This will allow targets more flexibility to replace the
register allocator core passes. In a future commit,
AMDGPU will run the core register assignment passes
twice, and will also want to disallow using the
standard -regalloc option.
llvm-svn: 356506