struct destructors

[Steven Schveighoffer]

On Mon, 09 May 2011 17:04:49 -0400, Sean Kelly <sean at invisibleduck.org>  
wrote:

> On May 9, 2011, at 12:51 PM, Alexander wrote:
>
>> On 09.05.2011 19:38, Sean Kelly wrote:
>>
>>> Not currently.  I thought I wrote some explanation of why...
>>
>>  OK, thanks - I've read the ticket. Though, the problem can be solved  
>> relatively simple, IMHO - adding one more pointer to BlkInfo, where  
>> address of the finalizer is stored, which will be used only for cases  
>> when collected object is not a class.
>
> Well... BlkInfo is generated from a selection of internal sources.   
> You're right that the GC would need to store a finalizer reference for  
> structs, but that's just the simple case.  It would also need to  
> finalize structs in dynamic arrays and AA nodes as well (though it's  
> possible the latter already works, since the AA is now template code).   
> And of course the finalizer reference would have to be passed to  
> GC.malloc() or whatever so that portion of the API would change as  
> well.  I originally held off on fixing this because its requirements  
> seemed to fit nicely with the requirements for precise scanning (ie.  
> passing the GC some TypeInfo so it can be smarter about how it deals  
> with allocated memory blocks in general), and I wanted to do both at  
> once.  But I have a feeling that I came up with some other problem  
> regarding struct finalization, so I want to see if I can find that old  
> email or whatever to be sure.

A few things here:

1. the reason why a struct is not finalized when allocated on the heap is  
because a struct does not have a vtable pointer (which is how the GC is  
able to find the dtor for a class).
2. A struct allocated on the heap is actually transformed by the compiler  
into an allocation of an array of one struct instance.  So essentially  
there is one problem to solve -- how to finalize an array of structs.
3. AA's are templated, but the templated functions simply forward to the  
runtime functions.  The template allows us to add extra functionality to  
an AA without modifying the compiler, but everything is very much done by  
druntime.  However, I think we would have the ability to add this without  
modifying the compiler.

-Steve