I.e., pass and return vectors in registers. With the default C calling
convention (and the Win64 TargetABI not touching any vectors), vectors
are returned in XMM0, but inefficiently passed as ref to hidden copy
(automatically by LLVM).
Microsoft's vector calling convention, introduced opt-in with VS 2013,
additionally puts HFAs and HVAs (Homogeneous Float/Vector Aggregates)
into registers, see
https://blogs.msdn.microsoft.com/vcblog/2013/07/11/introducing-vector-calling-convention/.
