More precise GC

[Steven Schveighoffer]

On Sun, 28 Mar 2010 12:50:20 -0400, bearophile <bearophileHUGS at lycos.com>  
wrote:

> Robert Jacques:
>
>> What I think you're forgetting is that all compile-time type info is  
>> lost at runtime. [... etc]<
>
> Thank you very much for all your explanations, I didn't know that the  
> situation is so terrible. I suddenly like not-GC languages more :-)
> I think the compilation of D code must build a data structure that will  
> be used at runtime by the GC to know the type of all variables and  
> pointers, otherwise there's no hope in a GC that works well in  
> long-running programs.

The current GC has a simple "type info" if you will -- contains pointers  
or doesn't contain pointers.  It doesn't mean we cannot add to that.  In  
fact, I think dsimcha has provided a way to have precise scanning for  
heap-allocated types.  I don't think a reasonably precise GC is out of the  
question.  However, it may be too much to require the GC to do semantic  
analysis of enums for unions.  Not impossible, but probably not worth the  
effort and restrictions necessary.

-Steve