FAIR: Flow Type-Aware Pre-Training of Compiler Intermediate Representations

TL;DR

FAIR is a novel pre-training model for compiler IRs that effectively captures flow types and long-distance dependencies, improving performance on code-related tasks.

Contribution

Introduces a flow type-aware pre-training approach for IRs using a graph transformer and new tasks to better understand IR semantics.

Findings

Achieves state-of-the-art results on four downstream tasks.

Effectively models flow types and long-range dependencies.

Addresses over-smoothing and over-squashing in IR graph learning.

Abstract

While the majority of existing pre-trained models from code learn source code features such as code tokens and abstract syntax trees, there are some other works that focus on learning from compiler intermediate representations (IRs). Existing IR-based models typically utilize IR features such as instructions, control and data flow graphs (CDFGs), call graphs, etc. However, these methods confuse variable nodes and instruction nodes in a CDFG and fail to distinguish different types of flows, and the neural networks they use fail to capture long-distance dependencies and have over-smoothing and over-squashing problems. To address these weaknesses, we propose FAIR, a Flow type-Aware pre-trained model for IR that involves employing (1) a novel input representation of IR programs; (2) Graph Transformer to address over-smoothing, over-squashing and long-dependencies problems; and (3) five pre-training tasks that we specifically propose to enable FAIR to learn the semantics of IR tokens, flow type information, and the overall representation of IR. Experimental results show that FAIR can achieve state-of-the-art results on four code-related downstream tasks.