Authors: Frankle, Osera, Walker, Zdancewic
Conference: POPL 2016
This is a very interesting paper about a cool and rather hot topic: program synthesis. Since the paper has been accepted at POPL 2016, one expects a theoretical investigation of the subject. As the title of the paper suggests, it is indeed the case. A surprise comes from a quite detailed explanation on the implementation challenges the authors have faced in developing a prototype.
Program synthesis is about code generation with respect to some high-level specification of a computational problem. This sounds like magic and, although feasible only to some extent, can even be quite useful in programming.
In recent years, many specification mechanisms have been proposed, some of them inherited or influenced by research on program verification. For example, this is the case for methods based on function contracts (as pre/post conditions).
One of the most user-friendly and intuitive way of specifying behaviors is that of input-output examples. This has notably been proposed within typeful settings such as functional programming languages. The idea is to provide the programmer with a language to express the behavior (output) of a function on a certain argument (input). So, for instance in a ML-like language, one may write something like
 -> 0 &&  -> 1 && [2;1] -> 2
to specify the function
length on lists.
The paper discussed here deals with two major questions about example-driven synthesis in this kind of settings:
The authors argue that each example can be given a type-theoretical interpretation in the sense of a refinement type, hence synthesis the process of finding an inhabitant of such a type. This “[…] enables us to exploit decades of research in type theory as well as its correspondence with intuitionistic logic rather than designing ad hoc theoretical frameworks for synthesis.”.
The authors develop a “[…] synchronized sequent calculus with intersection, union, function and syngleton types” as the formalization for specifications based on input-output examples. They refine this calculus from a more standard unsynchronized sequent calculus, itself derived from the classic formulation of refinement types in natural deduction. Soundness among the calculi is proved to hold.
This choice is quite natural as sequent calculus is a better fit when dealing with proof search, since it provides some handy properties such as the subformula property (there is no cut rule in sequent calculus), invertible rules and focusing.
This is indeed a smart move as a well-established foundational theory and semantics is known to entail good properties in the originated framework. This type-theoretic interpretation pays off as they “[…] develop a prototype implementation that extends the core calculus with a variety of additional feature” (i.e. base types, algebraic datatypes, parametric polymorphism, etc.). They even integrate to some extent libraries into the synthesis process.
The authors show that their synthesizer “[…] is competitive with state-of-the-art example-directed systems”, while “[…] able to condense the specifications, commonly seeing decreases of 10-20% compared to previous work.”. When polymorphism is possible, they achieve “even more dramatic specification reductions sometimes nearing 75%.”.
This is mainly due to the support of base types (like
nat) as valid
refinements, union types and parametric polymorphism. The first permits to
describe generic datatype values in a concise way like, for instance,
non-empty lists as
Cons(nat x list). The second, unions, permits to
combine refinements into larger types, eliminating some redundancy. Using
disjunctions, the first two cases of the
decrement function specification
0 -> 0 && 1 -> 0 && 2 -> 1
can be narrowed down to
(0 || 1) -> 0. This is true even for inductively
defined datatypes, such as lists, when refinements share
constructors. “For instance, lists
 can be combined into
Cons(0 x (Nil || Cons(1 x Nil))).”. The latter, parametric polymorphism,
helps in describing terse specifications and eases synthesis because of the
“free theorems” that polymorphic types imply.
Negation is not part of the refinement types language here. They are
encoded with De-Morgan’s laws though, as negations are important when
refining specifications after a first synthesis failure. Indeed, consider
once again the above specification for the
length function. That can be
seen as the specification for the
head function as well (as GB has
explained sometime during my presentation). Authors “[…] offer a similar
capability” to counterexample-guided inductive synthesis (CEGIS) “in the
form of negation, which makes it possible to articulate that a program is
not an inhabitant of a particular refinement.”. (This is correct to some
extent since real CEGIS is automatic.) Having negation, the original
length can be constrained by adding
 -> not(0).
The last part of the paper is dedicated to the prototype implementation. In particular, when the context of a sequent allows different choices in how to proceed the proof/program construction, “[…] making the correct decision in a performance-friendly manner proved to be the most difficult aspect of designing our prototype.”. The authors detail different approaches with respective performances.
One may expect that an “intellingent”, refinement-guided proof search would perform the best. The authors show that this is not the case. Quite the opposite. They ended up implementing a “classic” type-based enumeration. This means enumerating all possible values of a given refinement type, followed by some filtering.
Why this phenomenon? “[…] part of the answer is that some spaces”, even if the largest possible as that given by typing only, “are easier to traverse than others. The remainder is that the irregular shapes of the refinement-constrained search spaces make their results harder to cache.”. Indeed, the authors explain that optimizations are of course necessary, such as caching results of term-enumeration, evaluation and typechecking along with deep hash-consing of refinements, expressions and types.
There are some limitations. The authors point out that: