[vtkusers] Bug with Shallow Copy and Deconstructors

[Brad Lowekamp]

The way that this bug has manifested itself to me is by the field data
disappearing during filters operating on it. I have encountered a similar
bug with the vtkDataObjectToDataSetFilter and posted about it. I have also
encountered similar problems with several other classes. If two data
objects share an object lets say from a shallow copy, and one of the data
objects is deleted the data is deleted in the other. I have frequently
encountered this with the field data. I have spent a significant amount of
time tracking down and understanding all the related code. This bug can be
easily seen in the following tcl code, when the source contains field
data.

vtkPolyDataReader reader
reader SetFileName test.vtk

vtkPolyData pd
pd ShallowCopy [reader GetOutput]

puts [[pd GetFieldData] GetNumberOffArrays]

reader Delete

puts [[reader GetFieldData] GetNumberOfArrays]

Have other people encountered their data disappearing too? I have searched
that archive and not found much. I can't image that I am the only one. To
me this type of bug seems very significant and that other people must have
encountered it before. Or do few people delete their classes before they
are finished?

The project that I have been working on involves using the fielddata in
vtkDataObject and dynamically creating and deleting vtk objects. As a
result I have had to create an abstract filter vtkDataObjectToDataObject
to do a lot of transformation. And I have needed to keep careful track of
Objects, making sure they get deleted and filters remain correct when the
input and outputs are changed. In working with this I have had much
frustration working with VTK. Because it seems as though neither of these
aspects of VTK have been tested or worked with much before. It is
infrequent when there is a task that I wish to get done where the filters
are neither designed to work with vtkDataObject nor handle the FieldData
in a vtkDataSet correctly.

For example I was working with the vtkSourcePort and vtkInputPort and
trying to transmit FieldData with it. It does not accept vtkDataObject as
a class to transmit. No problem, I will just add a dumpy point and send a
vtkPolyData across the network. But wait the field data mysteriously
disappears! After a day of searching, I have come to discover that the
vtkDataSetReader uses SHALLOWCOPY. Which can be traced to the source of
the disappearing data! I dont see how vtkDataSetReader can read FieldData!
(Try using vtkDataSetReader instead of vtkPolyData in the above example,
interesting!)

I have seen this bug occour in many other filters as well. ANY filter
which uses a ShallowCopy will likely have this same bug too! Perhaps
people have not encountered this bug more because few people use the
FieldData and still fewer realize what is sometime goes on is a bug. It is
wide spread and shows itself is many different places, but there us just a
few places that need modification to fix it and those are with the
interaction with the deconstructors and initialization function of
DataObjects and related classes.

There is a simple rule that should be adopted by deconstruction that would
work and solve this problem so that shallow copies can be used with out
fear of causing bugs! The rule is: For every vtkObject an objects owns
(has a registered pointer to) Unregister or Delete it, and then set the
variable to NULL, but DO NOT CALL THAT VTKOBJECTS INITIALIZATION function.


To get my project to work, I must fix this bug myself in vtkDataObject. I
don't know how people submit code, but certainly I think that this is a
good candidate for that. However because these fixes are so close to the
core and could affect most of the other classes I am sure there many
issues. I certainly hope that this issue well be addressed. And I am
willing to give my help to getting it fixed.