Add a very simple optimalisation

Piece() requires a hash lookup in an lru_cache, and is thus sort-of
expensive.  This patch reduces the amount of calls to Piece()