[X86][SSE] Avoid vector extraction/insertion for non-constant uniform shifts

As discussed on D51263, we're better off using byte shifts to clear the upper bits on pre-SSE41 hardware.

llvm-svn: 340810