Ictíneo: a Tool for Instruction Level Parallelism Research

Current Fortran optimizing compilers often include source to source transformations for automatic parallelization or vectorization of loops. Lower level optimizations, such as those that aim to exploit ILP, are performed at later stages at the assembly language level and do not profit from information available at the source code level, such as array subscripts for data dependence analysis. Low-level optimizations could generate better code if this high-level information were available. In this paper we describe a framework in which low-level code close to machine language maintains high-level constructs such as loops, source level variable names, etc. This allows for low and high-level optimizations to be performed in the same framework, giving way to more optimized code. Our low-level representation is a single-assignment virtual assembly language. It can optionally be refined to an even lower level form in which address calculations for memory accesses are made explicit. We have imp..