I would like to draw attention to my feature request (and corresponding pull request):
The aim is to improve the speed of morphologic operators by thresholding the exact euclidean distance transform of an image. In my (non-scientific) tests, the existing variant was 8,5x - 28x as slow. (I did not compare to decomposed footprints, though.)
The idea works for circular footprints, but could be extended to other shapes by using different distance metrics.