Stack overflow exception when creating second Fake of type that takes a parameter of its own type

We are facing an issue while running the following tests:

using System;
using FakeItEasy;
using NUnit.Framework;

namespace ConsoleApp1
{
    [TestFixture]
    public class ClassTest
    {
        [SetUp]
        public void SetUp()
        {
        }

        [Test]
        public void Test1()
        {
            var fake = A.Fake<Class>();
        }

        [Test]
        public void Test2()
        {
            var fake = A.Fake<Class>();
        }
    }
    
    public class Class
    {
        public Class(Type type)
        {
        }

        public Class(Type type, Class @class)
        {
        }
    }
}

Only when they are executed together, the second test (Test2) fails with a Stack Overflow Exception. The first test instead has no problem. Until version 4.6.0 of FakeItEasy this sample works correctly.

[blairconrad]

I have a cause and a fix (figured out just as I was leaving the house this mornining).

Short version is:

• we try to fake Class, which attempts to resolve Dummy parameters for each constructor, longest first.
• making a Dummy Type is fine, so we try to make a Dummy for Class
• the Dummy-creation logic falls back to its “make a Dummy by making a Fake” logic, which does the same sort of thing as above
• we can make a Type as before, but when we try to make a Dummy Class, we figure out that we’re already making a Class, so we back off and try the short constructor, with just Type arguments
• We make the Fake and in the “best constructor” cache record that the one-parameter constructor is preferred, so we never have to try all this again
• We pop back up the stack and then make a Fake Class at the “make a Fake” level, using our Dummy Type and Dummy Class and then… we record that the two-parameter constructor is best

When we try to fake the Class the second time, we see the two-parameter constructor is best, and try to use it, resulting in stackoverflow.

If we replae https://github.com/FakeItEasy/FakeItEasy/blob/02e6f4ef3e0ae49fdd3e712dd2e523a025a94d69/src/FakeItEasy/Creation/FakeObjectCreator.cs#L214

with

this.parameterTypesCache.TryAdd(typeOfFake, parameterTypes);

The tests pass. I think this is a good change, as we could have competing construction from different threads. There may be other caches that would benefit from this sort of change.

A larger change would be to enroll Fake creation in the loop-detection mechanism that we have for creating a Dummy, but I think that amounts to a performance optimization, since it wouldn’t help with a multi-threaded problem.

I can whip a PR up (with spec) for the reported problem certainly by tomorrow, or maybe today if work allows me a few minutes.

[blairconrad]

Oh, good. I ended up leaning the same way. PR incoming!