I've seen floating point 'optimizations' cause issues with iterative methods that depend on the specifics of strict IEEE754 behavior (causing massive spikes in

I've seen floating point 'optimizations' cause issues with iterative methods that depend on the specifics of strict IEEE754 behavior (causing massive spikes in the number of iterations and thus hurting overall performance).