[110011]

> Because only a vanishingly rare amount of the incorrectness in compiled C programs comes from a bug in the compiler.

True.

What I meant to say, but didn't, is, why aren't more people working on ensuring that this provability bubbles up from the compiler to the program itself. I come across a lot of work in specialized areas on this topic but haven't seen anyone have a go integrating formal proofs and a widely used programming language. At the moment this looks like a pipe dream because of the obvious objection (and I hope this disappears "soon") that currently it is too cumbersome to have to prove your programs, let alone write them in the first place.

[seanmcdirmid]

The biggest problem in formal program verification is the lack of a spec to verify against for most programs. There are projects like ur web that define correctness more broadly to get around that.

[pron]

That's certainly a problem, but it's not the biggest problem. Writing a formal spec is no harder than writing tests (and arguably easier). It doesn't have to be perfect -- the more correctness conditions the better. Of course, a complete spec -- one that says, anything that satisfies this is a correct implementation of the system -- is best, but it's really not required to start getting real benefits.

I'd say that the biggest problem is the intractability of verification. Xavier Leroy has said that his style of verification -- end-to-end (i.e., global correctness conditions verified all the way down to source- and machine-code) and employing formal proofs -- is only suitable to the smallest and simplest programs (he claims it's suitable to programs hundreds to thousands of lines long, and even he had to take a few shortcuts in CompCert, e.g., he eschewed some termination proofs). Model-checking fares only a little better, and static analysis -- which scales to arbitrarily-sized programs -- can only verify specific, and mostly local correctness properties. Other approaches can tractably verify global correctness conditions, but only against high-level descriptions of the algorithm/system, and don't reach all the way down to code.

AFAIK, the most robust, and still scalable, formal verification employed in real-world programs of considerable size is that employed by Altran UK. They verify global properties against a high-level spec (verified mostly with model-checking, and possibly some proofs), plus local properties at the code level (verified mostly automatically with SMT solvers with a small number of manual proofs), but leaving a gap in between. This could very well become easier and scale well, and could become affordable (as it automates relatively well) yet still extremely useful, as end-to-end verification is hardly ever required. They're using Z for the high-level spec and Ada SPARK for the code-level spec, but other tools could be used (e.g. TLA+/Alloy for the high level, and JML/ACSL/Clojure Spec/Spec# for the code level).

I've personally successfully used refinement mappings in TLA+, which allow you to move between different levels of the system description, with the different specs as well as the refinement mapping itself checked with a model checker. This is indeed harder than just checking a single description, but still in the realm of affordability. Taking refinements all the way down to the source level is still far from affordable, though.

[neel_k]

> [...] even he had to take a few shortcuts in CompCert, e.g., he eschewed some termination proofs.

This turned out to be a premature optimization, actually! The CakeML compiler did full functional correctness proofs for their register allocator and report that it wasn't significantly harder than Compcert's translation validation.

> Writing a formal spec is no harder than writing tests (and arguably easier).

This, I'm much less sure of. IME, the really hard thing about correctness proofs is that specifications and proofs often require genuinely new ideas that didn't occur in the code.

For example, the spec of something like a sorting routine is that the output returns a list in increasing order, and that the output is a permutation of the input. The concept of a permutation doesn't occur explicitly in the implementation of the sort, and so the need to make this concept explicit can stop learners dead in their tracks.

I taught Agda (a dependently-typed programming language) to undergraduates, and found that Agda-the-programming language was actually easier to teach than Haskell (the IDE is better and there's less magic). But Agda-the-proof-assistant was hard to teach, mostly AFAICT because I don't know how to teach having mathematical ideas well.

[pron]

> Agda-the-proof-assistant was hard to teach, mostly AFAICT because I don't know how to teach having mathematical ideas well.

One of the (several) problems I find with dependently-typed languages is that they don't easily allow you to separate specification from verification [1]. When using a language like TLA+, it's very easy to separate the two. I actually think it's easier to teach mathematical thinking with TLA+ than without any formal language at all, as you can get feedback as to what has gone wrong, i.e, you can "play" with the math, and use the model checker before writing a proof to get a counterexample. I agree that teaching how to think mathematically is still required and still hard, but there's much less accidental complexity in the way.

Another advantage is that complex concepts are only encountered later [2] (e.g., you can cleanly separate safety from the much more complex liveness). The only downside (for teaching) is that the proofs are declarative, so you don't get to teach actual syntactic manipulation of terms (which is, arguably, not important unless you want to teach low-level proofs). Of course, the computational theory is TLA rather than lambda calculus, so it's irrelevant if you want to teach that formal system in particular. On the other hand, advanced and very powerful concepts such as refinement are very elegant and relatively simple.

[1]: Another is that proof is pretty much the only form of verification (of specification given as types), which is both extremely costly and very rarely required, and yet another is that complicated matters are encountered very early on.

[2] I've been playing with Lean for some weeks, and I still don't understand what object `id` (the identity "function", polymorphic over all types in all universes) is; all I know is that it's not an actual function. I'm not sure whether it's the same or not in Agda.

[neel_k]

HN seems to have eaten my first reply, so here's an abbreviated version.

It's easy to separate specification from verification with dependent types. You can (and I did) show students how to take ordinary simply-typed functional programs, give a specification as a different type, and then do a verification by producing a term of the appropriate specification type. Then, you can use this to motivate intrinsically-typed strategies, and (a) show the students that intrinsic/extrinsic is a gradient rather than a dichotomy, and (b) show them how to exploit this gradient systematically.

As for liveness, in a purely functional language, basically the only interesting liveness property is termination, and this didn't pose a challenge for the students. Indeed, it was the opposite -- I got a lot of questions about why Haskell lacked a termination checker. (They had in their minds that termination checking was an impossible dream, and so seeing a system with a termination checker that was both usable and useful really fired their imaginations.)

Specs were still hard. A proof-based approach is actually desirable for teaching good spec-writing skills, because proofs are more sensitive to the details of the spec. Eg, property-based testcase generation is happy with the equation that the reverse of the reverse of a list is the same list, but it's too weak an induction hypothesis for a proof. So it strongly motivates you to get to the characterizing equation for reversal (that reversing the concatenation of two lists is the concatenation of their reverses in the opposite order). (And it's still good for testcase generation.)

Unfortunately, I don't know enough small, self-contained examples where you have to strengthen the induction hypothesis to get the proof to go through. Ideally I'd want about 60 or 70 good examples of this (to set problems for a course) and, alas, I'd be hard-pressed to come up with a tenth that many satisfying examples.

[pron]

> You can (and I did) show students how to take ordinary simply-typed functional programs, give a specification as a different type, and then do a verification by producing a term of the appropriate specification type.

Sure, but disentangling the program from the spec and the proof is not quite what I meant. By separating specification from verification I mean using the ability to specify (which is a much more important skill), completely separately from how that specification is verified, whether by proof or by other, weaker means.

> Eg, property-based testcase generation is happy with the equation that the reverse of the reverse of a list is the same list, but it's too weak an induction hypothesis for a proof.

Ah, but that's exactly the benefit you gain from separating specification from verification. To write a proof of an algorithm in TLA+ you indeed need to find an inductive invariant. But in real life, when specifying large and complex systems (as I've experienced myself, and as reported by Amazon), finding an inductive invariant can be really, really hard. However, that does not make the specification useless, because it can still be checked with a model-checker. Indeed, most of the time, writing proofs of real-world programs is neither affordable nor required. So, I agree that teaching writing proofs is valuable, but it is no doubt secondary to teaching how to write specs. Being able to separate the two makes teaching and learning easier, and allows to focus on the more important skill first.

It's been reported by Amazon (and my experience was the same), that even an entry-level engineer can learn TLA+ and put it to good use in roughly two weeks, without any help outside the many tutorials. Those weeks are spent entirely on learning mathematical thinking in the context of computation, and make use only of the model-checker. After two weeks of learning TLA+, you're well into specifying complex, real-life systems, while after two weeks of learning, say, Lean, you're barely beyond learning the basics of quantification. Then, those who want can dive more deeply into writing proofs, but even there, the model-checker helps. For example, the model-checker can verify that your inductive invariant is, indeed, an inductive invariant (see below). In fact, Lamport's tutorial (the "hyperbook") is divided into a "principles track", a "specification track" and a "proof track". Inductive invariants are taught in the principles track because, despite being necessary for a proof, they are very useful for insight (which you can get even without a formal proof) as they form the core of why the algorithm works.

> Unfortunately, I don't know enough small, self-contained examples where you have to strengthen the induction hypothesis to get the proof to go through. Ideally I'd want about 60 or 70 good examples of this (to set problems for a course) and, alas, I'd be hard-pressed to come up with a tenth that many satisfying examples.

Yep, coming up with examples is very hard :) But TLA+ makes it easy to get an intuitive feel for the problem of the inductive invariant. In TLA+, the inductive invariant can be formally expressed generally. For a specification written in the "normal form" of:

    Spec ≜ Init ∧ □[Next] ∧ Liveness

Where `Init` specifies the initial condition, `Next` specifies the next-state relation (not a function!), and `Liveness` is the fairness condition which is only necessary if you care about liveness, an inductive invariant `I` of a correctness property `P` (i.e., we want to show `Spec ⇒ P`) is a predicate such that:

      I ⇒ P
    ∧ Init ⇒ I
    ∧ I ∧ Next ⇒ I'

(The logic contains the inference rule: `A ⇒ I, I ∧ B ⇒ I' ⊦ A ∧ □[B] ⇒ □I`)

The last conjunct (which states that if `I` holds in the current state, then any legal step would make `I` hold in the next) is the tricky part. Looking at it put in this way clearly shows the difficulty: a weak `I` may encompass too many states, many of which are unreachable, and so might not be an inductive invariant even though `P` is true. It is possible to check the inductive invariant in the model checker, by writing the specification, `I ∧ □[Next]` and checking whether it implies `I`. This, again, gives an intuition, as now `I` specifies the initial condition, and you can actually see (through the generated counterexamples) those "initial" states specified by `I` that are not reachable when starting from `Init`, and so you can see the weakness of the invariant in a very concrete way.

BTW, could you perhaps explain to me what kind of object `id` in Lean is? The problem is that there is no universe that contains all universes, so does that mean that `id` is merely syntax used to generate objects when types can be inferred, or does it represent some actual object in the theory?

[xgk]

> separate safety from the much more complex liveness

True, and with program logics, you can choose partial correctness vs total vs generalised correctness.

Itt's even worse in practise. With many contemporary dependently typed languages you'll have to worry about termination even during programming, even before thinking about verification.

The Curry-Howard prover advocates seem to regard this as a solved problem ("just distinguish programs from proofs by way of a typing system, problem solved") but has this solution been implemented in any mature systems as of 2017?

[catnaroek]

> use the model checker before writing a proof to get a counterexample.

Lovely, they should have taught me algebraic geometry that way. /s

Trying to write proofs without knowing a priori the truth value of the proposition to be proved builds character, and it's an important part of developing mathematical maturity. It forces you to develop intuition for what could or could not be true.

[pron]

Programmers on a deadline don't want to spend months building character on every software component.

[xgk]

> to separate specification from verification

Exactly.

Program logics are superior in this regard: the encourage rather than inhibit separation of concerns. (Not to mention: program logics are more developed and cover a large class of computing paradigms, while dependent types have not grown much beyond the pure functions ghetto.)

I put forward this point to dependent types luminaries all the time. Its difficult to get more than polite silence in response.

      * * *

It's partly a social problem: The community working on dependently typed languages and tools typically have most of their programming experience with pure functional languages. Exaggerating only a bit, here is the typical trajectory of a formal methods researcher:

- As an undergraduate writes a lambda-calculus interpreter in Lisp/Scheme/Racket/Haskell

- As a PhD Student writes their own pure language with wacky typing system, and demonstrate its usefulness by embedding a lambda-calculus interpreter in it.

- As a postdoc verifies a lambda-calculus interpreter in Coq/Agda.

- As an assistant professor writes their own Curry-Howard-based interactive prover in some dependently typed language and demonstrate its usefulness by verifying a lambda-calculus interpreter in it.

- After tenure, gets their students to do the above.

2018's POPL has a type-theory in type-theory paper, but none on how programming language theory can improve SAT solving ...

It's an echo chamber. I recently reviewed a grant proposal by some of the more famous dependent types researchers. They proposed using HoTT to verify computational effects. The grant proposal claimed (1) that there is currently no known way to verify effectful languages and (2) that embedding generalisations of effect monads in HoTT is the way to overcome this problem. It seems like these people have not ever heard of program logics?

   * * *

A second reason is more serious: the real problem of formal verification is not this-language/that-logic. In day-to-day verification, it doesn't really matter that much whether you are using this/that approach. You need to set up the right definitions and invariants. That's the hard part. Whether you express them in some program logic or dependently typed system doesn't matter much, you can usually transliterate between them.

The real problem is automation.

Better autmation means dramatically better SAT solvers (modern SMT solvers are heavily reliant of SAT solvers). It's just unclear where 10x, 100x, 1000x speedups in SAT solvers should come from. We don't even know how to parallelise SAT solvers well. In the absence of dramatic progress in SAT solvers, what can we do other than tinkering with dependently typed languages and their relatives?

[pron]

> here is the typical trajectory of a formal methods researcher

That's a programming language theory researcher. The vast majority of formal methods research is done neither using dependent types nor lambda calculus (and the vast majority of formal tools are similarly neither based on types nor on lambda calculus). However, I do agree that practitioners (i.e., programmers) who become interested in formal methods through functional programming, tend to miss out on the vast majority of progress in formal methods, which isn't related to functional programming at all, and indeed, is normally done in formalisms other than the lambda calculus (which introduces problems related to the common use of higher-order functions, which make some of the clever automation methods harder to apply). They also become familiar with equational reasoning, which works well in pure functional programs, and aren't aware of the more generally-applicable (and arguably more useful) assertional reasoning, which works well even in mainstream imperative languages. They may be aware of assertions, but may not know that their use is actually a formal method of sorts and is an application of Hoare logic.

> Better autmation means dramatically better SAT solvers (modern SMT solvers are heavily reliant of SAT solvers). It's just unclear where 10x, 100x, 1000x speedups in SAT solvers should come from. We don't even know how to parallelise SAT solvers well.

The problem is worse: we don't even know why SAT solvers work in practice at all.

> In the absence of dramatic progress in SAT solvers, what can we do other than tinkering with dependently typed languages and their relatives?

Well, much of formal methods research is done in the field of abstract interpretation (which forms the core of most sound static analysis tools). There is still a lot of interesting progress to be made there. Separation logic is an example of such relatively recent (and important!) progress.

[tom_mellior]

> This turned out to be a premature optimization, actually! The CakeML compiler did full functional correctness proofs for their register allocator and report that it wasn't significantly harder than Compcert's translation validation.

Not sure if you're saying that using a translation validation approach for register allocation was a premature optimization. But if that's what you're saying, you're wrong: The paper describing the register allocation validator explains that this replaced a register allocator in Coq that was proved directly. The validator approach is much smaller and simpler than this direct proof was. So it was an actual optimization after the fact.

It's possible that CakeML does it better, of course.

[neel_k]

That's interesting! I heard this from some CakeML people, who told me they decided to go with a full proof because it wasn't much more work for them. Maybe Compcert has a fancier register allocator architecture that benefits more from translation validation?

[tom_mellior]

Maybe, but on the other hand maybe the CakeML developers have some proof techniques that the CompCert people didn't think of. I don't know.

For reference, my only knowledge of this point comes from the CC 2010 paper at https://xavierleroy.org/publi/validation-regalloc.pdf, which says in section 5: "From a proof engineering viewpoint, the validator is a success. Its mechanized proof of correctness is only 900 lines of Coq, which is quite small for a 350-line piece of code. (The typical ratio for Coq program proofs is 6 to 8 lines of proof per line of code.) In contrast, 4300 lines of Coq proof were needed to verify the register allocation and spilling passes of the original CompCert compiler. Even this earlier development used translation validation on a sub-problem: the George-Appel graph coloring algorithm was implemented directly in untrusted Caml code, then followed by a verified validator to check that the resulting as- signment is a valid coloring of the interference graph. Later, Blazy, Robillard and Appel conducted a Coq proof of the graph coloring algorithm [13]. This is a fairly large proof: in total, more than 10000 lines of proof are needed to com- pletely verify the original CompCert register allocation and spilling passes."

(And hence my statement above was also inaccurate in that the original CompCert was not fully directly proven. I had misremembered the details of this paragraph. The point still stands that the newer validation solution is simpler.)

[catnaroek]

Clearly, the only way forward is to make programming attractive to mathematicians, since making mathematics attractive to programmers has largely failed.

[pron]

I'm not at all sure mathematicians would make better programmers. One of the mistakes, IMO, some proof assistants make in their design philosophy is that since, from the perspective of some formal logics, writing programs and proofs are "the same thing", their is a corresponding equivalence of the two activities. This is somewhat analogous to concluding that since both programming and writing involve entering words into a text editor, the two are the same. But the fact that, from some specific perspective, both mathematicians and programmers write proofs does not mean that the activities are similar. Both physicists and mathematicians solve equations, but no one would say that physics is math. Mathematicians manipulate objects that are very different from those that programmers manipulate. Mathematicians manipulate objects that are very regular, and reasoning about them is largely tractable, while programmers manipulate very irregular and very intractable objects. Ostensibly, the difference is only in measure, but a difference in any robust complexity measure is usually a difference in quality. Even the actual proofs are very different. Mathematical proofs are usually short but mathematically deep, while program proofs are mathematically shallow and uninteresting, but very, very long and full of detail. Not to mention that the motivations are completely different. The goal of a programmer is to write a program that meets the requirements, which may include some distribution of acceptable bugs that depends on their severity (i.e. a major, but non-catastrophic bug occurring once a month and some minor bugs occurring daily are acceptable), and increasing the cost of programming by an order of magnitude in order to make all bugs provably absent is the wrong thing to do. In math, the requirements are completely different.

[catnaroek]

> One of the mistakes, IMO, some proof assistants make in their design philosophy is that since, from the perspective of some formal logics, writing programs and proofs are "the same thing", their[sic] is a corresponding equivalence of the two activities.

Don't take this as an endorsement of proof assistants, but programming and proving are special cases of a single general activity: constructing a mathematical object.

> Both physicists and mathematicians solve equations, but no one would say that physics is math.

The crucial difference is that physicists, unlike mathematicians or programmers, don't create universes of their own. Their job is to explain the behavior of the one that already exists.

> Mathematicians manipulate objects that are very regular, and reasoning about them is largely tractable

Only because they have made them so. Also, insanely irregular and intractable mathematical objects do exist, yet nobody would be taken seriously (in the mathematical community) who uses this as an excuse for lowering the standards of proof.

> while programmers manipulate very irregular and very intractable objects.

Only because they have made them so. Dijkstra anticipated the need to devise program structures that are amenable to formal reasoning.

> The goal of a programmer is to write a program that meets the requirements, which may include some distribution of acceptable bugs that depends on their severity

Then quantify the bug, and make it a part of the specification.

> (i.e. a major, but non-catastrophic bug occurring once a month and some minor bugs occurring daily are acceptable)

Bugs occur in the program, not in its execution traces.

[catnaroek]

Ur/Web doesn't redefine “correctness”, since, by definition, your program specification is the definition of what it means for your program to be “correct”. Ur/Web merely guarantees the absence of certain behaviors in programs written in it.