This implementation was done to calculate buffer CRC on the go, as that was what the VPK file system uses for the file crc's. But it later turned out to be unnecessary, and was never reverted.
These methods where only used for and by the VPKDir_t structure.
Also renamed variable and method names to reflect new changes regarding the use of KeyValues by creating a structure with all values defined during file lookup rather than doing this during the pack process.
The CPackedStore class now heavily relies on the engine's FileSystem and KeyValues API. Since its part of the engine, it would make more sense to utilize the engine's features instead. It also allows for easier modifications.
* The logic behind the build process hasn't changed.
* Don't call copy constructor on entry blocks when building the directory file.
* Don't call copy constructor for string when we don't want to sanitize the directory file name.
The sanitizer can retrieve the directory file name from the block file name if this is passed instead. This was broken in the previous system. Sanitization can be opted out if the 3rd parameter of the 'fs_vpk_unpack' command is omitted.
* Use reference to array index.
* Only obtain array size once during loop.
* Use static buffer of 1024*1024 for decompressing.
* Removed extraneous goto.
* Fixed bug where shared data would be mapped wrong due to the last optimization.
* This increased packing performance by 40/60 seconds when rebuilding all vpk's.
* Hash chunk before compression to determine if it already exists or not (don't run compression if it does).
* Use the result of vector::insert for data sharing (insert already checks if it exists or not, and returns the iterator).
* This optimization saves around 20 seconds total when rebuilding all VPK's.
* Replaced the boolean 'fatal' parameter with a error code parameter, anything non-null will prompt a message (fatal) and terminate the process with given error code.
* Fixed bug where the global ostreamsink for spdlog did NOT get cleared in 'SQVM_PrintFunc' when cvar 'sq_showvmoutput' was < 3. Moved to global scope.
* Added error message for when detouring the process has failed, with the error code.
* Only call 'Plat_GetProcessUpTime()' once per log, (improves performance and fixes bug where the error message box would show a different time stamp than what is logged into the console or file).
* All TIER0 loggers only log to notify and console when the SDK engine has fully initialized and detoured all functions.
When fatal is set, the function will show a error dialogue, which will halt the process and close when the message is dismissed. Else we will end up in a crash.
More detailed operation statistics.
Discard VPK directory files who's header doesn't match engine requirements.
Separated manifest properly between language and context (you can now package for each language and context from a single workspace).
Cast to proper types.
Set the archive offset only instead of the entire block. The entry could still have its own flags. The only thing needed for saving data is sharing the same offset and size.
Fixed bug where the compressed buffer size equals the source buffer size getting packed into the VPK chunk. In all occurrences this resulted in corrupted data upon export.
The reason for data being corrupt upon export is because the engine checks the equality of the compressed and decompressed size fields in the directory file, if they are equal, the engine doesn't attempt to decompress the block. So we end up with a still compressed block on the disk (technically not corrupt in the context of the compression lib, but useless as-is on the disk).
If a compressed file doesn't get lower in size we are better out storing it rather than compressing it.
Added a new condition in lzham::lzham_lib_compress_memory which checks source and destination buffer size equality.
Initial proper implementation pending cleanup.
The new system builds a manifest file when a VPK is unpacked. The manifest files contains data such as the entry flags and texture flags. It also contains a field determining whether the file should be compressed or not.
When a user repacks a pack, the system attempts to load this manifest file and does a lookup to the object to retrieve the flags (most of these flags are unknown, but they are used by the engine and are necessary for stuff like cubemaps and texture files to work correctly. Cubemaps won't work with proper flags, and textures (decals, particle system components, etc..) will look washed out without them.
I think some also determine whether a file within the VPK should be cached or not, so simply marking everything as 0x101 will probably end up in more CPU time and higher filesystem cache usage (depot/ is only 0x1, I don't think anything there is getting cached ever without the 0x100 flag).
User could also repack a VPK while excluding anything that is not in the manifest file. So you could unpack all VPK's into a single directory (each VPK has its own manifest file tied to its level name), and rebuild all the VPK's with only the files that where originally in them.
fs_pack_vpk command usage: <locale> <context> <level_name> <manifest_only>
locale determines the pak language (default english), context determines whether is a server/client vpk, level_name determines the BSP name of the pak, manifest_only determines whether the pack system should only include files within the manifest (leaving this arg out will build all files into the vpk).
The VPK workspace path is determined with ConVar 'fs_packedstore_workspace'.
