Code behaves differently if compiled with MSVC vs GCC

[olologin]

VS 19 vs GCC 12 If you build and run the following sample with both compilers:

#include <fenv.h>
#include <iostream>
#pragma warning(push)
#pragma warning(disable : 4702)
#include <boost/numeric/interval.hpp>
#pragma warning(pop)

/// An interval of doubles [lower,upper)
using Interval = boost::numeric::interval<double,boost::numeric::interval_lib::policies<boost::numeric::interval_lib::rounded_math<double>,boost::numeric::interval_lib::checking_strict<double>>>;

/// An interval from -DBL_MAX to +DBL_MAX
const inline Interval InfiniteInterval =
	Interval(std::numeric_limits<double>::lowest(), std::numeric_limits<double>::max());

int main(void)
{
#ifdef _WIN32
	unsigned int current_word = 0;
	_controlfp_s(&current_word, 0, _MCW_EM);
#else
	feenableexcept(FE_ALL_EXCEPT);
#endif
	auto i = InfiniteInterval;
	boost::numeric::width(i);
	std::cerr << "The end" << std::endl;
}

You will see how it fails with FP exception on linux, but never fails on windows. The reason is simple: On linux we adjust only rounding mode https://github.com/boostorg/interval/blob/2eda7413ac16dd4158005446438daf8a7e435dd9/include/boost/numeric/interval/detail/c99sub_rounding_control.hpp#L27 But on windows in addition to rounding we also mask all the FP exceptions (turn off corresponding flags) https://github.com/boostorg/interval/blob/2eda7413ac16dd4158005446438daf8a7e435dd9/include/boost/numeric/interval/detail/msvc_rounding_control.hpp#L93 because we use _MCW_EM | _MCW_RC instead of _MCW_RC only.

Shouldn't both systems set the same FP settings? If yes - what should they adjust? Both systems should adjust rounding only, or both should adjust rounding and exception masking? If we only want to change rounding control bits - it is enough to remove _MCW_EM | part of that line and both systems will throw FP exception.

[olologin]

Maybe I can help with PR if you decide on what should be the preferred behavior.