Keep in mind that because the memory in banks $E0 and $E1 are used as the 128K of memory in 8-bit mode (yeah, it's counterintuitive), they are controlled by the MEGA II chip, which is in essence an Apple IIe on a single chip (aside from processor and a few other support circuits). Any time access happens within the MEGA II's domain, the system has to slow to 1 MHz. That's why shadowing and such is such a big win for graphics.
So there are two main data buses in the Apple IIgs (ignoring the slot I/O bus, which doesn't matter for this context): the MEGA II data bus, which operates at 1 MHz to support 8-bit emulation (and the bank $E0 and $E1 memory space), and the FPI (Fast Processor Interface) bus, which handles the 16-bit realm and all the rest of memory.
The two sides of the system operate largely independently; in fact, they even have their own memory refresh cycles for the areas of memory they're responsible for (the FPI side's refresh cycle causes an 8% loss of performance, which is why you often hear a 2.6 MHz speed for the IIgs).
Because of that independence, there's a need to synchronize between the two sides whenever data is transferred from one side to the other. Synchronization performs by the FPI basically stopping the FPI side's clock (thereby halting execution) and waiting until the MEGA II is in the right part of its cycle, then it starts running again, but synchronized to the MEGA II clock (and therefore temporarily running at 1 MHz for the duration of the operation).
I presume that adding the
NOP at appropriate intervals reduces the length of time spent waiting for synchronization to occur. I have not had time to sit down and do that math and would prefer not to. :)