Apparently the spec has overloads for fmin/fmax and ldexp with one of the operands as scalar. We need to broadcast the scalars to the vector type. https://reviews.llvm.org/D158077