perf: preserve space to prevent mem allocations when evaluate the bool expression

[st1page]

https://github.com/risingwavelabs/risingwave/blob/d4629c3fde1b3da8cfef91304913177ef63e1ad8/src/common/src/buffer/bitmap.rs#L597-L602

[wangrunji0408]

Some benchmark results:

• Collecting 1024 bools into a bitmap is ~90ns.
• Evaluating less_than(bigint, int) on 2 int arrays with 1024 elements and outputting the result as a boolean array is 138ns.
• Evaluating less_than(bigint, int) on a data chunk with 1024 rows is 378ns.

Some explanations:

• The reason why Bitmap::from_iter takes a majority overhead in the flamegraph is that it includes function evaluation.
• The reason why we construct an intermediate Vec<bool> in Bitmap::from_iter is that both steps can utilize auto-vectorization. But collecting bools directly into a bitmap does not have a fast way as far as I know.

Conclusions:

• The overhead is not caused by memory allocations because the input iterator has implemented TrustedLen.
• The performance of less_than(bigint, int) function is not optimal, but also not too bad (compared with other functions which can not utilize SIMD). The potential optimization is no more than 3x faster (378 -> 138ns). I think it's not a big bottleneck of the query.

[github-actions[bot]]

This issue has been open for 60 days with no activity. Could you please update the status? Feel free to continue discussion or close as not planned.