ARROW-5752: [Java] Improve the performance of ArrowBuf#setZero

[liyafan82]

The current implementation involves repeated calls of setLong, setInt & setByte. It is more efficient to directly call the native function.

Performance improve by 4.8x
before:
ArrowBufBenchmarks.setZero avgt 5 319616.675 ± 918.298 ns/op
after:
ArrowBufBenchmarks.setZero avgt 5 66616.553 ± 202.525 ns/op

[emkornfield]

LGTM, @praveenbingo @pravindra any concerns?

[siddharthteotia]

LGTM. I am just curious what the native operation does? I mean does it not write word by word?

[liyafan82]

LGTM. I am just curious what the native operation does? I mean does it not write word by word?

The native code also writes data word by word. However, the performance of native code is higher.
For details, please see Copy::fill_to_memory_atomic in https://github.com/JetBrains/jdk8u_hotspot/blob/master/src/share/vm/utilities/copy.cpp

[emkornfield]

@liyafan82 failure looks unrelated but I would like to make sure this doesn't affect gandiva in some way, could you retry?

[pravindra]

Nice one. thanks for fixing this !

[pravindra]

@liyafan82 failure looks unrelated but I would like to make sure this doesn't affect gandiva in some way, could you retry?

i retriggered the build, it finished successfully.

[emkornfield]

merging

[liyafan82]

@liyafan82 failure looks unrelated but I would like to make sure this doesn't affect gandiva in some way, could you retry?

i retriggered the build, it finished successfully.

@pravindra thanks a lot for your effort.