Huge reduce in logic usage when megafunction is instantiated

When you instantiate each component on their own, synthesis knows nothing about their inputs and outputs and generated a perfectly generic component.

However, when you begin putting them together on real circuits, synthesis is, sometimes, able to deduce that some output signals won't be used and that some input signals are constant (or related) and thus optimize the design.

Considering that you're using switches to change the inputs, I'm betting that your adder's inputs are far short of the ~(2^32)^2 possible combinations supported by a single precision fp adder.