Some tests in Speedometer2 are slower because they try the stringify fast path over and over again and it doesn't work. We should assume that if we call stringify repeatedly it is likely the calls will either all work with the fast path or all fail. We can create a heuristic where each time the stringify fast path fails we don’t try it again for some number of subsequent calls. That way if we are repeatedly doing calls that work with the fast path we'll keep using it every time, and if we are repeatedly doing calls that don't work with the fast path we'll only try it some small percentage of the time.

We can judge this heuristic with Speedometer2 itself, but it's likely the heuristic will also be effective on web content in the wild.