https://github.com/nikic approved this pull request. LGTM. I assume we don't need an argument guard here like in the align case -- presumably deref assumptions are used in contexts where there is no argument deref information. https://github.com/llvm/llvm-project/pull/202946