Anybody with experiences on this?
What magnitude is "render slightly faster" ?
I would not expect much difference, but it generally depends. For example, I recently read that ATI does not support unnormalized texture coordinates, so they have to add, behind the scenes, the proper instructions in a fragment shader to normalize the texcoords. Another possibility is that the texture gets resized to the next POT dimensions (you lose memory of course) and then texcoords get treated accordingly.
But rectangle textures may be faster in other cases. One example is render-to-texture. Because render-to-texture works better with unswizzled formats and rectangle textures are generally stored that way, it actually renders faster. I have verified this in Marathon.
I'm not sure if I fully undersand it, though. Can I still use
gl.glTexImage2D(GL_TEXTURE_2D, 0, GL_COMPRESSED_RGBA, ...)
as I did for power-of-two textures (provided the card supports the extension "GL_ARB_texture_compression") ?
Is it "just" that you can't use glCompressedTexImage ?
I haven't tried it but, if I read correctly, I think you may use GL_COMPRESSED_X in glTexImage, but most probably you won't get any compression (you should try though).