Development of an Automatic Vectorizing Intelligent Compiler

在这个财政年度，一切都按照提交的书面申请的计划进行。更具体地说，采取了以下步骤：1 .控制依赖模块的编码、测试与调试；2 .数据依赖模块的编码、测试与调试；3 .部分矢量化决策模块的编码、测试与调试。4 .矢量中间码生成模块的编码、测试与调试；5 .通过各模块的测试和调试；生成的目标代码的性能评估。在上述控制依赖模块中，实现了一种新的算法，该算法已发表（参见参考文献）。该算法可以分析任何嵌套循环，包括任何复杂的控制结构，如多入口循环、有出口循环等，并开发了一种新的数据依赖性分析算法。该算法是通过改进和结合原有的两种相同目的的算法而得到的。如果有必要，它会使用启发式方法自动为目标嵌套循环选择两种方法中最合适的方法。部分向量化模块可以对字符串操作和关系操作进行新的向量化。这一尝试可以说是为矢量超级计算机开辟新前景的第一步。此外，该编译器还具有更强大的检测和生成向量宏操作的功能。

In this fiscal year, everything went according to the plan of the submitted written application. To be more concrete, these steps was carried out: 1. Coding, testing and debugging of the control dependence module, 2. Coding, testing and debugging of the data dependence module, 3. Coding, testing and debugging of the partial vectorization decision module, 4. Coding, testing and debugging of the vector intermediate code generation module, 5. Testing and debugging through all modules, 6. Performance evaluation of generated object codes. In the above control dependence module, a new algorithm has been implemented, which had been published (see the reference). This algorithm can analyze any nested loops including any complicated control structures, for example, multi-entry loops, loops with exits, etc..A new algorithm for data dependence analysis, has been also developed. This algorithm is made by improving and combining our old two methods of the same purpose. It automatically selects the fittest one of the two methods for target nested loops, if necessary, using heuristics.The partial vectorization modules can newly vectorize string manipulations and relational operations. This attempt may be said to be the first step to open up a new vista for vector supercomputers. Moreover, this compiler has also more powerful feature of detecting and generating vector macro operations.