Test: simplify gc memmove

[EgorBo]

No description provided.

[EgorBo]

@EgorBot -intel -arm64 -profiler -nonativepgo

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

[GcServer(true)]
public class Bench
{
    static object[] Src0 = Enumerable.Range(0, 2).Select(i => (object)i).ToArray();
    static object[] Src1 = Enumerable.Range(0, 10).Select(i => (object)i).ToArray();
    static object[] Src2 = Enumerable.Range(0, 128).Select(i => (object)i).ToArray();
    static object[] Dst = new object[128];

    [Benchmark]public void CopyObjects2() => Src0.AsSpan().CopyTo(Dst);
    [Benchmark]public void CopyObjects10() => Src1.AsSpan().CopyTo(Dst);
    [Benchmark]public void CopyObjects128() => Src2.AsSpan().CopyTo(Dst);
}

[huoyaoyuan]

Is auto-vectorization by C++ compiler expected to kick in?

[EgorBo]

Is auto-vectorization by C++ compiler expected to kick in?

I am just curious about the impact of the logic we have here.

1. It seems that the current logic it's not 100% safe as it uses SIMD on x64. SIMD on x64 demands 16 byte alignment for both src and dest to be atomic, but we only align dest.
2. I've observed a lot of "Load with an additional latency due to misalignment" PMU counters pointing to this function on arm64 with perf

UPD: Results suggest the plain loop is faster? EgorBot/runtime-utils#123 (not sure about windows/msvc)

[huoyaoyuan]

How about the numbers at
#101699 (comment) ? I was basically doing the same thing for NoSSE.

[EgorBo]

@EgorBot -arm64 -profiler -nonativepgo -perf_events ld_align_lat

using BenchmarkDotNet.Attributes;

public class Bench
{
    static object[] Src = Enumerable.Range(0, 10).Select(i => (object)i).ToArray();
    static object[] Dst = new object[128];

    [Benchmark]public void CopyObjects10() => Src.AsSpan().CopyTo(Dst);
}

[EgorBo]

How about the numbers at #101699 (comment) ? I was basically doing the same thing for NoSSE.

do you mean AVX numbers? I guess the atomicity concern still stands

[EgorBo]

@EgorBot -intel -arm64 -profiler -nonativepgo

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

[GcServer(true)]
public class Bench
{
    static object[] Src0 = Enumerable.Range(0, 2).Select(i => (object)i).ToArray();
    static object[] Src1 = Enumerable.Range(0, 10).Select(i => (object)i).ToArray();
    static object[] Src2 = Enumerable.Range(0, 128).Select(i => (object)i).ToArray();
    static object[] Dst = new object[128];

    [Benchmark]public void CopyObjects2() => Src0.AsSpan().CopyTo(Dst);
    [Benchmark]public void CopyObjects10() => Src1.AsSpan().CopyTo(Dst);
    [Benchmark]public void CopyObjects128() => Src2.AsSpan().CopyTo(Dst);
}

[EgorBo]

With volatile it's obviosly a regression on large inputs, but the PMU counters for unaligned loads look happy:

Main:

 563591328      ld_align_lat
         0      st_align_lat
2801225037      unaligned_ld_spec
2780268499      unaligned_st_spec

PR:

     19048      ld_align_lat
         9      st_align_lat
         0      unaligned_ld_spec
         0      unaligned_st_spec