Flattening Basic Blocks

Abstract

Preaccumulation of Jacobians of low-level code sections is beneficial in certain automatic differentiation application scenarios. Cross-country vertex, edge, or face elimination can produce highly efficient Jacobian preaccumulation code but requires the code section's computational graph to be known. Such graphs can easily be derived for straight-line codes with scalar variables. Practical codes in languages such as Fortran, C, and C++ with array indexing, pointers, and references introduce ambiguities. Constructing unambiguous computational graphs is necessary to realize the full potential of cross-country elimination. We give a practical solution and investigate theoretical options for graph construction.