RFR: 8319883: Zero: Use atomic built-ins for 64-bit accesses

[Aleksey Shipilev]

On Fri, 10 Nov 2023 10:24:28 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> See the bug for reproducer and discussion.
> 
> This reimplements related Zero parts using `__atomic_*`: we already use some of those both in Atomic and around Hotspot code. As far as I can see from both GCC and LLVM docs, these would be translated to proper accesses on all target platforms. Doing even the relaxed atomic access for 64-bit platforms is probably an overkill, but for Zero we want baseline correctness first and foremost.
> 
> Example current uses of `__atomic_load` on both BSD (I think it assumes clang) and Linux (assumes GCC):
> https://github.com/openjdk/jdk/blob/a64fc48e2dbc650b31869dd79b1ba1012376fc51/src/hotspot/os_cpu/bsd_aarch64/atomic_bsd_aarch64.hpp#L115
> https://github.com/openjdk/jdk/blob/a64fc48e2dbc650b31869dd79b1ba1012376fc51/src/hotspot/os_cpu/linux_aarch64/atomic_linux_aarch64.hpp#L203
> https://github.com/openjdk/jdk/blob/a64fc48e2dbc650b31869dd79b1ba1012376fc51/src/hotspot/os_cpu/linux_riscv/atomic_linux_riscv.hpp#L172
> 
> Additional testing:
>  - [x] MacOS AArch64 Zero fastdebug; light jcstress tests still pass
>  - [x] Linux x86_32 Zero release; previously failing jcstress testing now passes
>  - [x] Linux x86_32 Zero fastdebug, `compiler/unsafe java/lang/invoke/VarHandles`
>  - [x] Linux x86_32 Zero fastdebug, bootcycle-images

Integrating to unblock other Zero bugfixing work.

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

PR Comment: https://git.openjdk.org/jdk/pull/16600#issuecomment-1810147516