MTWIST CHANGE LOG

    WARNING OF INCORRECT OUTPUT AND INCOMPATIBLE CHANGE: Versions of
    mtwist prior to 1.4 contained a bug that affects the random
    sequence generated if any of mt_llrand, mts_llrand, mt_ldrand, or
    mts_ldrand were used.  Versions prior to 1.4 generated sequences
    with an incorrectly short cycle (repeating approximately every 20
    billion random numbers).  They also had a probablity of saving
    incorrect state if mt_savestate or mts_savestate was called.

    Versions 1.4 and above may generate different sequences than prior
    versions if any of these functions are used.  If you are depending
    on reproducible PRNG sequences, you must decide whether to use the
    older versions for repeatability, or the newer ones for
    correctness.

VERSION 1.5, 23-Jan-2014

    - The second argument to gettimeofday is obsolete and, in modern
      libraries, must be given as NULL.  Since mtwist never used the
      second argument anyway, this correction has no run-time effect.

VERSION 1.4, 12-Jun-2013

    - INCOMPATIBLE CHANGE: correct a memory access error that resulted
      in incorrectly short cycles if mt_llrand or mt_ldrand was used.

VERSION 1.3, 11-Jun-2013

    - Try to fall back to /dev/random if /dev/urandom doesn't exist.
    - rd_empirical_setup now correctly declares its arguments as const.
    - Get rid of many compiler warnings.

VERSION 1.2, 30-Dec-2012

    - Make 32-bit autoseeders return the seeds used, for reproducibility.
    - Fix excess entropy usage in seeding functions.
    - Include unit tests in distribution.
    - Fix GCC warnings on Rcs_Id.
    - Fix GCC warnings on table of contents.
    - (Thanks to Markus Armbruster for all the above changes.)
    - INCOMPATIBLE CHANGE: correct a memory access error that resulted
      in incorrectly short cycles if mts_llrand or mts_ldrand was used.

VERSION 1.1, 11-Dec-2010

    - Fix inlining problems that prevented compilation under popular
      Windows compilers.
    - Rewrite empirical-distribution functions to use O(1) algorithm
      (incompatible change; old programs that used empirical
      distributions will need modification before they will compile).

VERSION 1.0, 24-Jun-2010

    - First production release.