c) Speedup vtkEventBroker::RemoveObservations():

• Don't look up SubjectMap[subjectKey] twice (SubjectMap can be as big as 10^4)
• push_back can be expensive when creating big vectors. Consider using "reserve"
• maybe doing "erase" would be faster than creating a new array each time.

d) Speedup vtkEventBroker::GetObservations():

• Don't lookup SubjectMap[subjectKey] for each loop iteration.
• construct observationList at the beginning of the function to avoid a copy in optimization mode.

e) Speedup vtkEventBroker::GetObservationsForSubjectByTag():

• Unnecessary vector copies are happening (e.g. subjectList).

f) Speedup ctkVTKConnection:

• Converting from const char to QString can be expensive, this is the case in ctkVTKObjectEventsObserverPrivate::findConnection() when the conversion happens many times. Moreover, the Qt signal/slot signature uses const char, so it should also probably be const char*.

g) Speedup vtkMRMLScene::RemoveReferencedNodeID

• By reserving the space, push_backs will be faster

h) Review vtkMRMLModelDisplayableManager::UpdateFromMRML()

• It probably doesn't need to remove the props each time a display[able] node is added: (called from vtkMRMLModelDisplayableManager::OnMRMLSceneNodeAdded in RequestRender)

i) Improve performance of vtkPiecewiseFunction::RemovePoint:
Just search the array once (use std::distance() to find if the point is found)
Speedup vtkPiecewiseFunction::DeepCopy:

• Fire only 1 modified event. Currently, each time a point is added, the range is changed and calls Modified().
• As a temporary solution, one can block Modified Event invocation on vtkMRMLVolumePropertyNode when the piecewise function of vtkVolumeProperty is modified: Call StartModify on vspNode->GetVolumePropertyNode() in vtkSlicerVolumeRenderingLogic::CopyLabelMapDisplayToVolumeRenderingDisplayNode and in vtkSlicerVolumeRenderingLogic::CopyScalarDisplayToVolumeRenderingDisplayNode
• Temporary solution implemented in r21278

a) fixed in r21580
c), d) and e) fixed in r21581
f) fixed in r21582
g) fixed in r21583

j) vtkEventBroker should have maximum 1 observation (AnyEvent) on a vtkObject. It should then redirect the events to observers.

k) vtkMRMLSliceLogic::UpdateSliceNode could be speed up by not calling vtkMRMLScene::GetNthNodeByClass. This method should be reviewed to improve performance.

i) Exposing vtkMRMLSliceNode::GetRASToXY() (have the matrix cached maybe) would prevent many matrix inversions in the code (e.g. in the displayable managers)

j) speed-up slice view zoom by wrapping SetFieldOfView and UpdateMatrices with Start/EndModify in vtkSliceViewInteractorStyle. This actually does not reduce the number of time observers are triggered because vtkMRMLSliceLogic::OnMRMLNodeModified() might actually trigger a new ModifiedEvent inside SetSliceExtentsToSliceNode() -> SetUVWExtentsAndDimensions() -> UpdateMatrices().

A second issue might comes from vtkSetAndObserveMRMLNodeEventsMacro in vtkMRMLAbstractLogic. If it is called twice, the 2nd time the old node value is not "unobserved". This can lead to errors or duplicated calls if it is done twice in a row.

k) Add vtkMRMLDisplayableNode::SetAndObserveDisplayNode()
convenient method that prevents a "search" in the mrml scene node collection

@Julien: Re-targetting minor issue from 4.4 to 4.5