Why doesn't GHC force a GC before program quits?

[Paul Liu]

Hi Simon,

Thanks for the explanation. So is this why there is a separation of
mkWeak# and mkWeakForeignEnv#, the latter being used for C finalizers
that is guaranteed to run before program quits?

I wonder if it's possible to put some constraint to the type of
finalizers in mkWeak# so as to guarantee that it won't do nasty
things. Or maybe even get rid of the finalizer argument in mkWeak#,
since they are not guaranteed to run anyway.

Regards,
Paul Liu

On Tue, Apr 10, 2012 at 5:54 AM, Simon Marlow <marlowsd at gmail.com> wrote:
> On 03/04/2012 21:40, Paul Liu wrote:
>>
>> It seems to violate the claim that "all finalizers are run exactly
>> once" mentioned in the weak pointer paper. Is there any reason not to
>> enforce this?
>
>
> GHC does do a GC before the program quits, but it doesn't wait for all the
> finalizers to finish.  It turns out to be difficult to do that: the runtime
> doesn't know the difference between a finalizer thread and an ordinary
> thread, and we don't normally wait for ordinary threads to finish before
> terminating the program.  What about blocked finalizers? What if the
> finalizers create more finalizers?  It's all a bit of a swamp.  So now we
> just don't guarantee that (Haskell) finalizers run before program exit; if
> you want guaranteed execution then you can use C finalizers.
>
> Some of the motivation for this change comes from Hans Boehm's seminal paper
> on finalizers:
>
> http://www.hpl.hp.com/techreports/2002/HPL-2002-335.html
>
> Cheers,
>        Simon



-- 
Regards,
Paul Liu