[cpp-threads] memory model

[Peter Dimov]

Peter Dimov wrote:

> Comments on the proposed memory model document:

It took me a while to realize this, but the definition of "canonical 
execution" is flawed.

The C++ abstract machine has no notion of reading from a non-volatile 
variable. That's why "side effects" only includes object modifications.

If you have:

    int a = 0;

the value of the expression 'a' is 0. There is no "read" involved in this 
evaluation; the value of 'a' is just zero, period. (The compiler is allowed 
to never touch the actual storage of 'a' if it can determine the value by 
other means.)

When a side effect modifies 'a' to be 1, the value of 'a' after this side 
effect is committed (after the first subsequent sequence point) is now 1.

So you have the flow:

...
a == 0
a == 0
<sequence point>
a == unspecified
side effect sets a to 1
a == unspecified
<sequence point>
a == 1
a == 1
...