C ++ 11是否允许（不要求）发布/获取volatile关键字的语义

Question

Since Visual C++ 2005, Microsoft has made additional ordering guarantees for accesses to volatile types which are not required by the C++ Standard.

Does anything in the C++ Standard actually forbid these guarantees? The Microsoft documentation seems to think so.

Please let me know whether the Standard allows the ordering implemented by Microsoft, and also vote on this bug report:

• /volatile:ms documentation mangles what "ISO compliant" means

Answer 1

An implementation is certainly allowed to do things beyond on what is required as long as it meets the requirements set forth by the standard. Adding release/acquire semantics to volatile object is definitely within scope. I don't think there is interest in the C++ committee to change the semantics (we are just starting a new week of discussing C++ in Portland with Herb right now talking about how to organize the meeting).

Answer 2

The standard requires that access to volatile objects is evaluated strictly according to the rules of the abstract machine , which means roughly "don't optimize", but no more than that. For example, the compiler is not allowed to cache a value in a register or perform common subexpression elimination. It has to do exactly what you tell it.

Thus, for all the standard cares, volatile has memory_order_relaxed semantics (it doesn't specify anything different). This doesn't mean you are not allowed to implement something more strict, of course.

The Microsoft compiler has always ( since 2005 as Ben Voigt points out) treated volatile as acquire/release, which lead to many people assuming "volatile == threadsafe", which in return led to many articles turning this around into "volatile is useless!" and "volatile is evil".

The likely reason why MS recommends to use the ISO implementation is that this makes their compiler behave like every other compiler, no more nasty surprises.