llvm-project

afe4cffc - DAG: Handle load in SimplifyDemandedVectorElts

DAG: Handle load in SimplifyDemandedVectorElts This improves some AMDGPU cases and avoids future regressions. The combiner likes to form shuffles for cases where an extract_vector_elt would do perfectly well, and this recovers some of the regressions from losing load narrowing. AMDGPU, Arch64 and RISCV test changes look broadly better. Other targets have some improvements, but mostly regressions. In particular X86 looks much worse. I'm guessing this is because it's shouldReduceLoadWidth is wrong. I mostly just regenerated the checks. I assume some set of them should switch to use volatile loads to defeat the optimization.