craig.topper added a comment. It feels like we should be testing non-splat vectors here so we get two calls to the library functions in the output. Rather than one call and a reuse of the result. Repository: rL LLVM https://reviews.llvm.org/D47491