log() function breaks for big floats

[JanHildenbrandBosch]

It seems like that the cmath.inl::get_fast_converge_params algorithm is using an integer intermediate (size_t) representation, where not all floats can be represented with.

    template<typename Float>
    static constexpr auto get_fast_converge_params(Float* const A, std::size_t* const n, Float x)
        -> void
    {
        auto _x = static_cast<std::size_t>(trunc(x));
        auto _n = static_cast<std::size_t>(logi_helper<10>(_x));
        *n = _n;
        *A = x / pow_helper(10, _n);
    }

The following code will lead to compile-time errors (compiled with gcc 10.3.0):
static_assert(smath::log2(1.0e8) < 27, "");

/workspace/test/static_math(master*) # make
Consolidate compiler generated dependencies of target test-cmath
[ 6%] Building CXX object CMakeFiles/test-cmath.dir/test/cmath.cpp.o
/workspace/test/static_math/test/cmath.cpp: In function ‘int main()’:
/workspace/test/static_math/test/cmath.cpp:132:38: error: non-constant condition for static assertion
132 | static_assert(smath::log2(1.0e8) < 27, "");
| ~~~~~~~~~~~~~~~~~~~^~~~
/workspace/test/static_math/test/cmath.cpp:132:30: in ‘constexpr’ expansion of ‘smath::log2(1.0e+8)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:664:38: in ‘constexpr’ expansion of ‘smath::detail::log2_helper(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:361:27: in ‘constexpr’ expansion of ‘smath::detail::log2_helper((std::is_floating_point(), std::integral_constant<bool, true>()), x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:354:35: in ‘constexpr’ expansion of ‘smath::detail::logf_helper(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:318:49: in ‘constexpr’ expansion of ‘smath::detail::fast_converge_log<0>::compute(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:266:37: in ‘constexpr’ expansion of ‘smath::detail::get_fast_converge_params((& a), (& n), x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:251:28: in ‘constexpr’ expansion of ‘smath::detail::pow_helper<int, long unsigned int>(10, _n)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:54:45: in ‘constexpr’ expansion of ‘smath::detail::pow_helper<int, long unsigned int>((x * x), (exponent / 2))’
/workspace/test/static_math/include/static_math/detail/cmath.inl:54:45: in ‘constexpr’ expansion of ‘smath::detail::pow_helper<int, long unsigned int>((x * x), (exponent / 2))’
/workspace/test/static_math/include/static_math/detail/cmath.inl:54:45: in ‘constexpr’ expansion of ‘smath::detail::pow_helper<int, long unsigned int>((x * x), (exponent / 2))’
/workspace/test/static_math/test/cmath.cpp:132:38: error: overflow in constant expression [-fpermissive]
make[2]: *** [CMakeFiles/test-cmath.dir/build.make:76: CMakeFiles/test-cmath.dir/test/cmath.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:396: CMakeFiles/test-cmath.dir/all] Error 2
make: *** [Makefile:146: all] Error 2

Changing this to

    template<typename Float> static constexpr auto get_fast_converge_params(Float* const A, std::size_t* const n, Float x)
        -> void
    {
        auto _x = static_cast<std::size_t>(trunc(x));
        auto _n = static_cast<std::size_t>(logi_helper<10>(_x));
        *n = _n;
        *A = x / pow_helper<int long long>(10, _n);
    }

will fix this issue for the particular case, but for values > std::numeric_limits<size_t> this will break.

static_assert(smath::log10(2.0e20) > 0, ""); leads to:

/workspace/test/static_math(master*) # make
[ 6%] Building CXX object CMakeFiles/test-cmath.dir/test/cmath.cpp.o
/workspace/test/static_math/test/cmath.cpp: In function ‘int main()’:
/workspace/test/static_math/test/cmath.cpp:133:40: error: non-constant condition for static assertion
133 | static_assert(smath::log10(2.0e20) > 0, "");
| ~~~~~~~~~~~~~~~~~~~~~^~~
/workspace/test/static_math/test/cmath.cpp:133:31: in ‘constexpr’ expansion of ‘smath::log10(2.0e+20)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:525:39: in ‘constexpr’ expansion of ‘smath::detail::log10_helper(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:278:24: in ‘constexpr’ expansion of ‘smath::detail::log10_helper((std::is_floating_point(), std::integral_constant<bool, true>()), x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:273:31: in ‘constexpr’ expansion of ‘smath::detail::logf_helper(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:262:45: in ‘constexpr’ expansion of ‘smath::detail::fast_converge_log<0>::compute(x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:219:33: in ‘constexpr’ expansion of ‘smath::detail::get_fast_converge_params((& a), (& n), x)’
/workspace/test/static_math/include/static_math/detail/cmath.inl:204:45: in ‘constexpr’ expansion of ‘smath::trunc(x)’
/workspace/test/static_math/test/cmath.cpp:133:40: error: overflow in constant expression [-fpermissive]
/workspace/test/static_math/test/cmath.cpp:133:40: error: overflow in constant expression [-fpermissive]
/workspace/test/static_math/test/cmath.cpp:133:40: error: overflow in constant expression [-fpermissive]
make[2]: *** [CMakeFiles/test-cmath.dir/build.make:76: CMakeFiles/test-cmath.dir/test/cmath.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:396: CMakeFiles/test-cmath.dir/all] Error 2
make: *** [Makefile:146: all] Error 2

[Morwenn]

As mentioned in #27 I don't maintain this project anymore: this kinds of errors keep piling up, I don't have the necessary skills to fix them, and whenever I dare to look back at the library I keep finding new issues. Sorry about that.

[JanHildenbrandBosch]

As mentioned in #27 I don't maintain this project anymore: this kinds of errors keep piling up, I don't have the necessary skills to fix them, and whenever I dare to look back at the library I keep finding new issues. Sorry about that.

No need to apologize, I fully understand your point. I was evaluating using this library to bridge the time until it's part of the standard (hopefully c++23 http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p0533r8.pdf).

From just inspecting the github page it seems everything works as intended and is tested (no open bugs/issue concerning functionality), and is referenced by some stack overflow posts. Hopefully, this issue will make it easier for others to evaluate, whether to use this library or not.

I'd encourage you to post your already know issues at least in #27, perhaps then a potential contributor/maintainer has a starting point to investigated/get started. Thanks for you work!