.NET 8 Update(6): CollectionsMarshal for ListFormatter

[pCYSl5EDgo]

List<T> for loop is slower than Span<T> for loop.
CollectionsMarshal.AsSpan converts List<T> to Span<T>
CollectionsMarshal.SetCount can be used for deserialize.

[AArnott]

Wow. Did you happen to measure perf impact of this change?

[pCYSl5EDgo]

The below benchmarks were done without Dynamic PGO.
With Dynamic PGO, there was no difference between new formatter and old formatter.

List<byte>

BenchmarkDotNet v0.13.12, Windows 11 (10.0.22621.3007/22H2/2022Update/SunValley2)
Intel Core i7-8750H CPU 2.20GHz (Coffee Lake), 1 CPU, 12 logical and 6 physical cores
.NET SDK 8.0.101
  [Host]     : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2
  DefaultJob : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2

Method	Size	Mean	Error	StdDev
SerializeNew	1	73.66 ns	1.510 ns	1.615 ns
SerializeOld	1	76.72 ns	1.452 ns	1.728 ns
DeserializeNew	1	89.63 ns	1.852 ns	2.535 ns
DeserializeOld	1	90.21 ns	1.846 ns	4.665 ns
SerializeNew	64	257.25 ns	5.140 ns	5.279 ns
SerializeOld	64	258.83 ns	4.092 ns	5.025 ns
DeserializeNew	64	441.35 ns	8.431 ns	7.474 ns
DeserializeOld	64	481.55 ns	7.090 ns	6.286 ns
SerializeNew	1024	3,420.87 ns	66.383 ns	71.029 ns
SerializeOld	1024	3,415.81 ns	13.653 ns	11.401 ns
DeserializeNew	1024	9,625.97 ns	191.207 ns	255.256 ns
DeserializeOld	1024	10,175.56 ns	203.576 ns	217.824 ns
SerializeNew	16777216	122,029,780.00 ns	1,344,153.681 ns	1,257,322.187 ns
SerializeOld	16777216	131,873,157.69 ns	2,493,291.890 ns	3,412,846.081 ns
DeserializeNew	16777216	160,013,514.81 ns	3,050,381.441 ns	4,276,208.758 ns
DeserializeOld	16777216	167,149,218.84 ns	3,336,039.530 ns	4,219,021.850 ns

List<Matrix4x4>

BenchmarkDotNet v0.13.12, Windows 11 (10.0.22621.3007/22H2/2022Update/SunValley2)
Intel Core i7-8750H CPU 2.20GHz (Coffee Lake), 1 CPU, 12 logical and 6 physical cores
.NET SDK 8.0.101
  [Host]     : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2
  DefaultJob : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2

Method	Size	Mean	Error	StdDev
SerializeNew	1	142.7 ns	2.87 ns	3.31 ns
SerializeOld	1	141.5 ns	2.56 ns	4.20 ns
DeserializeNew	1	220.5 ns	4.03 ns	5.51 ns
DeserializeOld	1	218.0 ns	3.89 ns	3.64 ns
SerializeNew	64	4,730.6 ns	68.85 ns	57.50 ns
SerializeOld	64	4,931.7 ns	98.49 ns	209.89 ns
DeserializeNew	64	8,431.1 ns	163.43 ns	194.56 ns
DeserializeOld	64	8,365.5 ns	146.44 ns	129.81 ns
SerializeNew	1024	149,412.6 ns	2,915.70 ns	4,364.08 ns
SerializeOld	1024	142,725.7 ns	2,752.63 ns	3,276.81 ns
DeserializeNew	1024	142,506.2 ns	2,848.95 ns	5,064.01 ns
DeserializeOld	1024	145,993.7 ns	2,861.05 ns	3,916.24 ns
SerializeNew	16777216	1,588,265,229.5 ns	38,374,748.24 ns	105,695,201.32 ns
SerializeOld	16777216	1,502,806,044.9 ns	28,744,116.00 ns	69,420,270.85 ns
DeserializeNew	16777216	2,413,118,764.5 ns	46,612,636.71 ns	71,182,419.37 ns
DeserializeOld	16777216	2,586,872,945.1 ns	51,537,327.26 ns	126,421,942.00 ns

[pCYSl5EDgo]

For smaller value types, Serialize of this PR is more performant than original one.
But for bigger value types Serialize of this PR is slower than original one.

Deserialize of this PR is in most cases better than original.

Considering the above facts, let's revert the change for Serialize.