Generic recursion and generic dictionaries

[jkotas]

Repro

Console.WriteLine(new G<string>().M());

class G<T>
{
    public object M()
    {
        return new G<List<T>>();
    }
}

Actual result:

AOT analysis warning IL3054: [repro]G<System.Collections.Generic.List<System.Collections.Generic.List<Syste m.Collections.Generic.List<System.Collections.Generic.List<string>>>>>: Generic expansion to 'G<List<List<Lis t<List<List<String>>>>>>' was aborted due to generic recursion.

Expected result:

No warning

Context: #85176 (comment)

Tagging subscribers to this area: @dotnet/area-system-collections
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Repro:

Console.WriteLine(new G<string>().M());

class G<T>
{
    public object M()
    {
        return new G<List<T>>();
    }
}

Actual result: AOT analysis warning IL3054: [repro]G<System.Collections.Generic.List<System.Collections.Generic.List<Syste m.Collections.Generic.List<System.Collections.Generic.List<string>>>>>: Generic expansion to 'G<List<List<Lis t<List<List<String>>>>>>' was aborted due to generic recursion.

Expected result: No warning

Context: #85176 (comment)

Author:	jkotas
Assignees:	-
Labels:	area-System.Collections
Milestone:	-

Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Repro

Console.WriteLine(new G<string>().M());

class G<T>
{
    public object M()
    {
        return new G<List<T>>();
    }
}

Actual result:

AOT analysis warning IL3054: [repro]G<System.Collections.Generic.List<System.Collections.Generic.List<Syste m.Collections.Generic.List<System.Collections.Generic.List<string>>>>>: Generic expansion to 'G<List<List<Lis t<List<List<String>>>>>>' was aborted due to generic recursion.

Expected result:

No warning

Context: #85176 (comment)

Author:	jkotas
Assignees:	-
Labels:	area-System.Collections, untriaged, area-NativeAOT-coreclr
Milestone:	-

[MichalStrehovsky]

Is there anything that can be done for this? Both Gstring and GListString are backed by GCanon. So they need the same generic dictionary layout.

We could in theory make a concrete dictionary with a hole in it (the slot still has to be there because we only have one version of code) but we no longer track specializations for this to be feasible. E.g. reflection and generic virtuals are analyzed purely on canon forms. There are also additional problems with inlining (RyuJIT inlines canon bodies, not instantiated bodies).

This will likely be one of those issues that is forever in the backlog.

[jkotas]

Or we can leave these recursive slots uninitialized and initialize them lazily at runtime.

[MichalStrehovsky]

Ah, so it's not about more precise analysis, just different cutoff for the generic recursion. Got confused by the title. Yep, this should not be too hard.

[jkotas]

Updated the title. More precise analysis was the first thing that came to my mind as a solution of the problem, but it is not the only possible solution.

[eiriktsarpalis]

My impression is that a fix for this would supersede #85176. Which would be nice since we'd get to keep the APIs that the workaround removes. I'm going to merge #85176 so that the issue is addressed for Preview 4 but we should try reverting once a compiler fix becomes available.

[MichalStrehovsky]

A potential NativeAOT fix won't help the customer who originally reported the issue on .NET Native (different product).

The potential fix for NativeAOT is still a deoptimization and we're going to report a warning. It just won't have the size impact, only runtime throughput impact.

Generic recursion is AOT cryptonite.