[lworld] Integrated: 8366717: [lworld] Cleanup defensive fixing of JDK-8365996

[Marc Chevalier]

On Thu, 18 Dec 2025 08:26:03 GMT, Marc Chevalier <mchevalier at openjdk.org> wrote:

> Let's proceed piece by piece.
> 
> # `G1BarrierSetAssembler::g1_write_barrier_pre` in `g1BarrierSetAssembler_aarch64.cpp`
> 
> https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/aarch64/gc/g1/g1BarrierSetAssembler_aarch64.cpp#L216-L220
> 
> `push_call_clobbered_registers`/`pop_call_clobbered_registers` should be enough around a runtime call, that's what clobbered registers are.
> 
> But let's dig a bit, that will be useful later!
> 
> 
>    push_call_clobbered_registers()
> => push_call_clobbered_registers_except(exclude = empty set)
> => push(call_clobbered_gp_registers() - exclude, sp)  // with exclude = empty set
> 
> So, we save at least `call_clobbered_gp_registers` which is defined as:
> 
> https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp#L3614-L3620
> 
> So, we save r0-r7 and r10-r17
> 
> # `CardTableBarrierSetAssembler::oop_store_at` in `cardTableBarrierSetAssembler_aarch64.cpp` and `cardTableBarrierSetAssembler_x86.cpp`
> 
> https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/aarch64/gc/shared/cardTableBarrierSetAssembler_aarch64.cpp#L116-L124
> 
> Digging the history, it seems it was added to fix a corruption of `dst.base()`. Yet, it might not be always necessary. It's rather nicer to save where it matter that the register is not corrupted than all the way down. So, there are a few such places:
> 
> - `gen_c2i_adapter_helper` in `sharedRuntime_x86_64.cpp`
>   `gen_c2i_adapter_helper` calls `store_heap_oop` that can corrupts `to.base()`. `gen_c2i_adapter_helper` is called, for instance from `gen_c2i_adapter`:
>   https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp#L1069-L1070
>   where we can see that `to.base()` is `r14`, which is set
>   https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp#L1026
>   before the loop enclosing the call to `gen_c2i_adapter_helper`
>   https://github.com/openjdk/valhalla/blob/1077e4f9f06336af30d95fc28cbab4d31c9f2a44/src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp#L1036
>   If we do more than one iteration of this loop when entering the `store_heap_oop` branch, `r14` is corrupted at the second call and we get a segfault.
> - `generate_buffered_inline_type_adapter` in `sharedRuntime_x86_64.cpp`: same with register `rax` being corrupted in a loop.
> - `pack_inline_helper` ...

This pull request has now been integrated.

Changeset: d8438ac7
Author:    Marc Chevalier <mchevalier at openjdk.org>
URL:       https://git.openjdk.org/valhalla/commit/d8438ac789f1425030dc61403384f9fbb548ee13
Stats:     118 lines in 9 files changed: 35 ins; 55 del; 28 mod

8366717: [lworld] Cleanup defensive fixing of JDK-8365996

Reviewed-by: thartmann

-------------

PR: https://git.openjdk.org/valhalla/pull/1824