Really, what I meant was the metaprogramming associated with homoiconicity can't be done in a useful and referentially transparent way. That's because "eval"-like things are not referentially transparent.

However, there are other uses for homoiconicity aside from executing the code in an eval-like way. For instance, a text editor might use that property to inspect and transform the code being edited more easily. Also, I'm sure there are plenty of times it just makes things easier.

I think metaprogramming is also at odds with static analysis (e.g. a good type system), simply because the code itself is not static enough to analyze.