The attached benchmark creates slices with both memory segments and byte buffers. When running the benchmark with this option:

-XX:CompileCommand=print,java/nio/HeapIntBuffer.slice -XX:CompileCommand=print,java/nio/DirectIntBufferU.slice -XX:CompileCommand=print,jdk/internal/foreign/AbstractMemorySegmentImpl.asSlice

Some memory barriers show up, of the kind:

  0x00007f76a4f83109: lock addl $0x0,-0x40(%rsp)

These barriers seem to be extremely sensitive to how fields are stored (see JDK-8279930), and they cause performance issues when said code is later inline into other methods.

Running with a GC other than G1 (e.g. -XX:+UseSerialGC) removes these barriers.