Because we use a lot of them to construct argument lists.
It would be possible to move to an pmt_vector based approach, which
would cut this down dramatically. I think we're still a bit early in
the game to start that kind of modification.
I think the first thing I would try is moving to the intrusive
implementation of the boost shared pointers for the pmt types.
Then I'd look at a data type specific alloc/free as well as see how
the default allocator is working across multiple threads. That is,
does it already use a separate allocation pool / thread. If it
doesn't, we could speed up the allocation/free and reduce the amount
of locking required in the typical case.
Before hacking away, I think we need to run the same test cases on
other machines besides the P4 Xeon and gather the oprofile data, as
well as the basic [ $ time <mytest> ] numbers. We may find wildy
different answers as f(microarchitecture). There's a reason intel
isn't featuring the P4 Xeon anymore ;)