MarkExplicitInterfaceImplementation doesn't work for generic interfaces

[MichalStrehovsky]

https://github.com/mono/linker/blob/9e3191cad279e945a4e2c99340454959fec637b0/src/linker/Linker.Steps/MarkStep.cs#L2469-L2492

This code is trying to ensure that if an explicit interface implementation method was marked, the appropriate member of the interface list is also preserved.

The code falls victim to Cecil's behavior where the act of resolving a TypeRef/MethodRef strips away the genericness information.

As a result, for code like this:

class Foo : IFoo<object>, IFoo<string>
{
    int IFoo<object> Frob() => 0;
    int IFoo<string> Frob() => 0;
}

Will end up preserving the first IFoo entry in the interface list, irrespective of whether IFoo<object>.Frob or IFoo<string>.Frob was the one responsible for MarkExplicitInterfaceImplementation getting called.

I think this is low impact, since we can only get there if the explicit impl was rooted through ILLinkTrim, but ILLinkTrim was the motivating factor why MarkExplicitInterfaceImplementation was added in the first place, so YMMV.

Found while investigating #1421. This issue gives the other issue a more weird failure mode.

[am11]

The code falls victim to Cecil's behavior where the act of resolving a TypeRef/MethodRef strips away the genericness information.

Was this PR trying to address the same issue: jbevain/cecil#229?

[MichalStrehovsky]

Was this PR trying to address the same issue: jbevain/cecil#229?

That looks like a different thing. Linker rolls its own override resolution logic and doesn't call into Cecil for this.

What I'm referring to is a fundamental design choice in Cecil and not a bugfix thing - it would break a lot of existing code based on Cecil. CCI has resolution logic that IMHO works better and doesn't come with this property/trap.