What patterns does one look for when building a system to decide which objects to protect coarsely and which to protect finely? Are there tools that assist in such profiling?

Do software/hardware implementations have a way of dynamically choosing between the three methods (objective, element/fields, $-line granularities)? I wonder if there is potential for performance improvement by dynamically choosing rather statically using one of the rules (similar to how hardware has relatively sophisticated branch predictors to improve performance on pipelined machines).