Evolutionary Compilation for Object Code Compatibility and Performance

SATHAYE, SUMEDH WASUDEO Evolutionary Compilation for Object Code Compatibility and Performance. (Under the direction of Prof. Thomas M. Conte) Optimized code for high-performance systems can be generated using (1) the knowledge of processor microarchitecture, and, (2) the use of representative program pro les for pro le-based optimization. Decisions made during compilation using such information may turn out to be erroneous when the code is run on a di erent machine model, or when program behavior changes due to a change in program inputs. The topic of this thesis, Evolutionary Compilation, is suggested as a solution. It helps the code adapt to such variations and thus preserves program performance. The evolutionary compiler (EC) runs as a part of the operating system. Program behavior is monitored using unobtrusive, fast, hardware-based pro ling techniques. A change in the program pro le or a di erence in the machine model assumed during compilation causes EC to be invoked on the object-code. One kind of variation in the machine model assumptions is the inter-generation compatibility problem in statically scheduled VLIW architectures. Dynamic Rescheduling (DR) is the technique used to reschedule code pages at rst-time page faults to overcome this problem. Any changes in code size while using such a scheme are avoided by using the TINKER encoding. Caching techniques to reduce the overhead incurred due to rescheduling are presented and studied. In order to adapt to variations in pro le information, recompilation requires more information than that available in the object code. Such information, e.g. live-outs across side-exits, is stored in the form of a non-loadable objectle annotation segment in the binary. The evolutionary compilation framework, the TINKER encoding, rescheduling algorithms and the properties of regions over which rescheduling is performed are discussed in detail. Evolutionary Compilation for Object Code Compatibility and Performance by Sumedh Wasudeo Sathaye A dissertation submitted to the Graduate Faculty of North Carolina State University in partial ful llment of the requirements for the Degree of Doctor of Philosophy Computer Engineering