The new way of transitioning to layer callbacks - done because of
portability (mainly to aarch64) - is a bit expensive.  This is a simple
way of recovering that cost.  Merge 603a24fc regressed speed a bit.