JIT: Bulk copy of byrefs

[EgorBo]

Closes #101699
Closes #8627

Use a bulk copy when we need to copy a struct with gc fields (non-null) instead of individual gc write barriers

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

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

public class MyBench
{
    GcStruct4 s4 = new GcStruct4(
        100001.ToString(), 100002.ToString(),
        100003.ToString(), 100004.ToString());

    GcStruct9 s9 = default;

    [Benchmark] public object BoxS4() => s4;

    [Benchmark] public object BoxS9() => s9;
}

record struct GcStruct4(
    string a1, string a2, 
    string a3, string a4);

record struct GcStruct9(
    string a1, string a2, string a3, 
    string a4, string a5, string a6,
    string a7, string a8, string a9);

Method	Toolchain	Mean	Ratio
BoxS4	\runtime-main...\corerun.exe	8.160 ns	1.00
BoxS4	\runtime-pr...\corerun.exe	5.254 ns	0.64
BoxS9	\runtime-main...\corerun.exe	13.546 ns	1.00
BoxS9	\runtime-pr...\corerun.exe	6.136 ns	0.45

[EgorBo]

@MihuBot

[EgorBo]

@MihuBot

[EgorBo]

@MihuBot

[EgorBo]

@jkotas PTAL

[jkotas]

Thanks

[EgorBo]

@jakobbotsch cc @dotnet/jit-contrib PTAL jit part, TryLowerBlockStoreAsGcBulkCopyCall method is based on the tricks we used in LowerBlockStoreAsHelperCall (morphing calls in lower).

[jakobbotsch]

Do we need to worry about any indir flags here, like GTF_IND_VOLATILE?

[EgorBo]

Good question, I think it's safer to just give up on those here, fixed.

[EgorBo]

Failures are known - #101721, #101618

[EgorBo]

Improvements on x64:

• [Perf] Linux/x64: 34 Improvements on 5/3/2024 3:54:35 PM perf-autofiling-issues#33896
• [Perf] Linux/arm64: 26 Improvements on 5/3/2024 3:54:35 PM perf-autofiling-issues#34054