RAND_MAX broken

[Bruce Evans]

On Wed, 3 Jul 2013, Andrey Chernov wrote:

> On 02.07.2013 20:33, Bruce Evans wrote:
>> I checked the values returned by rand().  The ACM part works as
>> intended, so it never returns RAND_MAX.  It also never returns 0.  So
>> the distribution of values in the documented range [0, RAND_MAX] is
>> very non-uniform.  It is uniform in [1, RAND_MAX - 1].  To use this
>> algorithm for rand(), 1 should have been subtracted, giving a range
>> of [0, 0x7ffffffe].
>
> Do you mean [0, 0x7ffffffd] (assuming 1 subtracted)?

Yes.

> See attached patch.
> I don't see compatibility problems at least from POSIX specs point of
> view - they don't say something specific about RAND_MAX.

I like the patch.  It is binary compatibility problems and maybe applications
assuming that (RAND_MAX + 1) is 0x80000000 so as to do magic bit operations
with it that may be problems.

Bruce