Re: [Sbcl-devel] 1.0.35.9: Add support for non-trivial random seeds

[Tobias C. R. <tc...@fr...>]

> 1.0.35.9: Add support for non-trivial random seeds
> 
> SBCL is using the popular MT19937 PRNG algorithm, but up until now,
> was only seeding the initial random state with a 32-bit seed, and
> choosing a seed subject to a lot of collisions (a second-precise timer)
> when called with (MAKE-RANDOM-STATE T).
> 
> This patch adds and documents an SBCL extension to MAKE-RANDOM-STATE
> that supports initializing a random-state based on an arbitrary UNSIGNED-BYTE
> or a (SIMPLE-ARRAY (UNSIGNED-BYTE 8)). Also supported (but documented as
> not officially so) is a (SIMPLE-ARRAY (UNSIGNED-BYTE 32)). Last but not least,
> (MAKE-RANDOM-STATE T) will try to initialize the random state by reading
> 256 bits from /dev/urandom, which should eliminate the collision problem and
> make SBCL's PRNG suitable for more applications than before.
> 
> Finally, we use in our random-state initialization routines the very same
> algorithms that the author of MT19937 recommends in the latest version
> of his C source, and we have tested the output to be identical (see
> November 2009 discussion in the sbcl-devel mailing-list).

It seems like this commit caused failings of

  MAKE-RANDOM-STATE.ERROR.4

and

  PRINT.RANDOM-STATE.1

from the ansi test suite.

  -T.