Quantitative global dataflow analysis on virtual instruction set simulators for hardware/software co-design

Abstract

One of the main challenges in system design whether for high performance computing or in embedded systems is to partition software for target architectures like multi-core, heterogeneous, or even hardware/software co-design systems. Several compiler techniques handle partitioning and related problems by using static analysis and therefor have no means to capture the global data flow in quantity and its dynamics which is essential for extracting tasks or exploiting coarse grained parallelism. We present a novel solution for capturing and analyzing an applicationpsilas quantitative data flow in this paper. The core part is the LLILA (Low Level Intermediate Language Analyzer) tool set, which automatically generates and augments self-profiling instruction set simulators from assembly level descriptions for a virtual machine. During run-time of the augmented program several properties (frequency, quantity and locality reflecting inter-procedural communication) of data exchange are captured at instruction level and as a consequence in the highest possible degree of accuracy.