fseek() in File::append() was invalidating the file buffer for each call
to append(). Better handling improves "import" performance test by 9%,
45% in "commit", on a system with a spinning disk.
This performance problem affects all operations where
{pending,completed}.data are rewritten. During normal operation a
garbage collection can be enough to trigger it. On storage with high
latency, e.g. networked, this previously took 20 seconds and more.
Reduce the amount of copies necessary for TDB2::gc() by moving the GC
processing to the Task object source - TF2::load_tasks().
This entangles TDB2 and TF2 more than previously, but leads to huge
performance benefits:
- "next" performance test down 21%
- "list" performance test down 11%
- "all" performance test down 4%
- "export" performance test down 9%
The "gc" measurement is down 96% for all performance tests. This is a
result of moving the actual processing into TF2::load_gc() and not
measuring the time taken by that function as "gc" time.
- Make the Task object's interface more explicit by removing the
std::map inheritance.
- Using this more explicit interface, remove unneeded ctors in order to
allow the compiler to "Do The Right Thing"(tm).
This leads to a performance improvement of 12% in the "add"
performance test, and 7% for "import".
- Improves "task import" performance test on test machine further from
13s to 1.9s.
- Active only during "task import". Might be worth using during other
mass lookups like "task sync", or could even be a good default.
- Switched Nibbler::getUUID to Nibbler::getPartialUUID, which caused partial
UUID mathcing to fail sometimes.
- Changed precedence to search for UUID before ID, which solves the numeric
UUID problem.
- Use variadic templates, with a slight twist for properly supporting
Taskwarrior's l10n strings.
- Adjust code that relied on implicit type conversion which led to
problems with template instantiation.
- Move some functionality out of format() in order to keep it generic.
