On Techniques in Vectorizing Compilers and Optimizing Program Transformations for Supercomputers

概要

論文の詳細を見る
Techniques in vectorizing compilers and optimizing program transformations for supercomputers are de-scribed. First we give a brief overview of programming languages for supercomputers, especially a historical review of languages for early vector processors in Japan in relation to languages for supercomputers in the pre-sent day. Then we give conditions for vectorization of basic loop programs. The order of vector instructions does not correspond, in general, to that of source statements in the program in order to attain high vectorization ratio. This can be considered as the reordering of statements in the source program. For dedugging a program, however, there are cases where it is convenient to force a compiler to keep the order of instructions with that of source statements, even if vectorization ratio is affected to some extent. We give conditions for vectorization with/without reordering of statements, in a easily verifiable way. Finally we describe an analysis of loop unrolling techniques for nested loops. This programming technique is one of few general and effective optimizing transformations. Reasons why the technique is effective depend on architecture of computers and compilers used. It is shown that the situation can be well understood using a rather simple machine model.
一般社団法人情報処理学会の論文
1988-03-31