This is highly experimental. It's defers most checkpointing
to a separate thread to see if we get better throughput and
less choppiness when applying effects.
This reenables synchronous mode by default. However, for processing
where a power cut or crash can be tolerated, synchronous and journaling
are disabled. Once the processing is complete, they are reenabled.
Types of processing that are like this are "Save As", "Backup Project",
and "Vacuuming". They all write to a separate project file while
running, so the real project file is safe.
Unfortunately, effects are back to be slow and sluggish.
I believe I've address all of the weird file corruption issues and
I'll continue to continue testing for these.
Optimizes a couple of sample block copy loops by only preparing
the statement once outside the loop.
The connection configuration ensure that all connections use the same
settings...assuming you remember to configure it after opening. :-)
The possibly controversial setting is the "PRAGMA synchronous = off"
This time it has the potential to produce much smaller
output files since it ONLY copies the active blocks and
not all of the blocks related to undo history. This is
done for "Save As" and "Backup Project". Normal save
can't take advantage of this, but then it really doesn't
need it as it has to depend on vacuuming.
The vacuuming at close has been adjusted to utilize CopyTo()
so it should produce similarly small files as long as the
vacuuming happens when the project is definitely closing.
This time it has the potential to produce much smaller
output files since it ONLY copies the active blocks and
not all of the blocks related to undo history.
* Fix the reading of autosave files...
... problem was in recreating strings from buffers, but copying too many because
null terminators were lacking.
* Autosave during recording backs up all tracks correctly...
... whether to new track, or appending; and it doesn't lose the other tracks
besides the recording.
It is also unnecessary when just starting to record, so remove one call.
It now uses VACUUM INTO instead of the SQLite backup API
in hopes that the copies will be smaller. And VACUUM INTO
is "supposed" to be faster, but time will tell. It's easy
to put the backup API usage back in.
This also fixes a bit I missed with redoing the orphan block
handling that was reported by Paul.
And finally, it renames the AutoRecovery.cpp/.h files and AutoSaveFile
class to ProjectSerializer since the AutoSaveFile class is being
used for regular project documents now and it doesn't write to a
file anymore.
If anyone has a better idea for a name other than ProjectSerializer
feel free to change it. I hate naming things.
* ProjectFileIO::SaveProject() won't close original db until done...
... So on the failure path, don't risk closing the original too early, failing
with the new database, and then failing to reopen the original.
Just keep the original open and a new connection open too, until it is
certain that the initial population of the new database is successful.
* Check return value from AutoSaveDelete when saving-as
* Fix the remaining unchecked sqlite3_bind* calls, which are in Autosave...
...similarly to checks in 8b3f9fa
* When saving-as or -copy, open the new database only once
* Check return value from sqlite_initialize
* Docs say: call sqlite3_close even when sqlite3_open returns error
* Be careful of possible failures of AutoSave...
... In ProjectHistory operations, do the AutoSave first and throw if there is
a failure.
Only then change UndoManager's state, confident that it remains consistent with
what was AutoSaved.
* Throw exceptions if lazy opening of project's database fails...
... because the calls to DB() as in Sqlite3SampleBlock may be in deeply nested
places that can't propagate the error codes; and besides, those functions
had been assuming non-null returns from DB(), which might have crashed before,
but now the assumption is correct when the throw didn't happen.
Note that this exception may also happen during attempted Autosave. Uses of
Autosave were reviewed and some changes made in the previous commit.
Preliminary tests show it to be a bit faster than the default
4KB. For a simple example, generate 2-hour chirp dropped from
11 seconds to 7 seconds. Not a lot, but...
* Define SampleBlockFactory replacing static members of SampleBlock...
... This will become an abstract base class
* Sequence and WaveTrack only store SampleBlockFactory not Project...
... This adds a dependency from Track to SampleBlock which temporarily enlarges
a cycle in the dependency graph
* Register a global factory of SampleBlockFactory...
... so that later we can make an abstract SampleBlockFactory, separate from the
concrete implementation in terms of sqlite, and inject the dependency at startup
avoiding static dependency
* New concrete classes SqliteSampleBlock, SqliteSampleBlockFactory...
... separated from abstract base classes and put into a new source file,
breaking dependency cycles, and perhaps allowing easy reimplementation for other
databases in the future.
Note that the new file is a header-less plug-in! Nothing depends on it. It
uses static initialization to influence the program's behavior.
* Compile dependency on sqlite3.h limited to just two .cpp files...
... these are ProjectFileIO.cpp and SqliteSampleBlock.cpp.
But there is still close cooperation of ProjectFileIO and SqliteSampleBlock.cpp.
This suggests that these files ought to be merged, and perhaps ProjectFileIO
also needs to be split into abstract and concrete classes, and there should be
another injection of a factory function at startup. That will make the choice
of database implementation even more modular.
Also removed one unnecessary inclusion of ProjectFileIO.h
* Fix crashes cutting and pasting cross-project...
... in case the source project is closed before the paste happens.
This caused destruction of the ProjectFileIO object and a closing of the sqlite
database with the sample data in it, leaving dangling references in the
SqliteSampleBlock objects.
The fix is that the SqliteSampleBlockFactory object holds a shared_ptr to the
ProjectFileIO object. So the clipboard may own WaveTracks, which own WaveClips,
which own Sequences, which own SqliteSampleBlockFactories, which keep the
ProjectFileIO and the database connection alive until the clipboard is cleared.
The consequence of the fix is delayed closing of the entire database associated
with the source project.
If the source project is reopened before the clipboard is cleared, will there
be correct concurrent access to the same persistent store? My preliminary
trials suggest this is so (reopening a saved project, deleting from it, closing
it again -- the clipboard contents are still unchanged and available).
* Eliminate unneeded back-pointer to project from non-wave Tracks...
... now that DirManager is gone
* Remove unused declarations
* SampleData::mProject was not used
* Correct ProjectFileIO::GetLibraryError
* Remove unnecessary #include directives
These mainly address the bugs that Steve reported and a couple
more I found along the way.
Corrected ProjectFileIO::GetMinMaxRMS() - It was still using the
original method of keep all block data in memory. I missed it
when I redid everything. Fixes his Amplify crash.
Temporary filenames should no longer be shown to the user.
Resaves will no longer present a Save As dialog.
Cleaned up duplicate pathname handling in ProjectFileIO.
Returned proper errors when loading a project
- SQLite added to libs in readme.txt
- User message 'aup not associated' updated to 'aup3'
- Typos strind -> string, in -> is
- No translation of debug message
!!! THERE WILL NO DOUBT BE BUGS !!!
This is a big one and there's still several things to
complete. Just want to get this in the wild to start
receiving feedback.
One big thing right now is that it will NOT load pre-aup3
files. An importer is on the way for that.
... Reimplemented without making dependency cycles.
Project and ProjectFileManager publish events for change of active project or
change of a project title.
WindowMenus.cpp can listen for those events, so that it can update the menu
appropriately. So it's all done nonintrusively in the rest of the code.
These changes were made too close to release and are considered too dangerous for 2.4.0.
We will use the changes Leland provided. We can review these proposed changes by Paul after 2.4.0 is released.
... Reimplemented without making dependency cycles.
Project and ProjectFileManager publish events for change of active project or
change of a project title.
WindowMenus.cpp can listen for those events, so that it can update the menu
appropriately. So it's all done nonintrusively in the rest of the code.
... The fix done at 2f9322e was undone at 43b1afc because of the unacceptable
file dependencies. This fixes it by other means. Make a null check in
SetProjectTitle in case it is invoked before ProjectWindow has been constructed;
also ensure in ProjectManager::New that it is called at least once after that.
