Tracking down a 25% Regression on LLVM RISC-V
Summary
This post analyzes a 25% regression in LLVM's RISCV backend caused by a recent InstCombine patch that improved isKnownExactCastIntToFP but inadvertently blocked a prior fpext-to-fptrunc optimization. The author investigates with llvm-mca, identifies the root cause, implements a patch to enable earlier narrowing, and demonstrates a ~25% performance improvement after the fix. Links to related GitHub issues and PRs are provided.