Re: [PATCH] msvcp120: implement _Cnd_* functions (try 2)

[Piotr Caban]

On 11/09/15 16:38, Daniel Lehman wrote:
> +static int (__cdecl*p__Mtx_init)(_Mtx_t*);

This function has different header.
> - if (0) /* causes crash in later_Cnd_* calls on Windows */
> -
>    p__Cnd_destroy(&cm.cnd);

This is not what I meant. I'm attaching a patch (generated on top of 
yours with the test I had in mind). It only shows the problem, the test 
needs more work.
> - for(i = 0; i < NUM_THREADS; i++)
> -
>    p__Thrd_create(&threads[i], cnd_bcast_thread, (void*)&cm);

...
> - p__Mtx_lock(&mtx);
> - cm.go = 1;
> - p__Mtx_unlock(&mtx);

Please add some kind of synchronization to make sure the thread is 
waiting on condition variable before setting cm.go to 1.
> - cm.go = 0;
> - p__Thrd_create(&threads[0], cnd_signal_thread, (void*)&cm);
> - p__Mtx_lock(&mtx);
> - while(!cm.go) {
> -
>    r = p__Cnd_wait(&cnd, &mtx);
> -
>    ok(!r, "failed to wait\n");
> - }

Here's a similar synchronization problem. In most cases p__Cnd_wait will 
not be called.
> - diff = p__Xtime_diff_to_millis2(&after, &before);
> - ok(r == 2, "should have timed out\n");
> - ok(diff > 2000 - TIMEDELTA, "got %d\n", diff);

This test failed once on my computer (in ~10 runs, diff == 1800).

Thanks,
Piotr