For example, let's assume I'm rendering cascaded shadow mapping, but for whatever reason, instead of one of the typical approach, I do the following:
Render the lowest resolution shadowmap
Copy part of the shadowmap to the next shadowmap, covering a smaller area(and thus more dense)
Compute a stencil mask based on different points' distance from camera, e.g. more detail on edges, and on points closer to the camera, reusing the old shadowmap elsewhere.
Render the next, higher-density shadowmap with the aforementioned stencil.
What performance can I realistically expect from that? If only 50% of the pixels(or should I say fragments here?) pass the stencil test, do I save 50% of the rendering time, less, none?
If this is a poor example, my fallback question scenario is a static scene, where the previous frame's result is reprojected, with an error accumulator in alpha, and a stencil mask is computed based on misses or too high error rates(e.g. position of the previous frame's reprojected pixel is too far from the new target pixel).