事务内存提议transactionalmemory-TransactionalLanguage.PDFVIP

Transactional Language Constructs for C++ Authors: Hans Boehm, HP, hans.boehm@ Justin Gottschlich, Intel, justin.e.gottschlich@ Victor Luchangco, Oracle, victor.luchangco@ Maged Michael, IBM, maged.michael@ Mark Moir, Oracle, mark.moir@ Clark Nelson, Intel, clark.nelson@ Torvald Riegel, Red Hat, triegel@ Tatiana Shpeisman, Intel, tatiana.shpeisman@ Michael Wong, IBM, michaelw@ Document number: N3341=12-0031 Date: 2012-01-11 Project: Programming Language C++, Evolution Working Group Reply-to: Michael Wong, IBM, michaelw@ Revision: 1 Introduction We propose that transactional language constructs be integrated into the C++ programming language, initially as a technical report (TR). We believe that this TR should be considered by the Concurrency Subgroup. It has impact to both the Core Language and the Library Working Group. This paper should be of interest to anyone who would like parallel programming in C++ to be easier, and to more easily support a modular programming paradigm. 1. The Problem Two types of synchronization supported by C++11 are mutexes, or locks, and atomics. Locks and atomics are basic abstractions used to access and mutate shared state, as defined by the Three Pillars of Concurrency [8]. From this view, these abstractions aim to avoid data races and synchronize objects in shared memory. Unfortunately, locks and atomics are notoriously difficult to use [9]. Simple coarse-grain locking strategies, where all program data are protected using one or few locks, lead to unnecessary serialization of program execution and loss of performance. Sophisticated fine-grain locking or use of atomics result in complex associat