On Wed, May 15, 2013 at 11:40:25AM +0100, Garth N. Wells wrote:
> On 15 May 2013 11:23, Anders Logg <[email protected]> wrote:
> > On Wed, May 15, 2013 at 10:48:31AM +0100, Garth N. Wells wrote:
> >> Chris and I are working on IO (including parallel) via HDF5 for mesh
> >> data structures. An issue that I've run into before and which is
> >> making things complicated now are the circular dependences in the Mesh
> >> class.
> >>
> >> Every Mesh has a MeshDomains object and a MeshData object. MeshDomains
> >> stores a reference to its Mesh, and it stores MeshFunctions, that hold
> >> a shared_ptr to a Mesh (it also has MeshValueCollections which we'll
> >> probably also let store a pointer to the Mesh).
> >
> > We should strive to avoid this (the owned object having a pointer to
> > the owner), but it often creeps in for convenience.
> >
> >> This all makes memory management complicated because the objects are
> >> constructed with the Mesh '*this' object, and leads to a lot of 'no
> >> deleter' shared pointers. Since MeshDomains are always associates with
> >> a Mesh and a Mesh always has a MeshDomains object, any objections to
> >> just using plain STL vectors to hold the MeshDomain data?
> >
> > What would the changes be in the interface?
> >
>
> There will be some internal changes, but I expect very little or
> nothing that the user will see. The only visible changes will be in
> the MeshDomains interface, which I've never used directly from
> application-level code.

Sounds good then.

--
Anders
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics

Reply via email to