的std :: uniform_int_distribution <int> g ++和msvc的范围

Question

While doing a minor school project, I just noticed a difference in the range of std::uniform_int_distribution<int> :

When using g++ the range is [a, b], however when using msvc (Visual Studio 2010) the range is (a, b], so the output of the following program:

#include <iostream>
#include <random>

using std::cout;
using std::cin;

int main()
{
    std::mt19937 random;
    std::uniform_int_distribution<int> intDist(-1, 1);

    for(int i = 0; i < 100; i++)
    {
        cout << intDist(random) << "\n";
    }

    cin.get();
}

Will display -1 at some point when using g++, but it will never display -1 when using msvc.

I know is common that such differences exists between compilers, but booth the MSDN documentation and the standard mark that the range should be [a, b].

What is the reason for this behavior?

Answer 1

It's a bug in the Visual Studio 2010 implementation of std::uniform_int_distribution .

I reported it here in December 2011 and got confirmation from Stephan T. Lavavej that:

1. It is indeed a bug.
2. It's been fixed in the next version of Visual Studio (including the current Developer Preview ).

The bug occurs when the first parameter to the std::uniform_int_distribution constructor is negative. To work around it in your example, construct the distribution like this:

std::uniform_int_distribution<int> intDist(0, 2);

And then call it like this:

cout << intDist(random) - 1 << "\n";  

Answer 2

What is the reason for this behavior?

If true, it's a bug in MSVC implementation.

Answer 3

MSDN says :

The first member operator uses the engine eng as a source of uniformly distributed integral values and returns integral values with each value i in the closed range [min(), max()] occurring with equal probability and values outside that range occurring with probability 0.

This means that your observation is wrong, or there is abug in the implementation. This should be easy to see in the source.

I also suggest seeding the rng to a different value. There is a very very very low possibility of the behaviour turning out better.