Nice, because it works for rank-deficient matrices.

And still fast enough for the systems I am considering here.

[[Imported from SVN: r9580]]