cpp_bin_float default value is NaN. But should be 0 according to the documentation.

[alex@…]

What should be: "Default constructed cpp_bin_floats have a value of zero." http://www.boost.org/doc/libs/1_61_0/libs/multiprecision/doc/html/boost_multiprecision/tut/floats/cpp_bin_float.html

This seems quite natural and convenient.

But in reality the default constructor seems to make NaN. And its code seems to confirm that as far as I can see:

cpp_bin_float() : m_data(), m_exponent(exponent_nan), m_sign(false) {}

As a result cpp_bin_float can't be used with boost::numeric::ublas::symmetric_matrix.

E.g. in boost\numeric\ublas\functional.hpp

in struct matrix_norm_inf

in real_type u = real_type ();

it is expected that the default value is 0.

[alex@…]

To be more precise the default constructor code is:

cpp_bin_float() BOOST_MP_NOEXCEPT_IF(noexcept(rep_type())) : m_data(), m_exponent(exponent_nan), m_sign(false) {}

After I have adjusted it in my local copy of boost:

cpp_bin_float() BOOST_MP_NOEXCEPT_IF(noexcept(rep_type())) : m_data(), m_exponent(exponent_zero), m_sign(false) {}

it seems to create zeroes. That is I've replaced exponent_nan with exponent_zero. I am just not sure if this adjustment is sufficient and consistent.

Thank you

[John Maddock]

Fixed in ​https://github.com/boostorg/multiprecision/commit/75dc2ebb331746e57b6f1ff313f3c6e3407658ab

Thanks!

[Michael Shatz]

It's a pity. I like a previous behavior. May be, it was contradiction docs, but it was great for debugging.