MSVC supports float, double and long double but long double is the same as double.

However with classic FPU instructions, x86 provides a native support for a more precise type than the usual double. Some implementations may use that as underlying implementation of long double. MSVC doesn't, but maybe Intel Compiler or Clang or CUDA has a switch for that support.

Does such a switch exist? Do we need to support that?

The question is for #3928 , where I assume that long double is the same as double. The vector algorithm may return a totally wrong pointer for array of long doubles of different implementation.