llvm-project

070b3e99 - [InlineSpiller][AMDGPU] Implement subreg reload during RA spill

[InlineSpiller][AMDGPU] Implement subreg reload during RA spill Currently, when a virtual register is partially used, the entire tuple is restored from the spilled location, even if only a subset of its sub-registers is needed. This patch introduces support for partial reloads by analyzing actual register usage and restoring only the required sub-registers. This improvement enhances register allocation efficiency, particularly for cases involving tuple virtual registers. For AMDGPU, this change brings considerable improvements in workloads that involve matrix operations, large vectors, and complex control flows.