Fix detection of ccmath usage in trunc

[mborland]

As reported by @robertramey on the ML

[mborland]

Errored on clone in drone. Merging.

[jzmaddock]

@mborland sorry to be late to this, I don't think this quite gets it - for one thing we have the same logic in round.hpp which also needs fixing, but the problem is more subtle - we have modern clang sitting on top of an ancient (pre-C++17) std library, so from the compilers perspective, everything that we're currently checking here is supported (or potentially so) where as the std library's <type_traits> has no C++17 _v suffixed constants. Other than checking for a random C++17 header as a proxy in addition to the existing compiler checks I'm not sure what to suggest. In the long run we could yet another macro to Boost.Config I guess... ah wait, I think this is __cpp_lib_bool_constant >= 201505L? The latter requires inclusion of either type_traits or version, so the logic had probably better somewhere central.

I'll see if I can put together a patch shortly, might be too late for this release though as I can see quite a few headers getting touched.

[mborland]

I thought clang didn't get a version of __builtin_constant_evalutated until 9? If the type_traits had compiled the ccmath ldexp would have errored on initializing a constexpr variable with a non-constexpr expression since it uses the STL at runtime. Unless someone is running say Clang-10 with libstdc++-5 shouldn't properly checking for BOOST_MATH_NO_CONSTEXPR_DETECTION be sufficient? I think there's another week until RC release so you should have time.

[jzmaddock]

This is what I had in mind: #1045

And yes, from a practical point of view, you may be correct that no one would use clang-10 on top of gcc-5, though I'm constantly surprised by what what people do do!

[robertramey]

FYI. It's not just mere "people". It's Boost Developers. Its from the test matrix which is our thing.