Speed of Kalman filter in unstable vs 4.4.3

Joris reports at http://www.dynare.org/phpBB3/viewtopic.php?f=1&t=6373&start=15#p37426 that 4.4.3 runs Bayesian estimation a lot faster than the current master. Profiling fs2000.mod, the culprit seems to be https://github.com/DynareTeam/dynare/pull/1088/commits/05fc096569e15e89d8d13b08799321c0313b168d where we made the Kalman filter more robust against ill-conditioned matrices. This seems to have considerable computational costs, because the function is called so often. I wonder if switching to the univariate filter if problems appear is not the better default.