If we have to write compiler specific code for micro-optimizations I am out.
The Posix compatible version nullified most the advantage on MSVC by writing out the XMM register to memory and then reading back the float.
That's not worth the hassle for an optimization that brings a few microseconds at best.
* dynamic lights also work in the true color software renderer and have been moved out of the OpenGL menu.
* created a separate software renderer menu and moved all relevant options there.
* delete non-applicable options when running in legacy mode.
* moved the OpenGL preferences menu one level up to eliminate a two-entry GL top level menu.
Removing this made me realize that calling the renderers' FakeFlat functions from the automap is inherently unsafe with the recent refactorings because there is absolutely no guarantee that the data may actually still be defined when the automap is being drawn.
So the best approach here is to give the automap its own FakeFlat function that runs independently of render data and assumptions of data preservation. This one can also be a lot simpler because it only needs the floor, not the ceiling info.
- optimized the math to get a plane equation from a linedef. The original code used a generic algorithm that knew nothing about the fact that Doom walls are always perfectly vertical. With this knowledge the plane calculation can be reduced to a lot less code because retrieving the normal is trivial in this special case.
- use the SSE2 rsqrtss instruction to calculate a wall's length, because this is by far the most frequent use of square roots in the GL renderer. So far this is only active on x64, it may be activated on 32 bit later as well, but only after it has been decided if 32 bit builds should be x87 or SSE2.
# Conflicts:
# src/gl/dynlights/gl_dynlight.cpp
# Conflicts:
# src/g_shared/a_dynlightdata.cpp
Since the true color software renderer also handles them there is no point keeping them on the GL side.
This also optimized how they are stored, because we no longer need to be aware of a base engine which doesn't have them.
- removed the LastCamera logic in RenderView. This code predates the first GZDoom release and apparently was only added because back then R_SetupFrame was not fully compatible with the hardware renderer. Today it is not needed anymore.
(Is there anyway to tone down GCC's warning level? It outputs too many false positives for potentially uninitialized variables in which the genuine errors get drowned.)
This one was particularly nasty because Windows also defines a DWORD, but in Windows it is an unsigned long, not an unsigned int so changing types caused type conflicts and not all could be removed.
Those referring to the Windows type have to be kept, fortunately they are mostly in the Win32 directory, with a handful of exceptions elsewhere.
Both files can now be included independently without causing problems.
This also required moving some inline functions into separate files and splitting off the GC definitions from dobject.h to ensure that r_defs does not need to pull in any part of the object hierarchy.
Most of those which still rely on ZDoom's own definition should be gone, unfortunately the code in files that include Windows headers is a gigantic mess with DWORDs being longs there intead of ints, so this needs to be done with care. DWORD should only remain where the Windows type is actually wanted.
This is one of two places that unconditionally pulled in all Windows headers into the GL code.
We also do not need the cruft for defining the standard integer types. GZDoom is C++11 which means that stdint.h will be present. So the madness with the definitions should be avoided to ensure that the types are always the same.
