Design and implementation of an optimizing type-centric compiler for a high-level language

Production compilers for programming languages face multiple requirements. They should be correct, as we rely on them to produce code. They should be fast, in order to provide a good developer experience. They should also be easy to maintain and evolve. This thesis shows how an expressive high level type system can be used to simplify the development of a compiler and demonstrates this on a compiler for Scala. First, it shows how expressive types of high level languages can be used to build internal data structures that provide a statically checked API, ensuring that important properties hold at compile time. Second, we also show how high level language features can be used to abstract the components of a compiler. We demonstrate this by introducing a type-safe layer on top of the bytecode emission phase. This makes it possible to abstract away the implementation details of the compiler frontend and run the same bytecode emission phase in two different Scala compilers. Third, it presents MiniPhases, a novel way to organize transformation passes in a compiler. MiniPhases impose constraints on the organization of passes that are beneficial for maintainability, performance, and testability. We include a detailed performance evaluation of MiniPhases which indicates that their speedup is due to improved cache friendliness and to a lower rate of promotions of objects into the old generations of garbage collectors. Finally, we demonstrate how the expressive type system of the language being compiled can be used for static analysis. We present a novel call graph construction algorithm which uses the typing context for context sensitivity. The resulting algorithm is both substantially faster and more precise than existing alternatives. We demonstrate the applicability of this analysis by extending common subexpression elimination to idempotent expression elimination.

Type-Safe Metaprogramming and Compilation Techniques For Designing Efficient Systems in High-Level Languages

Lionel Emile Vincent Parreaux

Software engineering practices have been steadily moving towards higher-level programming languages and away from lower-level ones. High-level languages tend to greatly improve safety, productivity, and code maintainability because they handle various implementation details automatically, allowing programmers to focus on their problem domains. However, the gains offered by high-level languages are usually made at the cost of reduced performance: higher-level languages usually consume more memory, run more slowly and require expensive garbage-collecting runtime systems. This trend has been worsening with the increasing adoption of the functional programming paradigm by the industry. Modern programmers are thus faced with a dilemma: should they favor productivity and lower maintenance costs at the expense of performance, or should they focus on performance, to the detriments of almost everything else? The main idea behind this thesis is that we can help solve this dilemma by making advances in type systems, metaprogramming, and compilers technology. In particular, we study how metaprogramming via statically-typed quasiquotation can let programmers define their own domain-specific optimizations in a safe way, while leveraging the latest advances in intermediate program representations. We present the design and implementation of the Squid metaprogramming framework, which extends the Scala programming language with multi-staged programming capabilities and more. We also present different application examples for Squid, including a polymorphic yet efficient library for linear algebra, a stream fusion engine improving on the state of the art, a demonstration of query compilation by rewriting, a staged SQL database system prototype, and a new embedded domain-specific language for expressing queries over collections of data.

EPFL2020

On Static Analysis for Expressive Pattern Matching

Mirco Dotta, Viktor Kuncak, Philippe Paul Henri Suter

Pattern matching is a widespread programming language construct that enables definitions of values by cases, generalizing if-then-else and case statements. The cases in a pattern matching expression should be exhaustive: when the value does not match any of the cases, the expression throws a run-time exception. Similarly, each pattern should be reachable, and, if possible, patterns should be disjoint to facilitate reasoning. Current compilers use simple analyses to check patterns. Such analyses ignore pattern guards, use static types to approximate possible expression values, and do not take into account properties of user-defined functions. We present a design and implementation of a new analysis of pattern matching expressions. Our analysis detects a wider class of errors and reports fewer false alarms than previous approaches. It checks disjointness, reachability, and exhaustiveness of patterns by expressing these conditions as formulas and proving them using decision procedures and theorem provers. It achieves precision by propagating possible values through nested expressions and approximating pattern-matching guards with formulas. It supports user-defined ``extractor'' functions in patterns by relying on specifications of relationships between the domains of such functions. The result is the first analysis that enables verified, declarative pattern matching with guards in the presence of data abstraction. We have implemented our analysis and describe our experience in checking a range of pattern matching expressions in a subset of the Scala programming language.

2008

Compile-Time Type-Driven Data Representation Transformations in Object-Oriented Languages

Vlad Ureche

High-level languages allow programmers to express data structures and algorithms that abstract over the type of data they handle. This improves code reuse and makes it possible to develop general-purpose libraries. Yet, data abstractions slow down program execution, as they require low-level indirection. In this thesis we explore three compile-time approaches that leverage type systems to reduce the cost of data abstractions, thus improving program performance. In the first part of the thesis we present miniboxing, a compile-time transformation that replaces generic classes by more efficient variants, optimized to handle primitive types. These variants use the miniboxed data encoding, producing speedups of up to 20x compared to generic classes. The miniboxing transformation is the main result of this thesis and motivates the other techniques. Generalizing miniboxing, we show the Late Data Layout (LDL) mechanism, which uses the type system to guide performance-oriented program rewritings. It can be instantiated to perform a host of transformations, such as miniboxing generics, inlining value classes and unboxing primitive types. The LDL mechanism has many desirable properties, such as provable correctness in handling different data representations, reduced number of conversions and built-in support for the object-oriented paradigm. Finally, we show Data-centric Metaprogramming, a technique that allows programmers to go beyond standard compiler optimizations by defining custom representations to be used for their data. These representations are then automatically introduced by the compiler when translating programs. This technique, similar in spirit to metaprogramming, opens new directions in programmer-driven optimizations and shows encouraging results, with speedups of up to 25x. Under the hood, Data-centric Metaprogramming uses the Late Data Layout mechanism.

EPFL2015

Compile-Time Type-Driven Data Representation Transformations in Object-Oriented Languages

Vlad Ureche

EPFL2015