Introduction

The Hipparch project [Cas94b] uses the Esterel language [Dio95] for describing protocol building blocks. These blocks can be configured to form an application specific protocol implementation. Using application specific knowledge should improve the performance of such an implementation when compared to the use of a standard protocol. However, the C code generated by the currently existing Esterel compiler is not very fast, and moreover has a large code size [Cas94a]. This is a problem the compiler shares with many similar tools that generate executable code starting from a protocol defined in a formal description technique (FDT). This paper describes our design for an optimizer that we incorporated into the Esterel compiler.

After analyzing the code generated by the Esterel compiler, we decided that two optimization techniques merit further investigation:

• implementation of a fast path: this optimization concerns the order in which the different paths through the protocol configuration are sequentialized in the C code. This will increase the spatial locality of the code, and lead to benefits due to a better use of the memory hierarchy, e.g. with modern cache-based RISC-CPUs [Mos95]. Applying this optimization requires information on how often each of the paths will be executed, and which nodes on the path will be executed together.
• function inlining: the Esterel implementation of a protocol building block may contain calls to auxiliary C functions. In general, function inline expansion can lead to a considerable performance increase. However, function inlining must be applied in a controled way in order to avoid code size explosion and adversary effects on cache performance [Mos95]. Thus, only functions that will be used frequently at run time should be inlined.