remove accidental quadratic null_count

[ritchie46]

I noticed that the list builders in polars were extremely slow compared to the concat kernels, so I took a look why.

The flamegraph showed that this snippet was dominated by null_counts.

pub fn create_list(value: &Series, length: usize) -> ListChunked {
    let name = "foo";
    let mut builder =
        get_list_builder(value.dtype(), value.len() * length, length, name).unwrap();
    for _ in 0..length {
        builder.append_series(value)
    }
    builder.finish()
}


fn main() -> Result<()> {
    let value = Int32Chunked::full("", 10, 1000).into_series();

    for _ in 0..100 {
        create_list(&value, 1000);
    };
    Ok(())
}

The underlying code did a full bit count at every extend_slice call, leading to to O(n^2) behavior.

So the result improved drastically when I removed the redundant count.

bench_1                 time:   [129.73 us 129.86 us 130.01 us]
                        change: [-97.749% -97.737% -97.725%] (p = 0.00 < 0.05)
                        Performance has improved.

[ritchie46]

There was no validity and we extend with true so we know that null_count == 0 no need to compute it.

[ritchie46]

Before

After

[jorgecarleitao]

oof, that is an amazing find - a lot of code uses this.

[ritchie46]

oof, that is an amazing find - a lot of code uses this.

Yes, Almost all polars expressions in the groupby context. So I am really excited about finding this one. :D

[codecov]

Codecov Report

Merging #991 (b0006f8) into main (826a2b8) will increase coverage by 0.00%.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main     #991   +/-   ##
=======================================
  Coverage   71.38%   71.39%           
=======================================
  Files         357      357           
  Lines       19791    19787    -4     
=======================================
- Hits        14128    14126    -2     
+ Misses       5663     5661    -2     

Impacted Files	Coverage Δ
src/array/primitive/mutable.rs	88.79% <ø> (+0.22%)	⬆️
src/compute/arithmetics/time.rs	26.60% <0.00%> (+0.91%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 826a2b8...b0006f8. Read the comment docs.