JIT: Loop alignment does not kick in for a tight "sum" loop

[jakobbotsch]

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Configs;
using BenchmarkDotNet.Environments;
using BenchmarkDotNet.Jobs;
using BenchmarkDotNet.Running;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;

var config = DefaultConfig.Instance
    .AddJob(Job.Default.WithRuntime(CoreRuntime.Core90));

BenchmarkSwitcher.FromAssembly(typeof(Program).Assembly).Run(args, config);

public class Benchmark
{
    private int[] _arr = null!;

    [Params(100, 1000, 10000)]
    public int Count { get; set; }

    [GlobalSetup]
    public void GlobalSetup()
    {
        _arr = new int[Count]; 
    }

    [Benchmark]
    public void Sum2()
    {
        Sum_bounds_check(_arr);
    }

    [Benchmark]
    public void Sum3()
    {
        Sum_no_bounds_check(_arr);
    }

    private static int Sum_bounds_check(int[] arr)
    {
        int sum = 0;
        for (nint i = 0; i < arr.Length; i++)
        {
            sum += arr[i];
        }

        return sum;
    }

    private static int Sum_no_bounds_check(int[] arr)
    {
        int sum = 0;
        for (nint i = 0; i < arr.Length; i++)
        {
            sum += Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(arr), i);
        }

        return sum;
    }
}

In this benchmark Sum_bounds_check has a bounds check in it (#104659). That makes the loop body big enough for loop alignment to kick in, which aligns it so that the loop starts at 0x20.
OTOH Sum_no_bounds_check is not aligned by the JIT, and its loop ends up starting at 0xD. Overall this seems to cause Sum_bounds_check to be faster than Sum_no_bounds_check, even though it has the extra bounds check:

Method	Count	Mean	Error	StdDev
Sum_bounds_check	100	27.81 ns	0.172 ns	0.143 ns
Sum_no_bounds_check	100	29.53 ns	0.156 ns	0.146 ns
Sum_bounds_check	1000	219.52 ns	2.484 ns	2.202 ns
Sum_no_bounds_check	1000	239.76 ns	1.713 ns	1.518 ns
Sum_bounds_check	10000	2,163.16 ns	3.656 ns	3.420 ns
Sum_no_bounds_check	10000	2,232.11 ns	7.284 ns	6.813 ns

(from my 5950X).

cc @kunalspathak