Interprocedural compilation of Fortran D for MIMD distributed-memory machines

Abstract

Algorithms exist for compiling Fortran D for MIMD distributed-memory machines, but are significantly r e stricted in the presence of procedure calls. This paper presents interprocedural analysis, optimization, and code generation algorithms for Fortran D that limit compilation to only one pass over each procedure. This is accomplished by collecting summary information after edits, then compiling procedures in reverse topological order to propagate necessary information. Delaying instantiation of the computation partition, communication, and dynamic data decomposition is key to enabling interprocedural o p timization. Recompilation analysis preserves the benefits of separate compilation. Empirical results show that interprocedural optimization is crucial in achieving acceptable performance for a common application.

A global communication optimization technique based on data-flow analysis and linear algebra

“…It should be noted that there are several inter-procedural communication optimization algorithms (e.g., [25], [26], [15]) with different degrees of sophistication, and the detailed analysis of communication optimization across procedure boundaries is beyond the scope of this paper. However, we believe that for most The message vectorized program is obtained using the popular vectorization approach based on dependence-analysis.…”

Compiling dynamic mappings with array copies

“…Such optimizations to avoid useless remapping communications, especially interprocedural ones, have been discussed 11,18]. It is shown 11] that the best approach to handle subroutine calls is that callers must comply to callee requirements.…”

Interprocedural Compilation of Fortran D

“…In addition, we have also developed techniques for optimizing dynamic data decomposition and performing recompilation analysis using our interprocedural compilation framework [15].…”

“…Each dimension is distributed in a BLOCK, CYCLIC, or BLOCK -CYCLIC manner; the symbol '' : '' marks dimensions that are not distributed. Because the alignment and distribution statements are executable, dynamic data decomposition is possible; compiler analysis is required to determine whether data layout for each array may be changed at run time [15]. As in HPF, each array is implicitly aligned with a default decomposition.…”

“…Otherwise the compiler may choose to either utilize buffers or go back and modify array declarations in those procedures. If analysis is insufficient, overlap alternatives are also possible [15]. The algorithm for calculating overlaps is described in Fig.…”