From: Jens Maurer (Jens.Maurer_at_[hidden])
Date: 2002-03-11 15:33:47
Douglas Gregor wrote:
>
> This code in uniform_int:
> 120 if(_brange / _range > 4 /* quantization_cutoff */ ) {
> 121 // the new range is vastly smaller than the source range,
> 122 // so quantization effects are not relevant
> 123 return uniform_smallint<base_type,result_type>(_rng, _min, _max)();
> 124 } else {
>
> will make uniform_smallint do very bad things when _max-_min+1 is large
> enough to make 32*_range*_range (where _range = _max-_min+1 at line 95 of
> uniform_smallint.hpp) overflow and become negative. Then we get stuck in the
> following loop forever (uniform_smallint.hpp):
>
> 95 const base_result r = 32*_range*_range;
> 96 for(; r_base >= r; _factor *= 2)
> 97 r_base /= 2;
>
> I don't know enough about this to know if uniform_int should take care of
> this (and not deflect to uniform_smallint if it would overflow), or if
> uniform_smallint should deal with this overflow itself.
Indeed. uniform_smallint<> should cope with the input. I've fixed it in the
CVS.
Thanks for the report.
Jens Maurer