Unfortunately, CMUCL doesn't run on the machine
configuration I have been using for testing.
I did however, download CMUCL for Sun Solaris
and compare it to Allegro CL on two benchmarks.
1. hash-table based polynomial multiplication, with
all the coefficients being 1.
2. same test with coefficients being those in (1+x+y+z)^20.
For test 1, the difference was negligible.
For test 2, CMUCL was about 12 percent faster, I think.
The run time between Allegro CL and GCL is fairly similar,
but in some cases there are major differences in garbage
collection time.
As for running maxima on cmucl: Can't adopt it as the
sole foundation unless cmucl or sbcl runs on windows, also.
There are also some warnings associated with cmucl. Like
it may not recover from some problems (stack overflow?)
Writing a system this way can make it run much faster
until it falls off a cliff.
Actually, I suspect GCL has some of the same robustness
problems. At least it used to. By contrast, the
commercial systems (Allegro and the Lucid/Harlequin/Digitool...
lisps) are safer, in my experience.
RJF