More interesting numbers: 

STOCK:      64,483,627 bytecodes/sec.  Geepers, I had been reading this
number wrong, with the commas in place I can see that I was actually
geting 64 MIPS performance out of an interpriter!!! That's damn fast. It
probably beats my 486(5x86)/150 too, the machine I am writing this on. 

CUSTOM:     34,801,522 bytecodes/sec. (-53%)
CUSTOM -O0: 52,330,335 (-19%) -- BEST GCC 3.1.1 result yet!

STOCK:     2,044,447 sends/sec.
CUSTOM:    1,587,861 sends/sec. (-23%) (REVISED)
Custom -O0 1,432,081 (-30%)
 
STOCK:      39.270 seconds.
CUSTOM:     41.052 seconds (+4% (worse))
CUSTOM -O0: 50.598. (+23% worse).

It seems that the register allocation hypothesis holds. The -O0 setting
produced a 1,072,720 byte executable that is within striking distance of
the Gcc 2.95? numbers. 

-- 
Linux has more source code than my brain.
http://users.rcn.com/alangrimes/