FDummyTexture had a big problem: Whenever it was accessed by accident it crashed the app because it wasn't fully implemented.
What it should do is return empty pixels of the given size, and an unextended FImageTexture is doing just that.
This will mostly ensure that each patch used for composition is only loaded once and automatically unloaded once no longer needed.
So far only for paletted rendering, but the same logic can be used for true color as well.
* split up FMultiPatchTexture into a builder class and the actual image source.
* since images can now be referenced by multiple textures the old redirection mechanism has been removed. It can be done better and less intrusive now. Simple single patch textures already directly reference the underlying patch image now.
* allocate all image source related data from a memory arena. Since this is all static this makes it a lot easier to free this in bulk.
Previously it tried to copy all patches of composite sub-images directly onto the main image.
This caused massive complications throughout the entire true color texture code and made any attempt of caching the source data for composition next to impossible because the entire composition process operated on the raw data read from the texture and not some cacheable image. While this may cause more pixel data to be processed, this will be easily offset by being able to reuse patches for multiple textures, once a caching system is in place, which even for the IWADs happens quite frequently.
Removing the now unneeded arguments from the implementation also makes things a lot easier to handle.
* it's no longer the main texture objects managing the pixel buffer but FSoftwareTexture.
* create proper spans for true color textures. The paletted spans only match if the image does not have any translucent pixels.
* create proper warp textures instead of working off the paletted variants.
As a side effect, caching of pixel buffers for texture composition is temporarily disabled, as it management of texture redirections. These things will be reimplemented once things progress further. The existing methods here had their share of serious issues that should be fixed.
This was done to make reviewing easier, again because it is virtually impossible to search for the operators in the code.
Going through this revealed quite a few places where texture animations were on but shouldn't and even more places that did not check PASLVERS, although they were preparing some paletted rendering.
This class has only meaning for software-based warping so it doesn't have to be a part of the FTexture hierarchy.
Making it a subclass of FSoftwareTexture is fully sufficient.
As a bonus this already fixes several bugs caused by the botched texture scaling implementation the original texture manager came with.
System cursors are currently disabled because they rely on functionality that needs to be moved to different classes.
This reuses the FTexCoordInfo class the hardware renderer had been using to calculate wall texture offsetting.
The software renderers still need this sorted out to bring them in line with the rest of the code, though, but they do not have this code sufficiently well organized to make this a straightforward task.
This had absolutely no sanity checks and unconditionally picked the source texture if one existed.
It should only be done for wall textures, only for those defined in TEXTUREx and only for those where the scale is identical with the underlying texture.
This replaces the old redirection hackery that had to work differently for the legacy render path.
Overriding CopyTrueColorTranslated is far more robust and ensures that no edge cases can reach the GetPixels function, which wasn't the case before.
With the software renderer not being used for 2D anymore it would be quite annoying if every texture class had to implement these.
If done properly, the SW renderer should actually be forced to dynamic_cast any texture to a FWorldTexture before using this stuff, but that'd be some pointless overhead better saved.
This theoretically means that the software renderer could access this data as well - if it just had been written with a more flexible texture interface.
However, as things stand, this may require quite a bit of work to achieve.
The old organization made sense when ZDoom still was a thing but now it'd be better if all pure data with no dependence on renderer implementation details was moved out.
A separation between GL2 and GL3+4 renderers looks to be inevitable and the more data is out of the renderer when that happens, the better.
