Title: AI-PROPELLER: Leveraging AlphaEvolve for Warehouse-Scale Interprocedural Code Layout Optimization

Abstract:

While post-link optimizers (PLOs) like Propeller and BOLT have proven that precise, profile-guided code layout can yield substantial performance benefits for highly optimized binaries, their capabilities are currently confined to intraprocedural techniques. As a result, the broader potential of interprocedural layout remains largely unexplored. Historically, interprocedural code layout has been hindered by a combinatorially intractable search space and the complexity of modeling call-return semantics, leaving the practical performance gains of fine-grained interprocedural layout unverified.

To address these challenges, AI-PROPELLER employs Magellan, an agentic workflow designed to evolve the compiler heuristics within Propeller into a fine-grained interprocedural optimizer. This process also involves fine-tuning the resulting policy’s hyperparameters. To guarantee high-fidelity results, the system abandons approximate static cost models. Instead, the agentic workflow generates multiple layout variants that are executed on real hardware to capture actual performance counters, thereby delivering a precise reward signal for the evolutionary loop.

Evaluations of AI-PROPELLER across several benchmarks, including large warehouse-scale applications, demonstrate performance improvements ranging from 0.23% to 1.6%. These gains were achieved on binaries already optimized with state-of-the-art FDO and PLO technologies, marking a significant achievement for real-world software. Notably, this represents the first instance in which large warehouse-scale applications within industrial environments have been optimized using fine-grained interprocedural code layout.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC