Supporting SELL for High-Performance Computing

We briefly introduce the notion of Semantically Enhanced Library Languages, SELL, as a practical and economical alternative to special-purpose programming languages for high-performance computing. Then we describe the Pivot infrastructure for program analysis and transformation that is our main tool for supporting SELL. Finally, we outline how the IPR (The Pivot’s Internal Program Representation) can be used to represent central notions of high-performance computing, such as parallelizable array operations. Our focus is on a broad exposition of ideas, rather than technical details. 1 Languages and libraries For ease of programming, portability, and acceptable performance, we design and implement special-purpose programming languages for high-performance computing [15]. Alternatively, we can use a Semantically Enhanced Library Language. A SELL is a language created by extending a programming language (usually a popular general-purpose programming language) with a library providing the desired added functionality and then using a tool to provide the desired semantic guarantees needed to reach a goal (often a higher level semantics, absence of certain kinds of errors, or library-specific optimizations) [12]. This paper focuses on a tool, The Pivot, being developed to support SELLs in ISO C++ [11, 5] and its application to High-Performance Computing. 2 A brief overview of the Pivot The Pivot is a general framework for the analysis and transformation of C++ programs. It is designed to handle the complete ISO C++, especially more advanced uses of templates and including some proposed C++0x features. It is compiler independent. The central part of the Pivot is a fully typed abstract syntax tree called IPR (Internal Program Representation). There are lots of (more than 20) tools for static analysis and transformation of C++ programs, e.g. [7, 2, 8, 6]. However, few — if any — handle all of ISO 1 This is the "cut" or "abbreviated" version of this paper. For a full version, see http://www.research.att.com/ bs/papers.html Standard C++, most are specialized to particular forms of analysis or transformation, and few will work well in combination with other tools. We are particularly interested in advanced uses of templates as used in generic programming, template meta-programming, and experimental uses of libraries as the basis of language extension. For that, we need a representation that deals with types as first-class citizens and allows analysis and transformation based on their properties. In the C++ community, this is discussed under the heading of concepts and is likely to receive significant language support in the next ISO C++ standard (C++0x) [13, 9, 14, 3]. We use the word concept to a designate collection of properties that describes usage of values and types. From the point of view of support for HPC — and for the provision of special-purpose facilities in general — a concept can be seen as a way of specifying new types with associated semantics without the modification of compilers or new syntax. That done, the SELL approach then uses the concepts as a hook for semantic properties beyond what C++ offers. 2.1 System organization To get IPR from a program, we need a compiler. Only a compiler “knows” enough about a C++ program to represent it completely with syntactic and type information in a useful form. In particular, a simple parser doesn’t understand types well enough to do a credible general job. We interface to a compiler in some appropriate and minimally invasive fashion. A compiler-specific IPR generator produces IPR on a per-translation-unit basis. Applications interface to “code” through the IPR interface. So as not to run the compiler all the time and to be able to store and merge translation units without compiler intervention, we can produce a persistent form of IPR called XPR (eXternal Program Representation).