Are C# weak references in fact soft?

Question

The basic difference is that weak references are supposed to be claimed on each run of the GC (keep memory footprint low) while soft references ought to be kept in memory until the GC actually requires memory (they try to expand lifetime but may fail anytime, which is useful for e.g. caches especially of rather expensive objects).

To my knowledge, there is no clear statement as to how weak references influence the lifetime of an object in .NET. If they are true weak refs they should not influence it at all, but that would also render them pretty useless for their, I believe, main purpose of caching (am I wrong there?). On the other hand, if they act like soft refs, their name is a little misleading.

Personally, I imagine them to behave like soft references, but that is just an impression and not founded.

Implementation details apply, of course. I'm asking about the mentality associated with .NET's weak references - are they able to expand lifetime, or do they behave like true weak refs?

(Despite a number of related questions I could not find an answer to this specific issue yet.)

Answer 1

I have seen no information that indicates that they would increase the lifetime of the object they point to. And the articles I read about the algorithm the GC uses to determine reachability do not mention them in this way either. So I expect them to have no influence on the lifetime of the object.

Weak
This handle type is used to track an object, but allow it to be collected. When an object is collected, the contents of the GCHandle are zeroed. Weak references are zeroed before the finalizer runs, so even if the finalizer resurrects the object, the Weak reference is still zeroed.

WeakTrackResurrection
This handle type is similar to Weak, but the handle is not zeroed if the object is resurrected during finalization.

http://msdn.microsoft.com/en-us/library/83y4ak54.aspx

---

There are a few mechanism by which an object that's unreachable can survive a garbage collection.

• The generation of the object is larger than the generation of the GC that happened. This is particularly interesting for large objects, which are allocated on the large-object-heap and are always considered Gen2 for this purpose.
• Objects with a finalizer and all objects reachable from them survive the GC.
• There might be a mechanism where former references from old objects can keep young objects alive, but I'm not sure about that.

Answer 2

Yes
Weak references do not extend the lifespan of an object, thus allowing it to be garbage collected once all strong references have gone out of scope. They can be useful for holding on to large objects that are expensive to initialize but should be avaialble for garabage collection if they are not actively in use.

Answer 3

Are C# weak references in fact soft?

No.

am I wrong there?

You are wrong there. The purpose of weak references is absolutely not caching in the sense that you mean. That is a common misconception.

are they able to expand lifetime, or do they behave like true weak refs?

No, they do not expand lifetime.

Consider the following program (F# code):

do
  let x = System.WeakReference(Array.create 0 0)
  for i=1 to 10000000 do
    ignore(Array.create 0 0)
  if x.IsAlive then "alive" else "dead"
  |> printfn "Weak reference is %s"

This heap allocates an empty array that is immediately eligible for garbage collection. Then we loop 10M times allocating more unreachable arrays. Note that this does not increase memory pressure at all so there is no motivation to collect the array referred to by the weak reference. Yet the program prints "Weak reference is dead" because it was collected nevertheless. This is the behaviour of a weak reference. A soft reference would have been retained until its memory was actually needed.

Here is another test program (F# code):

open System

let isAlive (x: WeakReference) = x.IsAlive

do
  let mutable xs = []
  while true do
    xs <- WeakReference(Array.create 0 0)::List.filter isAlive xs
    printfn "%d" xs.Length

This keeps filtering out dead weak references and prepending a fresh one onto the front of a linked list, printing out the length of the list each time. On my machine, this never exceeds 1,000 surviving weak references. It ramps up and then falls to zero in cycles, presumably because all of the weak references are collected at every gen0 collection. Again, this is the behaviour of a weak reference and not a soft reference.

Note that this behaviour (aggressive collection of weakly referenced objects at gen0 collections) is precisely what makes weak references a bad choice for caches. If you try to use weak references in your cache then you'll find your cache getting flushed a lot for no reason.