RFR: 8008243: Zero: Implement fast bytecodes [v2]

[Coleen Phillimore]

On Tue, 19 Oct 2021 10:00:17 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> This work improves Zero performance considerably, without complicating the interpreter all that much. This work is based on much earlier patch by @rkennke, then edited by @jerboaa, and finally polished by me here :)
>> 
>> SPECjvm benchmarks improve across the board, with lowest improvement on `Derby` (+6%) and largest on `Compress` (+50%). Given that Zero is used to implement the JVM on emerging platforms that are still rather slow themselves, it makes sense to make Zero quantifiably faster. Even 10% improvement means more than 2 hours a day worth of CPU time.
>> 
>> Additional testing:
>>  - [x] Linux x86_64 Zero fastdebug bootcycle-images
>>  - [x] Linux x86_64 Zero fastdebug fuzzer tests (no failures = no new behaviors introduced)
>
> Aleksey Shipilev has updated the pull request incrementally with three additional commits since the last revision:
> 
>  - Remove shadowing "obj" to fix JVMTI tests
>  - Comment _fast_zputfield LSB
>  - MAYBE_POST_FIELD_MODIFICATION should check _putstatic

Looks good! We've had this request for so long now so thank you for doing it!

src/hotspot/share/interpreter/zero/bytecodeInterpreter.cpp line 394:

> 392:         }
> 393: 
> 394: #define MAYBE_POST_FIELD_ACCESS(obj) {                              \

I wonder if making these functions would reduce the size of the code in the big case statement and improve the compiler's ability to optimize it?

src/hotspot/share/interpreter/zero/bytecodeInterpreter.cpp line 2719:

> 2717: 
> 2718:         int field_offset = cache->f2_as_index();
> 2719:         obj->byte_field_put(field_offset, (STACK_INT(-1) & 1)); // only store LSB

Good!

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

Marked as reviewed by coleenp (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/1938