and recompile math/cmplx.Inf, you get similar generated code. The constant folding that did not occur during generic opt occurs in lowered opt. (The generated code isn't identical, but I'm working on that independently.)
However, if the constant folding occurs during lowered opt, there is no inlining nop at the beginning of the function.
Either we need one, or we don't; it shouldn't be sensitive to whether a particular generic constant folding optimization occurs.