This allows to avoid scratch use or indirect VGPR addressing for small vectors. Differential Revision: https://reviews.llvm.org/D54606 llvm-svn: 347231