This allows formation of OpcodeSwitch for top level patterns, in
particular on X86.  This saves about 1K of data space in the x86
table and makes the dispatch much more efficient.

llvm-svn: 97440