Code Size Optimization for Embedded Processors

TL;DR: This thesis develops the Value State Dependence Graph as a powerful intermediate form and shows how procedural abstraction can be advantageously applied to the VSDG, and presents a method for using these instructions to reduce code size by provisionally combining loads and stores before code generation.

Abstract: This thesis studies the problem of reducing code size produced by an optimizing compiler. We develop the Value State Dependence Graph (VSDG) as a powerful intermediate form. Nodes represent computation, and edges represent value (data) and state (control) dependencies between nodes. The edges specify a partial ordering of the nodes—sufficient ordering to maintain the I/O semantics of the source program, while allowing optimizers greater freedom to move nodes within the program to achieve better (smaller) code. Optimizations, both classical and new, transform the graph through graph rewriting rules prior to code generation. Additional (semantically inessential) state edges are added to transform the VSDG into a Control Flow Graph, from which target code is generated. We show how procedural abstraction can be advantageously applied to the VSDG. Graph patterns are extracted from a program’s VSDG. We then select repeated patterns giving the greatest size reduction, generate new functions from these patterns, and replace all occurrences of the patterns in the original VSDG with calls to these abstracted functions. Several embedded processors have loadand store-multiple instructions, representing several loads (or stores) as one instruction. We present a method, benefiting from the VSDG form, for using these instructions to reduce code size by provisionally combining loads and stores before code generation. The final contribution of this thesis is a combined register allocation and code motion (RACM) algorithm. We show that our RACM algorithm formulates these two previously antagonistic phases as one combined pass over the VSDG, transforming the graph (moving or cloning nodes, or spilling edges) to fit within the physical resources of the target processor. We have implemented our ideas within a prototype C compiler and suite of VSDG optimizers, generating code for the Thumb 32-bit processor. Our results show improvements for each optimization and that we can achieve code sizes comparable to, and in some cases better than, that produced by commercial compilers with significant investments in optimization technology.