job/flag: improved parallelism test

test/job/flag.cpp

@@ -34,5 +34,40 @@ main ()
     tap.expect_eq (value, 2, "second wait didn't appear to block");
     t2.join ();
 
+    {
+        // perform a stress test to (hopefully) discover deadlocks
+        //
+        // * create a large matrix of flag variables
+        // * create a bank of threads which:
+        //     * wait on each flag of each row, or
+        //     * notify if the flag index matches the thread index
+        constexpr int iterations = 1024;
+        constexpr int parallelism = 16;
+
+        std::vector<
+            std::array<util::job::flag,parallelism>
+        > flags (iterations);
+
+        const auto func = [&flags] (const int idx) {
+            for (auto &row: flags) {
+                for (int i = 0; i < parallelism; ++i) {
+                    if (i == idx)
+                        row[i].notify ();
+                    else
+                        row[i].wait ();
+                }
+            }
+        };
+
+        std::vector<std::thread> workers;
+        for (int i = 0; i < parallelism; ++i)
+            workers.emplace_back (func, i);
+
+        for (auto &t: workers)
+            t.join ();
+
+        tap.expect (true, "flag sequence did not block");
+    }
+
     return tap.status ();
-}
+}