On Nov 13, 2011, at 4:42 PM, Jed Brown wrote:

> On Sun, Nov 13, 2011 at 16:29, Barry Smith <bsmith at mcs.anl.gov> wrote:
>  I have PETSC_OPTIONS set in my .bashrc file to -malloc -malloc_debug 
> -malloc_dump by default.
> 
>  This means when I introduce memory leaks I find them immediately.  Like 
> someone just introduced in src/dm/examples/test/ex30.c but did not detect 
> because they did not have these options set.
> 
> This is my fault. In DMDestroy(), there is somewhat strange logic to count 
> the number of circular references in the "gotten" vectors.
> 
> DMSliced clones new Vecs off of it's own private Vec, which has yet another 
> circular reference. Should we just artificially decrement the reference count 
> on the DM, and then re-reference it in DMDestroy_Sliced before destroying 
> slice->globalvector? I wonder if there is a more elegant way to handle these 
> circular references.

  Why do it anyway different than DMDA?  Better consistent and not a great way 
then 10 different inconsistent and slightly different and better and worse ways.





Reply via email to