Think about that one for a second. An encrypted block is essentially supposed to look like random data. I.e. if two people encrypt the same file with different keys, you shouldn't be able to tell that they're the same file (or your encryption sucks). So your block-level de-duping then depends on incidental matches between random data.
What's the probability of two 4KB (or whatever) blocks of random data being identical? Basically zero even with petabytes of data.
The issue is that you don't get the full benefits of encryption.
If you upload the map to the rayiner family treasure that only you have seen you're good. No one else will be able to read it.
But if you upload the latest episode of Modern Family and Disney gets ahold of the same rip you used they (if they can get a government to help them out) can see what you did and charge you with copyright infringement (or whatever the appropriate crime would be).
After 3-4 years of high-profile CPA-2 attacks on TLS, .NET, Java, and other systems, you'd think we'd all be a lot more skeeved out about cryptosystems that demand known-plaintexts. There's already an obvious conceptual attack (beyond file confirmation) in naive "convergent encryption", which is that you can leverage small amounts of known plaintext to learn unknown plaintext.
What's the probability of two 4KB (or whatever) blocks of random data being identical? Basically zero even with petabytes of data.