Design and Implementation of an Automatic Vectorizing PASCAL Compiler

该编译器由五个阶段组成，由五个阶段组成。(1)阶段1：语法和语义分析，(2)阶段2：向量化的预处理，(3)阶段3：向量化和并行化，(4)阶段4：向量化的后处理，(5)阶段5：目标代码生成。在第一阶段，该编译器将给定的源程序转换为专门设计的中间代码。源语言是Pascal8000的子集。在阶段2中，编译器优化代码并检测别名。选择优化技术的方式是：(A)增加可矢量化部分和(B)降低分析成本。第三阶段由以下模块组成：数据流分析、控制流分析、支配者分析、循环检测、控制依赖分析、数据依赖分析、中间代码重新排序、控制结构转换、向量化和并行化。在阶段4中，编译器主要优化标量循环，在阶段5中，它生成机器码。该编译器的目标机器是HITAC S-810，它的机器语言使我们能够轻松地表达并行性。该编译器的第一个版本的大小约为32,000行。在该编译器中实现的算法可以向量化包含IF-THEN-ELSE结构和/或退出的整个嵌套循环，从而突破了传统的向量化技术。

The compiler, which has resulted from this research project, is made of five phases. (1) Phase 1 : syntax and semantics analysis, (2) Phase 2 : pre-processing for vectorization, (3) Phase 3 : vectorizing and parallelizing, (4) Phase 4 : post-processing for vectorization, (5) Phase 5 : object-code generation. In phase 1, this compiler converts a given source program to specially designed intermediate codes. The source language is a subset of PASCAL 8000. In phase 2, the compiler optimizes the code and detects aliases. The optimizing techniques are chosen in such a way that they can : (A) increase vectorizable parts and (B) reduce costs of analyses. The phase 3 is made up from those modules given to : data-flow-analysis, control-flow analysis, dominator analysis, loop detection, control-dopendence analysis, data-dependence analysis, re-ordering of the intermediate codes, conversion of control structures, vectorization and parallelization. In phase 4, the compiler mainly optimizes scalar loops, and in phase 5, it generates machine codes. The target machine of this compiler is HITAC S-810, whose machine language allows us to express parallelism with ease. The size of the first version of this compiler is about 32,000 lines. The algorithm implemented in this compiler can vectorize the whole nested loops which may include If-then-else structures and/or exits from them, thus giving a breakthrough to the conventional vectorization technique.