I remember reading an article about Java performance many years ago. It had an example for Java being faster than C++ and it was maths on arrays / matrices. I guess that kind of code least trips up HotSpot optimizations, so it wins due to availability of runtime information and a really good implementation of certain parts.

Though I doubt that Java usually wins against the Eigen C++ library. Eigen uses some template tricks to fold operations together (which can avoid intermediate storage and unlocks more types of restructuring optimizations) and uses SIMD extensively.