Presenter Information

Reuven MuellerFollow

Location

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

Document Type

Event

Start Date

24-11-2025 4:00 PM

Description

This study investigates whether modern state-of-the-art Large Language Models (LLMs) can interpret, optimize, and translate low-level intermediate representations (IR) used in compilers and binary translation software. We evaluate LLM performance on LLVM IR across three tasks: 1) interpreting the underlying algorithmic behavior, 2) identifying missed optimization opportunities, and generating improved IR variants, 3) Translating IR between AArch64 and x86-64 targets. To ensure correctness, all LLM generated IR is checked using a validation pipeline that verifies its syntax, structural correctness, compiles it into machine code, and executes it on randomized test data. Early results show that LLMs can perform non-trivial IR transformations on top of existing LLVM-O3 optimizations, highlighting their potential role in future compiler and binary translation workflows.

Share

COinS
 
Nov 24th, 4:00 PM

GRP-0197 Evaluating the Ability of LLMs to Interpret, Optimize and Translate LLVM IR

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

This study investigates whether modern state-of-the-art Large Language Models (LLMs) can interpret, optimize, and translate low-level intermediate representations (IR) used in compilers and binary translation software. We evaluate LLM performance on LLVM IR across three tasks: 1) interpreting the underlying algorithmic behavior, 2) identifying missed optimization opportunities, and generating improved IR variants, 3) Translating IR between AArch64 and x86-64 targets. To ensure correctness, all LLM generated IR is checked using a validation pipeline that verifies its syntax, structural correctness, compiles it into machine code, and executes it on randomized test data. Early results show that LLMs can perform non-trivial IR transformations on top of existing LLVM-O3 optimizations, highlighting their potential role in future compiler and binary translation workflows.