Currently the AArch64 version of G1BarrierSetAssembler::gen_write_ref_array_pre_barrier unconditionally calls into the G1 runtime to perform the pre-barrier. This is unnecessary when marking is not active.

Change the barrier code to check the marking active field and skip the call if not active (like X86). On AArch64 we get an extra win as this avoids the ISB when returning native->Java, the overhead of which can be high if the arrays are small.