Split overlapping_{inherent,trait}_impls This yielded some perf improvement for me. Reduces some calls to `impl_trait_header` query. But I think the llvm optimization is more relevant.