Don't agree fully. Denormalization is very usefull sometimes. For example, I have ~100 tables which is representing some entities. Few entity types are primary and there are 5 complex searches with ~50 criterias. We decided to create denormalized form of data to provide search. Note, we kept an old structure as primary but populate denormalized data periodically (denormalization takes ~4-5m per day). Search is working very fast...

Is any idea how to avoid it without of using 20-30 joins (5 of them are very conditional: CASE WHEN, subquries, etc)?

I think it is pretty good decision.