Abstract
We introduce shape cliques, a simple way to organize a subset of the arrays appearing in an array-language-based application into sets of identically shaped arrays - shape cliques - and show how a compiler can analyze an application to infer membership in those cliques. We describe an algorithm for performing shape clique inference (SCI), and demonstrate that shape cliques can improve the performance of generated code, by permitting extension of an optimization for removal of run-time checks, and by extending the set of arrays to which optimizations, such as Index Vector Elimination (IVE), can be applied. Implementation of SCI in the APEX APL compiler permitted removal of 25% of run-time checks remaining on 156 benchmarks remaining after other compiler optimizations had eliminated 72% of the 1251 checks present in the original code. In the SAC compiler, IVE using SCI produced typical speedups of 2-14X on benchmarks operating on arrays of non-fixed rank and shape, compared to the operation of IVE in a non-SCI environment. Shape clique inference data can be exploited to allow certain other optimizations, such as loop fusion and withloop folding, to be performed on arrays of statically unknown shape and rank, with the potential for significant reductions in execution time.
Publisher
Association for Computing Machinery (ACM)
Reference18 articles.
1. International Standards Organization: International Standard for Programming Language APL. ISO N8485 edn. (1984). International Standards Organization: International Standard for Programming Language APL. ISO N8485 edn. (1984).
2. Hui R. K. Iverson K. E.: J Dictionary. (1998). Hui R. K. Iverson K. E.: J Dictionary. (1998).
3. Single Assignment C: efficient support for high-level array operations in a functional setting
4. SAC—A Functional Array Language for Efficient Multi-threaded Execution
5. Cann D. C.: The optimizing SISAL compiler: Version 12.0. Technical Report UCRL-MA-110080 Lawrence Livermore National Laboratory (1992). Cann D. C.: The optimizing SISAL compiler: Version 12.0. Technical Report UCRL-MA-110080 Lawrence Livermore National Laboratory (1992).
Cited by
2 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献
1. Just-in-time Length Specialization of Dynamic Vector Code;Proceedings of ACM SIGPLAN International Workshop on Libraries, Languages, and Compilers for Array Programming;2014-06-09
2. From Contracts Towards Dependent Types: Proofs by Partial Evaluation;Implementation and Application of Functional Languages