[CodeGen][X86] Don't scalarize vector saturating add/sub

Now that we have vector support for [US](ADD|SUB)O we no longer
need to scalarize when expanding [US](ADD|SUB)SAT.

This matches what the cost model already does.

Differential Revision: https://reviews.llvm.org/D57348

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@353651 91177308-0d34-0410-b5e6-96231b3b80d8
5 files changed