C2 Doesn't hoist flat array reference load out of the loop for a simple loop.
That causes a huge performance difference (in comparison, when hoisting is done manually):
Benchmark Mode Cnt Score Error Units
V1Bench.sharpValueFlat avgt 10 588869.029 ± 6188.785 ns/op
V1Bench.sharpValueFlatArrayHoist avgt 10 249914.623 ± 991.713 ns/op

For the identical loop working with an array of identity classes, the load is hoisted successfully.