Re: [rvm-research] Adding object headers in VM_MiscHeader causes a crash

[Ian R. <rog...@gm...>]

Rhodes Brown wrote:
> I too have encountered problems trying to add to VM_MiscHeader. I
> don't have all the answers, but here is what I have learned...
>
> I don't need all the info provided in the GCTrace configuration, nor
> do I want a GC trace. For my purposes, all I want is a simple unique
> ID for each object. I tried following the instructions in
> VM_MiscHeader and VM_MiscHeaderConstants, but adding a simple (int)
> counter to the header was causing bizarre crashes--even if I never
> actually read the ID! I assumed that I must have missed something, so
> I started poking through the files under org/jikesrvm/objectmodel to
> see if there was something else I needed to tweak. During my
> exploration I discovered the mystery that is
> VM_JavaHeader.getObjectFromStartAddress(). I am at a loss to explain
> why this method does what it does--it seems to be somewhat of a kludge
> that evolved from a change made in rev 10384. Apparently, it is
> looking for a word with the low bit on as a landmark from which to
> begin GC scanning. Since my extra field in the MiscHeader was often
> odd, this was causing the GC scan to begin at the wrong offset which
> would inevitably lead to a GC crash. So, I just set my counter to only
> use even values. Problem (apparently) solved.
>
> Obviously, this isn't a general solution. Someone with a better
> understanding of the GC system ought to look at this, possibly in
> conjunction with the revision to the object model (for 2.9.4) that Ian
> mentioned earlier. Personally, I would like to see something along the
> lines of what was done with VM_TIB, VM_CodeArray, etc. That is to say,
> a stronger object-like abstraction over instances (VM_ObjectInstance?
> An enhanced ObjectReference?) that is compiled away--something that
> can be tracked a little easier than direct memory accesses.
>
> Cheers,
>
> Rhodes Brown
> PhD Candidate in Computer Science
> University of Victoria
>
>   
Hi Rhodes,

I have independently had the same issue as you with the misc header. I 
have a patch that does the object model as you say  and adds support for 
a bidirectional object model. I will try to commit it at least to a 
branch by the end of the week.

Regards,
Ian