When does using C# structs (value types) sacrifice performance?

Question

I have been playing with structs as a mechanism to implicitly validate complex value objects, as well as generic structs around more complex classes to ensure valid values. I am a little ignorant as to the performance consequences, so I am hoping you all can help me out. For example, if I were to do something like injecting a domain object into a value type wrapper, would that cause problems ? Why? I understand the difference between value types and reference types, and my goal here is to leverage the different behavior of value types. What exactly do I need to look into in order to do this responsibly?

Here is an extremely basic idea of something I was thinking.

public struct NeverNull<T>
    where T: class, new()
{

    private NeverNull(T reference)
    {
        _reference = reference;
    }

    private T _reference;

    public T Reference
    {
        get
        {
            if(_reference == null)
            {
                _reference = new T();
            }
            return _reference;
        }
        set
        {
            _reference = value;
        }
    }

    public static implicit operator NeverNull<T>(T reference)
    {
        return new NeverNull<T>(reference);
    }

    public static implicit operator T(NeverNull<T> value)
    {
        return value.Reference;
    }
}

Answer 1

Well, one nasty thing is that this doesn't behave as you might expect it to naively:

NeverNull<Foo> wrapper1 = new NeverNull<Foo>();
NeverNull<Foo> wrapper2 = wrapper1;

Foo foo1 = wrapper1;
Foo foo2 = wrapper2;

That's going to create two instances of Foo because the original version was copied, before wrapper1 created an instance.

Basically, you're dealing with a mutable struct - which is almost never a nice thing to have. Additionally, I'm not generally keen on implicit conversions.

It feels like you're trying to achieve magic-looking code here... and I'm generally against that sort of thing. Maybe it makes sense for your particular use case, but I can't think of where I'd personally want to use it.

Answer 2

As Jon correctly points out, the problem here is that the behaviour of the type is unexpected, not that it is slow. From a performance perspective, the overhead of the struct wrapper around the reference should be very low.

If what you want to do is to represent a non-nullable reference type then a struct is a reasonable way to do it; however, I would be inclined to make the struct immutable by losing the "auto creation" feature:

public struct NeverNull<T> where T: class 
{ 
    private NeverNull(T reference) : this()
    { 
        if (reference == null) throw new Exception(); // Choose the right exception
        this.Reference = reference; 
    } 

    public T Reference { get; private set; }

    public static implicit operator NeverNull<T>(T reference) 
    { 
        return new NeverNull<T>(reference); 
    } 

    public static implicit operator T(NeverNull<T> value) 
    { 
        return value.Reference; 
    } 
}

Make the caller responsible for providing a valid reference; if they want to "new" one up, let 'em.

Note also that generic conversion operators can give you unexpected results. You should read the spec on conversion operators and understand it thoroughly. For example, you cannot make a non-null wrapper around "object" and then have that thing implicitly convert to the unwrapping conversion; every implicit conversion to object will be a boxing conversion on the struct. You cannot "replace" a built-in conversion of the C# language.

Answer 3

The main penalty is with boxing for structs. Also they are passed by value so a large struct when passed to a method will have to be copied:

MyStruct st;
foo.Bar(st); // st is copied

Answer 4

Ok, just a note on the above.

MyStruct st; foo.Bar(st); // st is copied

This isn't boxing, unless the parameter of Bar is object for example.

void Bar(MyStruct parameter){}

would not box the value type.

Parameters are passed by value in c# by default, unless you use the ref or out keyword. Parameters passed by value are copied. The difference between passing the struct and the object is what is passed. With a value type, the actual value is copied in, which means a new value type is created, so you end up with a copy. With a reference type the reference to the reference type is passed in. Clues in the name I guess :)

So there is a performance hit for structs because the whole structure is copied unless you use the ref/out keyword, and if you're doing that extensively, I think your code needs looking at.

Boxing is the process of assigning a value type to a reference type variable. A new reference type (object) is created and a copy of the value type assigned into it.

I sort of got what you were doing in the original code, but it seems to be solving one simple problem with one that has many implicit rather than explicit complexities.

Answer 5

The answers in this question seem to have moved away from discussing performance and instead are addressing the dangers of mutable value types.

Just in case you find this useful, here is an implementation that i threw together that does something similar to your original example using an immutable value type wrapper.

The difference is that my value type doesn't directly reference the object to which it refers; instead it holds a key and references to delegates that perform either a lookup using the key (TryGetValueFunc) or a create using the key. (Note: my original implementation had the wrapper holding a reference to an IDictionary object, but I changed it to a TryGetValueFunc delegate just to make it a bit more flexible, though that may be more confusing, and I'm not 100% sure that doing so didn't open up some kind of flaw).

Note, though, that this could still result in unexpected behavior (depending on what you expect), if you are manipulating the underlying data structures that the wrapper accesses.

Below is a complete working example, along with a usage example for a console program:

public delegate bool TryGetValueFunc<TKey, TValue>(TKey key, out TValue value);

public struct KeyedValueWrapper<TKey, TValue>
{
    private bool _KeyHasBeenSet;
    private TKey _Key;
    private TryGetValueFunc<TKey, TValue> _TryGetValue;
    private Func<TKey, TValue> _CreateValue;

    #region Constructors

    public KeyedValueWrapper(TKey key)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _TryGetValue = null;
        _CreateValue = null;
    }

    public KeyedValueWrapper(TKey key, TryGetValueFunc<TKey, TValue> tryGetValue)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _TryGetValue = tryGetValue;
        _CreateValue = null;
    }

    public KeyedValueWrapper(TKey key, Func<TKey, TValue> createValue)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _TryGetValue = null;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(TKey key, TryGetValueFunc<TKey, TValue> tryGetValue, Func<TKey, TValue> createValue)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _TryGetValue = tryGetValue;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(TryGetValueFunc<TKey, TValue> tryGetValue)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _TryGetValue = tryGetValue;
        _CreateValue = null;
    }

    public KeyedValueWrapper(TryGetValueFunc<TKey, TValue> tryGetValue, Func<TKey, TValue> createValue)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _TryGetValue = tryGetValue;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(Func<TKey, TValue> createValue)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _TryGetValue = null;
        _CreateValue = createValue;
    }

    #endregion

    #region "Change" methods

    public KeyedValueWrapper<TKey, TValue> Change(TKey key)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, _TryGetValue, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, TryGetValueFunc<TKey, TValue> tryGetValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, tryGetValue, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, _TryGetValue, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, TryGetValueFunc<TKey, TValue> tryGetValue, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, tryGetValue, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TryGetValueFunc<TKey, TValue> tryGetValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, tryGetValue, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TryGetValueFunc<TKey, TValue> tryGetValue, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, tryGetValue, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, _TryGetValue, createValue);
    }

    #endregion

    public TValue Value
    {
        get
        {
            if (!_KeyHasBeenSet)
                throw new InvalidOperationException("A key must be specified.");

            if (_TryGetValue == null)
                throw new InvalidOperationException("A \"try get value\" delegate must be specified.");

            // try to find a value in the given dictionary using the given key
            TValue value;
            if (!_TryGetValue(_Key, out value))
            {
                if (_CreateValue == null)
                    throw new InvalidOperationException("A \"create value\" delegate must be specified.");

                // if not found, create a value
                value = _CreateValue(_Key);
            }
            // then return that value
            return value;
        }
    }
}

class Foo
{
    public string ID { get; set; }
}

class Program
{
    static void Main(string[] args)
    {
        var dictionary = new Dictionary<string, Foo>();

        Func<string, Foo> createValue = (key) =>
        {
            var foo = new Foo { ID = key };
            dictionary.Add(key, foo);
            return foo;
        };

        // this wrapper object is not useable, since no key has been specified for it yet
        var wrapper = new KeyedValueWrapper<string, Foo>(dictionary.TryGetValue, createValue);

        // create wrapper1 based on the wrapper object but changing the key to "ABC"
        var wrapper1 = wrapper.Change("ABC");
        var wrapper2 = wrapper1;

        Foo foo1 = wrapper1.Value;
        Foo foo2 = wrapper2.Value;

        Console.WriteLine("foo1 and foo2 are equal? {0}", object.ReferenceEquals(foo1, foo2));
        // Output: foo1 and foo2 are equal? True

        // create wrapper1 based on the wrapper object but changing the key to "BCD"
        var wrapper3 = wrapper.Change("BCD");
        var wrapper4 = wrapper3;

        Foo foo3 = wrapper3.Value;
        dictionary = new Dictionary<string, Foo>(); // throw a curve ball by reassigning the dictionary variable
        Foo foo4 = wrapper4.Value;

        Console.WriteLine("foo3 and foo4 are equal? {0}", object.ReferenceEquals(foo3, foo4));
        // Output: foo3 and foo4 are equal? True

        Console.WriteLine("foo1 and foo3 are equal? {0}", object.ReferenceEquals(foo1, foo3));
        // Output: foo1 and foo3 are equal? False
    }
}

Alternative implementation using IDictionary<string, Foo> instead of TryGetValueFunc<string, Foo>. Note the counter-example I put in the usage code:

public struct KeyedValueWrapper<TKey, TValue>
{
    private bool _KeyHasBeenSet;
    private TKey _Key;
    private IDictionary<TKey, TValue> _Dictionary;
    private Func<TKey, TValue> _CreateValue;

    #region Constructors

    public KeyedValueWrapper(TKey key)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _Dictionary = null;
        _CreateValue = null;
    }

    public KeyedValueWrapper(TKey key, IDictionary<TKey, TValue> dictionary)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _Dictionary = dictionary;
        _CreateValue = null;
    }

    public KeyedValueWrapper(TKey key, Func<TKey, TValue> createValue)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _Dictionary = null;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(TKey key, IDictionary<TKey, TValue> dictionary, Func<TKey, TValue> createValue)
    {
        _Key = key;
        _KeyHasBeenSet = true;
        _Dictionary = dictionary;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(IDictionary<TKey, TValue> dictionary)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _Dictionary = dictionary;
        _CreateValue = null;
    }

    public KeyedValueWrapper(IDictionary<TKey, TValue> dictionary, Func<TKey, TValue> createValue)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _Dictionary = dictionary;
        _CreateValue = createValue;
    }

    public KeyedValueWrapper(Func<TKey, TValue> createValue)
    {
        _Key = default(TKey);
        _KeyHasBeenSet = false;
        _Dictionary = null;
        _CreateValue = createValue;
    }

    #endregion

    #region "Change" methods

    public KeyedValueWrapper<TKey, TValue> Change(TKey key)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, _Dictionary, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, IDictionary<TKey, TValue> dictionary)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, dictionary, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, _Dictionary, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(TKey key, IDictionary<TKey, TValue> dictionary, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(key, dictionary, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(IDictionary<TKey, TValue> dictionary)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, dictionary, _CreateValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(IDictionary<TKey, TValue> dictionary, Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, dictionary, createValue);
    }

    public KeyedValueWrapper<TKey, TValue> Change(Func<TKey, TValue> createValue)
    {
        return new KeyedValueWrapper<TKey, TValue>(_Key, _Dictionary, createValue);
    }

    #endregion

    public TValue Value
    {
        get
        {
            if (!_KeyHasBeenSet)
                throw new InvalidOperationException("A key must be specified.");

            if (_Dictionary == null)
                throw new InvalidOperationException("A dictionary must be specified.");

            // try to find a value in the given dictionary using the given key
            TValue value;
            if (!_Dictionary.TryGetValue(_Key, out value))
            {
                if (_CreateValue == null)
                    throw new InvalidOperationException("A \"create value\" delegate must be specified.");

                // if not found, create a value and add it to the dictionary
                value = _CreateValue(_Key);
                _Dictionary.Add(_Key, value);
            }
            // then return that value
            return value;
        }
    }
}

class Foo
{
    public string ID { get; set; }
}

class Program
{
    static void Main(string[] args)
    {
        // this wrapper object is not useable, since no key has been specified for it yet
        var wrapper = new KeyedValueWrapper<string, Foo>(new Dictionary<string, Foo>(), (key) => new Foo { ID = key });

        // create wrapper1 based on the wrapper object but changing the key to "ABC"
        var wrapper1 = wrapper.Change("ABC");
        var wrapper2 = wrapper1;

        Foo foo1 = wrapper1.Value;
        Foo foo2 = wrapper2.Value;

        Console.WriteLine("foo1 and foo2 are equal? {0}", object.ReferenceEquals(foo1, foo2));
        // Output: foo1 and foo2 are equal? True

        // create wrapper1 based on the wrapper object but changing the key to "BCD"
        var wrapper3 = wrapper.Change("BCD");
        var wrapper4 = wrapper3;

        Foo foo3 = wrapper3.Value;
        Foo foo4 = wrapper4.Value;

        Console.WriteLine("foo3 and foo4 are equal? {0}", object.ReferenceEquals(foo3, foo4));
        // Output: foo3 and foo4 are equal? True

        Console.WriteLine("foo1 and foo3 are equal? {0}", object.ReferenceEquals(foo1, foo3));
        // Output: foo1 and foo3 are equal? False


        // Counter-example: manipulating the dictionary instance that was provided to the wrapper can disrupt expected behavior
        var dictionary = new Dictionary<string, Foo>();

        var wrapper5 = wrapper.Change("CDE", dictionary);
        var wrapper6 = wrapper5;

        Foo foo5 = wrapper5.Value;
        dictionary.Clear();
        Foo foo6 = wrapper6.Value;

        // one might expect this to be true:
        Console.WriteLine("foo5 and foo6 are equal? {0}", object.ReferenceEquals(foo5, foo6));
        // Output: foo5 and foo6 are equal? False
    }
}

Answer 6

Another performance problem comes when you put structs in collections. For example, imagine you have a List<SomeStruct>, and you want to modify the Prop1 property of the first item in the list. The initial inclination is to write this:

List<SomeStruct> MyList = CreateList();
MyList[0].Prop1 = 42;

That's not going to compile. In order to make this work you have to write:

SomeStruct myThing = MyList[0];
myThing.Prop1 = 42;
MyList[0] = myThing.Prop1;

This causes two problems (primarily). First, you end up copying the entire struct twice: once into your working myThing instance, and then back to the list. The second problem is that you can't do this in a foreach because it changes the collection and will cause the enumerator to throw an exception.

By the way, your NeverNull thing has a rather odd behavior. It's possible to set the Reference property to null. It strikes me as very very strange that this statement:

var Contradiction = new NeverNull<object>(null);

Is valid.

I would be interested to know the reasons you have for trying to create this type of struct.