Reducing constant loads for ARM executable files

Abstract

For several benefits, unignorable amount of read-only data has been embedded in the text sections of an executable file though there are separate data sections in the file. This, however, becomes serious hindrance when we come to the issues of directly manipulating an executable file for analyzing, profiling or optimizing it. It also incurs some overhead in code size and performance. With all this drawbacks, in a fixed-length instruction set architecture, it goes beyond our choice to embed data into text sections due to the limitation on the bitfield size of the immediate operands in instructions. More specifically, whenever we access data segment we must load the address of the target data in advance from near the current instruction. This makes embedding the address values into the text sections inevitable in such architecture. In this thesis we show that a minor addition in a fixed-length instruction set architecture can cover most of the cases of necessity for the big constants like data addresses without embedding them into the text sections. In addition, it results in code size reduction and improvement in performance. The discussion and experiments are targeted to the ARM instruction set architecture which is an instance of fixed length instruction set architecture and the most widespread one in embedded computing area. The experiments are done with self-constructed binary engineering tool which does necessary analyzing, profiling and rewriting of the ARM executable files and the Simplescalar-ARM simulator.