Limitations of the dynamic type in C#

问题:

Could you give me some reasons for limitations of the dynamic type in C#? I read about them in "Pro C# 2010 and the .NET 4 platform". Here is an excerpt (if quoting books is illegal here, tell me and I will remove the excerpt):

While a great many things can be defined using the dynamic keyword, there are some limitations regarding its usage. While they are not show stoppers, do know that a dynamic data item cannot make use of lambda expressions or C# anonymous methods when calling a method. For example, the following code will always result in errors, even if the target method does indeed take a delegate parameter which takes a string value and returns void.

dynamic a = GetDynamicObject(); 
// Error!  Methods on dynamic data can’t use lambdas! 
a.Method(arg => Console.WriteLine(arg));

To circumvent this restriction, you will need to work with the underlying delegate directly, using the techniques described in Chapter 11 (anonymous methods and lambda expressions, etc). Another limitation is that a dynamic point of data cannot understand any extension methods (see Chapter 12). Unfortunately, this would also include any of the extension methods which come from the LINQ APIs. Therefore, a variable declared with the dynamic keyword has very limited use within LINQ to Objects and other LINQ technologies:

dynamic a = GetDynamicObject(); 
// Error!  Dynamic data can’t find the Select() extension method! 
var data = from d in a select d;

Thanks in advance.

回答1:

Tomas's conjectures are pretty good. His reasoning on extension methods is spot on. Basically, to make extension methods work we need the call site to at runtime somehow know what using directives were in force at compile time. We simply did not have enough time or budget to develop a system whereby this information could be persisted into the call site.

For lambdas, the situation is actually more complex than the simple problem of determining whether the lambda is going to expression tree or delegate. Consider the following:

d.M(123)

where d is an expression of type dynamic. *What object should get passed at runtime as the argument to the call site "M"? Clearly we box 123 and pass that. Then the overload resolution algorithm in the runtime binder looks at the runtime type of d and the compile-time type of the int 123 and works with that.

Now what if it was

d.M(x=>x.Foo())

Now what object should we pass as the argument? We have no way to represent "lambda method of one variable that calls an unknown function called Foo on whatever the type of x turns out to be".

Suppose we wanted to implement this feature: what would we have to implement? First, we'd need a way to represent an unbound lambda. Expression trees are by design only for representing lambdas where all types and methods are known. We'd need to invent a new kind of "untyped" expression tree. And then we'd need to implement all of the rules for lambda binding in the runtime binder.

Consider that last point. Lambdas can contain statements. Implementing this feature requires that the runtime binder contain the entire semantic analyzer for every possible statement in C#.

That was orders of magnitude out of our budget. We'd still be working on C# 4 today if we'd wanted to implement that feature.

Unfortunately this means that LINQ doesn't work very well with dynamic, because LINQ of course uses untyped lambdas all over the place. Hopefully in some hypothetical future version of C# we will have a more fully-featured runtime binder and the ability to do homoiconic representations of unbound lambdas. But I wouldn't hold my breath waiting if I were you.

UPDATE: A comment asks for clarification on the point about the semantic analyzer.

Consider the following overloads:

class C {
  public void M(Func<IDisposable, int> f) { ... }
  public void M(Func<int, int> f) { ... }
  ...
}

and a call

d.M(x=> { using(x) { return 123; } });

Suppose d is of compile time type dynamic and runtime type C. What must the runtime binder do?

The runtime binder must determine at runtime whether the expression x=>{...} is convertible to each of the delegate types in each of the overloads of M.

In order to do that, the runtime binder must be able to determine that the second overload is not applicable. If it were applicable then you could have an int as the argument to a using statement, but the argument to a using statement must be disposable. That means that the runtime binder must know all the rules for the using statement and be able to correctly report whether any possible use of the using statement is legal or illegal.

Clearly that is not restricted to the using statement. The runtime binder must know all the rules for all of C# in order to determine whether a given statement lambda is convertible to a given delegate type.

We did not have time to write a runtime binder that was essentially an entire new C# compiler that generates DLR trees rather than IL. By not allowing lambdas we only have to write a runtime binder that knows how to bind method calls, arithmetic expressions and a few other simple kinds of call sites. Allowing lambdas makes the problem of runtime binding on the order of dozens or hundreds of times more expensive to implement, test and maintain.

回答2:

Lambdas: I think that one reason for not supporting lambdas as parameters to dynamic objects is that the compiler wouldn't know whether to compile the lambda as a delegate or as an expression tree.

When you use a lambda, the compiler decides based on the type of the target parameter or variable. When it is Func<...> (or other delegate) it compiles the lambda as an executable delegate. When the target is Expression<...> it compiles lambda into an expression tree.

Now, when you have a dynamic type, you don't know whether the parameter is delegate or expression, so the compiler cannot decide what to do!

Extension methods: I think that the reason here is that finding extension methods at runtime would be quite difficult (and perhaps also inefficient). First of all, the runtime would need to know what namespaces were referenced using using. Then it would need to search all classes in all loaded assemblies, filter those that are accessible (by namespace) and then search those for extension methods...

回答3:

Eric (and Tomas) says it well, but here is how I think of it.

This C# statement

a.Method(arg => Console.WriteLine(arg)); 

has no meaning without a lot of context. Lambda expressions themselves have no types, rather they are convertible to delegate (or Expression) types. So the only way to gather the meaning is to provide some context which forces the lambda to be converted to a specific delegate type. That context is typically (as in this example) overload resolution; given the type of a, and the available overloads Method on that type (including extension members), we can possibly place some context that gives the lambda meaning.

Without that context to produce the meaning, you end up having to bundle up all kinds of information about the lambda in the hopes of somehow binding the unknowns at runtime. (What IL could you possibly generate?)

In vast contrast, one you put a specific delegate type there,

a.Method(new Action<int>(arg => Console.WriteLine(arg))); 

Kazam! Things just got easy. No matter what code is inside the lambda, we now know exactly what type it has, which means we can compile IL just as we would any method body (we now know, for example, which of the many overloads of Console.WriteLine we're calling). And that code has one specific type (Action<int>), which means it is easy for the runtime binder to see if a has a Method that takes that type of argument.

In C#, a naked lambda is almost meaningless. C# lambdas need static context to give them meaning and rule out ambiguities that arise from many possible coercisons and overloads. A typical program provides this context with ease, but the dynamic case lacks this important context.