Compilation is the automatic translation of high-level programming languages, understandable by humans, to lower-level languages, often executable directly by hardware. It is an essential step in the efficient execution, and therefore in the adoption, of high-level languages. Compilation is at the interface between programming languages and computer architecture, and because of this position has had considerable influence on the designs of both. Compilers have also attracted considerable research interest as the oldest instance of symbolic processing on computers.
Compilation has been the topic of much research work in the last 40 years, focusing mostly on high-performance execution ("optimization") of low-level languages such as Fortran and C. Two major results came out of these efforts. One is a superb body of performance optimization algorithms, techniques and methodologies that cries for application to more exotic programming languages. The other is the whole field of static program analysis, which now serves not only to increase performance but also to increase reliability, through automatic detection of bugs and establishment of safety properties. The work on compilation carried out in the Gallium group focuses on two less investigated topics: compiler certification and efficient compilation of "exotic" languages.
Formal verification of compiler correctness.
While the algorithmic aspects of compilation (termination and complexity) have been well studied, its semantic correctness -- the fact that the compiler preserves the meaning of programs -- is generally taken for granted. In other terms, the correctness of compilers is generally established only through testing. This is adequate for compiling low-assurance software, themselves validated only by testing: what is tested is the executable code produced by the compiler, therefore compiler bugs are detected along with application bugs. This is not adequate for high-assurance, critical software which must be validated using formal methods: what is formally verified is the source code of the application; bugs in the compiler used to turn the source into the final executable can invalidate the guarantees so painfully obtained by formal verification of the source.
To establish strong guarantees that the compiler can be trusted not to change the behavior of the program, it is necessary to apply formal methods to the compiler itself. Several approaches in this direction have been investigated, including translation validation, proof-carrying code, and type-preserving compilation. The approach that we currently investigate, called compiler verification, applies program proof techniques to the compiler itself, seen as a program in particular, and use a theorem prover (the Coq system) to prove that the generated code is observationally equivalent to the source code. (See the CompCert project.) Besides its potential impact on the critical software industry, this line of work is also scientifically fertile: it improves our semantic understanding of compiler intermediate languages, static analyses and code transformations.
Efficient compilation of high-level languages.
High-level and domain-specific programming languages raise fascinating compilation challenges: on the one hand, compared with Fortran and C, the wider semantic gap between these languages and machine code makes compilation more challenging; on the other hand, the stronger semantic guarantees and better controlled execution model offered by these languages facilitate static analysis and enable very aggressive optimizations. A paradigmatic example is the compilation of the Esterel reactive language: the very rich control structures of Esterel can be resolved entirely at compile-time, resulting in software automata or hardware circuits of the highest efficiency.
We have been working for many years on the efficient compilation of functional languages. The native-code compiler of the Objective Caml system embodies our results in this area. By adapting relatively basic compilation techniques to the specifics of functional languages, we achieved up to 10-fold performance improvements compared with functional compilers of the 80s. We are currently considering more advanced optimization techniques that should help bridge the last factor of 2 that separates Caml performance from that of C and C++. We are also interested in applying our knowledge in compilation to domain-specific languages that have high efficiency requirements, such as modeling languages used for simulations.
X. Leroy. Formal verification of a realistic compiler, Communications of the ACM, 52(7):107-115, 2009.