Limitations of Graham-Glanville style code generation

In recent years, the Graham-Glanville approach to code generation has become an increasingly important influence on new compiler writing projects. Initial enthusiasm prompted by the early research results has now fueled projects with more practical goals and requirements. This paper reports on the early results of a long-term project that is applying the Graham-Glanville method to the creation of an industrial-quality multi-language, multi-target compiler system. Much of the resulting code generation system deals with matters that are not fully handled by the Graham-Glanville method. We present a derailed list of the limitations encountered and, for each, the extension or modification that was invented to deal with it.Our conclusion is that the Graham-Glanville method does not always fulfill its early promise of generating locally-optimal code. Production-quality retargetable compiler generation still requires clever organization and many sophisticated algorithms to achieve this goal in the presence of register allocation and cost analysis.