One could reasonably use llvm for the backend (assuming <cough> they had an actual reference grammar for the language</cough>) and then you'd have the same prob

One could reasonably use llvm for the backend (assuming <cough> they had an actual reference grammar for the language</cough>) and then you'd have the same problem, the output would just be some unreadable intermediate language that is far from transparent to the average user. The manual doesn't really go into enough detail to imply you'd need to fiddle with the outputted code but I'd imagine any case would be a bug rather than a feature.