language features - Why C# doesn't implement indexed properties?

Question

I know, I know... Eric Lippert's answer to this kind of question is usually something like "because it wasn't worth the cost of designing, implementing, testing and documenting it".

But still, I'd like a better explanation... I was reading this blog post about new C# 4 features, and in the section about COM Interop, the following part caught my attention :

By the way, this code uses one more new feature: indexed properties (take a closer look at those square brackets after Range.) But this feature is available only for COM interop; you cannot create your own indexed properties in C# 4.0.

OK, but why ? I already knew and regretted that it wasn't possible to create indexed properties in C#, but this sentence made me think again about it. I can see several good reasons to implement it :

• the CLR supports it (for instance, PropertyInfo.GetValue has an index parameter), so it's a pity we can't take advantage of it in C#
• it is supported for COM interop, as shown in the article (using dynamic dispatch)
• it is implemented in VB.NET
• it is already possible to create indexers, i.e. to apply an index to the object itself, so it would probably be no big deal to extend the idea to properties, keeping the same syntax and just replacing this with a property name

It would allow to write that kind of things :

public class Foo
{
    private string[] _values = new string[3];
    public string Values[int index]
    {
        get { return _values[index]; }
        set { _values[index] = value; }
    }
}

Currently the only workaround that I know is to create an inner class (ValuesCollection for instance) that implements an indexer, and change the Values property so that it returns an instance of that inner class.

This is very easy to do, but annoying... So perhaps the compiler could do it for us ! An option would be to generate an inner class that implements the indexer, and expose it through a public generic interface :

// interface defined in the namespace System
public interface IIndexer<TIndex, TValue>
{
    TValue this[TIndex index]  { get; set; }
}

public class Foo
{
    private string[] _values = new string[3];

    private class <>c__DisplayClass1 : IIndexer<int, string>
    {
        private Foo _foo;
        public <>c__DisplayClass1(Foo foo)
        {
            _foo = foo;
        }

        public string this[int index]
        {
            get { return _foo._values[index]; }
            set { _foo._values[index] = value; }
        }
    }

    private IIndexer<int, string> <>f__valuesIndexer;
    public IIndexer<int, string> Values
    {
        get
        {
            if (<>f__valuesIndexer == null)
                <>f__valuesIndexer = new <>c__DisplayClass1(this);
            return <>f__valuesIndexer;
        }
    }
}

But of course, in that case the property would actually return a IIndexer<int, string>, and wouldn't really be an indexed property... It would be better to generate a real CLR indexed property.

What do you think ? Would you like to see this feature in C# ? If not, why ?

Answer 1

Here's how we designed C# 4.

First we made a list of every possible feature we could think of adding to the language.

Then we bucketed the features into "this is bad, we must never do it", "this is awesome, we have to do it", and "this is good but let's not do it this time".

Then we looked at how much budget we had to design, implement, test, document, ship and maintain the "gotta have" features and discovered that we were 100% over budget.

So we moved a bunch of stuff from the "gotta have" bucket to the "nice to have" bucket.

Indexed properties were never anywhere near the top of the "gotta have" list. They are very low on the "nice" list and flirting with the "bad idea" list.

Every minute we spend designing, implementing, testing, documenting or maintaining nice feature X is a minute we can't spend on awesome features A, B, C, D, E, F and G. We have to ruthlessly prioritize so that we only do the best possible features. Indexed properties would be nice, but nice isn't anywhere even close to good enough to actually get implemented.