There is call overhead in which LDAR, issued by the original implementation of JDK-8189596, plays a significant role. (more details in https://mail.openjdk.org/pipermail/hotspot-compiler-dev/2024-July/078698.html ) According to the later discussion https://mail.openjdk.org/pipermail/hotspot-compiler-dev/2024-July/078715.html it is no longer needed.
There are 2 places to change: SharedRuntime::generate_native_wrapper() and TemplateInterpreterGenerator::generate_native_entry().
The same approach is used in RISC-V port but this RFE is for AArch64 only.