c++ bug in uniform real distribution?

Question

I'm trying to use std::uniform_real_distribution<float>(a, b) to generate random floats, and I found a case where the output is equal to the upper limit b. According to: http://www.cplusplus.com/reference/random/uniform_real_distribution http://en.cppreference.com/w/cpp/numeric/random/uniform_real_distribution this shouldn't happen.

With both gcc-4.9.2 and clang-3.5.0, the following crashes for me:

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

int main() {
    float a = 1.0f;
    float b = 1.001f;
    size_t seed = 293846;
    size_t n = 9830;
    std::mt19937 rg(seed);
    std::uniform_real_distribution< float > u(a, b);
    for (size_t i = 0; i < n; ++i) {
        float v = u(rg);
        if (not (v < b)) {
            std::cerr << "error: i=" << i
                      << " v=" << std::scientific << std::setprecision(std::numeric_limits< double >::max_digits10) << v
                      << " b=" << b << std::endl;
            abort();
        }
    }
}

The output I see is:

error: i=9829 v=1.00100004673004150e+00 b=1.00100004673004150e+00
Aborted (core dumped)

Is this a bug with the Standard Library? Am I missing something?

Edit. While not an exact duplicate, the issue raised in this question is indeed covered by the answer to the other question referenced here.

Most informative about the status of this issue is the LLVM bug thread here: https://llvm.org/bugs/show_bug.cgi?id=18767

Not every day I see a C programmer using `not` instead of `!` ;) — Lapshin Dmitry, Nov 04 '15 at 22:01
@Lapshin: I'm of split opinion on the practice. On the one hand the keywords logicals are nice, but on the other hand a decade of not using them makes them seem out of place. — , Nov 04 '15 at 22:03
@Hurkyl i am not against, actually like here it's more readable, but still. — Lapshin Dmitry, Nov 04 '15 at 22:04
:) how about spaces inside template parameters, a relic of `>>` parsing :) — Matei David, Nov 04 '15 at 22:05
@MateiDavid I am actually too young to got used to that ugly spacing, praise c++11) — Lapshin Dmitry, Nov 04 '15 at 22:07
This is another manifestation of the `generate_canonical` problem. — T.C., Nov 04 '15 at 22:08

score 0 · Answer 1 · answered Nov 04 '15 at 22:07

0

While [rand.dist.uni.real] mandates the uniform distribution on the half-open interval of real numbers, it doesn't contain any language regarding how those should be rounded to floating point representations.

So from the information I could find, it may very well be legal (and required) for a conforming implementation to produce floating point numbers that compare equal to the upper bound.

answered Nov 04 '15 at 22:07

2

I like your answer, but still, standart says `a <= x < b`, and that part keeps me from agreing with you... – Lapshin Dmitry Nov 04 '15 at 22:10

c++ bug in uniform real distribution?

1 Answers1