[Git][ghc/ghc][master] Narrow before optimising MUL/DIV/REM into shifts