Use X86ISD::MOVSS and MOVSD to represent the movl mask pattern, also fix the handling of those nodes when seeking for scalars inside vector shuffles

llvm-svn: 112570