The follwing numbers seem to second that:
FRAMES_PER_CHUNK time real user 4 1.2s 0.9 8 1.6 16 2.3 1.4
These are runs of parrot -j -Oc examples/benchmarks/fib.imc on an unoptimized built parrot on Athlon 800.
The main problem seems to be, that we are copying around a lot of memory: 16 chunks * 32 regs * 4/8 bytes
More fields for optimizations: - regstack_copy_chunk alwyas copies a full chunk (despite of used) - memory is allocated zeroed - the COW flags is only reset on one user of 2 provocing even more unneeded ocpies.
Comments welcome, leo