One for experts in std.parallelism

[Stanislav Blinov via Digitalmars-d]

On Monday, 12 June 2017 at 14:21:29 UTC, Andrei Alexandrescu 
wrote:
> On 06/12/2017 12:13 AM, Stanislav Blinov wrote:
>> On Sunday, 11 June 2017 at 19:06:48 UTC, Andrei Alexandrescu 
>> wrote:
>>> I tried to eliminate the static shared ~this as follows:
>>>
>>> https://github.com/dlang/phobos/pull/5470/commits/a4b2323f035b663727349d390719509d0e3247ba
>>>
>>>
>>> However, unittesting fails at src/core/thread.d(2042). I 
>>> suspect it's because the atexit call occurs too late, after 
>>> the shared static this in src/core/thread.d has already been 
>>> called.
>>>
>>> Thoughts on how to make this work?
>>>
>>>
>>> Thanks,
>>>
>>> Andrei
>> 
>> Which atexit call? The shared static ~this is still present in 
>> that commit.
>
> Eh, the link doesn't work anymore; I hoped it would be 
> immutable. Anyhow, I just tried this again:
>
> private final class ParallelismThread : Thread
> {
>     this(void delegate() dg)
>     {
>         super(dg);
>         static shared bool once;
>         import std.concurrency;
>         initOnce!once({
>             import core.stdc.stdlib;
>             atexit(&doThisAtExit);
>             return true;
>         }());
>     }
>
>     TaskPool pool;
> }
>
> // Kill daemon threads.
> //shared static ~this()
> extern(C) void doThisAtExit()
> {
>     foreach (ref thread; Thread)
>     {
>         auto pthread = cast(ParallelismThread) thread;
>         if (pthread is null) continue;
>         auto pool = pthread.pool;
>         if (!pool.isDaemon) continue;
>         pool.stop();
>         pthread.join();
>     }
> }
>
> There's no more shared static ~this(). Instead, we have an 
> atexit registration. It fails in src/core/thread.d(2042).
>
>
> Andrei

To me, it feels like the options are:
1. Take a performance hit and unregister "parallel" threads from 
the runtime at creation, and keep a separate lock-protected list.
2. Modify runtime to allow creating unregistered Threads
3. Push thread_term() call back if at all possible (i.e. by also 
registering it with atexit).

I'll put the PR for (1) up for discussion a bit later. (2) and 
(3) I'll need to dig into runtime somewhat first.