I don't buy the argument that what people want to do with smart contracts requires Turing-completeness. In fact, as time goes on I become more and more convinced that Turing completeness is highly overrated. Almost all code wants to do finite and/or structurally inductive operations. The only case that doesn't fall in that are things like request loops where you want to "infinitely" wait, process request, loop back to waiting. But even those things can be done in total languages via co-induction, so exactly why do we want to make our analysis/verification work more difficult by using a non-total language for things like Ethereum?
But those total systems are, of course, subsets of Ethereum, which means that if you write your program in an obviously finite way, you can use the same inductive proofs you would use for a total language.
Turing completeness makes it hard to prove properties automatically about arbitrary programs, but you can still prove properties about specific programs.
I'm pretty sure that MOST research on program verification has been done in the context of Turing complete systems, from the work of Floyd and Hoare, to the symbolic execution of Deutsch and King, to temporal logic, TLA+, etc.
I'm not an expert on Ethereum, but even if they did pick a total language, how would you deal with bounding the CPU cost of complex contracts? Even if you could formally verify a loop would eventually terminate, wouldn't long running loops or those with expensive computations have bad impacts on the network? Their "gas" system is their way of dealing with that and perhaps you would still need something similar even with a total language.
I can't think of program verification research where both termination and CPU cost is verified. I can think of some papers that deal with identifying Big Oh growth rates though.
Either way, smart contracts sound like one of the top tier areas where you want strong static typing that has a route to be formally verified.
Exactly, knowing that the contract code will eventually halt is nice to know but knowing that it will halt after a reasonable CPU cost seems much more important in this case (which is really interesting).
From what I've seen from example contracts, it doesn't seem like the chance of infinite loops is that high. Personally, I find that in mainstream languages, if you accidentally write an infinite loop you find out very quickly on the first few runs and code that could be hiding an infinite loop stands out. The vast majority of loops just iterate from the start of a collection to the end.
For what I've read, having the sender setting gas limits seems really awkward especially when the limit is exceeded. Are there practical alternatives? Even with O(1) you'd need to limit the computation time.
The Parity multisig contract is very complex and I wouldn't and didn't trust any funds with it. I hope that the newly deployed fixed multisigs are bug-free but how would you know?
Right now I'm working on tools for static analysis of bytecode and using such tools to verify the correctness of contracts written by hand in assembly, which I think makes a great deal of sense for the simple contracts that we need as utilities (multisigs, tokens, etc).
Yeah...any ideas why they didn't go with a language that had an easier route for formal verification? It seems like an ideal application and the contract specifications seem like they would be fairly straightforward to prove (compared to what you'd see in most journal papers for instance). I'm guessing the designers just weren't aware of theorem provers but I'm puzzled why they didn't go with a language with strong static typing where private access and immutability are the defaults plus avoiding anything to do with the mechanism they have for picking default handling functions when a message isn't understood. Great experiment to watch though!
General mathematical proofs? If you mean paper proofs they wouldn't be considered formal as they don't go down to the axiom level.
Basically, if I'm going to spend even 5 minutes playing with smart contracts, I want to feel like the language designers were incredibly paranoid and aware of just how good modern languages can be at provability and security.
But, I mean, there's no proof that the Idris compiler does what it's supposed to do, so you'd still have a huge trusted base.
Yeah, all my formal proof experience is with Coq, which is just a bit too idiosyncratic and difficult for this sort of thing. But I seem to remember that—especially if you're willing to certify specific results, and not necessarily the program in general—you can make your trusted computing base very small indeed.
I mean, if the CompCert C compiler can be as thoroughly verified as it is (http://compcert.inria.fr/compcert-C.html), there's no reason why it shouldn't be possible to verify smart contracts.
Jack Petterson and Robert Edström did a master's thesis about retargeting the Idris compiler to emit EVM code, with a custom effect type to express smart contract effects. They ended up modestly skeptical of functional programming as a paradigm for EVM, instead looking toward process calculi. https://publications.lib.chalmers.se/records/fulltext/234939...
Yoichi Hirai's summary of his research into Ethereum formal verification as a full-time formal methods researcher at the Foundation: https://github.com/pirapira/ethereum-formal-verification-ove...
See especially his formalization of the EVM as Hoare triples in Lem, which is usable from Isabelle: https://github.com/pirapira/eth-isabelle -- very cool stuff, although I have to say that the ostensible complexity and difficulty of his proof verifying an utterly simple wallet contract makes me a bit skeptical of this approach to smart contract correctness: https://github.com/pirapira/eth-isabelle/blob/master/example...
Hmm, so from the smart contract examples I've seen, I would have thought writing them in something like Coq would be quite feasible. The code examples I've seen tend to only have simple loops if any, basic conditions and the code is short (e.g. compared to something like web development or mobile apps). Writing anything inside a theorem prover is very challenging though but it seems a compelling application.
If you were implementing a smart contract in something like Coq, I'd be interested to read about what kind of specifications you would typically want to verify. I guess properties like "the owner of this wallet cannot be changed" and "the owner of this wallet can only accept money but not give it away" would be good ones.
Solidity doesn't have a problem with formal verification of programs. Check their bug tracker: you can't even trust the runtime to do what it claims to do. On several levels.
A disaster that could have been avoided if they had done a little more research on security and smart contract languages:
http://lambda-the-ultimate.org/node/5003
Solidity is still clearly focused on programming productivity instead of contract soundness. The built in dynamism is proof of this.
As a platform for experimentation that's fine. As a platform for doing real things with meaningful amounts of money, it's madness.
Of course there are limited formalisms that make certain types of verification easier, but proving programs has been possible since, like, the 1960s.
A multisig, for example, has a finite number of states when considered under symbolic execution. A model checker can rip through it and prove whatever properties you want.
Is the Ethereum world right now full of buggy contracts that haven't been proven correct? Yes, absolutely. It's a disaster. But it's a disaster that points the way fairly straightforwardly.
Should Ethereum not have been launched until it had solid methods for auditing and proving? Maybe, but that's not how the world works, typically. Worse is better and all that.