RSSA backend

[Matthew Fluet]

> Can't you just add a primitive which does the right thing (I guess the max
> of 1 and the size)?

The question isn't so much one of what to compute, as can you compute it
in assembly without requiring branches and new labels (i.e., labels that
didn't exist in RSSA).

BTW, that reminds me of something I discovered while looking at the
benchmarks.  mprof -d 1 can once again print out labels that do not appear
in SSA; instead those stubs for datatype variant extraction and multiple
blocks for branching on a datatype can appear.