Detection of inaccuracies in some specialized loop optimizations in the LCC compiler for “Elbrus” architecture

A. V. Ermolitckii^a, D. N. Levchenko^ba, M. I. Neiman-Zade<sup>ba

a</sup> AO "MCST"
^b Moscow Institute of Physics and Technology (National Research University)

Abstract: This work is devoted to the development of mixed methods for analyzing inaccuracies in compiler optimizations. The development of these methods is important for “Elbrus” microprocessors with a very long instruction word (VLIW) because of static scheduling. The article analyzes existing approaches to identifying inaccuracies in optimizations and highlights their disadvantages. In this article a method is developed for detecting inaccuracies in the work of two important for VLIW optimizations: software pipeline with hardware support (overlap) and optimization of moving unlikely executed code into new outer loop (nesting). The method is implemented by instrumenting loops in the user program and obtaining static information about loops from the compiler. The method was checked on SPEC CPU 2006 and 2017 rate suites on a computer with an Elbrus-8S processor and has proven its effectiveness. The method allows to achieve a speedup of 70.7% on test 523.xalancbmk with placing hints for overlap optimization and 4.71% on test 520.omnetpp with placing hints for nesting optimization. Tests are done in base mode without profile information.

Keywords: compiler optimizations; software pipeline; code instrumentation; VLIW.

DOI: 10.15514/ISPRAS-2025-37(4)-3