Improve BlockEncoding null bits packing

[pettyjamesm]

Description

Improve block serialization logic to prefer using primitive arrays over virtual calls when checking positions for nulls.

Benchmarks: BenchmarkBlockSerde.serialize*

Release notes

(x) This is not user-visible or is docs only, and no release notes are required.
( ) Release notes are required. Please propose a release note for me.
( ) Release notes are required, with the following suggested text:

[wendigo]

It this PR about reducing the network overhead for nulls or does it improve performance due to autovectorization as well?

[pettyjamesm]

It this PR about reducing the network overhead for nulls or does it improve performance due to autovectorization as well?

This PR doesn't change the serialized representation (and therefore has no impact on network overhead), the performance comes from better JIT code generation for the serialization loops, although not necessarily autovectorization specifically.

[wendigo]

@pettyjamesm since you've shared benchmark results, can you contribute benchmark itself too?

[wendigo]

Is it faster than:

            value |= block.isNull(position) ? 0b1000_0000 : 0;
            value |= block.isNull(position + 1) ? 0b0100_0000 : 0;
            value |= block.isNull(position + 2) ? 0b0010_0000 : 0;
            value |= block.isNull(position + 3) ? 0b0001_0000 : 0;
            value |= block.isNull(position + 4) ? 0b0000_1000 : 0;
            value |= block.isNull(position + 5) ? 0b0000_0100 : 0;
            value |= block.isNull(position + 6) ? 0b0000_0010 : 0;
            value |= block.isNull(position + 7) ? 0b0000_0001 : 0;

[pettyjamesm]

Yes, the new version performs better (and more consistently) than the previous implementation across all null rates in the benchmark including when branches are (nearly) perfectly predicted (null rate 0.01 and 0.99). The performance of the new version is nearly flat (with some variation in performance coming from the logic performing null-supression of the values array) where as we see a typical "M"-shaped curve with the previous version where performance relies on the branch prediction rate- which suggests the new version is branchless although unfortunately I don't have a JVM build with hdsdis available to show the generated assembly to prove it.

[pettyjamesm]

@pettyjamesm since you've shared benchmark results, can you contribute benchmark itself too?

The benchmark is already checked in as BenchmarkBlockSerDe. The results posted above only include the serialization benchmarks since the deserialization paths are unaffected by this change.

[dain]

Looks good