Charles Oliver Nutter wrote:
- say "nyah" to all other implementations that claimed they'd be a lot
faster because of selective variable boxing
It appears that boxing does help one aspect of local variables:
assignment. In our case, assignment of a local variable do a heap-based
scope stuffs it in an array, which requires Java to do boundschecking.
But with selective boxing, we avoid the boundchecking cost completely.
Fully heap-based:
10M * 100 local var assignments and retrievals
8.492000 0.000000 8.492000 ( 8.492000)
8.129000 0.000000 8.129000 ( 8.129000)
7.151000 0.000000 7.151000 ( 7.152000)
7.146000 0.000000 7.146000 ( 7.146000)
Boxed:
10M * 100 local var assignments and retrievals
1.598000 0.000000 1.598000 ( 1.598000)
1.783000 0.000000 1.783000 ( 1.783000)
1.500000 0.000000 1.500000 ( 1.501000)
1.530000 0.000000 1.530000 ( 1.530000)
So not quite the "nyah" I wanted. This also points to a potential
performance issue with our DynamicScope using an array; the cost of
stuffing variables into an array is pretty high. Perhaps for low numbers
of vars we could try custom DynamicScope impls that use fields? Work
exploring.
- Charlie
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
