The X86 implementation of VtableStubs::create_itable_stub calls load_klass() twice on the same oop because the destination registers needs to be reused as a temporary in the intervening __ lookup_interface_method(). But on AArch64 we have more registers and no callee-saves so we can just allocate an additional temporary register and keep the result of load_klass live across the whole function.

Also the estimate of the stub size needs to be recalculated as it was too low before the fix for JDK-8234794 and now it's too high.