Tracking down a 25% Regression on LLVM RISC-V

This is a really detailed post - it does feel a little repetitive in places, and I think it highlights the frustrating fragility of pattern matching based optimizations.