Analysis and improvement of a multi-pass compiler for a pipeline architecture

Abstract

In this thesis a parallel environment for the execution of a multi-pass Pascal compiler is considered. Some possible and appropriate ways to speed up each pass of the parallelized compiler are investigated. In addition, a new approach, using the concepts of software science, is explored for obtaining gross performance characteristics of a multi-pass compiler. A pipeline architecture is used for the parallel compilation. The performance characteristics of the pipelined compiler are determined by a trace-driven simulation of the pipelined compiler. The actions in the multi-processor system are synchronized by an event-driven simulation of the pipeline system. The pipelined compiler and possible improvements are analyzed in terms of the location of the bottleneck, queue size, overhead factor, and partition policy. The lexical analysis phase is found to be the initial bottleneck. The improvement of this phase and its effects on the other phases are presented. Also, possible methods for improving the non-lexical analysis phases are investigated based on a study of the data structures and operations of these phases. For obtaining gross performance characteristics of a multi-pass compiler, an analysis based only on the intermediate code files is performed. One of the key concepts in Halstead's software science, called the language level, is applied to this analysis. From the experimental results and statistical verification it is found that there exists a strong correlation between the stand-alone execution time and language level. 1 CHAPTER I. INTRODUCTION Many systems, both natural and artificial, are a combination of parallel and sequential activities. Among these systems the parallel and sequential aspects are harmoniously combined and are not exclusive of each other at all. The relationships between parallel and sequential operations depend on several factors. Two of these factors are resource limits and level of detail. Generally speaking, if the number of the resources is equal to (or more than) the number of users, the users can act in parallel with each other unless there Is some other constraint preventing free access to the available resources. An example of such a constraint on the resources are those imposed by "reusable" resources. For example, when certain information structures are shared by processes, such as records in a file, no two or more processes can simultaneously access these information structures. This limit is imposed in order to guarantee that the structures can be inspected or updated without interference from other processes, even if there are different records available for multiple processes (Holt, 1972). On the other hand, if there is only one resource, then users definitely must operate sequentially in using this resource. If the number of the resources is more than one but less than the number of users, then users may act in a degree of parallelism depending upon the ratio of the number of the resources to the number of the users and how the environment is arranged for the resources, ...