
- For convenience, the dynamic_string_buffer and dynamic_vector_buffer classes conform to both DynamicBuffer_v1 and DynamicBuffer_v2 requirements.
- When ASIO_NO_DYNAMIC_BUFFER_V1 is defined, all support for DynamicBuffer_v1 types and functions is #ifdef-ed out. Support for using basic_streambuf with the read, async_read, read_until, async_read_until, write, and async_write functions is also disabled as a consequence.
- Note: This change should have no impact on existing source code that simply uses dynamic buffers in conjunction with Asio's composed operations.
Added a new async_compose function that simplifies the implementation of user-defined asynchronous operations.
Added a make_strand function, which creates a strand with a deduced Executor template argument.
Relaxed the completion condition type requirements to only require move-constructibility rather than copy-constructibility.
Added a constructor for local::basic_endpoint that takes a string_view.
Added the noexcept qualifier to various member functions of the ip::address, ip::address_v4, ip::address_v6, ip::basic_endpoint, and executor_work_guard classes.
Added the noexcept qualifier to the buffer_sequence_begin and buffer_sequence_end functions.
Added a new ASIO_DISABLE_VISIBILITY configuration #define that allows visibility pragmas to be disabled. (Note: If symbols are hidden, extra care must be taken to ensure that Asio types are not passed across shared library API boundaries.)
Changed compile-time feature detection to define ASIO_STANDALONE automatically if C++11 or later is detected.
- Users should define ASIO_ENABLE_BOOST to explicitly disable standalone mode when compiling with C++11 or later.
- The configure script now defaults to a standalone build unless Boost is specified or detected.
Enabled recycling of the memory used to type-erase a function object with the polymorphic executor.
Changed receive operations to return the correct number of bytes transferred when truncation (error::message_size) occurs on a datagram-oriented socket.
Fixed multicast behaviour on QNX by automatically applying SO_REUSEPORT when the reuse_address option is set.
Added inclusion of unistd.h when targeting Haiku OS, to fix feature detection.
Added the network_v[46].hpp headers to the top-level convenience header.
Fixed calculation of absolute timeout when the backend uses pthread_cond_timedwait.
Changed the range-based asynchronous connect operation to deduce the EndpointSequence iterator type rather than assume the presence of a const_iterator typedef.
Fixed buffer_sequence_begin and buffer_sequence_end to prevent implicit conversion. This change addresses an issue where a call to buffer_sequence_begin or buffer_sequence_end could trigger an implicit conversion to const_buffer or mutable_buffer. Whenever this implicit conversion occurred, the return value of buffer_sequence_begin or buffer_sequence_end would point to a temporary object.
Ensured SSL handshake errors are propagated to the peer before the local operation completes.
Suppressed the eof error on SSL shutdown as it actually indicates success.
Added a fallback error code for when we OpenSSL produces an SSL_ERROR_SYSCALL result without an associated error.
Changed composed asynchronous read and write operations to move, rather than copy, the buffer sequence objects when the composed operation implementation is moved.
Changed to use <atomic> when targeting apple/clang/libc++ with recent Xcode versions, even for C++03. This fixes a warning about the deprecation of OSMemoryBarrier.
Fixed compile errors that occur when using the composed read and write operations with MSVC 11.0, by disabling decltype support for that compiler.
Increased the default value of _WIN32_WINNT to 0x0601 (Windows 7).
Fixed dispatch documentation to note that it may call the supplied function object in the current thread.
Updated post and defer documentation to clarify the the distinction between them.

Fixed a problem with the detection of std::future availability with libstdc++.
Fixed compile error in regex overload of read_until.
Fixed a timer heap corruption issue that can occur when moving a cancelled timer.
Fixed detection of std::experimental::string_view and std::string_view with newer clang/libc++.
Fixed MSVC version detection for availability of std::invoke_result.
Fixed the buffer sequence traits to test the new requirements, if decltype is available.
Fixed an MSVC issue when building with exceptions disabled.
Added SSL context options for TLS v1.3.
Added a compile-time test for TLS v1 support.
Fixed the macro used to test for TLS v1.2 support.
Prevented global objects from being created once per thread on Windows.
Fixed a crash when using size(), max_size() or empty() on default-constructed resolver results.
Changed to move the return value in basic_resolver_results::begin() to avoid copying.
Enabled move support for the Intel Compiler.
Fixed std::string_view detection issue when using clang-cl.
Fixed the handler tracking operation name for io_context::executor_type::dispatch.
Fixed a buffer overflow that could occur when parsing an address string with a 64-bit scope id.
Added examples showing how to write composed operations.
Added C++11 versions of the Timeouts, Timers, SOCKS4 and SSL examples.
Fixed minor issues in documentation and examples.

Added missing const qualifier to basic_socket_acceptor::get_option.
Worked around a parsing error that occurs with some versions of gcc.
Fixed broken code samples in tutorial.
Added new experimental features. (Note that "experimental" features may be changed without notice in subsequent releases.)
- Added experimental::detached completion token.
- Added experimental::redirect_error completion token.
- Added experimental::co_spawn facility for integration with the coroutines technical specification.
Updated timeout examples to use latest features.
- Used asio::steady_timer rather than asio::deadline_timer.
- Used asio::dynamic_buffer rather than asio::streambuf.
- Used timed asio::io_context::run_for() function for blocking clients.
- Added example showing a custom completion token for blocking with timeouts.
Fixed unit tests to compile when (BOOST_)ASIO_NO_DEPRECATED is defined.
Changed socket iostreams to use chrono by default, to fix compatibility with the Networking TS. Define (BOOST_)ASIO_USE_BOOST_DATE_TIME_FOR_SOCKET_IOSTREAM to enable the old Boost.Date_Time interface in basic_socket_streambuf and basic_socket_iostream.
Updated examples to use chrono rather than Boost.Date_Time.
Fixed an incorrect member function detector in the is_dynamic_buffer trait.
Fixed an async_result incompatibility with deprecated handler_type.
Added a missing move optimisation in the SSL stream implementation.
Fixed incorrect basic_resolver_results::value_type typedef.
Fixed a compile error with some OpenSSL versions when SSL_OP_NO_COMPRESSION is defined.
Changed add_certificate_authority to process multiple certificates in a bundle.
Eliminated deprecation warning with MSVC by using std::invoke_result rather than std::result_of.
Changed to use std::string_view for C++17 or later, and std::experimental::string_view for C++14. Define the preprocessor macro (BOOST_)ASIO_DISABLE_STD_STRING_VIEW to force the use of std::experimental::string_view (assuming it is available) when compiling in C++17 mode.
Ensured DynamicBuffer template arguments are decayed before using in enable_if tests.
Changed documentation to distinguish legacy completion handlers (which are still required to be CopyConstructible) from new MoveConstructible handlers.
Suppressed a discarded return value warning in the buffer debugging support.
Fixed basic_yield_context to work with completion signatures containing reference parameters.
Ensured that stackful coroutines launched using spawn() correctly store decayed copies of their function and handler arguments.
Fixed some compatibility issues with Android.
Fixed some minor portability issues in examples.

Completed the interface changes to reflect the Networking TS (N4656).
- See the list of new interfaces and, where applicable, the corresponding old interfaces that have been superseded.
- The service template parameters, and the corresponding classes, are disabled by default. For example, instead of basic_socket<Protocol, SocketService> we now have simply basic_socket<Protocol>. The old interface can be enabled by defining the (BOOST_)ASIO_ENABLE_OLD_SERVICES macro.
Added support for customised handler tracking.
Added reactor-related (i.e. descriptor readiness) events to handler tracking.
Added special concurrency hint values that may be used to disable locking on a per io_context basis.
Enabled perfect forwarding for the first ssl::stream<> constructor argument.
Added ability to release ownership of the underlying native socket. (Requires Windows 8.1 or later when using the I/O completion port backend.)

Implemented changes to substantially reflect the Networking Library Proposal (N4370).
- New Executor type requirements and classes to support an executor framework, including the execution_context base class, the executor_work class for tracking outstanding work, and the executor polymorphic wrapper. Free functions dispatch(), post() and defer() have been added and are used to submit function objects to executors.
- Completion handlers now have an associated executor and associated allocator. The free function wrap() is used to associate an executor with a handler or other object. The handler hooks for allocation, invocation and continuation have been deprecated.
- A system_executor class has been added as a default executor.
- The io_service class is now derived from execution_context and implements the executor type requirements in its nested executor_type class. The member functions dispatch(), post(), defer() and wrap() have been deprecated. The io_service::work class has been deprecated.
- The io_service member function reset() has been renamed to restart(). The old name is retained for backward compatibility but has been deprecated.
- The make_service<>() function is now used to add a new service to an execution context such as an io_service. The add_service() function has been deprecated.
- A new strand<> template has been added to allow strand functionality to be used with generic executor types.
- I/O objects (such as sockets and timers) now provide access to their associated io_service via a context() member function. The get_io_service() member function is deprecated.
- All asynchronous operations and executor operations now support move-only handlers. However, the deprecated io_service::post(), io_service::dispatch(), io_service::strand::post() and io_service::strand::dispatch() functions still require copyable handlers.
- Waitable timer objects are now movable.
- Waitable timers, socket iostreams and socket streambufs now provide an expiry() member function for obtaining the expiry time. The accessors expires_at() and expires_after() have been deprecated, though those names are retained for the mutating members.
- The std::packaged_task class template is now supported as a completion handler. The initiating operation automatically returns the future associated with the task. The package() function has been added as a convenient factory for packaged tasks.
- Sockets, socket acceptors and descriptors now provide wait() and async_wait() operations that may be used to wait for readiness. The null_buffers type has been deprecated.
- The proposed error code enum classes are simulated using namespaces. Existing asio error codes now have a correspondence with the standard error conditions.
- Conversion between IP address types, and conversion from string to address, is now supported via the address_cast<>(), make_address(), make_address_v4() and make_address_v6() free functions. The from_string(), to_v4(), to_v6() and v4_mapped() member functions have been deprecated.
- A default-constructed ip::address now represents an invalid address value that is neither IPv4 nor IPv6.
- New buffer() overloads that generate mutable buffers for non-const string objects.
- Support for dynamic buffer sequences that automatically grow and shrink to accomodate data as it is read or written. This is a generic facility similar to the existing asio::streambuf class. This support includes:
  - New dynamic_string_buffer and dynamic_vector_buffer adapter classes that meet the DynamicBufferSequence type requirements.
  - New dynamic_buffer() factory functions for creating a dynamic buffer adapter for a vector or string.
  - New overloads for the read(), async_read(), write() and async_write(), read_until() and async_read_until() free functions that directly support dynamic buffer sequences.
- Support for networks and address ranges. Thanks go to Oliver Kowalke for contributing to the design and providing the implementation on which this facility is based. The following new classes have been added:
  - address_iterator_v4 for iterating across IPv4 addresses
  - address_iterator_v6 for iterating across IPv6 addresses
  - address_range_v4 to represent a range of IPv4 addresses
  - address_range_v6 to represent a range of IPv6 addresses
  - network_v4 for manipulating IPv4 CIDR addresses, e.g. 1.2.3.0/24
  - network_v6 for manipulating IPv6 CIDR addresses, e.g. ffe0:/120
- New convenience headers in <asio/ts/*.hpp> that correspond to the headers in the proposal.
Added a new, executor-aware thread_pool class.
Changed spawn() to be executor-aware.
Added a new spawn() overload that takes only a function object.
Changed spawn() and yield_context to permit nested calls to the completion handler.
Removed previously deprecated functions.
Added options for disabling TLS v1.1 and v1.2.
Changed the SSL wrapper to call the password callback when loading an in-memory key.
Changed the tutorial to use std::endl to ensure output is flushed.
Fixed false SSL error reports by ensuring that the SSL error queue is cleared prior to each operation.
Fixed an ssl::stream<> bug that may result in spurious 'short read' errors.
Enabled perfect forwarding for the first ssl::stream<> constructor argument.
Added standalone Asio support for Clang when used with libstdc++ and C++11.
Fixed an unsigned integer overflow reported by Clang's integer sanitizer.
Added support for move-only return types when using a yield_context object with asynchronous operations.
Ensured errors generated by Windows' ConnectEx function are mapped to their portable equivalents.
Changed multicast test to treat certain join_group failures as non-fatal.

Fixed the kqueue reactor so that it works on FreeBSD.
Fixed an issue in the kqueue reactor which resulted in spinning when using serial ports on Mac OS.
Fixed kqueue reactor support for read-only file descriptors.
Fixed a compile error when using the /dev/poll reactor.
Changed the Windows backend to use WSASocketW, as WSASocketA has been deprecated.
Fixed some warnings reported by Visual C++ 2013.
Fixed integer type used in the WinRT version of the byte-order conversion functions.
Changed documentation to indicate that use_future and spawn() are not made available when including the asio.hpp convenience header.
Explicitly marked asio::strand as deprecated. Use asio::io_service::strand instead.

Stopped using certain Winsock functions that are marked as deprecated in the latest Visual C++ and Windows SDK.
Fixed a shadow variable warning on Windows.
Fixed a regression in the kqueue backend that was introduced in Asio 1.10.2.
Added a workaround for building the unit tests with gcc on AIX.

Worked around a gcc problem to do with anonymous enums.
Reverted the Windows HANDLE backend change to ignore ERROR_MORE_DATA. Instead, the error will be propagated as with any other (i.e. in an error_code or thrown as a system_error), and the number of bytes transferred will be returned. For code that needs to handle partial messages, the error_code overload should be used.
Fixed an off-by-one error in the signal_set implementation's signal number check.
Changed the Windows IOCP backend to not assume that SO_UPDATE_CONNECT_CONTEXT is defined.
Fixed a Windows-specific issue, introduced in Asio 1.10.2, by using VerifyVersionInfo rather than GetVersionEx, as GetVersionEx has been deprecated.
Changed to use SSE2 intrinsics rather than inline assembly, to allow the Cray compiler to work.

Fixed asio::spawn() to work correctly with new Boost.Coroutine interface.
Ensured that incomplete asio::spawn() coroutines are correctly unwound when cleaned up by the io_service destructor.
Fixed delegation of continuation hook for handlers produced by io_service::wrap() and strand::wrap().
Changed the Windows I/O completion port backend to use ConnectEx, if available, for connection-oriented IP sockets.
Changed the io_service backend for non-Windows (and non-IOCP Windows) platforms to use a single condition variable per io_service instance. This addresses a potential race condition when run_one() is used from multiple threads.
Prevented integer overflow when computing timeouts based on some boost::chrono and std::chrono clocks.
Made further changes to EV_CLEAR handling in the kqueue backend, to address other cases where the close() system call may hang on Mac OS X.
Fixed infinite recursion in implementation of resolver_query_base::flags::operator~.
Made the select reactor more efficient on Windows for large numbers of sockets.
Fixed a Windows-specific type-aliasing issue reported by gcc.
Prevented execution of compile-time-only buffer test to avoid triggering an address sanitiser warning.
Disabled the GetQueuedCompletionStatus timeout workaround on recent versions of Windows.
Changed implementation for Windows Runtime to use FormatMessageW rather than FormatMessageA, as the Windows store does not permit the latter.
Added support for string-based scope IDs when using link-local multicast addresses.
Changed IPv6 multicast group join to use the address's scope ID as the interface, if an interface is not explicitly specified.
Fixed multicast test failure on Mac OS X and the BSDs by using a link-local multicast address.
Various minor documentation improvements.

Implemented a limited port to Windows Runtime. This support requires that the language extensions be enabled. Due to the restricted facilities exposed by the Windows Runtime API, the port also comes with the following caveats:
- The core facilities such as the io_service, strand, buffers, composed operations, timers, etc., should all work as normal.
- For sockets, only client-side TCP is supported.
- Explicit binding of a client-side TCP socket is not supported.
- The cancel() function is not supported for sockets. Asynchronous operations may only be cancelled by closing the socket.
- Operations that use null_buffers are not supported.
- Only tcp::no_delay and socket_base::keep_alive options are supported.
- Resolvers do not support service names, only numbers. I.e. you must use "80" rather than "http".
- Most resolver query flags have no effect.
Extended the ability to use Asio without Boost to include Microsoft Visual Studio 2012. When using a C++11 compiler, most of Asio may now be used without a dependency on Boost header files or libraries. To use Asio in this way, define ASIO_STANDALONE on your compiler command line or as part of the project options. This standalone configuration has been tested for the following platforms and compilers:
- Microsoft Visual Studio 2012
- Linux with g++ 4.7 or 4.8 (requires -std=c++11)
- Mac OS X with clang++ / Xcode 4.6 (requires -std=c++11 -stdlib=libc++)
Fixed a regression (introduced in 1.10.0) where, on some platforms, errors from async_connect were not correctly propagated through to the completion handler.
Fixed a Windows-specific regression (introduced in 1.10.0) that occurs when multiple threads are running an io_service. When the bug occurs, the result of an asynchronous operation (error and bytes tranferred) is incorrectly discarded and zero values used instead. For TCP sockets this results in spurious end-of-file notifications.
Fixed a bug in handler tracking, where it was not correctly printing out some handler IDs.
Fixed the comparison used to test for successful synchronous accept operations so that it works correctly with unsigned socket descriptors.
Ensured the signal number is correctly passed to the completion handler when starting an async_wait on a signal that is already raised.
Suppressed a g++ 4.8+ warning about unused typedefs.
Enabled the move optimisation for handlers that use the default invocation hook.
Clarified that programs must not issue overlapping async_write_at operations.
Changed the Windows HANDLE backend to treat ERROR_MORE_DATA as a non-fatal error when returned by GetOverlappedResult for a synchronous read.
Visual C++ language extensions use generic as a keyword. Added a workaround that renames the namespace to cpp_generic when those language extensions are in effect.
Fixed some asynchronous operations that missed out on getting async_result support in 1.10.0. In particular, the buffered stream templates have been updated so that they adhere to current handler patterns.
Enabled move support for Microsoft Visual Studio 2012.
Added use_future support for Microsoft Visual Studio 2012.
Removed a use of std::min in the Windows IOCP backend to avoid a dependency on the <algorithm> header.
Eliminated some unnecessary handler copies.
Fixed support for older versions of OpenSSL that do not provide the SSL_CTX_clear_options function.
Fixed various minor and cosmetic issues in code and documentation.

Added new traits classes, handler_type and async_result, that allow the customisation of the return type of an initiating function.
Added the asio::spawn() function, a high-level wrapper for running stackful coroutines, based on the Boost.Coroutine library. The spawn() function enables programs to implement asynchronous logic in a synchronous manner. For example: size_t n = my_socket.async_read_some(my_buffer, yield);. For further information, see Stackful Coroutines.
Added the asio::use_future special value, which provides first-class support for returning a C++11 std::future from an asynchronous operation's initiating function. For example: future<size_t> = my_socket.async_read_some(my_buffer, asio::use_future);. For further information, see C++ 2011 Support - Futures.
Promoted the stackless coroutine class and macros to be part of Asio's documented interface, rather than part of the HTTP server 4 example. For further information, see Stackless Coroutines.
Added a new handler hook called asio_handler_is_continuation. Asynchronous operations may represent a continuation of the asynchronous control flow associated with the current executing handler. The asio_handler_is_continuation hook can be customised to return true if this is the case, and Asio's implementation can use this knowledge to optimise scheduling of the new handler. To cover common cases, Asio customises the hook for strands, spawn() and composed asynchronous operations.
Added four new generic protocol classes, generic::datagram_protocol, generic::raw_protocol, generic::seq_packet_protocol and generic::stream_protocol, which implement the Protocol type requirements, but allow the user to specify the address family (e.g. AF_INET) and protocol type (e.g. IPPROTO_TCP) at runtime. For further information, see Support for Other Protocols.
Added C++11 move constructors that allow the conversion of a socket (or acceptor) into a more generic type. For example, an ip::tcp::socket can be converted into a generic::stream_protocol::socket via move construction. For further information, see Support for Other Protocols.
Extended the basic_socket_acceptor<>'s accept() and async_accept() functions to allow a new connection to be accepted directly into a socket of a more generic type. For example, an ip::tcp::acceptor can be used to accept into a generic::stream_protocol::socket object. For further information, see Support for Other Protocols.
Moved existing examples into a C++03-specific directory, and added a new directory for C++11-specific examples. A limited subset of the C++03 examples have been converted to their C++11 equivalents.
Add the ability to use Asio without Boost, for a limited set of platforms. When using a C++11 compiler, most of Asio may now be used without a dependency on Boost header files or libraries. To use Asio in this way, define ASIO_STANDALONE on your compiler command line or as part of the project options. This standalone configuration has currently been tested for the following platforms and compilers:
- Linux with g++ 4.7 (requires -std=c++11)
- Mac OS X with clang++ / Xcode 4.6 (requires -std=c++11 -stdlib=libc++)
Various SSL enhancements. Thanks go to Nick Jones, on whose work these changes are based.
- Added support for SSL handshakes with re-use of data already read from the wire. New overloads of the ssl::stream<> class's handshake() and async_handshake() functions have been added. These accept a ConstBufferSequence to be used as initial input to the ssl engine for the handshake procedure.
- Added support for creation of TLSv1.1 and TLSv1.2 ssl::context objects.
- Added a set_verify_depth() function to the ssl::context and ssl::stream<> classes.
- Added the ability to load SSL certificate and key data from memory buffers. New functions, add_certificate_authority(), use_certificate(), use_certificate_chain(), use_private_key(), use_rsa_private_key() and use_tmp_dh(), have been added to the ssl::context class.
- Changed ssl::context to automatically disable SSL compression by default. To enable, use the new ssl::context::clear_options() function, as in my_context.clear_options(ssl::context::no_compression).
Fixed a potential deadlock in signal_set implementation.
Fixed an error in acceptor example in documentation.
Fixed copy-paste errors in waitable timer documentation.
Added assertions to satisfy some code analysis tools.
Fixed a malformed #warning directive.
Fixed a potential data race in the Linux epoll implementation.
Fixed a Windows-specific bug, where certain operations might generate an error_code with an invalid (i.e. NULL) error_category.
Fixed basic_waitable_timer's underlying implementation so that it can handle any time_point value without overflowing the intermediate duration objects.
Fixed a problem with lost thread wakeups that can occur when making concurrent calls to run() and poll() on the same io_service object.
Fixed implementation of asynchronous connect operation so that it can cope with spurious readiness notifications from the reactor.
Fixed a memory leak in the ssl::rfc2818_verification class.
Added a mechanism for disabling automatic Winsock initialisation. See the header file asio/detail/winsock_init.hpp for details.

Fixed some 64-to-32-bit conversion warnings.
Fixed various small errors in documentation and comments.
Fixed an error in the example embedded in basic_socket::get_option's documentation.
Changed to use long rather than int for SSL_CTX options, to match OpenSSL.
Changed to use _snwprintf to address a compile error due to the changed swprintf signature in recent versions of MinGW.
Fixed a deadlock that can occur on Windows when shutting down a pool of io_service threads due to running out of work.
Changed UNIX domain socket example to treat errors from accept as non-fatal.
Added a small block recycling optimisation to improve default memory allocation behaviour.

Fixed an incompatibility between ip::tcp::iostream and C++11.
Decorated GCC attribute names with underscores to prevent interaction with user-defined macros.
Added missing #include <cctype>, needed for some versions of MinGW.
Changed to use gcc's atomic builtins on ARM CPUs, when available.
Changed strand destruction to be a no-op, to allow strand objects to be destroyed after their associated io_service has been destroyed.
Added support for some newer versions of glibc which provide the epoll_create1() function but always fail with ENOSYS.
Changed the SSL implementation to throw an exception if SSL engine initialisation fails.
Fixed another regression in buffered_write_stream.
Implemented various minor performance improvements, primarily targeted at Linux x86 and x86-64 platforms.

Changed the epoll_reactor backend to do lazy registration for EPOLLOUT events.
Fixed the epoll_reactor handling of out-of-band data, which was broken by an incomplete fix in the last release.
Changed Asio's SSL wrapper to respect OpenSSL's OPENSSL_NO_ENGINE feature test #define.
Fixed windows::object_handle so that it works with Windows compilers that support C++11 move semantics (such as g++).
Improved the performance of strand rescheduling.
Added support for g++ 4.7 when compiling in C++11 mode.
Fixed a problem where signal_set handlers were not being delivered when the io_service was constructed with a concurrency_hint of 1.

Added a new class template basic_waitable_timer based around the C++11 clock type requirements. It may be used with the clocks from the C++11 <chrono> library facility or, if those are not available, Boost.Chrono. The typedefs high_resolution_timer, steady_timer and system_timer may be used to create timer objects for the standard clock types.
Added a new windows::object_handle class for performing waits on Windows kernel objects. Thanks go to Boris Schaeling for contributing substantially to the development of this feature.
On Linux, connect() can return EAGAIN in certain circumstances. Remapped this to another error so that it doesn't look like a non-blocking operation.
Fixed a compile error on NetBSD.
Fixed deadlock on Mac OS X.
Fixed a regression in buffered_write_stream.
Fixed a non-paged pool "leak" on Windows when an io_service is repeatedly run without anything to do.
Reverted earlier change to allow some speculative operations to be performed without holding the lock, as it introduced a race condition in some multithreaded scenarios.
Fixed a bug where the second buffer in an array of two buffers may be ignored if the first buffer is empty.

Implemented various performance improvements, including:
- Using thread-local operation queues in single-threaded use cases (i.e. when concurrency_hint is 1) to eliminate a lock/unlock pair.
- Allowing some epoll_reactor speculative operations to be performed without holding the lock.
- Improving locality of reference by performing an epoll_reactor's I/O operation immediately before the corresponding handler is called. This also improves scalability across CPUs when multiple threads are running the io_service.
- Specialising asynchronous read and write operations for buffer sequences that are arrays (boost::array or std::array) of exactly two buffers.
Fixed a compile error in the regex overload of async_read_until.
Fixed a Windows-specific compile error by explicitly specifying the signal() function from the global namespace.
Changed the deadline_timer implementation so that it does not read the clock unless the timer heap is non-empty.
Changed the SSL stream's buffers' sizes so that they are large enough to hold a complete TLS record.
Fixed the behaviour of the synchronous null_buffers operations so that they obey the user's non-blocking setting.
Changed to set the size of the select fd_set at runtime when using Windows.
Disabled an MSVC warning due to const qualifier being applied to function type.
Fixed a crash that occurs when using the Intel C++ compiler.
Changed the initialisation of the OpenSSL library so that it supports all available algorithms.
Fixed the SSL error mapping used when the session is gracefully shut down.
Added some latency test programs.
Clarified that a read operation ends when the buffer is full.
Fixed an exception safety issue in epoll_reactor initialisation.
Made the number of strand implementations configurable by defining (BOOST_)ASIO_STRAND_IMPLEMENTATIONS to the desired number.
Added support for a new (BOOST_)ASIO_ENABLE_SEQUENTIAL_STRAND_ALLOCATION flag which switches the allocation of strand implementations to use a round-robin approach rather than hashing.
Fixed potential strand starvation issue that can occur when strand.post() is used.

Improved support for C++0x move construction to further reduce copying of handler objects. In certain designs it is possible to eliminate virtually all copies. Move support is now enabled when compiling in -std=c++0x mode on g++ 4.5 or higher.
Added build support for platforms that don't provide either of signal() or sigaction().
Changed to use C++0x variadic templates when they are available, rather than generating function overloads using the Boost.Preprocessor library.
Ensured the value of errno is preserved across the implementation's signal handler.
On Windows, ensured the count of outstanding work is decremented for abandoned operations (i.e. operations that are being cleaned up within the io_service destructor).
Fixed behaviour of zero-length reads and writes in the new SSL implementation.
Added support for building with OpenSSL 1.0 when OPENSSL_NO_SSL2 is defined.
Changed most examples to treat a failure by an accept operation as non-fatal.
Fixed an error in the tick_count_timer example by making the duration type signed. Previously, a wait on an already-passed deadline would not return for a very long time.

Added a new, completely rewritten SSL implementation. The new implementation compiles faster, shows substantially improved performance, and supports custom memory allocation and handler invocation. It includes new API features such as certificate verification callbacks and has improved error reporting. The new implementation is source-compatible with the old for most uses. However, if necessary, the old implementation may still be used by defining (BOOST_)ASIO_ENABLE_OLD_SSL.
Added new asio::buffer() overloads for std::array, when available. The support is automatically enabled when compiling in -std=c++0x mode on g++ 4.3 or higher, or when using MSVC 10. The support may be explicitly enabled by defining (BOOST_)ASIO_HAS_STD_ARRAY, or disabled by defining (BOOST_)ASIO_DISABLE_STD_ARRAY.
Changed to use the C++0x standard library templates array, shared_ptr, weak_ptr and atomic when they are available, rather than the Boost equivalents.
Support for std::error_code and std::system_error is no longer enabled by default for g++ 4.5, as that compiler's standard library does not implement std::system_error::what() correctly.

Added support for C++0x move construction and assignment to sockets, serial ports, POSIX descriptors and Windows handles.
Added support for the fork() system call. Programs that use fork() must call io_service.notify_fork() at the appropriate times. Two new examples have been added showing how to use this feature.
Cleaned up the handling of errors reported by the close() system call. In particular, assume that most operating systems won't have close() fail with EWOULDBLOCK, but if it does then set the blocking mode and restart the call. If any other error occurs, assume the descriptor is closed.
The kqueue flag EV_ONESHOT seems to cause problems on some versions of Mac OS X, with the io_service destructor getting stuck inside the close() system call. Changed the kqueue backend to use EV_CLEAR instead.
Changed exception reporting to include the function name in exception what() messages.
Fixed an insufficient initialisers warning with MinGW.
Changed the shutdown_service() member functions to be private.
Added archetypes for testing socket option functions.
Added a missing lock in signal_set_service::cancel().
Fixed a copy/paste error in SignalHandler example.
Added the inclusion of the signal header to signal_set_service.hpp so that constants like NSIG may be used.
Changed the signal_set_service implementation so that it doesn't assume that SIGRTMAX is a compile-time constant.
Changed the Boost.Asio examples so that they don't use Boost.Thread's convenience header. Use the header file that is specifically for the boost::thread class instead.

Added support for signal handling, using a new class called signal_set. Programs may add one or more signals to the set, and then perform an async_wait() operation. The specified handler will be called when one of the signals occurs. The same signal number may registered with multiple signal_set objects, however the signal number must be used only with Asio.
Added handler tracking, a new debugging aid. When enabled by defining (BOOST_)ASIO_ENABLE_HANDLER_TRACKING, Asio writes debugging output to the standard error stream. The output records asynchronous operations and the relationships between their handlers. It may be post-processed using the included handlerviz.pl tool to create a visual representation of the handlers (requires GraphViz).
Fixed a bug in asio::streambuf where the consume() function did not always update the internal buffer pointers correctly. The problem may occur when the asio::streambuf is filled with data using the standard C++ member functions such as sputn(). (Note: the problem does not manifest when the streambuf is populated by the Asio free functions read(), async_read(), read_until() or async_read_until().)
Fixed a bug on kqueue-based platforms, where reactor read operations that return false from their perform() function are not correctly re-registered with kqueue.
Support for std::error_code and std::system_error is no longer enabled by default for MSVC10, as that compiler's standard library does not implement std::system_error::what() correctly.
Modified the buffers_iterator<> and ip::basic_resolver_iterator classes so that the value_type typedefs are non-const byte types.

Added support for timeouts on socket iostreams, such as ip::tcp::iostream. A timeout is set by calling expires_at() or expires_from_now() to establish a deadline. Any socket operations which occur past the deadline will put the iostream into a bad state.
Added a new error() member function to socket iostreams, for retrieving the error code from the most recent system call.
Added a new basic_deadline_timer::cancel_one() function. This function lets you cancel a single waiting handler on a timer. Handlers are cancelled in FIFO order.
Added a new transfer_exactly() completion condition. This can be used to send or receive a specified number of bytes even if the total size of the buffer (or buffer sequence) is larger.
Added new free functions connect() and async_connect(). These operations try each endpoint in a list until the socket is successfully connected.
Extended the buffer_size() function so that it works for buffer sequences in addition to individual buffers.
Added a new buffer_copy() function that can be used to copy the raw bytes between individual buffers and buffer sequences.
Added new non-throwing overloads of read(), read_at(), write() and write_at() that do not require a completion condition.
Added friendlier compiler errors for when a completion handler does not meet the necessary type requirements. When C++0x is available (currently supported for g++ 4.5 or later, and MSVC 10), static_assert is also used to generate an informative error message. This checking may be disabled by defining (BOOST_)ASIO_DISABLE_HANDLER_TYPE_REQUIREMENTS.
Added support for using std::error_code and std::system_error, when available. The support is automatically enabled when compiling in -std=c++0x mode on g++ 4.5 or higher, or when using MSVC 10. The support may be explicitly enabled by defining ASIO_HAS_STD_SYSTEM_ERROR, or disabled by defining ASIO_DISABLE_STD_SYSTEM_ERROR. (Available in non-Boost version of Asio only.)
Made the is_loopback(), is_unspecified() and is_multicast() functions consistently available across the ip::address, ip::address_v4 and ip::address_v6 classes.
Added new non_blocking() functions for managing the non-blocking behaviour of a socket or descriptor. The io_control() commands named non_blocking_io are now deprecated in favour of these new functions.
Added new native_non_blocking() functions for managing the non-blocking mode of the underlying socket or descriptor. These functions are intended to allow the encapsulation of arbitrary non-blocking system calls as asynchronous operations, in a way that is transparent to the user of the socket object. The functions have no effect on the behaviour of the synchronous operations of the socket or descriptor.
Added the io_control() member function for socket acceptors.
For consistency with the C++0x standard library, deprecated the native_type typedefs in favour of native_handle_type, and the native() member functions in favour of native_handle().
Added a release() member function to posix descriptors. This function releases ownership of the underlying native descriptor to the caller.
Added support for sequenced packet sockets (SOCK_SEQPACKET).
Added a new io_service::stopped() function that can be used to determine whether the io_service has stopped (i.e. a reset() call is needed prior to any further calls to run(), run_one(), poll() or poll_one()).
Reduced the copying of handler function objects.
Added support for C++0x move construction to further reduce copying of handler objects. Move support is enabled when compiling in -std=c++0x mode on g++ 4.5 or higher, or when using MSVC10.
Removed the dependency on OS-provided macros for the well-known IPv4 and IPv6 addresses. This should eliminate the annoying "missing braces around initializer" warnings.
Reduced the size of ip::basic_endpoint<> objects (such as ip::tcp::endpoint and ip::udp::endpoint).
Changed the reactor backends to assume that any descriptors or sockets added using assign() may have been dup()-ed, and so require explicit deregistration from the reactor.
Changed the SSL error category to return error strings from the OpenSSL library.
Changed the separate compilation support such that, to use Asio's SSL capabilities, you should also include 'asio/ssl/impl/src.hpp` in one source file in your program.
Removed the deprecated member functions named io_service(). The get_io_service() member functions should be used instead.
Removed the deprecated typedefs resolver_query and resolver_iterator from the ip::tcp, ip::udp and ip::icmp classes.
Fixed a compile error on some versions of g++ due to anonymous enums.
Added an explicit cast to the FIONBIO constant to int to suppress a compiler warning on some platforms.
Fixed warnings reported by g++'s -Wshadow compiler option.

Fixed an integer overflow problem that occurs when ip::address_v4::broadcast() is used on 64-bit platforms.
Fixed a problem on older Linux kernels (where epoll is used without timerfd support) that prevents timely delivery of deadline_timer handlers, after the program has been running for some time.

Fixed a problem on kqueue-based platforms where a deadline_timer may never fire if the io_service is running in a background thread.
Fixed a const-correctness issue that prevented valid uses of has_service<> from compiling.
Fixed MinGW cross-compilation.
Removed dependency on deprecated Boost.System functions (Boost.Asio only).
Ensured close()/closesocket() failures are correctly propagated.
Added a check for errors returned by InitializeCriticalSectionAndSpinCount.
Added support for hardware flow control on QNX.
Always use pselect() on HP-UX, if it is available.
Ensured handler arguments are passed as lvalues.
Fixed Windows build when thread support is disabled.
Fixed a Windows-specific problem where deadline_timer objects with expiry times set more than 5 minutes in the future may never expire.
Fixed the resolver backend on BSD platforms so that an empty service name resolves to port number 0, as per the documentation.
Fixed read operations so that they do not accept buffer sequences of type const_buffers_1.
Redefined Protocol and id to avoid clashing with Objective-C++ keywords.
Fixed a vector reallocation performance issue that can occur when there are many active deadline_timer objects.
Fixed the kqueue backend so that it compiles on NetBSD.
Fixed the socket io_control() implementation on 64-bit Mac OS X and BSD platforms.
Fixed a Windows-specific problem where failures from accept() are incorrectly treated as successes.
Deprecated the separate compilation header asio/impl/src.cpp in favour of asio/impl/src.hpp.

Reduced compile times. (Note that some programs may need to add additional #includes, e.g. if the program uses boost::array but does not explicitly include <boost/array.hpp>.)
Reduced the size of generated code.
Refactored deadline_timer implementation to improve performance.
Improved multiprocessor scalability on Windows by using a dedicated hidden thread to wait for timers.
Improved performance of asio::streambuf with async_read() and async_read_until(). These read operations now use the existing capacity of the streambuf when reading, rather than limiting the read to 512 bytes.
Added optional separate compilation. To enable, include asio/impl/src.cpp in one source file in a program, then build the program with (BOOST_)ASIO_SEPARATE_COMPILATION defined in the project/compiler settings. Alternatively, (BOOST_)ASIO_DYN_LINK may be defined to build a separately-compiled Asio as part of a shared library.
Added new macro (BOOST_)ASIO_DISABLE_FENCED_BLOCK to permit the disabling of memory fences around completion handlers, even if thread support is enabled.
Reworked timeout examples to better illustrate typical use cases.
Ensured that handler arguments are passed as const types.
Fixed incorrect parameter order in null_buffers variant of async_send_to.
Ensured unsigned char is used with isdigit in getaddrinfo emulation.
Fixed handling of very small but non-zero timeouts.
Fixed crash that occurred when an empty buffer sequence was passed to a composed read or write operation.
Added missing operator+ overload in buffers_iterator.
Implemented cancellation of null_buffers operations on Windows.

Improved performance.
Reduced compile times.
Reduced the size of generated code.
Extended the guarantee that background threads don't call user code to all asynchronous operations.
Changed to use edge-triggered epoll on Linux.
Changed to use timerfd for dispatching timers on Linux, when available.
Changed to use one-shot notifications with kqueue on Mac OS X and BSD platforms.
Added a bitmask type ip::resolver_query_base::flags as per the TR2 proposal. This type prevents implicit conversion from int to flags, allowing the compiler to catch cases where users incorrectly pass a numeric port number as the service name.
Added #define NOMINMAX for all Windows compilers. Users can define (BOOST_)ASIO_NO_NOMINMAX to suppress this definition.
Fixed a bug where 0-byte asynchronous reads were incorrectly passing an error::eof result to the completion handler.
Changed the io_control() member functions to always call ioctl on the underlying descriptor when modifying blocking mode.
Changed the resolver implementation so that it no longer requires the typedefs InternetProtocol::resolver_query and InternetProtocol::resolver_iterator, as neither typedef is part of the documented InternetProtocol requirements. The corresponding typedefs in the ip::tcp, ip::udp and ip::icmp classes have been deprecated.
Fixed out-of-band handling for reactors not based on select().
Added new (BOOST_)ASIO_DISABLE_THREADS macro that allows Asio's threading support to be independently disabled.
Minor documentation improvements.

Added a new HTTP Server 4 example illustrating the use of stackless coroutines with Asio.
Changed handler allocation and invocation to use boost::addressof to get the address of handler objects, rather than applying operator& directly.
Restricted MSVC buffer debugging workaround to 2008, as it causes a crash with 2010 beta 2.
Fixed a problem with the lifetime of handler memory, where Windows needs the OVERLAPPED structure to be valid until both the initiating function call has returned and the completion packet has been delivered.
Don't block signals while performing system calls, but instead restart the calls if they are interrupted.
Documented the guarantee made by strand objects with respect to order of handler invocation.
Changed strands to use a pool of implementations, to make copying of strands cheaper.
Ensured that kqueue support is enabled for BSD platforms.
Added a boost_ prefix to the extern "C" thread entry point function.
In getaddrinfo emulation, only check the socket type (SOCK_STREAM or SOCK_DGRAM) if a service name has been specified. This should allow the emulation to work with raw sockets.
Added a workaround for some broken Windows firewalls that make a socket appear bound to 0.0.0.0 when it is in fact bound to 127.0.0.1.
Applied a fix for reported excessive CPU usage under Solaris.
Added some support for platforms that use older compilers such as g++ 2.95.

Added a new ping example to illustrate the use of ICMP sockets.
Changed the buffered*_stream<> templates to treat 0-byte reads and writes as no-ops, to comply with the documented type requirements for SyncReadStream, AsyncReadStream, SyncWriteStream and AsyncWriteStream.
Changed some instances of the throw keyword to boost::throw_exception() to allow Asio to be used when exception support is disabled. Note that the SSL wrappers still require exception support.
Made Asio compatible with the OpenSSL 1.0 beta.
Eliminated a redundant system call in the Solaris /dev/poll backend.
Fixed a bug in resizing of the bucket array in the internal hash maps.
Ensured correct propagation of the error code when a synchronous accept fails.
Ensured correct propagation of the error code when a synchronous read or write on a Windows HANDLE fails.
Fixed failures reported when _GLIBCXX_DEBUG is defined.
Fixed custom memory allocation support for timers.
Tidied up various warnings reported by g++.
Various documentation improvements, including more obvious hyperlinks to function overloads, header file information, examples for the handler type requirements, and adding enum values to the index.

Implement automatic resizing of the bucket array in the internal hash maps. This is to improve performance for very large numbers of asynchronous operations and also to reduce memory usage for very small numbers. A new macro (BOOST_)ASIO_HASH_MAP_BUCKETS may be used to tweak the sizes used for the bucket arrays. (N.B. this feature introduced a bug which was fixed in Asio 1.4.3 / Boost 1.40.)
Add performance optimisation for the Windows IOCP backend for when no timers are used.
Prevent locale settings from affecting formatting of TCP and UDP endpoints.
Fix a memory leak that occurred when an asynchronous SSL operation's completion handler threw an exception.
Fix the implementation of io_control() so that it adheres to the documented type requirements for IoControlCommand.
Fix incompatibility between Asio and ncurses.h.
On Windows, specifically handle the case when an overlapped ReadFile call fails with ERROR_MORE_DATA. This enables a hack where a windows::stream_handle can be used with a message-oriented named pipe.
Fix system call wrappers to always clear the error on success, as POSIX allows successful system calls to modify errno.
Don't include termios.h if (BOOST_)ASIO_DISABLE_SERIAL_PORT is defined.
Cleaned up some more MSVC level 4 warnings.
Various documentation fixes.

Improved compatibility with some Windows firewall software.
Ensured arguments to windows::overlapped_ptr::complete() are correctly passed to the completion handler.
Fixed a link problem and multicast failure on QNX.
Fixed a compile error in SSL support on MinGW / g++ 3.4.5.
Drop back to using a pipe for notification if eventfd is not available at runtime on Linux.
Various minor bug and documentation fixes.

Enhanced CompletionCondition concept with the signature size_t CompletionCondition(error_code ec, size_t total), where the return value indicates the maximum number of bytes to be transferred on the next read or write operation. (The old CompletionCondition signature is still supported for backwards compatibility).
New windows::overlapped_ptr class to allow arbitrary overlapped I/O functions (such as TransmitFile) to be used with Asio.
On recent versions of Linux, an eventfd descriptor is now used (rather than a pipe) to interrupt a blocked select/epoll reactor.
Added const overloads of lowest_layer().
Synchronous read, write, accept and connect operations are now thread safe (meaning that it is now permitted to perform concurrent synchronous operations on an individual socket, if supported by the OS).
Reactor-based io_service implementations now use lazy initialisation to reduce the memory usage of an io_service object used only as a message queue.

Added support for serial ports.
Added support for UNIX domain sockets.
Added support for raw sockets and ICMP.
Added wrappers for POSIX stream-oriented file descriptors (excluding regular files).
Added wrappers for Windows stream-oriented HANDLEs such as named pipes (requires HANDLEs that work with I/O completion ports).
Added wrappers for Windows random-access HANDLEs such as files (requires HANDLEs that work with I/O completion ports).
Added support for reactor-style operations (i.e. they report readiness but perform no I/O) using a new null_buffers type.
Added an iterator type for bytewise traversal of buffer sequences.
Added new read_until() and async_read_until() overloads that take a user-defined function object for locating message boundaries.
Added an experimental two-lock queue (enabled by defining (BOOST_)ASIO_ENABLE_TWO_LOCK_QUEUE) that may provide better io_service scalability across many processors.
Various fixes, performance improvements, and more complete coverage of the custom memory allocation support.