The reason behind this change is that the DirectMath, XM* stuff, supports SIMD instructions which allow you to add 2 vector4 together in one operation. SIMD liberaries are constructed in such a way that they keep all of the data in XMM registers on the CPU. This means that the operations like +, -, * and / are implemented as standalone functions that return a XMVector and takes two XMVector arguments.

https://msdn.microsoft.com/en-us/library/windows/desktop/microsoft.directx_sdk.arithmetic.xmvectoradd%28v=vs.85%29.aspx

As I remember there are a lot of functions starting with XM* to manipulate XMFLOAT* (and XMVECTOR*, XMMATRIX*). Pay attention at the use of the Load/Store functions, I remember some headaches because of not using them properly.

Take this while waiting for a more expert answer

EDIT: Ninja

Also watch out for alignment on the XM* types.

XMVECTOR and XMMATRIX etc must be 16-byte aligned. When you place the value into the stack, this is handled for you by the compiler, but if you allocate a struct or class containing an XMVECTOR or XMMATRIX then the alignment isn't done for you, and any attempt to access a mis-aligned XM* struct will cause an access violation and break.

You should probably use XMFLOAT4 etc to store the actual values in your classes and structures, and only load the XMMATRIX or XMVECTOR when neccessary within a methor or function.