Add Span.Sort, and make Array.Sort span-based

[stephentoub]

Replaces #24419 from @nietras.

The first commit is @nietras', rebased on top of master.

The second commit cleans up a few things, including removing the native TrySZSort entirely in favor of just using the managed implementation, and passing around only spans rather than spans + lo index + hi index.

Fixes https://github.com/dotnet/coreclr/issues/27683
Contributes to https://github.com/dotnet/corefx/issues/15329

cc: @jkotas, @nietras

I ran the dotnet/performance Array.Sort benchmarks (old is master, new is this PR):

Type	Method	Toolchain	Size	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Sort	Array	\new\corerun.exe	512	8.657 us	0.8237 us	0.8814 us	8.393 us	7.790 us	11.029 us	0.85	0.10	-	-	-	-
Sort	Array	\old\corerun.exe	512	10.215 us	0.9868 us	1.0558 us	9.811 us	9.086 us	12.327 us	1.00	0.00	-	-	-	-
Sort	Array	\new\corerun.exe	512	3.898 us	0.1021 us	0.1093 us	3.899 us	3.753 us	4.152 us	0.82	0.03	-	-	-	-
Sort	Array	\old\corerun.exe	512	4.772 us	0.1108 us	0.1138 us	4.772 us	4.605 us	4.984 us	1.00	0.00	-	-	-	-
Sort	Array	\new\corerun.exe	512	31.986 us	0.1564 us	0.1386 us	31.980 us	31.778 us	32.264 us	0.98	0.00	-	-	-	-
Sort	Array	\old\corerun.exe	512	32.773 us	0.0791 us	0.0701 us	32.773 us	32.648 us	32.925 us	1.00	0.00	-	-	-	-
Sort	Array	\new\corerun.exe	512	5.014 us	1.6981 us	1.8874 us	3.986 us	3.901 us	8.970 us	1.07	0.25	-	-	-	-
Sort	Array	\old\corerun.exe	512	4.570 us	0.6052 us	0.6969 us	4.059 us	3.955 us	5.955 us	1.00	0.00	-	-	-	-
Sort	Array	\new\corerun.exe	512	287.612 us	0.9120 us	0.7120 us	287.485 us	286.617 us	288.895 us	1.03	0.00	-	-	-	-
Sort	Array	\old\corerun.exe	512	280.042 us	0.8725 us	0.7286 us	280.188 us	278.930 us	281.135 us	1.00	0.00	-	-	-	-
Sort	Array_ComparerClass	\new\corerun.exe	512	28.315 us	0.7971 us	0.8859 us	27.932 us	27.380 us	30.100 us	0.99	0.03	-	-	-	64 B
Sort	Array_ComparerClass	\old\corerun.exe	512	28.458 us	0.3567 us	0.3336 us	28.391 us	28.105 us	29.032 us	1.00	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\new\corerun.exe	512	19.676 us	0.2118 us	0.1981 us	19.624 us	19.480 us	20.077 us	0.99	0.01	-	-	-	64 B
Sort	Array_ComparerClass	\old\corerun.exe	512	19.832 us	0.0548 us	0.0458 us	19.815 us	19.761 us	19.917 us	1.00	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\new\corerun.exe	512	39.608 us	0.1484 us	0.1388 us	39.571 us	39.447 us	39.869 us	1.05	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\old\corerun.exe	512	37.876 us	0.0922 us	0.0720 us	37.866 us	37.759 us	37.996 us	1.00	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\new\corerun.exe	512	18.892 us	0.0575 us	0.0538 us	18.886 us	18.827 us	19.008 us	0.88	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\old\corerun.exe	512	21.416 us	0.0671 us	0.0628 us	21.414 us	21.314 us	21.550 us	1.00	0.00	-	-	-	64 B
Sort	Array_ComparerClass	\new\corerun.exe	512	269.803 us	3.3385 us	3.1228 us	270.197 us	265.775 us	275.033 us	0.86	0.01	-	-	-	64 B
Sort	Array_ComparerClass	\old\corerun.exe	512	315.011 us	1.4353 us	1.2723 us	315.079 us	312.248 us	317.227 us	1.00	0.00	-	-	-	64 B
Sort	Array_ComparerStruct	\new\corerun.exe	512	31.203 us	0.1804 us	0.1687 us	31.219 us	30.899 us	31.422 us	0.96	0.01	-	-	-	88 B
Sort	Array_ComparerStruct	\old\corerun.exe	512	32.334 us	0.2771 us	0.2456 us	32.297 us	32.012 us	32.902 us	1.00	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\new\corerun.exe	512	23.737 us	0.1534 us	0.1435 us	23.676 us	23.596 us	24.077 us	1.03	0.01	-	-	-	88 B
Sort	Array_ComparerStruct	\old\corerun.exe	512	23.099 us	0.0647 us	0.0605 us	23.086 us	23.014 us	23.230 us	1.00	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\new\corerun.exe	512	40.157 us	0.1432 us	0.1270 us	40.106 us	40.043 us	40.464 us	0.99	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\old\corerun.exe	512	40.561 us	0.1432 us	0.1339 us	40.512 us	40.419 us	40.925 us	1.00	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\new\corerun.exe	512	23.095 us	0.0537 us	0.0476 us	23.088 us	23.006 us	23.185 us	0.93	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\old\corerun.exe	512	24.750 us	0.0464 us	0.0411 us	24.751 us	24.641 us	24.822 us	1.00	0.00	-	-	-	88 B
Sort	Array_ComparerStruct	\new\corerun.exe	512	289.899 us	1.1267 us	0.9408 us	289.953 us	288.531 us	291.185 us	1.06	0.01	-	-	-	88 B
Sort	Array_ComparerStruct	\old\corerun.exe	512	274.180 us	1.5219 us	1.2709 us	274.031 us	272.874 us	276.486 us	1.00	0.00	-	-	-	88 B
Sort	Array_Comparison	\new\corerun.exe	512	27.693 us	0.3704 us	0.3283 us	27.576 us	27.335 us	28.225 us	0.97	0.02	-	-	-	-
Sort	Array_Comparison	\old\corerun.exe	512	28.425 us	0.4428 us	0.3925 us	28.362 us	27.895 us	29.052 us	1.00	0.00	-	-	-	-
Sort	Array_Comparison	\new\corerun.exe	512	18.507 us	0.1165 us	0.0972 us	18.508 us	18.390 us	18.732 us	1.00	0.01	-	-	-	-
Sort	Array_Comparison	\old\corerun.exe	512	18.542 us	0.0847 us	0.0661 us	18.528 us	18.444 us	18.689 us	1.00	0.00	-	-	-	-
Sort	Array_Comparison	\new\corerun.exe	512	38.252 us	0.1348 us	0.1195 us	38.251 us	38.015 us	38.461 us	1.02	0.01	-	-	-	-
Sort	Array_Comparison	\old\corerun.exe	512	37.534 us	0.1877 us	0.1756 us	37.528 us	37.295 us	37.861 us	1.00	0.00	-	-	-	-
Sort	Array_Comparison	\new\corerun.exe	512	19.358 us	0.1053 us	0.0934 us	19.344 us	19.255 us	19.582 us	0.94	0.01	-	-	-	-
Sort	Array_Comparison	\old\corerun.exe	512	20.618 us	0.0984 us	0.0920 us	20.578 us	20.522 us	20.812 us	1.00	0.00	-	-	-	-
Sort	Array_Comparison	\new\corerun.exe	512	275.570 us	1.2051 us	1.0683 us	275.188 us	274.710 us	277.900 us	0.91	0.00	-	-	-	-
Sort	Array_Comparison	\old\corerun.exe	512	303.104 us	1.0831 us	0.9045 us	303.470 us	301.982 us	304.219 us	1.00	0.00	-	-	-	-

I also ran a few additional benchmarks to try to measure impact on small arrays where other overheads might show up:

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Diagnosers;
using BenchmarkDotNet.Running;
using System;
using System.Linq;

[MemoryDiagnoser]
public class Program
{
    static void Main(string[] args) => BenchmarkSwitcher.FromAssemblies(new[] { typeof(Program).Assembly }).Run(args);

    [Params(1, 5, 10, 50)]
    public int Length { get; set; }

    private int[] _intArray;

    private static bool s_direction = false;

    private IntStruct[] _intStructArray;

    private readonly struct IntStruct : IComparable<IntStruct>
    {
        private readonly int _value;
        public IntStruct(int value) => _value = value;
        public int CompareTo(IntStruct other) => s_direction ? _value.CompareTo(other._value) : other._value.CompareTo(_value);
    }

    [GlobalSetup]
    public void Setup()
    {
        _intArray = Enumerable.Range(0, Length).ToArray();
        _intStructArray = Enumerable.Range(0, Length).Reverse().Select(i => new IntStruct(i)).ToArray();
    }

    [Benchmark]
    public void AlreadySorted() => Array.Sort(_intArray);

    [Benchmark]
    public void AlreadySorted_NonGeneric() => Array.Sort((Array)_intArray);

    [Benchmark]
    public void Inverse()
    {
        s_direction = !s_direction;
        Array.Sort(_intStructArray);
    }

    [Benchmark]
    public void Inverse_NonGeneric()
    {
        s_direction = !s_direction;
        Array.Sort(_intStructArray);
    }
}

Method	Toolchain	Length	Mean	Error	StdDev	Ratio	RatioSD
AlreadySorted	\new\corerun.exe	1	0.9878 ns	0.0083 ns	0.0073 ns	0.58	0.00
AlreadySorted	\old\corerun.exe	1	1.7149 ns	0.0022 ns	0.0020 ns	1.00	0.00
AlreadySorted_NonGeneric	\new\corerun.exe	1	12.8182 ns	0.0562 ns	0.0526 ns	1.98	0.01
AlreadySorted_NonGeneric	\old\corerun.exe	1	6.4623 ns	0.0116 ns	0.0109 ns	1.00	0.00
Inverse	\new\corerun.exe	1	0.9696 ns	0.0045 ns	0.0040 ns	0.43	0.00
Inverse	\old\corerun.exe	1	2.2722 ns	0.0014 ns	0.0011 ns	1.00	0.00
Inverse_NonGeneric	\new\corerun.exe	1	0.8297 ns	0.0022 ns	0.0019 ns	0.45	0.00
Inverse_NonGeneric	\old\corerun.exe	1	1.8548 ns	0.0028 ns	0.0026 ns	1.00	0.00
AlreadySorted	\new\corerun.exe	5	16.2019 ns	0.0233 ns	0.0206 ns	0.64	0.00
AlreadySorted	\old\corerun.exe	5	25.2719 ns	0.0340 ns	0.0301 ns	1.00	0.00
AlreadySorted_NonGeneric	\new\corerun.exe	5	51.9199 ns	0.0690 ns	0.0645 ns	1.32	0.00
AlreadySorted_NonGeneric	\old\corerun.exe	5	39.3405 ns	0.0528 ns	0.0468 ns	1.00	0.00
Inverse	\new\corerun.exe	5	28.9344 ns	0.0859 ns	0.0803 ns	0.86	0.00
Inverse	\old\corerun.exe	5	33.6274 ns	0.0996 ns	0.0931 ns	1.00	0.00
Inverse_NonGeneric	\new\corerun.exe	5	28.9753 ns	0.0835 ns	0.0652 ns	0.86	0.00
Inverse_NonGeneric	\old\corerun.exe	5	33.8191 ns	0.0875 ns	0.0776 ns	1.00	0.00
AlreadySorted	\new\corerun.exe	10	21.9842 ns	0.0601 ns	0.0532 ns	0.58	0.00
AlreadySorted	\old\corerun.exe	10	37.9305 ns	0.0944 ns	0.0737 ns	1.00	0.00
AlreadySorted_NonGeneric	\new\corerun.exe	10	57.3741 ns	0.1686 ns	0.1577 ns	1.10	0.00
AlreadySorted_NonGeneric	\old\corerun.exe	10	52.2404 ns	0.0567 ns	0.0473 ns	1.00	0.00
Inverse	\new\corerun.exe	10	90.1305 ns	0.1498 ns	0.1328 ns	0.95	0.00
Inverse	\old\corerun.exe	10	94.6786 ns	0.1084 ns	0.0905 ns	1.00	0.00
Inverse_NonGeneric	\new\corerun.exe	10	90.3232 ns	0.5252 ns	0.4656 ns	0.95	0.01
Inverse_NonGeneric	\old\corerun.exe	10	95.1975 ns	0.1188 ns	0.0928 ns	1.00	0.00
AlreadySorted	\new\corerun.exe	50	131.9916 ns	0.3596 ns	0.3187 ns	0.91	0.00
AlreadySorted	\old\corerun.exe	50	145.2568 ns	0.9565 ns	0.8947 ns	1.00	0.00
AlreadySorted_NonGeneric	\new\corerun.exe	50	172.3455 ns	0.5048 ns	0.4475 ns	1.08	0.00
AlreadySorted_NonGeneric	\old\corerun.exe	50	160.1318 ns	0.6221 ns	0.5515 ns	1.00	0.00
Inverse	\new\corerun.exe	50	390.3818 ns	0.9525 ns	0.8910 ns	0.95	0.02
Inverse	\old\corerun.exe	50	412.0761 ns	8.0010 ns	8.5609 ns	1.00	0.00
Inverse_NonGeneric	\new\corerun.exe	50	394.3612 ns	1.6422 ns	1.5361 ns	0.97	0.00
Inverse_NonGeneric	\old\corerun.exe	50	408.4902 ns	2.0410 ns	1.7043 ns	1.00	0.00

Non-generic takes a small hit in overhead, but the numbers here are also miniscule, and the non-generic methods are generally not used by anyone who cares about this level of perf.

[jkotas]

Looks awesome!

[nietras]

@stephentoub great! 👍

[jkotas]

Was it intentional to include String here?

I believe that there are subtle differences in how the current culture is handled between the non-generic and generic default comparer, so using this path for string may change the behavior.

[stephentoub]

I'll just remove this case then. Thanks.

[GSPP]

It is possible that typeof(T[]) != array.GetType(). This happens when passing in, for example, (uint[])(object)new int[10]. The CLR type system allows treating an int[] as a uint[].

I wonder if there is a behavior change here if previously the native code used the runtime type of the array and now the C# code uses the type of the generic argument.

[jkotas]

All these cases should work fine. The switch is done on TypeCode. It that prevents int[] to be confused with uint[].

[adamsitnik]

@nietras @stephentoub great work!

/cc @damageboy who is working on a vectorized version of Array.Sort ;)

[damageboy]

Definitely nice to have this back it managed land, this will make special casing my stuff much more friendly :)