[MLton] Optimizations done by GCC 4.0

[Henry Cejtin]

I agree that instruction branch prediction is probably too specialized to put
into MLton at this stage, but it seems like the huge gain possible in the
value-range propogation paper (I only read the abstract so far) is in the
inliner.  Note, the current scheme just does a simple leaf-to-root pass of
the call graph, which means that things that are only called conditionally
get inlined just as much as things called unconditionally.  This is the
reason that things like the buffer-re-load code in TextIO does get inlined
and so the input1 often does not.