The other answers are all good but I want to emphasize a few points here.
The question essentially boils down to: when is the garbage collector allowed to deduce that a given object is dead? and the answer is the garbage collector has broad latitude to use any technique it chooses to determine when an object is dead, and this broad latitude can lead to some surprising results.
So let's start with:
My naive understanding of garbage collection is: "Delete an object as soon as no references point to it."
This understanding is wrong wrong wrong. Suppose we have
class C { C c; public C() { this.c = this; } }
Now every instance of C
has a reference to it stored inside itself. If objects were only reclaimed when the reference count to them was zero then circularly referenced objects would never be cleaned up.
A correct understanding is:
Certain references are "known roots". When a collection happens the known roots are traced. That is, all known roots are alive, and everything that something alive refers to is also alive, transitively. Everything else is dead, and eligable for reclamation.
Dead objects that require finalization are not collected. Rather, they are kept alive on the finalization queue, which is a known root, until their finalizers run, after which they are marked as no longer requiring finalization. A future collection will identify them as dead a second time and they will be reclaimed.
Lots of things are known roots. Static fields, for example, are all known roots. Local variables might be known roots, but as we'll see below, they can be optimized away in surprising ways. Temporary values might be known roots.
I'm creating an instance of a class without assigning it to a variable.
Your question here is a good one but it is based on an incorrect assumption, namely that a local variable is always a known root. Assigning a reference to a local variable does not necessarily keep an object alive. The garbage collector is allowed to optimize away local variables at its whim.
Let's give an example:
void M()
{
var resource = OpenAFile();
int handle = resource.GetHandle();
UnmanagedCode.MessWithFile(handle);
}
Suppose resource
is an instance of a class that has a finalizer, and the finalizer closes the file. Can the finalizer run before MessWithFile
? Yes! The fact that resource
is a local variable with a lifetime of the entire body of M
is irrelevant. The runtime can realize that this code could be optimized into:
void M()
{
int handle;
{
var resource = OpenAFile();
handle = resource.GetHandle();
}
UnmanagedCode.MessWithFile(handle);
}
and now resource
is dead by the time MessWithFile
is called. It is unlikely but legal for the finalizer to run between GetHandle
and MessWithFile
, and now we're messing with a file that has been closed.
The correct solution here is to use GC.KeepAlive
on the resource after the call to MessWithFile
.
To return to your question, your concern is basically "is the temporary location of a reference a known root?" and the answer is usually yes, with the caveat that again, if the runtime can determine that a reference is never dereferenced then it is allowed to tell the GC that the referenced object might be dead.
Put another way: you asked if
new MyClass().DoSomething();
and
var c = new MyClass();
c.DoSomething();
are the same from the point of view of the GC. Yes. In both cases the GC is allowed to kill the object the moment that it determines it can do so safely, regardless of the lifetime of local variable c
.
The shorter answer to your question is: trust the garbage collector. It has been carefully written to do the right thing. The only times you need to worry about the GC doing the wrong thing are scenarios like the one I laid out, where timing of finalizers is important for the correctness of unmanaged code calls.