Fully Integrated Quantum Method for Classical Register Allocation in LLVM

Quantum computing devices are being installed alongside supercomputing clusters to serve as hardware accel-erators. This new type of architecture will require an integrated hybrid software stack. With this goal in mind, we have developed a fully integrated hybrid quantum-classical method in the hope of improving Register Allocation in the classical LLVM compiler. We propose a hybrid variational optimization algorithm for the PBQP formulation of Register Allocation. We implemented this algorithm in C++ inside LLVM using the NVIDIA CUDA-<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$Q$</tex> framework. The performance of the method is evaluated using NVIDIA CUDA-Q noiseless emulators and shows promising results while still needing further optimizations. Our work constitutes a demonstration of an end-to-end tight integration of a quantum subroutine inside an existing classical codebase of interest with potentially interesting performance in fault-tolerant hardware.