This optimisation actually make things a bit slower on average, as
measured by nofib. The example in #1136 in particular suffers from high
memory usage. Therefore we no longer do the optimisation.