Since there seemed to be some dispute over my original answer, I decided to do a few tests, including looking at the generated code and monitoring the performance.
First of all, here's our test bed, a class with a delegate and another class to consume it:
class EventProducer
{
public void Raise()
{
var handler = EventRaised;
if (handler != null)
handler(this, EventArgs.Empty);
}
public event EventHandler EventRaised;
}
class Counter
{
long count = 0;
EventProducer producer = new EventProducer();
public void Count()
{
producer.EventRaised += CountEvent;
producer.Raise();
producer.EventRaised -= CountEvent;
}
public void CountWithNew()
{
producer.EventRaised += new EventHandler(CountEvent);
producer.Raise();
producer.EventRaised -= new EventHandler(CountEvent);
}
private void CountEvent(object sender, EventArgs e)
{
count++;
}
}
First thing to do is look at the generated IL:
.method public hidebysig instance void Count() cil managed
{
.maxstack 8
L_0000: ldarg.0
L_0001: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_0006: ldarg.0
L_0007: ldftn instance void DelegateTest.Program/Counter::CountEvent(object, class [mscorlib]System.EventArgs)
L_000d: newobj instance void [mscorlib]System.EventHandler::.ctor(object, native int)
L_0012: callvirt instance void DelegateTest.Program/EventProducer::add_EventRaised(class [mscorlib]System.EventHandler)
L_0017: ldarg.0
L_0018: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_001d: callvirt instance void DelegateTest.Program/EventProducer::Raise()
L_0022: ldarg.0
L_0023: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_0028: ldarg.0
L_0029: ldftn instance void DelegateTest.Program/Counter::CountEvent(object, class [mscorlib]System.EventArgs)
L_002f: newobj instance void [mscorlib]System.EventHandler::.ctor(object, native int)
L_0034: callvirt instance void DelegateTest.Program/EventProducer::remove_EventRaised(class [mscorlib]System.EventHandler)
L_0039: ret
}
.method public hidebysig instance void CountWithNew() cil managed
{
.maxstack 8
L_0000: ldarg.0
L_0001: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_0006: ldarg.0
L_0007: ldftn instance void DelegateTest.Program/Counter::CountEvent(object, class [mscorlib]System.EventArgs)
L_000d: newobj instance void [mscorlib]System.EventHandler::.ctor(object, native int)
L_0012: callvirt instance void DelegateTest.Program/EventProducer::add_EventRaised(class [mscorlib]System.EventHandler)
L_0017: ldarg.0
L_0018: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_001d: callvirt instance void DelegateTest.Program/EventProducer::Raise()
L_0022: ldarg.0
L_0023: ldfld class DelegateTest.Program/EventProducer DelegateTest.Program/Counter::producer
L_0028: ldarg.0
L_0029: ldftn instance void DelegateTest.Program/Counter::CountEvent(object, class [mscorlib]System.EventArgs)
L_002f: newobj instance void [mscorlib]System.EventHandler::.ctor(object, native int)
L_0034: callvirt instance void DelegateTest.Program/EventProducer::remove_EventRaised(class [mscorlib]System.EventHandler)
L_0039: ret
}
So it turns out that, yes, these do generate identical IL. I was wrong originally. But that's not the whole story. It may be that I'm going off-topic here but I think that it's important to include this when talking about events and delegates:
Creating and comparing different delegates is not cheap.
When I wrote this, I was thinking that the first syntax was able to cast the method group as a delegate, but it turns out that it's just a conversion. But it's completely different when you actually save the delegate. If we add this to the consumer:
class Counter
{
EventHandler savedEvent;
public Counter()
{
savedEvent = CountEvent;
}
public void CountSaved()
{
producer.EventRaised += savedEvent;
producer.Raise();
producer.EventRaised -= savedEvent;
}
}
You can see that this has very different characteristics, performance-wise, from the other two:
static void Main(string[] args)
{
const int TestIterations = 10000000;
TimeSpan countTime = TestCounter(c => c.Count());
Console.WriteLine("Count: {0}", countTime);
TimeSpan countWithNewTime = TestCounter(c => c.CountWithNew());
Console.WriteLine("CountWithNew: {0}", countWithNewTime);
TimeSpan countSavedTime = TestCounter(c => c.CountSaved());
Console.WriteLine("CountSaved: {0}", countSavedTime);
Console.ReadLine();
}
static TimeSpan TestCounter(Action<Counter> action, int iterations)
{
var counter = new Counter();
Stopwatch sw = new Stopwatch();
sw.Start();
for (int i = 0; i < TestIterations; i++)
action(counter);
sw.Stop();
return sw.Elapsed;
}
The results consistently come back as something similar to:
Count: 00:00:02.4742007
CountWithNew: 00:00:02.4272702
CountSaved: 00:00:01.9810367
That's nearly a 20% difference when using a saved delegate vs. creating a new one.
Now obviously not every program is going to be adding and removing this many delegates in such a small amount of time, but if you're writing library classes - classes that might be used in ways you cannot predict - then you really want to keep this difference in mind if you ever need to add and remove events (and I've written a lot of code that does this, personally).
So the conclusion of this is, writing SomeEvent += new EventHandler(NamedMethod) compiles to the same thing as just SomeEvent += NamedMethod. But if you plan to remove that event handler later, you really should save the delegate. Even though the Delegate class has some special-case code that allows you to remove a referentially-different delegate from the one you added, it has to do a non-trivial amount of work to pull this off.
If you're not going to save the delegate, then it makes no difference - the compiler ends up creating a new delegate anyway.
