RSSA backend

[Stephen Weeks]

> > Another possibility is to simply add POINTER_SIZE to the frontier
> > uncoditionally.  Then you don't need extra labels.
> 
> Doesn't that screw up walking the from heap in the GC?  Or do we just make
> GC_foreachPointerInObject understand that arrays always have an extra
> POINTER_SIZE of dead bytes at the end of a non-zero length array?

Something like that.  There's a test I see in
GC_foreachPointerInObject (line 453) that can be eliminated as well.

> I guess that is fairly simple, although it means we're wasting a
> little bit of space on arrays.

That is the reason the decision was made to include the test.  I lean
slightly towards saving the space and paying a little time.

Another way to fix the problem, while getting the space saving, is to
require that calls to Array_array have numElts > 0.  We can use
another primitive, Array_array0, for creating zero length arrays.
Then the test is expressed in the SML and RSSA.  The only problem is
for arrays that have 0 bytesPerElt, which isn't known until RSSA --
but those can be turned into Array_array0 at that time.