Note that Stree and DTM used a single FastStringBuffer as a string pool, by recording start and length indices within this single buffer. This minimizes heap overhead, but of course requires more work when retrieving the data.
FastStringBuffer operates as a "chunked buffer". Doing so reduces the need to recopy existing information when an append exceeds the space available; we just allocate another chunk and flow across to it. (The array of chunks may need to grow, admittedly, but that's a much smaller object.) Some excess recopying may arise when we extract Strings which cross chunk boundaries; larger chunks make that less frequent.
The size values are parameterized, to allow tuning this code. In theory, Result Tree Fragments might want to be tuned differently from the main document's text.
%REVIEW% An experiment in self-tuning is included in the code (using nested FastStringBuffers to achieve variation in chunk sizes), but this implementation has proven to be problematic when data may be being copied from the FSB into itself. We should either re-architect that to make this safe (if possible) or remove that code and clean up for performance/maintainability reasons.
| |