Remove Interlocked.{Compare}Exchange generic constraint

[GSPP]

EDITED by @stephentoub on 7/9/2024 to update the proposal:

Remove the T : class constraint from the existing Interlocked.{Compare}Exchange generic methods. They will be usable with not only reference type Ts but also primitive and enum Ts. Unsupported Ts will result in a NotSupportedException.

namespace System.Threading;

public static class Interlocked
{
    public static T CompareExchange<T> (ref T location1, T value, T comparand)
-       where T : class

    public static T Exchange<T>(ref T location1, T value)
-       where T : class
}

(We could also consider removing / making internal the new public overloads added in .NET 9 for byte/sbyte/ushort/short and just have the generic version to handle those.)

---

In principle, Interlocked.CompareExchange can be made to work on enums. Enums are just integers in disguise.

Motivating for this request is a repeated need to update a "state" field in a class. The state is most cleanly tracked as an enum. Since this API does not exist yet, I use integers and constants. That's of course viable but it would be cleaner to be able to use enums.

A workaround is to use Unsafe. There is nothing wrong with that in principle. It could be "nicer", though.

There's a recent discussion about small integer types that is pertinent to this issue:

• [API Proposal]: Interlocked.CompareExchange for byte/short/sbyte/ushort #64658

This is related because enums can have small integer types. If we don't allow small types, then the generic constraint : enum would not ensure runtime safety.

So maybe we can add enum overloads iff we add small integer types. The alternative would be to add it even without that and fail at runtime.

Design issues:

1. What APIs should be augmented like this in addition to CompareExchange. I can tell from the existing discussion that the team only wants to add API scope if there is demand. In principle, all Interlocked APIs could be upgraded.
2. Should there be an API to process any unmanaged type of at most 64 bits in size?
   a. This would mean that there could be a runtime error depending on T. If T : unmanaged, then the JIT can ensure that this validation is without runtime overhead because each struct type generates specialized code.
   b. Could this result in unaligned access?
3. Taking this further: There could be a generic type constraint "interlocked". This would then even work with things like ImmutableArray. Most likely, this is not going to happen, but I'm mentioning this for completeness.

---

For visitors looking for a workaround:

static T InterlockedCompareExchange32<T>(ref T location, T value, T comparand) where T : unmanaged
{
    var result = Interlocked.CompareExchange(ref Unsafe.As<T, int>(ref location), Unsafe.As<T, int>(ref value), Unsafe.As<T, int>(ref comparand));
    return Unsafe.As<int, T>(ref result);
}

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[MihaZupan]

For reference / as an example, we are working around the lack of such an API in Uri:

• Interlocked.CompareExchange: here
• Interlocked.Or: here, here and here

Tagging subscribers to this area: @mangod9
See info in area-owners.md if you want to be subscribed.

Issue Details

---

In principle, Interlocked.CompareExchange can be made to work on enums. Enums are just integers in disguise.

Motivating for this request is a repeated need to update a "state" field in a class. The state is most cleanly tracked as an enum. Since this API does not exist yet, I use integers and constants. That's of course viable but it would be cleaner to be able to use enums.

A workaround is to use Unsafe. There is nothing wrong with that in principle. It could be "nicer", though.

There's a recent discussion about small integer types that is pertinent to this issue:

• [API Proposal]: Interlocked.CompareExchange for byte/short/sbyte/ushort #64658

This is related because enums can have small integer types. If we don't allow small types, then the generic constraint : enum would not ensure runtime safety.

So maybe we can add enum overloads iff we add small integer types. The alternative would be to add it even without that and fail at runtime.

Design issues:

1. What APIs should be augmented like this in addition to CompareExchange. I can tell from the existing discussion that the team only wants to add API scope if there is demand. In principle, all Interlocked APIs could be upgraded.
2. Should there be an API to process any unmanaged type of at most 64 bits in size?
   a. This would mean that there could be a runtime error depending on T. If T : unmanaged, then the JIT can ensure that this validation is without runtime overhead because each struct type generates specialized code.
   b. Could this result in unaligned access?
3. Taking this further: There could be a generic type constraint "interlocked". This would then even work with things like ImmutableArray. Most likely, this is not going to happen, but I'm mentioning this for completeness.

---

For visitors looking for a workaround:

static T InterlockedCompareExchange32<T>(ref T location, T value, T comparand) where T : unmanaged
{
    var result = Interlocked.CompareExchange(ref Unsafe.As<T, int>(ref location), Unsafe.As<T, int>(ref value), Unsafe.As<T, int>(ref comparand));
    return Unsafe.As<int, T>(ref result);
}


<table>
  <tr>
    <th align="left">Author:</th>
    <td>GSPP</td>
  </tr>
  <tr>
    <th align="left">Assignees:</th>
    <td>-</td>
  </tr>
  <tr>
    <th align="left">Labels:</th>
    <td>

`api-suggestion`, `area-System.Threading`, `untriaged`

</td>
  </tr>
  <tr>
    <th align="left">Milestone:</th>
    <td>-</td>
  </tr>
</table>
</details>

[quixoticaxis]

Is it even possible to implement this API effectively without making it a special case in the compiler?
AFAIU, you want the API to implicitly cast the enum to its underlying type, and it cannot be done without accessing the type information through Enum.GetUnderlyingType and switching on the result in runtime which may not be fast enough.
The only other way is to force the compiler to inject no-op casts at CompareExhange call sites.

And sure the shorter integer types would have to be supported first anyway.

[jkotas]

Is it even possible to implement this API effectively without making it a special case in the compiler?

Right, the implementation of this API would have to be special-cased as runtime intrinsic. It is not something that can be done efficiently in pure C# today.

[NinoFloris]

This results in the desired codegen though, we've been using this for a while in our codebase.
Optionally you could add the Enum constraint, does this approach have some fundamental flaw I'm unaware of?

using System;
using System.Runtime.CompilerServices;
using System.Threading;

[MethodImpl(MethodImplOptions.AggressiveInlining)]
static unsafe T CompareExchange<T>(ref T location, T value, T comparand)
    where T: unmanaged
{
    switch (sizeof(T))
    {
        case sizeof(int):
            var x = Interlocked.CompareExchange(ref Unsafe.As<T, int>(ref location), Unsafe.As<T, int>(ref value), Unsafe.As<T, int>(ref comparand));
            return Unsafe.As<int,T>(ref x);
        default:
            ThrowUnsupported();
            return default;
    }

    static void ThrowUnsupported() => throw new NotSupportedException();
}

Sharplab against StringComparison:
https://sharplab.io/#v2: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=