Uoppi, you got it exactly the wrong way around.
Just image downscaling: downscale in LL to preserve original brightness (LL checkbox for downscaling ticked)
When there is downscaling because of image doubling and SuperRes is used: downscale in GL (LL checkbox for downscaling unticked) and use SuperRes in GL (LL checkbox for SuperRes unticked as well)
However, the effect for SuperRes of LL/GL used for image downscaling after image doubling probably is very small in most cases. It's much more important how the image of the video (untouched by madVR) has been downscaled before.
We can't say this for sure and can only assume that downscaling in GL by content creators/providers is most common. The results with SuperRes
seem to confirm this assumption, but I don't think anyone tested lots of material with this.