Friday, 04 July 2008
From a concurrency point of view, there are several papers of interest. Firstly, a few have been accepted into the working draft, notably:
- N2661: A Foundation to Sleep On
- This paper provides a generalised time
point and duration library, which is used by the thread functions that
take times or durations. These have been updated to use these new
types and renamed to make their purpose clearer: functions that wait
for a duration are now called
xxx_for, and take a value of type
std::chrono::duration<Rep,Period>, whereas those that take absolute time points are now called
xxx_untiland take a value of type
- N2668: Concurrency Modifications to Basic String
- Update: This paper has not actually been approved, I was
mistaken. Though the majority of the committee were in favour, there
was not consensus, so this paper will be discussed at a future
meeting. Thanks to Herb Sutter for
picking me up on this.
The changes in this paper ensure that it is safe for two threads to access the same
std::stringobject at the same time, provided they both perform only read operations. They also ensure that copying a string object and then modifying that copy is safe, even if another thread is accessing the original. This essentially disallows copy-on-write implementations since the benefits are now severely limited.
- N2660: Dynamic Initialization and Destruction with Concurrency
- With the changes from this paper, if an application uses multiple threads then the initialization and destruction of objects with static storage duration (such as global variables) may run concurrently on separate threads. This can provide faster start-up and shut-down times for an application, but it can also introduce the possibility of race conditions where none existed previously. If you use threads in your application, it is now even more important to check the initialization order of objects with static storage duration.
- N2514: Implicit Conversion Operators for Atomics
- With this change, the atomic types such as
std::atomic_intare implicitly convertible to their corresponding fundamental types. This means, for example, that:
std::atomic_int x; int y=x;is well-formed where it wasn't previously. The implicit conversions are equivalent to calling the
load()member function, and have
- N2674: Shared_ptr atomic access, revision 1
- This paper introduces a new set of overloads of the free functions
for atomic operations (such as
atomic_store), which operate on instances of
std::shared_ptr<>. This allows one thread to read an instance of
std::shared_ptrwhilst another thread is modifying that same instance if they both use the new atomic functions.
This paper also renames
atomic_exchange(and likewise for
atomic_compare_swapand the corresponding member functions) for all atomic types, in order to avoid confusion with other types that provide
swapfunctions. The atomic exchange operations only alter the value of a single object, replacing the old value with a new one, they do not exchange the values of two objects in the way that
- N2664: C++ Data-Dependency Ordering: Atomics and Memory Model
- With the adoption of this paper the memory model gets a new
memory_order_consume. This is a limited form of
memory_order_acquirewhich allows for data-dependent ordering. If a thread uses
memory_order_consume, then it is not guaranteed to see modifications to other variables made by the thread that performed the releasing operation unless those variables are accessed in conjunction with the consumed variable. This means, for example, that member variables of an object are visible if the consumed value is a pointer to that object, but that values of independent objects are not necessarily visible. This allows the compiler to perform some optimizations that are forbidden by
memory_order_acquire, and reduces the synchronization overhead on some hardware architectures.
- N2678: Error Handling Specification for Chapter 30 (Threads)
- This paper brings the exceptions thrown by the thread under the
system_errorumbrella, with corresponding error codes and error categories.
- N2669: Thread-Safety in the Standard Library (Rev 2)
- Now the standard supports threads, we need to say which standard library operations are thread-safe, and which are not. This paper basically says that non-modifying operations on the same object are safe, and any operations on separate objects are also safe. Also, separate threads may call the same library functions on separate objects without problems. As you might expect, concurrent modifications to the same object are data races and undefined behaviour.
The committee also voted to include N2659:
Thread-Local Storage in C++0x, but it doesn't appear to be in the
current draft. This paper introduces the
keyword to indicate that each thread should have its own copy of a
Finally, N2657: Local and Unnamed Types as Template Arguments has been incorporated in the working paper. Though this isn't directly concurrency related, it is something I've been campaigning for since N1427 back in 2003.
Design and Content Copyright © 2005-2015 Just Software Solutions Ltd. All rights reserved.