[PATCH v2] stdlib: Simplify arc4random_uniform

[Adhemerval Zanella Netto]

On 02/08/22 09:08, Yann Droneaud wrote:
> Hi,
> 
> Le 01/08/2022 à 21:20, Noah Goldstein a écrit :
>>> diff --git a/stdlib/arc4random_uniform.c b/stdlib/arc4random_uniform.c
>>> index 1326dfa593..5aa98d1c13 100644
>>> --- a/stdlib/arc4random_uniform.c
>>> +++ b/stdlib/arc4random_uniform.c
>>>
>>>   uint32_t
>>>   __arc4random_uniform (uint32_t n)
>>>   {
>>> @@ -57,83 +38,33 @@ __arc4random_uniform (uint32_t n)
>>> +  while (1)
>>>       {
>>> +      uint32_t value = __arc4random ();
>>> +
>>> +      /* Return if the lower power of 2 minus 1 satisfy the condition.  */
>>> +      uint32_t r = value & mask;
>>> +      if (r < n)
>>> +       return r;
>>> +
>>> +      /* Otherwise check if remaining bits of entropy provides fits in the
>>> +        bound.  */
>>> +      for (int bits_left = z; bits_left >= bits; bits_left -= bits)
>>> +       {
>>> +         value >>= bits;
>> Can this just be shift by 1 and repeat (32 - z) times or does that
>> introduce bias (not seeing exactly why it would)?
> 
> 
> That was bothering me too, as I was thinking a rotation would be possible instead of shift.
> 
> I posted the question https://crypto.stackexchange.com/questions/101325/uniform-rejection-sampling-by-shifting-or-rotating-bits-from-csprng-output-safe
> 
> The answer: there's indeed a bias.
> 
> This explains why my attempt with rotation leads to dieharder complaining. It was so obvious ... Damn
> 


Thanks, I will remove it then.  We might evaluate later if using the mask
and compare is indeed better than the other methods (as using by OpenBSD).