The old condition happened to work, because SFLAG_INTEXTURE was not
set(we're loading it), and neither was SFLAG_INDRAWABLE(otherwise
INTEXTURE would be set too). If the fbo INDRAWABLE == INTEXTURE is
replaced by INDRAWABLE == INSRGBTEX this doesn't hold true any longer
because SFLAG_INDRAWABLE is set, but not used because the drawable
readback is never used for fbos.
FBO stretchrect marks SFLAG_INDRAWABLE up to date, which currently
also marks SFLAG_INTEXTURE up to date. However, this will change when
we enable rendering to the srgb copy, in which case the drawable could
be equal to the srgb copy, not the rgb copy.
This prevents for example a d3d9 depth stencil from being destroyed when it
has no external references but is still in use by the device/stateblock. A
nice side effect is that it simplifies handling of "implicit" surfaces like
the frontbuffer and backbuffers, as well as the forwarding of reference counts
for surfaces that are part of a texture.
The MSDN is not totally correct: A8R8G8B8 and A1R5G5B5 also allow GetDC. The
main differences that have to be filtered out in d3d9.dll are GetDC on
A8B8G8R8, X8B8G8R8 and P8.
Passing NULL to glCompressedTexImage2DARB() doesn't work particularly well
when both client storage and PBOs are enabled. Also, two years is long enough
to give driver developers a chance to fix the bug this was supposed to work
around.
On ATI cards we use the driver's GL_ARB_texture_non_power_of_two emulation to
support conditional NP2 textures without having to deal with the denormalized
coordinates. The default GL_TEXTURE_2D address mode is incompatible with
GL_ARB_texture_rectangle however.
In general it might not be a bad idea to group GL specific data into a
separate struct, but IWineD3DSurfaceImpl is currently the only thing that
bothers. It doesn't get it quite right either, since e.g. the PBO isn't in
glDescriptor either.
Although these formats aren't lockable, we still explicitly set the initial
surface data when creating the texture. Unfortunately that means we'll need
the conversion functions, even though all they'll ever convert will be zeroes.