Use movaps instead of movups to spill 16-byte vector values when default alignment is >= 16. This fixes some massive performance regressions.

llvm-svn: 53844