2026 , Volume 31, № 5, p.102-116

This paper presents a tool for automatic analysis of inaccuracies in function in-lining and software data pre-fetching with hardware support optimizations in the LCC compiler for the “Elbrus” VLIW architecture. The proposed approach combines static and dynamic methods to identify cases where compiler heuristics fail due to the lack of profiling information. The tool automatically incorporates program code to collect statistical information, analyzes optimization behavior, and generates detailed reports with recomendations for the software developer. These reports indicate where compiler hints should be manually added to improve performance. The tool consists of two main components: a compiler module that incorporates a program and a runtime library that gathers and analyzes statistical data. It provides guidance on when to apply or disable optimizations. This approach mitigates typical drawbacks of profile-guided optimization (PGO), such as the need to recompile program after codebase update and distortion of profiling data during later compilation stages.

Experimental evaluation was performed on the SPEC CPU 2006 and 2017 benchmarks using the Elbrus v5 processor. The results demonstrate that applying the recommended hints leads to measurable performance improvements of up to 87.9 % based on the geometric mean (gmean) for inline optimization and up to 3.8 % gmean for software pre-fetch optimization. The proposed tool significantly simplifies the tuning of compiler optimizations and contributes to more efficient software development for the Elbrus platform.