MethodTable::IsClassInited returns false for <PrivateImplementationDetails> class

[EgorBo]

See #104054 for discussion.

public class Program
{
    static void Main()
    {
        Test();
    }

    static ReadOnlySpan<int> RvaData => [1, 2, 3, 4];

    [MethodImpl(MethodImplOptions.NoInlining)]
    static int Test() => RvaData[2];
}

This code (TieredCompilation=0) used to fold RvaData[2] into 3, it's no longer happening because getStaticFieldContent gives up on <PrivateImplementationDetails> being non-initialized statically (despite the fact it has no cctor)

Previously, it used to return true and everything was correctly folded, it's no longer the case after #99183

[dotnet-policy-service]

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

[jkotas]

I think the problem is more in the JIT/EE interface.

Before the change, CORINFO_FLG_FIELD_INITCLASS means initialization is required.

After the change, CORINFO_FLG_FIELD_INITCLASS means initialization may be required, call initClass to be sure.

We either need to fix JIT/EE interface to have the old behavior; or we need fix the JIT to account for the new behavior.

Alternatively, we can get rid of CORINFO_FLG_FIELD_INITCLASS and require JIT to always call initClass to figure out whether the initialization is required.

[EgorBo]

and require JIT to always call initClass to figure out whether the initialization is required.

so who (and when) is going to call initClass in case of CreateSpan(ldtoken) ?

[jkotas]

Why does initClass need to be called for CreateSpan(ldtoken) in the first place? The unexpanded implementation of CreateSpan(ldtoken) does not trigger static constructors, so I do not see why we need to call initClass in the intrinsic expansion of CreateSpan(ldtoken).

[EgorBo]

Why does initClass need to be called for CreateSpan(ldtoken) in the first place? The unexpanded implementation of CreateSpan(ldtoken) does not trigger static constructors, so I do not see why we need to call initClass in the intrinsic expansion of CreateSpan(ldtoken).

um.. so it means nobody calls initClass for <PrivateImplementationDetails> (since the only way it's referenced is via CreateSpan(ldtoken)) and then getStaticFieldContent always refuses to fold anything since IsClassInited will return false. I feel like I am missing something 😐 how is JIT-EE fixup going to help here?

[jkotas]

getStaticFieldContent always refuses to fold anything since IsClassInited will return false.

getStaticFieldContent may want to call AttemptToPreinit() before calling IsClassInitialized.

[davidwrighton]

I think we would prefer to make a new function... I'm calling it IsClassInitedOrPreinited which will will do the IsClassInited check + the AttemptToPreinit + a final IsClassInited. It looks like there may be several places in the JIT interface which should be calling it and just using the more complete function is a more reliable fix. I'm taking a look at making that change, which seems fairly simple.