This addresses a regression where pretty much all C++ compilations using

-frounding-math now fail, due to rounding being performed in constexpr

function definitions in the standard library.

This follows the "manifestly constant evaluated" approach described in

https://reviews.llvm.org/D87528#2270676 -- evaluations that are required

to succeed at compile time are permitted even in regions with dynamic

rounding modes, as are (unfortunately) the evaluation of the

initializers of local variables of const integral types.