A llvm.vector.reduce.fadd(float, <1 x float>) will be translated to G_VECREDUCE_SEQ_FADD with two scalar operands, which is illegal according to the verifier. This makes sure we generate a fadd/fmul instead.