I think that running the test suite you are using on your machine and
system
is probably inadequate to demonstrate the improvement.
Running the simple f(10^6) test on my computer gives 46.50 seconds
without the hack,
and 43.58 with the hack, for a whopping 6% improvement.
I know of no disadvantage to the change in location of the
simplification program
on the property list.
I have difficulty explaining the (slight) slowdown on SBCL unless
perhaps there is higher use of some OTHER property (e.g. derivatives?)
or the
locality of the property list is disturbed for purposes of the test suite.
I wonder if it is even repeatable.
I think this hack belongs in the code base.
RJF