Jesse Natalie authored Jul 30, 2020 and Erik Faye-Lund committed Oct 26, 2020

This series supports native 16bit casts in LLVM (e.g. f2f16) for
when 16bit types are supported by the target hardware. Note that
only the rtz variant is supported by DXIL, and rtne (and default,
which CL requires to be rtne) is lowered to a custom cast.

It also supports lowering the specific cast patterns we see used
by the vload/vstore_half implementations into the special DXIL
intrinsic for casting between fp32 and fp16, using an i32 value
to hold the fp16 data in its lower 16 bits.