LLVM-MCA Performance Analysis: Pitfalls of Vectorization Optimization

The author encountered a performance degradation issue when vectorizing code using ARM NEON. The initial code used five load instructions (5L), while the optimized version used two loads and three extensions (2L3E) to reduce memory accesses. Surprisingly, the 2L3E version was slower. Using LLVM-MCA for performance analysis revealed that 2L3E caused bottlenecks in CPU execution units, unbalanced resource utilization, and stronger instruction dependencies, leading to performance regression. The 5L version performed better due to its more balanced resource usage and independent load instructions. This case study highlights how seemingly sound optimizations can result in performance degradation if CPU resource contention and instruction dependencies aren't considered; LLVM-MCA proves a valuable tool for analyzing such issues.