Concurrency ISO C++ library
Design
Interface to Locks and Mutexes The file <ext/concurrence.h> contains all the higher-level constructs for playing with threads. In contrast to the atomics layer, the concurrence layer consists largely of types. All types are defined within namespace __gnu_cxx. These types can be used in a portable manner, regardless of the specific environment. They are carefully designed to provide optimum efficiency and speed, abstracting out underlying thread calls and accesses when compiling for single-threaded situations (even on hosts that support multiple threads.) The enumerated type _Lock_policy details the set of available locking policies: _S_single, _S_mutex, and _S_atomic. _S_single Indicates single-threaded code that does not need locking. _S_mutex Indicates multi-threaded code using thread-layer abstractions. _S_atomic Indicates multi-threaded code using atomic operations. The compile-time constant __default_lock_policy is set to one of the three values above, depending on characteristics of the host environment and the current compilation flags. Two more datatypes make up the rest of the interface: __mutex, and __scoped_lock. The scoped lock idiom is well-discussed within the C++ community. This version takes a __mutex reference, and locks it during construction of __scoped_locke and unlocks it during destruction. This is an efficient way of locking critical sections, while retaining exception-safety.
Interface to Atomic Functions Two functions and one type form the base of atomic support. The type _Atomic_word is a signed integral type supporting atomic operations. The two functions functions are: _Atomic_word __exchange_and_add_dispatch(volatile _Atomic_word*, int); void __atomic_add_dispatch(volatile _Atomic_word*, int); Both of these functions are declared in the header file <ext/atomicity.h>, and are in namespace __gnu_cxx. __exchange_and_add_dispatch Adds the second argument's value to the first argument. Returns the old value. __atomic_add_dispatch Adds the second argument's value to the first argument. Has no return value. These functions forward to one of several specialized helper functions, depending on the circumstances. For instance, __exchange_and_add_dispatch Calls through to either of: __exchange_and_add Multi-thread version. Inlined if compiler-generated builtin atomics can be used, otherwise resolved at link time to a non-builtin code sequence. __exchange_and_add_single Single threaded version. Inlined. However, only __exchange_and_add_dispatch and __atomic_add_dispatch should be used. These functions can be used in a portable manner, regardless of the specific environment. They are carefully designed to provide optimum efficiency and speed, abstracting out atomic accesses when they are not required (even on hosts that support compiler intrinsics for atomic operations.) In addition, there are two macros _GLIBCXX_READ_MEM_BARRIER _GLIBCXX_WRITE_MEM_BARRIER Which expand to the appropriate write and read barrier required by the host hardware and operating system.
Implementation
Using Builtin Atomic Functions The functions for atomic operations described above are either implemented via compiler intrinsics (if the underlying host is capable) or by library fallbacks. Compiler intrinsics (builtins) are always preferred. However, as the compiler builtins for atomics are not universally implemented, using them directly is problematic, and can result in undefined function calls. (An example of an undefined symbol from the use of __sync_fetch_and_add on an unsupported host is a missing reference to __sync_fetch_and_add_4.) In addition, on some hosts the compiler intrinsics are enabled conditionally, via the -march command line flag. This makes usage vary depending on the target hardware and the flags used during compile. If builtins are possible for bool-sized integral types, _GLIBCXX_ATOMIC_BUILTINS_1 will be defined. If builtins are possible for int-sized integral types, _GLIBCXX_ATOMIC_BUILTINS_4 will be defined. For the following hosts, intrinsics are enabled by default. alpha ia64 powerpc s390 For others, some form of -march may work. On non-ancient x86 hardware, -march=native usually does the trick. For hosts without compiler intrinsics, but with capable hardware, hand-crafted assembly is selected. This is the case for the following hosts: cris hppa i386 i486 m48k mips sparc And for the rest, a simulated atomic lock via pthreads. Detailed information about compiler intrinsics for atomic operations can be found in the GCC documentation. More details on the library fallbacks from the porting section.
Thread Abstraction A thin layer above IEEE 1003.1 (i.e. pthreads) is used to abstract the thread interface for GCC. This layer is called "gthread," and is comprised of one header file that wraps the host's default thread layer with a POSIX-like interface. The file <gthr-default.h> points to the deduced wrapper for the current host. In libstdc++ implementation files, <bits/gthr.h> is used to select the proper gthreads file. Within libstdc++ sources, all calls to underlying thread functionality use this layer. More detail as to the specific interface can be found in the source documentation. By design, the gthread layer is interoperable with the types, functions, and usage found in the usual <pthread.h> file, including pthread_t, pthread_once_t, pthread_create, etc.
Use Typical usage of the last two constructs is demonstrated as follows: #include <ext/concurrence.h> namespace { __gnu_cxx::__mutex safe_base_mutex; } // anonymous namespace namespace other { void foo() { __gnu_cxx::__scoped_lock sentry(safe_base_mutex); for (int i = 0; i < max; ++i) { _Safe_iterator_base* __old = __iter; __iter = __iter-<_M_next; __old-<_M_detach_single(); } } In this sample code, an anonymous namespace is used to keep the __mutex private to the compilation unit, and __scoped_lock is used to guard access to the critical section within the for loop, locking the mutex on creation and freeing the mutex as control moves out of this block. Several exception classes are used to keep track of concurrence-related errors. These classes are: __concurrence_lock_error, __concurrence_unlock_error, __concurrence_wait_error, and __concurrence_broadcast_error.