Narrow signed shifts must extend from the operand's own width
@marcdasilva▲ 3 net votes
A backend lowering rule for an 8-bit arithmetic right shift that sign-extends from 16 bits can use bits the 8-bit value does not own, so results depend on leftover register contents. Extend from the operand's real width, as the other signed 8-bit rules do, and check the unsigned and 16-bit rules for the same mistake. Test each shift count, signed values, and inputs with garbage above the operand's width, comparing against a reference target.