runtime: .Net 7 performance regression with struct fields

Description

I noticed a ~5-15% performance regression when benchmarking my library ProtoPromise in .Net 6 vs .Net 7. I wasn’t sure what the cause could be, so I tried making a simplified benchmark.

public class PromiseBenchmarks
{
    private Promise.Deferred deferred;

    [Benchmark]
    public void PromiseVoid()
    {
        deferred = Promise.NewDeferred();
        deferred.Promise.Forget();
        deferred.Resolve();
        deferred = default;
    }
}

Then I realized I didn’t need the field and used a local instead.

public class PromiseBenchmarks
{
    private Promise.Deferred deferred;

    [Benchmark]
    public void PromiseField()
    {
        deferred = Promise.NewDeferred();
        deferred.Promise.Forget();
        deferred.Resolve();
        deferred = default;
    }

    [Benchmark]
    public void PromiseLocal()
    {
        var deferred = Promise.NewDeferred();
        deferred.Promise.Forget();
        deferred.Resolve();
    }
}

And that yielded these surprising results

Method	Runtime	Mean	Error	StdDev	Ratio	Code Size
PromiseField	.NET 6.0	93.05 ns	1.227 ns	1.148 ns	1.00	448 B
PromiseField	.NET 7.0	100.94 ns	2.033 ns	2.342 ns	1.08	425 B
PromiseLocal	.NET 6.0	97.22 ns	1.916 ns	1.792 ns	1.00	438 B
PromiseLocal	.NET 7.0	92.66 ns	1.617 ns	1.513 ns	0.95	361 B

Clearly .Net 7 is generating better code for the actual work, but there’s something weird with using the field rather than local. The Promise.Deferred struct is simply this:

public struct Deferred
{
    internal readonly Internal.PromiseRefBase.DeferredPromise<Internal.VoidResult> _ref; // class reference.
    internal readonly short _promiseId;
    internal readonly int _deferredId;
}

Configuration

I ran this benchmark with BenchmarkDotNet arguments --runtimes net6.0 net7.0 --disasm --disasmDepth 10000

OS is Windows 10 x64 CPU is AMD Phenom II X6 1055T