Allow users to improve MemoryMeter performance for specific use cases

[blerer]

By default Jamm should provided correct measurements while being as simple as possible to use. Nevertheless for some usecases people might wish to make some trade off to get better faster measurements out of Jamm.

There are different areas on which user could play to get better performance:

• eliminating the cost of reflection for their own classes
• caching measurements of object shallow size
• using the more efficient measurements even if the belongs to different strategies

Eliminating the cost of reflection

The cost of reflection could be eliminated by providing a way for classes to add their children objects directly to the stack instead of having to rely on reflection. The drawback of this approach is that it push the responsibility to the user to ensure that all children are taken into account.

Caching measurement

Jamm should allow users to define a simple caching strategy for measurements

Using the more efficient measurements even if the belongs to different strategies

The INSTRUMENTATION strategy outperform other strategies for non array measurements but other strategies are more performant for array measurements. As array measurements do not suffer from correctness issue we could create a new strategy that use Instrumentation to measure non array objects and the SPEC approach for array objects

[blerer]

After looking into it, we decided to stay away from caching as it has some impact on memory that are difficult to plan for.

[blerer]

The final changes are:

• Deep measurement of String and measurement of arrays has been optimized
• A new INSTRUMENTATION_AND_SPECIFICATION strategy has been introduced. It improves significantly the speed of array measurements compare to the INSTRUMENTATION strategy, for java version < 17.
• A new Measurable interface has been introduce that allow user to avoid the cost of reflection and provide a more controlled crawling.
• The logic for dealing with ByteBuffer slabs has been revisited to provide more correct behavior based on how slabs are created.