I cannot understand random generated value: why infinity?

[Enoizat]

Example code:

#include <chrono>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <limits>
#include <random>


// tune the iterations number according to your pc performances: -------------//
using mytype = int;                     // int, long, long long...        <---//
constexpr mytype Maxval {                                                     //
    std::numeric_limits<mytype>::max() / 100                //            <---//
};


int main()
{
    std::mt19937 gen {
        static_cast<unsigned>(
            std::chrono::high_resolution_clock::now().time_since_epoch().count()
        )
    };
    std::uniform_real_distribution<> dst (
        std::numeric_limits<double>::lowest(),
        std::numeric_limits<double>::max()
    );

    for(mytype m {}; m < Maxval; ++m) {
        auto d { dst(gen) };
        if ( d != std::numeric_limits<double>::infinity() ) {
            std::cout << "Generated number: " << std::fixed << d << '\n';
        }
    }
    std::cout << '\n';
}

What I get (after a fairly long time) is... nothing!
It seems, with those parameters, the std::uniform_real_distribution only generates 'inf', while I thought it would have generated a variety of doubles (13.666, -12548795364.5248, ...).
What’s wrong in my code?

[Peter87]

https://en.cppreference.com/w/cpp/numeric/random/uniform_real_distribution/uniform_real_distribution

Requires that a ≤ b and b-a ≤ std::numeric_limits<RealType>::max()

[Enoizat]

Wow, thank you, Peter87, that’s very tricky.
So, since std::numeric_limits<double>::lowest() is negative, max() - lowest() turns out to be an incredibly large positive number...

I admit that I had read that page, but in my mind I was just doing the ‘simple’ (=wrong) math and calculating max() - lowest() as 0!
Thanks again.

May I ask your opinion about this? Doesn’t it sound like a little bug?
Shouldn’t it be better if it were b - |a| < max()? Or would it be there another downside?

[Peter87]

No, I don't think it's a bug, but I agree that it's "tricky" and easy to make a mistake.

b-a is the size of the range. I don't know how uniform_real_distribution is normally implemented but I'm guessing that b-a > max() makes it harder to implement, or at least in an efficient manner, because you wouldn't be able to calculate b-a without getting infinity.

[Enoizat]

Yes, Peter87, you are right, I just didn't want to admit I have proven so dumb :-)
Please, let me live this thread down... ;-)