This paper describes a system that generates compiler back ends from a strictly declarative specification of the code generation process. The generated back ends use tree pattern matching for code selection. Two methods for register allocation supporting a wide range of target architectures are provided. A general bottom-up pattern matching method avoids problems that occurred with previous systems using LR-parsing. The performance of compilers using generated back ends is comparable to very fast production compilers. Some figures are given about the results of using the system to generate the back end of a Modula-2 compiler.

In this paper we describe the design of a global machine independent low level optimizer for the Karlsruhe Ada Compiler. We give a short overview on the optimizations and data structures used in the optimizer as well as some experiences with the optimizer. Detailed measurements are provided for a collection of benchmarks. The average improvement of code speed is 40%.