Currently, the arith shift ops are documented as returning poison of the shift amount is greater than the bitwidth - this is in line with the LLVM Support lib which defines the result of shifting by an amount equal to the bitwdith as zero. However, the LLVM instructions return poison even when the shift amount is equal to the bitwidth. I would suggest adapting the poison behavior from the LLVM instructions because otherwise we have to catch this special case in the lowering which seems suboptimal for fast bitshift operations.
I have created a PR which changes the documentation accordingly but leaves the folders alone for now, since producing Poison in the folders is not yet implemented.