RSSA backend

[Matthew Fluet]

> > > Why am I plagued with inexplicable regression errors?  
> 
> Ah, I just figured it out.

O.k.  I checked in a fix.  The checked in sources all pass regressions (on
the one hand, I think the regression script is broken with respect to the
-exn-history true flag, but on the other hand, it did reveal that bug in
the flattener), compile all benchmarks, and successfully self-compiles.

Self-compile times look good:
MLton finished in 284.19 + 123.57 (30% GC)
380.16user 27.72system 8:04.05elapsed 84%CPU (0avgtext+0avgdata
0maxresident)k 0inputs+0outputs (57200major+1992433minor)pagefaults 0swaps

(The low %CPU usage is due to lots of processor intensive programs running
on the linux cluster.)

I'm currently running benchmarks against the 2001-10-06 release.  IIRC,
that release has the broken limit-check coalescing, so we're likely to see
some slowdowns.