Re: Questions

[James Youngman]

On 3/13/07, Leslie P. Polzer <address@hidden> wrote:
> I see.  But I believe we could stay with integers in normal operation
> (since we need only number_of_predicates numbers to reorder the tests)

This is true.

> and use floats only for performance measurement.

What's the problem with using floats?

> Or perhaps you could tell me more about the architecture of the
> optimizer you had in mind.  Should it be some adaptive system?

No, not yet.  At the moment, the optimiser is just the functions
opt_expr() and do_arm_swaps().  Before we change the decisions the
optimiser does, we need to (1) gather information about actual success
rates and (2) measure the performance changes.

Going on to adaptively reoptimise the operations needs to wait until
after the potential performance benefit (if any) has been measured.

James.