X86: Generalize the x << (y & const) optimization to also catch masks with more set bits set than 31 or 63.

llvm-svn: 148024