boost::math::cstdfloat::detail::float_internal128_t

[gpeterhoff]

Hello,
in boost::math::cstdfloat there are several inconsistencies with __float128/_Quad. Generally these are only available with gcc and intel. Since I don't have an intel compiler myself, you might have to modify this path; for now it should behave like the previous implementations.

types <boost/math/cstdfloat/cstdfloat_types.hpp>
intel https://community.intel.com/t5/Intel-C-Compiler/C-Compiler-128-bit-floating-point/m-p/1216949

• WINDOWS: float_internal128_t = _quad
• OTHER: float_internal128_t = __float128

defines/constants <boost/math/cstdfloat/cstdfloat_types.hpp>
gcc

• the constants from <quadmath.h> or builtins can be used
• signaling NaN is supported

functions <boost/math/cstdfloat/cstdfloat_cmath.hpp>
modified:

• gcc: nanq does not support payload https://github.com/gcc-mirror/gcc/blob/master/libquadmath/math/nanq.c
  --> #define BOOST_CSTDFLOAT_FLOAT128_NAN __builtin_nanq
• hypot(x, y, z) correct result even with large values; slower than previous naive implementation
• isfinite(x) branch-free

limits <boost/math/cstdfloat/cstdfloat_limits.hpp>
All functions can now be provided constexpr noexcept.

future

• branch-free isgreater, isgreaterequal, ... I don't understand the BOOST_PREVENT_MACRO_SUBSTITUTION inside the functions, otherwise I could have adjusted that right away.
• With Matt's boost::charconv the IO functions can be provided more efficiently (operator<<, operator>>, to_string, ...).
• It might be useful to provide quadmath as a separate sublibrary in boost::math, e.g. to avoid the #define/#undef orgies in <boost/math/cstdfloat/cstdfloat_cmath.hpp>.

cstdfloat_types.hpp.txt
cstdfloat_cmath.hpp.txt
cstdfloat_limits.hpp.txt

thx
Gero

[gpeterhoff]

spelling error
BOOST_CSTDFLOAT_FLOAT128_QUIET_NAN (instead of BOOST_CSTDFLOAT_FLOAT128_QUIT_NAN)