Regression from first-class spans with overload picking

[colejohnson66]

Version Used: 9.0.200.preview

Steps to Reproduce:

1. Use MemoryMarshal.Cast with an array
2. Expect Span<T> overload, but get ReadOnlySpan<T> overload

https://sharplab.io/#v2:EYLgtghglgdgNAExAagD4AEBMBGAsAKHQAYACdbAOgCUBXGAFyjAFMKBJB5gJwHsAHAMrcAblADGzAM4BuAgXQBmMphIBhADYRJkggG8CJQ2SXoALCQCyzegAseCABQBKA0f34jnkgh41g65gBtAF0Segh/ZhIAXhIYGnV1AEJZDy9DAT4IGAAeBJ4YAHMAPm8pehjLZjAeLgBPCwguSRsIdQpVLXocnz8AuBJ8ouKHcMinVM8AXwIpoA===

using System;
using System.Runtime.InteropServices;

public class Class
{
    public void Method()
    {
        double[] table = null!;
        Span<ulong> dest = MemoryMarshal.Cast<double, ulong>(table);
    }
}

Diagnostic Id:

error CS0029: Cannot implicitly convert type 'System.ReadOnlySpan' to 'System.Span'

Expected Behavior:

The Span<T> overload is used, as it was on .NET 8 and .NET 9 RC2.

Actual Behavior:

The ReadOnlySpan<T> overload is used instead.

[colejohnson66]

Forcing the call to use the Span<T> overload (with either a cast or MemoryExtensions.AsSpan) causes the issue to go away.

[jcouv]

Tagging @jjonescz to confirm, but I think this will be fixed by #76300

[colejohnson66]

If this is intended as a breaking change, that’s fine with me (I’ve already fixed the code). It just isn’t documented as one.

[jjonescz]

This is an expected breaking change. ReadOnlySpan is now preferred over Span to avoid ArrayTypeMismatchExceptions that would occur at runtime if you used covariant arrays.

See also https://github.com/dotnet/csharplang/blob/main/meetings/2024/LDM-2024-12-04.md#preferring-readonlyspant-over-spant-conversions.

I think this will be fixed by #76300

That PR doesn't fix it, it makes it broader in fact, per LDM discussion linked above.

It just isn’t documented as one.

I will open a PR to document the break, thanks.