[cpp-threads] Prism 0.9.1, and further on the effects of races

[Peter Dimov]

Boehm, Hans wrote in message c++std-ext-8018:

> My reaction:
>
> - The interaction of vtables and unordered atomics that Herb pointed
> out bothers me.  We now have the potential of races on objects that
> are introduced by the implementation and not visible to the
> programmer.

The vtable is not an object in the C++ sense (1.8).

In:

// thread 1

px = new X;

// thread 2

px->f();

there is a race on the object *px of type X, regardless of whether this 
object is polymorphic (may have a vtable.)

The lifetime of *px, when X has a non-trivial constructor, starts when its 
constructor completes (3.8/1).

Accessing the object *px before its lifetime has started is undefined 
behavior (3.8/5).

If the store to px doesn't have release semantics, thread 2 is not 
guaranteed to observe the effects of the constructor, and hence, its 
behavior is undefined.

When X has virtual functions, its default constructor is not trivial 
(12.1/5). However, this is not the only way in which a type can have a 
nontrivial default constructor.

In short, I believe that the vtable is a distraction; the problem is not 
limited to types with vtables.

Even if you "fix" atomic<class_type_with_nontrivial_default*>, this doesn't 
help with

// thread 1

px = new X( 1, 2 );

// thread 2

px->f();

since it uses a non-trivial constructor.