> And how is the situation different from DSP or other unusual architectures with
> 8-bit bytes? The effort is the same to support <simd>: massive, per
> architecture, and provided by those who are interested.

There's a much greater chance that you can piggy-back off of existing
implementations. If your architecture has 8-bit bytes, you can write
an LLVM backend. I don't think LLVM supports non-8-bit bytes right
now, or will do so in the near future. With LLVM, <simd> can use
clang's vector_size attribute, and you get the frontend of <simd> for
free.

The story would be similar for maintaining a GCC backend compared to
maintaining ... everything yourself.

If you don't have 8-bit bytes, you operate outside of the widespread
software stack into which thousands, if not millions of work hours
have been sunk; you start from nothing.