Use uninitialized array allocation

[voidzcy]

JVM has the optimization for eliminating zeroing when the allocation is followed by System.arraycopy. For codepaths that uses UnsafeUtil.copyMemory the allocation is allocated and the zeroing takes a fair amount of CPU cycles when array size is large.

Java 9+ has jdk.internal.misc.Unsafe.allocateUninitializedArray which does what we want.

Related to gRPC's optimization for passing Iterable<ByteBuffer> into Protobuf.

I've run the gRPC transport benchmark with 1 thread and 2MB response messages on Java 11:

• Baseline

Benchmark                                        (direct)  (transport)   Mode  Cnt          Score          Error  Units
TransportBenchmark.streamingCallsByteThroughput      true        NETTY  thrpt  100  655102314.930 ± 61603907.509  ops/s

• With changes in this PR

Benchmark                                        (direct)  (transport)   Mode  Cnt          Score          Error  Units
TransportBenchmark.streamingCallsByteThroughput      true        NETTY  thrpt  100  718053472.436 ± 45752070.587  ops/s

(The full benchmark results are here and here. The baseline commit is 70b0286)

We can see this change gives about 9.6% performance improvement for gRPC running on Java 11.

[voidzcy]

Gently ping @acozzette.

[voidzcy]

Close as discussions happening internally.