On x86: (shl V, 1) -> add V,V Hardware support for vector-shift is sparse and in many cases we scalarize the result. Additionally, on sandybridge padd is faster than shl. llvm-svn: 143311
Nadav Rotem