Re: [Sbcl-devel] Re: [Sbcl-commits] CVS: sbcl/src/pcl braid.lisp,1.25,1.26 low.lisp,1.23,1.24

[Raymond T. <to...@rt...>]

>>>>> "William" == William Harold Newman <wil...@ai...> writes:

    William> this idea in terms of existing SBCL code would be
    William>    (defun get-instance-hash-code ()
    William>      (mix 1234567
    William>           (mod (incf hash-code delta)
    William>                (1+ most-positive-fixnum))))
    William> And if we had some stronger hash function than MIX (imagine using MD5
    William> or something of the sort, though presumably neither so insanely
    William> thorough nor so computationally expensive), that'd work better. But
    William> for now I thought the bang-for-the-buck of just linear congruential
    William> distribution was sufficient.

Why is this linear congruential?  I always thought it would be
something like

        (mod (* hash-code a) (1+ most-positive-fixnum))

for some suitably selected a, except this conses quite a bit.

A reasonable alternative might be the Weyl generator used in the
:new-random generator (deprecated, now):

        (mod (- x a) 1.0)

where x and a are floating-point numbers.  I think this can all be
done without consing, and the conversions between fixnum and floats is
fast too.

Ray