Ross Tate

Equality saturation: a new approach to optimization

Optimizations in a traditional compiler are applied sequentially, with each optimization destructively modifying the program to produce a transformed program that is then passed to the next optimization. We present a new approach for structuring the optimization phase of a compiler. In our approach, optimizations take the form of equality analyses that add equality information to a common intermediate representation. The optimizer works by repeatedly applying these analyses to infer equivalences between program fragments, thus saturating the intermediate representation with equalities. Once saturated, the intermediate representation encodes multiple optimized versions of the input program. At this point, a profitability heuristic picks the final optimized program from the various programs represented in the saturated representation. Our proposed way of structuring optimizers has a variety of benefits over previous approaches: our approach obviates the need to worry about optimization ordering, enables the use of a global optimization heuristic that selects among fully optimized programs, and can be used to perform translation validation, even on compilers other than our own. We present our approach, formalize it, and describe our choice of intermediate representation. We also present experimental results showing that our approach is practical in terms of time and space overhead, is effective at discovering intricate optimization opportunities, and is effective at performing translation validation for a realistic optimizer.

Equality-based translation validator for LLVM

We updated our Peggy tool, previously presented in [6], to perform translation validation for the LLVM compiler using a technique called Equality Saturation. We present the tool, and illustrate its effectiveness at doing translation validation on SPEC 2006 benchmarks.

Generating compiler optimizations from proofs

We present an automated technique for generating compiler optimizations from examples of concrete programs before and after improvements have been made to them. The key technical insight of our technique is that a proof of equivalence between the original and transformed concrete programs informs us which aspects of the programs are important and which can be discarded. Our technique therefore uses these proofs, which can be produced by translation validation or a proof-carrying compiler, as a guide to generalize the original and transformed programs into broadly applicable optimization rules. We present a category-theoretic formalization of our proof generalization technique. This abstraction makes our technique applicable to logics besides our own. In particular, we demonstrate how our technique can also be used to learn query optimizations for relational databases or to aid programmers in debugging type errors. Finally, we show experimentally that our technique enables programmers to train a compiler with application-specific optimizations by providing concrete examples of original programs and the desired transformed programs. We also show how it enables a compiler to learn efficient-to-run optimizations from expensive-to-run super-optimizers.

Getting F-bounded polymorphism into shape

We present a way to restrict recursive inheritance without sacrificing the benefits of F-bounded polymorphism. In particular, we distinguish two new concepts, materials and shapes, and demonstrate through a survey of 13.5 million lines of open-source generic-Java code that these two concepts never actually overlap in practice. With this Material-Shape Separation, we prove that even naïve type-checking algorithms are sound and complete, some of which address problems that were unsolvable even under the existing proposals for restricting inheritance. We illustrate how the simplicity of our design reflects the design intuitions employed by programmers and potentially enables new features coming into demand for upcoming programming languages.

Taming wildcards in Java's type system

Wildcards have become an important part of Javas type system since their introduction 7 years ago. Yet there are still many open problems with Javas wildcards. For example, there are no known sound and complete algorithms for subtyping (and consequently type checking) Java wildcards, and in fact subtyping is suspected to be undecidable because wildcards are a form of bounded existential types. Furthermore, some Java types with wildcards have no joins, making inference of type arguments for generic methods particularly difficult. Although there has been progress on these fronts, we have identified significant shortcomings of the current state of the art, along with new problems that have not been addressed. In this paper, we illustrate how these shortcomings reflect the subtle complexity of the problem domain, and then present major improvements to the current algorithms for wildcards by making slight restrictions on the usage of wildcards. Our survey of existing Java programs suggests that realistic code should already satisfy our restrictions without any modifications. We present a simple algorithm for subtyping which is both sound and complete with our restrictions, an algorithm for lazily joining types with wildcards which addresses some of the shortcomings of prior work, and techniques for improving the Java type system as a whole. Lastly, we describe various extensions to wildcards that would be compatible with our algorithms.

The sequential semantics of producer effect systems

Effects are fundamental to programming languages. Even the lambda calculus has effects, and consequently the two famous evaluation strategies produce different semantics. As such, much research has been done to improve our understanding of effects. Since Moggi introduced monads for his computational lambda calculus, further generalizations have been designed to formalize increasingly complex computational effects, such as indexed monads followed by layered monads followed by parameterized monads. This succession prompted us to determine the most general formalization possible. In searching for this formalization we came across many surprises, such as the insufficiencies of arrows, as well as many unexpected insights, such as the importance of considering an effect as a small component of a whole system rather than just an isolated feature. In this paper we present our semantic formalization for producer effect systems, which we call a productor, and prove its maximal generality by focusing on only sequential composition of effectful computations, consequently guaranteeing that the existing monadic techniques are specializations of productors.

Inferable object-oriented typed assembly language

A certifying compiler preserves type information through compilation to assembly language programs, producing typed assembly language (TAL) programs that can be verified for safety independently so that the compiler does not need to be trusted. There are two challenges for adopting certifying compilation in practice. First, requiring every compiler transformation and optimization to preserve types is a large burden on compilers, especially when adopting certifying compilation into existing optimizing non-certifying compilers. Second, type annotations significantly increase the size of assembly language programs. This paper proposes an alternative to traditional certifying compilers. It presents iTalX, the first inferable TAL type system that supports existential types, arrays, interfaces, and stacks. We have proved our inference algorithm is complete, meaning if an assembly language program is typeable with iTalX then our algorithm will infer an iTalX typing for that program. Furthermore, our algorithm is guaranteed to terminate even if the assembly language program is untypeable. We demonstrate that it is practical to infer such an expressive TAL by showing a prototype implementation of type inference for code compiled by Bartok, an optimizing C# compiler. Our prototype implementation infers complete type annotations for 98% of functions in a suite of realistic C# benchmarks. The type-inference time is about 8% of the compilation time. We needed to change only 2.5% of the compiler code, mostly adding new code for defining types and for writing types to object files. Most transformations are untouched. Type-annotation size is only 17% of the size of pure code and data, reducing type annotations in our previous certifying compiler [4] by 60%. The compiler needs to preserve only essential type information such as method signatures, object-layout information, and types for static data and external labels. Even non-certifying compilers have most of this information available.

Java and scala's type systems are unsound: the existential crisis of null pointers

We present short programs that demonstrate the unsoundness of Java and Scalas current type systems. In particular, these programs provide parametrically polymorphic functions that can turn any type into any type without (down)casting. Fortunately, parametric polymorphism was not integrated into the Java Virtual Machine (JVM), so these examples do not demonstrate any unsoundness of the JVM. Nonetheless, we discuss broader implications of these findings on the field of programming languages.

Sound gradual typing is nominally alive and well

Recent research has identified significant performance hurdles that sound gradual typing needs to overcome. These performance hurdles stem from the fact that the run-time checks gradual type systems insert into code can cause significant overhead. We propose that designing a type system for a gradually typed language hand in hand with its implementation from scratch is a possible way around these and several other hurdles on the way to efficient sound gradual typing. Such a design process also highlights the type-system restrictions required for efficient composition with gradual typing. We formalize the core of a nominal object-oriented language that fulfills a variety of desirable properties for gradually typed languages, and present evidence that an implementation of this language suffers minimal overhead even in adversarial benchmarks identified in earlier work.

Strict and Lazy Semantics for Effects: Layering Monads and Comonads

Two particularly important classes of effects are those that can be given semantics using a monad and those that can be given semantics using a comonad. Currently, programs with both kinds of effects are usually given semantics using a technique that relies on a distributive law. While it is known that not every pair of a monad and a comonad has a distributive law, it was previously unknown if there were any realistic pairs of effects that could not be given semantics in this manner. This paper answers that question by giving an example of a pair of effects that cannot be given semantics using a distributive law. Our example furthermore is intimately tied to the duality of strictness and laziness. We discuss how to view this duality through the lens of effects.