Depends on where your bottleneck is. (And that means profiling)

If it is drawing, as small a boundingrect as possible.
If it is the calculation of what has to be redrawn, bigger roundingrect is better.

But as a general rule: as small boundingrect as possible. This way you do not get in the way when optimization deeper in the framework make all your little workarounds obsolete ;-)


As 4.3, it "felt" faster to me (in valgrind). But I did not compare 1:1, and it might be that I even compared an optimized build to an debug built. Best test yourself ;-)