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

[Aleksey Shipilev]

On Sun, 12 Nov 2023 22:11:46 GMT, David Holmes <dholmes 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
>
> Seems reasonable. Can I assume that the correct volatile access memory semantics are handled at a higher-level?
> 
> Thanks

Thanks for review, @dholmes-ora! I think this formally requires another Reviewer, but I am not sure who else is maintaining Zero these days. Maybe @phohensee or @tstuefe can formally ack this.

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

PR Comment: https://git.openjdk.org/jdk/pull/16600#issuecomment-1807705337