Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Smeet Chheda and Barbara Chapman (Stony Brook University)
Abstract: Branching hurts performance on GPU architectures, so unmerging control flow inside a loop looks like a poor optimization. Prior work showed otherwise: unmerging recovers information that a control-flow merge destroys, and pairing it with unrolling helps further.
The static heuristic developed in prior work performs poorly on many applications.
On an NVIDIA A100 across HeCBench, it loses to doing nothing: -4.9% mean improvement where +10% was available, and 23.9% of loops made slower. Its rule cannot express a quarter of the best decisions.
We port the transform to a newer LLVM and build a reinforcement learning pipeline that decides per loop whether to unmerge and which unroll factor to use, using measured kernel time as feedback. Two agents fit their training loops but neither transfers, because profitability follows the application rather than the loop. Adapting on two loops of the target cuts regressions to 9.5% and reaches +4.0% at best.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing