Question on shared memory concurrency

[Andy Valencia via Digitalmars-d-learn]

I tried a shared memory parallel increment.  Yes, it's basically 
a cache line thrasher, but I wanted to see what's involved in 
shared memory programming.  Even though I tried to follow all the 
rules to make true shared memory (not thread local) it appears I 
failed, as the wait loop at the end only sees its own local 250 
million increments?

import core.atomic : atomicFetchAdd;
import std.stdio : writeln;
import std.concurrency : spawn;
import core.time : msecs;
import core.thread : Thread;
const uint NSWEPT = 1_000_000_000;
const uint NCPU = 4;

void
doadd(ref shared(uint) val)
{
    for (uint count = 0; count < NSWEPT/NCPU; ++count) {
        atomicFetchAdd(val, 1);
    }
}

void
main()
{
    shared(uint) val = 0;

    for (int x = 0; x < NCPU-1; ++x) {
        spawn(&doadd, val);
    }
    doadd(val);
    while (val != NSWEPT) {
        Thread.sleep(1.msecs);
    }
}