While studying the result of a profiling session with Marko, we noticed a few things:

• There's far too much time spent in LocalizedMessage#equals(). We thought about 2 things: interverting the 2 equals condition to test the message first: it's far more discriminating than the locale in most cases, and (maybe) try to avoid creating LocalizedMessage instances using a nested map (but the benefit should be measured, it's not obvious).
• ConstraintTree#mainConstraintNeedsEvaluation() can be tuned to test early if the constraintViolations set is empty.
• There is a lot of time spent in ReflectionHelper#getValue(). Maybe we could optimize things by storing the constraint per location in BeanMetaData, thus calling ReflectionHelper#getValue() only once per location, instead of calling it once per constraint.
• Less pressing (it's pretty low in the profile) but might be worth some investigation: ConstraintTree#passesCompositionTypeRequirement() appears in the profile whereas it's only useful if we have composing constraints (which is not the common case). Maybe we could try to have a fast path in validateConstraints() when we don't have any composing constraints.