When working with the FFM API we noticed that using malloc/free directly seems to be always faster than using Unsafe::allocateMemory and Unsafe::freeMemory.

See the performance chart in this PR:

https://github.com/openjdk/panama-foreign/pull/855#issuecomment-1660112775

When investigating this gap, we realized that the issue is caused by the double thread state transition (from Java -> Native, then from Native -> VM) which occurs at each Unsafe call from Java code.

Using UNSAFE_LEAF instead of UNSAFE_ENTRY largely removes this issue.

Unsafe (UNSAFE_ENTRY): 60.815 ns/op
Unsafe (UNSAFE_LEAF): 41.771 ns/op
Malloc: 35.379 ns/op

Can UNSAFE_LEAF be used for Unsafe::allocateMemory/freeMemory? Or does NativeMemoryTracking (which is used by Unsafe) needs VM mode?