A comparison of three approaches to language, compiler, and library support for multidimensional arrays in Java

• Authors: José E. Moreira;Samuel P. Midkiff;Manish Gupta
• Venue: Proceedings of the 2001 joint ACM-ISCOPE conference on Java Grande
• Year: 2001

The lack of direct support for multidimensional arrays in Java™ has been recognized as a major deficiency in the language's applicability to numerical computing. The typical approach to adding multidimensional arrays to Java has been through class libraries that implement these structures. It has been shown that the class library approach can achieve very high-performance for numerical computing, through the use of compiler techniques and efficient implementations of aggregate array operations. Because of the inconvenience of accessing array elements through method invocations, it is advocated by many that class libraries for multidimensional arrays should be combined with new language syntax to facilitate manipulation of those multidimensional arrays. Another approach that has been discussed in the literature is that of relying exclusively on the JVM to recognize those arrays of arrays that are being used to stimulate multidimensional arrays. This approach can also deliver good performance, but it does not improve the existing interfaces for numerical computing. There is yet a third approach: extending the Java language with new syntactic constructs for multidimensional arrays and directly compiling those constructs to bytecode. The new constructs provide a more convenient interface for numerical computing, without requiring a matching class library. This paper is a comparative discussion or the three approaches to adding multidimensional arrays to Java mentioned above. We present a description of the three approaches, listing the pros and cons of each. We give a more detailed description of the third approach — language constructs translated to bytecode — as it is a new contribution. We compare each of the approaches with regards to functionality, impact on the language and virtual machine specification, implementation efforts, and typical achievable performance. We show that the best choice depends on the relative importance attached to the above metrics.